bpo-46323: Use PyObject_Vectorcall while calling ctypes callback function

[corona10]

https://bugs.python.org/issue46323

With v3 benchmark

Mean +- std dev: [base] 1.32 us +- 0.02 us -> [opt] 1.18 us +- 0.02 us: 1.13x faster

[corona10]

@erlend-aasland Can you please take a look?

[corona10]

Raised RLIMIT_NOFILE: 256 -> 1024
0:00:00 load avg: 4.50 Run tests sequentially
0:00:00 load avg: 4.50 [1/1] test_ctypes
beginning 6 repetitions
123456
......

== Tests result: SUCCESS ==

1 test OK.

Total duration: 1.9 sec
Tests result: SUCCES

[corona10]

Tests / Address sanitizer (pull_request)

             ^^^^^^^^^^^
  File "/home/runner/work/cpython/cpython/Lib/urllib/request.py", line 643, in http_error_default
    raise HTTPError(req.full_url, code, msg, hdrs, fp)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
urllib.error.HTTPError: HTTP Error 404: Not Found

Unrelated

[vstinner]

Mean +- std dev: [base] 552 ns +- 6 ns -> [vectorcall] 541 ns +- 5 ns: 1.02x faster

-11 ns is not really impressive :-(

[corona10]

@vstinner Updated!

[erlend-aasland]

@erlend-aasland Can you please take a look?

Sorry, I don't have time until earliest next weekend :(

(I remember doing similar stuff with some of the sqlite3 callbacks some months ago, but I discarded those changes because the added complexity did not outweigh the performance gain.)

[corona10]

@vstinner cc @erlend-aasland

Yeah the performance improvement itself is not impressive,
but considering _ctypes callback feature is widely used, doing something is better than doing nothing since it shows consistently 1-2 % improvement.
And we are fighting for 1-2 % improvement for these days :)

[vstinner]

Yeah the performance improvement itself is not impressive, but considering _ctypes callback feature is widely used, doing something is better than doing nothing since it shows consistently 1-2 % improvement. And we are fighting for 1-2 % improvement for these days :)

My rule is that an micro-optimization is worth it if it's at least 10% faster on a micro-benchmark. On a macro-benchmark like pyperformance, smaller speedup are worth it, but I have no global rule for that.

Here I would expect to save 20 ns by avoiding the creating a tuple for positional arguments, but it's about -11 ns. Maybe "20 ns" is what I had in mind for my laptop CPU, but on your CPU, it's closer to 11 ns?

[vstinner]

My rule is that an micro-optimization is worth it if it's at least 10% faster on a micro-benchmark.

It's not a strict rule and you're free to not follow it ;-)

[vstinner]

LGTM.

[erlend-aasland]

LGTM (post merge 😆), good job! (I'll revisit my sqlite3 vectorcall branches now.)

[erlend-aasland]

This comment is now misleading; it should be removed. Previously, PySequence_Length could return -1 on error, but PyTuple_GET_SIZE always succeeds (or, more correct: it does no error checking).