Design And Implementation Of The Vectorizing Complier Of High-Performance DSP Matrix

As DSP processor widely used in fields of communication and computer, people have higher requirements on real-time DSP processing. Matrix is a high-performance DSP with independent intellectual property rights. It contains vector component and scalar component. Matrix combines VLIW and SIMD features. To be really available for using, a new architecture must have a complete software supporting. Researching on the Matrix's vectorizing compiler is very important for the convenience of application development and making use of Matrix's architectural features. Currently, vectorization is still a difficulty in the research area of compiler technology because of critical conditions of auto-vectorization and differences between various SIMD platforms.In this paper, we research on the internal structure of GCC and techniques of auto-vectorization. We design and implement a vectorizing compiler for Matrix. The main contributions of this paper are as follows:We propose a language-extend vectorizing algorithm for Matrix. We design a set of C language extensions, which is implemented in GCC. The language-extend vectorizing algorithm can make good use of the architecture of Matrix for vectorizing the application.We implement a vector platform of Matrix in GCC back-end, including machine mode, command mode, and registers description. What's more we add a lot of new special operations in GCC, in order to meet the demand of Matrix's special commands.We vectorize FFT in Matrix. We develop the FFT application using the language-extend vectorizing algorithm, compile and execute it with Matrix's Compiler, Assembler, Linker, and Simulator. The experiment result proves the correctness and effectiveness of our language-extend vectorizing algorithm.