Disjointness based on associated types.

0000-disjoint_associated_types.md

@@ -0,0 +1,100 @@
+- Feature Name: disjoint_associated_types
+- Start Date: 2016-07-10
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+[summary]: #summary
+
+During coherence checking, when determining if the receivers of two impls are
+disjoint, treat bounds with disjoint associated types as mutually exclusive
+bounds.
+
+# Motivation
+[motivation]: #motivation
+
+Consider this set of impls:
+
+```rust
+impl<T> Foo for T where T: Iterator<Item=u64> { }
+impl<T> Foo for T where T: Iterator<Item=i64> { }
+```
+
+Both of these are "blanket impls" - they implement the trait `Foo` for an open
+set of types - any type which implements `Iterator`, with the item `u64` in
+the first, or the item `i64` in the second.
+
+Blanket impls are a tricky beast, because often blanket impls will overlap with
+other impls of the same trait, because the type being implemented for also
+falls into the blanket impl (or could).
+
+However, these two blanket impls, intuitively, do not overlap. Because you can
+only implement `Iterator` once for each type, you cannot have a type which
+implements `Iterator` with both of these types.
+
+Another example of how this could be useful are the return types of functions.
+Consider this code:
+
+```rust
+trait Applicator {
+    fn apply(&mut self, Event) -> io::Result<()>;
+}
+
+fn apply_it<T>(T) -> io::Result<()> where T: Applicator { ... }
+
+impl<F> Applicator for F where F: Fn(Event) -> io::Result<()> { ... }
+
+impl<F> Applicator for F where F: Fn(Event) -> () { ... }
+```
+
+The return type here is an associated type, a function cannot return both
+`io::Result<()>` and `()`. I had code a lot like this in a library I was
+writing; the idea was that quick and dirty stateless implementations of the
+analog trait to `Applicator` could be written as closures, returning either
+an `io::Result` or unit. However, these impls are regarded as overlapping by
+rustc today.
+
+# Detailed design
+[design]: #detailed-design
+
+When considering whether two type variables are disjoint, these (informal)
+rules prove that they are disjoint:
+
+1. If they are both concrete types, and they are not the same type (this rule
+already exists).
+2. If they are both bound by the same trait, and both specify the same
+associated type for that trait, and the types they specify are disjoint.
+
+Additional rules could be added in separate RFCs, such as rules based on a
+syntax for mutual exclusion.
+
+Note that the second rule is recursive.
+
+
+# How will we teach this?
+[teach]: #teach
+
+This will need to be documented in the reference or some other detailed
+document, along with the general description of how Rust coherence works.
+Otherwise, this doesn't particularly need to be called out separate from other
+aspects of the coherence system.
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+This adds more rules to coherence, making it more complicated. However, they
+are intuitive rules, which align with the expected behavior of coherence, so
+hopefully they will not add to the learning burden.
+
+# Alternatives
+[alternatives]: #alternatives
+
+Explicit mutual exclusion of bounds ("negative bound syntax") could provide
+some, but not all, of the benefits of this feature.
+
+We could always do nothing and leave coherence as it is.
+
+# Unresolved questions
+[unresolved]: #unresolved-questions
+
+None.