Veit's Blog

Pattern Matching, A Thing Of The Past


When I started my reading list in January of this year, I decided to give rereading the Lambda Papers a shot. To the uninitiated: in the latter half of the 1970s, Guy Steele and Gerald J. Sussman worked on Scheme. This was an extremely productive time for both of them, and they published a series of now-famous papers collectively dubbed the “Lambda Papers”. They are mostly concerned with Scheme and how it compares to other languages of the time, and neat tricks that they discovered were possible while working on it.

I've reread a good amount of the papers by now, and they are all as excellent as I remembered—maybe even more so now that I've gathered a fair amount of knowledge of what makes a Scheme and how to program in it. But most importantly I've gained an appreciation for sections that went unappreciated when I first read the papers. Today I want to talk about one of them, found in “Scheme, an Interpreter for Extended Lambda Calculus”, pages 10–12. This paper is startling for a bunch of reasons, the most anxiety-inducing of which is that Guy Steele was 21 years old when this paper was published, and it is ground-breaking in more than one respect.

The part I want to talk about is not about one person's gifts, however. I want to talk about an algorithm that implements a very capable pattern matching system in 38—admittedly very dense—lines. I didn't understand it the first time I read the paper, but the second time it completely blew my mind. In this post, I want to attempt to walk you through the original source, which I do not think is still runnable in modern implementations of Scheme. If you want to follow along using a slightly more modernized version, you can download an attempt at a transcript into zepto. Excuse the plug, but I like to put my language to use every once in a while. If you're familiar with Racket and prefer to look at things from a test-driven angle, you might want to check out this repository, where Carl Douglas and I re-explored the algorithm step by step through testing.

Consumer information

Before I explain the algorithm I want to show you the API. It exposes only one function, aptly called match, that is relatively simple and intuitive. Let's first look at the examples from the paper, and then construct our own, to make sure we understood the interface.

This first example is fairly involved and it's okay if it doesn't make sense at first glance–especially since the authors didn't choose a very simple example–, but I'm using it because here it's the only example in the paper.

; from the paper
(match '(A !B ?C ?C !B !E)
       '(A X Y Q Q X Y Z Z X Y Q Q X Y R))
; => (((E (Z Z X Y Q Q X Y R))
;      (C Q)
;      (B X Y))
;     <continuation1>)

; => (((E (R))
;      (C Z)
;      (B (X Y Q Q X Y)))
;      <continuation2>)

; => nil

Fig. 1: A few examples of function invocations and their result.

Let's go through the invocations and the return values one by one.

First, let's observe that the prefixes hold real semantic meaning. The ? character will match any one character, similar to . in regular expressions. The ! character, on the other hand, holds the same meaning as .* in regular expressions, i.e. it will match zero or more of any character. But, unlike in regular expressions, it will match non-greedily. This means that the match expression, transcribed into PCRE, would be A (.*) (.) \2 \1 (.*). Don't worry if that doesn't tell you much yet, I have another example that you can try to wrap your heads around.

It is important to note at this point, however, that these special prefixed symbols are similar to capture groups. These groups will make up the return value of the match function.

Fig. 2: The match groups.

For now, let's try to decypher the return values. It gives us back a two-element list, the first element of which is another list. The second element is a continuation that we'll ignore for now. The nested list represents the matched captured groups and contains pairs of names and the matched values. It will tell us that it bound C to Q, E to a long list, and B to X Y in the first call. Figure 2 tries to glue those back together, which will, unsurprisingly, result in the input list.

input: (A !B ?C ?C !B !E)
substituting B: (A X Y ?C ?C X Y !E)
substituting C: (A X Y Q Q X Y !E)
substituting E: (A X Y Q Q X Y Z Z X Y Q Q X Y R) == input
Fig. 3: Rebuilding the input from the matches.

Now, let's look at that continuation business. The contiunation will actually give us the facilities to try the same algorithm again, and try to match different values. If that's possible, calling the continuation will give us the next possible match. In case of failure it will return nil.

; Exercise: Can you figure out what the function
; should return, given the following inputs?
(match '(A !C B) '(A B C D E X X S B))

(match '(A !C !C ?B !D ?B) '(A F D A D F))

(match '(?X !Y ?Z ?Z !Y) '(A B B C D D B B C))
Fig. 4: An exercise for the reader.

That just about covers the API of the algorithm, and, looking at just that, we could be tempted to assume that the algorithm is long, complex, and daunting to understand—at least that's what I did. It's rather short, though, and completely understandable. Let's try to reimplement it, shall we?

A setup of sorts

Before we start with the actual algorithm, we need a few helper functions. Those are given up front in the paper, and so I shall do the same, because they're generally useful. The helper functions are nfirst and nrest, respectively, which people familiar with functional languages may know as take and drop. Everyone else will be happy to hear that these algorithms, given a list and a number n, just take the first n elements of the input list, or drop them and return the rest. I'll give to alternative implementations of each, one from the paper, and one in zepto.

; Original: nfirst
(define (nfirst e n)
  (if (= n 0)
    (cons (car e) (nfirst (cdr e) (- n 1)))))
; zepto: nfirst
(define nfirst (flip take))

; Original: nrest
(define (nrest e n)
  (if (= n 0) e (nrest (cdr e) (- n 1))))
; zepto: nfirst
(define nrest (flip drop))

Fig. 5: Two implementations of nfirst and ndrop each.

A few notes seem in order at this point: firstly, I'd like to say that I modernized the example given in the paper very slightly. Specifically, I introduced the first-class definition of functions, as opposed to binding a lambda to a name, because I consider it readable. Secondly, the zepto version is cheating a bit by using modern functional primitives to do the dirty work. Those functions are, however, tail-recursive and safe by default, so it is a big win. It's also shorter to use them.

Before moving on, I want to address why more modern languages decided to flip the argument list, moving from having the list as the first argument to having the number first. The issue being addressed here is composability. It turns out that you want to express give me the first 5 elements of a given list more often than give me the first n elements of list [1..n]. We can curry—that is, partially evaluate—functions, and define first5 as (curry take 5). In short, we learned how to be more modular and pattern-oriented, which is what object-oriented functional programming is all about.

Now, depending on which Scheme you are running, you either have reader macros at your disposal or you don't, both of which are fine. The original paper assumes there is a reader macro that transforms symbols with leading ? characters into a list of the form (THV <symbol>) and symbols with leading ! characters into a list of the form (THV* <symbol>). For reasons of compatibility I will instead define a few matching functions, that, in zepto form, look like this:

(define (? sym) (eq? #\? (car (symbol->string sym))))
(define (! sym) (eq? #\! (car (symbol->string sym))))
(define (plain? sym) (not (or (? sym) (! sym))))
(define to-plain (compose string->symbol

Fig. 6: Distinction functions.

The Scheme of your choice should provide functions that are at least equivalent. If you just follow along, you will need to keep in mind that the function plain? matches regular symbols, and the functions ? and ! match symbols that start with these letters, respectively. to-plain creates a plain symbol from a special symbol by stripping away the leading special character.

Now, let's go ahead and implement the actual pattern matching!

The meat of it

Take a deep breath now, 'cause we're diving in heads-first. If you're overwhelmed or feel like you don't understand what's going on, do something else for a bit, come back, reread the bits that were unclear to you and try again. If you don't feel like this post makes any sense to you, don't worry about it. The implementation of the algorithm—simple as the idea itself is—is dense and complex, and might not make sense on the first read-through.

First, let's define a function that shares the API of match as seen above. It will take a pattern and an expression, do some magic, and return a result.

(define (match pattern expression)
  (define (matchfun p e res cont)
    ;do some magic here
  (matchfun pattern expression '() (lambda () nil)))

Fig. 7: A setup for the match function.

You might ask yourself why there is an inner function defined within match, and for good reason. While the gain might not be immediately obvious, for now just think about it like this: it's easier to make everything a tail call if we carry around the state in arguments, while it also makes initializing the state we might need obvious. In this case, we initialize the result list to be an empty list—passed as res—andt the continuation to return as a simple function that returns nil, passed as cont. The latter makes it easy to tell the user there are no more valid matches: remember that when all the possible patterns have been returned, we want to return nil.

Let's begin fleshing out our inner function matchfun. First we want to check whether or not we've consumed all of the patterns.

(define (matchfun p e res cont)
  (if (null? p)
    (if (null? e)
      (list res cont)
    ; more magic...

Fig. 8: Our base case that tells us when to stop matching.

When we have consumed all of the patterns, a valid match requires us to have consumed all of the given expressions as well. If we have, we return a list of our result list res and our continuation cont. If, on the other hand, there are more expressions to match, the match isn't valid and we have to call into our continuation directly, which will either fail and return nil, or backtrack into a previous state—a mechanism we haven't defined yet.

Let's move on to the easy part of our pattern matcher: if we encounter a plain symbol in our pattern list, we just try to match it verbatim.

(define (matchfun p e res cont)
  ; base case goes here
    ((plain? (car p))
        ((null? e) (cont))
        ((eq? (car e) (car p))
          (matchfun (cdr p) (cdr e) res cont))
        (else (cont))))
  ; other cases here

Fig. 9: The plain case.

We use the function plain? we defined in the previous section to see whether the symbol is in fact a plain old symbol. If it is, we have to check whether we have anything to match against—and if not call the continuation—, and if we do call matchfun again with the rest of the lists, passing the other two arguments merrily along. If we don't, we defer to the continuation again. From now on, I will stop highlighting when we call the continuation. As a general rule we'll always do that if there is no way to progress in our current state and we have to yield.

It's time to move on to the second least-complicated case, matching one of any character.

(define (matchfun p e res cont)
  ; base case
    ; literal case
    ((? (car p))
      (if (null? e)
        ((lambda (v)
          (if (eq? v #f)
            (matchfun (cdr p) (cdr e)
                      (cons (list (to-plain (car p))
                                  (car e))
            (if (eq? (car e) (cadr v))
              (matchfun (cdr p) (cdr e) res cont)
          (assq (to-plain (car p)) res))))
    ; more cases

Fig. 10: Matching any one character.

Okay, this is an order of magnitude more difficult to understand than the plain case. Let me explain what we had to do, and then associate that with how we actually implemented that behavior. If we find a symbol p that matches any one character, we see whether there is anything left to match. If there is, we have to check whether the match will align with our previous matches, i.e. whether the p we matched previously was matched to the same symbol. If there was no previous occurrence of p, we just record the current match and try to match the rest with a recursive call. If it was there but did not resolve to the same match, we're trapped in a dead end and have to yield. If it was there and matches our previous occurrences, we just move along and don't need to record any new symbol.

How do we actually implement this in the above code? The answer is sadly not as elegant as I would like it to be, because regular Scheme—especially not the first iteration of it—did not have hash maps or dictionaries or any fast key-value data structure. Instead, the above code is implemented in terms of association lists. Association lists are basically a poor man's implementation of a key-value data structure, with O(n) time for everything except insertion. It's nothing more than a list of pairs of the structure (key value) with a fancy name.

I would love to implement this in terms of a faster key-value data structure, but sadly there is no big unifying structure that I could use—zepto does have hash maps, even hash map literals, but that's beside the point if we want to stick to a format that every Scheme implementation can run.

If you already know Scheme's API for association lists, then the above code snippet should be fairly straight-forward. If you do not, maybe pointing out that assq is a function that takes in a key and an association list and returns the key value pair will help. In the code above, this pair is handed to a lambda where it is bound to the name v. We then check whether v is truthy and either insert it and move on or try to match the current value to the old value. All in all, this is a big hack around a clunky data structure which happens to work quite well.

Now we've arrived at the last part of our matcher, the !. This means we need to match zero or more characters.

(define (matchfun p e res cont)
  ; base case
    ; literal and ? cases
    ((! (car p))
      ((lambda (v)
        (if (eq? v #f)
          (letrec ((match* (lambda (n)
                  (if (> n (length e))
                    (matchfun (cdr p) (nrest e n)
                              (cons (list (car p)
                                          (nfirst e n))
                              (lambda ()
                                 (match* (+ n 1))))))))
            (match* 0))
          (if (< (length e) (length (cadr v)))
            (if (eq? (nfirst e (length (cadr v)))
                     (cadr v))
              (matchfun (cdr p)
                        (nrest e (length (cadr v)))
                        res cont)
        (assq (car p) res)))
    ; more cases

Fig. 11: Matching any zero or more characters.

The structure of the case is similar to the one we just examined, with a few important differences. We don't need to check whether or not the list is empty, because that would be a valid match, so we move into the closure immediately to find out if we've already encountered this symbol. If we have, we simply check whether the two lists are equal, and then either progress or yield.

The really interesting bit is what happens if we encounter the symbol for the first time, in which case we define a function called match* that will progressively try to call the match function, at first with an empty list bound to our symbol. It will also set cont to a version of itself with a longer list. If we ponder that for a while we see that this means it will try to go on with a minimal list and then come back and grow the list if we can't proceed otherwise. This is a fairly simple idea, although the implementation tends to disagree and admittedly is a little hard to read.

We're almost done, all we have to do is add another case to our top-level expression that will call cont if none of the cases we handled makes sense. This should never happen provided the user hands us valid arguments, but we shouldn't rely on that.

(define (matchfun p e res cont)
  ; base case
    ; regular matchers
    (else (cont))))

Fig. 12: A catch-all matcher.

This is it! We've just built a pattern matching system with back-tracking in Scheme.

What remains to be done

The code I presented above is dense and very conservative. The Scheme or Lisp of your choice will probably provide many more advanced features that are suited to simplify the code—hash maps come to mind, for which even a SRFI with reference implementations exists. I limited myself to pure Scheme as specified in the R5RS standard, though—except for the helper functions, where I took some liberties. It would be a worthwile exercise to transcribe the algorithm into a more modern version of Scheme. I'm positive it would make the code more readable, and quite certainly more performant.

Other worthwile exercises I could think of:

There's plenty of work that could be done to improve this algorithm, but I believe that the fundamental simplicity of the idea is what makes the algorithm appealing to begin with. Thus any kind of work should strive for the same conceptual simplicity instead of just tacking on features blindly, at least in my book.


When I first understood this algorithm, I was stunned. Surely it can't be that simple, but somehow it is. It took me by surprise because it is not the paper's main event, rather, it was meant to be an illustration of Scheme as a language. I'm not sure if anyone ever took it at face value and tried to play around with it more, whether it had any impact on the implementation of logic or functional programming languages, or if everyone silently agreed that it was just a toy. Prolog had been around for three years at that point, providing a capable unification and pattern matching engine already. I've never ventured to understand the implementation of logic programming, but having seen the algorithm I presented today—and having enjoyed it thoroughly—I believe I might appreciate looking at their underpinning.

This has been my longest blog post to date by far. I hope you enjoyed reading it as much as I enjoyed writing it. I tried to make the post accessible to a lot of the programming world, though I understand both Lisp and the algorithm at hand might be confusing for a non-negligible percentage of programmers. If you have any questions, comments, or criticism, please reach out to me via e-mail or any medium you can get a hold of me in.

Lastly, I want to thank all of the awesome people who reviewed my blog post and without whose feedback this would've been much more of a mess! You know who you are and what you did and I'm really grateful for all of it.

See you next time!