Re: [RFC PATCH 2/2] gitlab-ci: Add a job building TCI with Clang

[Alex Bennée]

Wataru Ashihara <wataash@wataash.com> writes:

> On 2021/01/21 22:27, Philippe Mathieu-Daudé wrote:
>> On 1/21/21 1:02 PM, Daniel P. Berrangé wrote:
>>> On Thu, Jan 21, 2021 at 12:48:21PM +0100, Philippe Mathieu-Daudé wrote:
>>>> On 1/21/21 12:21 PM, Daniel P. Berrangé wrote:
>>>>> On Thu, Jan 21, 2021 at 12:18:18PM +0100, Philippe Mathieu-Daudé wrote:
>>>>>> On 1/21/21 11:32 AM, Daniel P. Berrangé wrote:
>>>>>>> On Thu, Jan 21, 2021 at 11:08:29AM +0100, Thomas Huth wrote:
>>>>>>>> On 10/01/2021 17.27, Philippe Mathieu-Daudé wrote:
>>>>>>>>> Split the current GCC build-tci job in 2, and use Clang
>>>>>>>>> compiler in the new job.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
>>>>>>>>> ---
>>>>>>>>> RFC in case someone have better idea to optimize can respin this 
>>>>>>>>> patch.
>>>>>>>>>
>>>>>>>>>   .gitlab-ci.yml | 22 ++++++++++++++++++++--
>>>>>>>>>   1 file changed, 20 insertions(+), 2 deletions(-)
>>>>>>>>
>>>>>>>> I'm not quite sure whether we should go down this road ... if we 
>>>>>>>> wanted to
>>>>>>>> have full test coverage for clang, we'd need to duplicate *all* jobs 
>>>>>>>> to run
>>>>>>>> them once with gcc and once with clang. And that would be just 
>>>>>>>> overkill.
>>>>>>>>
>>>>>>>> I think we already catch most clang-related problems with the clang 
>>>>>>>> jobs
>>>>>>>> that we already have in our CI, so problems like the ones that you've 
>>>>>>>> tried
>>>>>>>> to address here should be very, very rare. So I'd rather vote for not
>>>>>>>> splitting the job here.
>>>>>>>
>>>>>>> We can't possibly cope with the fully expanded matrix of what are
>>>>>>> theoretically possible combinations. Thus I think we should be guided
>>>>>>> by what is expected real world usage by platforms we target.
>>>>>>>
>>>>>>> Essentially for any given distro we're testing on, our primary focus
>>>>>>> should be to use the toolchain that distro will build QEMU with.
>>>>>>>
>>>>>>> IOW, for Windows and Linux distros our primary focus should be GCC,
>>>>>>> while for macOS, and *BSD, our focus should be CLang.
>>>>>>
>>>>>> Sounds good.
>>>>>>
>>>>>> Do we need a TCI job on macOS then?
>>>>>
>>>>> TCI is only relevant if there is no native TCG host impl.
>>>>>
>>>>> macOS only targets aarch64 and x86_64, both of which have TCG, so there
>>>>> is no reason to use TCI on macOS  AFAICT
>>>>
>>>> Yes, fine by me, but Wataru Ashihara reported the bug... ¯\_(ツ)_/¯
>>>
>>> It doesn't look like they were using macOS - the message suggests
>>> Ubuntu host, and AFAIK, all Ubuntu architectures have support
>>> for TCG, so using TCI shouldn't have been required in the first
>>> place.
>>>
>>> I guess we could benefit from a TCI job of some kind that uses
>>> CLang on at least 1 platform, since none exists.
>>>
>>> This does yet again open up the question of whether we should be
>>> supporting TCI at all in this particular user's scenario though,
>>> since both KVM and TCG are available on Ubuntu x86 hosts already.
>> 
>> I understand Stefan envisions other use cases for TCI, which is
>> why it is still maintained. See:
>> https://www.mail-archive.com/qemu-devel@nongnu.org/msg461131.html
>> 
>> I agree with your previous comment:
>>> we should be guided by what is expected real world usage by
>>> platforms we target. Essentially for any given distro we're
>>> testing on, our primary focus should be to use the toolchain
>>> that distro will build QEMU with.
>> 
>> This rarely used config does not justify adding yet another CI job.
>> 
>> Thanks,
>> 
>> Phil.
>> 
>> 
>
> Actually I use TCI also on macOS. Like the use case quoted by Philippe,
> there're even other reasons to use TCI:
>
> 1. Learning TCG ops.

Except it's only a subset of ops. Really interesting newer ones using
the TCGv_vec types are entirely absent.

> 2. Debugging QEMU with gdb. e.g. diagnose codegen or stepping into
>    helper functions from tci.c:tcg_qemu_tb_exec().

I do this quite often with TCG so I'm curious as to what the difference
is here?

> 3. Guest instruction tracing. TCI is faster than TCG or KVM when tracing
>    the guest ops [1]. I guess qira is using TCI for this reason [2].

How are you doing instruction tracing with TCG? Using the plugin
interface?

I think there probably is a roll for a *guest* interpreter given the
amount of code that is translated only to be run once. However it would
be a fairly large undertaking.

> [1]: https://twitter.com/wata_ash/status/1352899988032942080
> [2]: https://github.com/geohot/qira/blob/v1.3/tracers/qemu_build.sh#L55


-- 
Alex Bennée