I think(!) this patch is reasonable.  It seems to be in line with what
Jim has in mind
(based on my interpretation of your comments Jim on the gdb list -
please correct
me if I'm wrong of course).

When gdb resolves type information for class "derived" from objfile
gdb2384 (from the testcase), it fills in the TYPE_VPTR_BASETYPE field
with class "base"
from objfile gdb2384-base.so.  When the program is rerun the type
information for base-in-so-base.so is discarded leaving
TYPE_VPTR_BASETYPE dangling.

Ok to check in?  Or any suggestions for what's needed instead?