`append = mylist.append` optimization now harmful? - Python Help - Discussions on Python.org
`append = mylist.append` optimization now harmful? Python Help Stefan2 (Stefan Pochmann) October 12, 2023, 11:09am 1 Python 3.11 optimized loading methods for calls (see the table there, third row from bottom). In the case of list.append , that now seems to beat the optimization of storing the method in a local variable. For dict.get and set.add , it now seems about equally fast: 15.4 ± 0.1 ns a.append(0) 21.2 ± 0.1 ns a_append(0) 37.5 ± 0.2 ns d.get(0) 37.3 ± 0.1 ns d_get(0) 28.5 ± 0.1 ns s.add(0) 27.6 ± 0.1 ns s_add(0) Python: 3.11.4 (main, Sep 9 2023, 15:09:21) [GCC 13.2.1 20230801] With Python 3.10 (on another machine): 38.2 ± 0.1 ns a.append(0) 28.3 ± 0.1 ns a_append(0) 46.9 ± 0.1 ns d.get(0) 34.4 ± 0.1 ns d_get(0) 46.0 ± 0.5 ns s.add(0) 34.3 ± 0.3 ns s_add(0) Python: 3.10.9 (main, Jan 23 2023, 22:32:48) [GCC 10.2.1 20210110] Benchmark script Attempt This Online! from timeit import timeit from statistics import mean, stdev import sys setup = ''' a = [] a_append = a.append d = {0: 0} d_get = d.get s = set() s_add = s.add ''' codes = [ 'a.append(0)', 'a_append(0)', 'd.get(0)', 'd_get(0)', 's.add(0)', 's_add(0)', ] times = {c: [] for c in codes} def stats(c): ts = [t * 1e9 for t in sorted(times[c])[:5]] return f'{mean(ts):4.1f} ± {stdev(ts):3.1f} ns ' for _ in range(500): for c in codes: t = timeit(c, setup, number=10**5) / 1e5 times[c].append(t) for c in codes: print(stats(c), c) print('\nPython:', sys.version) The standard library uses that list.append optimization in various places, for example for copy.deepcopy : cpython/Lib/copy.py at 7dd3c2b80064c39f1f0ebbc1f8486897b3148aa5 · python/cpython · GitHub And removing that optimization does get me faster times, for example for deepcopying list(range(10000)) : 2.83 ± 0.01 ms current 2.75 ± 0.01 ms deoptimized ( Attempt This Online! , the copy module is copied&pasted into the Header section.) So should the standard library remove that optimization everywhere? effigies (Chris Markiewicz) October 12, 2023, 12:51pm 2 Does this persist for the nogil branch? guido (Guido van Rossum) October 12, 2023, 2:18pm 3 I recommend that you read through the source code to find out whether we intentionally did something about this.
Also, 3.11, 3.12 and main (3.13-to-be) might be different. Please don’t. Maybe we can recommend no longer doing this. But let’s not solicit mass PRs removing this. Such PRs tend to introduce bugs. Stefan2 (Stefan Pochmann) October 13, 2023, 7:39pm 4 Not sure what you mean, as that seems obvious from the announcement about PEP 659 that I linked to. Looking at the byte/source code, I see that a.append(0) ends up with LOAD_METHOD_NO_DICT and PRECALL_NO_KW_LIST_APPEND , whereas a_append(0) ends up with LOAD_FAST__LOAD_CONST and PRECALL_NO_KW_BUILTIN_O .
And I guess PRECALL_NO_KW_LIST_APPEND is just significantly faster than PRECALL_NO_KW_BUILTIN_O as it is more special (its implementation looks simpler and hardcodes a _PyList_AppendTakeRef call). Although both are followed by CALL_ADAPTIVE , maybe the difference is in there, I don’t understand enough. No idea, and I’m not equipped to compile/install CPython (also can’t test 3.12 or main). Stefan2 (Stefan Pochmann) October 13, 2023, 7:54pm 5 Moved this to Help, as the suggested action (removing the optimization from the standard library) was rejected and I’m now more interested to see whether others and other versions can reproduce this and could use help for that. storchaka (Serhiy Storchaka) October 14, 2023, 8:11am 6 Also, in the next Python release a new optimization can be added that will make the old trick relevant again (or at least make it not slower that more direct code).
Stefan2 (Stefan Pochmann) October 14, 2023, 12:01pm 7 Yes, that’s partly why I asked whether it’s now harmful, as I don’t know enough to predict future developments but I imagine the people behind the interpreter optimizations might be.