how to backport without breaking ABI?

[floris]

On Sat, 2011-02-12 at 21:12 +0100, Andreas Pakulat wrote:
> On 12.02.11 20:49:40, Milian Wolff wrote:
> > http://commits.kde.org/kdevplatform/288c92e122776b6310b3862d02372e465fd16837
> > 
> > the commit message says it all, I just forgot the CCMAIL...
> > 
> > Input?
> 
> You can't, period. Thats why Qt uses the dptr stuff (including
> dptr-subclasses) for its public API.
> 
> Andreas
> 
> -- 
> Someone is speaking well of you.
> 
isn't there a reasonable chance that KSharedPtr has the same layout as a
simple pointer? I mean the only member of KSharedPtr is the raw pointer
(reference counting happens apparently inside the pointed object)
therefore the change won't break this pointer, the only problem i see
would be when this happens:

struct foo {
	int data;
	bar* inst;
	int more_data
};

struct foo {
	int data;
	KSharedPtr<bar> inst;
	int more_data;
};

i think the padding after inst is allowed to change, but i'm quite sure
that dereferencing the second int as if it were the first should work
the more_data could be moved, i guess, but it doesn't happen here. 

to test my hypothesis i made a little test:

#include <cassert>
int main(){
	struct AA{
		int* ptr;
	};
	
	struct BB{
		int m1;
		int* m2;
		int m3;
	}b;
	
	struct CC{
		int m1;
		AA m2;
		int m3;
	}c;
	assert(sizeof(BB)==sizeof(CC));
	assert((long long)&b.m1-(long long)&b ==(long long)&c.m1-(long long)&c
);
	assert((long long)&b.m2-(long long)&b ==(long long)&c.m2.ptr - (long
long)&c );
	assert((long long)&b.m3-(long long)&b ==(long long)&c.m3-(long long)&c
);
	
}

of course older kdevelops which treat that member as a simple pointer
can still store a raw pointer which gets deleted prematurely, but
there's really nothing you can do about that.

Floris