False positive dead_code for const used as const generic argument during impl marker trait

[datdenkikniet]

In the following snippet, CONST gets a dead_code warning. However, CONST is most definitely used and the trait is also pub, so this warning seems incorrect.

Removing CONST causes the crate to no longer compile, and removing the impl would be a breaking change.

Adding any function (i.e. fn method();) to the trait and implementing it removes the warning as well. However, for marker traits this is not an option. It also clearly demonstrates that CONST is considered to be used despite being unaffected by fn method().

Rust version: rustc 1.80.0 (0514789 2024-07-21)

This code compiles without warnings on 1.79.0 (rustc 1.79.0 (129f3b9 2024-06-10))

pub struct Value<const NUMBER: u8> {}

pub trait Trait {}

const CONST: u8 = 11;

impl Trait for Value<CONST> {}

Warning:

warning: constant `CONST` is never used
 --> lib.rs:5:7
  |
5 | const CONST: u8 = 11;
  |       ^^^^^
  |
  = note: `#[warn(dead_code)]` on by default

Adding method() removes the warning:

pub struct Value<const NUMBER: u8> {}

pub trait Trait {
    fn method();
}

const CONST: u8 = 11;

impl Trait for Value<CONST> {
    fn method() {}
}

[datdenkikniet]

Potentially fixed by #128404. Have no opportunity to test RN, but saw this come by and it seemed potentially relevant.

[compiler-errors]

@datdenkikniet: This is indeed fixed by that PR.

[compiler-errors]

cc @mu001999: This is another regression that should be fixed first if the dead code changes are re-landed.

[datdenkikniet]

This is fixed in the new point release, 1.80.1. Should we close this issue, or keep it open for tracking?

[jieyouxu]

This is probably worth keeping open for tracking purposes, and probably should be closed by the re-land PR(s) that add this case as a regression(?) test.