08-Mar-2025
This is my first blog post ever. I want to use it to say I’m truly sorry for communicating inaccurate results for Python’s tail-calling interpreter. I take full personal responsibility for the oversight that led to it.
About a month ago, I merged a new tail-calling interpreter into Python. That interpreter reported a 9-15% performance boost on Python 3.14’s Whats New page.
These figures turned out to be inaccurate. Long story short, the compiler we were using (Clang 19), had a bug that worsened our baseline performance. We (the CPython developers) were completely unaware of this bug.
The real performance uplift one can expect by upgrading to the tail-calling interpreter is between the 3-5% range. We are not too sure about this figure as well, because we had to compare across different compilers.
Thanks to Nelson Elhage for their excellent investigation into this issue and bringing it up. For more information, you can read their blog post here.
Upon receiving news from Nelson confirming that the Clang 19 bug caused a 10% performance regression on our baseline. I did the following:
If you feel there’s more I could do, please let me know.
This completely blindsided me and I’ve learnt to never trust the compiler when the performance results are too good to be true. That, and to carefully investigate our baselines.
At the time of writing, the Clang 19 bug I talked about is not yet fixed, and it exists in Clang 19, 20, maybe 21-beta. I do not want to blame the LLVM developers for this. Like me, they are probably volunteer contributors as well. Sometimes we make mistakes.
In short, a compiler bug in Clang 19 that we were unaware of resulted in worse baselines. I reported these figures believing they were true. I should have done more investigation into the compiler before reporting these figures. I’m deeply sorry for mistakenly reporting inaccurate numbers.