Cyclic queries in chalk

12 September 2017

In my last post about chalk queries, I discussed how the query model in chalk. Since that writing, there have been some updates, and I thought it’d be nice to do a new post covering the current model. This post will also cover the tabling technique that scalexm implemented for handling cyclic relations and show how that enables us to implement implied bounds and other long-desired features in an elegant way. (Nice work, scalexm!)

What is a chalk query?

A query is simply a question that you can ask chalk. For example, we could ask whether Vec<u32> implements Clone like so (this is a transcript of a cargo run session in chalk):

?- load libstd.chalk
?- Vec<u32>: Clone
Unique; substitution [], lifetime constraints []

As we’ll see in a second, the answer “Unique” here is basically chalk’s way of saying “yes, it does”. Sometimes chalk queries can contain existential variables. For example, we might say exists<T> { Vec<T>: Clone } – in this case, chalk actually attempts to not only tell us if there exists a type T such that Vec<T>: Clone, it also wants to tell us what T must be:

?- exists<T> { Vec<T>: Clone }
Ambiguous; no inference guidance

The result “ambiguous” is chalk’s way of saying “probably it does, but I can’t say for sure until you tell me what T is”.

So you think can think of a chalk query as a kind of subroutine like Prove(Goal) = R that evaluates some goal (the query) and returns a result R which has one of the following forms:

  • Unique: indicates that the query is provable and there is a unique value for all the existential variables.
    • In this case, we give back a substitution saying what each existential variable had to be.
    • Example: exists<T> { usize: PartialOrd<T> } would yield unique and return a substitution that T = usize, at least today (since there is only one impl that could apply, and we haven’t implemented the open world modality that aturon talked about yet).
  • Ambiguous: the query may hold but we could not be sure. Typically, this means that there are multiple possible values for the existential variables.
    • Example: exists<T> { Vec<T>: Clone } would yield ambiguous, since there are many T that could fit the bill).
    • In this case, we sometimes give back guidance, which are suggested values for the existential variables. This is not important to this blog post so I’ll not go into the details.
  • Error: the query is provably false.

(The form of these answers has changed somewhat since my previous blog post, because we incorporated some of aturon’s ideas around negative reasoning.)

So what is a cycle?

As I outlined long ago in my first post on lowering Rust traits to logic, the way that the Prove(Goal) subroutine works is basically just to iterate over all the possible ways to prove the given goal and try them one at a time. This often requires proving subgoals: for example, when we were evaluating ?- Vec<u32>: Clone, internally, this would also wind up evaluating u32: Clone, because the impl for Vec<T> has a where-clause that T must be clone:

impl<T> Clone for Vec<T>
where
  T: Clone,
  T: Sized,
{ }

Sometimes, this exploration can wind up trying to solve the same goal that you started with! The result is a cyclic query and, naturally, it requires some special care to yield a valid answer. For example, consider this setup:

trait Foo { }
struct S<T> { }
impl<U> Foo for S<U> where U: Foo { }

Now imagine that we were evaluating exists<T> { T: Foo }:

  • Internally, we would process this by first instantiating the existential variable T with an inference variable, so we wind up with something like ?0: Foo, where ?0 is an as-yet-unknown inference variable.
  • Then we would consider each impl: in this case, there is only one.
    • For that impl to apply, ?0 = S<?1> must hold, where ?1 is a new variable. So we can perform that unification.
      • But next we must check that ?1: Foo holds (that is the where-clause on the impl). So we would convert this into “closed” form by replacing all the inference variables with exists binders, giving us something like exists<T> { T: Foo }. We can now perform this query.
        • Only wait: This is the same query we were already trying to solve! This is precisely what we mean by a cycle.

In this case, the right answer for chalk to give is actually Error. This is because there is no finite type that satisfies this query. The only type you could write would be something like

S<S<S<S<...ad infinitum...>>>>: Foo

where there are an infinite number of nesting levels. As Rust requires all of its types to have finite size, this is not a legal type. And indeed if we ask chalk this query, that is precisely what it answers:

?- exists<T> { S<T>: Foo }
No possible solution: no applicable candidates

But cycles aren’t always errors of this kind. Consider a variation on our previous example where we have a few more impls:

trait Foo { }

// chalk doesn't have built-in knowledge of any types,
// so we have to declare `u32` as well:
struct u32 { }
impl Foo for u32 { }

struct S<T> { }
impl<U> Foo for S<U> where U: Foo { }

Now if we ask the same query, we get back an ambiguous result, meaning that there exists many solutions:

?- exists<T> { T: Foo }
Ambiguous; no inference guidance

What has changed here? Well, introducing the new impl means that there is now an infinite family of finite solutions:

  • T = u32 would work
  • T = S<u32> would work
  • T = S<S<u32>> would work
  • and so on.

Sometimes there can even be unique solutions. For example, consider this final twist on the example, where we add a second where-clause concerning Bar to the impl for S<T>:

trait Foo { }
trait Bar { }

struct u32 { }
impl Foo for u32 { }

struct S<T> { }
impl<U> Foo for S<U> where U: Foo, U: Bar { }
//                                 ^^^^^^ this is new

Now if we ask the same query again, we get back yet a different response:

?- exists<T> { T: Foo }
Unique; substitution [?0 := u32], lifetime constraints []

Here, Chalk figured out that T must be u32. How can this be? Well, if you look, it’s the only impl that can apply – for T to equal S<U>, U must implement Bar, and there are no Bar impls at all.

So we see that when we encounter a cycle during query processing, it doesn’t necessarily mean the query needs to result in an error. Indeed, the overall query may result in zero, one, or many solutions. But how does should we figure out what is right? And how do we avoid recursing infinitely while doing so? Glad you asked.

Tabling: how chalk is handling cycles right now

Naturally, traditional Prolog interpreters have similar problems. It is actually quite easy to make a Prolog program spiral off into an infinite loop by writing what seem to be quite reasonable clauses (quite like the ones we saw in the previous section). Over time, people have evolved various techniques for handling this. One that is relevant to us is called tabling or memoization – I found this paper to be a particularly readable introduction. As part of his work on implied bounds, scalexm implemented a variant of this idea in chalk.

The basic idea is as follows. When we encounter a cycle, we will actually wind up iterating to find the result. Initially, we assume that a cycle means an error (i.e., no solutions). This will cause us to go on looking for other impls that may apply without encountering a cycle. Let’s assume we find some solution S that way. Then we can start over, but this time, when we encounter the cyclic query, we can use S as the result of the cycle, and we would then check if that gives us a new solution S'.

If you were doing this in Prolog, where the interpreter attempts to provide all possible answers, then you would keep iterating, only this time, when you encountered the cycle, you would give back two answers: S and S’. In chalk, things are somewhat simpler: multiple answers simply means that we give back an ambiguous result.

So the pseudocode for solving then looks something like this:

  • Prove(Goal):
    • If goal is ON the stack already:
      • return stored answer from the stack
    • Else, when goal is not on the stack:
      • Push goal on to the stack with an initial answer of error
      • Loop
        • Try to solve goal yielding result R (which may generate recursive calls to Solve with the same goal)
        • Pop goal from the stack and return the result R if any of the following are true:
          • No cycle was encountered; or,
          • the result was the same as what we started with; or,
          • the result is ambiguous (multiple solutions).
        • Otherwise, set the answer for Goal to be R and repeat.

If you’re curious, the real chalk code is here. It is pretty similar to what I wrote above, except that it also handles “coinductive matching” for auto traits, which I won’t go into now. In any case, let’s apply this to our three examples of proving exists<T> { T: Foo }:

  • In the first example, where we only had impl<U> Foo for S<U> where U: Foo, the cyclic attempt to solve will yield an error (because the initial answer for cyclic alls is errors). There is no other way for a type to implement Foo, and hence the overall attempt to solve yields an error. This is the same as what we started with, so we just return and we don’t have to cycle again.
  • In the second example, where we added impl Foo for u32, we again encounter a cycle and return error at first, but then we see that T = u32 is a valid solution. So our initial result R is Unique[T = u32]. This is not what we started with, so we try again.
    • In the second iteration, when we encounter the cycle trying to process impl<U> Foo for S<U> where U: Foo, this time we will give back the answer U = u32. We will then process the where-clause and issue the query u32: Foo, which succeeds. Thus we wind up yielding a successful possibility, where T = S<u32>, in addition to the result that T = u32. This means that, overall, our second iteration winds up producing ambiguity.
  • In the final example, where we added a where clause U: Bar, the first iteration will again produce a result of Unique[T = u32]. As this is not what we started with, we again try a second iteration.
    • In the second iteration, we will again produce T = u32 as a result for the cycle. This time however we go on to evaluate u32: Bar, which fails, and hence overall we still only get one successful result (T = u32).
    • Since we have now reached a fixed point, we stop processing.

Why do we care about cycles anyway?

You may wonder why we’re so interested in handling cycles well. After all, how often do they arise in practice? Indeed, today’s rustc takes a rather more simplistic approach to cycles. However, this leads to a number of limitations where rustc fails to prove things that it ought to be able to do. As we were exploring ways to overcome these obstacles, as well as integrating ideas like implied bounds, we found that a proper handling of cycles was crucial.

As a simple example, consider how to handle “supertraits” in Rust. In Rust today, traits sometimes have supertraits, which are a subset of their ordinary where-clauses that apply to Self:

// PartialOrd is a "supertrait" of Ord. This means that
// I can only implement `Ord` for types that also implement
// `PartialOrd`.
trait Ord: PartialOrd { }

As a result, whenever I have a function that requires T: Ord, that implies that T: PartialOrd must also hold:

fn foo<T: Ord>(t: T) {
  bar(t); // OK: `T: Ord` implies `T: PartialOrd`
}  

fn bar<T: PartialOrd>(t: T) {
  ...
}  

The way that we handle this in the Rust compiler is through a technique called elaboration. Basically, we start out with a base set of where-clauses (the ones you wrote explicitly), and then we grow that set, adding in whatever supertraits should be implied. This is an iterative process that repeats until a fixed-point is reached. So the internal set of where-clauses that we use when checking foo() is not {T: Ord} but {T: Ord, T: PartialOrd}.

This is a simple technique, but it has some limitations. For example, RFC 1927 proposed that we should elaborate not only supertraits but arbitrary where-clauses declared on traits (in general, a common request). Going further, we have ideas like the implied bounds RFC. There are also just known limitations around associated types and elaboration.

The problem is that the elaboration technique doesn’t really scale gracefully to all of these proposals: often times, the fully elaborated set of where-clauses is infinite in size. (We somewhat arbitrarily prevent cycles between supertraits to prevent this scenario in that special case.)

So we tried in chalk to take a different approach. Instead of doing this iterative elaboration step, we push that elaboration into the solver via special rules. The basic idea is that we have a special kind of predicate called a WF (well-formed) goal. The meaning of something like WF(T: Ord) is basically “T is capable of implementing Ord” – that is, T satisfies the conditions that would make it legal to implement Ord. (It doesn’t mean that T actually does implement Ord; that is the predicate T: Ord.) As we lower the Ord and PartialOrd traits to simpler logic rules, then, we can define the WF(T: Ord) predicate like so:

// T is capable of implementing Ord if...
WF(T: Ord) :-
  T: PartialOrd. // ...T implements PartialOrd.

Now, WF(T: Ord) is really an “if and only if” predicate. That is, there is only one way for WF(T: Ord) to be true, and that is by implementing PartialOrd. Therefore, we can define also the opposite direction:

// T must implement PartialOrd if...
T: PartialOrd :-
  WF(T: Ord). // ...T is capable of implementing Ord.

Now if you think this looks cyclic, you’re right! Under ordinary circumstances, this pair of rules doesn’t do you much good. That is, you can’t prove that (say) u32: PartialOrd by using these rules, you would have to use other rules for that (say, rules arising from an impl).

However, sometimes these rules are useful. In particular, if you have a generic function like the function foo we saw before:

fn foo<T: Ord>() { .. }

In this case, we would setup the environment of foo() to contain exactly two predicates {T: Ord, WF(T: Ord)}. This is a form of elaboration, but not the iterative elaboration we had before. We simply introduce WF-clauses. But this gives us enough to prove that T: PartialOrd (because we know, by assumption, that WF(T: Ord)). What’s more, this setup scales to arbitrary where-clauses and other kinds of implied bounds.

Conclusion

This post covers the tabling technique that chalk currently uses to handle cycles, and also the key ideas of how Rust handles elaboration.

The current implementation in chalk is really quite naive. One interesting question is how to make it more efficient. There is a lot of existing work on this topic from the Prolog community, naturally, with the work on the well-founded semantics being among the most promising (see e.g. this paper). I started doing some prototyping in this direction, but I’ve recently become intrigued with a different approach, where we use the techniques from Adapton (or perhaps other incremental computation systems) to enable fine-grained caching and speed up the more naive implementation. Hopefully this will be the subject of the next blog post!