Re: [PATCH 4/4] pc-bios/s390-ccw: Allow building with Clang, too

[Thomas Huth <thuth@redhat.com>]

On 03/05/2021 11.14, Cornelia Huck wrote:
> On Mon, 03 May 2021 10:23:20 +0200
> Markus Armbruster <armbru@redhat.com> wrote:
> 
>> David Hildenbrand <david@redhat.com> writes:
>>
>>> On 03.05.21 07:17, Thomas Huth wrote:
>>>> On 03/05/2021 06.58, Markus Armbruster wrote:
>>>>> Thomas Huth <thuth@redhat.com> writes:
>>>>>   
>>>>>> Clang unfortunately does not support generating code for the z900
>>>>>> architecture level and starts with the z10 instead. Thus to be able
>>>>>> to support compiling with Clang, we have to check for the supported
>>>>>> compiler flags. The disadvantage is of course that the bios image
>>>>>> will only run with z10 guest CPUs upwards (which is what most people
>>>>>> use anyway), so just in case let's also emit a warning in that case.
>>>>>
>>>>> What happens when you try to use this bios with an old CPU anyway?
>>>>
>>>> Interesting question. I was expecting the guest to crash since it would be
>>>> using a CPU instruction that is not supported on the old CPU model. But I
>>>> just gave it a try, and there was no crash. The guest booted just fine.
>>>> Either Clang only emits instructions that work with the old z900 anyway, or
>>>> our emulation in QEMU is imprecise and we allow newer instructions to be
>>>> executed on old models, too.
>>>
>>> Yes, that's currently still done. We once thought about disabling that
>>> (there was a patch from Richard), but decided against it because -- back
>>> then -- the default QEMU model was still very basic and would have
>>> essentially disabled all more recent instructions as default.
>>>
>>> We can most probably do that change soon as we have a "fairly new"
>>> default QEMU CPU model. I can glue it to my z14 change.
>>
>> In case this makes the BIOS crash with old CPUs: when a guest refuses to
>> start because the BIOS was compiled the wrong way for it, configure
>> having told you so back then is not a nice user experience.  Can we do
>> better, with reasonable effort?
> 
> I fear the experience will be as bad as for any guest that is using
> features from a newer cpu level (i.e. random crashes when the guest
> actually tries to use that newer instruction.)
> 
> I see two options:
> - Just try to start and hope that it works.
> - Deprecate any cpu model older than z10.
> 
> Anyone have a better idea? I don't particularly like any of the two.

I think we should simply continue to build the default bios with GCC and 
-mz900. So the normal user (who does not explicitly use the freshly compiled 
binaries but the pre-built ones) will never experience any problem here. The 
Clang builds are (at least right now) rather only meant for us developers to 
check the sources from time to time with this compiler, to see whether it 
detects some additional issues compared to GCC.

  Thomas