Towards Optimizing Code Generation by Domain-Sensitive Macro Expansion

Many modern code generation methods use tree pattern matching with dynamic programming. However, especially in the case of an irregular special-purpose processor architecture their lack of transparency and stability may be problematic: it is diicult to predict the exact code generation result in advance, and the eeects of a modiication in the code generation rules may be surprisingly wide. In contrast, macro expansion techniques are intuitively transparent. When global variables are disallowed, macro expansion typically has the Church-Rosser property: the nal expansion result is independent of the expansion order of the individual intermediate macro calls. Besides enabling parallel implementation, order-independence means stability: the eeects of modifying a macro deenition are guaranteed to remain local. The locality is actually the problem with macro expansion; code optimization is improved when an assembly language macro is sensitive to its context. For instance , it should know which registers are free and which ones contain useful values. A traditional mechanism for propagating contextual information is the use of global variables. But relying on global variables means that the expansion should take place strictly from left to right, so that order-independence is lost while it is still not possible to pass information in the opposite direction. In this work a novel assembly-language-level macro expansion technique with both order-independence and rather general context-sensitivity is proposed. The technique supports modular and hierarchical code libraries. Both a formal model and a code generator prototype demonstrating the technique are presented. Foreword This work has been carried out at Digital Systems Laboratory of Helsinki University of Technology. I wish to thank Professor Leo Ojala for his continuous support and guidance; in particular, I appreciate with gratitude the opportunity to concentrate on the present subject. To all my co-workers at the laboratory, I certainly owe a lot for the inspiring atmosphere. I am indebted especially to Dr. Johan Lilius and Dr. Ilkka Niemell a for their valuable suggestions on the coordination of my eeorts.