samskivert: Ramblings on programming language design

13 June 2018

Mario asked me an interesting question the other day. As is often my tendency, I gave him a multi-thousand word answer. Then he foolishly asked more interesting questions, so I hit him with additional walls of text. Be careful what you ask for kids. In the end, we figured we'd share with the class because perhaps some useful explanations happened along the way.

[Note: because I'm lazy I did not edit out all the profanity. Please excuse two guys talking to one another like sailors. When forced to deal with the failings of overly complex programming languages and tools on a daily basis, one has a tendency to cope by swearing a lot.]

Mario

Date: Tue, 12 Jun 2018 22:33:38 +0200
From: Mario Zechner
Subject: Re: Un-Reason-able
To: Michael Bayne

...blah blah...

Speaking of which. Assume I built a tiny little typed scripting language with generics. Would you consider Kotlin's flavor of variance a good compromise compared to other langs?

https://kotlinlang.org/docs/reference/generics.html

MDB

Yes, definitely. You want declaration site variance, rather than use site variance. And you definitely want what they call generic functions (which the programming language people call f-bounded polymorphism).

Kotlin's solution to variance refinement is reasonable, though it doesn't solve all the variance problems. It allows you to do what they show in their example (write a method that reads from one array and writes to another), but it doesn't solve the "classic" variance problem, which is Collection.contains.

Say you have a Collection trait (interface) which is immutable and thus covariant in its type parameter. That allows you to say that a collection of Bs is a subtype of a collection of As as long as B is a subtype of A.

In Scala (where + is out and - is in), you would naturally want a contains method for your Collection:

trait Coll[+A] {
  def contains (elem :A) :Boolean
}

but you can't have it:

<console>:12: error: covariant type A occurs in contravariant position in type A of value elem
         def contains (elem :A) :Boolean
                       ^

because covariant interfaces "produce" things but the contains method is "consuming" an A not producing it. If instead of contains we were writing add, you can see how this could lead to breakage:

trait Coll[+A] {
  def add (elem :A) :Unit
  // implementation omitted, but assume it puts elem into some internal array
}

// a Dog is an Animal, so this is OK
val thing :Animal = new Dog()
// A is covariant so Coll[Cat] is a Coll[Animal], this is OK
val things :Coll[Animal] = new Coll[Cat]()
// bad! this adds a dog to cats
things.add(thing)

So in general it is illegal to use a covariant type parameter in "contravariant position". However, we know that in this case we're not doing anything naughty with elem (like adding it to an array internally) so we want to somehow express that what we're doing is OK. Before we look at how Scala solves it, let's see how Kotlin solves it:

abstract operator fun contains(element: @UnsafeVariance E): Boolean (source)

Ha. So we can see that Kotlin has admitted defeat in this particular case. Which is not to say that their solution to variance is bad, just that it has limits. Maybe those limits are fine, this doesn't really come up that often. Even in Scala where they have a "better" solution, they still have an @uncheckedVariance annotation which I have had to use on a number of occasions, so I don't know that there is a "perfect" solution.

Back to our example, let's try it in Kotlin:

interface Coll<out A> {
  fun contains (elem :A) :Boolean
}

gives us:

Test.kt:2:23: error: type parameter A is declared as 'out' but occurs in 'in' position in type A
  fun contains (elem :A) :Boolean
                      ^

Same as Scala.

The "right" thing to do is to use an existential type, which neither Java nor Kotlin "fully" supports.

Java supports existential "bounds" which is part of its use site variance solution:

Coll<? super A>

which means "a collection of something, and I don't know quite what, but either an A or some supertype of A". This is basically what we want for contains, in that we want to accept our A or any supertype of A (because if we're treated as a collection of our supertype then someone could pass a supertype element into contains). And we still want to be prevented from using what we're passed as if it were an A (like by sticking it into an internal array of As), which would be true because a ? super A is not an A so we couldn't treat it as one.

But we can't write that type down. You can't say (we're in Java here):

boolean contains (? super A elem);

If if you were clever, you might try to use f-bounds to solve it (which swaps the existential for a universal, more on that later):

boolean contains<B super A> (B elem);

but Java doesn't allow super type constraints in f-bound methods, only subtype constraints. Thwarted again!

Kotlin uses what it calls "type projections" in the same way. At a use site, you can say Coll<in A> to indicate that you have a collection of A or some supertype of A, and then you won't be able to call any methods on the collection that return an A.

The way they describe this working is that you "project out" only the subset of Coll<in A>'s API that is compatible with the supplied variance annotation (in this case contravariance). So in this case, any methods that return an A would be omitted (that's covariant position), whereas methods that take an A are OK (that's contravariant position).

But another way of thinking about this is that A becomes some existential type. So you can pass values in for that existential type, you have a "nameable" type that meets the criteria of the existential, but you can't get values out that have the existential type because you can't write down (name) the type of a variable to hold them. Even if you let type inference propagate the unnamable type to a variable, you wouldn't be able to do much with it because it's this weird "A or any possible supertype of A" type and about the only thing you can do with that is call Object methods on it since that's the only thing you can prove that it is.

Anyway, we can't write down a simple existential type in Kotlin either (and it would be weird to write down existentials using this variance projection syntax anyway). This doesn't work:

fun contains (elem :in A) :Boolean

Like Java, Kotlin only supports subtype constraints on f-bounded methods. They don't even use a syntax that can be "inverted". You just write foo<A : B>(...) which means A is a subtype of B. I guess you could try to be sneaky and write:

interface Coll<out A> {
  fun <A : B> contains (elem :B) :Boolean
}

But kotlinc tells you to fuck right off because the thing after the : is supposed to be a reference to a known type, not the place where you declare a new type:

Test.kt:2:12: error: unresolved reference: B
  fun <A : B> contains (elem :B) :Boolean
           ^

Now we can go back to how Scala solves this, which I've hinted at above with my attempts to use f-bounds to solve the problem in Kotlin and Java. In Scala, the method is actually defined like so:

trait Coll[+A] {
  def contains[B >: A] (elem :B) :Boolean
}

Which means "Hey, we take any type B as long as it's a supertype of A" (which includes A itself). This using a universal instead of an existential, but we could also use an existential:

def contains (elem :_ >: A) :Boolean

where _ :> A means "some type that is a supertype of A", just like Java's ? super A. But you can't use that existential syntax outside of square brackets (yay parsers!), so we have to use the more verbose existential syntax:

def contains (elem :B forSome { B >: A }) :Boolean

which the parser is happier about.

As you can see, existentials and universals are duals of one another in a deep mathy way. This is neat but not usually super relevant when designing programming languages. However, in the case of how we solve variance problems, this turns out to be an important distinction. Hence my rambling on for this whole long ass email. :)

With an existential type, you're saying "hey, there's a type that fits the bill here, and I'm going to declare it, but I can never talk about it ever again". So every time you write down an existential variable, it's a new one. If you say:

def foo (as :Coll[_ >: A], bs :Coll[_ >: A])

then the things in as could be totally different than the things in bs and you cannot interchange them. Indeed you could pass two different collections of supertypes of A to the two different arguments.

But if you write:

def foo[B >: A] (as :Coll[B], bs :Coll[B])

then as and bs must have the same type of thing in them. You've named this hypothetical type and now you can talk about it.

This has an important influence on how you design APIs. For example, in our hypothetical read-only collection class, we want to be able to append another collection to this one and return a new collection that is the concatenation of both:

trait Coll[+A] {
  def append (as :Coll[A]) :Coll[A]
}

Well, we can't write that because A is used in a contravariant position, so like before we have to use existentials or universals. If we use existentials (which were arguably the 'right' solution to the contains problem):

def append (as :Coll[_ >: A]) :Coll[A]

then we're still in trouble because the things inside as are not As they're some existential type (which might be a supertype of A). So we can't combine them with our As and claim to be returning a new collection of As. That's a lie.

But if we use universals:

def append[B >: A] (as :Coll[B]) :Coll[B]

we accomplish a few things. First, we declare a type B that we know A is a subtype of, so our elements are also Bs. Second, we specifically accept a collection of Bs, so we know we can safely put them together with our elements (which are Bs too). Finally, we also tell the caller what they're getting back: a collection of Bs, not a collection of As.

So if you do:

trait Animal
trait Cat extends Animal
trait Dog extends Animal

val cats :Coll[Cat] = ...
val dogs :Coll[Dog] = ...
val animals = cats.append(dogs)

then the inferred type for animals will be a collection of the most specific supertype of Cat and Dog, in this case Animal. Showing it all in action in the Scala repl:

scala> :paste
// Entering paste mode (ctrl-D to finish)

class Coll[+A] {
  def append[B >: A] (as :Coll[B]) :Coll[B] = null
}
trait Animal
trait Cat extends Animal
trait Dog extends Animal
val cats :Coll[Cat] = new Coll[Cat]()
val dogs :Coll[Dog] = new Coll[Dog]()
val animals = cats.append(dogs)

// Exiting paste mode, now interpreting.

defined class Coll
defined trait Animal
defined trait Cat
defined trait Dog
cats: Coll[Cat] = Coll@1e1232cf
dogs: Coll[Dog] = Coll@6f6efa4f
animals: Coll[Animal] = null
// see right ^^^^^^^^ here, it inferred Animal for B

This allows you to concatenate two collections of different types as long as they have a common supertype, and everything just works. The existential solution is a non-starter, and the "make everything invariant" solution (i.e. Java) means you end up wanting to add cats and dogs but have no way to do it that doesn't involve casts and bullshit.

This may or may not be the reason that Kotlin basically doesn't have an append() method on their read-only collections. They just decided "oops, can't solve this problem". Even if they tried to fake it with the @UnsafeVariance annotation, there's no way in Kotlin's type system to name this inferred supertype so that we can say what kind of collection you get back. You don't get back cats when you add dogs to cats, and you don't get back (the existential) dogs. You get back their LUB (least upper bounds).

So the moral of this story is: declaration site variance is good, f-bounded polymorphism is good, but you should also consider supporting supertype f-bound constraints in addition to subtype f-bound constraints. They allow you to solve a bunch of the problems that existentials are used to solve in other type systems, along with even more problems that other type systems can't solve.

You can in theory get away completely without existentials. The new version of Scala (called Dotty) has no existentials. They do have some other crazy shit though, so it's possible that some of that crazy shit is needed to fully avoid existentials. But I don't think so, because existentials almost never appear even in current Scala APIs.

That concludes your type theory lesson for the day. :)

I can't write all this without also mentioning that this is one of the reasons that OO needs to die in a fire. Almost all of this variance stuff just goes away when you don't have subtyping. Poof! So much simpler.

Mario

That all made a lot of sense actually. Thanks for taking the time to write it all up.

I too have a feeling that OO needs to die. I need to read up on how non-OO languages do polymorphism without subtyping (in the OO sense).

MDB

Type classes! :)

[We had been talking about type classes earlier in the conversation.]

Mario

I think between subtyping, adts and type classes, type classes seem to be a good compromise. maybe i should start with that.

[Then shortly after because he hadn't yet seen my "Type classes!" email.]

Haha, great minds...

MDB

ADTs are also needed in combination with type classes. Well, not strictly. You could just have named records. But then you'd have to use type classes for everything which would be annoying.

There are two things people are usually doing when they say they want subtyping: unions (solved with ADTs) and ad-hoc/subtype polymorphism (solved with type classes).

For example, if I'm reading in JSON in an OO language, I'm gonna have some sort of type-lattice:

abstract class JsonValue
class JsonNumber extends JsonValue
class JsonList extends JsonValue
class JsonObject extends JsonValue
// etc.

You could have methods on those classes, but you don't have to. In a language where you don't have subclasses, that's not how you would solve it. You use ADTs (warning pseudo-Haskell syntax here):

data JsonValue = JsonNumber Int
               | JsonList (List JsonValue)
               | JsonObject (Map String JsonValue)
               -- etc.

Then when you want to do something to a JsonValue, you pattern match on it:

case value of
  JsonNumber num => ...
  JsonList list => ...
  JsonObject map => ...
  // etc.

So what would be in methods on subclasses is instead in functions that pattern match on the ADT type.

All peachy keen and no subtypes anywhere. Technically the "cases" of an ADT are kind of like subtypes of the ADT type, but many languages don't give them their own type at all, there's only the ADT type (e.g. JsonValue). You can never say "Hey, I have a variable of type JsonList".

I think it's useful to have types for the cases, but this is still a very limited form of subtyping and doesn't cause problems. Indeed you actually don't have subtypes in the type checker, you just have a rule that says "If I ever have a case type, and I want the ADT type, I'm good." They call it a "coercion". That's it. If you do have types for the cases, you never want to infer them, always infer the ADT type. This is a problem that Scala has because it uses classes to model ADTs and type inference ended up doing the undesired thing.

OK, so that's one use case for subtyping. The other is ad-hoc/subtype polymorphism. For that you use type classes. A simple example is "How do I do toString?" (which in Java etc. is a method on Object that you override). In Haskell and type classy languages, they call this typeclass Show:

class Show a where
  show :: a -> String

Then I can implement that for some type by saying:

instance Show JsonValue
  show :: JsonValue -> String -- optional type signature just to show what's going on
  show value = case value of
    JsonNumber num -> show num
    JsonList list -> "[" ++ show list ++ "]"
    JsonObject map -> -- i'll skip this one because it's complicated
    -- etc.

So now my JsonValue ADT "implements" the Show type class and can be turned into a string. I might have a log function that wants to toString something, it could look like this:

logShowable :: Show a => a -> IO Unit
logShowable thing = log (show thing)

This assumes I have some log :: String -> IO Unit function that logs something that's already a string. And ignore the IO stuff, that's just Haskell's way of saying "the thing you're doing has side effects".

The key here is the Show a => type class constraint. That says, this function works for anything as long as it has an implementation of Show.

This is, in certain ways, "better" than OO polymorphism because I can declare an instance of Show for any type I like, whether I wrote the type or not (modulo caveats!). It's "open extension" rather than "closed extension" (you can't make some class implement an interface after the fact, you would have to modify the class). But it also doesn't have the "magic" of the this pointer (and implementation inheritance). But I think the magic of the this pointer is bad magic, so good riddance.

Anyway, that's a quick overview of how you can solve the same problems people usually do with subtyping, using ADTs and type classes.

MDB

One more thing and then I'll shut up¹. This is a super good paper that brought a whole lot of haphazard ideas about OO that were rolling around in my head into clarity:

On Understanding Data Abstraction, Revisited

It also talks about how module systems play into all of this, which is also important for an industrial strength language.

Highly recommended.

[¹ Editor's note: MDB does not in fact shut up.]

MDB

One important note about that paper: don't confuse "abstract data type" with "algebraic data type" (both tragically called ADTs). What I call unions up above are "algebraic data types", not abstract. Abstract data types are about hiding the implementation, like in ML modules, and also possible in Haskell though slightly less fancy than with ML. In the paper he unfortunately conflates these two things because most of the languages of the day also conflated them.

An abstract type is just one that you can't look inside of. It might just be an int, or it could be an array of ints, or some crazy algebraic data type. You don't know and it doesn't matter. You can only pass values of the type around and pass them to functions which can "see behind the curtain" (because they're defined in the same module as the abstract data type).

Mario

I love these "Michael shows Mario the real programming world" email threads. I was aware of ADTs and type classes (TS has a primitive (?) form of ADTs that I actually used in anger. I'm aware that it's not real ADTs, but simply a more civilized way to handle function parameters that could be one of many types, because JS is funny that way). But your explanation made a few more things click.

One thing that mildly irks me about ADTs is that they are "closed" (if I understand correctly).

Now, in practice that's probably a good thing in many respects. In OO land you end up with something like this all the time

AstInterpreter.java#L762

(ignore that these should be instanceofs, I was JMHing a bit to hard...). There's no language construct that ensures I match all subtypes of Node, so you end up with a shitty ass big fat if/else statement. ADTs would be great for this, and I can imagine one could even wrangle out a bit of performance but essentially doing a computed goto style match based on the adt subtypes.

However, I think that breaks down as soon as I let that code out into the wild, to be used and extended by others. For a user to extend an ADT, they'd have to modify my source (unless I'm missing something). That sucks deeply. Extendable types are a must in many APIs.

OTOH, not having implementation inheritance isn't such a big deal. Most of that can be factored out into functions anyways that can be reused when pattern matching to implement a "method" for an ADT. The same is true for deep type hierarchies. Fuck em, I can easily live without them.

However, my OO brain is somewhat opposed to how the proximity of types and type methods (in OO speak) is lost with ADTs. It is nice having the type lattice in one place. But it's less nice to stuff the implementations of a function for all adt "subtypes" into a single big match statement. When implementing a type and functions operating on that type, all that (temporary) noise is distracting. Ideally, a sufficiently smart IDE would let me switch into OO mode and only display what I need when implementing a specific ADT "subtype". Or I can tell my OO brain to just rotate itself by 90°.

Now type classes I really like! Mapped to my OO brain, they are a kind of mix between Interfaces and extension methods (and sort of similar to Go's structural typing, though I think structural typing is a bad idea). Which is great in terms of being able to extend an API.

Maybe I should try my hands on writing a little language with records, ADTs and type classes, but using familiar C style syntax. I'm not quite sure how parameterized types fit into this yet, but I'll find out!

Also, thanks for the recommended reading. I didn't wanna sleep tonight anyways :)

MDB

One thing that mildly irks me about ADTs is that they are "closed" (if I understand correctly).

You get right to all the good problems! :)

That tension: between a closed data type which is easy to add functions to (ADTs), versus a closed set of functions which is easy to add new "types" to (OO interfaces/classes) is known as the "expression problem". And enough ink has been spilled describing solutions to it to fill all the swimming pools in Southern California.

The first thing to acknowledge is that there is in fact a tradeoff. Being used to OO style, you might take it for granted that someone can't come along and add new methods to a class hierarchy, but that's a thing someone might want to do (there are so many methods I would love to add to Java's standard libraries...). So in OO style, it's easy to make a new type that implements a fixed set of methods, but a cumbersome totally external process (more on that later) to add a new method that's implemented by all the types. Whereas in functional/ADT style, it's super easy to add a new method that is implemented by all the types: just write a function that pattern matches on the ADT. Done! But if you want to add a new type case, you're fucked.

You're probably familiar with some of the ways the OO people try to have their cake and eat it too: extension methods, the visitor pattern, other crazier shit like partial classes and the mother of all OO battleships: a research language called Familia (https://dl.acm.org/citation.cfm?doid=3152284.3133894) which tries to solve all the things and ends up being so complex that I doubt even the authors could use it sensibly.

As it turns out, there are a bunch of ways that functional/ADT people try to have their cake and eat it too. To name a few: GADTs (generalized algebraic data types, which are actually useful for a bunch of different things, but solving the expression problem is one of them), multimethods, type classes (via a variety of creative approaches).

There are even problems you don't realize you have (haha) which have vast numbers of papers exploring not very compelling solutions to not very pressing problems. Family polymorphism is a popular one (which is also tackled by Familia since it tries to do all the things). This paper provides an interesting summary of a bunch of "extensibility problems" with the particular slant of how they can be solved with type classes: https://dl.acm.org/citation.cfm?id=1173732 Whether or not the type classes are the best solution, the overview of the problems is nice.

However, most of these "solutions" make the assumption that these sorts of problems are so common that we absolutely must make our languages bend over backwards to support some concise, efficient encoding of a solution. But I'm not so sure. I think that making it too easy to extend things (and accidentally allowing things to be extended that were never anticipated) is actually a bad idea. I want to tackle the problem from a radically different angle. In cases where the full matrix of extensions really are needed, I just want to make it super easy for you to modify the library you're depending on.

Effectively, no matter what the approach: you are modifying the library by extending it, whether the library authors planned for it or not. And usually they didn't. And even if they did, they probably did a shitty job of it. So the best solution IMO is for you to acknowledge that you are getting deep into bed with this library, and to just see all of its dirty laundry. Go in there and add the code you need. Change the behavior that you need. And read the damned code along the way so that you know what you're doing. Don't just override a method and sneakily change shit and hope for the best.

I want to provide good tool support for maintaining patches against your dependencies. For a start, this is often the first thing you want when you're trying to use a library and you just want to put a goddamned print statement in some library code, or make some other exploratory change to try to better understand how this library code works. Yes if you really just want to inspect a value, you can fire up a debugger, and I want to make improvements to debugging too so you can easily trace values without making code edits, or graph values over time and other useful inspections, but sometimes you need to change the code.

Sometimes you find bugs, now you have to fix those bugs. Oh god, I have to fork this library, change my build system, deploy patched versions of artifacts to company-wide Maven repositories, deal with dependency conflicts elsewhere, etc. etc. Someone please put a bullet in my head.

Instead, I want my build system to express both a dependency and any patches that I want applied to that dependency. Now I can easily fix bugs or make little tweaks. Well hell, lets make it easy to version control those patches along with my project. And then when I want to upgrade my dependency, I see exactly where my patches need to change based on upstream changes. I don't discover later, oops, they rerouted this method I was overriding, or changed this internal behavior I was subtly depending on. I see it right there when I update the dependency. And since Compose is all about making it easier to cope with change, all of the things that make it easier to update code that depends on a library API, also make it easier to update code that patches the library.

Anyway, it's a radical idea and I have no idea if it will pan out, but I think it has a lot of potential and also happens to be a great solution to the expression problem because you can go in and add things exactly where you want them. Then everything is just as easy to use as if the library anticipated your needs from the beginning instead of having some bolted on extra crap that's kind of annoying to use, versus the built in stuff.

All that said, type classes are like the OO approach, so sometimes you want to use type classes plus ADTs when you anticipate that people will be extending your code both with new types and with new functions. As an example of a solution to your interpretNode problem, you can use an approach called "object algebras" (when applied to OO languages) but which applies just as well for type classes (where they call it 'tagless final interpreters').

The idea is that you create a thing called an "algebra": a set of operations on abstract things, which captures the structure of the computations you're going to be performing, without pinning you down to particular data structures. Best explained with an example.

First we define an ADT to represent the "output" of the interpreter, I'm going to make this super simple and say we can only have Int, but you can support multiple kinds of values, you just have to then deal with the fact that when you're interpreting you might get a runtime type error. I don't want to deal with that here.

data Value = IntVal Int

In real Haskell you'd probably use: newtype Value = Value { value :: Int }, but using an ADT is what you'd do if you had more than one type of value, so I'll use this degenerate ADT instead. None of this is particularly relevant to the technique I'm describing. Because we're using type classes to drive teh boat, we do need some sort of type to wrap Int so that we can say "when you ask for this type, use this algebra" because all of this ends up being driven by the type you ask for in the end. You could technically make Int mean "run the interpreter algebra" but that's a bad idea.

Now we define our abstract "node algebra":

class NodeAlg a where
  lit :: Int -> a
  add :: a -> a -> a
  mult :: a -> a -> a

This enumerates all the kinds of AST nodes we can create and in the mathematical sense "is" the AST. It is akin to the "intensional vs. extensional" distinction in set theory. You can define a set "extensionally" as the list of things in the set, or you can define a set "intensionally" as a function that says yes or no when you want to know whether something is in the set. Programmers are used to defining things "extensionally": enumerate all the cases, or enumerate all the values, but this algebra defines things "intensionally" by saying "this is the structure of the computations you can perform over this tree". It requires turning your brain inside out a bit, but when you squint, you see that all the exact same "things" are being said, just in a different way.

Now we can write an interpreter as one kind of concrete algebra:

instance NodeAlg Value where
  lit value = IntValue value
  add (IntValue a) (IntValue b) = IntValue (a + b)
  mult (IntValue a) (IntValue b) = IntValue (a * b)

Now if I want to create an expression in the "AST", I can write something like:

myexp = mult (add (lit 5) (lit 3)) (lit 3)

That expression is "unevaluated", it's basically a function that takes an algebra and "runs" the algebra through the computation. If you asked Haskell, it would have the type:

myexp :: NodeAlg a => a

which means "give me a NodeAlg for any type a and I'll give you back an a".

The way type classes work, the algebra that's going to get passed in is implicitly defined by the type you ask for, so if I write:

let IntValue value = mult (add (lit 5) (lit 3)) (lit 3)

then it's going to use the "interpreter" algebra that I defined above (NodeAlg Value). That will interpret the AST and give me back 24. IntValue value pattern matches on the single case of the Value ADT and extracts the Int inside so value will be bound to 24.

But I can also define a "pretty printer" algebra (which is one way of extending this code, this pretty printer algebra can be written by anyone, not just the original library author):

data Printed = Printed String

instance NodeAlg Printed where
  lit value = Printed (show value)
  add (Printed a) (Printed b) = "(" ++ a ++ "+" ++ b ++ ")"
  mult (Printed a) (Printed b) = a ++ "*" ++ b

And I can run this in the exact same way:

let Printed text = mult (add (lit 5) (lit 3)) (lit 3)

And the typeclass machinery will say, oh send in the Printed node algebra, and text will get bound to the string "(5 + 3) * 3".

So then some guy can come along and say, dude, what kind of barbarian created a language that doesn't have subtraction!? Christ, let's fix this shit up.

class NodeAlg a => MyAwesomeNodeAlg a where
  sub :: a -> a -> a

which "extends" the NodeAlg type class with an additional method "sub" which represents subtraction. In the type class world, extension is really more like logical implication. It says "if you have a MyAwesomeNodeAlg" for some type, you gotta have a NodeAlg for that type too. So you can "imply" multiple parents with no problem.

Under the hood, parents and children still often get consolidated to avoid having to pass around huge numbers of typeclass dictionaries (which are roughly like vtables), but that's an implementation detail rather than a fundamental property of the language semantics (compared to class inheritance, say, where a single vtable is part of the semantics and is necessary to make dynamic dispatch via this work).

The lack of the diamond inheritance problem here is partly due to a property called "coherence" which means that there can only ever be one implementation of a typeclass for a particular type. This is a deep topic and there are people who are pro-coherence and people who are against coherence (or rather want to sometimes break it). A good design probably allows a little of both, but I haven't figured out the exact right thing on that front yet. TBD.

Anyway, if we try to just use our awesome node algebra right away, like so:

let IntValue value = mult (sub (lit 5) (lit 3)) (lit 3)

Haskell will say, "can't find an instance of MyAwesomeNodeAlg for IntValue" because we haven't provided one yet. So we whip that up:

instance MyAwesomeNodeAlg Value where
  sub (IntValue a) (IntValue b) = IntValue (a - b)

and now it just works. Because there's already an instance of NodeAlg for Value, the MyAwesomeNodeAlg automatically implies that one for all the methods defined by NodeAlg and we only have to define the new ones. If, in addition to adding sub, we were the ones to add the pretty printer concrete algebra, we could define everything at once:

data Printed = Printed(text :String)

instance NodeAlg Printed where
  lit value = Printed (show value)
  add (Printed a) (Printed b) = "(" ++ a ++ "+" ++ b ++ ")"
  mult (Printed a) (Printed b) = a ++ "*" ++ b

instance MyAwesomeNodeAlg Printed where
  sub (Printed a) (Printed b) = "(" ++ a ++ "-" ++ b ++ ")"

So we can extend on both axes as needed and without too much ceremony in the extended definitions or in the uses of those extensions.

These algebras are a little bit mind bending because we never create an AST data structure. So it's hard to understand how it even works (it's functions all the way down!). Again its the intensional extensional thing. A list is a data structure and it is also a fold function. They are mathematical duals. We've just written a set of "fold functions" over our tree.

To be fair, there are situations where using an algebra directly is not ideal. Sometimes you really want the data structure. But it's trivially easy to "reify" the fold by making an algebra that creates an ADT when you need it:

data Node = Lit Int | Add Node Node | Mult Node Node

instance NodeAlg Node where
  lit value = Lit value
  add a b = Add a b
  mult a b = Mult a b

So now you can say:

node :: Node
node = mult (sub (lit 5) (lit 3)) (lit 3)

and you get back an AST.

Also, in reality you're not going to write these algebra functions directly, you'll have a parser that takes a NodeAlg constraint and then calls that whenever it parses the appropriate AST node. In very simplified form:

parse :: NodeAlg a => String -> a
parse source = ...

The crazy thing is that if you pass the evaluation algebra to parse, the parser is literally interpreting the code as it parses. But you can pass the Node algebra and have the parser construct an AST as it parses, which is less crazy.

This folding stuff gets into a really interesting topic called "generalized recursion schemes" (also called f-algebras, functional programmers love to f things :). This is a rabbit hole I've gotten sucked into on multiple occasions. If you want to expand your brain, take this red pill: http://blog.sumtypeofway.com/an-introduction-to-recursion-schemes/

Of course, once you've reified things into an AST data structure, then you're back to the expression problem because the data structure can't be easily extended, but there are other techniques for dealing with that situation (with various tradeoffs), but at least you've got everything structured in an extensible way so then you can make smart decisions about which technique to use (or potentially allow the library user to make those decisions so they can use whatever works best for their use case).

But I still hold out hope for the "make it easy to hack the upstream library" approach! It's so crazy it just might work. :)

OTOH, not having implementation inheritance isn't such a big deal. Most of that can be factored out into functions anyways that can be reused when pattern matching to implement a "method" for an ADT.

Yeah for sure, and reusing straight functions can be made easier too. There are lots of ways to avoid busywork once you have a nice conceptual framework where the programmer understands what busywork is being automated for them. The danger I think is when there's too much magic and the programmers don't really understand what the wiring is under the hood.

However, my OO brain is somewhat opposed to how the proximity of types and type methods (in OO speak) is lost with ADTs. It is nice having the type lattice in one place. But it's less nice to stuff the implementations of a function for all adt "subtypes" into a single big match statement. When implementing a type and functions operating on that type, all that (temporary) noise is distracting. Ideally, a sufficiently smart IDE would let me switch into OO mode and only display what I need when implementing a specific ADT "subtype". Or I can tell my OO brain to just rotate itself by 90°.

As you can see above, when you use type classes, you do end up aggregating the methods rather than the data type variants, so that's an OO style. When I first said that ADTs weren't strictly necessary but that if you didn't have them, it would mean that you'd have to do everything with type classes, I was alluding to the fact that often you can choose whether you want to structure your operations as pattern matching on ADTs or as type classes implemented for a collection of otherwise unrelated types. So you have some flexibility there.

I agree that it would be interesting to have support in the IDE for "focusing" on a particular ADT case. Haskell already supports a style of writing functions that would work very nicely with such a "focusing" mechanism. You can write:

data Foo = Bar Int | Baz String | Bing Boolean

foozle :: Foo -> Int
foozle Bar value = value
foozle Baz text = length text
foozle Bing yes = if yes then 1 else 0

The pattern match is implicit in the definition of the function. It would be easy to "hide" all the cases that were not the one you were interested in.

[Editor's note: after this, MDB finally shuts up.]

©1999–2018 Michael Bayne