Solibs and objfile BFD ownership

[Daniel Jacobowitz <drow@false.org>]

I'm seeing ARM test failures caused by this patch:

2009-03-05  Paul Pluzhnikov  <ppluzhnikov@google.com>

        * printcmd.c (do_one_display): Reparse exp_string.
        (display_uses_solib_p): New function.
        (clear_dangling_display_expressions): New function.
        (_initialize_printcmd): Add observer.
        * solib.c (no_shared_libraries): Swap order of calls to
        clear_solib and objfile_purge_solibs.

Specifically this bit:

  /* The order of the two routines below is important: clear_solib notifies
     the solib_unloaded observers, and some of these observers might need
     access to their associated objfiles.  Therefore, we can not purge the
     solibs' objfiles before clear_solib has been called.  */

  clear_solib ();
  objfile_purge_solibs ();

The problem with doing things in this order is that both the solib and
the objfile have a reference to the same BFD.  The solib is
responsible for releasing it (OBJF_KEEPBFD).  arm_objfile_data_cleanup
accesses objfile->obfd during free_objfile, to find the number
of sections; at that point it's already been free'd.

Any thoughts?  I could change the ARM code to work around this, but
that means banning access to objfile->obfd during a not well specified
range of functions.  Handing off the ownership of the BFD seems like
it would be messy.

-- 
Daniel Jacobowitz
CodeSourcery