Re: ICF on PowerPC Bug

[Sriraman Tallam <tmsriram at google dot com>]

On Tue, Apr 1, 2014 at 11:54 PM, Alan Modra <amodra@gmail.com> wrote:
> On Tue, Apr 01, 2014 at 03:06:37PM -0700, Sriraman Tallam wrote:
>> because fold2 is folded onto fold1.    The only way to differentiate
>> fold1 and fold2 which have the same object code is via the relocation
>> type and the Info value is different.  However, with powerpc this is
>> rewritten here in icf.cc:
>>
>> // Look through function descriptors
>> parameters->target().function_location(&loc);
>>  if (loc.shndx != it_v->second)
>> {
>>  it_v->second = loc.shndx;
>>  // Modify symvalue/addend to the code entry.
>>  it_a->first = loc.offset;
>>  it_a->second = 0;
>> }
>>
>> I am not sure how to fix it.
>
> The following patch fixes the problem, but did you really mean to copy
> all the vectors here?

Yikes!, no did not intend that. I will make these references and send a patch.

Thanks
Sri

>
>       Icf::Sections_reachable_info v =
>         (it_reloc_info_list->second).section_info;
>       // Stores the information of the symbol pointed to by the reloc.
>       Icf::Symbol_info s = (it_reloc_info_list->second).symbol_info;
>       // Stores the addend and the symbol value.
>       Icf::Addend_info a = (it_reloc_info_list->second).addend_info;
>       // Stores the offset of the reloc.
>       Icf::Offset_info o = (it_reloc_info_list->second).offset_info;
>       Icf::Reloc_addend_size_info reloc_addend_size_info =
>         (it_reloc_info_list->second).reloc_addend_size_info;
>
>
> diff --git a/gold/icf.cc b/gold/icf.cc
> index f30eb41..920514c 100644
> --- a/gold/icf.cc
> +++ b/gold/icf.cc
> @@ -288,8 +288,7 @@ get_section_contents(bool first_iteration,
>
>        for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size)
>          {
> -         if (first_iteration
> -             && it_v->first != NULL)
> +         if (it_v->first != NULL)
>             {
>               Symbol_location loc;
>               loc.object = it_v->first;
>
> --
> Alan Modra
> Australia Development Lab, IBM