TAIT: it's possible to impl a trait for a tait by using projections

[lcnr]

#![feature(type_alias_impl_trait)]
type Alias = impl Sized;

fn constrain() -> Alias {
    1i32
}

trait HideIt {
    type Assoc;
}

impl HideIt for () {
    type Assoc = Alias;
}

pub trait Yay {}

impl Yay for <() as HideIt>::Assoc {}
// impl Yay for i32 {} // this errors
// impl Yay for u32 {} // this also errors

As coherence checking keeps opaque types opaque this should not be unsound. Considering that a direct impl for opaque types is forbidden, this is an inconsistency we should resolve though.

cc @oli-obk

[oli-obk]

The two commented out lines do not cause any errors (https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=8366951aa5afd218745dd0ebc4564387) So I'd say this is unsound or at least causes ICEs: https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=07fbbfb3b0ecc1dc59ffc71b18b06f1e

[lcnr]

what, is that a recent regression? :o these do error with rustc 1.63.0-nightly (5435ed691 2022-06-07)

[lqd]

It looks to be from last week's #99506, and out of that rollup #99345 looks likely.

[oli-obk]

Hmm... that one got reverted in #99079

cc @compiler-errors

[lqd]

Was that a revert, or a targeted fix for closures ?

[oli-obk]

revert + actual fix of the original issue

[lqd]

The behavior is the same on master, no errors + ICEs. I'll try to manually bisect within the rollup to confirm cargo-bisect-rustc's findings.

[compiler-errors]

Isn't this just another form of #99554?

[lqd]

(I wish we had try builds for each rolled up PR 😅) It was indeed introduced in the same rollup: by eecfdfb / #99383.

Reverting that does indeed bring the conflicting implementation errors back, and fixes the ICE.

[lcnr]

Isn't this just another form of #99554?

no, #99554 is using projections in ways which should fail the orphan check, here we have projections which should pass the orphan check, but can be normalized to opaque types.

These issues need two distinct fixes

the fix for this being: add tests for auto trait stuff with opaque types but allow them in impls. As coherence doesn't use Reveal::All, having one impl trait ref with impl A<X, opaque, Y> for Z should conflict with all impl A<X, whatever, Y> for Z. Have to think more about this though

[oli-obk]

I have a proper fix locally, but we should investigate how #99383 affected this

[ouz-a]

(I wish we had try builds for each rolled up PR 😅) It was indeed introduced in the same rollup: by eecfdfb / #99383.

Reverting that does indeed bring the conflicting implementation errors back, and fixes the ICE.

Although it looks suspicious this is not caused by #99383 I reverted all changes made by it and tested it, this still passes without any errors. Here https://github.com/ouz-a/rust/tree/undo_formalize