Re: [RFC PATCH v2 1/2] qapi: Inline qmp_marshal_output() functions

[Markus Armbruster]

Philippe Mathieu-Daudé <philmd@redhat.com> writes:

> On 6/10/21 11:33 AM, Markus Armbruster wrote:
>> Eric Blake <eblake@redhat.com> writes:
>> 
>>> On Wed, Jun 09, 2021 at 08:49:54PM +0200, Philippe Mathieu-Daudé wrote:
>>>> In case we need to use QAPI types but no QAPI command / QAPI event
>>>> actually use them, the generated qmp_marshal_output() function will
>>>> trigger the compiler 'unused-function' warnings.
>>>> To prevent that, emit these functions inlined: the compiler will
>>>> ignore such unused functions.
>>>>
>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com>
>>>> ---
>>>> RFC: No clue about QAPI...
>>>> Tested with GCC. If the compiler is picky we could use the 'unused'
>>>> function attribute.
>>>
>>> And I have no clue if clang will warn about an unused inline function.
>>> Going with the compiler attribute seems safer and just as easy to do
>>> in the same two-line change (remember, the "unused" attribute merely
>>> means "suppress warnings if I don't use this", and not "warn me if I
>>> use it in spite of calling it unused").
>>>
>>>> ---
>>>>  scripts/qapi/commands.py | 4 ++--
>>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/scripts/qapi/commands.py b/scripts/qapi/commands.py
>>>> index 0e13d510547..bbed776a909 100644
>>>> --- a/scripts/qapi/commands.py
>>>> +++ b/scripts/qapi/commands.py
>>>> @@ -91,8 +91,8 @@ def gen_call(name: str,
>>>>  def gen_marshal_output(ret_type: QAPISchemaType) -> str:
>>>>      return mcgen('''
>>>>  
>>>> -static void qmp_marshal_output_%(c_name)s(%(c_type)s ret_in,
>>>> -                                QObject **ret_out, Error **errp)
>>>> +static inline void qmp_marshal_output_%(c_name)s(%(c_type)s ret_in,
>>>> +                                        QObject **ret_out, Error **errp)
>>>
>>> On the other hand, the qapi generator is smart enough to only output
>>> introspection data for qapi types that were actually used by a command
>>> or event, so how is that working, and why is it not also being used to
>>> elide the generation of unused qmp_marshal_output_FOO functions?  This
>>> is where I'll have to defer to Markus.
>> 
>> This is a QAPI generator restriction.  Let me explain.
>> 
>> The qmp_marshal_output_T() are shared by all commands returning T.
>> 
>> The commands may be conditional.  The user is responsible for making T's
>> 'if' the conjunction of the commands'.  See the FIXME in commands.py.
>
> Yes, I noticed the FIXME:
>
>     # FIXME: If T is a user-defined type, the user is responsible
>     # for making this work, i.e. to make T's condition the
>     # conjunction of the T-returning commands' conditions.  If T
>     # is a built-in type, this isn't possible: the
>     # qmp_marshal_output_T() will be generated unconditionally.
>
> Using inline / unused attributes don't invalidate this :)

Generating the unused attribute lets us keep types unconditional even
when the commands returning them are conditional (also takes care of the
built-in case, where we cannot make the type conditional).

However, conditional commands returning an unconditional type is a bit
of a code smell.  In this particular case, the smell seems to lead to a
(minor) issue: too much TPM code is compiled even when CONFIG_TPM is
off.  With the attribute in place, we wouldn't have learned this.

We may still find non-smelly instances of this pattern.  Until then, I'm
a bit reluctant to generate the attribute.

>> If I do this for tpm.json (appended), then tpm.h misses TpmModel when
>> CONFIG_TPM is off, and tpm_backend.h misses TpmType and TpmInfo.  I
>> suspect more TPM code needs to be guarded by CONFIG_TPM.
>
> Yes, this is what I did first, use the code below and add #ifdef'ry,
> but the code becomes ugly and harder to maintain because the enums
> are used in middle of a QOM interface structure:
>
> include/sysemu/tpm.h-37-struct TPMIfClass {
> include/sysemu/tpm.h-38-    InterfaceClass parent_class;
> include/sysemu/tpm.h-39-
> include/sysemu/tpm.h:40:    enum TpmModel model;
> include/sysemu/tpm.h-41-    void (*request_completed)(TPMIf *obj, int ret);
> include/sysemu/tpm.h-42-    enum TPMVersion (*get_version)(TPMIf *obj);
> include/sysemu/tpm.h-43-};
> include/sysemu/tpm.h-44-
>
> If you think using inline / unused attributes is not an option for
> QAPI, then the #ifdef'ry isn't worth it and I'd prefer use v1 which
> doesn't use conditional QAPI suggested by Marc-André.

Ignorant question: why do we want to define QOM type "tpm-if" when
CONFIG_TPM is off?

[...]