Abstracts Category : Other

Reliably composable language extensions

by Ted Kaminski

Many programming tasks are dramatically simpler when an appropriate domain-specific language can be used to accomplish them. These languages offer a variety of potential advantages, including programming at a higher level of abstraction, custom analyses specific to the problem domain, and the ability to generate very efficient code. But they also suffer many disadvantages as a result of their implementation techniques. Fully separate languages (such as YACC, or SQL) are quite flexible, but these are distinct monolithic entities and thus we are unable to draw on the features of several in combination to accomplish a single task. That is, we cannot compose their domain-specific features. "Embedded" DSLs (such as parsing combinators) accomplish something like a different language, but are actually implemented simply as libraries within a flexible host language. This approach allows different libraries to be imported and used together, enabling composition, but it is limited in analysis and translation capabilities by the host language they are embedded within. A promising combination of these two approaches is to allow a host language to be directly extended with new features (syntactic and semantic.) However, while there are plausible ways to attempt to compose language extensions, they can easily fail, making this approach unreliable. Previous methods of assuring reliable composition impose onerous restrictions, such as throwing out entirely the ability to introduce new analysis. This thesis introduces reliably composable language extensions as a technique for the implementation of DSLs. This technique preserves most of the advantages of both separate and "embedded" DSLs. Unlike many prior approaches to language extension, this technique ensures composition of multiple language extensions will succeed, and preserves strong properties about the behavior of the resulting composed compiler. We define an analysis on language extensions that guarantees the composition of several extensions will be well-defined, and we further define a set of testable properties that ensure the resulting compiler will behave as expected, along with a principle that assigns "blame" for bugs that may ultimately appear as a result of composition. Finally, to concretely compare our approach to our original goals for reliably composable language extension, we use these techniques to develop an extensible C compiler front-end, together with several example composable language extensions.