I created a branch to try to fix the star data packing problem.
In this branch I removed all the bit fields of the star structures, and
instead use bit operations to get the values directly from the mapped
memory. With this we should never need to repack the structures no
matter the architecture.
I just tested it on linux 64 bit so far, where it seems to work with no
noticeable performance penalty (in theory the code emitted should be the
same as before).