We used to look through the event queue in QCoreApplication::postEvent,
and if we found an existing DeferredDelete event for the receiver we
would compress the two events into one.
This was changed in 99b89d30fa, as the
logic was causing O(n^2) for deleteLater, by using one of the bits
in QObjectData to track whether the object had already been deleted.
But it kept the logic for tracking this in QCoreApplication::postEvent,
and QCoreApplication::compressEvent would still do the work of deleting
the additional QDeferredDeleteEvents.
To avoid the unnecessary heap allocation of the QDeferredDeleteEvents
we can move the debouncing/compression to QObject::deleteLater().
We use the same mutex as in QCoreApplication::postEvent to guard
concurrent access to deleteLaterCalled.
A note has been added about the (preexisting) issue that the mutex
is not sufficient to prevent data races, as the deleteLaterCalled
flag is part of a bit-field, and we're not guarding any of our
other accesses to other bits.
As QDeferredDeleteEvents is private API, we can rely on no-one else
posting it than QObject::deleteLater(), which should be the case now
that tst_QApplication::sendPostedEvents() was fixed.
The documentation has been clarified as well. It's safe to call
deleteLater() more than once, but that's not _because_ other
pending events for the object are cleared. The latter behavior
is normal ~QObject() behavior. The documentation was probably
written at a point we didn't do any event compression at all
for QDeferredDeleteEvents.
Task-number: QTBUG-120124
Task-number: QTBUG-119918
Change-Id: I2a733095b7cb066ba494b1335aa40200c749cb0c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
13074a967f
