Re: void pointer and map (fwd)

[Ted Byers <r dot ted dot byers at rogers dot com>]

--- Anitha Boyapati <anithab@sankhya.com> wrote:

> 
> Hi,
> 
> On Fri, 21 Dec 2007, Andrew Haley wrote:
> 
> > Anitha Boyapati writes:
> >  > 
> > You're passing an instance of Value to
> Map::value_type, which creates
> > a pair instance by copying both the key and the
> value.  Note that we
> > now have two instances of Value, both of which are
> pointing to the
> > same buf.  operator delete will be called on both
> of these instances.
> 
> Right! I missed a copy constructor then. That solves
> it then. Thanks.
> 
That is a simple way to look at it, but perhaps too
simple.

You have to bear in mind the semantics maps require of
objects contained in them.  Here is what you wrote:

struct Value {
	int id;
	void *buf;
	Value (long l) { id = 4; buf = operator new(128); 
		          long *c = reinterpret_cast<long *>(buf);
		    	  *c = l;
		  	}
	~Value() { operator delete(buf); }
}; 

bool operator< (const Value &v1, const Value &v2) {
	bool result = (*(reinterpret_cast<long *>(v1.buf))) <
 (*(reinterpret_cast<long *>(v2.buf)));
 	return result;
}

bool operator== (const Value &v1, const Value &v2) {
	bool result = (*(reinterpret_cast<long *>(v1.buf)))
==
 (*(reinterpret_cast<long *>(v2.buf)));
 	return result;
}

typedef map<Value, string> Map;

First, I don't see the point of reinterpret cast here,
or even use of a void pointer.  Unless this is a
really dumbed down example, you probably ought to
think about the types you'll be using for this.

Second, why use a naked pointer instead of a smart
pointer (such as that found in boost)?  You can use
the latter wherever you'd be tempted to use a naked
pointer, and it becomes all the easier to avoid the
sort of problem you posted about.  With a smart
pointer, you could even get rid of the destructor
without losing anything, unless you need to make it
virtual, should there be a chance another class could
be derived from Value, which has a need for a
non-trivial destructor.

HTH

Ted