Meta-tracing makes a fast Racket

Tracing just-in-time (JIT) compilers record and optimize the instruction sequences they observe at runtime. With some modifications, a tracing JIT can perform well even when the executed program is itself an interpreter, an approach called meta-tracing. The advantage of meta-tracing is that it separates the concern of JIT compilation from language implementation, enabling the same JIT compiler to be used with many different languages. The RPython meta-tracing JIT compiler has enabled the efficient interpretation of several dynamic languages including Python (PyPy), Prolog, and Smalltalk. In this paper we present initial findings in applying the RPython JIT to Racket. Racket comes from the Scheme family of programming languages for which there are mature static optimizing compilers. We present the result of spending just a couple person-months implementing and tuning an implementation of Racket written in RPython. The results are promising, with a geometric mean equal to Racket’s performance and within a factor of 2 slower than Gambit and Larceny on a collection of standard Scheme benchmarks. The results on individual benchmarks vary widely. On the positive side, our interpreter is sometimes up to two to five times faster than Gambit, three times faster than Larceny, and two orders of magnitude faster than the Racket JIT compiler when making heavy use of continuations. On the negative side, our interpreter is sometimes two times slower than Racket, eight times slower than Gambit, and six times slower than Larceny.