Demand-Driven, Symbolic Range Propagation

To e ectively parallelize real programs, parallelizing compilers need powerful symbolic analysis techniques [13, 6]. In previous work we have introduced an algorithm called range propagation [4] that provides such capabilities. Range propagation has been implemented in Polaris, a parallelizing Fortran compiler being developed at the University of Illinois [7]. Because the algorithm is relatively expensive and several compilation passes make use of it, we have studied opportunities for increasing its e ciency. In this paper we present an algorithm that provides range analysis on-demand. We have implemented this algorithm in Polaris as well, and have measured its e ectiveness.