Phoenix Singletons

[stephan beal]

Good morning, C++ users,

i've been hesitating to post this, primarily because i know that most of you
here are *way* ahead of me in C++ and i'm a little embarassed about the
possibility of some gross errors in what i'll be posting...

That said, please go easy on me. :)

About 2 weeks ago i wrote a paper, called Context Singletons, where i
discuss some of the uses of context-specific Singleton-like objects. During
the development of the paper i came up with a Phoenix class for providing
Phoenixing behaviour (as defined in Alexandrescu's MC++D) to my Context
Singletons. (http://s11n.net/misccode/)

So what, right?

i'm here to ask the Gurus to please take a look at my Phoenix, and see if
i'm doing something Inherently Wrong here. My gut tells me that the
implementation is simply too straightforward to solve the problem
adequately, but practice shows that it works quite well :/.

The source code is here:

template < typename BaseType, typename ContextType = BaseType >
class phoenix : public BaseType {
public:
typedef ContextType context_type;
typedef BaseType base_type;
static base_type & instance() {
static this_type meyers;
if( this_type::m_destroyed ) {
new( &meyers ) this_type;
atexit( this_type::do_atexit );
}
return meyers;
}
private:
typedef phoenix<base_type,context_type> this_type;
static bool m_destroyed;
phoenix() { m_destroyed = false; }
~phoenix() { m_destroyed = true; }
static void do_atexit() {
if( m_destroyed ) return;
static_cast<this_type SPMamp;>(instance()).~phoenix();
}
};
template <typename T, typename C> bool phoenix<T,C>::m_destroyed = false;

This code is complete for purposes of the questions posed here, but the code
at http://s11n.net/misccode/phoenix.h is "more complete", with docs,
comments, one additional (optional) template parameter and debugging code.

Regarding the implementation, i have these questions:

- What happens when BaseType has a non-virtual destructor? We never use
delete(), relying on static destruction instead, so i think this is safe. i
have successfully used this class with, e.g., std::map<> as BaseType, but
maybe i just have a remarkably accomodating C++ platform.

- Is it possible that m_destroyed gets destroyed at a point such that the
above usages of it becomes invalid? If so, is there a way to phoenix that
bool, so to speak?
Any insights on these points would be greatly appreciated.
--
----- stephan beal
http://s11n.net/
Registered Linux User #71917 http://counter.li.org
I speak for myself, not my employer. Contents may
be hot. Slippery when wet. Reading disclaimers makes
you go blind. Writing them is worse. You have been Warned.

[Nick Hounsome]

It's undefined behaviour.
You obviously know that it is undefined behaviour.
Getting someone on this newsgroup to say it's OK wont make it so.

"stephan beal" <st*****@s11n.net> wrote in message
news:bs**********@ork.noris.net...

Good morning, C++ users,

i've been hesitating to post this, primarily because i know that most of you here are *way* ahead of me in C++ and i'm a little embarassed about the
possibility of some gross errors in what i'll be posting...

That said, please go easy on me. :)

About 2 weeks ago i wrote a paper, called Context Singletons, where i
discuss some of the uses of context-specific Singleton-like objects. During the development of the paper i came up with a Phoenix class for providing
Phoenixing behaviour (as defined in Alexandrescu's MC++D) to my Context
Singletons. (http://s11n.net/misccode/)

So what, right?

i'm here to ask the Gurus to please take a look at my Phoenix, and see if
i'm doing something Inherently Wrong here. My gut tells me that the
implementation is simply too straightforward to solve the problem
adequately, but practice shows that it works quite well :/.

The source code is here:

template < typename BaseType, typename ContextType = BaseType >
class phoenix : public BaseType {
public:
typedef ContextType context_type;
typedef BaseType base_type;
static base_type & instance() {
static this_type meyers;
if( this_type::m_destroyed ) {
new( &meyers ) this_type;
atexit( this_type::do_atexit );
}
return meyers;
}
private:
typedef phoenix<base_type,context_type> this_type;
static bool m_destroyed;
phoenix() { m_destroyed = false; }
~phoenix() { m_destroyed = true; }
static void do_atexit() {
if( m_destroyed ) return;
static_cast<this_type SPMamp;>(instance()).~phoenix();
}
};
template <typename T, typename C> bool phoenix<T,C>::m_destroyed = false;

This code is complete for purposes of the questions posed here, but the code at http://s11n.net/misccode/phoenix.h is "more complete", with docs,
comments, one additional (optional) template parameter and debugging code.

Regarding the implementation, i have these questions:

- What happens when BaseType has a non-virtual destructor? We never use
delete(), relying on static destruction instead, so i think this is safe. i have successfully used this class with, e.g., std::map<> as BaseType, but
maybe i just have a remarkably accomodating C++ platform.

- Is it possible that m_destroyed gets destroyed at a point such that the
above usages of it becomes invalid? If so, is there a way to phoenix that
bool, so to speak?
Any insights on these points would be greatly appreciated.
--
----- stephan beal
http://s11n.net/
Registered Linux User #71917 http://counter.li.org
I speak for myself, not my employer. Contents may
be hot. Slippery when wet. Reading disclaimers makes
you go blind. Writing them is worse. You have been Warned.

[Martijn Lievaart]

On Tue, 30 Dec 2003 09:16:12 +0000, Nick Hounsome wrote:

[ Please don't toppost. Thx, M4 ]

It's undefined behaviour.
You obviously know that it is undefined behaviour.
Getting someone on this newsgroup to say it's OK wont make it so.

Where is the UB? Why do you think he knows this?

M4

[Victor Bazarov]

"stephan beal" <st*****@s11n.net> wrote...

[...]
template < typename BaseType, typename ContextType = BaseType >
class phoenix : public BaseType {
public:
typedef ContextType context_type;
typedef BaseType base_type;
static base_type & instance() {
static this_type meyers;
if( this_type::m_destroyed ) {
new( &meyers ) this_type;
atexit( this_type::do_atexit );
}
return meyers;
}
private:
typedef phoenix<base_type,context_type> this_type;
static bool m_destroyed;
phoenix() { m_destroyed = false; }
~phoenix() { m_destroyed = true; }
static void do_atexit() {
if( m_destroyed ) return;
static_cast<this_type SPMamp;>(instance()).~phoenix();
}
};
template <typename T, typename C> bool phoenix<T,C>::m_destroyed = false;

I have only a few questions: have you checked what happens when you call
'atexit' to register the function that is being executed at that moment?
I mean, the rest of it is working allegedly, but have you tried using
'exit' in your program? Does it actually exit or does it fall into
an endless loop?

Have you checked whether the number of d-tor calls matches the number of
c-tor calls?

In your 'instance' function, if the object has been destroyed, you never
set the flag back to 'false'. Is that intentional?

[stephan beal]

Victor Bazarov wrote:

I have only a few questions: have you checked what happens when you call
Hi, Victor! i appologize for the delayed response: i was out for the
holidays longer than i expected.
'atexit' to register the function that is being executed at that moment?
I mean, the rest of it is working allegedly, but have you tried using
'exit' in your program? Does it actually exit or does it fall into
an endless loop?
That's a good question. Now that i think about it, i never call exit()
manually - i return from main() instead. i'll give that a try.
Have you checked whether the number of d-tor calls matches the number of
c-tor calls?
Yes, from what i've seen they do match.
In your 'instance' function, if the object has been destroyed, you never
set the flag back to 'false'. Is that intentional?

That is set via the phoenix's ctor and dtor, so it will be togged to false
when the object is initially created or any time that the object is
phoenixed, and to true when the object is destroyed normally or destroyed
via do_atexit().
--
----- stephan beal
http://s11n.net/
Registered Linux User #71917 http://counter.li.org
I speak for myself, not my employer. Contents may
be hot. Slippery when wet. Reading disclaimers makes
you go blind. Writing them is worse. You have been Warned.

[stephan beal]

Nick Hounsome wrote:

It's undefined behaviour.
You obviously know that it is undefined behaviour.
Getting someone on this newsgroup to say it's OK wont make it so.

You'r mistaken: what i KNOW is that deleting a POINTER to a derived type
when the parent has a non-virtual dtor is undefined behaviour. What i DON'T
know is what happens for NON-POINTERS which are never delete()ed.

--
----- stephan beal
http://s11n.net/
Registered Linux User #71917 http://counter.li.org
I speak for myself, not my employer. Contents may
be hot. Slippery when wet. Reading disclaimers makes
you go blind. Writing them is worse. You have been Warned.