On Fri, 24 Aug 2007 20:20:28 +0200, Daniel Jacobowitz wrote:
> On Fri, Aug 24, 2007 at 08:04:50PM +0200, Jan Kratochvil wrote:
...
> > This part may be questionable:
> > -  debugfile = find_separate_debug_file (objfile); 
> > +  /* If the file has its own symbol tables it has no separate debug info.  */
> > +  if (objfile->psymtabs == NULL)
> > +    debugfile = find_separate_debug_file (objfile);
...
> > Not sure if `PSYMTABS == NULL' is the right condition.  According to my
> > experiments on GNU/Linux it should be correct.
> 
> In some cases a library will be left with .symtab but not .debug_info;
> sometimes elfutils does this, also sometimes it is done deliberately
> for ld.so / libpthread.so so that the separate debug file is not
> necessary for minimal debugging to work.

You are right about the content of `/lib64/ld-linux-x86-64.so.2'.
Still in such case both `.dynsym' and `.symtab' gets read into MSYMBOLS but
PSYMTABS and SYMTABS are left forever NULL for the main executable (as there is
no .debug_info).

While I would still expect the former patch was right I changed it according to
your advice.  It is definitely the safe way but with a tiny performance hit:

The reason of my check was that build-id (contrary to the debug-link) is always
present there, even for locally built unsplit files.  With the patch attached
below GDB will try to open a separate debug file in /usr/lib/debug/.debug-id/
even for the unsplit files.  This step did not happen with my former patch.


> We should either pass a flag down so that we know we're already
> looking at the separate debug file (more efficient, saves us the
> second search, more work to implement)

Chose this way as the intention of the whole patch is performance.
(Aware OBJF_DEBUG_FILE is not an OBJFILE flag - just an additional
SYMBOL_FILE_ADD_WITH_ADDRS_OR_OFFSETS parameter - but it may get one day useful
for later retrieval from OBJFILE.)


> > +/* Locate NT_GNU_BUILD_ID and return its content.
> > +   Separate debuginfo files have corrupted PHDR but SHDR is correct there.  */
> > +
> > +static struct build_id *
> > +build_id_bfd_shdr_get (bfd *abfd)
> > +{
> > +  struct build_id *retval;
> > +
> > +  if (!bfd_check_format (abfd, bfd_object)
> > +      || bfd_get_flavour (abfd) != bfd_target_elf_flavour
> > +      || elf_tdata (abfd)->build_id == NULL)
> > +    return NULL;
> > +
> > +  retval = xmalloc (sizeof *retval - 1 + elf_tdata (abfd)->build_id_size);
> > +  retval->size = elf_tdata (abfd)->build_id_size;
> > +  memcpy (retval->data, elf_tdata (abfd)->build_id, retval->size);
> > +
> > +  return retval;
> > +}
> 
> This function isn't doing anything with program headers or section
> headers now, maybe it should be renamed (and the comment doesn't need
> to talk about corrupted phdrs in that case).

Not agreed as the PHDR<->SHDR difference is important (there are issues about
it for readelf/eu-readelf - whether PHDR or SHDR should be chosen there).
But changed according to your advice.


> I believe the manual needs an update to mention the new search path.

I see now, sorry, implemented.



Thanks,
Jan