NLL may be mishandling `_` within type-ascription expression form

[pnkfelix]

Adapted from examples given on #55748. Note that observing this bug requires use of feature(type_ascription) and therefore is not high priority.

consider the following code (play):

#![feature(nll, type_ascription)]

#![allow(dead_code, unused_mut)]
type PairCoupledTypes<T> = (T, T);

#[cfg(this_is_just_here_for_comparison_it_behaves_same_with_and_without_nll)]
fn coupled_types_rhs<'a>(_x: &'a u32, s: &'static u32) -> &'static u32 {
    let ((y, _z),) = ((s, &_x),): (PairCoupledTypes<&u32>,);
    y //~ ERROR unsatisfied lifetime constraints
}

fn coupled_wilds_rhs<'a>(_x: &'a u32, s: &'static u32) -> &'static u32 {
    let ((y, _z),) = ((s, &_x),): (PairCoupledTypes<_>,);
    y //~ ERROR unsatisfied lifetime constraints
}

fn main() {}

Right now, as written, this compiles without error. (Note that fn coupled_types_rhs is not included; that function will error with and without NLL.)

If you remove the nll in the feature declaration, the code for fn coupled_wilds_rhs is rejected.

This is an indication that either:

1. NLL has a bug in how it is handling _ in the type-ascription expression form EXPR: TYPE, or
2. there exists some argument that fn coupled_wilds_rhs should be accepted (and that NLL is correct in accepting it), but I do not see how such an argument exists that would not also imply that we should be seeing the same behavior in fn coupled_types_rhs.

[matthewjasper]

I had a brief look, the annotation is dropped before we get to NLL. Probably here:

rust/src/librustc_typeck/check/mod.rs

Lines 2442 to 2446 in f613dc1

	if ty.has_free_regions() || ty.has_projections() {
	let c_ty = self.infcx.canonicalize_response(&UserTypeAnnotation::Ty(ty));
	debug!("to_ty_saving_user_provided_ty: c_ty={:?}", c_ty);
	self.tables.borrow_mut().user_provided_types_mut().insert(ast_ty.hir_id, c_ty);
	}

[matthewjasper]

This also applies to casts, so it is visble on stable:

fn coupled_wilds_rhs<'a>(_x: &'a u32, s: &'static u32) -> &'static u32 {
    let ((y, _z),) = ((s, _x),): (PairCoupledTypes<_>,);
    y //~ ERROR unsatisfied lifetime constraints
}

[pnkfelix]

so it is visble on stable

type ascription is feature-gated though, no?

[pnkfelix]

(or did you mean to use as, as in ((s, _x),) as (PairCoupledTypes<_>,), in your example?)

[matthewjasper]

Yes, it should be as in the example.