Unlike traditional languages, Haskell has neither a
for
loop nor a while
loop. If we’ve got a lot
of data to process, what do we use instead? There are several possible
answers to this question.
A straightforward way to make the jump from a language that has loops to one that doesn’t is to run through a few examples, looking at the differences. Here’s a C function that takes a string of decimal digits and turns them into an integer:
int as_int(char *str) { int acc; /* accumulate the partial result */ for (acc = 0; isdigit(*str); str++) { acc = acc * 10 + (*str - '0'); } return acc; }
Given that Haskell doesn’t have any looping constructs, how should we think about representing a fairly straightforward piece of code such as this?
We don’t have to start off by writing a type signature, but it helps to remind us of what we’re working with:
-- file: ch04/IntParse.hs import Data.Char (digitToInt) -- we'll need digitToInt shortly asInt :: String -> Int
The C code computes the result
incrementally as it traverses the string; the Haskell code can do the
same. However, in Haskell, we can express the equivalent of a loop as
a function. We’ll call ours loop
just to keep things nice and explicit:
-- file: ch04/IntParse.hs loop :: Int -> String -> Int asInt xs = loop 0 xs
That first parameter to loop
is the accumulator variable we’ll be
using. Passing zero into it is equivalent to initializing the
acc
variable in C at the beginning of the
loop.
Rather than leap into blazing code,
let’s think about the data we have to work with. Our familiar
String
is just a synonym for [Char]
, a list
of characters. The easiest way for us to get the traversal right is to
think about the structure of a list: it’s either empty or a single
element followed by the rest of the list.
We can express this structural thinking directly by pattern matching on the list type’s constructors. It’s often handy to think about the easy cases first; here, that means we will consider the empty list case:
-- file: ch04/IntParse.hs loop acc [] = acc
An empty list doesn’t just mean “the input string is empty”; it’s also the case that we’ll encounter when we traverse all the way to the end of a nonempty list. So we don’t want to “error out” if we see an empty list. Instead, we should do something sensible. Here, the sensible thing is to terminate the loop and return our accumulated value.
The other case we have to consider arises when the input list is not empty. We need to do something with the current element of the list, and something with the rest of the list:
-- file: ch04/IntParse.hs loop acc (x:xs) = let acc' = acc * 10 + digitToInt x in loop acc' xs
We compute a new value for the
accumulator and give it the name acc'
. We then call
the loop
function
again, passing it the updated value acc'
and the
rest of the input list. This is equivalent to the loop starting
another round in C.
Remember, a single quote is a legal
character to use in a Haskell variable name, and it is pronounced
“prime.” There’s a common idiom in Haskell programs
involving a variable—say, foo
—and another
variable—say, foo'
. We can usually assume that
foo'
is somehow related to
foo
. It’s often a new value for
foo
, as just shown in our code.
Sometimes we’ll see this idiom
extended, such as foo''
. Since keeping track of
the number of single quotes tacked onto the end of a name rapidly
becomes tedious, use of more than two in a row is thankfully rare.
Indeed, even one single quote can be easy to miss, which can lead to
confusion on the part of readers. It might be better to think of the
use of single quotes as a coding convention that you should be able
to recognize, and less as one that you should actually
follow.
Each time the loop
function calls itself, it has a new
value for the accumulator, and it consumes one element of the input
list. Eventually, it’s going to hit the end of the list, at which time
the []
pattern will match and the recursive calls will
cease.
How well does this function work? For positive integers, it’s perfectly cromulent:
ghci>
asInt "33"
33
But because we were focusing on how to traverse lists, not error handling, our poor function misbehaves if we try to feed it nonsense:
ghci>
asInt ""
0ghci>
asInt "potato"
*** Exception: Char.digitToInt: not a digit 'p'
We’ll defer fixing our function’s shortcomings to Exercises.
Because the last thing that loop
does is simply call itself, it’s an
example of a tail recursive function. There’s another common idiom in
this code, too. Thinking about the structure of the list, and handling the
empty and nonempty cases separately, is a kind of approach called structural recursion.
We call the nonrecursive case (when the list is empty) the base case (or sometimes the terminating case). We’ll see people refer to the case where the function calls itself as the recursive case (surprise!), or they might give a nod to mathematical induction and call it the inductive case.
As a useful technique, structural recursion is not confined to lists; we can use it on other algebraic data types, too. We’ll have more to say about it later.
In an imperative language, a loop executes in constant space. Lacking loops, we use tail recursive functions in Haskell instead. Normally, a recursive function allocates some space each time it applies itself, so it knows where to return to.
Clearly, a recursive function would be at a huge disadvantage relative to a loop if it allocated memory for every recursive application—this would require linear space instead of constant space. However, functional language implementations detect uses of tail recursion and transform tail recursive calls to run in constant space; this is called tail call optimization (TCO).
Few imperative language implementations perform TCO; this is why using any kind of ambitiously functional style in an imperative language often leads to memory leaks and poor performance.
Consider another C function, square
,
which squares every element in an array:
void square(double *out, const double *in, size_t length) { for (size_t i = 0; i < length; i++) { out[i] = in[i] * in[i]; } }
This contains a straightforward and common kind of loop, one that does exactly the same thing to every element of its input array. How might we write this loop in Haskell?
-- file: ch04/Map.hs square :: [Double] -> [Double] square (x:xs) = x*x : square xs square [] = []
Our square
function consists of two
pattern-matching equations. The first “deconstructs” the
beginning of a nonempty list, in order to get its head and tail. It
squares the first element, then puts that on the front of a new list,
which is constructed by calling square
on the remainder of the empty list.
The second equation ensures that square
halts when it reaches the end of the
input list.
The effect of square
is to construct a new list that’s
the same length as its input list, with every element in the input
list substituted with its square in the output list.
Here’s another such C loop, one that ensures that every letter in a string is converted to uppercase:
#include <ctype.h> char *uppercase(const char *in) { char *out = strdup(in); if (out != NULL) { for (size_t i = 0; out[i] != '\0'; i++) { out[i] = toupper(out[i]); } } return out; }
Let’s look at a Haskell equivalent:
-- file: ch04/Map.hs import Data.Char (toUpper) upperCase :: String -> String upperCase (x:xs) = toUpper x : upperCase xs upperCase [] = []
Here, we’re importing the toUpper
function from the standard Data.Char
module, which
contains lots of useful functions for working with Char
data.
Our upperCase
function follows a similar
pattern to our earlier square
function. It terminates with an empty list when the input list is
empty; when the input isn’t empty, it calls toUpper
on the first element, then
constructs a new list cell from that and the result of calling itself
on the rest of the input list.
These examples follow a common pattern for writing recursive functions over lists in Haskell. The base case handles the situation where our input list is empty. The recursive case deals with a nonempty list; it does something with the head of the list and calls itself recursively on the tail.
The square
and
upperCase
functions that we just
defined produce new lists that are the same lengths as their input
lists, and they do only one piece of work per element. This is such a
common pattern that Haskell’s Prelude
defines a function, map
, in order to make it easier. map
takes a function and applies it to
every element of a list, returning a new list constructed from the
results of these applications.
Here are our square
and upperCase
functions rewritten to use
map
:
-- file: ch04/Map.hs square2 xs = map squareOne xs where squareOne x = x * x upperCase2 xs = map toUpper xs
This is our first close look at a
function that takes another function as its argument. We can learn a
lot about what map
does by simply
inspecting its type:
ghci>
:type map
map :: (a -> b) -> [a] -> [b]
The signature tells us that map
takes two arguments. The first is a
function that takes a value of one type, a
, and returns a value of another type, b
.
Because map
takes a function as an argument, we
refer to it as a higher-order function. (In
spite of the name, there’s nothing mysterious about higher-order
functions; it’s just a term for functions that take other functions as
arguments, or return functions.)
Since map
abstracts out the pattern common to our square
and upperCase
functions so that we can reuse it
with less boilerplate, we can look at what those functions have in
common and figure out how to implement it ourselves:
-- file: ch04/Map.hs myMap :: (a -> b) -> [a] -> [b] myMap f (x:xs) = f x : myMap f xs myMap _ _ = []
If you’re new to functional
programming, the reasons for matching patterns in certain ways won’t always
be obvious. For example, in the definition of myMap
in the preceding code, the first
equation binds the function
we’re mapping to the variable f
, but the second
uses wild cards for both parameters. What’s going on?
We use a wild card in place of
f
to indicate that we aren’t calling the function
f
on the righthand side of the equation. What
about the list parameter? The list type has two constructors. We’ve
already matched on the nonempty constructor in the first equation
that defines myMap
. By
elimination, the constructor in the second equation is necessarily
the empty list constructor, so there’s no need to perform a match to
see what its value really is.
As a matter of style, it is fine to use wild cards
for well-known simple types such as lists
and
Maybe
. For more complicated or less familiar types, it
can be safer and more readable to name constructors
explicitly.
We try out our myMap
function to give ourselves some
assurance that it behaves similarly to the standard map
:
ghci>
:module +Data.Char
ghci>
map toLower "SHOUTING"
"shouting"ghci>
myMap toUpper "whispering"
"WHISPERING"ghci>
map negate [1,2,3]
[-1,-2,-3]
This pattern of spotting a repeated idiom, and then abstracting it so we can reuse (and write less!) code, is a common aspect of Haskell programming. While abstraction isn’t unique to Haskell, higher-order functions make it remarkably easy.
Another common operation on a sequence of data is to comb through it for elements that satisfy some criterion. Here’s a function that walks a list of numbers and returns those that are odd. Our code has a recursive case that’s a bit more complex than our earlier functions—it puts a number in the list it returns only if the number is odd. Using a guard expresses this nicely:
-- file: ch04/Filter.hs oddList :: [Int] -> [Int] oddList (x:xs) | odd x = x : oddList xs | otherwise = oddList xs oddList _ = []
ghci>
oddList [1,1,2,3,5,8,13,21,34]
[1,1,3,5,13,21]
Once again, this idiom is so common that the Prelude
defines a function, filter
,
which we already introduced. It removes the need for boilerplate code
to recurse over the list:
ghci>
:type filter
filter :: (a -> Bool) -> [a] -> [a]ghci>
filter odd [3,1,4,1,5,9,2,6,5]
[3,1,1,5,9,5]
The filter
function takes a predicate and
applies it to every element in its input list, returning a list of
only those for which the predicate evaluates to True
.
We’ll revisit filter
again later
in this chapter in Folding from the Right.
It is also common to reduce a collection to a single value. A simple example of this is summing the values of a list:
-- file: ch04/Sum.hs mySum xs = helper 0 xs where helper acc (x:xs) = helper (acc + x) xs helper acc _ = acc
Our helper
function is tail-recursive and uses
an accumulator parameter, acc
, to hold the current
partial sum of the list. As we already saw with asInt
, this is a “natural” way
to represent a loop in a pure functional language.
For something a little more complicated, let’s take a look at the Adler-32 checksum. It is a popular checksum algorithm; it concatenates two 16-bit checksums into a single 32-bit checksum. The first checksum is the sum of all input bytes, plus one. The second is the sum of all intermediate values of the first checksum. In each case, the sums are computed modulo 65521. Here’s a straightforward, unoptimized Java implementation (it’s safe to skip it if you don’t read Java):
public class Adler32 { private static final int base = 65521; public static int compute(byte[] data, int offset, int length) { int a = 1, b = 0; for (int i = offset; i < offset + length; i++) { a = (a + (data[i] & 0xff)) % base; b = (a + b) % base; } return (b << 16) | a; } }
Although Adler-32 is a simple checksum, this code isn’t particularly easy to read on account of the bit-twiddling involved. Can we do any better with a Haskell implementation?
-- file: ch04/Adler32.hs import Data.Char (ord) import Data.Bits (shiftL, (.&.), (.|.)) base = 65521 adler32 xs = helper 1 0 xs where helper a b (x:xs) = let a' = (a + (ord x .&. 0xff)) `mod` base b' = (a' + b) `mod` base in helper a' b' xs helper a b _ = (b `shiftL` 16) .|. a
This code isn’t exactly easier to follow
than the Java code, but let’s look at what’s going on. First of all,
we’ve introduced some new functions. The shiftL
function implements a logical shift left; (.&.)
provides a bitwise
“and”; and (.|.)
provides a bitwise “or”.
Once again, our helper
function is tail-recursive. We’ve
turned the two variables that we updated on every loop iteration in
Java into accumulator parameters. When our recursion terminates on the
end of the input list, we compute our checksum and return it.
If we take a step back, we can
restructure our Haskell adler32
to more closely resemble our earlier mySum
function. Instead of two accumulator
parameters, we can use a pair as the accumulator:
-- file: ch04/Adler32.hs adler32_try2 xs = helper (1,0) xs where helper (a,b) (x:xs) = let a' = (a + (ord x .&. 0xff)) `mod` base b' = (a' + b) `mod` base in helper (a',b') xs helper (a,b) _ = (b `shiftL` 16) .|. a
Why would we want to make this seemingly
meaningless structural change? Because as we’ve already seen with
map
and filter
, we can extract the common behavior
shared by mySum
and adler32_try2
into a higher-order function.
We can describe this behavior as “do something to every element
of a list, updating an accumulator as we go, and returning the
accumulator when we’re done.”
This kind of function is called a
fold, because it “folds up” a list. There are two kinds
of fold-over lists: foldl
for
folding from the left (the start), and foldr
for folding
from the right (the end).
Here is the definition of foldl
:
-- file: ch04/Fold.hs foldl :: (a -> b -> a) -> a -> [b] -> a foldl step zero (x:xs) = foldl step (step zero x) xs foldl _ zero [] = zero
The foldl
function takes a “step”
function, an initial value for its accumulator, and a list. The
“step” takes an accumulator and an element from the list
and returns a new accumulator value. All foldl
does is call the
“stepper” on the current accumulator and an element of
the list, and then passes the new accumulator value to itself
recursively to consume the rest of the list.
We refer to foldl
as a left fold
because it consumes the list from left (the head) to
right.
Here’s a rewrite of mySum
using foldl
:
-- file: ch04/Sum.hs foldlSum xs = foldl step 0 xs where step acc x = acc + x
That local function step
just adds two numbers, so let’s simply
use the addition operator instead, and eliminate the unnecessary
where
clause:
-- file: ch04/Sum.hs niceSum :: [Integer] -> Integer niceSum xs = foldl (+) 0 xs
Notice how much simpler this code is
than our original mySum
. We’re no
longer using explicit recursion, because foldl
takes care of that for us. We’ve
simplified our problem down to two things: what the initial value of
the accumulator should be (the second parameter to foldl
) and how to update the
accumulator (the (+)
function). As an added bonus, our code is now shorter, too, which
makes it easier to understand.
Let’s take a deeper look at what
foldl
is doing here, by manually
writing out each step in its evaluation when we call niceSum
[1,2,3]
:
-- file: ch04/Fold.hs foldl (+) 0 (1:2:3:[]) == foldl (+) (0 + 1) (2:3:[]) == foldl (+) ((0 + 1) + 2) (3:[]) == foldl (+) (((0 + 1) + 2) + 3) [] == (((0 + 1) + 2) + 3)
We can rewrite adler32_try2
using foldl
to let us focus on the details that
are important:
-- file: ch04/Adler32.hs adler32_foldl xs = let (a, b) = foldl step (1, 0) xs in (b `shiftL` 16) .|. a where step (a, b) x = let a' = a + (ord x .&. 0xff) in (a' `mod` base, (a' + b) `mod` base)
Here, our accumulator is a pair, so the
result of foldl
will be, too. We
pull the final accumulator apart when foldl
returns, and then bit-twiddle it into
a “proper” checksum.
A quick glance reveals that adler32_foldl
isn’t really any shorter than
adler32_try2
. Why should we use a
fold in this case? The advantage here lies in the fact that folds are
extremely common in Haskell, and they have regular, predictable
behavior.
This means that a reader with a little experience will have an easier time understanding a use of a fold than code that uses explicit recursion. A fold isn’t going to produce any surprises, but the behavior of a function that recurses explicitly isn’t immediately obvious. Explicit recursion requires us to read closely to understand exactly what’s going on.
This line of reasoning applies to other
higher-order library functions, including those we’ve already seen,
map
and filter
. Because they’re library functions
with well-defined behavior, we need to learn what they do only once,
and we’ll have an advantage when we need to understand any code that
uses them. These improvements in readability also carry over to
writing code. Once we start to think with higher-order functions in
mind, we’ll produce concise code more quickly.
The counterpart to foldl
is foldr
, which folds from the right of a list:
-- file: ch04/Fold.hs foldr :: (a -> b -> b) -> b -> [a] -> b foldr step zero (x:xs) = step x (foldr step zero xs) foldr _ zero [] = zero
Let’s follow the same manual evaluation
process with foldr (+) 0 [1,2,3]
as we did with niceSum
earlier in
the section The Left Fold:
-- file: ch04/Fold.hs foldr (+) 0 (1:2:3:[]) == 1 + foldr (+) 0 (2:3:[]) == 1 + (2 + foldr (+) 0 (3:[]) == 1 + (2 + (3 + foldr (+) 0 [])) == 1 + (2 + (3 + 0))
The difference between foldl
and foldr
should be clear from looking at where
the parentheses and the empty list elements show up. With foldl
, the empty list element is on the
left, and all the parentheses group to the left. With foldr
, the zero
value is
on the right, and the parentheses group to the right.
There is a lovely intuitive explanation
of how foldr
works: it replaces
the empty list with the zero
value, and replaces
every constructor in the list with an application of the step
function:
-- file: ch04/Fold.hs 1 : (2 : (3 : [])) 1 + (2 + (3 + 0 ))
At first glance, foldr
might seem less useful than foldl
: what use is a function that folds
from the right? But consider the Prelude
’s filter
function, which we last encountered
earlier in this chapter in Selecting Pieces of Input. If we write
filter
using explicit recursion,
it will look something like this:
-- file: ch04/Fold.hs filter :: (a -> Bool) -> [a] -> [a] filter p [] = [] filter p (x:xs) | p x = x : filter p xs | otherwise = filter p xs
Perhaps surprisingly, though, we can
write filter
as a fold, using
foldr
:
-- file: ch04/Fold.hs myFilter p xs = foldr step [] xs where step x ys | p x = x : ys | otherwise = ys
This is the sort of definition that could cause us a
headache, so let’s examine it in a little depth. Like foldl
, foldr
takes a function and a base case
(what to do when the input list is empty) as arguments. From reading
the type of filter
, we know that
our myFilter
function must return
a list of the same type as it consumes, so the base case should be a
list of this type, and the step
helper function must return a list.
Since we know that foldr
calls step
on one
element of the input list at a time, then with the accumulator as its
second argument, step
’s actions
must be quite simple. If the predicate returns True
, it pushes that element onto the
accumulated list; otherwise, it leaves the list untouched.
The class of functions that we can
express using foldr
is called primitive recursive. A
surprisingly large number of list manipulation functions are primitive
recursive. For example, here’s map
written in terms of foldr
:
-- file: ch04/Fold.hs myMap :: (a -> b) -> [a] -> [b] myMap f xs = foldr step [] xs where step x ys = f x : ys
In fact, we can even write foldl
using foldr
!
-- file: ch04/Fold.hs myFoldl :: (a -> b -> a) -> a -> [b] -> a myFoldl f z xs = foldr step id xs z where step x g a = g (f a x)
If you want to set yourself a solid
challenge, try to follow our definition of foldl
using foldr
. Be warned: this is not trivial!
You might want to have the following tools at hand: some headache
pills and a glass of water, ghci
(so that you can find out what the id
function
does), and a pencil and paper.
You will want to follow the same
manual evaluation process as we just outlined to see what foldl
and foldr
were really doing. If you get
stuck, you may find the task easier after reading Partial Function Application and Currying.
Returning to our earlier intuitive
explanation of what foldr
does,
another useful way to think about it is that it
transforms its input list. Its first two
arguments are “what to do with each head/tail element of the
list,” and “what to substitute for the end of the
list.”
The “identity”
transformation with foldr
thus
replaces the empty list with itself and applies the list constructor
to each head/tail pair:
-- file: ch04/Fold.hs identity :: [a] -> [a] identity xs = foldr (:) [] xs
It transforms a list into a copy of itself:
ghci>
identity [1,2,3]
[1,2,3]
If foldr
replaces the end of a list with some
other value, this gives us another way to look at Haskell’s list append function, (++)
:
ghci>
[1,2,3] ++ [4,5,6]
[1,2,3,4,5,6]
All we have to do to append a list onto another is substitute that second list for the end of our first list:
-- file: ch04/Fold.hs append :: [a] -> [a] -> [a] append xs ys = foldr (:) ys xs
ghci>
append [1,2,3] [4,5,6]
[1,2,3,4,5,6]
Here, we replace each list constructor with another list constructor, but we replace the empty list with the list we want to append onto the end of our first list.
As our extended treatment of folds
should indicate, the foldr
function is nearly as important a member of our list-programming
toolbox as the more basic list functions we saw in Working with Lists. It can consume and produce a list
incrementally, which makes it useful for writing lazy data-processing
code.
To keep our initial discussion simple, we use foldl
throughout most of this section. This is convenient for testing, but
we will never use foldl
in
practice. The reason has to do with Haskell’s nonstrict evaluation. If
we apply foldl (+) [1,2,3]
, it evaluates to the
expression (((0 + 1) + 2) + 3)
. We can see this occur if
we revisit the way in which the function gets expanded:
-- file: ch04/Fold.hs foldl (+) 0 (1:2:3:[]) == foldl (+) (0 + 1) (2:3:[]) == foldl (+) ((0 + 1) + 2) (3:[]) == foldl (+) (((0 + 1) + 2) + 3) [] == (((0 + 1) + 2) + 3)
The final expression will not be evaluated to
6
until its value is demanded. Before it is evaluated, it
must be stored as a thunk. Not surprisingly, a thunk is more expensive
to store than a single number, and the more complex the thunked
expression, the more space it needs. For something cheap such as
arithmetic, thunking an expression is more computationally expensive
than evaluating it immediately. We thus end up paying both in space
and in time.
When GHC is evaluating a thunked expression, it uses an internal stack to do so. Because a thunked expression could potentially be infinitely large, GHC places a fixed limit on the maximum size of this stack. Thanks to this limit, we can try a large thunked expression in ghci without needing to worry that it might consume all the memory:
ghci>
foldl (+) 0 [1..1000]
500500
From looking at this expansion, we can
surmise that this creates a thunk that consists of 1,000 integers and
999 applications of (+)
. That’s a
lot of memory and effort to represent a single number! With a larger
expression, although the size is still modest, the results are more
dramatic:
ghci>
foldl (+) 0 [1..1000000]
*** Exception: stack overflow
On small expressions, foldl
will work correctly but slowly, due
to the thunking overhead that it incurs. We refer to this invisible
thunking as a space leak, because our code is
operating normally, but it is using far more memory than it
should.
On larger expressions, code with a
space leak will simply fail, as above. A space leak with
foldl
is a classic roadblock for new Haskell programmers.
Fortunately, this is easy to avoid.
The Data.List
module
defines a function named foldl'
that is similar to foldl
, but
does not build up thunks. The difference in behavior between the two
is immediately obvious:
ghci>
foldl (+) 0 [1..1000000]
*** Exception: stack overflowghci>
:module +Data.List
ghci>
foldl' (+) 0 [1..1000000]
500000500000
Due to foldl
’s thunking behavior, it is wise to
avoid this function in real programs, even if it doesn’t fail
outright, it will be unnecessarily inefficient. Instead, import
Data.List
and use foldl'
.
The article “A tutorial on the universality and expressiveness of fold” by Graham Hutton (http://www.cs.nott.ac.uk/~gmh/fold.pdf) is an excellent and in-depth tutorial that covers folds. It includes many examples of how to use simple, systematic calculation techniques to turn functions that use explicit recursion into folds.