Compile-time Optimization of a Constraint-based Compiler Back-end

Generating optimal code is a challenging problem. Traditional compilers break down the problem complexity by solving the main interdependent compilation tasks (that is, instruction selection, register allocation and instruction scheduling) in stages, and optimizing each stage individually using heuristics. While this approach achieves good results, it clearly compromises on the generated code quality. In the last decades, combinatorial optimization approaches for code generators have emerged that open up the potential for optimal code generation. One such approach is the Unison code generator. It captures the tasks of register allocation and instruction scheduling in one combinatorial model, and hence allowing for task-interdependent code optimizations, achieving performance improvements of up to 40%2 for generated codes. However, combinatorial approaches suffer from slow compilation times due to the exponentially growing search effort. This paper investigates refinements of Unison’s combinatorial model, aiming at cutting down the search effort for faster compilation times.