Be Like Water

Shapeless is a Scala library for generic programming. The name “Shapeless” comes from a famous Bruce Lee quote:

Don’t get set into one form, adapt it and build your own, and let it grow, be like water. Empty your mind, be formless, shapeless — like water. Now you put water in a cup, it becomes the cup; You put water into a bottle it becomes the bottle; You put it in a teapot it becomes the teapot. Now water can flow or it can crash. Be water, my friend.

There have been many blog posts devoted to explaining the very basics of Shapeless, but, in my opinion, fewer posts show how to truly understand how to use Shapeless. In this post I am trying to give an overview, starting from a concrete situation. Hopefully, in the end, you will have a better picture of how to read a piece of code that makes extensive use of Shapeless and a Shapeless-like coding-style.

Table of Contents

Running Example: Stream Processing

Reactive stream frameworks are a fancy new way to do Dataflow programming, where a computation is modeled after the flow of data through the nodes of a pipeline, or, more generally, of a flow graph.

Each node of this graph computes a function on the data that comes in, and returns a new value that will be written onto an output stream. In the simplest case of a pipeline, each node has exactly one input and one output. That is, it computes a function T => R.

The initial and final nodes are an exception:

  • the initial node produces data without an actual input (a function () => R);

  • the final node consumes input without producing output (a function T => Unit);

Beside the initial and final nodes, which can be special-cased, the node of a pipeline may be generally described by the case class:

1
2
case class Node[T, R](f: T => R)

But, in a stream processing framework, the more interesting case is that of flow-graph rather than that of a simple pipeline. In this case, each node would compute a function of N>=1 parameters, where each parameter would be an in-edge into the node. In pseudo-code:

1
2
case class Node[T1, T2, ..., TK, R](f: (T1, T2, ..., TK) => R)

The challenge now would be supporting K-ary function with arbitrary K, without writing K different implementations. As you may know, tuples and functions in Scala are implemented by explicitly defining 22 variants, with up to 22 distinct type parameters, like so:

1
2
3
4
5
Function1[T,R] 
Function2[T1,T2,R] Tuple2[T1,T2]
... ...
Function22[T1,T2,...,T22,R] Tuple22[T1,T2,...,T22]

The limit is known to be somewhat arbitrary, but we are not debating this now. The point is, although the code for this could be auto-generated, we would like to write this once and for all, and

  1. without resorting to meta-programming (macros and the like)

  2. without needless code repetition

In other words, we would like to be able to abstract over the arity (number of arguments) of a function.

Abstracting over Arity

The Shapeless wiki includes a single, interesting example that actually solves our problem.

Conversions between tuples and HList’s, and between ordinary Scala functions of arbitrary arity and functions which take a single corresponding HList argument allow higher order functions to abstract over the arity of the functions and values they are passed

The following incantation does the trick:

1
2
3
4
5
6
7
8
9
10
11
12
13
import syntax.std.function._
import ops.function._

def applyProduct[P <: Product, F, L <: HList, R](p: P)(f: F)
 (implicit gen: Generic.Aux[P, L], fp: FnToProduct.Aux[F, L => R]) =
 f.toProduct(gen.to(p))

scala> applyProduct(1, 2)((_: Int)+(_: Int))
res0: Int = 3

scala> applyProduct(1, 2, 3)((_: Int)*(_: Int)*(_: Int))
res1: Int = 6

Now, we may copy and paste the piece of code above and call it a day. Instead, this entire post is devoted to achieve a more deep understanding of this short, but very dense snippet. First of all, let us rule out the type parameters. What is an HList?

HLists and Product Types

Many blog posts have been devoted to describe the ins and the outs of HLists. This is not one such post. The point is, HLists are not that hard a construct to understand and even to implement. The neat part about Shapeless is that many operations on HLists are provided out of the box.

I want to refer you to the many posts that describe how HLists can be implemented in Scala to find more detail: see the References section at the bottom of this post for a list. In this section I will instead explain how Shapeless HLists can be used as a substitute for tuples. In order to understand what will follow, the simple mental model that equates HLists to an alternative tuple implementation should suffice.

HLists (short for Heterogeneous List) are lists of objects of arbitrary types, where the type information for each object is kept. In fact, in Scala, we may do:

1
2
3
import shapeless._
val l = 10 :: "string" :: 1.0 :: Nil

but the type of l then would be List[Any], because the common super type of the elements there would be, in fact, only Any. A HList is declared similarly to a List:

1
2
val hl = 10 :: "string" :: 1.0 :: HNil

except for the terminator HNil. But the type of hl is actually Int :: String :: Double :: HNil. As you can see, no type information is lost.

You can probably think of another family of types in vanilla Scala that act as a container for values of different types; these types retain type information for each of their component, and the order in which they are declared: tuples.

HLists can be seen as an alternative implementation of the concept of Tuple. Or, more generally, of the concept of Product. It is not by accident that Scala tuples all extend the Product trait (in fact, so do case classes; more on this later). A Product is really just this: for some types T1, T2, … , TN (N>=1) the product type is the n-tuple (v1, v2, ..., vN) (with v1 an instance of type T1, etc.). When each field is named then the tuple is called a record (in vanilla Scala, case classes could be seen as some sort of record).

You already know that Scala tuples are a first-class concept; tuples are constructed through convenient literal syntax (v1, v2, ..., vN); using Shapeless, HLists are constructed using the syntax v1 :: v2 :: ... :: vN :: HNil.

The benefit of using HLists instead of tuples is that they can be used in all those situations where a tuple would work just as well, but without the 22-elements limit.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
val t = (1, "String", 2.0)
val hl = 1 :: "String" :: 2.0 :: HNil

val (a, b, c) = t
val x :: y :: z :: HNil = hl

t match {
 case (1, s, _) => ...
}

hl match {
 case 1 :: s :: _ :: HNil => ...
}

Moreover, Shapeless HLists provide many of the operations that can be usually applied to Lists such as map, flatMap etc. But, as a simple substitute for tuples, HLists are already quite valuable; for instance, it is not necessary to know the size of an HList before being able to match against it, rendering possible to do things like:

1
2
3
4
5
6
7
hl match {
 case 1 :: rest => ... // matches if first element 1, regardless the size of the list
 // and binds the tail to `rest`
 case x :: y :: HNil => ... // matches when it is a pair
 // etc. 
}

From Tuples to HLists and Back Again

Because Tuples and HLists are essentially the implementation of a similar concept, they are also isomorphic to eachother, that is, there is a morphism (a function) to go back and forth between them.

The way we go from HLists to Tuples is through the tupled method.

1
2
3
4
import syntax.std.product._
val hl = 1 :: "String" :: 2.0 :: HNil
val t = hl.tupled

The way we go from Tuples to HLists is through the productElements method. In fact, if hl is the HList, and t is the equivalent tuple, then it is always true that:

1
2
hl == t.productElements

Of course, a simple consequence is that you are able to abstract over the arity of tuples by transforming them into HLists

1
2
3
4
aTuple.productElements match {
 case a :: b :: HNil => ...
}

Whatsmore, because all tuples implement the Product trait and case classes implement the Product trait, many of the operations working for tuples also work for case classes.

1
2
3
4
5
case class Person(name: String, age: Int)
Person("John", 40).productElements match {
 case name :: age :: HNil => ...
}

It is even possible to create instances of a case class from an HList. Let us see how.

The Generic[T] object

A Generic[T] object implements the methods to(T) and from(HList) for a given product type T (usually, a case class or a tuple). For instance, the Generic object to convert back and forth between a Person and an HList is created and used as follows:

1
2
3
4
5
6
7
val gp = Generic[Person]
val john = Person("John", 40)
val hl: String :: Int :: HNil = gp.to(john)
val p: Person = gp.from(hl)

assert( john == p )

This is likewise true for tuples.

1
2
3
4
5
6
7
val gp = Generic[Tuple2[String,Int]]
val johnTuple = ("John", 40)
val hl: String :: Int :: HNil = gp.to(johnTuple)
val tp: Tuple2[String,Int] = gp.from(hl)

assert( johnTuple == tp )

In fact, the t.productElements method is really short-hand syntax for the code snippet above.

The FnToProduct[F] object

Suppose you have function f that takes K arguments, and that you want to conflate those arguments into one single product-type argument.

Scala provides f.tupled which turns a K-ary function into a unary function of one K-tuple argument. For instance:

1
2
3
val f = (s: String, i: Int) => s"$s $i"
val ft: Tuple2[String, Int] => String = f.tupled

But then, again, as we have seen already, you are not really able to abstract over the arity of Scala tuples. However, we can import FnToProduct[F] to turn a K-ary function into a function of K-sized HList of the same argument types.

1
2
3
4
import ops.function._
val fp = FnToProduct[(String, Int) => String]
val fhl: String::Int::HNil => String = fp.apply(f) // or, equivalently, fp(f)

in fact, there is even syntactic sugar for this:

1
2
3
import syntax.std.function._
val fhl: String::Int::HNil => String = f.toProduct

So now we have already sufficient elements to understand the following line of the Shapeless wiki.

1
2
f.toProduct(gen.to(p)) // could have been written as f.toProduct.apply(gen.to(p))

First of all, let us consider a special case; imagine f is the function we have just defined above, and let p = Person("John", 40). Now, the line above is a bit terse. Let us expand it a little bit:

1
2
3
4
5
val gen = Generic[Person]
val fhl: String :: Int :: HNil => String = f.toProduct
val hl: String :: Int :: HNil = gen.to(p)
fhl(hl)

  • we get gen, a Generic[Person] instance

  • we convert f to an HList-accepting function (f.toProduct);

  • we convert p to the HList hl by applying the to(Person) method of gen

There is still something missing, though. In the expanded version we have explicitly requested the Generic[Person] instanced to be created. The wiki uses implicit parameters. How do these implicits work? What is their relation to Generic and FnToProduct? And, do we really need them there?

Implicit Value Resolution

Implicits are a controversial feature of the Scala programming language. They are hard to grasp for beginners, and they can be cause of headaches even to the more experienced users. The choice of the word implicit is arguable. It probably contributed to add to Scala’s black magic fame.

You are probably already aware that implicits are values that, when in scope, are automatically inserted as required by the compiler. When one such implicit value is in scope, then the implicit arguments can be completely omitted by who is writing the code. For instance:

1
2
3
4
5
implicit val implicitJohn = Person("John", 40)
def somePersonString(implicit p: Person) = p.toString

somePersonString // returns "Person(John, 40)"

Implicits can be brought into scope by explicitly declaring them, as in the example above, by importing them from a library. The most simple use case for implicits is to provide fallback values, but they can be exploited for other advanced use, such as validating nontrivial constraints. An example of this is the encoding of dependent types.

implicit defs can be used to generate implicit values as needed. The most simple use for implicit defs is to provide implicit type conversions that are often used to define extension methods. Implicit defs are also used when implicit parameters of a function are being resolved to see if a matching value can be generated on-the-fly.

For instance, consider this example:

1
2
3
4
5
implicit def personProvider = Person("Random Guy", (math.random * 100).toInt)
def somePersonString(implicit p: Person) = p.toString

> somePersonString // Person("Random Guy", and a random integer between 0 and 100)

Let me repeat this once again: implicit parameter resolution is a compile-time procedure that does not affects run-time performance. Implicit values that are generated by implicit defs, on the other hand, may obviously affect run-time since they are instances of classes which may have arbitrary code in their constructors; on the other hand, the JVM is pretty good at optimizing out final classes, singletons and the like, but be advised that a matching implicit def is code that will actually execute at run time!

1
2
3
4
5
6
7
8
implicit def aSlowPersonProvider = {
Thread.sleep(3000) // faking a slow computation here
Person("Random Guy", (math.random * 100).toInt)
}
def somePersonString(implicit p: Person) = p.toString

> somePersonString // sits 3 secs, then returns the Person instance

Implicit parameter resolution can be twisted to enforce stronger constraints than simple type-checking. Parameterized types such as Generic.Aux[P,L] in the Shapeless wiki can be used to encode predicates about types, and relations between them. Your compiler checks for conformity to type signatures when you invoke methods or create class instances. In the case of implicits, though, the compiler attempts to fill in the voids automatically. The voids will be filled if and only if these relations hold.

In the previous example, we were asking the compiler to execute method somePersonString if and only if a value of type Person could be brought into scope. To put it in another way, we wanted the method invocation to compile if and only if the compiler could prove that a matching value (a value of that type) may exist.

When I saw this, I had one of those “aha!”-moments. The key to understanding implicit resolution in Scala, to me, was finding the parallel that exist with the evaluation of a program in a logic programming language such as Prolog.

A Short Prolog Digression

In the Prolog programming language you may state facts and provide rules to derive new facts. These facts and rules are kept in a space called a knowledge base (or KB), a database of all the facts and rules that have been defined.

The knowledge base, can be queried, like you would do on a SQL database, although with a different query language. And, like in any other query language, the execution of a Prolog program consists in solving the constraints that are found in the query.

For instance, let us say that John, Carl and Tom are all persons. We do this by writing the Prolog listing:

1
2
3
4
person(john).
person(carl).
person(tom).

These are all facts that we are stating about the atomic values john, carl, and tom (notice that they are all in small-letters; capitals are reserved to variables). You can now check whether john is a person with the query

You can also get all the persons in the KB:

Where X is a free variable (Prolog use capitals to denote variables), that is, a fresh variable that is not bound to any value. In this case we are asking the interpreter to find a concrete proof (or evidence or witness) that the condition holds in the KB. In other words, we want to find at least one binding of X for which person(X) is true; By hitting ; we can request the next binding.

1
2
3
4
5
?- person(X).
X = john ;
X = carl ;
X = tom

Let us now provide relations; that is K-ary facts about these persons.

1
2
3
child(john, carl).
child(carl, tom).

we can now instruct Prolog on how to derive further relations between these persons using rules. For instance, let us define the grandchild rule.

1
2
3
4
grandchild(A, B) :-
child(A, X),
child(X, B).

That is, A and B are in a grandchild relation if A is child of some X and X is a child of B. A rule describes an implication relation, in fact, this, in logic may be written as:

1
2
∀ A, B, X: child(A,X) ∧ child(X,B) → grandchild(A,B)

Notice that the implication is written in reverse, compared to the Prolog version. The Prolog version results more readable in code, because it makes more prominent the fact that the implication will derive. The interpreter can be queried in a number of ways. Again, small letters denote atoms, while capitals indicate free variables, we can choose any combination of the two to produce different results.

1
2
3
4
5
6
7
8
9
10
11
12
13
% find the pair(s) X,Y for which the grandchild relation holds
?- grandchild(X, Y).
X = john
Y = tom

% find the john's grandchildren
?- grandchild(john, Y).
Y = tom

% find tom's grandparents
?- grandchild(X, tom).
X = john

The system will also fail in impossible cases, that is, when we try to find something that is not derivable in the knowledge base. For instance, looking for someone who is his/her own grandchild

or looking for someone who is grandchild of tom (who has no grandchildren)

1
2
3
?- grandchild(tom, X)
no

or looking for the granparent of john, which is unknown in this knowledge base

1
2
3
?- grandchild(X, john)
no

GrandChild In Scala

In order to better understand the relation between Prolog and Scala, let us first introduce a dialect of Scala that we will call Logic Scala.

Logic Scala is a superset of Scala where two new keywords are introduced:

Let us see how an implementation of the grandchild example may look in Logic Scala. First, we have to introduce facts about children. The relation may be declared as a parameterized type:

1
2
trait Child[A,B] // notice that a class would work as well

where A,B will be substituted by atoms. Atoms are, again, types, so we may declare them as follows

1
2
3
4
trait John
trait Carl
trait Tom

We can now declare facts (remember, this is Logic Scala):

1
2
3
fact johncarl = new Child[John,Carl]{} // an instance of type Child
fact carltom = new Child[Carl,Tom ]{}

The rule for grandchild is made of two parts. The type declaration,

and the semantics of the rule, which is given using the rule construct. Just like in Prolog, our imaginary Scala dialect would describe how to derive new instances of the type GrandChild[A,B] from facts that are found (or that can be derived) in the current knowledge base. A rule is written similarly to a def:

1
2
3
4
5
6
rule grandChild[A,B,X]( 
facts
xy: Child[A,X],
yz: Child[X,B]
) = new GrandChild[A,B] {}

Notice that, because a rule declaration is syntactically similar to a def declaration, it is written in the same order of the abstract logic declaration, while the Prolog version is written in reverse. However, if we remove the syntactic noise, this is really stating the same:

1
2
rule[A,B,X](Child[A,X], Child[X,B]): GrandChild[A,B]

that is, for some A, B, X, if there exist Child[A,X], Child[X,B], then a GrandChild[A,B] can be derived.

we can now write a query. The query will compile if and only if the compiler is able to satisfy the contraints using the facts and the rules in the knowledge base.

1
2
3
4
5
6
7

> query[ GrandChild[John, Tom] ]
// (compiles; returns the fact instance)

> query[ GrandChild[John, Carl] ]
Compilation Failed

You should now be more convinced that there is a strong correspondence between the Logic Scala version and the Prolog version. Truth is, the only difference between Logic Scala and real-world Scala is that we renamed a few keywords. The code and machinery are actually the same!

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
trait John
trait Carl
trait Tom
trait Child[T,U]
implicit val john_carl = new Child[John,Carl]{}
implicit val carl_tom = new Child[Carl,Tom ]{}

trait GrandChild[T,U]
implicit def grandChild[X,Y,Z](
implicit
xy: Child[X,Y],
yz: Child[Y,Z]
) = new GrandChild[X,Z] {}


> implicitly[ GrandChild[John, Tom] ]
// (compiles; returns the fact instance)

> implicitly[ GrandChild[John, Carl] ]
Compilation Failed

You should now be convinced that there is a correspondence between the way logic programs are evaluated and the way implicits are resolved at compile-time. In fact,

  • implicit vals can be seen as facts

  • implicit defs can be seen as rules

  • type parameters in a def correspond to variables in a rule definition

  • implicit parameter lists are bodies of a rule

Please recall that in Prolog we could even write

1
2
?- grandchild(john, Y).

The query above is asking for evidence that there exist a granchild of John in the KB. The interpreter would return the Y binding for which the condition holds; i.e., Y=tom. Something similar can be achieved in Scala through existential types. The query above could be encoded as:

1
2
> implicitly [ GrandChild[John, Y forSome { type Y }] ]

this encodes exactly the same query: find evidence (an object instance) that there exist Y such that GrandChild[John, Y] is a valid statement (a concrete type).

This can be also more concisely expressed as:

1
2
> implicitly [ GrandChild[John, _] ]

There are caveats, though.

First of all, Scala will not return the actual binding for which the condition holds. In fact, this does not make sense, because, in general, when you write GrandChild[John, _] you are representing the type of something that is in a GrandChild relation with John. It may be an instance of GrandChild[John, Tom], but there might be others at some point in time. You do not want GrandChild[John, _] to always be an alias for GrandChild[John, Tom]

Second, the symmetric query GrandChild[_, Tom] will not work in this case. The deep reasons for which are a bit technical and would go beyond the scope of this essay. In short, Scala evaluates the implicit parameters in the def signature in the order they are defined; thus, when you write GrandChild[_, Tom] the system looks for an implicit instance of Child[_,Y], but this cannot be derived, because Y is still unknown at that time. In fact, if we define another implicit def where the implicit values are re-ordered:

1
2
3
4
5
6
7
8
9
10
11
12
13
implicit def grandChildReordered[X,Y,Z](
implicit
 yz: Child[Y,Z],
 xy: Child[X,Y]
) = new GrandChild[X,Z] {}


> implicitly[ GrandChild[_, Tom] ]
// compiles

> implicitly[ GrandChild[John, _] ]
// compiles

GrandChild[_,_] will only work if you explicitly “assert” it:

1
2
3
implicit val fallback: GrandChild[_,_] = new GrandChild[Nothing,Nothing]{}
> implicitly[ GrandChild[_, _] ]

We are now ready to understand how type parameter inference and implicit resolution work in applyProduct.

Understanding applyProduct: evidences and typeclasses

We already saw what the body of the applyProduct function did; we left for later the description of the implicit parameters. Because now we have found the parallel between Scala and Prolog, we can easily describe what is happening in the function:

1
2
3
4
5
6
def applyProduct[P <: Product, F, L <: HList, R](p: P)(f: F)
 (implicit
 gen: Generic.Aux[P, L],
 fp: FnToProduct.Aux[F, L => R]
 ) = f.toProduct(gen.to(p))

  • Let be P a Product, that is, a tuple or a case class

  • Let be F an unconstrained type parameter

  • Let be L an HList

  • Let be R an unconstrained type parameter

then,

  • For a given product of type P

  • For a given object of type F (the function)

we can invoke applyProduct if the following relations hold:

Generic.Aux[P, L]; this is the built-in “predicate” that Shapeless provides to encode the relation between a product type P and an HList L. It holds when it is possible to derive a Generic[P] instance that converts P into L FnToProduct.Aux[F, L => R]; is he built-in “predicate” that Shapeless provides to encode the relation that holds when F can be converted into a function from HList L to return type R; it holds when it is possible to derive an FnToProduct[F] instance called that converts F into L => R

Congratulations, you have just learned how to use typeclasses ! In fact, the values gen and fp are instances of the Generic[P] and FnToProduct[F] typeclasses.

As we saw in the previous section, the value of an implicit parameter can be seen as evidence that a predicate holds. In Scala, we implement predicates through types, and evidences are object instances. If these object instances provide methods, then they are called typeclasses.

For instance, we saw before that Generic[P] provides the methods to(P) and from(HList).

The Aux Pattern

We have left only one little detail. Why are the predicates called Generic.Aux and FnToProduct.Aux but the type classes are just Generic and FnToProduct ? Again, this is a bit technical.

Generic.Aux and FnToProduct.Aux are actually type aliases. You may know that scala includes a type alias feature. Type aliases are also called type members because, just like properties and methods, they occur as members of another type. For instance, the Generic[P] trait:

1
2
3
4
trait Generic[P] {
 type Out
}

The type alias Generic.Aux is defined in the companion object of Generic

1
2
3
4
5
object Generic {
 type Aux[P,L] = Generic[P] { type Out = L } 
 ...
}

and it reads as follows: Generic.Aux[P,L] is the type of Generic[P] when the type alias Generic.Out within it equals to L. Again, this is a predicate. When we write (implicit gen: Generic.Aux[P,L]) or equivalently (implicit gen: Generic[P] { type Out = L }); we want the compiler to prove that there exists (or that it can be derived) in current scope an evidence that Generic[P] { type Out = L }; in particular, in the case Generic.Aux[Person, L] we want proof that there exist some HList L such that Person can be converted into L.

But then, you may ask, why can we just write Generic[P,L], and be done with it ? Recall that in one of our first examples we wanted to convert a Person(String,Int) into the corresponding HList:

1
2
3
val gen = Generic[Person]
val hl = gen.to(p)

If L were part of the type signature of Generic then we would need to write:

1
2
val gen = Generic[Person, L]

but then L would be a fresh type parameter, which cannot occur in that position!

1
2
3
4
5
scala> val gen = Generic[Person, L]
<console>:19: error: not found: type L
 val gen = Generic[Person, L]
 ^

The type member trick is useful to “hide” the type parameter in those situations where it is not only unnecessary (the type L of the HList is univocally determined by Person) but also a problem, because we would now be required to explicitly give the HList type for all the Generic instances we wanted to create!

This is what we usually mean when we say that type members can be used to hold the result of a type-level computation; in this case type member Generic.Out contained the result of automaticaly deriving L from Person.

Visavis, Generic.Aux[P,L] is sometimes called a type extractor because it extracts the type member value Generic.Out by raising it into the signature of type Generic.Aux.

Bonus: applyProduct encoding in Prolog

We will show how the applyProduct works by encoding it in Prolog.

The type-level constraints of our running example can be re-implemented in Prolog. A tuple can be defined just like in Scala:

1
2
(arg_type1, arg_type2, ..., arg_typeN)

Since we know that in Scala we can always transform a FunctionN into a Function1 of a TupleN for all N>1, we can represent a function type as:

1
2
fn((arg_type1, arg_type2, ..., arg_typeN), ret_type)

and, when there is only one arg, simply:

1
2
fn(arg_type1, ret_type)

for instance, the type signature for function def stringify(s: Any): String could be represented by fn(any, string)

Now, let us try and translate the constraints in the signature of applyProduct into Prolog predicates.

1
2
3
4
5
6
def applyProduct[P <: Product, F, L <: HList, R](p: P)(f: F)
 (implicit
 gen: Generic.Aux[P, L],
 fp: FnToProduct.Aux[F, L => R]
 ) = ...

In the previous section we saw that type parameters in a def can be seen as the variables in a Prolog rule, while the rule body can be expressed using type parameters.

We may define a can_apply_product(P,F,L,R) predicate as follows:

P must be a Product. So, let us define the product predicate which holds when P is a Product. For instance, in the case of a pair: product(X):- X = (,).

Or, in short: product((,)). product((,,_)). … % for >3 tuples

F is unconstrained (in the sense that there is no unary predicate on F)

L must be an HList, which in Prolog we can represent as a list. We can define the predicate hlist, which holds when L is a list. hlist(X):- X = [|]. % [Head | Tail] product([|]). % in short.

R is unconstrained (in the sense that there is no unary predicate on F)

Generic.Aux[P,L] puts in relation product P with HList L. In Prolog, this would a predicate. Let us see it for the case of a pair P: generic(P, L) :- product(P), (X, Y) = P, % deconstruct P into X and Y (the correct term is «unify») hlist(L), [X, Y] = L. % deconstruct L into X and Y

But, really, in Prolog we can drop the product and hlist constraints, because they are really implied by the syntax (X,Y) and [X,Y], respectively. So this can be further shortened:

FnToProduct.Aux[F, L => R] puts in relation the type parameter F with the function type L => R. In Prolog, this could be described as the predicate fn_to_product(F, fn(L, R)):

in fact, we decided to represent any T => R function type using fn(T,R). The fn_to_product predicate shall hold when

  1. F is a function

  2. the tuple of the types of the arguments of F can be converted into a (h)list

This can be written as follows:

1
2
3
4
fn_to_product(F, fn(L, R)) :-
 fn(Args,R) = F,
 generic(Args, L). 

or, in short:

1
2
3
fn_to_product(fn(Args,R), L, R) :-
 generic(Args, L). 

The can_apply_product predicate, which encodes all of the type constraints in the applyProduct Scala function, puts in relation:

  • the product P

  • the function F

  • the list L of the arg types of F

  • the return type R of F

Let’s see Prolog and Scala side-by-side: |can_apply_product(P, F, L, R) :- generic(P,L),fn_to_product(F, fn(L, R)).|def applyProductP <: Product, F, L <: HList, R(f: F) (implicit gen: Generic.Aux[P, L], fp: FnToProduct.Aux[F, L => R])|

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

def applyProduct[P <: Product, F, L <: HList, R](p: P)(f: F)
 (implicit
 gen: Generic.Aux[P, L],
 fp: FnToProduct.Aux[F, L => R])



So, let’s see it as a whole:
generic((X,Y), [X,Y]).

fn_to_product(fn(Args,R), L, R) :-
 generic(Args, L).

can_apply_product(P, F, L, R) :-
 generic(P,L),
fn_to_product(F, fn(L, R)).

Conclusions and References

This long blog post has only scratched the surface of Shapeless. If this has whetted your appetite, then I suggest you to delve further into Shapeless’s inner guts. I have compiled for you a short list of resources that you may want to have a look at.

Longer Essays

Slides and Courses

Misc. Resources