Program Memory Redundancy Analysis and Elimination to Improve Application Performance

As a consequence of current evolution trend of both software development paradigm and processor architecture, memory operations have become not only more frequent during program execution, but also more expensive in terms of hardware resource usage. A useful phenomenon which can be exploited to optimize memory operations is memory redundancy: loads and stores that have the same effects as previous memory operations. Elimination of these memory redundancies can reduce program running time and improve usage of hardware resources. In this article, we first presents an efficient and powerful static memory redundancy detection algorithm. Previous techniques separate the task of scalar and memory redundancy detection, thus they fail to discover the redundancies due to the interaction between scalar and memory values. The new algorithm unifies the scalar and memory redundancy detection by an integrated value numbering process, and is more powerful than the separate approach. A natural application of the algorithm is to enable program transformations to remove the redundancies detected. We show that traditional scalar redundancy removal frameworks can be easily adapted to remove memory redundancies. Experimental results show that the techniques can remove up to 40% of dynamic loads in benchmark applications from SPEC2000 and MediaBench. To understand the run-time memory redundancy phenomenon, we conduct a limit study to evaluate the effectiveness of memory redundancy removal methods with different level of capabilities. The study shows that many run-time memory redundancies still exist in highly optimized code, even after applying the techniques we have described in this article. This suggests that more aggressive methods to detect and remove dynamic memory redundancies should be profitable. We also look at the interaction between register allocation and memory redundancy elimination. The results show that register spills due to memory redundancy removal are rare in the benchmark applications, and there is little negative impact on performance caused by spilling overhead.