BWDSP10x上地址和数据谓词执行的编译优化

Compilation Optimization of Address and Data Predicated Execution on BWDSP10x

传统的谓词优化技术是在冯·诺伊曼体系结构计算机上实施的,仅对数据流进行优化,并没有考虑哈佛体系结构下指令和数据分开的情况.BWDSP10x是指令和数据分开的哈佛体系结构,它支持超长指令字,不仅提供了对数据谓词执行的支持也提供了对地址谓词执行的支持.特此提出了一种在区域上对两种谓词模式优化支持的方法,在进行两种比较之前,通过判断比较操作的两个操作数类型来分别实施两种模式的谓词优化,使得对地址的比较不用传输到通用寄存器中.实验结果表明该优化方法能显著地节省CPU的时间和带宽,大大减少了分支指令,使程序性能提高了28.4%.

The traditional predicate optimization technique is based on the Von Neumann architecture, which considers the data flow optimization only. However, BWDSP10 x is based on the Harvard architecture, which data and instructions are physically separated. It provides VLIW and supports not only data predicated execution but also address predicated execution. Hence, we present an optimization method for the two predicated execution technology based on the region. In this method, types of the two operands of comparison operation will be identified before the two kinds of operations are executed, and when addresses are compared, the two operands don’t need to transfer to general registers. Experimental result shows that the optimization method can highly reduce the time and bandwidth of CPU, and reduce large numbers of branch instructions. The performance of programs tested is increased by 28.4 percent after the optimization.