Interprocedural dependence analysis and parallelization

Abstract

We present a method that combines a deep analysis of program dependences with a broad analysis of the interaction among procedures. The method is more efficient than existing methods: we reduce many tests, performed separately by existing methods, to a single test. The method is more precise than existing methods with respect to references to multi-dimensional arrays and dependence information hidden by procedure calls. The method is more general than existing methods: We accommodate potentially aliased variables and structures of differing shapes that share storage. We accomplish the above through a unified approach that integrates subscript analysis with aliasing and interprocedural information.