• Derek Parnell (7/7) Sep 28 2005 Why is a Box a struct and not a class object?
• Burton Radons (7/9) Sep 29 2005 "there aren't any real opportunities for overloading, it's not mutable
• Derek Parnell (12/23) Sep 29 2005 I was playing around with the reference counting class that Regan Heath
• Larry Evans (9/12) Oct 15 2005 [snip]
• Regan Heath (16/28) Oct 15 2005 Attached.
• Larry Evans (22/42) Oct 17 2005 Thanks!
• Regan Heath (24/57) Oct 17 2005 The reference count is not part of the resource because that would requi...
• Larry Evans (16/24) Oct 18 2005 Couldn't:
• Regan Heath (6/30) Oct 18 2005 Perhaps. However, you need to synchronize access to that integer. My cod...
• Larry Evans (37/56) Oct 19 2005 [snip]
• Regan Heath (12/67) Oct 19 2005 What advantage do you see in using 'another' class as opposed to doing
• Larry Evans (14/55) Oct 19 2005 It makes the purpose of the RefPtr members clearer. I was confused,
• Regan Heath (6/52) Oct 19 2005 Nah. I would re-organise as follows:
• Larry Evans (8/26) Oct 19 2005 Now, I'm guessing a "deep copy" of the resource would be is done with:
• Regan Heath (7/29) Oct 19 2005 No. D has no built in deep copy functionality. Instead it's customary to...

Derek Parnell <derek psych.ward> writes:

Why is a Box a struct and not a class object?

It sure makes generic coding more difficult.

-- 
Derek
(skype: derek.j.parnell)
Melbourne, Australia
29/09/2005 4:31:23 PM

Burton Radons <burton-radons smocky.com> writes:

Derek Parnell wrote:
 Why is a Box a struct and not a class object?

"there aren't any real opportunities for overloading, it's not mutable 
anyway, and many people will be repelled by the allocation".  Boxing 
necessarily resulting in an allocation is one of the reasons people 


 It sure makes generic coding more difficult.

Could you give an example?  I can't even begin to say anything without 
knowing what you're trying to do.

Derek Parnell <derek psych.ward> writes:

On Thu, 29 Sep 2005 13:13:37 -0700, Burton Radons wrote:

 Derek Parnell wrote:
 Why is a Box a struct and not a class object?

 
 "there aren't any real opportunities for overloading, it's not mutable 
 anyway, and many people will be repelled by the allocation".  Boxing 
 necessarily resulting in an allocation is one of the reasons people 


Okay, that makes good sense.

 It sure makes generic coding more difficult.

 
 Could you give an example?  I can't even begin to say anything without 
 knowing what you're trying to do.

I was playing around with the reference counting class that Regan Heath
supplied in an earlier post. In that case, the only things that could be
reference-counted were class instances. In order to count strings or
numbers I first tried to 'box' them without realizing that Box was a
struct. Of course that failed. Modifying Regan's code to allow things other
than objects was a PITA.

-- 
Derek Parnell
Melbourne, Australia
30/09/2005 6:39:35 AM

Larry Evans <cppljevans cos-internet.com> writes:

On 09/29/2005 03:46 PM, Derek Parnell wrote:
[snip]
 I was playing around with the reference counting class that Regan Heath
 supplied in an earlier post. In that case, the only things that could be

[snip]
 

How can I find Regan's reference counting class, please?
Would this class contradict the statement:

   it is impossible to implement a useful smart_ptr in D.

from:

   http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/29135

?

"Regan Heath" <regan netwin.co.nz> writes:

On Sat, 15 Oct 2005 16:38:04 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 09/29/2005 03:46 PM, Derek Parnell wrote:
 [snip]
 I was playing around with the reference counting class that Regan Heath
 supplied in an earlier post. In that case, the only things that could be

 [snip]

 How can I find Regan's reference counting class, please?

Attached.

 Would this class contradict the statement:

    it is impossible to implement a useful smart_ptr in D.

 from:

    http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/29135

 ?

That's debatable. It does 'work' but it's not as "fool proof" as you'd  
like it to be, in other words it is more likely you (by this I mean anyone  
using it) will accidently missuse it than a version written in say  
C++ (which allows you you overload the assignment operator)

For example, you must remember to 'assign' a new RefPtr with:

   RefPtr p = new RefPtr(old_one);

you cannot simply say:

   RefPtr p = old_one;

which means that when adding RefPtr's to an existing app you have to find  
all assignments of all references to the shared resource and replace them  
with constructor calls like those above. This makes it significantly more  
work, and significantly more prone to error.

Regan

Larry Evans <cppljevans cos-internet.com> writes:

On 10/15/2005 05:05 PM, Regan Heath wrote:
 On Sat, 15 Oct 2005 16:38:04 -0500, Larry Evans  

[snip]
 How can I find Regan's reference counting class, please?

 
 Attached.

Thanks!

[snip]
 anyone  using it) will accidently missuse it than a version written in 
 say  C++ (which allows you you overload the assignment operator)
 
 For example, you must remember to 'assign' a new RefPtr with:
 
   RefPtr p = new RefPtr(old_one);
 
 you cannot simply say:
 
   RefPtr p = old_one;
 
 which means that when adding RefPtr's to an existing app you have to 
 find  all assignments of all references to the shared resource and 
 replace them  with constructor calls like those above. This makes it 
 significantly more  work, and significantly more prone to error.
 

I'll certainly miss that from c++ :(

There are two questions I have about refcnt.d:


1) The reference count, declared as:

    int refs = 0;

at refcnt.d:9, is part of the RefPtr, not the
resource:

    Object resource = null;

at recfnt.d:10.  Hence, it's a detached refcount.
However, it's declared as a non-pointer type, and
so it cannot be shared between 2 RefPtr's pointing
to the same resource.  What am I missing?

2) There's the RefPtr field:

    RefPtr parent;

What is the purpose of parent?  I can't think of
a counterpart in c++ code.


Thanks for any clarification.

-best regards,
Larry

"Regan Heath" <regan netwin.co.nz> writes:

On Mon, 17 Oct 2005 12:21:38 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 10/15/2005 05:05 PM, Regan Heath wrote:
 On Sat, 15 Oct 2005 16:38:04 -0500, Larry Evans

 [snip]
 How can I find Regan's reference counting class, please?

  Attached.

 Thanks!

 [snip]
 anyone  using it) will accidently missuse it than a version written in  
 say  C++ (which allows you you overload the assignment operator)
  For example, you must remember to 'assign' a new RefPtr with:
    RefPtr p = new RefPtr(old_one);
  you cannot simply say:
    RefPtr p = old_one;
  which means that when adding RefPtr's to an existing app you have to  
 find  all assignments of all references to the shared resource and  
 replace them  with constructor calls like those above. This makes it  
 significantly more  work, and significantly more prone to error.

 I'll certainly miss that from c++ :(

 There are two questions I have about refcnt.d:


 1) The reference count, declared as:

     int refs = 0;

 at refcnt.d:9, is part of the RefPtr, not the
 resource:

     Object resource = null;

 at recfnt.d:10.  Hence, it's a detached refcount.

The reference count is not part of the resource because that would require  
all countable resources to be derived from a common base, D has single  
inheritance so this would be very limiting. The idea was to keep the  
reference couting seperate from the resource itself. It is likely possible  
to create a mixin of some sort and derive a countable class from the class  
you want to count using the mixin. I am not sure what the benefit of that  
might be.

 However, it's declared as a non-pointer type, and
 so it cannot be shared between 2 RefPtr's pointing
 to the same resource.  What am I missing?

The count is shared using the 'parent' reference. This is why you must  
create a RefCnt using an existing RefCnt (and not the resource itself) in  
all cases besides the first. That is why I used a static ctor to create  
the initial RefCnt, because that must be used in place of the resource for  
the rest of the program.

 2) There's the RefPtr field:

     RefPtr parent;

 What is the purpose of parent?  I can't think of
 a counterpart in c++ code.

There likely isn't. This was the only way I could think of to share the  
reference count and ensure it was incremented and decremented at the  
appropriate times.

Note also 'auto' or explicit delete is required on all RefCnt references  
or the count will not be decremented when that reference is collected by  
the GC.

It is possible my implementation is naive. To be honest I am not an expert  
on reference counting, I created this as an example of what was possible  
in D, to show how deficient D was in this area.

Regan

Larry Evans <cppljevans cos-internet.com> writes:

On 10/17/2005 04:01 PM, Regan Heath wrote:
 On Mon, 17 Oct 2005 12:21:38 -0500, Larry Evans  

[snip]
 The reference count is not part of the resource because that would 
 require  all countable resources to be derived from a common base, D has 
 single  inheritance so this would be very limiting. The idea was to keep 
 the  reference couting seperate from the resource itself. It is likely 
 possible  to create a mixin of some sort and derive a countable class 
 from the class  you want to count using the mixin. I am not sure what 
 the benefit of that  might be.

Couldn't:

   int refs = 0;

be changed to:

   int* refs = null;

Then, change this(RefPtr rhs) to:

	this(RefPtr rhs)
	{
		resource = rhs.resource;
		refs = rhs.refs;
                 (*refs)++;
	}

and eliminate:

   RefPtr parent;

altogether?

"Regan Heath" <regan netwin.co.nz> writes:

On Tue, 18 Oct 2005 11:32:49 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 10/17/2005 04:01 PM, Regan Heath wrote:
 On Mon, 17 Oct 2005 12:21:38 -0500, Larry Evans

 [snip]
 The reference count is not part of the resource because that would  
 require  all countable resources to be derived from a common base, D  
 has single  inheritance so this would be very limiting. The idea was to  
 keep the  reference couting seperate from the resource itself. It is  
 likely possible  to create a mixin of some sort and derive a countable  
 class from the class  you want to count using the mixin. I am not sure  
 what the benefit of that  might be.

 Couldn't:

    int refs = 0;

 be changed to:

    int* refs = null;

 Then, change this(RefPtr rhs) to:

 	this(RefPtr rhs)
 	{
 		resource = rhs.resource;
 		refs = rhs.refs;
                  (*refs)++;
 	}

 and eliminate:

    RefPtr parent;

 altogether?

Perhaps. However, you need to synchronize access to that integer. My code  
synchronizes on the class, you would have to use some sort of  
InterlockedIncrement function or similar.

Regan

Larry Evans <cppljevans cos-internet.com> writes:

On 10/18/2005 03:08 PM, Regan Heath wrote:
 On Tue, 18 Oct 2005 11:32:49 -0500, Larry Evans  

[snip]
 Couldn't:

    int refs = 0;

 be changed to:

    int* refs = null;


[snip]
 and eliminate:

    RefPtr parent;

 altogether?

 
 
 Perhaps. However, you need to synchronize access to that integer. My 
 code  synchronizes on the class, you would have to use some sort of  
 InterlockedIncrement function or similar.

Ah!  I hadn't considered synchronization.  But then couldn't the
refs be encapsulated in another class to do this:

class RefPtr
{
   RefPtr parent;
   Object resource = null;

   class RefCount
   {
     int num_refs = 0;

     int increment()
     {
        //sychronized increment
     }

     int decrement()
     {
        //synchronized decrement
     }


   }//end RefCount class

   RefCount refs = null;

   ...

}//end RefPtr class

OK, now I'm probably showing my ignorance of how D differs from C++.  My
understanding is that, since RefCount is a class, it's allocated on
the heap.  Also, based on the original code, which contained:

   resource = res;

I'm assuming that the assert in the following code passes:

   this(RefPtr rhs)
   {
     resource = rhs.resource;
     refs = rhs.refs;
     refs.increment();
     assert(refs is rhs.refs);
   }

IOW, the reference count is shared between rhs and this.

"Regan Heath" <regan netwin.co.nz> writes:

On Wed, 19 Oct 2005 10:33:07 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 10/18/2005 03:08 PM, Regan Heath wrote:
 On Tue, 18 Oct 2005 11:32:49 -0500, Larry Evans

 [snip]
 Couldn't:

    int refs = 0;

 be changed to:

    int* refs = null;


 [snip]
 and eliminate:

    RefPtr parent;

 altogether?

   Perhaps. However, you need to synchronize access to that integer. My  
 code  synchronizes on the class, you would have to use some sort of   
 InterlockedIncrement function or similar.

 Ah!  I hadn't considered synchronization. But then couldn't the
 refs be encapsulated in another class to do this:

What advantage do you see in using 'another' class as opposed to doing  
what I did?

 class RefPtr
 {
    RefPtr parent;
    Object resource = null;

    class RefCount
    {
      int num_refs = 0;

      int increment()
      {
         //sychronized increment
      }

      int decrement()
      {
         //synchronized decrement
      }


    }//end RefCount class

    RefCount refs = null;

    ...

 }//end RefPtr class

The only difference I see is that the shared reference count no longer  
contains the reference to resource. To my mind the count and the reference  
to the resource belong together. That said, my implementation did have an  
extra reference to the resource hanging around.

 OK, now I'm probably showing my ignorance of how D differs from C++.  My
 understanding is that, since RefCount is a class, it's allocated on
 the heap.

Correct.

 Also, based on the original code, which contained:

    resource = res;

 I'm assuming that the assert in the following code passes:

    this(RefPtr rhs)
    {
      resource = rhs.resource;
      refs = rhs.refs;
      refs.increment();
      assert(refs is rhs.refs);
    }

Yes, it should.

 IOW, the reference count is shared between rhs and this.

Yes.

Regan

Larry Evans <cppljevans cos-internet.com> writes:

On 10/19/2005 03:43 PM, Regan Heath wrote:
 On Wed, 19 Oct 2005 10:33:07 -0500, Larry Evans  

[snip]
 Ah!  I hadn't considered synchronization. But then couldn't the
 refs be encapsulated in another class to do this:

 
 What advantage do you see in using 'another' class as opposed to doing  
 what I did?

It makes the purpose of the RefPtr members clearer.  I was confused,
initially, about the purpose of parent.  Maybe my OOPS (see next few
lines) obscured this advantage.
 
 class RefPtr
 {
    RefPtr parent;


OOPS.  The above should be removed.
    Object resource = null;

    class RefCount
    {
      int num_refs = 0;

      int increment()
      {
         //sychronized increment
      }

      int decrement()
      {
         //synchronized decrement
      }


    }//end RefCount class

    RefCount refs = null;

    ...

 }//end RefPtr class

 
 
 The only difference I see is that the shared reference count no longer  
 contains the reference to resource. To my mind the count and the 
 reference  to the resource belong together. 

Well, they are, in RefPtr.  OTOH, I can see where you could put them
together in RefCount and rename it to something like ResourceCount,
which would essentially be the type of your parent.  Then you could
just remove RefPtr.resource since it's now
RefPtr.refs.resource.  But then that's an extra indirection
which, AFAICT, has no advantage.

 That said, my implementation 
 did have an  extra reference to the resource hanging around.

And that extra reference was confusing to me as well as
taking extra space.

"Regan Heath" <regan netwin.co.nz> writes:

On Wed, 19 Oct 2005 16:06:02 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 10/19/2005 03:43 PM, Regan Heath wrote:
 On Wed, 19 Oct 2005 10:33:07 -0500, Larry Evans

 [snip]
 Ah!  I hadn't considered synchronization. But then couldn't the
 refs be encapsulated in another class to do this:

  What advantage do you see in using 'another' class as opposed to  
 doing  what I did?


 It makes the purpose of the RefPtr members clearer.  I was confused,
 initially, about the purpose of parent.  Maybe my OOPS (see next few
 lines) obscured this advantage.

Nah. I would re-organise as follows:

 class RefPtr
 {
    class ResCount
    {
      Object resource = null;
      int num_refs = 0;

      int increment()
      {
         //sychronized increment
      }

      int decrement()
      {
         //synchronized decrement
      }
    }//end ResCount class

    ResCount ref = null;

    ...

 }//end RefPtr class

   The only difference I see is that the shared reference count no  
 longer  contains the reference to resource. To my mind the count and  
 the reference  to the resource belong together.

 Well, they are, in RefPtr.  OTOH, I can see where you could put them
 together in RefCount and rename it to something like ResourceCount,
 which would essentially be the type of your parent.  Then you could
 just remove RefPtr.resource since it's now
 RefPtr.refs.resource.  But then that's an extra indirection
 which, AFAICT, has no advantage.

Indeed, I have done just that above.

 That said, my implementation did have an  extra reference to the  
 resource hanging around.

 And that extra reference was confusing to me as well as
 taking extra space.

True.

Regan

Larry Evans <cppljevans cos-internet.com> writes:

On 10/19/2005 03:43 PM, Regan Heath wrote:
 On Wed, 19 Oct 2005 10:33:07 -0500, Larry Evans  

[snip]
 I'm assuming that the assert in the following code passes:

    this(RefPtr rhs)
    {
      resource = rhs.resource;
      refs = rhs.refs;
      refs.increment();
      assert(refs is rhs.refs);
    }

 
 
 Yes, it should.
 
 IOW, the reference count is shared between rhs and this.

 
 
 Yes.

Now, I'm guessing a "deep copy" of the resource would be is done with:

     resource = new Object(rhs.resource)

although this isn't relevant to the current discusssion about
smart pointers.  I'm just checking my understanding.  Or maybe
the above just does a "level-1" copy.  Maybe nows the time
to actually try some D and see ;)

"Regan Heath" <regan netwin.co.nz> writes:

On Wed, 19 Oct 2005 16:22:56 -0500, Larry Evans  
<cppljevans cos-internet.com> wrote:
 On 10/19/2005 03:43 PM, Regan Heath wrote:
 On Wed, 19 Oct 2005 10:33:07 -0500, Larry Evans

 [snip]
 I'm assuming that the assert in the following code passes:

    this(RefPtr rhs)
    {
      resource = rhs.resource;
      refs = rhs.refs;
      refs.increment();
      assert(refs is rhs.refs);
    }

   Yes, it should.

 IOW, the reference count is shared between rhs and this.

   Yes.

 Now, I'm guessing a "deep copy" of the resource would be is done with:

      resource = new Object(rhs.resource)

No. D has no built in deep copy functionality. Instead it's customary to  
define a 'dup' method the class, so the above would read:

resource = rhs.resource.dup();

 although this isn't relevant to the current discusssion about
 smart pointers.  I'm just checking my understanding.  Or maybe
 the above just does a "level-1" copy.  Maybe nows the time
 to actually try some D and see ;)

Sounds like a plan. :)

Regan