That was part of the lesson, really. In practice any of the implementations would be "fast enough". The speed of this function does make a measurable difference to the run-time, but measurable isn't the same as significant; the entire program currently spends maybe two minutes a day running, so even if that was all spent in these routines, it still wouldn't make much difference.

(I didn't profile before I started; I went straight to optimisation mode because this particular routine involved Other People's Code and it Looked Inefficient. Hubris again.)

I wrote the final version purely out of curiosity, and it put things so firmly in perspective that I just had to write it up.