Hello,

This patch is a prototype towards converting all the ad hoc code
and various deprecated features used to implement relocations to
using a target_so_ops vector instead. One of the nice side-effects
is that we are able to get rid of nm-rs6000.h entirely.

This is a fairly lengthy email, which I am not too happy about,
because I think it's going to make it harder for anyone willing
to take a look and comment. But I thought it was important to provide
some general info on it too. The reason for posting the patch
generate comments, but also to ask a couple of questions on topics
open for discussion regarding missing pieces in the core part of GDB.
I will ask these questions again at the next submission, when the
patch is a little cleaner, if no decision is made with that patch.

The interface to the relocation data is itself relatively simple,
and documented through a combination of the PT_LDINFO request in
the ptrace man page, and the ldr.h system include. Basically,
the PT_LDINFO gives us access to a raw buffer which is an array
of structures whose type is defined in ldr.h.

One of the small difficulties came from a specificity of the AIX
platform regarding shared libraries: What we view as shared libraries
on, say, GNU/Linux, is an archive file. For instance, /lib/libc.a.
The linker then searches inside the archive file for dynamic objects
(Eg: shr.o) on which the program depends, and links it against those
specific object files, not just the archive. We observe that through
the use of ldd on a simple program:

    % ldd simple_main
    simple_main needs:
         /lib/libc.a(shr.o)
         /lib/libcrypt.a(shr.o)

This means that the current model where an objfile is a filename
doesn't work very well. In this case, it's a filename and a member
name. This prototype tip-toes around the issue for now, to limit
its scope.

This is a relatively messy prototype, where I have not paid
any attention to memory management or process separation
(Eg: -tdep code calling -nat code). This is a big part of
the next iteration for this patch.

Some elements I would like to have some feedback on:

  - One important bit is something I alluded to above: a solib
    name is now no longer sufficient to identify it; we need
    the filename, which is usually an archive, and a member
    name (which may be NULL). Right now, I store a composite
    into the struct so_list->so_name field which follows the
    same convention as AIX's ldd: Eg. "/lib/lib.c(shr.o)".
    Some minor issues:

      - struct objfile's name is the name from the bfd, and
        thus Eg "shr.o"; any display of that name in the debugger
        output and warning/error messages is incomplete

      - the bfd_open target_so_ops method only takes a pathname
        for getting the bfd; but that's not sufficient.
        What the current does is just parse the pathname
        and split it into real pathname, and member name.
        Acceptable for a prototype, but perhaps we should
        consider passing a reference to the so_list object?

      - See also the FIXME in solib.c:solib_map_sections,
        which overwrites the so_name in the so_list struct.
        This of course overwrites the composite name. Two
        consequences:
           - "info sharedlibrary" does not show the composite name
             (missing piece of info)
           - the bfd_open hook is missing the composite name also,
             at least until the so_list is passed instead of just
             the pathname.

    Some of the options I have considered:

       - For the "info sharedlibrary" command, enhance the solib
         vector to provide a "so_name" function that, given a
         struct so_list, returns the solib's "print" name.
         solib-aix would access the lm_info to do so. Or rather
         than returning a string, with its usual memory management
         issues, would print it to a stream?

       - Stop encoding the archive-name+member-name into the
         so_list so_name. We need to be careful there, because
         even if we were able to print the solib name correct
         (as discussed above), I think we have some mechanism
         in the solib code that deals with duplicates using
         the so_name. Since it is very common for two archives
         to have members with the same name, we need to make
         sure that this identification mechanism keeps working.
         The so_name compositing helps achieving that.

       - Issue when printing the objfile name: This one is going
         to be messy, I think. I was thinking of defining the
         concept of an objfile's "print" name, which could work
         as follow: If the objfile's bfd has a my_archive, then
         it would be "<my_archive->name>(<objfile->name>)";
         otherwise, the usual "<objfile->name>". I don't think
         we really need to address that issue, especially not
         right now.

  - xcoff_symfile_offsets was greatly simplified, and in fact could be
    entirely replaced by default_symfile_offsets, if it wasn't for
    some code which defaults some section indices in the objfile
    to zero even when the section actually does not exist. I could
    probably work with that because this seems to only affect
    the rodata sect index in practice, and that section does not
    exist on AIX (yet). But I think that's taking a chance.
    The code that does that was added a very long time ago, and
    was probably meant for ELF. For now, I've added code in
    xcoff_symfile_offsets to just call default_symfile_offsets
    followed by the undoing of the sect index zero'ing. Fine
    for now, but something we might want to look at eventually?

Some questions I asked myself (with some answers):

  - Because of the dependency on the system ldr.h file, there is
    a temporary dependency of solib-aix on rs6000-nat, which we
    do not want, especially since rs6000-aix-tdep also depends
    on solib-aix (to get the TOC value in order to setup inferior
    function calls). I am planning on fixing this by adding
    a new xfer object.

  - The changes in rs6000-aix-tdep.c are adjustments to the new
    method towards getting the TOC value. It calls a function
    in solib-aix rather than relying on a hook setup by the -nat
    code. This means that it can also be made to work when configured
    as a cross debugger.

  - The new code in rs6000-nat is a little hard to read. It basically
    just gets the raw ld_info data from ptrace (PT_LDINFO), and
    transforms it into a structured VEC. This will likely change
    when integrated into the xfer_memory framework, but the idea
    will remain the same.

  - I am also planning on caching the ld_info data, to avoid
    re-fetching the data multiple times when not necessary.
    But I think that this will be the subject of a followup patch,
    to avoid mixing everything in the same bag.

  - xcoffsolib.[hc] is now gone, as it is redundant with the solib.c
    code;

  - We will need to update the ARI in order to forbid the macros
    finally deleted.

Voila voila,
-- 
Joel

PS: The patch is against AdaCore's HEAD, which is off a fairly recent
    set of sources, but modified to contain AdaCore's changes too.
    This was simpler for me to allow for testing, and may not apply
    cleanly to current FSF HEAD. Sorry if that is the case.