llvm dev - Jul 2021 - [lld][ELF] Regarding removeUnusedSyntheticSections()

Hi,

I am working on an LLD port for a research project based on the AARCH64 based
architecture which requires adding a new SyntheticSection to InStruct, but doing
so gives me 2 failed x86 tests:
  lld :: ELF/shuffle-sections-init-fini.s
  lld :: ELF/shuffle-sections.s
This new SyntheticSection is empty for X86, so in principle those tests should
be unaffected.

The immediate reason is that the shuffle implementation, maybeShuffle(), is
dependent on inputSections. I think the real cause is that
removeUnusedSyntheticSections() does not remove empty sections from the
inputSections vector (although there is a comment that says that it does that).
AFAICT elements are removed only from isd->sections and
script->orphanSections.

I could not find any uses of empty sections after this point, but you never know
so thought of asking here. Is there a reason empty SyntheticSections in
inputSections is preserved after removeUnusedSyntheticSections()? Or is there
another approach that will not require maintaining the shuffle tests everytime a
Synthetic section is added/removed?

If it is safe to remove empty SyntheticSections from inputSections I can send
out a patch to do that. Locally I have a simple patch that confirms that the
shuffle tests do not fail after adding a new SyntheticSection. I can't think
of new regression tests to add because this behavior cannot be demonstrated
without adding a new SyntheticSection to InStruct.

Thanks
Amilendra

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210717/70dcdc1b/attachment.html>

On 2021-07-17, Amilendra Kodithuwakku via llvm-dev
wrote:
>Hi,
>
>I am working on an LLD port for a research project based on the AARCH64
based architecture which requires adding a new SyntheticSection to InStruct, but
doing so gives me 2 failed x86 tests:
>  lld :: ELF/shuffle-sections-init-fini.s
>  lld :: ELF/shuffle-sections.s
>This new SyntheticSection is empty for X86, so in principle those tests
should be unaffected.
>
>The immediate reason is that the shuffle implementation, maybeShuffle(), is
dependent on inputSections. I think the real cause is that
removeUnusedSyntheticSections() does not remove empty sections from the
inputSections vector (although there is a comment that says that it does that).
AFAICT elements are removed only from isd->sections and
script->orphanSections.
>
>I could not find any uses of empty sections after this point, but you never
know so thought of asking here. Is there a reason empty SyntheticSections in
inputSections is preserved after removeUnusedSyntheticSections()? Or is there
another approach that will not require maintaining the shuffle tests everytime a
Synthetic section is added/removed?
>
>If it is safe to remove empty SyntheticSections from inputSections I can
send out a patch to do that. Locally I have a simple patch that confirms that
the shuffle tests do not fail after adding a new SyntheticSection. I can't
think of new regression tests to add because this behavior cannot be
demonstrated without adding a new SyntheticSection to InStruct.
It is safe. The comment is about the references (isd->sections;
script->orphanSections).

(I can fix it after one week if I haven't seen a patch yet:) )
>Thanks
>Amilendra