[PATCH 3/4] powerpc: Runtime selection between sc and scv for syscalls

[Matheus Castanho]

On 12/1/20 10:35 AM, Adhemerval Zanella via Libc-alpha wrote:
> 
> 
> On 01/12/2020 10:11, Florian Weimer via Libc-alpha wrote:
>> * Matheus Castanho via Libc-alpha:
>>
>>> Hi Florian,
>>>
>>> On 11/19/20 5:35 PM, Florian Weimer wrote:
>>>> * Matheus Castanho:
>>>>
>>>>> That should work for shared libc, but in the static case we may also
>>>>> hit the same problem: trying to access the TLS to read hwcap2 before
>>>>> it has been initialized, but this time in csu/libc-tls.c
>>>>
>>>> Ahh.
>>>>
>>>>> Is there a way to also check if we are in the static startup code at
>>>>> compile time? If not, I'm afraid I'll have to keep the check for the
>>>>> thread pointer.
>>>>
>>>> Is the thread pointer in a regular register?  Then you could install a
>>>> fake TCB early on that has a zero bit in the right place.
>>>>
>>>> The other option would be to always use the old interface for !SHARED.
>>>> Just saying. 8-)
>>>>
>>>
>>> I believe adding the fake TCB is a bit out of the scope of this patch, and I'd
>>> also prefer to keep the same behavior for both static and shared libcs, so we
>>> avoid surprises in the future.
>>>
>>> Do you see this as a blocker for merging this patch?
>>
>> I think the run-time check in the shared builds is unnecessary.  I don't
>> have an opinion on the static case, but I think ld.so should use the
>> legacy interface unconditionally, and shared libc.so should use a
>> dynamic check while assuming that the TCB is initialized.  If we can
>> avoid making things harder for the branch predict, we should do so.
> I agree, for static case it might be more complicate to disentangle the
> loader code so the check might be required.
> 

[snip]

Ok, thanks for the feedback. I believe I was able to address this in v2 [0].

[0] https://sourceware.org/pipermail/libc-alpha/2020-December/120353.html

Thanks,
Matheus Castanho