* pcre2el: convert between PCRE, Emacs and rx regexp syntax

** Overview
   =pcre2el= or =rxt= (RegeXp Translator or RegeXp Tools) is a utility
   for working with regular expressions in Emacs, based on a
   recursive-descent parser for regexp syntax. In addition to
   converting (a subset of) PCRE syntax into its Emacs equivalent, it
   can do the following:

   - convert Emacs syntax to PCRE
   - convert either syntax to =rx=, an S-expression based regexp
     syntax
   - untangle complex regexps by showing the parse tree in =rx= form
     and highlighting the corresponding chunks of code
   - show the complete list of strings (productions) matching a
     regexp, provided the list is finite
   - provide live font-locking of regexp syntax (so far only for
     Elisp buffers -- other modes on the TODO list)

** Usage
   Enable =rxt-mode= or its global equivalent =rxt-global-mode=
   to get the default key-bindings. There are three sets of commands:
   commands that take a PCRE regexp, commands which take an Emacs
   regexp, and commands that try to do the right thing based on the
   current mode. Currently, this means Emacs syntax in
   =emacs-lisp-mode= and =lisp-interaction-mode=, and PCRE syntax
   everywhere else.

   The default key bindings all begin with =C-c /= and have a mnemonic
   structure: =C-c / <source> <target>=, or just =C-c / <target>= for
   the "do what I mean" commands. The complete list of key bindings is
   given here and explained in more detail below:

   - "Do-what-I-mean" commands:
     - =C-c / /= :: =rxt-explain=
     - =C-c / c= :: =rxt-convert-syntax=
     - =C-c / x= :: =rxt-convert-to-rx=
     - =C-c / ′= :: =rxt-convert-to-strings=

   - Commands that work on a PCRE regexp:
     - =C-c / p e= :: =rxt-pcre-to-elisp=
     - =C-c / %= :: =pcre-query-replace-regexp=
     - =C-c / p x= :: =rxt-pcre-to-rx=
     - =C-c / p s= :: =rxt-pcre-to-sre=
     - =C-c / p ′= :: =rxt-pcre-to-strings=
     - =C-c / p /= :: =rxt-explain-pcre=

   - Commands that work on an Emacs regexp:
     - =C-c / e /= :: =rxt-explain-elisp=
     - =C-c / e p= :: =rxt-elisp-to-pcre=
     - =C-c / e x= :: =rxt-elisp-to-rx=
     - =C-c / e s= :: =rxt-elisp-to-sre=
     - =C-c / e ′= :: =rxt-elisp-to-strings=
     - =C-c / e t= :: =rxt-toggle-elisp-rx=
     - =C-c / t= :: =rxt-toggle-elisp-rx=

*** Interactive input and output
    When used interactively, the conversion commands can read a regexp
    either from the current buffer or from the minibuffer. The output
    is displayed in the minibuffer and copied to the kill-ring.

    - When called with a prefix argument (=C-u=), they read a regular
      expression from the minibuffer literally, without further
      processing -- meaning there's no need to double the backslashes if
      it's an Emacs regexp.  This is the same way commands like
      =query-replace-regexp= read input.

    - When the region is active, they use they the region contents,
      again literally (without any translation of string syntax).

    - With neither a prefix arg nor an active region, the behavior
      depends on whether the command expects an Emacs regexp or
      a PCRE one.

      Commands that take an Emacs regexp behave like =C-x C-e=: they
      evaluate the sexp before point (which could be simply a string
      literal) and use its value. This is designed for use in Elisp
      buffers. As a special case, if point is *inside* a string, it's
      first moved to the string end, so in practice they should work
      as long as point is somewhere within the regexp literal.

      Commands that take a PCRE regexp try to read a Perl-style
      delimited regex literal *after* point in the current buffer,
      including its flags. For example, putting point before the =m=
      in the following example and doing =C-c / p e=
      (=rxt-pcre-to-elisp=) displays =\(?:bar\|foo\)=, correctly
      stripping out the whitespace and comment:

      : $x =~ m/  foo   |  (?# comment) bar /x

      The PCRE reader currently only works with =/ ... /= delimiters. It
      will ignore any preceding =m=, =s=, or =qr= operator, as well as
      the replacement part of an =s= construction.

      Readers for other PCRE-using languages are on the TODO list.

    The translation functions display their result in the minibuffer
    and copy it to the kill ring. When translating something into
    Elisp syntax, you might need to use the result either literally
    (e.g. for interactive input to a command like
    =query-replace-regexp=), or as a string to paste into Lisp code.
    To allow both uses, =rxt-pcre-to-elisp= copies both versions
    successively to the kill-ring. The literal regexp without string
    quoting is the top element of the kill-ring, while the Lisp string
    is the second-from-top. You can paste the literal regexp somewhere
    by doing =C-y=, or the Lisp string by =C-y M-y=.

*** Syntax conversion commands
    =rxt-convert-syntax= (=C-c / c=) converts between Emacs and PCRE
    syntax, depending on the major mode in effect when called.
    Alternatively, you can specify the conversion direction explicitly
    by using either =rxt-pcre-to-elisp= (=C-c / p e=) or
    =rxt-elisp-to-pcre= (=C-c / e p=).

    Similarly, =rxt-convert-to-rx= (=C-c / x=) converts either kind of
    syntax to =rx= form, while =rxt-convert-pcre-to-rx= (=C-c / p x=)
    and =rxt-convert-elisp-to-rx= (=C-c / e x=) convert to =rx= from a
    specified source type.

    In Elisp buffers, you can use =rxt-toggle-elisp-rx= (=C-c / t= or
    =C-c / e t=) to switch the regexp at point back and forth between
    string and =rx= syntax. Point should either be within an =rx= or
    =rx-to-string= form or a string literal for this to work.

*** PCRE mode (experimental)
    If you want to use emulated PCRE regexp syntax in all Emacs
    commands, try =pcre-mode=, which uses Emacs's advice system to
    make all commands that read regexps using the minibuffer use
    emulated PCRE syntax.  It should also work with Isearch.

    This feature is still fairly experimental.  It may fail to work or
    do the wrong thing with certain commands.  Please report bugs.

    =pcre-query-replace-regexp= was originally defined to do
    query-replace using emulated PCRE regexps, and is now made
    somewhat obsolete by =pcre-mode=.  It is bound to =C-c / %= by
    default, by analogy with =M-%=.  Put the following in your
    =.emacs= if you want to use PCRE-style query replacement
    everywhere:

    : (global-set-key [(meta %)] 'pcre-query-replace-regexp)

*** Explain regexps
    When syntax-highlighting isn't enough to untangle some gnarly
    regexp you find in the wild, try the 'explain' commands:
    =rxt-explain= (=C-c / /=), =rxt-explain-pcre= (=C-c / p=) and
    =rxt-explain-elisp= (=C-c / e=). These display the original regexp
    along with its pretty-printed =rx= equivalent in a new buffer.
    Moving point around either in the original regexp or the =rx=
    translation highlights corresponding pieces of syntax, which can
    aid in seeing things like the scope of quantifiers.

    I call them "explain" commands because the =rx= form is close to a
    plain syntax tree, and this plus the wordiness of the operators
    usually helps to clarify what is going on.  People who dislike
    Lisp syntax might disagree with this assessment.

*** Generate all matching strings (productions)
    Occasionally you come across a regexp which is designed to match a
    finite set of strings, e.g. a set of keywords, and it would be
    useful to recover the original set. (In Emacs you can generate
    such regexps using =regexp-opt=). The commands
    =rxt-convert-to-strings= (=C-c / ′=), =rxt-pcre-to-strings= (=C-c
    / p ′=) or =rxt-elisp-to-strings= (=C-c / e ′=) accomplish this by
    generating all the matching strings ("productions") of a regexp.
    (The productions are copied to the kill ring as a Lisp list).

    An example in Lisp code:

   : (regexp-opt '("cat" "caterpillar" "catatonic"))
   :    ;; => "\\(?:cat\\(?:atonic\\|erpillar\\)?\\)"
   : (rxt-elisp-to-strings "\\(?:cat\\(?:atonic\\|erpillar\\)?\\)")
   :     ;; => '("cat" "caterpillar" "catatonic")

    For obvious reasons, these commands only work with regexps that
    don't include any unbounded quantifiers like =+= or =*=. They also
    can't enumerate all the characters that match a named character
    class like =[[:alnum:]]=. In either case they will give a (hopefully
    meaningful) error message. Due to the nature of permutations, it's
    still possible for a finite regexp to generate a huge number of
    productions, which will eat memory and slow down your Emacs. Be
    ready with =C-g= if necessary.

*** RE-Builder support
    The Emacs RE-Builder is a useful visual tool which allows using
    several different built-in syntaxes via =reb-change-syntax= (=C-c
    TAB=). It supports Elisp read and literal syntax and =rx=, but it
    can only convert from the symbolic forms to Elisp, not the other
    way. This package hacks the RE-Builder to also work with emulated
    PCRE syntax, and to convert transparently between Elisp, PCRE and
    rx syntaxes. PCRE mode reads a delimited Perl-like literal of the
    form =/ ... /=, and it should correctly support using the =x= and
    =s= flags.
    
*** Use from Lisp

   Example of using the conversion functions:
   : (rxt-pcre-to-elisp "(abc|def)\\w+\\d+")
   :    ;; => "\\(\\(?:abc\\|def\\)\\)[_[:alnum:]]+[[:digit:]]+"

   All the conversion functions take a single string argument, the
   regexp to translate:

   - =rxt-pcre-to-elisp=   
   - =rxt-pcre-to-rx=
   - =rxt-pcre-to-sre=
   - =rxt-pcre-to-strings=
   - =rxt-elisp-to-pcre=
   - =rxt-elisp-to-rx=
   - =rxt-elisp-to-sre=
   - =rxt-elisp-to-strings=
    
** Bugs and Limitations
*** Limitations on PCRE syntax
    PCRE has a complicated syntax and semantics, only some of which
    can be translated into Elisp. The following subset of PCRE should
    be correctly parsed and converted:
    
    - parenthesis grouping =( .. )=, including shy matches =(?: ... )=
    - backreferences (various syntaxes), but only up to 9 per expression    
    - alternation =|=
    - greedy and non-greedy quantifiers =*=, =*?=, =+=, =+?=, =?= and =??=
      (all of which are the same in Elisp as in PCRE)
    - numerical quantifiers ={M,N}=
    - beginning/end of string =\A=, =\Z=
    - string quoting =\Q .. \E=
    - word boundaries =\b=, =\B= (these are the same in Elisp)
    - single character escapes =\a=, =\c=, =\e=, =\f=, =\n=, =\r=,
      =\t=, =\x=, and =\octal digits= (but see below about non-ASCII
      characters)
    - character classes =[...]= including Posix escapes
    - character classes =\d=, =\D=, =\h=, =\H=, =\s=, =\S=, =\v=, =\V=
      both within character class brackets and outside
    - word and non-word characters =\w= and =\W=
      (Emacs has the same syntax, but its meaning is different)
    - =s= (single line) and =x= (extended syntax) flags, in regexp
      literals, or set within the expression via =(?xs-xs)= or =(?xs-xs:
      .... )= syntax
    - comments =(?# ... )=
    
    Most of the more esoteric PCRE features can't really be supported
    by simple translation to Elisp regexps. These include the
    different lookaround assertions, conditionals, and the
    "backtracking control verbs" =(* ...)= . OTOH, there are a few
    other syntaxes which are currently unsupported and possibly could be:

    - =\L=, =\U=, =\l=, =\u= case modifiers
    - =\g{...}= backreferences
  
*** Other limitations
   - The order of alternatives and characters in char classes
     sometimes gets shifted around, which is annoying.
   - Although the string parser tries to interpret PCRE's octal and
     hexadecimal escapes correctly, there are problems with matching
     8-bit characters that I don't use enough to properly understand,
     e.g.:
     : (string-match-p (rxt-pcre-to-elisp "\\377") "\377") => nil
     A fix for this would be welcome.

   - Most of PCRE's rules for how =^=, =\A=, =$= and =\Z= interact
     with newlines are not implemented, since they seem less relevant
     to Emacs's buffer-oriented rather than line-oriented model.
     However, the different meanings of the =.= metacharacter *are*
     implemented (it matches newlines with the =/s= flag, but not
     otherwise).

   - Not currently namespace clean (both =rxt-= and a couple of
     =pcre-= functions).

*** TODO:
   - Python-specific extensions to PCRE?
   - Language-specific stuff to enable regexp font-locking and
     explaining in different modes. Each language would need two
     functions, which could be kept in an alist:

     1. A function to read PCRE regexps, taking the string syntax into
        account. E.g., Python has single-quoted, double-quoted and raw
        strings, each with different quoting rules.  PHP has the kind
        of belt-and-suspenders solution you would expect: regexps are
        in strings, /and/ you have to include the =/ ...  /=
        delimiters!  Duh.

     2. A function to copy faces back from the parsed string to the
        original buffer text. This has to recognize any escape
        sequences so they can be treated as a single character.


** Internal details
   Internally, =rxt= defines an abstract syntax tree data type for
   regular expressions, parsers for Elisp and PCRE syntax, and
   "unparsers" from to PCRE, rx, and SRE syntax. Converting from a
   parsed syntax tree to Elisp syntax is a two-step process: first
   convert to =rx= form, then let =rx-to-string= do the heavy lifting.
   See =rxt-parse-re=, =rxt-adt->pcre=, =rxt-adt->rx=, and
   =rxt-adt->sre=, and the section beginning "Regexp ADT" in
   pcre2el.el for details.

   This code is partially based on Olin Shivers' reference SRE
   implementation in scsh, although it is simplified in some respects
   and extended in others. See =scsh/re.scm=, =scsh/spencer.scm= and
   =scsh/posixstr.scm= in the =scsh= source tree for details. In
   particular, =pcre2el= steals the idea of an abstract data type for
   regular expressions and the general structure of the string regexp
   parser and unparser. The data types for character sets are extended
   in order to support symbolic translation between character set
   expressions without assuming a small (Latin1) character set. The
   string parser is also extended to parse a bigger variety of
   constructions, including POSIX character classes and various Emacs
   and Perl regexp assertions. Otherwise, only the bare minimum of
   scsh's abstract data type is implemented.

** Soapbox
   Emacs regexps have their annoyances, but it is worth getting used
   to them. The Emacs assertions for word boundaries, symbol
   boundaries, and syntax classes depending on the syntax of the mode
   in effect are especially useful. (PCRE has =\b= for word-boundary,
   but AFAIK it doesn't have separate assertions for beginning-of-word
   and end-of-word). Other things that might be done with huge regexps
   in other languages can be expressed more understandably in Elisp
   using combinations of `save-excursion' with the various searches
   (regexp, literal, skip-syntax-forward, sexp-movement functions,
   etc.).

   There's not much point in using =rxt-pcre-to-elisp= to use PCRE
   notation in a Lisp program you're going to maintain, since you
   still have to double all the backslashes.  Better to just use the
   converted result (or better yet, the =rx= form).

** History and acknowledgments
   This was originally created out of an answer to a stackoverflow
   question:
   http://stackoverflow.com/questions/9118183/elisp-mechanism-for-converting-pcre-regexps-to-emacs-regexps

   Thanks to:

   - Wes Hardaker (hardaker) for the initial inspiration and
     subsequent hacking
   - priyadarshan for requesting RX/SRE support
   - Daniel Colascione (dcolascione) for a patch to support Emacs's
     explicitly-numbered match groups
   - Aaron Meurer (asmeurer) for requesting Isearch support
   - Philippe Vaucher (silex) for a patch to support
     =ibuffer-do-replace-regexp= in PCRE mode