This post is a continuation of my posts discussing the topic of associated type constructors (ATC) and higher-kinded types (HKT):

  1. The first post focused on introducing the basic idea of ATC, as well as introducing some background material.
  2. The second post showed how we can use ATC to model HKT, via the “family” pattern.
  3. This post dives into what it would mean to support HKT directly in the language, instead of modeling them via the family pattern.

The story thus far (a quick recap)

In the previous posts, we had introduced a basic Collection trait that used ATC to support an iterate() method:

trait Collection<Item> {
    fn empty() -> Self;
    fn add(&mut self, value: Item);
    fn iterate<'iter>(&'iter self) -> Self::Iter<'iter>;
    type Iter<'iter>: Iterable<Item=&'iter Item>;
}

And then we were discussing this function floatify, which converts a collection of integers to a collection of floats. We started with a basic version using ATC:

fn floatify<I, F>(ints: &I) -> F
    where I: Collection<i32>, F: Collection<f32>

However, this version does not constrain the inputs and outputs to be the same “sort” of collection. For example, it can be used to convert a Vec<i32> to a List<f32>. Sometimes that is desirable, but maybe not. To compensate, we augmented Collection with an associated “family” trait, so that if we have (say) a Foo<i32>, we can convert to a Foo<f32>:

trait Collection<Item> {
    ... // as before
    type Family: CollectionFamily;
}

trait CollectionFamily {
    type Coll<Item>: Collection<Item>;
}

This let us write a floatify_family like so, which does enforce that the input and output collections belong to the same “family”:

fn floatify_family<C>(ints: &C) -> C::Family::Member<f32>
    where C: Collection<i32> //    ^^^^^^^^^^^^^^^^^ another collection,
{                            //                      in same family
    ...
}    

A common question in response to the previous post was whether the CollectionFamily trait was actually necessary. The answer is that it is not, one could also have augmented the Collection trait to just have a Sibling member:

trait Collection<Item> {
    ...
    type Sibling<AnotherItem>: Collection<AnotherItem>;
}

And then we could write floatify_sibling as follows:

fn floatify_sibling<C>(ints: &C) -> C::Sibling<f32>
    where C: Collection<i32> //     ^^^^^^^^^^^^^^^ another collection,
{                            //                     in same family
    ...
}    

For some more thoughts on that, see my comment on internals.

In any case, where I want to go today is to start exploring what it might mean to encode this family pattern directly into the language itself. This is what people typically mean when they talk about higher-kinded types.

Supporting families directly in the language via HKT

The family trait idea is very powerful, but in a way it’s a bit indirect. Now for each collection type (e.g., List<T>), we wind up adding another “family type” (ListFamily) that effectively corresponds to the List part without the <T>:

struct List<T> { ... }
impl Collection for List<T> { type Family = ListFamily; ... }
struct ListFamily;
impl CollectionFamily for ListFamily { ... }

The idea of HKT is that can make it possible to just refer to List (without proving a <T>), instead of introducing a “family type”. So for example we might write floatify_hkt() like so:

fn floatify_hkt<I<_>>(ints: &I<i32>) -> I<f32>
//              ^^^^ the notation `I<_>` signals that `I` is
//                   not a complete type

Here you see that we declared a different kind of parameter I – normally I would represent a complete type, like List<i32>. But because we wrote I<_> (I’m pilfering a bit from Scala’s syntax here), we have declared that I represents a type constructor, meaning something like List. To be a bit more explicit, I’m going to write List<_>, where the _ indicates an “unspecified” type parameter.

So this signature is effectively saying that it takes as input a I<i32> (for some I) and returns an I<f32> – the intention is to mean that it takes a collection of integers and returns the same sort of collection, but applied to floats (so, e.g., I might be mapped to List<_> or Vec<_>, yielding List<i32>/List<f32> or Vec<i32>/Vec<f32> respectively). But is that what it really says? It turns out that this question is bit more subtle than you might think; let’s dig in.

Trait bounds, higher-ranked and otherwise

The first thing to notice is that floatify_hkt() is missing some where-clauses. In particular, nowhere do we declare that I<i32> is supposed to be a collection. To do that, we would need something like this:

fn floatify_hkt<I<_>>(ints: &I<i32>) -> I<f32>
    where for<T> I<T>: Collection<T>
    //    ^^^^^^^^^^^^^^^^^^^^^^^^^^ "higher-ranked trait bound"

Here I am using the “higher-ranked trait bounds (HRTB) applied to types” introduced by RFC 1598, and discussed in the previous post. Basically we are saying that I<T> is always a Collection, regardless of what T is.

So we just saw that we need HRTB to declare that any type I<T> is a collection (otherwise, we just know it is some type). But (as far as I know) Haskell doesn’t have anything like HRTB – in Haskell, trait bounds cannot be higher-ranked, so you could only write a declaration that uses explicit types, like so:

fn floatify_hkt<I<_>>(ints: &I<i32>) -> I<f32>
    where I<i32>: Collection<i32>,
          I<f32>: Collection<f32>,

In this case, that’s a perfectly adequate declaration. But in some cases, being forced to write out explicit types like this can cause you to expose information in your interface you might otherwise prefer to keep secret. Consider this function process()1, which takes a collection of inputs (of type Input) and returns a collection of outputs (of type – wait for it – Output). The interesting thing about this function is that, internally, it creates a temporary collection of some intermediate type called MyType:

fn process<I<_>>(inputs: &I<Input>) -> I<Output>
    where I<Input>: Collection<Input>, I<Output>: Collection<Output>,
{
    struct MyType { ... }
    
    // create an intermediate collection for some reason or other
    let mut shapes: I<MyType> = points.iter().map(|p| ...).collect();
    //              ^^^^^^^^ wait, how do I know I<MyType> is a collection?

    ...
}

Now you can see the problem! We know that I<Input> is a collection, and we know that I<Output> is a collection, but without some form of HRTB, we can’t declare that I<MyType> is a collection without moving MyType outside of the fn body2. So being able to say something like “I<T> is a collection no matter what T is” is actually crucial to our ability to encapsulate the internal processing that we are doing.

So, if Haskell lacks HRTB, how do they handle a case like this anyway?

Higher-kinded self types

If you have higher-kinded types at your disposal, you can use them to achieve something very similar to higher-ranked trait bounds, but we would have to change how we defined our Collection trait. Currently, we have a trait Collection<T> which is defined for some collection type C; the type C is then considered a collection of items of type T. So for example C might be List<Foo> (in which case T would be Foo). The new idea would be to redefine Collection to be defined over collection type constructors (like List<_>). So we might write something like this:

trait HkCollection for Self<_> {
//    ^^               ^^^^^^^ declare that `Self` is a type constructor
//    stands for "higher-kinded"

    fn empty<T>() -> Self<T>;
    fn add<T>(self: &mut Self<T>, value: T);
    //    ^^^ the `T` effectively moved from the trait to the methods
    ...
}

Now I might implement this not for List<T> but rather for List<_>:

impl HkCollection for List<_> {
    fn empty<T>() -> List<T> {
        List::new()
    }
    ...
}

And, finally, instead of writing where for<T> I<T>: Collection<T>, we can write where I: HkCollection. Note that here I bounded I, not I<_>, since I am applying this trait not to any particular type, but rather to the type constructor.

At first it may appear that these two setups are analogous, but it turns out that the “higher-kinded self types” approach has some pretty big limitations. Perhaps the most obvious is that it rules out collections like BitSet, which can only store values of one particular type:

impl HkCollection for BitSet { ... }
//                    ^^^^^^ not a type constructor

Note that with the older, non-higher-kinded collection trait, we could easily do something like this:

impl Collection<usize> for BitSet { ... }

The same problem also confronts collections like HashSet or BTreeSet that require bounds – that is, even though these are generic types, you can’t actually make a HashSet of just any old type T. It must be a T: Hash. In other words, when I write something like Self<_>, I am actually leaving out some important information about what kinds of types the _ can be:

trait HkCollection for HashSet<_>
//                     ^^^^^^^ how can we restrict `_` to `Hash` types?

In Haskell, at least, if I have a HKT, that means I can apply this type constructor to any type and get a result. But all collections in Rust tend to apply some bounds on that. For example, Vec<T> and List<T> both (implicitly) require that T: Sized. Or, if you have HashMap<K,V>, you might consider it to be a collection of pairs (K, V), except that it only works if K: Hash + Eq + Sized and V: Sized.

So, really, if we did want to support a syntax like Foo<_>, we would actually need some way of constraining this _.

SPJ’s “Type Classes: Exploring the Design Space”

Naturally, Haskell has encountered all of these problems as well. One of my favorite papers is “Type Classes: Exploring the Design Space” by Jones et al., published way back in 1997. They motivate “multiparameter type classes” (which in Rust would be “generic traits” like Collection<T>) by reviewing the various shortcomings of traits defined with a higher-kinded Self type (like HkCollection):

  • Section 2.1, “Overloading with coupled parameters” basically talks about the idea that impls might not always apply to all types. So something like impl Collection<usize> for BitSet is a simple example – if you choose the “collection family” to be BitSet, you can then forced to pick usize as your element type.
    • In these situations, it is often (but not always) the case that the “second” parameter could (and perhaps should) be an associated type. For example, we might have changed trait Collection<Item> { ... } to trait Collection { type Item; ... }. This would have meant that, for any given collection type, there is a fixed Item type.
      • So, for example, the BitSet imply that applied to any integral type would be illegal, because the type BitSet alone does not define the item type T:
        • impl<T: Integer> Collection for BitSet { type Item = T; ... }
      • I talked some about this tradeoff in the “Things I learned” section from my post on Rayon; the rule of thumb I describe there seems to suggest Collection<T> would be better, though I think you could argue it the other way. We’ll have to experiment.
  • Section 2.2, “Overloading with constrained parameters” covers the problem of wanting constraints like T: Sized or T: Hash. In Haskell, the Sized bound isn’t necessary, but certainly things like HashSet<T> wanting T: Hash still applies.

Obviously this paper is pretty old (1997!), and a lot of new things in Haskell have been developed since then (e.g., I think the paper predates associated types in Haskell). I think this core tradeoff is still relevant, however. Let me know though if you think I’m out of date and I need to read up on feature X which tries to address this trade-off. (For example, is there any treatment of higher-kinded types 5Bthat adds the ability to constrain parameters in some way?)

Time to get a bit more formal

OK, I want to get a bit more formal in terms of how I am talking about HKT. In particular, I want to talk more about what a kind is and why we could call a type constructor like List<_> higher-kinded. The idea is that just like types tell us what sort of value we have (e.g., i32 vs f32), kinds tell us what sort of generic parameter we have.

In fact, Rust already has two kinds: lifetimes and types. Consider the item ListIter that we saw earlier:

struct ListIter<'iter, T> { ... }`

Here we see that there are two parameters, 'iter and T, and the first one represents a lifetime and the second a type. Let’s say that 'iter has the kind lifetime and T has the kind type (in Haskell, people would write type as *) .

Now what is the kind of ListIter<'foo, i32>? This is also a type.

So what is the kind of a type constructor like ListIter<'foo, _>? This is something which, if you give it a type, you get a type. That sounds like a function, right? Well, the idea is to write that kind as type -> type.

And so higher-kinded type parameters are kind of like functions, except that instead of calling them at runtime (foo(22)), you apply them to types (Foo<i32>). In general, when we can talk about something “callable”, we tend to call it “higher-“, so in this case we say “higher-kinded”.

You can also imagine higher-kinded type parameters that abstract over lifetimes. We might write this like ListIter<'_, i32>, which would correspond to the kind lifetime -> type. If you had a parameter I<'_>, then you could apply it like I<'foo>, and – assuming I = ListIter<'_, i32> – you would get ListIter<'foo, i32>.

Speaking more generally, we can say that the kind K of a type parameter can fit this grammar:

K = type | lifetime | K -> K

Note that this supports all kinds of crazy kinds, like I<_<_>>, which would be (type -> type) -> type. This is like a Foo that is not parameterized by another type, but rather by a type constructor, so one would not write Foo<i32>, but rather Foo<Vec<_>>. Wow, meta.

Note that everything here assumes that if you have a type constructor I of kind type -> type, we can apply I to any type. There’s no way to say “types that are hashable”. In later posts, I hope to dig into this a bit more, and show that HRTB (and traits) can provide us a means to express things like that.

Decidability and inference

So you may have noticed that, in the previous paragraph, I was making all kinds of analogies to higher-kinded types being like functions. And certainly you can imagine defining “general type lambdas”, so that if you have a type parameter of kind type -> type, you could supply any kind of function which, given one type, yields another. But it turns out this is likely not what we want, for a couple of reasons:

  1. It doesn’t actually express what we wanted.
  2. It makes inference imposssible.

To get some intuition here, Let’s go back to our first example:

fn floatify_hkt<I<_>>(ints: &I<i32>) -> I<f32>

Here, I is declared a parameter of kind type -> type. Now remember that our intention was to say that these two parameters were the same “sort” of collection (e.g., we take/return a Vec<i32>/Vec<f32> or a List<i32>/List<f32>, but not a Vec<i32>/List<f32>). If however I can be any “type lambda”, then I could be a lambda that returns Vec<i32> if given an i32, and List<f32> is given an f32. We might imagine pseudo-code that uses if and talks about types, like this:

type I<T> = if T == i32 { Vec<i32> } else { List<f32> };

At this point, if you’ve been carefully reading along, this should be striking a memory. This sounds a lot like our first attempt at family traits from the previous post! Let’s go back in time to that first take on floatify_family():

fn floatify_family<F>(ints: &F::Collection<i32>) -> F::Collection<f32>
    where F: CollectionFamily

Basically here the F is playing exactly this “type lambda” role. F::Collection<T> is the same as I<T>. Moreover, using impls and traits, we can write arbitrary, turing-complete functions on types!

(Note: it sounds like being turing-complete is hard; it’s not. It’s actually hard to avoid once you start adding in any reasonably expressive system. You essentially have to add some special-cases and limitations to do it.)

This implies that if we permit higher-kinded type parameters like I to be mapped to just any old kind of “type lambda”, our inference is going to get stuck. So whenever you called floatify_hkt(), you would need to explicitly annotate the “type lambda” I. Note that this is worse than something like collect(), where all we need to know is what the return type is, and we can figure everything out. Here, even if we know the argument/return types, we can’t figure out the function that maps between them, at least not uniquely.

As an analogy, it’d be like if I told you “ok, so f(1) = 2 and f(2) = 3, what is the function f?”. Naturally there is no unique answer. You might think that the answer is f(x) = 1 + x, and that does fit the data, but of course that’s not the only answer. It could also be f(x) = min(x + 1, 10), and so forth.

Limiting higher-kinded types via currying, like Haskell

The way that Haskell solves this problem is by limiting higher-kinded types. In particular, they say that a higher-kinded type has to be (the equivalent of) a struct or enum name with some suffix of parameters left blank.

So that means that if you have a kind like type -> type, it could be satisfied with Vec<_> or Result<i32, _>, but not Result<_, i32> and certainly not some more complex function. It also means that if you have aliases (like type PairVec<T> = Vec<(T, T)> in Rust), you can’t make an HKT from PairVec<_>.

This scheme has a lot of advantages! In particular, let’s go back to our type inference problem. As you recall, the fundamental kind of constraint we end up with is type equalities. In that case, we wind up knowing the “inputs” to a HKT and the “output”. So I might have something like:

?1<?2> = Result<i32, u32>

Since ?1 can’t be just any function, I can uniquely determine that ?1 = Result<i32, _> and ?2 = u32. There is just nothing else it could be!

(This scheme is called currying in Haskell and it’s actually really quite elegant, at least in terms of how it fits into the whole abstract language. It’s basically a universal principle in Haskell that any sort of “function” can be converted into a lambda by leaving off a suffix of its parameters. I won’t say more because (a) converting the examples into Rust syntax doesn’t really give you as good a feeling for its elegance and (b) this post is long enough without explaining Haskell too!)

In fact, we can go even futher. Imagine that we have an equality like this, where we don’t really know much at all about either side:

?1<?2> = ?3<?4>

Even here, we can make progress, because we can infer that ?1 = ?3 and ?2 = ?4. This is a pretty strong and useful property.

Problems with currying for Rust

So there are a couple of reasons that a currying approach wouldn’t really be a good fit for Rust. For one thing, it wouldn’t fit the & “type constructor” very well. If you think of types like &'a T, you effectively have a type of kind lifetime -> type -> type (well, not exactly; the type T must outlive 'a, giving rise to the same matter of constrained types I raised earlier, but this problem is not unique to & and applies to most any generic Rust type). Essentially, give me a lifetime ('a) and a type (T) and I will give you a new combined type &'a T. OK, so far so good, but if we follow a currying-based approach, then this means that you can partially apply & to a particular lifetime ('a), yielding a HKT like type -> type. This is good for those cases where you wish to treat &'a T interchangeably with other pointer-like types, such as Rc<T>.

But then there are times like Iterable, where you might like to be able to take a base type like &'a T and plugin other lifetimes to get &'b T. In other words, you might want lifetime -> type. But using a Haskell-like currying approach you basically have to pick one or the other.

Another problem with currying is that you always have to leave a suffix of type parameters unapplied, and that is just (in practice) unlikely to be a good choice in Rust. Imagine we wanted to use a map-like type parameter M<_,_>, so that (say) we could take in a M<i32, T> and convert it to a map M<f32, T> of the same basic kind. Now consider the definition of HashMap, which actually has three parameters (one of which is defaulted):

pub struct HashMap<K, V, S = RandomState>

We would have wanted M = HashMap<_, _, S>, but we can’t do that, because that’s a prefix of the types we need, not a suffix.

One strategy that might work ok in practice is to say, in Rust, you can name a HKT by putting an _ on some prefix of the parameters for any given kind. So e.g. we can do the following:

  • &'a T yielding type
  • &'_ T yielding lifetime -> type
  • &'a _ yielding type -> type
  • Ref<'_, T> yielding lifetime -> type
  • Ref<'a, _> yielding type -> type
  • HashMap<_, i32, S> yielding type -> type (where the first type is the key)
  • HashMap<_, _, S> yielding type -> type -> type
  • Result<_, Err> yielding type -> type

but we could not do any of these, because in each case the _ is not a prefix:

  • Foo<'a, '_, i32>
  • HashMap<i32, _, S>

Obviously it’s unfortunate that the _ would have to be a prefix, but that’s basically a necessary limitation to support type inference. If you permitted _ to appear anywhere, then only the most basic constraints become solveable – essentially in practice you wind up with a scenario where all type parameters must become _, and partial application never works. To see what I mean, consider some examples:

  • ?T<?U> = Rc<i32>, solvable:
    • could be ?T = Rc<_>, ?U = i32
    • but not that this only works because all type parameters of Rc were made into _
  • ?T<?U> = Result<i32, u32>, unsolvable:
    • could be ?T = Result<_, u32>, ?U = i32
    • could be ?T = Result<i32, _>, ?U = u32
  • ?T<?U, ?V> = Result<i32, u32>, solvable if we assume that ordering must be respected:
    • could be ?T = Result<_, _>, ?U = i32, ?V = u32
    • again, this only works because all type parameters of Result were made into _

So, essentially, choosing a prefix (or suffix) is actually more expressive in practice than allowing _ to go anywhere, since the latter would cripple inference and require manual type annotation.

So, what do you do if you’d like to be able to put the _ anywhere? Say, because you want the choice of Result<_, E> or Result<T, _>? The answer is that you build “wrapper types”, like:

struct Unresult<E, T> {
    result: Result<T, E>
}

(Note that this won’t work with a plain type alias, as you can’t partially apply a type alias. This is precisely because Unresult<X, Y> is a distinct type from Result<Y, X>, which is not the case with a type alias.)

I find this kind of interesting because it starts to resemble the “dummy” types that we made for families. But this is not really a “win” for ATC or family traits in particular. After all, you only need said dummy types when the default order isn’t working for you; and, if you wanted to make one collection type (like List<T>) participate in two different collection families, you’d need a wrapper there too.

Side note: Alternatives to currying

I am not sure of the full space of alternatives here.

For example, it may be possible to permit higher-kinded types to be assigned to more complex functions, but only if the user provides explicit type hints in those cases. This would be perhaps analogous to higher-ranked types in Haskell, which sometimes require a certain amount of type annotation since they can’t be fully inferred in general.

Another fruitful area to explore is the branch of logic programming, where this sort of inference is referred to as higher-order unification – basically, solving unification problems where you have variables that are functions. Unsurprisingly, unrestricted higher-order unification is a pretty thorny problem, lacking most of the nice properties of first-order unification. For example, there can be an infinite number of solutions, none of which is more general than the other; in fact, in general, it’s not even decidable whether there is a solution or not!

Now, none of this means that there don’t exist algorithms for solving higher-order unification. In particular, there is a core solution called Huet’s algorithm; it’s just that it is not guaranteed to terminate and may generate an infinite number of solutions. Nonetheless, in some settings it can work quite well.

There is also a subset of the higher-order unification called higher-order pattern matching. In this subset, if I understand correctly, we can solve unification constraints with higher-kinded variables, but only if they look like this:

for<T> (?1<T> = U<T>)

The idea here is that we are constraining ?1<T> to be equal to U<T> no matter what T is. In this case, clearly, ?1 must be equal to U. Apparently, this subset appears often in higher-order logic programming languages like Lambda Prolog, but sadly it doesn’t seem that relevant to Rust.

Conclusions

This concludes our first little tour of what HKT is, and what it might mean for Rust. Here is a little summary of the some of the highlights:

  • Higher-kinded types let you use a type constructor as a parameter; so you might have a parameter declared like I<_> whose value is Vec<_>; that is, the Vec type without specifying a particular kind of element.
  • Higher-ranked trait bounds (which Haskell doesn’t offer, but which are part of the ATC RFC) permit functions to declare something like “I<T> is a collection of T elements, regardless of what T is”.
    • Otherwise, you have to have a series of constraints like I<i32>: Collection<i32>, I<f32>: Collection<f32>.
    • This can reveal implementation details you might prefer to hide.
  • What Haskell does offer as an alternative is traits whose Self type is higher-kinded.
    • However, because HKT in Haskell do not permit where-clauses or conditions, such a trait would not be usable for collections that impose limitations on their element types (i.e., basically all collections):
      • BitSet might require that the element is a usize;
      • HashSet<T> requires T: Hash;
      • BTreeSet<T> requires T: Ord;
      • heck, even Vec<T> requires T: Sized in Rust!
    • Thus, a tradeoff is born between multi-parameter type classes, which permit such conditions, and type classes based around higher-kinded types.
    • To be usable in Rust, we would have to extend the concept of HKT to include where clauses, since almost all Rust types include some condition, even if only T: Sized.
    • Note that collection families naturally permit one to apply where conditions and side clauses.
  • Higher-kinded types in Haskell are limited to a “curried type declaration”:
    • This makes type inference tractable and feels natural in Haskell.
    • Exporting this scheme to Rust feels awkward.
    • One thing that might work is that one can omit a prefix of the type parameters of any given kind.

OK, that’s enough for one post! In the next post, I plan to tackle the following question:

  • What is the difference between an “associated type constructor” and an “associated HKT”?
    • What might it mean to unify those two worlds?

Comments

Please leave comments on this internals thread.

Footnotes

  1. I was not able to find an even-mildly-convincing variant on floatify for this. =)

  2. This problem is not specific to types declared in fn bodies; one can easily construct similar examples where it would be necessary to make private structs public.