size[0-3] are not initialized (and can get random negative trash

values), so the comparison with vs_total_ac_bits is messed up on the
first couple loop iterations, leading to AC underflows.

the b->prev[] pointers were not being maintained correctly. We
potentially have to update b->prev[] both before and after the area
whose VLC length is getting adjusted.

this also might fix the ppc regression failure?

patch by (Dan Maas <dmaas maasdigital com>)

Originally committed as revision 5064 to svn://svn.ffmpeg.org/ffmpeg/trunk