rustc_resolve: Move ReducedGraphParent building into its own module

[gereeter]

This also gets rid of a bunch of unnecessary .clones.

cc @eddyb

[rust-highfive]

r? @eddyb

(rust_highfive has picked a reviewer for you, use r? to override)

[gereeter]

I rearranged the commits to make the more controversial free function change last, killed ReducedGraphParent, and added an explaination of my motivation to the free function commit.

[eddyb]

I see, but why not remove just the parent field in the visitor and leave the code in methods?
Also, I'd have someone else double-check the clone changes, it's not obvious to me that it's a net improvement.
Otherwise this looks very good to me, thanks a lot!

[nikomatsakis]

So this looks good to me. I don't have a strong opinion about the move to free fns from methods. It is true that self.foo(...) reads better than foo(resolver, ...) -- helps to focus the eye on the relevant arguments. Another possibility would be to use an extension trait on Resolver or, probably better, impl the methods on a newtype'd Resolver like struct GraphBuilder { resolver: Resolver }. The visitor would thus carry around a pointer to that instead of the resolver itself, and it would invoke self.builder.foo() rather than foo(self.resolver, ...).

[nikomatsakis]

Another option might be to just rename the parent field to something less tempting, like most_recently_visited_item

[gereeter]

Okay, rebased. I now use the idea that @nikomatsakis suggested of having a separate GraphBuilder struct.

P.S. Do you know of any good way to rebase refactors like this? The most effective method I've found is just scrapping everything and starting from the beginning.