New lint: manual_is_multiple_of

[samueltardieu]

I've added a min_divisor configuration option, default to 4, to not trigger on smaller divisibility operations. I would prefer not to lint if a & 1 == 0 as if a.is_multiple_of(2) by default because the former is very idiomatic in systems (and embedded) programming.

A min_and_mask_size option, defaulting to 3, sets the default bits to be and-masked before this lint triggers; that would be n in v & ((1 << n) - 1) == 0. The form v % 10 == 0 is always linted.

This PR will stay in draft mode until the next rustup which will mark unsigned_is_multiple_of stable for Rust 1.87.0, and the feature flags will be removed.

What should its category be? I've used "complexity", but "pedantic" might be suitable as well.

Close #14289
changelog: [manual_is_multiple_of]: new lint
r? ghost

[joshtriplett]

Having a min_divisor configuration option seems reasonable, but I think that option should default to flagging every instance of this. Large codebases that want to keep the & 1 == 0 pattern can set that option.

[samueltardieu]

Having a min_divisor configuration option seems reasonable, but I think that option should default to flagging every instance of this. Large codebases that want to keep the & 1 == 0 pattern can set that option.

I've added an extra commit to test this with lintcheck. There was only one hit in Clippy sources (except for tests).

[samueltardieu]

@joshtriplett The lintcheck output seems quite reasonable indeed.

[y21]

I agree with linting the a % b == 0 pattern, but I'm unsure about the & one... Something like (x & 0b111) == 0 feels like a very general expression.
I'm a bit concerned that linting here has the opposite effect: emitting warnings on code that has a totally different semantic meaning where x.is_multiple_of(8) would be a very obfuscated way of checking that the last 3 bits aren't set.

[joshtriplett]

First of all, please don't treat that as a blocker; it's not as important as the % N == 0 case.

That said: I've seen many, many codebases which use & (PAGE_SIZE - 1) == 0 and similar instead of % PAGE_SIZE == 0, when they know the value is a power of two.

We may need to tune the heuristic to make sure it doesn't produce false positives, and for instance we may want a different threshold for & vs % (& 1 == 0 might be a flag test), but I think things like & 4095 == 0 or & (4096 - 1) == 0 or & (CONST_EQUAL_TO_4096 - 1) == 0 are pretty clear.

Looking at the lintcheck output, I think I agree with not flagging & 1 == 0 by default.

[samueltardieu]

What about applying the min-divisor option (maybe using another name, like min-one-bits to represent the number of ones) only to the &(2^n-1) form of the lint, with a sensible default such as 5 bits? The % form would be linted unconditionally.

[joshtriplett]

@samueltardieu Sounds reasonable. I would say the default should flag 7 and 15 by default, and maybe 3, but not 1. Because 1 might be a flag, but 3 is definitely two bits.

[samueltardieu]

Done, see the toplevel comment for the updated option.