[PATCH v1 1/1] Set Prefer_No_AVX512 flag for hygon platform

[Jiamei Xie]

Hi Florlan,

> -----Original Message-----
> From: Florian Weimer <fweimer@redhat.com>
> Sent: 2025年10月14日 16:48
> To: Jiamei Xie <xiejiamei@hygon.cn>
> Cc: libc-alpha@sourceware.org; Jing Li <lijing@hygon.cn>
> Subject: Re: [PATCH v1 1/1] Set Prefer_No_AVX512 flag for hygon platform
> 
> * Jiamei Xie:
> 
> > Hi Florian，
> >
> >> -----Original Message-----
> >> From: Florian Weimer <fweimer@redhat.com>
> >> Sent: 2025年10月14日 16:26
> >> To: Jiamei Xie <xiejiamei@hygon.cn>
> >> Cc: libc-alpha@sourceware.org; Jing Li <lijing@hygon.cn>
> >> Subject: Re: [PATCH v1 1/1] Set Prefer_No_AVX512 flag for hygon
> >> platform
> >>
> >> * Jiamei Xie:
> >>
> >> >> And surely this must be under the HygonGenuine condition?
> >> >>
> >> >
> >> > Yes. This patch is for the Hygon Platform, whose model is 0x7.
> >>
> >> Ahh.  Is that really unique?  I would have expected that this check
> >> would be
> >> reuse:
> >>
> >>   /* This spells out "HygonGenuine".  */
> >>   else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx ==
> 0x6e65476e)
> >>     {
> >>
> >
> > AFAIK, "HygonGenuine" is the vendor ID, and the model number identifies a
> specific CPU model. Hygon has multiple models.
> >
> > In our testing, Hygon model 0x7 does not perform well on AVX-512.
> > In the future, there may be models that support AVX-512 and outperform
> EVEX.
> >
> > Therefore, I believe it’s appropriate to guard this logic under a model check.
> 
> Then do both the vendor ID check and the model check?  Aren't think model
> IDs vendor-specific?

> 

Yes — we should check both the vendor ID and the model number. I will modify and
resend. Thanks.

> Thanks,
> Florian
>