Dependency Injection Container for C++, with DI, next-gen and an 'o' in a name. Header only.
Tested with:
- C++17, C++20, C++23
- GCC 9-12
- Clang 11-15
- Visual Studio 2019, 2022
Features Overview:
- Non-intrusive Class Registration
- Customizable Scopes
- Customizable Factories
- Service Locator Pattern
- Runtime-based Resolution
- Customizable RTTI
- Container Nesting
- Static and Dynamic Containers
- Multibindings
- Named Resolution
- Annotated Types
- Customizable Allocation
Dingo is a dependency injection library that preforms recursive instantiation of registered types. It has no dependencies except for the C++ standard library. History of the project is summarized in Motivation and History.
Tests are using google/googletest, benchmarks are using google/benchmark.
Example code included from examples/quick.cpp:
// Class types to be managed by the container. Note that there is no special
// code required for a type to be used by the container and that conversions
// are applied automatically based on an registered type scope.
struct A {
A() {}
};
struct B {
B(A&, std::shared_ptr<A>) {}
};
struct C {
C(B*, std::unique_ptr<B>&, A&) {}
};
container<> container;
// Register struct A with a shared scope, stored as std::shared_ptr<A>
container.register_type<scope<shared>, storage<std::shared_ptr<A>>>();
// Register struct B with a shared scope, stored as std::unique_ptr<B>
container.register_type<scope<shared>, storage<std::unique_ptr<B>>>();
// Register struct C with an unique scope, stored as plain C
container.register_type<scope<unique>, storage<C>>();
// Resolving the struct C will recursively instantiate required dependencies
// (structs A and B) and inject the instances based on their scopes into C.
// As C is in unique scope, each resolve<C> will return new C instance.
// As A and B are in shared scope, each C will get the same instances
// injected.
C c = container.resolve<C>();
struct D {
A& a;
B* b;
};
// Construct an un-managed struct using dependencies from the container
D d = container.construct<D>();
// Invoke callable
D e = container.invoke([&](A& a, B* b) { return D{a, b}; });To use the library from CMake-based projects, FetchContent or vcpkg can be used. Alias dingo::dingo is added as an INTERFACE library without any other dependencies.
Example code included from test/fetchcontent/CMakeLists.txt:
include(FetchContent)
FetchContent_Declare(dingo
GIT_REPOSITORY https://github.com/romanpauk/dingo.git
GIT_TAG master)
FetchContent_MakeAvailable(dingo)
target_link_libraries(dingo_fetchcontent_test PRIVATE dingo::dingo)As creating a generic DI library is tricky, libraries end up different in many aspects (API, supported C++ constructs, compile-time/run-time, intrusiveness). Please take a look at few other libraries to find out which one suits the intended use-case the best:
Managed types do not need to adhere to any special protocol. Constructor auto-detection is attempted, but it is possible to provide a custom factory to use for type construction. During the type registration, user can customize the following: factory to create an instance, instance scope, type of the stored instance, allowed conversions for resolution and interfaces that can be used to resolve the instance. The customization is done through policies applied during a register_type call. Some of the policies can be ommited if it is possible to deduce them from other policies.
Example code included from examples/non_intrusive.cpp:
container<> container;
// Registration of a struct A
container.register_type<scope<unique>, // using unique scope
factory<constructor<A>>, // using constructor-detecting
// factory
storage<std::unique_ptr<A>>, // stored as unique_ptr<A>
interface<A> // resolvable as A
>();
// As some policies can be deduced from the others, the above
// registration simplified
container.register_type<scope<unique>, storage<std::unique_ptr<A>>>();See dingo/type_registration.h for details.
Type instances are stored in a form that is specified during a registration. Usual semantics of the stored type is respected (example: if a type is stored as std::shared_ptr, it will be created using std::make_shared). Stored type, together with type's scope, determines what conversions can be applied to an instance during resolution of other instance's dependencies. Allowed conversions are different for each scope, as not all conversions can be performed safely or without unwanted side-effects. Conversions that could be applied are dereferencing, taking an address, down-casting and referencing down-casted types.
Factories parametrize how a type is instantiated. Factory type is deduced and does not have to be specified during a type registration. Factories can generally be stateless or stateful.
Constructor to create a type instance is determined automatically by attempting a list initialization, going from a pre-configured number of arguments down to zero. In a case the constructor is not determined, compile assert occurs. This factory type is a default one so it does not have to be specified.
Example code included from examples/factory_constructor_deduction.cpp:
struct A {
A(int); // Definition is not required as constructor is not called
A(double, double) {} // Definition is required as constructor is called
};
container<> container;
container.register_type<scope<external>, storage<double>>(1.1);
// Constructor with a highest arity will be used (factory<> is deduced
// automatically)
container.register_type<scope<unique>,
storage<A> /*, factory<constructor_deduction<A>> */>();See dingo/factory/constructor_deduction.h for details.
In a case an automatic deduction fails due to an ambiguity, it is possible to override the constructor by specifying an constructor overload that will be used.
Example code included from examples/factory_constructor.cpp:
struct A {
A(int); // Definition is not required as constructor is not called
A(double) {} // Definition is required as constructor is called
};
container<> container;
container.register_type<scope<external>, storage<double>>(1.1);
// Register class A that will be constructed using manually selected
// A(double) constructor. Manually disambiguation is required to avoid
// compile time assertion
container.register_type<scope<unique>, storage<A>,
factory<constructor<A(double)>>>();See dingo/factory/constructor.h for details.
Static function factory allows a static function to be used for an instance creation.
Example code included from examples/factory_function.cpp:
// Declare struct A that has an inaccessible constructor
struct A {
// Provide factory function to construct instance of A
static A factory() { return A(); }
private:
A() = default;
};
container<> container;
// Register A that will be instantiated by calling A::factory()
container
.register_type<scope<unique>, storage<A>, factory<function<&A::factory>>>();See dingo/factory/function.h for details.
Using a stateful factory, any functor can be used to construct a type.
Example code included from examples/factory_callable.cpp:
struct A {
int value;
};
container<> container;
// Register double that will be passed to lambda function below
container.register_type<scope<external>, storage<int>>(2);
// Register A that will be instantiated by calling provided lambda function
// with arguments resolved using the container.
container.register_type<scope<unique>, storage<A>>(callable([](int value) {
return A{value * 2};
}));
assert(container.resolve<A>().value == 4);See dingo/factory/callable.h for details.
Scopes parametrize stored instances' lifetime. Based on a stored type, each scope defines a list of allowed conversions that are applied to the stored type during resolution to satisfy dependencies of other types. Both cyclical and a-cyclical dependency graphs are supported.
An already existing instance is referred. The scope can eventually take the ownership by moving the instance inside into the scope storage.
Example code included from examples/scope_external.cpp:
struct A {
} instance;
container<> container;
// Register existing instance of A, stored as a pointer.
container.register_type<scope<external>, storage<A*>>(&instance);
// Resolution will return an existing instance of A casted to required type.
assert(&container.resolve<A&>() == container.resolve<A*>());See dingo/storage/external.h for allowed conversions for accessing external instances.
An unique instance is created for each resolution. In a case dependencies form a cycle, an exception is thrown. See dingo/storage/unique.h for allowed conversions for accessing unique instances.
Example code included from examples/scope_unique.cpp:
struct A {};
container<> container;
// Register struct A with unique scope
container.register_type<scope<unique>, storage<A>>();
// Resolution will get a unique instance of A
container.resolve<A>();The instance is cached for a subsequent resolutions. In a case dependencies form a cycle, an exception is thrown. See dingo/storage/shared.h for allowed conversions for accessing shared instances.
Example code included from examples/scope_shared.cpp:
struct A {};
container<> container;
// Register struct A with shared scope
container.register_type<scope<shared>, storage<A>>();
// Resolution will always return the same A instance
assert(container.resolve<A*>() == &container.resolve<A&>());The instance is cached for a subsequent resolutions and allows to create object graphs with cycles. This is implemented using two-phase construction and thus has some limitations: injected types from a cyclical storage can't be used safely in constructors, and as some conversions require a type to be fully constructed, those are disallowed additionally (injecting an instance through it's virtual base class). The functionality is quite a hack, but not all code bases have a clean, a-cyclical dependencies, so it could be handy.
Example code included from examples/scope_shared_cyclical.cpp:
// Forward-declare structs that have cyclical dependency
struct A;
struct B;
// Declare struct A, note that its constructor is taking arguments of struct B
struct A {
A(B& b, std::shared_ptr<B> bptr) : b_(b), bptr_(bptr) {}
B& b_;
std::shared_ptr<B> bptr_;
};
// Declare struct B, note that its constructor is taking arguments of struct A
struct B {
B(A& a, A* aptr) : a_(a), aptr_(aptr) {}
A& a_;
A* aptr_;
};
container<> container;
// Register struct A with cyclical scope
container.register_type<scope<shared_cyclical>, storage<A>>();
// Register struct B with cyclical scope
container.register_type<scope<shared_cyclical>, storage<std::shared_ptr<B>>>();
// Returns instance of A that has correctly set b_ member to an instance of
// B, and instance of B has correctly set a_ member to an instance of A.
// Conversions are supported with cycles, too.
A& a = container.resolve<A&>();
B& b = container.resolve<B&>();
// Check that the instances are constructed as promised
assert(&a.b_ == &b);
assert(&a.b_ == a.bptr_.get());
assert(&b.a_ == &a);
assert(b.aptr_ == &a);See dingo/storage/shared_cyclical.h for allowed conversions for accessing shared cycle-aware instances.
Container can be used from multiple modules, specifically it can be filled in one module and used in other modules. This is a requirement for larger projects with not so clean dependencies. Compared to the compile-time DI solutions, errors are propagated in runtime as exceptions.
When an exception occurs in a user-provided code during instance resolution, the state of a container is valid with all successfully resolved instances intact. Note that as a resolution is a recursive process, the container state can be modified due to successful resolutions being made before reaching a resolution that will fail.
Template class specializations are used to tweak the behavior and functionality for different storage types, instance creation or passing instances through type-erasure boundary. Standard RTTI is optional and an emulation can be used. Single-module-single-container use-case is optimized with static type maps.
It is possible to register a binding to the type under a key that is a base class of the type. As an upcast is compiled at the time of a registration so multiple inheritance is correctly supported even without using dynamic_cast operator. One instance can be resolved through multiple interfaces.
Example code included from examples/service_locator.cpp:
// Interface that will be resolved
struct IA {
virtual ~IA() {}
};
// Struct implementing the interface
struct A : IA {};
container<> container;
// Register struct A, resolvable as interface IA
container.register_type<scope<shared>, storage<A>, interface<IA>>();
// Resolve instance A through interface IA
IA& instance = container.resolve<IA&>();
assert(dynamic_cast<A*>(&instance));Multibindings allow to resolve a collection of types that are resolvable using the same interface. The collection can participate in a resolution of other types. Lambda function can be used to transform the resolved types.
Example code included from examples/multibindings.cpp:
struct IProcessor {
virtual ~IProcessor() {}
};
template <size_t N> struct Processor : IProcessor {};
container<> container;
// Register multi-bindings collection
container.template register_type_collection<
scope<shared>, storage<std::vector<IProcessor*>>>();
// Register types under the same interface
container.template register_type<scope<shared>, storage<Processor<0>>,
interface<IProcessor>>();
container.template register_type<scope<shared>, storage<Processor<1>>,
interface<IProcessor>>();
// Resolve the collection
container.template resolve<std::vector<IProcessor*>>();Traits used during container construction can specify one or more indexes allowing each type to be indexed under additional user-specified key to the type interface. The index data structure can be overridden using template specialization, with the library providing specializations for std::map, std::unordered_map and std::array.
Example code included from examples/index.cpp:
struct IAnimal {
virtual ~IAnimal() {}
};
struct Dog : IAnimal {};
struct Cat : IAnimal {};
// Declare traits with std::string based index
struct container_traits : dynamic_container_traits {
using index_definition_type =
std::tuple<std::tuple<std::string, index_type::unordered_map>>;
};
container<container_traits> container;
container.template register_indexed_type<
scope<shared>, storage<Dog>, interface<IAnimal>>(std::string("dog"));
container.template register_indexed_type<
scope<shared>, storage<Cat>, interface<IAnimal>>(std::string("cat"));
// Resolve an instance of a dog
auto dog = container.template resolve<IAnimal>(std::string("dog"));It is possible to register different implementations with the same interface, disambiguating the registration with an user-provided tag. See test/annotated.cpp.
Unregistered types can be constructed using registered dependencies with construct() member function.
Example code included from examples/construct.cpp:
// struct A that will be registered with the container
struct A {};
// struct B that will be constructed using container
struct B {
A& a;
};
container<> container;
// Register struct A with shared scope
container.register_type<scope<shared>, storage<A>>();
// Construct instance of B, injecting shared instance of A
B b = container.construct<B>();Callable objects can be called using invoke() member function with arguments provided by the container. Supported callable types are lambdas, std::function and function pointers.
Example code included from examples/invoke.cpp:
// struct B that will be constructed using container
struct B {
A& a;
static B factory(A& a) { return B{a}; }
};
// Construct instance of B, injecting shared instance of A
B b1 = container.invoke([&](A& a) { return B{a}; });
B b2 = container.invoke(std::function<B(A&)>([](auto& a) { return B{a}; }));
B b3 = container.invoke(B::factory);For non-RTTI enabled builds, it is possible to parametrize the container with custom RTTI implementation. No library functionality depends on dynamic_cast conversion. Two RTTI providers are available: dynamic one based on typeid operator and static one, based on template specializations.
Static and dynamic container are just differently parametrized containers using container traits.
Container type can be parametrized with a custom memory allocator, RTTI provider and type map implementation. Dynamic containers use standard allocator, operator typeid() and type map implemented using standard map. Static containers use template specializations to implement required features, as all types that are ever present are always different types, allowing static member variables to be used to store and retrieve required data quickly.
While dynamic containers can be created completely freely, with static containers it is a user's responsibility to use type tag to distinguish the containers. The benefit of static containers is a performance gain. The drawback is slightly harder and limited use.
Note that "static" in this context does not mean "compile-time", both container parametrizations are fully runtime-based.
The container functions as a cache for already resolved types with shared scope, returning stable references to those types. The initial resolution needs to happen through some form of dynamic dispatch, as due to runtime nature of the container, some informations are not present on the caller side. But as soon as the type is resolved using factory through dynamic dispatch, the result is cached on the caller side, so subsequent resolution will retrieve it directly without having to use factory. For static containers, the cache is implemented using template specializations, so the lookup is as instant as it can be, lets just say it is much much faster than doing virtual function call.
As this slightly degrades performance for types with unique scope because on the caller side, we do not know the scope type from T, the feature can be turned on/off using traits.
Containers can form a parent-child hierarchy and resolution will traverse the container chain from child to last parent. Calling register_type() returns an implicitly created child container for the type being registered, allowing a per-type configuration to override a global configuration in the container. Nesting is supported for both dynamic and static type maps based containers.
Example code included from examples/nesting.cpp:
struct A {
int value;
};
struct B {
int value;
};
container<> base_container;
base_container.register_type<scope<external>, storage<int>>(1);
base_container.register_type<scope<unique>, storage<A>>();
// Resolving A will use A{1} to construct A
assert(base_container.resolve<A>().value == 1);
base_container.register_type<scope<unique>, storage<B>>()
.register_type<scope<external>, storage<int>>(
2); // Override value of int for struct B
// Resolving B will use B{2} to construct B
assert(base_container.resolve<B>().value == 2);
// Create a nested container, overriding B (effectively removing override in
// base_container)
container<> nested_container(&base_container);
nested_container.register_type<scope<unique>, storage<B>>();
// Resolving B using nested container will use B{1} as provided by the
// parent container to construct B
assert(nested_container.resolve<B>().value == 1);See test/nesting.cpp for details.
To customize memory management, container constructor can take an optional user supplied allocator that is used to allocate all memory required for container operations. Note that resolved instances do not use this allocator as the container does not have to own instances it constructs.
Example code included from examples/allocator.cpp:
// Define a container with user-provided allocator type
std::allocator<char> alloc;
container<dingo::dynamic_container_traits, std::allocator<char>>
container(alloc);The functionality is covered with tests written using google test. See available test coverage.