Jeffrey Pang and Jimeng Sun
Abstract: Many features of modern architectures, such as Performance Monitering Units (PMUs), can be leveraged for efficient program profiling. We present several simple heuristics to estimate path profiles from sampled partial paths which can be collected at runtime. Like recent work, we show that we can find 80-90% of the hot paths within a program using these techniques. However, we also find that despite this high accuracy in identifying hot paths, using estimated paths to perform superblock formation does not necessarily yield an equivalent boost in performance. We evaluate how estimated path profiles impact superblock formation in the CASH compiler. In addition, we compare superblocks to hyperblocks and demonstrate that predicate calculation overhead can cause maximal hyperblocks to have inferior performance.