Confusing error reported for trait impls on DST with the last field being Sized

[BoxyUwU]

Given the following code: https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=d9a00dc0e329f0ea73b037f3af9f26db

use std::marker::PhantomData;

trait Trait {}

struct Unsized([u8], ());
struct Foo<T: ?Sized>(PhantomData<T>);

impl<T> Trait for Foo<T> {}
impl Trait for Foo<Unsized> {}

The current output is:

error[E0119]: conflicting implementations of trait `Trait` for type `Foo<Unsized>`:
 --> src/main.rs:9:1
  |
8 | impl<T> Trait for Foo<T> {}
  | ------------------------ first implementation here
9 | impl Trait for Foo<Unsized> {}
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^ conflicting implementation for `Foo<Unsized>`

Ideally the output should look like:

error[E0277]: the size for values of type `[u8]` cannot be known at compilation time
 --> src/lib.rs:1:16
  |
5 | struct Unsized([u8], ());
  |                ^^^^ doesn't have a size known at compile-time
  |
  = help: the trait `Sized` is not implemented for `[u8]`
  = note: only the last field of a struct may have a dynamically sized type
  = help: change the field's type to have a statically known size
help: borrowed types always have a statically known size
  |
5 | struct Unsized(&[u8], ());
  |                ^
help: the `Box` type always has a statically known size and allocates its contents in the heap
  |
5 | struct Unsized(Box<[u8]>, ());
  |                ^^^^    ^

error: aborting due to previous error

Note: this error is particularly confusing because the trait impls *dont* overlap after fixing the field ordering error

cc @estebank

[eddyb]

FWIW I don't think we can avoid the trait impl overlap false positive error, without introducing cyclic dependencies.

cc @nikomatsakis but my understanding is that to prove Unsized: Sized, only the last field of Unsized has to be considered, and this may be used sometimes before Unsized's definition was checked (as in this case, during coherence-checking).

However, the fact that we get "conflicting implementations" errors and no other errors suggests to me that we stop compilation early, which is something that I know has been slowly rectified.

So my opinion is that for now we likely can't do better than "keep going and emit both errors", which should be simple enough to try - tho I am a bit worried its effects on tests' expected output (more errors/warnings from later on), and it's possible the trait system might get a bit wonky if there are overlapping impls - do they all apply and cause ambiguity errors down the line?

[kadiwa4]

Fixed since 1.78.0:

error[E0277]: the size for values of type `[u8]` cannot be known at compilation time
 --> src/lib.rs:5:16
  |
5 | struct Unsized([u8], ());
  |                ^^^^ doesn't have a size known at compile-time
  |
  = help: the trait `Sized` is not implemented for `[u8]`
  = note: only the last field of a struct may have a dynamically sized type
  = help: change the field's type to have a statically known size
help: borrowed types always have a statically known size
  |
5 | struct Unsized(&[u8], ());
  |                +
help: the `Box` type always has a statically known size and allocates its contents in the heap
  |
5 | struct Unsized(Box<[u8]>, ());
  |                ++++    +

error[E0119]: conflicting implementations of trait `Trait` for type `Foo<Unsized>`
 --> src/lib.rs:9:1
  |
8 | impl<T> Trait for Foo<T> {}
  | ------------------------ first implementation here
9 | impl Trait for Foo<Unsized> {}
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^ conflicting implementation for `Foo<Unsized>`

Some errors have detailed explanations: E0119, E0277.
For more information about an error, try `rustc --explain E0119`.

The "conflicting implementations" error still appears, but I thinks that's fine, especially because rustc suggests that you should turn the [u8] field into a sized one (&[u8] or Box<[u8]>). So at least the diagnostics are consistent with each other.

@rustbot label: +E-needs-test

[zakrad]

@rustbot claim