Fix PR12762 using linkonce sections in plugin bfd

[H.J. Lu]

On Thu, Aug 11, 2011 at 6:08 AM, Alan Modra <amodra@gmail.com> wrote:
> This patch undoes HJ's changes to bfd_section_already_linked in
> http://sourceware.org/ml/binutils/2011-07/msg00070.html, and some of
> my later changes.  HJ's patch effectively added another type of entry
> to the already_linked list for plugin comdat symbols.  This works fine
> most of the time, but there are parts of the bfd and linker code that
> really want separate sections for comdat symbols.  For example,
> ldmain.c:multiple_definition just after the FIXME comment.  So
> this patch simply provides dummy linkonce sections for plugin comdat
> symbols.
>
> Ironically, the FIXME advocates doing exactly what we were attempting
> in plugin.c:add_symbols, so in some respects this patch is a backwards
> step.  It also uses more memory.  The reason I didn't apply further
> tweaks like the patch in PR13066 is that bfd_section_already_linked
> was getting quite messy.  If we are going to go the way this FIXME
> suggests then we really aught to do so for all symbols, not just
> plugin symbols.  I'd like to see that happen but don't have time at
> the moment to fix all the fallout.
>

> Index: ld/plugin.c
> ===================================================================
> RCS file: /cvs/src/src/ld/plugin.c,v
> retrieving revision 1.38
> diff -u -p -r1.38 plugin.c
> --- ld/plugin.c 9 Aug 2011 09:27:34 -0000       1.38
> +++ ld/plugin.c 11 Aug 2011 06:24:21 -0000
> @@ -32,7 +32,6 @@
>  #include "plugin.h"
>  #include "plugin-api.h"
>  #include "elf-bfd.h"
> -#include "libbfd.h"
>  #if !defined (HAVE_DLFCN_H) && defined (HAVE_WINDOWS_H)
>  #include <windows.h>
>  #endif
> @@ -240,7 +239,7 @@ plugin_get_ir_dummy_bfd (const char *nam
>        {
>          flagword flags;
>
> -         /* Create sections to own the symbols.  */
> +         /* Create section to own the symbols.  */
>          flags = (SEC_CODE | SEC_HAS_CONTENTS | SEC_READONLY
>                   | SEC_ALLOC | SEC_LOAD | SEC_KEEP | SEC_EXCLUDE);
>          if (bfd_make_section_anyway_with_flags (abfd, ".text", flags))
> @@ -285,7 +284,25 @@ asymbol_from_plugin_symbol (bfd *abfd, a
>       /* FALLTHRU */
>     case LDPK_DEF:
>       flags |= BSF_GLOBAL;
> -      section = bfd_get_section_by_name (abfd, ".text");
> +      if (ldsym->comdat_key)
> +       {
> +         section = bfd_get_section_by_name (abfd, ldsym->comdat_key);
> +         if (section == NULL)
> +           {
> +             flagword sflags;
> +
> +             sflags = (SEC_CODE | SEC_HAS_CONTENTS | SEC_READONLY
> +                       | SEC_ALLOC | SEC_LOAD | SEC_KEEP | SEC_EXCLUDE
> +                       | SEC_LINK_ONCE | SEC_LINK_DUPLICATES_DISCARD);
> +             section = bfd_make_section_anyway_with_flags (abfd,
> +                                                           xstrdup (ldsym->comdat_key),
> +                                                           sflags);
> +             if (section == NULL)
> +               return LDPS_ERR;
> +           }
> +       }
> +      else
> +       section = bfd_get_section_by_name (abfd, ".text");
>       break;
>

This is the first approach I tried and failed with GCC 4.7 LTO
tests on Linux/x86-64 since adding new sections to dummy
LTO bfd at this stage didn't work very well.

-- 
H.J.