* To compile r{i,j} we need a sequence that does not match epsilon
  (or a constructor around an expression telling that this expression
  does not match epsilon)
* A subexpression repeated by an asterisk ( '*' ) or an interval
  expression shall not match a null expression unless this is the only
  match for the repetition or it is necessary to satisfy the exact or
  minimum number of occurrences for the interval expression.
* There might be a typo in deriv_1/delta_1: should we generate 'TMatch
  mark' or 'TMatch mark'? (neither is correct!)

POSIX:
   "(a?)*"   "b"    ""
   "(a?)*"   "ab"   "a"
   "((a)|(b))*"   "ab"  -> "b" none "b"
Str
   "(a?)*"   "b"    no submatch
   "(a?)*"   "ab"   "a"
   "((a)|(b))*"   "ab"  -> "b" "a" "b"
Javascript
   "(a?)*"   "b"    no submatch
   "(a?)*"   "ab"   "a"
   "((a)|(b))*"   "ab"  -> "b" none "b"
PCRE
   "(a?)*"   "b"    ""
   "(a?)*"   "ab"   ""
   "(a?)*?"  "b"    ""
   "(a?)*?"  "ab"   "a"
   "((a)|(b))*"   "ab"  -> "b" "a" "b"
Emacs
   "(a?)*"   "b"    ""
   "(a?)*"   "ab"   ""
   "(a?)*?"   "b"    ""
   "(a?)*?"   "ab"   "a"
   "((a)|(b))*"   "ab"  -> "b" "a" "b"


  r{0,0}  = eps
  r{i+1,j+1} = r,r{i,j}
  r{0,j+1} = r,r{0,j} | eps            PCRE/Emacs
  r{0,j+1} = (r-eps},r{0,j} | eps      JavaScript

* Rewrite sequences of sequences when possible...

High priority
=============
* Improve the Perl regular expressions parser
* Character classes (in the three regular expression parsers)

* Reduce memory usage
  - More compact representation of character sequences
  - Special notation for "anything but this set of characters"
  (more generally, optimize the compilation of regular expressions)
* Simple optimisations
  - alt containing alt
  - epsilon elimination
  - Seq (Seq (x,y), z) => Seq (x, Seq (y, z)) under some circumstances
    (x or y has a fixed length)
  ...

* Test suite

Medium priority
===============
* Implement back-references
* Implement look-ahead and look-behind assertions

Low priority
============
* Optimize the main loop for processor that are not register starved
* Rewrite the main loops in C
  (but keep the option to compile a pure OCaml version)
* Limit the size of the cached DFAs by removing states that have not
  been used recently
* Documentation

Other ideas
===========
* It would be great to have a more generic interface (parameterized
  over some abstract tokens).
* Compile checked printers parameterized over match groups (DRY for
  literal subexpressions)