The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.

NAME

Synopsis_05 - Regexes and Rules

AUTHORS

Damian Conway <damian@conway.org> and Allison Randal <al@shadowed.net>

VERSION

   Maintainer: Patrick Michaud <pmichaud@pobox.com> and
               Larry Wall <larry@wall.org>
   Date: 24 Jun 2002
   Last Modified: 20 Feb 2008
   Number: 5
   Version: 73

This document summarizes Apocalypse 5, which is about the new regex syntax. We now try to call them regex rather than "regular expressions" because they haven't been regular expressions for a long time, and we think the popular term "regex" is in the process of becoming a technical term with a precise meaning of: "something you do pattern matching with, kinda like a regular expression". On the other hand, one of the purposes of the redesign is to make portions of our patterns more amenable to analysis under traditional regular expression and parser semantics, and that involves making careful distinctions between which parts of our patterns and grammars are to be treated as declarative, and which parts as procedural.

In any case, when referring to recursive patterns within a grammar, the terms rule and token are generally preferred over regex.

New match result and capture variables

The underlying match result object is now available as the $/ variable, which is implicitly lexically scoped. All user access to the most recent match is through this variable, even when it doesn't look like it. The individual capture variables (such as $0, $1, etc.) are just elements of $/.

By the way, unlike in Perl�5, the numbered capture variables now start at $0 instead of $1. See below.

Unchanged syntactic features

The following regex features use the same syntax as in Perl�5:

  • Capturing: (...)

  • Repetition quantifiers: *, +, and ?

  • Alternatives: |

  • Backslash escape: \

  • Minimal matching suffix: ??, *?, +?

While the syntax of | does not change, the default semantics do change slightly. We are attempting to concoct a pleasing mixture of declarative and procedural matching so that we can have the best of both. In short, you need not write your own tokener for a grammar because Perl will write one for you. See the section below on "Longest-token matching".

Simplified lexical parsing of patterns

Unlike traditional regular expressions, Perl�6 does not require you to memorize an arbitrary list of metacharacters. Instead it classifies characters by a simple rule. All glyphs (graphemes) whose base characters are either the underscore (_) or have a Unicode classification beginning with 'L' (i.e. letters) or 'N' (i.e. numbers) are always literal (i.e. self-matching) in regexes. They must be escaped with a \ to make them metasyntactic (in which case that single alphanumeric character is itself metasyntactic, but any immediately following alphanumeric character is not).

All other glyphs--including whitespace--are exactly the opposite: they are always considered metasyntactic (i.e. non-self-matching) and must be escaped or quoted to make them literal. As is traditional, they may be individually escaped with \, but in Perl�6 they may be also quoted as follows.

Sequences of one or more glyphs of either type (i.e. any glyphs at all) may be made literal by placing them inside single quotes. (Double quotes are also allowed, with the same interpolative semantics as the current language in which the regex is lexically embedded.) Quotes create a quantifiable atom, so while

    moose*

quantifies only the 'e' and matches "mooseee", saying

    'moose'*

quantifies the whole string and would match "moosemoose".

Here is a table that summarizes the distinctions:

                 Alphanumerics        Non-alphanumerics         Mixed

 Literal glyphs   a    1    _        \*  \$  \.   \\   \'       K\-9\!
 Metasyntax      \a   \1   \_         *   $   .    \    '      \K-\9!
 Quoted glyphs   'a'  '1'  '_'       '*' '$' '.' '\\' '\''     'K-9!'

In other words, identifier glyphs are literal (or metasyntactic when escaped), non-identifier glyphs are metasyntactic (or literal when escaped), and single quotes make everything inside them literal.

Note, however, that not all non-identifier glyphs are currently meaningful as metasyntax in Perl�6 regexes (e.g. \1 \_ - !). It is more accurate to say that all unescaped non-identifier glyphs are potential metasyntax, and reserved for future use. If you use such a sequence, a helpful compile-time error is issued indicating that you either need to quote the sequence or define a new operator to recognize it.

Modifiers

  • The extended syntax (/x) is no longer required...it's the default. (In fact, it's pretty much mandatory--the only way to get back to the old syntax is with the :Perl5/:P5 modifier.)

  • There are no /s or /m modifiers (changes to the meta-characters replace them - see below).

  • There is no /e evaluation modifier on substitutions; instead use:

         s/pattern/{ doit() }/

    or:

         s[pattern] = doit()

    Instead of /ee say:

         s/pattern/{ eval doit() }/

    or:

         s[pattern] = eval doit()
  • Modifiers are now placed as adverbs at the start of a match/substitution:

         m:g:i/\s* (\w*) \s* ,?/;

    Every modifier must start with its own colon. The delimiter must be separated from the final modifier by whitespace if it would otherwise be taken as an argument to the preceding modifier (which is true if and only if the next character is a left parenthesis.)

  • The single-character modifiers also have longer versions:

             :i        :ignorecase
             :b        :basechar
             :g        :global
  • The :i (or :ignorecase) modifier causes case distinctions to be ignored in its lexical scope, but not in its dynamic scope. That is, subrules always use their own case settings.

    The :ii variant may be used on a substitution to change the substituted string to the same case pattern as the matched string.

    If the pattern is matched without the :sigspace modifier, case info is carried across on a character by character basis. If the right string is longer than the left one, the case of the final character is replicated. Titlecase is carried across if possible regardless of whether the resulting letter is at the beginning of a word or not; if there is no titlecase character available, the corresponding uppercase character is used. (This policy can be modified within a lexical scope by a language-dependent Unicode declaration to substitute titlecase according to the orthographic rules of the specified language.) Characters that carry no case information leave their corresponding replacement character unchanged.

    If the pattern is matched with :sigspace, then a slightly smarter algorithm is used which attempts to determine if there is a uniform capitalization policy over each matched word, and applies the same policy to each replacement word. If there doesn't seem to be a uniform policy on the left, the policy for each word is carried over word by word, with the last pattern word replicated if necessary. If a word does not appear to have a recognizable policy, the replacement word is translated character for character as in the non-sigspace case. Recognized policies include:

        lc()
        uc()
        ucfirst(lc())
        lcfirst(uc())
        capitalize()

    In any case, only the officially matched string part of the pattern match counts, so any sort of lookahead or contextual matching is not included in the analysis.

  • The :b (or :basechar) modifier scopes exactly like :ignorecase except that it ignores accents instead of case. It is equivalent to taking each grapheme (in both target and pattern), converting both to NFD (maximally decomposed) and then comparing the two base characters (Unicode non-mark characters) while ignoring any trailing mark characters. The mark characters are ignored only for the purpose of determining the truth of the assertion; the actual text matched includes all ignored characters, including any that follow the final base character.

    The :bb variant may be used on a substitution to change the substituted string to the same accent pattern as the matched string. Accent info is carried across on a character by character basis. If the right string is longer than the left one, the remaining characters are substituted without any modification. (Note that NFD/NFC distinctions are usually immaterial, since Perl encapsulates that in grapheme mode.) Under :sigspace the preceding rules are applied word by word.

  • The :c (or :continue) modifier causes the pattern to continue scanning from the specified position (defaulting to $/.to):

         m:c($p)/ pattern /     # start scanning at position $p

    Note that this does not automatically anchor the pattern to the starting location. (Use :p for that.) The pattern you supply to split has an implicit :c modifier.

    String positions are of type StrPos and should generally be treated as opaque.

  • The :p (or :pos) modifier causes the pattern to try to match only at the specified string position:

         m:pos($p)/ pattern /  # match at position $p

    If the argument is omitted, it defaults to $/.to. (Unlike in Perl 5, the string itself has no clue where its last match ended.) All subrule matches are implicitly passed their starting position. Likewise, the pattern you supply to a Perl macro's is parsed trait has an implicit :p modifier.

    Note that

         m:c($p)/pattern/

    is roughly equivalent to

         m:p($p)/.*? <( pattern )> /
  • The new :s (:sigspace) modifier causes whitespace sequences to be considered "significant"; they are replaced by a whitespace matching rule, <.ws>. That is,

         m:s/ next cmd '='   <condition>/

    is the same as:

         m/ <.ws> next <.ws> cmd <.ws> '=' <.ws> <condition>/

    which is effectively the same as:

         m/ \s* next \s+ cmd \s* '=' \s* <condition>/

    But in the case of

         m:s{(a|\*) (b|\+)}

    or equivalently,

         m { (a|\*) <.ws> (b|\+) }

    <.ws> can't decide what to do until it sees the data. It still does the right thing. If not, define your own ws and :sigspace will use that.

    In general you don't need to use :sigspace within grammars because the parser rules automatically handle whitespace policy for you. In this context, whitespace often includes comments, depending on how the grammar chooses to define its whitespace rule. Although the default <.ws> subrule recognizes no comment construct, any grammar is free to override the rule. The <.ws> rule is not intended to mean the same thing everywhere.

    It's also possible to pass an argument to :sigspace specifying a completely different subrule to apply. This can be any rule, it doesn't have to match whitespace. When discussing this modifier, it is important to distinguish the significant whitespace in the pattern from the "whitespace" being matched, so we'll call the pattern's whitespace sigspace, and generally reserve whitespace to indicate whatever <.ws> matches in the current grammar. The correspondence between sigspace and whitespace is primarily metaphorical, which is why the correspondence is both useful and (potentially) confusing.

    The :s modifier is considered sufficiently important that match variants are defined for them:

        mm/match some words/                        # same as m:sigspace
        ss/match some words/replace those words/    # same as s:sigspace
  • New modifiers specify Unicode level:

         m:bytes  / .**2 /       # match two bytes
         m:codes  / .**2 /       # match two codepoints
         m:graphs / .**2 /       # match two language-independent graphemes
         m:chars  / .**2 /       # match two characters at current max level

    There are corresponding pragmas to default to these levels. Note that the :chars modifier is always redundant because dot always matches characters at the highest level allowed in scope. This highest level may be identical to one of the other three levels, or it may be more specific than :graphs when a particular language's character rules are in use. Note that you may not specify language-dependent character processing without specifying which language you're depending on. [Conjecture: the :chars modifier could take an argument specifying which language's rules to use for this match.]

  • The new :Perl5/:P5 modifier allows Perl�5 regex syntax to be used instead. (It does not go so far as to allow you to put your modifiers at the end.) For instance,

         m:P5/(?mi)^(?:[a-z]|\d){1,2}(?=\s)/

    is equivalant to the Perl�6 syntax:

        m/ :i ^^ [ <[a..z]> || \d ] ** 1..2 <before \s> /
  • Any integer modifier specifies a count. What kind of count is determined by the character that follows.

  • If followed by an x, it means repetition. Use :x(4) for the general form. So

         s:4x [ (<.ident>) '=' (\N+) $$] = "$0 => $1";

    is the same as:

         s:x(4) [ (<.ident>) '=' (\N+) $$] = "$0 => $1";

    which is almost the same as:

         s:c[ (<.ident>) '=' (\N+) $$] = "$0 => $1" for 1..4;

    except that the string is unchanged unless all four matches are found. However, ranges are allowed, so you can say :x(1..4) to change anywhere from one to four matches.

  • If the number is followed by an st, nd, rd, or th, it means find the Nth occurrence. Use :nth(3) for the general form. So

         s:3rd/(\d+)/@data[$0]/;

    is the same as

         s:nth(3)/(\d+)/@data[$0]/;

    which is the same as:

         m/(\d+)/ && m:c/(\d+)/ && s:c/(\d+)/@data[$0]/;

    Lists and junctions are allowed: :nth(1|2|3|5|8|13|21|34|55|89).

    So are closures: :nth({.is_fibonacci})

  • With the new :ov (:overlap) modifier, the current regex will match at all possible character positions (including overlapping) and return all matches in a list context, or a disjunction of matches in a scalar context. The first match at any position is returned. The matches are guaranteed to be returned in left-to-right order with respect to the starting positions.

         $str = "abracadabra";
    
         if $str ~~ m:overlap/ a (.*) a / {
             @substrings = @@();    # bracadabr cadabr dabr br
         }
  • With the new :ex (:exhaustive) modifier, the current regex will match every possible way (including overlapping) and return all matches in a list context, or a disjunction of matches in a scalar context. The matches are guaranteed to be returned in left-to-right order with respect to the starting positions. The order within each starting position is not guaranteed and may depend on the nature of both the pattern and the matching engine. (Conjecture: or we could enforce backtracking engine semantics. Or we could guarantee no order at all unless the pattern starts with "::" or some such to suppress DFAish solutions.)

         $str = "abracadabra";
    
         if $str ~~ m:exhaustive/ a (.*?) a / {
             say "@()";    # br brac bracad bracadabr c cad cadabr d dabr br
         }

    Note that the ~~ above can return as soon as the first match is found, and the rest of the matches may be performed lazily by @().

  • The new :rw modifier causes this regex to claim the current string for modification rather than assuming copy-on-write semantics. All the captures in $/ become lvalues into the string, such that if you modify, say, $1, the original string is modified in that location, and the positions of all the other fields modified accordingly (whatever that means). In the absence of this modifier (especially if it isn't implemented yet, or is never implemented), all pieces of $/ are considered copy-on-write, if not read-only.

    [Conjecture: this should really associate a pattern with a string variable, not a (presumably immutable) string value.]

  • The new :keepall modifier causes this regex and all invoked subrules to remember everything, even if the rules themselves don't ask for their subrules to be remembered. This is for forcing a grammar that throws away whitespace and comments to keep them instead.

  • The new :ratchet modifier causes this regex to not backtrack by default. (Generally you do not use this modifier directly, since it's implied by token and rule declarations.) The effect of this modifier is to imply a : after every construct that could backtrack, including bare *, +, and ? quantifiers, as well as alternations. (Note: for portions of patterns subject to longest-token analysis, a : is ignored in any case, since there will be no backtracking necessary.)

  • The new :panic modifier causes this regex and all invoked subrules to try to backtrack on any rules that would otherwise default to not backtracking because they have :ratchet set. Never panic unless you're desperate and want the pattern matcher to do a lot of unnecessary work. If you have an error in your grammar, it's almost certainly a bad idea to fix it by backtracking.

  • The :i, :s, :Perl5, and Unicode-level modifiers can be placed inside the regex (and are lexically scoped):

         m/:s alignment '=' [:i left|right|cent[er|re]] /

    As with modifiers outside, only parentheses are recognized as valid brackets for args to the adverb. In particular:

        m/:foo[xxx]/        Parses as :foo [xxx]
        m/:foo{xxx}/        Parses as :foo {xxx}
        m/:foo<xxx>/        Parses as :foo <xxx>
  • User-defined modifiers will be possible:

             m:fuzzy/pattern/;
  • User-defined modifiers can also take arguments, but only in parentheses:

             m:fuzzy('bare')/pattern/;
  • To use parens for your delimiters you have to separate:

             m:fuzzy (pattern);

    or you'll end up with:

             m:fuzzy(fuzzyargs); pattern ;

Changed metacharacters

  • A dot . now matches any character including newline. (The /s modifier is gone.)

  • ^ and $ now always match the start/end of a string, like the old \A and \z. (The /m modifier is gone.) On the right side of an embedded ~~ or !~~ operator they always match the start/end of the indicated submatch because that submatch is logically being treated as a separate string.

  • A $ no longer matches an optional preceding \n so it's necessary to say \n?$ if that's what you mean.

  • \n now matches a logical (platform independent) newline not just \x0a.

  • The \A, \Z, and \z metacharacters are gone.

New metacharacters

  • Because /x is default:

    • An unescaped # now always introduces a comment. If followed by an opening bracket character (and if not in the first column), it introduces an embedded comment that terminates with the closing bracket. Otherwise the comment terminates at the newline.

    • Whitespace is now always metasyntactic, i.e. used only for layout and not matched literally (but see the :sigspace modifier described above).

  • ^^ and $$ match line beginnings and endings. (The /m modifier is gone.) They are both zero-width assertions. $$ matches before any \n (logical newline), and also at the end of the string if the final character was not a \n. ^^ always matches the beginning of the string and after any \n that is not the final character in the string.

  • . matches an anything, while \N matches an anything except newline. (The /s modifier is gone.) In particular, \N matches neither carriage return nor line feed.

  • The new & metacharacter separates conjunctive terms. The patterns on either side must match with the same beginning and end point. Note: if you don't want your two terms to end at the same point, then you really want to use a lookahead instead.

    As with the disjunctions | and ||, conjuctions come in both & and && forms. The & form is considered declarative rather than procedural; it allows the compiler and/or the run-time system to decide which parts to evaluate first, and it is erroneous to assume either order happens consistently. The && form guarantees left-to-right order, and backtracking makes the right argument vary faster than the left. In other words, && and || establish sequence points. The left side may be backtracked into when backtracking is allowed into the construct as a whole.

    The & operator is list associative like |, but has slightly tighter precedence. Likewise && has slightly tighter precedence than ||. As with the normal junctional and short-circuit operators, & and | are both tighter than && and ||.

  • The ~~ and !~~ operators cause a submatch to be performed on whatever was matched by the variable or atom on the left. String anchors consider that submatch to be the entire string. So, for instance, you can ask to match any identifier that does not contain the word "moose":

        <ident> !~~ 'moose'

    In contrast

        <ident> !~~ ^ 'moose' $

    would allow any identifier (including any identifier containing "moose" as a substring) as long as the identifier as a whole is not equal to "moose". (Note the anchors, which attach the submatch to the beginning and end of the identifier as if that were the entire match.) When used as part of a longer match, for clarity it might be good to use extra brackets:

        [ <ident> !~~ ^ 'moose' $ ]

    The precedence of ~~ and !~~ fits in between the junctional and sequential versions of the logical operators just as it does in normal Perl expressions (see S03). Hence

        <ident> !~~ 'moose' | 'squirrel'

    parses as

        <ident> !~~ [ 'moose' | 'squirrel' ]

    while

        <ident> !~~ 'moose' || 'squirrel'

    parses as

        [ <ident> !~~ 'moose' ] || 'squirrel'

Bracket rationalization

  • (...) still delimits a capturing group. However the ordering of these groups is hierarchical rather than linear. See "Nested subpattern captures".

  • [...] is no longer a character class. It now delimits a non-capturing group.

  • {...} is no longer a repetition quantifier. It now delimits an embedded closure. It is always considered procedural rather than declarative; it establishes a sequence point between what comes before and what comes after. (To avoid this use the <?{...}> assertion syntax instead.)

  • You can call Perl code as part of a regex match by using a closure. Embedded code does not usually affect the match--it is only used for side-effects:

         / (\S+) { print "string not blank\n"; $text = $0; }
            \s+  { print "but does contain whitespace\n" }
         /

    An explicit reduction using the make function sets the result object for this match:

            / (\d) { make $0.sqrt } Remainder /;

    This has the effect of capturing the square root of the numified string, instead of the string. The Remainder part is matched but is not returned unless the first make is later overridden by another make.

    These closures are invoked with a topic ($_) of the current match state (a Cursor object). Within a closure, the instantaneous position within the search is denoted by the .pos method on that object. As with all string positions, you must not treat it as a number unless you are very careful about which units you are dealing with.

    The Cursor object can also return the original item that we are matching against; this is available from the ._ method, named to remind you that it probably came from the user's $_ variable. (But that may well be off in some other scope when indirect rules are called, so we mustn't rely on the user's lexical scope.)

    The closure is also guaranteed to start with a $/ Match object representing the match so far. However, if the closure does its own internal matching, its $/ variable will be rebound to the result of that match until the end of the embedded closure.

  • It can affect the match if it calls fail:

         / (\d+) { $0 < 256 or fail } /

    Since closures establish a sequence point, they are guaranteed to be called at the canonical time even if the optimizer could prove that something after them can't match. (Anything before is fair game, however. In particular, a closure often serves as the terminator of a longest-token pattern.)

  • The general repetition specifier is now ** for maximal matching, with a corresponding **? for minimal matching. (All such quantifier modifiers now go directly after the **.) Space is allowed on either side of the complete quantifier. This space is considered significant under :sigspace, and will be distributed as a call to <.ws> between all the elements of the match but not on either end.

    The next token will determine what kind of repetition is desired:

    If the next thing is an integer, then it is parsed as either as an exact count or a range:

        . ** 42                  # match exactly 42 times
        <item> ** 3..*           # match 3 or more times

    This form is considered declarational.

    If you supply a closure, it should return either an Int or a Range object.

        'x' ** {$m}              # exact count returned from closure
        <foo> ** {$m..$n}        # range returned from closure
    
        / value was (\d **? {1..6}) with ([ <alpha>\w* ]**{$m..$n}) /

    It is illegal to return a list, so this easy mistake fails:

        / [foo] ** {1,3} /

    The closure form is always considered procedural, so the item it is modifying is never considered part of the longest token.

    If you supply any other atom (which may be quantified), it is interpreted as a separator (such as an infix operator), and the initial item is quantified by the number of times the separator is seen between items:

        <alt> ** '|'            # repetition controlled by presence of character
        <addend> ** <addop>     # repetition controlled by presence of subrule
        <item> ** [ \!?'==' ]   # repetition controlled by presence of operator
        <file>**\h+             # repetition controlled by presence of whitespace

    A successful match of such a quantifier always ends "in the middle", that is, after the initial item but before the next separator. Therefore

        / <ident> ** ',' /

    can match

        foo
        foo,bar
        foo,bar,baz

    but never

        foo,
        foo,bar,

    It is legal for the separator to be zero-width as long as the pattern on the left progresses on each iteration:

        . ** <?same>   # match sequence of identical characters

    The separator never matches independently of the next item; if the separator matches but the next item fails, it backtracks all the way back through the separator. Likewise, this matching of the separator does not count as "progress" under :ratchet semantics unless the next item succeeds.

    When significant space is used under :sigspace with the separator form, it applies on both sides of the separator, so

        mm/<element> ** ','/
        mm/<element>** ','/
        mm/<element> **','/

    all allow whitespace around the separator like this:

        / <element>[<.ws>','<.ws><element>]* /

    while

        mm/<element>**','/

    excludes all significant whitespace:

        / <element>[','<element>]* /

    Of course, you can always match whitespace explicitly if necessary, so to allow whitespace after the comma but not before, you can say:

        / <element>**[','\s*] /
  • <...> are now extensible metasyntax delimiters or assertions (i.e. they replace Perl�5's crufty (?...) syntax).

Variable (non-)interpolation

  • In Perl�6 regexes, variables don't interpolate.

  • Instead they're passed raw to the regex engine, which can then decide how to handle them (more on that below).

  • The default way in which the engine handles a scalar is to match it as a '...' literal (i.e. it does not treat the interpolated string as a subpattern). In other words, a Perl�6:

         / $var /

    is like a Perl�5:

         / \Q$var\E /

    However, if $var contains a Regex object, instead of attempting to convert it to a string, it is called as a subrule, as if you said <$var>. (See assertions below.) This form does not capture, and it fails if $var is tainted.

    However, a variable used as the left side of an alias or submatch operator is not used for matching.

        $x = <ident>
        $0 ~~ <ident>

    If you do want to match $0 again and then use that as the submatch, you can force the match using double quotes:

        "$0" ~~ <ident>

    On the other hand, it is non-sensical to alias to something that is not a variable:

        "$0" = <ident>     # ERROR
        $0 = <ident>       # okay
        $x = <ident>       # okay, temporary capture
        $<x> = <ident>     # okay, persistent capture
        <x=ident>          # same thing

    Variables declared in capture aliases are lexically scoped to the rest of the regex. You should not confuse this use of = with either ordinary assignment or ordinary binding. You should read the = more like the pseudoassignment of a declarator than like normal assignment. It's more like the ordinary := operator, since at the level regexes work, strings are immutable, so captures are really just precomputed substr values. Nevertheless, when you eventually use the values independently, the substr may be copied, and then it's more like it was an assignment originally.

    Capture variables of the form $<ident> may persist beyond the lexical scope; if the match succeeds they are remembered in the Match object's hash, with a key corresponding to the variable name's identifier. Likewise bound numeric variables persist as $0, etc.

    The capture performed by = creates a new lexical variable if it does not already exist in the current lexical scope. To capture to an outer lexical variable you must supply an OUTER:: as part of the name, or perform the assignment from within a closure.

        $x = [...]                       # capture to our own lexical $x
        $OUTER::x = [...]                # capture to existing lexical $x
        [...] -> $tmp { let $x = $tmp }  # capture to existing lexical $x

    Note however that let (and temp) are not guaranteed to be thread safe on shared variables, so don't do that.

  • An interpolated array:

         / @cmds /

    is matched as if it were an alternation of its elements. Ordinarily it matches using junctive semantics:

         / [ @cmds[0] | @cmds[1] | @cmds[2] | ... ] /

    However, if it is a direct member of a || list, it uses sequential matching semantics, even it's the only member of the list. Conveniently, you can put || before the first member of an alternation, hence

         / || @cmds /

    is equivalent to

         / [ @cmds[0] || @cmds[1] || @cmds[2] || ... ] /

    Or course, you can also

         / | @cmds /

    to be clear that you mean junctive semantics.

    As with a scalar variable, each element is matched as a literal unless it happens to be a Regex object, in which case it is matched as a subrule. As with scalar subrules, a tainted subrule always fails. All string values pay attention to the current :ignorecase and :basechar settings, while Regex values use their own :ignorecase and :basechar settings.

    When you get tired of writing:

        token sigil { '$' | '@' | '@@' | '%' | '&' | '::' }

    you can write:

        token sigil { < $ @ @@ % & :: > }

    as long as you're careful to put a space after the initial angle so that it won't be interpreted as a subrule. With the space it is parsed like angle quotes in ordinary Perl�6 and treated as a literal array value.

  • Alternatively, if you predeclare a proto regex, you can write multiple regexes for the same category, differentiated only by the symbol they match. The symbol is specified as part of the "long name". It may also be matched within the rule using <sym>, like this:

        proto token sigil { }
        multi token sigil:sym<$>  { <sym> }
        multi token sigil:sym<@>  { <sym> }
        multi token sigil:sym<@@> { <sym> }
        multi token sigil:sym<%>  { <sym> }
        multi token sigil:sym<&>  { <sym> }
        multi token sigil:sym<::> { <sym> }

    (The multi is optional and generally omitted with a grammar.)

    This can be viewed as a form of multiple dispatch, except that it's based on longest-token matching rather than signature matching. The advantage of writing it this way is that it's easy to add additional rules to the same category in a derived grammar. All of them will be matched in parallel when you try to match /<sigil>/.

    If there are formal parameters on multi regex methods, matching still proceeds via longest-token rules first. If that results in a tie, a normal multiple dispatch is made using the arguments to the remaining variants, assuming they can be differentiated by type.

  • An interpolated hash provides a way of inserting various forms of run-time table-driven submatching into a regex. An interpolated hash matches the longest possible token (typically the longest combination of key and value). The match fails if no entry matches. (A "" key will match anywhere, provided no other entry takes precedence by the longest token rule.)

    In a context requiring a set of initial token patterns, the initial token patterns are taken to be each key plus any initial token pattern matched by the corresponding value (if the value is a string or regex). The token patterns are considered to be canonicalized in the same way as any surrounding context, so for instance within a case-insensitive context the hash keys must match insensitively also.

    Subsequent matching depends on the hash value:

    • If the corresponding value of the hash element is a closure, it is executed.

    • If the value is a string, it is matched literally, starting after where the key left off matching. As a natural consequence, if the value is "", nothing special happens except that the key match succeeds.

    • If it is a Regex object, it is executed as a subrule, with an initial position after the matched key. (This is further described below under the <%hash> notation.) As with scalar subrules, a tainted subrule always fails, and no capture is attempted.

    • If the value is a number, this entry represents a "false match". The match position is set back to before the current false match, and the key is rematched using the same hash, but this time ignoring any keys longer than the number. (This is measured in the default Unicode level in effect where the hash was declared, usually graphemes. If the current Unicode level is lower, the results are as if the string to be matched had been upconverted to the hash's Unicode level. If the current Unicode level is higher, the results are undefined if the string contains any characters whose interpretation would be changed by the higher Unicode level, such as language-dependent ligatures.)

    • Any other value causes the match to fail.

    All hash keys, and values that are strings, pay attention to the :ignorecase and :basechar settings. (Subrules maintain their own case settings.)

    You may combine multiple hashes under the same longest-token consideration by using declarative alternation:

        %statement | %prefix | %term

    This means that, despite being in a later hash, %term<food> will be selected in preference to %prefix<foo> because it's the longer token. However, if there is a tie, the earlier hash wins, so %statement<if> hides any %prefix<if> or %term<if>.

    In contrast, if you use a procedural alternation:

        [ %prefix || %term ]

    a %prefix<foo> would be selected in preference to a %term<food>. (Which is not what you usually want if your language is to do longest-token consistently.)

  • Variable matches are considered provisionally declarative, on the assumption that the contents of the variable will not change frequently. If it does change, it may force recalculation of any analysis relying on its supposed declarative nature. (If you know this is going to happen too often, put some kind of sequence point before the variable to disable static analysis such as the generation of longest-token automata.)

Extensible metasyntax (<...>)

Both < and > are metacharacters, and are usually (but not always) used in matched pairs. (Some combinations of metacharacters function as standalone tokens, and these may include angles. These are described below.) Most assertions are considered declarative; procedural assertions will be marked as exceptions.

For matched pairs, the first character after < determines the nature of the assertion:

  • If the first character is whitespace, the angles are treated as an ordinary "quote words" array literal.

        < adam & eve >   # equivalent to [ 'adam' | '&' | 'eve' ]
  • A leading alphabetic character means it's a capturing grammatical assertion (i.e. a subrule or a named character class - see below):

         / <sign>? <mantissa> <exponent>? /

    The first character after the identifier determines the treatment of the rest of the text before the closing angle. The underlying semantics is that of a function or method call, so if the first character is a left parenthesis, it really is a call:

        <foo('bar')>

    If the first character after the identifier is an =, then the identifier is taken as an alias for what follows. In particular,

        <foo=bar>

    is just shorthand for

        $<foo> = <bar>

    If the first character after the identifier is whitespace, the subsequent text (following any whitespace) is passed as a regex, so:

        <foo bar>

    is more or less equivalent to

        <foo(/bar/)>

    To pass a regex with leading whitespace you must use the parenthesized form.

    If the first character is a colon, the rest of the text (following any whitespace) is passed as a string, so the previous may also be written as:

        <foo: bar>

    To pass a string with leading whitespace, or to interpolate any values into the string, you must use the parenthesized form.

    No other characters are allowed after the initial identifier.

    Subrule matches are considered declarative to the extent that the front of the subrule is itself considered declarative. If a subrule contains a sequence point, then so does the subrule match. Longest-token matching does not proceed past such a subrule, for instance.

  • The special named assertions include:

         / <?before pattern> /    # lookahead
         / <?after pattern> /     # lookbehind
    
         / <?same> /              # true between two identical characters
    
         / <.ws> /                # match "whitespace":
                                  #   \s+ if it's between two \w characters,
                                  #   \s* otherwise
    
         / <?at($pos)> /          # match only at a particular StrPos
                                  # short for <?{ .pos === $pos }>
                                  # (considered declarative until $pos changes)

    The after assertion implements lookbehind by reversing the syntax tree and looking for things in the opposite order going to the left. It is illegal to do lookbehind on a pattern that cannot be reversed.

    Note: the effect of a forward-scanning lookbehind at the top level can be achieved with:

        / .*? prestuff <( mainpat )> /
  • A leading . causes a named assertion not to capture what it matches (see "Subrule captures". For example:

         / <ident>  <ws>  /      # $/<ident> and $/<ws> both captured
         / <.ident> <ws>  /      # only $/<ws> captured
         / <.ident> <.ws> /      # nothing captured

    The non-capturing behavior may be overridden with a :keepall.

  • A leading $ indicates an indirect subrule. The variable must contain either a Regex object, or a string to be compiled as the regex. The string is never matched literally.

    Such an assertion is not captured. (No assertion with leading punctuation is captured by default.) You may always capture it explicitly, of course.

    A subrule is considered declarative to the extent that the front of it is declarative, and to the extent that the variable doesn't change. Prefix with a sequence point to defeat repeated static optimizations.

  • A leading :: indicates a symbolic indirect subrule:

         / <::($somename)> /

    The variable must contain the name of a subrule. By the rules of single method dispatch this is first searched for in the current grammar and its ancestors. If this search fails an attempt is made to dispatch via MMD, in which case it can find subrules defined as multis rather than methods. This form is not captured by default. It is always considered procedural, not declarative.

  • A leading @ matches like a bare array except that each element is treated as a subrule (string or Regex object) rather than as a literal. That is, a string is forced to be compiled as a subrule instead of being matched literally. (There is no difference for a Regex object.)

    This assertion is not automatically captured.

  • A leading % matches like a bare hash except that a string value is always treated as a subrule, even if it is a string that must be compiled to a regex at match time. (Numeric values may still indicate "false match". and a closure may do whatever it likes.)

    This assertion is not automatically captured.

    As with bare hash, the longest key matches according to the venerable longest-token rule.

  • A leading { indicates code that produces a regex to be interpolated into the pattern at that point as a subrule:

         / (<.ident>)  <{ %cache{$0} //= get_body_for($0) }> /

    The closure is guaranteed to be run at the canonical time; it declares a sequence point, and is considered to be procedural.

  • A leading & interpolates the return value of a subroutine call as a regex. Hence

         <&foo()>

    is short for

         <{ foo() }>

    This is considered procedural.

  • In any case of regex interpolation, if the value already happens to be a Regex object, it is not recompiled. If it is a string, the compiled form is cached with the string so that it is not recompiled next time you use it unless the string changes. (Any external lexical variable names must be rebound each time though.) Subrules may not be interpolated with unbalanced bracketing. An interpolated subrule keeps its own inner match result as a single item, so its parentheses never count toward the outer regexes groupings. (In other words, parenthesis numbering is always lexically scoped.)

  • A leading ?{ or !{ indicates a code assertion:

         / (\d**1..3) <?{ $0 < 256 }> /
         / (\d**1..3) <!{ $0 < 256 }> /

    Similar to:

         / (\d**1..3) { $0 < 256 or fail } /
         / (\d**1..3) { $0 < 256 and fail } /

    Unlike closures, code assertions are considered declarative; they are not guaranteed to be run at the canonical time if the optimizer can prove something later can't match. So you can sneak in a call to a non-canonical closure that way:

         token { foo .* <?{ do { say "Got here!" } or 1 }> .* bar }

    The do block is unlikely to run unless the string ends with "bar".

  • A leading [ indicates an enumerated character class. Ranges in enumerated character classes are indicated with ".." rather than "-".

         / <[a..z_]>* /

    Whitespace is ignored within square brackets:

         / <[ a..z _ ]>* /
  • A leading - indicates a complemented character class:

         / <-[a..z_]> <-alpha> /
         / <- [a..z_]> <- alpha> /  # whitespace allowed after -

    This is essentially the same as using negative lookahead and dot:

        / <![a..z_]> . <!alpha> . /

    Whitespace is ignored after the initial -.

  • A leading + may also be supplied to indicate that the following character class is to matched in a positive sense.

         / <+[a..z_]>* /
         / <+[ a..z _ ]>* /
         / <+ [ a .. z _ ] >* /      # whitespace allowed after +
  • Character classes can be combined (additively or subtractively) within a single set of angle brackets. Whitespace is ignored. For example:

         / <[a..z] - [aeiou] + xdigit> /      # consonant or hex digit

    A named character class may be used by itself:

        <alpha>

    However, in order to combine classes you must prefix a named character class with + or -.

  • The special assertion <.> matches any logical grapheme (including a Unicode combining character sequences):

         / seekto = <.> /  # Maybe a combined char

    Same as:

         / seekto = [:graphs .] /
  • A leading ! indicates a negated meaning (always a zero-width assertion):

         / <!before _ > /    # We aren't before an _

    Note that <!alpha> is different from <-alpha>. /<-alpha>/ is a complemented character class equivalent to /<!before <alpha>> ./, whereas <!alpha> is a zero-width assertion equivalent to a /<!before <alpha>>/ assertion.

    Note also that as a metacharacter ! doesn't change the parsing rules of whatever follows (unlike, say, + or -).

  • A leading ? indicates a positive zero-width assertion, and like ! merely reparses the rest of the assertion recursively as if the ? were not there. In addition to forcing zero-width, it also suppresses any named capture:

        <alpha>     # match a letter and capture to $alpha (eventually $<alpha>)
        <.alpha>    # match a letter, don't capture
        <?alpha>    # match null before a letter, don't capture
  • A leading ~~ indicates a recursive call back into some or all of the current rule. An optional argument indicates which subpattern to re-use, and if provided must resolve to a single subpattern. If omitted, the entire pattern is called recursively:

        <~~>       # call myself recursively
        <~~0>      # match according to $0's pattern
        <~~foo>    # match according to $foo's pattern

    Note that this rematches the pattern associated with the name, not the string matched. So

        $_ = "foodbard"
    
        / ( foo | bar ) d $0 /      # fails; doesn't match "foo" literally
        / ( foo | bar ) d <$0> /    # fails; doesn't match /foo/ as subrule
        / ( foo | bar ) d <~~0> /   # matches using rule associated with $0

    The last is equivalent to

        / ( foo | bar ) d ( foo | bar) /

    Note that the "self" call of

        / <term> <operator> <~~> /

    calls back into this anonymous rule as a subrule, and is implicitly anchored to the end of the operator as any other subrule would be. Despite the fact that the outer rule scans the string, the inner call to it does not.

    Note that a consequence of previous section is that you also get

        <!~~>

    for free, which fails if the current rule would match again at this location.

The following tokens include angles but are not required to balance:

  • A <( token indicates the start of a result capture, while the corresponding )> token indicates its endpoint. When matched, these behave as assertions that are always true, but have the side effect of setting the .from and .to attributes of the match object. That is:

        / foo <( \d+ )> bar /

    is equivalent to:

        / <after foo> \d+ <before bar> /

    except that the scan for "foo" can be done in the forward direction, while a lookbehind assertion would presumably scan for \d+ and then match "foo" backwards. The use of <(...)> affects only the meaning of the result object and the positions of the beginning and ending of the match. That is, after the match above, $() contains only the digits matched, and $/.to is pointing to after the digits. Other captures (named or numbered) are unaffected and may be accessed through $/.

    These tokens are considered declarative, but may force backtracking behavior.

  • A « or << token indicates a left word boundary. A » or >> token indicates a right word boundary. (As separate tokens, these need not be balanced.) Perl�5's \b is replaced by a <?wb> "word boundary" assertion, while \B becomes <!wb>. (None of these are dependent on the definition of <.ws>, but only on the \w definition of "word" characters.)

Backslash reform

  • The \p and \P properties become intrinsic grammar rules such as (<alpha> and <-alpha>). They may be combined using the above-mentioned character class notation: <[_]+alpha+digit>. Regardless of the higher-level character class names, low-level Unicode properties are always available with a prefix of is. Hence, <+isLu+isLt> is equivalent to <+upper+title>. If you define your own "is" properties they hide any Unicode properties of the same name.

  • The \L...\E, \U...\E, and \Q...\E sequences are gone. In the rare cases that need them you can use <{ lc $regex }> etc.

  • The \G sequence is gone. Use :p instead. (Note, however, that it makes no sense to use :p within a pattern, since every internal pattern is implicitly anchored to the current position.) See the at assertion below.

  • Backreferences (e.g. \1, \2, etc.) are gone; $0, $1, etc. can be used instead, because variables are no longer interpolated.

    Numeric variables are assumed to change every time and therefore are considered procedural, unlike normal variables.

  • New backslash sequences, \h and \v, match horizontal and vertical whitespace respectively, including Unicode.

  • \s now matches any Unicode whitespace character.

  • The new backslash sequence \N matches anything except a logical newline; it is the negation of \n.

  • A series of other new capital backslash sequences are also the negation of their lower-case counterparts:

    • \H matches anything but horizontal whitespace.

    • \V matches anything but vertical whitespace.

    • \T matches anything but a tab.

    • \R matches anything but a return.

    • \F matches anything but a formfeed.

    • \E matches anything but an escape.

    • \X... matches anything but the specified character (specified in hexadecimal).

Regexes are now first-class language, not strings

  • The Perl�5 qr/pattern/ regex constructor is gone.

  • The Perl�6 equivalents are:

         regex { pattern }    # always takes {...} as delimiters
         rx    / pattern /    # can take (almost any) chars as delimiters

    You may not use whitespace or alphanumerics for delimiters. Space is optional unless needed to distinguish from modifier arguments or function parens. So you may use parens as your rx delimiters, but only if you interpose whitespace:

         rx ( pattern )      # okay
         rx( 1,2,3 )         # tries to call rx function

    (This is true for all quotelike constructs in Perl�6.)

  • If either form needs modifiers, they go before the opening delimiter:

         $regex = regex :g:s:i { my name is (.*) };
         $regex = rx:g:s:i     / my name is (.*) /;    # same thing

    Space is necessary after the final modifier if you use any bracketing character for the delimiter. (Otherwise it would be taken as an argument to the modifier.)

  • You may not use colons for the delimiter. Space is allowed between modifiers:

         $regex = rx :g :s :i / my name is (.*) /;
  • The name of the constructor was changed from qr because it's no longer an interpolating quote-like operator. rx is short for regex, (not to be confused with regular expressions, except when they are).

  • As the syntax indicates, it is now more closely analogous to a sub {...} constructor. In fact, that analogy runs very deep in Perl�6.

  • Just as a raw {...} is now always a closure (which may still execute immediately in certain contexts and be passed as an object in others), so too a raw /.../ is now always a Regex object (which may still match immediately in certain contexts and be passed as an object in others).

  • Specifically, a /.../ matches immediately in a value context (void, Boolean, string, or numeric), or when it is an explicit argument of a ~~. Otherwise it's a Regex constructor identical to the explicit regex form. So this:

         $var = /pattern/;

    no longer does the match and sets $var to the result. Instead it assigns a Regex object to $var.

  • The two cases can always be distinguished using m{...} or rx{...}:

         $match = m{pattern};    # Match regex immediately, assign result
         $regex = rx{pattern};   # Assign regex expression itself
  • Note that this means that former magically lazy usages like:

         @list = split /pattern/, $str;

    are now just consequences of the normal semantics.

  • It's now also possible to set up a user-defined subroutine that acts like grep:

         sub my_grep($selector, *@list) {
             given $selector {
                 when Regex { ... }
                 when Code  { ... }
                 when Hash  { ... }
                 # etc.
             }
         }

    Using {...} or /.../ in the scalar context of the first argument causes it to produce a Code or Regex object, which the switch statement then selects upon.

  • Just as rx has variants, so does the regex declarator. In particular, there are two special variants for use in grammars: token and rule.

    A token declaration:

        token ident { [ <alpha> | _ ] \w* }

    never backtracks by default. That is, it likes to commit to whatever it has scanned so far. The above is equivalent to

        regex ident { [ <alpha>: | _: ]: \w*: }

    but rather easier to read. The bare *, +, and ? quantifiers never backtrack in a token unless some outer regex has specified a :panic option that applies. If you want to prevent even that, use *:, +:, or ?: to prevent any backtracking into the quantifier. If you want to explicitly backtrack, append either a ? or a ! to the quantifier. The ? forces minimal matching as usual, while the ! forces greedy matching. The token declarator is really just short for

        regex :ratchet { ... }

    The other is the rule declarator, for declaring non-terminal productions in a grammar. Like a token, it also does not backtrack by default. In addition, a rule regex also assumes :sigspace. A rule is really short for:

        regex :ratchet :sigspace { ... }
  • The Perl�5 ?...? syntax (succeed once) was rarely used and can be now emulated more cleanly with a state variable:

        $result = do { state $x ||= m/ pattern /; }    # only matches first time

    To reset the pattern, simply say $x = 0. Though if you want $x visible you'd have to avoid using a block:

        $result = state $x ||= m/ pattern /;
        ...
        $x = 0;

Backtracking control

Within those portions of a pattern that are considered procedural rather than declarative, you may control the backtracking behavior.

  • By default, backtracking is greedy in rx, m, s, and the like. It's also greedy in ordinary regex declarations. In rule and token declarations, backtracking must be explicit.

  • To force the preceding atom to do eager backtracking, append a :? or ? to the atom. If the preceding token is a quantifier, the : may be omitted, so *? works just as in Perl�5.

  • To force the preceding atom to do greedy backtracking in a spot that would default otherwise, append a :! to the atom. If the preceding token is a quantifier, the : may be omitted. (Perl�5 has no corresponding construct because backtracking always defaults to greedy in Perl�5.)

  • To force the preceding atom to do no backtracking, use a single : without a subsequent ? or !. Backtracking over a single colon causes the regex engine not to retry the preceding atom:

         mm/ \( <expr> [ , <expr> ]*: \) /

    (i.e. there's no point trying fewer <expr> matches, if there's no closing parenthesis on the horizon)

    To force all the atoms in an expression not to backtrack by default, use :ratchet or rule or token.

  • Backtracking over a double colon causes the immediately surrounding group (usually but not always a group of alternations) to immediately fail:

         mm/ [ if :: <expr> <block>
             | for :: <list> <block>
             | loop :: <loop_controls>? <block>
             ]
         /

    (i.e. there's no point trying to match a different keyword if one was already found but failed). Note that you can still back into such an alternation, so you may also need to put : after it if you also want to disable that. If an explicit or implicit :ratchet has disabled backtracking by supplying an implicit :, you need to put an explicit ! after the alternation to enable backing into another alternative if the first pick fails.

    The :: also has the effect of hiding any constant string on the right from "longest token" processing by |. Only the left side is evaluated for initial constancy.

  • Backtracking over a triple colon causes the current regex to fail outright (no matter where in the regex it occurs):

         regex ident {
               ( [<alpha>|_] \w* ) ::: { fail if %reserved{$0} }
             || " [<alpha>|_] \w* "
         }
    
         mm/ get <ident>? /

    (i.e. using an unquoted reserved word as an identifier is not permitted)

  • Backtracking over a <commit> assertion causes the entire match to fail outright, no matter how many subrules down it happens:

         regex subname {
             ([<alpha>|_] \w*) <commit> { fail if %reserved{$0} }
         }
         mm/ sub <subname>? <block> /

    (i.e. using a reserved word as a subroutine name is instantly fatal to the surrounding match as well)

    If commit is given an argument, it's the name of a calling rule that should be committed:

        <commit('infix')>
  • A <cut> assertion always matches successfully, and has the side effect of logically deleting the parts of the string already matched. Whether this actually frees up the memory immediately may depend on various interactions among your backreferences, the string implementation, and the garbage collector. In any case, the string will report that it has been chopped off on the front. It's illegal to use <cut> on a string that you do not have write access to.

    Attempting to backtrack past a <cut> causes the complete match to fail (like backtracking past a <commit>). This is because there's now no preceding text to backtrack into. This is useful for throwing away successfully processed input when matching from an input stream or an iterator of arbitrary length.

Regex Routines, Named and Anonymous

  • The analogy between sub and regex extends much further.

  • Just as you can have anonymous subs and named subs...

  • ...so too you can have anonymous regexes and named regexes (and tokens, and rules):

         token ident { [<alpha>|_] \w* }
    
         # and later...
    
         @ids = grep /<ident>/, @strings;
  • As the above example indicates, it's possible to refer to named regexes, such as:

         regex serial_number { <[A..Z]> \d**8 }
         token type { alpha | beta | production | deprecated | legacy }

    in other regexes as named assertions:

         rule identification { [soft|hard]ware <type> <serial_number> }

    These keyword-declared regexes are officially of type Method, which is derived from Routine.

    In general, the anchoring of any subrule call is controlled by context. When a regex, token, or rule method is called as a subrule, the front is anchored to the current position (as with :p), while the end is not anchored, since the calling context will likely wish to continue parsing. However, when such a method is smartmatched directly, it is automatically anchored on both ends to the beginning and end of the string. Thus, you can do direct pattern matching by using an anonymous regex routine as a standalone pattern:

        $string ~~ regex { \d+ }
        $string ~~ token { \d+ }
        $string ~~ rule { \d+ }

    and these are equivalent to

        $string ~~ m/^ \d+ $/;
        $string ~~ m/^ \d+: $/;
        $string ~~ m/^ <.ws> \d+: <.ws> $/;

    The basic rule of thumb is that the keyword-defined methods never do implicit .*?-like scanning, while the m// and s// quotelike forms do such scanning in the absence of explicit anchoring.

    The rx// and // forms can go either way: they scan when used directly within a smartmatch or boolean context, but when called indirectly as a subrule they do not scan. That is, the object returned by rx// behaves like m// when used directly, but like regex {} when used as a subrule:

        $pattern = rx/foo/;
        $string ~~ $pattern;                  # equivalent to m/foo/;
        $string ~~ /'[' <$pattern> ']'/       # equivalent to /'[foo]'/

Nothing is illegal

  • The empty pattern is now illegal.

  • To match whatever the prior successful regex matched, use:

         / <prior> /
  • To match the zero-width string, you must use some explicit representation of the null match:

        / '' /;
        / <?> /;

    For example:

         split /''/, $string

    splits between characters. But then, so does this:

         split '', $string
  • Likewise, to match a empty alternative, use something like:

         /a|b|c|<?>/
         /a|b|c|''/

    This makes it easier to catch errors like this:

        /a|b|c|/

    As a special case, however, the first null alternative in a match like

         mm/ [
             | if :: <expr> <block>
             | for :: <list> <block>
             | loop :: <loop_controls>? <block>
             ]
         /

    is simply ignored. Only the first alternative is special that way. If you write:

         mm/ [
                 if :: <expr> <block>              |
                 for :: <list> <block>             |
                 loop :: <loop_controls>? <block>  |
             ]
         /

    it's still an error.

  • However, it's okay for a non-null syntactic construct to have a degenerate case matching the null string:

         $something = "";
         /a|b|c|$something/;

    In particular, <?> always matches the null string successfuly, and <!> always fails to match anything.

Longest-token matching

Instead of representing temporal alternation, | now represents logical alternation with declarative longest-token semantics. (You may now use || to indicate the old temporal alternation. That is, | and || now work within regex syntax much the same as they do outside of regex syntax, where they represent junctional and short-circuit OR. This includes the fact that | has tighter precedence than ||.)

Historically regex processing has proceeded in Perl via a backtracking NFA algorithm. This is quite powerful, but many parsers work more efficiently by processing rules in parallel rather than one after another, at least up to a point. If you look at something like a yacc grammar, you find a lot of pattern/action declarations where the patterns are considered in parallel, and eventually the grammar decides which action to fire off. While the default Perl view of parsing is essentially top-down (perhaps with a bottom-up "middle layer" to handle operator precedence), it is extremely useful for user understanding if at least the token processing proceeds deterministically. So for regex matching purposes we define token patterns as those patterns containing no whitespace that can be matched without side effects or self-reference. Basically, Perl automatically derives a lexer from the grammar without you having to write one yourself.

To that end, every regex in Perl�6 is required to be able to distinguish its "pure" patterns from its actions, and return its list of initial token patterns (transitively including the token patterns of any subrule called by the "pure" part of that regex, but not including any subrule more than once, since that would involve self reference, which is not allowed in traditional regular expressions). A logical alternation using | then takes two or more of these lists and dispatches to the alternative that matches the longest token prefix. This may or may not be the alternative that comes first lexically. (However, in the case of a tie between alternatives, the textually earlier alternative does take precedence.)

This longest token prefix corresponds roughly to the notion of "token" in other parsing systems that use a lexer, but in the case of Perl this is largely an epiphenomenon derived automatically from the grammar definition. However, despite being automatically calculated, the set of tokens can be modified by the user; various constructs within a regex declaratively tell the grammar engine that it is finished with the pattern part and starting in on the side effects, so by inserting such constructs the user controls what is considered a token and what is not. The constructs deemed to terminate a token declaration and start the "action" part of the pattern include:

  • Any :: or ::: backtracking control (but not the : possessive modifier).

  • Any atom that is quantified with a minimal match (using the ? modifier).

  • Any {...} action, but not an assertion containing a closure. The closure form of the general **{...} quantifier terminates the longest token, but not the closureless forms.

  • Any sequential control flow operator such as || or &&.

  • As a consequence of the previous point, and because the standard grammar's <ws> rule defines whitespace using ||, the longest token is also terminated by any part of the regex or rule that might match whitespace using that rule, including whitespace implicitly matched via :sigspace. (However, token declarations are specifically allowed to recognize whitespace within a token by using such lower-level primitives as \h+ or other character classes.)

Subpatterns (captures) specifically do not terminate the token pattern, but may require a reparse of the token to find the location of the subpatterns. Likewise assertions may need to be checked out after the longest token is determined. (Alternately, if DFA semantics are simulated in any of various ways, such as by Thompson NFA, it may be possible to know when to fire off the assertions without backchecks.)

Greedy quantifiers and character classes do not terminate a token pattern. Zero-width assertions such as word boundaries are also okay.

For a pattern that starts with a positive lookahead assertion, the assertion is assumed to be more specific than the subsequent pattern, so the lookahead's pattern is treated as the longest token; the longest-token matcher will be smart enough to rematch any text traversed by the lookahead when (and if) it continues the match.

Oddly enough, the token keyword specifically does not determine the scope of a token, except insofar as a token pattern usually doesn't do much matching of whitespace. In contrast, the rule keyword (which assumes :sigspace) defines a pattern that tends to disqualify itself on the first whitespace. So most of the token patterns will end up coming from token declarations. For instance, a token declaration such as

    token list_composer { \[ <expr> \] }

considers its "longest token" to be just the left square bracket, because the first thing the expr rule will do is traverse optional whitespace.

The initial token matcher must take into account case sensitivity (or any other canonicalization primitives) and do the right thing even when propagated up to rules that don't have the same canonicalization. That is, they must continue to represent the set of matches that the lower rule would match.

The || form has the old short-circuit semantics, and will not attempt to match its right side unless all possibilities (including all | possibilities) are exhausted on its left. The first || in a regex makes the token patterns on its left available to the outer longest-token matcher, but hides any subsequent tests from longest-token matching. Every || establishes a new longest-token matcher. That is, if you use | on the right side of ||, that right side establishes a new top level scope for longest-token processing for this subexpression and any called subrules. The right side's longest-token automaton is invisible to the left of the || or outside the regex containing the ||.

Return values from matches

Match objects

  • A match always returns a Match object, which is also available as $/, which is a contextual lexical declared in the outer subroutine that is calling the regex. (A regex declares its own lexical $/ variable, which always refers to the most recent submatch within the rule, if any.) The current match state is kept in the regex's $_ variable which will eventually get processed into the user's $/ variable when the match completes.

  • Notionally, a match object contains (among other things) a boolean success value, a scalar result object, an array of ordered submatch objects, and a hash of named submatch objects. To provide convenient access to these various values, the match object evaluates differently in different contexts:

    • In boolean context it evaluates as true or false (i.e. did the match succeed?):

           if /pattern/ {...}
           # or:
           /pattern/; if $/ {...}

      With :global or :overlap or :exhaustive the boolean is allowed to return true on the first match. The Match object can produce the rest of the results lazily if evaluated in list context.

    • In string context it evaluates to the stringified value of its result object, which is usually the entire matched string:

           print %hash{ "{$text ~~ /<.ident>/}" };
           # or equivalently:
           $text ~~ /<.ident>/  &&  print %hash{~$/};

      But generally you should say ~$/ if you mean ~$/.

    • In numeric context it evaluates to the numeric value of its result object, which is usually the entire matched string:

           $sum += /\d+/;
           # or equivalently:
           /\d+/; $sum = $sum + $/;
    • When used as a scalar, a Match object evaluates to its underlying result object. Usually this is just the entire match string, but you can override that by calling make inside a regex:

          my $moose = $(m:{
              <antler> <body>
              { make Moose.new( body => $<body>.attach($<antler>) ) }
              # match succeeds -- ignore the rest of the regex
          });

      $() is a shorthand for $($/). The result object may be of any type, not just a string.

      You may also capture a subset of the match as the result object using the <(...)> construct:

          "foo123bar" ~~ / foo <( \d+ )> bar /
          say $();    # says 123

      In this case the result object is always a string when doing string matching, and a list of one or more elements when doing array matching.

      Additionally, the Match object delegates its coerce calls (such as +$match and ~$match) to its underlying result object. The only exception is that Match handles boolean coercion itself, which returns whether the match had succeeded at least once.

      This means that these two work the same:

          / <moose> { make $moose as Moose } /
          / <moose> { make $$moose as Moose } /
    • When used as an array, a Match object pretends to be an array of all its positional captures. Hence

           ($key, $val) = mm/ (\S+) => (\S+)/;

      can also be written:

           $result = mm/ (\S+) '=>' (\S+)/;
           ($key, $val) = @$result;

      To get a single capture into a string, use a subscript:

           $mystring = "{ mm/ (\S+) '=>' (\S+)/[0] }";

      To get all the captures into a string, use a zen slice:

           $mystring = "{ mm/ (\S+) '=>' (\S+)/[] }";

      Or cast it into an array:

           $mystring = "@( mm/ (\S+) '=>' (\S+)/ )";

      Note that, as a scalar variable, $/ doesn't automatically flatten in list context. Use @() as a shorthand for @($/) to flatten the positional captures under list context. Note that a Match object is allowed to evaluate its match lazily in list context. Use eager @() to force an eager match.

    • When used as a hash, a Match object pretends to be a hash of all its named captures. The keys do not include any sigils, so if you capture to variable @<foo> its real name is $/{'foo'} or $/<foo>. However, you may still refer to it as @<foo> anywhere $/ is visible. (But it is erroneous to use the same name for two different capture datatypes.)

      Note that, as a scalar variable, $/ doesn't automatically flatten in list context. Use %() as a shorthand for %($/) to flatten as a hash, or bind it to a variable of the appropriate type. As with @(), it's possible for %() to produce its pairs lazily in list context.

    • The numbered captures may be treated as named, so $<0 1 2> is equivalent to $/[0,1,2]. This allows you to write slices of intermixed named and numbered captures.

    • In ordinary code, variables $0, $1, etc. are just aliases into $/[0], $/[1], etc. Hence they will all be undefined if the last match failed (unless they were explicitly bound in a closure without using the let keyword).

  • Match objects have methods that provide additional information about the match. For example:

         if m/ def <ident> <codeblock> / {
             say "Found sub def from index $/.from.bytes ",
                 "to index $/.to.bytes";
         }
  • All match attempts--successful or not--against any regex, subrule, or subpattern (see below) return an object of class Match. That is:

         $match_obj = $str ~~ /pattern/;
         say "Matched" if $match_obj;
  • This returned object is also automatically assigned to the lexical $/ variable of the current surroundings. That is:

         $str ~~ /pattern/;
         say "Matched" if $/;
  • Inside a regex, the $_ variable holds the current regex's incomplete Match object, known as a match state. Generally this should not be modified unless you know how to create and propagate match states. All regexes actually return match states even when you think they're returning something else, because the match states keep track of the success and failures of the pattern for you.

    Fortunately, when you just want to return a different result object instead of the default Match object, you may associate your return value with the current match state using the make function, which works something like a return, but doesn't clobber the match state:

        $str ~~ / foo                 # Match 'foo'
                   { make 'bar' }     # But pretend we matched 'bar'
                 /;
        say $();                      # says 'bar'

Subpattern captures

  • Any part of a regex that is enclosed in capturing parentheses is called a subpattern. For example:

            #               subpattern
            #  _________________/\___________________
            # |                                      |
            # |       subpattern  subpattern         |
            # |          __/\__    __/\__            |
            # |         |      |  |      |           |
          mm/ (I am the (walrus), ( khoo )**2  kachoo) /;
  • Each subpattern in a regex produces a Match object if it is successfully matched.

  • Each subpattern's Match object is pushed onto the array inside the outer Match object belonging to the surrounding scope (known as its parent Match object). The surrounding scope may be either the innermost surrounding subpattern (if the subpattern is nested) or else the entire regex itself.

  • Like all captures, these assignments to the array are hypothetical, and are undone if the subpattern is backtracked.

  • For example, if the following pattern matched successfully:

            #                subpat-A
            #  _________________/\__________________
            # |                                     |
            # |         subpat-B  subpat-C          |
            # |          __/\__    __/\__           |
            # |         |      |  |      |          |
          mm/ (I am the (walrus), ( khoo )**2 kachoo) /;

    then the Match objects representing the matches made by subpat-B and subpat-C would be successively pushed onto the array inside subpat- A's Match object. Then subpat-A's Match object would itself be pushed onto the array inside the Match object for the entire regex (i.e. onto $/'s array).

  • As a result of these semantics, capturing parentheses in Perl�6 are hierarchical, not linear (see "Nested subpattern captures").

Accessing captured subpatterns

  • The array elements of a Match object are referred to using either the standard array access notation (e.g. $/[0], $/[1], $/[2], etc.) or else via the corresponding lexically scoped numeric aliases (i.e. $0, $1, $2, etc.) So:

         say "$/[1] was found between $/[0] and $/[2]";

    is the same as:

         say "$1 was found between $0 and $2";
  • Note that, in Perl�6, the numeric capture variables start from $0, not $1, with the numbers corresponding to the element's index inside $/.

  • The array elements of the regex's Match object (i.e. $/) store individual Match objects representing the substrings that were matched and captured by the first, second, third, etc. outermost (i.e. unnested) subpatterns. So these elements can be treated like fully fledged match results. For example:

         if m/ (\d\d\d\d)-(\d\d)-(\d\d) (BCE?|AD|CE)?/ {
               ($yr, $mon, $day) = $/[0..2];
               $era = "$3" if $3;                    # stringify/boolify
               @datepos = ( $0.from() .. $2.to() );  # Call Match methods
         }

Nested subpattern captures

  • Substrings matched by nested subpatterns (i.e. nested capturing parens) are assigned to the array inside the nested subpattern's parent Match object, not to the array of $/.

  • This behavior is quite different from Perl�5 semantics:

          # Perl 5...
          #
          # $1---------------------  $4---------  $5------------------
          # |   $2---------------  | |          | | $6----  $7------  |
          # |   |         $3--   | | |          | | |     | |       | |
          # |   |         |   |  | | |          | | |     | |       | |
         m/ ( A (guy|gal|g(\S+)  ) ) (sees|calls) ( (the|a) (gal|guy) ) /x;
  • In Perl�6, nested parens produce properly nested captures:

          # Perl 6...
          #
          # $0---------------------  $1---------  $2------------------
          # |   $0[0]------------  | |          | | $2[0]-  $2[1]---  |
          # |   |       $0[0][0] | | |          | | |     | |       | |
          # |   |         |   |  | | |          | | |     | |       | |
         m/ ( A (guy|gal|g(\S+)  ) ) (sees|calls) ( (the|a) (gal|guy) ) /;

Quantified subpattern captures

  • If a subpattern is directly quantified (using any quantifier), it no longer produces a single Match object. Instead, it produces a list of Match objects corresponding to the sequence of individual matches made by the repeated subpattern.

  • Because a quantified subpattern returns a list of Match objects, the corresponding array element for the quantified capture will store a (nested) array rather than a single Match object. For example:

         if m/ (\w+) \: (\w+ \s+)* / {
             say "Key:    $0";         # Unquantified --> single Match
             say "Values: @($1)";      # Quantified   --> array of Match
         }

Indirectly quantified subpattern captures

  • A subpattern may sometimes be nested inside a quantified non-capturing structure:

          #       non-capturing       quantifier
          #  __________/\____________  __/\__
          # |                        ||      |
          # |   $0         $1        ||      |
          # |  _^_      ___^___      ||      |
          # | |   |    |       |     ||      |
         m/ [ (\w+) \: (\w+ \h*)* \n ] ** 2..* /

    Non-capturing brackets don't create a separate nested lexical scope, so the two subpatterns inside them are actually still in the regex's top-level scope, hence their top-level designations: $0 and $1.

  • However, because the two subpatterns are inside a quantified structure, $0 and $1 will each contain an array. The elements of that array will be the submatches returned by the corresponding subpatterns on each iteration of the non-capturing parentheses. For example:

         my $text = "foo:food fool\nbar:bard barb";
    
                   #   $0--     $1------
                   #   |   |    |       |
         $text ~~ m/ [ (\w+) \: (\w+ \h*)* \n ] ** 2..* /;
    
         # Because they're in a quantified non-capturing block...
         # $0 contains the equivalent of:
         #
         #       [ Match.new(str=>'foo'), Match.new(str=>'bar') ]
         #
         # and $1 contains the equivalent of:
         #
         #       [ Match.new(str=>'food '),
         #         Match.new(str=>'fool' ),
         #         Match.new(str=>'bard '),
         #         Match.new(str=>'barb' ),
         #       ]
  • In contrast, if the outer quantified structure is a capturing structure (i.e. a subpattern) then it will introduce a nested lexical scope. That outer quantified structure will then return an array of Match objects representing the captures of the inner parens for every iteration (as described above). That is:

         my $text = "foo:food fool\nbar:bard barb";
    
                   # $0-----------------------
                   # |                        |
                   # | $0[0]    $0[1]---      |
                   # | |   |    |       |     |
         $text ~~ m/ ( (\w+) \: (\w+ \h*)* \n ) ** 2..* /;
    
         # Because it's in a quantified capturing block,
         # $0 contains the equivalent of:
         #
         #       [ Match.new( str=>"foo:food fool\n",
         #                    arr=>[ Match.new(str=>'foo'),
         #                           [
         #                               Match.new(str=>'food '),
         #                               Match.new(str=>'fool'),
         #                           ]
         #                         ],
         #                  ),
         #         Match.new( str=>'bar:bard barb',
         #                    arr=>[ Match.new(str=>'bar'),
         #                           [
         #                               Match.new(str=>'bard '),
         #                               Match.new(str=>'barb'),
         #                           ]
         #                         ],
         #                  ),
         #       ]
         #
         # and there is no $1
  • In other words, quantified non-capturing parens collect their components into handy flattened lists, whereas quantified capturing parens collect their components in a handy hierarchical structure.

Subpattern numbering

  • The index of a given subpattern can always be statically determined, but is not necessarily unique nor always monotonic. The numbering of subpatterns restarts in each lexical scope (either a regex, a subpattern, or the branch of an alternation).

  • In particular, the index of capturing parentheses restarts after each | or || (but not after each & or &&). Hence:

                      # $0      $1    $2   $3    $4           $5
         $tune_up = rx/ ("don't") (ray) (me) (for) (solar tea), ("d'oh!")
                      # $0      $1      $2    $3        $4
                      | (every) (green) (BEM) (devours) (faces)
                      /;

    This means that if the second alternation matches, the @$/ array will contain ('every', 'green', 'BEM', 'devours', 'faces') rather than (undef, undef, undef, undef, undef, undef, 'every', 'green', 'BEM', 'devours', 'faces') (as the same regex would in Perl�5).

  • Note that it is still possible to mimic the monotonic Perl�5 capture indexing semantics. See "Numbered scalar aliasing" below for details.

Subrule captures

  • Any call to a named <regex> within a pattern is known as a subrule, whether that regex is actually defined as a regex or token or rule or even an ordinary method or multi.

  • Any bracketed construct that is aliased (see "Aliasing" below) to a named variable is also a subrule.

  • For example, this regex contains three subrules:

          # subrule       subrule     subrule
          #  __^__    _______^_____    __^__
          # |     |  |             |  |     |
         m/ <ident>  $<spaces>=(\s*)  <digit>+ /
  • Just like subpatterns, each successfully matched subrule within a regex produces a Match object. But, unlike subpatterns, that Match object is not assigned to the array inside its parent Match object. Instead, it is assigned to an entry of the hash inside its parent Match object. For example:

          #  .... $/ .....................................
          # :                                             :
          # :              .... $/[0] ..................  :
          # :             :                             : :
          # : $/<ident>   :        $/[0]<ident>         : :
          # :   __^__     :           __^__             : :
          # :  |     |    :          |     |            : :
          mm/  <ident> \: ( known as <ident> previously ) /

Accessing captured subrules

  • The hash entries of a Match object can be referred to using any of the standard hash access notations ($/{'foo'}, $/<bar>, $/«baz», etc.), or else via corresponding lexically scoped aliases ($<foo>, $«bar», $<baz>, etc.) So the previous example also implies:

          #    $<ident>             $0<ident>
          #     __^__                 __^__
          #    |     |               |     |
          mm/  <ident> \: ( known as <ident> previously ) /
  • Note that it makes no difference whether a subrule is angle-bracketed (<ident>) or aliased internally (<ident=name>) or aliased externally ($<ident>=(<alpha>\w*)). The name's the thing.

Repeated captures of the same subrule

  • If a subrule appears two (or more) times in any branch of a lexical scope (i.e. twice within the same subpattern and alternation), or if the subrule is quantified anywhere within a given scope, then its corresponding hash entry is always assigned an array of Match objects rather than a single Match object.

  • Successive matches of the same subrule (whether from separate calls, or from a single quantified repetition) append their individual Match objects to this array. For example:

         if mm/ mv <file> <file> / {
             $from = $<file>[0];
             $to   = $<file>[1];
         }

    (Note, for clarity we are ignoring whitespace subtleties here--the normal sigspace rules would require space only between alphanumeric characters, which is wrong. Assume that our file subrule deals with whitespace on its own.)

    Likewise, with a quantified subrule:

         if mm/ mv <file> ** 2 / {
             $from = $<file>[0];
             $to   = $<file>[1];
         }

    And with a mixture of both:

         if mm/ mv <file>+ <file> / {
             $to   = pop @($<file>);
             @from = @($<file>);
         }
  • However, if a subrule is explicitly renamed (or aliased -- see "Aliasing"), then only the new name counts when deciding whether it is or isn't repeated. For example:

         if mm/ mv <file> <dir=file> / {
             $from = $<file>;  # Only one subrule named <file>, so scalar
             $to   = $<dir>;   # The Capture Formerly Known As <file>
         }

    Likewise, neither of the following constructions causes <file> to produce an array of Match objects, since none of them has two or more <file> subrules in the same lexical scope:

         if mm/ (keep) <file> | (toss) <file> / {
             # Each <file> is in a separate alternation, therefore <file>
             # is not repeated in any one scope, hence $<file> is
             # not an Array object...
             $action = $0;
             $target = $<file>;
         }
    
         if mm/ <file> \: (<file>|none) / {
             # Second <file> nested in subpattern which confers a
             # different scope...
             $actual  = $/<file>;
             $virtual = $/[0]<file> if $/[0]<file>;
         }
  • On the other hand, unaliased square brackets don't confer a separate scope (because they don't have an associated Match object). So:

         if mm/ <file> \: [<file>|none] / { # Two <file>s in same scope
             $actual  = $/<file>[0];
             $virtual = $/<file>[1] if $/<file>[1];
         }

Aliasing

Aliases can be named or numbered. They can be scalar-, array-, or hash-like. And they can be applied to either capturing or non-capturing constructs. The following sections highlight special features of the semantics of some of those combinations.

Named scalar aliasing to subpatterns

  • If a named scalar alias is applied to a set of capturing parens:

            #         _____/capturing parens\_____
            #        |                            |
            #        |                            |
          mm/ $<key>=( (<[A..E]>) (\d**3..6) (X?) ) /;

    then the outer capturing parens no longer capture into the array of $/ as unaliased parens would. Instead the aliased parens capture into the hash of $/; specifically into the hash element whose key is the alias name.

  • So, in the above example, a successful match sets $<key> (i.e. $/<key>), but not $0 (i.e. not $/[0]).

  • More specifically:

    • $/<key> will contain the Match object that would previously have been placed in $/[0].

    • $/<key>[0] will contain the A-E letter,

    • $/<key>[1] will contain the digits,

    • $/<key>[2] will contain the optional X.

  • Another way to think about this behavior is that aliased parens create a kind of lexically scoped named subrule; that the contents of the parentheses are treated as if they were part of a separate subrule whose name is the alias.

Named scalar aliases applied to non-capturing brackets

  • If a named scalar alias is applied to a set of non-capturing brackets:

            #         __/non-capturing brackets\__
            #        |                            |
            #        |                            |
          mm/ $<key>=[ (<[A..E]>) (\d**3..6) (X?) ] /;

    then the corresponding $/<key> Match object contains only the string matched by the non-capturing brackets.

  • In particular, the array of the $/<key> entry is empty. That's because square brackets do not create a nested lexical scope, so the subpatterns are unnested and hence correspond to $0, $1, and $2, and not to $/<key>[0], $/<key>[1], and $/<key>[2].

  • In other words:

    • $/<key> will contain the complete substring matched by the square brackets (in a Match object, as described above),

    • $0 will contain the A-E letter,

    • $1 will contain the digits,

    • $2 will contain the optional X.

Named scalar aliasing to subrules

  • If a subrule is aliased, it assigns its Match object to the hash entry whose key is the name of the alias. And it no longer assigns anything to the hash entry whose key is the subrule name. That is:

         if m/ ID\: <id=ident> / {
             say "Identified as $/<id>";    # $/<ident> is undefined
         }
  • Hence aliasing a subrule changes the destination of the subrule's Match object. This is particularly useful for differentiating two or more calls to the same subrule in the same scope. For example:

         if mm/ mv <file>+ <dir=file> / {
             @from = @($<file>);
             $to   = $<dir>;
         }

Numbered scalar aliasing

  • If a numbered alias is used instead of a named alias:

         m/ $1=(<-[:]>*) \:  $0=<ident> /

    the behavior is exactly the same as for a named alias (i.e. the various cases described above), except that the resulting Match object is assigned to the corresponding element of the appropriate array rather than to an element of the hash.

  • If any numbered alias is used, the numbering of subsequent unaliased subpatterns in the same scope automatically increments from that alias number (much like enum values increment from the last explicit value). That is:

          #  --$1---    -$2-    --$6---    -$7-
          # |       |  |    |  |       |  |    |
         m/ $1=(food)  (bard)  $6=(bazd)  (quxd) /;
  • This follow-on behavior is particularly useful for reinstituting Perl5 semantics for consecutive subpattern numbering in alternations:

         $tune_up = rx/ ("don't") (ray) (me) (for) (solar tea), ("d'oh!")
                      | $6 = (every) (green) (BEM) (devours) (faces)
                      #              $7      $8    $9        $10
                      /;
  • It also provides an easy way in Perl�6 to reinstitute the unnested numbering semantics of nested Perl�5 subpatterns:

          # Perl 5...
          #               $1
          #  _____________/\___________
          # |    $2        $3      $4  |
          # |  __/\___   __/\___   /\  |
          # | |       | |       | |  | |
         m/ ( ( [A-E] ) (\d{3,6}) (X?) ) /x;
    
    
          # Perl 6...
          #                $0
          #  ______________/\______________
          # |   $0[0]       $0[1]    $0[2] |
          # |  ___/\___   ____/\____   /\  |
          # | |        | |          | |  | |
         m/ ( (<[A..E]>) (\d ** 3..6) (X?) ) /;
    
    
          # Perl 6 simulating Perl 5...
          #                 $1
          #  _______________/\________________
          # |        $2          $3       $4  |
          # |     ___/\___   ____/\____   /\  |
          # |    |        | |          | |  | |
         m/ $1=[ (<[A..E]>) (\d ** 3..6) (X?) ] /;

    The non-capturing brackets don't introduce a scope, so the subpatterns within them are at regex scope, and hence numbered at the top level. Aliasing the square brackets to $1 means that the next subpattern at the same level (i.e. the (<[A..E]>)) is numbered sequentially (i.e. $2), etc.

Scalar aliases applied to quantified constructs

  • All of the above semantics apply equally to aliases which are bound to quantified structures.

  • The only difference is that, if the aliased construct is a subrule or subpattern, that quantified subrule or subpattern will have returned a list of Match objects (as described in "Quantified subpattern captures" and "Repeated captures of the same subrule"). So the corresponding array element or hash entry for the alias will contain an array, instead of a single Match object.

  • In other words, aliasing and quantification are completely orthogonal. For example:

         if mm/ mv $0=<file>+ / {
             # <file>+ returns a list of Match objects,
             # so $0 contains an array of Match objects,
             # one for each successful call to <file>
    
             # $/<file> does not exist (it's pre-empted by the alias)
         }
    
    
         if m/ mv \s+ $<from>=(\S+ \s+)* / {
             # Quantified subpattern returns a list of Match objects,
             # so $/<from> contains an array of Match
             # objects, one for each successful match of the subpattern
    
             # $0 does not exist (it's pre-empted by the alias)
         }
  • Note, however, that a set of quantified non-capturing brackets always returns a single Match object which contains only the complete substring that was matched by the full set of repetitions of the brackets (as described in "Named scalar aliases applied to non-capturing brackets"). For example:

         "coffee fifo fumble" ~~ m/ $<effs>=[f <-[f]> ** 1..2 \s*]+ /;
    
         say $<effs>;    # prints "fee fifo fum"

Array aliasing

  • An alias can also be specified using an array as the alias instead of a scalar. For example:

         m/ mv \s+ @<from>=[(\S+) \s+]* <dir> /;
  • Using the @alias= notation instead of a $alias= mandates that the corresponding hash entry or array element always receives an array of Match objects, even if the construct being aliased would normally return a single Match object. This is useful for creating consistent capture semantics across structurally different alternations (by enforcing array captures in all branches):

         mm/ Mr?s? @<names>=<ident> W\. @<names>=<ident>
            | Mr?s? @<names>=<ident>
            /;
    
         # Aliasing to @names means $/<names> is always
         # an Array object, so...
    
         say @($/<names>);
  • For convenience and consistency, @<key> can also be used outside a regex, as a shorthand for @( $/<key> ). That is:

         mm/ Mr?s? @<names>=<ident> W\. @<names>=<ident>
            | Mr?s? @<names>=<ident>
            /;
    
         say @<names>;
  • If an array alias is applied to a quantified pair of non-capturing brackets, it captures the substrings matched by each repetition of the brackets into separate elements of the corresponding array. That is:

         mm/ mv $<files>=[ f.. \s* ]* /; # $/<files> assigned a single
                                         # Match object containing the
                                         # complete substring matched by
                                         # the full set of repetitions
                                         # of the non-capturing brackets
    
         mm/ mv @<files>=[ f.. \s* ]* /; # $/<files> assigned an array,
                                         # each element of which is a
                                         # Match object containing
                                         # the substring matched by Nth
                                         # repetition of the non-
                                         # capturing bracket match
  • If an array alias is applied to a quantified pair of capturing parens (i.e. to a subpattern), then the corresponding hash or array element is assigned a list constructed by concatenating the array values of each Match object returned by one repetition of the subpattern. That is, an array alias on a subpattern flattens and collects all nested subpattern captures within the aliased subpattern. For example:

         if mm/ $<pairs>=( (\w+) \: (\N+) )+ / {
             # Scalar alias, so $/<pairs> is assigned an array
             # of Match objects, each of which has its own array
             # of two subcaptures...
    
             for @($<pairs>) -> $pair {
                 say "Key: $pair[0]";
                 say "Val: $pair[1]";
             }
         }
    
    
         if mm/ @<pairs>=( (\w+) \: (\N+) )+ / {
             # Array alias, so $/<pairs> is assigned an array
             # of Match objects, each of which is flattened out of
             # the two subcaptures within the subpattern
    
             for @($<pairs>) -> $key, $val {
                 say "Key: $key";
                 say "Val: $val";
             }
         }
  • Likewise, if an array alias is applied to a quantified subrule, then the hash or array element corresponding to the alias is assigned a list containing the array values of each Match object returned by each repetition of the subrule, all flattened into a single array:

         rule pair { (\w+) \: (\N+) \n }
    
         if mm/ $<pairs>=<pair>+ / {
             # Scalar alias, so $/<pairs> contains an array of
             # Match objects, each of which is the result of the
             # <pair> subrule call...
    
             for @($<pairs>) -> $pair {
                 say "Key: $pair[0]";
                 say "Val: $pair[1]";
             }
         }
    
    
         if mm/ mv @<pairs>=<pair>+ / {
             # Array alias, so $/<pairs> contains an array of
             # Match objects, all flattened down from the
             # nested arrays inside the Match objects returned
             # by each match of the <pair> subrule...
    
             for @($<pairs>) -> $key, $val {
                 say "Key: $key";
                 say "Val: $val";
             }
         }
  • In other words, an array alias is useful to flatten into a single array any nested captures that might occur within a quantified subpattern or subrule. Whereas a scalar alias is useful to preserve within a top-level array the internal structure of each repetition.

  • It is also possible to use a numbered variable as an array alias. The semantics are exactly as described above, with the sole difference being that the resulting array of Match objects is assigned into the appropriate element of the regex's match array rather than to a key of its match hash. For example:

         if m/ mv  \s+  @0=((\w+) \s+)+  $1=((\W+) (\s*))* / {
             #          |                |
             #          |                |
             #          |                 \_ Scalar alias, so $1 gets an
             #          |                    array, with each element
             #          |                    a Match object containing
             #          |                    the two nested captures
             #          |
             #           \___ Array alias, so $0 gets a flattened array of
             #                just the (\w+) captures from each repetition
    
             @from     = @($0);      # Flattened list
    
             $to_str   = $1[0][0];   # Nested elems of
             $to_gap   = $1[0][1];   #    unflattened list
         }
  • Note again that, outside a regex, @0 is simply a shorthand for @($0), so the first assignment above could also have been written:

         @from = @0;

Hash aliasing

  • An alias can also be specified using a hash as the alias variable, instead of a scalar or an array. For example:

         m/ mv %<location>=( (<ident>) \: (\N+) )+ /;
  • A hash alias causes the corresponding hash or array element in the current scope's Match object to be assigned a (nested) Hash object (rather than an Array object or a single Match object).

  • If a hash alias is applied to a subrule or subpattern then the first nested numeric capture becomes the key of each hash entry and any remaining numeric captures become the values (in an array if there is more than one).

  • As with array aliases it is also possible to use a numbered variable as a hash alias. Once again, the only difference is where the resulting Match object is stored:

         rule one_to_many {  (\w+) \: (\S+) (\S+) (\S+) }
    
         if mm/ %0=<one_to_many>+ / {
             # $/[0] contains a hash, in which each key is provided by
             # the first subcapture within C<one_to_many>, and each
             # value is an array containing the
             # subrule's second, third, fourth, etc. subcaptures...
    
             for %($/[0]) -> $pair {
                 say "One:  $pair.key()";
                 say "Many: { @($pair.value) }";
             }
         }
  • Outside the regex, %0 is a shortcut for %($0):

             for %0 -> $pair {
                 say "One:  $pair.key()";
                 say "Many: @($pair.value)";
             }

External aliasing

  • Instead of using internal aliases like:

         m/ mv  @<files>=<ident>+  $<dir>=<ident> /

    the name of an ordinary variable can be used as an external alias, like so:

         m/ mv  @OUTER::files=<ident>+  $OUTER::dir=<ident> /
  • In this case, the behavior of each alias is exactly as described in the previous sections, except that any resulting capture is bound directly (but still hypothetically) to the variable of the specified name that must already exist in the scope in which the regex is declared.

Capturing from repeated matches

  • When an entire regex is successfully matched with repetitions (specified via the :x or :g flag) or overlaps (specified via the :ov or :ex flag), it will usually produce a series of distinct matches.

  • A successful match under any of these flags still returns a single Match object in $/. However, this object may represent a partial evaluation of the regex. Moreover, the values of this match object are slightly different from those provided by a non-repeated match:

    • The boolean value of $/ after such matches is true or false, depending on whether the pattern matched.

    • The string value is the substring from the start of the first match to the end of the last match (including any intervening parts of the string that the regex skipped over in order to find later matches).

    • Subcaptures are returned as a multidimensional list, which the user can choose to process in either of two ways. If you refer to @(), the multidimensionality is ignored and all the matches are returned flattened (but still lazily). If you refer to @@(), you can get each individual sublist as a Capture object. (That is, there is a @@() coercion operator that happens, like @(), to default to $/.) As with any multidimensional list, each sublist can be lazy separately.

    For example:

         if $text ~~ mm:g/ (\S+:) <rocks> / {
             say "Full match context is: [$/]";
         }

    But the list of individual match objects corresponding to each separate match is also available:

         if $text ~~ mm:g/ (\S+:) <rocks> / {
             say "Matched { +@@() } times";    # Note: forced eager here
    
             for @@() -> $m {
                 say "Match between $m.from() and $m.to()";
                 say 'Right on, dude!' if $m[0] eq 'Perl';
                 say "Rocks like $m<rocks>";
             }
         }

:keepall

  • All regexes remember everything if :keepall is in effect anywhere in the outer dynamic scope. In this case everything inside the angles is used as part of the key. Suppose the earlier example parsed whitespace:

         / <key> <.ws> '=>' <.ws> <value> { %hash{$key} = $value } /

    The two instances of <.ws> above would store an array of two values accessible as @<.ws>. It would also store the literal match into $<'=\>'>. Just to make sure nothing is forgotten, under :keepall any text or whitespace not otherwise remembered is attached as an extra property on the subsequent node. (The name of that property is "pretext".)

Grammars

  • Your private ident rule shouldn't clobber someone else's ident rule. So some mechanism is needed to confine rules to a namespace.

  • If subs are the model for rules, then modules/classes are the obvious model for aggregating them. Such collections of rules are generally known as grammars.

  • Just as a class can collect named actions together:

         class Identity {
             method name { "Name = $.name" }
             method age  { "Age  = $.age"  }
             method addr { "Addr = $.addr" }
    
             method desc {
                 print &.name(), "\n",
                       &.age(),  "\n",
                       &.addr(), "\n";
             }
    
             # etc.
         }

    so too a grammar can collect a set of named rules together:

         grammar Identity {
             rule name { Name '=' (\N+) }
             rule age  { Age  '=' (\d+) }
             rule addr { Addr '=' (\N+) }
             rule desc {
                 <name> \n
                 <age>  \n
                 <addr> \n
             }
    
             # etc.
         }
  • Like classes, grammars can inherit:

         grammar Letter {
             rule text     { <greet> <body> <close> }
    
             rule greet { [Hi|Hey|Yo] $<to>=(\S+?) , $$}
    
             rule body     { <line>+? }   # note: backtracks forwards via +?
    
             rule close { Later dude, $<from>=(.+) }
    
             # etc.
         }
    
         grammar FormalLetter is Letter {
    
             rule greet { Dear $<to>=(\S+?) , $$}
    
             rule close { Yours sincerely, $<from>=(.+) }
    
         }
  • Just like the methods of a class, the rule definitions of a grammar are inherited (and polymorphic!). So there's no need to respecify body, line, etc.

  • Perl�6 will come with at least one grammar predefined:

         grammar Perl {    # Perl's own grammar
    
             rule prog { <statement>* }
    
             rule statement {
                       | <decl>
                       | <loop>
                       | <label> [<cond>|<sideff>|;]
             }
    
             rule decl { <sub> | <class> | <use> }
    
             # etc. etc. etc.
         }
  • Hence:

         given $source_code {
             $parsetree = m:keepall/<Perl.prog>/;
         }

Syntactic categories

For writing your own backslash and assertion subrules or macros, you may use the following syntactic categories:

     token rule_backslash:<w> { ... }     # define your own \w and \W
     token rule_assertion:<*> { ... }     # define your own <*stuff>
     macro rule_metachar:<,> { ... }     # define a new metacharacter
     macro rule_mod_internal:<x> { ... } # define your own /:x() stuff/
     macro rule_mod_external:<x> { ... } # define your own m:x()/stuff/

As with any such syntactic shenanigans, the declaration must be visible in the lexical scope to have any effect. It's possible the internal/external distinction is just a trait, and that some of those things are subs or methods rather than subrules or macros. (The numeric regex modifiers are recognized by fallback macros defined with an empty operator name.)

Pragmas

Various pragmas may be used to control various aspects of regex compilation and usage not otherwise provided for. These are tied to the particular declarator in question:

    use s :foo;         # control s defaults
    use m :foo;         # control m defaults
    use rx :foo;        # control rx defaults
    use regex :foo;     # control regex defaults
    use token :foo;     # control token defaults
    use rule :foo;      # control rule defaults

(It is a general policy in Perl�6 that any pragma designed to influence the surface behavior of a keyword is identical to the keyword itself, unless there is good reason to do otherwise. On the other hand, pragmas designed to influence deep semantics should not be named identically, though of course some similarity is good.)

Transliteration

  • The tr/// quote-like operator now also has a method form called trans(). Its argument is a list of pairs. You can use anything that produces a pair list:

         $str.trans( %mapping.pairs.sort );

    Use the .= form to do a translation in place:

         $str.=trans( %mapping.pairs.sort );

    (Perl�6 does not support the y/// form, which was only in sed because they were running out of single letters.)

  • The two sides of any pair can be strings interpreted as tr/// would:

         $str.=trans( 'A..C' => 'a..c', 'XYZ' => 'xyz' );

    As a degenerate case, each side can be individual characters:

         $str.=trans( 'A'=>'a', 'B'=>'b', 'C'=>'c' );
  • The two sides of each pair may also be Array objects:

         $str.=trans( ['A'..'C'] => ['a'..'c'], <X Y Z> => <x y z> );
  • The array version can map one-or-more characters to one-or-more characters:

         $str.=trans( [' ',      '<',    '>',    '&'    ] =>
                      ['&nbsp;', '&lt;', '&gt;', '&amp;' ]);

    In the case that more than one sequence of input characters matches, the longest one wins. In the case of two identical sequences the first in order wins.

Substitution

There are also method forms of m// and s///:

     $str.match(/pat/);
     $str.subst(/pat/, "replacement");
     $str.subst(/pat/, {"replacement"});
     $str.=subst(/pat/, "replacement");
     $str.=subst(/pat/, {"replacement"});

There is no syntactic sugar here, so in order to get deferred evaluation of the replacement you must put it into a closure. The syntactic sugar is provided only by the quotelike forms. First there is the standard "triple quote" form:

    s/pattern/replacement/

Only non-bracket characters may be used for the "triple quote". The right side is always evaluated as if it were a double-quoted string regardless of the quote chosen.

As with Perl�5, a bracketing form is also supported, but unlike Perl�5, Perl�6 uses the brackets only around the pattern. The replacement is then specified as if it were an ordinary item assignment, with ordinary quoting rules. To pick your own quotes on the right just use one of the q forms. The substitution above is equivalent to:

    s[pattern] = "replacement"

or

    s[pattern] = qq[replacement]

This is not a normal assigment, since the right side is evaluated each time the substitution matches (much like the pseudo-assignment to declarators can happen at strange times). It is therefore treated as a "thunk", that is, as if it has implicit curlies around it. In fact, it makes no sense at all to say

    s[pattern] = { doit }

because that would try to substitute a closure into the string.

Any scalar assignment operator may be used; the substitution macro knows how to turn

    $target ~~ s:g[pattern] op= expr

into something like:

    $target.subst(rx:g[pattern], { $() op expr })

So, for example, you can multiply every dollar amount by 2 with:

    s:g[\$ <( \d+ )>] *= 2

(Of course, the optimizer is free to do something faster than an actual method call.)

You'll note from the last example that substitutions only happen on the "official" string result of the match, that is, the $() value. (Here we captured $() using the <(...)> pair; otherwise we would have had to use lookbehind to match the $.)

Positional matching, fixed width types

  • To anchor to a particular position in the general case you can use the <at($pos)> assertion to say that the current position is the same as the position object you supply. You may set the current match position via the :c and :p modifiers.

    However, please remember that in Perl�6 string positions are generally not integers, but objects that point to a particular place in the string regardless of whether you count by bytes or codepoints or graphemes. If used with an integer, the at assertion will assume you mean the current lexically scoped Unicode level, on the assumption that this integer was somehow generated in this same lexical scope. If this is outside the current string's allowed Unicode abstraction levels, an exception is thrown. See S02 for more discussion of string positions.

  • Buf types are based on fixed-width cells and can therefore handle integer positions just fine, and treat them as array indices. In particular, buf8 (also known as buf) is just an old-school byte string. Matches against Buf types are restricted to ASCII semantics in the absence of an explicit modifier asking for the array's values to be treated as some particular encoding such as UTF-32. (This is also true for those compact arrays that are considered isomorphic to Buf types.) Positions within Buf types are always integers, counting one per unit cell of the underlying array. Be aware that "from" and "to" positions are reported as being between elements. If matching against a compact array @foo, a final position of 42 indicates that @foo[42] was the first element not included.

Matching against non-strings

  • Anything that can be tied to a string can be matched against a regex. This feature is particularly useful with input streams:

         my $stream := cat =$fh;       # tie scalar to filehandle
    
         # and later...
    
         $stream ~~ m/pattern/;         # match from stream
  • Any non-compact array of mixed strings or objects can be matched against a regex as long as you present them as an object with the Str interface, which does not preclude the object having other interfaces such as Array. Normally you'd use cat to generate such an object:

        @array.cat ~~ / foo <,> bar <elem>* /;

    The special <,> subrule matches the boundary between elements. The <elem> assertion matches any individual array element. It is the equivalent of the "dot" metacharacter for the whole element.

    If the array elements are strings, they are concatenated virtually into a single logical string. If the array elements are tokens or other such objects, the objects must provide appropriate methods for the kinds of subrules to match against. It is an assertion failure to match a string-matching assertion against an object that doesn't provide a stringified view. However, pure object lists can be parsed as long as the match (including any subrules) restricts itself to assertions like:

         <.isa(Dog)>
         <.does(Bark)>
         <.can('scratch')>

    It is permissible to mix objects and strings in an array as long as they're in different elements. You may not embed objects in strings, however. Any object may, of course, pretend to be a string element if it likes, and so a Cat object may be used as a substring with the same restrictions as in the main string.

    Please be aware that the warnings on .from and .to returning opaque objects goes double for matching against an array, where a particular position reflects both a position within the array and (potentially) a position within a string of that array. Do not expect to do math with such values. Nor should you expect to be able to extract a substr that crosses element boundaries. [Conjecture: Or should you?]

  • To match against every element of an array, use a hyper operator:

         @array».match($regex);
  • To match against any element of the array, it suffices to use ordinary smartmatching:

        @array ~~ $regex;

When $/ is valid

To provide implementational freedom, the $/ variable is not guaranteed to be defined until the pattern reaches a sequence point that requires it (such as completing the match, or calling an embedded closure, or even evaluating a submatch that requires a Perl expression for its argument). Within regex code, $/ is officially undefined, and references to $0 or other capture variables may be compiled to produce the current value without reference to $/. Likewise a reference to $<foo> does not necessarily mean $/<foo> within the regex proper. During the execution of a match, the current match state is likely to be stored in a $_ variable lexically scoped to an appropriate portion of the match, but that is not guaranteed to behave the same as the $/ object, because $/ is of type Match, while the match state is of type Cursor. (It really depends on the implementation of the pattern matching engine.)

In any case this is all transparent to the user for simple matches; and outside of regex code (and inside closures within the regex) the $/ variable is guaranteed to represent the state of the match at that point. That is, normal Perl code can always depend on $<foo> meaning $/<foo>, and $0 meaning $/[0], whether that code is embedded in a closure within the regex or outside the regex after the match completes.