Reducing the latency of division operations with partial caching

We seek to reduce the latency of division operations. In many programs arithmetic instructions are frequently executed on the same inputs. Previous work has exploited this property by caching the quotient, bypassing the divider in the case of a cache hit. We propose caching a portion of the quotient, allowing a reduction in the cache size and an increase in cache hit rates. We call this approach partial caching. We present modifications to digit-recurrence division methods to accommodate partial caching. Using several benchmarks, we measure the hit rate and speedup which can be obtained.