In preparation of Qt6 move away from pre-C++11 macros.
Change-Id: I44126693c20c18eca5620caab4f7e746218e0ce3
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Removing connections and resizing the signal vector now
happens in a way that will not interfere with activate(),
as the old objects won't get deleted if we are somewhere
inside a signal emission.
This means that we don't need to lock the senders mutex
in activate anymore, as long as the reference counting
on the ConnectionData is atomic and we are in the senders
thread.
This implies that we now need to lock the receivers mutex in
queued_activate() abd blocking queued activation to ensure
it hasn't been deleted while we are emitting.
In addition, some precautions need to be taken to not read
from the receiver without holding the lock, as it could get
deleted while we're activating (if it's in a different thread).
To make that possible store the receivers thread id in the
connection data.
Use atomic pointers for all variables that can get modified
with the signalSlotLock() held and that are being read
without the lock being held.
This gives us a very nice additional speed improvement
for signal emissions.
without change with change
string based connect: 3287 2436
pointer based connect: 3941 3265
not connected: 403 400
disconnected: 460 489
5 slots connected: 9112 4515
Change-Id: Ib7324bb74c389dcc3b6581a03c31469a6e589fc2
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
And use it to clean up the reimplementations in Qt DBus.
Change-Id: I8e3fe35e8db6405cbcbfb45b42a8f2efecc1cef0
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Put all connections that get disconnected into a singly
linked orphaned list.
Whenever the refcount on the connectionData drops down to
one, this list can safely be cleared, even with the planned
removal of locking in activate().
Use an id integer in the connection to acoid activating newly
added connections.
Fixes: QTBUG-72649
Change-Id: Ide3d116ae7fc9ca497598c1c2b71d43b4339c92d
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
When checking whether a slot is connected to a signal, we need to
hold the signalSlotLock to be sure about the answer, or we can
get crashes when a connection gets removed while doing the check.
The check in activate() can handle some uncertainty as it's only
a shortcut to the longer path.
Fixes: QTBUG-74604
Change-Id: I3fc822455fbadc0223ef68632f5fb3df3ff3e86d
Reviewed-by: Aapo Keskimolo <aapo.keskimolo@qt.io>
It Q_ASSERTs that argv is not a nullptr, so it would fail when an
unconnected signal with 0 arguments was emitted.
Change-Id: I5dd810fbeea5b6b511eff4705efdaa6a55739604
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Conflicts:
tests/auto/widgets/itemviews/qheaderview/tst_qheaderview.cpp
Added tests/auto/testlib/selftests/expected_crashes_5.txt to work
round the output of the crashes test (which exercises UB, see
QTBUG-73903) being truncated on one test platform.
Change-Id: I9cd3f2639b4e50c3c4513e14629a40bdca8f8273
Add a simple private QBasicMutexLocker class, and let the
QOrderedMutexLocker operate on a QBasicMutex.
This allows the compiler to inline more things when handling
connections and speeds up activate() a bit more.
without change with change
string based connect: 3621 3368
pointer based connect: 4341 3919
not connected: 433 437
disconnected: 551 538
Change-Id: If979337891178aaeb0b3340b6d4f68b6f86b0260
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
The main difference is that QObject itself also holds on reference
on the structure.
Also rename the orphaned flag to objectDeleted for clarity.
Change-Id: Ief9b9ff9c8b9cc3630dcfd29806ed24cd07150e4
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Adn create that data structure on demand on the heap.
This reduces the size of QObjectPrivate if there are no
connections. If we have connections, it'll use the same
amount of allocations and memory as before.
Change-Id: I900f6980a2cd8a5f72c3ad18697b5dd49100217d
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Squeezes another percent of performance out of
QMetaObject::activate().
Change-Id: I620b8c578681280efcc9bec50cfb1020d2afc928
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Refactor activate(), so that we eliminate almost all
checks for signal hooks in the common case.
Here are the benchmark numbers showing the improvement
for 100M signal emissions
without change with change
string based connect: 3836 3693
pointer based connect: 4571 4510
not connected: 479 433
disconnected: 559 522
Change-Id: I394e6ea5d5bc96e298e8cc0c763eed78c8041876
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Measurements show that it's just almost as fast to simply query
the connectionlist directly and avoid both the memory
overhead of the bitfield and the associated bookkeeping.
For connected signals, the difference is not relevant at all.
With a signal that was never connected, removing the bitfield will
cause signal emission to be ~2.5% faster. And if you ever disconnect
from a signal, the bitfields might not be accurate and this can
cause a major slowdown.
Here are some numbers to validate this. All times are measured
in ms for 100M signal emissions:
without change with change
string based connect: 3817 3836
pointer based connect: 4552 4571
not connected: 493 479
disconnected: 2113 559
Change-Id: Ia2c85036afaa7f991b883c8ff812f69cf4580f7e
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: hjk <hjk@qt.io>
And simply emit the signal spy and tracing callbacks in
that code path as well.
Change-Id: I17f65055c7044caf1be58fac94bb7fe3487f3060
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Does not make much sense now that we can connect to lambda functions
[ChangeLog][QtCore][QSignalMapper] QSignalMapper is now marked
as deprecated.
Change-Id: I89135f23fdf16b42142a125eb7c9a86084c90bfc
Reviewed-by: Sze Howe Koh <szehowe.koh@gmail.com>
Compiler support for lambda functions and variadic templates is
required since Qt 5.7, so no need to mention in the documentation
what happens if the compiler doesn't support it.
Change-Id: I5caeaa0bd7f0edce81e22e22964e0b7dd042c719
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Solves a data race found by TSan.
Since thread and threadId are QAtomicPointer, I've removed the explicit
initialization in the QThreadData constructor
Task-number: QTBUG-58855
Change-Id: I4139d5f93dcb4b429ae9fffd14a34082f2683f76
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
In client code I often see code like:
startTimer(1000); //ms
Let the code to be self-explaining. So provide overload
method that takes std::chrono::milliseconds as arg.
QTimer already has std::chrono support, but QObject does not.
Change-Id: Ib348612ce35f1a997b4816fe9e864775cbcbec16
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
We now have qtHookData, which is a better (and cheaper) way to add hooks
for object creation and deletion. For binary-compatibility reasons we
cannot remove it in Qt5.
Change-Id: Iecd9f4e1195f90279c395845fa26c6301b67b9a1
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
qt_find_obj_child is a vestige from Qt3. It is not used nor
exported. It can safely be removed.
Change-Id: Ief6164c361a369559e067b99880c357e421e6342
Reviewed-by: Martin Smith <martin.smith@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
They don't modify "this", so they should be const. Mark
the existing non-const overloads for removal in Qt 6.
[ChangeLog][QtCore][QObject] QObject::dumpObjectInfo and
QObject::dumpObjectTree now have const overloads.
Change-Id: If9fb15692d2d1536930f86d043d688236d4b778a
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
It makes very little sense that one needs to have a debug build
of Qt to debug QObject issues in application code. At this date
we don't even offer debug builds for Linux systems, and anyhow
one might want to debug an application running against a
release build of Qt.
[ChangeLog][QtCore][QObject] QObject::dumpObjectInfo and
QObject::dumpObjectTree are now fully functional even in a non-debug
build of Qt.
Change-Id: Ifddd3023ffc82f3dc3928a7a94d4970e2fb1b44a
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Q_NAMESPACE is useful to add Q_ENUM_NS/Q_ENUMS, Q_FLAG_NS/Q_FLAGS and
Q_CLASSINFO to a namespace.
[ChangeLog] Added Q_NAMESPACE which can be used to add Q_ENUM_NS/
Q_ENUMS, Q_FLAG_NS/Q_FLAGS and Q_CLASSINFO to a namespace
Task-number: QTBUG-54981
Change-Id: Ic61b972794063e77134681fb347d6c4acddcdb44
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
One of the good features of the new connection style is that
implicit conversion is performed for the connection arguments.
However, this is also a bad feature when it comes to the old
C remnants in the C++ language: for instance, doubles implicitly
convert to ints, possibly losing precision (and GCC/Clang do not
even warn about those under -Wall, only MSVC does) or even
triggering undefined behavior.
For this reason, when using braced initialization, C++11
disables narrowing conversions or floating/integral conversions.
Use this feature when checking the arguments of a PMF-style
signal/slot connection. Technically this makes the program
ill-formed, however GCC still accepts it (but at least
warns under -Wall).
Hence, add a way to disable these implicit conversions.
This is a opt-in and guarded by a macro, as it's a source
incompatible change.
[ChangeLog][QtCore][QObject] The
QT_NO_NARROWING_CONVERSIONS_IN_CONNECT macro has been added.
When using the new connection syntax (PMF-based) this macro
makes it illegal to narrow the arguments carried by the signal,
and/or to perform floating point to integral implicit
conversions on them. When the macro is defined,
depending on your compiler a QObject::connect() statement
triggering such conversions will now fail to compile.
Change-Id: Ie17eb3e66ce0cd780138e60d8bb7da815a4ada83
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Conflicts:
configure
5.7 now supports clang on android; but dev re-worked configure
src/gui/kernel/qevent.h
One side renamed a parameter of a constructor; the other added an
alternate constructor on the next line. Applied the rename to both
for consistency.
tests/auto/tools/moc/tst_moc.cpp
Each side added a new test at the end.
.qmake.conf
Ignored 5.7's change to MODULE_VERSION.
configure.json
No conflict noticed by git; but changes in 5.7 were needed for the
re-worked configure to accommodate 5.7's stricter handling of C++11.
Change-Id: I9cda53836a32d7bf83828212c7ea00b1de3e09d2
There was a race condition between QObject::disconnect() and
QMetaObject::activate() which can occur if there are multiple
BlockingQueued connections to one signal from different threads and
they connect/disconnect their connections often.
What can happen in this case is:
T1 is in activate() method and T2 is in disconnect() method
T1 T2
locks sender mutex
selects next connection
unlocks sender mutex
locks sender mutex
sets isSlotObject to false
creates QMetaCallEvent derefs connection
posts event
Two things can happen here:
1. Connection can still be valid, but it will have isSlotObject==false
and callFunction will be used instead of slotObj
2. Connection can already be invalid
To fix it mutex unlock should be moved after QMetaCallEvent creation.
Also there is another case, when we don't disconnect but delete the
receiver object. In this case it can already be invalid during
postEvent, so we need to move mutex unlock after postEvent.
Change-Id: I8103798324140ee11de5b4e10906562ba878ff8b
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
In all other forms of disconnecting this is done. We also know the
signal index, so there is no reason not to do this.
Change-Id: Ic8b042cd8f45dbff74b42ee30c384a84bef78b20
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
...if a PMF connection had already happened. Since UniqueConnection
isn't implemented for non-PMFs (functors and lambdas aren't comparable,
even if static member functions or non-member functions are), we pass a
null pointer for comparison argument. The disconnect() code already
protected against a null pointer there, but not the connect code path
with Qt::UniqueConnection
Change-Id: I87e17314d8b24ae983b1fffd145324beced0494d
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Dario Freddi <dario.freddi@ispirata.com>
as in other examples which come with plugins, use an additional
hierarchy level which contains the app and plugin subdirs.
Change-Id: I2487755967aa3474c337c8c8af10be49627b63d0
Reviewed-by: Topi Reiniö <topi.reinio@theqtcompany.com>
When using QML, it quite often happens that only the QML engine is
connected to a signal, and no C++ handlers. By splitting up the
fast-exit case and handling QML separately, we can prevent a call to
QThread::currentThreadId, and locking+unlocking the mutex.
On x86 this saves ~130 instructions according to valgrind.
Change-Id: I947fe42afe351922339ac982a6d498bc2f7b5192
Reviewed-by: Simon Hausmann <simon.hausmann@theqtcompany.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
