Exhibit #1: the lock to protect lazy initialization. That just needs a CAS on the slow path and a load-load fence on the fast path. Delete the object you created if you lose the race and try again.
Exhibit #2: you can probably do dirty things to make loading from dicts/lists not require locks even though storing to them does.
template<typename T, typename... Args>
void constructSingleton(Atomic<T*>& singletonRef, Args&&... args)
{
for (;;) {
T* oldValue = singletonRef.load();
if (oldValue)
return oldValue;
T* newValue = new T(std::forward<Args>(args)...);
if (singletonRef.compareExchangeWeak(nullptr, newValue))
return newValue;
delete newValue;
}
}