Modulo scheduling with cache reuse information

Software pipelining for instruction-level parallel computers with non-blocking caches usually assigns memory access latency by assuming either all accesses are cache hits or all are cache misses. We contend setting memory latencies by cache reuse analysis leads to better software pipelining than either an all-hit or all-miss assumption. Using a simple cache-reuse model, our software pipelining optimization achieved 10% improved execution performance over assuming all-cache-hits and used 18% fewer registers than required by an all-cache-miss assumption. We conclude that software pipelining for architectures with non-blocking cache should incorprate a memory-reuse model