Assessing the Overhead of ML Exceptions by Selective CPS Transformation

Abstract

ML's exception handling makes it possible to describe exceptional execution flows conveniently, but it also forms a performance bottleneck. Our goal is to reduce this overhead by source-level transformation.
To this end, we transform source programs into continuation-passing style (CPS), replacing handle and raise expressions by continuation-catching and throwing expressions, respectively. CPS-transforming every expression, however, introduces a new cost. We therefore use an exception analysis to transform expressions selectively: if an expression is statically determined to involve exceptions then it is CPS-transformed; otherwise, it is left in direct style.
In this article, we formalize this selective CPS transformation, prove its correctness, and present early experimental data indicating its effect on ML programs.