inhabitedness cache handles cycles incorrectly

[nikomatsakis]

In this example, I get an (incorrect) compilation error:

#![feature(never_type)]

struct Foo { 
    field1: !,
    field2: Option<&'static Bar>,
}

struct Bar {
    field1: &'static Foo
}

fn test_a() {
  let x: Option<Foo> = None;
  match x { None => () }
}

fn test_b() {
  let x: Option<Bar> = None;
  match x { None => () }
}
    
fn main() { }

But if you comment out test_a, you don't. This is because the inhabitedness code, if it starts from Foo, incorrectly caches a result for Bar -- the result is incorrect because Bar did not fully explore Foo, since we found a cycle.

Encountered while fixing #44137 -- I will fix en passante.

[bstrie]

I know the text of the bug implies that this will be fixed relatively soon, yet I'm still nominating because of my perpetual quest to let no soundness bug go unprioritized. :) If this happens to get closed before you get around to discussing it, all the better!

[nikomatsakis]

Will be fixed by #44501. I'm going to categorize this as P-medium -- I think it only affects unstable code, and I think it just causes things to fail to compile that should have compiled.

[nikomatsakis]

triage: P-medium