I concur. There is sustained and heavy memory allocation going on in the second loop for which I can see no apparent reason. There is no autovivification of hash elements occuring. You have discovered another memory leak and the culprit is join(*).

If you replace your second loop with

while ( my( $key1, $val1 ) = each %hash1 ) {
    while ( my( $key2, $val2 ) = each %$val1 ) {
        print( ##join( "\t",
            $key1,
            $key2,
            $hash1{ $key1 }{ $key2 },
            $hash2{ $key1 }{ $key2 },
            $hash3{ $key1 }{ $key2 }
        );
    }
}
[download]

The code runs to completion and shows zero memory growth during the second loop.

The problem also exists in 5.6.2. Raise a perlbug.

*Update: It's more complicated than just join. Seems it needs both join and a reference to a compound hash plus some unknown factor to cause the leak.

If you remove the join, or the hash reference, or make either of the keys a constant and the leak does not occur!

while ( my( $key1, $val1 ) = each %hash1 ) {
    while ( my( $key2, $val2 ) = each %$val1 ) {
        die 'Autovivify' unless exists $hash1{ $key1 }
                            and exists $hash1{ $key1 }{ $key2 };
        print( join "\t", $key1, $key2, $hash1{ $key1 }{ $key2 }, );
    }
}
[download]