Variance of type and lifetime parameters

For a more general background on variance, see the background appendix.

During type checking we must infer the variance of type and lifetime parameters. The algorithm is taken from Section 4 of the paper "Taming the Wildcards: Combining Definition- and Use-Site Variance" published in PLDI'11 and written by Altidor et al., and hereafter referred to as The Paper.

This inference is explicitly designed not to consider the uses of types within code. To determine the variance of type parameters defined on type X, we only consider the definition of the type X and the definitions of any types it references.

We only infer variance for type parameters found on data types like structs and enums. In these cases, there is a fairly straightforward explanation for what variance means. The variance of the type or lifetime parameters defines whether T<A> is a subtype of T<B> (resp. T<'a> and T<'b>) based on the relationship of A and B (resp. 'a and 'b).

We do not infer variance for type parameters found on traits, functions, or impls. Variance on trait parameters can indeed make sense (and we used to compute it) but it is actually rather subtle in meaning and not that useful in practice, so we removed it. See the addendum for some details. Variances on function/impl parameters, on the other hand, doesn't make sense because these parameters are instantiated and then forgotten, they don't persist in types or compiled byproducts.

Notation

We use the notation of The Paper throughout this chapter:

  • + is covariance.
  • - is contravariance.
  • * is bivariance.
  • o is invariance.

The algorithm

The basic idea is quite straightforward. We iterate over the types defined and, for each use of a type parameter X, accumulate a constraint indicating that the variance of X must be valid for the variance of that use site. We then iteratively refine the variance of X until all constraints are met. There is always a solution, because at the limit we can declare all type parameters to be invariant and all constraints will be satisfied.

As a simple example, consider:

enum Option<A> { Some(A), None }
enum OptionalFn<B> { Some(|B|), None }
enum OptionalMap<C> { Some(|C| -> C), None }

Here, we will generate the constraints:

1. V(A) <= +
2. V(B) <= -
3. V(C) <= +
4. V(C) <= -

These indicate that (1) the variance of A must be at most covariant; (2) the variance of B must be at most contravariant; and (3, 4) the variance of C must be at most covariant and contravariant. All of these results are based on a variance lattice defined as follows:

   *      Top (bivariant)
-     +
   o      Bottom (invariant)

Based on this lattice, the solution V(A)=+, V(B)=-, V(C)=o is the optimal solution. Note that there is always a naive solution which just declares all variables to be invariant.

You may be wondering why fixed-point iteration is required. The reason is that the variance of a use site may itself be a function of the variance of other type parameters. In full generality, our constraints take the form:

V(X) <= Term
Term := + | - | * | o | V(X) | Term x Term

Here the notation V(X) indicates the variance of a type/region parameter X with respect to its defining class. Term x Term represents the "variance transform" as defined in the paper:

If the variance of a type variable X in type expression E is V2 and the definition-site variance of the corresponding type parameter of a class C is V1, then the variance of X in the type expression C<E> is V3 = V1.xform(V2).

Constraints

If I have a struct or enum with where clauses:

struct Foo<T: Bar> { ... }

you might wonder whether the variance of T with respect to Bar affects the variance T with respect to Foo. I claim no. The reason: assume that T is invariant with respect to Bar but covariant with respect to Foo. And then we have a Foo<X> that is upcast to Foo<Y>, where X <: Y. However, while X : Bar, Y : Bar does not hold. In that case, the upcast will be illegal, but not because of a variance failure, but rather because the target type Foo<Y> is itself just not well-formed. Basically we get to assume well-formedness of all types involved before considering variance.

Dependency graph management

Because variance is a whole-crate inference, its dependency graph can become quite muddled if we are not careful. To resolve this, we refactor into two queries:

  • crate_variances computes the variance for all items in the current crate.
  • variances_of accesses the variance for an individual reading; it works by requesting crate_variances and extracting the relevant data.

If you limit yourself to reading variances_of, your code will only depend then on the inference of that particular item.

Ultimately, this setup relies on the red-green algorithm. In particular, every variance query effectively depends on all type definitions in the entire crate (through crate_variances), but since most changes will not result in a change to the actual results from variance inference, the variances_of query will wind up being considered green after it is re-evaluated.

Addendum: Variance on traits

As mentioned above, we used to permit variance on traits. This was computed based on the appearance of trait type parameters in method signatures and was used to represent the compatibility of vtables in trait objects (and also "virtual" vtables or dictionary in trait bounds). One complication was that variance for associated types is less obvious, since they can be projected out and put to myriad uses, so it's not clear when it is safe to allow X<A>::Bar to vary (or indeed just what that means). Moreover (as covered below) all inputs on any trait with an associated type had to be invariant, limiting the applicability. Finally, the annotations (MarkerTrait, PhantomFn) needed to ensure that all trait type parameters had a variance were confusing and annoying for little benefit.

Just for historical reference, I am going to preserve some text indicating how one could interpret variance and trait matching.

Variance and object types

Just as with structs and enums, we can decide the subtyping relationship between two object types &Trait<A> and &Trait<B> based on the relationship of A and B. Note that for object types we ignore the Self type parameter – it is unknown, and the nature of dynamic dispatch ensures that we will always call a function that is expected the appropriate Self type. However, we must be careful with the other type parameters, or else we could end up calling a function that is expecting one type but provided another.

To see what I mean, consider a trait like so:

#![allow(unused)]
fn main() {
trait ConvertTo<A> {
    fn convertTo(&self) -> A;
}
}

Intuitively, If we had one object O=&ConvertTo<Object> and another S=&ConvertTo<String>, then S <: O because String <: Object (presuming Java-like "string" and "object" types, my go to examples for subtyping). The actual algorithm would be to compare the (explicit) type parameters pairwise respecting their variance: here, the type parameter A is covariant (it appears only in a return position), and hence we require that String <: Object.

You'll note though that we did not consider the binding for the (implicit) Self type parameter: in fact, it is unknown, so that's good. The reason we can ignore that parameter is precisely because we don't need to know its value until a call occurs, and at that time (as you said) the dynamic nature of virtual dispatch means the code we run will be correct for whatever value Self happens to be bound to for the particular object whose method we called. Self is thus different from A, because the caller requires that A be known in order to know the return type of the method convertTo(). (As an aside, we have rules preventing methods where Self appears outside of the receiver position from being called via an object.)

Trait variance and vtable resolution

But traits aren't only used with objects. They're also used when deciding whether a given impl satisfies a given trait bound. To set the scene here, imagine I had a function:

fn convertAll<A,T:ConvertTo<A>>(v: &[T]) { ... }

Now imagine that I have an implementation of ConvertTo for Object:

impl ConvertTo<i32> for Object { ... }

And I want to call convertAll on an array of strings. Suppose further that for whatever reason I specifically supply the value of String for the type parameter T:

let mut vector = vec!["string", ...];
convertAll::<i32, String>(vector);

Is this legal? To put another way, can we apply the impl for Object to the type String? The answer is yes, but to see why we have to expand out what will happen:

  • convertAll will create a pointer to one of the entries in the vector, which will have type &String

  • It will then call the impl of convertTo() that is intended for use with objects. This has the type fn(self: &Object) -> i32.

    It is OK to provide a value for self of type &String because &String <: &Object.

OK, so intuitively we want this to be legal, so let's bring this back to variance and see whether we are computing the correct result. We must first figure out how to phrase the question "is an impl for Object,i32 usable where an impl for String,i32 is expected?"

Maybe it's helpful to think of a dictionary-passing implementation of type classes. In that case, convertAll() takes an implicit parameter representing the impl. In short, we have an impl of type:

V_O = ConvertTo<i32> for Object

and the function prototype expects an impl of type:

V_S = ConvertTo<i32> for String

As with any argument, this is legal if the type of the value given (V_O) is a subtype of the type expected (V_S). So is V_O <: V_S? The answer will depend on the variance of the various parameters. In this case, because the Self parameter is contravariant and A is covariant, it means that:

V_O <: V_S iff
    i32 <: i32
    String <: Object

These conditions are satisfied and so we are happy.

Variance and associated types

Traits with associated types – or at minimum projection expressions – must be invariant with respect to all of their inputs. To see why this makes sense, consider what subtyping for a trait reference means:

<T as Trait> <: <U as Trait>

means that if I know that T as Trait, I also know that U as Trait. Moreover, if you think of it as dictionary passing style, it means that a dictionary for <T as Trait> is safe to use where a dictionary for <U as Trait> is expected.

The problem is that when you can project types out from <T as Trait>, the relationship to types projected out of <U as Trait> is completely unknown unless T==U (see #21726 for more details). Making Trait invariant ensures that this is true.

Another related reason is that if we didn't make traits with associated types invariant, then projection is no longer a function with a single result. Consider:

trait Identity { type Out; fn foo(&self); }
impl<T> Identity for T { type Out = T; ... }

Now if I have <&'static () as Identity>::Out, this can be validly derived as &'a () for any 'a:

<&'a () as Identity> <: <&'static () as Identity>
if &'static () < : &'a ()   -- Identity is contravariant in Self
if 'static : 'a             -- Subtyping rules for relations

This change otoh means that <'static () as Identity>::Out is always &'static () (which might then be upcast to 'a (), separately). This was helpful in solving #21750.