bpo-37207: Use PEP 590 vectorcall to speed up set()

[corona10]

PEP 590

./python.exe -m pyperf timeit "set([1, 2, 3, 4, 5])"
.....................
Mean +- std dev: 2.00 us +- 0.04 us
./python.exe -m pyperf timeit "set((1, 2, 3, 4, 5))"
.....................
Mean +- std dev: 1.65 us +- 0.04 us
./python.exe -m pyperf timeit "set()"
.....................
Mean +- std dev: 448 ns +- 15 ns

Master

./python.exe -m pyperf timeit "set([1, 2, 3, 4, 5])"
.....................
Mean +- std dev: 2.30 us +- 0.04 us
./python.exe -m pyperf timeit "set((1, 2, 3, 4, 5))"
.....................
Mean +- std dev: 1.84 us +- 0.05 us
./python.exe -m pyperf timeit "set()"
.....................
Mean +- std dev: 528 ns +- 16 ns

https://bugs.python.org/issue37207

[vstinner]

When doing benchmarks, you can use two Python binaries and use pyperf timeit --compare-to=REF_PYTHON_BINARY (...). It checks that the difference is significant.
https://pyperf.readthedocs.io/en/latest/cli.html#pyperf-timeit

Or you can store results in JSON files and use pyperf compare_to command.

[vstinner]

I merged PR #18980, please rebase your PR on top of master and modify your PR to use _PyArg_NoKwnames().

[vstinner]

LGTM.

[vstinner]

IMHO it's worth it to add a few lines of C code to make tuple() and set() constructors faster.

[corona10]

IMHO it's worth it to add a few lines of C code to make tuple() and set() constructors faster

Yeah, I think the PEP 590 implemented very well to bring performance enhancement very easily.

[vstinner]

Yeah, I think the PEP 590 implemented very well to bring performance enhancement very easily.

What I mean is that vectorcall should not be used for everything. But set() and tuple() are very commonly used, so here it's worth it.