Re: Make check failure: ARM in ld-bootstrap/bootstrap.exp...

[scottb <scottb@netwinder.org>]

Ok, thanks for the tips.  I'll investigate next week and post my 
findings.

Ian Lance Taylor wrote:
> 
>    Date: Fri, 3 Sep 1999 18:04:26 -0400
>    From: Roland McGrath <roland@frob.com>
> 
>    > There is no immediately obvious linker bug here.  The error messages
>    > mean that a 24 bit PC relative relocation is being used to call a
>    > symbol which is more than 24 bits away.  You, or somebody, will need
>    > to investigate further to see whether the relocation should success,
>    > or whether the linker is using some bogus linker script which places
>    > some code at a bad address, or whether the libraries should not be
>    > assuming that a 24 bit PC relative relocation is always enough.
> 
>    Given the context of those messages, I suspect that the issue here is that
>    those symbols (_pthread_*) are weak references that found no definition (in
>    the absence of -lpthread) and so resolve to value zero.  If the load
>    address is above 0x01000000 (I think it's 0x02000000 on ARM), then the PC
>    in question is more than 24 bits away from zero.
> 
>    This doesn't come up in the shared library, since the relocation is of a
>    32-bit element in the GOT rather than a 26-bit field in a jump instruction.
> 
> That makes sense.
> 
>    Probably you need to resort to something like this in the libc sources:
> 
>            rettype (*volatile foo_ptr) (argtypes) = &foo;
>            (*foo_ptr) (args...);
> 
>    so that the relocation that's produced is for 32-bit immediate data.
>    I don't know the ARM details to suggest in further detail.
> 
> We can simply change the ARM linker so that it does not complain about
> an overflow in a PC relative relocation to an undefined weak symbol.
> That's what we did for the PowerPC:
> 
> 1999-06-04  Franz Sirl <Franz.Sirl-kernel@lauterbach.com>
> 
>         * elf32-ppc.c (ppc_elf_relocate_section): Don't barf on out of
>         range undefweak symbols.
> 
>       else if (r == bfd_reloc_overflow)
>         {
>           const char *name;
> 
>           if (h != NULL)
>             {
>               if (h->root.type == bfd_link_hash_undefweak
>                   && howto->pc_relative)
>                 {
>                   /* Assume this is a call protected by other code that
>                      detect the symbol is undefined.  If this is the case,
>                      we can safely ignore the overflow.  If not, the
>                      program is hosed anyway, and a little warning isn't
>                      going to help.  */
> 
>                   continue;
>                 }
> 
>               name = h->root.root.string;
>             }
> 
> Ian