On 02/06/2025 4:50 pm, Jan Beulich wrote:
> On 02.06.2025 17:36, Andrew Cooper wrote:
>> On 02/06/2025 4:09 pm, Michal Orzel wrote:
>>> These tests are supposed to run on target. HOSTCC can be different than
>>> CC (when cross-compiling). At the moment, tests installation would put
>>> a binary of a wrong format in the destdir.
>>>
>>> Fixes: e90580f25bd7 ("vpci: introduce basic handlers to trap accesses to
>>> the PCI config space")
>>> Signed-off-by: Michal Orzel <michal.or...@amd.com>
>> Oh. This didn't explode in GitlabCI because there's no ARM version of
>> *-tools-tests-*.
>>
>> Can we fix that too please, seeing as there is a real ARM board?
>>
>> Also, I guess we have to finally sort out the CC vs HOSTCC debate.
> I think the situation here makes pretty clear that HOSTCC is almost always
> wrong to use for tests/. The emulator test harness is special in that it (in
> principle) needs a target compiler (CC) and additionally an x86 one (with no
> present representation). The present way of (partly) distinguishing the two
> by using CC and HOSTCC was assigning wrong meaning to one (perhaps both) of
> them. The (or maybe just my) problem is that in the toolchain world it is
> build, host, and target which are distinguished. As per Michal's description
> my understanding is that HOSTCC is matching "build" there, not "host".
vPCI is a true unit test. It can be run in the build environment.
Installing it here was to get it run somewhere.
x86_emulator is weird. It can be run in the build environment, but it
is also dependent on CPUID so wants running on target environments too.
Everything else (IIRC) needs a real dom0 to run in, because they make
hypercalls.
~Andrew
