面向分簇结构的编译器后端设计与优化-计算机体系结构专业毕业论文.docx

AbstractAbstract Abstract Abstract Within traditional VLIW processors，many computing resources ar e often integrated in one functional unit． With the increasing of issue width in VLIW processor,the number of accessing to the functional unit increases significantly, resulting in the decrease of processor frequency and the increase of power consumption．In such a case，the clustered structure，using distributed functional units instead of integrated ones and parallel executing instructions under the help of the compiler assigning stream of instructions onto different clusters，thus advances the 丘equeney of processor,decreases the power consumption and design complexity． The clustered structure call be described as following：multiple individual register files are used to replace single integrated register file，and each individual register file is connected to its private functional units，composing a cluster．Data 舶m different clusters Can be tranfered via the dedicated inter-cluster bus． In clustered structure，the compiler assigns each instruction to a cluster in its code generation phase．The quality of cluster-assignment designing affects the efficiency of clustered hardware．Based on the IMPACT compiler infrastructure，this thesis has explored the compiling technology for cluster processor．The main work includes： 1．Study the procedure of transporting IMPACT to BWDSP cluster-processor． 2．Study the address operation identifying method for the AGU，which is a heterogeneous cluster in BWDSP 3．Study the cluster-assignment algorithm for computing clusters，which are homogeneous clusters in BWDSE The work includes： 1)for loops without inter-iteration dependency,we study the optimization of SIMD instruction crossing over multi—clusters． 2)for the rest program，we study the cluster-assign algorithm based on component． 4．Study the software pipeline optimization for BWDSP,showing the importance of the silent-load instruction usage for this optimization． Keywords：clustered—flgofithrn，o