Re: TLS tpoff alignment question

[Alan Modra <amodra at gmail dot com>]

On Wed, May 21, 2014 at 10:59:41AM +0100, Will Newton wrote:
> Lots of bfd beckends have code similar to this, from aarch64, to
> handle TLS relocations:

Right, the ones that didn't learn from previous mistakes.  :)

In this case, the "mistake" is to define the thread pointer as the
address of the TCB and allocate static TLS blocks after the TCB.

> static bfd_vma
> tpoff_base (struct bfd_link_info *info)
> {
>   struct elf_link_hash_table *htab = elf_hash_table (info);
> 
>   /* If tls_sec is NULL, we should have signalled an error already.  */
>   if (htab->tls_sec == NULL)
>     return 0;
> 
>   bfd_vma base = align_power ((bfd_vma) TCB_SIZE,
>                   htab->tls_sec->alignment_power);
>   return htab->tls_sec->vma - base;
> }

The comment on the function says "address which should be subtracted".
So, you'll be calculating

  STT_TLS_sym_val - tls_sec->vma + base

for the offset from the thread pointer to the tls sym.

The adjustment using align_power is necessary to keep the TLS blocks
aligned in memory.  You'll have

____________________________ .. ______
| TCB | pad     | static TLS         |
~~~~~~~~~~~~~~~~~~~~~~~~~~~~ .. ~~~~~~
 ^               ^              ^
 |               |              |
tp              dtv1           dtvn

Now if tp was at the *end* of the TCB, as used by powerpc, then there
is no need for the "base" adjustment.  It doesn't really save you
anything, just hides the pad-for-alignment in the run time TLS
support..

-- 
Alan Modra
Australia Development Lab, IBM