A 0-1 integer linear programming based approach for global locality optimizations

Compiler optimizations aimed at improving cache locality are critical in realizing the performance potential of memory subsystem. For scientific programs, loop and data transformations are two important compiler optimization methods to improve cache locality. In this paper, we combine loop and data transformations and present a 0-1 integer linear programming (0-1 ILP) based approach that attempts to solve global locality optimization problems. We use the treelike memory layout graph (TMLG) to describe it program's locality characteristics, formulate the locality optimization problems as the problems of finding the optimal path sets in TMLGs, and then use 0-1 ILP to find the optimal path sets. Our approach is applicable not only to perfectly nested loops but also to non-perfectly nested loops. Moreover, the approach is suitable for handling the circumstances that arrays are accessed not only along dimensions but also along diagonal-like directions. The experimental results show the effectiveness of our approach.