std::cmp::max vs std::num::Float vs NaN

[steveklabnik]

Issue by pornel
Saturday Jan 31, 2015 at 18:28 GMT

For earlier discussion, see rust-lang/rust#21816

This issue was labelled with: A-libs in the Rust repository

---

It's werid that std::cmp::max can't compare f64 numbers. I know that strictly speaking IEEE floats don't have total order this function expects, but still it's surprising (and partial_max is awkward to use).

And there's std::num::Float::max which works with f64 just fine (the docs don't even say how NaN is handled).

It bugs me that the two versions of max are not consistent in their strictness, and that the first-and-most-obvious max function in the stdlib "doesn't work" with a basic type in the language.

My suggestion:

• Rename the max version that only allows Ord to something else, like total_max or strict_max.
• Implement std::cmp::max for floating point numbers, so that a.max(b) is consistent with max(a,b).

[barosl]

I stumbled upon this issue after reading a comment from HN (which was posted by the issue author, by the way). While I'm more inclined to not have max and min impls for the primitive types f64 and f32, they are now stable and will not change in the future.

Regarding the documentation concerns, rust-lang/rust#25951 made it clear how NaN is handled in those functions. So at least their behaviors are well-defined now.

Also reading rust-lang/rust#25663, @alexcrichton said that cmp::max and cmp::min (and their PartialOrd counterparts) should be more recommended methods for numeric comparisons. Does it mean that we should also document the existence of those methods in {f64, f32}::{max, min} and encourage their uses? cc @bluss @aturon

[xaviershay]

I was confused by this. It's not immediately obvious why max shouldn't work for floats - in practical usage you don't come across NaN all that often.

[xaviershay]

std::num::Float::max not a thing anymore apparently, but 0.0.max(1.0) works.

[scottmcm]

I'm going to deem this resolved with the addition of Ord::min/max, as now all code can use the min method consistently -- the inherent for floats, the trait method for Ord.

Feel free to comment or reopen if you disagree.