I am somewhat confused by why we need this weird flag at all. This feels very wrong to have. Why did the earlier version not suffice? (where we just introduced a reborrow)

This test is the problem:

#[test]
fn method_resolution_by_value_before_autoref() {
    check_types(
        r#"
trait Clone { fn clone(&self) -> Self; }
struct S;
impl Clone for S {}
impl Clone for &S {}
fn test() { (S.clone(), (&S).clone(), (&&S).clone()); }
          //^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (S, S, &S)
"#,
    );
}

We should continue for the reborrow only one time, and this flag serve that propose. Without it, we will get (S, S, S).

Ah, now I see what we want to do here. We need to set the autoref field of the ReceiverAdjustments first_adjustment as well as increment the autoderefs field once for this iterate_method_candidates_by_receiver call when the receiver_ty is a reference

That then also mimics what rustc does here