C[omp]ute

From: andrew cooke <andrew@...>

Date: Sun, 29 May 2011 11:22:09 -0400

RXPY is a Python 2 project that implements the re2 approach to regexps in pure
Python.  Lepl is a recursive descent parser that can delegate to a regular
expression library in various areas (for example, it can compile some
sub-parsers to regular expressions).

Lepl contains two regexp implementations - NFA and DFA - but they have some
problems:
 - They don't implement all regexp features
 - They don't implement the standard Python interface for regular
   expressions (re package)

RXPY addresses both of these issues - all the engines are feature complete and
have the Python interface (note that the docs for RXPY are very incomplete -
the package is in a much better state than described).

Both Lepl and RXPY support both backtracking and "stable" algorithms (stable
in that they don't suffer from combinatorial explosion on certain patterns).
However, the Lepl backtracking engine is particular slow at handling at
pre-compiling a common pattern (for float values).  RXPY has not been tested
on this.

However, Lepl needs more than the standard Python interface.  It requires:
 1 Input to be wrapped in a "stream" abstraction.  This allows "infinite"
   data to be handled with finite resources.
 2 Python 3 support
 3 Multiple and ordered matches
 4 Fast matching of entire groups and lazy matching of sub-groups

The last two points need more explanation:

(3) Multiple and ordered matches are used by the tokenizer.  A set of
different regular expressions are matched against the input and the longest
match wins.  This match must be associated with input pattern (ie we must know
which pattern won); if multiple patterns tie with the longest match we want to
know all patterns; sometimes we may want to tie this back to the ordering of
patterns (in particular, we want the "first" pattern that matches).

This can be achieved by combining named groups with "or" (the | symbol), but
named groups require a slower approach in RXPY.  It may be that the special
case of groups that match the entire string can be optimised into the faster
approach (the main RXPY matcher uses a fast initial match and the re-processes
for groups).  For example, the group name can be embedded in the final (exit)
node(s) of the graph that describe the regular expression.

(4) People tend to use "(...)" to group patterns in regular expressions, even
when they do not want to access the results.  Above I described how RXPY uses
two approaches - a faster, group-free stage and then re-processing for groups.
Currently the second stage is triggered by the expression itself, as it
matches.  Instead, I believe that we can wait until the user requests a named
group.  Since, in many situations, the fast stage can tell whether the
expression matched, this avoids the second stage unless it is "really needed".

One final aim: I would like to make the code particularly well documented.
According to my website logs there is quite a lot of interest in implementing
regular expressions (from students, I guess).  So a clear, well documented
implementation could be useful.

Andrew