Checking the API status for the variable “Private d” of the QVariant class

elfring

I have noticed the following source code.

…
// ### Qt6: FIXME: Remove the special Q_CC_MSVC handling, it was introduced to maintain BC for QTBUG-41810 .
#if !defined(Q_NO_TEMPLATE_FRIENDS) && !defined(Q_CC_MSVC)
    template<typename T>
    friend inline T qvariant_cast(const QVariant &);
    template<typename T> friend struct QtPrivate::QVariantValueHelper;
protected:
#else
public:
#endif
    Private d;
…

I have noticed also that the member variable “d” is not mentioned in the Qt documentation so far. So I get doubts if it can really belong to the public programming interface by this class.
How will this software evolve further?

raven-worx

@elfring
the d-pointer is the private implementation of Qt classes and guarantee binary compatibility over Qt minor versions.
I don't know when Q_NO_TEMPLATE_FRIENDS will really be defined so that the d-pointer becomes public.

But anyway d-pointers are not meant to be accessed.

Christian Ehrlicher

d does not belong to the public interface so you don't have to worry

sierdzio

@Christian-Ehrlicher said in Checking the API status for the variable “Private d” of the QVariant class:

d does not belong to the public interface so you don't have to worry

Guys but take a look at the code - this is not the d pointer, it's and instance of Private struct, called d. Looks like it is indeed part of the public interface here.

elfring

Looks like it is indeed part of the public interface here.

Conditional compilation can eventually make this variable “protected”, can't it?

sierdzio

@elfring said in Checking the API status for the variable “Private d” of the QVariant class:

Looks like it is indeed part of the public interface here.

Conditional compilation can eventually make this variable “protected”, can't it?

Yes, but since it's not a pointer, it can affect memory layout => damage binary compatibility. Even if it is private. Some more info: https://stackoverflow.com/questions/7699259/how-does-adding-a-private-member-variable-break-c-abi-compatibility

VRonin

Q_NO_TEMPLATE_FRIENDS is only defined in old or unusual compilers, if you trust the comment (i.e. the Q_CC_MSVC only exist for binary compatibility but it should be removed) I'd say the base case is for d to be protected

Christian Ehrlicher

And you can be sure this will not change inbetween a Qt X release cycle. It's a 32+64 bit struct which was adjusted to match the requirements. Noone will ever touch this in a form that the size will change.

elfring

I'd say the base case is for d to be protected

Would you like to encapsulate accesses for the member variable “d” (of the class “QVariant”) by any other function?

Christian Ehrlicher

d is protected so no other function will get access to this variable and there is also no need for it. What do you want to achieve?

VRonin

@elfring said in Checking the API status for the variable “Private d” of the QVariant class:

Would you like to encapsulate accesses for the member variable “d” (of the class “QVariant”) by any other function?

it is already [kind of]* exposed by data() but as you see it's explicitly set to return a const void*.

*if you don't save basic types in the variant then it is exposed

elfring

What do you want to achieve?

The mentioned access specifiers will influence possible software extensions around classes like “QVariant”.

I would like to get safe access to the members in the data structure “QVariant::Private”.
Would you like to support in-place modification for referenced objects then?

Christian Ehrlicher

You will never get any safe access to QVariant::Private - it's a private class and therefore can change any time. And no I don't need in-place modifications (this was discussed on the mailing list some weeks ago).

https://lists.qt-project.org/pipermail/interest/2018-June/030216.html

If you want in-place modifications then store a pointer in QVariant so you can modify the data without accessing the QVariant.

elfring

it's a private class

Please take another look.
The class declaration is using the access specifier “public” for this data structure so far, doesn't it?

this was discussed on the mailing list some weeks ago

Thanks for your link of the discussion topic “Why no T& QVariant::value()?”.

If you want in-place modifications then store a pointer in QVariant …

I became interested once more to perform some data processing for contents (which could be provided over variables like “QObject *o” and “PrivateShared *shared”) without extra copies.

VRonin

You are looking for an unnecessary hack to solve a very simple problem.
If all the classes you have stored in a QVariant are implicitly shared* and the members of those classes are also all implicitly shared* and so on down the object tree, then your get-modify-set is already as efficient as you can get short of memcpy. If you want I can explain in detail why if it's not clear.

' * or fundamental types

elfring

You will never get any safe access to QVariant::Private -

I suggest to reconsider this information. - I guess that we got different expectations for the “safety” of such data structures.

it's a private class

The privacy concerns can vary also if you would like to implement possibly useful functionality for this class.

and therefore can change any time.

This aspect is reasonable for software evolution in general. I am curious on how change acceptance (or resistance?) will evolve here.

And no I don't need in-place modifications

Do other Qt users and software developers would like to apply this technique?

If you want in-place modifications then store a pointer in QVariant so you can modify the data without accessing the QVariant.

This is another useful software design approach. But these pointers should refer to valid data.
This class can be treated as a container for a single item (with a varying internal data type). You would occasionally prefer direct write access to the contained element.
I guess that I am looking again for Qt support of an operation which became standardised as “std::get”.

elfring

You are looking for an unnecessary hack to solve a very simple problem.

It can take a while to get acceptance for another useful programming interface, can't it?

VRonin

@elfring said in Checking the API status for the variable “Private d” of the QVariant class:

You would occasionally prefer direct write access to the contained element.

My point is exactly that implicit sharing removes this need

It can take a while to get acceptance for another useful programming interface, can't it?

And it can take even longer if it's not necessary. Move constructor/assignments are another great example of a feature introduced in the standard that is basically useless when dealing with implicitly shared classes as they do it better and seamlessly.

elfring

My point is exactly that implicit sharing removes this need

This technique depends on the management of object reference counters which can be useful functionality and can trigger related software development challenges.

And it can take even longer if it's not necessary.

Do you care for improvements around generic data processing algorithms?

VRonin

@elfring said in Checking the API status for the variable “Private d” of the QVariant class:

This technique depends on the management of object reference counters

Are we complaining about loss of performance due to an additional integer increment and decrement?

Do you care for improvements around generic data processing algorithms?

Not if they interfere heavily with my objects functionality isolation.

Can I ask, really, what is causing your aversion to how QVariant works? The only scenario I'd imagine is if you are developing microsecond-critical applications (like automatic stock trading) but you mention you are using QStandardItemModel and the inefficiencies inside that model implementation dwarf any moot point about optimising the editing of a QVariant