中文摘要：

猜测并行化编译,即线程级猜测（TLS）编译,可将原来顺序运行的程序并行化。但由于猜测数据的不确定性引起的数据管理开销过大,以及猜测线程失败引起的线程回滚开销,使得并行后的执行性能较低。针对上述问题,提出一种HL-TLS并行化编译优化框架。HL-TLS能有效地标记并行化的循环体为热点循环体,采用对最高层次热点循环体进行更激进的并行化的方式提高性能,而对非热点循环体采用保守的顺序执行以减少开销。实验结果表明,使用HL-TLS编译优化框架,实验程序的执行效率可以提高20%。

英文摘要：

Thread Level Speculation （TLS） compiling can effectively improve the parallel efficiency. But the overheads,caused by the management of the speculative data and the failure of speculative thread＇ s rollback, decreases the improvement of the parallel performance. Aiming at the too big overhead of data management and thread rollback, the Hot Loops-TLS（HL-TLS） framework is proposed. HL-TLS marks the loops which can be efficiently paralleled as HL, using a more eager parallel way on HL to improve performance, while using conservative sequence way on non-HL to reduce the overheads. Experimental result shows that HL-TLS improves 20% performance.