We extend the ML language by alowing a sorted regular equational theory on types for which unification is decidable and unitary. We prove that the extension keeps principal typings and subject reduction. A new set of typing rules is proposed so that type generalization is simpler and more efficient. We consider typing problems as general unification problems, which we solve with a formalism of unificands. Unificands naturally deal with sharing between types and lead to a more efficient type inference algorithm by splitting it into more elementary steps.