Optimize QSharedPointer::getAndRef with the three-operand testAndSet

Originally QAtomicPointer didn't have the three-operand version,
resulting in code like:
  if (!atomic.testAndSetXxx(expected, newvalue))
       expected = atomic.load();

The three-operand version gives us the current value of the atomic in
case the test failed and it's free in all architectures, unlike the
extra load.

I have to use testAndSetOrdered here because I need the failing load to
use the Acquire memory order, even though that has an extra Acquire for
the successful case we don't need. QAtomicPointer does not have
testAndSetReleaseAcquire.

Change-Id: I1bd327aeaf73421a8ec5fffd1560fe30d3bfd9b8
Reviewed-by: Romain Pokrzywka <romain.pokrzywka@gmail.com>
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@qt.io>

1 files changed, 7 insertions, 4 deletions

diff --git a/src/corelib/tools/qsharedpointer.cpp b/src/corelib/tools/qsharedpointer.cpp
index 622b03f42d..a1caeeb135 100644
--- a/src/corelib/tools/qsharedpointer.cpp
+++ b/src/corelib/tools/qsharedpointer.cpp
@@ -1383,15 +1383,18 @@ QtSharedPointer::ExternalRefCountData *QtSharedPointer::ExternalRefCountData::ge
     ExternalRefCountData *x = new ExternalRefCountData(Qt::Uninitialized);
     x->strongref.store(-1);
     x->weakref.store(2);  // the QWeakPointer that called us plus the QObject itself
-    if (!d->sharedRefcount.testAndSetRelease(0, x)) {
+
+    ExternalRefCountData *ret;
+    if (d->sharedRefcount.testAndSetOrdered(nullptr, x, ret)) {     // ought to be release+acquire; this is acq_rel+acquire
+        ret = x;
+    } else {
         // ~ExternalRefCountData has a Q_ASSERT, so we use this trick to
         // only execute this if Q_ASSERTs are enabled
         Q_ASSERT((x->weakref.store(0), true));
         delete x;
-        x = d->sharedRefcount.loadAcquire();
-        x->weakref.ref();
+        ret->weakref.ref();
     }
-    return x;
+    return ret;
 }
 
 /**