Implement DerefImm and DerefMut #7141

[eddyb]

Add the Deref and DerefMut traits and implement overloading explicit dereferences.

[eddyb]

*x and &*x both work as expected, and I've figured that DerefMut would only be used by an explicit dereference in a limited set of contexts: &mut *x, *x = y and *x OP= y, all of which are easy to detect in typeck.

But then there's record_candidate_traits_for_expr_if_necessary in resolve, which assigns one lang-item trait per overloadable expression, so I would have to implement the DerefImm/DerefMut choosing logic there and that doesn't seem right to me.

Oh, and how could #12402 happen while I can't seem to call a method on the right trait (if it's not in trait_map), which is also in the prelude? I'm a bit confused.

EDIT: I didn't realize it at first, but trait_map holds a vector for each entry, so I can have both DerefImm and DerefMut available (not that we should do this in resolve, but I don't want to fix #12402 right now).

[nikomatsakis]

Is this ready for review?

[eddyb]

Ended up fixing #12402 and removing callee_id in the process, I'm happy with how things are turning out.

[nikomatsakis]

/me will review

[pnkfelix]

@eddyb please put a description into the issue at some point, since bors will not pick up its title when it merges.

[nikomatsakis]

This looks excellent. You're missing a few tests. Here is a patch to apply:

https://gist.github.com/nikomatsakis/9348385

r+ once the new tests are applied and nit corrected (and rebased, of course ;)

[nikomatsakis]

Actually, the one other thing I'd like to see:

Some new run-pass tests that:

(1) do not depend on Rc (I prefer to avoid dependencies between tests and the standard library if at all possible, to isolate the tests from refactorings and so forth)

(2) are targeted at checking that the correct method (deref vs deref_mut) is invoked in the correct times. I'm envisioning a type which either modifies a static mut each time the deref vs deref_mut method is called, or perhaps has a Rc<Cell<int>>, something like that. The point is, I want the test to assert that the right method was called at the right times and only once per deref.

[nikomatsakis]

Regarding the existing overloaded operator test, arguably it should be moved to #[test] blocks in Rc but... I don't know. I guess this is really not testing Rc so much as the compiler, even though it draws on Rc. It seems a bit silly to copy the Rc code into the test. So I'd leave the existing test as is.

[nikomatsakis]

Updated my patch with some fns for the compile-fail tests.