proper subtyping/LUB/GLB for region fn types

[nikomatsakis]

It occurs to me that our universal quantification step rather expects to know the type which is being universally quantified. But, in some cases, there may be variables present in this type. I am specifically concerned about inference of the parameter types for fn expressions. It's possible that we do something wrong there, though I can't say specifically what, so this issue is kind of a mental note to come back to this and think about it.

[nikomatsakis]

In fact the inference is wrong. We need to adjust our handling of functions and we need to correctly implement the algorithm for handling this sort of thing---

• replace bound regions in the subtype with fresh region variables;
• replace bound regions in the supertype with "skolemized" variants;
• perform unification;
• scan the resulting bindings to make sure that the skolemized variants do not leak out

[nikomatsakis]

One important point: how do we handle GLB/LUB computations?

[lkuper]

@nikomatsakis suggested this as a starter project, so I'm assigning to myself...

[lkuper]

Some notes on what I'm doing for this: https://github.com/mozilla/rust/wiki/Bikeshed-function-subtyping-wrt-regions

[nikomatsakis]

#2981 will help with this.

[nikomatsakis]

This is becoming more problematic as demoding proceeds. I have flagged some of the awkward spots in the code that would benefit from having this bug fixed with FIXMEs.

[nikomatsakis]

This situation is not perfect but I am going to close the issue for now. It's working well enough for the time being.