After tinkering about for a while trying to make a pattern match produce sensible results, I stepped back and tried to think of the problem at core. What you really want to find is a "longest common subsequence" between the strings. There's a very good module on CPAN which does just this - Algorith::Diff.

#!/usr/bin/perl -w
use strict;

use Algorithm::Diff qw(traverse_sequences);

# construct arrayrefs containing an array of single chars
my ($A, $B) = map [/(.)/sg], qw(
    ATGGAGTCGACGAATTTGAAGAAT
    xxxxxxATGGAGyxxxTCGAzxxxxCGAATTTGAAxxwGAAT
);

my $prev = '';
my @seq;

traverse_sequences(
    $A, $B,
    {
        MATCH => sub {
            my ($aidx, $bidx) = @_;
            if('=' ne $prev) {
                push @seq, '';
                $prev = '=';
            }
            $seq[-1] .= $A->[$aidx];
        },
        DISCARD_A => sub { die "Sequence in A is not fully contained i
+n B" },
        DISCARD_B  => sub {
            my ($aidx, $bidx) = @_;
            if('!' ne $prev) {
                push @seq, '';
                $prev = '!';
            }
            $seq[-1] .= $B->[$bidx];
        },
    },
);

print "@seq\n";

__END__
xxxxxx ATGGAG yxxx TCGA zxxxx CGAATTTGAA xxw GAAT
[download]

Even this falls short on "actual" data though: working against GATAGCATGGAGGCCATCGATAACGCGAATTTGAATTTGAAT it produces G AT A G CAT G G AG GCCA TCGA TAA CG CG AATTTGAA TTT GAAT..

Makeshifts last the longest.