`unreachable_patterns` lint due to `min_exhaustive_patterns`

[kpreid]

Code

fn foo() -> Result<(), core::convert::Infallible> {
    todo!()
}

fn main() {
    match foo() {
        Ok(()) => {},
        Err(e) => match e {}
    }
}

Current output

warning: unreachable pattern
 --> src/lib.rs:8:9
  |
8 |         Err(e) => match e {}
  |         ^^^^^^
  |
  = note: this pattern matches no values because `Infallible` is uninhabited
  = note: `#[warn(unreachable_patterns)]` on by default

Desired output

Until the 2024 edition, or the code's MSRV is ≥ 1.82, no warning.

Rationale and extra context

Deleting the Err(e) => match e {} arm will produce code which does not warn on nightly/1.82, but which does not compile on 1.80. This means that users which wish to keep a warning-free-ish nightly (or beta, eventually) build will have to #[allow(unreachable_patterns)] from now until 1.82 is released, or perhaps even until their next MSRV increase.

Deferring the warning until an edition or explicit MSRV increase seems to me to be a good way to avoid churn in the form of #[allow]s. Such #[allow]s are particularly dangerous because unreachable_patterns is also the lint that detects binding patterns that were intended to be matching against constants instead, so declaring it on a wide scope might hide serious bugs, especially if (as is likely) they're not promptly removed upon the release of 1.82.

Rust Version

1.82.0-nightly (2024-08-11 41dd149fd6a6a06795fc)

Anything else?

I noticed this due to my CI against nightly, but comments on the stabilization PR #122792 (comment) mentioned it first. I'm filing this issue to make sure it is not overlooked by being only visible in that comment thread.

[compiler-errors]

@Nadrieril: Do you mind preparing a PR to split out these new cases from the unreachable_patterns, so we can mark it as allow for a few versions (or at least so people can allow it separately from the unreachable pats lint)? Perhaps call it impossible_pattern or something?

[compiler-errors]

Such #[allow]s are particularly dangerous because unreachable_patterns is also the lint that detects binding patterns that were intended to be matching against constants instead

This is a great point IMO, and feels like enough justification to warrant splitting this lint out.

[traviscross]

@rustbot labels +T-lang +I-lang-nominated

Nominating to handle the question of whether bumping the lint to warn-by-default should be done now, done never, done in a few versions, or done in an edition-dependent way (and then maybe done in all editions later).

[Nadrieril]

Yep, will prepare a PR when I get a moment