Fix inference for projections with associated traits with multiple bounds; fixes #34792

[rozbb]

The problem was due to the fact that no particular trait bound was specified when a projection was matched with a trait bound, even when multiple bounds existed. The compiler consistently picked the first bound on the associated type that it found that was able to match the type's obligation (which, in the regression test, is <F as Foo> or <Self as Foo>).

The fix involved modifying the SelectionCandidate::ProjectionCandidate variant to carry a trait ref to refer to its matching bound. And instead of stopping on the first match we find, we add all matching bounds to the candidates vector. Also in the winnowing step, a ProjectionCandidate cannot trump another ProjectionCandidate unless they are identical.

[rust-highfive]

r? @pnkfelix

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

[rozbb]

r? @eddyb

[rozbb]

I recall someone saying that this might require a crater test. Is that still the case?

[eddyb]

@doomrobo Indeed, although you first need to fix the two compile-fail tests Travis reports - looks like the error messages changed a bit there.

[eddyb]

This doesn't need to be a probe, I believe in_snapshot will do.

[rozbb]

Hmm. It appears that the error messages for test/compile-fail/associated-types-project-from-hrtb-in-fn-body.rs and test/compile-fail/associated-types-subtyping-1.rs have gotten significantly worse.

Here is the gist for the first one. The second differs in the same way. I have no intuitions so far, but I'm looking into it.

[eddyb]

cc @nikomatsakis@jonathandturner What could cause those errors? It looks like the Sized bound necessary on local variables, but what is it mismatching on?

[bors]

☔ The latest upstream changes (presumably #35091) made this pull request unmergeable. Please resolve the merge conflicts.

[alexcrichton]

Closing due to inactivity, but feel free to resubmit with a rebase!