relax adt unsizing requirements

[lcnr]

Changes unsizing of structs in case the last struct field shares generic params with other adt fields which do not change.
This change is currently insta stable and changes the language, so it at least requires a lang fcp. I feel like the current state is fairly unintuitive.

An example for what's now allowed would be https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=6dd331d23f5c9ffc8c978175aae2e967

struct A<T, U: ?Sized>(T, B<T, U>); // previously ERR
// struct A<T, U: ?Sized>(T, B<[u32; 1], U>); // ok
struct B<T, U: ?Sized>(T, U);

fn main() {
    let x = A([0; 1], B([0; 1], [0; 1]));
    let y: &A<[u32; 1], [u32]> = &x;
    assert_eq!(y.1.1.len(), 1);
}

[rust-highfive]

r? @estebank

(rust-highfive has picked a reviewer for you, use r? to override)

[joshtriplett]

We discussed this in the @rust-lang/lang meeting today, but didn't come to any conclusions.

We felt that it would help to have a little more information about what exactly this is changing, beyond an example. The example helped, but some background on what exactly was allowed before and what is allowed now (rather than just the delta from the former to the latter) would make us much more confident making this change.

[lcnr]

wasn't able to get to summarize the current status here, removing nomination and renominating after I finish the summary

[lcnr]

The only ADT we implement Unsize for are structs. All other coercion remain the same so I am going to focus on that here.

Current implementation

A generic struct implements Unsize if and only if:

• the tail field depends on at least one type or const parameter
• all type and const parameters used in the last field must only be used in the last field
• the target struct can be created from the source by replacing only the parameters found in the last struct field
• the tail field implements Unsize from source to target

This PR

A generic struct implements Unsize if and only if:

• the tail field depends on at least one type or const parameter not used in any other field
• the target struct can be created from the source by replacing only the parameters only found in the last struct field
• the tail field implements Unsize from source to target

[pnkfelix]

Thanks for the write-up, @lcnr . The T-lang design team discussed this in our meeting today.

We're generally in favor of this change, but we wanted to push back on it being insta-stable. From skimming the PR, it seemed to us like there might be reasonable strategies for feature-gating this change. (Either by running old and new copies of the algorithm, or potentially switching from a boolean flag to a tri-state one when determining whether Unsize can be applied.)

Can you look into putting a feature gate in here?

[lcnr]

added a feature gate, note that this means that activating feature(relaxed_struct_unsize) changes which trait impls exist which afaik is generally unsound to do.

I think this should mostly be a theoretical concern as Unsize isn't really used like an ordinary trait so that unsoundness would have to be intentionally caused. Or maybe we don't even care as one has to explicitly enable a feature idk 🤷

[oli-obk]

@bors r+

seems fine to me.

[bors]

📌 Commit 031cce8 has been approved by oli-obk