bounds on trait impls are used in implied bounds

[aliemjay]

The following compiles although it shouldn't:

trait Trait {
    type Assoc;
}

impl<T: 'static> Trait for Box<T> {
    type Assoc = ();
}

struct MyTy<U>(U)
where
    U: Trait,
    U::Assoc: Sized, // any predicate naming U::Assoc
;

fn fn_test<T>(_: MyTy<Box<T>>) {}

fn_test should fail with an error requiring an explicit T: 'static bound.

This is a compiler bug for the following reasons:

• It certainly does not follow from the rules of RFC 1214 and RFC 2089. The implied bounds from a trait reference (Box<T>: Trait) includes only the where-clauses of the trait itself, not the trait impl.
• It breaks stability guarantees in the sense that relaxing region constraints on trait impls is now a breaking change. Removing the useless bound U::Assoc: Sized is also a breaking change.
• It requires remote reasoning when dealing with implied bounds.
• It is a surprising behavior: why does the trivial bound U::Assoc: Sized make any meaningful difference?

[jackh726]

Just for reference: This is a stable-to-stable regression, first passing in 1.29.0.

🤷‍♀️ on adding the label

[lcnr]

this still applies to bevy and during MIR typeck due to #119956, reopening.