Research On Compiling Optimization For Embedded Processors

With the requirement of memory increasing, code density has gradually become a vital impediment to the development of embedded system. Code density, which is determined by not only compiling optimization methods but also instruction set architecture, is a key parameter of memory consumption of executable codes. For example, code density in X86 is often higher than ARM when compiling the same program. In addition, higher code density may result in better cache behavior and bandwidth utility, hence the program can gain better performance as the whole. Due to these reasons, this thesis focuses on the two aspects as below:1. Address Constants Compiling OptimizationAs some address constants may have similarities, this thesis raises a novel strategy, which is called separating-sharing method. In order to reduce the numbers of constants, we first separate address constants to extract the base part of them. Then, by merging those literal pools which have common constants into one, we can make them shared by different functions. This step may need reordering functions to get a more optimized result when doing similarity analysis.2. Multiple Memory Access OptimizationThe traditional method is called SOLVEMMA algorithm, which is based on directed access graph, and the direction of nodes is determined when searching the optimizing path. This algorithm originates from SOA algorithm in DSP, and it is not well suitable for embedded application field. This paper show a more concise method based on undirected graph, by classifying nodes, the time complexity of sorting can reduced to a great extent.At last, with the help of GCC compiling framework, the methods in this thesis are well implemented. And the result shows that they can raise the code density when compared with the original ones.