Feature Specification: Interface Conflict Management

Owner: eernst@

Status: Background material, normative text is now in dartLangSpec.tex. 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.

Version: 0.3 (2018-04-24)

This document is a Dart 2 feature specification which specifies how to handle conflicts among certain program elements associated with the interface of a class. In particular, it specifies that multiple occurrences of the same generic class in the superinterface hierarchy must receive the same type arguments, and that no attempts are made at synthesizing a suitable method signature if multiple distinct signatures are provided by the superinterfaces, and none of them resolves the conflict.

Motivation

In Dart 1, the management of conflicts during the computation of the interface of a class is rather forgiving. On page 42 of ECMA-408, we have the following:

However, if the above rules would cause multiple members m1, ..., mk with the same name n to be inherited (because identically named members existed in several superinterfaces) then at most one member is inherited.

...

Then I has a method named n, with r required parameters of type dynamic, h positional parameters of type dynamic, named parameters s of type dynamic and return type dynamic.

In particular, the resulting class interface may then contain a method signature which has been synthesized during static analysis, and which differs from all declarations of the given method in the source code. In the case where some superintenfaces specify some optional positional parameters and others specify some named parameters, any attempt to implement the synthesized method signature other than via a user-defined noSuchMethod would fail (it would be a syntax error to declare both kinds of parameters in the same method declaration).

For Dart 2 we modify this approach such that more emphasis is given to predictability, and less emphasis is given to convenience: No class interface will ever contain a method signature which has been synthesized during static analysis, it will always be one of the method interfaces that occur in the source code. In case of a conflict, the developer must explicitly specify how to resolve the conflict.

To reinforce the same emphasis on predictability, we also specify that it is a compile-time error for a class to have two superinterfaces which are instantiations of the same generic class with different type arguments.

Syntax

The grammar remains unchanged.

Static Analysis

We introduce a new relation among types, more interface-specific than, which is similar to the subtype relation, but which treats top types differently.

  • The built-in class Object is more interface-specific than void.
  • The built-in type dynamic is more interface-specific than void.
  • None of Object and dynamic is more interface-specific than the other.
  • All other subtype rules are also valid rules about being more interface-specific.

This means that we will express the complete rules for being ‘more interface-specific than’ as a slight modification of subtyping.md and in particular, the rule ‘Right Top’ will need to be split in cases such that Object and dynamic are more interface-specific than void and mutually unrelated, and all other types are more interface-specific than both Object and dynamic.

For example, List<Object> is more interface-specific than List<void> and incomparable to List<dynamic>; similarly, int Function(void) is more interface-specific than void Function(Object), but the latter is incomparable to void Function(dynamic).

It is a compile-time error if a class C has two superinterfaces of the form D<T1 .. Tk> respectively D<S1 .. Sk> such that there is a j in 1 .. k where Tj and Sj denote types that are not mutually more interface-specific than each other.

This means that the (direct and indirect) superinterfaces must agree on the type arguments passed to any given generic class. Note that the case where the number of type arguments differ is unimportant because at least one of them is already a compile-time error for other reasons. Also note that it is not sufficient that the type arguments to a given superinterface are mutual subtypes (say, if C implements both I<dynamic> and I<Object>), because that gives rise to ambiguities which are considered to be compile-time errors if they had been created in a different way.

This compile-time error also arises if the type arguments are not given explicitly.

They might be obtained via instantiate-to-bound or, in case such a mechanism is introduced, they might be inferred.

The language specification already contains verbiage to this effect, but we mention it here for two reasons: First, it is a recent change which has been discussed in the language team together with the rest of the topics in this document because of their similar nature and motivation. Second, we note that this restriction may be lifted in the future. It was a change in the specification which did not break many existing programs because dart2js always enforced that restriction (even though it was not specified in the language specification), so in that sense it just made the actual situation explicit. However, it may be possible to lift the restriction: Given that an instance of a class that has List<int> among its superinterfaces can be accessed via a variable of type List<num>, it seems unlikely that it would violate any language invariants to allow the class of that instance to have both List<int> and List<num> among its superinterfaces. We may then relax the rule to specify that for each generic class G which occurs among superinterfaces, there must be a unique superinterface which is the most specific instantiation of G.

During computation of the interface of a class C, it may be the case that multiple direct superinterfaces have a declaration of a member of the same name n, and class C does not declare member named n. Let D1 .. Dn denote this set of declarations.

It is a compile-time error if some declarations among D1 .. Dn are getters and others are non-getters.

Otherwise, if all of D1 .. Dn are getter declarations, the interface of C inherits one, Dj, whose return type is more interface-specific than that of every declaration in D1 .. Dn. It is a compile-time error if no such Dj exists.

For example, it is an error to have two declarations with the signatures Object get foo and dynamic get foo, and no others, because none of these is more interface-specific than the other. This example illustrates why it is unsatisfactory to rely on subtyping alone: If we had accepted this kind of ambiguity then it would be difficult to justify the treatment of o.foo.bar during static analysis where o has type C: If it is considered to be a compile-time error then dynamic get foo is being ignored, and if it is not an error then Object get foo is being ignored, and each of these behaviors may be surprising and/or error-prone. Hence, we require such a conflict to be resolved explicitly, which may be done by writing a signature in the class which overrides both method signatures from the superinterfaces and explicitly chooses Object or dynamic.

Otherwise, (when all declarations are non-getter declarations), the interface of C inherits one, Dj, where its function type is more interface-specific than that of all declarations in D1 .. Dn. It is a compile-time error if no such declaration Dj exists.

In the case where more than one such declaration exists, it is known that their parameter list shapes are identical, and their return types and parameter types are pairwise mutually more interface-specific than each other (i.e., for any two such declarations Di and Dj, if Ui is the return type from Di and Uj is the return type from Dj then Ui is more interface-specific than Uj and vice versa, and similarly for each parameter type). This still allows for some differences. We ignore differences in metadata on formal parameters (we do not consider method signatures in interfaces to have metadata). But we need to consider one more thing:

In this decision about which declaration among D1 .. Dn the interface of the class C will inherit, if we have multiple possible choices, let Di and Dj be such a pair of possible choices. It is a compile-time error if Di and Dj declare two optional formal parameters p1 and p2 such that they correspond to each other (same name if named, or else same position) and they specify different default values.

Discussion

Conflicts among distinct top types may be considered to be spurious in the case where said type occurs in a contravariant position in the method signature. Consider the following example:

abstract class I1 {
  void foo(dynamic d);
}

abstract class I2 {
  void foo(Object o);
}

abstract class C implements I1, I2 {}

In both situations—when foo accepts an argument of type dynamic and when it accepts an Object—the acceptable actual arguments are exactly the same: Every object can be passed. Moreover, the formal parameters d and o are not in scope anywhere, so there will never be an expression like d.bar or o.bar which is allowed respectively rejected because the receiver is or is not dynamic. In other words, it does not matter for clients of C whether that argument type is dynamic or Object.

During inference, the type-from-context for an actual argument to foo will depend on the choice: It will be dynamic respectively Object. However, this choice will not affect the treatment of the actual argument.

One case worth considering is the following:

abstract class I1 {
  void foo(dynamic f());
}

abstract class I2 {
  void foo(Object f());
}

If a function literal is passed in at a call site, it may have its return type inferred to dynamic respectively Object. This will change the type-from-context for any returned expressions, but just like the case for the actual parameter, that will not change the treatment of such expressions. Again, it does not matter for clients calling foo whether that type is dynamic or Object.

Conversely, the choice of top type matters when it is placed in a contravariant location in the parameter type:

abstract class I1 {
  void foo(int f(dynamic d));
}

abstract class I2 {
  void foo(int f(Object o));
}

In this situation, a function literal used as an actual argument at a call site for foo would receive an inferred type annotation for its formal parameter of dynamic respectively Object, and the usage of that parameter in the body of the function literal would then differ. In other words, the developer who declares foo may decide whether the code in the body of the function literal at the call sites should use strict or relaxed type checking—and it would be highly error-prone if this decision were to be made in a way which is unspecified.

All in all, it may be useful to “erase” all top types to Object when they occur in contravariant positions in method signatures, such that the differences that may exist do not create conflicts; in contrast, the top types that occur in covariant positions are significant, and hence the fact that we require such conflicts to be resolved explicitly is unlikely to be relaxed.

Updates

  • Apr 24th 2018, version 0.3: Renamed ‘override-specific’ to ‘interface-specific’, to avoid giving the impression that it can be used to determine whether a given signature can override another one (the override check must use different rules, e.g., it must allow dynamic foo(); to override Object foo(); and vice versa).

  • Apr 16th 2018, version 0.2: Introduced the relation ‘more override-specific than’ in order to handle top types more consistently and concisely.

  • Feb 8th 2018, version 0.1: Initial version.