基于资源重用的嵌入式处理器指令发射机制研究

发布时间：2018-04-05 06:40

本文选题：资源重用　切入点：操作数旁路　出处：《浙江大学》2012年硕士论文

【摘要】：指令发射机制的研究主要包括指令相关性检查和操作数旁路机制、寄存器重命名方法以及指令发射队列,本文分别对这三个方面进行研究并提出了基于计数方式的操作数旁路方法、基于存储资源迭代重用的寄存器重命名方法以及基于最老最先发射的集中分布式结构指令发射队列。基于计数方式的操作数旁路方法通过对指令执行周期数进行递减计数来判断指令能够获得源操作数的时间,并通过对流水线级数的信息来确定旁路的路径,缓解了指令间数据真相关性带来的流水线停顿。针对超标量深流水线中物理寄存器资源冲突造成的流水线阻塞问题,本文提出了一种多指令共享同一物理寄存器资源的非阻塞指令发射方法。该方法可在物理寄存器资源冲突下继续分配物理寄存器,利用发射缓冲队列临时缓冲冲突的指令,增加发射流水级实际可分配的物理寄存器数量,释放发射窗口,提高物理寄存器使用的并行性。实验结果表明,相对于传统重命名方法,该方法可减少27.3%的物理寄存器资源实现传统方法相同的性能。基于最老最先发射的集中分布式指令发射队列分别从队列的发射行为和结构出发,通过移动队列表项内容的方式实现了最老最先发射的机制,提高了指令的执行和退休效率,而集中分布式的指今发射队列结构在提高资源利用效率的同时并没有给队列的控制逻辑带来复杂度。
[Abstract]:The research of instruction launching mechanism mainly includes instruction correlation checking and operand bypass mechanism , register renaming method and instruction transmission queue . In this paper , we study the three aspects and put forward an operand bypass method based on counting mode , register renaming method based on iterative reuse of storage resources , and distributed structure instruction transmission queue based on the oldest first transmission .

The operand bypass method based on the counting mode determines the time of the instruction to obtain the source operand by decreasing the number of instruction execution cycles , and determines the path of the bypass by information on the pipeline stage , and relieves the pipeline stall caused by the true correlation of the data between the instructions .

In order to solve the problem of pipeline congestion caused by the conflict of physical register resources in the superscalar deep pipeline , a non - blocking instruction transmission method for sharing the same physical register resources is presented in this paper . The method can continue to allocate the physical register under the conflict of physical register resources , increase the number of physical registers which can actually be allocated by transmitting the buffer queue , release the emission window and improve the parallelism of the physical register . The experimental results show that compared with the traditional renaming method , the method can reduce 27.3 % of the physical register resources to realize the same performance of the traditional method .

Based on the transmitting behavior and the structure of the queue , the oldest transmission mechanism is realized by the way of the contents of the mobile unit list , which improves the execution and retirement efficiency of the instruction , and the centralized distributed index transmission queue structure brings complexity to the control logic of the queue while improving the resource utilization efficiency .

【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP332

【参考文献】