Integer Linear Programming vs. Graph-based Methods in Code Generation Integer Linear Programming vs. Graph-based Methods in Code Generation

A common characteristic of many embedded applications is that they are aimed at the high-volume consumer market, which is extremely cost-sensitive. However many of them impose stringent performance demands on the underlying system. Therefore, the code generation must take into account the restrictions and features given by the target architecture while satisfying these performance demands. High-level language compilers often are unable to generate code meeting these requirements. One reason is the phase coupling problem between instruction scheduling and register allocation. Many compilers perform these tasks separately with each phase ignorant of the requirements of the other. Commonly, each task is accomplished by using heuristic methods. As the goals of the two phases often connict, whichever phase is performed rst imposes constraints on the other, sometimes producing ineecient code. Integer linear programming (ILP) provides an integrated approach to the combined instruction scheduling and register allocation problem. This way, optimal solutions can be found|albeit at the cost of high compilation times. In our experiments, we considered as target processor the 32-bit DSP ADSP-2106x. We have examined two diierent ILP formulations and compared them with conventional approaches including list scheduling and the critical path method. Moreover, we have investigated approximations based on the ILP formulations; this way, compilation time can be reduced considerably while still producing near-optimal results. From the results of our implementation, we have concluded that integrating ILP formulations in conventional global algorithms is a promising method for generating high-quality code.