December 04, 2004
Thanks, Pragma, for the introduction :-)

The status is that signals and slots as defined by dcouple are currently usable in single-threaded applications. The declaration of signals is a tad more verbose than for example Trolltech's Qt framework, but that is just syntactic sugar. Connecting/disconnecting them looks actually nicer as compared to Qt, and since signals and slots get connected more often than declared, we are not too far off.

I should clean up the code a bit, because there are some work-arounds in there for compiler limitations that have since been fixed. On the todo-list are thread-safety, for which I see no problems. I am also interested in benchmarks against C++ implementations.

The big hindrance I see before we can call dcouple a state-of-the-art signal and slots implementation, is run-time identification of signals and slots in dynamically loaded classes. This is necessary for visual GUI designers/builders. Other properties in the classes will also need this kind of RTTI. This still needs some thinking, and possibly extension of the D language.

A language feature that has been discussed before, called "structed classes" (thread "Dropping the distinction between objects and references may hinder performance", started 10/18/04 23:12), if it ever gets implemented, would take away a good deal of the current complexity of dcouple, and improve performance.

By all means, check out dcouple and share your thoughts!
http://svn.dsource.org/svn/projects/dcouple/trunk/managed/doc/index.html

Bastiaan.


pragma wrote:
> I thought this sounded familiar, so I went looking.  Apparently, it's already
> out there.
> 
> http://www.dsource.org/projects/dcouple/
> 
> From the Readme.txt in the svn/trunk:
> 
> The dcouple project is a place for experimenting towards an effective
> implementation of the "signal and slot" call-back concept in D. As the basics
> have already been illustrated by others, the main contribution of this project
> will be usage issues (syntax) and performance in a garbage collected system. The
> ability to break all connections to a particular slot or from a particular
> signal is essential in this respect, and is reflected in the name of this
> project. Hopefully, a library will result that can be included into some
> "standard" library (phobos, deimos, phoenix, whatever) and form the basis of a
> Qt-like GUI library. 
> 
> In article <conii3$tkv$1@digitaldaemon.com>, Georg Wrede says...
> 
>>Hi!
>>
>>How hard would it be to have signals&slots included in the D spec?
>>
>>An intro at http://doc.trolltech.com/3.3/signalsandslots.html describes the idea as far as Qt is concerned, but this might be a generally useful thing in any OO language?
>>
>>Since they have implemented this with a couple of preprocessor definitions,
>>it probably wouldn't be too hard to incorporate in the D language core?
>>
>>Code shops usually have their staff writing stock classes between assignments,
>>and having signals&slots would make these classes handier to use. And large
>>programming projects would benefit from looser coupling between objects.
>>
>>I'm not saying SS should be used _instead_ of regular object methods, just
>>that they add value and versatility to a class. And adding some SS to already existing classes could make them usable in new contexts.
>>
>>(I know, I know) this might be a little too late for D 1.0. :-(
>>
>>Then again, having SS in D 1.0 could make writing usable libraries easier.
>>By D 2.0 we'd have a real mean package!
>>
>>Writing games, UI, multithreaded, real-time, control, OS, ... you name it,
>>SS would be of help.
>>
>>
>>
> 
> 
> - Pragma
December 04, 2004
In article <cor3q6$1ls$1@digitaldaemon.com>, Russ Lewis says...
>Georg Wrede wrote:
..
>That would require that the dying object know about all delegates which point to it, and have the ability to zero out their pointers.  That's

I take it that you have connected all sending objects with delegates to all those interested in their signals? Thus, 100 sending objects emitting 10 different signals, and 100 receiving objects, each interested in all of them, gives 100*100*10 = 100_000 connections. Additionally, using delegates, every sending object would "know" the recipient?

If all signals are sent via SS Dispatch, then we'd have 100*10 + 10*100 = 2000 connections, only 2% of the former.

If the pointers in SS Dispatch are not reference counted, then SS objects should signal it from their destructor. A semaphore in the Dispatch table could guarantee that the destructor does not return before same-time signals have returned.

BTW, the Dispatch is only interested in signal receiving objects.

>doable, but again, it's complexity better left to an (optional) library.

I agree that new things should be made library, and only later considered for inclusion into the language. But I can't figure out how you could end up with a convenient syntax for SS then. Example:

#   // (C++ from QT)
#   class Foo : public QObject
#   {
#       Q_OBJECT
#   public:
#       Foo();
#       int value() const { return val; }
#   public slots:
#       void setValue( int );
#   signals:
#       void valueChanged( int );
#   private:
#       int val;
#   };
#

Of course, one could always run the code through a preprocessor, but that would be against the spirit of D, right?

>> Race Conditions -- how does D currently handle simultaneous calling of a method in a dying object? How, or whether, it is currently taken care of, is good enough for SS. Nothing more is needed.
>
>Simple.  We don't GC an object until all references are gone, so that never becomes an issue.
>
>If you manually delete an object, then you are responsible to have already cleaned up all of the lingering references...which is exactly like the old C/C++ way of doing things.

Hmm. This suggests that (if not else, then for Asserts to be possible) you can ask the GC how many references currently point to an object. Like when you think there's only one reference to an object and it turns out there are 5.


December 04, 2004
"Georg Wrede" <Georg_member@pathlink.com> wrote in message news:conii3$tkv$1@digitaldaemon.com...
> Hi!
>
> How hard would it be to have signals&slots included in the D spec?
>
> An intro at http://doc.trolltech.com/3.3/signalsandslots.html describes the idea as far as Qt is concerned, but this might be a generally useful thing in any OO language?
>
> Since they have implemented this with a couple of preprocessor
> definitions,
> it probably wouldn't be too hard to incorporate in the D language core?
>
> Code shops usually have their staff writing stock classes between
> assignments,
> and having signals&slots would make these classes handier to use. And
> large
> programming projects would benefit from looser coupling between objects.
>
> I'm not saying SS should be used _instead_ of regular object methods, just that they add value and versatility to a class. And adding some SS to already existing classes could make them usable in new contexts.
>
> (I know, I know) this might be a little too late for D 1.0. :-(
>
> Then again, having SS in D 1.0 could make writing usable libraries easier. By D 2.0 we'd have a real mean package!
>
> Writing games, UI, multithreaded, real-time, control, OS, ... you name it, SS would be of help.
>
>

Delegates come close but they only fire a single callback instead of
allowing multiple callbacks. This thread got me thinking about multicast
delegates (ala C#) so I whipped up another chunk of D code to manage
multiple delegates. Here's the example from the supplied Qt page modified to
use multi-delegates, with output:
  class Foo {
    MultiDelegate!(int) valueChanged;
    private int val;
    int value() {
      return val;
    }
    void value(int v) {
      if (val != v) {
        val = v;
        valueChanged(v);
      }
    }
  }

  import std.stdio;
  int main() {
    Foo a = new Foo;
    Foo b = new Foo;
    a.valueChanged ~= &b.value;
    writefln(a.value," ",b.value);
    b.value = 11;
    writefln(a.value," ",b.value);
    a.value = 79;
    writefln(a.value," ",b.value);
    a.valueChanged.remove(&b.value);
    a.value = 100;
    writefln(a.value," ",b.value);
    return 0;
  }

output:

C:\d>dmd multidg.d -version=MultiDgExample C:\dmd\bin\..\..\dm\bin\link.exe multidg,,,user32+kernel32/noi;

C:\d>multidg
0 0
0 11
79 79
100 79

It's probably almost the same as other implementations only I use dynamic
arrays instead of associative arrays since:
1) dynamic array take up less space and are faster and most multi-delegates
will have 1 or 2 delegates attached
2) the order of execution of the delegates matches the order in which they
were attached.

anyhow, enjoy!

-Ben




December 04, 2004
"Georg Wrede" <Georg_member@pathlink.com> wrote in message news:cosgvu$23g0$1@digitaldaemon.com...
> In article <cor3q6$1ls$1@digitaldaemon.com>, Russ Lewis says...
>>Georg Wrede wrote:
> ..
>>That would require that the dying object know about all delegates which point to it, and have the ability to zero out their pointers.  That's
>
> I take it that you have connected all sending objects with delegates to all those interested in their signals? Thus, 100 sending objects emitting 10 different signals, and 100 receiving objects, each interested in all of them, gives 100*100*10 = 100_000 connections. Additionally, using delegates, every sending object would "know" the recipient?
>
> If all signals are sent via SS Dispatch, then we'd have 100*10 + 10*100 = 2000 connections, only 2% of the former.

I don't know how SS stores the connection information so it is possible it compresses it for special cases but I would guess in reality the set of connections is not very compressable. In my experience you typically only have 0, 1 or 2 objects listening for a given signal from a given object so who cares how it compresses?

> If the pointers in SS Dispatch are not reference counted, then SS objects should signal it from their destructor. A semaphore in the Dispatch table could guarantee that the destructor does not return before same-time signals have returned.
>
> BTW, the Dispatch is only interested in signal receiving objects.
>
>>doable, but again, it's complexity better left to an (optional) library.
>
> I agree that new things should be made library, and only later considered for inclusion into the language. But I can't figure out how you could end up with a convenient syntax for SS then. Example:
>
> #   // (C++ from QT)
> #   class Foo : public QObject
> #   {
> #       Q_OBJECT
> #   public:
> #       Foo();
> #       int value() const { return val; }
> #   public slots:
> #       void setValue( int );
> #   signals:
> #       void valueChanged( int );
> #   private:
> #       int val;
> #   };
> #
>
> Of course, one could always run the code through a preprocessor, but that would be against the spirit of D, right?

see my post using multicast delegates. I think the user code is cleaner than the SS example.

>>> Race Conditions -- how does D currently handle simultaneous calling of a method in a dying object? How, or whether, it is currently taken care of, is good enough for SS. Nothing more is needed.
>>
>>Simple.  We don't GC an object until all references are gone, so that never becomes an issue.
>>
>>If you manually delete an object, then you are responsible to have already cleaned up all of the lingering references...which is exactly like the old C/C++ way of doing things.
>
> Hmm. This suggests that (if not else, then for Asserts to be possible) you can ask the GC how many references currently point to an object. Like when you think there's only one reference to an object and it turns out there are 5.

There isn't an API to ask the GC how many references point to an object. The current implementation of the GC uses a mark-sweep algorithm so it only figures out if an object has live references during the GC run and otherwise an object has no idea if it is live or not - kindof a bummer to not know if you are alive but so it goes :-)

-Ben


December 04, 2004
So you create an intermediate object between the various senders and the various receievers.  My argument still holds: whoever calls the receivers (in this case, the intermediate object) needs to get *immediate* notification when an object dies so as to prevent race conditions.

Anyhow, it seems that you're missing the point of what I wrote.  I think that SS sound really cool!  I would like to try them out!  But they are complex enough that they should not be a mandatory part of the language. 
 They should be implemented in a library for now.  Years later, if after long experience with it in a library, we all agree that it should be part of the next generation language, then we'll do it then.

Russ

Georg Wrede wrote:
> In article <cor3q6$1ls$1@digitaldaemon.com>, Russ Lewis says...
> 
>>Georg Wrede wrote:
> 
> ..
> 
>>That would require that the dying object know about all delegates which point to it, and have the ability to zero out their pointers.  That's 
> 
> 
> I take it that you have connected all sending objects with delegates
> to all those interested in their signals? Thus, 100 sending objects
> emitting 10 different signals, and 100 receiving objects, each interested in all of them, gives 100*100*10 = 100_000 connections.
> Additionally, using delegates, every sending object would "know" the recipient?
> 
> If all signals are sent via SS Dispatch, then we'd have
> 100*10 + 10*100 = 2000 connections, only 2% of the former.
> 
> If the pointers in SS Dispatch are not reference counted, then
> SS objects should signal it from their destructor. A semaphore
> in the Dispatch table could guarantee that the destructor does
> not return before same-time signals have returned.
> 
> BTW, the Dispatch is only interested in signal receiving objects.
> 
> 
>>doable, but again, it's complexity better left to an (optional) library.
> 
> 
> I agree that new things should be made library, and only later
> considered for inclusion into the language. But I can't figure out
> how you could end up with a convenient syntax for SS then. Example:
> 
> #   // (C++ from QT)
> #   class Foo : public QObject
> #   {
> #       Q_OBJECT
> #   public:
> #       Foo();
> #       int value() const { return val; }
> #   public slots:
> #       void setValue( int );
> #   signals:
> #       void valueChanged( int );
> #   private:
> #       int val;
> #   };
> #
> 
> Of course, one could always run the code through a preprocessor,
> but that would be against the spirit of D, right?
> 
> 
>>>Race Conditions -- how does D currently handle simultaneous calling
>>>of a method in a dying object? How, or whether, it is currently taken care of, is good enough for SS. Nothing more is needed.
>>
>>Simple.  We don't GC an object until all references are gone, so that never becomes an issue.
>>
>>If you manually delete an object, then you are responsible to have already cleaned up all of the lingering references...which is exactly like the old C/C++ way of doing things.
> 
> 
> Hmm. This suggests that (if not else, then for Asserts to be possible)
> you can ask the GC how many references currently point to an object.
> Like when you think there's only one reference to an object and it
> turns out there are 5.
> 
> 

December 05, 2004
In article <conii3$tkv$1@digitaldaemon.com>, Georg Wrede says...
>
>How hard would it be to have signals&slots included in the D spec?
>

Hi!
Walter is in trouble because he has two "twins" in D: function pointers and
delegates. He doesn't know what to do with them. He doesn't understand S&S well,
they OOP-ish and power. There is must be delegates and S&S instead of function
pointers and delegates. But Walter still love C too much... shame! He froze any
move in this way.
:-(


December 07, 2004
In article <cosich$255j$1@digitaldaemon.com>, Ben Hinkle says...
>
>
>"Georg Wrede" <Georg_member@pathlink.com> wrote in message news:conii3$tkv$1@digitaldaemon.com...
>> Hi!
>>
>> How hard would it be to have signals&slots included in the D spec?
>output:
>
>C:\d>dmd multidg.d -version=MultiDgExample C:\dmd\bin\..\..\dm\bin\link.exe multidg,,,user32+kernel32/noi;
>
>C:\d>multidg
>0 0
>0 11
>79 79
>100 79
>
>It's probably almost the same as other implementations only I use dynamic
>arrays instead of associative arrays since:
>1) dynamic array take up less space and are faster and most multi-delegates
>will have 1 or 2 delegates attached
>2) the order of execution of the delegates matches the order in which they
>were attached.
>
>anyhow, enjoy!
>
>-Ben

Cool code! I'll use this.



December 07, 2004
In article <cot4qh$2soj$1@digitaldaemon.com>, Russ Lewis says...
>
>Anyhow, it seems that you're missing the point of what I wrote.  I think that SS sound really cool!  I would like to try them out!  But they are complex enough that they should not be a mandatory part of the language.
>
>They should be implemented in a library for now.  Years later, if after long experience with it in a library, we all agree that it should be part of the next generation language, then we'll do it then.

What can I say, I'm getting convinced!

Somehow this reminds me of the times everyone did their own linked lists and trees again and again. It sure took a long time for mainstream languages to render that daily chore obsolete.

Actually, the only thing why I thought they should be in the language core was to get a nice syntax. But on second thought that is a minor issue.



December 09, 2004
Dr.Dizel wrote:
> In article <conii3$tkv$1@digitaldaemon.com>, Georg Wrede says...
> 
>>How hard would it be to have signals&slots included in the D spec?
>>
> 
> 
> Hi!
> Walter is in trouble because he has two "twins" in D: function pointers and
> delegates. He doesn't know what to do with them. He doesn't understand S&S well,
> they OOP-ish and power. There is must be delegates and S&S instead of function
> pointers and delegates. But Walter still love C too much... shame! He froze any
> move in this way.
> :-(
> 
> 

Sorry Dizel, you must be trolling.
December 10, 2004
Dr.Dizel wrote:

> In article <conii3$tkv$1@digitaldaemon.com>, Georg Wrede says...
>>
>>How hard would it be to have signals&slots included in the D spec?
>>
> 
> Hi!
> Walter is in trouble because he has two "twins" in D: function pointers
> and delegates. He doesn't know what to do with them. He doesn't understand
> S&S well, they OOP-ish and power. There is must be delegates and S&S
> instead of function pointers and delegates. But Walter still love C too
> much... shame! He froze any move in this way.
> :-(

Function pointers are still very useful for interfacing with C projects. They serve a different purpose than delegates.  I think you're hacking away at a straw man.
Next ›   Last »
1 2
Top | Discussion index | About this forum | D home