The Hindley-Milner Algorithm

by Nikita Borisov

My implementation of the Hindley-Milner algorithm was heavily based on the Cardelli paper "Basic Polymorphic Typechecking"; if you find that the code structure is similar to the code in his Appendix, don't be surprised. Having found a bug in the code, I feel reasonably confident that I understand the algorithm well enough to be able to reproduce it, at least to a first order.

The implementation consists of a Perl program that implements type inference on Abstract Syntax Trees. I felt that implementing a parser for the sample language was unnecessary, seeing is it's a task orthogonal to type inference. The language processed by the program is a subset of the one defined in the Cardelli paper. I only support single variable let, and letrec is an explicit statement. Also, the conditional construct is not part of the language, but instead defined using via the cond function. Our belief is that the language has equivalent expressiveness. The following grammar is used:

Exp ::= 
	Identifier |
	Exp Exp | 		[function application]
	fun Identifier Exp |
	let Identifier  = Exp in Exp |
	letrec Identifier  = Exp in Exp |
	( Exp )

The main function, tryexp, takes an AST and prints either the inferred type of the expression, or an error.

Code structure

AST package

The ast package has constructors for AST expressions. It has one constructor for each of the productions in the above grammar; with the exception of the "(" one, since it doesn't correspond to an AST node. It also has a print routine which recursively prints out a fully parenthesized expression corresponding to the AST

Type package

This package has constructors for new types. The basic type classes are variables, and operators (composite types); the latter take zero or more types as arguments. In general, the type system uses the following types:

type ::=
		Identifier | 		[type variable]
		Operator type... |
		Operator |

Commonly defined operators are "->" and "X" (for pair). The atomic types, such as int and bool are represented as operators with no arguments (a la Cardelli).

The new_var procedure defines a new type variable. Due to Perl's "magical" ++ operator, there is little worry of ever running out of names.

A type variable may have an "instance" field, which points to the The print procedure recursively prints a fully parenthesized type expression, following the instance pointers to get to the best currently known type.

The main package

The functions mirror Cardelli's paper. I won't reiterate the comments describing each function. analyze and unify perhaps formulate the interesting component of this work. unify has a "fix" from the Cardelli version: a generic variable won't be unified with a non-generic variable (instead, the reverse unification is performed). This allows us to "taint" the generic variable, and partially implement the rule that:

In unifying a non-generic variable to a term, all the type variables contained in that term become non-generic.

It is my belief that this rule isn't fully implemented even in my code, but the "fix" enforces this rule for all the examples that I've tried (which was not the case in Cardelli's code.

For convenience gettype function "knows" about integer literals and assigns them all the type int.

The main body of code sets up the environment with some basic functions. It then constructs several examples and calls tryexp on them. The reader is encouraged to construct his own examples to verify the behavior of the algorithm.

The code

... is available here

Future work

A fun addition to this algorithm would be to implement polymorphic references using the construct of imperative and applicative type variables. Another important task is to ensure that the non-generic variable rule (above) is actually fully enforced by this program (my current belief is that it still breaks sometimes). Implementing a parser, especially one with some syntactic sugar added, would simplify testing quite a bit.

Things learned

Other than the algorithm, an interesting item of knowledge discerned from my investigation is the dual view of type inference as either a way to solve a system of type constraint equations or proving theorems using derivations defined by the inference system. This connection makes the way the PCC system uses a type system to represent and check proofs more natural.


Valid HTML 4.0!

Nikita Borisov