6% faster on core2 and k8, no change on p4.

Originally committed as revision 8984 to svn://svn.ffmpeg.org/ffmpeg/trunk