In a later commit WorkerThread should also replace classes similar to
ImageLoader: PageLoader, ImageLoaderGL and ImageLoaderByteArrayGL.
Bugs fixed:
1. Eliminated a data race between ImageLoader::run() and
ComicFlow::updateImageData()->ImageLoader::result(). Specifically when
ImageLoader::busy() returns false, then ImageLoader::run() sets
ImageLoader::working to true, loads the image and starts assigning it to
ImageLoader::img, while ImageLoader::result() is accessed without
locking from updateImageData().
Making ImageLoader::working atomic is clearly insufficient to eliminate
this data race. The fix is to set 'working' to true immediately and
synchronously as soon as a new task is assigned to the worker.
2. Replaced thread termination with graceful thread exit. ComicFlow
destructor called QThread::terminate(), using which is discouraged by Qt
documentation. The application exited without errors in Release mode.
In Debug mode, however, it received the SIG32 signal on exit and printed
the following warning - "QWaitCondition: mutex destroy failure:
Device or resource busy".
The loop in WorkerThread::run() is no longer endless. The worker thread
properly ends and is joined in WorkerThread destructor.
Design decisions:
1. WorkerThread could emit a signal when it completes a task.
Thus updateTimer could be removed from ComicFlow and GoToFlow. However,
there is no obvious way to use this new signal in the two GL classes.
Also I don't know whether updateTimer is just an inefficient polling
substitute for notification or an intentional animation mechanism.
2. The index variable is no longer stored in the worker class, but in
ComicFlow directly. Thing is, this data member was never actually
accessed by the worker, but ComicFlow went so far as to lock worker's
mutex to "protect" access to the index.
3. The common ImageLoader implementation turned out to be very general.
So I converted it into the WorkerThread class template that is not
restricted to producing QImage results and can be reused elsewhere.
4. I used standard classes (such as std::thread) instead of their Qt
equivalents (e.g. QThread) because they are more thoroughly documented.
The standard classes should also be more efficient as they were more
carefully designed and provide much fewer unnecessary features.
5. Release-Acquire ordering is safe for the WorkerThread::working
use case and is more efficient than the std::atomic-default
Sequentially-consistent ordering.
6. condition.notify_one() is called while the mutex is unlocked
to improve performance. This is safe in both cases:
a) if the worker thread exits due to a spurious wakeup just before
the condition.notify_one() call in WorkerThread destructor, so much the
better;
b) if a spurious wakeup lets the worker thread finish the task and
start waiting on the condition again just before the
condition.notify_one() call in WorkerThread::performTask(), the second
waking will be ignored by the worker thread as 'working' and 'abort'
will be false then.
Qt documentation recommends calling saveGeometry() in closeEvent().
This commit fixes the following bug on my GNU/Linux with Xfce system:
1. Move the top of the YACReader window to the top of the screen.
2. Restart YACReader (exit and run again).
2. Enter full screen mode.
4. Restart YACReader.
5. Exit full screen mode.
At this point YACReader's title bar is hidden beyond the top of the
screen, i.e. the window has moved up.
We previously had to disable antialiasing because z-buffer
picking did not work with multisampling enabled. Now we
have raytracing, so we can restore it.
There were many run-time warnings in YACReaderLibrary built in Debug
mode with hardware acceleration disabled and with ClassicComicsView
as the active comics view:
QWidget::paintEngine: Should no longer be called
QPainter::begin: Paint device returned engine == 0, type: 1
The ComicFlowWidgetSW::paintEvent() implementation now corresponds to
ComicFlowWidgetGL::paintEvent(), which has been fixed earlier.
QWidget::repaint() calls paintEvent() immediately.
PictureFlow::paintEvent() calls d->renderer->paint(). So the
d->renderer->paint() call in PictureFlow::updateMarks() was redundant.
