Category Archives: Languages

Programming Languages


The Y-Combinator

I’ve struggled a bit in the past to explain why letrec was necessary to allow recursion in a language with first class functions. All we’re trying to achieve is:

But without the use of a global subroutine name, or in fact any environment assignments. If you remember, letrec created a recursive function by creating a symbol naming the function first, with a dummy value,  then evaluated the function in the environment where it’s name was already present, then assigned the resulting closure to the symbol so the function could “see itself”. But in a purely functional setting, assignment is bad, right?

There is a little bit of programming language magic called the “Y-Combinator” that does the job. It’s very succinctly expressed in the λ calculus as:

That is to say, a function taking a function as argument  applying that function to itself, and given (a copy of) itself as argument.

In case this seems all a bit too esoteric, here it is in F♮:

And if that’s still too esoteric here it is in Perl:

Notice that we haven’t named any subroutine, so on the face of it recursion is impossible, but nonetheless, if you give the above code to perl it will very slowly rattle your discs until an out of memory exception, without even a deep recursion error because there’s no function name for perl to attribute the recursion to.

Beore going any further I should point out that none of this is of any value to you whatsoever, other than to assuage your curiosity. Most all modern languages allow recursion, if not support or encourage it (supporting as opposed to just allowing recursion is a fine but important point: scheme supports recursion, Perl and its ilk merely allow it.) Anyway we can use the Y-combinator to calculate a factorial:

Once the inner sub has got hold of itself in  $factorial  it can call  $factorial  as a subref. The outer anonymous sub bootstraps the whole thing by:

  1. Capturing the inner sub in its $factorial
  2. Both calling  $factorial  and passing  $factorial  to it
  3. Passing an extra argument, 5, the number we require the factorial of.

Algebraic Data Types and Pattern Matching

What may not be clear to readers in a lot of the previous discussions is the use of Algebraic Data Types in combination with patterm matching to define functions. It’s really quite simple, conceptually (implementation may be a different matter, we’ll see.) Here’s an example we’ve seen before, I’ll just be more descriptive this time:

This declaration achieves two things:

  1. It defines a type  list(t)  (list of t) where  t is a type variable that can stand for any type.
  2. It creates two constructor functions, called  cons and null, that accept arguments of the specified types (none in the case of null,) and return data of type list(t).

Reading it aloud, it says define a type list of some unspecified type t which is either a cons of a  t and a  list of t, or a null.

Once defined, we use these type costructors to create lists of a concrete type:

After the above definition, a has type list(bool). The following, on the other hand, would fail to type check:

It fails because:

  • cons('x', null)  is of type list(char) .
  • The outer cons expects arguments  <t>  and list(<t>) , but it gets  bool  and list(char) .
  • The outer cons cannot reconcile  <t> = bool  with  <t> = char  so the type check fails.

That’s all very nice, but how can we use Algeraic Data Types? It turns out that they become very useful in combination with pattern matching in case statements. Consider:

In that case statement, a must match either  cons(head, tail)  or null. Now if it matches cons(head, tail), the (normal) variables  head and  tail are automatically created and instantiated as the relevant components of the  cons in the body of the case statement. This kind of behaviour is so commonplace in languages like ML that special syntax for functions has evolved, which I’m borrowing for F♮:

This version of length, instead of having a single formal argument list outside the body, has alternative formal argument lists inside the body, with mini bodies of their own, just like a case statement. It’s functionally identical to the previous version, but a good deal more concise and readable.

One thing to bear in mind, in both versions, is that  length  has type list(t) int. That is to say, each of the formal argument lists inside the body of a function, or the alternative cases in a case statement, must agree in the number and types of the arguments, and must return the same type of result.

Now, it becomes obvious that, just as we can rewrite a  let to be a lambda, this case statement is in fact just syntactic sugar for an anonymous function call. The earlier definition of  length  above, using a case statement, can be re-written as:

so we get case statements: powerful, pattern matching ones, allowing more than one argument, for free if we take this approach.

length is polymorphic. It does not do anything to the value of head so does not care about its type. Therefore the type of length, namely  list(t) int actually contains a type variable t.

Here’s a function that does care about the type of the list:

Assuming strlen has type string int, that would constrain  sum_strlen to have type list(string) int. Of course that’s a rather silly function, we would be better passing in a function like this:

That would give sum a type:

and we could call it like:

or even, with a Curried application:

This is starting to look like map-reduce. More on that later.

Real-World Applications

Algebraic Data Types really come in to their own when it comes to tree walking. Consider the following definitions:

Given that, we can write an evaluator for arithmetic expressions very easily:

So eval has type expr(int) int . We can call it like:

to get 17.

Pattern matching not only covers variables and type constructors, it can also cope with constants. For example here’s a definition offactorial:

For this and other examples to work, the cases must be checked in order and the first case that matches is selected. so the argument to  factorial  would only match  n  if it failed to match .

As another example, here’s member:

Here I’m using F♮’s built-in list type constructors @, (pronounced cons,) and  [] (pronounced null,) and a wildcard  _ to indicate a don’t care variable that always unifies, but apart from that it’s just the same as the  cons and  null constructors. Anyway, the cases say:

  • member(item, list)  is  true if  item is at the head of the list.
  • member(item, list) is  true if item is a member of the tail of the list.
  • item is not a member of the empty list.

Problems and Solutions

You’ve probably realised that given a type like  list(t) above, it’s not possible to directly create lists of mixed type. That is because it is usually a very bad idea to do so. However if you need to do so, you can get around the restriction without breaking any rules, as follows:

  1. Create a container type for your mixed types:
  2. Create lists of that type:

After the above definition, a has type list(either(string, int)), and you can’t get at the data without knowing its type:

Here,  sum_numbers has type [either(<t>, int)] int. e.g. it doesn’t care what type  first holds. We could have written  first(s) instead of first(_), but the use of a wildcard  _explicitly says we don’t care, stops any potential warnings about unused variables, and is more efficient.


Evaluating Partial Function Application

I’ve mentioned Currying and partial function application already. The idea is that given a function with more than one argument:

if we call it with less than the arguments it expects, then it will return a function that accepts the rest:

(The trailing comma is just a syntactic convention that I’ve come up with that lets the compiler know that we know what we are doing, and lets the reader know that there is Currying going on.) Now setting aside how we might type-check that, it turns out that it’s actually pretty easy to evaluate.

Normal application of a closure looks something like this (Java-ish pseudocode):

For those of you that don’t know Java, List<Symbol> means List of Symbol. And yes, we’re ignoring the possibility that we’re passed the wrong number of arguments, the type checker should deal with that.

Now if we are expecting that we might get fewer than the full set of arguments, we can instead create a new closure that expects the rest:

Note that the dictionary that we have been building is used to extend the environment of the new closure with the values we know already, and that the formal_args we’ve been chaining down is now precisely the remaining arguments that we haven’t seen yet.

Of course this allows for silly definitions like:

But presumably our type checker (if not our grammar) would disallow that sort of thing, because there’s nothing to put a trailing comma after.

[Edit] You could alternatively add a guard clause to  apply() that says if this closure is expecting arguments and doesn’t get any, just return the original closure. That way, something like:

while still silly, would at least not be unnecessarily inefficient.

Addendum – over-complete function application

So I got the above working in F♮ easily enough, then I noticed an anomaly. The type of:

is definately int → int → int, which means that the type checker is allowing it to be called like:  adder(2, 3). Why can’t I call it like that? It turns out I can:

Assuming the type checker has done its job, then if we have any actual arguments left over then they must be destined for the function that must be the result of evaluating the body. So instead of just evaluating the body in the new env, we additionally call  apply()  on the result, passing in the left-over arguments.

This is pretty cool. We can have:

and call it like  adder(2, 3) or  adder(2)(3), and we can have:

and call it like  add(2, 3) or  add(2)(3).

One or the Other, or Both?

The question arises: if we have implicit Currying, (partial function application) then do we need explicit Currying (explicitly returning a function from a function)? The answer is a resounding yes! Consider:

We’ve only called  bigfn once, when evaluating the first argument to map, so expensive_calculation only got called once, and the explicit closure calling either cheap_op_1 or  cheap_op_2 gets called on each element of the list.

If instead we had written:

Then the call to  expensive_calculation would get deferred until the  map actually called its argument function, repeatedly, for each element of the  long_list.


The Hindley-Milner Algorithm

The Hindley-Milner Algorithm is an algorithm for determining the types of expressions. Basically it’s a formalisation of this earlier post. There’s an article on Wikipedia which is frustratingly terse and mathematical. This is my attempt to explain some of that article to myself, and to anyone else who may be interested.


The Hindley-Milner algorithm is concerned with type checking the lambda calculus, not any arbitrary programming language. However most (all?) programming language constructs can be transformed into lambda calculus. For example the lambda calculus only allows variables as formal arguments to functions, but the declaration of a temp variable:

can be replaced by an anonymous function call with argument:

Similarily the lambda calculus only treats on functions of one argument, but a function of more than one argument can be curried, etc.


We start by defining the expressions (e) we will be type-checking:

e::=EA primitive expression, i.e. 3.
|sA symbol.
|λs.eA function definition. s is the formal argument symbol and e is the function body (expression).
|(e e)The application of an expression to an expression (a function call).


Next we define our types (τ):

τ::=TA primitive type, i.e. int.
|τ0 → τ1A function of one argument taking type τ0 and returning type τ1


We need a function:

[1]f(ε, e)=τ


εA type environment.
eAn expression.
τA type


We assume we already have:

[2]f(ε, E)=TA set of mappings from primitive expressions to their primitive types (from 3 to int, for example.)

The following equations are logic equations. They are easy enough to read, Everything above the line are assumptions. The statement below the line should follow if the assumptions are true.

Our second assumption is:

[3](s, τ)εIf (s, τ) is in ε (i.e. if ε has a mapping from s to τ)
f(ε, s)=τThen in the context of ε, s is a τ

Informally symbols are looked up in the type environment.


[4]f(ε, g)=τ1 → τIf g is a function mapping a τ1 to a τ
f(ε, e)=τ1and e is a τ1
f(ε, (g e))=τThen the application of g to e is a τ

That is just common sense.

[5]ε1=ε ∪ (s, τ)If ε1 is ε extended by (s, τ), e.g. if s is a τ
f(ε, λs.e)=τ → f(ε1, e)Then the output type of a function with argument s of type τ, and body e, is the type of the body e in the context of ε1

This is just a bit tricky. We don’t necessarily know the value of τ when evaluating this expression, but that’s what logic variables are for.


  • We extend the set T of primitive types with an infinite set of type variables α1, α2 etc.
  • We have a function new which returns a fresh type variable each time it is called.
  • We have a function eq which unifies two equations.

We modify our function, part [4] (function application) as follows:

[6]τ0=f(ε, e0)If τ0 is the type of e0
τ1=f(ε, e1)and τ1 is the type of e1
τ=newand τ is a fresh type variable
f(ε, (e0 e1))=eq(τ0, τ1 → τ); τThen after unifying τ0 with τ1 → τ, the type of (e0 e1) is τ.

That deserves a bit of discussion. We know e0 is a function, so it must have a type τa → τb for some types τa and τb. We calculate τ0 as the provisional type of e0 and τ1 as the type of e1, then create a new type variable τ to hold the type of (e0 e1).


Suppose e0 is the function length (the length of a list of some unspecified type τ2), then τ0 should come out as [τ2] → int (using [x] to mean list of x.)

Suppose further that τ1 is char.

We therefore unify:


Which correctly infers that the type of (length [‘c’]) is int. Unfortunately, in doing so, we permanently unify τ2 with char, forcing length to have permanent type [char] → int so this algorithm does not cope with polymorphic functions such as length.


Types, Type Checking, Type Variables and Type Environments

This was the bit of Comp. Sci. I always thought looked uninteresting, but in fact when you delve in to it it’s really fascinating and dynamic. What we’re actually talking about here is implicit, strong type checking. Implicit means there is not (usually) any need to declare the type of a variable or function, and strong means that there is no possibility of a run-time type error (so there is no need for run-time type checking.)

Take a look at the following code snippet:

You and I can both infer a lot of information about doublexy and z from that piece of code, if we just assume the normal meaning for +. If we assume that + is an operator on two integers, then we can infer that x is an integer, and therefore the argument to doublemust be an integer, and therefore y must be an integer. Likewise since + returns an integer, double must return an integer, and therefore z must be an integer (in most languages + etc. are overloaded to operate on either ints or floats, but we’ll ignore that for now.)

Before diving in to how we might implement our intuition, we need a formal way of describing types. For simple types like integers we can just say int, but functions and operators are just a bit more tricky. All we’re interested in are the argument types and the return type, so we can describe the type of + as:

I’m flying in the face of convention here, as most text books would write that as (int * int) → int. No, that * isn’t a typo, it is meant to be some cartesian operator for tuples of types, but I think it’s just confusing so I’ll stick with commas.

To pursue a more complex example, let’s take that adder function from a previous post:

So adder is a function that takes an integer x and returns another function that takes an integer y and adds x to it, returning the result. We can infer that x and y are integers because of + just like above. We’re only interested for the moment in the formal type ofadder, which we can write as:

We’ll adopt the convention that → is right associative, so we don’t need parentheses.

Now for something much more tricky, the famous map function. Here it is again in F♮:

map takes a function and a list, applies the function to each element of the list, and returns a list of the results. Let’s assume as an example, that the function being passed to map is some sort of strlenstrlen‘s type is obviously:

so we can infer that in this case the argument list given to map must be a list of string, and that map must return a list of int:

(using [x] as shorthand for list of x). But what about mapping square over a list of int? In that case map would seem to have a different type signature:

In fact, map doesn’t care that much about the types of its argument function and list, or the type of its return list, as long as the function and the lists themselves agree. map is said to be polymorphic. To properly describe the type of map we need to introducetype variables which can stand for some unspecified type. Then we can describe map verbally as taking as arguments a function that takes some type a and produces some type b, and a list of a, producing a list of b. Formally this can be written:

where <a> and <b> are type variables.

So, armed with our formalisms, how do we go about type checking the original example:

Part of the trick is to emulate evaluation of the code, but only a static evaluation (we don’t follow function calls). Assume that all we know initially is the type of +. We set up a global type environment, completely analogous to a normal interpreter environment, but mapping symbols to their types rather than to their values. So our type environment would look like:

On seeing the function declaration, before we even begin to inspect the function body, we can add another entry to our global environment, analogous to the def of double (we do this first in case the function is recursive):

Note that we are using type variables already, to stand for types we don’t know yet. Now the second part of the trick is that these type variables are actually logic variables that can unify with other data.

As we descend into the body of the function, we do something else analogous to evaluation: we extend the environment with a binding for x. But what do we bind x to? Well, we don’t know the value of x, but we do have a placeholder fot its type, namely the type variable <a>. We have a tiny choice to make here. Either we bind x to a new type variable and then unify that type variable with <a>, or we bind x directly to <a>. Since unifying two unassigned logic variables makes them the same logic variable, the outcome is the same:

With this extended type environment we descend into the body and come across the application of + to x and x.

Pursuing the analogy with evaluation further, we evaluate the symbol x to get <a>. We know also that all operations return a value, so we can create another type variable <c> and build the structure (<a>, <a>) → <c>. We can now look up the type of + andunify the two types:

In case you’re not familiar with unification, we’ll step through it. Firstly <a> gets the value int:

Next, because <a> is int, the second comparison succeeds:

Finally, <c> is also unified with int:

So <a> has taken on the value (and is now indistinguishable from) int. This means that our environment has changed:

Now we know <c> (now int) is the result type of double, so on leaving double we unify that with <b>, and discard the extended environment. Our global environment now contains:

We have inferred the type of double!

Proceeding, we next encounter def y = 10;. That rather uninterestingly extends our global environment to:

Lastly we see the line def z = double(y);. Because of the def we immediately extend our environment with a binding of z to a new placeholder <d>:

We see the form of a function application, so we look up the value of the arguments and result and create the structure:

Next we look up the value of double and unify the two:

<d> gets the value int and our job is done, the code type checks successfully.

What if the types were wrong? suppose the code had said def y = "hello"? That would have resulted in the attempted unification:

That unification would fail and we would report a type error, without having even run the program!


Preliminary Sketch for F Natural

I’ve gone back to F♮, but decided to abandon the Perl implementation in favour of Java. I might be able to target the Java VM, but I’m not sure yet. In any case it’s a good chance to brush up my Java and learn IntelliJ IDEA (CE). I’m using SableCC to parse, and I’ve just got Log4j working. I’m long overdue writing some unit tests but for the moment, after a week or so, I have a working interpreter and the following sample script that runs as advertised:

Apart from the env directive, and the fact that strings are lists of chars, this is still very much a scheme interpreter with a more syntax laden parser in front of it, but it’s early days. Next steps:

  • Get unit tests in place. I’ve delayed too long.
  • Implement an implicit strong type-checking visitor (I’m falling out of love with the Visitor pattern, but SableCC gives me no choice.)
  • Replace the variable implementation with the same logic variables used by the type checker.
  • Add algebraic data types a la ML. This should look like:

    (except that List is already predefined.) t is a type variable, so this says List of some type t is either a Pair of a t and a List of t, or it is Null. Pair and Null become type constructors for type List(t), so Pair('c', Null) creates a list of length 1 with type List(char). Like I said, we already have lists, and h @ t is (cons h t), e.g. Pair(h, t).
  • Extend the function definition to allow pattern matching (actually unification). This would look like:

    so the formal arguments to the function can optionally be moved inside the function body, and repeated with alternative formats, like a case statement. The format that unifies with the actual arguments has its body evaluated with the relevant variables bound.
  • Rewrite to CPS, add failure continuations, implement amb as a binary operator then:
  • Implement fail as a keyword of no arguments:

    (That last one might give the type checker a headache.)



Garbage Collection

One of the defining features of a high-level language like Scheme, Perl, PHP or almost any other recent language is that they have built-in garbage collection (GC), which makes the programmers life much easier because they don’t have to worry about memory management too much. However there are different GC strategies, and costs and benefits depending on the choices you make. I’d like to talk a little about the common GC strategies here, so you can see what the trade-offs are.

Reference Counting Garbage Collection

Unfortunately for both Perl and PHP they made a bad choice of garbage collector; they both opted for the simplest possible mechanism, that is reference counting. Why it is a bad choice I hope to demonstrate.

Reference counting is simple. Each object being memory-managed has a reference-count field which is incremented whenever a new variable refers to the object, and decremented when a variable stops referring to that object. If the reference count reaches zero, nothing can be referring to that object and it can be immediately returned to the free store, decrementing the reference counts of any objects it may contain in turn.

Now this is initially quite attractive: one of the very attractive features, beyond its simplicity, is that garbage collection happens in tiny little steps, and does not usually interfere with the flow of the program, in contrast to Mark-and-Sweep GC discussed below. The real problem with it is that it doesn’t work, at least not in all cases.

The cases where it doesn’t work are where there are circular references between structures. If we imagine a situation where A refers to B, B to C and C back to A, and we have an additional reference X also referring to A, and nothing else, then A will have a reference count of 2 (one from X, one from C), and B and C a reference count of 1. Now when X stops referring to A, A’s reference count will be decremented to 1, but now that entire cyclic structure of A, B and C is cut adrift: since nothing external refers to it, no reference counts can ever be decremented again, and A B and C all still have reference counts of 1.

There are some fancy tricks in Perl to try to work around the problem, specifically weak references which are references that do not disturb the reference counts, but they are probably more difficult to reason about than simple memory management in low-level languages and so are not a valid solution. Another solution is to have a “guard” object acting as a wrapper around the cyclic structure. If the guard object’s reverence count goes to zero its DESTROY method will be invoked, and it then explicitly tells the cyclic structure to remove the circular references so that the components can be recovered. Again this is making memory management the programmers concern.

Mark and Sweep Garbage Collection

This is a strategy that at least works in all cases, but has significant drawbacks. With mark and sweep, a “Free List” of objects is maintained that can be used for their space. A normal request for memory will retrieve an object from the free list and put it on to an “in Use” list. If there are no objects available on the free list, then Mark and Sweep Garbage collection is invoked. Mark and Sweep, as its name suggests, proceeds in two stages. The first is to follow all pointers from the current execution context, recursively, and to “mark” all objects found as “in use”. The next phase, the “Sweep” phase, traverses the “In Use” list and moves any objects not marked as in use to the free list (also resetting the “mark” on all objects).

The problem is that in a typical application the majority of the objects on the “In Use” list will not be actually in use, and the sweep phase will use a lot of resources moving unused objects back to the free list. This can produce a pronounced “stutter” in interactive programms, where they appear to hang for seconds at a time. This behaviour was a common failing of early Lisp implementations, and one solution there was to provide a (gc) command that the programmer could sprinkle around the code in order that the number of unused objects on the used list never got too big. Again this is passing memory management back to the programmer, albeit at a higher level.

Copying Garbage Collection

The realisation that most supposedly in-use objects actually are not in use is a clue to a more time efficient (but not space-efficient) garbage collection strategy.

To get it to work we have to drop the high-level notion of lists, and instead get down to the machine level and consider free pools and in-use pools, where a pool is just a contiguous region of memory on the heap.

Basically, instead of moving everything that isn’t in use from the in-use pool to the free pool, we move what is in use, as we find it during the mark phase, to the “free pool” then we swap the ponters to the two pools: so the free pool becomes the in-use pool and vice versa. We know that the free pool was initially empty otherwise garbage collection would not have been invoked. Of course that means we can no longer have an explicit (gc) command, but that’s not a bad thing. It also means we can dispose of the Sweep phase altogether.

The remaining details are fairly simple: when moving an object from the in-use pool to the free pool, we must leave a note behind to say that it has been moved, and where it has been moved to. We replace the old copy of the object with a “forwarding pointer” with a flag to say that it has been forwarded. That way any further encounters of the object during garbage collection merely update their pointers to refer to the new copy and then stop, because the object and its contents have already been moved.

Generational Garbage Collection

Another observation about the memory usage of typical applications provides a more time efficient variation on Copying Garbage Collection. That observation is “the longer an object has persisted, the longer it is likely to persist”.

To leverage that observation, we divide our heaps into “generations”. The most recent generation being the objects that have not yet been through a garbage collection. Generational Garbage collection inspects the most recent generation of in-use objects first, and if it can reclaim enough space there its job is done. It moves objects still in use to the next generation, and so on (the choice of the number of generations is presumably configurable, and the oldest generation falls back to standard Copying Garbage collection behaviour.)

More details needed, later…


A Bit on the Lambda Calculus

In my book I briefly mention (in a footnote I think) the lambda calculus. It’s probably time to expand on that. As I said, the lambda calculus is a mathematical discipline that predates computers, but is amazingly relevant to computer language design. Basically it assumes that there are nothing but functions, each taking only one argument, and it derives all possible computations from them.

I won’t bother describing the specific syntax right now, rather let’s take a look at a translation of some of it into Perl.

 First off, functions can only take one argument. So how do we imitate functions, such as addition, that require more than one argument? The answer is a technique called “Currying” (after the mathematician Haskell Curry who gave the better part of his name to a programming language.) To summarise, a “Curried” function of two arguments is a function that takes one argument and returns a function that takes a second argument and returns the result. So for example Currying a perl function add:

(using anonymous subs throughout for consistency) produces:

which you can call like:

to get 7.

More interestingly, we can represent truth and falsehood as functions. Assuming implicit Currying from now on, true can be represented as a function that takes two arguments and returns its first one:

and false similarily as a function that returns its second argument:

Now we can represent the conditional if as a function of three arguments:

It just applies the truth value to the consequent and alternative arguments, so if is just syntactic sugar.

What about data structures? The simplest composite data structure is a pair, and we can do that with functions too:

so a pair is a function that takes a truth value and returns the car if it is true and the cdr if it is false, and car and cdr are functions that pass the appropriate truth value to the pair.

(I need to talk here about Church Numerals, when I can figure them out.)

Currying all of the above is left as an exercise for the reader.

So what does the lambda calculus actually look like? Well, it’s very terse. Functions are created by λ.x body where x is the argument, and are called like fa where f is the function and a the argument. Of course you can use brackets to disambiguate, but that’s it, nothing more to it. Here’s my attempt at a definition for if:

I hope I got that right: a function taking argument t (test) returning a function taking argument c (consequent) returning a function taking argument a (alternative) which applies t to c, and applies the result to a (remember t is Curried). Phew!

Currying is quite a useful thing to be able to do given support from a language, but it can be difficult to read. For F♮ I hit upon the idea of using a trailing comma to indicate a partial function call:

Which provides a visual clue to the reader that add() in this example is not being called with all of its arguments and therefore returns a function that expects the rest. Perhaps more practically, something like this is more visually appealing, once you get what the trailing comma means:

where we map eval with a supplied first argument env over the items of lst. This avoids having to create a closure:



Purely Functional Environments

Here’s a different take on environments. In a pure functional language all data is immutable, end of story, and that would apply to any environments implemented in such a paradigm too. Operators like define would not be allowed to modify an environment, instead such operators would have to return a copy of the environment with the modified value. This has some interesting implications for any language using such an environment. For example:

So each fn captures the current environment, and since nothing is allowed to change, the environment captured is precisely as it was when it was captured.

So how can we make a functional environment that is even remotely efficient? A standard technique is to implement it as a binary tree. Let’s assume symbols have unique numeric identifiers, so that symbol lookup can be reduced to numeric lookup. Then our environment can be a tree who’s left nodes all have indexes less than the current node, and who’s right nodes all have indexes greater than the current node. Here’s a pair of classes that implement such a tree:

So far so good, but what about extension?

extend() only makes a copy of the path through the tree that was traversed in order to find the node to add or replace. this means that for a balanced tree, extend() is O(log n) of the number of items in the environment, which is not too bad.

Note also how the EmptyNode behaves as a singleton without explicitly requiring a static instance variable. This makes the implementation quite space efficient too.

The implementation does not support deletion of values, but that is ok since only a non-functional language with backtracking would require it, in order to undo side-effects.

Lastly, note that because these environments are immutable, we do not need any notion of an environment frame to encapsulate local changes, the old environment is still present and will be reverted to outside of a closure just as if we were using environment frames.

Of course this completely ignores the issues of maintaining balanced binary trees, for the sake of simplicity.


A Bit More on Compiler Environments

Rather than a single “Environment” type, a compiler distinguishes between compile-time and run-time environments.

In an interpreter, the value of a symbol is the result of looking it up in the current environment. That lookup process involves traversing the environment, frame by frame, until the symbol is found and the associated value is retrieved. The interpreter’s environment is constructed at run-time by closures extending a previous environment with bindings of formal to actual arguments.

A compiler can’t know the values of variables at compile time. It can however, determine the location those variables will have in the environment at run-time. A compiler will perform lexical analysis of code analagous to evaluation. It passes and (when entering a closure) extends a compile-time environment that contains only the symbols, not their values. When it comes across a variable access it looks up the symbol in the compile time environment and replaces the variable with its address in the environment (i.e.three frames back, two variables in).

At run time that data can rapidly be retrieved from the run-time environment (which contains only the data) without any laborious traversal and comparison.

That gives us three classes and a dumb data object for the location (all of this is pseudocode)

  1. A compile-time environment
  2. A run-time environment:
  3. And of course a terminating Null object for the CTE

Extenders for the CTE and RTE take arrays of symbols and values respectively, otherwise they are just the same as interpreter extend() methods.