#[repr(X)] should inhibit enum layout optimization

[bluss]

This is a regression stable to nightly (breaking nodrop and arrayvec at the moment; not their compilation but their tests, which indicate their implementation is not sound for element types like &T or Vec<T>).

Found by cargobomb report

Note that the regression applies to enums with #[repr(u8)] (and similar) but not #[repr(C)].

PR #25651 was never merged, but it would document this aspect of #[repr(u8)]

Evaluating #39595 for its impact. Thanks @eddyb for the pointers!

[bluss]

As usual filing an issue for every regression; it doesn't mean that the change in behaviour must be fixed. In this case it probably should, though.

[bluss]

First nightly with change is rustc 1.17.0-nightly (24a70eb59 2017-02-09) (rustup's nightly-2017-02-10)

[bluss]

All crates affected by the nodrop breakage are fixed by just released nodrop 0.1.9.

It changed from this:

    #[repr(u8)]
    enum Flag<T> {
        Alive(T),
        Dropped,
    }

to this:

    #[repr(u8)]
    enum Flag<T> {
        Alive(T),
        Dropped(u8),
    }

I didn't find a reason to not use this workaround.

[nikomatsakis]

I believe we should restore the old behavior. I feel like if you request a representation with a u8 discriminant, we should give you a u8 discriminant, not no discriminant at all. =) More specifically, I've always thought of the u8 repr as being struct { u8, X } where X is the union of all arguments. (Same for u16 etc.)

I could see other choices, but given that this is what we've done till now (right?) I'd say we ought to stick with it. If you want the option for us to apply further optimizations (use "u8" if you need a discriminant, but otherwise who cares...) we should have people opt in to that.

[brson]

I agree this is a regression that should be fixed.

[nikomatsakis]

on it.

[nikomatsakis]

See #40188