Fix GAT lifetime bounds

[PlasmaPower]

This makes GAT lifetimes in traits (not impls) late-bound regions, which matches the RFC, stating that bounds placed on the GAT should be interpreted as HRTBs.

This fixes most of #62521, but the last code block in that issue fails because GAT impls are unimplemented (pun not intended).

[rust-highfive]

r? @eddyb

(rust_highfive has picked a reviewer for you, use r? to override)

[PlasmaPower]

I was flying blind here, and just tried to implement what the previous comment said, with a bit of trial and error. If someone who knows this code could review this that'd be appreciated.

[eddyb]

r? @nikomatsakis cc @rust-lang/wg-traits

[nikomatsakis]

Hmm. I have to think about this. I do not believe these should be late-bound, but rather early-bound -- we basically want to substitute them. In general, it'd be good @PlasmaPower if you wanted to pop into the #wg-traits channel on Zulip. We've been discussing on and off whether to pursue GATs in the context of Rust's current trait solver or postpone that work until we've made more progress on chalk (with an inclination to the latter).

[PlasmaPower]

According to the RFC, specifically the Iterable example here, type GAT<'a>: 'a should be roughly interpreted as the following bound on the trait where for<'a> Self::GAT<'a>: 'a which is a late bound region from my understanding. However, this approach doesn't work for type generics, which is a gap in the RFC I think, or at least a gap in how to implement it with the current trait solver. I'll hop onto that channel/stream to discuss this further.

[eddyb]

@PlasmaPower You can always synthesize where for<'lt> X<'lt>: 'lt, anonymous or not, but X needs to take an early-bound lifetime for X<'lt> to work at all.

Presumably the only thing missing is converting the (early-bound) lifetime parameter of a GAT to late-bound when generating the bounds.

I think we have a similar problem with constructors and their lifetimes not being turned into late-bound ones when computing the signature (#30904).

[PlasmaPower]

To clarify, you're talking about something like fn f<'b>() -> Self::X<'b> later on, not reusing the lifetime with fn f() -> X<'lt> right? (the latter isn't supposed to work)

[eddyb]

For a struct constructor, at least, all parameters must be used in the fields, so it would always be valid to use late-bound lifetimes (but this is getting a bit offtopic).

(To be clear, I was talking about the ADT constructors provided by the language, not user-written methods).

[PlasmaPower]

Well it sounds like this approach isn't correct, and that it's planned to wait on chalk for GAT implementation, so I'll close this PR.

[alexreg]

How close is Chalk to being ready for things like this?

[semtexzv]

GAT and all the other goodness that comes with chalk is one of the things I'm most interested. I'd be glad to help if I can.