NSInvocation in Objective-C++

UPDATE: This article was initially about some difficulties that I had calling some Objective-C++ code from C++ in a dynamic way using NSInvocation, which failed in the first place. I switched to directly calling the IMP pointer instead while passing C++ object pointers as arguments. Thanks to the comments that I received from bob, an obviously very experienced Objective-C++ developer, I could change the code and make it work also with NSInvocation.

Here is a C++ class that represents the details of a notifier/observer association called Observation:

    class Observation: public Object {
    public:

        void setObjcObserver(NSObject *theObserver) { _nsObserver = theObserver; }
        void setObjcSlot(SEL theSlot) { _nsSlot = theSlot; }
                
        void notify(bool complete);
        
        Observation() : _nsObserver(nil), _nsSlot(nil) {}
        
    private:

        Object *_notifier;
        NSObject *_nsObserver;
        SEL _nsSlot;
    };

First Attempt Using NSInvocation Failed

    void Observation::notify(bool complete) {

        NSMethodSignature *aSignature = [[_nsObserver class]
                        instanceMethodSignatureForSelector:_nsSlot];

        assert(aSignature);
        assert([_nsObserver respondsToSelector:_nsSlot]);

        NSInvocation *anInvocation = [NSInvocation invocationWithMethodSignature:aSignature];
        [anInvocation setSelector:_nsSlot];
        [anInvocation setArgument:this atIndex:2];
        [anInvocation setArgument:&complete atIndex:3];
            
        [anInvocation invokeWithTarget:_nsObserver];
    }

this code compiled fine but it crashed as soon as I tried to access members of the received Observation object in Objective-C.

IMP Based Solution

My approach to solve the issue is that I am directly calling the IMP of the method in the observer object. For this, I slightly change the C++ Observation model to cache the IMP pointer:

    class Observation: public Object {
    public:

        void setObjcObserver(NSObject *theObserver) { _nsObserver = theObserver; }
        void setObjcSlot(SEL theSlot) { _nsSlot = theSlot; }
                
        void notify(bool complete);
        
        Observation() : _nsObserver(nil), _nsSlot(nil), _nsSlotImpl(nil) {}
        
    private:

        Object *_notifier;
        NSObject *_nsObserver;
        SEL _nsSlot;

        typedef void (*slotImpl_t)(__strong id, SEL, lang::Observation *, BOOL);
        
        // used to cache the IMP of the objcSlot
        slotImpl_t _nsSlotImpl;

    };

Here, I simply add a typedef of the desired observer method signature called “slotImpl_t” and a member to cache the IMP function pointer. The cache would not be necessary but in my case the observer IMP will not change, so I cache it in order to save the look-up on frequent execution of the code. The notify() method now looks different:

    void Observation::notify(bool complete) {
        if(!_nsSlotImpl) {
            _nsSlotImpl = (slotImpl_t)[_nsObserver methodForSelector:_nsSlot];
            assert(_nsSlotImpl);
        }
        _nsSlotImpl(_nsObserver, _nsSlot, this, complete);
    }

This is working like a charm now. The address of the Observation object at the receiving end is correct as it should be, even in case of multiple-inheritance combined with virtual inheritance.

Correct Solution with NSInvocation

It turned out that in my original solution using NSInvocation I did it wrong. NSInvocation needs the address of the this pointer and not the this pointer itself. To make it worse, when simply doing

        [anInvocation setArgument:&this atIndex:2];

the compiler complains with the error:

Address expression must be an lvalue or a function designator

So, instead of &this it has to be an lvalue which essentially requires to copy the this pointer to a temporary variable, ‘temp’ in this case. Here is the complete and working code:

    void Observation::notify(bool complete) {

        NSMethodSignature *aSignature = [[_nsObserver class]
                        instanceMethodSignatureForSelector:_nsSlot];

        assert(aSignature);
        assert([_nsObserver respondsToSelector:_nsSlot]);

        NSInvocation *anInvocation = [NSInvocation invocationWithMethodSignature:aSignature];
        [anInvocation setSelector:_nsSlot];
        Observation *temp = this;
        [anInvocation setArgument:&temp atIndex:2];
        [anInvocation setArgument:&complete atIndex:3];
            
        [anInvocation invokeWithTarget:_nsObserver];
    }

Conclusion

There is expert knowlegde about Objective-C++ available on the planet. But it’s not easy to get access to it. Sometimes, it helps to blog about a problem and the solution will come and find you;) Again, bob, thank you very much

Java Like Programming in C++

C++ is a nice and very flexible language, but this comes at the cost that it forces you to think about many programming details before you can even think about solving your actual problem. Examples would be:

  • object oriented or generic programming
  • when to use references, value types and pointers
  • memory management rules
  • casting rules, const correctness, virtual methods
  • STL? boost?

This list can become quite endless. Being in the same boat as every other C++ programmer but having grasped some of the look&feel of other languages like Java and Objective-C (in its Cocoa incarnation) or even Qt (as a good C++ OO style example), I am continuously and unconsciously thinking about the pros and cons of each of them and recently I thought: wouldn’t it be nice to be able to program in C++ but make it look like Java?

Sure it would! Even it it were only just for fun…

So, I came up with a working code sample which looks quite like Java but is actually C++. I show you the code sample first before I elaborate on any details. First come some classes which you can consider the “framework”:

#include <iostream>
#include <sstream>
#include <assert.h>
#include <vector>


size_t globalRefCount = 0;

class Object;
class String;
std::ostream &operator<<(std::ostream&, Object&);
std::ostream &operator<<(std::ostream &, const String&);



class Object {    
protected:
    
    struct Impl {
        size_t _refCount;
        Impl() : _refCount(0) {}
        virtual ~Impl() {}
    } *data;
    
    inline void retain() {
        if(data) {
            data->_refCount++;
            globalRefCount++;
        }
    }
    inline void release() {
        if(data) {
            data->_refCount--;
            globalRefCount--;
            if(data->_refCount == 0) {
                delete data;
                data = 0;
            }
        }
    }
    
    Object(): data(new Impl) {
        retain();
    }
    Object(Impl *imp) : data(imp) {
        retain();
    }
    Object(const Object& other): data(0) {
        operator=(other);
    }
    
public:
    virtual ~Object() {
        release();
    }
    virtual const char *type() const { return "Object"; }
    
    void operator=(const Object& other) {
        if(data!=other.data) {
            release();
            data = other.data;
            retain();
        }
    }
    
    // make a heap clone of this object for usage in containers
    virtual Object *clone() const {
        return new Object(*this);
    }
    
    virtual String toString() const;
        
};




class String : public Object {
protected:
    struct Impl: public Object::Impl {
        std::string str;
    };
public:
    String(): Object(new Impl) {}
    String(const char *s): Object(new Impl) {
        static_cast<Impl*>(data)->str = s;
    }
    String(const String &other): Object(other) {}
    String operator+(const String& s) const {
        String result;
        static_cast<Impl*>(result.data)->str = static_cast<Impl*>(data)->str;
        static_cast<Impl*>(result.data)->str += static_cast<Impl*>(s.data)->str;
        return result;
    }
    String operator+(const Object& s) const {
        String result;
        static_cast<Impl*>(result.data)->str = static_cast<Impl*>(data)->str;
        static_cast<Impl*>(result.data)->str += static_cast<Impl*>(s.toString().data)->str;
        return result;
    }
    String operator+(long l) const {
        std::ostringstream oss;
        oss << static_cast<Impl*>(data)->str << l;
        String result;
        static_cast<Impl*>(result.data)->str = oss.str();
        return result;
    }
    const char *c_str() const {
        return static_cast<Impl*>(data)->str.c_str();
    }
    bool operator==(const String &other) const {
        return static_cast<Impl*>(data)->str == static_cast<Impl*>(other.data)->str;
    }
    
    // must have's
    const char *type() const { return "String"; }
    Object *clone() const { return new String(*this); }
    
    String toString() const {
        return *this;
    }
};

std::ostream &operator<<(std::ostream &os, const String& s) {
    os << s.c_str();
    return os;
}

String Object::toString() const {
    std::ostringstream os;
    os << this->type() << "@" << (void *)this << "[" << (data ? data->_refCount : 0) << "]";
    return String(os.str().c_str());
}

std::ostream &operator<<(std::ostream &os, Object& o) {
    os << o.toString().c_str();
    return os;
}




class ClassCastException: public Object {
    struct Impl: public Object::Impl {
        String message;
    };
public:
    ClassCastException() : Object(new Impl) {}
    ClassCastException(const String& msg) : Object(new Impl) {
        static_cast<Impl*>(data)->message = msg;
    }
    const char *type() const { return "ClassCastException"; }
    Object *clone() const { return new ClassCastException(*this); }
    String message() const {
        return static_cast<Impl*>(data)->message;
    }
    String toString() const {
        return message();
    }

};



class ArrayList: public Object {
    struct Impl: public Object::Impl {
        std::vector<Object*> _data;
    };
    
public:
    ArrayList(): Object(new Impl) {}
    ~ArrayList() {
        Impl *self = static_cast<Impl*>(data);
        for (std::vector<Object*>::iterator it = self->_data.begin(); it!=self->_data.end(); it++) {
            delete *it;
        }
    }
    void add(const Object& element) {
        static_cast<Impl*>(data)->_data.push_back(element.clone());
    }
    size_t size() const {
        return static_cast<Impl*>(data)->_data.size();
    }
    
    Object &at(size_t index) const {
        return *static_cast<Impl*>(data)->_data.at(index);
    }
    template<class T> const T &at(size_t index) const {
        Object *o = static_cast<Impl*>(data)->_data.at(index);
        T *t = dynamic_cast<T*>(o);
        if(t) return *t;
        throw ClassCastException(o->type());
    }
    
    const char *type() const { return "ArrayList"; }
    Object *clone() const { return new ArrayList(*this); }

    String toString() const {
        std::ostringstream oss;
        oss << Object::toString() << "(";
        for (size_t i = 0; i<size(); i++) {
            oss << at(i).toString();
            if(i+1<size()) {
                oss << ",";
            }
        }
        oss << ")";
        return oss.str().c_str();
    }

};


class OutputStream: public Object {
    struct Impl: public Object::Impl {
        std::ostream &stream;
        Impl(std::ostream &os): stream(os) {}
    };
    OutputStream() {}
public:
    
    OutputStream(std::ostream& os): Object(new Impl(os)) {}
    OutputStream(const OutputStream &other): Object(other) {}

    void println(const Object &object) {
        static_cast<Impl*>(data)->stream << object.toString() << std::endl;
    }
    
    const char *type() const { return "OutputStream"; }
    Object *clone() const { return new OutputStream(*this); }

};

Now let’s see how to use it in client code. I supply only a main() here but I have some anonymous blocks to show the effects of scoping:


struct system {
    OutputStream out;
    system(): out(std::cout) {}
};

struct system System;


int main (int argc, const char * argv[]) {
    
    assert(globalRefCount==1); // 1 is for System.out
    
    {
        String s;
        assert(globalRefCount==2);
    }
    assert(globalRefCount==1);
    
    
    {
        String s = "Connecting...";
        System.out.println(String("s = ") + s);
        
        String dots = "the dots";
        String t = s + " " + dots + ".";
        System.out.println(String("t = ") + t);
        
        ArrayList l;
        l.add(s);
        assert(l.size() == 1);
        assert(globalRefCount==6);
        
        l.add(t);
        l.add(ArrayList());
        System.out.println(String("l = ") + l);
        assert(globalRefCount==8);
        
        try {
            System.out.println(String("l[0] as String = ") + l.at<String>(0)); // ok
            System.out.println(String("l[1] as String = ") + l.at<String>(1)); // ok
            System.out.println(String("l[2] as String = ") + l.at<String>(2)); // this throws!
            assert(false);
        } catch(ClassCastException e) {
            System.out.println(String("ClassCastException: ") + e);
            l.add(e);
        }
        
        
        // adding to the list in other scope will keep the object valid
        {
            String other = "Created in other scope";
            l.add(other);
        }
        System.out.println(String("l[3] as String = ") + l.at(3));
        assert(l.size()==5);
        assert(l.at<String>(4) == "Created in other scope");
        System.out.println(String("l = ") + l);
    }
    assert(globalRefCount == 1);
    
    return 0;
}

Pretty much like Java, isn’t it?

I have written this just as a proof of concept. As such, it is fully working and I like it so far. It could serve as a good starting point for a complete implementation. Here’s the output when executed:

s = Connecting...
t = Connecting... the dots.
l = ArrayList@0x7fff5fbff8c0[1](Connecting...,Connecting... the dots.,ArrayList@0x100100b00[1]())
l[0] as String = Connecting...
l[1] as String = Connecting... the dots.
l[2] as String = ClassCastException: ClassCastException@0x100100cb8[1]
l[3] as String = ClassCastException@0x100100b20[1]
l = ArrayList@0x7fff5fbff8c0[1](Connecting...,Connecting... the dots.,ArrayList@0x100100b00[1](),ClassCastException@0x100100b20[1],Created in other scope)
Program ended with exit code: 0

Fundamental Design

Java is (with the exception of primitive types like int, double etc. and their array forms) an object-oriented language. Everything in Java or Objective-C derives from a common and well known base class. So there is a base class Object in this example as well and there is an equivalent of the platform type String as well as a sample collection called ArrayList which is holding just Objects.

One thing is very important: There is no need for pointers as memory management is part of the solution! For instance, the ClassCastException that is thrown can be added to the ArrayList without having to worry about leaking memory afterwards. It’s not complete yet though as the notion of “weak” pointers is still missing (think: ARC), but the strong part it is fully working here.

The whole idea of the implementation is based on the well-known PIMPL idiom but it also throws the idea of the smart_ptr into the mix but it even goes further. Firstly, all behavior and data are strictly separated, not just the private members. Each conceptual class like String, ArrayList or ClassCastException has a functional class that implements behavior only and no data at all, it acts like a fully functional proxy to the data. This makes it possible to clone (copy assign) these proxy objects very cheaply, because they consist only of 2 pointers (data and _vtable). The actual data is implemented in the nested class “Impl”. There is one specialized Impl class for each conceptual class (1v1 mapping). Both the conceptual classes and the Impl classes span two parallel type hierarchies. As one picture says more like 1000 words, here it is:

In the base Impl (Object::Impl) a smart_ptr like reference counting is implemented (_refCount). I have explicitly added two methods Object::retain() and Object::release() in the code to express the similarity to Objective-C’s NSObject, but this is all handled internally during copy construction or assignment.

Conclusion

I have still to decide wether an approach like this is generally feasible. What I like is to be able to clone good concepts and class library designs from other languages like Java or Objective-C into C++ and continue coding without having to worry about the aforementioned detailed C++ design decisions that trouble me every day.

Of course, I would have to implement ARC style memory management completely before it can be used, otherwise cyclical references would leak. Also, I’d like to mention that I’m fully aware that a coding style like this leads to immediate code bloat. But so does the PIMPL idiom. In order to mitigate that, I have a flexible code generator on my side which let’s me do most part of the actual coding in UML as opposed to hand-crafting it where I would definitely think twice or even more before traveling down this road…

View the full source: https://gist.github.com/2279561

Multiple Inheritance in Objective-C / Core Data vs C++

Multiple inheritance is hard. In fact it is so hard, that only very few programming languages support it. Objective-C is one for instance, where support for multiple inheritance is limited to the conformance to @protocols. Behavior can only be inherited from one single base class.

If you need multiple inheritance in Objective-C you have several options to choose from. Most of the time when you are looking for answers to the question of how to do multiple inheritance in Objective-C the right way, you will be pointed into one of two directions: #1 don’t use it because it implies a flawed design. #2 do it using composition (via delegates).

Option #1 I do not like at all. There are cases where MI is very well suited and only because Objective-C doesn’t support it doesn’t mean it is bad. It just means that the language designers considered it way too complicated to implement for the benefit of a few cases where it makes sense.

For instance, my current use case with strong need for multiple inheritance support is the UML specification. UML makes heavy use of multiple inheritance and if you study the UML model you will find that the abstractions found in there make very well sense because they eliminate redundancy and the need to explain what’s going on. All those abstractions are basically orthogonal classifications which can be combined in a subclass to express things very precise and in a type-safe manner.

So, if you are forced to deal with multiple inheritance in your program you can do so with option #2 in Objective-C. However, in my opinion, this has limitations. I will give you an example: Imagine a model like this:

Let’s say we map this to the following physical implementation in Objective-C. Here, the greenish elements are Objective-C @protocol and the yellowish elements are Objective-C @class:

It follows the often heard recommendation to map inheritance using composition. Here, the Class part of an AssociationClass is mapped to a delegate called “theClassImpl”, whereas the Association base class is mapped to plain Objective-C inheritance.

Suppose now we want to map this structure to CoreData. We need to model NSManagedEntity with NSManagedProperty. CoreData does not work on top of @protocols but on actual @classes. Therefore, we have one physical implementation of the association between Class and Property (owningClass-properties).

But here comes the big BUT: This can only work if we have full control over the OR mapping! CoreData on the other hand, does not rely on interfaces but on the actual implementations. That means, we must publish the otherwise internal composition mapping of theClassImpl to CoreData. If we then have a client of Class (for instance: Property::owningClass) then it will not be possible to downcast such a Class obtained from the persistence layer into an AssociationClass. Instead, it would be necessary to navigate backwards from the Class to the actual AssociationClass. But this kind of “alternative” cast can not be implemented transparently using Objective-C language constructs. An [aClass isKindOfClass:[AssociationClassImpl class]] would yield a technical “NO” and it’s not possible to extend the language to make it yield “YES”.

Such an MI -> SI mapping scenario can only work if every consumer solely relies on the Interfaces only and makes no assumption about internal structures. This would imply that the ORM uses factories instead of instantiating from its meta information. In CoreData, this is not the case.

This is why you can pretty much ignore the advice how to map multiple inheritance at the language level if you don’t also consider the APIs you’re dealing with because those APIs will render the easy sounding solution in the context of reality useless pretty quickly. In my case, I was forced to implement the model part of the system in C++ because C++ has awesomely good support for multiple inheritance. All problems related to the mapping of multiple inheritance to the microprocessor architecture had been solved by Bjarne Stroustrup in C++ since day 1. Read here why and how: http://drdobbs.com/184402074

Here’s how the dreaded diamond from the example above would be implemented in C++:

class Element {
};

class Association: public virtual Element {
};

class Class: public virtual Element {
public:
    std::vector<class Property*> properties;
};

class Property: public Element {
public:
    Class *owningClass;
};

class AssociationClass: public Association, public Class {
};

A straight 1:1 mapping of the concept to the language. Here, class “AssociationClass” would fully inherit the behavior of Class::properties without the need to implement something special. It just works. But, in comparison to Objective-C the C++ implementation lacks support for Core Data. But: so does multiple inheritance in Objective-C with CoreData! So no real difference here.

Conclusion

Multiple inheritance with CoreData is close to impossible except for very simple cases. With C++, besides all its ugliness and controversy, at least you get multiple inheritance in the language and usually in an implementation quality without the need to waste your time thinking around the whole concept.

Why I completely dropped Qt and QtWebKit from my Hybrid App

In my recent articles I described certain aspects related to the use of QtWebKit in a hybrid C++ application. In the meantime, it turned out that Qt and QtWebKit are not such a good bet as I thought they would be at the beginning. Here’s why:

Apple

First and foremost, the most interesting question in a hybrid app written in C++ is: where do you want it to run? In my case, it is on Mac OS X in the first place, then on Windows and very likely also on Apple’s iPhone or iPad. However, this puts me in a situation where I have to decide which subset of these platforms I support primarily and which ones secondarily as there is no such thing as a common layer to access all of them from a single code base. It is either a combination of Mac OS X and Windows via the Qt port or a combination of Mac OS X and iPad/iPhone that can go as the primary target.

For a certain time, I concentrated on the first combination as shown above while hoping that time works for me and somebody ports Qt to the iPhone, and there already is a project for that at http://www.qt-iphone.com/Introduction.html. But, in the meantime, Apple introduced iPhone OS 4.0 and *bam* there will never be such a thing, because http://daringfireball.net/2010/04/iphone_agreement_bans_flash_compiler “and only code written in C, C++, and Objective-C may compile and directly link against the Documented APIs (e.g., Applications that link to Documented APIs through an intermediary translation or compatibility layer or tool are prohibited)”, a rule introduced by Apple which is actually targeted against Adobe and its Flash technology but will also affect Qt, especially because that comes from Nokia, another one of Apple’s (future arch) rivals. This means that Qt can never be the common layer on all of my targeted platforms. Given that, and the fact I target Mac OS X first, Windows second, and that I also tend to shift Windows one step further down and promote the iPad version instead which completely changes the porting picture:

This looks much cleaner now as it provides unified, access to all Apple platforms. For Windows and Android Google Chromium is a good choice.

But what about Qt in general?

Qt / Nokia

The first part was the rational part, here comes the emotional one:

In the recent weeks I had enough time to dive into the details of Qt in all aspects. Its visual quality, its API quality, its implementation quality and the quality and structure of their development process. My conclusion from that is that it is the wrong way of doing things as evaluated from a broader perspective.

What is very important in a visual app is visual quality. Qt does a good job in replicating functionality of controls on each addressed platform. But, it can never, by its nature, take care about the special controls certain platforms offer because then it wouldn’t be cross-platform any more. Qt certainly has to focus on the lowest common denominator across all supported platforms. By driving the GUI via Qt controls, an application loses the capability to provide the eye-candy to the platform user like he is used to. The most prominent example of this is Google Maps which is based on Qt and which totally lacks platform L&F on the Mac up to a degree where it certainly loses style. In Google Maps this is not so important as interaction with the Earth requires a special user interface anyway and the rest of the app is unimportant, but it shows how difficult it is even for Google to provide attractive visual quality on the platform. Gaining access to the platform specialities is difficult as everything is hidden by Qt in private implementations. You have no chance to do “that special thing”.

Qt has a comprehensive “fat” API that consumes everything from all platforms and which has to take care about the specialities on all platforms, which is of course its nature but which is also very difficult to maintain and to improve over years. Development of Qt is distributed across Europe with Nokia focussing on minimizing cost not maximizing quality. Their development centers are mostly in cheaper parts of Europe and they try to cooperate with educational institutions to bring the costs down. This might be a legacy from Trolltech but is still in place and shows that quality is not their first concern.

So far I have touched Qt releases 4.5, 4.6 and 4.7 and came across older API documentations. From there I can tell that many things they have done are not very well and strategically thought out. They introduced too many APIs doing similar things and deprecating previous ones (QML, QtQuick, Scripting with QtScript and in QtWebKit, Qt3 vs. Qt4, QGraphicsScene, QItemView, QtDesigner, QtCreator). To me this is too much feature driven and dictated by the progress of other market players. Then I saw these super new technologies contrasted by very basic demo videos in YouTube showing excitement over very basic features which is not very professional – you can check out their channel on http://www.youtube.com/user/QtStudios to get an idea. What I want as a consumer of platform technology building a commercial grade application is that the selected platform is very stable and well thought out. But there is just too much noise coming from Qt and to make matters worse, none of the new features really feels final, it feels more like work in progress and you can never know how long something survives. Last not least, one day down the road Nokia might become a rival again to Apple just as relevant as Google now is. Economically speaking, Nokia already is, but Apple sees Google as its main competitor, not Microsoft and not Nokia. They are just not in the same technology league any more. Apple and Google have the brains now.

Don’t get me wrong, the Qt guys are doing a good job and they surely do it with a lot of enthusiasm, all I want to say is that Qt is just not the best bet in a commercial hybrid application. That is why I gave up on it.

Accessing the Original WebKit API in QtWebKit Hybrid Apps

Everybody wraps WebKit

For each and every port of WebKit, be it Safari, Chrome, QtWebkit, there is always a wrapper around the WebKit object model. These wrappers function as language mappers, e.g., in case of Safari C++ API is mapped to Objective-C, and as a simplification layer around the complex WebKit API. The first case makes total sense, as no developer familiar with Cocoa and Objective-C will like to take an excursion (back) to C++. On the other hand, the Chrome port as well as the Qt port expose a C++ API towards their developers and completely hide the WebKit API despite that this is also already written in C++. Within the C++ universe, I do not understand why the original WebKit API is such a bad bet and must be hidden under all circumstances (I derive this from the fact that everybody hides it)? Is it really that evil?

WebKit API is very stable as it is a reflection of HTML and XML and all the things used in the browser for years and which do rarely change, so a dependency should be acceptable. The only benefit is that the Qt developer works solely with the Qt API or the Chromium developer with a Chromium API. This is ok in case of very simple integration with WebKit, like displaying some external web content in an otherwise traditional application or when implementing a browser. But not in a hybrid application! In a hybrid application I want much more control over the internals of the browser. I do not consider it evil. I consider it as a very nice work horse doing most of the rendering and layout work. I want to reference DOM elements from my application and I want the DOM to interact with my application. I want to replace the excessive use of JavaScript found in Web 2.0 apps with excessive C++ – I do not want JavaScript in a hybrid app, I want C++ only!

In case of QtWebKit, how hard is to get access to the original WebKit API? It is very hard, as the necessary headers are all eliminated from the QtWebKit API and all WebKit symbols are local in the compiled DLL or OS X framework However, I have taken the pain and hacked me into it. The steps required to expose the WebCore API via QtWebKit are:

Get the source tree from Qt git

You will need to build from source. So, get it from http://qt.gitorious.org/qt and switch to the desired branch or tag.

Configure Qt just as normal

configure it but do not run make now!

Adjust the compiler flags for WebCore

run the following commands in the source tree

sed -i -e 's/-fvisibility=hidden//g' src/3rdparty/webkit/{Web,JavaScript}Core/Makefile.*
sed -i -e 's/-fvisibility-inlines-hidden//g' src/3rdparty/webkit/{Web,JavaScript}Core/Makefile.*

(or change these the Makfiles manually) This will compile WebCore in way to expose all internal symbols later in the DLL or framework.

Gain access to the WebKit API

In src/3rdparty/webkit/WebKit/qt/Api/qwebelement.h make QWebElement::m_element public

This is just one way to do it (see later for an alternative)

Build Qt

Now run the normal make; make install

Add WebKit header files and defines to your .pro file

The application code needs to include the WebCore header files.

DEFINES += QT_SHARED ...
release:DEFINES += NDEBUG
INCLUDEPATH += /Users/andre/src/qt/git/src/3rdparty/webkit/WebCore/bridge/qt \
    /Users/andre/src/qt/git/src/3rdparty/webkit/WebCore/page/qt \
    ...

This is a very shortened version to save space here. See the full set of defines in this file: myapp.pri

The order is important. Replace /Users/andre/src/qt/git with the home of your Qt source tree. Some defines are very important and some might be just optional, I just took all of them from the effective compile command when Qt was built.

Include WebCore headers in application code

#include <WebCore/html/HTMLElement.h>
#include <WebCore/platform/text/PlatformString.h>
#include <WebCore/platform/text/CString.h>
#include <WebCore/svg/SVGElement.h>
...

Alternative access to QWebElement::m_element

Instead of changing the visibility of the “m_element” member from private to public, it would also be possible to use a fake subclass in the application code with a public member m_element like here:

class HackWebElement: public QWebElement {
public:
WebCore::HTMLElement *m_element;
};

and then gain access to it by down-casting a QWebElement to a HackWebElement:

QWebElement webElement = ...;
HackWebElement *hwe = (HackWebElement*)&webElement;
// now we can access hwe->m_element

You can also use your preferred method to access a private member in C++ 😉

Using the WebKit API

Now I can do fancy stuff with the WebKit API. Here is an example where I create an SVG element programmatically, i.e., without the need to have WebKit parse some HTML (or XML in this case):

WebCore::QualifiedName svg("svg", "svg", "http://www.w3.org/2000/svg");
RefPtr<WebCore::SVGElement> nel = WebCore::SVGElement::create(svg, hel->document());
WebCore::String s = nel->tagName();
debug(string(s.utf8().data()));
if(nel) {
    debug("it's a WebCore::SVGElement!!!");
}

Conclusion

So far, I have only gained initial access to the internal browser DOM as a first step. This way I can create DOM elements programmatically. In the future, I will experiment with event handlers on DOM objects written in C++ in order handle all UI events in the C++ part of my application. This will reduce the browser part to act as a pure layout and rendering engine which I think should be its sole role in a hybrid application.

Regarding the little “hacky” approach I can say that there is only one place on the whole application code where access to WebKit is gained. This can be in the document, in the frame, or like here in the general element. From then on, no further hacks are required, it is just used and QtWebKit is not needed any more.

I would also wish that in ports like QtWebKit the WebCore API is preserved and exposed to the application developer. I think there are 2 aspects in a WebKit port: One is the physical rendering and display and the other one is the kind how WebKit embeds into the application and is access. Both concerns should be considered separately and I want to be free to consume only the first one and live with the original in the latter one.

Advanced Object Lifecycle in OO Systems applied to implementations in C++

C++ certainly provides sophisticated mechanisms to create and destroy objects via it’s constructors and destructors. There are, however, certain aspects to these which make them hardly usable in advanced OO systems. First and foremost, a C++ constructor implementation lacks one very important feature: polymorphic calls! In a constructor, a call to a virtual function of the object constructed, is not a polymorphic call but a call to the method as overridden at the level of the class currently constructing the object! This makes it impossible to give the more abstract constructor information from the more concrete class. Imagine you have a UI element that need size information during construction, but that size information can only be provided by it’s specific subclasses, then you will need to implement a 2-phased approach with construction first and using (displaying in this case) second.

C++ constructors also lack in contrast for instance to Java the possibility to call another constructor from the same class. You can only call constructors from super classes. This makes it harder to initialize a large number of members because the initializers have to be implemented redundantly in every constructor. Without a code generator this is very error prone.

There is another problem which is the rather weird logic of which members get initialized and which do not. If you don’t have a constructor which initializes pointers, then these pointers are not initialized to 0. However, if a class uses some struct as a member and this struct has a pointer as member, then this pointer is initialized to 0.

So, a general approach to get over all these problems would be to not use constructors at all and delegate all responsibilities to some generic virtual methods. Let’s first look at the requirements to an OO lifecycle in general:

Creation

  • Create all mandatory child objects upon creation of the parent
  • If an instance requiring a parent element is created, the parent is already present

Transfer Ownership

  • Take a partial tree out of an object tree and connect it to a different object tree (reparenting)

Deletion

  • Isolate a subtree and destroy it

These requirements can be translated into a general lifecycle model in OO systems:

Objects are, once allocated via operator new() in an intermediate state “allocated”, where they can be attached to a parent object, i.e., the parent becomes known to the child and can be used in the subsequent init() call which initializes the object structure in it’s depth. Instead of deleting an object via operator delete() it is destroyed by the destroy() method. In the destroy method, all child objects get destroyed and all destroyed objects are returned to the object pool as isolated instances. Only the pool deletes these instances if needed. Instances are never created via operator new() directly, they are always obtained from the pool via the factory method instance() of the class. This factory method creates a new instance if none is found to be recycled from the pool. Both the default constructor and the destructor are declared as private to ensure the sole usage of the factory method.

This would further lead to the 2 virtual methods init() and destroy() to be implemented on each class level. These methods would have the following behavioral characteristics:

init()

  • is called only after business parent has been set
  • is the constructor equivalent
  • init() calls super::init() first, just like a constructor
  • init() initializes members of the class
  • init() can be called repeatedly
  • memory allocation is a separated concern
  • init() is virtual
  • init() can call virtual methods

destroy()

  • is called before business parent is removed
  • is the destructor equivalent
  • destroy() first resets own members
  • destroy() calls super::destroy() second like a destructor
  • destroy() can be called repeatedly
  • memory deallocation is a separated concern
  • destroy() is virtual
  • destroy() can call virtual methods

Conclusion

  • Constructors/destructors are not needed any more, except for memory allocation
  • Instances are created in the pool only (i.e., in the class)
  • Objects have either a business parent or are referenced by the pool during destroy()
  • objects are returned to the pool
  • if objects do not have a business parent, then a parent is also not needed for initialization

These feature improve the handling of complex object structures in a meta-driven OO environment applied to C++. In fact, they are so generic that they can be easily adapted in a completely different programming and runtime environment. There are, of course, some drawbacks which is basically the more complex protocol for initialization and the implementation of the object pool.

Embedding 2D Drawing Objects in QWebView

One question not so easily answered is how to display high-frequently updated dynamic content in a QWebView. QWebView is a great tool to build hybrid web applications in C++. But because it is basically only a port of WebKit and Qt tries to wrap and encapsulate as much as possible, it is also only a plain web browser, just capable of displaying content that comes in HTML form. What if you want to display content within a web page like a stock chart or even a video…?

The options WebKit browsers offer are basically:

  • <object>, , <applet>: they each require you to implement a specific API and in case of applet even a different programming language
  • html 5.0 canvas: this is interesting, however, it requires JavaScript to program against it which might be to slow due to frequent parsing; it would be much more interesting if a QWebElement style wrapper would be available for drawing directly from C++
  • http://code.google.com/p/nativeclient/ is very interesting in normal browsers, but not in a hybrid application, as it is just another form of the <object> tag

This is why I have been thinking about a different approach, which is closer to the machine’s (i.e., Qt’s) drawing and came up with this idea:

Here we place a special, empty <div> in the document which can later be accessed by it’s ID using a QWebElement. We overwrite the QWebView::render() method and check if the region overlaps our dynamic content as specified by the QWebElement::geometry() data. The actual dynamic content can for instance be pre-rendered into an internal buffer, which we can use later as the source of a pixel copy operation towards the QWebKit paint device. Whenever our dynamic content changes, we can force a repaint via QWebView::repaint() exactly on the region that has changed.

Architecture-wise this just means that the main presentation layer of the hybrid app is not accessed in an outbound fashion only. I have no problem with that, as long as the dynamic content can be displayed and the module doing that can later be replaced with some less “hacked” approach…

03/05/2010 Update: HTML 5.0 Canvas is working in QWebView. I don’t know to what exact extent but some initial drawing tests completed successfully.