Mostly seems to help in the MC code, which gets a hundred cycles faster.

Originally committed as revision 24463 to svn://svn.ffmpeg.org/ffmpeg/trunk