Below is a refactoring that avoids using stuff from main. Using stuff from main prevents you from precompiling your grammar into a module. Also, using stuff from main makes the script incompatible with mod_perl. (I also lined up the productions.)

use Parse::RecDescent;

my $g = Parse::RecDescent->new(<<'END_GRAMMAR');
  {
     use List::Util 'sum';
     use vars '%weights';

     %weights = qw( C 12 O 16 Pb 207 );
  }

  weight   : compound         { $item[1] }
  compound : group(s)         { sum( @{$item[1]} ) }
  group    : element /\d+/    { $item[1] * $item[2] }
           | element          { $item[1] }
  element  : /[A-Z][a-z]*/    { $weights{ $item[1] } }
           | "(" compound ")" { $item[2] }
END_GRAMMAR

print $g->weight("Pb(CO3)2"), $/;
# prints 327
[download]

I was planning on doing P::RD solution for the OP, but I abandonned the idea when others pointed to existing specialized modules. Thanks for filling in the gap.

Update: The common start of both group productions is very innefficient. Fix:

  weight   : compound         { $item[1] }
  compound : group(s)         { sum( @{$item[1]} ) }
  group    : element factor   { $item[1] * $item[2] }
  factor   : /\d+/            { $item[1] }
           |                  { 1 }
  element  : /[A-Z][a-z]*/    { $weights{ $item[1] } }
           | "(" compound ")" { $item[2] }
[download]