Optimized implementation of QReadWriteLock

QReadWriteLock is supposed to be a better alternative to QMutex when there
are only a few writers but potentially lots of reads. However, in practice
the previous implementation was much slower, unless you really do a lot of
work with the lock for read and you have lots of contention.
Indeed, the previous implementation was locking a QMutex both for lock, and
unlock (making it already at least twice as slow as QMutex).

This new implementation brings QReadWriteLock back to the same level as QMutex:
 - No memory allocations in the uncontended case (almost no overhead allowing to
   create many of them in classes)
 - Lock-free if there is no contention

Should support up to 2^31 concurrent readers on 64 bit platforms, and 2^28
on 32 bit platforms

Change-Id: Ifa2fc999075cbb971088f4ee8e6fde78ce262da3
Reviewed-by: Edward Welbourne <edward.welbourne@theqtcompany.com>
Reviewed-by: Sean Harmer <sean.harmer@kdab.com>
Reviewed-by: Robin Burchell <robin.burchell@viroteck.net>

1 files changed, 4 insertions, 2 deletions

diff --git a/src/corelib/thread/qreadwritelock.h b/src/corelib/thread/qreadwritelock.h
index b7e2092e38..777efdb3bf 100644
--- a/src/corelib/thread/qreadwritelock.h
+++ b/src/corelib/thread/qreadwritelock.h
@@ -47,7 +47,7 @@ QT_BEGIN_NAMESPACE
 
 #ifndef QT_NO_THREAD
 
-struct QReadWriteLockPrivate;
+class QReadWriteLockPrivate;
 
 class Q_CORE_EXPORT QReadWriteLock
 {
@@ -69,8 +69,10 @@ public:
 
 private:
     Q_DISABLE_COPY(QReadWriteLock)
-    QReadWriteLockPrivate *d;
+    QAtomicPointer<QReadWriteLockPrivate> d_ptr;
 
+    enum StateForWaitCondition { LockedForRead, LockedForWrite, Unlocked, RecursivelyLocked };
+    StateForWaitCondition stateForWaitCondition() const;
     friend class QWaitCondition;
 };