Although counting curlies might succeed quite foten, it will fail if you somehwre have a string constant containing a an unbalanced curly or acomment.

public void readExternal () {
    ... '}' /* this } is NOT the end as well */
}
[download]

So it seems you can't get around a basic tokenizer that can differenciate code curlies from string-constant curlies. At least Java syntax is not as hard to parse as Perl, the only way to create a place where a curly does not close a block is in "" and '', after // and between /* and */, there's no qq,q,qr,qw or s///,m//,tr/// ....

A quick solution (demonstrating only basic functionaility):

use Regexp::Common;

local $_ = join '',<>;

my $code = '';
my $curlies = 1; #one curly open

while( m# \G ( [^{}'"/]* ) #xg ){

        $code .= $1;

        my $p = pos;

        # '' + ""
        if( m# \G ( $RE{quoted} ) #xg ){

                $code .= $1;
                next;
        }

        pos = $p;

        # /* */
        if( m# \G ( $RE{balanced}{-begin=>'/*'}{-end=>'*/'} ) #xg ){

                $code .= $1;
                next;
        }

        pos = $p;

        #//

        if( m# \G ( // [^\n]* ) #xg ){

                $code .= $1;
                next;
        }

        pos = $p;

        # {
        if( m# \G \{ #xg ){

                $code .= '{';
                ++ $curlies;
                next;
        }

        pos = $p;

        # }
        if( m# \G \} #xg ){

                $code .= '}';
                -- $curlies;

                last unless $curlies;
                next;
        }

        pos = $p;

        m# \G ( . ) #sxg or last;

        $code .= $1;
}
print "CODE\n$code\n";
[download]

--
http://fruiture.de