Increasing usage for C++ new operators based on data model indexes?

kshegunov

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

Must a C++ new operator provide a non-void pointer data type?

It can provide whatever type you want it to, however I still don't see how you're going to have a generic type (i.e. QAbstractItemModel) that deals with concrete types it does know nothing about ...

It provides the generic programming interface for the desired data sources.

Indeed.

Derived classes will contain member variables which will manage storage in expected ways, won't they?

They may or may not. It's up to the user code to decide where and how the storage will be done. There may be no storage at all and the data to be fetched on the fly from someplace.

C++ programmers are using this operator for various resource allocations.

Many C++ programmers overuse this operator, especially those with limited experience.

I propose to increase the usage of the construct “placement new” also together with customised data models.

https://www.youtube.com/watch?v=iUQCFI02zZA

@J.Hilk said in Increasing usage for C++ new operators based on data model indexes?:

Just a question out of curiosity, because I remeber by instructor also telling me c++ heap allocation is slower than stack

It is. It's a call to the heap manager in the OS. The heap manager has to find a free place for your object before it can return you an address ...
A stack allocation is nothing on the other hand - you (rather the compiler) move the stack pointer to the appropriate offset from the stack base pointer and voila - you have memory. That also is the reason that you must know the size of the allocated object at compile time.

and slower than heap allocation in other languages.

Nope!

Is there a different method in c++ to allocation memory instead of the standard new?

You have *alloc from the C runtime, the placement new if you're crazy enough to build your own heap manager (or for some very specific similar purposes), and of course you have my buddy - the stack.

J.Hilk

@kshegunov thanks for the info.

IIRC boost also has special allocation methods for faster instantiation of objects. But I'm not sure, never used that libary much.

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

The concrete target data types can vary while model indexes are resolved to QVariant objects as the default data.
A C++ new operator can provide a known pointer data type, can't it?

Isn't dynamic_cast what you want?

@kshegunov said in Increasing usage for C++ new operators based on data model indexes?:

You have *alloc from the C runtime

Note for passers-by: If you use those to allocate complex types the constructor may/will not be called

elfring

…, however I still don't see how you're going to have a generic type (i.e. QAbstractItemModel) that deals with concrete types it does know nothing about ...

Various data structures are accessible over pointers (or customised indexes?).

Does the class “QVariant” provide also a programming interface for pointer data types?

…, the placement new if you're crazy enough to build your own heap manager (or for some very specific similar purposes), …

I am trying to increase the software development attention for the latter.

elfring

Isn't dynamic_cast what you want?

Yes.

I prefer to avoid null pointers which can eventually be returned by such a cast operation.

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

Does the class “QVariant” provide also a programming interface for pointer data types?

Yes using QVariant::fromValue or via the constructor if the type is a QObject subclass.
hic sunt leones though as you need to make clear in your logic who owns the items pointed. In general a QVariant containing a owning pointer is a ticking time bomb toward memory leak and/or double deletion.

I prefer to avoid null pointers which can eventually be returned by such a cast operation.

If you are certain of the downcast type you can even use static_cast and forget about null pointers

elfring

In general a QVariant containing a owning pointer is a ticking time bomb toward memory leak and/or double deletion.

The safe handling of object lifetimes is a general software development challenge.
But this class is the only way to get data from a model so far, isn't it?

If you are certain of the downcast type you can even use static_cast and forget about null pointers

Can it be more convenient to let a C++ new operator (which can work also with extra allocation parameters directly) perform the desired data type conversion?

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

The safe handling of object lifetimes is a general software development challenge.
But this class is the only way to get data from a model so far, isn't it?

Yes, what I'm saying is that if you want to store a pointer in a QVariant you probably want a non-owning pointer. Keep the owning pointer somewhere else.

Can it be more convenient to let a C++ new operator (which can work also with extra allocation parameters directly) perform the desired data type conversion?

No, new allocates memory. static_cast just tell the compiler to treat a piece of memory as it was a different type.

But this class is the only way to get data from a model so far, isn't it?

And here we come back to the fact that having your objects implicitly shared stored in a QVariant directly is much more practical and has a negligible impact on performance

elfring

new allocates memory.

Only if you do not use the construct “placement new”.

And here we come back to the fact that having your objects implicitly shared stored in a QVariant directly is much more practical

I guess that the run time consequences are more interesting then for the distinction if only pointers are transferred or complete “value objects” are copied.

and has a negligible impact on performance

There are software applications where more data transfers might not have a directly noticeable effect. But I guess that some C++ programmers have got a strong focus on software efficiency.

kshegunov

This is getting pretty ridiculous ...

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

Only if you do not use the construct “placement new”.

Yeah, which I have a creeping suspicion you haven't. The so called placement new is nothing more than going around calling constructors on a preallocated memory block. So where is this preallocated memory going to come from? It's going to materialize from the ether?

If you're intent on pushing this make-you-own-heap-for-models, I advise to create a proof-of-concept first, then we can have something to base a discussion on, otherwise - thanks but no thanks, I'm out of this conversation.

elfring

The so called placement new is nothing more than going around calling constructors on a preallocated memory block.

I agree on this aspect.

But can you pass a “model index” as another allocation parameter (for existing data) to this C++ operator?

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

Only if you do not use the construct “placement new”.

Is your plan to do exactly what dynamic_cast does but using a new operator? I mean it's possible but I don't see why you'd want to do it. How would you implement it?

guess that the run time consequences are more interesting then for the distinction if only pointers are transferred or complete “value objects” are copied.
There are software applications where more data transfers might not have a directly noticeable effect. But I guess that some C++ programmers have got a strong focus on software efficiency.

@VRonin said in Increasing usage for C++ new operators based on data model indexes?:

having your objects implicitly shared

With implicitly shared objects, the distinction from a pointer to a complete value object when copying is just an operator++ on an int. 1 atomic instruction. 100000 times more efficient than a dynamic_cast.
If you want to go into memory difference it's 32bits more. If 32 bits break your design then it's the design itself that comes into question

elfring

Is your plan to do exactly what dynamic_cast does but using a new operator?

Maybe.

But these C++ operations provide different functionality, don't they?

With implicitly shared objects, the distinction from a pointer to a complete value object when copying is just an operator++ on an int.

Will it become nicer to avoid (or reduce) even the influence of object reference counting?

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

Will it become nicer to avoid (or reduce) even the influence of object reference counting?

C++ is moving in the opposite direction actually. If you try to suggest replacing std::shared_ptr (which is a pointer + a reference counter) with raw pointers on Stack Overflow you'd better bring a helmet and kevlar vest because you are going to get shot.
The ISO C++ style guide goes one step further and even discourages the use of owning raw pointer and suggest ownership encapsulation (an implicitly shared object behaves as a smart pointer). If you don't like it, good luck convincing Herb Sutter and Bjarne Stroustrup himself.

But these C++ operations provide different functionality, don't they?

Yes but so far you did not highlight any need for custom allocation, your problems all seem to come from casting.

elfring

but so far you did not highlight any need for custom allocation,

I suggest to reconsider the software situation once more.

A pair of row and column values (data model index, address or coordinate) can be passed as a parameter to a C++ new operator.
You can choose then if you need to work with fresh objects (using dynamic memory allocation) or would like to reuse existing data.

Is the usage of the construct “placement new” a kind of customised operation?

your problems all seem to come from casting.

I hope that specific software development challenges can be adjusted when the new operator call takes care of casting to the desired target (pointer) data type already.

VRonin

@elfring said in Increasing usage for C++ new operators based on data model indexes?:

A pair of row and column values (data model index, address or coordinate) can be passed as a parameter to a C++ new operator.
You can choose then if you need to work with fresh objects (using dynamic memory allocation) or would like to reuse existing data.
Is the usage of the construct “placement new” a kind of customised operation?

Once again, the model implementation might or might not allocate data. This should not be a problem of the API user.

Could you please provide a concrete example of what you are suggesting?
Otherwise we are just discussing of theory and nobody is gaining any value out of it.
I will not continue the discussion until you provide at least a code snippet

elfring

Could you please provide a concrete example of what you are suggesting?

Does the software situation become really more challenging for passing a few extra parameters to a (member) function when it is “accidentally” (or intentionally) called “new”?

elfring

… concrete example of what you are suggesting?

sierdzio

The first link clearly states:

ADVICE: Don’t use this “placement new” syntax unless you have to. Use it only when you really care that an object is placed at a particular location in memory. For example, when your hardware has a memory-mapped I/O timer device, and you want to place a Clock object at that memory location.

In case of model indexes, there seems to be no need to use placement new.

elfring

In case of model indexes, there seems to be no need to use placement new.

Do you care if customised model data were “placed at a particular location in memory”?
How do you think about to get direct access to object data which are managed by a class like QVariant?