Should we use R_ARM_MOVT_PREL for ARM PIC?

[Peter Smith via llvm-commits]

I think it is worth opening a bug report, at least to record it as a
suggestion. My understanding is that MOVT/MOVW is generally faster
than loading a literal on most ARM implementations, on the other hand
a literal can be shared more easily which can give smaller code-size.
For PC-relative offsets I think sharing is less likely so in theory
MOVT/MOVW should be profitable in this case.

I don't have any benchmarks to back any of that up, but it would be
worth an investigation.

Peter



On 21 June 2016 at 14:05, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
> On 21 June 2016 at 06:18, Peter Smith <peter.smith at linaro.org> wrote:
>> Hello Rafael,
>>
>> I don't have a particular good answer for why there is a difference in
>> code-generation right now.
>
> Do you think it would be profitable to use movw+movt? Should I open a
> bug about it?
>
>> With respect to the relocations, I think there might already be
>> movw/movt relocations that can do some of what you want.
>> The base relative versions of the relocations are:
>> R_ARM_MOVT_BREL
>> R_ARM_MOVW_BREL
>> R_ARM_MOVW_BREL_NC
>> R_ARM_THM_MOVT_BREL
>> R_ARM_THM_MOVW_BREL
>> R_ARM_THM_MOVW_BREL_NC
>>
>> These are all of the form ((S + A) | T) – B(S)
>>
>> where B(S) is defined as:
>>
>> B(S) is the addressing origin of the output segment defining the
>> symbol S. The origin is not required to be the base address of the
>> segment. This value must always be word-aligned.
>>
>> The wording of B(S) is vague enough for platforms such as Linux to say
>> what the addressing origin of B(S). It could be defined in terms of
>> the GOT.
>
> It doesn't seem to be implemented like that in gold.
>
>> If you've got something different in mind and want to make a proposal
>> for a new relocation type the ABI documents have an email address on
>> the front of the document that can be used.
>>
>> Link to ARM ELF for convenience:
>> http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044f/IHI0044F_aaelf.pdf
>
> Nah, it is just a passing curiosity.
>
> I am refactoring the pic/preemption code and noticed that ELF codegen
> avoids using movw/movt when MachO is able to. If it was easy I would
> change it to use it, but adding to the ABI would be an yak too many
> :-)
>
> Cheers,
> Rafael