The CPython project recently made significant improvements to their bytecode interpreter resulting in a 10-15% performance boost. However, it was later discovered that the gains were primarily due to working around an LLVM 19 regression, leading to a more modest improvement of 1-5% when compared to a better baseline. The introduction of the tail-calling interpreter sparked curiosity and investigation into its effectiveness, revealing complexities in benchmarking and performance engineering. Surprisingly, it was found that the whole computed-goto interpreter may have been unnecessary complexity, as some compilers can replicate the intended optimization without it. The post delves into the challenges of benchmarking, performance engineering, and software complexity.
https://blog.nelhage.com/post/cpython-tail-call/