ICE - trait lifetime polymorphism + associated type

[idupree]

trait Trait1<T> {}
trait Trait2<'a> {
  type Ty;
}
fn ice(param: Box<for <'a> Trait1<<() as Trait2<'a>>::Ty>>) {}
fn main() {}

stable: error: internal compiler error: ../src/librustc_trans/debuginfo/metadata.rs:306: get_unique_type_id_of_type() - unexpected type: <() as Trait2<'static>>::Ty

nightly: thread 'rustc' panicked at 'assertion failed: ty.is_normalized_for_trans()', ../src/librustc_trans/collector.rs:534

The stable version's ICE error message also appears in nightly if I write a transmute inside ice():

use std::mem;
fn ice(param: Box<for <'a> Trait1<<() as Trait2<'a>>::Ty>>) {
  let e: (usize, usize) = unsafe{mem::transmute(param)};
}

(This came up while trying to find ways to convince Rust that two complicated type expressions are the same type. I got this ICE, simplified here.)

[TimNN]

cc #36381

[talchas]

A related one that fails with get_unique_type_id_of_type in both cases:

trait Lifetime<'a> {
    type Out;
}
impl<'a> Lifetime<'a> for () {
    type Out = &'a ();
}
fn foo<'a>(x: &'a ()) -> <() as Lifetime<'a>>::Out {
    x
}

fn takes_lifetime(f: for<'a> fn(&'a ()) -> <() as Lifetime<'a>>::Out) {
}
fn main() {
    takes_lifetime(foo);
}

Replacing the explicit fn() with a desired F: for<'a> FnOnce... gives the problematic errors reported in #34430 (at least it doesn't ICE in that case!).

Also apparently this is a regression somewhat, as it doesn't crash on 1.10 (and https://is.gd/NOhNNU doesn't error, etc) (this was incorrect, I was testing the wrong variant of the code)

[Mark-Simulacrum]

Neither example ICEs today. E-needstest.