Code generation using tree matching and dynamic programming

Reviewer: Martin Joseph Jourdan

This paper describes a locally optimal code generation technique based on top-down tree pattern matching and dynamic programming, together with its implementation in the form of a language and system called twig. Although both subtechniques have long been recognized as effective aids for code generation, their combination is original and quite attractive; in particular, it is a step toward the purely declarative specification of code generators. The value of this approach is confirmed by practical experiments with twig, which meets its goals with quite acceptable efficiency. The paper is well written and generally convincing, especially the theoretical material, but it suffers from several defects. First, the publication delays were exceptionally long, so the results appear somewhat out-of-date. Second, the approach is only compared with parsing- and attribute-grammar-based approaches, which is unfair since their weaknesses were recognized long ago; comparisons with other tree matching techniques, especially bottom-up approaches, would be more significant. Third, a twig specification loses some declarativeness because many aspects of the code generation process must be performed by hand; in addition, the use of the TOPDOWN modifier, which allows the designer to influence the pattern-matching process, is insufficiently documented and seems a bit dangerous. Fourth, twig does not specify complete code generators but merely instruction selectors: the authors completely overlook other aspects, such as register allocation and instruction scheduling, that are becoming more and more important on today's machines. Nevertheless, the basic idea of paper is valuable, and engineers working on practical problems in retargetable code generation could find a starting point for good solutions.