Re: [PATCH v4 03/16] x86emul: support AVX10.1

[Jan Beulich <jbeulich@xxxxxxxx>]

On 21.04.2026 09:32, Oleksii Kurochko wrote:
> 
> 
> On 4/9/26 5:03 PM, Jan Beulich wrote:
>> This requires relaxing various pre-existing AVX512* checks, as AVX10.1
>> covers all AVX512* except PF, ER, 4FMAPS, 4VNNIW (support for all of
>> which was removed meanwhile anyway), and VP2INTERSECT. At the same time
>> all vector widths are always supported.
>>
>> Note that visa_check(), replacing host_and_vcpu_must_have() uses, checks
>> only the guest capability: We wouldn't expose AVX512* (nor AVX10)
>> without the hardware supporting it. Similarly in avx512_vlen_check() the
>> original host_and_vcpu_must_have() is reduced just vcpu_must_have().
>> This also simplifies (resulting) code in the test and fuzzing harnesses,
>> as there the XCR0 checks that are part of cpu_has_avx512* are only
>> needed in local code, not in the emulator itself (where respective
>> checking occurs elsewhere anyway, utilizing emul_test_read_xcr()).
>>
>> While in most cases the changes to x86_emulate() are entirely
>> mechanical, for opmask insns earlier unconditional AVX512F checks are
>> converted into "else" clauses to existing if/else-if ones.
>>
>> To be certain that no uses remain, also drop respective cpu_has_avx512*
>> (except in the test harness) and vcpu_has_avx512*().
>>
>> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
>> ---
>> Probably avx512_vlen_check() should have the avx512_ prefix dropped, now
>> that it also covers AVX10. But if so that wants to be either a prereq or
>> a follow-on patch.
>>
>> visa_check() won't cover AVX10.2 and higher, but probably we will want
>> independent checking logic for that anyway.
>>
>> Spec version 2 still leaves unclear what the xstate components are which
>> would need enabling for AVX10/256. x86emul_get_fpu() is therefore
>> untouched for now.
>>
>> Since it'll be reducing code size, we may want to further convert
>> host_and_vcpu_must_have() to just vcpu_must_have() where appropriate
>> (should be [almost?] everywhere).
>> ---
>> v4: Switch to using fallthrough pseudo-keyword. Drop new _vlen_check()
>>      helper again. Re-base.
>> v3: Add ChangeLog entry.
>> v2: Drop use of vsz128 field. Re-base, in particular over dropping of
>>      Xeon Phi support.
>>
>> --- a/CHANGELOG.md
>> +++ b/CHANGELOG.md
>> @@ -16,6 +16,7 @@ The format is based on [Keep a Changelog
>>        mitigate (by rate-limiting) the system wide impact of an HVM guest
>>        misusing atomic instructions.
>>      - Support for CPIO microcode in discrete multiboot modules.
>> +   - Support for AVX10.1. (Experimental)
> 
> Acked-by: Oleksii Kurochko <oleksii.kurochko@xxxxxxxxx>.

Thanks.

> What should be done else to have not experimental support of AVX10.1?

Someone to actually try it out on real hardware. All my testing was with
Intel's SDE.

Jan