After more thought, I must admit that the behavior change in splitext, I proposed with patch 536120 is not acceptable. So I would instead propose this one which should only improve performances without modifying behavior.
The following bench says that patched splitext is between 2x(for l1) and 25x(for l2) faster than the original one.

The diff patch also test_posixpath.py to check the pitfall described by Tim comments in patch 536120 page.

def splitext(p):
    root, ext = '', ''
    for c in p:
        if c == '/':
            root, ext = root + ext + c, ''
        elif c == '.':
            if ext:
                root, ext = root + ext, c
            else:
                ext = c
        elif ext:
            ext = ext + c
        else:
            root = root + c
    return root, ext

def splitext2(p):
    i = p.rfind('.')
    if i<=p.rfind('/'):
        return p, ''
    else:
        return p[:i], p[i:]

l1 = ('t','.t','a.b/','a.b','/a.b','a.b/.c','a.b/c.d')

l2 = (
'usr/tmp.doc/list/home/sebastien/foo/bar/hghgt/yttyutyuyuttyuyut.tyyttyt',
'usr/tmp.doc/list/home/sebastien/foo/bar/hghgt/yttyutyuyuttyuyut.',
'usr/tmp.doc/list/home/sebastien/foo/bar/hghgt/.tyyttyt',
'usr/tmp.doc/list/home/sebastien/foo/bar/hghgt/yttyutyuyuttyuyut',
'reeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeyttyutyuyuttyuyut.tyyttyt',
'/iuouiiuuoiiuiikhjzekezhjzekejkejkzejkhejkhzejzehjkhjezhjkehzkhjezh.tyyttyt'
    )

for i in l1+l2:
    assert splitext2(i) == splitext(i)

import time

def test(f,args):
    t = time.clock()
    for p in args:
        for i in range(1000):
            f(p)
    return time.clock() - t

def f(p):pass

a=test(splitext, l1)
b=test(splitext2, l1)
c=test(f,l1)
print a,b,c,(a-c)/(b-c)

a=test(splitext, l2)
b=test(splitext2, l2)
c=test(f,l2)
print a,b,c,(a-c)/(b-c)