Let's say I have a large, between 50 and 200, pool of individual functions whose job it is to operate on a single object and modify it. The pool of functions is selectively put into a single array and arranged in an arbitrary order.
The functions themselves take no arguments outside of the values present within the object it is modifying, and in this way the object's behavior is determined only by which functions are executed and in what order.
A way I have tentatively used so far is this, which might explain better what my goal is:
class Behavior{
public:
virtual void act(Object * obj) = 0;
};
class SpecificBehavior : public Behavior{
// many classes like this exist
public:
void act(Object * obj){ /* do something specific with obj*/ };
};
class Object{
public:
std::list<Behavior*> behavior;
void behave(){
std::list<Behavior*>::iterator iter = behavior.front();
while(iter != behavior.end()){
iter->act(this);
++iter;
};
};
};
My Question is, what is the most efficient way in C++ of organizing such a pool of functions, in terms of performance and maintainability. This is for some A.I research I am doing, and this methodology is what most closely matches what I am trying to achieve.
edits: The array itself can be changed at any time by any other part of the code not listed here, but it's guaranteed to never change during the call to behave(). The array it is stored in needs to be able to change and expand to any size
If the behaviour functions have no state and only take one Object argument, then I'd go with a container of function objects:
#include <functional>
#include <vector>
typedef std::function<void(Object &)> BehaveFun;
typedef std::vector<BehaveFun> BehaviourCollection;
class Object {
BehaviourCollection b;
void behave() {
for (auto it = b.cbegin(); it != b.cend(); ++it) *it(*this);
}
};
Now you just need to load all your functions into the collection.
if the main thing you will be doing with this collection is iterating over it, you'll probably want to use a vector as dereferencing and incrementing your iterators will equate to simple pointer arithmetic.
If you want to use all your cores, and your operations do not share any state, you might want to have a look at a library like Intel's TBB (see the parallel_for example)
I'd keep it exactly as you have it.
Perofmance should be OK (there may be an extra indirection due to the vtable look up but that shouldn't matter.)
My reasons for keeping it as is are:
You might be able to lift common sub-behaviour into an intermediate class between Behaviour and your implementation classes. This is not as easy using function pointers.
struct AlsoWaveArmsBase : public Behaviour
{
void act( Object * obj )
{
start_waving_arms(obj); // Concrete call
do_other_action(obj); // Abstract call
end_waving_arms(obj); // Concrete call
}
void start_waving_arms(Object*obj);
void end_waving_arms(Object*obj);
virtual void do_other_actions(Object * obj)=0;
};
struct WaveAndWalk : public AlsoWaveArmsBase
{
void do_other_actions(Object * obj) { walk(obj); }
};
struct WaveAndDance : pubic AlsoWaveArmsBase
{
void do_other_actions(Object * obj) { walk(obj); }
}
You might want to use state in your behaviour
struct Count : public Behavior
{
Behaviour() : i(0) {}
int i;
void act(Object * obj)
{
count(obj,i);
++i;
}
}
You might want to add helper functions e.g. you might want to add a can_act like this:
void Object::behave(){
std::list<Behavior*>::iterator iter = behavior.front();
while(iter != behavior.end()){
if( iter->can_act(this) ){
iter->act(this);
}
++iter;
};
};
IMO, these flexibilities outweigh the benefits of moving to a pure function approach.
For maintainability, your current approach is the best (virtual functions). You might get a tiny little gain from using free function pointers, but I doubt it's measurable, and even if so, I don't think it is worth the trouble. The current OO approach is fast enough and maintainable. The little gain I'm talking about comes from the fact that you are dereferencing a pointer to an object and then (behind the scenes) dereferencing a pointer to a function (which happening as the implementation of calling a virtual function).
I wouldn't use std::function, because it's not very performant (though that might differ between implementations). See this and this. Function pointers are as fast as it gets when you need this kind of dynamism at runtime.
If you need to improve the performance, I suggest to look into improving the algorithm, not this implementation.
Related
I have a base class, BaseObject, and two derived class DerivedObject1 and DerivedObject2. They share a common behavior and methods, but DerivedObject1 has an additional method. My main class MyClass stores (in std::vector) boost::shared_ptr of instances of those classes. MyClass needs to call commonMethod() for all the BaseObject, and sometimes call additionalMethod() for all DerivedObject1.
class BaseObject
{
virtual void commonMethod();
}
Class DerivedObject1 : public BaseObject
{
void commonMethod();
void additionalMethod();
}
Class DerivedObject2 : public BaseObject
{
void commonMethod();
}
Are there any disadvantages of having two vectors in MyClass, one that stores ALL the pointers of DerivedObject1 and DerivedObject2, and another vector that stores only the pointers of DerivedObject1 ? Meaning I would have all the DerivedObject1 pointers twice. But I think the call to the different methods would be clear, at least.
class MyClass
{
typedef std::vector<std::shared_ptr<BaseObject>> BaseObjectVector;
typedef std::vector<std::shared_ptr<DerivedObject1>> DerivedObject1Vector;
BaseObjectVector everything;
DerivedObject1Vector only_derived1;
void doSomething()
{
for (BaseObjectVector::iterator iter = everything.begin(); iter != everything.end(); ++iter)
{
(*iter)->commonMethod();
}
}
void doSomethingForDerivedObject1()
{
for (DerivedObject1Vector::iterator iter = only_derived1.begin(); iter != only_derived1.end(); ++iter)
{
(*iter)->additionalMethod();
}
}
}
I can think of other ways to do this, mainly having one vector for DerivedObject1 and one vector for DerivedObject2, but to call commonMethod(), I would have to iterate over both vectors. My original solution seems the best to me, except that some pointers are stored twice. What are the disadvantages of this ?
Interesting question.
We sometimes meet such a situation in maintenance of legacy codes which we don’t know who wrote.
Are there any disadvantages of having two vectors in MyClass … ?
I think there are no mechanical (or performance) disadvantages.
If we are struggled to meet the release deadline, we have no choice but to pick such an simple way.
However, storing same vectors twice actually lowers maintainability and we should consider improving it in the future.
std::dynamic_pointer_cast (or boost::dynamic_pointer_cast) [Demo]
If you need dynamic_pointer_cast to implement functions like #Caleth ’s push_back / remove,
how about removing only_derived1 and reluctantly applying just one dynamic_pointer_cast in doSomethingForDerivedObject1() from the get-go as follows ?
It will make MyClass simpler. Required modification will not be complicated if DerivedObject3 is defined in the future.
void MyClass::doSomethingForDerivedObject1()
{
for (const auto& obj_i : everything)
{
if (auto derived1 = std::dynamic_pointer_cast<DerivedObject1>(obj_i))
{
derived1->additionalMethod();
}
}
}
void MyClass::doSomethingForDerivedObject3()
{
for (const auto& obj_i : everything)
{
if (auto derived3 = std::dynamic_pointer_cast<DerivedObject3>(obj_i))
{
derived3->additionalMethod();
}
}
}
Dummy Method as proposed by #sagi [Demo]
Declaring virtual function BaseObject::additionalMethod() and implementing
void DerivedObject2::additionalMethod()
{
/* nothing to do */
}
then you can again remove only_derived1.
In this method you must implement DerivedObject3::additionalMethod() only if DerivedObject3 is defined.
But, although it depends on your constructor or setter code, if the following case will also occur
everything; ->derived2
only_derived1; ->derived1
this method is still insufficient.
Changing Design [Demo]
Ideally, we should not use public inheritance to implement objects within "IS-ALMOST-A“ relations, as Herb Sutter says.
The relation between BaseObject, DerivedObject1 and DerivedObject2 looks like this one.
Since I don’t know whole code of your application I may be wrong, but it will be worthwhile to consider extracting DerivedObject1::additionalMethod() as another class or a function pointer and putting its vector in MyClass as a private member.
I can suggest this : Store everything in one array , and make a dummy additionalMethod() in DerivedObject2 . Then - just invoke additionalMethod for every object .
Alternativly :
They share a common behavior and methods, but DerivedObject1 has an additional method
Make DerivedObject1 inherit from DerivedObject2
Yes, your first method is good, the main problem is that you end up duplicating the public members of vector to ensure that the consistency of the Derived vector.
class MyClass
{
typedef std::vector<std::shared_ptr<BaseObject>> BaseObjectVector;
typedef std::vector<std::shared_ptr<DerivedObject1>> DerivedObject1Vector;
BaseObjectVector everything;
DerivedObject1Vector only_derived1;
public:
void push_back(shared_ptr<Base> ptr)
{
everything.push_back(ptr);
if (shared_ptr<Derived1> derived = dynamic_ptr_cast<Derived1>(ptr))
{
only_derived1.push_back(derived);
}
}
void remove(shared_ptr<Base> ptr)
{
base.remove(ptr);
only_derived1.remove(dynamic_ptr_cast<Derived1>(ptr));
}
// dozens more...
};
What you can instead do is provide a view of your bases using something like boost::range's adaptors
shared_ptr<Derived1> convert(shared_ptr<Base> ptr)
{
return dynamic_ptr_cast<Derived1>(ptr);
}
bool not_null(shared_ptr<Derived1> ptr)
{
return ptr.get();
}
boost::for_each(bases
| boost::adaptors::transformed(convert) // Base to Derived
| boost::adaptors::filtered(not_null) // Eliminate null
| boost::adaptors::indirected, // dereference
boost::bind(&Derived1::additionalMethod, boost::placeholders::_1));
This is probably very basic but somehow I cannot figure it out.
Say I have a class A which embeds 42 Things, plus some common data:
class A {
Thing things[42];
int common_data[1024];
}
I would like each thing to have access to the common data, but I don't want to copy the data in each Thing object, nor pay the price of a pointer to it in each thing. In other word, I would like Thing to look like this:
class Thing {
int ident;
int f() {
return common_data[ident];
}
}
Of course here common_data is unbound. What is the canonical way to make this work?
FWIW I am working with a subset of C++ with no dynamic allocation (no "new", no inheritance, basically it's C with the nice syntax to call methods and declare objects); I am ideally looking for a solution that fits in this subset.
You can solve your issue by making the common_data attribute of Class A static. Static variables are shared by all members of class A, and will be accessible if you make it public.
class A
{
private:
Thing things[42];
public:
static int common_data[1024];
}
It can be accessed by doing...
A::common_data[index];
I am not sure if I understand the question correctly, but maybe this helps:
struct A {
Thing things[42];
int common_data[1024];
void foo(int index) {
things[index].doSomeThingWithCommonData(int* common_data);
}
};
struct Thing {
void doSomeThinWithCommonData(int* common_data) {
/* now you have access to common_data */
}
};
Your reasons for avoiding pointers/reference is based on irrational fears. "Copying" a pointer 42 times is nothing (read this word carefully) for the machine. Moreover this is definitely not the bottleneck of the application.
So the idiomatic way is to simply use dependency injection, which is indeed a slightly more costly action for you (if passing an array can be considered costly), but allows for a much more decoupled design.
This is therefore the solution I recommend:
struct Thing {
using data = std::shared_ptr<std::array<int, 1024>>;
data common_data;
Thing(data arg)
: common_data(arg)
{}
// ...
};
If the system is costrained, then you should benchmark your program. I can tell you already with almost absolutely certainty that the bottleneck won't be the copying of those 42 pointers.
I have a ClientInterface class, that uses the Strategy pattern to organize two complex algorithms conforming to interfaces Abase and Bbase, respectively. The ClientInterface agglomerates (via composition) the data on which the algorithms operate, which needs to conform to the Data interface.
What I tried to do is to have a single ClientInterface class, which is able to choose different Strategies and Data implementations at run-time. The algorithms and data implementations are chosen using the Factory Method which reads the strings from an input file and selects the algorithm and data implementation in the ClientInterface constructor. The run-time choice of data and algorithms is not provided in the code model below.
The Data implementation can be based on a map, a list, an unordered_map , etc. to test how does the efficiency of two complex algorithms (Abase and Bbase implemented Strategies) change with different containers used for the Data.
Additionally, the Data agglomerates different Elements (ElementBase implementations). Different element implementations will also have significant impact on the efficiency of theh ClientInterface, but the Elements are really disjoint types with implementations coming from different libraries. I know this for a fact, since profiling the existing application shows the Element operation to be one of the bottlenecks.
I know that if I use polymorphism with containers, there is "boost/ptr_container" out there, but the Data will store hundreds of thousands, if not millions of Elements. Using polymorphism for Elements in this case will have a significant overhead on the ClientInterface, but if I choose to make data a class Template for the Element type, I will end up statically defining the ClientInterface class, which means producing a client application per each Element type at least.
Can I assume that for the same number of Elements and the ClientInterface configuration obtained at run-time, the overhead induced by the use of polymorphism for the Element type will have the same impact on all configurations of the Data and Algorithm implementations? In this case, I can run the automated tests, decide on the configuration of the Data implementation and the Element implementation, and define a statically configured EfficientClientInterface to be used in the productive code?
Goal: I have a test harness prepared, and what I am trying to do is to automatize the testing on the family of test cases, since changing the Algorithms and Elements at run-time, allows me to use a single application in a loop, which is configured at run-time and whose output is measured for efficiency. In the real implementation, I am dealing with at least 6 algorithm interfaces, 3-4 Data implementations, and I estimate 3 Element implementations at least.
So, my questions are:
1) How can an Element support different operations when overloading is not working for return types? If I make the operation a template, it needs to be defined at compile-time, which messes with my automated testing procedure.
2) How can I design this code better to achieve the goal?
3) Is there a better overall approach to this problem?
Here is the code model:
#include <iostream>
#include <memory>
class ElementOpResultFirst
{};
class ElementOpResultSecond
{};
class ElementBase
{
public:
// Overloading does not allow different representation of the solution for the element operation.
virtual ElementOpResultFirst elementOperation() = 0;
//virtual ElementOpResultSecond elementOperation() = 0;
};
class InterestingElement
:
public ElementBase
{
public:
ElementOpResultFirst elementOperation()
{
// Implementation dependant operation code.
return ElementOpResultFirst();
}
//ElementOpResultSecond elementOperation()
//{
//// Implementation dependant operation code.
//return ElementOpResultSecond();
//}
};
class EfficientElement
:
public ElementBase
{
public:
ElementOpResultFirst elementOperation()
{
// Implementation dependant operation code.
return ElementOpResultFirst();
}
//ElementOpResultSecond elementOperation()
//{
//// Implementation dependant operation code.
//return ElementOpResultSecond();
//}
};
class Data
{
public:
virtual void insertElement(const ElementBase&) = 0;
virtual const ElementBase& getElement(int key) = 0;
};
class DataConcreteMap
:
public Data
{
// Map implementation
public:
void insertElement(const ElementBase&)
{
// Insert element into the Map implementation.
}
const ElementBase& getElement(int key)
{
// Get element from the Map implementation.
}
};
class DataConcreteVector
:
public Data
{
// Vector implementation
public:
void insertElement(const ElementBase&)
{
// Insert element into the vector implementation.
}
const ElementBase& getElement(int key)
{
// Get element from the Vector implementation
}
};
class Abase
{
public:
virtual void aFunction() = 0;
};
class Aconcrete
:
public Abase
{
public:
virtual void aFunction()
{
std::cout << "Aconcrete::function() " << std::endl;
}
};
class Bbase
{
public:
virtual void bFunction(Data& data) = 0;
};
class Bconcrete
:
public Bbase
{
public:
virtual void bFunction(Data& data)
{
data.getElement(0);
std::cout << "Bconcrete::function() " << std::endl;
}
};
// Add a static abstract factory for algorithm and data generation.
class ClientInterface
{
std::unique_ptr<Data> data_;
std::unique_ptr<Abase> algorithmA_;
std::unique_ptr<Bbase> algorithmB_;
public:
ClientInterface()
:
// A Factory Method is defined for Data, Abase and Bbase that
// produces the concrete type based on an entry in a text-file.
data_ (std::unique_ptr<Data> (new DataConcreteMap())),
algorithmA_(std::unique_ptr<Abase> (new Aconcrete())),
algorithmB_(std::unique_ptr<Bbase> (new Bconcrete()))
{}
void aFunction()
{
return algorithmA_->aFunction();
}
void bFunction()
{
return algorithmB_->bFunction(*data_);
}
};
// Single client code: both for testing and final version.
int main()
{
ClientInterface cli;
cli.aFunction();
cli.bFunction();
return 0;
};
What I tried to do is to have a single ClientInterface class, which is
able to choose different Strategies and Data implementations at
run-time. The algorithms and data implementations are chosen using the
Factory Method which reads the strings from an input file and selects
the algorithm and data implementation in the ClientInterface
constructor. The run-time choice of data and algorithms is not
provided in the code model below.
Sounds like you have the basis for some of it here: Either just produce a set of files to test with that produce the right different sets of inputs. Or refactor the Factory function so that the reading of the file and the strings are separate, so you can call your factory function [internals] with a a string from the code.
Can I assume that for the same number of Elements and the
ClientInterface configuration obtained at run-time, the overhead
induced by the use of polymorphism for the Element type will have the
same impact on all configurations of the Data and Algorithm
implementations? In this case, I can run the automated tests, decide
on the configuration of the Data implementation and the Element
implementation, and define a statically configured
EfficientClientInterface to be used in the productive code?
I don't think you can make that assumption. Different implementations may well have different effects on the algorithms - copying a 100 byte string is significantly harder than copying a 4 byte integer, for example. So what the data is, and how it's organized will have some effect on the work you do. Of course, since you haven't described in much detail what your Elements actually contain, it's all guesswork.
1) How can an Element support different operations when overloading is
not working for return types? If I make the operation a template, it
needs to be defined at compile-time, which messes with my automated
testing procedure.
Make a factory class that returns an ElementBase reference or pointer? That's my immediate reaction to this question, but again, the detail in your question is sufficiently vague that it's hard to say for sure.
In the real application, how does this work? Is it done by templates, then you'd better implement the testcode by templates, and fill it out with a selection of realistic variations on what you think the real system is likely to do.
2) How can I design this code better to achieve the goal?
Try to reuse the production code?
3) Is there a better overall approach to this problem?
Not sure yet.
I've read through this article, and what I take from it is that when you want to call a pointer to a member function, you need an instance (either a pointer to one or a stack-reference) and call it so:
(instance.*mem_func_ptr)(..)
or
(instance->*mem_func_ptr)(..)
My question is based on this: since you have the instance, why not call the member function directly, like so:
instance.mem_func(..) //or: instance->mem_func(..)
What is the rational/practical use of pointers to member functions?
[edit]
I'm playing with X-development & reached the stage where I am implementing widgets; the event-loop-thread for translating the X-events to my classes & widgets needs to start threads for each widget/window when an event for them arrives; to do this properly I thought I needed function-pointers to the event-handlers in my classes.
Not so: what I did discover was that I could do the same thing in a much clearer & neater way by simply using a virtual base class. No need whatsoever for pointers to member-functions. It was while developing the above that the doubt about the practical usability/meaning of pointers to member-functions arose.
The simple fact that you need a reference to an instance in order to use the member-function-pointer, obsoletes the need for one.
[edit - #sbi & others]
Here is a sample program to illustrate my point:
(Note specifically 'Handle_THREE()')
#include <iostream>
#include <string>
#include <map>
//-----------------------------------------------------------------------------
class Base
{
public:
~Base() {}
virtual void Handler(std::string sItem) = 0;
};
//-----------------------------------------------------------------------------
typedef void (Base::*memfunc)(std::string);
//-----------------------------------------------------------------------------
class Paper : public Base
{
public:
Paper() {}
~Paper() {}
virtual void Handler(std::string sItem) { std::cout << "Handling paper\n"; }
};
//-----------------------------------------------------------------------------
class Wood : public Base
{
public:
Wood() {}
~Wood() {}
virtual void Handler(std::string sItem) { std::cout << "Handling wood\n"; }
};
//-----------------------------------------------------------------------------
class Glass : public Base
{
public:
Glass() {}
~Glass() {}
virtual void Handler(std::string sItem) { std::cout << "Handling glass\n"; }
};
//-----------------------------------------------------------------------------
std::map< std::string, memfunc > handlers;
void AddHandler(std::string sItem, memfunc f) { handlers[sItem] = f; }
//-----------------------------------------------------------------------------
std::map< Base*, memfunc > available_ONE;
void AddAvailable_ONE(Base *p, memfunc f) { available_ONE[p] = f; }
//-----------------------------------------------------------------------------
std::map< std::string, Base* > available_TWO;
void AddAvailable_TWO(std::string sItem, Base *p) { available_TWO[sItem] = p; }
//-----------------------------------------------------------------------------
void Handle_ONE(std::string sItem)
{
memfunc f = handlers[sItem];
if (f)
{
std::map< Base*, memfunc >::iterator it;
Base *inst = NULL;
for (it=available_ONE.begin(); ((it != available_ONE.end()) && (inst==NULL)); it++)
{
if (it->second == f) inst = it->first;
}
if (inst) (inst->*f)(sItem);
else std::cout << "No instance of handler for: " << sItem << "\n";
}
else std::cout << "No handler for: " << sItem << "\n";
}
//-----------------------------------------------------------------------------
void Handle_TWO(std::string sItem)
{
memfunc f = handlers[sItem];
if (f)
{
Base *inst = available_TWO[sItem];
if (inst) (inst->*f)(sItem);
else std::cout << "No instance of handler for: " << sItem << "\n";
}
else std::cout << "No handler for: " << sItem << "\n";
}
//-----------------------------------------------------------------------------
void Handle_THREE(std::string sItem)
{
Base *inst = available_TWO[sItem];
if (inst) inst->Handler(sItem);
else std::cout << "No handler for: " << sItem << "\n";
}
//-----------------------------------------------------------------------------
int main()
{
Paper p;
Wood w;
Glass g;
AddHandler("Paper", (memfunc)(&Paper::Handler));
AddHandler("Wood", (memfunc)(&Wood::Handler));
AddHandler("Glass", (memfunc)(&Glass::Handler));
AddAvailable_ONE(&p, (memfunc)(&Paper::Handler));
AddAvailable_ONE(&g, (memfunc)(&Glass::Handler));
AddAvailable_TWO("Paper", &p);
AddAvailable_TWO("Glass", &g);
std::cout << "\nONE: (bug due to member-function address being relative to instance address)\n";
Handle_ONE("Paper");
Handle_ONE("Wood");
Handle_ONE("Glass");
Handle_ONE("Iron");
std::cout << "\nTWO:\n";
Handle_TWO("Paper");
Handle_TWO("Wood");
Handle_TWO("Glass");
Handle_TWO("Iron");
std::cout << "\nTHREE:\n";
Handle_THREE("Paper");
Handle_THREE("Wood");
Handle_THREE("Glass");
Handle_THREE("Iron");
}
{edit] Potential problem with direct-call in above example:
In Handler_THREE() the name of the method must be hard-coded, forcing changes to be made anywhere that it is used, to apply any change to the method. Using a pointer to member-function the only additional change to be made is where the pointer is created.
[edit] Practical uses gleaned from the answers:
From answer by Chubsdad:
What: A dedicated 'Caller'-function is used to invoke the mem-func-ptr;Benefit: To protect code using function(s) provided by other objectsHow: If the particular function(s) are used in many places and the name and/or parameters change, then you only need to change the name where it is allocated as pointer, and adapt the call in the 'Caller'-function. (If the function is used as instance.function() then it must be changed everywhere.)
From answer by Matthew Flaschen:
What: Local specialization in a classBenefit: Makes the code much clearer,simpler and easier to use and maintainHow: Replaces code that would conventionally be implement using complex logic with (potentially) large switch()/if-then statements with direct pointers to the specialization; fairly similar to the 'Caller'-function above.
The same reason you use any function pointer: You can use arbitrary program logic to set the function pointer variable before calling it. You could use a switch, an if/else, pass it into a function, whatever.
EDIT:
The example in the question does show that you can sometimes use virtual functions as an alternative to pointers to member functions. This shouldn't be surprising, because there are usually multiple approaches in programming.
Here's an example of a case where virtual functions probably don't make sense. Like the code in the OP, this is meant to illustrate, not to be particularly realistic. It shows a class with public test functions. These use internal, private, functions. The internal functions can only be called after a setup, and a teardown must be called afterwards.
#include <iostream>
class MemberDemo;
typedef void (MemberDemo::*MemberDemoPtr)();
class MemberDemo
{
public:
void test1();
void test2();
private:
void test1_internal();
void test2_internal();
void do_with_setup_teardown(MemberDemoPtr p);
};
void MemberDemo::test1()
{
do_with_setup_teardown(&MemberDemo::test1_internal);
}
void MemberDemo::test2()
{
do_with_setup_teardown(&MemberDemo::test2_internal);
}
void MemberDemo::test1_internal()
{
std::cout << "Test1" << std::endl;
}
void MemberDemo::test2_internal()
{
std::cout << "Test2" << std::endl;
}
void MemberDemo::do_with_setup_teardown(MemberDemoPtr mem_ptr)
{
std::cout << "Setup" << std::endl;
(this->*mem_ptr)();
std::cout << "Teardown" << std::endl;
}
int main()
{
MemberDemo m;
m.test1();
m.test2();
}
My question is based on this: since you have the instance, why not call the member function directly[?]
Upfront: In more than 15 years of C++ programming, I have used members pointers maybe twice or thrice. With virtual functions being around, there's not all that much use for it.
You would use them if you want to call a certain member functions on an object (or many objects) and you have to decide which member function to call before you can find out for which object(s) to call it on. Here is an example of someone wanting to do this.
I find the real usefulness of pointers to member functions comes when you look at a higher level construct such as boost::bind(). This will let you wrap a function call as an object that can be bound to a specific object instance later on and then passed around as a copyable object. This is a really powerful idiom that allows for deferred callbacks, delegates and sophisticated predicate operations. See my previous post for some examples:
https://stackoverflow.com/questions/1596139/hidden-features-and-dark-corners-of-stl/1596626#1596626
Member functions, like many function pointers, act as callbacks. You could manage without them by creating some abstract class that calls your method, but this can be a lot of extra work.
One common use is algorithms. In std::for_each, we may want to call a member function of the class of each member of our collection. We also may want to call the member function of our own class on each member of the collection - the latter requires boost::bind to achieve, the former can be done with the STL mem_fun family of classes (if we don't have a collection of shared_ptr, in which case we need to boost::bind in this case too). We could also use a member function as a predicate in certain lookup or sort algorithms. (This removes our need to write a custom class that overloads operator() to call a member of our class, we just pass it in directly to boost::bind).
The other use, as I mentioned, are callbacks, often in event-driven code. When an operation has completed we want a method of our class called to handle the completion. This can often be wrapped into a boost::bind functor. In this case we have to be very careful to manage the lifetime of these objects correctly and their thread-safety (especially as it can be very hard to debug if something goes wrong). Still, it once again can save us from writing large amounts of "wrapper" code.
There are many practical uses. One that comes to my mind is as follows:
Assume a core function such as below (suitably defined myfoo and MFN)
void dosomething(myfoo &m, MFN f){ // m could also be passed by reference to
// const
m.*f();
}
Such a function in the presence of pointer to member functions, becomes open for extension and closed for modification (OCP)
Also refer to Safe bool idiom which smartly uses pointer to members.
The best use of pointers to member functions is to break dependencies.
Good example where pointer to member function is needed is Subscriber/Publisher pattern :
http://en.wikipedia.org/wiki/Publish/subscribe
In my opinion, member function pointers do are not terribly useful to the average programmer in their raw form. OTOH, constructs like ::std::tr1::function that wrap member function pointers together with a pointer to the object they're supposed to operate on are extremely useful.
Of course ::std::tr1::function is very complex. So I will give you a simple example that you wouldn't actually use in practice if you had ::std::tr1::function available:
// Button.hpp
#include <memory>
class Button {
public:
Button(/* stuff */) : hdlr_(0), myhandler_(false) { }
~Button() {
// stuff
if (myhandler_) {
delete hdlr_;
}
}
class PressedHandler {
public:
virtual ~PressedHandler() = 0;
virtual void buttonPushed(Button *button) = 0;
};
// ... lots of stuff
// This stores a pointer to the handler, but will not manage the
// storage. You are responsible for making sure the handler stays
// around as long as the Button object.
void setHandler(const PressedHandler &hdlr) {
hdlr_ = &hdlr;
myhandler_ = false;
}
// This stores a pointer to an object that Button does not manage. You
// are responsible for making sure this object stays around until Button
// goes away.
template <class T>
inline void setHandlerFunc(T &dest, void (T::*pushed)(Button *));
private:
const PressedHandler *hdlr_;
bool myhandler_;
template <class T>
class PressedHandlerT : public Button::PressedHandler {
public:
typedef void (T::*hdlrfuncptr_t)(Button *);
PressedHandlerT(T *ob, hdlrfuncptr_t hdlr) : ob_(ob), func_(hdlr) { }
virtual ~PressedHandlerT() {}
virtual void buttonPushed(Button *button) { (ob_->*func_)(button); }
private:
T * const ob_;
const hdlrfuncptr_t func_;
};
};
template <class T>
inline void Button::setHandlerFunc(T &dest, void (T::*pushed)(Button *))
{
PressedHandler *newhandler = new PressedHandlerT<T>(&dest, pushed);
if (myhandler_) {
delete hdlr_;
}
hdlr_ = newhandler;
myhandler_ = true;
}
// UseButton.cpp
#include "Button.hpp"
#include <memory>
class NoiseMaker {
public:
NoiseMaker();
void squee(Button *b);
void hiss(Button *b);
void boo(Button *b);
private:
typedef ::std::auto_ptr<Button> buttonptr_t;
const buttonptr_t squeebutton_, hissbutton_, boobutton_;
};
NoiseMaker::NoiseMaker()
: squeebutton_(new Button), hissbutton_(new Button), boobutton_(new Button)
{
squeebutton_->setHandlerFunc(*this, &NoiseMaker::squee);
hissbutton_->setHandlerFunc(*this, &NoiseMaker::hiss);
boobutton_->setHandlerFunc(*this, &NoiseMaker::boo);
}
Assuming Button is in a library and not alterable by you, I would enjoy seeing you implement that cleanly using a virtual base class without resorting to a switch or if else if construct somewhere.
The whole point of pointers of pointer-to-member function type is that they act as a run-time way to reference a specific method. When you use the "usual" syntax for method access
object.method();
pointer->method();
the method part is a fixed, compile-time specification of the method you want to call. It is hardcoded into your program. It can never change. But by using a pointer of pointer-to-member function type you can replace that fixed part with a variable, changeable at run-time specification of the method.
To better illustrate this, let me make the following simple analogy. Let's say you have an array
int a[100];
You can access its elements with fixed compile-time index
a[5]; a[8]; a[23];
In this case the specific indices are hardcoded into your program. But you can also access array's elements with a run-time index - an integer variable i
a[i];
the value of i is not fixed, it can change at run-time, thus allowing you to select different elements of the array at run-time. That is very similar to what pointers of pointer-to-member function type let you do.
The question you are asking ("since you have the instance, why not call the member function directly") can be translated into this array context. You are basically asking: "Why do we need a variable index access a[i], when we have direct compile-time constant access like a[1] and a[3]?" I hope you know the answer to this question and realize the value of run-time selection of specific array element.
The same applies to pointers of pointer-to-member function type: they, again, let you to perform run-time selection of a specific class method.
The use case is that you have several member methods with the same signature, and you want to build logic which one should be called under given circumstances. This can be helpful to implement state machine algorithms.
Not something you use everyday...
Imagine for a second you have a function that could call one of several different functions depending on parameters passed.
You could use a giant if/else if statement
You could use a switch statement
Or you could use a table of function pointers (a jump table)
If you have a lot of different options the jump table can be a much cleaner way of arranging your code ...
Its down to personal preference though. Switch statement and jump table correspond to more or less the same compiled code anyway :)
Member pointers + templates = pure win.
e.g. How to tell if class contains a certain member function in compile time
or
template<typename TContainer,
typename TProperty,
typename TElement = decltype(*Container().begin())>
TProperty grand_total(TContainer& items, TProperty (TElement::*property)() const)
{
TProperty accum = 0;
for( auto it = items.begin(), end = items.end(); it != end; ++it) {
accum += (it->*property)();
}
return accum;
}
auto ship_count = grand_total(invoice->lineItems, &LineItem::get_quantity);
auto sub_total = grand_total(invoice->lineItems, &LineItem::get_extended_total);
auto sales_tax = grand_total(invoice->lineItems, &LineItem::calculate_tax);
To invoke it, you need a reference to an instance, but then you can call the func direct & don't need a pointer to it.
This is completely missing the point. There are two indepedent concerns here:
what action to take at some later point in time
what object to perform that action on
Having a reference to an instance satisfies the second requirement. Pointers to member functions address the first: they are a very direct way to record - at one point in a program's execution - which action should be taken at some later stage of execution, possibly by another part of the program.
EXAMPLE
Say you have a monkey that can kiss people or tickle them. At 6pm, your program should set the monkey loose, and knows whom the monkey should visit, but around 3pm your user will type in which action should be taken.
A beginner's approach
So, at 3pm you could set a variable "enum Action { Kiss, Tickle } action;", then at 6pm you could do something like "if (action == Kiss) monkey->kiss(person); else monkey->tickle(person)".
Issues
But that introducing an extra level of encoding (the Action type's introduced to support this - built in types could be used but would be more error prone and less inherently meaningful). Then - after having worked out what action should be taken at 3pm, at 6pm you have to redundantly consult that encoded value to decide which action to take, which will require another if/else or switch upon the encoded value. It's all clumsy, verbose, slow and error prone.
Member function pointers
A better way is to use a more specialised varibale - a member function pointer - that directly records which action to perform at 6pm. That's what a member function pointer is. It's a kiss-or-tickle selector that's set earlier, creating a "state" for the monkey - is it a tickler or a kisser - which can be used later. The later code just invokes whatever function's been set without having to think about the possibilities or have any if/else-if or switch statements.
To invoke it, you need a reference to an instance, but then you can call the func direct & don't need a pointer to it.
Back to this. So, this is good if you make the decision about which action to take at compile time (i.e. a point X in your program, it'll definitely be a tickle). Function pointers are for when you're not sure, and want to decouple the setting of actions from the invocation of those actions.
In Java, you can have a List of Objects. You can add objects of multiple types, then retrieve them, check their type, and perform the appropriate action for that type.
For example: (apologies if the code isn't exactly correct, I'm going from memory)
List<Object> list = new LinkedList<Object>();
list.add("Hello World!");
list.add(7);
list.add(true);
for (object o : list)
{
if (o instanceof int)
; // Do stuff if it's an int
else if (o instanceof String)
; // Do stuff if it's a string
else if (o instanceof boolean)
; // Do stuff if it's a boolean
}
What's the best way to replicate this behavior in C++?
boost::variant is similar to dirkgently's suggestion of boost::any, but supports the Visitor pattern, meaning it's easier to add type-specific code later. Also, it allocates values on the stack rather than using dynamic allocation, leading to slightly more efficient code.
EDIT: As litb points out in the comments, using variant instead of any means you can only hold values from one of a prespecified list of types. This is often a strength, though it might be a weakness in the asker's case.
Here is an example (not using the Visitor pattern though):
#include <vector>
#include <string>
#include <boost/variant.hpp>
using namespace std;
using namespace boost;
...
vector<variant<int, string, bool> > v;
for (int i = 0; i < v.size(); ++i) {
if (int* pi = get<int>(v[i])) {
// Do stuff with *pi
} else if (string* si = get<string>(v[i])) {
// Do stuff with *si
} else if (bool* bi = get<bool>(v[i])) {
// Do stuff with *bi
}
}
(And yes, you should technically use vector<T>::size_type instead of int for i's type, and you should technically use vector<T>::iterator instead anyway, but I'm trying to keep it simple.)
Your example using Boost.Variant and a visitor:
#include <string>
#include <list>
#include <boost/variant.hpp>
#include <boost/foreach.hpp>
using namespace std;
using namespace boost;
typedef variant<string, int, bool> object;
struct vis : public static_visitor<>
{
void operator() (string s) const { /* do string stuff */ }
void operator() (int i) const { /* do int stuff */ }
void operator() (bool b) const { /* do bool stuff */ }
};
int main()
{
list<object> List;
List.push_back("Hello World!");
List.push_back(7);
List.push_back(true);
BOOST_FOREACH (object& o, List) {
apply_visitor(vis(), o);
}
return 0;
}
One good thing about using this technique is that if, later on, you add another type to the variant and you forget to modify a visitor to include that type, it will not compile. You have to support every possible case. Whereas, if you use a switch or cascading if statements, it's easy to forget to make the change everywhere and introduce a bug.
C++ does not support heterogenous containers.
If you are not going to use boost the hack is to create a dummy class and have all the different classes derive from this dummy class. Create a container of your choice to hold dummy class objects and you are ready to go.
class Dummy {
virtual void whoami() = 0;
};
class Lizard : public Dummy {
virtual void whoami() { std::cout << "I'm a lizard!\n"; }
};
class Transporter : public Dummy {
virtual void whoami() { std::cout << "I'm Jason Statham!\n"; }
};
int main() {
std::list<Dummy*> hateList;
hateList.insert(new Transporter());
hateList.insert(new Lizard());
std::for_each(hateList.begin(), hateList.end(),
std::mem_fun(&Dummy::whoami));
// yes, I'm leaking memory, but that's besides the point
}
If you are going to use boost you can try boost::any. Here is an example of using boost::any.
You may find this excellent article by two leading C++ experts of interest.
Now, boost::variant is another thing to look out for as j_random_hacker mentioned. So, here's a comparison to get a fair idea of what to use.
With a boost::variant the code above would look something like this:
class Lizard {
void whoami() { std::cout << "I'm a lizard!\n"; }
};
class Transporter {
void whoami() { std::cout << "I'm Jason Statham!\n"; }
};
int main() {
std::vector< boost::variant<Lizard, Transporter> > hateList;
hateList.push_back(Lizard());
hateList.push_back(Transporter());
std::for_each(hateList.begin(), hateList.end(), std::mem_fun(&Dummy::whoami));
}
How often is that sort of thing actually useful? I've been programming in C++ for quite a few years, on different projects, and have never actually wanted a heterogenous container. It may be common in Java for some reason (I have much less Java experience), but for any given use of it in a Java project there might be a way to do something different that will work better in C++.
C++ has a heavier emphasis on type safety than Java, and this is very type-unsafe.
That said, if the objects have nothing in common, why are you storing them together?
If they do have things in common, you can make a class for them to inherit from; alternately, use boost::any. If they inherit, have virtual functions to call, or use dynamic_cast<> if you really have to.
I'd just like to point out that using dynamic type casting in order to branch based on type often hints at flaws in the architecture. Most times you can achieve the same effect using virtual functions:
class MyData
{
public:
// base classes of polymorphic types should have a virtual destructor
virtual ~MyData() {}
// hand off to protected implementation in derived classes
void DoSomething() { this->OnDoSomething(); }
protected:
// abstract, force implementation in derived classes
virtual void OnDoSomething() = 0;
};
class MyIntData : public MyData
{
protected:
// do something to int data
virtual void OnDoSomething() { ... }
private:
int data;
};
class MyComplexData : public MyData
{
protected:
// do something to Complex data
virtual void OnDoSomething() { ... }
private:
Complex data;
};
void main()
{
// alloc data objects
MyData* myData[ 2 ] =
{
new MyIntData()
, new MyComplexData()
};
// process data objects
for ( int i = 0; i < 2; ++i ) // for each data object
{
myData[ i ]->DoSomething(); // no type cast needed
}
// delete data objects
delete myData[0];
delete myData[1];
};
Sadly there is no easy way of doing this in C++. You have to create a base class yourself and derive all other classes from this class. Create a vector of base class pointers and then use dynamic_cast (which comes with its own runtime overhead) to find the actual type.
Just for completeness of this topic I want to mention that you can actually do this with pure C by using void* and then casting it into whatever it has to be (ok, my example isn't pure C since it uses vectors but that saves me some code). This will work if you know what type your objects are, or if you store a field somewhere which remembers that. You most certainly DON'T want to do this but here is an example to show that it's possible:
#include <iostream>
#include <vector>
using namespace std;
int main() {
int a = 4;
string str = "hello";
vector<void*> list;
list.push_back( (void*) &a );
list.push_back( (void*) &str );
cout << * (int*) list[0] << "\t" << * (string*) list[1] << endl;
return 0;
}
While you cannot store primitive types in containers, you can create primitive type wrapper classes which will be similar to Java's autoboxed primitive types (in your example the primitive typed literals are actually being autoboxed); instances of which appear in C++ code (and can (almost) be used) just like primitive variables/data members.
See Object Wrappers for the Built-In Types from Data Structures and Algorithms with Object-Oriented Design Patterns in C++.
With the wrapped object you can use the c++ typeid() operator to compare the type.
I am pretty sure the following comparison will work:
if (typeid(o) == typeid(Int)) [where Int would be the wrapped class for the int primitive type, etc...]
(otherwise simply add a function to your primitive wrappers that returns a typeid and thus:
if (o.get_typeid() == typeid(Int)) ...
That being said, with respect to your example, this has code smell to me.
Unless this is the only place where you are checking the type of the object,
I would be inclined to use polymorphism (especially if you have other methods/functions specific with respect to type). In this case I would use the primitive wrappers adding an interfaced class declaring the deferred method (for doing 'do stuff') that would be implemented by each of your wrapped primitive classes. With this you would be able to use your container iterator and eliminate your if statement (again, if you only have this one comparison of type, setting up the deferred method using polymorphism just for this would be overkill).
I am a fairly inexperienced, but here's what I'd go with-
Create a base class for all classes you need to manipulate.
Write container class/ reuse container class.
(Revised after seeing other answers -My previous point was too cryptic.)
Write similar code.
I am sure a much better solution is possible. I am also sure a better explanation is possible. I've learnt that I have some bad C++ programming habits, so I've tried to convey my idea without getting into code.
I hope this helps.
Beside the fact, as most have pointed out, you can't do that, or more importantly, more than likely, you really don't want to.
Let's dismiss your example, and consider something closer to a real-life example. Specifically, some code I saw in a real open-source project. It attempted to emulate a cpu in a character array. Hence it would put into the array a one byte "op code", followed by 0, 1 or 2 bytes which could be a character, an integer, or a pointer to a string, based on the op code. To handle that, it involved a lot of bit-fiddling.
My simple solution: 4 separate stacks<>s: One for the "opcode" enum and one each for chars, ints and string. Take the next off the opcode stack, and the would take you which of the other three to get the operand.
There's a very good chance your actual problem can be handled in a similar way.
Well, you could create a base class and then create classes which inherit from it. Then, store them in a std::vector.
The short answer is... you can't.
The long answer is... you'd have to define your own new heirarchy of objects that all inherit from a base object. In Java all objects ultimately descend from "Object", which is what allows you to do this.
RTTI (Run time type info) in C++ has always been tough, especially cross-compiler.
You're best option is to use STL and define an interface in order to determine the object type:
public class IThing
{
virtual bool isA(const char* typeName);
}
void myFunc()
{
std::vector<IThing> things;
// ...
things.add(new FrogThing());
things.add(new LizardThing());
// ...
for (int i = 0; i < things.length(); i++)
{
IThing* pThing = things[i];
if (pThing->isA("lizard"))
{
// do this
}
// etc
}
}
Mike