Re: [Buildroot] [PATCH] package/qemu: host-qemu does not support AVX512 CPUs

[Thomas Petazzoni via buildroot <buildroot@buildroot.org>]

Hello Yegor,

On Thu, Sep 17, 2026 at 08:56:46AM +0200, Yegor Yefremov wrote:

>  No, I don't think it does anymore, and I believe it can simply be dropped.
> 
>   1. The exclusion comes from commit 1f7efaf89f03 ("package/qemu: do not
>   support x86_steamroller or x86_core_avx2", Feb 2020), i.e. from the Qemu
>   4.2 era, when TCG had no AVX support at all. Since Qemu 7.2, TCG
>   implements AVX, AVX2, F16C, FMA3 and VAES. In Qemu 11.0.0,
>   target/i386/cpu.c:
> 
>     #define TCG_EXT_FEATURES (... | CPUID_EXT_AVX | CPUID_EXT_F16C | \
>               CPUID_EXT_FMA | ...)
>     #define TCG_7_0_EBX_FEATURES (... | CPUID_7_0_EBX_BMI1 | \
>               CPUID_7_0_EBX_BMI2 | CPUID_7_0_EBX_AVX2 | ...)
> 
>   which covers everything -march=core-avx2 can generate.

Cool!

>   2. You are right that core-avx2 and haswell are equivalent: gcc's
>   core-avx2 is a deprecated alias for haswell, and "gcc -march=core-avx2
>   -Q --help=target" yields exactly the same feature bits as

TIL about "gcc -march=core-avx2 -Q --help=target", nice!

>   -march=haswell. In arch/Config.in.x86 both also select the very same
>   BR2_X86_CPU_HAS_* symbols. So excluding one and not the other never
>   made much sense. The same applies to broadwell, skylake, zen*,
>   x86-64-v3, ..., which all enable AVX2 and were never excluded - and
>   they do build fine on the autobuilders, which is another hint that
>   AVX2 is not the problem.

ACK.

>   3. The steamroller exclusion, however, is still needed - but it is
>   incomplete. steamroller is bdver3, and what TCG cannot emulate there is
>   not AVX, but XOP/FMA4/TBM/LWP. Qemu 11.0.0 has:
> 
>     #define TCG_EXT3_FEATURES (CPUID_EXT3_LAHF_LM | CPUID_EXT3_SVM | \
>               CPUID_EXT3_CR8LEG | CPUID_EXT3_ABM | CPUID_EXT3_SSE4A  | \
>               CPUID_EXT3_3DNOWPREFETCH | CPUID_EXT3_KERNEL_FEATURES | \
>               CPUID_EXT3_CMP_LEG)
> 
>   CPUID_EXT3_XOP, CPUID_EXT3_FMA4, CPUID_EXT3_TBM and CPUID_EXT3_LWP are
>   defined in target/i386/cpu.h, but are deliberately not part of that
>   mask.
> 
>   And the whole Bulldozer family is affected, not only steamroller
>   (checked with gcc -march=<variant> -Q --help=target):
> 
>     bdver1 (bulldozer):   XOP, FMA4, LWP
>     bdver2 (piledriver):  XOP, FMA4, TBM, LWP
>     bdver3 (steamroller): XOP, FMA4, TBM, LWP
>     bdver4 (excavator):   XOP, FMA4, TBM, LWP (and AVX2)
> 
>   So BR2_x86_bulldozer, BR2_x86_piledriver and BR2_x86_excavator ought to
>   be excluded as well.
> 
>   If you agree, I'll send a follow-up patch that:
> 
>    - drops the !BR2_x86_core_avx2 dependency,
>    - adds a hidden BR2_X86_CPU_HAS_XOP symbol in arch/Config.in.x86,
>      selected by the four bdver variants, and replaces
>      !BR2_x86_steamroller with !BR2_X86_CPU_HAS_XOP in
>      BR2_PACKAGE_HOST_QEMU_ARCH_SUPPORTS - i.e. the same pattern as the
>      existing BR2_X86_CPU_HAS_AVX512 dependency.

This looks perfect!

>   If you'd rather not add a new BR2_X86_CPU_HAS_* symbol for a single
>   user, I can just list the four variants explicitly instead.

I think adding BR2_X86_CPU_HAS_XOP is better, even if it has a single
user. Indeed, it makes it explicit what is the feature that isn't
implemented by Qemu.

Thanks a lot for this very thorough answer!

Best regards,

Thomas
-- 
Thomas Petazzoni, co-owner and CEO, Bootlin
Embedded Linux and Kernel engineering and training
https://bootlin.com
_______________________________________________
buildroot mailing list
buildroot@buildroot.org
https://lists.buildroot.org/mailman/listinfo/buildroot