Summary

Require "anonymous traits", i.e. impl MyStruct to occur only in the same module that MyStruct is defined.

Motivation

Before I can explain the motivation for this, I should provide some background as to how anonymous traits are implemented, and the sorts of bugs we see with the current behaviour. The conclusion will be that we effectively already only support impl MyStruct in the same module that MyStruct is defined, and making this a rule will simply give cleaner error messages.

The compiler first sees impl MyStruct during the resolve phase, specifically in Resolver::build_reduced_graph(), called by Resolver::resolve() in src/librustc/middle/resolve.rs. This is before any type checking (or type resolution, for that matter) is done, so the compiler trusts for now that MyStruct is a valid type.
If MyStruct is a path with more than one segment, such as mymod::MyStruct, it is silently ignored (how was this not flagged when the code was written??), which effectively causes static methods in such impls to be dropped on the floor. A silver lining here is that nothing is added to the current module namespace, so the shadowing bugs demonstrated in the next bullet point do not apply here. (To locate this bug in the code, find the match immediately following the FIXME (#3785) comment in resolve.rs.) This leads to the following

mod break1 {
    pub struct MyGuy;

    impl MyGuy {
        pub fn do1() { println!("do 1"); }
    }
}

impl break1::MyGuy {
    fn do2() { println!("do 2"); }
}

fn main() {
    break1::MyGuy::do1();
    break1::MyGuy::do2();
}

<anon>:15:5: 15:23 error: unresolved name `break1::MyGuy::do2`.
<anon>:15     break1::MyGuy::do2();

as noticed by @huonw in https://github.com/rust-lang/rust/issues/15060 .

If one does not exist, the compiler creates a submodule MyStruct of the current module, with kind ImplModuleKind. Static methods are placed into this module. If such a module already exists, the methods are appended to it, to support multiple impl MyStruct blocks within the same module. If a module exists that is not ImplModuleKind, the compiler signals a duplicate module definition error.
Notice at this point that if there is a use MyStruct, the compiler will act as though it is unaware of this. This is because imports are not resolved yet (they are in Resolver::resolve_imports() called immediately after Resolver::build_reduced_graph() is called). In the final resolution step, MyStruct will be searched in the namespace of the current module, checking imports only as a fallback (and only in some contexts), so the use MyStruct is effectively shadowed. If there is an impl MyStruct in the file being imported from, the user expects that the new impl MyStruct will append to that one, same as if they are in the original file. This leads to the original bug report https://github.com/rust-lang/rust/issues/15060 .
In fact, even if no methods from the import are used, the name MyStruct will not be associated to a type, so that

trait T {}
impl<U: T> Vec<U> {
    fn from_slice<'a>(x: &'a [uint]) -> Vec<uint> {
        fail!()
    }
}
fn main() { let r = Vec::from_slice(&[1u]); }

error: found module name used as a type: impl Vec<U>::Vec<U> (id=5)
impl<U: T> Vec<U>

which @Ryman noticed in https://github.com/rust-lang/rust/issues/15060 . The reason for this is that in Resolver::resolve_crate(), the final step of Resolver::resolve(), the type of an anonymous impl is determined by NameBindings::def_for_namespace(TypeNS). This function searches the namespace TypeNS (which is not affected by imports) for a type; failing that it tries for a module; failing that it returns None. The result is that when typeck runs, it sees impl [module name] instead of impl [type name].

The main motivation of this RFC is to clear out these bugs, which do not make sense to a user of the language (and had me confused for quite a while).

A secondary motivation is to enforce consistency in code layout; anonymous traits are used the way that class methods are used in other languages, and the data and methods of a struct should be defined nearby.

Detailed design

I propose three changes to the language:

impl on multiple-ident paths such as impl mymod::MyStruct is disallowed. Since this currently surprises the user by having absolutely no effect for static methods, support for this is already broken.
impl MyStruct must occur in the same module that MyStruct is defined. This is to prevent the above problems with impl-across-modules. Migration path is for users to just move code between source files.

Drawbacks

Static methods on impls-away-from-definition never worked, while non-static methods can be implemented using non-anonymous traits. So there is no loss in expressivity. However, using a trait where before there was none may be clumsy, since it might not have a sensible name, and it must be explicitly imported by all users of the trait methods.

For example, in the stdlib src/libstd/io/fs.rs we see the code impl path::Path to attach (non-static) filesystem-related methods to the Path type. This would have to be done via a FsPath trait which is implemented on Path and exported alongside Path in the prelude.

It is worth noting that this is the only instance of this RFC conflicting with current usage in the stdlib or compiler.

Alternatives

Leaving this alone and fixing the bugs directly. This is really hard. To do it properly, we would need to seriously refactor resolve.

Unresolved questions

None.

The Rust RFC Book