16% faster on core2, 5% faster on p4. 10% slower (and thus disabled) on k8.

Originally committed as revision 8992 to svn://svn.ffmpeg.org/ffmpeg/trunk