Maximum and sorted cache occupation using array padding

Abstract

The paper describes a framework for analyzing the cache content on affine references to arrays in loops. The framework is based on a small set of key cache parameters. We study the relation between these cache parameters and the data memory layout of arrays to demonstrate how to use array padding (static array re-dimensioning) to optimize the use of the cache. Based on the cache model we present a method to carry out intra-array padding for a maximum cache occupation and for a maximum sorted cache occupation, and a simple method to carry out inter-array padding. We also present an experimental evaluation of our techniques using a cache simulator and actual code executions on the MIPS R10K processor.