Characterizing Coarse-Grained Reuse of Computation

Value locality is the phenomenon that a small number of values occur repeatedly in the same register or memory location. Non-speculative reuse of computation [21] is one of the methods that has been proposed to exploit value locality. However, reuse becomes profitable only when multiple instructions are reused simultaneously. Identifying suitable chains of reusable instructions requires a global view of the program and is therefore difficult to be accomplished with hardware alone. This paper investigates the properties of reuse in the context of a dynamic optimization setting. We focus on characterizing the available computation reuse in programs at coarse granularities, and in determining the relative applicability of specialization and memoization, two commonly used techniques for exploiting coarse-grained reuse. Our study suggests that a reuse technique based on an online optimization system is feasible due to the following reasons. First, programs contain many large regions: on average, regions of 16 or more dynamic instructions represent 54% of all reuse, or 26% of all dynamic instructions. Second, large regions are stable over time and hence can likely be identified cost-effectively with a dynamic optimizer. Third, we observed that many reuse opportunities cannot be exploited with memoization, the currently used reuse technique, but must be supplemented with specialization, a technique that requires run-time code generation.