A compiler-based approach for improving intra-iteration data reuse

Intra-iteration data reuse occurs when multiple array references exhibit data reuse in a single loop iteration. An optimizing compiler can exploit this reuse by clustering (in the loop body) array references with data reuse as much as possible. This reduces the number of intervening references between references to the same array and improves overall execution time and energy consumption. In this paper, we present a strategy where inter-statement and intra-statement optimizations are used in concert for optimizing intra-iteration data reuse. The objective is to cluster (within the loop body) the array references with spatial or temporal reuse. Using four array-intensive applications from image processing domain, we show that our approach improves the cache behavior of programs by 13.8% on the average.