Re: load of in-place relocation addend in AArch64

["Frédéric Rivière" <frederic.riviere@microej.com>]

Hi Nick,

Thanks for the quick reply.

My compiler generates REL relocations because I was just naively 
following the official AArch64 ABI.

I understand that RELA is more easy to manage since there is no need to 
retrieve the source section for getting the addend.

However, I believe that ld AArch64 should not support REL entries at all 
(with a clear error message) instead of silently generating a wrong 
object file.

Finally, I just updated my compiler to generate RELA and now it works fine.

Best regards,

--Frédéric

On 12/6/2023 12:14 PM, Nick Clifton wrote:
> Hi Frédéric,
>
>> My compiler generates REL relocations with an in-place addend but the 
>> addend is not taken into account for computing the final relocated 
>> symbol address (S+A).
>
> Is there any particular reason for this ?
>
> Ie, if your compiler generated RELA relocations then all of your problems
> with the binutils should go away...
>
>> So it seems the that I'm the first one that generates REL relocations 
>> for AArch64 target.
>
> True.
>
>> However, I don't see any limitation on using REL relocations in the 
>> AARCH64 ABI.
>
> Correct.  There are - in theory - allowed.
>
>> I would appreciate any comment/feedback that would help me 
>> understanding the reasons behind that.
>
> As a general rule RELA relocations are preferable in that they
> are simpler - it is easier to extract the addend - and in theory
> they have the possibility to express a wider range of addends
> than would be available to REL relocations.
>
> Have you tried modifying the AArch64 code in the BFD library to
> support REL relocations ?  In theory all that you need to do is
> to change a few lines in bfd/elfnn-aarch64.c:
>
>   #define elf_backend_may_use_rel_p      0
>
> and maybe:
>
>   #define elf_backend_may_use_rela_p     1
>   #define elf_backend_default_use_rela_p 1
>   #define elf_backend_rela_normal     1
>
> In practice I am pretty sure that there will be assumptions of RELA
> relocs elsewhere in the AArch64 specific code, so I doubt it will be
> as simple as I have suggested.
>
> Cheers
>   Nick