Syntactic language extensions can introduce new facilities into a programming language while requiring little implementation effort and modest changes to the compiler. It is typical to desugar language extensions in a distinguished compiler phase after parsing or type checking, not affecting any of the later compiler phases. If desugaring happens before type checking, the desugaring cannot depend on typing information and type errors are reported in terms of the generated code. If desugaring happens after type checking, the code generated by the desugaring is not type checked and may introduce vulnerabilities. Both options are undesirable. We propose a system for syntactic extensibility where desugaring happens after type checking and desugarings are guaranteed to only generate well-typed code. A major novelty of our work is that desugarings operate on typing derivations instead of plain syntax trees. This provides desugarings access to typing information and forms the basis for the soundness guarantee we provide, namely that a desugaring generates a valid typing derivation. We have implemented our system for syntactic extensibility in a language-independent fashion and instantiated it for a substantial subset of Java, including generics and inheritance. We provide a sound Java extension for Scala-like for-comprehensions.

Language extensions introduce high-level programming constructs that protect programmers from low-level details and repetitive tasks. For such an abstraction barrier to be sustainable, it is important that no errors are reported in terms of generated code. A typical strategy is to check the original user code prior to translation into a low-level encoding, applying the assumption that the translation does not introduce new errors. Unfortunately, such assumption is untenable in general, but in particular in the context of extensible programming languages, such as Racket or SugarJ, that allow regular programmers to define language extensions.