> On Oct 14, 2024, at 14:12, Kito Cheng <kito.ch...@gmail.com> wrote:
>
> On Mon, Oct 14, 2024 at 1:40 PM Yangyu Chen <chenyan...@isrc.iscas.ac.cn>
> wrote:
>>
>>> On Oct 14, 2024, at 12:32, Kito Cheng <kito.ch...@gmail.com> wrote:
>>>
>>> I prefer not to handle the extension implication rules, as it's easy
>>> for them to get out of sync, and this should be the linux kernel's
>>> responsibility rather than libgcc's.
>> I still prefer to handle the extension here since users may use old
>> versions / LTS versions of the Linux kernel, which have no some
>> implied extensions.
>>
>> I have another patch to let GCC set the minimal required bits to
>> provide the best compatibility. However, in some cases, such as
>> some targets requiring Zve32x but not full V and users who have
>> kernel version < v6.11, the resolver will not probe the Zve32x
>> extension. Thus, we will not use the corresponding version. If we
>> check V in zve32x ifunc resolver, building it in gimple will make
>> it harder to maintain. Thus, I think both handling minimal extension
>> in libgcc and having the minimal feature bits to generate the IFUNC
>> resolver will be a better way to handle this.
>>
>> Meanwhile, handling it in this way does not increase the maintenance
>> overhead since we don’t need to handle the extensions and implied
>> extensions appear in the same version of the Linux kernel. So implied
>> extensions like Zbc->Zbkc does not need to be handled here. Till now,
>> we only need the handle C->Zca, Zcd, Zcf and V->Zve32x, Zve32f,
>> Zve64x, Zve64f, Zve64d.
>
> I am OK with minimal extension implication rules, like what you implement now,
> but I am still concerned about implementing full rules.
I didn't know the reason for the concerns. Maybe I didn't tell it
clearly.
I want both sets of implied extensions in libgcc here and also to
use a minimal feature bit for the IFUNC resolver to solve. I think
this is the safest way to provide compatibility for both older
versions of Linux kernels and users who only set a minimal extension
(like Zve32x but no V).
Would this be acceptable? If OK, I will fix the C->Zcd for RV64 only
and submit the next revision.
Thanks,
Yangyu Chen
