Albert Albert - 4 months ago 55
C++ Question

How to execute a functor in a given thread in Qt, GCD-style?

In ObjC with GCD, there is a way of executing a lambda in any of the threads that spin an event loop. For example:

dispatch_sync(dispatch_get_main_queue(), ^{ /* do sth */ });


or:

dispatch_async(dispatch_get_main_queue(), ^{ /* do sth */ });


It executes something (equivalent to
[]{ /* do sth */ }
in C++) in the main thread's queue, either blocking or asynchronously.

How can I do the same in Qt?

From what I have read, I guess the solution would be somehow to send a signal to some object of the main thread. But what object? Just
QApplication::instance()
? (That is the only object living in the main thread at that point.) And what signal?




From the current answers and my current research, it really seems that I need some dummy object to sit in the main thread with some slot which just waits to get in some code to execute.

So, I decided to subclass
QApplication
to add that. My current code, which doesn't work (but maybe you can help):

#include <QApplication>
#include <QThread>
#include <QMetaMethod>
#include <functional>
#include <assert.h>

class App : public QApplication
{
Q_OBJECT

public:
App();

signals:

public slots:
void genericExec(std::function<void(void)> func) {
func();
}

private:
// cache this
QMetaMethod genericExec_method;
public:
void invokeGenericExec(std::function<void(void)> func, Qt::ConnectionType connType) {
if(!genericExec_method) {
QByteArray normalizedSignature = QMetaObject::normalizedSignature("genericExec(std::function<void(void)>)");
int methodIndex = this->metaObject()->indexOfSlot(normalizedSignature);
assert(methodIndex >= 0);
genericExec_method = this->metaObject()->method(methodIndex);
}
genericExec_method.invoke(this, connType, Q_ARG(std::function<void(void)>, func));
}

};

static inline
void execInMainThread_sync(std::function<void(void)> func) {
if(qApp->thread() == QThread::currentThread())
func();
else {
((App*) qApp)->invokeGenericExec(func, Qt::BlockingQueuedConnection);
}
}

static inline
void execInMainThread_async(std::function<void(void)> func) {
((App*) qApp)->invokeGenericExec(func, Qt::QueuedConnection);
}

Answer

It is certainly possible. Any solution will center on delivering an event that wraps the functor to a consumer object residing in the desired thread. We shall call this operation metacall posting. The particulars can be executed in several ways.

TL;DR for the main thread or the thread of a given QObject

// Qt 5/4 - preferred, has least allocations
template <typename F>
static void postToThread(F && fun, QObject * obj = qApp) {
  struct Event : public QEvent {
    F fun;
    Event(F && fun) : QEvent(QEvent::None), fun(std::move(fun)) {}
    ~Event() { fun(); }
  };
  QCoreApplication::postEvent(obj, new Event(std::forward(fun)));
}

// Qt 5 - alternative version
template <typename F>
static void postToThread(F && fun, QObject * obj = qApp) {
   QObject src;
   QObject::connect(&src, &QObject::destroyed, obj, std::move(fun), Qt::QueuedConnection);
}

QThread t;
QObject o;
o.moveToThread(&t);

// Execute in given object's thread
postToThread([&]{ o.setObjectName("hello"); }, &o);

// Execute in the main thread
postToThread([]{ qDebug() << "hello"; });

Common Code

Let's define our problem in terms of the following common code. The simplest solutions will post the event to either the application object, iff the target thread is the main thread, or to an event dispatcher for any other given thread. Since the event dispatcher will exist only after QThread::run has been entered, we indicate the requirement for the thread to be running by returning true from needsRunningThread.

#ifndef HAS_FUNCTORCALLCONSUMER
namespace FunctorCallConsumer {
   bool needsRunningThread() { return true; }
   QObject * forThread(QThread * thread) {
      Q_ASSERT(thread);
      QObject * target = thread == qApp->thread()
            ? static_cast<QObject*>(qApp) : QAbstractEventDispatcher::instance(thread);
      Q_ASSERT_X(target, "postMetaCall", "the receiver thread must have an event loop");
      return target;
   }
}
#endif

The metacall posting functions, in their simplest form, require the functor call consumer to provide object for a given thread, and instantiate the functor call event. The implementation of the event is still ahead of us, and is the essential difference between various implementations.

The second overload takes a rvalue reference for the functor, potentially saving a copy operation on the functor. This is helpful if the continuation contains data that is expensive to copy.

#ifndef HAS_POSTMETACALL
void postMetaCall(QThread * thread, const std::function<void()> & fun) {
   auto receiver = FunctorCallConsumer::forThread(thread);
   QCoreApplication::postEvent(receiver, new FunctorCallEvent(fun, receiver));
}

void postMetaCall(QThread * thread, std::function<void()> && fun) {
   auto receiver = FunctorCallConsumer::forThread(thread);
   QCoreApplication::postEvent(receiver,
                               new FunctorCallEvent(std::move(fun), receiver));
}
#endif

For demonstration purposes, the worker thread first posts a metacall to the main thread, and then defers to QThread::run() to start an event loop to listen for possible metacalls from other threads. A mutex is used to allow the thread user to wait in a simple fashion for the thread to start, if necessitated by the consumer's implementation. Such wait is necessary for the default event consumer given above.

class Worker : public QThread {
   QMutex m_started;
   void run() {
      m_started.unlock();
      postMetaCall(qApp->thread(), []{
         qDebug() << "worker functor executes in thread" << QThread::currentThread();
      });
      QThread::run();
   }
public:
   Worker(QObject * parent = 0) : QThread(parent) { m_started.lock(); }
   ~Worker() { quit(); wait(); }
   void waitForStart() { m_started.lock(); m_started.unlock(); }
};

Finally, we start the above worker thread that posts a metacall to the main (application) thread, and the application thread posts a metacall to the worker thread.

int main(int argc, char *argv[])
{
   QCoreApplication a(argc, argv);
   a.thread()->setObjectName("main");
   Worker worker;
   worker.setObjectName("worker");
   qDebug() << "worker thread:" << &worker;
   qDebug() << "main thread:" << QThread::currentThread();
   if (FunctorCallConsumer::needsRunningThread()) {
      worker.start();
      worker.waitForStart();
   }
   postMetaCall(&worker, []{ qDebug() << "main functor executes in thread" << QThread::currentThread(); });
   if (!FunctorCallConsumer::needsRunningThread()) worker.start();
   QMetaObject::invokeMethod(&a, "quit", Qt::QueuedConnection);
   return a.exec();
}

The output will look approximately as follows in all implementations. The functors cross the threads: the one created in the main thread is executed in the worker thread, and vice-versa.

worker thread: QThread(0x7fff5692fc20, name = "worker") 
main thread: QThread(0x7f86abc02f00, name = "main") 
main functor executes in thread QThread(0x7fff5692fc20, name = "worker") 
worker functor executes in thread QThread(0x7f86abc02f00, name = "main") 

Qt 5 Solution Using a Temporary Object as The Signal Source

The simplest approach for Qt 5 is to use a temporary QObject as a signal source, and connect the functor to its destroyed(QObject*) signal. When postMetaCall returns, the signalSource gets destructed, emits its destroyed signal, and posts the metacall to the proxy object.

This is perhaps the most concise and straightforward implementation in the C++11 style. The signalSource object is used in the C++11 RAII fashion for the side effects of its destruction. The phrase "side effects" has a meaning within C++11's semantics and should not be interpreted to mean "unreliable" or "undesirable" - it's anything but. QObject's contract with us is to emit destroyed sometime during the execution of its destructor. We're more than welcome to use that fact.

#include <QCoreApplication>
#include <QThread>
#include <QAbstractEventDispatcher>
#include <QDebug>
#include <functional>

namespace FunctorCallConsumer { QObject * forThread(QThread*); }

#define HAS_POSTMETACALL
void postMetaCall(QThread * thread, const std::function<void()> & fun) {
   QObject signalSource;
   QObject::connect(&signalSource, &QObject::destroyed,
                    FunctorCallConsumer::forThread(thread), [=](QObject*){ fun(); });
}
// Common Code follows here

If we only intend to post to the main thread, the code becomes almost trivial:

void postToMainThread(const std::function<void()> & fun) {
  QObject signalSource;
  QObject::connect(&signalSource, &QObject::destroyed, qApp, [=](QObject*){
    fun();
  });
}

Qt 4/5 Solution Using QEvent Destructor

The same approach can be applied to QEvent directly. The event's virtual destructor can call the functor. The events are deleted right after they are delivered by the consumer object's thread's event dispatcher, so they always execute in the right thread. This will not change in Qt 4/5, and is unlikely to change in Qt 6 as well.

#include <QCoreApplication>
#include <QThread>
#include <QEvent>
#include <QAbstractEventDispatcher>
#include <QDebug>
#include <functional>

class FunctorCallEvent : public QEvent {
   std::function<void()> m_fun;
   QThread * m_thread;
public:
   FunctorCallEvent(const std::function<void()> & fun, QObject * receiver) :
      QEvent(QEvent::None), m_fun(fun), m_thread(receiver->thread()) {}
   FunctorCallEvent(std::function<void()> && fun, QObject * receiver) :
      QEvent(QEvent::None), m_fun(std::move(fun)), m_thread(receiver->thread()) { qDebug() << "move semantics"; }
   ~FunctorCallEvent() {
      if (QThread::currentThread() == m_thread)
         m_fun();
      else
         qWarning() << "Dropping a functor call destined for thread" << m_thread;
   }
};
// Common Code follows here

To post to main thread only, things become even simpler:

class FunctorCallEvent : public QEvent {
   std::function<void()> m_fun;
public:
   FunctorCallEvent(const std::function<void()> & fun) :
      QEvent(QEvent::None), m_fun(fun) {}
   FunctorCallEvent(std::function<void()> && fun, QObject * receiver) :
      QEvent(QEvent::None), m_fun(std::move(fun)) {}
   ~FunctorCallEvent() {
      m_fun();
   }
};

void postToMainThread(const std::function<void()> & fun) {
   QCoreApplication::postEvent(qApp, new FunctorCallEvent(fun);
}

void postToMainThread(std::function<void()> && fun) {
   QCoreApplication::postEvent(qApp, new FunctorCallEvent(std::move(fun)));
}

Qt 5 Solution Using the Private QMetaCallEvent

The functor can be wrapped in the Qt 5 slot object payload of the QMetaCallEvent. The functor will be invoked by QObject::event, and thus can be posted to any object in the target thread. This solution uses the private implementation details of Qt 5.

#include <QCoreApplication>
#include <QThread>
#include <QAbstractEventDispatcher>
#include <QDebug>
#include <private/qobject_p.h>
#include <functional>

class FunctorCallEvent : public QMetaCallEvent {
public:
   template <typename Functor>
   FunctorCallEvent(const Functor & fun, QObject * receiver) :
      QMetaCallEvent(new QtPrivate::QFunctorSlotObject<Functor, 0, typename QtPrivate::List_Left<void, 0>::Value, void>(fun),
                     receiver, 0, 0, 0, (void**)malloc(sizeof(void*))) {}
   // Metacalls with slot objects require an argument array for the return type, even if it's void.
};
// Common Code follows here

Qt 4/5 Solution Using a Custom Event and Consumer

We reimplement the event() method of the object, and have it call the functor. This calls for an explicit event consumer object in each thread that the functors are posted to. The object is cleaned up when its thread is finished, or, for the main thread, when the application instance is destructed. It works on both Qt 4 and Qt 5. The use of rvalue references avoids copying of the temporary functor.

#include <QCoreApplication>
#include <QThread>
#include <QEvent>
#include <QMap>
#include <QMutex>
#include <QDebug>
#include <functional>

class FunctorCallEvent : public QEvent {
   std::function<void()> m_fun;
public:
   FunctorCallEvent(const std::function<void()> & fun, QObject *) :
      QEvent(QEvent::None), m_fun(fun) {}
   FunctorCallEvent(std::function<void()> && fun, QObject *) :
      QEvent(QEvent::None), m_fun(fun) { qDebug() << "move semantics"; }
   void call() { m_fun(); }
};

#define HAS_FUNCTORCALLCONSUMER
class FunctorCallConsumer : public QObject {
   typedef QMap<QThread*, FunctorCallConsumer*> Map;
   static QObject * m_appThreadObject;
   static QMutex m_threadObjectMutex;
   static Map m_threadObjects;
   bool event(QEvent * ev) {
      if (!dynamic_cast<FunctorCallEvent*>(ev)) return QObject::event(ev);
      static_cast<FunctorCallEvent*>(ev)->call();
      return true;
   }
   FunctorCallConsumer() {}
   ~FunctorCallConsumer() {
      qDebug() << "consumer done for thread" << thread();
      Q_ASSERT(thread());
      QMutexLocker lock(&m_threadObjectMutex);
      m_threadObjects.remove(thread());
   }
   static void deleteAppThreadObject() {
      delete m_appThreadObject;
      m_appThreadObject = nullptr;
   }
public:
   static bool needsRunningThread() { return false; }
   static FunctorCallConsumer * forThread(QThread * thread) {
      QMutexLocker lock(&m_threadObjectMutex);
      Map map = m_threadObjects;
      lock.unlock();
      Map::const_iterator it = map.find(thread);
      if (it != map.end()) return *it;
      FunctorCallConsumer * consumer = new FunctorCallConsumer;
      consumer->moveToThread(thread);
      if (thread != qApp->thread())
         QObject::connect(thread, SIGNAL(finished()), consumer, SLOT(deleteLater()));
      lock.relock();
      it = m_threadObjects.find(thread);
      if (it == m_threadObjects.end()) {
         if (thread == qApp->thread()) {
            Q_ASSERT(! m_appThreadObject);
            m_appThreadObject = consumer;
            qAddPostRoutine(&deleteAppThreadObject);
         }
         m_threadObjects.insert(thread, consumer);
         return consumer;
      } else {
         delete consumer;
         return *it;
      }
   }
};

QObject * FunctorCallConsumer::m_appThreadObject = nullptr;
QMutex FunctorCallConsumer::m_threadObjectMutex;
FunctorCallConsumer::Map FunctorCallConsumer::m_threadObjects;
// Common Code follows here