x86: Incorrect CPUID leaf in Zen 5 detection

[Jussi Kivilinna]

Hello,


On 08/10/2026 12:22, HMBSbige via Gcrypt-devel wrote:
> Hi,
> The AVX512_BMM check in src/hwf-x86.c has two errors:
> 
> 1. The leaf should be 0x80000021, not 0x21.
> 2. The availability check should use the maximum extended leaf
>     returned by CPUID(0x80000000).EAX, not the maximum basic leaf
>     from CPUID(0).EAX. The file currently never queries 0x80000000.

Thanks for reporting this.

> 
> The minimal fix would be:
> 
>      unsigned int max_ext_level, amd_feat;
> 
>      get_cpuid(0x80000000, &max_ext_level, NULL, NULL, NULL);
>      if (max_ext_level >= 0x80000021)
>        {
>          get_cpuid(0x80000021, &amd_feat, NULL, NULL, NULL);
>          if (amd_feat & (1 << 23))
>            has_avx512bmm = 1;
>        }
> 
> Alternatively, I suggest identifying Zen 5 directly by vendor,
> family and model. Family 0x1A also includes Zen 6, so the absence
> of AVX512_BMM is not sufficient to identify Zen 5.
> 
> Following Linux's classification in arch/x86/kernel/cpu/amd.c,
> the Zen 5 condition would be:
> 
>      if (is_amd_cpu && family == 0x1a
>          && (model <= 0x2f
>              || (model >= 0x40 && model <= 0x4f)
>              || (model >= 0x60 && model <= 0x7f)
>              || (model >= 0xd0 && model <= 0xd7)))
>        {
>          cpu_details->int_vector_latency = 2;
>          cpu_details->prefer_gpr_over_scalar_int_vector = 1;
>        }
> 
> This reuses the already decoded family/model and eliminates the
> AVX512_BMM probe. Unknown models would retain the default tuning
> until evaluated.

I think this approach would make more sense. I'll switch to using it.

-Jussi