Computing a hash on 1MB of data is slower than computing a hash of 6-8 bytes of data. Brute-force attacks are based on trying different passwords and seeing that after being salted they generate the same hash as in the database. Therefore, adding a large string to the password before hashing would force the attacker to hash that string. The question is, can this be pre-computed once or efficiently parallelized?
You're advocating creating a 1MB "salt" string to slow down hashes? That's the same as simply iterating your hash function enough times to invoke the block function repeatedly.
Just use bcrypt, scrypt, or PBKDF2. People have already figured this problem out.
First, I do not advocate anything here. I asked a question.
Second, working with a large string of bits is the same as recursive hashing only if you can pre-compute some small intermediate state of the hash function for that string independently from the password you're trying to guess. If you can't, you would have to work with the entire string for every new password tried.
