Re: [RFC v1 00/38] arm cleanup experiment for kvm-only build

[Philippe Mathieu-Daudé]

On 3/4/21 6:19 PM, Claudio Fontana wrote:
> On 3/4/21 5:39 PM, Philippe Mathieu-Daudé wrote:
>> On 3/3/21 7:54 PM, Richard Henderson wrote:
>>> On 3/3/21 10:45 AM, Claudio Fontana wrote:
>>>> It is a bit weird that qemu-system-aarch64 runs with a cortex-a15
>>>> model tbh, as cortex-a15 is not capable of aarch64.
>>>
>>> No, but qemu-system-aarch64 is capable of 32-bit emulation (because most
>>> 64-bit cpus retain 32-bit mode).  It takes no extra effort to run
>>> cortex-a15 than it does a cortex-a57.
>>
>> IIRC qemu-system-arm starts the aa64 cores in 32-bit mode, while
>> qemu-system-aarch64 in 64-bit (this gave me trouble because the
>> kernels for the raspi 64-bit SoCs are in 32-bit mode -- because
>> the GPU starts them in this mode).
>>
>>> I have wondered if we should have just one qemu-system-arm that does it
>>> all and drop the separate qemu-system-aarch64 -- or vice versa.  But
>>> we've had the separation around so long I'm sure someone would be confused.
>>
>> That would be great cleanup IMHO.
>>
> 
> Would we still be able to configure a lean AARCH64-only qemu that only 
> contains the cpu models we want,
> (via board configuration / KConfig?),

We shouldn't even worry about this IMO. Because ...

> for example, a kvm-only build that only has a few 64-bit cpu models in it, 
> plus max/host and removes all the rest?

... since host word-size becomes irrelevant, you only have to consider
the accelerator possibilities. Restricting to KVM directly select a
subset of CPUs usable.