NLL injects type infer failures in presence of subsequently-masked lifetime errors

[pnkfelix]

Look at https://github.com/rust-lang/rust/blob/master/src/test/ui/lifetime-errors/ex2a-push-one-existing-name-early-bound.rs

trait Foo<'a> {}
impl<'a, T> Foo<'a> for T {}

fn baz<'a, 'b, T>(x: &mut Vec<&'a T>, y: &T)
    where i32: Foo<'a>,
          u32: Foo<'b>
{
    x.push(y); //~ ERROR explicit lifetime required
}
fn main() {
let x = baz;
}

its expected error output is (https://github.com/rust-lang/rust/blob/master/src/test/ui/lifetime-errors/ex2a-push-one-existing-name-early-bound.stderr)

error[E0621]: explicit lifetime required in the type of `y`
  --> $DIR/ex2a-push-one-existing-name-early-bound.rs:17:12
   |
LL | fn baz<'a, 'b, T>(x: &mut Vec<&'a T>, y: &T)
   |                                       - consider changing the type of `y` to `&'a T`
...
LL |     x.push(y); //~ ERROR explicit lifetime required
| ^ lifetime `'a` required

but under NLL it yields (https://github.com/rust-lang/rust/blob/master/src/test/ui/lifetime-errors/ex2a-push-one-existing-name-early-bound.nll.stderr):

warning: not reporting region error due to nll
  --> $DIR/ex2a-push-one-existing-name-early-bound.rs:17:12
   |
LL |     x.push(y); //~ ERROR explicit lifetime required
   |            ^

error[E0282]: type annotations needed
  --> $DIR/ex2a-push-one-existing-name-early-bound.rs:20:9
   |
LL | let x = baz;
   |     -   ^^^ cannot infer type for `T`
   |     |
   |     consider giving `x` a type

error: aborting due to previous error

The latter is amazingly different from the former.

[nikomatsakis]

Hmm. We actually do try to give the nicer errors, but for whatever reason the heuristics we use to decide what is the "main error" to report are failing here. Seems like an interesting case to dig into a bit.

[nikomatsakis]

But I'll mark it as a sub-issue of #51169

[nikomatsakis]

OK so -- the problem here I think is the TEST. In particular, the let x = baz line serves no purpose. It does however cause typeck to fail and hence we never get to MIR borrowck. If you remove that line and enable #![feature(nll)], we get the same error message.