In code that supports both byte and unicode strings,
mixing unicode strings with plain character constants
is frequent. E.g. both sre_compile and xmlproc look for
specific characters in an input string. Every usage of
such a character requires default decoding, which will
create a temporary Unicode object.

This patch caches Unicode objects that represent ASCII
characters. On the benchmark

import time
u = u""
t=time.time()
for i in xrange(1000000):
    u+"("
print time.time()-t

it shows a 10% speed-up.

Logged In: YES 
user_id=21627

Attach patch.

Logged In: YES 
user_id=38388

I knew this would come one day :-) 

The patch looks OK, but please also add proper init and
finalize code so that unicode_ascii[] gets cleared up
properly when the interpreter shuts down (this is important
for uses of Python in e.g. mod_snake).

Logged In: YES 
user_id=21627

Committed as 2.83 of unicodeobject.c, with the requested
addition of init/fini code.

Logged In: YES 
user_id=21627

Reopened, since the previous patch broke test_unicodedata.

In this version, the cache is only consulted in DecodeASCII,
since PyUnicode_FromUnicode must not share objects. It also
has the requested init/fini code.

Logged In: YES 
user_id=21627

I've added an alternative patch, which does return shared
objects from PyUnicode_FromUnicode, and corrects the two
places where the result of PyUnicode_FromUnicode did modify
the resulting object.

Logged In: YES 
user_id=38388

Thanks for the update. Digging a little deeper into the
possibilities of sharing Unicode objects I found that there
are some important issues to be taken into consideration
which require a little more work on the sharing code.

I will work on this during the week and get back to you next
week.

Logged In: YES 
user_id=38388

Checked in a modified patch.