[blog] | [projects] | [about] | [imprint]

Functional Programming in (Common) Lisp

29 May 2021


Functional programming (FP) has again been getting popular in recent years. Why again? Because the paradigm of FP is very old. Early languages like Lisp, APL from the 1950/60s were already functional. Later, imperative languages like C, Pascal, etc., replaced functional languages to some degree. Now, for some years functional languages have again become more popular.

This article will describe some key concepts of functional programming. While many pure functional programming languages exist (like ML, Erlang, Haskell, Clojure), which lock you in into a pure functional style of programming, this article will describe some techniques to allow functional programming with multi-paradigm languages. The language of choice for this article is Common Lisp, a well-established industry standard that, despite its age, provides a broad range of modern language features. The techniques apply to any language that provides lambdas and functions as objects (like Java, Python, C#, Scala, etc.), which is a key feature for functional programming.

What is the difference between the two styles?

Functional programming

As the name suggests, FP is all about functions. About composing functions and about applying those functions to data to transform the data. Functions can be anonymous functions created at runtime or named functions. FP is declarative. It is more important to say what should be done than how. For instance, FP languages have higher-order functions that operate on lists. You tell this function what transformation you want to have for each list element by supplying a function that is called with each list element and transforms it, rather than manually writing a loop construct (while, for, etc.) that iterates over a list and implicitly does the transformation as part of the loop construct.
Pure functions are a key element of FP. Pure functions are referencially transparent which means that they always produce the same output for the same input and, therefore, this function could be replaced with just the value of the function output (wiki). This means that functions that have side-effects are not referencially transparent. Referential transparency then also implies that the parameters supplied to the function (the input set) are not changed. A pure function produces new data, but doesn't change the old data that was used to produce the new data. Functions that don't alter the input set are not destructive.
Additionally, this implies that there exists some immutability of data structures, whether the data structures are immutable by design or functions just don't mutate the input data structures (the latter requires a lot of discipline from the programmer). Those characteristics are often also mentioned together with data-driven development. Pure functions, as you can imagine, are thread-safe and can, therefore, be easily used in multi-threaded environments.
Aside from the practical benefits of the above, like: it's a lot easier to reason on pure functions and they tend to be easier to maintain (especially in multi-core and multi-threaded environments) and a (subjective) nicer way to code. Though it's, of course, still possible to write crap code that no one understands and is hard to maintain, the skill of the programmer to write readable and maintainable code is still imperative.
There are also disadvantages: FP is more memory consuming and requires more CPU cycles. Because, practically, due to the immutability of the objects and functions producing new data structures rather than modifying existing ones, more data is produced. This means more memory allocations are necessary and more memory must be cleaned up, which costs garbage collector time. However, with current multi-core hardware, those disadvantages are neglectable. Developer time and a maintainable code base that is easy to reason on are much more valuable for a business.
For people interested in mathmatics, FP involves quite a bit of it with category theory, morphisms, functors, etc. But this is not something that is part of this article.

Design considerations

From a software design perspective, FP programms usually have a purity core where only data transformations are taking place in the form of pure functions (a 'functional core'). But even FP programms have to deal with side-effects and, of course, there is state. Languages deal differently with this. Haskell, for example, knows 'monads' to deal with side-effects and state. Erlang/Elixir have efficient user-space processes where state is locked in. Side-effects, state, as well as threading should happen on the system's boundaries. The hexagonal architecture is an architectural style that can nicely fit to FP. It can be used with an 'imperative shell' and a 'functional core'.


Due to the immutability, pure FP languages allow assignments to variables only once. Erlang/Elixir for example have a = operator, but it's not an assignment operator as in languages like C or Java. It is a match operator, similar as in mathmatics where you can say x = y which means that x is equal to y. Let's have a quick look at Erlang:

4> X = 1.
5> Y = 2.
6> X = Y.
** exception error: no match of right hand side value 2

The X and Y variables here take the corresponding values of 1 and 2 because at this point, they are unset. Since they are unset, the = matches and assigns X with 1.
But when both variables are set and = is used, the match fails because 1 is not equal to 2.

Elixir is a little less strict on the last part:

iex(1)> x = 1
iex(2)> y = 2
iex(3)> x = y

As you can see, the last part works in Elixir. But there is still an important difference to a normal assignment. The new x has a different memory location than the previous x, which means the value in the old memory location is not altered.


Immutability is also a key characteristic in FP languages. FP languages provide immutable datatypes like tuples, lists, arrays, maps, trees, etc.
The provided functions that operate on the datatypes are pure functions. For example, removing or adding items to a list do create a new list.
This behavior is built into FP languages (ML, Haskell, Erlang/Elixir, just to name but a few). So, using those languages, your are operating in an immutability bubble. To maintain immutability inside the immutable bubble, a 'shallow' copy of the data is usually sufficient. This is much more efficient for the runtime system and means that functions operating on datatypes create a copy of the datatype but share the instances that the datatype holds. For instance, adding a new head to a list (cons) will create a new list that is constructed from the given head element and the given input list as the tail.

CL-USER> (defvar *list* '(1 2 3 4))
CL-USER> (defun prepend-head (head list)
           (cons head list))
CL-USER> (prepend-head 0 *list*)
(0 1 2 3 4)
CL-USER> *list*
(1 2 3 4)

What is not built-in are deep copies that are necessary when one leaves the immutable bubble, i.e., to do I/O.

Design patterns (GoF)

Just a few words on this. Some of the design patterns of object-oriented programming, as captured in the Gang of Four book, also apply for FP, but their implementation is, in many cases, much simpler, i.e., a strategy pattern in FP is just a higher-order function. A visitor pattern in a simple form could just be a reduce function.

Imperative programming

Imperative programming (IP), on the other hand, is more about mutating the state of a machine with every instruction/statement. I would tend to say that also Object-Oriented Programming (OOP) is also imperative. The difference being that OOP allows you to give more structure to the programm. From a memory and CPU perspective, the paradigm of IP is more efficient. And I've read somewhere (can't remember where, but it makes sense to me) that IP did replace FP in the latter half of the last century because memory was expensive and CPUs were not fast, and IP clearly has an advantage there when the value of a memory location is just changed instead of a new memory location being allocated and the old one having to be cleaned up.
But in today's multi-core and multi-threaded computing world, state is a problem. It's not possible without state, but how it is dealt with is important and different in FP.

Functional style in multi-paradigm language, Common Lisp

Common Lisp is a Lisp (obviously). Lisps have always had functions as first-class elements of the language. Functions can be anonymous (lambdas), or can have a name, they can be created at runtime and they can be passed around in the same way as strings or other objects can. Every function returns something, even if is it's just nil (for an empty function).
Common Lisp allows you to program in multiple paradigms. Historically, it had to capture, consolidate and modernize all the Lisp implementations flowing around at the time. So, it has a rather large but well-balanced feature spec and it allows both OOP (with CLOS) and FP. Common Lisp doesn't have the very strict functional characteristics, like the above-mentioned assignments. Functions in Common Lisp can have side effects and Common Lisp has assignments in the form of set, setq and setf. For FP, this means that the programmer has to be more disciplined in how they programm and which elements of the language they use. But it's, of course, possible. This applies in a similar way to Java, Scala, C#, etc.
When we look at the important characteristics of FP, they are:

  • first-class functions
  • pure functions
  • non-destructive functions
  • immutable data structures

Then, we have to be a bit careful which of all the things available in Common Lisp we use. All built-in data structures in Lisp are mutable. Lists are a bit of an exception because they are usually used in a non-destructive way. Like the function cons (construct) creates a new list by prepending a new element to the head of a list, in which case, the old list is not modified or destroyed but a new list is created. But delete-if (in contrast to remove-if) is destructive because it modifies the input list. The array/vector and hashmap data structures and their functions are destructive and shouldn't be used when doing FP. So, when developing pure and non-destructive functions (i.e. for a functional core), you need to make sure you use the right higher-order functions for operating on lists or other data structures. Depending on which data structure it is, it could also mean manually making deep copies of the input parameters, operating on the copy and returning the copy. modf can help with this, see later info.

Immutable data structures - FSet

The immutable data structures library, FSet should be a good fit when doing FP in Common Lisp. FSet defines a large set of functions for all sorts of operations.

There are more alternatives offering this functionality. See, for example, Sycamore.

Other languages, like Java and Scala, offer a range of immutable data structures out of the box.

Custom immutable types

Immutable maps are commonly used in FP instead of classes or structure types. They have one disadvantage. They don't create a specific type. In some FP languages, like Erlang/Elixir, it's possible to dispatch a function based on destructuring of the function arguments, like a map or list. But Elixir also allows you to define a type for a map structure, which then can be used for dispatching on a function level.
In Common Lisp, it would also be cool to use generic functions also for FP because it allows dynamic/multi-dispatch and is generally a nice feature. But it can't destructure lists or maps on function arguments. It can only check on a type or equality of objects using eql. Both won't work well when using FSet with just maps, or sets.
So, in addition to the data structures available in FSet the standard structure type in Common Lisp (defstruct) could still be usable. It defines a new type, so, we can use it with generic functions, we can check equality on the slots/instance vars with equalp, and we can set the slots/instance vars to :read-only which prevents from changing the slot/variable values. defstruct automatically generates a 'copier' function that copies the structure. This copy is just a flat copy and it doesn't allow you to change values while copying. Let's have a quick look at some of the structure items:

CL-USER> (defstruct foo (bar "" :read-only t))
CL-USER> (defstruct bar (foo "" :read-only t))

The option :read-only has the effect that defstruct doesn't generate 'setter' functions to change the slot. (Though it is still possible to change the slot values using lower-level API, i.e., the slot-value function, but the public interface does disallows it.)

The next snippet shows how the dynamic dispatch works with the created new structure types.

CL-USER> (defgeneric m-dispatch (arg))

CL-USER> (defmethod m-dispatch ((arg foo))
           (format t "me: foo~%"))

CL-USER> (defmethod m-dispatch ((arg bar))
           (format t "me: bar~%"))

CL-USER> (m-dispatch (make-foo :bar "bar"))
me: foo
CL-USER> (m-dispatch (make-bar :foo "foo"))
me: bar

The above shows the dynamic dispatch on the different structure types 'foo and 'bar. This works quite well. To use the structure type in FP, we'd 'just' have to come up with a copy function that allows changing the values when copying the object.

In Scala, this works quite nicely with case classes where a copy of an immutable object can be performed like this:

case class MyObject(arg1: String, arg2: Int)

val myObj1 = MyObject("foo", 1)

val myObj2 = myObj1.copy(arg1 = "bar", arg2 = 2)
Modf to the rescue

The Modf library does exactly that for Common Lisp. modf has to be used instead of setf. But it works in the same way as setf, except that it creates a new instance of the structure instead of modifying the existing structure. Let's see this in action:

CL-USER> (defstruct foo (x 1) (y 2))
CL-USER> (defparameter *foo* (make-foo))
CL-USER> *foo*
#S(FOO :X 1 :Y 2)
CL-USER> (modf (foo-x *foo*) 5)
#S(FOO :X 5 :Y 2)
CL-USER> *foo*
#S(FOO :X 1 :Y 2)

Following this little example, we can see that modf doesn't touch the original *foo* instance but creates a new one with x = 5. This is pretty cool. It's getting better. This also works for standard CLOS objects:

CL-USER> (defclass my-class () 
           ((x :initform 1)
            (y :initform 2)))
CL-USER> (defparameter *my-class* (make-instance 'my-class))
CL-USER> *my-class*
#<MY-CLASS #x302002101F8D>
CL-USER> (slot-value *my-class* 'x)
1 (1 bit, #x1, #o1, #b1)
CL-USER> (slot-value *my-class* 'y)
2 (2 bits, #x2, #o2, #b10)
CL-USER> (modf (slot-value *my-class* 'x) 5)
#<MY-CLASS #x302002250CFD>
CL-USER> (slot-value * 'x)
5 (3 bits, #x5, #o5, #b101)
CL-USER> (slot-value *my-class* 'x)
1 (1 bit, #x1, #o1, #b1)

We can see from the memory reference that a new instance was created: #x302002101F8D vs. #x302002250CFD.

So now we basically have our immutable custom types. The only important thing to remember, which again requires discipline, is to use modf instead of setf.

Even though modf also works on the built-in data structures like lists, arrays and hashmaps I would probably still tend to use a library like FSet.

One thing to mention here is that modf only makes a 'shallow' copy of the data, which means that only a new instance of the 'container' is created while the internal objects (if they are references) are shared.

More things that help doing FP
Function composition

Common Lisp does not have a construct to compose functions other than just wrapping function calls like: (f(g(h())). But that is not pleasant to read and it actually turns around the logical order of function calls. The defacto standard Common Lisp library alexandria has a function for that that can ben used, like (compose h g f):

CL-USER> (funcall (alexandria:compose #'1+ #'1+ #'1+) 1)
4 (3 bits, #x4, #o4, #b100)

This generates a composition function of the three functions 1+ like: (1+ (1+ (1+ 1))) which then can be called using funcall or provided as a higher-order function. The call order is still from right to left.

There is another alternative way of composing functions which comes from Clojure. It's actually rather a piping than a composition. Elixir also knows this as the |> operator. In Clojure it's called 'threading'. In Common Lisp exist 3 third-party libraries that implement this. The one used here is binding-arrows. There are a few more operators (macros) available for threading with slightly different features than the used ->. I like this a lot and use it often.

CL-USER> (binding-arrows:->
4 (3 bits, #x4, #o4, #b100)

The normal 'thread' arrow -> passes the previous value as the first argument to the next function. There is also a ->> arrow operator which passes the value as the last argument.

Pattern matching

Pattern matching is kind of standard for languages that have FP features. In Common Lisp, pattern matching is not part of the standard language features. But the Trivia library fills that gap. Trivia has an amazing feature set. It can match (and capture) on all native Common Lisp data structures, including structure and class slots. There are extensions for pattern matching on regular expressions and also for the before-mentioned FSet library. It can be relatively easily expanded with new patterns. The documentation is OK but could be more and better structured.

Here a simple example:

;; matching on an FSet map
(match (map (:x 5) (:y 10))
   ((fset-map :x x :y y)
    (list x y)))
=> (5 10)
;; matching on a list with capturing the tail
(match '(1 2 3)
   ((list* 1 tail)
=> (2 3)

Currying is something you see in most FP languages. It is a way to decompose one function with multiple arguments into a sequence of functions with fewer arguments. In practical terms, it reduces the dimension of available inputs to a function. For example, say you have the function coords that takes two arguments and produces a coordinate in an x-y coordinate system. With currying, we can lock one dimension, x or y.

Say we have the function:

CL-USER> (defun coords (x y)
           (cons x y))

Now, I want to lock the x coordinate to a value, say 1:

CL-USER> (curry #'coords 1)

The curry function here creates a new function that locks the x coordinate to 1 and now supports only one argument. Calling this now produces:

CL-USER> (funcall * 2)
(1 . 2)
CL-USER> (funcall ** 5)
(1 . 5)

(* denotes the last, ** the second-from-last result in the REPL.)
So, currying destructured the coord function call into two function calls. But the curried function can be stored and reused. It represents only a single dimention from the original two-dimentional set.

Common Lisp also doesn't have currying built-in. But it's easy to create. The following function performed the trick above:

CL-USER> (defun curry (fun &rest cargs)
           (lambda (&rest args)
             (apply fun (append cargs args))))

Though there is no need to create this. This is also part of the Alexandria library. Which in addition to this, also provides rcurry to curry from the right.


It is possible to do functional programming in languages that are not made for pure FP. It is important to separate the areas where side-effects may happen ('imperative-shell') and where not ('functional-core'). Functions in the 'functional core' should be pure functions that don't modify input parameters. Using immutable data structures is a big help in doing that. But immutable data structures are not always available. In that case you have to manually copy mutable data structures and operate on the copies. This requires discipline. In multi-threaded environments, it might still be worth the effort for the gain of simplicity and reasonability.

[atom/rss feed]