Re: [RFC PATCH-for-9.1] qapi: Do not generate commands/events/introspect code for user emulation

["Philippe Mathieu-Daudé" <philmd@linaro.org>]

Hi Markus,

On 5/4/24 07:35, Markus Armbruster wrote:
> Philippe Mathieu-Daudé <philmd@linaro.org> writes:
> 
>> User emulation requires the QAPI types. Due to the command
>> line processing, some visitor code is also used. The rest
>> is irrelevant (no QMP socket).
>>
>> Add an option to the qapi-gen script to allow generating
>> the minimum when only user emulation is being built.
>>
>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
>> ---
>> RFC: Quick PoC for Markus. It is useful for user-only builds.
>> ---
>>   qapi/meson.build     |  6 +++++-
>>   scripts/qapi/main.py | 16 +++++++++++-----
>>   2 files changed, 16 insertions(+), 6 deletions(-)


>> @@ -50,9 +51,10 @@ def generate(schema_file: str,
>>       schema = QAPISchema(schema_file)
>>       gen_types(schema, output_dir, prefix, builtins)
>>       gen_visit(schema, output_dir, prefix, builtins)
>> -    gen_commands(schema, output_dir, prefix, gen_tracing)
>> -    gen_events(schema, output_dir, prefix)
>> -    gen_introspect(schema, output_dir, prefix, unmask)
>> +    if not gen_types_only:
>> +        gen_commands(schema, output_dir, prefix, gen_tracing)
>> +        gen_events(schema, output_dir, prefix)
>> +        gen_introspect(schema, output_dir, prefix, unmask)
> 
> This is the behavior change, everything else is plumbing.  You suppress
> generation of source code for commands, events, and introspection, i.e.
> 
>      qapi-commands*.[ch]
>      qapi-init-commands.[ch]
>      qapi-events*[ch]
>      qapi-introspect.[ch]
> 
> and the associated .trace-events.
> 
> But none of these .c get compiled for a user-only build.
> 
> So, all we save is a bit of build time and disk space: less than 0.1s on
> my machine, ~1.6MiB in ~220 files.  My linux-user-only build tree clocks
> in at 317MiB in ~4900 files, a full build takes me around 30s (real
> time, -j 14 with ccache), so we're talking about 0.5% in disk space and
> 0.3% in build time.

What I want to catch is invalid uses of these headers in user-only
units.

See for example:
https://lore.kernel.org/qemu-devel/20240404194757.9343-5-philmd@linaro.org/

(Actually I have this patch based on that series).

> Moreover, the patch needs work:
> 
>      FAILED: tests/unit/test-qobject-input-visitor.p/test-qobject-input-visitor.c.o
>      cc [...] -c ../tests/unit/test-qobject-input-visitor.c
>      ../tests/unit/test-qobject-input-visitor.c:27:10: fatal error: qapi/qapi-introspect.h: No such file or directory
>         27 | #include "qapi/qapi-introspect.h"
>            |          ^~~~~~~~~~~~~~~~~~~~~~~~

I'd simply skip these tests on user-only builds.

>      FAILED: libqemuutil.a.p/stubs_monitor-core.c.o
>      cc [...] -c ../stubs/monitor-core.c
>      ../stubs/monitor-core.c:3:10: fatal error: qapi/qapi-emit-events.h: No such file or directory
>          3 | #include "qapi/qapi-emit-events.h"
>            |          ^~~~~~~~~~~~~~~~~~~~~~~~~

Doh, this is 
https://lore.kernel.org/qemu-devel/20240404194757.9343-4-philmd@linaro.org/, 
again I forgot:

Based-on: <20240404194757.9343-1-philmd@linaro.org>

> 
> I don't think it's worth the bother.

OK, I'll keep it locally until I finish my full exec/ rework then.

Regards,

Phil.