Scheduling problems while compiling the real-time language Lustre on the digital signal processor ST18930

An experiment in optimizing the computations of loops on a digital signal processor is reported. Working with a simple pipelined monoprocessor, the DSP ST18930, the goal was to optimize the code generated by a dedicated version of the Lustre compiler. If there is a computation loop whose end may be overlapped with the beginning of the next loop, the problem is to minimize the delay between two successive initiations of such loops. After briefly summarizing Lustre, the source real-time language, the DSP ST18930 is briefly described. The pipeline used in the execution of instructions is highlighted. The method is verified using a lattice filter example