Peter Dimov wrote:
Tobias Schwinger:
Peter Dimov wrote:
What is the reason to prefer the current instance/lease formulation instead of
typedef ... instance_type; instance_type instance(); // automatically serves as a lease
? As we can't expose the instance when automatic locking is done, it is supposed to provide a means to not have to access the instance directly.
I'm not sure what this reformulation cannot do that the original can. An ordinary singleton can define instance_type as T*. A synchronized singleton can define instance_type as the moral equivalent of shared_ptr<T> (with a hidden unlocking deleter). What am I missing?
It was one goal to enforce client code not to make any assumptions about the type of the pointer. The reason was easy migration from an unsynchronized to a synchronized scenario by enforcing threading-agnostic client code. Further, I found 'T::instance->foo()' syntacticly nicer than 'T::instance()->foo()' or 'T::instance().foo()' . Regards, Tobias