Most memory references in numerical codes correspond to array references whose indices are affine functions of surrounding loop indices. These array references follow a regular predictable memory pattern that can be analysed at compile time. This analysis can provide valuable information like the locality exhibited by the program, which can be used to implement more intelligent caching strategy. In this paper we propose a static locality analysis oriented to the management of data caches. We show that previous proposals on locality analysis are not appropriate when the proposals have a high conflict miss ratio. This paper examines those proposals by introducing a compile-time interference analysis that significantly improve the performance of them. We first show how this analysis can be used to characterize the dynamic locality properties of numerical codes. This evaluation show for instance that a large percentage of references exhibit any type of locality. This motivates the use of a dual data cache, which has a module specialized to exploit temporal locality, and a selective cache respectively. Then, the performance provided by these two cache organizations is evaluated. In both organizations, the static locality analysis is responsible for tagging each memory instruction accordingly to the particular type(s) of locality that it exhibits.

Peer Reviewed