7 Conclusion 6 Experimental Results 5.2 Moving a Load across Multiple Stores

In many code streams, a load instruction is on a critical path and is followed by a chain of operations depending on the load. The execution of a load from memory often yields a long latency on a modern microprocessor. To compensate for this latency and reduce the length of the critical path, one way is to trigger the load as early as possible in a non-blocking way. However, to preserve the semantics of a program on a conventional architecture, a load from a memory location must be executed after all of the corresponding aliasing stores potentially writing to this very location. This constraint often hampers an instruction scheduler to move the load sufficiently early to reduce the critical path length. This aliasing problem can be greatly alleviated with an architectural support of data speculation, which allows to speculate on a load and other instructions using the result of the load, that is to initiate a load before being sure all of the previous aliasing stores have completed. Clearly, when the aliasing does occur, the result of the load has to be reloaded and other speculated instructions have to be re-executed, thus incurring a penalty. Therefore, the benefit of data speculation greatly depends on the probability of such an alias that actually occurs. This paper presents a general probabilistic memory disambiguation (PMD) framework to automatically derive the symbolic probabilities of the aliases among array references in application programs at compile time. We show how to apply this framework to build a profitability cost model to be incorporated into a compiler for data speculation. We also provide a set of heuristics to quickly approximate the aliasing probabilities in many common cases. Experimental results show that data speculation aided with the PMD framework is an effective way to maximize program performance.