Mismatched type errors have become useless

[nagisa]

A minimal test case. My actual case is much more complex, involves functions/callbacks with multiple params and several layers of typedefs.

fn takes_cb(f: fn(i8)) {}

fn main() {
    fn callback(x: i32) {}
    takes_cb(callback)
}

The error reported on 1.11.0 is this:

test.rs:5:14: 5:22 error: mismatched types [E0308]
test.rs:5     takes_cb(callback)
                       ^~~~~~~~
test.rs:5:14: 5:22 help: run `rustc --explain E0308` to see a detailed explanation
test.rs:5:14: 5:22 note: expected type `fn(i8)`
test.rs:5:14: 5:22 note:    found type `fn(i32) {main::callback}`

Useful and obvious. What error looks like now (stable (1.14.0)/beta/nightly (1.16.0))?

error[E0308]: mismatched types
 --> test.rs:5:14
  |
5 |     takes_cb(callback)
  |              ^^^^^^^^ expected i8, found i32

Cute but is as worthless and confusing as it gets.

[eddyb]

This actually looks like a change in coercion? That is, a mismatch in the signature used to prevent the coercion, resulting in the subtyping fallback which simply fails, at the othermost level, whereas now the coercion is attempted and the error from inside the signature mismatch is being reported.

[nikomatsakis]

@eddyb any clues what change might have caused that?

cc @estebank @jonathandturner -- this doesn't look familiar, does it?

[sophiajt]

@nikomatsakis - not to me. I thought we still always showed the expected/found full type since sometimes you need it to understand the error (this one included)

[estebank]

@nikomatsakis sadly, it doesn't for me either.

@jonathandturner it doesn't if the type mismatch is between primitive types, which leads me to think that somewhere a TypeError is being constructed for the arguments alone, and not for the method signature. Probably a callee of report_mismatched_types.

[arielb1]

triage: P-high - serious diagnostic regression, assign to @pnkfelix.

[estebank]

While doing something else around type errors, I realized by chance why this was happening.

[pnkfelix]

(seems like #39905 will fix this.)