3种提高软件流水有效性的算法:比较和结合

软件流水是开发循环程序指令级并行性的技术,它通过并行执行连续的多个循环体来加快循环的执行速度.在软件流水中,循环体的重叠增加了寄存器需求,导致寄存器压力增大,当目标处理机所提供的寄存器不足时,软件流水可能失败.在Itanium处理机上评估了NAS和SPEC2000基准程序中的软件流水循环的寄存器需求,发现静态寄存器不足是造成软件流水失败的主要原因,提出了3种增加软件流水个数、提高软件流水有效性的算法:限制循环展开因子的算法(register sensitive unrolling,简称RSU)、堆栈寄存器分配算法(stacked registerallocation,简称SRA)以及变量类型转换的算法(variabletype conversion,简称VTC).RSU根据静态寄存器需求确定一个合理的展开因子,增加了软件流水的成功率;SRA和VTC分别使用空闲的堆栈寄存器和旋转寄存器来充当静态寄存器,提高了寄存器的利用率.在面向Itanium处理器的开放源码编译器ORC(open research compiler)上实现了这3种算法,通过NAS程序的测试比较了这3种算法的有效性,同时对它们的结合应用进行了研究和实验.

Three Algorithms for Improving the Effectiveness of Software Pipelining:Comparison and Combination

Software pipelining is a loop scheduling technique that extracts instruction level parallelism by overlapping the execution of several consecutive iterations. One of its drawbacks is the high register requirements, which may lead to software pipelining failure due to insufficient static general registers in Itanium. This paper evaluates the register requirements of software-pipelined loops and presents three new methods for software pipelining loops that require more static general registers than those available in Itanium processor. They reduce register pressure by either reducing instructions in the loop body or allocating stacked non-rotating registers or rotating register in register stack to serve as static registers. These methods are better than the existing techniques in that they further improve performance gain from software pipelining by increasing software-pipelined loops. These methods have been implemented in open research compiler (ORC) targeted for Itanium processor, and they perform well on loops of the programs in NAS Benchmarks. For some benchmarks, the performance is improved by more than 21%.