Alternative compilation method for regular expressions
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

Alternative Regular Expression Engine

NodeJS javascript library with an alternative compilation method for regular expressions.

1. Installation

alt-regex-engine can be installed for Node using npm.

Using npm:

npm install alt-regex-engine

2. Introduction

In order to execute a regular expression, the expression needs to be compiled into a set of state-transitions. For example, let us investigate the following regular expression:

    a b c d

Each time after receiving a character, the regular expression walker will move to a new state, expecting a new character:

    [0] a [1] b [2] c [3] d [4]F

The initial state for the walker is state [0]. According the our example regular expression, in state zero, it expects the character a. After receiving character a, the walker proceeds to state [1]. If the walker receives any other character in state [0], it will not walk to state [1] but terminate with a FAIL. The walker will keep walking until he reaches the final state [4]F and return SUCCESS in that stage.

In table format, the walker keeps querying the following table:

state/from character state/to final
0 a 1
1 b 2
2 c 3
3 d 4 F

In JSON format, the table looks like this:


Such state-transition table is traditionally also called a DFA, a Deterministic Finite Automaton. It is called deterministic because there is no confusion possible for the walker if he knows his current state and he knows what character he received. There will be only one new state to go to.

In other words, the function newState(currentState,character) returns a single state number for a DFA and not an array of them. When the function newState could return an array instead of a single number, it is not a DFA but an NFA.

3. Kleene operators

In his work throughout the 1950s, Stephen Kleene introduced the concept of regular language operators.

Note: Let's represent nothing by something: Traditionally, we represent nothing by ε. It stands for: nothing at all. So, yes, it is a bit paradoxical that we need something to represent nothing.

3.1. Kleene Star: zero or more repetitions of a pattern

(ab)* ε
abababab ...

3.2. Kleene Plus: One or more repetitions of a pattern

(ab)+ ab
abababab ...

3.3. Kleene Option: Zero or one times the pattern

(ab)? ε

3.4. Kleene OR: One pattern or another

(ab|cd) ab

4. State transitions in the presence of Kleene operators

Let us investigate the following regular expression:


The state decoration rules are the following:

  • in front of each character we add another state
  • at the end of the regular expression, we add the final state

Therefore, the decorated expression becomes:

    [0] r( [1] a [2] b)* [3] c [4]F

According the Kleene expansion rule, the expression contains the following transitions:

    0: [0] r [3] c [4]F
    1: [0] r [1] a [2] b [1] a [2] b [3] c [4]F
    2: [0] r [1] a [2] b [1] a [2] b [3] [1] a [2] b [3] c [4]F
    3: [0] r [1] a [2] b [1] a [2] b [3] [1] a [2] b [3] [1] a [2] b [3]c [4]F

If you look carefully at the expanded expressions, you will notice that no new transitions have been introduced during repetition 3 (or 4 or 5 if you enumerate them too). All possible transitions in a Kleene Star are fully contained in repetitions 0,1, and 2.

It is this property that allows for a simplification of Glushkov's algorithm.

In general terms, we can state that the following reduction rules apply to Kleene operators:

    operator        alternative rules

    (ab)*           ε  ab  abab
    (ab)+              ab  abab
    (ab)?           ε  ab 

We can compute the transitions in an expression containing a Kleene operator by computing the transitions for its reduced rules.

For the example, computing the transitions in:

    [0] r( [1] a [2] b)* [3] c [4]F

amounts to computing the transitions in the following flattened expressions:

    [0] r [3] c [4]F
    [0] r [1] a [2] b [1] a [2] b [3] c [4]F
    [0] r [1] a [2] b [1] a [2] b [3] [1] a [2] b [3] c [4]F

The results are equivalent to computing them on the full expression.

The program in Javascript for NodeJS is a practical demonstration for the statement that we can compute Kleene's closure by applying systematically the reduction rules mentioned above. Using the simple technique demonstrated in the first example, you can derive manually the following transitions:


The result is an NFA. Using a simple disambiguation technique (see below), you can from there derive the DFA:


5. Compilation steps

5.1. lexer

The lexer is a simple program that accepts a regular expression as input and returns a set of tokens as output. The tokens are either a state or a character. For example:



    [0]  r  (  [1]  a  [2]  b  )  *  [3]  c  [4]F

5.2. parser

The parser removes all subexpressions in brackets from the expression or its subexpressions and replaces them by an expression token. For example:

    [0]  r  (  [1]  a  [2]  b  )  *  [3]  c  [4]F


    [0]  r  exp1  *  [3]  c  [4]F 

With the collection of stored expressions:

    1: [1]  a  [2]  b

The parser does this recursively. For example:



    [0]  r  (  [1]  t  (  [2]  a  [3]  b  )  *  [4]  y  )  [5]F 

And after parsing:

    [0]  r  exp2  [5]F 

With the collection of stored expressions:

    1: [2]  a  [3]  b 
    2: [1]  t  exp1  *  [4]  y 

5.3. flattener

The main algorithm to compute the NFA is the flattener. It works as following. It takes as input the parsed expression. From there, it looks for Kleene operators. If it finds one, it reduces the operator using the reduction rules and stores the new expressions in a queue; and starts processing the queue again. If it cannot find operators in a queued expression, it brings back the subexpressions it finds and puts the expression back in the queue. The flattener keeps processing the queue until no operators nor expressions can be found in an expression. Then, it is ready to leave the queue and joined the flattened expressions.

Contrary to Thompson's classical algorithm, my algorithm does not use a stack but a queue. I do not know, however, without further investigation whether it is actually any faster than the Thompson-McNaughton-Yamada approach.

For the example:


The flattened expressions look like this:

    [0]  r  [1]  t  [2]  a  [3]  b  [4]  y  [5]F 
    [0]  r  [1]  t  [4]  y  [5]F 
    [0]  r  [1]  t  [2]  a  [3]  b  [2]  a  [3]  b  [4]  y  [5]F 

5.4. transition deriver

The transition deriver will just go through each flattened expression and produce the transitions. For example:

    ... [1]  t  [2] ...

yields the transition:

    state/from: 1
    character: t 
    state/to: 2

For the example:


The NFA transitions look like this:


5.5. transition compressor

In order to prepare the disambiguation of the NFA into a DFA, the transition compressor will create one record per combination of from/char:


5.6. transition disambiguator

The example contains two ambiguous transitions. For example, the following transition:


indicates two different states, state [4] and state [2] that can be reached by the walker when he sees a char b in state [3]. Therefore, the walker will reach a new, combined state [6], which is the combination of both state [2] and state [4].

The disambiguator will replace the transition itself by:


Since neither state [4] nor state [2] are final states, the new state is not final either. All transitions departing from either state [2] or state [4] will also depart from state [6]. Therefore, the disambiguator creates the following new transitions out from state [6]:


After disambiguating all ambiguous transitions, we end with the following disambiguated transitions:


5.7. transition purger

The transition purger simply removes the transition fields that were only needed for disambiguation and are no longer needed any further:


The transition purger yields the final DFA.

6. Using the transitions to match

In the test folder in the demonstration library, you can find a simplistic walker implementation that matches a regular expression pattern to a given text:

    $ ./test-walking.js 
    pattern: (ax)*b
    text: ztaxaxbc        match: axaxb
    text: ewrwere         match: 
    text: axb             match: axb
    text: b               match: b
    text: trbtr           match: b
    text:                 match: 

7. Performance

It is probably a bit naive to state that the compilation time increases with the size of the regular expression pattern. This is not really true. Performance degrades especially with the complexity of the expressions. The more Kleene operators -- embedded in subexpressions or not -- the larger the number of flattened expressions to compute. I do not think that this is a property tied to this algorithm. It is tied to the fact that the more operators there can be found in the expression, the more transitions there will be to derive.

8. License

Copyright (c) 2012 Erik Poupaert.
Licensed under the Library General Public License (LGPL).