Spans for errors in try operator expressions are not great

[nrc]

E.g.,

error: the type of this value must be known in this context
   --> src/librustc_save_analysis/sig.rs:281:39
    |
281 |             let generics: Signature = generics.make(offset + text.len(), id, scx)?;
    |                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

the expression with unknown type is the variable use generics, but the error highlight covers the entire sub-expression of expr?. I assume this is due to us doing something sub-optimal with the spans here in the HIR lowering.

[aturon]

@nrc is worried that similar issues are arising with other desugarings, i.e. that this might be an across-the-board regression. Worth investigating.

[nikomatsakis]

Standalone example:

fn foo() -> Result<(), ()> {
    let x = None;
    Ok(x.unwrap().make()?)
}

fn main() { }

gives

rustc 1.17.0 (56124baa9 2017-04-24)
error: the type of this value must be known in this context
 --> <anon>:3:8
  |
3 |     Ok(x.unwrap().make()?)
  |        ^^^^^^^^^^^^^^^^^

error: aborting due to previous error

[nikomatsakis]

triage: P-medium

Doesn't seem like a "stop everything" sort of problem, but would be good to figure it out. However, it does not appear to be a regression: e.g., the above test case behaves the same in 1.16.0 as well.