Dynamic dispatch

Short description: Selecting which implementation of a method or function to call at run time

In computer science, dynamic dispatch is the process of selecting which implementation of a polymorphic operation (method or function) to call at run time. It is commonly employed in, and considered a prime characteristic of, object-oriented programming (OOP) languages and systems.

Object-oriented systems model a problem as a set of interacting objects that enact operations referred to by name. Polymorphism is the phenomenon wherein somewhat interchangeable objects each expose an operation of the same name but possibly differing in behavior. As an example, a File object and a Database object both have a StoreRecord method that can be used to write a personnel record to storage. Their implementations differ. A program holds a reference to an object which may be either a File object or a Database object. Which it is may have been determined by a run-time setting, and at this stage, the program may not know or care which. When the program calls StoreRecord on the object, something needs to choose which behavior gets enacted. If one thinks of OOP as sending messages to objects, then in this example the program sends a StoreRecord message to an object of unknown type, leaving it to the run-time support system to dispatch the message to the right object. The object enacts whichever behavior it implements.

Dynamic dispatch contrasts with static dispatch, in which the implementation of a polymorphic operation is selected at compile time. The purpose of dynamic dispatch is to defer the selection of an appropriate implementation until the run time type of a parameter (or multiple parameters) is known.

Dynamic dispatch is different from late binding (also known as dynamic binding). Name binding associates a name with an operation. A polymorphic operation has several implementations, all associated with the same name. Bindings can be made at compile time or (with late binding) at run time. With dynamic dispatch, one particular implementation of an operation is chosen at run time. While dynamic dispatch does not imply late binding, late binding does imply dynamic dispatch, since the implementation of a late-bound operation is not known until run time.

Single and multiple dispatch

The choice of which version of a method to call may be based either on a single object, or on a combination of objects. The former is called single dispatch and is directly supported by common object-oriented languages such as Smalltalk, C++, Java, C#, Objective-C, Swift, JavaScript, and Python. In these and similar languages, one may call a method for division with syntax that resembles

dividend.divide(divisor)  # dividend / divisor

where the parameters are optional. This is thought of as sending a message named divide with parameter divisor to dividend. An implementation will be chosen based only on dividend's type (perhaps rational, floating point, matrix), disregarding the type or value of divisor.

By contrast, some languages dispatch methods or functions based on the combination of operands; in the division case, the types of the dividend and divisor together determine which divide operation will be performed. This is known as multiple dispatch. Examples of languages that support multiple dispatch are Common Lisp, Dylan, and Julia.

Dynamic dispatch mechanisms

A language may be implemented with different dynamic dispatch mechanisms. The choices of the dynamic dispatch mechanism offered by a language to a large extent alter the programming paradigms that are available or are most natural to use within a given language.

Normally, in a typed language, the dispatch mechanism will be performed based on the type of the arguments (most commonly based on the type of the receiver of a message). Languages with weak or no typing systems often carry a dispatch table as part of the object data for each object. This allows instance behaviour as each instance may map a given message to a separate method.

Some languages offer a hybrid approach.

Dynamic dispatch will always incur an overhead so some languages offer static dispatch for particular methods.

C++ implementation

C++ uses early binding and offers both dynamic and static dispatch. The default form of dispatch is static. To get dynamic dispatch the programmer must declare a method as virtual.

C++ compilers typically implement dynamic dispatch with a data structure called a virtual function table (vtable) that defines the name-to-implementation mapping for a given class as a set of member function pointers. This is purely an implementation detail, as the C++ specification does not mention vtables. Instances of that type will then store a pointer to this table as part of their instance data, complicating scenarios when multiple inheritance is used. Since C++ does not support late binding, the virtual table in a C++ object cannot be modified at runtime, which limits the potential set of dispatch targets to a finite set chosen at compile time.

Type overloading does not produce dynamic dispatch in C++ as the language considers the types of the message parameters part of the formal message name. This means that the message name the programmer sees is not the formal name used for binding.

Go, Rust and Nim implementation

In Go, Rust and Nim, a more versatile variation of early binding is used. Vtable pointers are carried with object references as 'fat pointers' ('interfaces' in Go, or 'trait objects' in Rust).

This decouples the supported interfaces from the underlying data structures. Each compiled library needn't know the full range of interfaces supported in order to correctly use a type, just the specific vtable layout that they require. Code can pass around different interfaces to the same piece of data to different functions. This versatility comes at the expense of extra data with each object reference, which is problematic if many such references are stored persistently.

The term fat pointer simply refers to a pointer with additional associated information. The additional information may be a vtable pointer for dynamic dispatch described above, but is more commonly the associated object's size to describe e.g. a slice.

Smalltalk implementation

Smalltalk uses a type-based message dispatcher. Each instance has a single type whose definition contains the methods. When an instance receives a message, the dispatcher looks up the corresponding method in the message-to-method map for the type and then invokes the method.

Because a type can have a chain of base types, this look-up can be expensive. A naive implementation of Smalltalk's mechanism would seem to have a significantly higher overhead than that of C++ and this overhead would be incurred for each and every message that an object receives.

Real Smalltalk implementations often use a technique known as inline caching that makes method dispatch very fast. Inline caching basically stores the previous destination method address and object class of the call site (or multiple pairs for multi-way caching). The cached method is initialized with the most common target method (or just the cache miss handler), based on the method selector. When the method call site is reached during execution, it just calls the address in the cache. (In a dynamic code generator, this call is a direct call as the direct address is back patched by cache miss logic.) Prologue code in the called method then compares the cached class with the actual object class, and if they don't match, execution branches to a cache miss handler to find the correct method in the class. A fast implementation may have multiple cache entries and it often only takes a couple of instructions to get execution to the correct method on an initial cache miss. The common case will be a cached class match, and execution will just continue in the method.

Out-of-line caching can also be used in the method invocation logic, using the object class and method selector. In one design, the class and method selector are hashed, and used as an index into a method dispatch cache table.

As Smalltalk is a reflective language, many implementations allow mutating individual objects into objects with dynamically generated method lookup tables. This allows altering object behavior on a per object basis. A whole category of languages known as prototype-based languages has grown from this, the most famous of which are Self and JavaScript. Careful design of the method dispatch caching allows even prototype-based languages to have high-performance method dispatch.

Many other dynamically typed languages, including Python, Ruby, Objective-C and Groovy use similar approaches.

Example in Python

class Cat:
    def speak(self):
        print("Meow")

class Dog:
    def speak(self):
        print("Woof")


def speak(pet):
    # Dynamically dispatches the speak method
    # pet can either be an instance of Cat or Dog
    pet.speak()

cat = Cat()
speak(cat)
dog = Dog()
speak(dog)

Example in C++

#include <iostream>

// make Pet an abstract virtual base class
class Pet {
public:
    virtual void speak() = 0;
};

class Dog : public Pet {
public:
    void speak() override
    {
        std::cout << "Woof!\n";
    }
};

class Cat : public Pet {
public:
    void speak() override
    {
        std::cout << "Meow!\n";
    }
};

// speak() will be able to accept anything deriving from Pet
void speak(Pet& pet)
{
    pet.speak();
}

int main()
{
    Dog fido;
    Cat simba;
    speak(fido);
    speak(simba);
    return 0;
}

See also

• Function multi-versioning
• Function overloading
• Message passing
• Method overriding
• Double dispatch
• Name binding

Further reading

• Inside the C++ Object Model. Addison-Wesley. 1996. ISBN 0-201-83454-5. 
• Groeber, Marcus; Di Geronimo, Jr., Edward "Ed"; Paul, Matthias R. (2002-03-02) [2002-02-24]. "GEOS/NDO info for RBIL62?". Newsgroup: comp.os.geos.programmer. Archived from the original on 2019-04-20. Retrieved 2019-04-20. […] The reason Geos needs 16 interrupts is because the scheme is used to convert inter-segment ("far") function calls into interrupts, without changing the size of the code. The reason this is done so that "something" (the kernel) can hook itself into every inter-segment call made by a Geos application and make sure that the proper code segments are loaded from virtual memory and locked down. In DOS terms, this would be comparable to an overlay loader, but one that can be added without requiring explicit support from the compiler or the application. What happens is something like this: […] 1. The real mode compiler generates an instruction like this: CALL <segment>:<offset> -> 9A <offlow><offhigh><seglow><seghigh> with <seglow><seghigh> normally being defined as an address that must be fixed up at load time depending on the address where the code has been placed. […] 2. The Geos linker turns this into something else: INT 8xh -> CD 8x […] DB <seghigh>,<offlow>,<offhigh> […] Note that this is again five bytes, so it can be fixed up "in place". Now the problem is that an interrupt requires two bytes, while a CALL FAR instruction only needs one. As a result, the 32-bit vector (<seg><ofs>) must be compressed into 24 bits. […] This is achieved by two things: First, the <seg> address is encoded as a "handle" to the segment, whose lowest nibble is always zero. This saves four bits. In addition […] the remaining four bits go into the low nibble of the interrupt vector, thus creating anything from INT 80h to 8Fh. […] The interrupt handler for all those vectors is the same. It will "unpack" the address from the three-and-a-half byte notation, look up the absolute address of the segment, and forward the call, after having done its virtual memory loading thing... Return from the call will also pass through the corresponding unlocking code. […] The low nibble of the interrupt vector (80h–8Fh) holds bit 4 through 7 of the segment handle. Bit 0 to 3 of a segment handle are (by definition of a Geos handle) always 0. […] all Geos API run through the "overlay" scheme […]: when a Geos application is loaded into memory, the loader will automatically replace calls to functions in the system libraries by the corresponding INT-based calls. Anyway, these are not constant, but depend on the handle assigned to the library's code segment. […] Geos was originally intended to be converted to protected mode very early on […], with real mode only being a "legacy option" […] almost every single line of assembly code is ready for it […]
• "Re: [fd-dev ANNOUNCE: CuteMouse 2.0 alpha 1"]. freedos-dev. 2002-04-11. https://marc.info/?l=freedos-dev&m=101853130816682&w=2. "[…] in case of such mangled pointers […] many years ago Axel and I were thinking about a way how to use *one* entry point into a driver for multiple interrupt vectors (as this would save us a lot of space for the multiple entry points and the more or less identical startup/exit framing code in all of them), and then switch to the different interrupt handlers internally. For example: 1234h:0000h […] 1233h:0010h […] 1232h:0020h […] 1231h:0030h […] 1230h:0040h […] all point to exactly the same entry point. If you hook INT 21h onto 1234h:0000h and INT 2Fh onto 1233h:0010h, and so on, they would all go through the same "loophole", but you would still be able to distinguish between them and branch into the different handlers internally. Think of a "compressed" entry point into a A20 stub for HMA loading. This works as long as no program starts doing segment:offset magics. […] Contrast this with the opposite approach to have multiple entry points (maybe even supporting IBM's Interrupt Sharing Protocol), which consumes much more memory if you hook many interrupts. […] We came to the result that this would most probably not be save in practise because you never know if other drivers normalize or denormalize pointers, for what reasons ever. […]"  (NB. Something similar to "fat pointers" specifically for Intel's real-mode segment:offset addressing on x86 processors, containing both a deliberately denormalized pointer to a shared code entry point and some info to still distinguish the different callers in the shared code. While, in an open system, pointer-normalizing 3rd-party instances (in other drivers or applications) cannot be ruled out completely on public interfaces, the scheme can be used safely on internal interfaces to avoid redundant entry code sequences.)
• "C's Biggest Mistake". Digital Mars. 2009-12-22. https://digitalmars.com/articles/C-biggest-mistake.html. Retrieved 2022-07-11.  [1]
• "A Fat Pointer Library". Cello: High Level C. 2015. https://libcello.org/learn/a-fat-pointer-library.