Re: Question: deep - shallow - move semantics

Hi,

To avoid misunderstandings I start with repeating what you certainly  
know. Deep copy has no side effects but is expensive. Shallow copy is  
fast but isn't really a copy to be honest. It is more an aliasing and  
if you change the copy the original will be also changed and this is  
not always what you want. Most swap implementations of data structures  
pointing to some constructs in memory are done with shallow copy but  
there it is okay since the two constructs in memory are referred by one  
object each. Swapping only the pointers does not establish any  
aliasing.

The idea behind move semantics is that an object that is computed in a  
function or by an operator or by any expression only exist once.  
Assigning this to a variable in C++ is done by a deep copy and the  
temporary object in the function/operator/expression is thrown away.  
Since this temporary object isn't used by anybody else (after deletion  
:-D) it can be copied shallowly in the assignment. The same can be done  
for intermediate results, e.g., summing three vectors (or matrices)  
a+b+c the temporary vector containing the result of a+b can be  
shallowly copied to the 1st operand of the second addition. Assignments  
from one variable to another are copied deeply to avoid the side  
effects. David Abrahams proposed this to the C++ standard committee and  
hopefully C++ compiler will be able to do it one day. The MTL should  
have this mechanism earlier.

Unfortunately, the current implementation of MTL does not provide move  
semantics. Normal assignments of vectors and matrices are shallow  
copies. Deep copies are provided in the 'copy' command. I would in your  
position starting by implementing everything with the copy command  
(instead of assignments) and later replace some deep copies with  
shallow ones by being very careful which aliasing can cause trouble.

Good luck!

Peter

On Jul 21, 2005, at 3:21 AM, 杨 铭 wrote:

> Hi all,
>
>   I'm a newbie (both of C++ and Matrix Computation).
>
>   I'm currently working on a small project which contains some matrix  
> manipulation. I decide to write a matrix class. I have read some code  
> of MTL, and it's a great work, I think.
>   What bothers me is how to avoid deep copy. Now MTL uses a reference  
> counting stratage, as I know. But the plan of future version say that  
> deep copy semantics
> will be adopt ( see   
> http://boost-consulting.com/projects/mtl4/libs/numeric/mtl/doc/html/ 
> plan.html )
>
>   After implementing a simple reference counting based matrix class, I  
> realized that shallow copy semantics is quite confusing but alleviates  
> the efficiency problem.
>
>   My question is how to ensure high performance in the forthcoming MTL  
> 4? (Sorry, I found section 1.7.2 of the plan is hard to understand,  
> especially
> "move semantics").
>
>   And would you like to give me some advise about my matrix class?  
> Should
> I use shallow copy semantics or a deep one?
>
>   thank you!
>
> _________________________________________________________________
> 享用世界上最大的电子邮件系统— MSN Hotmail。  http://www.hotmail.com
> _______________________________________________
> This list is archived at http://www.osl.iu.edu/MailArchives/mtl-devel/
------------
Peter Gottschling
Research Associate
Open Systems Laboratory
Indiana University
301i Lindley Hall
Bloomington, IN 47405
Tel.: +1 812 855-8898   Fax: +1 812 856 0853
http://www.osl.iu.edu/~pgottsch

_______________________________________________
This list is archived at http://www.osl.iu.edu/MailArchives/mtl-devel/