A New Compiler Framework based on Superword Level Parallel

Superword level parallel (SLP) algorithm is an automatic vectorization method that is suitable for the applications including parallel codes. Existing SLP algorithm could not efficiently deal with the applications that contain few parallel codes. In the present study, a new compile framework based on the improved SLP algorithm is presented. The framework contains three phases: isomorphic processing for isomeric statements, establishment of superword statements, and data layout optimization. Firstly, isomeric statements with similar instruction in the codes were transformed to isomorphic statements by the improved SLP algorithm. Secondly, the superwords reuse patterns were obtained before making the optimization decisions from a global point of view. Finally, data layout optimization was combined for further performance improvement. The experimental results indicated that the optimization of the compile framework was better than existing SLP algorithm.