[cfe-dev] Parser Stmt/Expr Owning Pointer

[Howard Hinnant]

On Dec 9, 2008, at 11:16 AM, Sebastian Redl wrote:

> In terms of C++0x rvalue references, I've basically implemented an  
> implicit conversion from "unique_ptr<void, ASTDeleter>&&" to void*  
> that is equivalent to release(). Obviously, this is not possible in C 
> ++0x, but since my rvalue reference is a special object, it's very  
> easy.
>
> The advantage of this is that I can pass my pointer's mover to a  
> function that is updated to take my pointer as easily as I can pass  
> it to a function that still takes a raw pointer, with no change in  
> the caller's syntax. Since the update is the business of the callee,  
> I don't see this as unsafe, I see it as convenient.
> We're not redefining ownership here, we're just trying to enforce  
> the existing rules through the type system. So the caller moves the  
> pointer to the callee, no matter how the callee decides to handle it  
> further.

People have been warning against implicit conversions from smart  
pointers to raw pointers for a very long time...

http://www.aristeia.com/Papers/C++ReportColumns/jun96.pdf

> The bottom line is simple: don’t provide implicit conversion  
> operators to dumb point-
> ers unless there is a compelling reason to do so.

Scott Meyers, 1996.

I realize you're only converting implicitly from an ASTMove (result of  
member move).  However this is at least conceptually the same as  
implicitly converting from an rvalue smart pointer.

>> I'll try to get it in soon.  I'm not yet positive what "in" means.
> For starters, you could send me a link. :-) I'm more interested in  
> the technique, though. Using unique_ptr directly is difficult  
> because we need to store the success flag alongside the pointer (and  
> if I wrap unique_ptr, I need to write all the emulation stuff again  
> anyway).

Info sent offline.

-Howard