[DUG] How's this for inconsistent
Todd
todd.martin.nz at gmail.com
Thu Nov 25 11:18:17 NZDT 2010
>
>
> On a more general point, I'd say that whilst referencing "self" as an
> interface is an almost unavoidable part of any interface based
> framework, the same is not true in the destructor. Anything
> referenced or notified during destruction of an object that requires
> being passed a reference TO that object, could and most likely should,
> have been passed a reference to that object during construction or
> during some later operation.
>
> However, if those other objects were holding on to those references,
> that would prevent the object from being destroyed in the first place.
>
During construction an object may interact with another object via its
interface, so that the contained object does not know what exact object
is calling its methods. There is no reason why it shouldn't do that in
its destructor as well.
>
> Another thing that is missing is a safe, common implementation of a
> weak reference (an interface reference that does not hold on to --
> i.e. does not contribute +1 toward -- the reference count of the
> referenced object). You can use an interface reference cast as a
> Pointer, but the catch is that the reference must be sure that it will
> be notified if the referenced object is destroyed, so that it can NIL
> itself. Using a plain Pointer won't do... you have to wrap it inside
> some object that can register for those notifications (and be sure
> that the object referenced will implement the required notification
> for the wrapper to respond to).
>
This is another scenario which could invoke a call to the container
object's interface.
>
>
> But, that is not a general purpose solution as it requires that
> objects exposing interfaces that may be encapsulated in my weak
> reference implementation also implement IOn_Destroy, which is not
> built into the VCL. In fact, the VCL doesn't have any such general
> purpose system (and neither should it have imho). The
> FreeNotification() mechanism is not supported on TObject, and neither
> imho should it be.
>
I certainly wouldn't advocate putting FreeNotification on TObject. It is
far too convoluted. I was thinking more of an external observer, so that
TObject doesn't even know it is being watched. Since TObject's only task
would be to broadcast a message when it was being destroyed, any
observers would be responsible for severing the link.
>
> Not every application -- or indeed every object - needs these things,
> nor the overhead that it would incur, adding housekeeping code to the
> tear-down of every object.
>
> Most applications simply don't need these sorts of exotica (as
> evidenced by the fact that these sort of facilities either don't exist
> at all or took so long to be introduced into the core VCL, following
> the introduction of interfaces waaaaaay back in Delphi 3 -- Delphi 2
> did things slightly differently you may recall).
>
I started with Delphi 4.
>
> The very beauty of Delphi and the VCL is that we can introduce these
> things into the applications (or indeed the small parts of our
> applications) that need them, without imposing them on everything else.
>
> When we want the convenience of having everything to hand whether you
> want it or not, that's what managed code is for. J
>
> *From:* delphi-bounces at delphi.org.nz
> [mailto:delphi-bounces at delphi.org.nz] *On Behalf Of *Todd
> *Sent:* Wednesday, 24 November 2010 8:08 p.m.
> *To:* NZ Borland Developers Group - Delphi List
> *Subject:* Re: [DUG] How's this for inconsistent
>
>
>
> Yep -- I remember that my fix was to set the "destructing" state
> indicator in a BeforeDestruction() override. This was then tested in
> _Release() to render it a NO-OP during execution of the destructor
> chain (incomplete, obviously, just to give the idea):
>
> Procedure BeforeDestruction;
>
> SetState(csDestroying);
>
> Function _Release;
>
> If csDestroying in State then EXIT;
>
> Upon reflection, that would only work for a TComponent descendant.
>
> Nothing else needs be done, as long as any further BeforeDestruction
> overrides call inherited before doing their work, which they should do
> (in my framework I introduced another virtual to be overridden in my
> descendants, in case there were occasions when work was done to
> generate references during the destructor execution -- even in your
> case, the FreeInstance() override is redundant I think, other than as
> a sanity/safety check and so could be made subject to some conditional
> compilation flag.
>
> *From:* delphi-bounces at delphi.org.nz
> <mailto:delphi-bounces at delphi.org.nz>
> [mailto:delphi-bounces at delphi.org.nz] *On Behalf Of *Todd
> *Sent:* Wednesday, 24 November 2010 6:15 p.m.
> *To:* NZ Borland Developers Group - Delphi List
> *Subject:* Re: [DUG] How's this for inconsistent
>
> Actually, this would be better
>
> function TamObject._Release: Integer;
> begin
> Result := InterlockedDecrement(FCount);
> if (FCount = 0) then
> begin
> //add a reference count, incase an interface is acquired and
> released during destruction
> InterlockedIncrement(FCount);
> self.Destroy;
> end;
> end;
>
> procedure TamObject.FreeInstance;
> begin
> //remove the reference count added in _Release
> InterlockedDecrement(FCount);
> assert(FCount = 0,'Destroying object with non-zero reference count');
> inherited FreeInstance;
> end;
>
>
> I spotted that they fixed that a while ago -- I remember having to fix
> the issue myself many years ago so was quite pleased to see that it
> was now taken care of in TInterfaceObject as a matter of course. For
> some reason I never noticed the omission of the same facility in the
> destructor. And yes, it's a [potentially] big problem.
>
> I need to think about this tho... setting a fake ref count during
> execution of the constructor is safe enough as you know precisely when
> construction is done and to restore the ref count back to zero.
>
> Setting a fake ref count during destruction strikes me as more
> problematic and makes me nervous, but I can't quite put my finger on
> why. It might be nothing. That doesn't mean it can't be fixed, only
> that the solution put in place for construction might not work for
> destruction and it wasn't felt necessary to do any extra work for a
> more comprehensive solution.
>
> Certainly in the case of my code where I fixed this I had specific
> "constructing" / "destructing" state markers (it wasn't a general
> purpose interfacedobject class but a base class in a far richer
> framework that happened to also implement its own version of IUnknown)
> -- I know I didn't rely on side effects of a faked ref count.
>
> *From:* delphi-bounces at delphi.org.nz
> <mailto:delphi-bounces at delphi.org.nz>
> [mailto:delphi-bounces at delphi.org.nz] *On Behalf Of *Todd
> *Sent:* Wednesday, 24 November 2010 16:55
> *To:* NZ Borland Developers Group - Delphi List
> *Subject:* [DUG] How's this for inconsistent
>
> The Delphi developer who implemented TInterfacedObject obviously
> considered the case when an interface reference is grabbed during
> construction......
>
> // Set an implicit refcount so that refcounting
> // during construction won't destroy the object.
> class function TInterfacedObject.NewInstance: TObject;
> begin
> Result := inherited NewInstance;
> TInterfacedObject(Result).FRefCount := 1;
> end;
>
> procedure TInterfacedObject.AfterConstruction;
> begin
> // Release the constructor's implicit refcount
> InterlockedDecrement(FRefCount);
> end;
>
> but didn't consider applying the same logic during destruction. So
> grabing an interface reference during destruction causes all hell to
> break loose, as the _Release method tries to free the object again and
> again recursively.
>
> procedure TInterfacedObject.BeforeDestruction;
> begin
> if RefCount <> 0 then
> Error(reInvalidPtr);
> end;
>
> function TInterfacedObject._Release: Integer;
> begin
> Result := InterlockedDecrement(FRefCount);
> if Result = 0 then
> Destroy;
> end;
>
> Todd.
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post:delphi at delphi.org.nz <mailto:delphi at delphi.org.nz>
> Admin:http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email todelphi-request at delphi.org.nz <mailto:delphi-request at delphi.org.nz> with Subject: unsubscribe
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post:delphi at delphi.org.nz <mailto:delphi at delphi.org.nz>
> Admin:http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email todelphi-request at delphi.org.nz <mailto:delphi-request at delphi.org.nz> with Subject: unsubscribe
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post: delphi at delphi.org.nz
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email to delphi-request at delphi.org.nz with Subject: unsubscribe
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserver.123.net.nz/pipermail/delphi/attachments/20101125/7533e241/attachment-0001.html
More information about the Delphi
mailing list