STINNER Victor added the comment: Ok, I fixed the PyObject_Call() bug, and tried a new approach to help the compiler for code placement: I moved all "call" functions into a new Objects/call.c file. It should help the compiler to inline more code, and I move functions which call themself closer: _PyObject_FastCall(), _PyCFunction_FastCall() and _PyFunction_FastCall() are now very close in the same file for example.
Not only the stack usage is better, but the performance seems also better: see benchmark results below. Again, if you like the _PyObject_FastCall() idea (to reduce the stack consumption), I will split my big patch into smaller patches to have a "smoother" hg/git history. Benchmark results on speed-python, CPython compiled with LTO+PGO: haypo@speed-python$ python3 -m perf compare_to /home/haypo/benchmarks/2017-02-06_07-15-default-e06af4027546.json pyobject_fastcall-4_ref_e06af4027546.json -G --min-speed=4 Faster (11): - sympy_expand: 949 ms +- 12 ms -> 878 ms +- 7 ms: 1.08x faster (-8%) - chameleon: 21.9 ms +- 0.2 ms -> 20.5 ms +- 0.3 ms: 1.07x faster (-7%) - sympy_str: 425 ms +- 5 ms -> 399 ms +- 5 ms: 1.07x faster (-6%) - sympy_integrate: 40.8 ms +- 0.3 ms -> 38.3 ms +- 0.4 ms: 1.06x faster (-6%) - sympy_sum: 192 ms +- 6 ms -> 181 ms +- 7 ms: 1.06x faster (-6%) - scimark_lu: 352 ms +- 11 ms -> 332 ms +- 13 ms: 1.06x faster (-6%) - xml_etree_generate: 208 ms +- 2 ms -> 197 ms +- 3 ms: 1.05x faster (-5%) - nbody: 233 ms +- 2 ms -> 222 ms +- 3 ms: 1.05x faster (-5%) - telco: 14.7 ms +- 0.3 ms -> 14.0 ms +- 0.3 ms: 1.05x faster (-5%) - scimark_fft: 662 ms +- 10 ms -> 633 ms +- 8 ms: 1.05x faster (-4%) - genshi_text: 71.0 ms +- 0.8 ms -> 68.3 ms +- 0.7 ms: 1.04x faster (-4%) Benchmark hidden because not significant (53): (...) ---------- Added file: http://bugs.python.org/file46559/pyobject_fastcall-4.patch _______________________________________ Python tracker <rep...@bugs.python.org> <http://bugs.python.org/issue29465> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
