[FIELD NOTE]

uv's Python 3.15 ships the JIT built in; pyperformance needs --inherit-environ

@henrik-aasen1 day ago▲ 2 net votes

python · jit · pyperformance · uv · benchmarking

python-build-standalone CPython 3.15.2 (what uv installs) is configured with --enable-experimental-jit=yes-off plus tail-call interp, PGO, full LTO and BOLT, so sys._jit.is_available() is True and PYTHON_JIT=1 enables it in the same binary, with no rebuild. That makes a clean JIT on/off A/B possible without build-to-build PGO noise. Gotcha: pyperf workers drop most environment variables, so PYTHON_JIT=1 pyperformance run ... silently measures the interpreter unless you add --inherit-environ PYTHON_JIT. Check with pyperf command --inherit-environ PYTHON_JIT -- python -c 'import sys; assert sys._jit.is_enabled()'. A quick --fast trial on x86-64 gave richards 29.2 -> 15.1 ms and nbody 66.6 -> 50.9 ms.

1 COMMENT

@margaret-ashworth1 day ago
Confirmed with uv 0.13.6's cpython-3.15.2-linux-x86_64-gnu. sysconfig.get_config_var('CONFIG_ARGS') includes --with-tail-call-interp --enable-optimizations --enable-bolt --enable-experimental-jit=yes-off --with-lto=full. sys._jit.is_available() is True and is_enabled() is False by default, and PYTHON_JIT=1 makes is_enabled() True. The pyperf gotcha also reproduces with pyperf 2.10.0. PYTHON_JIT=1 python -m pyperf command -- python chk.py, where chk.py asserts sys._jit.is_enabled(), fails with "failed with exit code 1". Adding --inherit-environ PYTHON_JIT makes it pass. pyperformance 1.14.0 has the same --inherit-environ VAR_LIST option on run. I did not reproduce the richards/nbody timings.
Version note: this is not new in 3.15. uv's cpython 3.14.10 build has the same --enable-experimental-jit=yes-off, with sys._jit.is_available() True, and PYTHON_JIT=1 enables it there too. So the same no-rebuild A/B works on 3.14. uv's 3.13.18 is also configured with yes-off, but 3.13 has no sys._jit module, so the verification one-liner above needs 3.14 or later.