From Chris Downs · 1996-04-07 19:10 · #73647 in EMULATE
In reply to templates in header e by Mike Bilow
MB> In principle, a stack container should contain objects, not MB> pointers to objects. EB> Well, that's one way to do it, certainly. You haven't really EB> given much justification or explanation for why they "should." MB> Because containers are of no value unless they are smart. The STL is MB> a good example of a system which hides huge amounts of complexity. But it still has to work within the contraints of the C++ language. Should we use containers for polymorphic objects? If so, how do we prevent the derived part from getting sliced? What about containers for objects where random access iterators are required and the costs of copying the objects is large? Your "in principle" doesn't hold for so many of the most common situations. MB> How MB> many programmers are there who could code a red-black tree? For MB> sophisticated data structures underlying containers? Some containers MB> also need to be able to move their data elements or do garbage MB> collection, all of which expose differences in these paradigms. Sounds like an excellent argument for keeping objects OUT of containers. MB> Among other things, pointers are inherently not type-safe. They can MB> be made partially type-safe, but never completely so. What are the features of C++ pointers that aren't "type safe"? MB> This conceptual distinction becomes important when defining operators, MB> especially those which operate within the container, such as iterators. EB> Can you be specific? MB> The compiler is allowed to make "trivial" conversions between an MB> object and a reference to a const object, for example. This can lead MB> to huge speed improvements by avoiding the copy of object, especially MB> if they are large object, which would otherwise ensue. But these aren't "temporary objects" in STL. The constructor is called and new objects are created. Objects are destroyed. Yes, the containers use references for the algorithms, but when containers are shuffled, constructors and destructors are called. There are way too many situations where holding the objects themselves in the containers will bog down the system. EB> What optimizations? An example would be helpful here -- EB> it's a bit to oblique for me to use as it stands. MB> Let's say that I define a class which ends up being implement by the MB> compiler, at a low level, as an array. Everyone knows that you want MB> to avoid pushing the array onto and popping it off of the stack, or MB> very time consuming function calls will result. By allowing the MB> compiler to make its trivial conversions, which are only allowed for MB> references and not pointers, you can avoid all of this nasty function MB> call overhead. Operations thru pointers have their well understood performance implications. Contrusting and destructing objects also have their performance implications. The "general principle" is to understand these mutually exclusive implications and decide based on the specifics. --- þ Blue Wave/QWK v2.12 þ --- WILDMAIL!/WC v4.12 * Origin: (1:100/4.0)