Of all the problems on 4Clojure, I continue to revisit the challenge of implementing
map. It's such a basic function, yet writing it from scratch has some subtle problems. Let's look at the process in detail.
To begin, here are the assertions that our custom
map function needs to handle:
;; assertion 1 (= [3 4 5 6 7] (my-map inc [2 3 4 5 6])) ;; assertion 2 (= (repeat 10 nil) (my-map (fn [_] nil) (range 10))) ;; assertion 3 (= [1000000 1000001] (->> (my-map inc (range)) (drop (dec 1000000)) (take 2))
For now, we will focus on the first assertion. A naive solution might make use of
(defn my-map [f col] (for [x col] (f x)))
In fact, the implementation above passes all three assertions. To make the problem more interesting, though, 4Clojure disallows the use of
for in a solution. So the question is, how do we implement map without using
Again, let's concern ourselves with just the first assertion. For anyone who has worked through The Little Schemer, the following recursive solution might come to mind:
(defn my-map [f col] (loop [acc  c coll] (if (empty? c) acc (recur (conj acc (f (first c))) (rest c)))))
Using an accumulator (i.e.,
acc), we recursively work our way through a collection, accumulating the values which have been calculated using the mapping function (i.e.,
f). This solves the constraint of avoiding
for, but when we run this implementation as part of the third assertion, we have a performance disaster.
Our implementation of
map will never finish incrementing an infinite range, and so, the only way to remove the performance problem is to start being lazy.
(defn my-map [f col] (if (seq col) (lazy-seq (cons (f (first col)) (my-map f (rest col))))))
By wrapping our recursive call to the map function with a
lazy-seq, we ensure that only the necessary values will be calculated and nothing else. And with the solution here, we have a working implementation of
map which accepts a mapping function and a single collection as an argument.