Module 14: Monads
This module is due Monday, December 3 at 4:00pm.
import Control.Monad
import Control.Monad.State
import Control.Monad.Except
import Control.Monad.Writer
Recall that we’ve seen the (>>=)
operator with the following type:
(>>=) :: Either Error a -> (a -> Either Error b) -> Either Error b
This is actually an instance of a more general pattern called a monad. A monad is a type m
which supports the following operations:
return :: a -> m a
(>>=) :: m a -> (a -> m b) -> m b
Intuitively, a value of type m a
represents a computation or program that computes a value of type a
, and might have some “effects” (depending on m
).
return
creates a computation that immediately results in the given value, with no effects. (In the case ofEither Error
,return = Right
.)(>>=)
sequences two computations together into one bigger one, where the second computation can depend on the output of the first.
Let’s look at a simple example. The following Possibly
type is the same as the standard Maybe
, type, but redefined so you can practice making a Monad
instance for it (of course, Maybe
already has a Monad
instance defined in the standard library).
Note that every Monad
is also a Functor
(which generalizes the <$>
operator we saw for parsers), as well as an Applicative
(which generalizes the <*>
operator). (Because of time constraints, we won’t spend time talking about those but see here if you’re interested in learning more about them, or take CSCI 365, Functional Programming.) GHC can automatically generate a default instance of Functor
for us (see deriving Functor
above), and we can make an Applicative
instance for free once we have a Monad
instance: we just define pure = return
and (<*>) = ap
(ap
is a function from the standard library which is defined in terms of return
and (>>=)
).
- Your turn: fill in a defintion for the
Monad Possibly
instance below!
Let’s look at a simple example. First, checkEven
returns Just n
when its argument n
is even, and Nothing
otherwise. That is, it is a sort of “filter” that only “lets through” even numbers.
addEvens
takes two Int
values as input and adds them if both are even (and fails otherwise). Here’s a straightforward first version:
addEvens :: Int -> Int -> Maybe Int
addEvens m n =
checkEven m >>= (\me -> checkEven n >>= (\ne -> return (me + ne)))
We actually don’t need the parentheses, since anonymous functions extend as far to the right as possible:
addEvens2 :: Int -> Int -> Maybe Int
addEvens2 m n =
checkEven m >>= \me -> checkEven n >>= \ne -> return (me + ne)
It’s annoying to have everything on the same line; it’s hard to read, and quickly gets unwieldy with larger compositions of actions. So we can introduce some newlines to make things more readable, like this:
addEvens3 :: Int -> Int -> Maybe Int
addEvens3 m n =
checkEven m >>= \me ->
checkEven n >>= \ne ->
return (me + ne)
This style became so common that it was decided to add some special syntax sugar, called do-notation, to Haskell. We can rewrite addEvens
like so:
addEvens4 :: Int -> Int -> Maybe Int
addEvens4 m n = do
me <- checkEven m
ne <- checkEven n
return (me + ne)
addEvens4
desugars to exactly the same thing as addEvens3
.
Another monad: State
Let’s look at another example of a Monad
, one we haven’t seen before. The type State s
is a Monad
which represents the use of a mutable state of type s
. That is, a value of type State s a
is a computation that results in an a
and has access to a mutable state of type s
.
In addition to return :: a -> State s a
and (>>=) :: State s a -> (a -> State s b) -> State s b
, the State
monad supports two additional operations:
get :: State s s
returns the current state as its output.put :: s -> State s ()
sets the state to a new value. It returns()
(“unit”), which is an uninformative type with only a single value (also called()
). Essentially a return type of()
means it “does not return anything”, i.e. it is the same as avoid
method in C or Java.
Let’s try some examples.
- Write a function called
tick
which adds one to the current integer state and returns the old state. For example, if the state is 6,tick
will update the state to 7 and then return 6.
THINGS WE HAVE:
- get :: State s s
- put :: s -> State s ()
- return :: a -> State s a
- (>>=) :: State s a -> (a -> State s b) -> State s b
tick''' :: State Int Int
tick''' =
get >>= \i ->
put (i+1) *>
return i
tick'''' :: State Int Int
tick'''' = do
i <- get
put (i+1)
return i
To test your implementation of tick
and other exercises below, you can use the function
runState :: State s a -> s -> (a,s)
which takes a State
computation to run and an initial state value, and yields the return value and final state value. For example,
ghci> runState tick 4
(4,5)
This runs the tick
computation with an initial state of 4
, which results in the return value 4 and an updated state of 5. (You may also be interested in evalState
and execState
; ask GHCi for their types to see what they do.)
- Now write a function
tick3 :: State Int Int
which increments the state by 3 and returns the original state. For example, runningtick3
with an initial state of4
will return4
and update the state to7
. Do not use the number 3 or the+
operator in your solution; just calltick
three times.
State
also supports an operation calledmodify
, which applies a given function to the mutable state. Show how to implementmodify
in terms ofget
andput
:
Essentially, we can think of State
as a mini-EDSL for expressing programs that manipulate a single mutable variable.
A tiny stack EDSL
- ROTATE ROLES and write the name of the new driver here:
Let’s use the State
monad to build a tiny embedded DSL for operating on a mutable stack of integers.
- Complete the definition of the
StackProgI
EDSL below.
-- A 'StackProgI a' is a program which returns an 'a' and has access
-- to a mutable stack of Ints.
type StackProgI a = State [Int] a
-- Pop the top Int from the stack and return it. (For now, fail by
-- calling 'error' the stack is empty.)
popI :: StackProgI Int
popI = undefined
-- Look at the top Int on the stack without popping it. (Fail with 'error'
-- if the stack is empty.)
peekI :: StackProgI Int
peekI = undefined
-- Run a 'StackProgI a' starting with the empty stack, returning the
-- produced value of type 'a' along with the final stack state.
runStackProgI :: StackProgI a -> (a, [Int])
runStackProgI = undefined
Now let’s write a few programs using this new EDSL.
- Write
opI
, which pops the top two values from the stack, performs a given binary operation on them, and pushes the result back on the stack. (Hint: usepopI
andpushI
instead of manipulating the stack directly!)
- Write
pushListI
, which pushes all the elements in a list onto the stack. (Challenge: can you implementpushListI
with one line of code?)
- Write
crushI
, which repeatedly performs an operation on the top two items on the stack until there is only one item left. (Hint: don’t manipulate the stack directly; usesizeI
,opI
, and recursion to get the job done. You may also find thewhen
function useful.)
Be sure to test your programs using runStackProgI
. You will find pushListI
helpful for testing opI
and crushI
.
Error handling and monad transformers
- ROTATE ROLES and write the name of the new driver here:
It’s unsatisfactory having our stack language crash with error
on a stack underflow. In this section we’ll explore a more principled approach to error handling via monad transformers, which allow us to combine the effects of several monads into one.
Here’s a type for runtime errors that can arise from stack programs (for now there’s only Underflow
):
And now for the definition of StackProgE
, which adds the possibility of errors. While we’re at it, let’s generalize from stacks of Int
s to stacks of any type, represented by the type variable el
. That is, something of type StackProgE el a
is a computation which:
- returns a value of type
a
, - has a mutable stack of type
[el]
, and - can fail with a
StackError
.
ExceptT
is a monad transformer which adds the possibility of StackErrors
on top of the existing monad State [el]
.
Reimplement the stack EDSL by implementing
sizeE
,pushE
,popE
, andpeekE
. They should work like the corresponding methods forStackProgI
, but modified to throwStackError
s as appropriate instead of callingerror
.Here’s what you need to know:
StackProgE
is still aMonad
, of course.StackProgE
still has aState
component, so you can useput
,get
, andmodify
as before.You can use
throwError :: StackError -> StackProgE el a
to signal an error. Essentially, usethrowError
anywhere you would have usedLeft
, as if we were still working in terms of a type likeEither StackError
.
sizeE :: StackProgE el Int
sizeE = undefined
pushE :: el -> StackProgE el ()
pushE = undefined
popE :: StackProgE el el
popE = undefined
peekE :: StackProgE el el
peekE = undefined
- Now implement
runStackProgE
below. Follow the types! You will find therunExceptT
function helpful.
Test the operations that can throw an error to make sure they work properly. For example, what happens when you call
runStackProgE popE
?Finally, reimplement operations analogous to
opI
,pushListI
, andcrushI
. Ideally, if you implemented them elegantly in the first place, their implementations will not need to change other than updatingpopI
topopE
and so on:
opE :: (el -> el -> el) -> StackProgE el ()
opE = undefined
pushListE :: [el] -> StackProgE el ()
pushListE = undefined
crushE :: (el -> el -> el) -> StackProgE el ()
crushE = undefined
A deep stack EDSL
- ROTATE ROLES and write the name of the new driver here:
It was annoying that we had to reimplement everything when we switched from StackProgI
to StackProgE
. The solution is to use a deep embedding, which we can then interpret via multiple semantics. The cool thing is that we can make a Monad
instance for our deep embedding, and continue to write exactly the same programs as before, using do
-notation and so on. The difference is that our programs will now construct ASTs, which we can separately optimize, interpret, and so on.
Here is the AST for our deep embedding:
data StackProgAST el a where
-- A simple return value.
Return :: a -> StackProgAST el a
-- Push a value on the stack. This instruction stores the value
-- to push, and the rest of the program (i.e. it's a node with a
-- single child node).
Push :: el -> StackProgAST el a -> StackProgAST el a
-- Pop a value from the stack. Stores a function which, when
-- given the element that is popped, determines the rest of the
-- program. Another way to think of it is that a Pop node is like
-- an infinitely-branching tree node: there is one child AST node
-- for every possible element that could be popped.
Pop :: (el -> StackProgAST el a) -> StackProgAST el a
-- Peek at the value on the top of the stack.
Peek :: (el -> StackProgAST el a) -> StackProgAST el a
-- Get the size of the stack.
Size :: (Int -> StackProgAST el a) -> StackProgAST el a
deriving Functor
-- We get an Applicative instance for free from the Monad instance.
instance Applicative (StackProgAST el) where
pure = return
(<*>) = ap
Write a
Monad
instance forStackProgAST el
below. You can do it just by following the types, but here is the intuitive way to think about it: given an ASTt :: StackProgAST el a
, which is like a big (infinitely-branching) tree with values of typea
at all the leaves (wrapped inReturn
), and a functiona -> StackProgAST el b
which says how to continue from any value of typea
, the bind operatort >>= f
callsf
on every leaf value and splices in the resulting trees.This instance is definitely tricky to think about, although the code does not end up being that long in the end. If you are stuck, try carefully writing out all the types involved. Treat it like a jigsaw puzzle, figuring out how to put all the different pieces together so the types fit properly.
- Now write
size
,push
,pop
, andpeek
below, which each build a single AST node immediately followed by aReturn
node.
size :: StackProgAST el Int
size = undefined
push :: el -> StackProgAST el ()
push = undefined
pop :: StackProgAST el el
pop = undefined
peek :: StackProgAST el el
peek = undefined
Now reimplement
op
,pushList
, andcrush
. The implementations ofop
andcrush
should still be very similar to your previous implementations of these operations, even though they are now doing something very different (namely, splicing together ASTs instead of actually operating on stacks).The implementation of
pushList
, however, might have to change a bit. Previously, since you were implementingpushList
directly in terms of a state-based semantics with a stack, you could implementpushList
simply by modifying the state. Now, however,pushList
has to build an AST, and we don’t know in advance how the AST will be interpreted: the ultimate semantics might not even involve a mutable state at all! So instead, if you didn’t already, you will have to implementpushList
by turning it into a series of calls topush
(which will in turn build an AST containing a bunch ofPush
nodes).
op :: (el -> el -> el) -> StackProgAST el ()
op = undefined
pushList :: [el] -> StackProgAST el ()
pushList = undefined
crush :: (el -> el -> el) -> StackProgAST el ()
crush = undefined
- Now write
interpStackProgE
, which interprets an AST as aStackProgE
program.
- Finally, write
runAsStackProgE
, which combinesinterpStackProgE
withrunStackProgE
to directly run an AST. Test it by running programs such aspushList [3,2,6,4] >> crush (*)
.
An alternative semantics
- ROTATE ROLES and write the name of the new driver here:
Now that we have a deep EDSL, we can use different semantics to interpret it without having to rewrite all our existing programs.
Consider the following definition of StackProgW
, which adds WriterT [String]
as compared to StackProgE
. WriterT [String]
adds a write-only log of type [String]
, which we can use to generate a sequence of log messages.
You can use the function tell :: [String] -> StackProgW el ()
to add some logging messages. In particular, each pop or push operation should log a message like "Popped 3"
or "Pushed 6"
.
- Implement
interpStackProgW
, which interprets aStackProgAST
as aStackProgW
computation. (Hint: you will probably want to implementsizeW
,pushW
,popW
, andpeekW
functions which implement the operations specifically forStackProgW
, with appropriate logging operations added. You will probably need to addShow el
constraints, e.g.popW :: Show el => StackProgW el el
.)
- Finally, implement
runAsStackProgW
. Test it. For example, here is what my implementation does:>>> runAsStackProgW (pushList [1,2,3] >> crush (*)) ((Right (),["Pushed 1","Pushed 2","Pushed 3","Popped 3","Popped 2","Pushed 6","Popped 6","Popped 1","Pushed 6"]),[6])