Since PyMapping_Keys always return a list or tuple and most __all__s are list (all in stdlib), I think we can avoid calling PySequence_GetItem for every key and use PySequence_Fast* APIs instead. This doesn't help much since other operations involved are expensive.

With patch:

[bin]$ ./python3 -m perf timeit -s 'import _ast; code=compile("from _ast import *", "<string>", "exec")' 'exec(code)'
....................
Median +- std dev: 11.3 us +- 0.3 us
[bin]$ ./python3 -m perf timeit -s 'import zipfile; code=compile("from zipfile import *", "<string>", "exec")' 'exec(code)'
....................
Median +- std dev: 2.98 us +- 0.04 us

Without patch:

[bin]$ ./python3 -m perf timeit -s 'import _ast; code=compile("from _ast import *", "<string>", "exec")' 'exec(code)'
....................
Median +- std dev: 12.5 us +- 0.3 us
./python3 -m perf timeit -s 'import zipfile; code=compile("from zipfile import *", "<string>", "exec")' 'exec(code)'
....................
Median +- std dev: 3.09 us +- 0.06 us