Author: eernst@.
Version: 0.8 (2018-10-16).
Status: Background material. The language specification has the normative text on this topic. Note that the rules have changed, which means that this document cannot be used as a reference, it can only be used to get an overview of the ideas; please refer to the language specification for all technical details.
This document is an informal specification of the support in Dart 2 for using certain generic types where the declared bounds are violated. The feature described here, super-bounded types, consists in allowing an actual type argument to be a supertype of the declared bound, as long as a consistent replacement of Object
, dynamic
, and void
by Null
produces a traditional, well-bounded type. For example, if a class C
takes a type argument X
which must extend C<X>
, C<Object>
, C<dynamic>
, and C<void>
are correct super-bounded types. This is useful because there is no other way to specify a type which retains the knowledge that it is a C
based type, and at the same time it is a supertype of C<T>
for all the T
that satisfy the specified bound. In other words, it allows developers to specify that they want to work with a C<T>
, except that they don't care which T
, and every such T
must be allowed.
This permission to use a super-bounded type is only granted in some situations. For instance, super-bounded types are allowed as type annotations, but they are not allowed in instance creation expressions like new C<Object>()
(assuming that Object
violates the bound of C
). Similarly, a function declared as
void foo<X extends List<num>>(X x) { ... }
cannot be invoked with foo<List<dynamic>>([])
, nor can the type argument be inferred to List<dynamic>
in an invocation like foo([])
. But C<void> x = new C<int>();
is OK, and so is x is C<Object>
.
Many well-known classes have a characteristic typing structure:
abstract class num implements Comparable<num> {...} class Duration implements Comparable<Duration> {...} class DateTime implements Comparable<DateTime> {...} ...
The class Comparable<T>
has a method int compareTo(T other)
, which makes it possible to do things like this:
int comparison = a.compareTo(b);
This works fine when a
and b
both have type num
, or both have type Duration
, but it is not so easy to describe the situation where the comparable type can vary. For instance, consider the following:
class ComparablePair<X extends Comparable<X>> { X a, b; } main() { ComparablePair<MysteryType> myPair = ... int comparison = myPair.a.compareTo(myPair.b); }
We could replace MysteryType
by num
and then work on pairs of num
only. But can we find a type to replace MysteryType
such that myPair
can hold an instance of ComparablePair<T>
, no matter which T
it uses?
We would need a supertype of all those T
where T extends Comparable<T>
; but we cannot use the obvious ones like Object
or dynamic
, because they do not satisfy the declared bound for the type argument to ComparablePair
. There is in fact no such type in the Dart type system!
This is an issue that comes up in various forms whenever a type parameter bound uses the corresponding type variable itself (or multiple type parameters mutually depend on each other), that is, whenever we have one or more F-bounded type parameters. Here is an example which is concise and contains the core of the issue:
class C<X extends C<X>> { X next; }
For each given type T
it is possible to determine whether T
is a subtype of C<T>
, in which case that T
would be an admissible actual type argument for C
. This means that the set of possible values for X
is a well-defined set.
However, there is no type S
such that the set of possible values for X
is equal to the set of subtypes of S
; that is, the set of types we seek to express is not the set of subtypes of anything. Sure, those types are a subset of all subtypes of Object
, but we need to express that exact set of types, not a superset.
Hence, we cannot correctly characterize “all possible values for X
” as a single type argument. This means that we cannot find a U
such that C<U>
is the least upper bound of all possible types on the form C<T>
.
But that's exactly what we must find, if we are to safely express the greatest possible amount of information about the set of objects whose type is on the form C<T>
for some T
. In particular, we cannot express the type which “should be” the result of instantiate-to-bound on the raw type C
.
We can make an attempt to approximate the least supertype of all correct generic instantiations of C
(that is, a supertype of all types on the form C<T>
). Assume that T
is an admissible actual type argument for C
(that is, we can rely on T extends C<T>
):
// Because `T extends C<T>`, and due to generic covariance: C<T> <: C<C<T>> // Same facts used on the nested type argument `T`: C<C<T>> <: C<C<C<T>>> ... // Same at the next level ... C<C<C<T>>> <: C<C<C<C<T>>>> ...
We can continue ad infinitum, and this means that a good candidate for the “least upper bound of all C<T>
” would be the infinite type W
where W = C<W>
. Basically, W = C<C<C<C<C<C<...>>>>>>
, nesting to an infinite depth.
Note that T
“disappears” when we extend the nesting ad infinitum, which means that W
is the result we find for every T
. Conversely, we cannot hope to find a different type V
(not equal to C<R>
for any R
) such that V
is both a supertype of all types on the form C<T>
for some T
and V
is a proper subtype of W
. In other words, if the “least upper bound of all C<T>
” exists, it must be W
.
However, we do not wish to introduce these infinite types into the Dart type universe. The ability to express types on this form will inevitably introduce the ability to express many new kinds of types, and we do not expect this generalization to improve the expressive power of the language in a manner that compensates sufficiently for the burden of managing the added complexity.
Instead, we give developers the responsibility to make the choice explicitly: They can use super-bounded types to express a range of supertypes of these infinite types (as well as other types, if they wish). When they do that with an infinite type, they can make the choice to unfold it exactly as many times as they want. At the same time, they will be forced to maintain a greater level of awareness of the nature of these types than they would, had we chosen to model infinite types, e.g., by unfolding them to some specific, finite level.
Here are some examples of finite unfoldings, and the effect they have on types of expressions:
class C<X extends C<X>> { X next; C(this.next); } class D extends C<D> { D(D next): super(next); } main() { D d = new D(new D(null)); C<dynamic> c0 = d; C<C<dynamic>> c1 = d; C<C<C<dynamic>>> c2 = d; c0.next.unknown(42); // Statically OK, `c0.next` is `dynamic`. c1.next.unknown(43); // Compile-time error. c1.next.next.unknown(44); // Statically OK. c2.next.next.unknown(45); // Compile-time error. c2.next.next.next.unknown(46); // Statically OK. // With type `D`, the static analysis is aware of the cyclic // structure of the type, and every level of nesting is handled // safely. But `D` may be less useful because there may be a // similar type `D2`, and this code will only work with `D`. d.next.next.next.next.next.next.next.unknown(46); // Compile-time error. }
We can make a choice of how to deal with the missing type information. When we use C<dynamic>
, C<C<dynamic>>
and C<C<C<dynamic>>>
we will implicitly switch to dynamic member access after a few steps of navigation.
If we choose to use C<Object>
, C<C<Object>>
and so on then we will have to use explicit downcasts in order to access all non-Object
members. We will still be able to pass c0.next
as an argument to a function expecting a C<S>
(where S
can be anything), but we could also pass it where a String
is expected, etc.
Finally, if we choose to use C<void>
and so on then we will not even be able to access the object where the type information ends: we cannot use the value of an expression like c0.next
without an explicit cast (OK, void v = c0.next;
is accepted, but it is mostly impossible to use the value of an expression of type void
). This means that we cannot pass c0.next
as an argument to a function that accepts a C<S>
(for any S
) without an explicit cast.
In summary, the choice of dynamic
, Object
, and void
offers a range of approaches to the lack of typing information, but the amount of information remains the same.
This feature does not require any modifications to the Dart grammar.
We say that the parameterized type G<T1..Tk> is regular-bounded when Tj <: [T1/X1 .. Tk/Xk]Bj for all j, 1 <= j <= k, where X1..Xk are the formal type parameters of G in declaration order, and B1..Bk are the corresponding upper bounds.
This means that each actual type argument satisfies the declared upper bound for the corresponding formal type parameter.
We extend covariance for generic class types such that it can be used also in cases where a type argument violates the corresponding bound.
For instance, assuming the classes C
and D
as declared in the Motivation section, C<D>
is a subtype of C<Object>
. This is new because C<Object>
used to be a compile-time error, which means that no questions could be asked about its properties. Note that this is a straightforward application of the usual covariance rule: C<D> <: C<Object>
because D <: Object
. We need this relaxation of the rules in order to be able to define which violations of the declared bounds are admissible.
Let G denote a generic class, X1..Xk the formal type parameters of G in declaration order, and B1..Bk the types in the corresponding upper bounds, using Object
when the upper bound is omitted. The parameterized type G<T1..Tk> is then a super-bounded type iff the following two requirements are satisfied:
There is a j, 1 <= j <= k, such that Tj is not a subtype of [T1/X1..Tk/Xk]Bj.
Let Sj, 1 <= j <= k, be the result of replacing every covariant occurrence of Object
, dynamic
, and void
in Tj by Null
, and every contravariant occurrence of Null
by Object
. It is then required that Sj <: [S1/X1..Sk/Xk]Bj for all j, 1 <= j <= k.
In short, at least one type argument violates its bound, and the type is regular-bounded after replacing all occurrences of an extreme type by the opposite extreme type, according to their variance.
For instance, assuming the declarations of C
and D
as in the Motivation section, C<Object>
is a super-bounded type, because Object
violates the declared bound and C<Null>
is regular-bounded.
Here is an example that involves contravariance:
class E<X extends void Function(X)> {}
With this declaration, E<void Function(Null)>
is a super-bounded type because E<void Function(Object)>
is a regular-bounded type. Note that the contravariance can also be eliminated, yielding a simpler super-bounded type: E<dynamic>
is a super-bounded type because E<Null>
is a regular-bounded type.
We say that a parameterized type T is well-bounded if it is regular-bounded or super-bounded.
Note that it is possible for a super-bounded type to be nested in another type which is super-bounded, and it can also be nested in another type which is not super-bounded. For example, assuming C
as in the Motivation section, C<C<Object>>
is a super-bounded type which contains a super-bounded type; in contrast, List<C<Object>>
is a regular type (a generic instantiation of List
) which contains a super-bounded type (C<Object>
).
It is a compile-time error if a parameterized type is not well-bounded.
That is, a parameterized type is regular-bounded, or it is super-bounded, or it is an error. This rule replaces and relaxes the rule in the language specification that constrains parameterized types to be regular-bounded.
It is a compile-time error if a type used as the type in an instance creation expression (that is, the T
in expressions of the form new T(...)
, new T.id(...)
, const T(...)
, or const T.id(...)
) is super-bounded. It is a compile-time error if the type in a redirection of a redirecting factory constructor (that is, the T
in a phrase of the form T
or T.id
after =
in the constructor declaration) is super-bounded. It is a compile-time error if a super-bounded type is specified as a superinterface for a class. (This implies that a super-bounded type cannot appear in an extends
, implements
, or with
clause, or in a mixin application; e.g., T
in class C = T with M;
cannot be super-bounded). Finally, it is a compile-time error if a bound in a formal type parameter declaration is super-bounded.
This means that we allow super-bounded types as function return types, as type annotations on variables (all of them: library, static, instance, and local variables, and formal parameters of functions), in type tests (e is T
), in type casts (e as T
), in on
clauses, and as type arguments.
Let F denote a parameterized type alias, X1..Xk the formal type parameters of F in declaration order, and B1..Bk the types in the corresponding upper bounds, using Object
when the upper bound is omitted. The parameterized type F<T1..Tk> is then a super-bounded type iff the following three requirements are satisfied:
There is a j, 1 <= j <= k, such that Tj is not a subtype of [T1/X1..Tk/Xk]Bj.
Let Sj, 1 <= j <= k, be the result of replacing every covariant occurrence of Object
, dynamic
, and void
in Tj by Null
, and every contravariant occurrence of Null
by dynamic
. It is then required that Sj <: [S1/X1..Sk/Xk]Bj for all j, 1 <= j <= k.
Let T be the right hand side of the declaration of F, then [T1/X1..Tk/Xk]T is a well-bounded type.
In short, a parameterized type based on a type alias, F<...>
, must pass the super-boundedness checks in itself, and so must the body of F
.
For instance, assume that F
and G
are declared as follows:
class A<X extends C<X>> { ... } typedef F<X extends C<X>> = A<X> Function(); typedef G<X extends C<X>> = void Function(A<X>);
The type F<Object>
is then a super-bounded type, because F<Null>
is regular-bounded (Null
is a subtype of C<Null>
) and because A<Object> Function()
is well-bounded, because A<Object>
is super-bounded. Similarly, G<Object>
is a super-bounded type because void Function(A<Object>)
is well-bounded because A<Object>
is super-bounded.
Note that it is necessary to require that the right hand side of a type alias declaration is taken into account when determining that a given application of a type alias to an actual type argument list is correctly super-bounded. That is, we do not think that it is possible for a (reasonable) constraint specification mechanism on the formal type parameters of a type alias declaration to ensure that all arguments satisfying those constraints will also be suitable for the type on the right hand side. In particular, we may use simple upper bounds and F-bounded constraints (as we have always done), perform and pass the ‘correctly super-bounded’ check on a given parameterized type based on a type alias, and still have a right hand side which is not well-bounded:
class B<X extends List<num>> {} typedef H<Y extends num> = void Function(B<List<Y>>); typedef K<Y extends num> = B<List<Y>> Function(B<List<Y>>); H<Object> myH = null; // Error!
H<Object>
is a compile-time error because it is not regular-bounded (Object <: num
does not hold), and it is also not correctly super-bounded: Null
does satisfy the constraint in the declaration of Y
, but H<Object>
gives rise to the right hand side void Function(B<List<Object>>)
, and that is not a well-bounded type: It is not regular-bounded (List<Object> <: List<num>
does not hold), and it does not become a regular-bounded type by the type replacements (that yield void Function(B<List<Object>>)
because that occurrence of Object
is contravariant).
Semantically, this failure may be motivated by the fact that H<Object>
, were it allowed, would not be a supertype of H<T>
for all the T
where H<T>
is regular-bounded. So it would not be capable of playing the role as a “default type” that abstracts over all the possible actual types that are expressible using H
. For example, a variable declared like List<H<Object>> x;
would not be allowed to hold a value of type List<H<num>>
because the latter is not a subtype of the former.
In the given situation it is possible to express such a default type: H<Null>
is actually a common supertype of H<T>
for all T
such that H<T>
is regular-bounded. However, K
shows that this is not always the case: There is no type S
such that K<S>
is a common supertype of K<T>
for all those T
where K<T>
is regular-bounded. Facing this situation, we prefer to bail out rather than implicitly allow some kind of super-bounded type (assuming that we amend the rules such that it is not an error) which would not abstract over all possible instantiations anyway.
The subtype relations for super-bounded types follow directly from the extension of generic covariance to include actual type arguments that violate the declared bounds. For the example in the Motivation section, D
is a subtype of C<D>
which is a subtype of C<C<D>>
, which is a subtype of C<C<C<D>>>
, continuing with C<C<C<Object>>>>
, C<C<Object>>
, C<Object>
, and Object
, respectively, and similarly for dynamic
and void
.
Types of members from super-bounded class types are computed using the same rules as types of members from other types. Types of function applications involving super-bounded types are computed using the same rules as types of function applications involving other types.
For instance, using the example class C
again, if c1
has static type C<C<dynamic>>
then c1.next
has static type C<dynamic>
and c1.next.next
has static type dynamic
. Similarly, if List<X> foo(X)
were the signature of a method in C
, c1.foo
would have static type List<C<dynamic>> Function(C<dynamic>)
. Note that the argument type X
makes that parameter of foo
covariant, which implies that the reified type of the tear-off c1.foo
would have argument type Object
, which ensures that the expression c1.foo
evaluates to a value whose dynamic type is a subtype of the static type, as it should.
Similarly, if we invoke an instance method with statically known argument type C<void>
whose argument is covariant, there will be a dynamic type check on the actual argument (which might require that it is, say, of type D
); that check may fail at run time, but this is no different from the situation with types that are not super-bounded. In general, the introduction of super-bounded types does not introduce new soundness considerations around covariance.
Super-bounded function types do not have to be only in the statically known types of first class functions, they can also be part of the actual type of a function at run time. For instance, a function may be declared as follows:
List<C<dynamic>> foo(C<dynamic> x) { ... }
It would then have type exactly List<C<dynamic>> Function(C<dynamic>)
, and this means that it will accept an object which is an instance of a subtype of C<T>
for any T
, and it will return a list whose element type is some subtype of C<dynamic>
, which could be D
or C<C<D>>
at run time.
The reification of a super-bounded type (e.g., as a parameter type in a reified function type) uses the types as specified.
For instance void foo(C<Object> x) => print(x);
will have reified type void Function(C<Object>)
. It is allowed for a run-time entity to have a type which contains a super-bounded type, it is only prohibited for run-time entities to have a super-bounded type themselves. So there can be an instance whose dynamic type is List<C<Object>>
but no instance whose dynamic type is C<Object>
.
The subtype rules used for run-time type tests, casts, and generated type checks are the same as the subtype rules used during static analysis.
If an implementation applies an optimization that is only valid when super-bounded types cannot exist, or in other ways relies on the (no longer valid) assumption that super-bounded types cannot exist, it will need to stop using that optimization or making that assumption. We do not expect this to be a common situation, nor do we expect significant losses in performance due to the introduction of this feature.
The super-bounded type feature is all about violating bounds, in a controlled manner. But what is the motivation for enforcing bounds in the first place? The answer to that question serves to justify why it must be ‘controlled’. We have at least two reasons, one internal and one external.
The internal reason is that the bound of each formal type parameter is relied upon during type checking of the body of the corresponding generic declaration. For instance:
class C<X extends num> { X x; bool get foo => x.isNegative; // Statically safe invocation. }
If we ever allow an instance of C<Object>
to be created, or even an instance of a subclass which has C<Object>
as a (possibly indirect) superclass, then we could end up executing that implementation of foo
in a situation where x
does not have an isNegative
getter. In other words, the internal issue is that super-bounding may induce a plain soundness violation in the scope of the type parameter.
This motivates the ban on super-bounding in instance creation expressions, e.g., the ban on new C<Object>()
.
However, it does not suffice to justify banning super-bounded implements
clauses: There will not be any inherited method implementations from a type that a given class implements, and hence no code will ever be executed in the above situation (where a formal type parameter is in scope, and its actual value violates the bound). In fact, code which could be executed in this context would have static knowledge of the super-bound, and hence there is no soundness issue in the body of such a class, nor in its subclasses or subtypes.
// A thought experiment (explaining why this is a compile-time error). class D implements C<Object> { Object x; bool get foo => false; }
It is reasonable to expect a C<Object>
to have a field x
of type Object
and a foo
getter of type bool
, and we can easily implement that. There is no soundness issue, because no code is inherited from C
.
But there is also an external reason: It is reasonable to expect that every instance will satisfy declared bounds, e.g., whenever an object is accessed under the type C<T>
for any T
, it should be true that T
is a subtype of num
. This is not a soundness issue per se; the class D
is perfectly consistent in its behavior with a typing as C<Object>
, and its implementation is type safe.
However, it seems reasonable for developers to reckon as follows: When an object o has a static type like C<Object>
it must satisfy the expectations associated with C
. So there exists an actual type argument T
which satisfies the declared bound, and o must then behave like an instance of C<T>
. In the example, with the given bound num
and using covariance, o would then be guaranteed to be typable as a C<num>
. So the following contains downcasts, but it is “reasonable” to expect them to be guaranteed to succeed at run time:
C<Object> cObject = ...; // Complex creation. C<num> cNum = cObject; // Safe, right? bool b = (cObject.x as num).isNegative; // Also safe, right?
If D
is allowed to exist then we can have a consistent language, and the above would be OK, but the “safe” downcasts would in fact fail at run time. The point is that when we know something is a C<Object>
then we know that it satisfies the constraints of C<Object>
, and we can't assume that it satisfies any stronger constraints (such as those of C<num>
).
This is not a soundness issue in the traditional sense, but it is an issue about how important it is to allow developers to make that extra assumption that all implementations of a given generic class G must be just as picky about their actual type arguments as G itself.
We think that it is indeed justified to make these extra assumptions, and hence we have banned super-bounded implements
clauses.
The extra assumptions which are now supported could be stated as: We can rely on the declared bounds on type parameters of generic classes and functions, also for code which is outside the scope of those type parameters.
In short, the underlying principle is that “there cannot be an instance of a generic class (including instances of subtypes), nor an invocation of a generic function, whose actual type arguments violate the declared bounds”.
Super-bounded function types are possible and useful. Consider the following example:
// If bound on `X` holds then `C<X>` is regular-bounded. typedef F<X extends C<X>> = C<X> Function(); main() { F<C<dynamic>> f = ...; // OK, checking `F<C<Null>>` and `C<dynamic> Function()`. var c0 = f(); // `c0` has type `C<C<dynamic>>`. var c1 = c0.next; // `c1` has type `C<dynamic>` var c2 = c1.next; // `c2` has type `dynamic` ... }
In this example, an unfolding of C
to a specific level is supported in a function type, and application of such a function immediately brings out class types like C<C<dynamic>>
that we have already argued are useful.
Version 0.8 (2018-10-16), emphasized that this document is no longer specifying the current rules, it is for background info only.
Version 0.7 (2018-06-01), marked as background material: The normative text on variance and on super-bounded types is now part of the language specification.
Version 0.6 (2018-05-25), added example showing why we must check the right hand side of type aliases.
Version 0.5 (2018-01-11), generalized to allow replacement of top types covariantly and bottom types contravariantly. Introduced checks on parameterized type aliases (such that bounds declared for the type alias itself are taken into account).
Version 0.4 (2017-12-14), clarified several points and corrected locations where super-bounded types were prohibited, but we should just say that the bounds must be satisfied.
Version 0.3 (2017-11-07), dropping super
, instead allowing Object
, dynamic
, or void
for super-bounded types, with a similar treatment as super
used to get.
Version 0.2 (2017-10-31), introduced keyword super
as a type argument.
Version 0.1 (2017-10-20), initial version of this informal specification.