Re: [PATCH v2] Add support for the --readnever command-line option (DWARF only)

[Pedro Alves <palves@redhat.com>]

On 11/24/2017 08:26 PM, Sergio Durigan Junior wrote:
> On Friday, November 24 2017, Pedro Alves wrote:
> 
>> On 11/24/2017 04:54 AM, Sergio Durigan Junior wrote:
>>> On Thursday, November 23 2017, Pedro Alves wrote:
>>>
>>>> On 11/23/2017 05:21 PM, Sergio Durigan Junior wrote:
>>>>
>>>>>>> +This option is currently limited to debug information in DWARF format.
>>>>>>> +For all other format, this option has no effect.
>>>>>>
>>>>>> How hard would it be to just make it work?  There's only stabs and mdebug
>>>>>> left, I think?  There should be a single a function somewhere that we can
>>>>>> add an early return.  And then we don't need to document this limitation...
>>>>>>
>>>>>> For example, in elf_symfile_read, we could just skip the elf_locate_sections
>>>>>> call.  In coffread.c we could skip reading stabs right after 
>>>>>>   bfd_map_over_sections (abfd, coff_locate_sections....);
>>>>>>
>>>>>> Looking for:
>>>>>>
>>>>>>  $ grep -h "^[a-z]*_build_psymtabs" gdb/
>>>>>>  coffstab_build_psymtabs (struct objfile *objfile,
>>>>>>  elfstab_build_psymtabs (struct objfile *objfile, asection *stabsect,
>>>>>>  stabsect_build_psymtabs (struct objfile *objfile, char *stab_name,
>>>>>>  mdebug_build_psymtabs (minimal_symbol_reader &reader,
>>>>>>  elfmdebug_build_psymtabs (struct objfile *objfile,
>>>>>>
>>>>>> finds all the relevant places.
>>>>>>
>>>>>> Maybe it wouldn't be that hard to make this be an objfile flag
>>>>>> afterall (like OBJF_READNOW is).  That'd make it possible
>>>>>> to add the location "-readnever" counterpart switch to add-symbol-file
>>>>>> too, BTW:
>>>>
>>>> I meant "logical" instead of "location".  I was staring at
>>>> gdb/location.c at that time.  :-P
>>>>
>>>>>>
>>>>>>  symfile.c:        if (strcmp (arg, "-readnow") == 0)
>>>>>>  symfile.c:        else if (strcmp (arg, "-readnow") == 0)
>>>>>
>>>>> Hm, I'll look into this.  Just to make it clear: the idea is to have
>>>>> both a --readnever global option and also a OBJF_READNEVER specific to
>>>>> each objfile?
>>>>
>>>> Sure, the idea is to do something similar to what's done for --readnow.
>>>
>>> Sorry, but I guess I need a few more details on this.
>>>
>>> The way I understand the code at elf_symfile_read, the very first thing
>>> to do would be to check if OBJF_READNEVER is set and return early if it
>>> is.  But it seems that you're proposing something a bit different when
>>> you say that we should "... just skip the elf_locate_sections call."  It
>>> doesn't seem to me that is worth continuing on that function if
>>> OBJF_READNEVER is present.
>>
>> No, you can't return early the very first thing, because
>>  --readnever is supposed to skip _debug_ info, not ELF/minimal symbols...
>> So the "return early" would have to be _after_ the
>> elf_read_minimal_symbols call:
> 
> Hm, OK, it makes sense and I confess I thought "why is Pedro mentioning
> elf_symfile_read if the feature is about skipping DWARF, not ELF?".
> 
>>  static void
>>  elf_symfile_read (struct objfile *objfile, symfile_add_flags symfile_flags)
>>  {
>>    bfd *abfd = objfile->obfd;
>>    struct elfinfo ei;
>>
>>    memset ((char *) &ei, 0, sizeof (ei));
>>    bfd_map_over_sections (abfd, elf_locate_sections, (void *) & ei);
>>  
>>    elf_read_minimal_symbols (objfile, symfile_flags, &ei);
>>
>> I don't know whether we can reorder that.  Maybe we can.
>>
>> When I looked at this quickly yesterday, I saw that elf_location_sections
>> is what finds the stabs and mdebug sections:
>>
>> static void
>> elf_locate_sections (bfd *ignore_abfd, asection *sectp, void *eip)
>> {
>>   struct elfinfo *ei;
>>
>>   ei = (struct elfinfo *) eip;
>>   if (strcmp (sectp->name, ".stab") == 0)
>>     {
>>       ei->stabsect = sectp;
>>     }
>>   else if (strcmp (sectp->name, ".mdebug") == 0)
>>     {
>>       ei->mdebugsect = sectp;
>>     }
>> }
>>
>> and it seemed to be that skipping the section location would make
>> the parts of  elf_symfile_read that actually read the symbols
>> be no-ops, because the stabsect/mdebusect pointers would be NULL.
>>
>> But if returning early or something else works, that's fine.
> 
> OK, thanks for clarifying.
> 
>>> As for the *_build_psymtabs functions, I am doing exactly that: if
>>> objfile->flags contains OBJF_READNEVER, then just return and do nothing.
>>
>> Sure, that should work too.  It's just the difference between
>> skipping checking whether debug info is available (skipping before
>> calling into those), vs letting gdb do the work to figure out whether
>> debug info is available, but then ignore it.
>> The grep for "*_build_psymtabs" was intended as a pointer to find
>> what the relevant code is, including to look at the code that
>> is calling those functions, see if there's something to be done there.
> 
> I think it makes more sense, logically speaking, to not mess with
> elf_symfile_read and instead modify the *_build_psymtabs functions.

Yet, somehow that logic didn't carry to the DWARF reader
changes?  ;-)  dwarf2_has_info is the place that checks
whether the objfile has DWARF sections, 
see dwarf2_per_objfile::dwarf2_per_objfile.
So for DWARF, the patch is checking for readnever _before_
dwarf2_build_psymtabs is reached, by essentially skipping
the "locate_sections" call.  

A small advantage of checking before is that you can skip a little bit
more work.  See the comment in elf_read_minimal_symbols about skipping
work unless stabs, or this bit here:

  if (info->stabsects)
    {
      if (!info->stabstrsect)
	{
	  error (_("The debugging information in `%s' is corrupted.\nThe "
		   "file has a `.stabs' section, but no `.stabstr' section."),
		 name);
	}
      ...

It's really not a bit deal, but to me it'd be more
consistent have all readers do the same logically.

On another note: there's a symbol_file_add_separate call at the tail
end of elf_symfile_read, where we seem to read separate info.  It seems
like that'd end up reading debug info, with "add-symbol-file -readnever"?
Same in read_symbols, I guess.  Or is OBJF_READNEVER etc. somehow
propagated to the separate objfile?

Thanks,
Pedro Alves