Be Especially Careful with non‐ASCII in Python
There are some really subtle issues with how Python deals with, or fails to deal with, non-ASCII in patterns and strings. Worse, these disparities vary substantially according, not just to which version of Python you are using, but also whether you have a “wide build”.
In general, when you’re doing Unicode stuff, Python 3 with a wide build works best and Python 2 with a narrow build works worst, but all combinations are still a pretty far cry far from how Perl regexes work vis‐à‐vis Unicode. If you’re looking for ???? patterns in Python, you may have to look a bit further afield than its old re module.
The vexing “wide-build” issues have finally been fixed once and for all — provided you use a sufficiently advanced release of Python. Here’s an excerpt from the v3.3 release notes:
Functionality
Changes introduced by PEP 393 are the following:
- Python now always supports the full range of Unicode codepoints, including non-BMP ones (i.e. from U+0000 to U+10FFFF). The distinction between narrow and wide builds no longer exists and Python now behaves like a wide build, even under Windows.
- With the death of narrow builds, the problems specific to narrow builds have also been fixed, for example:
len() now always returns 1 for non-BMP characters, so len('U0010FFFF') == 1;- surrogate pairs are not recombined in string literals, so
'uDBFFuDFFF' != 'U0010FFFF';
- indexing or slicing non-BMP characters returns the expected value, so
'U0010FFFF'[0] now returns 'U0010FFFF' and not 'uDBFF';
- all other functions in the standard library now correctly handle non-BMP codepoints.
- The value of
sys.maxunicode is now always 1114111 (0x10FFFF in hexadecimal). The PyUnicode_GetMax() function still returns either 0xFFFF or 0x10FFFF for backward compatibility, and it should not be used with the new Unicode API (see issue 13054).
The ./configure flag --with-wide-unicode has been removed.
The Future of Python Regexes
In contrast to what’s currently available in the standard Python distribution’s re library, Matthew Barnett’s regex module for both Python 2 and Python 3 alike is much, much better in pretty much all possible ways and will quite probably replace re eventually. Its particular relevance to your question is that his regex library is far more ???? (i.e. it’s much more Perl‐compatible) in every way than re now is, which will make porting Perl regexes to Python easier for you. Because it is a ground‐up rewrite (as in from‐scratch, not as in hamburger :), it was written with non-ASCII in mind, which re was not.
The regex library therefore much more closely follows the (current) recommendations of UTS#18: Unicode Regular Expressions in how it approaches things. It meets or exceeds the UTS#18 Level 1 requirements in most if not all regards, something you normally have to use the ICU regex library or Perl itself for — or if you are especially courageous, the new Java 7 update to its regexes, as that also conforms to the Level One requirements from UTS#18.
Beyond meeting those Level One requirements, which are all absolutely essential for basic Unicode support, but which are not met by Python’s current re library, the awesome regex library also meets the Level Two requirements for RL2.5 Named Characters (N{...})), RL2.2 Extended Grapheme Clusters (X), and the new RL2.7 on Full Properties from revision 14 of UTS#18.
Matthew’s regex module also does Unicode casefolding so that case insensitive matches work reliably on Unicode, which re does not.
The following is no longer true, because regex now supports full Unicode casefolding, like Perl and Ruby.
One super‐tiny difference is that for now, Perl’s case‐insensitive patterns use full string‐oriented casefolds while his regex module still uses simple single‐char‐oriented casefolds, but this is something he’s looking into. It’s actually a very hard problem, one which apart from Perl, only Ruby even attempts.
Under full casefolding, this means that (for example) "?" now correct matches "SS", "ss", "??", "?s" (etc.) when case-insensitive matching is selected. (This is admittedly more important in the Greek script than the Latin one.)
See also the slides or doc source code from my third OSCON2011 talk entitled “Unicode Support Shootout: The Good, the Bad, and the (mostly) Ugly” for general issues in Unicode support across JavaScript, PHP, Go, Ruby, Python, Java, and Perl. If can’t use either Perl regexes or possibly the ICU regex library (which doesn’t have named captures, alas!), then Matthew’s regex for Python is probably your best shot.
N??? B??? s.?.?. (= s’il vous pla?t, et même s’il ne vous pla?t pas :) The following unsolicited noncommercial nonadvertisement was not actually put here by the author of the Python regex library. :)
Cool regex Features
The Python regex library has a cornucopeia of superneat features, some of which are found in no other regex system anywhere. These make it very much worth checking out no matter whether you happen to be using it for its ????‐ness or its stellar Unicode support.
A few of this module’s outstanding features of interest are:
- Variable‐width lookbehind, a feature which is quite rare in regex engines and very frustrating not to have when you really want it. This may well be the most frequently requested feature in regexes.
- Backwards searching so you don’t have to reverse your string yourself first.
- Scoped
ismx‐type options, so that (?i:foo) only casefolds for foo, not overall, or (?-i:foo) to turn it off just on foo. This is how Perl works (or can).
- Fuzzy matching based on edit‐distance (which Udi Manber’s
agrep and glimpse also have)
- Implicit shortest‐to‐longest sorted named lists via
L<list> interpolation
- Metacharacters that specifically match only the start or only the end of a word rather than either side (
m, M)
- Support for all Unicode line separators (Java can do this, as can Perl albeit somewhat begrudgingly with
R per RL1.6.
- Full set operations — union, intersection, difference, and symmetric difference — on bracketed character classes per RL1.3, which is much easier than getting at it in Perl.
- Allows for repeated capture groups like
(w+s+)+ where you can get all separate matches of the first group not just its last match. (I believe C# might also do this.)
- A more straightforward way to get at overlapping matches than sneaky capture groups in lookaheads.
- Start and end positions for all groups for later slicing/substring operations, much like Perl’s
@+ and @- arrays.
- The branch‐reset operator via
(?|...|...|...|) to reset group numbering in each branch the way it works in Perl.
- Can be configured to have your coffee waiting for you in the morning.
- Support for the more sophisticated word boundaries from RL2.3.
- Assumes Unicode strings by default, and fully supports RL1.2a so that
w, s, and such work on Unicode.
- Supports
X for graphemes.
- Supports the
G continuation point assertion.
- Works correctly for 64‐bit builds (
re only has 32‐bit indices).
- Supports multithreading.
Ok, that’s enough hype. :)
Yet Another Fine Alternate Regex Engine
One final alternative that is worth looking at if you are a regex geek is the Python library bindings to Russ Cox’s awesome RE2 library. It also supports Unicode natively, including simple char‐based casefolding, and unlike re it notably provides for both the Unicode General Category and the Unicode Script character properties, which are the two key properties you most often need for the simpler kinds of Unicode processing.
Although RE2 misses out on a few Unicode features like N{...} named character support found in ICU, Perl, and Python, it has extremely serious computational advantages that make it the regex engine of choice whenever you’re concern with starvation‐based denial‐of‐service attacks through regexes in web queries and such. It manages this by forbidding backreferences, which cause a regex to stop being regular and risk super‐exponential explosions in time and space.
Library bindings for RE2 are available not just for C/C++ and Python, but also for Perl and most especially for Go, where it is slated to very shortly replace the standard regex library there.
