Regression for #![feature(const_in_array_repeat_expressions)]

[josephlr]

The following code used to compile on nightly, but now does not:

#![feature(const_in_array_repeat_expressions)]

struct T;

impl T {
    const fn new() -> Self {
        Self
    }
}

const fn bar() -> [T; 2] {
    [T::new(); 2]
}

Note that

static FOO: [T; 2] = [T::new(); 2];

still compiles. Adding #![feature(const_fn)] doesn't affect this issue.

I think this regression was introduced in either #76807 or #76850

CC: @ecstatic-morse @oli-obk

[josephlr]

Actually @RalfJung this might be caused by #76411, as this seems to be a promotion issue.

[RalfJung]

No, this is caused by #75502. We used to incorrectly apply const-promotion rules inside const fn, but since const fn are also fn, that has all the same problems as applying those rules in fn. We still allow this in static and const as those are never subject to codegen, which makes promotion much less risky. (Less risky, but still problematic -- this particular feature will require some annoying quirks all over the compiler if we cannot take it back.)

So, this was never intended to work, and worked by accident. To make this work we'd have to change array initializers to explicit promotion rules, which IIRC we already have an issue for?

See https://github.com/rust-lang/const-eval/blob/master/promotion.md for some background.

[RalfJung]

Hm, looks like we do not have an issue open for this, but the tracking issue probably makes sense as a place for this discussion: #49147.

[josephlr]

Hm, looks like we do not have an issue open for this, but the tracking issue probably makes sense as a place for this discussion: #49147.

Makes sense, we can continue discussion in the tracking issue.