Executing subordinate activities by pushing return addresses on the stack is the most e cient working mode for sequential programs. It is supported by all current processors, yet in most cases is inappropriate for parallel execution of indepented threads of control. This paper describes an approach of dynamically switching between e cient on-stack execution of sequential threads and o -stack spawning of parallel activities. The presented method allows to incorporate work-stealing into the scheduler, letting the system pro t from its near-to-optimal loadbalancing properties.