Partial Evaluation, Whole-Program Compilation

TL;DR

This paper introduces a partial evaluator that automatically derives optimized compiled code from an interpreter, simplifying the compilation process and improving runtime performance in dynamic language engines.

Contribution

The work presents a novel partial evaluation technique that generates efficient compiled code directly from interpreters, reducing complexity and maintaining correctness.

Findings

Achieved 2.17x speedup on SpiderMonkey JavaScript engine

Achieved 1.84x speedup on PUC-Rio Lua interpreter

Requires only three hours of effort to implement

Abstract

There is a tension in dynamic language runtime design between speed and correctness: state-of-the-art JIT compilation, the result of enormous industrial investment and significant research, achieves heroic speedups at the cost of complexity that can result in serious correctness bugs. Much of this complexity comes from the existence of multiple tiers and the need to maintain correspondence between these separate definitions of the language's semantics; also, from the indirect nature of the semantics implicitly encoded in a compiler backend. One way to address this complexity is to automatically derive, as much as possible, the compiled code from a single source-of-truth; for example, the interpreter tier. In this work, we introduce a partial evaluator that can derive compiled code ``for free'' by specializing an interpreter with its bytecode. This transform operates on the interpreter body at a basic-block IR level and is applicable to almost unmodified existing interpreters in systems languages such as C or C++. We show the effectiveness of this new tool by applying it to the interpreter tier of an existing industrial JavaScript engine, SpiderMonkey, yielding $2.17\times$ speedups, and the PUC-Rio Lua interpreter, yielding $1.84\times$ speedups with only three hours' effort. Finally, we outline an approach to carry this work further, deriving more of the capabilities of a JIT backend from first principles while retaining semantics-preserving correctness.