Added a smart pointers subject to the wiki

[Thaddy]

I added a smart pointer entry to the wiki that describes three different ways to create smartpointers.
Please review.
https://wiki.freepascal.org/smartpointers

Comments and additions are welcome.

[cdbc]

Hi Thaddy
That's a very good write-up, "Me Likey" =^
Clear & to the point -- Well done mate 
Regards Benny

[marcov]

I see RAII like behaviour in these examples and code, but not really ARC.

Nowhere references are passed around and refcounted.

[Thaddy]

Nowhere references are passed around and refcounted.

Well, it works because all three examples are refcounted types. Makes them not only RAII but ARC too.
Or is that not what you mean?

[marcov]

So what is the refcounted type in the advanced record example ? I see a free at the end, but no reference counting, so big fat exception if you add the object to some global list since that will lead to a double free.

The other two interface based once have reference count somewhere, but non of the demos really demonstrate that by increasing references beyond the one instance used locally in that procedure. All three examples seem to only care about skipping finalization code for the simplest case.

That has its place, certainly, (*) but is it really ARC/Smartpointer?

(*) e.g. in web request processing. Anything that simplifies the endless try-finally blocks is welcome.

[Thaddy]

So what is the refcounted type in the advanced record example ? I see a free at the end, but no reference counting, so big fat exception if you add the object to some global list since that will lead to a double free.

The other two interface based once have reference count somewhere, but non of the demos really demonstrate that by increasing references beyond the one instance used locally in that procedure. All three examples seem to only care about skipping finalization code for the simplest case.

That has its place, certainly, (*) but is it really ARC/Smartpointer?

(*) e.g. in web request processing. Anything that simplifies the endless try-finally blocks is welcome.

The managed record itself is refcounted. That's how they work.

[marcov]

Well, I can't seem to get it to work then. I didn't see reference counting when I compiled to assembler (other than for strings), and assigning an instance to another variable of the type crashes at finalization:

Code: [Select]

program smartrec;
{ A simple smart pointer based on advanced records.
  Works for classes with parameterless constructors }
{$mode delphi}
uses classes;
type
  TSmartPointer<T:class, constructor> = record
  FInvoke:T;
  public
  class operator initialize(var value: TSmartPointer<T>);
  class operator finalize(var value: TSmartPointer<T>);
  function invoke:T;
  end;
 
  class operator TSmartPointer<T>.initialize(var value: TSmartPointer<T>);
  begin
    Value.FInvoke := T.Create;
  end;
 
  class operator TSmartPointer<T>.finalize(var value: TSmartPointer<T>);
  begin
    Value.FInvoke.Free;
  end;
 
  function TSmartPointer<T>.invoke:T;
  begin
    Result := FInvoke;
  end;

Type  TSmartStringList=TSmartPointer<TStringlist>;
var
  smList:TSmartStringList; 
  sm2list:TSmartStringList;
  List:TStringlist;
begin
  { Use the list }
  smList.Invoke.Add('some text');
  sm2list:=smlist;                                  // copy instance. If reference counted this should be fine. If only initialized and finalizated it will bomb.
  { It is also possible to assign to an
    uninitialized stringlist variable }
  list := smList.Invoke;
  writeln(List.Text);
  readln;
  { no leaks }
end.


And it bombs.

[jamie]

Something tells me that using the "Class Copy" operator may help in that regard to clean up the existing one before overwriting it with a new one?

Jamie

[marcov]

Something tells me that using the "Class Copy" operator may help in that regard to clean up the existing one before overwriting it with a new one?

But that might make it more dependent of the type you nest in it, and less general. Depending on if it is a class, an interface, an array/record etc. Maybe things are possible (e.g. with gettypekind, ismanagedtype etc), but I feel this is all incomplete.

[Thaddy]

Yes, the record based approach needs addref and copy. I had them in, but left them out. Will  put them back in.
Anyway, I only use the last (ref) solution.

[avra]

I wish you kept naming as in your initial example from 2019:
Auto<> instead of TSmartPointer<>

If FPC ever gets an official unit for smart pointers, I would like to see "auto" naming because I prefer naming by functionality over naming by implementation detail as I find it much more clear.

[Thaddy]

I thought "Auto" to be a good but maybe too C++ like idea. (_auto_ is indeed similar)
Whereas "SmartPointer", although a faux description, seems well settled in the Object Pascal community.
Since I will have to do the repairs on the first entry, I might return to Auto too (or my very first: Box).

[avra]

I thought "Auto" to be a good but maybe too C++ like idea. (_auto_ is indeed similar)

I didn't know that, so not so C++ like, I guess. It simply sounded more natural and more self explaining when looking at examples to a beginner. Smart pointers term needs a little research before understanding.

Just my 2 cents on what resonates better with my ears. 

[Thaddy]

Nehhh, we'll make it auto, it's  us.

[nixbody]

I use function references for objects/pointers with ARC and call the type TShared (inspired by std::shared_ptr) for objects and TSharedPointer for pointers. TAuto resembles to me a variable with automatic storage duration (a C/C++ term I believe) which is a block-scoped variable that cease to exist when goes out of its scope which is more akin to std::unique_ptr (a better version of now deprecated std::auto_ptr).