behavior of `'_` around `dyn Trait` is wrong

[nikomatsakis]

Generally speaking, '_ is equivalent to "eliding" a lifetime completely. That is, &'_ T and &T are equivalent, as are Foo<'_> and Foo (unfortunately). However, there are two cases where this is not true:

• dyn Trait + '_ -- the object-lifetime-defaulting rules are different, and '_ needs to interact with them (same for Trait, of course)
• impl Trait + '_ -- same, but for different reasons

@cramertj pointed out that our current behavior around dyn Trait is not what we really want:

#![feature(dyn_trait)]
#![feature(underscore_lifetimes)]

// These two give the same error that the iterator must be valid for the 'static lifetime:
fn a<T>(items: &[T]) -> Box<dyn Iterator<Item=&T>> {
    Box::new(items.iter()) //~ ERROR cannot infer an appropriate lifetime
}

fn b<T>(items: &[T]) -> Box<dyn Iterator<Item=&T> + '_> {
    Box::new(items.iter()) // *Should* be OK, currently errors
}

// But this one works:
fn c<'a, T>(items: &'a [T]) -> Box<dyn Iterator<Item=&'a T> + 'a> {
    Box::new(items.iter()) // OK
}

fn main() { }

[ExpHP]

fn b<T>(items: &[T]) -> Box<dyn Iterator<Item=&T> + '_>;

I find this fascinating. Who would have predicted prior to implementation that introducing '_ would allow us to elide even more lifetimes?