Effective breakage to `jiff` due to `ambiguous_negative_literals`

[traviscross]

We should discuss this bit of evidence that came in after the FCP completed on:

• Implement lint against ambiguous negative literals #121364

@BurntSushi said:

This completely broke Jiff once the change landed. Because it specifically supports negating spans like -1.hour(). In the case of Jiff, there is no ambiguity because -1.hour() and (-1).hour() and -(1.hour()) are all precisely equivalent.

cc @BurntSushi @rust-lang/lang

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

[workingjubilee]

This lint indeed assumes the operations are ordering-sensitive. Like abs. Perhaps the lint should be on the function, not the operator?

[Urgau]

The partially-uplifted Clippy clippy::precedence lint (which made up this lint) had a basic heuristic for not linting in cases where the precedence of the operations didn't matter. It was part of the original PR, but @RalfJung preferred consistent linting [comment]:

Personally I'd prefer it to warn consistently, even if the semantics of the function mean it doesn't matter. (a) that seems more consistent and less surprising, and (b) I am writing tests which ensure that sin behaves correctly, and I care about which sign the argument actually has.

[Urgau]

If it's decided that it's not worth linting on these cases, I would propose using the #[diagnostic] namespace by introducing an opt-out, that library authors could use on to suppress the lint for downstream users.

Something like #[diagnostic::irrelevant_negative_literal_precedence] (better name to be found) which could applied to all the relevant function definitions.

As for my rational of being opt-out instead of opt-in, I think it will avoid discrepancy in the coverage as (the wast majority) of users won't know to annotate the functions that are sensitive to negative literals, which is the point of the lint, bring awareness to this surprising aspect of Rust.

[traviscross]

We discussed this in our lang triage call today.

@Urgau: If you would, we'd like this lint changed to allow-by-default while we consider this question further.

cc @compiler-errors