Type inference regression on nightly-2024-05-20

[oblique]

Code

I tried this code:

[package]
name = "regression"
version = "0.1.0"
edition = "2021"

[dependencies]
time = { version = "=0.3.34", features = ["parsing"] }

fn main() {}

I expected to see this happen: Compile successfully.

Instead, this happened: Compilation failed with the following.

error[E0282]: type annotations needed for `Box<_>`
  --> /home/oblique/.cargo/registry/src/index.crates.io-6f17d22bba15001f/time-0.3.34/src/format_description/parse/mod.rs:83:9
   |
83 |     let items = format_items
   |         ^^^^^
...
86 |     Ok(items.into())
   |              ---- type must be known at this point
   |
help: consider giving `items` an explicit type, where the placeholders `_` are specified
   |
83 |     let items: Box<_> = format_items
   |              ++++++++

For more information about this error, try `rustc --explain E0282`.
error: could not compile `time` (lib) due to 1 previous error

Version it worked on

It most recently worked on:

rustc --version --verbose:

rustc 1.80.0-nightly (b1ec1bd65 2024-05-18)
binary: rustc
commit-hash: b1ec1bd65f89c1375d2cf2fb733a87ef390276d3
commit-date: 2024-05-18
host: x86_64-unknown-linux-gnu
release: 1.80.0-nightly
LLVM version: 18.1.4

Version with regression

rustc --version --verbose:

rustc 1.80.0-nightly (d84b90375 2024-05-19)
binary: rustc
commit-hash: d84b9037541f45dc2c52a41d723265af211c0497
commit-date: 2024-05-19
host: x86_64-unknown-linux-gnu
release: 1.80.0-nightly
LLVM version: 18.1.4

@rustbot modify labels: +regression-from-stable-to-nightly-regression-untriaged

[lukas-code]

Regression in #99969.

This code used to infer _ as [char], but now it is ambiguous between [char] and str. (playground)

fn main() {
    let _: Box<_> = ['a'].into_iter().collect();
}

@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged T-libs-api

[kpreid]

Per #99969 (comment) this is a known breakage, which can be addressed by updating your project's time dependency to version 0.3.36 or later.

[oblique]

Does this mean it will be enabled by 2024 edition?

[workingjubilee]

A type inference regression is permitted breakage according to Rust RFC 1122.

This is not even a change in type inference though, it's rather because of type inference. In particular, when we add an impl and that means now 2 possible impls resolve, instead of 1, it can break things. Customarily, if the library teams expect a breakage may be particularly widespread, rejecting it is discussed and there is a formal decision to accept it or not. However it's not clear to me from the FCP's comments (or lack thereof) if T-libs-api discussed this breakage before/during the FCP, beyond the chatter that occurred there? #99969 (comment)

@rust-lang/libs-api Can I just assume that if T-libs-api FCPs a new impl of something like From or FromIterator, and then a team member r+'s landing something that they said "maybe a crater run" about, that you have discussed the breakage possibility and said "worth it"? It may be helpful if you just formally note "probably will break someone relying on a particularly strained .into() somewhere, but that seems fine" on the FCP.

Spicy option: r+ it the moment it hits the final comment period so we might actually get any such reports in the 10 day period.

[dtolnay]

We discussed this regression in today's libs-api team meeting, and are inclined not to revert.

1.80 is in beta already so we'll start seeing fallout in the beta crater runs. The team will re-evaluate if the beta crater runs look disastrous with many crates other than time impacted.