Documentation, naive_count_32 vs naive_count

[ndmitchell]

These two functions are presented, with naive_count_32 having the weakness that it can overflow. That's a trade-off I might be happy to make, but the documentation doesn't say what the performance difference is. Could that be added, at least in one benchmark, to enable a more informed decision?

[Veedrac]

Per a cargo bench run, naive_count_32 can be some three times as fast from 40 characters upward, due to better vectorization.

I'm not sure I actually want to document this, since performance is so temperamental given its reliance on compiler behaviours of the month.

[ndmitchell]

Without that information, how can I be expected to make a choice between the two alternatives? Short of running my own benchmarks, and that's way too much effort in most cases. Just having that one data point gives me good information.

[llogiq]

As @Veedrac said, this pretty much relies on compiler behavior, which can be a fickle beast. So you should run your own benchmarks before deciding to risk overflow for unknown gains.

[LoganDark]

So you should run your own benchmarks before deciding to risk overflow for unknown gains.

It's worth adding that, should you go this route, you need to also re-run the benchmarks each compiler update to track regressions. It is a lot of effort to rely on the optimizer in a safe and reliable way.

[Veedrac]

Closing wontfix. I am sympathetic to wanting guidance in the documentation, but in this case, if you haven't benchmarked your code as integrated, then I would be much more comfortable if you just use the one that is more correct for the context.