Refining traits and impls

4 October 2012

Currently, the Rust compiler accepts all manner of trait, impl, and bound declarations. In fact, it accepts plenty of declarations that later phases of the compiler are not sophisticated enough to handle. In other words, the syntax is writing checks the semantics can’t cash. (An aside: I just love saying that phrase for some perverse reason. I really wish however that checks, like rotary dial telephones, were something that younger people vaguely understood but which no longer had relevance in the modern era. The Swiss Einzahlungschein truly opened my eyes! Anyhow.)

I would like to solve this. In particular, I would very much like to have rules with the following properties:

  • Coherent and unambiguous: There is at most one impl of a given trait for a given type.
  • Safe linking: Linking in a new crate cannot cause a static error nor violate any of the above properties.
  • Overloading- and backtracking-free: Given an expression a.m(...), the choice of which method m to invoke should be based only on the type of a. More generally, given a type T and a trait N, we should be able to decide whether T implements N without considering other types or other traits.
  • Termination: Compilation always terminates.

What follows is a justifiation of those principles and a description of how to achieve them. If you don’t want to read it, here is the summary:

Aside from the coherence requirement, the main restriction is that a given type can only implement or extend a given trait once (and hence with one set of type parameters).

Why these principles?

None of these principles is strictly necessary. Haskell for example maintains coherence but does not (necessarily) guarantee safe linking; you can have modules that do not link together due to incompatible typeclass declarations (actually, given the right configuration settings, Haskell can either violate coherence or guarantee safe-linking, so I guess what I mean is “based on my folklore idea of which specific settings are commonly used”). Scala does not guarantee coherence, though they can use inheritance to do so when it is important. Haskell’s multiparameter type classes are not overloading-free in general, though they can be made overloading-free through the use of functional dependencies or type functions.

I won’t bother to justify coherence nor monotonicity, but I do want to discuss overloading freedom a bit. It has several nice implications. For one, the implementation never needs to backtrack: that is, it can consider whether a type implements a trait simply by searching through the impls for a trait, rather than having to consider whether a set of types implement a trait. This simplifies the implementation; it also fits better with the object-oriented flavor of our traits/impls: when you write <A: N<B, C, D>>, it is the type A which decides which impl of N will be used, not the types B, C, or D. Another nice property is that it is consistent with the rest of the language, which doesn’t support overloading (except through the trait mechanism).

Interestingly, in the paper “Type classes: an exploration of the design space”, Jones et al. discuss several uses for multiparameter type classes which they find compelling. If I am not mistaken, all of those examples are compatible with the requirements I propose here, with the exception of overloading for mathematical operations, and I describe a workaround for that case below.

What restrictions are necessary?

Aside from the coherence requirement, the main restriction is that a given type can only implement or extend a given trait once (and hence with one set of type parameters).

More formally phrased, the full set of restrictions on traits and impls are:

  1. You can only provide one implementation of any particular trait for a given type. In other words, for any two impls with the same trait name:

    impl<...> T1: N<...> {...}
    impl<...> T2: N<...> {...}
    

    the self types T1 and T2 must not have a common subtype for any substitutions of the various impl parameters. Note in particular that it does not matter what type parameters are provided to the trait N. In other words, the following set of impls would be illegal:

     impl ~str: Iterable<~str> {...} // by word
     impl ~str: Iterable<char> {...} // by character
    

    These are illegal because both implement the Iterable trait but with the same type.

  2. There are no “orphan” implementations. This means that for every impl<...> T: N<...> defined in the current crate, at least one of the following two conditions holds:

    • the trait N is declared in the current crate;
    • the self type T is either a nominal type declared in the current crate or an arity-1 type constructor whose argument (transitively) is declared in the current crate (e.g., Foo, @Foo, ~[Foo], or—for that matter—~[@Foo], where Foo is declared in the current crate).
  3. If the self type of an impl for trait N<...> is a type parameter T, T must not be bounded (directly or indirectly) by the trait N.

  4. Trait inheritance must form a tree and no trait N may inherit from another trait M more than once with multiple values for M’s type parameters.

Restriction 1 is concerned with ensuring freedom from both ambiguity and overloading. Restriction 2 guarantees safe linking. Restriction 3 is required to ensure that compilation always terminates, I believe. Otherwise you might have some impl like impl<T: Eq> T: Eq and we could wind up searching forever trying to find out whether T implements Eq (though I imagine one could write an algorithm that knows to terminate, sort of like subtyping for recursive structural types, I just don’t care to, since an impl like this has no purpose in practice).

It is interesting to compare this to Haskell. The situation is loosely similar to a multiparameter type class specified as follows:

 class C self a ... z | self -> a ... z where ...

I have a “read-only” relationship with Haskell, so there’s probably a mistake in the syntax somewhere, but basically I mean that the first type parameter to the type class (self) determines the values for the other type parameters (a...z). However, the analogy is not exact. With functional dependencies, knowing self would tell you a...z precisely. In our system it’s not quite that simple:

  1. The self type in our system in general can bound the later type parameters a...z but due to variance does not strictly determine them.
  2. The rules I gave permit an impl such as impl<A> Foo: Bar<A>. In this case, the self type (Foo) does not even bound th type variable A. I don’t think this has any ill effects, but it does make me nervous, so we could add another restriction that every type variable appearing in the trait type must appear in the self type.

One interesting distinction between multiparameter typeclasses in Haskell and the rules I propose is that we support an impl like impl<A: Foo> A: Bar: that is, an impl that allows you to implement a trait Bar given an instance of the trait Foo. This pattern of impl is useful in some cases as I discuss below, though it can also be abused (also discussed below). One must be cautious here to ensure termination. I believe that restriction #3 is sufficient, though it seems somewhat less restrictive than Haskell’s rules (but we are not trying to infer as much as Haskell is). I probably ought to read these papers to be sure, since they apparently discuss the matter in more detail.

UPDATE: I discuss this analogy to Haskell further in the next post.

UPDATE: I am pretty sure that the termination condition is insufficient. More to come.

What if I want overloading?

Despite the fact that I want to prevent it, overloading can sometimes be very useful. For example, consider the trait Add that is used to define the meaning of the + operator:

trait Add<R,S> {
    pure fn add(&self, rhs: &R) -> S;
}

As you can see, this trait is parameterized by the type of the right-hand side (R) and the type of the sum (S).

Now, imagine that we wished to permit the addition of ints with ints but also ints with floats (a classic example of overloading). We might have the following implementations of Add:

impl int: Add<int, int> { ... }
impl int: Add<float, float> { ... }

We have effectively overloaded the add() method now. I can add ints to ints or ints to floats. So if I write x + y (which is basically shorthand for (&x).add(&y)), the compiler must use both the type of x and y to decide which impl will be invoked.

Interestingly, it is possible to encode something very much like overloading by making use of double-dispatch in a style somewhat similar to the visitor pattern:

trait IntRhs<S> {
    pure fn add_to_int(lhs: int) -> S;
}
impl<S, R: IntRhs<S>> int: Add<R, S> {
    pure fn add(&self, rhs: &R) -> S {
        rhs.add_to_int(*self)
    }
}
impl int: IntRhs<int> {...}
impl float: IntRhs<float> {...}

Although this pattern is wordier, it is also more flexible: it is now possible for other types to implement IntRhs and hence be usable as the right-hand-side of an addition with an integer. In other words, this pattern is probably how you should be doing it anyhow.

When should I use an impl whose self type is a type parameter?

I spent some time looking through the impls that appear in our standard library. Most are fairly straightforward. There is one pattern though that stands out as unusal. You can have an impl which has the form:

impl<A: Foo> A: Bar { ... }

This impl implements the trait Bar for any type A that implements the trait Foo. This kind of impl is supported under the rules I propose; however, it is only permitted if (1) Bar is defined in the current crate and (2) there are no other impls of the trait Bar.

The reason that there can be no other impls is that if there were another impl, no matter what self type it had, it would violate restriction #1 from my list, since you could substitute that type for A. Note that restriction #1 did not require that the substituted type meet the bounds placed on the type variable—in other words, it is a violation to implement the trait Bar for any other type T, even if T does not implement Foo! This is an intentional limitation: it helps ensure that the implementation does not need to backtrack. It also has other nice properties: for example, adding an implementation for the trait Foo can never cause ambiguity errors with respect to implementations of the trait Bar.

It turns out that we use this impl in two distinct pattern. The first of those patterns is compatible with restriction #1 and the second is not. However, there is a better way to express the second pattern.

Pattern #1: Extension traits

Sometimes we simply wish to extend an interface with “convenience” methods:

trait Hash {
    pure fn hash_keyed(k0: u64, k1: u64) -> u64;
}
trait HashUtil {
    pure fn hash() -> u64;
}
impl<A: Hash> A: HashUtil {
    pure fn hash() -> u64 { self.hash_keyed(0,0) }
}

In this particular case, a default method implementation would likely be a better choice. However, this pattern can still be useful if the trait you are extending is outside of your control.

The key point here is that there will never be another implementor of this trait. That is, you could just as well have written this impl as a standalone function (but then you couldn’t use dot notation):

fn hash<A: Hash>(a: &A) -> u64 {
    a.hash_keyed(0, 0)
}

Pattern #2: Simulating inheritance

We have one impl which says that everything which can be converted into a stream of bytes is also hashable:

impl<A: IterBytes> A: Hash {
    pure fn hash_keyed(k0: u64, k1: u64) -> u64 {
        /* apply a super awesome cryptographic hashing algorithm
           that goes way over my head */
    }
}

Technically speaking, this example is precisely the same as the previous section, but it’s purpose is very different. I believe that the intention was to define a kind of “default” method that says: for every type that is byte-iterable, it can be hashed as follows, if no better way is provided."

There are two problems with this solution. The first is that it raises interesting ambiguity questions if there is a type that implies Hash and also IterBytes—which version of Hash should we prefer? The explicit one or the “default-y” one based on IterBytes?

Even if we assume that all types will either implement Hash OR IterBytes, impls like this make resolution difficult. Imagine we have some type Foo that implements Hash and not IterBytes. When we are searching for implementations of Hash that apply to Foo, we might chance upon the impl above (written in terms of IterBytes) first. In that case, we would see that if we instantiate the type variable A to be the type Foo, then it matches the receiver type—but we would also inherit the additional obligation of finding an implementation of IterBytes for Foo. Of course this search would fail, and then we would (presumably) have to keep searching for another impl that matches Foo. This is what I mean by backtracking—even though unification succeeded, we don’t know that the impl applies.

In general I am not interesting in supporting this use-case. A better way to implement this pattern would be to model the relationship using inheritance and default methods. For example, the definition of the IterBytes trait would be:

trait IterBytes : Hash {
    pure fn hash_keyed(k0: u64, k1: u64) -> u64 {
        ...
    }
}

What else can we crib from Haskell?

Browsing through the list of type class extensions in the Haskell’s user manual, I have some thoughts about enabling something equivalent to “Class method types”. That’ll have to wait for another blog post. We already support many of the “relaxed rules” that are listed there (such as “relaxed rules for the instance head” and “relaxed rules for the instance contexts”). We will never support overlapping or incoherent impls, I believe.

Appendix: Some implementation details

Here are some incomplete thoughts on how each algorithm will work. These are intended more as notes for myself than anything else. I wanted to think out some of the issues in detail to be sure I wasn’t missing anything.

Checking coherence, non-ambiguity, and overloading-freedom

The coherence check consists of two parts. The first check ensures that there are no cases of overlapping or overloadded impls. To do this, we iterate over all pairs of impls that target the same trait N (regardless of the type parameters). We replace all variables bound on each impl with fresh type/region variables and try to unify them. Put more formally, if T1 and T2 are the self types of the two impls, the check fails if there exists a substitution S such that S(T1) <: S(T2) or S(T2) <: S(T1).

Note that we must be careful about about trait inheritance. It is possible to have an impl for trait N where N extends M that conflicts with an impl for trait M.

The second part of the coherence check ensures that there are no “orphan” impls, to use Haskell-speak. The orphan check reports an error for any impl that does not meet (at least) one of the two following criteria:

  • the trait N is declared in the current crate;
  • the self type T is either a nominal type declared in the current crate or an arity-1 type constructor whose argument (transitively) is declared in the current crate (e.g., Foo, @Foo, ~[Foo], or—for that matter—~[@Foo], where Foo is declared in the current crate).

Checking a bounded type parameter

When a type variable X: N<U1...Un> is bound to a type T, we must ensure that T implements the trait N<U1...Un>. To do so we iterate through all implementations associated with the trait N, instantiate their bound variables, and attempt to unify their self type Ts with T. If this unification is successful (formally, if there exists a substitution S where T <: S Ts), then we must check that:

  1. the trait type N<V1...Vn> associated with the impl is a subtype of N<U1...Un> (formally, that N<S V1...S Vn> <: N<U1...Un>).
  2. the bounds associated with the bound variables on the impl are themselves satisfied.

Note that the non-ambiguity check guarantees us that at most one trait will have a self type that is unifiable with T. Therefore, we do not need to consider backtracking should property (2) above not hold.

Checking a method call

Checking a call to a trait method is much like checking a bounded type parameter. If we find a call a.m(...), and m() is defined on an in-scope trait T, then we can find all implementations of T. For each implementation, we instantiate the bound variables to yield a self-type. The procedure is basically identical.

Notes on F-bounds

I would like to support F-bounded polymorphism, meaning that the bound of a type parameter X may refer to X. For example, something like X: ComparableTo<X> is legal. In Rust, F-bounds are not needed as frequently as you might think due to the special self type available in every trait, but they can still play a role. Today I think our implementation does not support F-bounds but that is a shortcoming we should correct. As it happens, I do not believe this adds an particular difficulty; you can implement it in the “usual way”.

Just to spell out one particular such case for reference, imagine a trait and impl pair such as:

trait Foo<A: Foo<A>> { ... }
impl int: Foo<int> { ... }

In this case, while sanity checking the impl declaration we would see that int is used as the value for Foo’s type parameter, and hence need to validate that, yes, int does implement the trait Foo<int>. We do this simply by searching through the impls in scope; we will find the impl int: Foo<int>. But now we are not checking the impl but rather using it and hence we can trust it to be true to its word (coinduction FTW!), therefore we know that int implements Foo<int> (presuming the whole program typechecks).