Thanks for the clarification and the pointers! I will have a look. Regarding building a parse tree with a recursive regexp, I found an inspiring example at ppp from demerphq (which seems a tiny bit hackish though). It uses a different technique with assertions that evaluate Perl code:

use strict;
use warnings;
use Data::Dumper;

$^R=undef;
my $re = qr/
    (?{[$^R]})                      # initialize a new node in the par
+setree
                                    #   preserving the old as the 0th 
+element
    \(                              # match a paren
    (?>                             # atomic match
        ([^()]+)                    # capture
        (?{
            push @{$^R},$1;         # push the capture string into the
+ current node
            $^R                     # propagate the current node
        })
    |
        (?0)                        # recurse
        (?{
            push @{$^R->[0]},$^R;   # push the current node into the o
+ld node
            shift @{$^R}            # and restore the old node to be t
+he current
                                    # (while removing it from the new)
        })
    )*
    \)
/x;

"(a(b)(c(d)e)(f)g)" =~ m/$re/x
           and
print "matched\n"; 
print Dumper $^R;
[download]

which gives (with perl 5.10.0):

matched
$VAR1 = [
          undef,
          'a',
          [
            'b'
          ],
          [
            'c',
            [
              'd'
            ],
            'e'
          ],
          [
            'f'
          ],
          'g'
        ];
[download]

But my current goal is not parsing, it is optimization of stringified output from parsed productions. It would have been cool to do optimizing transforms with recursive pattern substitutions. Then optimizations could be applied on the most nested level first, and propagate to higher levels.