Re: [PATCH, bfin] Clear relocs for removed entries in .eh_frame section

["Jie Zhang" <jzhang918 at gmail dot com>]

On 4/26/06, Alan Modra <amodra@bigpond.net.au> wrote:
> On Wed, Apr 26, 2006 at 11:45:19AM +0800, Jie Zhang wrote:
> > 1. I saw this in elflink.c:elf_link_input_bfd ()
> >
> >                     /* This is a reloc for a deleted entry or somesuch.
> >                        Turn it into an R_*_NONE reloc, at the same
> >                        offset as the last reloc.  elf_eh_frame.c and
> >                        elf_bfd_discard_info rely on reloc offsets
> >                        being ordered.  */
> >
> > In my patch I just set the offset to zero. It seems works. I don't
> > know if we need do the same.
> >
> > Alan: The comment was from you. Can you help me understand it?
>
> bfd_elf_discard_info indirectly calls bfd_elf_reloc_symbol_deleted_p,
> which scans relocs looking for a particular offset.
>
>   for (; rcookie->rel < rcookie->relend; rcookie->rel++)
>     {
>       unsigned long r_symndx;
>
>       if (! rcookie->bad_symtab)
>         if (rcookie->rel->r_offset > offset)
>           return FALSE;
>       if (rcookie->rel->r_offset != offset)
>         continue;
>
> There are some macros in elf-eh-frame.c, ENSURE_NO_RELOCS and
> SKIP_RELOCS that similarly test r_offset.  The above code won't work in
> general if relocs are unordered.  However, I see that the way the code
> is written, setting r_offset to zero is OK.  You just need to ensure
> r_offset is not set to some larger value than r_offset in following
> relocs.
>
Thanks for explain this. Now I feel safe to set the offset of deleted
reloc to 0.

> > 2. In bfinfdpic_relocate_section () there are several calls to
> > _bfd_elf_section_offset () before calling _bfinfdpic_add_dyn_reloc ().
> > Only in one place the return value of _bfd_elf_section_offset () is
> > checked in my patch. It seems works. But I'm not sure if we need add
> > such check for all these calls.
>
> If _bfd_elf_section_offset can return -1 or -2 then you need to add the
> check.  If it cannot, eg. because the particular reloc type is not used
> in eh_frame or stabs section, then you don't need the check.  Of course,
> if you don't need the check, then you also don't need to call
> _bfd_elf_section_offset.  :-)
>
I'll review the related code in Blackfin FDPIC to see if there are
more cases need check. I also would like to hear from Alex before
doing any checkin.

> > 3. Is it possible to not emit the reloc instead of set its type to R_*_NONE?
>
> It's too late.  You have already allocated space for the dynamic relocs.
>
OK.

Thanks,
Jie