From: Florian Zumbiehl <redacted>
Date: Tue, 31 Jul 2007 13:05:47 +0200

A few variations I tried back when I created the patch, using larger
things than a char for accumulating the pieces and then folding down
from that, turned out to be slower than what I finally submitted, at
least on the machines I tested it on. I didn't do comprehensive testing,
as it really doesn't matter, after all, but I think the version I
submitted is pretty fast, plus it's quite readable, and it keeps the
flexibility of different hash sizes, but still should allow the
compiler to optimize away the loops that allow for this flexibility.

Therefore, I've put your original patch into the tree :-)

Thanks.