Optimize QTimer::singleShot(0, ...) when taking PMF or Functor callable

QTimer::singleShot is optimized for zero timeouts when using the API
taking a string method name. This optimization was not used for the API
taking a PMF or functor. This patch adds it, making the various API
calls behave similarly from a performance point of view.

The approach taken here requires a QObject context object. If none is
available, e.g. a nullptr was passed explicitly, or the
QTimer::singleShot(O, Functor) API was used, the optimization could
not easily be applied. This is not only bad from a performance POV,
but also poses as a potential source for heisenbugs: Using the
different API versions of QTimer::singleShot would use different code
paths internally, which then would not ensure the expected slot call
order. This problem actually existed already when mixing the
string-based slot syntax with PMF/functors in the QTimer::singleShot
API.

This patch overcomes this hurdle and fixes all of the above: When we
encounter a 0ms single shot timer, and no QObject context object is
available, we fall back to the main thread, or create a temporary
QObject for any other thread. The updated and extended benchmark
shows that this is still a significant performance improvement
over using a timer:

********* Start testing of qtimer_vs_qmetaobject *********
Config: Using QtTest library 5.14.0, Qt 5.14.0 (x86_64-little_endian-lp64 shared (dynamic) release build; by GCC 8.2.1 20181127)
PASS   : qtimer_vs_qmetaobject::initTestCase()
PASS   : qtimer_vs_qmetaobject::bench(singleShot_slot)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_slot":
     7.48 msecs per iteration (total: 748, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_pmf)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_pmf":
     7.20 msecs per iteration (total: 720, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_functor)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_functor":
     6.79 msecs per iteration (total: 679, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_functor_noctx)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_functor_noctx":
     6.92 msecs per iteration (total: 693, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_string)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_string":
     7.34 msecs per iteration (total: 735, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_pmf)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_pmf":
     6.90 msecs per iteration (total: 690, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_functor)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_functor":
     6.62 msecs per iteration (total: 662, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_slot)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_slot":
     7.45 msecs per iteration (total: 745, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_pmf)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_pmf":
     7.46 msecs per iteration (total: 747, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_functor)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_functor":
     6.70 msecs per iteration (total: 671, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_functor_noctx)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_functor_noctx":
     13.75 msecs per iteration (total: 1,376, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_string)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_string":
     7.05 msecs per iteration (total: 706, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_pmf)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_pmf":
     6.70 msecs per iteration (total: 670, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_functor)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_functor":
     6.58 msecs per iteration (total: 658, iterations: 100)
PASS   : qtimer_vs_qmetaobject::cleanupTestCase()
Totals: 16 passed, 0 failed, 0 skipped, 0 blacklisted, 20977ms
********* Finished testing of qtimer_vs_qmetaobject *********

Without the change to qtimer.cpp, the results are:

********* Start testing of qtimer_vs_qmetaobject *********
Config: Using QtTest library 5.14.0, Qt 5.14.0 (x86_64-little_endian-lp64 shared (dynamic) release build; by GCC 8.2.1 20181127)
PASS   : qtimer_vs_qmetaobject::initTestCase()
PASS   : qtimer_vs_qmetaobject::bench(singleShot_slot)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_slot":
     7.45 msecs per iteration (total: 745, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_pmf)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_pmf":
     112.84 msecs per iteration (total: 11,285, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_functor)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_functor":
     115.62 msecs per iteration (total: 11,563, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(singleShot_functor_noctx)
RESULT : qtimer_vs_qmetaobject::bench():"singleShot_functor_noctx":
     110.81 msecs per iteration (total: 11,082, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_string)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_string":
     7.04 msecs per iteration (total: 704, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_pmf)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_pmf":
     6.62 msecs per iteration (total: 662, iterations: 100)
PASS   : qtimer_vs_qmetaobject::bench(invokeMethod_functor)
RESULT : qtimer_vs_qmetaobject::bench():"invokeMethod_functor":
     6.62 msecs per iteration (total: 662, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_slot)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_slot":
     7.45 msecs per iteration (total: 746, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_pmf)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_pmf":
     118.42 msecs per iteration (total: 11,842, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_functor)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_functor":
     119.35 msecs per iteration (total: 11,936, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(singleShot_functor_noctx)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"singleShot_functor_noctx":
     130.96 msecs per iteration (total: 13,096, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_string)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_string":
     8.08 msecs per iteration (total: 808, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_pmf)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_pmf":
     6.79 msecs per iteration (total: 680, iterations: 100)
PASS   : qtimer_vs_qmetaobject::benchBackgroundThread(invokeMethod_functor)
RESULT : qtimer_vs_qmetaobject::benchBackgroundThread():"invokeMethod_functor":
     7.49 msecs per iteration (total: 749, iterations: 100)
PASS   : qtimer_vs_qmetaobject::cleanupTestCase()
Totals: 16 passed, 0 failed, 0 skipped, 0 blacklisted, 153995ms
********* Finished testing of qtimer_vs_qmetaobject *********

Additionally, this patch adds a unit test to verify that the slot call
order for 0ms single shot timers is followed while mixing the various
API versions. It fails without this patch but passes now.

Finally, another test is added to verify that using QTimer::singleShot
before a QCoreApplication was constructed is still working properly.

Change-Id: I0d6211554b6198cb3e527be9ec3adc572b1b54ee
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

2 files changed, 30 insertions, 0 deletions

diff --git a/src/corelib/kernel/qobjectdefs.h b/src/corelib/kernel/qobjectdefs.h
index 28674193c3..418b8cf1d2 100644
--- a/src/corelib/kernel/qobjectdefs.h
+++ b/src/corelib/kernel/qobjectdefs.h
@@ -576,6 +576,7 @@ struct Q_CORE_EXPORT QMetaObject
 
 private:
     static bool invokeMethodImpl(QObject *object, QtPrivate::QSlotObjectBase *slot, Qt::ConnectionType type, void *ret);
+    friend class QTimer;
 };
 
 class Q_CORE_EXPORT QMetaObject::Connection {
diff --git a/src/corelib/kernel/qtimer.cpp b/src/corelib/kernel/qtimer.cpp
index 13f027074a..188d529f04 100644
--- a/src/corelib/kernel/qtimer.cpp
+++ b/src/corelib/kernel/qtimer.cpp
@@ -42,6 +42,8 @@
 #include "qabstracteventdispatcher.h"
 #include "qcoreapplication.h"
 #include "qobject_p.h"
+#include "qthread.h"
+#include "qcoreapplication_p.h"
 
 QT_BEGIN_NAMESPACE
 
@@ -343,6 +345,33 @@ void QTimer::singleShotImpl(int msec, Qt::TimerType timerType,
                             const QObject *receiver,
                             QtPrivate::QSlotObjectBase *slotObj)
 {
+    if (msec == 0) {
+        bool deleteReceiver = false;
+        // Optimize: set a receiver context when none is given, such that we can use
+        // QMetaObject::invokeMethod which is more efficient than going through a timer.
+        // We need a QObject living in the current thread. But the QThread itself lives
+        // in a different thread - with the exception of the main QThread which lives in
+        // itself. And QThread::currentThread() is among the few QObjects we know that will
+        // most certainly be there. Note that one can actually call singleShot before the
+        // QApplication is created!
+        if (!receiver && QThread::currentThread() == QCoreApplicationPrivate::mainThread()) {
+            // reuse main thread as context object
+            receiver = QThread::currentThread();
+        } else if (!receiver) {
+            // Create a receiver context object on-demand. According to the benchmarks,
+            // this is still more efficient than going through a timer.
+            receiver = new QObject;
+            deleteReceiver = true;
+        }
+
+        QMetaObject::invokeMethodImpl(const_cast<QObject *>(receiver), slotObj,
+                                      Qt::QueuedConnection, nullptr);
+
+        if (deleteReceiver)
+            const_cast<QObject *>(receiver)->deleteLater();
+        return;
+    }
+
     new QSingleShotTimer(msec, timerType, receiver, slotObj);
 }