rustc_typeck: unify expected return types with formal return types to propagate coercions through calls of generic functions.

[eddyb]

This allows passing Some(free_fn) to something expecting Option<fn()>, dealing with most of the fallout from #19891.
Fixes #20178.

[rust-highfive]

r? @luqmana

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

[nikomatsakis]

assigning myself as reviewer

[nikomatsakis]

This looks really nice but not quite there. Look at what I wrote and let me know if you have questions.

[nikomatsakis]

One last comment: I'd rather not see the Combine trait used outside of infer. You can just invoke infer::mk_eqty() instead, it's equivalent and shorter.

[nikomatsakis]

(Better yet, move it to a method :)

[eddyb]

It's not equivalent, what I have here could be with_eqty or resolve_with_eqty.

[nikomatsakis]

in what way is it not equivalent? it looks equivalent to me.

[eddyb]

I execute code in the Ok case, before rolling back the snapshot, while mk_eqty would commit it. Unless you mean it's okay to have nested snapshots like this, in which case I have to wonder what the cost is.

[nikomatsakis]

It is fine to have nested snapshots and the cost is nothing.

[nikomatsakis]

Commit within a nested snapshot in particular is basically a no-op.

[eddyb]

I've switched back to subtyping and used a modified version of commit_if_ok that commits region vars (which can be created during subtyping) but discards the type bindings (which could be incorrect and are only usable for expectations).

[nikomatsakis]

We probably want a more detailed comment here. "...commit only the region bindings if successful. The function f must be very careful not to leak any non-region variables that get created."

[nikomatsakis]

@eddyb ok I have to admit, I have my reservations but this looks pretty nice.

r+ modulo nits (and maybe a cleaned up commit history?)

[Ms2ger]

"expected type"