Re: [committed, PATCH] Always create dynsym section with dynamic sections

[Faraz Shahbazker <faraz dot shahbazker at imgtec dot com>]

On 04/23/16 08:36, H.J. Lu wrote:
> On Sat, Apr 23, 2016 at 8:24 AM, Faraz Shahbazker
> <faraz.shahbazker@imgtec.com> wrote:
>> On 04/23/16 05:27, H.J. Lu wrote:
>>> On Fri, Apr 22, 2016 at 7:31 PM, Faraz Shahbazker
>>> <faraz.shahbazker@imgtec.com> wrote:
>>>> On 04/22/16 19:05, H.J. Lu wrote:
>>>>> On Fri, Apr 22, 2016 at 6:37 PM, Faraz Shahbazker
>>>>> <faraz.shahbazker@imgtec.com> wrote:
>>>>>> On 04/22/16 16:24, H.J. Lu wrote:
>>>>>>> On Fri, Apr 22, 2016 at 2:49 PM, Faraz Shahbazker
>>>>>>> <Faraz.Shahbazker@imgtec.com> wrote:
>>>>>>>> On 04/22/2016 12:28 PM, H.J. Lu wrote:
>>>>>>>>> On Fri, Apr 22, 2016 at 11:55 AM, Faraz Shahbazker
>>>>>>>>>> + if (dynsymcount != 0 || elf_hash_table (info)->dynamic_sections_created)
>>>>>>>>>> +     ++dynsymcount;
>>>>>>>>>
>>>>>>>>> Are you saying dynamic_sections_created is 0 for MIPS here
>>>>>>>>> and will become 1 later?
>>>>>>>>
>>>>>>>> No, it will remain 0. The link is static, but the hash_table is still used to
>>>>>>>> record global symbols that have GOT relocations against them. Ofc, this
>>>>>>>> hash_table does not result in creation of a dynsym section, because well,
>>>>>>>> dynamic_sections_created is 0.
>>>>>>>>
>>>>>>>> Check the list of callers to bfd_elf_link_record_dynamic_symbol(), a number of
>>>>>>>> architectures use the link_hash_table in situations where it is not clear whether it is
>>>>>>>> being used to track dynamic symbols for a dynamic executable, as it is for x86.
>>>>>>>>
>>>>>>>
>>>>>>> So MIPS doesn't have dynamic symbols in this case.  It just borrows
>>>>>>> dynsymcount for different purpose.  Is this correct?
>>>>>>
>>>>>> Not quite! MIPS is expecting dynsymcount to count the number of symbols
>>>>>> that would have gone in to the .dynsym, even for a static executable. That way
>>>>>> parts of the arch-specific code can remain agnostic to the static/dynamic nature
>>>>>> of the link. It may not be used exactly as documented, but its not being used
>>>>>> for what one would call a different purpose.
>>>>>>
>>>>>> All we need is for handling of dynsymcount when renumbering to be consistent with
>>>>>> its initialization. If the initial increment for a NULL symbol was not gated by
>>>>>> dynamic_sections_created, then the increment when renumbering should also not.
>>>>>> If the increment when renumbering has to be gated by dynamic_sections_created,
>>>>>> then the initial increment must also be so.
>>>>>
>>>>> From what you are saying, shouldn't dynsymcount be incremented
>>>>> unconditionally?
>>>> No. Always, when the table is non-empty + whatever else you need.
>>>>
>>>
>>> You said dynsym should be treated treated the same for static and
>>> dynamic executables.   dynsymcount is number of dynsym + 1 in
>>> dynamic executable.  Why isn't it true for static executable?
>>
>> It is, or at least used to be, before this patch. It still is for both,
>> before renumbering. But now the +1 only happens for dynamic executables
>> when renumbering.
> 
> Then what is wrong to always +1 for both dynamic and static
> executables?

Aah, now I see! Ever since the table was created, the count was at least 1. 
You are saying it should always remain at least 1 no matter what. Had a quick
look through other uses of the field and it looks safe enough to me.

Regards,
Faraz Shahbazker