[PATCH,plugins,head+2.21] Handle COMDAT groups on ELF.

[H.J. Lu]

On Sat, May 7, 2011 at 7:52 AM, Dave Korn <dave.korn.cygwin@gmail.com> wrote:
>
>    Hi again,
>
>  One of the advantages of the two-stage-link tear-down-and-rebuild approach
> was that we got COMDAT group selection correct for free, as a consequence of
> resetting the already-linked table.  Since we're not doing that now, we have
> to handle COMDAT in the linker plugin code by creating synthetic .group
> sections in the dummy IR BFD, but making sure that these sections don't
> actually win in group selection decisions.
>
>  That's what this patch implements, and on x86_64-linux, with this, we now
> get a completely clean GCC testrun (entire c-family and fortran), which brings
> ld.bfd into parity with ld.gold.  :-)
>
> bfd/ChangeLog:
>
> 2011-05-07  Dave Korn  <dave.korn.cygwin@...
>
>        * elflink.c (elf_link_add_object_symbols): Don't mark symbols undef
>        when the kept section comes from a plugin BFD that we expect to be
>        replaced later.
>        (_bfd_elf_section_already_linked): When there is a duplicate section,
>        but the initial kept section comes from a plugin BFD and the newer
>        duplicate section does not, keep the newer one and discard the initial
>        kept section.
>
> ld/ChangeLog:
>
>        * plugin.c (struct group_sec_hash_table): New derived hash table type.
>        (struct group_sec_hash_entry): Derived entry type for the above.
>        (group_sec_hash_newfunc): Derived ctor for the table entries creates
>        an ELF comdat group section with the given signature.
>        (init_group_sec_hash_table): Derived ctor for the table.
>        (group_sec_hash_lookup): Derived lookup function.
>        (group_sec_hash_traverse): Derived traverse function.
>        (group_sec_hash_table_free): Derived dtor.
>        (struct add_syms_info): Helper struct to wrap args to callback.
>        (get_elf_comdat_section): New helper function for ELF targets.
>        (get_comdat_section): Likewise for all targets.
>        (asymbol_from_plugin_symbol): Take an add_syms_info rather than a
>        bfd pointer argument and adjust usage accordingly.  Generate comdat
>        sections or groups for defined symbols with non-NULL comdat_key.
>        (add_symbols):  Wrap parameters to pass down in an add_syms_info, and
>        handle ELF comdat group section creation.
>
>  Tested on x86_64-unknown-linux-gnu as mentioned.  I'm going to start running
> a GCC testsuite on i686-pc-cygwin, but that could take a couple of days for
> the final results to come in.  If this patch is OK, I'll wait until that's got
> far enough to be sure it's working before I commit.
>
>  Any comments on the implementation, while we're waiting?  Or OK as it is?
>

Do you have a testcase to show it fails without your change?

Thanks.


-- 
H.J.