Nominating for discussion at lang team meeting.

1. Does this deserve an formal major change proposal, or is this a small enough tweak that we could just amend the RFC and the reference as well as revise the implementation?
2. Is this worth doing at all?
   • I was about to claim that it seems like a minor issue, but then I realized of course that we accept #[repr(transparent)] struct R([usize; 1]);, so one can imagine getting annoyed by this corner case.

@rustbot label: +I-lang-nominated

That is, we could loosen the repr(transparent) rules to say that it "requires at most one field that occupies memory"; this is a loosening of our current rules on two levels:

I think this formulation makes it not at all clear that alignment matters.

I would just say, it can have at most one field that is not a 1-ZST.

(Also Cc #96921)

we accept #[repr(transparent)] struct R([usize; 1]);, so one can imagine getting annoyed by this corner case.

Yeah, and it is possible to construct repr(transparent) where all fields are ZST via generics. So if we have logic bugs in the compiler because of this (which I kind of expect we do, a bunch of code will be looking for "the first non-ZST field" and I am not sure if it always has code to handle the situation where no field is found), they are already exposed, just harder to hit.

I would just say, it can have at most one field that is not a 1-ZST.

I think this rule makes the most sense to me.

After all, as you said, one can do

#[repr(transparent)]
struct Rtemp<const N: usize>([usize; N]);
type R = Rtemp<0>;

which I think is fine.

Or, more generally, I think that repr(transparent) on a one-field struct should always work, since it does on single field of generic type.

Discussed in lang-team meeting to general happiness and accord. Based on current attendance (not everyone), we would welcome a PR implementing this change and we could FCP that PR to confirm.

An alternate phrasing of the rule is:

• If you ignore any 1-ZST fields, you can have at most one field in your struct remaining.

@tmandry proposes:

"You can have at most one field in your struct, ignoring any 1-ZST fields."

Anything that doesn't include a double negative works for me :P

I was going to add an induction definition (one-field is fine; adding 1-ZST fields is fine), but it turns out that this is legal:

#[repr(transparent)]
struct Foo {}

Which means that the current rule isn't actually

Today we require exactly one non-ZST field in a repr(transparent) struct

And thus I'm even more in favour of the "at most one" version of the rule, since that seems to be what we're already doing in other situations.

removing nomination; we discussed this last week.

@rustbot label: -I-lang-nominated