- 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 adds three different representations, copied from the Chalk model:
- `Ty::Projection` is an associated type projection written somewhere in the
code, like `<Foo as Trait>::Bar`.
- `Ty::UnselectedProjection` is similar, but we don't know the trait
yet (`Foo::Bar`).
- The above representations are normalized to their actual types during type
inference. When that isn't possible, for example for `T::Item` inside an `fn
foo<T: Iterator>`, the type is normalized to an application type with
`TypeCtor::AssociatedType`.
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
This gives a significant speedup, because chalk will call these
functions several times even withing a single revision. The only
significant one here is `impl_data`, but I figured it might be good to
cache others just for consistency.
The results I get are:
Before:
from scratch: 16.081457952s
no change: 15.846493ms
trivial change: 352.95592ms
comment change: 361.998408ms
const change: 457.629212ms
After:
from scratch: 14.910610278s
no change: 14.934647ms
trivial change: 85.633023ms
comment change: 96.433023ms
const change: 171.543296ms
Seems like a nice win!
For Send/Sync/Sized, we don't handle auto traits correctly yet and because they
have a lot of impls, they can easily lead to slowdowns. In the case of
Fn/FnMut/FnOnce, we don't parse the special Fn notation correctly yet and don't
handle closures yet, so we are very unlikely to find an impl.
This is slightly hacky, but maybe more elegant than alternative solutions: We
just use a hardcoded Chalk trait ID which we special-case to have no impls.
