[Qt-creator] Couple of questions about the design of Qt Creator

[Elvis Stansvik]

2017-09-11 9:33 GMT+02:00 Eike Ziller <Eike.Ziller at qt.io>:
>
>> On Sep 10, 2017, at 13:09, Elvis Stansvik <elvstone at gmail.com> wrote:
>>
>> Right, the two-phase plugin initialization and object pool is quite
>> well described in the docs:
>>
>>    https://doc-snapshots.qt.io/qtcreator-extending/plugin-lifecycle.html
>>
>> I think my question was a little vague, sorry about that.
>>
>> Consider an example: Some functionality "foo" is to be added to the
>> core plugin, and a plugin Bar written to implement that functionality.
>>
>> Approach 1:
>>
>> - Add an interface IFoo to the core plugin.
>> - Bar implements IFoo and add an instance of its implementation to the
>> pool in its initialize().
>> - The core plugin get the object from the pool (in its
>> extensionsInitialized()), and make use of it.
>>
>> Approach 2:
>>
>> - Add a singleton Foo to the core plugin.
>> - Bar carries out its work using direct calls on Foo::instance().
>>
>> I was just wondering if the second approach is ever used, or if the
>> mechanism set up by the object pool + two-phase initialization is
>> always used. If approach 2 is used, I was interested in what the
>> deciding factor is between the two approaches.
>
> Many good things have already been said, I’m late :)

You're very welcome any time Eike :)

>
> Basically there is also
>
> Approach 3:
>
> - Add an interface IFoo to the core plugin
> - Add a static method for registering instances of IFoo to core plugin
> - Bar implements IFoo and registers an instance of its implementation via the method above
> - The core plugin gets the object(s) that have been registered through the static method (in its extensionsInitialized or later)

Ah yes of course.

>
> Approach 1 works when there should not be a hard runtime dependency between the plugins (at compile time the interface header is needed). For making this work without a link dependency, the interface must use Q_DECLARE_INTERFACE. I think the only place where it actually used this way is with CodePaster::Service, to avoid a runtime dependency between CodePaster and DiffEditor plugins. Historically we used that a lot also when there actually already was a hard runtime dependency between the plugins, which is problematic because of the performance issue that André hinted at.

Aha, I was actually going to ask if plugin dependencies (as declared
in their metadata) always implies a link time dependency as well,
because when I looked at qtcreator.pri (the loops at the bottom that
add to LIBS), it looked to me like it does. But you're saying that if
Q_DECLARE_INTERFACE is used in the depended-upon plugin's interface
class, then there will be no link time dependency introduced, despite
-l<dep> being always being added to the link line for the depending
plugin?

>
> In Approach 2 we actually moved towards using mostly classes with static methods, only using the instance() when we actually need an instance (e.g. for connecting to signals). It works well when a hard runtime dependency is ok and Bar is the driving force for the work to be done. I would say that we do this a lot, e.g. ActionManager, EditorManager, …, but I don’t know if you meant that kind of interaction ;)

Yes, my description of 2 was a bit vague, but this was what I meant I think.

Having static methods that delegate to the instance() does make the
calls more readable, so I like that.

>
> Approach 3 works well if a hard runtime dependency between the plugins is ok. We still experiment with ways for avoiding the monotonous work of adding add/registerXXX, remove/unregisterXXX, methods for managing the list of implementations. One experiment is Utils::ObjectPool. One could add a public static instance of that to the core plugin in your example, and then call Core::foos.addObject(myIFooImpl) from Bar plugin.

Ah, I hadn't seen that one. That makes sense. I see that it also
defaults to destroying the objects when the pool destructs, which
makes sense since looking at the number of IPlugin::addObject calls vs
IPlugin::addAutoReleasedObject, it looks like most of the time you
want the auto-destruction behavior.

>
> I’d also prefer having fewer hard runtime dependencies, but it is a struggle… Qt helps a bit, but it is still not nice.
> There is the Q_DECLARE_INTERFACE/Q_INTERFACES combo + object pool, there is invokeMethod + object pool, and the other thing that we also use is custom QObject properties (e.g. CodePaster looks for a custom property “plainText” in the current IEditor, as a fallback for finding pasteable text).

Alright, thanks for the insight!

Elvis

>
> Br, Eike
>