One thing that I find interesting in all of this is that even before optimising the code, version 2 still worked faster than version 1, even though it does more things. A lot of that is because the pre-built looping material in expl3 does a much better job than my own attempts in version 1 of siunitx. Of course, good programmers will always use fast loops, but for the rest of us I think this shows how a sensible tool-kit can bring benefits “behind the scenes”.