Relational compilation: Functional-to-imperative code generation for performance-critical applications

Abstract

Purely functional programs verified using interactive theorem provers typically need to be translated to run: either by extracting them to a similar language (like Coq to OCaml) or by proving them equivalent to deeply embedded implementations (like C programs). Traditionally, the first approach is automated but produces unverified programs with average performance, and the second approach is manual but produces verified, high-performance programs.

This thesis shows how to recast program extraction as a proof-search problem to automatically derive correct-by-construction, high-performance code from shallowly embedded functional programs. It introduces a unifying framework, relational compilation, to capture and extend recent developments in program extraction, with a focus on modularity and sound extensibility. To demonstrate the value of this approach, it then presents Rupicola, a relational compiler-construction toolkit designed to extract fast, verified, idiomatic low-level code from annotated functional models.

The originality of this approach lies in its combination of foundational proofs, extensibility, and performance, backed by an unconventional take on compiler extensions: unlike traditional compilers, Rupicola generates good code not because of clever built-in optimizations, but because it allows expert users to plug in domain- and sometimes program-specific extensions that allow them to generate exactly the low-level code that they want. This thesis demonstrates the benefits of this approach through case studies and performance benchmarks that highlight how easy Rupicola makes it to create domain-specific compilers that generate code with performance comparable to that of handwritten C programs.