Accurate analysis of array references

Modern computer systems are increasingly relying on parallelism to improve performance. Automatic parallelization techniques offer the hope that users can simply and portably exploit parallelism. This thesis addresses the problem of data dependence analysis, the base step in detecting loop level parallelism in scientific programs. Exploiting parallelism can change the order of memory operations. Data dependence analysis involves analyzing the dynamic memory reference behavior of array operations so that compilers will only parallelize loops in the cases where any resultant reordering of memory references does not change the sequential semantics of the program. In general, data dependence analysis is undecidable, and compilers must conservatively approximate array reference behavior, thus sequentializing parallel loops. Traditional data dependence analysis research has concentrated on the simpler problem of affine memory disambiguation. Many algorithms have been developed that conservatively approximate even this simpler problem. By using a series of algorithms, each one guaranteed to be exact for a certain class of input, we are able to devise a new method that in practice solves exactly and efficiently the affine memory disambiguation problem. Because our affine memory disambiguator is exact in practice, we can devise an experiment to test the effectiveness of affine memory disambiguation at approximating the full data dependence problem. We discover that the lack of data-flow information on array elements is the key limitation of affine memory disambiguators. We develop a new representation and algorithm to efficiently calculate these data-flow dependences. Finally, we address the problem of interprocedural data dependence analysis. By using an array summary representation that is guaranteed to be exact when applicable, we can combine summary information with inlining to exactly and efficiently analyze affine array references across procedure boundaries. Taken together, our algorithms generate the more accurate information that will be needed to exploit parallelism in the future.