Generics

Generics can be used to define functions and structs over different input data types. This language feature is sometimes referred to as parametric polymorphism. In Move, we will often use the term generics interchangeably with type parameters and type arguments.

Generics are commonly used in library code, such as in vector, to declare code that works over any possible instantiation (that satisfies the specified constraints). In other frameworks, generic code can sometimes be used to interact with global storage many different ways that all still share the same implementation.

Declaring Type Parameters

Both functions and structs can take a list of type parameters in their signatures, enclosed by a pair of angle brackets <...>.

Generic Functions

Type parameters for functions are placed after the function name and before the (value) parameter list. The following code defines a generic identity function that takes a value of any type and returns that value unchanged.

fun id<T>(x: T): T {
    // this type annotation is unnecessary but valid
    (x: T)
}

Once defined, the type parameter T can be used in parameter types, return types, and inside the function body.

Generic Structs

Type parameters for structs are placed after the struct name, and can be used to name the types of the fields.

struct Foo<T> has copy, drop { x: T }

struct Bar<T1, T2> has copy, drop {
    x: T1,
    y: vector<T2>,
}

Note that type parameters do not have to be used

Type Arguments

Calling Generic Functions

When calling a generic function, one can specify the type arguments for the function's type parameters in a list enclosed by a pair of angle brackets.

fun foo() {
    let x = id<bool>(true);
}

If you do not specify the type arguments, Move's type inference will supply them for you.

Using Generic Structs

Similarly, one can attach a list of type arguments for the struct's type parameters when constructing or destructing values of generic types.

fun foo() {
    let foo = Foo<bool> { x: true };
    let Foo<bool> { x } = foo;
}

If you do not specify the type arguments, Move's type inference will supply them for you.

Type Argument Mismatch

If you specify the type arguments, and they conflict with the actual values supplied, an error will be given:

fun foo() {
    let x = id<u64>(true); // error! true is not a u64
}

and similarly:

fun foo() {
    let foo = Foo<bool> { x: 0 }; // error! 0 is not a bool
    let Foo<address> { x } = foo; // error! bool is incompatible with address
}

Type Inference

In most cases, the Move compiler will be able to infer the type arguments, so you don't have to write them down explicitly. Here's what the examples above would look like if we omit the type arguments:

fun foo() {
    let x = id(true);
    //        ^ <bool> is inferred

    let foo = Foo { x: true };
    //           ^ <bool> is inferred

    let Foo { x } = foo;
    //     ^ <bool> is inferred
}

Note: when the compiler is unable to infer the types, you'll need annotate them manually. A common scenario is to call a function with type parameters appearing only at return positions.

address 0x2 {
module m {
    using std::vector;

    fun foo() {
        // let v = vector::new();
        //                    ^ The compiler cannot figure out the element type.

        let v = vector::new<u64>();
        //                 ^~~~~ Must annotate manually.
    }
}
}

However, the compiler will be able to infer the type if that return value is used later in that function:

address 0x2 {
module m {
    using std::vector;

    fun foo() {
        let v = vector::new();
        //                 ^ <u64> is inferred
        vector::push_back(&mut v, 42);
    }
}
}

Unused Type Parameters

For a struct definition, an unused type parameter is one that does not appear in any field defined in the struct, but is checked statically at compile time. Move allows unused type parameters so the following struct definition is valid:

struct Foo<T> {
    foo: u64
}

This can be convenient when modeling certain concepts. Here is an example:

address 0x2 {
module m {
    // Currency Specifiers
    struct Currency1 {}
    struct Currency2 {}

    // A generic coin type that can be instantiated using a currency
    // specifier type.
    //   e.g. Coin<Currency1>, Coin<Currency2> etc.
    struct Coin<Currency> has store {
        value: u64
    }

    // Write code generically about all currencies
    public fun mint_generic<Currency>(value: u64): Coin<Currency> {
        Coin { value }
    }

    // Write code concretely about one currency
    public fun mint_concrete(value: u64): Coin<Currency1> {
        Coin { value }
    }
}
}

In this example, struct Coin<Currency> is generic on the Currency type parameter, which specifies the currency of the coin and allows code to be written either generically on any currency or concretely on a specific currency. This genericity applies even when the Currency type parameter does not appear in any of the fields defined in Coin.

Phantom Type Parameters

In the example above, although struct Coin asks for the store ability, neither Coin<Currency1> nor Coin<Currency2> will have the store ability. This is because of the rules for Conditional Abilities and Generic Types and the fact that Currency1 and Currency2 don't have the store ability, despite the fact that they are not even used in the body of struct Coin. This might cause some unpleasant consequences. For example, we are unable to put Coin<Currency1> into a wallet in the global storage.

One possible solution would be to add spurious ability annotations to Currency1 and Currency2 (i.e., struct Currency1 has store {}). But, this might lead to bugs or security vulnerabilities because it weakens the types with unnecessary ability declarations. For example, we would never expect a resource in the global storage to have a field in type Currency1, but this would be possible with the spurious store ability. Moreover, the spurious annotations would be infectious, requiring many functions generic on the unused type parameter to also include the necessary constraints.

Phantom type parameters solve this problem. Unused type parameters can be marked as phantom type parameters, which do not participate in the ability derivation for structs. In this way, arguments to phantom type parameters are not considered when deriving the abilities for generic types, thus avoiding the need for spurious ability annotations. For this relaxed rule to be sound, Move's type system guarantees that a parameter declared as phantom is either not used at all in the struct definition, or it is only used as an argument to type parameters also declared as phantom.

Declaration

In a struct definition a type parameter can be declared as phantom by adding the phantom keyword before its declaration. If a type parameter is declared as phantom we say it is a phantom type parameter. When defining a struct, Move's type checker ensures that every phantom type parameter is either not used inside the struct definition or it is only used as an argument to a phantom type parameter.

More formally, if a type is used as an argument to a phantom type parameter we say the type appears in phantom position. With this definition in place, the rule for the correct use of phantom parameters can be specified as follows: A phantom type parameter can only appear in phantom position.

The following two examples show valid uses of phantom parameters. In the first one, the parameter T1 is not used at all inside the struct definition. In the second one, the parameter T1 is only used as an argument to a phantom type parameter.

struct S1<phantom T1, T2> { f: u64 }
                  ^^
                  Ok: T1 does not appear inside the struct definition


struct S2<phantom T1, T2> { f: S1<T1, T2> }
                                  ^^
                                  Ok: T1 appears in phantom position

The following code shows examples of violations of the rule:

struct S1<phantom T> { f: T }
                          ^
                          Error: Not a phantom position

struct S2<T> { f: T }

struct S3<phantom T> { f: S2<T> }
                             ^
                             Error: Not a phantom position

Instantiation

When instantiating a struct, the arguments to phantom parameters are excluded when deriving the struct abilities. For example, consider the following code:

struct S<T1, phantom T2> has copy { f: T1 }
struct NoCopy {}
struct HasCopy has copy {}

Consider now the type S<HasCopy, NoCopy>. Since S is defined with copy and all non-phantom arguments have copy then S<HasCopy, NoCopy> also has copy.

Phantom Type Parameters with Ability Constraints

Ability constraints and phantom type parameters are orthogonal features in the sense that phantom parameters can be declared with ability constraints. When instantiating a phantom type parameter with an ability constraint, the type argument has to satisfy that constraint, even though the parameter is phantom. For example, the following definition is perfectly valid:

struct S<phantom T: copy> {}

The usual restrictions apply and T can only be instantiated with arguments having copy.

Constraints

In the examples above, we have demonstrated how one can use type parameters to define "unknown" types that can be plugged in by callers at a later time. This however means the type system has little information about the type and has to perform checks in a very conservative way. In some sense, the type system must assume the worst case scenario for an unconstrained generic. Simply put, by default generic type parameters have no abilities.

This is where constraints come into play: they offer a way to specify what properties these unknown types have so the type system can allow operations that would otherwise be unsafe.

Declaring Constraints

Constraints can be imposed on type parameters using the following syntax.

// T is the name of the type parameter
T: <ability> (+ <ability>)*

The <ability> can be any of the four abilities, and a type parameter can be constrained with multiple abilities at once. So all the following would be valid type parameter declarations:

T: copy
T: copy + drop
T: copy + drop + store + key

Verifying Constraints

Constraints are checked at call sites so the following code won't compile.

struct Foo<T: key> { x: T }

struct Bar { x: Foo<u8> }
//                  ^ error! u8 does not have 'key'

struct Baz<T> { x: Foo<T> }
//                     ^ error! T does not have 'key'

struct R {}

fun unsafe_consume<T>(x: T) {
    // error! x does not have 'drop'
}

fun consume<T: drop>(x: T) {
    // valid!
    // x will be dropped automatically
}

fun foo() {
    let r = R {};
    consume<R>(r);
    //      ^ error! R does not have 'drop'
}

struct R {}

fun unsafe_double<T>(x: T) {
    (copy x, x)
    // error! x does not have 'copy'
}

fun double<T: copy>(x: T) {
    (copy x, x) // valid!
}

fun foo(): (R, R) {
    let r = R {};
    double<R>(r)
    //     ^ error! R does not have 'copy'
}

For more information, see the abilities section on conditional abilities and generic types.

Limitations on Recursions

Recursive Structs

Generic structs can not contain fields of the same type, either directly or indirectly, even with different type arguments. All the following struct definitions are invalid:

struct Foo<T> {
    x: Foo<u64> // error! 'Foo' containing 'Foo'
}

struct Bar<T> {
    x: Bar<T> // error! 'Bar' containing 'Bar'
}

// error! 'A' and 'B' forming a cycle, which is not allowed either.
struct A<T> {
    x: B<T, u64>
}

struct B<T1, T2> {
    x: A<T1>
    y: A<T2>
}

Advanced Topic: Type-level Recursions

Move allows generic functions to be called recursively. However, when used in combination with generic structs, this could create an infinite number of types in certain cases, and allowing this means adding unnecessary complexity to the compiler, vm and other language components. Therefore, such recursions are forbidden.

Allowed:

address 0x2 {
module m {
    struct A<T> {}

    // Finitely many types -- allowed.
    // foo<T> -> foo<T> -> foo<T> -> ... is valid
    fun foo<T>() {
        foo<T>();
    }

    // Finitely many types -- allowed.
    // foo<T> -> foo<A<u64>> -> foo<A<u64>> -> ... is valid
    fun foo<T>() {
        foo<A<u64>>();
    }
}
}

Not allowed:

address 0x2 {
module m {
    struct A<T> {}

    // Infinitely many types -- NOT allowed.
    // error!
    // foo<T> -> foo<A<T>> -> foo<A<A<T>>> -> ...
    fun foo<T>() {
        foo<A<T>>();
    }
}
}

address 0x2 {
module n {
    struct A<T> {}

    // Infinitely many types -- NOT allowed.
    // error!
    // foo<T1, T2> -> bar<T2, T1> -> foo<T2, A<T1>>
    //   -> bar<A<T1>, T2> -> foo<A<T1>, A<T2>>
    //   -> bar<A<T2>, A<T1>> -> foo<A<T2>, A<A<T1>>>
    //   -> ...
    fun foo<T1, T2>() {
        bar<T2, T1>();
    }

    fun bar<T1, T2> {
        foo<T1, A<T2>>();
    }
}
}

Note, the check for type level recursions is based on a conservative analysis on the call sites and does NOT take control flow or runtime values into account.

address 0x2 {
module m {
    struct A<T> {}

    fun foo<T>(n: u64) {
        if (n > 0) {
            foo<A<T>>(n - 1);
        };
    }
}
}

The function in the example above will technically terminate for any given input and therefore only creating finitely many types, but it is still considered invalid by Move's type system.