Abstract: The rapid development of SMP and new proposed DHMP bring new challenges for program performance optimization.We raised two performance tuning problems and the solutions were given by phase analysis.The first problem is to find theperformance bottlenecks in each phase.We proposed a static phase analysis method,which finds performance bottlenecks in each phases by analyzing architecture features and its similar matrix.The second problem is to give the proper time to reconfigure for DHMP.We proposed dynamic phase detection algorithms,namely DPDA and HTPA.DPDA archives effective performance in a relative low software/hardware cost,and HTPD is the first phase detection algorithm using statistics theory.Our results show that comparing with BBV,DPDA and HTPD can avoid its limitation of offline,plus additional hardware in online algorithm and compiler’s effect,while they offer a comparable stability and correctness.Since DPDA and HTPD do not relay on additional hardware support,they can be implemented directly in mainstream processors and DHMP.

Key words: Program analysis,Program phase,Static program analysis,Phase detection