path: root/pod/perlreguts.pod

=head1 NAME

perlreguts - Description of the Perl regular expression engine.

=head1 DESCRIPTION

This document is an attempt to shine some light on the guts of the regex
engine and how it works. The regex engine represents a signifigant chunk
of the perl codebase, but is relatively poorly understood. This document
is a meagre attempt at addressing this situation. It is derived from the
authors experience, comments in the source code, other papers on the
regex engine, feedback in p5p, and no doubt other places as well.

B<WARNING!> It should be clearly understood that this document
represents the state of the regex engine as the author understands it at
the time of writing. It is B<NOT> an API definition, it is purely an
internals guide for those who want to hack the regex engine, or
understand how the regex engine works. Readers of this document are
expected to understand perls regex syntax and its usage in detail, if
you are a beginner you are in the wrong the place.

=head1 OVERVIEW

=head2 A quick note on terms

There is some debate as to whether to say 'regexp' or 'regex'. In this
document we will use the term "regex" unless there is a special reason
not to, and then we will explain why.

When speaking about regexes we need to distinguish between their source
code form and their internal form. In this document we will use the term
"pattern" when we speak of their textual, source code form, the term
"program" when we speak of their internal representation. These
correspond to the terms C<S-regex> and C<B-regex> that Mark Jason
Dominus employs in his paper on "Rx"[1].

=head2 What is a regular expression engine?

A regular expression engine is a program whose job is to efficiently
find a section of a string that matches a set criteria of criteria. The
criteria is expressed in text using a formal language. See perlre for a
full definition of the language.

So the job in less grandiose terms is to some turn a pattern into
something the computer can efficiently use to find the matching point in
the string.

To do this we need to produce a program by parsing the text. We then
need to execute the program to find the point in the string that
matches. And we need to do the whole thing efficiently.

=head2 Structure of a Regexp Program

=head3 High Level

Although it is a bit confusing and some object to the terminology it
is worth taking a look at a comment that has
been in regexp.h for years:

I<This is essentially a linear encoding of a nondeterministic
finite-state machine (aka syntax charts or "railroad normal form" in
parsing technology).>

The term "railroad normal form" is a bit esoteric, with "syntax
diagram/charts", or "railroad diagram/charts" being more common terms.
Nevertheless it provides a useful mental image of a regex program: Each
node can be thought of as a unit of track, with a single entry and in
most cases a single exit point (there are pieces of track that fork, but
statistically not many),  and the total forms a system of track with a
single entry and single exit point. The matching process can be thought
of as a car that moves on the track, with the particular route through
the system being determined by the character read at each possible
connector point. A car can roll off the track at any point but it may
not procede unless it matches the track...

Thus the pattern C</foo(?:\w+|\d+|\s+)bar/> can be thought of as the
following chart:

                  [start]
                     |
                   <foo>
                     |
                 +---+---+
                 |   |   |
               <\w+> | <\s+>
                 | <\d+> |
                 |   |   |
                 +---+---+
                     |
                   <bar>
                     |
                   [end]

The truth of the matter is that perls regular expressions these days are
way beyond such models, but they can help when trying to get your
bearings, and they do match pretty closely with the current
implementation.

To be more precise we will say that a regex program is an encoding
of a graph.  Each node in the graph corresponds to part of
the original regex pattern, such as a literal string or a branch,
and has a pointer to the nodes representing the next component
to be matched. Since "node" and opcode are overloaded terms in the
perl source, we will call the nodes in a regex program 'regops'.

The program is represented by an array of regnode structures, one or
more of which together represent a single regop of the program. Struct
regnode is the smallest struct needed and has a field structure which is
shared with all the other larger structures.

"Next" pointers of all regops except BRANCH implement concatenation; a
"next" pointer with a BRANCH on both ends of it is connecting two
alternatives.  [Here we have one of the subtle syntax dependencies:  an
individual BRANCH (as opposed to a collection of them) is never
concatenated with anything because of operator precedence.

The operand of some types of regop is a literal string; for others,
it is a regop leading into a sub-program.  In particular, the operand
of a BRANCH node is the first regop of the branch.

B<NOTE>: As the railroad metaphor suggests this is B<not> a tree
structure:  the tail of the branch connects to the thing following the
set of BRANCHes.  It is a like a single line of railway track that
splits as it goes into a station or railway yard and rejoins as it comes
out the other side.

=head3 Regops

The base structure of a regop is defined in regexp.h as follows:

    struct regnode {
        U8  flags;    /* Various purposes, sometimes overriden */
        U8  type;     /* Opcode value as specified by regnodes.h */
        U16 next_off; /* Offset in size regnode */
    };

Other larger regnode-like structures are defined in regcomp.h. They
are almost like subclasses in that they have the same fields as
regnode, with possibly additional fields following in
the structure, and in some cases the specific meaning (and name)
of some of base fields are overriden. The following is a more
complete description.

=over 4

=item regnode_1

=item regnode_2

regnode_1 structures have the same header, followed by a single
four-byte argument; regnode_2 structures contain two two-byte
arguments instead:

    regnode_1                U32 arg1;
    regnode_2                U16 arg1;  U16 arg2;

=item regnode_string

regnode_string structures, used for literal strings, follow the header
with a one-byte length and then the string data. Strings are padded on
the end with zero bytes so that the total length of the node is a
multiple of four bytes:

    regnode_string           char string[1];
                             U8 str_len; (overides flags)

=item regnode_charclass

character classes are represented by regnode_charclass structures,
which have a four-byte argument and then a 32-byte (256-bit) bitmap
indicating which characters are included in the class.

    regnode_charclass        U32 arg1;
                             char bitmap[ANYOF_BITMAP_SIZE];

=item regnode_charclass_class

There is also a larger form of a char class structure used to represent
POSIX char classes called regnode_charclass_class which contains the
same fields plus an additional 4-byte (32-bit) bitmap indicating which
POSIX char class have been included.

    regnode_charclass_class  U32 arg1;
                             char bitmap[ANYOF_BITMAP_SIZE];
                             char classflags[ANYOF_CLASSBITMAP_SIZE];

=back

regnodes.h defines an array called regarglen[] which gives the size
of each opcode in units of size regnode (4-byte). A macro is used
to calculate the size of an EXACT node based on its C<str_len> field.

The opcodes are defined in regnodes.h which is generated from
regcomp.sym by regcomp.pl. Currently the maximum possible number
of distinct opcodes is restricted to 256, with about 1/4 already
used.

There's a set of macros provided to make accessing the fields
easier and more consistent. These include C<OP()> which is used to tell
the type of a regnode-like structure, NEXT_OFF() which is the offset to
the next node (more on this later), ARG(), ARG1(), ARG2(), ARG_SET(),
and equivelents for reading and setting the arguments, STR_LEN(),
STRING(), and OPERAND() for manipulating strings and regop bearing
types.

=head3 What opcode is next?

There are three distinct concepts of "next" in the regex engine, and
it is important to keep them clear.

=over 4

=item *

There is the "next regnode" from a given regnode, a value which is
rarely useful except that sometimes it matches up in terms of value
with one of the others, and that sometimes the code assumes this to
always be so.

=item *

There is the "next opcode" from a given opcode/regnode. This is the
opcode physically located after the the current one, as determined by
the size of the current opcode. This is often useful, such as when
dumping the structure we use this order to traverse. Sometimes the code
assumes that the "next regnode" is the same as the "next opcode", or in
other words assumes that the sizeof a given opcode type is always going
to be 1 regnode large.

=item *

There is the "regnext" from a given opcode. This is the opcode which
is reached by jumping forward by the value of NEXT_OFF(),
or in a few cases for longer jumps by the arg1 field of the regnode_1
structure. The subroutine regnext() handles this transparently.
This is the logical successor of the node, which in some cases, like
that of the BRANCH opcode, has special meaning.

=back

=head1 PROCESS OVERVIEW

Broadly speaking performing a match of a string against a pattern
involves the following steps

    A. Compilation
        1. Parsing for size
        2. Parsing for construction
        3. Peep-hole Optimisation and Analysis
    B. Execution
        4. Start position and no-match optimisations
        5. Program execution

Where these steps occur in the actual execution of a perl program is
determined by whether the pattern involves interpolating any string
variables. If it does then compilation happens at run time. If it
doesn't then it happens at compile time. (The C</o> modifier changes this,
as does C<qr//> to a certain extent.) The engine doesn't really care that
much.

=head2 Compilation

This code exists primarily in regcomp.c, along with the header files
regcomp.h, regexp.h, regnodes.h.

Compilation starts with C<pregcomp()>, which is mostly an initialization
wrapper which farms out two other routines for the heavy lifting. The
first being C<reg()> which is the start point for parsing, and
C<study_chunk()> which is responsible for optimisation.

Initialization in C<pregcomp()> mostly involves the creation and data
filling of a special structure RExC_state_t, (defined in regcomp.c).
Almost all internally used routines in regcomp.h take a pointer to one
of these structures as their first argument, with the name *pRExC_state.
This structure is used to store the compilation state and contains many
fields. Likewise their are many macros defined which operate on this
variable. Anything that looks like RExC_xxxx is a macro that operates on
this pointer/structure.

=head3 Parsing for size

In this pass the input pattern is parsed in order to calculate how much
space is needed for each opcode we would need to emit. The size is also
used to determine whether long jumps will be required in the program.

This stage is controlled by the macro SIZE_ONLY being set.

The parse procedes pretty much exactly as it does during the
construction phase except that most routines are shortcircuited to
change the size field RExC_size and not actually do anything.

=head3 Parsing for construcution

Once the size of the program has been determine the pattern is parsed
again, but this time for real. Now SIZE_ONLY will be false, and the
actual construction can occur.

C<reg()> is the start of the parse process. It is responsible for
parsing an arbitrary chunk of pattern up to either the end of the
string, or the first closing parenthesis it encounters in the pattern.
This means it can be used to parse the toplevel regex, or any section
inside of a grouping parenthesis. It also handles the "special parens"
that perls regexes have. For instance when parsing C</x(?:foo)y/> C<reg()>
will at one point be called to parse from the '?' symbol up to and
including the ')'.

Additionally C<reg()> is responsible for parsing the one or more
branches from the pattern, and for "finishing them off" by correctly
setting their next pointers. In order to do the parsing it repeatedly
calls out to C<regbranch()> which is responsible for handling up to the
first C<|> symbol it sees.

C<regbranch()> in turn calls C<regpiece()> which is responsible for
handling "things" followed by a quantifier. In order to parse the
"things" C<regatom()> is called. This is the lowest level routine which
is responsible for parsing out constant strings, char classes, and the
various special symbols like C<$>. If C<regatom()> encounters a '('
character it in turn calls C<reg()>.

The routine C<regtail()> is called by both C<reg()>, C<regbranch()>
in order to "set the tail pointer" correctly. When executing and
we get to the end of a branch we need to go to node following the
grouping parens. When parsing however we don't know where the end will
be until we get there, so when we do we must go back and update the
offsets as appropriate. C<regtail> is used to make this easier.

A subtlety of the parse process means that a regex like C</foo/> is
originally parsed into an alternation with a single branch. It is only
afterwards that the optimizer converts single branch alternations into the
simpler form.

=head3 Parse Call Graph and a Grammar

The call graph looks like this:

    reg()                        # parse a top level regex, or inside of parens
        regbranch()              # parse a single branch of an alternation
            regpiece()           # parse a pattern followed by a quantifier
                regatom()        # parse a simple pattern
                    regclass()   #   used to handle a class
                    reg()        #   used to handle a parenthesized subpattern
                    ....
            ...
            regtail()            # finish off the branch
        ...
        regtail()                # finish off the branch sequence. Tie each
                                 # branches tail to the tail of the sequence
                                 # (NEW) In Debug mode this is
                                 # regtail_study().

A grammar form might be something like this:

    atom  : constant | class
    quant : '*' | '+' | '?' | '{min,max}'
    _branch: piece
           | piece _branch
           | nothing
    branch: _branch
          | _branch '|' branch
    group : '(' branch ')'
    _piece: atom | group
    piece : _piece
          | _piece quant

=head3 Debug Output

In bleadperl you can C<< use re Debug => 'PARSE'; >> to see some trace
information about the parse process. We will start with some simple
patterns and build up to more complex patterns.

So when we parse C</foo/> we see something like the following table. The
left shows whats being parsed, the number indicates where the next regop
would go. The stuff on the right is the trace output of the graph. The
names are chosen to be short to make it less dense on the screen. 'tsdy'
is a special form of C<regtail()> which does some extra analysis.

 >foo<             1            reg
                                  brnc
                                    piec
                                      atom
 ><                4              tsdy~ EXACT <foo> (EXACT) (1)
                                      ~ attach to END (3) offset to 2

The resulting program then looks like:

   1: EXACT <foo>(3)
   3: END(0)

As you can see, even though we parsed out a branch and a piece, it was ultimately
only an atom. The final program shows us how things work. We have an EXACT regop,
followed by an END regop. The number in parens indicates where the 'regnext' of
the node goes. The 'regnext' of an END regop is unused, as END regops mean
we have successfully matched. The number on the left indicates the position of
the regop in the regnode array.

Now lets try a harder pattern. We will add a quantifier so we have the pattern
C</foo+/>. We will see that C<regbranch()> calls C<regpiece()> regpiece twice.

 >foo+<            1            reg
                                  brnc
                                    piec
                                      atom
 >o+<              3                piec
                                      atom
 ><                6                tail~ EXACT <fo> (1)
                   7              tsdy~ EXACT <fo> (EXACT) (1)
                                      ~ PLUS (END) (3)
                                      ~ attach to END (6) offset to 3

And we end up with the program:

   1: EXACT <fo>(3)
   3: PLUS(6)
   4:   EXACT <o>(0)
   6: END(0)

Now we have a special case. The EXACT regop has a regnext of 0. This is
because if it matches it should try to match itself again. The PLUS regop
handles the actual failure of the EXACT regop and acts appropriately (going
to regnode 6 if the EXACT matched at least once, or failing if it didn't.)

Now for something much more complex: C</x(?:foo*|b[a][rR])(foo|bar)$/>

 >x(?:foo*|b...    1            reg
                                  brnc
                                    piec
                                      atom
 >(?:foo*|b[...    3                piec
                                      atom
 >?:foo*|b[a...                         reg
 >foo*|b[a][...                           brnc
                                            piec
                                              atom
 >o*|b[a][rR...    5                        piec
                                              atom
 >|b[a][rR])...    8                        tail~ EXACT <fo> (3)
 >b[a][rR])(...    9                      brnc
                  10                        piec
                                              atom
 >[a][rR])(f...   12                        piec
                                              atom
 >a][rR])(fo...                                 clas
 >[rR])(foo|...   14                        tail~ EXACT <b> (10)
                                            piec
                                              atom
 >rR])(foo|b...                                 clas
 >)(foo|bar)...   25                        tail~ EXACT <a> (12)
                                          tail~ BRANCH (3)
                  26                      tsdy~ BRANCH (END) (9)
                                              ~ attach to TAIL (25) offset to 16
                                          tsdy~ EXACT <fo> (EXACT) (4)
                                              ~ STAR (END) (6)
                                              ~ attach to TAIL (25) offset to 19
                                          tsdy~ EXACT <b> (EXACT) (10)
                                              ~ EXACT <a> (EXACT) (12)
                                              ~ ANYOF[Rr] (END) (14)
                                              ~ attach to TAIL (25) offset to 11
 >(foo|bar)$<                       tail~ EXACT <x> (1)
                                    piec
                                      atom
 >foo|bar)$<                            reg
                  28                      brnc
                                            piec
                                              atom
 >|bar)$<         31                      tail~ OPEN1 (26)
 >bar)$<                                  brnc
                  32                        piec
                                              atom
 >)$<             34                      tail~ BRANCH (28)
                  36                      tsdy~ BRANCH (END) (31)
                                              ~ attach to CLOSE1 (34) offset to 3
                                          tsdy~ EXACT <foo> (EXACT) (29)
                                              ~ attach to CLOSE1 (34) offset to 5
                                          tsdy~ EXACT <bar> (EXACT) (32)
                                              ~ attach to CLOSE1 (34) offset to 2
 >$<                                tail~ BRANCH (3)
                                        ~ BRANCH (9)
                                        ~ TAIL (25)
                                    piec
                                      atom
 ><               37                tail~ OPEN1 (26)
                                        ~ BRANCH (28)
                                        ~ BRANCH (31)
                                        ~ CLOSE1 (34)
                  38              tsdy~ EXACT <x> (EXACT) (1)
                                      ~ BRANCH (END) (3)
                                      ~ BRANCH (END) (9)
                                      ~ TAIL (END) (25)
                                      ~ OPEN1 (END) (26)
                                      ~ BRANCH (END) (28)
                                      ~ BRANCH (END) (31)
                                      ~ CLOSE1 (END) (34)
                                      ~ EOL (END) (36)
                                      ~ attach to END (37) offset to 1<div></div>

Resulting in the program

   1: EXACT <x>(3)
   3: BRANCH(9)
   4:   EXACT <fo>(6)
   6:   STAR(26)
   7:     EXACT <o>(0)
   9: BRANCH(25)
  10:   EXACT <ba>(14)
  12:   OPTIMIZED (2 nodes)
  14:   ANYOF[Rr](26)
  25: TAIL(26)
  26: OPEN1(28)
  28:   TRIE-EXACT(34)
        [StS:1 Wds:2 Cs:6 Uq:5 #Sts:7 Mn:3 Mx:3 Stcls:bf]
          <foo>
          <bar>
  30:   OPTIMIZED (4 nodes)
  34: CLOSE1(36)
  36: EOL(37)
  37: END(0)

Here we can see a much more complex program, with various optimisations in
play. At regnode 10 we can see an example where a char class with only
one character in it was turned into an EXACT node. We can also see where
an entire alternation was turned into a TRIE-EXACT node. As a consequence
some of the regnodes have been marked as optimised away. We can see that
the C<$> symbol has been converted into an EOL regop, a special piece of
code that looks for \n or the end of a string.

The next pointer for BRANCHes is interesting in that it points at where
execution should go if the branch fails. When executing if the engine
tries to traverse from a branch to a regnext that isnt a branch then
the engine will know the overall series of branches have failed.

=head3 Peep-hole Optimisation and Analysis

The regular expression engine can be a weighty tool to wield. On long
strings and complex patterns it can end up having to do a lot of work
to find a match, and even more to decide that no match is possible.
Consider a situation like the following pattern.

   'ababababababababababab' =~ /(a|b)*z/

The C<(a|b)*> part can match at every char in the string, and then fail
every time because there is no C<z> in the string. So obviously we can
not bother to use the regex engine unless there is a 'z' in the string.
Likewise in a pattern like:

   /foo(\w+)bar/

In this case we know that the string must contain a C<foo> which must be
followed by C<bar>. We can use Fast Boyer-More matching as implemented
in fbm_instr() to find the location of these strings. If they dont exist
then we dont need to resort to the much more expensive regex engine.
Even better if they do exist then we can use their positions to
reduce the search space that the regex engine needs to cover to determine
if the entire pattern does match.

There are various aspects of the pattern that can be used to facilitate
optimisations along these lines:

    * anchored fixed strings
    * floating fixed strings
    * minimum and maximum length requirements
    * start class
    * Beginning/End of line positions

Another form of optimisation that can occur is post-parse "peep-hole"
optimisations, where inefficient constructs are modified so that they
are more efficient. An example of this is TAIL regops which are used
during parsing to mark the end of branches and the end of groups. These
regops are used as place holders during construction and "always match"
so they can be "optimised away" by making the things that point to the
TAIL point to thing the TAIL points to, in essence "skipping" the node.

Another optimisation that can occur is that of "EXACT merging" which is
where two consecutive EXACT nodes are merged into a single more efficient
to execute regop. An even more agressive form of this is that a branch
sequence of the form EXACT BRANCH ... EXACT can be converted into a TRIE
regop.

All of this occurs in the routine study_chunk() which uses a special
structure scan_data_t to store the analysis that it has performed, and
as it goes does the "peep-hole" optimisations.

The code involved in study_chunk() is extremely cryptic. Be careful. :-)

=head2 Execution

Execution of a regex generally involves two phases, the first being
finding the start point in the string where we should match from,
and the second being running the regop interpreter.

If we can tell that there is no valid start point we don't bother running
interpreter at all. Likewise if we know from the analysis phase that we
can not optimise detection of the start position we go straight to the
interpreter.

The two entry points are re_intuit_start() and pregexec(). These routines
have a somewhat incestuous relationship with overlap between their functions,
and pregexec() may even call re_intuit_start() on its own. Nevertheless
the perl source code may call into either, or both.

Execution of the interpreter itself used to be recursive. Due to the
efforts of Dave Mitchel in blead perl it no longer is. Instead an
internal stack is maintained on the heap and the routine is fully
iterative. This can make it tricky as the code is quite conservative
about what state it stores which means that two consecutive lines in the
code can actually be running in totally different contexts due to the
simulated recursion.

=head3 Start position and no-match optimisations

re_intuit_start() is responsible for handling start point and no match
optimisations as determined by the results of the analysis done by
study_chunk() (and described in L<Peep-hole Optimisation and Analysis>).

The basic structure of this routine is to try to find the start and/or
end points of where the pattern could match, and to ensure that the string
is long enough to match the pattern. It tries to use more efficent
methods over less efficient methods and may involve considerable cross
checking of constraints to find the place in the string that matches.
For instance it may try to determine that a given fixed string must be
not only present but a certain number of chars before the end of the
string, or whatever.

It calls out into several other routines, like fbm_instr() which does
"Fast Boyer More" matching and find_byclass() which is responsible for
finding the start using the first mandatory regop in the program.

When the optimisation criteria have been satisfied reg_try() is called
to perform the match.

=head3 Program execution

C<pregexec()> is the main entry point for running a regex. It contains
support for initializing the regex interpreters state, running
re_intuit_start() if needed, and running the intepreter on the string
from various start positions as needed. When its necessary to use
the regex interpreter C<pregexec()> calls C<regtry()>.

C<regtry()> is the entry point into the regex interpreter. It expects
as arguments a pointer to a regmatch_info structure and a pointer to
a string.  It returns an integer 1 for success and a 0 for failure.
It is basically a setup wrapper around C<regmatch()>.

C<regmatch> is the main "recursive loop" of the interpreter. It is
basically a giant switch statement that executes the regops based on
their type. A few of the regops are implemented as subroutines but
the bulk are inline code.

=head1 MISCELLANEOUS

=head2 UNICODE and Localization Support

No matter how you look at it unicode support is going to be a pain in a
regex engine. Tricks that might be fine when you have 256 possible
characters often won't scale to handle the size of the 'utf8' character
set.  Things you can take for granted with ASCII may not be true with
unicode. For instance in ASCII its safe to assume that
C<sizeof(char1) == sizeof(char2)>, in utf8 it isn't. Unicode case folding is
vastly more complex than the simple rules of English, and even when not
using unicode but only localized single byte encodings things can get
tricky (technically GERMAN-SHARP-ESS should match 'ss' in localized case
insensitive matching.)

Making things worse is that C<utf8> support was a later addition to the
regex engine (as it was to perl) and necessarily this made things a lot
more complicated. Obviously its easier to design a regex engine with
unicode support from the beginning than it is to retrofit one that
wasn't designed with it in mind.

Pretty well every regop that involves looking at the input string has
two cases, one for 'utf8' and one not. In fact its often more complex
than that, as the pattern may be 'utf8' as well.

Care must be taken when making changes to make sure that you handle
utf8 properly both at compile time and at execution time, including
when the string and pattern are mismatched.

The following comment in regcomp.h gives an example of exactly how
tricky this can be:

    Two problematic code points in Unicode casefolding of EXACT nodes:

    U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
    U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS

    which casefold to

    Unicode                      UTF-8

    U+03B9 U+0308 U+0301         0xCE 0xB9 0xCC 0x88 0xCC 0x81
    U+03C5 U+0308 U+0301         0xCF 0x85 0xCC 0x88 0xCC 0x81

    This means that in case-insensitive matching (or "loose matching",
    as Unicode calls it), an EXACTF of length six (the UTF-8 encoded
    byte length of the above casefolded versions) can match a target
    string of length two (the byte length of UTF-8 encoded U+0390 or
    U+03B0). This would rather mess up the minimum length computation.

    What we'll do is to look for the tail four bytes, and then peek
    at the preceding two bytes to see whether we need to decrease
    the minimum length by four (six minus two).

    Thanks to the design of UTF-8, there cannot be false matches:
    A sequence of valid UTF-8 bytes cannot be a subsequence of
    another valid sequence of UTF-8 bytes.

=head1 AUTHOR

by Yves Orton, 2006.

With excerpts from Perl, and contributions and suggestions from
Ronald J. Kimball, Dave Mitchell, Dominic Dunlop, Mark Jason Dominus,
and Stephen McCamant.

=head1 LICENSE

Same terms as Perl.

=head1 REFERENCES

[1] http://perl.plover.com/Rx/paper/

=cut