Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @petrochenkov (or someone else) soon.

If any changes to this PR are deemed necessary, please add them as extra commits. This ensures that the reviewer can see what has changed since they last reviewed the code. Due to the way GitHub handles out-of-date commits, this should also make it reasonably obvious what issues have or haven't been addressed. Large or tricky changes may require several passes of review and changes.

Please see the contribution instructions for more information.

r? @eddyb

cc @arielb1 Should we do this or should we just never consider unions uninhabited?

What should ManuallyDrop<!> do? This reminds me of the "does it need a () field" discussion from back in #40559 (comment)...

(MaybeUninit<T> has the () field already in rust-lang/rfcs#1892)

I would expect ManuallyDrop<!> to be uninhabited both just intuitively and because if I'm using ManuallyDrop, I'm probably doing it for optimization purposes, so taking an optimization away seems problematic. If unions with no inhabited fields are inhabited, then there are certain (admittedly fairly minor, probably) optimization that just can't be done. I don't think a () field in ManuallyDrop is necessary because drop_in_place already leaves its target invalid and I would expect that to be true of the method on ManuallyDrop as well.

That's not to say that there isn't a use for a union

union MaybeInitialized<T> {
    init: T,
    uninit: ()
}

That should be used whenever cleaned up data sticks around, that is when "uninitialized" is a valid state to be in.

All that said, I think arrayvec is a good example of why this might not be desirable. Currently, an ArrayVec<[T; 20]> is represented as (ManuallyDrop<[T; 20]>, usize). This is totally invalid if ManuallyDrop<!> is uninhabited. I personally believe it is invalid anyway, since the type does not at all encode the fact that the [T; 20] might be partially initialized. Even if ManuallyDrop had a () variant, I would expect any union sanitizer to object to this code - the union is in the initialized state, but the inner value is invalid. However, at the very least, it shows that people today are writing code that assumes that values stored in unions can be invalid.

The safe way to write ArrayVec, under the stricter reasoning I'm arguing for, would be to use an associated type parameter so that <[T; 20]>::PartiallyInitialized == [MaybeInitialized<T>; 20] and store the PartiallyInitialized version of an array.

@gereeter Part of the problem is that people have conflated MaybeInitialized and ManuallyDrop.

@eddyb Oh, absolutely. I guess my point is that people have gone further and collapsed some PartiallyInitialized and ManuallyDrop, so I'm not sure that declaring MaybeInitialized to be the same thing as ManuallyDrop will actually help. Therefore, I'd rather just declare that code incorrect.

@gereeter I still prefer not having unions able to be uninhabited but maybe there's a point to it?

Hmm, what's an empty union? In the "a union is an enum without an explicit discriminant" model it's uninhabited; in the "a union is a struct where all the fields overlap" model it's a ZST.

Though in both of those models, a union with one uninhabited field is uninhabited...

Edit: To answer my own question, 0-field unions are an error.

I admit that the biggest motivation for me is that it feels more consistent. We call enums tagged unions, but an enum with all uninhabited variants is uninhabited. Therefore, a tag with a union with all uninhabited variants should be uninhabited. We have safe functions like ManuallyDrop::into_inner. Therefore, if ManuallyDrop<T> is inhabited, we can get a T, so T is inhabited. Also, as I said above, I shouldn't lose an optimization just because of an implementation detail, and uninhabitedness is a useful optimization.

@scottmcm Is the following a valid implementation of transmute?

union Transmuter<T, U> {
    start: T,
    end: U
}

unsafe fn transmute<T, U>(value: T) -> U {
    Transmuter { start: value }.end
}

In the "overlapping struct" model, I'd say definitely yes. In the "tagless enum" model, I'm not sure.

just because of an implementation detail

This isn't really about implementation details, which don't really matter here either way.
The real question is the model that we actually want to expose to the user.

@eddyb Sorry, I meant an implementation detail of a data structure library. If I have a data structure like NonEmptyList<T> and I want it to be uninhabited if T is uninhabited, then I am forbidden from using T only inside of a union (if unions are always inhabited). Unless PhantomData induces uninhabitedness, but that seems very wrong to me. Therefore, the optimization I want to give to the users of my library forces my hand for implementation.

@gereeter What does a sound usecase look like? If you can optimize it out, how do you use union?

Uninhabitedness is the first thing I've seen where the "a union is an enum" model is clearly better than the"a union is a struct" model, since a struct is uninhabited if any field is uninhabited (aka the issue at hand) whereas an enum is uninhabited only if all variants are uninhabited.

@gereeter That's a question for the smantics-of-unsafe-code folks, I think -- it seems open as of https://github.com/petrochenkov/rfcs/blob/8f1a960844b6ae37ec19fd89d39355fa98b1bb2b/text/1444-union.md#delayed-and-unresolved-questions

I think that we want unions (or, at least, ManuallyDrop) to always be inhabited - unions are used in unsafe code, and introducing basically unneeded and randomly-firing footguns into unsafe code is a stupid thing.

@eddyb I'm thinking of some sort of wrapper like

struct PermutedDrop<A: Array> {
    data: [ManuallyDrop<A::Item>; A::LENGTH],
    order: [usize; A::LENGTH]
}

impl<A: Array> Drop for PermutedDrop<A> {
    fn drop(&mut self) {
        for idx in self.order {
            unsafe {
                ManuallyDrop::drop(&mut self.data[idx]);
            }
        }
    }
}

that rearranges the drop order of the things it contains. If T is uninhabited, PermutedDrop<[T; 20]> should also be uninhabited. If unions are always uninhabited, I believe that PermutedDrop cannot be written to work and have that property. I admit, however, that this is a very niche use case where the optimization is unlikely to matter - why would you put uninhabited things in PermutedDrop anyway?

@arielb1 I'd posit that almost all code that assumes that unions cannot be uninhabited is broken in some other way. In that case, this being a footgun is not a problem - it just makes it more likely that the unsoundness will be noticed. That said, I may be wrong about this, and I think the situation is a bit different with ManuallyDrop, so it seems reasonable to at least make ManuallyDrop always be inhabited.

I don't even think we treat [!; N] as uninhabited, just ZST.

@eddyb Oh. Well, that also seems wrong to me (for N greater than 0). However, it does largely invalidate my argument.

Since it seems to be the most accepted opinion, doesn't break code, and isn't setting a precedent that hasn't already been set, I switched this PR to always making unions inhabited.

@bors r+

📌 Commit da97917 has been approved by eddyb

⌛ Testing commit da97917 with merge 34c65e22a3c9b0610bfd98e9cf2244df4656ccc3...

💔 Test failed - status-travis

@bors retry

[01:49:45] test net::tcp::tests::clone_accept_smoke has been running for over 60 seconds

No output has been received in the last 30m0s, this potentially indicates a stalled build or something wrong with the build itself.
Check the details on how to adjust your build configuration on: https://docs.travis-ci.com/user/common-build-problems/#Build-times-out-because-no-output-was-received

The build has been terminated

(check x86_64-apple-darwin)

⌛ Testing commit da97917 with merge 4ce6b9a...

☀️ Test successful - status-appveyor, status-travis
Approved by: eddyb
Pushing 4ce6b9a to master...

Looks like we forgot to backport this to 1.23.0 (sorry about that!) so removing the beta tags

@alexcrichton #45225 was backed out from that beta (IIUC) so there wasn't anything to backport.

Oh yay!