@data1 = (
           'one',
           'won',
           [
             'two',
             'too',
             'to',
             \@data1
           ]
         );
@data2 = @{$data1[2]};
[download]

Where should Perl get the information that the address of @data2 is the same as $data1[2] ? Interestingly, even specifying Data::Dumper->Purity(1) doesn't fix the problem. It changes the code to

@data1 = (
           'one',
           'won',
           [
             'two',
             'too',
             'to',
             []
           ]
         );
$data1[2][3] = \@data1;
@data2 = @{$data1[2]};
[download]

If you don't use Dump but Dumper , you have to store array references, and everything works correctly:

package Refs {
    my $data1 = [qw[ one won ]];
    my $data2 = [qw[ two too to ]];
    push @$data2, $data1;
    push @$data1, $data2;
    open my $FH, '>', \ our $file;
    local $Data::Dumper::Varname = 'data';
    local $Data::Dumper::Purity = 1;
    print {$FH} Data::Dumper::Dumper($data1, $data2);
    close $FH;

    Testing::recursive_data($data1, $data2);
}

package RefsBack {
    my ($data1, $data2);
    open my $FH, '<', \ $Refs::file;
    eval do { local $/; <$FH> };
    Testing::recursive_data($data1, $data2);
}
[download]

The corresponding code becomes

$data1 = [
           'one',
           'won',
           [
             'two',
             'too',
             'to',
             []
           ]
         ];
$data1->[2][3] = $data1;
$data2 = $data1->[2];
[download]

You might try reporting a bug for the Dump method.

Update: Simulating the correct dumped string:

package Correct {
    my $string = '@data1 = ("one", "won", \@data2);
                  @data2 = ( "two", "too", "to", \@data1);';
    my (@data1, @data2);
    eval $string . ';1' or die $@;
    Testing::recursive_data(\@data1, \@data2);
}
[download]

($q=q:Sq=~/;[c](.)(.)/;chr(-||-|5+lengthSq)`"S|oS2"`map{chr |+ord
}map{substrSq`S_+|`|}3E|-|`7**2-3:)=~y+S|`+$1,++print+eval$q,q,a,
[download]