by Marcial Gaißert, Carl-Friedrich Bolz-Tereick, and Jonathan Brachthäuser

In Proceedings of the International Conference on Object-Oriented Programming, Systems, Languages and Applications (OOPSLA), 2025.

Abstract

Effect handlers are a programming language feature that has recently gained popularity. They allow for non-local yet structured control flow and subsume features like generators, exceptions, asynchronicity, etc. However, implementations of effect handlers currently often sacrifice features to enable efficient implementations. Meta-tracing just-in-time (JIT) compilers promise to yield the performance of a compiler by implementing an interpreter. They record execution in a trace, dynamically detect hot loops, and aggressively optimize those using information available at runtime. They excel at optimizing dynamic control flow, which is exactly what effect handlers introduce. We present the first evaluation of tracing JIT compilation specifically for effect handlers. To this end, we developed RPython-based tracing JIT implementations for Eff, Effekt, and Koka by compiling them to a common bytecode format. We evaluate the performance, discuss which classes of effectful programs are optimized well and how our additional optimizations influence performance. We also benchmark against a baseline of state-of-the-art mainstream language implementations.