const_err does not trigger on associated constants

[Nemo157]

I tried this code:

pub struct Foo;

pub const FOO: Foo = Foo::new(1);

impl Foo {
    pub const FOO: Self = Self::new(1);

    pub const fn new(i: usize) -> Foo {
        [][i] = 5;
        Foo
    }
}

I expected to see this happen: Both FOO and Foo::FOO should trigger const_err errors.

Instead, this happened: Only FOO was marked with const_err

Meta

Checked on current stable (1.42) and nightly (2020-04-17) on the playground

[Nemo157]

This was discussed at #67191 (comment) and below, so might be covered by #70820 but I didn't really look into the details.

[RalfJung]

Cc @rust-lang/wg-const-eval

I would say this is expected. The error comes from actually computing the const (not some kind of static analysis). Basically rustc iterates over every const in the crate and computes it, to make sure the errors are triggered even if the const is not used. Associated consts cannot in general be evaluated eagerly as they could depend on generic parameters, unlike global consts.

You will still get the error if you actually use Foo::FOO.

I suppose we could try to still eagerly evaluate monomorphic associated consts like this one, but that scheme will also quickly reach its limit, and it feels rather ad-hoc. Currently, from what I recall, const-prop and the unused-const-err lint carefully balance who is running on what to avoid duplicate warnings; any such change would need careful testing to make sure we do not regress that.