Extending Traditional Graph-Coloring Register Allocation Exploiting Meta-heuristics for Embedded Systems

This paper presents a hybrid meta-heuristic algorithm for graph-coloring register allocation problems for embedded systems. This technique uses a classic graph-coloring register allocator as the frontend to preprocess a interference graph and uses a meta-heuristic component to deal with the remaining complex graph. It applies some genetic operators to evolve a population of solutions and to preserve diversity between these solutions. It augments Tabu Search with the power of diversification and intensification based on the probabilistic solution construction and the vocabulary building principle of Tabu Search. This technique and several classic graph-coloring register allocation algorithms have been implemented in the SUIF/MachineSUIF compiler research framework, and experimental results show that the former can often generate much better solutions than the latter for a group of benchmarks for embedded systems.