High-Level Loop Optimizations for GCC

This paper will present a design for loop optimizations using high-level loop transformations. We will describe a loop optimization infrastructure based on improved induction variable, scalar evolution, and data dependence analysis. We also will describe loop transformation opportunities that utilize the information discovered. These transformations increase data locality and eliminate data dependencies that prevent optimization. The transformations also can be used to enable automatic vectorization and automatic parallelization functionality. The TreeSSA infrastructure in GCC provides an opportunity for high level loop transforms to be implemented. Prior to the Loop Nest Optimization effort described in this paper, GCC has performed no cache reuse, data locality, parallelization, or loop vectorization optimizations. It also had no infrastructure to perform data dependence analysis for array accesses that are necessary to apply these transformations safely. We have implemented data dependence analysis and linear loop transforms on top of TreeSSA, which provides the following features: 1. A data dependence framework for determining whether two data references have a dependence. The core of the dependence analysis is a new, low-complexity algorithm for the recognition of scalar evolutions that tracks induction variables across a def-use graph. It is used to determine the legality of various transformations, including the vectorization transforms being implemented, and the matrix based transformations. 2. A matrix-based transformation method for rearranging loop nests to optimize locality, cache reuse, and remove inner loop dependencies (to help vectorization and parallelization). This method can perform any legal combination of loop interchange, scaling, skewing, and reversal to a loop nest, and provides a simple interface to doing it.