For the last month or so, I’ve gotten kind of obsessed with exploring a new evaluation model for Chalk. Specifically, I’ve been looking at adapting the SLG algorithm, which is used in the XSB Prolog engine. I recently opened a PR that adds this SLG-based solver as an alternative, and this blog post is an effort to describe how that PR works, and explore some of the advantages and disadvantages I see in this approach relative to the current solver that I described in my previous post.

TL;DR

For those who don’t want to read all the details, let me highlight the things that excite me most about the new solver:

  • There is a very strong caching story based on tabling.
  • It handles negative reasoning very well, which is important for coherence.
  • It guarantees termination without relying on overflow, but rather a notion of maximum size.
  • There is a lot of work on how to execute SLG-based designs very efficiently (including virtual machine designs).

However, I also have some concerns. For one thing, we have to figure out how to include coinductive reasoning for auto traits and a few other extensions. Secondly, the solver as designed always enumerates all possible answers up to a maximum size, and I am concerned that in practice this will be very wasteful. I suspect both of these problems can be solved with some tweaks.

What is this SLG algorithm anyway?

There is a lot of excellent work exploring the SLG algorithm and extensions to it. In this blog post I will just focus on the particular variant that I implemented for Chalk, which was heavily based on this paper “Efficient Top-Down Computation of Queries Under the Well-formed Semantics” by Chen, Swift, and Warren (JLP ‘95), though with some extensions from other work (and some of my own).

Like a traditional Prolog solver, this new solver explores all possibilities in a depth-first, tuple-at-a-time fashion, though with some extensions to guarantee termination1. Unlike a traditional Prolog solver, however, it natively incorporates tabling and has a strong story for negative reasoning. In the rest of the post, I will go into each of those bolded terms in more detail (or you can click on one of them to jump directly to the corresponding section).

All possibilities, depth-first, tuple-at-a-time

One important property of the new SLG-based solver is that it, like traditional Prolog solvers, is complete, meaning that it will find all possible answers to any query2. Moreover, like Prolog solvers, it searches for those answers in a so-called depth-first, tuple-at-a-time fashion. What this means is that, when we have two subgoals to solve, we will fully explore the implications of one answer through multiple subgoals before we turn to the next answer. This stands in contrast to our current solver, which rather breaks down goals into subgoals and processes each of them entirely before turning to the next. As I’ll show you now, our current solver can sometimes fail to find solutions as a result (but, as I’ll also discuss, our current solver’s approach has advantages too).

Let me give you an example to make it more concrete. Imagine this program:

// sour-sweet.chalk
trait Sour { }
trait Sweet { }

struct Vinegar { }
struct Lemon { }
struct Sugar { }

impl Sour for Vinegar { }
impl Sour for Lemon { }

impl Sweet for Lemon { }
impl Sweet for Sugar { }

Now imagine that we had a query like:

exists<T> { T: Sweet, T: Sour }

That is, find me some type T that is both sweet and sour. If we plug this into Chalk’s current solver, it gives back an “ambiguous” result (this is running on my PR):

> cargo run -- --program=sour-sweet.chalk
?- exists<T> { T: Sour, T: Sweet }
Ambiguous; no inference guidance

This is because of the way that our solver handles such compound queries; specifially, the way it breaks them down into individual queries and performs each one recursively, always looking for a unique result. In this case, it would first ask “is there a unique type T that is Sour?” Of course, the answer is no – there are two such types. Then it asks about Sweet, and gets the same answer. This leaves it with nowhere to go, so the final result is “ambiguous”.

The SLG solver, in contrast, tries to enumerate individual answers and see them all the way through. If we ask it the same query, we see that it indeed finds the unique answer Lemon (note the use of --slg in our cargo run command to enable the SLG-based solver):

> cargo run -- --program=sour-sweet.chalk --slg
?- exists<T> { T: Sour, T: Sweet }     
1 answer(s) found:
- ?0 := Lemon

This result is saying that the value for the 0th (i.e., first) existential variable in the query (i.e., T) is Lemon.3

In general, the way that the SLG solver proceeds is kind of like a sort of loop. To solve a query like exists<T> { T: Sour, T: Sweet }, it is sort of doing something like this:

for T where (T: Sour) {
  if (T: Sweet) {
    report_answer(T);
  }
}

(The actual struct is a bit complex because of the possibility of cycles; this is where tabling, the subject of a later section, comes in, but this will do for now.)

As we have seen, a tuple-at-a-time strategy finds answers that our current strategy, at least, does not. If we adopted this strategy wholesale, this could have a very concrete impact on what the Rust compiler is able to figure out. Consider these two functions, for example (assuming that the traits and structs we declared earlier are still in scope):

fn foo() {
  let vec: Vec<_> = vec![];
  //           ^
  //           |
  // NB: We left the element type of this vector
  // unspecified, so the compiler must infer it.

  bar(vec);
  //   ^
  //   |
  // This effectively generates the two constraints
  //
  //     ?T: Sweet
  //     ?T: Sour
  //
  // where `?T` is the element type of our vector.
}

fn bar<T: Sweet + Sour>(x: Vec<T>) {
}

Here, we wind up creating the very sort of constraint I was talking about earlier. rustc today, which follows a chalk-like strategy, will fail compilation, demanding a type annotation:

error[E0282]: type annotations needed
  --> src/main.rs:15:21
     |
  15 |   let vec: Vec<_> = vec![];
     |       ---           ^^^^^^ cannot infer type for `T`
     |       |
     |       consider giving `vec` a type

An SLG-based solver of course could find a unique answer here. (Also, rustc could give a more precise error message here regarding which type you ought to consider giving.)

Now, you might ask, is this a realistic example? In other words, here there happens to be a single type that is both Sour and Sweet, but how often does that happen in practice? Indeed, I expect the answer is “quite rarely”, and thus the extra expressiveness of the tuple-at-a-time approach is probably not that useful in practice. (In particular, the type-checker does not want to “guess” types on your behalf, so unless we can find a single, unique answer, we don’t typically care about the details of the result.) Still, I could imagine that in some narrow circumstances, especially in crates like Diesel that use traits as a complex form of meta-programming, this extra expressiveness may be of use. (And of course having the trait solver fail to find answers that exist kind of sticks in your craw a bit.)

There are some other potential downsides to the tuple-at-a-time approach. For example, there may be an awfully large number of types that implement Sweet, and we are going to effectively enumerate them all while solving. In fact, there might even be an infinite set of types! That brings me to my next point.

Guaranteed termination

Imagine we extended our previous program with something like a type HotSauce<T>. Naturally, if you add hot sauce to something sour, it remains sour, so we can also include a trait impl to that effect:

struct HotSauce<T> { }
impl<T> Sour for HotSauce<T> where T: Sour { }

Now if we have the query exists<T> { T: Sour }, there are actually an infinite set of answers. Of course we can have T = Vinegar and T = Lemon. And we can have T = HotSauce<Vinegar> and T = HotSauce<Lemon>. But we can also have T = HotSauce<HotSauce<Lemon>>. Or, for the real hot-sauce enthusiast4, we might have:

T = HotSauce<HotSauce<HotSauce<HotSauce<Lemon>>>>

In fact, we might have an infinite number of HotSauce types wrapping either Lemon or Vinegar.

This poses a challenge to the SLG solver. After all, it tries to enumerate all answers, but in this case there are an infinite number! The way that we handle this is basically by imposing a maximum size on our answers. You could measure size various ways. A common choice is to use depth, but the total size of a type can still grow exponentially relative to the depth, so I am instead limiting the maximum size of the tree as a whole. So, for example, our really long answer had a size of 5:

T = HotSauce<HotSauce<HotSauce<HotSauce<Lemon>>>>

The idea then is that once an answer exceeds that size, we start to approximate the answer by introducing variables.5 In this case, if we imposed a maximum size of 3, we might transform that answer into:

exists<U> { T = HotSauce<HotSauce<U>> }

The original answer is an instance of this – that is, we can substitute U = HotSauce<HotSauce<Lemon>> to recover it.

Now, when we introduce variables into answers like this, we lose some precision. We can now only say that exists<U> { T = HotSauce<HotSauce<U>> } might be an answer, we can’t say for sure. It’s a kind of “ambiguous” answer6.

So let’s see it in action. If I invoke the SLG solver using a maximum size of 3, I get the following:7

> cargo run -- --program=sour-sweet.chalk --slg --overflow-depth=3
7 answer(s) found:
- ?0 := Vinegar
- ?0 := Lemon
- ?0 := HotSauce<Vinegar>
- ?0 := HotSauce<Lemon>
- exists<U0> { ?0 := HotSauce<HotSauce<?0>> } [ambiguous]
- ?0 := HotSauce<HotSauce<Vinegar>>
- ?0 := HotSauce<HotSauce<Lemon>>

Notice that middle answer:

- exists<U0> { ?0 := HotSauce<HotSauce<?0>> } [ambiguous]

This is precisely the point where the abstraction mechanism kicked in, introducing a variable. Note that the two instances of ?0 here refer to different variables – the first one, in the “key”, refers to the 0th variable in our original query (what I’ve been calling T). The second ?0, in the “value” refers, to the variable introduced by the exists<> quantifier (the U0 is the “universe” of that variable, which has to do with higher-ranked things and I won’t get into here). Finally, you can see that we flagged this result as [ambiguous], because we had to truncate it to make it fit the maximum size.

Truncating answers isn’t on its own enough to guarantee termination. It’s also possible to setup an ever-growing number of queries. For example, one could write something like:

trait Foo { }
impl<T> Foo for T where HotSauce<T>: Foo { }

If we try to solve (say) Lemon: Foo, we will then have to solve HotSauce<Lemon>, and HotSauce<HotSauce<Lemon>>, and so forth ad infinitum. We address this by the same kind of tweak. After a point, if a query grows too large, we can just truncate it into a shorter one8. So e.g. trying to solve

exists<T> HotSauce<HotSauce<HotSauce<HotSauce<T>>>>: Foo

with a maximum size of 3 would wind up “redirecting” to the query

exists<T> HotSauce<HotSauce<HotSauce<T>>>: Foo

Interestingly, unlike the “answer approximation” we did earlier, redirecting queries like this doesn’t produce imprecision (at least not on its own). The new query is a generalization of the old query, and since we generate all answers to any given query, we will find the original answers we were looking for (and then some more). Indeed, if we try to perform this query with the SLG solver, it correctly reports that there exists no answer (because this recursion will never terminate):

> cargo run -- --program=sour-sweet.chalk --slg --overflow-depth=3
?- Lemon: Foo
No answers found.

(The original solver panics with an overflow error.)

Tabling

The key idea of tabling is to keep, for each query that we are trying to solve, a table of answers that we build up over time. Tabling came up in my previous post, too, where I discussed how we used it to handle cyclic queries in the current solver. But the integration into SLG is much deeper.

In SLG, we wind up keeping a table for every subgoal that we encounter. Thus, any time that you have to solve the same subgoal twice in the course of a query, you automatically get to take advantage of the cached answers from the previous attempt. Moreover, to account for cyclic dependencies, tables can be linked together, so that as new answers are found, the suspended queries are re-awoken.

Tables can be in one of two states:

  • Completed: we have already found all the answers for this query.
  • Incomplete: we have not yet found all the answers, but we may have found some of them.

By the time the SLG processing is done, all tables will be in a completed state, and thus they serve purely as caches. These tables can also be remembered for use in future queries. I think integrating this kind of caching into rustc could be a tremendous performance enhancement.

Variant- versus subsumption-based tabling

I implemented “variant-based tabling” – in practical terms, this means that whenever we have some subgoal G that we want to solve, we first convert it into a canonical form. So imagine that we are in some inference context and ?T is a variable in that context, and we want to solve HotSauce<?T>: Sour. We would replace that variable ?T with ?0, since it is the first variable we encountered as we traversed the type, thus giving us a canonical query like:

HotSauce<?0>: Sour

This is then the key that we use to lookup if there exists a table already. If we do find such a table, it will have a bunch of answers; these answers are in the form of substitutions, like

  • ?0 := Lemon
  • ?0 := Vinegar

and so forth. At this point, this should start looking familiar: you may recall that earlier in the post I was showing you the output from the chalk repl, which consisted of stuff like this:

> cargo run -- --program=sour-sweet.chalk --slg
?- exists<T> { T: Sour, T: Sweet }     
1 answer(s) found:
- ?0 := Lemon

This printout is exactly dumping the contents of the table that we constructed for our exists<T> { T: Sour, T: Sweet } query. That query would be canonicalized to ?0: Sour, ?0: Sweet, and hence we have results in terms of this canonical variable ?0.

However, this form of tabling that I just described has its limitations. For example, imagine that I we have the table for exists<T> { T: Sour, T: Sweet } all setup, but then I do a query like Lemon: Sour, Lemon: Sweet. In the solver as I wrote it today, this will create a brand new table and begin computation again. This is somewhat unfortunate, particularly for a setting like rustc, where we often solve queries first in the generic form (during type-checking) and then later, during trans, we solve them again for specific instantiations.

The paper about SLG that I pointed you at earlier describes an alternative approach called “subsumption-based tabling”, in which you can reuse a table’s results even if it is not an exact match for the query you are doing. This extension is not too difficult, and we could consider doing something similar, though we’d have to do some more experiments to decide if it pays off.

(In rustc, for example, subsumption-based tabling might not help us that much; the queries that we perform at trans time are often not the same as the ones we perform during type-checking. At trans time, we are required to “reveal” specialized types and take advantage of other details that type-checking does not do, so the query results are somewhat different.)

Negative reasoning and the well-founded semantics

One last thing that the SLG solver handles quite well is negative reasoning. In coherence – and maybe elsewhere in Rust – we want to be able to support negative queries, such as:

not { exists<T> { Vec<T>: Foo } }

This would assert that there is no type T for which Vec<T>: Foo is implemented. In the SLG solver, this is handled by creating a table for the positive query (Vec<?0>: Foo) and letting that execute. Once it completes, we can check whether the table has any answers or not.

There are some edge cases to be careful of though. If you start to allow negative reasoning to be used more broadly, there are logical pitfalls that start to arise. Consider the following Rust impls, in a system where we supported negative goals:

trait Foo { }
trait Bar { }

impl<T> Foo for T where T: !Bar { }
impl<T> Bar for T where T: !Foo { }

Now consider the question of whether some type T implements Foo and Bar. The trouble with these two impls is that the answers to these two queries (T: Foo, T: Bar) are no longer independent from one another. We could say that T: Foo holds, but then T: Bar does not (because T: !Foo is false). Alternatively, we could say that T: Bar holds, but then T: Foo does not (because T: !Bar is false). How is the compiler to choose?

The SLG solver chooses not to choose. It is based on the well-founded semantics, which ultimately assigns one of three results to every query: true, false, or unknown. In the case of negative cycles like the one above, the answer is “unknown”.

(In contrast, our current solver will answer that both T: Foo and T: Bar are false, which is clearly wrong. I imagine we could fix this – it was an interaction we did not account for in our naive tabling implementation.)

Extensions and future work

The SLG papers themselves describe a fairly basic set of logic programs. These do not include a number of features that we need to model Rust. My current solver already extends the SLG work to cover first-order hereditary harrop clauses (meaning the ability to have queries like forall<T> { if (T: Clone) { ... } }) – this was relatively straight-forward. But I did not yet cover some of the other things that the current solver handles:

  • Coinductive predicates: To handle auto traits, we need to support coinductive predicates like Send. I am not sure yet how to extend SLG to handle this.
  • Fallback clauses: If you normalize something like <Vec<u32> as IntoIterator>::Item, the correct result is u32. The SLG solver gives back two answers, however: u32 or the unnormalized form <Vec<u32> as IntoIterator>::Item. This is not wrong, but the current solver understands that one answer is “better” than the other.
  • Suggested advice: in cases of ambiguity, the current solver knows to privilege where clauses and can give “suggestions” for how to unify variables based on those.

The final two points I think can be done in a fairly trivial fashion, though the full implications of fallback clauses may require some careful thought, but coinductive predicates seem a bit harder and may require some deeper tinkering.

Conclusions

I’m pretty excited about this new SLG-based solver. I think it is a big improvement over the existing solver, though we still have to work out the story for auto traits. The things that excited me the most:

  • The deeply integrated use of tabling offers a very strong caching story.
  • There is a lot of work on efficienctly executing the SLG solving algorithm. The work I did is only the tip of the iceberg: there are existing virtual machine designs and other things that we could adapt if we wanted to.

I am also quite keen on the story around guaranteed termination. I like that it does not involve a concept of overflow – that is, a hard limit on the depth of the query stack – but rather simply a maximum size imposed on types. The problem with overflow is that it means that the results of queries wind up dependent on where they were executed, complicating caching and other things. In other words, a query that may well succeed can wind up failing just because it was executed as part of something else. This does not happen with the SLG-based solver – queries always succeed or fail in the same way.

However, I am also worried – most notably about the fact that the current solver is designed to always enumerate all the answers to a query, even when that is unhelpful. I worry that this may waste a ton of memory in rustc processes, as we are often asked to solve silly queries like ?T: Sized during type-checking, which would basically wind up enumerating nearly all types in the system up to the maximum size[^ms]. Still, I am confident that we can find ways to address this shortcoming in time, possibly without deep changes to the algorithm.

Credit where credit is due

I also want to make sure I thank all the authors of the many papers on SLG whose work I gleefully stole built upon. This is a list of the papers that I papers that described techniques that went into the new solver, in no particular order; I’ve tried to be exhaustive, but if I forgot something, I’m sorry about that.

Footnotes

  1. True confessions: I have never (personally) managed to make a non-trivial Prolog program terminate. I understand it can be done. Just not by me.

  2. Assuming termination. More on that later.

  3. Some might say that lemons are not, in fact, sweet. Well fooey. I’m not rewriting this blog post now, dang it.

  4. Try this stuff, it’s for real.

  5. This technique is called “radial restraint” by its authors.

  6. In terms of the well-formed semantics that we’ll discuss later, its truth value is considered “unknown”.

  7. Actually, in the course of writing this blog post, I found I sometimes only see 5 answers, so YMMV. Some kind of bug I suppose. (Update: fixed it.)

  8. This technique is called “subgoal abstraction” by its authors.