[pve-devel] [PATCH v4 qemu-server 1/1] cpuconfig: add new x86-64-vX models

[Fiona Ebner]

Am 02.06.23 um 12:05 schrieb Alexandre Derumier:
> https://lists.gnu.org/archive/html/qemu-devel/2021-06/msg01592.html
> "
> In 2020, AMD, Intel, Red Hat, and SUSE worked together to define
> three microarchitecture levels on top of the historical x86-64
> baseline:
> 
>   * x86-64:    original x86_64 baseline instruction set
>   * x86-64-v2: vector instructions up to Streaming SIMD
>                Extensions 4.2 (SSE4.2)  and Supplemental
>                Streaming SIMD Extensions 3 (SSSE3), the
>                POPCNT instruction, and CMPXCHG16B
>   * x86-64-v3: vector instructions up to AVX2, MOVBE,
>                and additional bit-manipulation instructions.
>   * x86-64-v4: vector instructions from some of the
>                AVX-512 variants.
> "

Can we also link to https://gitlab.com/x86-psABIs/x86-64-ABI/ because
table 3.1 in the PDF contains a bit more precise information?

I used the following to test for some instructions, but feel free to
tell me something better ;)

> int main()
> {
>     __asm__ __volatile__
>     (
>         "xgetbv"
>     );
>     return 0;
> }


I also found
https://gist.github.com/rindeal/81198b1cf8f55c356743
and after adding #include <stdbool.h> to the cpuid-dump2.c it seemed to
work.

> 
> This patch add new builtin model derivated from qemu64 model,
> to be compatible between intel/amd.
> 
> x86-64-v1 : I'm skipping it, as it's basicaly qemu64|kvm64 -vme,-cx16 for compat Opteron_G1 from 2004
>             so will use it as qemu64|kvm64 is higher are not working on opteron_g1 anyway
> 
> x86-64-v2 : Derived from qemu, +popcnt;+pni;+sse4.1;+sse4.2;+ssse3
> 
> min intel: Nehalem
> min amd : Opteron_G3
> 
> x86-64-v2-AES : Derived from qemu, +aes;+popcnt;+pni;+sse4.1;+sse4.2;+ssse3

Could you describe how you chose/tested the flags?

>From the table in the PDF, most flags are clear (with pni apparently
being SSE3). There are also CMPXCHG16B and LAHF-SAHF CPU
features/instructions, but they are already part of qemu64 AFAICT. At
least the example instructions cmpxchg16b (needs an operand) and lahf
seem to work with my little test program and the cpuid-dump2 program
also outputs -mcx16 -msahf indicating support.

> min intel: Westmere
> min amd : Opteron_G3
> 
> x86-64-v3 : Derived from qemu64 +aes;+popcnt;+pni;+sse4.1;+sse4.2;+ssse3;+avx;+avx2;+bmi1;+bmi2;+f16c;+fma;+abm;+movbe

Again comparing with the table in the PDF all flags are clear (with abm
adding the LZCNT feature/instruction).

But isn't the OSXSAVE feature missing? At least if I try with my little
test program above I get "illegal hardware instruction" for xgetbv (that
is the example instruction for the OSXSAVE CPU Feature mentioned in the
PDF) and the has_osxsave variable in the cpuid-dump2 program is also false.

> 
> min intel: Haswell
> min amd : EPYC_v1
> 
> x86-64-v4 : Derived from qemu64 +aes;+popcnt;+pni;+sse4.1;+sse4.2;+ssse3;+avx;+avx2;+bmi1;+bmi2;+f16c;+fma;+abm;+movbe;+avx512f;+avx512bw;+avx512cd;+avx512dq;+avx512vl
> 

Those match up exactly with the table in the PDF, so should be fine
(except the OSXSAVE feature will also be missing as v4 extends v3).

> min intel: Skylake
> min amd : EPYC_v4
> 

(...)

> @@ -302,6 +321,17 @@ sub get_cpu_models {
>  	};
>      }
>  
> +    for my $model (keys %{$builtin_models}) {
> +	my $reported_model = $builtin_models->{$model}->{'reported-model'};
> +	$reported_model //= $cpu_fmt->{'reported-model'}->{default};

Nit: we could just avoid this and similar fallbacks in the rest of the
patch, because we know that 'reported-model' is set for the built-in
models. But it's not a big deal and in a language like Perl we don't
have compile-time checks for it, so I'm fine with either way. If you
drop these fallbacks, please add a comment above $builtin_models stating
that 'reported-model' has to be set.