Re: Another MIPS multigot patch

[Daniel Jacobowitz <drow@mvista.com>]

On Sat, Nov 22, 2003 at 07:14:06PM +0000, Richard Sandiford wrote:
> Daniel Jacobowitz <drow@mvista.com> writes:
> > I'm using GCC 3.3.1, FYI.  So it may be fixed in HEAD which I'm not
> > really set up to try at the moment.  The relocation comes from:
> >         lw      $2,flag_dump_unnumbered
> 
> Ah, never mind then. ;)  I assumed this was gcc explicit-reloc output.
> 
> > We're in mips_elf_global_got_index, looking at this symbol.  It has
> > h->dynindx of around 3500, and global_got_dynindx is 5.  g->local_gotno
> > is about 5000.  So we have an overflow; that's more than 8K 8-byte
> > entries.  Are you saying that the fact that there are R_MIPS_GOT_PAGE
> > entries against flag_dump_unnumbered using the primary GOT means it
> > should have a lower dynindx?
> 
> Well, there are two ways we can handle the relocation:
> 
>     1. Decay to R_MIPS_GOT_DISP.  In that case, yes, the index
>        should be lower.
> 
>     2. Use a local page entry.
> 
> (2) should be possible in this case, and looking at the code,
> it does seem like we try to support it.
> 
> It's difficult to be sure without access to the test case, so this is
> probably completely wrong.  But I'm guessing the problem is:

I can package up the testcase if you want.  It's a bit large.  In the
mean time, I'll look at this more tomorrow or Monday but:

>     (a) _bfd_mips_elf_check_relocs assumes print-rtl.c can use a local
>         page entry for GOT_PAGE/GOT_OFST accesses to flag_dump_unnumbered.
>         f_d_u is therefore not entered into print-rtl's got_entries.
> 
>     (b) toplev uses GOT_DISP, so we end up needing a global GOT
>         entry for f_d_u.

000000001028  00a900000012 R_MIPS_64         0000000000000000 flag_dump_unnumbered + 0
                    Type2: R_MIPS_NONE      
                    Type3: R_MIPS_NONE      

So, close enough for our purposes.

>     (c) Because of (a), no bfd that uses the primary GOT seems to need
>         an entry for f_d_u.  We therefore give it a higher index than
>         symbols that _do_ need an entry.
> 
>     (d) Because of (c), mips_elf_calculate_relocation assumes that f_d_u
>         is !local_p, and it therefore decays the GOT_PAGE reloc into a
>         GOT_DISP.  So print-rtl uses f_d_u's GOT entry after all.
> 
> If so, then IMO the bug is with (d).  We can still use local pages
> for print-rtl, despite the entry in .dynsym.

This all makes sense to me.  I'll think about what the local_p
condition really ought to be.  This:
    case R_MIPS_GOT_PAGE:
    case R_MIPS_GOT_OFST:
      /* If this symbol got a global GOT entry, we have to decay
         GOT_PAGE/GOT_OFST to GOT_DISP/addend.  */
      local_p = local_p || ! h
        || (h->root.dynindx
            < mips_elf_get_global_gotsym_index (elf_hash_table (info)
needs to become... I'm not quite sure what.  Is it reasonable to use
SYMBOL_REFERENCES_LOCAL here?

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer