weak type inference on late-bound lifetime within Option

[pnkfelix]

Spawned off of #10846 and #15061 (comment)

Example code:

fn main() {
    fn test1(_x: Option< <'a> |y: &'a int|>) {}
    fn test2(_x: Option< |y: &'static int|>) {}
    #[cfg(show_bug)]
    fn siegelord_fn() { test1(Some(|_y      | {})); }
    #[cfg(not(show_bug))]
    fn siegelord_fn() { test1(Some(|_y: &int| {})); }
    fn static_serfn() { test2(Some(|_y      | {})); }

    siegelord_fn();
    static_serfn();
}

Transcript of results:

% ./objdir-dbgopt/x86_64-apple-darwin/stage1/bin/rustc bar.rs 
% ./objdir-dbgopt/x86_64-apple-darwin/stage1/bin/rustc --cfg show_bug bar.rs 
bar.rs:5:31: 5:50 error: mismatched types: expected `core::option::Option<|&'a int|>` but found `core::option::Option<|<generic #1>|>` (expected concrete lifetime, but found bound lifetime parameter 'a)
bar.rs:5     fn siegelord_fn() { test1(Some(|_y      | {})); }
                                       ^~~~~~~~~~~~~~~~~~~
note: expected concrete lifetime is lifetime ReInfer(ReSkolemized(0u, BrNamed(syntax::ast::DefId{krate: 0u32, node: 20u32}, 59u32)))
error: aborting due to previous error
% 

[pnkfelix]

Niko had some ideas on friday for ways we might try to improve things here.

Some of them were integrated into more long-term cleanup of the type-inference code, such as moving to a cleaner model of constraint generation + solving. Others were, for lack of a better term, "hacks" that we might adopt in the short-term, such as attempting to unify the return types of call-expressions first (that is, before doing type inference for the argument expressions), and using the constraints garnered from that to inform the types inferred for the arguments.

[steveklabnik]

https://botbot.me/mozilla/rust/2015-02-11/?msg=31713574&page=1

@huonw: steveklabnik: seems to work now