The difference is that the old code was using something considerably more heavyweight to start with that has a lot of capabilities that aren't needed for this feature. My excuse is that this was already there and needed for other things and hundreds of microseconds is still plenty fast enough for this, whereas the new code does a lot less and it was a few hundred SLoC that I had to write, though I did encapsulate that such that it's almost exactly the same code at the call site.