Compile-time detection for calling unsupported virtual method override

[arnetheduck]

• QAbstractListModel inherits from QAbstractItemModel
• creating a QAbstractListModel instantiates a MiqtVirtualQAbstractListModel
• QAbstractItemModel contains https://doc.qt.io/qt-6/qabstractitemmodel.html#columnCount as a pure virtual function
• the generated bindings for QAbstractListModel do not contain an override function for columnCount
• Instead, QAbstractListModel_override_virtual_columnCount is calledvia inheritance which fails its dynamic_cast<MiqtVirtualQAbstractItemModel> (because we created MiqtVirtualQAbstractListModel).

[mappu]

There is an override function for QAbstractListModel::columnCount - https://github.com/mappu/miqt/blob/master/qt6/gen_qabstractitemmodel.cpp#L547

There is also a working example for this Qt class, in the examples folder. Can you repro this on miqt master?

[arnetheduck]

What you linked actually is part of MiqtVirtualQAbstractItemModel and as a consequence calls miqt_exec_callback_QAbstractItemModel_ColumnCount - notice Item vs List (basically QAbstractListModel_override_virtual_ColumnCount is nowhere to be found)

[arnetheduck]

Ok, I think I understand this one now - columnCount has a private override in QAbstractListModel - so when you're using C++ it's "off limits" if you have a QAbstractListModel pointer.

But there is nothing in the exported API that prevents it from being called through the inheritance structure, and this is where the failure happens.

I think this happens in the go bindings as well (don't know go enough), but to fix and "block" onColumnCount for the QAbstractListModel, the IL would have to retain information about private method overrides.

Normally, users wouldn't try to override this way, it doesn't make much sense, but this happened in a generic override section of nimqml that ideally would have seen a compile-time error.

[mappu]

Right, that makes sense - columnCount is virtual public/protected in the base QAbstractItemModel, and then the derived QAbstractListModel changes it to private.

The Miqt IL does track this, and only generates an OnColumnCount function for the base QAbstractItemModel, there is no generated QAbstractListModel.OnColumnCount function. You can see a comment about this in intermediate.go: https://github.com/mappu/miqt/blob/master/cmd/genbindings/intermediate.go#L477

In the Go projection, the base struct is embedded in the derived struct, allowing you to seamlessly call methods on the base type. However, only the directly constructed type has the isSubclass flag set. When calling QAbstractListModel.OnColumnCount, Go transparently calls (QAbstractListModel).(*QAbstractItemModel).OnColumnCount, but the interior QAbstractItemModel has isSubclass=false inside its Go struct, which prevents you from calling the method and segfaulting.

The change in #138 moves this defense into the C ABI, which can help for other language projections. This will detect the dynamic_cast failure and, instead of dereferencing a null pointer, it can return false to the language projection, which can then produce an error message at runtime. But this has no effect for the Go projection other than saving 1x bool.

---

Another way to fix this would be to prevent downcasting, but rather, expose language projection methods for all of the public methods on the base class. But, this causes problems if some other function wants to take a base class pointer.

Another way to fix this would be to distinguish QFoo and MiqtVirtualQFoo in the ABI, and only allow the override methods if you have a strictly typed subclass pointer. This would also allow compiling with -fno-rtti for smaller binary sizes. But that's a really massive change.

[arnetheduck]

there is no generated QAbstractListModel.OnColumnCount function.

What I'd be looking to generate is a "deleted" function that would catch the attempt at compile time as long as you make the attempt through a QAbstractListModel (ie it would still be possible to fail at runtime if via the base class, but as long as you use the derived List class, the function call would simply not be available to you.

One could almost achieve this with PrivateMethods, as long as there's no overloading (since there's only the name to go by).

[mappu]

What I'd be looking to generate is a "deleted" function that would catch the attempt at compile time

That would be fantastic.

I can see how it would be possible in C++, but I think Go is not expressive enough to do this (maybe with some go:linkname tricks?).