gh-74598: add fnmatch.filterfalse for excluding names

[picnixz]

In this implementation, I did not use a lambda function as it was proposed in the original patch nor did I use auxiliary functions. The reason is:

# without lambda function
python -m pyperf timeit -s "import fnmatch; data=['Python', 'Ruby', 'Perl', 'Tcl']" 'fnmatch.filterfalse(data, "P*")'
.....................
Mean +- std dev: 1.72 us +- 0.04 us

# with lambda function
./python -m pyperf timeit -s "import fnmatch; data=['Python', 'Ruby', 'Perl', 'Tcl']" 'fnmatch.filterfalse(data, "P*")'
.....................
Mean +- std dev: 2.24 us +- 0.06 us

Therefore, I think it's better not to use a lambda function. I have a small question but would it make sense to re-implement fnmatch in C? (this may be a question for @barneygale)

Additional benchmarks

small_match_none_loop: Mean +- std dev: 1.54 us +- 0.05 us
small_match_none_iter: Mean +- std dev: 1.61 us +- 0.07 us

small_match_some_loop: Mean +- std dev: 1.71 us +- 0.04 us
small_match_some_iter: Mean +- std dev: 1.89 us +- 0.06 us

small_match_all_loop: Mean +- std dev: 1.76 us +- 0.03 us
small_match_all_iter: Mean +- std dev: 2.00 us +- 0.05 us

medium_match_none_loop: Mean +- std dev: 10.3 us +- 0.2 us
medium_match_none_iter: Mean +- std dev: 8.75 us +- 0.26 us

medium_match_some_loop: Mean +- std dev: 12.1 us +- 0.1 us
medium_match_some_iter: Mean +- std dev: 11.1 us +- 0.1 us

medium_match_all_loop: Mean +- std dev: 12.9 us +- 0.2 us
medium_match_all_iter: Mean +- std dev: 12.4 us +- 0.3 us

large_match_none_loop: Mean +- std dev: 98.2 us +- 1.3 us
large_match_none_iter: Mean +- std dev: 80.6 us +- 2.7 us

large_match_some_loop: Mean +- std dev: 119 us +- 4 us
large_match_some_iter: Mean +- std dev: 104 us +- 2 us

large_match_all_loop: Mean +- std dev: 127 us +- 2 us
large_match_all_iter: Mean +- std dev: 115 us +- 2 us

The benchmark script is:

import os
import posixpath
import pyperf
import itertools
from fnmatch import _compile_pattern

def filterfalse_loop(names, pat):
    result = []
    pat = os.path.normcase(pat)
    match = _compile_pattern(pat)
    if os.path is posixpath:
        # normcase on posix is NOP. Optimize it away from the loop.
        for name in names:
            if match(name) is None:
                result.append(name)
    else:
        for name in names:
            if match(os.path.normcase(name)) is None:
                result.append(name)
    return result

def filterfalse_iter(names, pat):
    pat = os.path.normcase(pat)
    match = _compile_pattern(pat)
    if os.path is posixpath:
        # normcase on posix is NOP. Optimize it away from the loop.
        return list(itertools.filterfalse(match, names))

    result = []
    for name in names:
        if match(os.path.normcase(name)) is None:
            result.append(name)
    return result

runner = pyperf.Runner()
base_data = ['Python', 'Ruby', 'Perl', 'Tcl']
for data_label, data_size in [('small', 1), ('medium', 10), ('large', 100)]:
    for pat_label, pattern in [('none', 'A*'), ('some', 'P*'), ('all', '*')]:
        data = base_data * data_size
        runner.bench_func(f'{data_label}_match_{pat_label}_loop', filterfalse_loop, data, pattern)
        runner.bench_func(f'{data_label}_match_{pat_label}_iter', filterfalse_iter, data, pattern)

• Issue: Add fnmatch.filterfalse function #74598

---

📚 Documentation preview 📚: https://cpython-previews--121185.org.readthedocs.build/

[picnixz]

FT failure is known so I'll wait until the fix (#130724) is merged.