Scaling Existing Lock-based Applications with Lock Elision
queue.acm.org
queue.acm.org
It sounds like the new hardware support allows you to speculatively continue execution and fail later? Seems like the wrapping approach described in this paper doesn't take advantage of that?
EDIT: http://en.wikipedia.org/wiki/Transactional_Synchronization_E...
With lock elision, two cores can both optimistically execute code within the critical section and only abort when a dynamic data race (or some other abort condition, like running out of space in a hardware buffer) occurs. In that case, one core may take the lock and force the other to wait, just like the traditional lock.