Item used in (unused) item named _name is still deemed dead

[DoubleHyphen]

Code

const _MY_CODE_IS_CORRECT: bool = true;

const fn sanity_check() {
    let true = _MY_CODE_IS_CORRECT else {
        panic!();
    };
}

const _SANITY_CHECK: () = sanity_check();

Current output

warning: function `sanity_check` is never used
 --> src/lib.rs:3:10
  |
3 | const fn sanity_check() {
  |          ^^^^^^^^^^^^
  |
  = note: `#[warn(dead_code)]` on by default

Desired output

Compiling with no warnings.

Rationale and extra context

I have created some modules. I want each of them to fail to compile if it is not correct. To that end, I instituted a sanity_check in the crate, which is called within each module as const _SANITY_CHECK: () = crate::sanity_check();

The problem is that, because the _SANITY_CHECK consts are obviously not used anywhere, the entire sanity_check function is considered unused too. As one can tell, it is in fact very much used, and serves a very important purpose.

I guess this could be worked around by renaming it to _sanity_check, but in this case leaving the function actually unused would return no warnings, which is less than ideal.

Other cases

Rust Version

I tried it in the play-ground. I tried to find the version there but failed—that said, it's always the most recent one, isn't it?

Anything else?

No response

[hanna-kruppe]

I would expect const _: () = …; to remove the warning but it doesn’t. Curiously, pub const _ doesn’t get the warning even though there’s no name to make public so the pub shouldn’t make a difference. That definitely feels like a bug to me.

[fmease]

To recap this reduces to

const fn f() {} //~ WARN function `f` is never used
const _X: () = f();

However, we apparently don't do this for any other cases either (namely, propagating liveness via _ident):

fn f() {} //~ WARN function `f` is never used
fn _g() { f() }

[fmease]

Also CC #118424

[hanna-kruppe]

I think it’s a consistent, defensible choice to treat an underscore prefix as only suppressing the dead code lint for that one item, without counting as a “root” for the purpose of computing which things are reachable from exports/entry points. It could be done differently but there’s arguments for it as well.

However, const _ not counting as “root” seems wrong because that syntax exists only for the side effects of type checking and evaluating the constant expression. It’s the best solution for “I only want a compile time assertion, not define a value” (at the module level; within functions a const block is more flexible) and was specifically added after const _FOO already existed. So it should work for @DoubleHyphen’s use case and for #118424 — even and especially if we stick with the current behavior for const _FOO for consistency with other items.

[DoubleHyphen]

Curiously, pub const _ doesn’t get the warning

…it does when it's inside a module, though!

const fn foo() {} // warning: function `foo` is never used

mod whatever {
    pub const _: () = crate::foo();
}