Re: [GOLD] PowerPC TLS fixes

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Sun, Sep 9, 2012 at 7:42 AM, Alan Modra <amodra@gmail.com> wrote:
> This patch corrects TLS GOT entries for powerpc.  We were generating
> unnecessary relocations for symbols that resolve locally, and TLSGD
> module/offset GOT entries were just plain wrong.  TLSGD entries for
> local symbols consist of two words, the first value doesn't matter but
> this word needs a DTPMOD dynamic relocation against any local tls
> symbol or even the zero index symbol.  The second word is the DTPREL
> offset for the local symbol, a constant, but powerpc differs from
> other targets in that the dynamic thread pointer is offset by 0x8000.
> A DTPREL relocation thus evaluates to the symbol value minus 0x8000.
> If we remove all the unnecessary relocations then we also need to
> support TPREL GOT entries.  These also are offset, but by 0x7000.
>
> Offsets from the symbol are not supported by current gold GOT handling
> infrastructure, and trying to use the typical BFD scheme of
> initializing the .got section entries while relocating other sections
> isn't a good idea when section relocation happens in parallel.  I
> originally thought I'd need to crib another bit from Got_entry
> local_sym_index_, but then realized a tls symbol will never need a plt
> offset, so we can use that bit to signify a target dependent offset.
>
> I also tidy the existing add_local_pair_with_rel.  No target uses the
> second reloc, and I can't see any possible use of a DTP relative reloc
> against an output section symbol.  A DTP relative reloc in a tls_index
> (module/offset) entry always refers to the symbol of interest in the
> associated __tls_get_addr call (contrasted to the module reloc which
> ought to be happy with any symbol in the same object).
>
> OK, to apply?
>
>         * output.h (Output_data_got::add_global_tls, add_local_tls,
>         add_local_tls_pair): New functions.
>         (Output_data_got::add_local_pair_with_rel): Remove second
>         reloc param.  Expand comment.
>         (Output_data_got::Got_entry::write): Add got_index param.
>         * output.cc (Output_data_got::add_global_tls, add_local_tls,
>         add_local_tls_pair): New functions.
>         (Output_data_got::Got_entry::write): Handle tls symbols
>         with use_plt_offset_ set specially.
>         (Output_data_got::add_local_pair_with_rel): Only one reloc.
>         (Output_data_got::do_write): Replace iterator with index, pass
>         index to entry write function.
>         * target.h (Target::tls_offset_for_local, tls_offset_for_global,
>         do_tls_offset_for_local, do_tls_offset_for_global): New functions.
>         * arm.cc (Target_arm::Scan::local): Update add_local_pair_with_rel
>         call.
>         * i386.cc (Target_i386::Scan::local): Likewise.
>         * sparc.cc (Target_sparc::Scan::local): Likewise.
>         * x86_64.cc (Target_x86_64::Scan::local): Likewise.
>         * powerpc.cc (Target_powerpc::do_tls_offset_for_local,
>         do_tls_offset_for_global): New functions.
>         (Target_powerpc::Scan::local): Correct TLS relocations and got
>         entry values.
>         (Target_powerpc::Scan::global): Don't emit unnecessary
>         dynamic relocations on TLS GOT entries.
>

>
> Index: gold/output.cc
> ===================================================================
> RCS file: /cvs/src/src/gold/output.cc,v
> retrieving revision 1.171
> diff -u -p -r1.171 output.cc
> --- gold/output.cc      5 Sep 2012 00:34:20 -0000       1.171
> +++ gold/output.cc      9 Sep 2012 11:45:27 -0000
> @@ -1369,7 +1369,9 @@ Output_data_group<size, big_endian>::do_
>
>  template<int size, bool big_endian>
>  void
> -Output_data_got<size, big_endian>::Got_entry::write(unsigned char* pov) const
> +Output_data_got<size, big_endian>::Got_entry::write(
> +    unsigned int got_indx,
> +    unsigned char* pov) const
>  {
>    Valtype val = 0;
>
> @@ -1392,6 +1394,9 @@ Output_data_got<size, big_endian>::Got_e
>             // as small as possible.
>             sgsym = static_cast<Sized_symbol<size>*>(gsym);
>             val = sgsym->value();
> +           if (this->use_plt_offset_ && gsym->type() == elfcpp::STT_TLS)
> +             val += parameters->target().tls_offset_for_global(gsym,
> +                                                               got_indx);
>           }
>        }
>        break;
> @@ -1409,19 +1414,24 @@ Output_data_got<size, big_endian>::Got_e
>
>      default:
>        {
> -       const Relobj* object = this->u_.object;
> +       const Sized_relobj_file<size, big_endian>* object
> +         = static_cast<Sized_relobj_file<size, big_endian>*>(this->u_.object);
>          const unsigned int lsi = this->local_sym_index_;

This breaks x32.  X32 uses 32-bit ELF format with 64-bit GOT entry.

H.J.