LGAT associated type bounds diagnostics are not great

[CAD97]

Tested on nightly 2022-07-21.

Illustrating as a edit-compile cycle:

#![feature(generic_associated_types)]

use core::future::Future;

trait Task {
    type Output<'a>: 'a + Future<Output = ()>
    where
        Self: 'a;
    fn call(&mut self) -> Self::Output<'_>;
}

type DynTask = dyn Task<Output = Box<dyn Future<Output = ()>>>;

error[E0107]: missing generics for associated type `Task::Output`
  --> src/lib.rs:12:25
   |
12 | type DynTask = dyn Task<Output = Box<dyn Future<Output = ()>>>;
   |                         ^^^^^^ expected 1 lifetime argument
   |
note: associated type defined here, with 1 lifetime parameter: `'a`
  --> src/lib.rs:6:10
   |
6  |     type Output<'a>: 'a + Future<Output = ()>
   |          ^^^^^^ --
help: add missing lifetime argument
   |
12 | type DynTask = dyn Task<Output<'a> = Box<dyn Future<Output = ()>>>;
   |                         ~~~~~~~~~~

Apply the suggestion.

type DynTask = dyn Task<Output<'a> = Box<dyn Future<Output = ()>>>;

error[E0261]: use of undeclared lifetime name `'a`
  --> src/lib.rs:12:32
   |
12 | type DynTask = dyn Task<Output<'a> = Box<dyn Future<Output = ()>>>;
   |                                ^^ undeclared lifetime
   |
   = note: for more information on higher-ranked polymorphism, visit https://doc.rust-lang.org/nomicon/hrtb.html
help: consider making the bound lifetime-generic with a new `'a` lifetime
   |
12 | type DynTask = dyn for<'a> Task<Output<'a> = Box<dyn Future<Output = ()>>>;
   |                    +++++++
help: consider introducing lifetime `'a` here
   |
12 | type DynTask<'a> = dyn Task<Output<'a> = Box<dyn Future<Output = ()>>>;
   |             ++++

That looks wrong, let me try:

type DynTask = dyn Task<for<'a> Output<'a> = Box<dyn Future<Output = ()>>>;

error: `for<...>` is not allowed on associated type bounds
  --> src/lib.rs:12:25
   |
12 | type DynTask = dyn Task<for<'a> Output<'a> = Box<dyn Future<Output = ()>>>;
   |                         ^^^^^^^^^^^^^^^^^^

Okay, let's do what the compiler suggested

type DynTask = dyn for<'a> Task<Output<'a> = Box<dyn Future<Output = ()>>>;

warning: type alias `DynTask` is never used
  --> src/lib.rs:12:6
   |
12 | type DynTask = dyn for<'a> Task<Output<'a> = Box<dyn Future<Output = ()>>>;
   |      ^^^^^^^
   |
   = note: `#[warn(dead_code)]` on by default

warning: `playground` (lib) generated 1 warning

Success! Time to use it...

error[E0038]: the trait `Task` cannot be made into an object
  --> src/lib.rs:14:18
   |
14 | fn takes(_: &mut DynTask) {}
   |                  ^^^^^^^ `Task` cannot be made into an object
   |
note: for a trait to be "object safe" it needs to allow building a vtable to allow the call to be resolvable dynamically; for more information visit <https://doc.rust-lang.org/reference/items/traits.html#object-safety>
  --> src/lib.rs:6:10
   |
5  | trait Task {
   |       ---- this trait cannot be made into an object...
6  |     type Output<'a>: 'a + Future<Output = ()>
   |          ^^^^^^ ...because it contains the generic associated type `Output`
   = help: consider moving `Output` to another trait

🙃

[CAD97]

@rustbot modify labels +requires-nightly +F-generic_associated_types

[compiler-errors]

Check out #![feature(generic_associated_types_extended)] -- GATs are not allowed in object types currently.

[compiler-errors]

Though fair warning that #![feature(generic_associated_types_extended)] is unsound.

[compiler-errors]

I guess there are a few actionable diagnostics improvements here:

1. The "add missing lifetime argument" should try to look for a binder or set of generics to add for<'a> to if the suggestion is being made.
2. The "for<...> is not allowed on associated type bounds" suggestion could probably suggest moving the binder outside of the dyn.

Probably more.

[jackh726]

That looks wrong, let me try:
type DynTask = dyn Task<for<'a> Output<'a> = Box<dyn Future<Output = ()>>>;

I'm confused. Was this done this was because the suggestions don't make sense? Either one of the suggestions would have worked.

help: add missing lifetime argument
|
12 | type DynTask = dyn Task<Output<'a> = Box<dyn Future<Output = ()>>>;

This certainly isn't great. We should be better (not just with GATs) about making sure we suggest lifetimes in scope or suggest creating a new one. (We obviously are smart enough because of the next suggestion, just need to "apply" it earlier.)

the trait Task cannot be made into an object

🤷‍♀️ not much that can be done for this atm. I wouldn't even touch generic_associated_types_extended with a 10 foot pole for anything but playing around. It's not just unsound, but the changes it makes aren't well thought out yet.

[CAD97]

That looks wrong, let me try:
type DynTask = dyn Task<for<'a> Output<'a> = Box<dyn Future<Output = ()>>>;

I'm confused. Was this done this was because the suggestions don't make sense?

It was done out of the (misguided?) assumption¹ that this would be the correct place to put the binder; for<'a> dyn Task<Output<'a> = ...> feels² like the wrong place, since the associated type is the generic introducer, not the trait. A suggestion on that form to move the binder out would be sufficient.

(This was the form I originally wrote, and only backed up to not providing it after hitting that error.)

Footnotes

1. Partially because the feature is not stable, I'm more willing to apply my own logic to overrule what the compiler is telling me. ↩

2. Unfortunately I'm not sure I can qualify this feeling much further. ↩