This repository contains nix configuration to build CPython against various versions of GCC and clang/LLVM, using various configuration options. I used this to explore the performance of the new tail-call interpreter and to explore the impact of [an LLVM regression I found][regression] starting with LLVM 19.
See data for the raw data and full comparisons.
I benchmarked several configurations, on both an Intel server (a Raptor Lake i5-13500 I maintain in Hetzner), and on my Apple M1 Macbook Air.
These configurations are:
clang18: Built using clang 18.1.8, using computed gotos.gcc(Intel only): Built with GCC 14.2.1, using computed gotos.clang19: Built using clang 19.1.7, using compute gotos.clang19.tc: Built using clang 19.1.7, using the new tail-call interpreter.clang19.taildup: Built using clang 19.1.7, computed gotos and some magic-mllvmoptions which work around the regression.
All builds use LTO and PGO.
| Platform | clang18 | clang19 | clang19.taildup | clang19.tc | gcc |
|---|---|---|---|---|---|
| Raptor Lake i5-13500 | (ref) | 1.09x slower | 1.01x faster | 1.03x faster | 1.02x faster |
| Apple M1 Macbook Air | (ref) | 1.12x slower | 1.02x slower | 1.00x slower | N/A |
Based on these results, I conclude that the tail-call interpreter represents a small (1-2%) speedup relative to the computed-goto interpreter, except for the fact that clang 19 (and 20) contain a regression that merges the computed gotos too-eagerly, which incurs a ~10% performance impact(!)