Ambivalent Types for Principal Type Inference with GADTs

GADTs, short for Generalized Algebraic DataTypes, which allow constructors of algebraic datatypes to be non-surjective, have many useful applications. However, pattern matching on GADTs introduces local type equality assumptions, which are a source of ambiguities that may destroy principal types— and must be resolved by type annotations. We introduce ambivalent types to tighten the definition of ambiguities and better confine them, so that type inference has principal types, remains monotonic, and requires fewer type annotations.