PATCH: PR ld/13302: IRELATIVE relocation should come last

[H.J. Lu]

On Tue, Mar 6, 2012 at 1:00 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> On Fri, Oct 21, 2011 at 12:49:12PM -0700, H.J. Lu wrote:
>> Here is a patch to replace IRELATIVE with RELATIVE in .rel.dyn. We now
>> use PLT entry, instead of real function, for non-GOT reference to locally
>> defined IFUNC symbols.
>
>> --- a/bfd/ChangeLog
>> +++ b/bfd/ChangeLog
>> @@ -1,3 +1,12 @@
>> +2011-10-21  H.J. Lu  <hongjiu.lu@intel.com>
>> +
>> +     PR ld/13302
>> +     * elf32-i386.c (elf_i386_relocate_section): Replace
>> +     R_386_IRELATIVE with R_386_RELATIVE.
>> +
>> +     * elf64-x86-64.c (elf_x86_64_relocate_section): Replace
>> +     R_X86_64_IRELATIVE with R_X86_64_RELATIVE.
>
> This patch apparently broke prelink, which on x86_64 expects that the added
> for R_X86_64_RELATIVE is always applied (i.e. the same number is in both
> addend and memory and it normally is for all other kinds of
> R_X86_64_RELATIVE, just not this spot now).
>
>> --- a/bfd/elf64-x86-64.c
>> +++ b/bfd/elf64-x86-64.c
>> @@ -3161,6 +3161,7 @@ elf_x86_64_relocate_section (bfd *output_bfd,
>>               {
>>                 Elf_Internal_Rela outrel;
>>                 asection *sreloc;
>> +               bfd_boolean relocate;
>>
>>                 /* Need a dynamic relocation to get the real function
>>                    address.  */
>> @@ -3180,15 +3181,15 @@ elf_x86_64_relocate_section (bfd *output_bfd,
>>                     || info->executable)
>>                   {
>>                     /* This symbol is resolved locally.  */
>> -                   outrel.r_info = htab->r_info (0, R_X86_64_IRELATIVE);
>> -                   outrel.r_addend = (h->root.u.def.value
>> -                                      + h->root.u.def.section->output_section->vma
>> -                                      + h->root.u.def.section->output_offset);
>> +                   outrel.r_info = htab->r_info (0, R_X86_64_RELATIVE);
>> +                   outrel.r_addend = relocation;
>> +                   relocate = FALSE;
>>                   }
>>                 else
>>                   {
>>                     outrel.r_info = htab->r_info (h->dynindx, r_type);
>>                     outrel.r_addend = 0;
>> +                   relocate = FALSE;
>>                   }
>>
>>                 sreloc = htab->elf.irelifunc;
>> @@ -3199,7 +3200,8 @@ elf_x86_64_relocate_section (bfd *output_bfd,
>>                    we need to include the symbol value so that it
>>                    becomes an addend for the dynamic reloc.  For an
>>                    internal symbol, we have updated addend.  */
>> -               continue;
>> +               if (! relocate)
>> +                 continue;
>>               }
>>             /* FALLTHROUGH */
>>           case R_X86_64_PC32:
>
> The above looks like typo to me, if relocate was intentionally
> always false, why add the boolean at all?  Here is a fix.  Ok for trunk?
>
> 2012-03-06  Jakub Jelinek  <jakub@redhat.com>
>
>        * elf64-x86-64.c (elf_x86_64_relocate_section): For R_X86_64_RELATIVE
>        set relocate to TRUE.
>
> --- bfd/elf64-x86-64.c.jj       2012-02-08 22:12:43.000000000 +0100
> +++ bfd/elf64-x86-64.c  2012-03-06 09:52:16.760752214 +0100
> @@ -1,6 +1,6 @@
>  /* X86-64 specific support for ELF
>    Copyright 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009,
> -   2010, 2011
> +   2010, 2011, 2012
>    Free Software Foundation, Inc.
>    Contributed by Jan Hubicka <jh@suse.cz>.
>
> @@ -3178,7 +3178,7 @@ elf_x86_64_relocate_section (bfd *output
>                      /* This symbol is resolved locally.  */
>                      outrel.r_info = htab->r_info (0, R_X86_64_RELATIVE);
>                      outrel.r_addend = relocation;
> -                     relocate = FALSE;
> +                     relocate = TRUE;
>                    }
>                  else
>                    {
>
>
>        Jakub

It is a typo.  OK for trunk.  I think it should also be applied on 2.22 branch.

Thanks.


-- 
H.J.