We have covered some of the reasons why automatic annotations are useful in my previous post, now let’s build up the infrastructure necessary to perform it from scratch.

This work is part of a crowdfunding effort, please support the campaign here!

At the center of our approach is the track function. It wraps values and remembers a ‘path’ to report the source of the value.

How is track defined? Let’s add support for flat values.

Inferring flat values

Here’s how you might define track to remember the type of integers.

(defn track [v path]
  (cond
    (integer? v) (do (type-at-path (-class (class v)) path)
                     v)
    :else v))

-class creates a core.typed type of a given class; type-at-path associates the given type with the current path.

A ‘path’ is a sequence of ‘path elements’. Our first path element is the ‘var path element’. To track the value of a def, we rewrite it with the following transformation.

The we track a def by tracking its right-hand side as a singleton path containing just the var name.

Recall from the previous post, we are interested in two kind of annotations: user-level vars and libraries.

We can track user-level vars by directly transforming the def expressions before they are compiled. To intercept library functions, we need to wrap each library var dereference in a track.

Here’s how the evaluator step through an evaluation of tracking a def.

1. The initial def is what the programmer actually writes.
2. We rewrite the plain def into a track that remembers the value as coming from forty-two.
3. We remember that forty-two is of type Long.
4. Using this type environment, we can then output the following correct annotation:

(ann forty-two Long)

Extensions

From here, we have many options to extend our algorithm. In the following posts, we’ll talk about:

• inferring function types,
• inferring sequence types,
• different merging strategies for types,

The basics, however, remain the same, so this help should help as a reference when we discuss these extensions.

This work is part of a crowdfunding effort, please support the campaign here!