I think this question is still relevant today. Using the C++11 standard you are now able to implement a instanceof
function without using dynamic_cast
like this:
if (dynamic_cast<B*>(aPtr) != nullptr) {
// aPtr is instance of B
} else {
// aPtr is NOT instance of B
}
But you're still reliant on RTTI
support. So here is my solution for this problem depending on some Macros and Metaprogramming Magic. The only drawback imho is that this approach does not work for multiple inheritance.
InstanceOfMacros.h
#include <set>
#include <tuple>
#include <typeindex>
#define _EMPTY_BASE_TYPE_DECL() using BaseTypes = std::tuple<>;
#define _BASE_TYPE_DECL(Class, BaseClass) \
using BaseTypes = decltype(std::tuple_cat(std::tuple<BaseClass>(), Class::BaseTypes()));
#define _INSTANCE_OF_DECL_BODY(Class) \
static const std::set<std::type_index> baseTypeContainer; \
virtual bool instanceOfHelper(const std::type_index &_tidx) { \
if (std::type_index(typeid(ThisType)) == _tidx) return true; \
if (std::tuple_size<BaseTypes>::value == 0) return false; \
return baseTypeContainer.find(_tidx) != baseTypeContainer.end(); \
} \
template <typename... T> \
static std::set<std::type_index> getTypeIndexes(std::tuple<T...>) { \
return std::set<std::type_index>{std::type_index(typeid(T))...}; \
}
#define INSTANCE_OF_SUB_DECL(Class, BaseClass) \
protected: \
using ThisType = Class; \
_BASE_TYPE_DECL(Class, BaseClass) \
_INSTANCE_OF_DECL_BODY(Class)
#define INSTANCE_OF_BASE_DECL(Class) \
protected: \
using ThisType = Class; \
_EMPTY_BASE_TYPE_DECL() \
_INSTANCE_OF_DECL_BODY(Class) \
public: \
template <typename Of> \
typename std::enable_if<std::is_base_of<Class, Of>::value, bool>::type instanceOf() { \
return instanceOfHelper(std::type_index(typeid(Of))); \
}
#define INSTANCE_OF_IMPL(Class) \
const std::set<std::type_index> Class::baseTypeContainer = Class::getTypeIndexes(Class::BaseTypes());
You can then use this stuff (with caution) as follows:
DemoClassHierarchy.hpp*
#include "InstanceOfMacros.h"
struct A {
virtual ~A() {}
INSTANCE_OF_BASE_DECL(A)
};
INSTANCE_OF_IMPL(A)
struct B : public A {
virtual ~B() {}
INSTANCE_OF_SUB_DECL(B, A)
};
INSTANCE_OF_IMPL(B)
struct C : public A {
virtual ~C() {}
INSTANCE_OF_SUB_DECL(C, A)
};
INSTANCE_OF_IMPL(C)
struct D : public C {
virtual ~D() {}
INSTANCE_OF_SUB_DECL(D, C)
};
INSTANCE_OF_IMPL(D)
The following code presents a small demo to verify rudimentary the correct behavior.
InstanceOfDemo.cpp
#include <iostream>
#include <memory>
#include "DemoClassHierarchy.hpp"
int main() {
A *a2aPtr = new A;
A *a2bPtr = new B;
std::shared_ptr<A> a2cPtr(new C);
C *c2dPtr = new D;
std::unique_ptr<A> a2dPtr(new D);
std::cout << "a2aPtr->instanceOf<A>(): expected=1, value=" << a2aPtr->instanceOf<A>() << std::endl;
std::cout << "a2aPtr->instanceOf<B>(): expected=0, value=" << a2aPtr->instanceOf<B>() << std::endl;
std::cout << "a2aPtr->instanceOf<C>(): expected=0, value=" << a2aPtr->instanceOf<C>() << std::endl;
std::cout << "a2aPtr->instanceOf<D>(): expected=0, value=" << a2aPtr->instanceOf<D>() << std::endl;
std::cout << std::endl;
std::cout << "a2bPtr->instanceOf<A>(): expected=1, value=" << a2bPtr->instanceOf<A>() << std::endl;
std::cout << "a2bPtr->instanceOf<B>(): expected=1, value=" << a2bPtr->instanceOf<B>() << std::endl;
std::cout << "a2bPtr->instanceOf<C>(): expected=0, value=" << a2bPtr->instanceOf<C>() << std::endl;
std::cout << "a2bPtr->instanceOf<D>(): expected=0, value=" << a2bPtr->instanceOf<D>() << std::endl;
std::cout << std::endl;
std::cout << "a2cPtr->instanceOf<A>(): expected=1, value=" << a2cPtr->instanceOf<A>() << std::endl;
std::cout << "a2cPtr->instanceOf<B>(): expected=0, value=" << a2cPtr->instanceOf<B>() << std::endl;
std::cout << "a2cPtr->instanceOf<C>(): expected=1, value=" << a2cPtr->instanceOf<C>() << std::endl;
std::cout << "a2cPtr->instanceOf<D>(): expected=0, value=" << a2cPtr->instanceOf<D>() << std::endl;
std::cout << std::endl;
std::cout << "c2dPtr->instanceOf<A>(): expected=1, value=" << c2dPtr->instanceOf<A>() << std::endl;
std::cout << "c2dPtr->instanceOf<B>(): expected=0, value=" << c2dPtr->instanceOf<B>() << std::endl;
std::cout << "c2dPtr->instanceOf<C>(): expected=1, value=" << c2dPtr->instanceOf<C>() << std::endl;
std::cout << "c2dPtr->instanceOf<D>(): expected=1, value=" << c2dPtr->instanceOf<D>() << std::endl;
std::cout << std::endl;
std::cout << "a2dPtr->instanceOf<A>(): expected=1, value=" << a2dPtr->instanceOf<A>() << std::endl;
std::cout << "a2dPtr->instanceOf<B>(): expected=0, value=" << a2dPtr->instanceOf<B>() << std::endl;
std::cout << "a2dPtr->instanceOf<C>(): expected=1, value=" << a2dPtr->instanceOf<C>() << std::endl;
std::cout << "a2dPtr->instanceOf<D>(): expected=1, value=" << a2dPtr->instanceOf<D>() << std::endl;
delete a2aPtr;
delete a2bPtr;
delete c2dPtr;
return 0;
}
Output:
a2aPtr->instanceOf<A>(): expected=1, value=1
a2aPtr->instanceOf<B>(): expected=0, value=0
a2aPtr->instanceOf<C>(): expected=0, value=0
a2aPtr->instanceOf<D>(): expected=0, value=0
a2bPtr->instanceOf<A>(): expected=1, value=1
a2bPtr->instanceOf<B>(): expected=1, value=1
a2bPtr->instanceOf<C>(): expected=0, value=0
a2bPtr->instanceOf<D>(): expected=0, value=0
a2cPtr->instanceOf<A>(): expected=1, value=1
a2cPtr->instanceOf<B>(): expected=0, value=0
a2cPtr->instanceOf<C>(): expected=1, value=1
a2cPtr->instanceOf<D>(): expected=0, value=0
c2dPtr->instanceOf<A>(): expected=1, value=1
c2dPtr->instanceOf<B>(): expected=0, value=0
c2dPtr->instanceOf<C>(): expected=1, value=1
c2dPtr->instanceOf<D>(): expected=1, value=1
a2dPtr->instanceOf<A>(): expected=1, value=1
a2dPtr->instanceOf<B>(): expected=0, value=0
a2dPtr->instanceOf<C>(): expected=1, value=1
a2dPtr->instanceOf<D>(): expected=1, value=1
The most interesting question which now arises is, if this evil stuff is more efficient than the usage of dynamic_cast
. Therefore I've written a very basic performance measurement app.
InstanceOfPerformance.cpp
#include <chrono>
#include <iostream>
#include <string>
#include "DemoClassHierarchy.hpp"
template <typename Base, typename Derived, typename Duration>
Duration instanceOfMeasurement(unsigned _loopCycles) {
auto start = std::chrono::high_resolution_clock::now();
volatile bool isInstanceOf = false;
for (unsigned i = 0; i < _loopCycles; ++i) {
Base *ptr = new Derived;
isInstanceOf = ptr->template instanceOf<Derived>();
delete ptr;
}
auto end = std::chrono::high_resolution_clock::now();
return std::chrono::duration_cast<Duration>(end - start);
}
template <typename Base, typename Derived, typename Duration>
Duration dynamicCastMeasurement(unsigned _loopCycles) {
auto start = std::chrono::high_resolution_clock::now();
volatile bool isInstanceOf = false;
for (unsigned i = 0; i < _loopCycles; ++i) {
Base *ptr = new Derived;
isInstanceOf = dynamic_cast<Derived *>(ptr) != nullptr;
delete ptr;
}
auto end = std::chrono::high_resolution_clock::now();
return std::chrono::duration_cast<Duration>(end - start);
}
int main() {
unsigned testCycles = 10000000;
std::string unit = " us";
using DType = std::chrono::microseconds;
std::cout << "InstanceOf performance(A->D) : " << instanceOfMeasurement<A, D, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "InstanceOf performance(A->C) : " << instanceOfMeasurement<A, C, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "InstanceOf performance(A->B) : " << instanceOfMeasurement<A, B, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "InstanceOf performance(A->A) : " << instanceOfMeasurement<A, A, DType>(testCycles).count() << unit
<< "\n"
<< std::endl;
std::cout << "DynamicCast performance(A->D) : " << dynamicCastMeasurement<A, D, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "DynamicCast performance(A->C) : " << dynamicCastMeasurement<A, C, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "DynamicCast performance(A->B) : " << dynamicCastMeasurement<A, B, DType>(testCycles).count() << unit
<< std::endl;
std::cout << "DynamicCast performance(A->A) : " << dynamicCastMeasurement<A, A, DType>(testCycles).count() << unit
<< "\n"
<< std::endl;
return 0;
}
The results vary and are essentially based on the degree of compiler optimization. Compiling the performance measurement program using g++ -std=c++11 -O0 -o instanceof-performance InstanceOfPerformance.cpp
the output on my local machine was:
InstanceOf performance(A->D) : 699638 us
InstanceOf performance(A->C) : 642157 us
InstanceOf performance(A->B) : 671399 us
InstanceOf performance(A->A) : 626193 us
DynamicCast performance(A->D) : 754937 us
DynamicCast performance(A->C) : 706766 us
DynamicCast performance(A->B) : 751353 us
DynamicCast performance(A->A) : 676853 us
Mhm, this result was very sobering, because the timings demonstrates that the new approach is not much faster compared to the dynamic_cast
approach. It is even less efficient for the special test case which tests if a pointer of A
is an instance ofA
. BUT the tide turns by tuning our binary using compiler otpimization. The respective compiler command is g++ -std=c++11 -O3 -o instanceof-performance InstanceOfPerformance.cpp
. The result on my local machine was amazing:
InstanceOf performance(A->D) : 3035 us
InstanceOf performance(A->C) : 5030 us
InstanceOf performance(A->B) : 5250 us
InstanceOf performance(A->A) : 3021 us
DynamicCast performance(A->D) : 666903 us
DynamicCast performance(A->C) : 698567 us
DynamicCast performance(A->B) : 727368 us
DynamicCast performance(A->A) : 3098 us
If you are not reliant on multiple inheritance, are no opponent of good old C macros, RTTI and template metaprogramming and are not too lazy to add some small instructions to the classes of your class hierarchy, then this approach can boost your application a little bit with respect to its performance, if you often end up with checking the instance of a pointer. But use it with caution. There is no warranty for the correctness of this approach.
Note: All demos were compiled using clang (Apple LLVM version 9.0.0 (clang-900.0.39.2))
under macOS Sierra on a MacBook Pro Mid 2012.
Edit:
I've also tested the performance on a Linux machine using gcc (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 20160609
. On this platform the perfomance benefit was not so significant as on macOs with clang.
Output (without compiler optimization):
InstanceOf performance(A->D) : 390768 us
InstanceOf performance(A->C) : 333994 us
InstanceOf performance(A->B) : 334596 us
InstanceOf performance(A->A) : 300959 us
DynamicCast performance(A->D) : 331942 us
DynamicCast performance(A->C) : 303715 us
DynamicCast performance(A->B) : 400262 us
DynamicCast performance(A->A) : 324942 us
Output (with compiler optimization):
InstanceOf performance(A->D) : 209501 us
InstanceOf performance(A->C) : 208727 us
InstanceOf performance(A->B) : 207815 us
InstanceOf performance(A->A) : 197953 us
DynamicCast performance(A->D) : 259417 us
DynamicCast performance(A->C) : 256203 us
DynamicCast performance(A->B) : 261202 us
DynamicCast performance(A->A) : 193535 us