Re: [gold] enable sorting of text sections with the same prefix

[Alexander Ivchenko <aivchenk at gmail dot com>]

>The documentation of --sort-section=name for GNU ld is unfortunately
>tied to the notion of a default linker script, which gold does not
>share.  Still, we ought to be able to come up with some plausible
>meaning for gold.  And restricting the behaviour to .data and .bss
>does not make sense to me

that's true, e.g for a testcase with .data and .sdata here what ld -M gave me:

.data           0x0000000000600188        0xc
 *(SORT(.data) SORT(.data.*) SORT(.gnu.linkonce.d.*))
 .data          0x0000000000600188        0x0 section_sorting_name_3.o
 .data          0x0000000000600188        0x0 section_sorting_name_1.o
 .data          0x0000000000600188        0x0 section_sorting_name_2.o
 .data.0001     0x0000000000600188        0x4 section_sorting_name_1.o
                0x0000000000600188                vdata_0001
 .data.0002     0x000000000060018c        0x4 section_sorting_name_2.o
                0x000000000060018c                vdata_0002
 .data.0003     0x0000000000600190        0x4 section_sorting_name_3.o
                0x0000000000600190                vdata_0003

.sdata.0003     0x0000000000600194        0x4
 .sdata.0003    0x0000000000600194        0x4 section_sorting_name_3.o
                0x0000000000600194                vsdata_0003

.sdata.0001     0x0000000000600198        0x4
 .sdata.0001    0x0000000000600198        0x4 section_sorting_name_1.o
                0x0000000000600198                vsdata_0001

.sdata.0002     0x000000000060019c        0x4
 .sdata.0002    0x000000000060019c        0x4 section_sorting_name_2.o
                0x000000000060019c                vsdata_0002

It seems like there is no particular reason for not sorting .sdata
sections: just BFD script doesn't do it. Therefore I agree with Sri
when he said:

>Why not sort all output sections when --sort-section=name is
>specified? However, for special output sections like ctors, dtors,
>init_array, fini_array, etc. the original sort compare function will
>be used. For all other sections, use the new sort compare can be used.
>No need to hard code any names.

that approach looks reasonable. Also thank you very much, Sri, for
your fix. I attached the patch with your
changes and also with updated ChangLog.

thanks,
Alexander

2013/2/12 Sriraman Tallam <tmsriram@google.com>:
> On Tue, Feb 12, 2013 at 10:22 AM, Sriraman Tallam <tmsriram@google.com> wrote:
>> On Tue, Feb 12, 2013 at 9:21 AM, Ian Lance Taylor <iant@google.com> wrote:
>>> On Tue, Feb 12, 2013 at 7:01 AM, Alexander Ivchenko <aivchenk@gmail.com> wrote:
>>>> 2013/2/9 Ian Lance Taylor <iant@google.com>:
>>>>> On Fri, Feb 8, 2013 at 5:11 PM, Ian Lance Taylor <iant@google.com> wrote:
>>>>>> On Thu, Feb 7, 2013 at 3:56 AM, Alexander Ivchenko <aivchenk@gmail.com> wrote:
>>>>>>> thank you for your help, Sri. I fixed help string and deleted warning.
>>>>>>>
>>>>>>> Ian, could you please take a look at the attached patch?
>>>>>>
>>>>>> As far as I can see, with this patch, when you use
>>>>>> --sort-section=name, gold will only sort sections that start with
>>>>>> .data. and .bss. by name.  Other sections remain unsorted.  This does
>>>>>> not appear to be what the GNU linker does.  The GNU linker appears to
>>>>>> sort all input sections by name when using --sort-section=name.
>>>>>
>>>>> Sorry, I guess that's not quite right.  You call the sorting code one
>>>>> every section.  But you only set must_sort_attached_input_sections on
>>>>> the .data and .bss sections.  How can you get away with that?
>>>>>
>>>>> Ian
>>>>
>>>> You mean from hardcoding those names (.bss and .data) there? I'm not
>>>> sure so far,
>>>> but I know that BFD sorts them by name when we have -sort-section=name
>>>> and at the
>>>> same time, BFD doesn't sort, say,.sdata and .sbss.
>>>> Do we need to fully mimic the behavior of BFD for this option?
>>>
>>> We do not need to fully mimic GNU ld.  However, we need to understand
>>> how and why GNU ld behaves the way it does.  When I look at the GNU ld
>>> code, I don't see anything that restricts the effect of
>>> --sort-section=name to the .data and .bss sections.  Nor is it
>>> documented to behave that way.
>>>
>>> The documentation of --sort-section=name for GNU ld is unfortunately
>>> tied to the notion of a default linker script, which gold does not
>>> share.  Still, we ought to be able to come up with some plausible
>>> meaning for gold.  And restricting the behaviour to .data and .bss
>>> does not make sense to me.
>>
>> Why not sort all output sections when --sort-section=name is
>> specified? However, for special output sections like ctors, dtors,
>> init_array, fini_array, etc. the original sort compare function will
>> be used. For all other sections, use the new sort compare can be used.
>> No need to hard code any names.
>>
>>>
>>> Also, in order for this to work correctly, you must call
>>> set_may_sort_attached_input_sections when you create the output
>>> section.  You aren't doing that, and I'm surprised that your code is
>>> working reliably.
>>
>> I missed this part completely when I was reviewing his code, sorry!. I
>> am not surprised his patch works for ".text" because the input
>> sections are retained as he can piggy back on default text sorting.
>> But, how does his test pass for bss and data? I will apply his patch
>> and find out.
>
> I figured out how input sections are kept for some ".data" sections.
> When the first object is seen, the isecn entries for its .data and
> .bss are not kept. But after the first object, Layout::layout is
> called which sets must_sort for .data and .bss. From then on, .data
> and .bss are saved. This is definitely wrong. Infact, this patch does
> not work correctly on the test case included and produces an assert in
> reloc.cc:830 when I tried it. This is because some input section
> entries have isecn and some do not. We encountered an instance of this
> problem earlier with the text reordering patch.
>
> This can be fixed by removing the lines which set_must_sort in
> Layout::layout and simply set_must_sort to all output sections in
> Layout::make_output_section. Please note that setting may_sort and
> then must_sort later is not necessary here since we know for sure that
> we are going to sort this.
>
> I have modified this patch accordingly and attached a new patch that
> sorts all output sections by name when --sort-section=name is passed.
> For special output sections like .ctors, it will still use the
> original sort compare function. I have not special cased if for
> ".data" and ".bss" but I am not sure if ".sdata" and ".sbss" must be
> ignored.
>
> Thanks
> Sri
>
>
>
>>
>> Sri
>>
>>>
>>> Ian

Attachment: enable_sorting_sections_by_name_7.patch
Description: Binary data