I believe that your 'utf8' case is mostly benchmarking the pulling out of the character count cached in the magic, thus completely missing the original problem.

I was mostly concerned with pointing out that it's important to consider exactly what you're benchmarking.

I'm also not convinced that there is enough of a description of "the original problem" to say whether I missed or not. I'm finding it quite hard to think of an application where the time taken to obtain the length of a string would be very significant?

Except maybe when sorting strings by length, at which point the caching would pretty much negate the first-time cost.

I don't see how your benchmark provides any justification for not using eq ''

Um...cos I modified the original benchmark and didn't think about it. Having just swapped the 'ne's for 'eq's, it does make a suprising difference. Though I haven't thought through whether that's down the efficiency of the operator or the change to the boolean logic.

It did surprise me greatly that ord worked out much faster than bytes::length.