gh-125159: Speed up posixpath.realpath() in the common case

[nineteendo]

Originally proposed in #120146.

Benchmark

Show script

# patch-1.sh
mkdir -p "tmp/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/\
a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a/a"

echo 200 .
main/python.exe    -m timeit -s "import os; path = './' * 200"   "os.path.realpath(path)"
patch-1/python.exe -m timeit -s "import os; path = './' * 200"   "os.path.realpath(path)"

echo 200 ..
main/python.exe    -m timeit -s "import os; path = '../' * 200"   "os.path.realpath(path)"
patch-1/python.exe -m timeit -s "import os; path = '../' * 200"   "os.path.realpath(path)"

echo 200 dirs
main/python.exe    -m timeit -s "import os; path = os.getcwd() + '/tmp' + '/a' * 200" "os.path.realpath(path)"
patch-1/python.exe -m timeit -s "import os; path = os.getcwd() + '/tmp' + '/a' * 200" "os.path.realpath(path)"

200 . - no difference

10000 loops, best of 5: 26.9 usec per loop
10000 loops, best of 5: 26.7 usec per loop

200 .. - 1.02x faster

5000 loops, best of 5: 50.3 usec per loop
5000 loops, best of 5: 49.2 usec per loop

200 dirs - no difference

100 loops, best of 5: 2.19 msec per loop
100 loops, best of 5: 2.16 msec per loop

$ python -m timeit -s "name = '.'" "pass" "if name is None: pass"
50000000 loops, best of 5: 9.03 nsec per loop
$ python -m timeit -s "name = '.'" "pass"
50000000 loops, best of 5: 6.08 nsec per loop

• Issue: Speed up posixpath.realpath() in the common case #125159

[nineteendo]

cc @barneygale

[ericvsmith]

Does this make a measurable difference?

[Wulian233]

Could you provide the results of the speed test?

[nineteendo]

It's not a lot faster, 2.95 nsec/iteration.

[ericvsmith]

I think this change isn't worth the churn.

[nineteendo]

Just as expected.

[erlend-aasland]

Just as expected.

If I understand you correctly, you expected this PR to be rejected.

The next time you consider opening a PR that you "expect to be rejected", please consider again the following:

• is it worth the reviewer and CI churn to post a "ready-to-be-rejected" PR?
• is there anything you can improve pre proposing the PR that can increase the chances of it being accepted? (benchmarks, seeking consensus about the change, backwards compatibility, etc.) See also the devguide: Making good PRs

[barneygale]

@erlend-aasland it's my fault - see #120146 (comment)

@nineteendo really sorry for giving you bad advice and wasting your time

[nineteendo]

Thanks for being honest, Barney, but it's also my fault. I knew it wasn't a lot faster (much less than the 5% Alex mentioned in the past) and therefore probably wasn't worth a PR to change just that. That doesn't mean that making the code 3 nanoseconds slower for speeding up a rare case is fine though.

[vstinner]

In general, an optimization is worth it if it's at least 10% faster on a microbenchmark. It doesn't seem to be the case here.