C[omp]ute

From: andrew cooke <andrew@...>

Date: Thu, 8 Jul 2010 09:04:14 -0400

Extended regular expressions (particularly) have empty transitions that can
occur in loops.  For example, (?(1)a) only matches "a" if group 1 exists, so
(?(1)a)* could be a repated matching od the empty string.

To some extent this is already avoided at compile time, by refusing to parse
things like a**, but there are many possible cases.  The problems for an
implementation are then:

- Whether to warn or reject such cases
- If not rejected, whether to try avoid infinite loops during evaluation

I am currently working on this with RXPY.  In general, I want to (i) provide a
safe system as a default, but (ii) allow the user complete control.  So it
seeems that two flags are necessary: one to disable compile time errors and
one to disable run time safety.

Implementation must also consider efficiency and ease of maintenance /
impementation.  It seems to me that many (but not all) cases could be
automatically rewritten to a safer version, but I don't currently have good
graph rewriting support (I talk about graphs here because the "opcodes" in
RXPY are nodes on a graph; the regular expression is "compiled" to a graph of
these nodes that is then "evaluated" against the input).

Since I do not have rewriting, and because that is not a complete solution
anyway, I need some other runtime scheme.  The best I have found so far is to
add additional nodes that "break" any dangerous loops.  A machine can then
verify that input has beenconsumed between each encounter with such a node.
This keeps almost all the cost to those expressions that need such a feature.

Also, the logic to generate these nodes can also be used to generate compile
time errors.  The graph API includes "consumer(lenient)" where lenient is a
boolean.  If lenient is true then consumer returns True except when a node (or
sequence of nodes) *cannot* consume input.  Repeating such nodes (or
sequences) gives a compile time error.  If lenient is false then consumer
returns True only when a node (or sequence) *guarantees* consumption (in all
cases).  This can be used to detect when to add the runtime check node (ie
when False).

This is not completely implemented, but an initial attempt is looking very
positive - "spinning" in empty loops is avoided at very little cost, and the
integration of logic for compile and runtime checks reduces the code / detail
required.

Andrew