Re: [PATCH v2] gdb: do not add const sections to the section map

[Andrew Burgess <aburgess@redhat.com>]

Andrew Burgess <aburgess@redhat.com> writes:

> I know there's been a little discussion of this patch on the v1 thread,
> but I wanted to record my thoughts, and here seemed the better place.
>
> Ilya Leoshkevich via Gdb-patches <gdb-patches@sourceware.org> writes:
>
>> From: Ulrich Weigand <ulrich.weigand@de.ibm.com>
>>
>> build_objfile_section_table () creates four synthetic sections, which
>> significantly slow down section map sorting.  This is especially
>> noticeable when debugging JITs that report a lot of objfiles.  Since
>> these sections are not useful for find_pc_section (), do not add them
>> to the section map.
>
> This description could really be fleshed out a little more.
>
> You say "which significantly slow down section map sorting", but I'd
> like this to say which sort(s), in which function(s), otherwise I'm
> expected to either know, or go figure it out myself.
>
> You then jump to say the sections are not useful for "find_pc_section",
> but it's not immediately obvious how that relates to the change you're
> making.
>
> I think you should spell out that insert_section_p is only used by
> update_section_map, which updates the struct objfile_pspace_info
> sections table, which is only used from find_pc_section.  Then you'd
> need to explain why non of these sections can ever be returned from
> find_pc_section, though it's not clear (from the discussion on the v1
> thread) if the ABS section might be returned in some cases or not..
>
> I tracked down the patch which I think originally added these synthetic
> sections, though I don't know if this helps much:
>
>   https://sourceware.org/pipermail/gdb-patches/2013-February/100257.html
>
>
>> ---
>> v1: https://sourceware.org/pipermail/binutils/2022-May/120863.html
>> v1 -> v2: Fix code style, post to the correct mailing list (Andrew).
>>
>>  gdb/objfiles.c | 5 +++++
>>  1 file changed, 5 insertions(+)
>>
>> diff --git a/gdb/objfiles.c b/gdb/objfiles.c
>> index 80f68fda1c1..8a297c57530 100644
>> --- a/gdb/objfiles.c
>> +++ b/gdb/objfiles.c
>> @@ -1005,6 +1005,11 @@ insert_section_p (const struct bfd *abfd,
>
> You'll need to update the comment just before this function.
>
>>    if ((bfd_section_flags (section) & SEC_THREAD_LOCAL) != 0)
>>      /* This is a TLS section.  */
>>      return 0;
>> +  if (bfd_is_const_section (section))
>> +    {
>> +      /* This is one of the global *ABS*, *UND*, *IND*, or *COM* sections.  */
>> +      return 0;
>> +    }
>>  
>>    return 1;
>>  }
>
> The final thing I think you need to add with this patch is some
> testing.  We don't have much (that I'm aware of) in the way of
> performance testing, but what we can do, is add a mechanism by which we
> can gather performance data.
>
> Below you will find a patch that extends one of the existing JIT tests
> to gather performance data.  I tried this before and after applying your
> patch, and I can confirm that the performance improvement with your
> change is significant.  I think something like this should be included
> with this patch.

FYI, here's a graph of the performance change I'm seeing when loading
500 JIT ELF files:

  https://ibb.co/kQrGjtb

Thanks,
Andrew