Re: [Qemu-devel] [RFC qom-cpu v2 2/2] target-i386: Turn Haswell into subclass of SandyBridge

[Alexander Graf <agraf@suse.de>]

On 12/12/2012 06:29 PM, Andreas Färber wrote:
> Am 12.12.2012 16:05, schrieb Eduardo Habkost:
>> On Wed, Dec 12, 2012 at 03:47:49PM +0100, Andreas Färber wrote:
>>> Am 12.12.2012 13:45, schrieb Eduardo Habkost:
>>>> On Mon, Dec 10, 2012 at 11:59:32PM +0100, Andreas Färber wrote:
>>>>>    ehabkost: "When adding the Haswell CPU model, I intended to make it
>>>>>    a superset of the features present on the SandyBridge model"
>>>>>
>>>>> Inherit from SandyBridge to keep only the delta for Haswell.
>>>> Most CPUs have a superset of the features of their predecessors. Are you
>>>> simply using SandyBridge->Haswell as an example, or you think their
>>>> relationship is special somehow?
>>>>
>>>> I believe we don't want to make externally-visible class inheritance,
>>>> but probably just reuse constans or init functions internally. A Haswell
>>>> CPU is not a type of SandyBridge CPU, it just happens to contain a
>>>> superset of the features present in SandyBridge.
>>>>
>>>> I mean: Haswell also has a superset of features of 486, but we don't
>>>> want to make the hierarchy look like the following, do we?
>>> I don't see why we would want to use a #define-based inheritence as
>>> suggested for the PPRO when we have QOM. QOM inheritence reduces lines
>>> of code significantly compared to just taking the values from elsewhere.
>> The reuse doesn't need to be #define-based (although maybe a
>> #define-based approach would work too), it could be function-call-based.
>>
>>> For the Haswell you said what I quoted, for the other models I said I
>>> need your or someone's help to verify whether a hierarchy such as below
>>> is semantically right or just a coincidence. I was at least considering
>>> an abstract intel-/amd-*-cpu to avoid repeating the three value
>>> assignments over and over.
>> Creating X86IntelCPU and X86AMDCPU classes make sense to me, because
>> Haswell is a kind of Intel CPu. Making Haswell a subclass of 486 (like
>> below) wouldn't.
>>
>>> At this time I believe the parents of a type are not (yet) exposed via
>>> QMP, just the "type" string property.
>> Even if they are not exposed externally, it's a confusing usage of
>> inheritance for me. I mean: a Haswell CPU is not a type of 486 CPU, it's
>> simply a different kind of CPU that happens to have a superset of the
>> 486 features.
> I concur with your is-a argument. My patch took a pragmatic approach.

Considering that x86 is very straight forward in its CPU definitions, I 
think we can easily make this a flat model based on CPUID features + a 
few variables.

If you really want to model something, I would rather say

386
   -> 386SX
   -> 386DX
opteron-G3
   -> phenom-9550
   -> phenom-9560

Any inheritance above that level sounds excessive to me.


Alex

> I'd like to wait for some more review comments on how to share code
> between models then - I remember Paul, Anthony and Alex discussing the
> x86 models a while back on IRC, CC'ing. (Summary: reading CPU models
> from config files has been dropped, we only have built-in models left -
> now how to design name-class mappings below X86CPU)
>
> Andreas
>
>>>> - X86CPU
>>>>    ->  X86IntelCPU
>>>>       ->  486
>>>>          ->  pentium
>>>>             ->  pentium2
>>>>                ->  pentium3
>>>>                   ->  Conroe
>>>>                      ->  Penryn
>>>>                         ->  Nehalem
>>>>                            ->  Westmere
>>>>                               ->  SandyBridge
>>>>                                  ->  Haswell