Andre,
Bravo!
I've only had a quick opportunity to test your code on a real ARM processor, but preliminary results seem to show about a 2 percentage point reduction in CPU utilization on my SA-1100, which is very good. It may be enough to bring Layer III performance on par with Xaudio's integer decoder under ARM.
I'll continue to do some tests, and may very well include the code in a future release.
Many thanks, -rob