Type inference fails even though all types are known

[SeeSpring]

The following code fails to compile, even though all types are known
https://play.rust-lang.org/?version=stable&mode=debug&edition=2015&gist=8de9e57e8de5cd8f2896f422d693f453

pub fn f(s: &str) {
    let mut it = s.split(|c: char| c == ',').fuse();
    let t = it.next().unwrap().parse::<usize>().unwrap();
}

   Compiling playground v0.0.1 (/playground)
error[E0282]: type annotations needed
 --> src/main.rs:3:13
  |
3 |     let t = it.next().unwrap().parse::<usize>().unwrap();
  |             ^^^^^^^^^^^^^^^^^^ cannot infer type for `T`
  |
  = note: type must be known at this point

error: aborting due to previous error

For more information about this error, try `rustc --explain E0282`.
error: Could not compile `playground`.

To learn more, run the command again with --verbose.

[talchas]

Bisecting releases says it worked in 1.12.0, broke in 1.13.0

[jonas-schievink]

Technically a stable regression, so marking as such - even though nobody seems to have been hit by it. It does pretty clearly look like a bug though.

[ExpHP]

Ran cargo-bisect-rustc on the nightlies:

3c5a0fa...e9bc1ba

This is still a pretty broad range...

[jonas-schievink]

Only thing in that range that touches the typesystem seems to be #35883

[Mark-Simulacrum]

May also be related to #35656

[jonas-schievink]

Ah, I missed that one. That might be more likely, yeah (since specialization might now be involved, but before wasn't).

[pnkfelix]

triage, P-medium. (Not a soundness issue, and appears to have a straight-forward if annoying work-around of introducing a local with an explicit type.) Leaving nominated label for discussion at next T-compiler meeting.

[pnkfelix]

Discussed briefuly at T-compiler meeting. assigning to self in attempt to ensure this doesn't just get lost. (I don't want to leave it nominated for discussion again at next week's meeting.)