This is the mail archive of the mailing list for the binutils project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] HPPA/PR16082 : Remove version information on hidden symbols

On 10/25/2013 03:54 AM, Guy Martin wrote:
> Hi,
> We've run into a bug on Gentoo/hppa when linking :
> It turns out that the generated .gnu.version section is invalid
> because the version information of _GLOBAL_OFFSET_TABLE_ from
> isn't removed when making the symbol local in
> elf32_hppa_hide_symbol().
> The attached patch fixes this. The check-ld test suite doesn't
> show any regression.

Is there an existing test in the testsuite that checks for this?

> diff --git a/bfd/ChangeLog b/bfd/ChangeLog
> index 8977400..3923a7c 100644
> --- a/bfd/ChangeLog
> +++ b/bfd/ChangeLog
> @@ -1,3 +1,8 @@
> +2013-10-24  Guy Martin <>
> +
> +	* elf32-hppa.c (elf32_hppa_hide_symbol): Remove old version
> +	information when forcing a symbol to be local.
> +


>  2013-10-18  Hans-Peter Nilsson  <>
>  	* elf32-cris.c (cris_elf_check_relocs): Don't assume
> diff --git a/bfd/elf32-hppa.c b/bfd/elf32-hppa.c
> index dfffbcb..5da3b5a 100644
> --- a/bfd/elf32-hppa.c
> +++ b/bfd/elf32-hppa.c
> @@ -1771,6 +1771,9 @@ elf32_hppa_hide_symbol (struct bfd_link_info *info,
>  	  _bfd_elf_strtab_delref (elf_hash_table (info)->dynstr,
>  				  eh->dynstr_index);
>  	}
> +      /* Remove version information from hidden symbol */
> +      eh->verinfo.verdef = NULL;
> +      eh->verinfo.vertree = NULL;

Why doesn't the generic elflink.c (_bfd_elf_link_hash_hide_symbol)
need to do this?

It could be that symbols we are about to hide never make it this
far in other targets and aren't even added to the list.

>      }
>    /* STT_GNU_IFUNC symbol must go through PLT.  */

>From a first principles perspective a hidden symbol will have
been converted to a local symbol and will never require version
information. I see no reason that verdef or vertree need be preserved.

There is quite a lot of code which uses verdef == NULL to validate
that the symbol has no version information so it seems valid to set
it to NULL here.

One is under the assumption here that whatever called
elf32_hppa_hide_symbol has already removed the version definition
from the appropriate hashes.

I would be OK with this patch if you can show that gcc (and a couple
of languages) and glibc all build without failure using the new

Can you try using the new bintuils to build gcc and glibc?


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]