class pointer and non pointer member function question..

JonB

QLabel myLabel;  //this should be fine, anyway will be on heap

This class member variable is created when any Alpha is created, and destroyed when that Alpha instance is deleted or goes out of scope. The nice thing is that nobody has to delete this variable, because its destruction is automatic, when Alpha is destroyed.

QLabel* myLabelPointer {nullptr}; //this should not be necessary.

This means that somebody somewhere has to go myLabelPointer = new QLabel(). And that means it will need explicit deleteing. Either you can put delete myLabelPointer into Alpha::~Alpha() destructor, or when you create it go myLabelPointer = new QLabel(this) (or use a call which puts it on the dialog, which will do this at some point). Now because of the parent = this, when the QDialog gets destroyed it will have Qt code which goes down all the widgets it owns this way and call delete on them for you.

Chris Kawa

It's worth noting that while it's fine to use a non pointer variable in a simple case like this, an average ui is far more complex and it becomes harder to track the order of the children. Keep in mind that placing a widget in a layout of another widget or manually reparenting it will also make the parent try to delete that widget.

So, if you have something like

class Alpha : public QDialog
{
   QLabel myLabel;
};

the lifetime timeline is:

Alpha is constructed
myLabel is constructed
myLabel is placed somewhere in Alpha -> Alpha becomes parent
Alpha destructor hits
myLabel is destroyed -> detaches from Alpha
Alpha base QObject destructor hits, no children to delete
Alpha is destroyed

but, if you do something like

class Alpha : public QDialog
{
   QGroupBox myGroupBox; //just an example
   QLabel myLabel;
};

and do something like:

myGroupBox.layout()->addWidget(&myLabel);

it becomes problematic:

Alpha is constructed
myGroupBox is constructed
myLabel is constructed
myGroupBox is placed somewhere in Alpha -> Alpha becomes parent
myLabel is placed somewhere in myGroupBox-> myGroupBox becomes parent
Alpha destructor hits
myGroupBox destructor hits -> tries to free all children, including calling delete on myLabel
myLabel destructor hits -> oh no! crash

You could avoid this particular problem by reversing order of the variables, but you can see how easily it can get messy when talking about dozens of widgets in an average dialog. So while dynamic allocation is not something to be taken lightly entire Qt is designed in a way that makes it a lot easier and less error prone if you use it.

JonB

@Chris-Kawa
That's frightening :(

So.... keeping it simple the rule is: it's OK to declare member instances in, say, QDialog provided all of them are put directly onto the dialog, so it stays their parent. But here you want to re-parent the label onto the group box, so all hell breaks loose. And you'd have been better doing these via pointers and

myGroupBoxPointer = new QGroupBox(this);
myLabelPointer = new QLabel;
myGroupBoxPointer->layout()->addWidget(myLabelPointer);

Is that about right?

When I see @SGaist often recommending people declare a QTimer someone needs as a member instance (not pointer), this turns out to be "safe" without having to worry about it because you're not going to re-parent a timer. The problem of having instance variables will come, in practice, when you are declaring QWidgets, because there you do re-parent.

Is that about right too?

Chris Kawa

@JonB Yup, about right. It's not only with widgets though. It's the same with any QObject but maybe easier to trip over QWidgets because it's easier to reparent them implicitly without calling setParent.

In a nutshell:

// any scope
{
    QObject a;
    QObject b;
}

it's ok for a to become the parent of b, but for reverse to work you'd have to explicitly de-parent a before the end of scope, otherwise there's double-delete problem.

My super relative, super simplified, super context dependent rule of thumb: if it has a parent allocate dynamically.

JonB

@Chris-Kawa
Funnily enough, a whole convo related to this is going on in https://forum.qt.io/topic/124322/avoid-dangling-qtimer-object atm.

U7Development

thank you very much for such rich information..

kshegunov

@Chris-Kawa said in class pointer and non pointer member function question..:

My super relative, super simplified, super context dependent rule of thumb: if it has a parent allocate dynamically.

Eh! I hate allocating a single pointer on the heap, so as a consequence I hate that rule of thumb. Nothing personal, but it's just terribly inefficient, even if manually detaching the parent feels so clumsy.

Chris Kawa

@kshegunov Sure, if it's one. That's where the super * part of the rule comes in :)

Nothing personal, but it's just terribly inefficient

Agreed, but otherwise it's really fragile. Even if you have it perfectly under control and all 50 widgets in your dialog play nice there's always that intern that comes in and tries to be smart by rearranging your class members so the name alignment looks nicer :)

It's the usual design struggle - it depends on oh so many factors, not all technical.
But yeah, I'm usually the one complaining about performance, so I'd be a hypocrite not to agree ;)

kshegunov

@Chris-Kawa said in class pointer and non pointer member function question..:

Agreed, but otherwise it's really fragile.

Maybe. If I may suggest a better rule of thumb:
Use member variables (QObjects) and parent them to the containing one, like this:

class A : public QObject
{
   QObject b, c;

   A(QObject * parent = nullptr)
       : QObject(parent), b(this), c(this)
   {
   }
};

PS. Don't get me wrong, the pain is real. Currently I'm tinkering with Qt3D and QML. Everybody and their mother is a QObject there. Need a struct with 3 plain members for configuration, take this QObject on the heap instead ...

Chris Kawa

@kshegunov said:

Use member variables (QObjects) and parent them to the containing one, like this:

Yeah, kinda, but it's hairy with widgets. Often you have a class with bunch of members that parent each other through layouts, menus, views and what not.

I'm tinkering with Qt3D and QML. Everybody and their mother is a QObject there

Sigh, I feel you. I mean... that's one of the reasons that, although I'm all about graphics, I've never even tried Qt3D. I just can't see how this could ever have any reasonable performance for complexity beyond maybe a car dashboard. QObject is fat and it was designed back in the day as a building block for widgets based static ui interfaces. If I had to have an overhead of a QObject for every mesh, texture or a friggin buffer attribute I would need 128GB of RAM and a beefy top shelve GPU to even start.

And don't even get me started on QML :P Maybe I'm getting old but I feel very left behind with Qt 6.

kshegunov

@Chris-Kawa said in class pointer and non pointer member function question..:

Yeah, kinda, but it's hairy with widgets. Often you have a class with bunch of members that parent each other through layouts, menus, views and what not.

That is fair. When they're on the same hierarchy it's useful to declare the ownership explicitly through the constructor. Even if reparenting happens down the tree this ensures that the parent always outlives the child and you don't double-axe the object. Something like this:

QDialog someDialog;
QAction action(&someDialog);
// More children here, so we ensure the children are destroyed before the ancestor ...

Sigh, I feel you. I mean... that's one of the reasons that, although I'm all about graphics, I've never even tried Qt3D. I just can't see how this could ever have any reasonable performance for complexity beyond maybe a car dashboard. QObject is fat and it was designed back in the day as a building block for widgets based static ui interfaces. If I had to have an overhead of a QObject for every mesh, texture or a friggin buffer attribute I would need 128GB of RAM and a beefy top shelve GPU to even start.

For a thousand meshes with 6 vertices and normals each with a simple phong material I got 20-30 fps for my RX580, so don't hold your breath (not going to even mention the load times ...). But honestly I'm not entirely convinced it's the Qt3D itself, more like the default QML integration. The damn thing renders the scene on a texture which it blits after than onto the QML scene. And to be honest I'm rather annoyed by the lack of any reasonable documentation or an example how one integrates the two in a fashion beyond the proof-of-concept level. It all feels like stitched with white thread ... I'm left with reading sources for few days now, which is beyond ridiculous ...

And don't even get me started on QML :P Maybe I'm getting old but I feel very left behind with Qt 6.

Nah. It's the same as the widgets underneath. You can even build it directly from C++ if you don't mind the verbosity. Thankfully the QtQuick scenegraph has nothing to do with the QML itself, the latter being just a glorified factory for the QQuickItems.

Chris Kawa

@kshegunov said:

For a thousand meshes with 6 vertices and normals each with a simple phong material I got 20-30 fps for my RX580

In the meantime I'm staring at a blinking red counter showing ~16'000 render items, each consisting of a vertex and index buffer of few thousands elements, with multi layer materials and heavy post process pipeline and scratching my head how to fit it in 16ms on an 8 year old console hardware :P Nope, Qt3D is definitely not for me.

kshegunov

@Chris-Kawa said in class pointer and non pointer member function question..:

... scratching my head how to fit it in 16ms on an 8 year old console hardware :P Nope, Qt3D is definitely not for me.

You have to give me some hints about the engine(s) you use when I'm ready to give up on the Qt3D. ;)