Inference regression in 1.26

[pietroalbini]

There is an inference regression for a rust project.

• jonhoo/rust-at-sunrise regressed from stable to beta (build log) cc @jonhoo

[Mark-Simulacrum]

cc @nikomatsakis -- I think you wanted to at least be aware of these regressions. We'll probably be closing this though as it looks minimally impactful.

[nikomatsakis]

Hmm. Thanks. I'm not sure what would cause that off top of my head.

[pnkfelix]

visiting for triage.

@pietroalbini are you in a position to provide a standalone, minimized test case to speed up our ability to jump on this?

[pnkfelix]

triage: P-high

(for now, since its a stable-to-beta regression, though @Mark-Simulacrum commented above that the impact may be "minimal" so who knows, it might get downgraded real soon...)

[pietroalbini]

Bisect points at #49163 (wut?), cc @SimonSapin

[SimonSapin]

Apr 12 11:35:07.538 INFO kablam!    Compiling sunrise v0.1.0 (file:///source)
Apr 12 11:35:17.931 INFO kablam! error[E0282]: type annotations needed
Apr 12 11:35:17.931 INFO kablam!    --> src/main.rs:270:23
Apr 12 11:35:17.932 INFO kablam!     |
Apr 12 11:35:17.932 INFO kablam! 270 |     let closest_to = |d| {
Apr 12 11:35:17.932 INFO kablam!     |                       ^ consider giving this closure parameter a type
Apr 12 11:35:17.932 INFO kablam! 
Apr 12 11:35:18.055 INFO kablam! error: aborting due to previous error

https://github.com/jonhoo/rust-at-sunrise/blob/52fcbf516f1f051d8cacad852f4a800b264194fc/src/main.rs#L270-L272

    let closest_to = |d| {
        let gt = perfs.range(d..).next();
        let lt = perfs.range(..d).next_back();

This closure passes RangeFrom<D> values to BTreeMap::range whose signature involves R: RangeBounds<T> (formerly R: RangeArgument<T>). Before #49163, the only corresponding impl were RangeFrom<T>: RangeArgument<T> so this was enough to deduce T = D and infer the type of d. #49163 added RangeFrom<&T>: RangeArgument<T>, making &T = D another possibility and the type of d now ambiguous.

[SimonSapin]

I think this is business as usual when adding new impls. It is a minor change per https://github.com/rust-lang/rfcs/blob/master/text/1105-api-evolution.md#minor-change-implementing-any-non-fundamental-trait since the closure can be changed to |d| {…} to |d: DateTime<Utc>| {…} to fix this and still work in older versions.

[pietroalbini]

Since the crate is now fixed and the breakage was minor and expected, I'm closing this. Thanks to everyone involved on this!