A Framework for Programming Using Non-Atomic Variables

27 alized data race free program runs on alpha consistent memories as if they were sequentially consistent. We present a general method for transforming a non-cooperative mutual exclusion algorithm based on sequential consistency into an algorithm based on alpha consistency. We also show that for two well-known mutual exclusion algorithms, a speciic analysis of the code provides a more eecient algorithm. Gharachorloo, Horning, Saxe, Sites and Yu are currently working on further formalization of the Alpha memory model 18]. Given such a formalization it would be interesting to verify that it is indeed captured by alpha consistency. An important question left open by our work is how to transform programs written for sequentially consistent memories to work under alpha consistency with minimum amount of synchronization, i.e., using a minimal number of MB operations. Since this problem is probably quite hard, it might be more realistic to concentrate on methods that approximate this goal. The PowerPC is another new microprocessor architecture which includes a speciication of the shared memory model for PowerPC based multiprocessors. A formal deenition of this model using axioms appears in 12], where it is shown that if all accesses to shared data are protected by lock/unlock then a PowerPC based multiprocessor behaves as if it was sequentially consistent. A natural question is whether this result holds for any kind of release and acquire sequences. It is also interesting to check whether there is a simple way of transforming mutual exclusion algorithms based on sequential consistency to PowerPC based multiprocessors. Acknowledgements: Marc Snir suggested the Alpha as a test case for our framework; this paper would have never been written if it was not for his suggestion. We would like to thank Kourosh Gharachorloo, Maurice Herlihy and Burkhard Neidecker-Lutz for helping us understand the Alpha memory model. Thanks also to Jennifer Welch and Bill Weihl for helpful comments.