"Regression" - some static bounds incorrectly compiled in 1.63..=1.67

[QuineDot]

Heads-up

The example below should not compile and currently does not compile, so this is going to be low priority.

However, the example compiled for 1.63 -- 1.67 inclusive (it doesn't compile in 1.62 or in 1.68 to current).

I didn't see anything that looked relevant in the release notes, and I figure there should at least be a test somewhere, so it doesn't start compiling again.

So this issue is primarily about making sure a test exists.

Code

I tried this code:

pub trait Trait {}

pub trait Assoc<'x> {
    type A: ?Sized;
}

impl<'x> Assoc<'x> for str { type A = &'x dyn Trait; }

pub fn assoc<T: 'static>() {}
pub fn assoc_helper<'a: 'a>() {
    assoc::< <str as Assoc<'a>>::A >();
}

I expected to see this happen: Inexplicably compile like it did during some exploration months ago. (Search for "this all works" for more examples with probably irrelevant variations if you're curious.)

Instead, this happened: Failed to compile, which makes more sense.

Version it worked on

It most recently worked on: Rust 1.67, and also 1.63, 1.64, 1.65, and 1.66.

Version with regression

1.68 to present.

[Jules-Bertholet]

@rustbot label E-needs-test regression-from-stable-to-stable -regression-untriaged

[compiler-errors]

Oh right, forgot to click enter--

searched nightlies: from nightly-2023-01-01 to nightly-2023-06-01
regressed nightly: nightly-2023-01-10
searched commit range: cc47b06...3020239
regressed commit: af58fc8

bisected with cargo-bisect-rustc v0.6.6

Host triple: aarch64-apple-darwin
Reproduce with:

cargo bisect-rustc 2023-01-01 --end 2023-06-01 --regress error

[QuineDot]

The tests from PR #101947 look sufficient to me, so I'm going to close this.

Thanks!