On 9/3/21 11:11 PM, Philippe Mathieu-Daudé wrote:
> On 9/3/21 10:42 PM, Richard Henderson wrote:
>> On 9/3/21 2:50 AM, David Gibson wrote:
>>> On Thu, Sep 02, 2021 at 06:15:34PM +0200, Philippe Mathieu-Daudé wrote:
>>>> Each POWER cpu has its own has_work() implementation. Instead of
>>>> overloading CPUClass on each PowerPCCPUClass init, register the
>>>> generic ppc_cpu_has_work() handler, and have it call the POWER
>>>> specific has_work().
>>>
>>> I don't quite see the rationale for introducing a second layer of
>>> indirection here. What's wrong with switching the base has_work for
>>> each cpu variant?
>>
>> We're moving the hook from CPUState to TCGCPUOps.
>> Phil was trying to avoid creating N versions of
>>
>> static const struct TCGCPUOps ppc_tcg_ops = {
>> ...
>> };
>
> Ah yes this is the reason! Too many context switching so
> I forgot about it.
>
>> A plausible alternative is to remove the const from this struct and
>> modify it, just as we do for CPUState, on the assumption that we cannot
>> mix and match ppc cpu types in any one machine.
>
> I thought about this case and remembered how it works on the ARM arch,
> i.e. ZynqMP machine uses both Cortex-R5F and Cortex-A53. Even if no
> similar PPC machine exists, IMHO we should try to generally allow to
> possibility to experiment machine with different CPUs. Restricting it
> on PPC goes the other way around. Thoughts?
I'm running out of ideas to do avoid the indirection and multiple
copies of TCGCPUOps. I'm not giving up, I suppose I'm simply not
seeing it... David, any suggestions?
