Out-Fibbing CPython with the Plush Interpreter
18 days ago
- #programming
- #benchmarking
- #optimization
- Plush is a toy programming language with actor-based parallelism.
- The author aims to optimize the Plush interpreter for better performance.
- The recursive Fibonacci (fib) benchmark is used to measure performance.
- Plush initially runs fib(38) in 9.10s, slower than Python's 5.70s.
- Optimizations include combining push and call into call_direct, reducing time to 8.44s.
- Using perf tool, the author identifies hash lookup overhead in function calls.
- Introducing call_pc instruction with self-modifying code reduces time to 5.13s.
- Adding add_i64 instruction for integer operations further reduces time to 4.57s.
- Optimizations show no impact on a parallel raytracer program, highlighting microbenchmark limitations.