a tutorial

A Bayesian Neural NetworkOn linear types and

exceptionsTweag Fellowships:

Funding for Open Source ContributorsSafe memory management in inline-java

using linear typesLocating Performance Bottlenecks in

Large Haskell codebases

21 December 2017 |

|An important device in the tool belt I carry around everyday is type class reflection. I don't reach for it often, but it can be very useful. Reflection is a little known device. And for some reason it is often spoken of with a hint of fear.

In this post, I want to convince you that reflection is not hard and that you ought to know about it. To that end, let me invite you to join me on a journey to sort a list:

```
sortBy :: (a->a->Ordering) -> [a] -> [a]
```

Type class reflection is an extension of Haskell which makes it possible to use a value as a type class instance. There is a package on Hackage, implementing type class reflection for GHC, which I will use for this tutorial. Type class reflection being an extension of Haskell (that is, it can't be defined from other Haskell features), this implementation is GHC-specific and will probably not work with another compiler.

This blog post was generated from literate Haskell sources. You can find an extracted Haskell source file here.

There is a bit of boilerplate to get out of the way before we start.

```
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE UndecidableInstances #-}
module Reflection where
import Data.Proxy
import Data.Reflection
```

`UndecidableInstances`

...
scary, I know. It is unfortunately required. It means that we could
technically send the type checker into an infinite loop. Of course, we
will be careful not to introduce such loops.

My goal, today, is to sort a list. In order to make the exercise
a tiny bit interesting, I will use types to enforce invariants. I'll
start by introducing a type of *sorted* lists.

```
newtype SortedList a = Sorted [a]
```

Obviously, a `SortedList`

is a list: we can just forget about its
sortedness.

```
forget :: SortedList a -> [a]
forget (Sorted l) = l
```

But how does one construct a sorted list? Well, at the very least, the empty lists and the lists of size 1 are always sorted.

```
nil :: SortedList a
nil = Sorted []
singleton :: a -> SortedList a
singleton a = Sorted [a]
```

What about longer lists though? We could go about it in several ways. Let's decide to take the union of two sorted list:

```
merge :: Ord a => SortedList a -> SortedList a -> SortedList a
merge (Sorted left0) (Sorted right0) = Sorted $ mergeList left0 right0
where
-- 'mergeList l1 l2' returns a sorted permutation of 'l1++l2' provided
-- that 'l1' and 'l2' are sorted.
mergeList :: Ord a => [a] -> [a] -> [a]
mergeList [] right = right
mergeList left [] = left
mergeList [email protected](a:l) [email protected](b:r) =
if a <= b then
a : (mergeList l right)
else
b : (mergeList left r)
```

We *need* `Ord a`

to hold in order to define `merge`

. Indeed, type
classes are *global* and coherent: there is only one `Ord a`

instance,
and it is *guaranteed* that `merge`

always uses the same comparison
function for `a`

. This enforces that if `Ord a`

holds, then
`SortedList a`

represents lists of `a`

sorted according to the order
defined by the unique `Ord a`

instance. In contrast, a function argument
defining an order is *local* to this function call. So if `merge`

were
to take the ordering as an extra argument, we could change the order for
each call of `merge`

; we couldn't even state that `SortedList a`

are
sorted.

That's it! we are done writing unsafe code. We can sort lists with the
`SortedList`

interface: we simply need to split the list in two parts,
sort said parts, then merge them (you will have recognised merge
sort).

```
fromList :: Ord a => [a] -> SortedList a
fromList [] = nil
fromList [a] = singleton a
fromList l = merge orderedLeft orderedRight
where
orderedLeft = fromList left
orderedRight = fromList right
(left,right) = splitAt (div (length l) 2) l
```

Composing with `forget`

, this gives us a sorting function

```
sort :: Ord a => [a] -> [a]
sort l = forget (fromList l)
```

Though that's not quite what we had set out to write. We wanted

```
sortBy :: (a->a->Ordering) -> [a] -> [a]
```

It is easy to define `sort`

from `sortBy`

(`sort = sortBy compare`

). But
we needed the type class for type safety of the `SortedList`

interface.
What to do? We would need to use a value as a type class instance. Ooh!
What may have sounded excentric when I first brought it up is now
exactly what we need!

As I said when I discussed the type of `merge`

: one property of type
classes is that they are globally attached to a type. It may seem
impossible to implement `sortBy`

in terms of `sort`

: if I use
`sortBy myOrd :: [a] -> [a]`

and `sortBy myOtherOrd :: [a] -> [a]`

on the
same type, then I am creating two different instances of `Ord a`

. This
is forbidden.

So what if, instead, we created an *entirely new* type each time we need
an order for `a`

. Something like

```
newtype ReflectedOrd a = ReflectOrd a
```

Except that we can't do a `newtype`

every time we call `sortBy`

. So
let's make one `newtype`

once and for all, with an additional parameter.

```
newtype ReflectedOrd s a = ReflectOrd a
-- | Like `ReflectOrd` but takes a `Proxy` argument to help GHC with unification
reflectOrd :: Proxy s -> a -> ReflectedOrd s a
reflectOrd _ a = ReflectOrd a
unreflectOrd :: ReflectedOrd s a -> a
unreflectOrd (ReflectOrd a) = a
```

Now, we only have to create a new parameter `s`

locally at each `sortBy`

call. This is done like this:

```
reifyOrd :: (forall s. Ord (ReflectedOrd s a) => …) -> …
```

What is happening here? The `reifyOrd`

function takes an argument which
works for *any* `s`

. In particular, if every time we called `reifyOrd`

we were to actually use a different `s`

then the program would be
correctly typed. Of course, we're not actually creating types: but it is
safe to reason just as if we were! For instance if you were to call
`reifyOrd (reifyOrd x)`

then `x`

would have two distinct parameters `s1`

and `s2`

: `s1`

and `s2`

behave as names for two different types.
Crucially for us, this makes `ReflectOrded s1 a`

and `ReflectOrded s2 a`

two distinct types. Hence their `Ord`

instance can be different. This is
called a rank 2 quantification.

In order to export a single `reify`

function, rather than one for every
type class, the `reflection`

package introduces a generic type class so
that you have:

```
reify :: forall d r. d -> (forall s. Reifies s d => Proxy s -> r) -> r
```

Think of `d`

as a *dictionary* for `Ord`

, and `Reifies s d`

as a way to
retrieve that dictionary. The `Proxy s`

is only there to satisfy the
type-checker, which would otherwise complain that `s`

does not appear
anywhere. To reiterate: we can read `s`

as a unique generated type which
is valid only in the scope of the `reify`

function. For completeness,
here is the the `Reifies`

type class, which just gives us back our `d`

:

```
class Reifies s d | s -> d where
reflect :: proxy s -> d
```

The `| s -> d`

part is called a functional
dependency. It is
used by GHC to figure out which type class instance to use; we won't
have to think about it.

All that's left to do is to use reflection to give an `Ord`

instance to
`ReflectedOrd`

. We need a dictionary for `Ord`

: in order to build an
`Ord`

instance, we need an equality function for the `Eq`

subclass, and
a comparison function for the instance proper:

```
data ReifiedOrd a = ReifiedOrd {
reifiedEq :: a -> a -> Bool,
reifiedCompare :: a -> a -> Ordering }
```

Given a dictionary of type `ReifiedOrd`

, we can define instances for
`Eq`

and `Ord`

of `ReflectedOrd`

. But since type class instances only
take type class instances as an argument, we need to provide the
dictionary as a type class. That is, using `Reifies`

.

```
instance Reifies s (ReifiedOrd a) => Eq (ReflectedOrd s a) where
(==) (ReflectOrd x) (ReflectOrd y) =
reifiedEq (reflect (Proxy :: Proxy s)) x y
instance Reifies s (ReifiedOrd a) => Ord (ReflectedOrd s a) where
compare (ReflectOrd x) (ReflectOrd y) =
reifiedCompare (reflect (Proxy :: Proxy s)) x y
```

Notice that because of the `Reifies`

on the left of the instances GHC
does not know that it will for sure terminate during type class
resolution (hence the use of `UndecidableInstances`

). However, these are
indeed global instances: by definition, they are the only way to have an
`Ord`

instances on the `ReflectedOrd`

type! Otherwise GHC would
complain.

We are just about done: if we `reify`

a `ReifiedOrd a`

, we have
a scoped instance of `Ord (ReflectedOrd s a)`

(for some locally
generated `s`

). To sort our list, we simply need to convert between
`[a]`

and `ReflectedOrd s a`

.

```
sortBy :: (a->a->Ordering) -> [a] -> [a]
sortBy ord l =
reify (fromCompare ord) $ \ p ->
map unreflectOrd . sort . map (reflectOrd p) $ l
-- | Creates a `ReifiedOrd` with a comparison function. The equality function
-- is deduced from the comparison.
fromCompare :: (a -> a -> Ordering) -> ReifiedOrd a
fromCompare ord = ReifiedOrd {
reifiedEq = \x y -> ord x y == EQ,
reifiedCompare = ord }
```

We've reached the end of our journey. And we've seen along the way that
we can enjoy the safety of type classes, which makes it safe to write
function like `merge`

in Haskell, while still having the flexibility to
instantiate the type class from a function argument, such as options
from the command line. Since type class instances are global, such local
instances are defined globally for locally generated types. This is what
type class reflection is all about.

If you want to delve deeper into the subject of type class reflection, let me, as I'm wrapping up this tutorial, leave you with a few pointers to further material:

- A talk by Edward Kmett, the author of the reflection package, on the importance of the global coherence of type classes and about reflection
- There is no built-in support for reflection in GHC, this tutorial
by Austin
Seipp
goes over the
*very unsafe*, internal compiler representation dependent, implementation of the library - John Wiegley discusses an application of reflection in relation with QuickCheck.
- You may have noticed, in the definition of
`sortBy`

, that we`map`

the`reflectOrd`

and`unreflectOrd`

in order to convert between`a`

and`ReflectedOrd s a`

. However, while,`reflectOrd`

and`unreflectOrd`

, have no computational cost, using them in combination with`map`

will traverse the list. If you are dissatified with this situation, you will have to learn about the Coercible type class. I would start with this video from Simon Peyton Jones.