Static analysis of upper and lower bounds on dependences and parallelism

Reviewer: Benjamin Rayborn Seyfarth

The authors present a two-step approach to searching source code for possible parallelism. The basic problem is to accurately compute array data dependences. Existing compilers employ conservative algorithms and sometimes fail to detect parallelism, leaving it to the programmer to study source code, typically without assistance from the compiler. The first step of the authors' algorithm is to construct constraints that describe which iterations of a statement may be executed concurrently. Following this, their algorithm searches for program transformations that provide parallelism. In some cases where the algorithm cannot find a parallelizing transformation, it suggests to the programmer what might be possible if certain apparent dependences could be ignored. The authors discuss the effectiveness of their algorithm on several short code segments and on one routine of the Perfect Club b enchmark. In addition, they present timings that illustrate that the computation of memory- and value-based dependences is both valuable and practical. The authors present a useful discussion of the fundamentals of dependence analysis along with the discussion of their algorithm. They also provide an extensive bibliography of related work. The paper is moderately long and discusses difficult material. It is directed toward researchers and compiler writers interested in automatic parallelization.