Virtual Structs Part 2: Classes strike back

29 May 2015

This is the second post summarizing my current thoughts about ideas related to “virtual structs”. In the last post, I described how, when coding C++, I find myself missing Rust’s enum type. In this post, I want to turn it around. I’m going to describe why the class model can be great, and something that’s actually kind of missing from Rust. In the next post, I’ll talk about how I think we can get the best of both worlds for Rust. As in the first post, I’m focusing here primarily on the data layout side of the equation; I’ll discuss virtual dispatch afterwards.

(Very) brief recap

In the previous post, I described how one can setup a class hierarchy in C++ (or Java, Scala, etc) with a base class and one subclass for every variant:

class Error { ... };
class FileNotFound : public Error { ... };
class UnexpectedChar : public Error { ... };

This winds up being very similar to a Rust enum:

enum ErrorCode {
    FileNotFound,
    UnexpectedChar
}

However, there are are some important differences. Chief among them is that the Rust enum has a size equal to the size of its largest variant, which means that Rust enums can be passed “by value” rather than using a box. This winds up being absolutely crucial to Rust: it’s what allows us to use Option<&T>, for example, as a zero-cost nullable pointer. It’s what allows us to make arrays of enums (rather than arrays of boxed enums). It’s what allows us to overwrite one enum value with another, e.g. to change from None to Some(_). And so forth.

Problem #1: Memory bloat

There are a lot of use cases, however, where having a size equal to the largest variant is actually a handicap. Consider, for example, the way the rustc compiler represents Rust types (this is actually a cleaned up and simplified version of the real thing).

The type Ty represents a rust type:

// 'tcx is the lifetime of the arena in which we allocate type information
type Ty<'tcx> = &'tcx TypeStructure<'tcx>;

As you can see, it is in fact a reference to a TypeStructure (this is called sty in the Rust compiler, which isn’t completely up to date with modern Rust conventions). The lifetime 'tcx here represents the lifetime of the arena in which we allocate all of our type information. So when you see a type like &'tcx, it represents interned information allocated in an arena. (As an aside, we added the arena back before we even had lifetimes at all, and used to use unsafe pointers here. The fact that we use proper lifetimes here is thanks to the awesome eddyb and his super duper safe-ty branch. What a guy.)

So, here is the first observation: in practice, we are already boxing all the instances of TypeStructure (you may recall that the fact that classes forced us to box was a downside before). We have to, because types are recursively structured. In this case, the ‘box’ is an arena allocation, but still the point remains that we always pass types by reference. And, moreover, once we create a Ty, it is immutable – we never switch a type from one variant to another.

The actual TypeStructure enum is defined something like this:

enum TypeStructure<'tcx> {
    Bool,                                      // bool
    Reference(Region, Mutability, Type<'tcx>), // &'x T, &'x mut T
    Struct(DefId, &'tcx Substs<'tcx>),         // Foo<..>
    Enum(DefId, &'tcx Substs<'tcx>),           // Foo<..>
    BareFn(&'tcx BareFnData<'tcx>),            // fn(..)
    ...
}

You can see that, in addition to the types themselves, we also intern a lot of the data in the variants themselves. For example, the BareFn variant takes a &'tcx BareFnData<'tcx>. The reason we do this is because otherwise the size of the TypeStructure type balloons very quickly. This is because some variants, like BareFn, have a lot of associated data (e.g., the ABI, the types of all the arguments, etc). In contrast, types like structs or references have relatively little associated data. Nonetheless, the size of the TypeStructure type is determined by the largest variant, so it doesn’t matter if all the variants are small but one: the enum is still large. To fix this, Huon spent quite a bit of time analyzing the size of each variant and introducing indirection and interning to bring it down.

Consider what would have happened if we had used classes instead. In that case, the type structure might look like:

typedef TypeStructure *Ty;
class TypeStructure { .. };
class Bool : public TypeStructure { .. };
class Reference : public TypeStructure { .. };
class Struct : public TypeStructure { .. };
class Enum : public TypeStructure { .. };
class BareFn : public TypeStructure { .. };

In this case, whenever we allocated a Reference from the arena, we would allocate precisely the amount of memory that a Reference needs. Similarly, if we allocated a BareFn type, we’d use more memory for that particular instance, but it wouldn’t affect the other kinds of types. Nice.

Problem #2: Common fields

The definition for Ty that I gave in the previous section was actually somewhat simplified compared to what we really do in rustc. The actual definition looks more like:

// 'tcx is the lifetime of the arena in which we allocate type information
type Ty<'tcx> = &'tcx TypeData<'tcx>;

struct TypeData<'tcx> {
    id: u32,
    flags: u32,
    ...,
    structure: TypeStructure<'tcx>,
}

As you can see, Ty is in fact a reference not to a TypeStructure directly but to a struct wrapper, TypeData. This wrapper defines a few fields that are common to all types, such as a unique integer id and a set of flags. We could put those fields into the variants of TypeStructure, but it’d be repetitive, annoying, and inefficient.

Nonetheless, introducing this wrapper struct feels a bit indirect. If we are using classes, it would be natural for these fields to live on the base class:

typedef TypeStructure *Ty;
class TypeStructure {
    unsigned id;
    unsigned flags;
    ...
};
class Bool : public TypeStructure { .. };
class Reference : public TypeStructure { .. };
class Struct : public TypeStructure { .. };
class Enum : public TypeStructure { .. };
class BareFn : public TypeStructure { .. };

In fact, we could go further. There are many variants that share common bits of data. For example, structs and enums are both just a kind of nominal type (“named” type). Almost always, in fact, we wish to treat them the same. So we could refine the hierarchy a bit to reflect this:

class Nominal : public TypeStructure {
    DefId def_id;
    Substs substs;
};
class Struct : public Nominal {
};
class Enum : public Nominal {
};

Now code that wants to work uniformly on either a struct or enum could just take a Nominal*.

Note that while it’s relatively easy in Rust to handle the case where all variants have common fields, it’s a lot more awkward to handle a case like Struct or Enum, where only some of the variants have common fields.

Problem #3: Initialization of common fields

Rust differs from purely OO languages in that it does not have special constructors. An instance of a struct in Rust is constructed by supplying values for all of its fields. One great thing about this approach is that “partially initialized” struct instances are never exposed. However, the Rust approach has a downside, particularly when we consider code where you have lots of variants with common fields: there is no way to write a fn that initializes only the common fields.

C++ and Java take a different approach to initialization based on constructors. The idea of a constructor is that you first allocate the complete structure you are going to create, and then execute a routine which fills in the fields. This approach to constructos has a lot of problems – some of which I’ll detail below – and I would not advocate for adding it to Rust. However, it does make it convenient to separately abstract over the initialization of base class fields from subclass fields:

typedef TypeStructure *Ty;
class TypeStructure {
    unsigned id;
    unsigned flags;
    
    TypeStructure(unsigned id, unsigned flags)
      : id(id), flags(flags)
    { }
};

class Bool : public TypeStructure {
    Bool(unsigned id)
      : TypeStructure(id, 0) // bools have no flags
    { }
};

Here, the constructor for TypeStructure initializes the TypeStructure fields, and the Bool constructor initializes the Bool fields. Imagine we were to add a field to TypeStructure that is always 0, such as some sort of counter. We could do this without changing any of the subclasses:

class TypeStructure {
    unsigned id;
    unsigned flags;
    unsigned counter; // new
    
    TypeStructure(unsigned id, unsigned flags)
      : id(id), flags(flags), counter(0)
    { }
};

If you have a lot of variants, being able to extract the common initialization code into a function of some kind is pretty important.

Now, I promised a critique of constructors, so here we go. The biggest reason we do not have them in Rust is that constructors rely on exposing a partially initialized this pointer. This raises the question of what value the fields of that this pointer have before the constructor finishes: in C++, the answer is just undefined behavior. Java at least guarantees that everything is zeroed. But since Rust lacks the idea of a “universal null” – which is an important safety guarantee! – we don’t have such a convenient option. And there are other weird things to consider: what happens if you call a virtual function during the base type constructor, for example? (The answer here again varies by language.)

So, I don’t want to add OO-style constructors to Rust, but I do want some way to pull out the initialization code for common fields into a subroutine that can be shared and reused. This is tricky.

Problem #4: Refinement types

Related to the last point, Rust currently lacks a way to “refine” the type of an enum to indicate the set of variants that it might be. It would be great to be able to say not just “this is a TypeStructure”, but also things like “this is a TypeStructure that corresponds to some nominal type (i.e., a struct or an enum), though I don’t know precisely which kind”. As you’ve probably surmised, making each variant its own type – as you would in the classes approach – gives you a simple form of refinement types for free.

To see what I mean, consider the class hierarchy we built for TypeStructure:

typedef TypeStructure *Ty;
class TypeStructure { .. };
class Bool : public TypeStructure { .. };
class Reference : public TypeStructure { .. };
class Nominal : public TypeStructure { .. }
class Struct : public Nominal { .. };
class Enum : public Nominal { .. };
class BareFn : public TypeStructure { .. };

Now, I can pass around a TypeStructure* to indicate “any sort of type”, or a Nominal* to indicate “a struct or an enum”, or a BareFn* to mean “a bare fn type”, and so forth.

If we limit ourselves to single inheritance, that means one can construct an arbitrary tree of refinements. Certainly one can imagine wanting arbitrary refinements, though in my own investigations I have always found a tree to be sufficient. In C++ and Scala, of course, one can use multiple inheritance to create arbitrary refinements, and I think one can imagine doing something similar in Rust with traits.

As an aside, the right way to handle ‘datasort refinements’ has been a topic of discussion in Rust for some time; I’ve posted a different proposal in the past, and, somewhat amusingly, my very first post on this blog was on this topic as well. I personally find that building on a variant hierarchy, as above, is a very appealing solution to this problem, because it avoids introducing a “new concept” for refinements: it just leverages the same structure that is giving you common fields and letting you control layout.

Conclusion

So we’ve seen that there also advantages to the approach of using subclasses to model variants. I showed this using the TypeStructure example, but there are lots of cases where this arises. In the compiler alone, I would say that the abstract syntax tree, the borrow checker’s LoanPath, the memory categorization cmt types, and probably a bunch of other cases would benefit from a more class-like approach. Servo developers have long been requesting something more class-like for use in the DOM. I feel quite confident that there are many other crates at large that could similarly benefit.

Interestingly, Rust can gain a lot of the benefits of the subclass approach—namely, common fields and refinement types—just by making enum variants into types. There have been proposals along these lines before, and I think that’s an important ingredient for the final plan.

Perhaps the biggest difference between the two approaches is the size of the “base type”. That is, in Rust’s current enum model, the base type (TypeStructure) is the size of the maximal variant. In the subclass model, the base class has an indeterminate size, and so must be referenced by pointer. Neither of these are an “expressiveness” distinction—we’ve seen that you can model anything in either approach. But it has a big effect on how easy it is to write code.

One interesting question is whether we can concisely state conditions in which one would prefer to have “precise variant sizes” (class-like) vs “largest variant” (enum). I think the “precise sizes” approach is better when the following apply:

  1. A recursive type (like a tree), which tends to force boxing anyhow. Examples: the AST or types in the compiler, DOM in servo, a GUI.
  2. Instances never change what variant they are.
  3. Potentially wide variance in the sizes of the variants.

The fact that this is really a kind of efficiency tuning is an important insight. Hopefully our final design can make it relatively easy to change between the ‘maximal size’ and the ‘unknown size’ variants, since it may not be obvious from the get go which is better.

Preview of the next post

The next post will describe a scheme in which we could wed together enums and structs, gaining the advantages of both. I don’t plan to touch virtual dispatch yet, but intead just keep focusing on concrete types.