Report Number: CSL-TR-95-665
Institution: Stanford University, Computer Systems Laboratory
Title: Interprocedural Parallelization Analysis: Preliminary Results
Author: Hall, Mary W.
Author: Amarasinghe, Saman P.
Author: Murphy, Brian R.
Author: Liao, Shih-Wei
Author: Lam, Monica S.
Date: March 1995
Abstract: This paper describes a fully interprocedural automatic parallelization system for Fortran programs, and presents the results of extensive experiments obtained using this system. The system incorporates a comprehensive and integrated collection of analyses including dependence, privatization and reduction recognition for both array and scalar variables, and scalar symbolic analysis to support these. All the analyses have been implemented in the SUIF (Stanford University Intermediate Format) compiler system, with the aid of an interprocedural analysis construction tool known as FIAT. Our interprocedural analysis is uniquely designed to provide the same quality of information as if the program were analyzed as a single procedure, while managing the complexity of the analysis. We have implemented a robust system that has parallelized, completely automatically, loops containing over a thousand lines of code. This work makes possible the first comprehensive empirical evaluation of state-of-the-art automatic parallelization technology. This paper reports evaluation numbers on programs from standard benchmark suites. The results demonstrate that all the interprocedural analyses taken together can substantially advance the capability of current automatic parallelization technology.