[PATCH v9 06/19] bfd: write Object Attributes v2

[Jan Beulich]

On 13.11.2025 20:51, Matthieu Longo wrote:
> On 07/11/2025 08:23, Jan Beulich wrote:
>> On 05.11.2025 11:18, Matthieu Longo wrote:
>>> On 30/10/2025 15:55, Jan Beulich wrote:
>>>> On 01.09.2025 18:56, Matthieu Longo wrote:
>>>>> --- a/gas/write.c
>>>>> +++ b/gas/write.c
>>>>> @@ -1919,6 +1919,8 @@ create_obj_attrs_section (void)
>>>>>      bfd_set_section_flags (s, SEC_READONLY | SEC_DATA);
>>>>>      frag_now_fix ();
>>>>>      char *p = frag_more (size);
>>>>> +
>>>>> +  bfd_elf_obj_attr_finalize_contents (stdoutput);
>>>>>      bfd_elf_set_obj_attr_contents (stdoutput, (bfd_byte *)p, size);
>>>>
>>>> Why is another export needed from libbfd? Can't bfd_elf_set_obj_attr_contents()
>>>> simply do whatever it takes? The other caller likely needs doing the same anyway?
>>>
>>> bfd_elf_set_obj_attr_contents() is also used in objcpy, and objcpy
>>> should not change the order of subsections and attributes, but simply
>>> make a copy of the data as they were in the original object.
>>
>> Is this the case already prior to your series? I can't seem to be able to
>> identify the call path for that case. Besides the direct call from gas,
>> bfd_elf_set_obj_attr_contents() is called only from
>> _bfd_elf_write_section_build_attributes(), which in turn is called solely by
>> bfd_elf_final_link(). And I can't see this being used by objcopy.
>>
> 
> Sorry, I said something wrong. It is not used for objcopy but by the 
> linker in _bfd_elf_write_section_object_attributes().
> 
> In the case of the link, all the subsections and attributes are already 
> sorted by the merge, so no need to sort them again.
> 
>>> Another possibility might be to add a parameter to
>>> bfd_elf_set_obj_attr_contents() to enable/disable the sorting.
>>
>> Can the distinction not be made by what struct bfd_link_info contains?
>>
>> Jan
> 
> Here is the new approach after having eliminated 
> bfd_elf_obj_attr_finalize_contents(). It relies on is_linker_output 
> being set to 1 in the case of the linker. Is it the approach that you 
> had in mind ?
> 
> /* Write the contents of the object attributes section to CONTENTS.  */
> void
> bfd_elf_set_obj_attr_contents (bfd *abfd, bfd_byte *buffer, bfd_vma size)
> {
>    if ((! abfd->is_linker_output)
>        && elf_obj_attr_version (abfd) == OBJ_ATTR_V2)
>      /* Before dumping the data, sort subsections in alphabetical order, and
>         attributes according to their tag in numerical order.  This is 
> useful
>         for diagnostic tools so that they dump the same output even if the
>         subsections or their attributes were not declared in the same 
> order in
>         different files.
>         This sorting is only performed in the case of the assembler.  In the
>         case of the linker, the subsections and attributes are already 
> sorted
>         by the merge process.  */
>      oav2_sort_subsections (&elf_obj_attr_subsections (abfd));
> 
>    obj_attr_version_t version = elf_obj_attr_version (abfd);
>    if (version == OBJ_ATTR_V1)
>      oav1_write_section (abfd, buffer, size);
>    else if (version == OBJ_ATTR_V2)
>      oav2_write_section (abfd, buffer, size);
>    else
>      abort ();
> }

Yes, this looks good to me. Thanks.

Jan