Increasing Perfect Nests in Scientific Programs

Loop optimizations for data locality often require perfect loop nests. In this paper, we report on the extent to which perfect nests are present in 23 applications from 4 standard benchmark suites. Further, we evaluate the effectiveness of 3 transformations for converting imperfect nests into perfect ones: code sinking loop distribution and loop fusion. We find that on average, perfect nests contribute to only 39% of the execution time of the benchmarks. Code sinking causes the largest increase in the relative contribution to execution time of perfect nests (to 56%) followed by loop distribution (to 45%). Loop fusion results in no significant improvement. We also evaluate the overhead of applying the transformations. Based on our results, we suggest a strategy for using the transformations in a compiler.