Hi,

Thanks a lot for that advice, I was stupid not to try that before. I
got a 4 fold improvement too on my PIV desktop machine, which brings
the specialized ML code about as fast as the C code. This does not
change when I pass -O3 to gcc.

However, I ran the test on a dedibox, which has a VIA processor. The
machine is loaded so it's difficult to conclude anything, but I get
quite always the same surprising figures: without -O3 I get 0.57 for
the C code, 0.91 for the generic ML and 0.25 for the specialized ML;
with -O3 I get 0.12 for the C code, same for the others.

Optimizing is boring...
-- 
David