- don't return the receiver type from method resolution; instead just
return the autorefs/autoderefs that happened and repeat them. This
ensures all the effects like trait obligations and whatever we learned
about type variables from derefing them are actually applied. Also, it
allows us to get rid of `decanonicalize_ty`, which was just wrong in
principle.
- Autoderef itself now directly works with an inference table. Sadly
this has the effect of making it harder to use as an iterator, often
requiring manual `while let` loops. (rustc works around this by using
inner mutability in the inference context, so that things like unifying
types don't require a unique reference.)
- We now record the adjustments (autoref/deref) for method receivers
and index expressions, which we didn't before.
- Removed the redundant crate parameter from method resolution, since
the trait_env contains the crate as well.
- in the HIR API, the methods now take a scope to determine the trait env.
`Type` carries a trait env, but I think that's probably a bad decision
because it's easy to create it with the wrong env, e.g. by using
`Adt::ty`. This mostly didn't matter so far because
`iterate_method_candidates` took a crate parameter and ignored
`self.krate`, but the trait env would still have been wrong in those
cases, which I think would give some wrong results in some edge cases.
Fixes#10058.
In code like this:
```rust
impl<T> Option<T> {
fn as_deref(&self) -> T::Target where T: Deref {}
}
```
when trying to resolve the associated type `T::Target`, we were only
looking at the bounds on the impl (where the type parameter is defined),
but the method can add additional bounds that can also be used to refer
to associated types. Hence, when resolving such an associated type, it's
not enough to just know the type parameter T, we also need to know
exactly where we are currently.
This fixes#11364 (beta apparently switched some bounds around).
