[PATCH, cygwin] Internal error when loading dlls without .data section.

[Pedro Alves]

Hi all,

Now that shlib tests run on cygwin (with binutils-cvs),  the 
gdb.base/shreloc.exp
test caught a dll loading bug.  There is an attempt to get to 
SECT_OFF_DATA (objfile),
in an objfile without the sect_index_data set.

Here is what happens:

The dlls built from shreloc1.c and shreloc2.c don't have .data sections.
Also, they are both compiled into the same preferred load address using
-Wl,--image-base,0x04000000.  Since only one can load at that image-base,
the other dll will be relocated.

When the first dll is loaded, in symfile.c:init_objfile_sect_indices 
there is code that
sets objfile->sect_index_data to 0 if no section was found.

symfile.c:init_objfile_sect_indices:

 for (i = 0; i < objfile->num_sections; i++)
  {
    if (ANOFFSET (objfile->section_offsets, i) != 0)
      {
        break;
      }
  }
  if (i == objfile->num_sections)
    {
      if (objfile->sect_index_text == -1)
        objfile->sect_index_text = 0;
      if (objfile->sect_index_data == -1)
        objfile->sect_index_data = 0;
      if (objfile->sect_index_bss == -1)
        objfile->sect_index_bss = 0;
      if (objfile->sect_index_rodata == -1)
        objfile->sect_index_rodata = 0;
    }


But, when the second dll is loaded, and since it was relocated,
the 'ANOFFSET (objfile->section_offsets, i) != 0' will be true, thus 
skipping
the initialization to zero afterwards, thus leaving 
objfile->sect_index_data with
the default of -1.  The internal_error then is called on 
coffread.c:coff_symtab_read,
while reading the minsyms, where there is code like this:

 else if (sec == SECT_OFF_DATA (objfile))
      {
             ms_type =
                 cs->c_sclass == C_EXT || cs->c_sclass == C_THUMBEXT ?
                 mst_data : mst_file_data;
       }

The attached patch fixes it by not depending on the SECT_OFF_* macros to 
determine
the minimal_symbol_type of the minsymbol, but using bfd_section flags.

In the process I noted that bss symbols where recorded with mst_data or 
mst_file_data
types.  Fixed that too.

The shreloc.exp needs a little tweak since passing  
-Wl,--image-base,0x04000000 on
additional_flags when building the .o files ends up giving the 
unfiltered warning:
  "linker input file unused because linking not done"
which causes the test to end up being untested.

Tested on i686-pc-cygwin, where it fixes the crash, makes shreloc.exp 
pass perfectly, and
introduces no new regressions.

Please review and commit,

Cheers,
Pedro Alves

----
gdb/

2006-12-21  Pedro Alves  <pedro_alves@portugalmail.pt>

        * coffread.c (cs_to_section): If bfd_section is found, always
        return it's section index.
        (coff_symtab_read): Determine the minimal_symbol_type using the
        bfd_section flags.

gdb/testsuite/

2006-12-21  Pedro Alves  <pedro_alves@portugalmail.pt>

        * gdb.base/shreloc.exp: Use ldflags instead of additional_flags to
        pass --image-base to linker.

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: dllfix.diff
URL: <http://sourceware.org/pipermail/gdb-patches/attachments/20061221/2e771263/attachment.ksh>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: shreloc.exp.diff
URL: <http://sourceware.org/pipermail/gdb-patches/attachments/20061221/2e771263/attachment-0001.ksh>