Is all your XDATA random small variables, or do you have a few very large structures? In that case, you may consider manually banking them, by writing accessor functions for this data, i.e. get_records(n,&rec) and set_record(n,&rec). This could allow 99% of all variables be non-banked, reducing the code size and speed penalty. Just remember that such an approach requires all variables touched by interrupt handlers to be non-bankend.

Another thing is of course to before starting a project consider exactly what your needs are, and from the start select a processor baseed on these needs.