Edge profiling versus path profiling

Abstract

Edge profiles are the traditional control flow profile of choice for profile-directed compilation. They have been the basis of path-based optimizations that select paths, even though edge profiles contain strictly less information than path profiles. Recent work on path profiling has suggested that path profiles are superior to edge profiles in practice.We present theoretic and algorithmic results that may be used to determine when an edge profile is a good predictor of hot paths (and what those hot paths are) and when it is a poor predictor. Our algorithms efficiently compute sets of definitely and potentially hot paths in a graph annotated with an edge profile. A definitely hot path has a frequency greater than some non-zero lower bound in all path profiles that induce a given edge profile.Experiments on the SPEC95 benchmarks show that a huge percentage of the execution frequency in these programs is dominated by definitely hot paths (on average, 84% for FORTRAN benchmarks and 76% for C benchmarks). We also show that various hot path selection algorithms based on edge profiles work extremely well in most cases, but that path profiling is needed in some cases. These results indicate the usefulness of our algorithms for characterizing edge profiles and selecting hot paths.