[Bug gold/15662] New: gold (powerpc) internal error in do_relax() at gold/output.h:436

[jingyuuiuc at gmail dot com]

http://sourceware.org/bugzilla/show_bug.cgi?id=15662

            Bug ID: 15662
           Summary: gold (powerpc) internal error in do_relax() at
                    gold/output.h:436
           Product: binutils
           Version: unspecified
            Status: NEW
          Severity: normal
          Priority: P2
         Component: gold
          Assignee: ian at airs dot com
          Reporter: jingyuuiuc at gmail dot com
                CC: ccoutant at google dot com

I found a few link time failures, reporting

internal error in set_current_data_size_for_child, at
/usr/local/home/jingyu/opensource/binutils_trunk/src/gold/output.h:436

I am using top of trunk gold.

I investigated the failure and found that the error happens when
set_current_data_size_for_child is called through
    this->rel_->add_relative(elfcpp::R_POWERPC_RELATIVE, this, off, to);
through
   Target_powerpc<size, big_endian>::do_relax().

I notice that the data size for the failing object has been changed a
few times during relaxation. Every time before it is updated,
this->is_data_size_valid_ will be set to false. And after data size is
updated, is_data_size_valid_ will be set to true. However, when the
data size is updated through add_relative, the flag
is_data_size_valid_ is not set to false beforehand, which triggers the
assertion failure. I think it is a bug.

To verify my thought, I made a patch on powerpc.cc, though I am not sure it is
the proper way to fix. With the patch, the failing tests pass linking.

Index: powerpc.cc
===================================================================
RCS file: /cvs/src/src/gold/powerpc.cc,v
retrieving revision 1.91
diff -r1.91 powerpc.cc
2631a2632
>       this->brlt_section_->reset_rel_data_size();
2638a2640
>       this->brlt_section_->finalize_rel_data_size();
3015a3018,3029
>   void
>   reset_rel_data_size()
>   {
>     this->rel_->reset_data_size();
>   }
>
>   void
>   finalize_rel_data_size()
>   {
>     this->rel_->finalize_data_size();
>   }
>

Is the patch the proper way to fix the bug? Thanks!

-- 
You are receiving this mail because:
You are on the CC list for the bug.