Compiler Optimization Technique for Data Cache Prefetching Using a Small CAM Array

With advances in compiler optimization and program flow analysis, software assisted cache prefetching schemes using PREFETCH instructions are now possible. Although data can be prefetched accurately into the cache, the runtime overhead associated with these schemes often limits their practical use. In this paper, we propose a new scheme, called the Stride_CAM Data Prefetching (SCP), to prefetch array references with constant strides accurately. Compared to current software assisted data prefetching schemes, the SCP scheme has much lower runtime overhead without sacrificing prefetching accuracy. Our result showed that the SCP scheme is particularly suitable for computing intensive scientific applications where cache misses are mainly due to array references with constant strides and they can be prefetched very accurately by this SCP scheme.