Intraprocedural and interprocedural data dependence analysis for parallel computing

A parallelizing compiler relies on data dependence analysis to detect independent operations in a user's program. In scientific and engineering programs, it is important for the compiler to analyze data dependences involving array references. This dissertation addresses a few fundamental issues in such an analysis.The first part of the dissertation discusses algorithms for data dependence tests. A real-valued algorithm ($\lambda$-test) is proposed for testing dependences between multi-dimensional array references, since previous algorithms are either too time-consuming or too imprecise. Algorithms for an unconstrained integer test and an exact integer test are also presented for some common classes of array references. In order to evaluate the complexity of array references and data dependences in real programs, an empirical study was conducted, and some of its results are presented.The second part of the dissertation is devoted to interprocedural analysis for parallel computing. The key to successful parallelization of programs which contain procedure calls is an efficient and accurate interprocedural data dependence analysis. A scheme, called the atom image scheme, is proposed for this analysis. Unlike previous approaches, the atom image scheme preserves precise array reference details while allowing recursive calls to be analyzed. Further, its precise form for representing the subscript details allows flexibility in selecting various data dependence test algorithms, depending on their usage. Such flexibility is important to the accuracy and efficiency of data dependence analysis.U of I OnlyETDs are only available to UIUC Users without author permissio