RFC 0558: require-parentheses-for-chained-comparisons

lang (syntax | expressions)

Summary

Remove chaining of comparison operators (e.g. a == b == c) from the syntax. Instead, require extra parentheses ((a == b) == c).

Motivation

fn f(a: bool, b: bool, c: bool) -> bool {
    a == b == c
}

This code is currently accepted and is evaluated as ((a == b) == c). This may be confusing to programmers coming from languages like Python, where chained comparison operators are evaluated as (a == b && b == c).

In C, the same problem exists (and is exacerbated by implicit conversions). Styleguides like Misra-C require the use of parentheses in this case.

By requiring the use of parentheses, we avoid potential confusion now, and open up the possibility for python-like chained comparisons post-1.0.

Additionally, making the chain f < b > (c) invalid allows us to easily produce a diagnostic message: "Use ::< instead of < if you meant to specify type arguments.", which would be a vast improvement over the current diagnostics for this mistake.

Detailed design

Emit a syntax error when a comparison operator appears as an operand of another comparison operator (without being surrounded by parentheses). The comparison operators are < > <= >= == and !=.

This is easily implemented directly in the parser.

Note that this restriction on accepted syntax will effectively merge the precedence level 4 (< > <= >=) with level 3 (== !=).

Drawbacks

It's a breaking change.

In particular, code that currently uses the difference between precedence level 3 and 4 breaks and will require the use of parentheses:

if a < 0 == b < 0 { /* both negative or both non-negative */ }

However, I don't think this kind of code sees much use. The rustc codebase doesn't seem to have any occurrences of chained comparisons.

Alternatives

As this RFC just makes the chained comparison syntax available for post-1.0 language features, pretty much every alternative (including returning to the status quo) can still be implemented later.

If this RFC is not accepted, it will be impossible to add python-style chained comparison operators later.

A variation on this RFC would be to keep the separation between precedence level 3 and 4, and only reject programs where a comparison operator appears as an operand of another comparison operator of the same precedence level. This minimizes the breaking changes, but does not allow full python-style chained comparison operators in the future (although a more limited form of them would still be possible).

Unresolved questions

Is there real code that would get broken by this change? So far, I've been unable to find any.