Re: [PATCH v3 05/17] qapi: pass QAPISchemaModule to visit_module instead of str

From: John Snow <jsnow@redhat.com>
To: Markus Armbruster <armbru@redhat.com>
Cc: "Marc-André Lureau" <marcandre.lureau@redhat.com>,
	"Cleber Rosa" <crosa@redhat.com>,
	qemu-devel@nongnu.org, "Eduardo Habkost" <ehabkost@redhat.com>
Subject: Re: [PATCH v3 05/17] qapi: pass QAPISchemaModule to visit_module instead of str
Date: Wed, 20 Jan 2021 10:51:40 -0500	[thread overview]
Message-ID: <049e81a7-c7d7-65af-129d-4d0993aff559@redhat.com> (raw)
In-Reply-To: <87eeifu805.fsf@dusky.pond.sub.org>

On 1/20/21 7:07 AM, Markus Armbruster wrote:
> John Snow <jsnow@redhat.com> writes:
> 
>> Modify visit_module to pass the module itself instead of just its
>> name. This allows for future patches to centralize some
>> module-interrogation behavior within the QAPISchemaModule class itself,
>> cutting down on duplication between gen.py and schema.py.
> 
> We've been tempted to make similar changes before (don't worry, I'm not
> building a case for "no" here).
> 

It's fine: you'll probably notice later I don't go the full distance and 
rely on both object and class methods anyway, so this isn't strictly 
needed right now.

(It was not possible to go the full distance without heavier, more 
invasive changes, so...)

> When I wrote the initial version of QAPISchemaVisitor (commit 3f7dc21be,
> 2015), I aimed for a loose coupling of backends and the internal
> representation.  Instead of
> 
>      def visit_foo(self, foo):
>          pass
> 
> where @foo is a QAPISchemaFooBar, I wrote
> 
>      def visit_foo_bar(self, name, info, [curated attributes of @foo]):
>          pass
> 
> In theory, this is nice: the information exposed to the backends is
> obvious, and the backends can't accidentally mutate @foo.
> 
> In practice, it kind of failed right then and there:
> 
>      def visit_object_type(self, name, info, base, members, variants):
>          pass
> 
> We avoid passing the QAPISchemaObjectType (loose coupling, cool!), only
> to pass member information as List[QAPISchemaObjectTypeMember].
> 
> Morever, passing "curated atttibutes" has led to visit_commands() taking
> a dozen arguments.  Meh.
> 
> This had made Eric and me wonder whether we should write off the
> decoupling idea as misguided, and just pass the object instead of
> "curated attributes", always.  Thoughts?
> 

I'm not sure. Just taking the object would avoid a lot of duplicated 
interface typing, and type hints would allow editors to know what fields 
are available.

>>
>> Signed-off-by: John Snow <jsnow@redhat.com>
>> ---
>>   docs/sphinx/qapidoc.py         |  8 ++++----
>>   scripts/qapi/gen.py            | 16 ++++++++++------
>>   scripts/qapi/schema.py         |  4 ++--
>>   tests/qapi-schema/test-qapi.py |  4 ++--
>>   4 files changed, 18 insertions(+), 14 deletions(-)
>>
>> diff --git a/docs/sphinx/qapidoc.py b/docs/sphinx/qapidoc.py
>> index e03abcbb959..f754f675d66 100644
>> --- a/docs/sphinx/qapidoc.py
>> +++ b/docs/sphinx/qapidoc.py
>> @@ -463,11 +463,11 @@ def __init__(self, env, qapidir):
>>           self._env = env
>>           self._qapidir = qapidir
>>   
>> -    def visit_module(self, name):
>> -        if name is not None:
>> -            qapifile = self._qapidir + '/' + name
>> +    def visit_module(self, module):
>> +        if module.name:
> 
> Replacing the "is not None" test by (implicit) "is thruthy" changes
> behavior for the empty string.  Intentional?
> 

Instinctively it was intentional, consciously it wasn't. I was worried 
about what "qapifile" would produce if the string happened to be empty.

> I've had the "pleasure" of debugging empty strings getting interpreted
> like None where they should be interpreted like any other string.
> 

assert module.name, then?

>> +            qapifile = self._qapidir + '/' + module.name
>>               self._env.note_dependency(os.path.abspath(qapifile))
>> -        super().visit_module(name)
>> +        super().visit_module(module)
>>   
>>   
> [...]
> 