IO inside

Haskell I/O can be a source of confusion and surprises for new Haskellers - if that's you, a good place to start is the Introduction to IO which can help you learn the basics (e.g. the syntax of I/O expressions) before continuing on.

While simple I/O code in Haskell looks very similar to its equivalents in imperative languages, attempts to write somewhat more complex code often result in a total mess. This is because Haskell I/O is really very different in how it actually works.

The following text is an attempt to explain the inner workings of I/O in Haskell. This explanation should help you eventually learn all the smart I/O tips. Moreover, I've added a detailed explanation of various traps you might encounter along the way. After reading this text, you will be well on your way towards mastering I/O in Haskell.

Haskell is a pure language

Haskell is a pure language and even the I/O system can't break this purity. Being pure means that the result of any function call is fully determined by its arguments. Imperative routines like rand() or getchar() in C, which return different results on each call, are simply impossible to write in Haskell. Moreover, Haskell functions can't have side effects, which means that they can't make any changes "outside the Haskell program", like changing files, writing to the screen, printing, sending data over the network, and so on. These two restrictions together mean that any function call can be replaced by the result of a previous call with the same parameters, and the language guarantees that all these rearrangements will not change the program result! For example, the hyperbolic cosine function cosh can be defined in Haskell as:

cosh r = (exp r + 1/exp r)/2

using identical calls to exp, which is another function. So cosh can instead call exp once, and reuse the result:

cosh r = (x + 1/x)/2 where x = exp r

Let's compare this to C: optimizing C compilers try to guess which routines have no side effects and don't depend on mutable global variables. If this guess is wrong, an optimization can change the program's semantics! To avoid this kind of disaster, C optimizers are conservative in their guesses or require hints from the programmer about the purity of routines.

Compared to an optimizing C compiler, a Haskell compiler is a set of pure mathematical transformations. This results in much better high-level optimization facilities. Moreover, pure mathematical computations can be much more easily divided into several threads that may be executed in parallel, which is increasingly important in these days of multi-core CPUs. Finally, pure computations are less error-prone and easier to verify, which adds to Haskell's robustness and to the speed of program development using Haskell.

Haskell's purity allows the compiler to call only functions whose results are really required to calculate the final value of a top-level definition (e.g. main) - this is called lazy evaluation. It's a great thing for pure mathematical computations, but how about I/O actions? Something like

putStrLn "Press any key to begin formatting"

can't return any meaningful result value, so how can we ensure that the compiler will not omit or reorder its execution? And in general: How we can work with stateful algorithms and side effects in an entirely lazy language? This question has had many different solutions proposed while Haskell was developed (see A History of Haskell), with one solution eventually making its way into the current standard.

I/O in Haskell, simplified

So what is actually inside an I/O action? Let's look at how MicroHs defines the I/O type:

data IO a

just as described on page 95 of 329 in the Haskell 2010 Report: no visible data constructors. So someone who is implementing Haskell could in fact define functions and I/O actions in much the same way. The only difference that really matters is this:

• Only I/O actions are allowed to make changes "outside the Haskell program".

(Or to state it more formally: only I/O actions are allowed to have externally visible side effects).

It doesn't get much simpler than that :-D

The question of purity

So if Haskell uses the same side effects for I/O as an imperative language, how can it possibly be "pure"?

Because of what doesn't work in Haskell:

\msg x -> seq (putStrLn msg) x

No, that won't work. And neither will this:

getChar >>= \c -> c

Remember, Haskell functions can't have side effects - they can't make any changes "outside the Haskell program". Therefore if either of these examples really did work, then they would no longer be Haskell functions! (More importantly, just imagine if those two were being used in parallel somewhere in a program...)

So in Haskell, the result of running an I/O action must be another I/O action. This restriction ensures that Haskell functions really are pure.

Running with I/O

So, main just has the type IO (). Let's look at main calling getChar two times:

main :: IO ()
main = getChar >>= \a ->
       getChar >>= \b ->
       return ()

By defining a Monad instance for IO a:

unitIO :: a -> IO a                     -- implemented
bindIO :: IO a -> (a -> IO b) -> IO b   --  elsewhere

instance Monad IO where
    return = unitIO
    (>>=)  = bindIO

we can then expand main to get:

main = getChar `bindIO` (\a ->
       getChar `bindIO` (\b ->
       unitIO ()))

Now to run main:

1. main = getChar `bindIO` (\a -> ...) doesn't require evaluation, continuing;
2. run getChar to obtain a character c1 :: Char;
3. apply (\a -> ...) to c1;
4. then evaluate the result to obtain the next action getChar `bindIO` (\b -> ...);
5. run getChar to obtain another character c2 :: Char;
6. apply (\b -> ...) to c2;
7. then evaluate the result to obtain the next action unitIO ();
8. run unitIO () to obtain () :: (), which ends the program.

From that example we can see that:

• Each action is run - it doesn't matter if what's obtained from running it isn't actually used.

• Each action is run in the order it appears in the program - there is no reordering of actions.

• Each action is run once, then the next action is obtained or the program ends - if a program only uses an action once, then it is only run once.

Overall, in order to obtain the final value of main, each I/O action that is called from main - directly or indirectly - is run. This means that each action inserted in the chain will be performed just at the moment (relative to the other I/O actions) when we intended it to be called. Let's consider the following program:

main = do a <- ask "What is your name?"
          b <- ask "How old are you?"
          return ()

ask s = do putStr s
           readLn

Now you have enough knowledge to rewrite it in a low-level way and check that each operation that should be performed will really be performed with the arguments it should have and in the order we expect.

But what about conditional execution? No problem. Let's define the well-known when function:

when :: Bool -> IO () -> IO ()
when condition action =
    if condition
      then action
      else return ()

Because it's a function:

• it will be applied to two arguments;
• its result (the conditional expression) will be evaluated;
• then the chosen action will be run.

As you can see, we can easily include or exclude from the execution chain I/O actions depending on the data values. If condition will be False on the call of when, action will never be run.

Loops and more complex control structures can be implemented in the same way. Try it as an exercise!

(>>=) and do notation

All beginners (including me) start by thinking that do is some super-awesome statement that executes I/O actions. That's wrong - do is just syntactic sugar that simplifies the writing of definitions that use I/O (and also other monads, but that's beyond the scope of this manual). do notation eventually gets translated to a series of I/O actions much like we've manually written above. This simplifies the gluing of several I/O actions together. You don't need to use do for just one action; for example,

main = do putStr "Hello!"

is desugared to:

main = putStr "Hello!"

Let's examine how to desugar a do-expression with multiple actions in the following example:

main = do putStr "What is your name?"
          putStr "How old are you?"
          putStr "Nice day!"

The do-expression here just joins several I/O actions that should be performed sequentially. It's translated to sequential applications of one of the so-called "binding operators", namely (>>):

main = (putStr "What is your name?")
       >> ( (putStr "How old are you?")
            >> (putStr "Nice day!")
          )

Defining (>>) looks easy:

(>>) :: IO a -> IO b -> IO b
action1 >> action2 = action1 >>= \_ -> action2

Now you can substitute the definition of (>>) at the places of its usage and check that program constructed by the do desugaring is actually the same as we could write by using I/O actions manually.

A more complex example involves the binding of variables using <-:

main = do a <- readLn
          print a

This code is desugared into:

main = readLn
       >>= (\a -> print a)

where (>>=) corresponds to bindIO.

As you now know, the (>>) binding operator silently ignores the value of its first action and returns as an overall result the result of its second action only. On the other hand, the (>>=) binding operator (note the extra = at the end) allows us to use the result of its first action - it gets passed as an additional parameter to the second one!

You can use (>>) and (>>=) to simplify your program. For example, in the code above we don't need to introduce the variable, because the result of running readLn can be passed directly to print:

main = readLn >>= print

As you see, the notation:

 do x <- action1
    action2

where action1 has type IO a and action2 has type IO b, translates into:

 action1 >>= (\x -> action2)

where the second argument of (>>=) has the type a -> IO b. It's the way the <- binding is processed - the name on the left-hand side of <- just becomes a parameter of subsequent operations represented as one large I/O action. Note also that if action1 has type IO a then x will just have type a; you can think of the effect of <- as "unpacking" the I/O value of action1 into x. Note also that <- is not a true operator; it's pure syntax, just like do itself. Its meaning results only from the way it gets desugared.

Look at the next example:

main = do putStr "What is your name?"
          a <- readLn
          putStr "How old are you?"
          b <- readLn
          print (a,b)

This code is desugared into:

main = putStr "What is your name?"
       >> readLn
       >>= \a -> putStr "How old are you?"
       >> readLn
       >>= \b -> print (a,b)

I omitted the parentheses here; both the (>>) and the (>>=) operators are left-associative, but lambda-bindings always stretches as far to the right as possible, which means that the a and b bindings introduced here are valid for all remaining actions. As an exercise, add the parentheses yourself and translate this definition into action-level code. I think it should be enough to help you finally realize how the do translation and binding operators work.

Oh, no! I forgot the third monadic operator: return. But that's understandable - it does very little! The resulting I/O action immediately returns its given argument (when it is run).

How about translating a simple example of return usage? Say,

main = do a <- readLn
          return (a*2)

Programmers with an imperative language background often think that return in Haskell, as in other languages, immediately returns from the I/O definition. As you can see in its definition (and even just from its type!), such an assumption is totally wrong. The only purpose of using return is to "lift" some value (of type a) into the result of a whole action (of type IO a) and therefore it should generally be used only as the last executed action of some I/O sequence. For example try to translate the following definition into the corresponding low-level code:

main = do a <- readLn
          when (a>=0) $ do
              return ()
          print "a is negative"

and you will realize that the print call is executed even for non-negative values of a. If you need to escape from the middle of an I/O definition, you can use an if expression:

main = do a <- readLn
          if (a>=0)
            then return ()
            else print "a is negative"

Moreover, Haskell layout rules allow us to use the following layout:

main = do a <- readLn
          if (a>=0) then return ()
            else do
          print "a is negative"
          ...

that may be useful for escaping from the middle of a longish do-expression.

Last exercise: implement a function liftM that lifts operations on plain values to the operations on monadic ones. Its type signature:

liftM :: (a -> b) -> (IO a -> IO b)

If that's too hard for you, start with the following high-level definition and rewrite it in low-level fashion:

liftM f action = do x <- action
                    return (f x)

Mutable data (references, arrays, hash tables...)

As you should know, every name in Haskell is bound to one fixed (immutable) value. This greatly simplifies understanding algorithms and code optimization, but it's inappropriate in some cases. As we all know, there are plenty of algorithms that are simpler to implement in terms of updatable variables, arrays and so on. This means that the value associated with a variable, for example, can be different at different execution points, so reading its value can't be considered as a pure function. Imagine, for example, the following code:

main = do let a0 = readVariable varA
              _  = writeVariable varA 1
              a1 = readVariable varA
          print (a0, a1)

Does this look strange?

1. The two calls to readVariable look the same, so the compiler can just reuse the value returned by the first call.
2. The result of the writeVariable call isn't used so the compiler can (and will!) omit this call completely.
3. These three calls may be rearranged in any order because they appear to be independent of each other.

This is obviously not what was intended. What's the solution? You already know this - use I/O actions! Doing that guarantees:

1. the result of the "same" action (such as readVariable varA) will not be reused
2. each action will have to be executed
3. the execution order will be retained as written

So, the code above really should be written as:

import Data.IORef
main = do varA <- newIORef 0  -- Create and initialize a new variable
          a0 <- readIORef varA
          writeIORef varA 1
          a1 <- readIORef varA
          print (a0, a1)

Here, varA has the type IORef Int which means "a variable (reference) in the I/O monad holding a value of type Int". newIORef creates a new variable (reference) and returns it, and then read/write actions use this reference. The value returned by the readIORef varA action depends not only on the variable involved but also on the moment this operation is performed so it can return different values on each call.

Arrays, hash tables and any other mutable data structures are defined in the same way - for each of them, there's an operation that creates new "mutable values" and returns a reference to it. Then value-specific read and write operations in the I/O monad are used. The following code shows an example using mutable arrays:

import Data.Array.IO
main = do arr <- newArray (1,10) 37 :: IO (IOArray Int Int)
          a <- readArray arr 1
          writeArray arr 1 64
          b <- readArray arr 1
          print (a, b)

Here, an array of 10 elements with 37 as the initial value at each location is created. After reading the value of the first element (index 1) into a this element's value is changed to 64 and then read again into b. As you can see by executing this code, a will be set to 37 and b to 64.

Other state-dependent operations are also often implemented with I/O actions. For example, a random number generator should return a different value on each call. It looks natural to give it a type involving IO:

rand :: IO Int

Moreover, when you import a C routine you should be careful - if this routine is impure, i.e. its result depends on something "outside the Haskell program" (file system, memory contents, its own static internal state and so on), you should give it an IO type. Otherwise, the compiler can "optimize" repetitive calls to the definition with the same parameters!

For example, we can write a non-IO type for:

foreign import ccall
   sin :: Double -> Double

because the result of sin depends only on its argument, but

foreign import ccall
   tell :: Int -> IO Int

If you will declare tell as a pure function (without IO) then you may get the same position on each call!

Encapsulated mutable data: ST

If you're going to be doing things like sending text to a screen or reading data from a scanner, IO is the type to start with - you can then customise existing I/O operations or add new ones as you see fit. But what if that shiny-new (or classic) algorithm you're working on really only needs mutable state - then having to drag that IO type from main all the way through to wherever you're implementing the algorithm can get quite irritating.

Fortunately there is a better way! One that remains totally pure and yet allows the use of references, arrays, and so on - and it's done using, you guessed it, Haskell's versatile type system (and one extension).

Remember our definition of IO?

data IO a

Well, the new ST type makes just one change - in theory, it can be used with any suitable state type:

data ST s a

If we wanted to, we could even use ST to define IO:

type IO a = ST RealWorld a  -- RealWorld defined elsewhere

Let's add some extra definitions:

newSTRef     :: a -> ST s (STRef s a)      -- these are
readSTRef    :: STRef s a -> ST s a        --  usually
writeSTRef   :: STRef s a -> a -> ST s ()  -- primitive

newSTArray   :: Ix i => (i, i) -> ST s (STArray s i e) -- also usually primitive
              ⋮
unitST       :: a -> ST s a
bindST       :: ST s a -> (a -> ST s b) -> ST s b

instance Monad (ST s) where
    return = unitST
    (>>=)  = bindST

...that's right - this new ST type is also monadic!

So what's the big difference between the ST and IO types? In one word - runST:

runST :: (forall s . ST s a) -> a

Yes - it has a very unusual type. But that type allows you to run your stateful computation as if it was a pure definition!

The s type variable in ST is the type of the local state. Moreover, all the fun mutable stuff available for ST is quantified over s:

newSTRef  :: a -> ST s (STRef s a)
newArray_ :: Ix i => (i, i) -> ST s (STArray s i e)

So why does runST have such a funky type? Let's see what would happen if we wrote

makeSTRef :: a -> STRef s a
makeSTRef a = runST (newSTRef a)

This fails, because newSTRef a doesn't work for all state types s - it only works for the s from the return type STRef s a.

This is all sort of wacky, but the result is that you can only run an ST computation where the output type is functionally pure, and makes no references to the internal mutable state of the computation. In exchange for that, there's no access to I/O operations like writing to or reading from the console. The monadic ST type only has references, arrays, and such that are useful for performing pure computations.

Due to how similar IO and ST are internally, there's this function:

stToIO :: ST RealWorld a -> IO a

The difference is that ST uses the type system to forbid unsafe behavior like extracting mutable objects from their safe ST wrapping, but allowing purely functional outputs to be performed with all the handy access to mutable references and arrays.

For example, here's a particularly convoluted way to compute the integer that comes after zero:

oneST :: ST s Integer -- note that this works correctly for any s
oneST = do var <- newSTRef 0
           modifySTRef var (+1)
           readSTRef var

one :: Int
one = runST oneST

I/O actions as values

By this point you should understand why it's impossible to use I/O actions inside non-I/O (pure) functions: when needed, fully-applied functions are always evaluated - they aren't run like I/O actions. In addition, the prohibition of using I/O actions inside pure functions is maintained by the type system (as it usually is in Haskell).

But while pure code can't be used to run I/O actions, it can work with them as with any other value - I/O actions can be stored in data structures, passed as parameters, returned as results, collected in lists or in tuples. But what won't work is something like:

\ msg x -> case putStrLn msg of _ -> x

because it will be treated as a function, not an I/O action.

To run an I/O action, we need to make it part of main:

• either directly:

main = action

• or in the "action chain" of another action which is already a part of the main "chain":

main = ... >>= \ _ -> action >>= ...

Only then will the action be run. For example, in:

main = do let skip2chars = getChar >> getChar >> return ()
          putStr "Press two keys"
          skip2chars
          return ()

the non-let actions are run in the exact order in which they're written.

Example: a list of I/O actions

Let's try defining a list of I/O actions:

ioActions :: [IO ()]
ioActions = [(print "Hello!"),
             (putStr "just kidding"),
             (getChar >> return ())
            ]

I used additional parentheses around each action, although they aren't really required. If you still can't believe that these actions won't be run immediately, remember that in this expression:

\ b -> if b then (putStr "started...") (putStrLn "completed.")

both I/O actions won't immediately be run either.

Well, now we want to execute some of these actions. No problem, just insert them into the main chain:

main = do head ioActions
          ioActions !! 1
          last ioActions

Looks strange, right? Really, any I/O action that you write in a do-expression (or use as a parameter for the (>>)/(>>=) operators) is an expression returning a result of type IO a for some type a. Typically, you use some function that has the type x -> y -> ... -> IO a and provide all the x, y, etc. parameters. But you're not limited to this standard scenario - don't forget that Haskell is a functional language and you're free to evaluate any value as required (recall that IO a is really a function type) in any possible way. Here we just extracted several functions from the list - no problem. This value can also be constructed on-the-fly, as we've done in the previous example - that's also OK. Want to see this value passed as a parameter? Just look at the definition of when. Hey, we can buy, sell, and rent these I/O actions just like we can with any other values! For example, let's define a function that executes all the I/O actions in the list:

sequence_ :: [IO a] -> IO ()
sequence_ [] = return ()
sequence_ (x:xs) = do x
                      sequence_ xs

No smoke or mirrors - we just extract I/O actions from the list and insert them into a chain of I/O operations that should be performed one after another (in the same order that they occurred in the list) to obtain the end result of the entire sequence_ call.

With the help of sequence_, we can rewrite our last main action as:

main = sequence_ ioActions

Haskell's ability to work with I/O actions just like other values allows us to define control structures of arbitrary complexity. Try, for example, to define a control structure that repeats an action until it returns the False result:

while :: IO Bool -> IO ()
while action = ???

Most programming languages don't allow you to define control structures at all, and those that do often require you to use a macro-expansion system. In Haskell, control structures are just trivial functions anyone can write.

Example: returning an I/O action as a result

How about returning an I/O action as the result of a function? Well, we've done this for each I/O definition - they all return I/O actions built up from other I/O actions (or themselves, if they're recursive). While we usually just execute them as part of a higher-level I/O definition, it's also possible to just collect them without actual execution:

main = do let a = sequence ioActions
              b = when True getChar
              c = getChar >> getChar >> return ()
          putStr "These let-bindings are not executed!"

These assigned I/O actions can be used as parameters to other definitions, or written to global variables, or processed in some other way, or just executed later, as we did in the example with skip2chars.

But how about returning a parameterized I/O action from an I/O definition? Here's a definition that returns the i'th byte from a file represented as a Handle:

readi h i = do hSeek h AbsoluteSeek i
               hGetChar h

So far so good. But how about a definition that returns the i'th byte of a file with a given name without reopening it each time?

readfilei :: String -> IO (Integer -> IO Char)
readfilei name = do h <- openFile name ReadMode
                    return (readi h)

As you can see, it's an I/O definition that opens a file and returns...an I/O action that will read the specified byte. But we can go further and include the readi body in readfilei:

readfilei name = do h <- openFile name ReadMode
                    let readi h i = do hSeek h AbsoluteSeek i
                                       hGetChar h
                    return (readi h)

That's a little better. But why do we add h as a parameter to readi if it can be obtained from the environment where readi is now defined? An even shorter version is this:

readfilei name = do h <- openFile name ReadMode
                    let readi i = do hSeek h AbsoluteSeek i
                                     hGetChar h
                    return readi

What have we done here? We've build a parameterized I/O action involving local names inside readfilei and returned it as the result. Now it can be used in the following way:

main = do myfile <- readfilei "test"
          a <- myfile 0
          b <- myfile 1
          print (a,b)

This way of using I/O actions is very typical for Haskell programs - you just construct one or more I/O actions that you need, with or without parameters, possibly involving the parameters that your "constructor" received, and return them to the caller. Then these I/O actions can be used in the rest of the program without any knowledge of how you actually implemented them. One thing this can be used for is to partially emulate the OOP (or more precisely, the ADT) programming paradigm.

Example: a memory allocator generator

As an example, one of my programs has a module which is a memory suballocator. It receives the address and size of a large memory block and returns two specialised I/O operations - one to allocate a subblock of a given size and the other to free the allocated subblock:

memoryAllocator :: Ptr a -> Int -> IO (Int -> IO (Ptr b),
                                       Ptr c -> IO ())

memoryAllocator buf size = do ......
                              let alloc size = do ...
                                                  ...
                                  free ptr = do ...
                                                ...
                              return (alloc, free)

How this is implemented? alloc and free work with references created inside the memoryAllocator definition. Because the creation of these references is a part of the memoryAllocator I/O-action chain, a new independent set of references will be created for each memory block for which memoryAllocator is called:

memoryAllocator buf size =
   do start <- newIORef buf
      end <- newIORef (buf `plusPtr` size)
      ...

These two references are read and written in the alloc and free definitions (we'll implement a very simple memory allocator for this example):

      ...
      let alloc size = do addr <- readIORef start
                          writeIORef start (addr `plusPtr` size)
                          return addr

      let free ptr = do writeIORef start ptr

What we've defined here is just a pair of closures that use state available at the moment of their definition. As you can see, it's as easy as in any other functional language, despite Haskell's lack of direct support for impure routines.

The following example uses the operations returned by memoryAllocator, to simultaneously allocate/free blocks in two independent memory buffers:

main = do buf1 <- mallocBytes (2^16)
          buf2 <- mallocBytes (2^20)
          (alloc1, free1) <- memoryAllocator buf1 (2^16)
          (alloc2, free2) <- memoryAllocator buf2 (2^20)
          ptr11 <- alloc1 100
          ptr21 <- alloc2 1000
          free1 ptr11
          free2 ptr21
          ptr12 <- alloc1 100
          ptr22 <- alloc2 1000

Example: emulating OOP with record types

Let's implement the classical OOP example: drawing figures. There are figures of different types: circles, rectangles and so on. The task is to create a heterogeneous list of figures. All figures in this list should support the same set of operations: draw, move and so on. We will define these operations using I/O actions. Instead of a "class" let's define a structure from which all of the required operations can be accessed:

data Figure = Figure { draw :: IO (),
                       move :: Displacement -> IO ()
                     }

type Displacement = (Int, Int)  -- horizontal and vertical displacement in points

The constructor of each figure's type should just return a Figure record:

circle    :: Point -> Radius -> IO Figure
rectangle :: Point -> Point -> IO Figure

type Point = (Int, Int)  -- point coordinates
type Radius = Int        -- circle radius in points

We will "draw" figures by just printing their current parameters. Let's start with implementing simplified circle and rectangle constructors, without actual move support:

circle center radius = do
    let description = "  Circle at "++show center++" with radius "++show radius
    return $ Figure { draw = putStrLn description }

rectangle from to = do
    let description = "  Rectangle "++show from++"-"++show to)
    return $ Figure { draw = putStrLn description }

As you see, each constructor just returns a fixed draw operation that prints parameters with which the concrete figure was created. Let's test it:

drawAll :: [Figure] -> IO ()
drawAll figures = do putStrLn "Drawing figures:"
                     mapM_ draw figures

main = do figures <- sequence [circle (10,10) 5,
                               circle (20,20) 3,
                               rectangle (10,10) (20,20),
                               rectangle (15,15) (40,40)]
          drawAll figures

Now let's define "full-featured" figures that can actually be moved around. In order to achieve this, we should provide each figure with a mutable variable that holds each figure's current screen location. The type of this variable will be IORef Point. This variable should be created in the figure constructor and manipulated in I/O operations (closures) enclosed in the Figure record:

circle center radius = do
    centerVar <- newIORef center

    let drawF = do center <- readIORef centerVar
                   putStrLn ("  Circle at "++show center
                             ++" with radius "++show radius)

    let moveF (addX,addY) = do (x,y) <- readIORef centerVar
                               writeIORef centerVar (x+addX, y+addY)

    return $ Figure { draw=drawF, move=moveF }

rectangle from to = do
    fromVar <- newIORef from
    toVar   <- newIORef to

    let drawF = do from <- readIORef fromVar
                   to   <- readIORef toVar
                   putStrLn ("  Rectangle "++show from++"-"++show to)

    let moveF (addX,addY) = do (fromX,fromY) <- readIORef fromVar
                               (toX,toY)     <- readIORef toVar
                               writeIORef fromVar (fromX+addX, fromY+addY)
                               writeIORef toVar   (toX+addX, toY+addY)

    return $ Figure { draw=drawF, move=moveF }

Now we can test the code which moves figures around:

main = do figures <- sequence [circle (10,10) 5,
                               rectangle (10,10) (20,20)]
          drawAll figures
          mapM_ (\fig -> move fig (10,10)) figures
          drawAll figures

It's important to realize that we are not limited to including only I/O actions in a record that's intended to simulate a C++/Java-style interface. The record can also include values, IORefs, pure functions - in short, any type of data. For example, we can easily add to the Figure interface fields for area and origin:

data Figure = Figure { draw :: IO (),
                       move :: Displacement -> IO (),
                       area :: Double,
                       origin :: IORef Point
                     }

Exception handling (under development)

Although Haskell provides a set of exception raising/handling features comparable to those in popular OOP languages (C++, Java, C#), this part of the language receives much less attention. This is for two reasons:

• you just don't need to worry as much about them - most of the time it just works "behind the scenes".

• Haskell, lacking OOP-style inheritance, doesn't allow the programmer to easily subclass exception types, therefore limiting the flexibility of exception handling.

Haskell can raise more exceptions than other programming languages - pattern match failures, calls with invalid arguments (such as head []) and computations whose results depend on special values undefined and error "...." all raise their own exceptions:

• example 1:

main = print (f 2)

f 0 = "zero"
f 1 = "one"

• example 2:

main = print (head [])

• example 3:

main = print (1 + (error "Value that wasn't initialized or cannot be computed"))

This allows the writing of programs in a much more error-prone way.

Interfacing with C/C++ and foreign libraries (under development)

While Haskell is great at algorithm development, speed isn't its best side. We can combine the best of both languages, though, by writing speed-critical parts of program in C and the rest in Haskell. We just need a way to call C routines from Haskell and vice versa, and to marshal data between the two languages.

We also need to interact with C to use Windows/Linux APIs, linking to various libraries and DLLs. Even interfacing with other languages often requires going through C, which acts as a "common denominator". Chapter 8 of the Haskell 2010 report provides a complete description of interfacing with C.

We will learn to use the FFI via a series of examples. These examples include C/C++ code, so they need C/C++ compilers to be installed, the same will be true if you need to include code written in C/C++ in your program (C/C++ compilers are not required when you just need to link with existing libraries providing APIs with C calling convention). On Unix (and Mac OS?) systems, the system-wide default C/C++ compiler is typically used by GHC installation. On Windows, no default compilers exist, so GHC is typically shipped with a C compiler, and you may find on the download page a GHC distribution bundled with C and C++ compilers. Alternatively, you may find and install a GCC/MinGW version compatible with your GHC installation.

If you need to make your C/C++ code as fast as possible, you may compile your code by Intel compilers instead of GCC. However, these compilers are not free, moreover on Windows, code compiled by Intel compilers may not interact correctly with GHC-compiled code, unless one of them is put into DLLs (due to object file incompatibility).

More links:

C->Haskell

A lightweight tool for implementing access to C libraries from Haskell.

HSFFIG

The Haskell FFI Binding Modules Generator (HSFFIG) is a tool that takes a C library header (".h") and generates Haskell Foreign Function Interface import declarations for items (functions, structures, etc.) the header defines.

MissingPy

MissingPy is really two libraries in one. At its lowest level, MissingPy is a library designed to make it easy to call into Python from Haskell. It provides full support for interpreting arbitrary Python code, interfacing with a good part of the Python/C API, and handling Python objects. It also provides tools for converting between Python objects and their Haskell equivalents. Memory management is handled for you, and Python exceptions get mapped to Haskell Dynamic exceptions. At a higher level, MissingPy contains Haskell interfaces to some Python modules.

HsLua

A Haskell interface to the Lua scripting language

Foreign calls

We begin by learning how to call C routines from Haskell and Haskell definitions from C. The first example consists of three files:

main.hs:

{-# LANGUAGE ForeignFunctionInterface #-}

main = do print "Hello from main"
          c_routine

haskell_definition = print "Hello from haskell_definition"

foreign import ccall safe "prototypes.h"
    c_routine :: IO ()

foreign export ccall
    haskell_definition :: IO ()

vile.c:

#include <stdio.h>
#include "prototypes.h"

void c_routine (void)
{
  printf("Hello from c_routine\n");
  haskell_definition();
}

prototypes.h:

extern void c_routine (void);
extern void haskell_definition (void);

It may be compiled and linked in one step by ghc:

 ghc --make main.hs vile.c

Or, you may compile C module(s) separately and link in ".o" files (this may be preferable if you use make and don't want to recompile unchanged sources; ghc's --make option provides smart recompilation only for ".hs" files):

 ghc -c vile.c
 ghc --make main.hs vile.o

You may use gcc/g++ directly to compile your C/C++ files but I recommend to do linking via ghc because it adds a lot of libraries required for execution of Haskell code. For the same reason, even if main in your program is written in C/C++, I recommend calling it from the Haskell action main - otherwise you'll have to explicitly init/shutdown the GHC RTS (run-time system).

We use the foreign import declaration to import foreign routines into Haskell, and foreign export to export Haskell definitions "outside" for imperative languages to use. Note that import creates a new Haskell symbol (from the external one), while export uses a Haskell symbol previously defined. Technically speaking, both types of declarations create a wrapper that converts the names and calling conventions from C to Haskell or vice versa.

All about the foreign declaration

The ccall specifier in foreign declarations means the use of the C (not C++ !) calling convention. This means that if you want to write the external routine in C++ (instead of C) you should add export "C" specification to its declaration - otherwise you'll get linking errors. Let's rewrite our first example to use C++ instead of C:

prototypes.h:

#ifdef __cplusplus
extern "C" {
#endif

extern void c_routine (void);
extern void haskell_definition (void);

#ifdef __cplusplus
}
#endif

Compile it via:

 ghc --make main.hs vile.cpp

where "vile.cpp" is just a renamed copy of "vile.c" from the first example. Note that the new "prototypes.h" is written to allow compiling it both as C and C++ code. When it's included from "vile.cpp", it's compiled as C++ code. When GHC compiles "main.hs" via the C compiler (enabled by the -fvia-C option), it also includes "prototypes.h" but compiles it in C mode. It's why you need to specify ".h" files in foreign declarations - depending on which Haskell compiler you use, these files may be included to check consistency of C and Haskell declarations.

The quoted part of the foreign declaration may also be used to give the import or export another name - for example,

foreign import ccall safe "prototypes.h CRoutine"
    c_routine :: IO ()

foreign export ccall "HaskellDefinition"
    haskell_definition :: IO ()

specifies that:

• the C routine called CRoutine will become known as c_routine in Haskell,
• while the Haskell definition haskell_definition will be known as HaskellDefinition in C.

It's required when the C name doesn't conform to Haskell naming requirements.

Although the Haskell FFI standard tells about many other calling conventions in addition to ccall (e.g. cplusplus, jvm, net) current Haskell implementations support only ccall and stdcall. The latter, also called the "Pascal" calling convention, is used to interface with WinAPI:

foreign import stdcall unsafe "windows.h SetFileApisToOEM"
  setFileApisToOEM :: IO ()

And finally, about the safe/unsafe specifier: a C routine imported with the unsafe keyword is called directly and the Haskell runtime is stopped while the C routine is executed (when there are several OS threads executing the Haskell program, only the current OS thread is delayed). This call doesn't allow recursively entering back into Haskell by calling any Haskell definition - the Haskell RTS is just not prepared for such an event. However, unsafe calls are as quick as calls in C. It's ideal for "momentary" calls that quickly return back to the caller.

When safe is specified, the C routine is called in a safe environment - the Haskell execution context is saved, so it's possible to call back to Haskell and, if the C call takes a long time, another OS thread may be started to execute Haskell code (of course, in threads other than the one that called the C code). This has its own price, though - around 1000 CPU ticks per call.

You can read more about interaction between FFI calls and Haskell concurrency in [7].

Marshalling simple types

Calling by itself is relatively easy; the real problem of interfacing languages with different data models is passing data between them. In this case, there is no guarantee that Haskell's Int is represented in memory the same way as C's int, nor Haskell's Double the same as C's double and so on. While on some platforms they are the same and you can write throw-away programs relying on these, the goal of portability requires you to declare foreign imports and exports using special types described in the FFI standard, which are guaranteed to correspond to C types. These are:

import Foreign.C.Types (               -- equivalent to the following C type:
         CChar, CUChar,                --  char/unsigned char
         CShort, CUShort,              --  short/unsigned short
         CInt, CUInt, CLong, CULong,   --  int/unsigned/long/unsigned long
         CFloat, CDouble...)           --  float/double

Now we can typefully import and export to and from C and Haskell:

foreign import ccall unsafe "math.h"
    c_sin :: CDouble -> CDouble

Note that C routines which behave like pure functions (those whose results depend only on their arguments) are imported without IO in their return type. The const specifier in C is not reflected in Haskell types, so appropriate compiler checks are not performed.

All these numeric types are instances of the same classes as their Haskell cousins (Ord, Num, Show and so on), so you may perform calculations on these data directly. Alternatively, you may convert them to native Haskell types. It's very typical to write simple wrappers around foreign imports and exports just to provide interfaces having native Haskell types:

-- |Type-conversion wrapper around c_sin
sin :: Double -> Double
sin = fromRational . c_sin . toRational

Memory management

Marshalling strings

import Foreign.C.String (   -- representation of strings in C
         CString,           -- = Ptr CChar
         CStringLen)        -- = (Ptr CChar, Int)

foreign import ccall unsafe "string.h"
    c_strlen :: CString -> IO CSize     -- CSize defined in Foreign.C.Types and is equal to size_t

-- |Type-conversion wrapper around c_strlen
strlen :: String -> Int
strlen = ....

Marshalling composite types

A C array may be manipulated in Haskell as StorableArray.

There is no built-in support for marshalling C structures and using C constants in Haskell. These are implemented in the c2hs preprocessor, though.

Binary marshalling (serializing) of data structures of any complexity is implemented in the library module "Binary".

Dynamic calls

DLLs

because i don't have experience of using DLLs, can someone write into this section? Ultimately, we need to consider the following tasks:

• using DLLs of 3rd-party libraries (such as ziplib)
• putting your own C code into a DLL to use in Haskell
• putting Haskell code into a DLL which may be called from C code

The dark side of the I/O monad

Unless you are a systems developer, postgraduate CS student, or have alternate (and eminent!) verifiable qualifications you should have no need whatsoever for this section - here is just one tiny example of what can go wrong if you don't know what you are doing. Look for other solutions!

unsafePerformIO

Do you remember this definition?

getChar >>= \c -> c

Let's try to "define" something with it:

getchar :: Char
getchar = getChar >>= \c -> c

get2chars :: String
get2chars = [a, b] where a = getchar
                         b = getchar

But what makes all of that so wrong? Besides getchar and get2chars not being I/O actions:

1. Because the Haskell compiler treats all functions as pure (not having side effects), it can avoid "unnecessary" calls to getchar and use one returned value twice;
2. Even if it does make two calls, there is no way to determine which call should be performed first. Do you want to return the two characters in the order in which they were read, or in the opposite order? Nothing in the definition of get2chars answers this question.

Despite these problems, programmers coming from an imperative language background often look for a way to do this - disguise one or more I/O actions as a pure definition. Having seen procedural entities similar in appearance to:

void putchar(char c);

the thought of just writing:

putchar :: Char -> ()
putchar c = ...

would definitely be more appealing - for example, defining readContents as though it were a pure function:

readContents :: Filename -> String

will certainly simplify the code that uses it. However, those exact same problems are also lurking here:

1. Attempts to read the contents of files with the same name can be factored (i.e. reduced to a single call) despite the fact that the file (or the current directory) can be changed between calls. Haskell considers all non-IO functions to be pure and feels free to merge multiple calls with the same parameters.
2. This call is not inserted in a sequence of I/O actions all the way from main so the compiler doesn't know at what exact moment you want to execute this action. For example, if the file has one kind of contents at the beginning of the program and another at the end - which contents do you want to see? You have no idea when (or even if) this function is going to get invoked, because Haskell sees this function as pure and feels free to reorder the execution of any or all pure functions as needed.

So, implementing supposedly-pure functions that interact with the Real World is considered to be Bad Behavior. Nice programmers never do it ;-)

Nevertheless, there are (semi-official) ways to use I/O actions inside of pure functions - there is a (ahem) "special" definition that will (mis)use the Haskell implementation to run an I/O action. This particular (and dangerous) is:

unsafePerformIO :: IO a -> a

Using unsafePerformIO, you could easily write "pure-looking functions" that actually do I/O inside. But don't do this without a real need, and remember to follow this rule:

• the compiler doesn't know that you are cheating; it still considers each non-IO function to be a pure one. Therefore, all the usual optimization rules can (and will!) be applied to its execution.

So you must ensure that:

• The result of each call depends only on its arguments.
• You don't rely on side-effects of this function, which may be not executed if its results are not needed.

Let's investigate this problem more deeply. Function evaluation in Haskell is determined by a value's necessity - the language computes only the values that are really required to calculate the final result. But what does this mean with respect to the main action? To run it to completion, all the intermediate I/O actions that are included in main's chain need to be run. By using unsafePerformIO we call I/O actions outside of this chain. What guarantee do we have that they will be run at all? None. The only time they will be run is if running them is required to compute the overall function result (which in turn should be required to perform some action in the main chain). This is an example of Haskell's evaluation-by-need strategy. Now you should clearly see the difference:

• An I/O action inside an I/O definition is guaranteed to execute as long as it is (directly or indirectly) inside the main chain - even when its result isn't used (because it will be run anyway). You directly specify the order of the action's execution inside the I/O definition.

• An I/O action called by unsafePerformIO will be performed only if its result is really used. The evaluation order is not guaranteed and you should not rely on it (except when you're sure about whatever data dependencies may exist).

I should also say that inside the unsafePerformIO call you can organize a small internal chain of I/O actions with the help of the same binding operators and/or do syntactic sugar we've seen above. So here's how we'd rewrite our previous (pure!) definition of one using unsafePerformIO:

one :: Integer
one = unsafePerformIO $ do var <- newIORef 0
                           modifyIORef var (+1)
                           readIORef var

and in this case all the I/O actions in this chain will be run when the result of the unsafePerformIO call is needed.

inlinePerformIO

The internal code for inlinePerformIO is similar to that of unsafePerformIO, sometimes having an INLINE pragma. Semantically inlinePerformIO = unsafePerformIO in as much as either of those have any semantics at all.

The difference of course is that inlinePerformIO is even less safe than unsafePerformIO. While ghc will try not to duplicate or common up different uses of unsafePerformIO, we aggressively inline inlinePerformIO. So you can really only use it where the I/O content is really properly pure, like reading from an immutable memory buffer (as in the case of ByteStrings). However things like allocating new buffers should not be done inside inlinePerformIO since that can easily be floated out and performed just once for the whole program, so you end up with many things sharing the same buffer, which would be bad.

So the rule of thumb is that I/O actions wrapped in unsafePerformIO have to be externally pure while with inlinePerformIO it has to be really, really pure or it'll all go horribly wrong.

That said, here's some really hairy code. This should frighten any pure functional programmer...

write :: Int -> (Ptr Word8 -> IO ()) -> Put ()
write !n body = Put $ \c buf@(Buffer fp o u l) ->
  if n <= l
    then write</code> c fp o u l
    else write</code> (flushOld c n fp o u) (newBuffer c n) 0 0 0

  where {-# NOINLINE write</code> #-}
        write</code> c !fp !o !u !l =
          -- warning: this is a tad hardcore
          inlinePerformIO
            (withForeignPtr fp
              (\p -> body $! (p `plusPtr` (o+u))))
          `seq` c () (Buffer fp o (u+n) (l-n))

it's used like:

word8 w = write 1 (\p -> poke p w)

This does not adhere to my rule of thumb above. Don't ask exactly why we claim it's safe :-) (and if anyone really wants to know, ask Ross Paterson who did it first in the Builder monoid)

unsafeInterleaveIO

But there is an even stranger operation:

unsafeInterleaveIO :: IO a -> IO a

and here's one clear reason why:

{-# NOINLINE unsafeInterleaveIO #-}
unsafeInterleaveIO   :: IO a -> IO a
unsafeInterleaveIO a =  return (unsafePerformIO a)

So don't let that type signature fool you - unsafeInterleaveIO also has to be used carefully! It too sets up its unsuspecting parameter to run lazily, instead of running in the main action chain, with the only difference being the result of running the parameter can only be used by another I/O action. But this is of little benefit - ideally the parameter and the main action chain should have no other interactions with each other, otherwise things can get ugly!

At least you have some appreciation as to why unsafeInterleaveIO is, well unsafe! Just don't ask - to talk further is bound to cause grief and indignation. I won't say anything more about this ruffian I...use all the time (darn it!)

One can use unsafePerformIO (not unsafeInterleaveIO) to perform I/O operations not in some predefined order but by demand. For example, the following code:

do let c = unsafePerformIO getChar
   do_proc c

will perform the getChar I/O call only when the value of c is really required by the calling code, i.e. it this call will be performed lazily like any regular Haskell computation.

Now imagine the following code:

do let s = [unsafePerformIO getChar, unsafePerformIO getChar, unsafePerformIO getChar]
   do_proc s

The three characters inside this list will be computed on demand too, and this means that their values will depend on the order they are consumed. It is not what we usually want.

unsafeInterleaveIO solves this problem - it performs I/O only on demand but allows you to define the exact internal execution order for parts of your data structure.

• unsafeInterleaveIO accepts an I/O action as a parameter and returns another I/O action as the result:

do str <- unsafeInterleaveIO myGetContents
                    ⋮

• unsafeInterleaveIO doesn't perform any action immediately, it only creates a closure of type a which upon being needed will perform the action specified as the parameter.

• this action by itself may compute the whole value immediately...or use unsafeInterleaveIO again to defer calculation of some sub-components:

myGetContents = do
   c <- getChar
   s <- unsafeInterleaveIO myGetContents
   return (c:s)

This code will be executed only at the moment when the value of str is really demanded. In this moment, getChar will be performed (with its result assigned to c) and a new lazy-I/O closure will be created - for s. This new closure also contains a link to a myGetContents call.

The resulting list is then returned. It contains the Char that was just read and a link to another myGetContents call as a way to compute the rest of the list. Only at the moment when the next value in the list is required will this operation be performed again.

As a final result, we can postpone the read of the second Char in the list before the first one, but have lazy reading of characters as a whole - bingo!

PS: of course, actual code should include EOF checking; also note that you can read multiple characters/records at each call:

myGetContents = do
   l <- replicateM 512 getChar
   s <- unsafeInterleaveIO myGetContents
   return (l++s)

and we can rewrite myGetContents to avoid needing to use unsafeInterleaveIO where it's called:

myGetContents = unsafeInterleaveIO $ do
   l <- replicateM 512 getChar
   s <- myGetContents
   return (l++s)

Welcome to the machine: taking off the covers

A little disclaimer: I should say that I'm not describing here exactly what a monad is (I don't even completely understand it myself) and my explanation shows only one possible way to implement the I/O monad in Haskell. For example, the hbc compiler and the Hugs interpreter implements the I/O monad via continuations [9]. I also haven't said anything about exception handling, which is a natural part of the "monad" concept. You can read the All About Monads guide to learn more about these topics.

But there is some good news: the I/O monad understanding you've just acquired will work with any implementation and with many other monads.

The GHC implementation

newtype IO a = IO (State# RealWorld -> (# State# RealWorld, a #))

It uses the State# RealWorld type and the strict tuple type (# ... #) for optimization. It also uses an IO data constructor. Nevertheless, there are no significant changes from the standpoint of our explanation.

Of course, other compilers e.g. yhc/nhc (jhc, too?) define IO in other ways.

The Yhc/nhc98 implementation

data World = World
newtype IO a = IO (World -> Either IOError a)

This implementation makes the World disappear somewhat[10], and returns Either a result of type a, or if an error occurs then IOError. The lack of the World on the right-hand side of the function can only be done because the compiler knows special things about the IO type, and won't overoptimise it.

Further reading

[1] This manual is largely based on Simon Peyton Jones's paper Tackling the awkward squad: monadic input/output, concurrency, exceptions, and foreign-language calls in Haskell. I hope that my manual improves his original explanation of the Haskell I/O system and brings it closer to the point of view of new Haskell programmers. But if you need to learn about concurrency, exceptions and the FFI in Haskell/GHC, the original paper is the best source of information.

[2] You can find more information about concurrency, the FFI and STM at the GHC/Concurrency#Starting points page.

[3] The Arrays page contains exhaustive explanations about using mutable arrays.

[4] Look also at the Using monads page, which contains tutorials and papers really describing these mysterious monads.

[5] An explanation of the basic monad functions, with examples, can be found in the reference guide A tour of the Haskell Monad functions, by Henk-Jan van Tuyl.

[6] Official FFI specifications can be found on the page The Haskell 98 Foreign Function Interface 1.0: An Addendum to the Haskell 98 Report

[7] Using the FFI in multithreaded programs is described in Extending the Haskell Foreign Function Interface with Concurrency

[8] This particular behaviour is not a requirement of Haskell 2010, so the operation of seq may differ between various Haskell implementations - if you're not sure, staying within the I/O monad is the safest option.

[9] How to Declare an Imperative by Phil Wadler provides an explanation of how this can be done.

Do you have more questions? Ask in the haskell-cafe mailing list.

To-do list

If you are interested in adding more information to this manual, please add your questions/topics here.

Topics:

• fixIO and mdo
• Q monad

Questions:

• split (>>=)/(>>)/return section and do section, more examples of using binding operators
• IORef detailed explanation (==const*), usage examples, syntax sugar, unboxed refs
• explanation of how the actual data "in" mutable references are inside GHC's RealWorld, rather than inside the references themselves (IORef, IOArray & co.)
• control structures developing - much more examples
• unsafePerformIO usage examples: global variable, ByteString, other examples
• how unsafeInterLeaveIO can be seen as a kind of concurrency, and therefore isn't so unsafe (unlike unsafeInterleaveST which really is unsafe)
• discussion about different senses of safe/unsafe (like breaking equational reasoning vs. invoking undefined behaviour (so can corrupt the run-time system))
• actual code used by GHC - how to write low-level definitions based on example of how newIORef is implemented

This manual is collective work, so feel free to add more information to it yourself. The final goal is to collectively develop a comprehensive manual for using the I/O monad.

---