Cross-loop reuse analysis and its application to cache optimizations

Abstract

In this paper we describe the design of a data-flow framework for detecting cross-loop reuse. Cross-loop reuse takes place when a set of data items or cache lines is accessed in a given loop nest and then accessed again within some subsequent portion of the program, usually another outer loop nest. In contrast to intra-loop reuse, which occurs during the execution of a single loop nest, cross-loop reuse is hard to analyze using traditional dependence-based techniques. The framework we have constructed is based on a combination of array section analysis (to capture array access patterns at a high level) and data-flow analysis (to deal with intra-procedural control flow). The framework is designed to account for cache size when gathering reuse information, and when used in an interprocedural setting, the framework also provides a mechanism for summarizing the effects of procedure calls.

Cross-loop reuse information can be used to drive a number of transformations that enhance locality and improve cache utilization, including loop fusion and loop reversal. Although these transformations are not new, their impact on cache behavior has not always been possible to predict, making them difficult to apply. As part of this paper we report the results of a comprehensive experimental study in which we apply our techniques to a set of ten programs from the SPEC95 floating point benchmark suite. We were able to obtain modest performance gains overall for several of the programs, based mostly on improvements in cache utilization.

This work was supported in part by ARPA (Army Contract DABT63-95-C-0115).