Clarify and document region nesting rules

[nikomatsakis]

We currently infer some special relationships between regions based on region nesting. i.e., if you have fn foo<'a,'b>(x: &'a &'b int), we infer that 'a <= 'b. This is important for type checking impls, though it will become less so (and possibly irrelevant even) after DST (#6308) lands. In any case, in the meantime, as part of formalizing the type system, we should clarify precisely which kinds of nesting relationships yield subtyping relations. I have a FIXME in the code for the spot where this ought to be done.

[nikomatsakis]

Today btw we are pretty liberal: if you have &'a T, then 'a must be <= all lifetimes that appear free in T. This is I think stronger than we want to be. We should probably limit it to lifetimes that appear in & pointers or as contravariant/invariant type parameters (not sure about covariant ones). Still I think the current setup is not unsound -- so long as we enforce consistent "sanity check" rules on both the caller/callee, which we do.

[emberian]

cc @pnkfelix

Any update on this?

[steveklabnik]

I'll need some help to document this...

[nikomatsakis]

@steveklabnik let's talk ;)

[steveklabnik]

Okay, so after some discussions, this ticket feels like something we'd want well-defined for a formalism, but rarely comes up when you're writing Rust code. as such, i'm going to close it, because once we start formalizing, we'll want a systematic attack, but for now, random issues about it aren't helpful.