New submission from Inada Naoki <songofaca...@gmail.com>:
When PGO is not used, compilers don't know which part is hot.
So gcc failed to inline hot code in pymalloc_alloc and pymalloc_free
into _PyObject_Malloc and _PyObject_Free. For example, only this code is
inlined into _PyObject_Malloc.
if (nbytes == 0) {
return 0;
}
if (nbytes > SMALL_REQUEST_THRESHOLD) {
return 0;
}
But the hottest part is taking memory block from freelist in the pool.
To optimize it,
* make pymalloc_alloc and pymalloc_free inline functions
* Split code for rare / slow paths out to new functions
In PR 14674, pymalloc is now as fast as mimalloc in spectral_norm benchmark.
$ ./python bm_spectral_norm.py --compare-to=./python-master
python-master: ..................... 199 ms +- 1 ms
python: ..................... 176 ms +- 1 ms
Mean +- std dev: [python-master] 199 ms +- 1 ms -> [python] 176 ms +- 1 ms:
1.13x faster (-11%)
----------
components: Interpreter Core
messages: 347615
nosy: inada.naoki
priority: normal
severity: normal
status: open
title: Optimize pymalloc for non PGO build
type: performance
versions: Python 3.9
_______________________________________
Python tracker <rep...@bugs.python.org>
<https://bugs.python.org/issue37543>
_______________________________________
_______________________________________________
Python-bugs-list mailing list
Unsubscribe:
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
