The inspiration comes from Parse::RecDescent's "autotree" directive, which basically constructs a parse tree by collecting submatches into a hash (keyed by submatch name) and blessing that into a package named according to the parent rule, e.g. A ::= B C creates an object of type A with two fields B and C.

I think a much less blessed scheme would be appropriate for the regex engine, in which numbered submatches are collected into match variables @/ and %/. For @/, each capturing group generates a capture:

• If the capturing group is quantified, it is an arrayref of captures, one for each repetition.
• Otherwise, if the capturing group contains no submatches, the capture is a string containing the captured text.
• Otherwise, $/[0] contains the captured text, and @/[1..N] contain the captures for nested groups $1 ... $N, where the indices are numbered starting from the first nested submatch.

Similarly, for %/, each capturing group yields a hash capture:

• $/{0} is the entire text.
• If (?'NAME'...) has no subcaptures, $/{NAME} is the captured text.
• Otherwise if (?'NAME'...) is quantified, $/{NAME} is an arrayref of hash captures, one entry for each repetition.
• Otherwise, $/{NAME} is a hash capture.
• (maybe) $/{1} through $/{N} are like $/{NAME}, but for all captures, not just named ones.

For example, I believe the above rules should yield:

$sexp = qr/
    (?<sexp>
        \s* \( \s* (?&sexp)* \s* \) \s*
      | \s* (?<atom>\w+) \s*
    )/x;

'(A (B C))' =~ /$sexp/;

## AFTERWARDS

%/ = (sexp => {
    0 => '(A (B C))',
    sexp => [{ 0 => 'A', atom => 'A' },
             { 0 => '(B C)',
               sexp => [{ 0 => 'B', atom => 'B' },
                        { 0 => 'C', atom => 'C' }] }]};

@/ = ('(A (B C))',
      [['A', 'A'],
       ['(B C)', [['B', 'B'],
                  ['C', 'C']]]]);
[download]

There are probably some obvious oversights here, but I'll try to get the Regexp::Parser version of this working to shake the bugs out.