Trace-based just-in-time compilation for lazy functional programming languages

This thesis investigates the viability of trace-based just-in-time (JIT) compilation for optimising programs written in the lazy functional programming language Haskell. A trace-based JIT compiler optimises only execution paths through the program, which is in contrast to method-based compilers that optimise complete functions at a time. The potential advantages of this approach are shorter compilation times and more natural interprocedural optimisation. Trace-based JIT compilers have previously been used successfully to optimise programs written in dynamically typed languages such as JavaScript, Python, or Lua, but also statically typed languages like Java or the Common Language Runtime (CLR). Lazy evaluation poses implementation challenges similar to those of dynamic languages, so trace-based JIT compilation promises to be a viable approach. In this thesis we focus on program performance, but having a JIT compiler available can simplify the implementation of features like runtime inspection and mobile code. We implemented Lambdachine, a trace-based JIT compiler which implements most of the pure subset of Haskell. We evaluate Lambdachine’s performance using a set of micro-benchmarks and a set of larger benchmarks from the “spectral” category of the Nofib benchmark suite. Lambdachine’s performance (excluding garbage collection overheads) is generally about 10 to 20 percent slower than GHC on statically optimised code. We identify the two main causes for this slow-down: trace selection and impeded heap allocation optimisations due to unnecessary thunk updates.