Dynamic Function Result Reuse

Continuing recent architectural trends that rely on branch predictions, value predictions, speculative execution, and reuse of results from instruction execution, we investigate the reuse of results from previous function invocations. In this paper we show that for integer benchmarks, it is possible to eliminate some function invocations since the same function is executed repeatedly with the same arguments. We feel that along with compiler techniques such as function cloning and partial evaluations, dynamic, hardware based techniques to check if a function should be invoked or the results from a prior execution can be used will lead to dramatic performance gains. We use HP Alpha based instrumentation tool called ATOM in our experiments on SPEC2000 integer benchmarks.