librustc: Implement explicit self for objects

[pcwalton]

r? @nikomatsakis

[pcwalton]

I'm not sure that's simpler than just enforcing the method restrictions at the call site, on a per-method basis. I'm worried about the loss of flexibility that would result from "if one method is tainted, the entire trait is tainted".

[nikomatsakis]

r+ --- however, I believe that our type system does not currently enforce the various common-sense requirements that the trans code expects. We should open up another bug to draft and enforce those rules.

[nikomatsakis]

@pcwalton the main thing is that the rules get complicated due to the possibility of generic types. Keep in mind that one can always refactor one's traits to include a supertrait with only object-compatible methods and a subtype with the other methods.

In any case, I am also fine with extending the more fine-grained approach we have now. However, I have historically found it hard to explain, even just to other team members, much less the "general public". The fact that it took me an hour to explain it to dherman when it was just us two in a room with a whiteboard made me nervous. But maybe I just did a bad job. For reference, the full set of rules we enforce today are:

• if you have an object receiver, you cannot invoke the forbidden methods
• if a trait T contains forbidden methods, then an object @T is not considered to implement the trait T

So I could imagine a rule that says something like:

• given an object receiver @T, you can call &self or @self methods
• given an object receiver ~T, you can call &self or ~self methods
• given an object receiver &T, you can call &self methods
• if there are any methods that are not &self, then T objects do not implement T