Compiling Natural Semantics - ReadingSample

Natural semantics has become a popular tool among programming language researchers. It is used for specifying many aspects of programming languages, including type systems, dynamic semantics, translations between representations, and static analyses. The formalism has so far largely been limited to theoretical applications, due to the absence of practical tools for its implementation. Those who try to use it in applications have had to translate their speci cations by hand into existing programming languages, which can be tedious and error-prone. Hence, natural semantics is rarely used in applications. Compiling high-level languages to correct and e cient code is non-trivial, hence implementing compilers is di cult and time-consuming. It has become customary to specify parts of compilers using special-purpose speci cation languages, and to compile these speci cations to executable code. While this has simpli ed the construction of compiler front-ends, and to some extent their back-ends, little is available to help construct those parts that deal with semantics and translations between higher-level and lower-level representations. This is especially true for the Natural Semantics formalism. In this thesis, we introduce the Relational Meta-Language, RML, which is intended as a practical language for natural semantics speci cations. Runtime e ciency is a prerequisite if natural semantics is to be generally accepted as a practical tool. Hence, the main parts of this thesis deal with the problem of compiling natural semantics, actually RML, to highly e cient code. We have designed and implemented a compiler, rml2c, that translates RML to e cient low-level C code. The compilation phases are described in detail. High-level transformations are applied to reduce the usually enormous amount of non-determinism present in speci cations. The resulting forms are often completely deterministic. Pattern-matching constructs are expanded using a pattern-match compiler, and a translation is made into a continuation-passing style intermediate representation. Intermediate-level CPS optimizations are applied before low-level C code is emitted. A new and e cient technique for mapping tailcalls to C has been developed. We have compared our code with other alternative implementations. Our