gh-100762: Switch order of exception match in gen_close

[eendebakpt]

This PR optimizes the closing of generators by switching the match statements in the expression PyErr_ExceptionMatches(PyExc_StopIteration) || PyErr_ExceptionMatches(PyExc_GeneratorExit) from the gen_close method. The case of a PyExc_GeneratorExit is more common.

Benchmark (with optimizations including PGO) updated May 9th:

all: Mean +- std dev: [main2] 283 ns +- 11 ns -> [pr] 280 ns +- 6 ns: 1.01x faster
all with comparison: Mean +- std dev: [main2] 570 ns +- 14 ns -> [pr] 563 ns +- 9 ns: 1.01x faster

Benchmark hidden because not significant (3): generator, list comprehension, generator exhaust

Geometric mean: 1.00x faster

Benchmark script:

import pyperf
runner = pyperf.Runner()

setup="""
l=[1,2]
"""

runner.timeit(name=f"generator", stmt=f"(x for x in l)", setup=setup)
runner.timeit(name=f"list comprehension", stmt=f"[x for x in l]", setup=setup)
stmt="""
for x in (_ for _ in ()):
    pass
"""
runner.timeit(name=f"generator exhaust", stmt=stmt, setup=setup)
runner.timeit(name=f"all", stmt=f"all( (x for x in l) ) ", setup=setup)
runner.timeit(name=f"all with comparison", stmt=f"all( (x==1 for x in l) ) ", setup=setup)

Notes:

• Also see Generator finalization is slower in 3.11 vs 3.10 #100762 (comment)
• The cases for which the PyExc_StopIteration is triggered occur less than the PyExc_GeneratorExit. Cases from the python test suite are test_async_gen_asyncio_aclose_10 in test_asyncgen.py and the following code from test_generators.py:

def f():
    try: 
        yield None
        print('here')
    except GeneratorExit:
        print("exiting")
    print('done')

g = f()
next(g)
g.close()

• Issue: Generator finalization is slower in 3.11 vs 3.10 #100762

[brandtbucher]

Seems good. Let's wait for 3.13 in two weeks, since performance improvements are effectively frozen right now.

[eendebakpt]

Seems good. Let's wait for 3.13 in two weeks, since performance improvements are effectively frozen right now.

@brandtbucher Thanks for looking into this. The main branch has changed in the meantime, and there is currently only a single match instead of the two before. This makes this PR irrelevant, I will close it.