Hello everyone, I juste tried to launch the stringbench on python3.2 and python3.3 dev versions and some bytes tests run slower in python3.3 than in python3.2.

I cc the two raw output of both runs. I also extracted most interesting data (all the tests with more than 20% of performance regression):

• (b"A"*1000).rfind(b"A") (*1000): -70.103093%
• (b"A"*1000).find(b"B") (*1000): -48.372093%
• (b"A"*1000).rindex(b"A") (*1000): -68.888889%
• s=b"ABC"*33; (s+b"E"+(b"D"+s)*500).rfind(s+b"E") (*100): -28.982301%
• (b"C"+b"AB"*300).rfind(b"CA") (*1000): -29.565217%
• (b"AB"*1000).index(b"AB") (*1000): -68.539326%
• b"Andrew".endswith(b"w") (*1000): -21.212121%
• (b"A"*1000).index(b"A") (*1000): -71.111111%
• (b"BC"+b"AB"*300).rfind(b"BC") (*1000): -42.788462%
• b"Andrew".startswith(b"Andrew") (*1000): -20.588235%
• (b"AB"*1000).find(b"AB") (*1000): -69.318182%
• (b"AB"*1000).rfind(b"AB") (*1000): -69.791667%
• (b"A"*1000).rfind(b"B") (*1000): -37.988827%
• (b"AB"*300+"C").index(b"BC") (*1000): -28.750000%
• b"B" in b"A"*1000 (*1000): -24.479167%
• (b"AB"*300+"CA").find(b"CA") (*1000): -33.673469%
• (b"AB"*1000).rindex(b"AB") (*1000): -67.777778%
• (b"C"+"AB"*300).rindex(b"CA") (*1000): -29.017857%
• (b"AB"*300+"C").find(b"BC") (*1000): -28.451883%
• b"Andrew".startswith(b"A") (*1000): -21.212121%
• b"Andrew".startswith(b"Anders") (*1000): -21.212121%
• (b"A"*1000).partition(b"B") (*1000): -30.656934%
• (b"AB"*1000).rfind(b"CA") (*1000): -20.603015%
• (b"AB"*1000).rfind(b"BC") (*1000): -35.645472%
• (b"A"*1000).find(b"A") (*1000): -70.454545%

My environment is:
Mac OS X 10.6.8
GCC i686-apple-darwin10-gcc-4.2.1 (GCC) 4.2.1 (Apple Inc. build 5666) (dot 3)
CPython3.3 revision ea421c534305
CPython3.2 revision 0b86da9d6964

Grouped results.

find (first):

• (b"A"*1000).find(b"A") : -70%

• (b"A"*1000).rfind(b"A") : -70%

• (b"A"*1000).index(b"A") : -71%

• (b"A"*1000).rindex(b"A") : -68%

• (b"AB"*1000).index(b"AB") : -68%

• (b"AB"*1000).rindex(b"AB"): -67%

• (b"AB"*1000).find(b"AB") : -69%

• (b"AB"*1000).rfind(b"AB") : -69%

• b"Andrew".startswith(b"Andrew"): -20%

• b"Andrew".startswith(b"A") : -21%

• b"Andrew".startswith(b"Anders"): -21%

• b"Andrew".endswith(b"w"): -21%

find (last):

• (b"AB"*300+"CA").find(b"CA") : -33%
• (b"C"+"AB"*300).rindex(b"CA") : -29%
• (b"AB"*300+"C").find(b"BC") : -28%
• (b"AB"*300+"C").index(b"BC") : -28%
• (b"C"+b"AB"*300).rfind(b"CA") : -29%
• (b"BC"+b"AB"*300).rfind(b"BC"): -42%
• s=b"ABC"*33; (s+b"E"+(b"D"+s)*500).rfind(s+b"E"): -28%

find (not found):

• (b"A"*1000).find(b"B") : -48%
• (b"A"*1000).rfind(b"B") : -37%
• (b"AB"*1000).rfind(b"CA") : -20%
• (b"AB"*1000).rfind(b"BC") : -35%

others:

• b"B" in b"A"*1000 : -24%
• (b"A"*1000).partition(b"B") : -30%

See also the issue bpo-13621 for results on Unicode.

(b"A"*1000).find(b"A") : -70%

This one is a performance regression introduced by bpo-12170. Attached patch checks object type before trying a conversion to size_t instead of catching an exception.

bytes_find.patch only works for Python int, not object with the __index__ method. My new patch (bytes_find-2.patch) uses PyNumber_Check() instead of PyLong_Check() to be more generic. It fixes also a different issue: raise the same ValueError than bytes.find(-1) on overflow error.

New changeset 75648db1b3f3 by Victor Stinner in branch 'default':
Issue bpo-13623: Fix a performance regression introduced by issue bpo-12170 in
http://hg.python.org/cpython/rev/75648db1b3f3

I checked stringbench: there is no more performance regression (difference of more than 20%).