PyPy’s Number Crunching Optimization

PyPy is a widely known virtual machine for the Python programming language implemented in the RPython subset of Python which includes a tracing Just-In-Time (JIT) compiler. In this article the new auto vectorizer built into the RPython optimizing backend is presented. It uses the linear sequence of instructions of a trace to find parallelism. Dependency information is gathered and used to reschedule some of the instructions as vector statements. This optimization is neither tailored for the NumPy library nor for a specific hardware architecture. Every interpreter written in RPython can benefit from the new optimization. To empirically evaluate the optimization, the x86 assembler backend has been extended to emit SSE4 vector instructions for the optimized traces. Preliminary results show that it is indeed possible to leverage the speed gain SIMD instruction sets offer. The implementation is not very complex and the optimizer is reasonably fast.