gh-91146: More reduce allocation size of list from str.split/rsplit

[corona10]

Memory size

AS-IS 50b2261

>>> import sys
>>> s = "1 2".split()
>>> sys.getsizeof(s)
88
>>> s = "12345".split()
>>> sys.getsizeof(s)
104
>>> s = "1 2 3 4 5".split()
>>> sys.getsizeof(s)
136
>>> s = "1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5".split()
>>> sys.getsizeof(s)
280

TO-BE (Allocation is reduced & No regression)

>>> import sys
>>> s = "1 2".split()
>>> sys.getsizeof(s)
80
>>> s = "12345".split()
>>> sys.getsizeof(s)
88
>>> s = "1 2 3 4 5".split()
>>> sys.getsizeof(s)
104
>>> s = "1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5".split()
>>> sys.getsizeof(s)
280

Performance: No regression

Script

import pyperf

def bench_split_tiny():
    s = "1 2".split()

def bench_split_small():
    s = "1 2 3 4 5".split()

def bench_split_larger():
    s = "1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5".split()

def bench_split_tiny_append():
    s = "1 2".split()
    for e in range(100):
        s.append("9")

def bench_split_small_append():
    s = "1 2 3 4 5".split()
    for e in range(100):
        s.append("9")

def bench_split_larger_append():
    s = "1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5".split()
    for e in range(100):
        s.append("9")

runner = pyperf.Runner()
runner.bench_func('bench_split_tiny', bench_split_tiny)
runner.bench_func('bench_split_small', bench_split_small)
runner.bench_func('bench_split_larger', bench_split_larger)
runner.bench_func('bench_split_tiny_append', bench_split_tiny_append)
runner.bench_func('bench_split_small_append', bench_split_small_append)
runner.bench_func('bench_split_larger_append', bench_split_larger_append)

Result

Benchmark	origin (before 50b2261)	50b2261	PR
bench_split_small	63.3 ns	62.5 ns: 1.01x faster	62.5 ns: 1.01x faster
bench_split_larger	164 ns	166 ns: 1.01x slower	not significant
bench_split_small_append	1.61 us	1.56 us: 1.03x faster	not significant
Geometric mean	(ref)	1.00x faster	1.00x faster

Benchmark hidden because not significant (3): bench_split_tiny, bench_split_tiny_append, bench_split_larger_append

• Issue: Surprising list overallocation from .split() #91146

[corona10]

@methane san,

I adopt your suggestion from #95473 (comment).
When considering that focusing on memory usage is more important, there is no performance impact with this patch when comparing the performance before 50b2261.
IMO, your suggestion is more proper for this issue :)

summary:

1. Memory usage win
2. No performance impact

PTAL.

[corona10]

@methane san
Thanks for the sharp code review!