Toward efficient flow-sensitive induction variable analysis and dependence testing for loop optimization

Abstract

This paper presents a new approach to improve flow-sensitive induction variable analysis and data dependence testing on intermediate program representations, such as control-flow graphs with low-level operations representations in single static assignment forms. Current compiler techniques have difficulties optimizing loops that exhibit irregular control flow patterns. The inaccuracy of loop analysis results in conservative estimations of array-based data dependences in loops, which negatively affects the speedup of the loop via parallelization and vectorization. Our approach is based on a novel CR# (CR-sharp) algebra that effectively represents the value progressions of (conditionally) updated variables in loops. The CR# forms of induction variables are constructed with a new flow-sensitive induction variable recognition algorithm. We also developed a new CR#-based nonlinear data dependence test that enables loops to be more effectively optimized.