Panic "can only override virtual methods for directly constructed types" when calling a parent class override method

[marc-a-didier]

Hi, currently giving a try to miqt to port legacy gtk2 code to qt6 on ubuntu 22.04. Everything goes well, it compiles and get executables. But as soon as I try to do something less trivial than the miqt provided examples, I run into problems. Simple example: I built examples/modelview_color6, it runs ok. Then I changed the code from

	model.OnRowCount(func(parent *qt.QModelIndex) int {
		return 1000
	})

to

	model.OnColumnCount(func(parent *qt.QModelIndex) int {
		return 2
	})
	model.OnRowCount(func(parent *qt.QModelIndex) int {
		return 1000
	})

Just added a method to have more columns and all what I get is:

panic: miqt: can only override virtual methods for directly constructed types

Can someone explain me where the problem lies?

[mappu]

I can confirm the issue.

Diff

diff --git a/examples/modelview_color6/main.go b/examples/modelview_color6/main.go
index da48f246c..aa2441a29 100644
--- a/examples/modelview_color6/main.go
+++ b/examples/modelview_color6/main.go
@@ -12,6 +12,10 @@ func main() {
 
        model := qt.NewQAbstractListModel()
 
+       model.OnColumnCount(func(parent *qt.QModelIndex) int {
+               return 2
+       })
+
        model.OnRowCount(func(parent *qt.QModelIndex) int {
                return 1000
        })

Output: panic: miqt: can only override virtual methods for directly constructed types

---

In Qt/C++, you had to implement these rowCount and columnCount by subclassing. Miqt lets you do it easily at runtime with these methods, but in exchange, there is a runtime check that it's done correctly only in the same places Qt/C++ would have allowed it.

This check is misfiring i think, this should have been allowed.

QAbstractListModel defines OnRowCount, but OnColumnCount is defined by its parent, the QAbstractItemModel. I think this is a regression from #176 (or similar) regarding the way parent classes are handled.

[mappu]

qt.QAbstractListModel gets its parent pointer populated via the _virtbase() helper, but if you look in gen_qabstractitemmodel.cpp, the problem is right there -

void QAbstractListModel_virtbase(QAbstractListModel* src, QAbstractItemModel** outptr_QAbstractItemModel) {
	*outptr_QAbstractItemModel = static_cast<QAbstractItemModel*>(src);
}

We cannot static_cast to the virtual version of the parent class (MiqtVirtualQAbstractItemModel) because it's not strictly in the C++ class hierarchy. Miqt will need an update to (probably) generate all the additional override methods directly in this derived class.

Penny for your thoughts @arnetheduck , is there another way of doing this in C++ ? Does the vtable approach in seaqt affect this case?

[arnetheduck]

This is a funny one.

In a sense, it's right that trying to override columnCount results in an error for List but it's also bug that it's not allowed - as the docs state, lists are one-dimensional and overriding columnCount is wrong semantically. To mark this point, columnCount is private.

That said, being private is a restriction for who can call the function, not who can override it and it's docs and not C++ semantics that prevent the override. In general, overriding private functions should be allowed.

If miqt were to support this correctly, it would indeed need to provide an override point for the full inheritance tree in each type, and this is exactly what the vtable approach does!

You can see how the vtable for QAIM includes all "overridable" functions all the way down to QObject - ditto table which also has its own spot for overriding columnCount. It gets even more interesting with multiple inheritance because then you also need to consider which virtual functions are governed by virtual inheritance rules and which are separate.

The issue is funny because it caused me to spot a bug - the vtable for List is missing its columnCount override spot probably because I confused "calling" with "overriding" - only the former should be blocked, while the latter should be allowed - QAbstractListModel_VTable should have a columnCount.

Long story short, the go projection has the "problem" that it uses inheritance between List and Item to expose overrides when indeed it should repeat all "overridables" in each type - it's not necessary to use a vtable to fix this, but it might be easier/more convenient - this way, inheritance is for "calling" things and "duplication" is for overriding things and this might be easier to express separately. That said, I don't know go :)

[arnetheduck]

MiqtVirtualQAbstractItemModel

Actually, another alternative might be to have MiqtVirtualQAbstractListModel inherit from MiqtVirtualQAbstractItemModel - this might actually work if you're not using a vtable and I can't think of any downside right now. In seaqt the vtable is flat because this is easier to "import" in C and inheriting VirtualQAbstractListModel from VirtualQAbstractItemModel wouldn't make sense because it would result in two vtable fields.

[mappu]

That's definitely an interesting option.

MiqtVirtualQAbstractListModel would have to still inherit from QAbstractListModel - so adding MiqtVirtualQAbstractItemModel creates diamond inheritance on QAbstractItemModel (and so on, for every base class), so the virtual would have to come back (we got rid of it in 929e4e6 / #138).