This is only allowed for generic parameters (including `Self` in traits), and
special care needs to be taken to not run into cycles while resolving it,
because we use the where clauses of the generic parameter to find candidates for
the trait containing the associated type, but the where clauses may themselves
contain instances of short-hand associated types.
In some cases this is even fine, e.g. we might have `T: Trait<U::Item>, U:
Iterator`. If there is a cycle, we'll currently panic, which isn't great, but
better than overflowing the stack...
Type-relative paths (`<T>::foo`) also need to work in type context, for example
`<T>::Item` is legal. So rather than returning the type ref from the resolver
function, just check it before.
E.g. `fn foo<T: Iterator>() -> T::Item`. It seems that rustc does this only for
type parameters and only based on their bounds, so we also only consider traits
from bounds.
Nameres related types, like `PerNs<Resolution>`, can represent
unreasonable situations, like a local in a type namespace. We should
clean this up, by requiring that call-site specifies the kind of
resolution it expects.
- refactor bounds handling in the AST a bit
- add HIR for bounds
- add `Ty::Dyn` and `Ty::Opaque` variants and lower `dyn Trait` / `impl Trait`
syntax to them
This wasn't a right decision in the first place, the feature flag was
broken in the last rustfmt release, and syntax highlighting of imports
is more important anyway
- make it possible to get parent trait from method
- add 'obligation' machinery for checking that a type implements a
trait (and inferring facts about type variables from that)
- handle type parameters of traits (to a certain degree)
- improve the hacky implements check to cover enough cases to exercise the
handling of traits with type parameters
- basic canonicalization (will probably also be done by Chalk)
