T. Baba, H. Yanome, K. Ootsu, and T. Yokota (Japan)
Dynamic optimization, loop parallelization, hardware hot loop path detector
Dynamic parallelization and optimization of a loop is a crucial issue for enhancing the performance of sequential programs as loops account for a large fraction of execu tion time. Loop level parallelism can also be extracted efficiently due to its regular structure. Based on the ob servation that only a limited number of paths are executed frequently in hot loops, we propose a hardware hot loop path detector to specify such hot loops and their hot paths accurately so that the dynamic optimizer may utilize the detected information effectively. The detector consists of a stack structured bit-tracing unit that identifies loop paths at a subroutine level, a hot loop detector that detects hot loops by utilizing loop path information and a hot path accumu lator of loop paths. Experiments using SPEC CINT2000 show that loop paths occupy a small fraction (14.46%) of Ball-Larus paths but are detected frequently (64.45% of Ball-Larus paths). A combined small scale hot loop detector and hot path ac cumulator (32 entries each) attain a detection accuracy of 97.10% for the hottest loop path and 93.83% for the top 2 hottest loop paths and their order within hot loops.
Important Links:
Go Back
