rustdoc: Impl for boxed type doesn't show up on the type's page

[camelid]

You'd expect the impl for the boxed type to show up, but it doesn't. I seem to recall rustdoc displays impls anytime the type is contained in the generics, though, so it's odd that it's not happening.

pub struct MyType;

impl Iterator for Box<MyType> {
    type Item = ();

    fn next(&mut self) -> Option<Self::Item> {
        todo!()
    }
}

Originally reported by @danielhenrymantilla (with Pin<&mut MyType>) on Discord.

[camelid]

@danielhenrymantilla also reported a case involving a Box<dyn Trait> impl that didn't show up on the trait's page. IIRC this issue is somewhat known already though.

[GuillaumeGomez]

I can't find a case where the impl is added if the type is in the generics. Do you have an example maybe?

[camelid]

On Vec's page: https://doc.rust-lang.org/nightly/std/vec/struct.Vec.html#impl-PartialEq%3CVec%3CU%2C%20A%3E%3E-5

[notriddle]

That one's happening because of this thing. Notice that the generics on the trait count, but the generics on Self don't.

rust/src/librustdoc/formats/cache.rs

Lines 444 to 475 in 1c8966e

	let ret = if let clean::Item { kind: box clean::ImplItem(ref i), .. } = item {
	// Figure out the id of this impl. This may map to a
	// primitive rather than always to a struct/enum.
	// Note: matching twice to restrict the lifetime of the `i` borrow.
	let mut dids = FxHashSet::default();
	match i.for_ {
	clean::Type::Path { ref path }
	| clean::BorrowedRef { type_: box clean::Type::Path { ref path }, .. } => {
	dids.insert(path.def_id());
	}
	clean::DynTrait(ref bounds, _)
	| clean::BorrowedRef { type_: box clean::DynTrait(ref bounds, _), .. } => {
	dids.insert(bounds[0].trait_.def_id());
	}
	ref t => {
	let did = t
	.primitive_type()
	.and_then(|t| self.cache.primitive_locations.get(&t).cloned());
	if let Some(did) = did {
	dids.insert(did);
	}
	}
	}
	if let Some(generics) = i.trait_.as_ref().and_then(|t| t.generics()) {
	for bound in generics {
	if let Some(did) = bound.def_id(self.cache) {
	dids.insert(did);
	}
	}
	}

We definitely need to fix this, specifically, for #[fundamental] types, because that's how you wind up with impls that aren't actually documented anywhere.