NLL doesn't respect user types in closure signatures

[matthewjasper]

For example, the following code compiles

#![feature(nll)]
fn foo<'a>() {
    |x: &'a i32| -> &'static i32 {
        return &x;
    };
}

cc #47184

[nikomatsakis]

Tagging as RC2

[nikomatsakis]

So, the closure region requirements are as follows:

> rustc ~/tmp/issue-54692.rs   -Zverbose
note: External requirements
 --> /home/nmatsakis/tmp/issue-54692.rs:6:5
  |
6 | /     |x: &'a i32| -> &'static i32 {
7 | |         return &x;
8 | |     };
  | |_____^
  |
  = note: defining type: DefId(0/1:10 ~ issue_54692[317d]::foo[0]::{{closure}}[0]) with closure substs [
              i8,
              extern "rust-call" fn((&'_#1r i32,)) -> &'_#2r i32
          ]
  = note: late-bound region is '_#3r
  = note: number of external vids: 4
  = note: where '_#1r: '_#2r

[nikomatsakis]

You can see the '_#1r: '_#2r indicating that — indeed — the argument must outlive the return type. But we have lost the connection to 'a and 'static, indeed.

[nikomatsakis]

The user-given signature of the closure is computed here, during the HIR typeck:

rust/src/librustc_typeck/check/closure.rs

Lines 529 to 533 in 1c5e9c6

	// Get the signature S that the user gave.
	//
	// (See comment on `sig_of_closure_with_expectation` for the
	// meaning of these letters.)
	let supplied_sig = self.supplied_sig_of_closure(decl);

(In general, reconciling the expected signature and the user-given signature is a royal pain, as is described in detail in that file.)

I imagine what we have to do is probably to record the "supplied closure signature" somewhere, probably in a new table similar to user_substs.

The question is where it should be accessed from. I think it should probably be the closure's job to enforce these relations (if you read that code above, you might see why: in general, the user's signatures are written from the "POV" of the closure body, rather than its creator).

This implies that MIR may want to grow a user_fn_sig field, set to None for everything but closures. The type-checker can then unify those signatures with the types of the arguments.

[nikomatsakis]

Hmm this may be blocking me in other areas.

[pnkfelix]

adding self @spastorino to the assignee list to try to ensure this gets addressed soon

[pnkfelix]

Got a report from @mikhail-m1 that they are actively working on this

Unassigning @spastorino (since @mikhail-m1 is taking lead here)