[PATCH] D41105: [ELF] Fix placement of a sentinel entry in the .ARM.exidx section.

[Peter Smith via llvm-commits]

One possibility is to add a pointer to exidx section field to the
exidx sentinel that can be set to the last .ARM.exidx section after
we've processed the link order dependencies. We'd then not need to go
searching for it later.

Peter

On 13 December 2017 at 11:13, Igor Kudrin <ikudrin at accesssoftek.com> wrote:
> Maybe I can't fully understand your idea, but to put the relocation we still
> need to find the reference section, so I can't see why this approach is
> simpler than the current code. And we still have to write EXIDX_CANTUNWIND,
> so we can't get rid of  'ARMExidxSentinelSection::writeTo()' altogether. Am
> I right?
>
>
> Best Regards,
> Igor Kudrin
> C++ Developer, Access Softek, Inc.
> ________________________________
> From: Rafael Avila de Espindola <rafael.espindola at gmail.com>
> Sent: Wednesday, December 13, 2017 4:56 AM
> To: reviews+D41105+public+9917650116a4fc81 at reviews.llvm.org; Igor Kudrin;
> peter.smith at linaro.org; ruiu at google.com
> Cc: aemerson at apple.com; emaste at freebsd.org; javed.absar at arm.com;
> llvm-commits at lists.llvm.org; kristof.beyls at arm.com;
> adhemerval.zanella at linaro.org; t.p.northover at gmail.com; kanheim at a-bix.com;
> james.molloy at arm.com; diana.picus at linaro.org; florian.hahn at arm.com
> Subject: Re: [PATCH] D41105: [ELF] Fix placement of a sentinel entry in the
> .ARM.exidx section.
>
> Igor Kudrin via Phabricator <reviews at reviews.llvm.org> writes:
>
>> +    // Find the last existing InputSectionDescription.
>> +    // Add a new only if we cannot find one.
>> +    // As a result, if we have a description for an output section like
>> +    // .foo : { *(.foo) bar = . }
>> +    // we will place the new section before the assignment so that it
>> will
>> +    // affect the value of 'bar', which is, probably, expected.
>> +    // We also had a crash in the similar situation with an .ARM.exidx
>> sentinel,
>> +    // because 'ARMExidxSentinelSection::writeTo()' expected it to be
>> placed
>> +    // into a non-empty bucket.
>
> I would actually find the result surprising in the general case.
>
> I think we can just make the exidx code more resilient. Instead of
> having a special handling in writeTo, could we push a relocation to the
> Relocations vector of the sentinel entry?
>
> Cheers,
> Rafael