Welcome to the MacNN Forums.

Mithras · Apr 25, 2002, 08:06 AM

Hi all.

I'm a beginning CS student, just tackling intro C++. Everything has been fine thus far; however, a new unit on template classes is clobbering me.

I'm using Project Builder. If I make a very simple project that involves template classes, the project compiles, but I get linking errors about undefined symbols.

Example:

main.cpp
<BLOCKQUOTE>code:<HR><pre>
#include <iostream>
#include "myclass.h"

int main () {
Set<int> aPlaySet;
return 0;
}
[/code]

myclass.h
<BLOCKQUOTE>code:<HR><pre>
#include <iostream>
#include <vector>

template <class T>
class Set
{
public:
Set();
private:
vector<T> theItems;
};
[/code]

myclass.cpp
<BLOCKQUOTE>code:<HR><pre>
#include "myclass.h"

template<class T>
Set<T>::Set()
{
//no-op
}
[/code]

I did some web searching that indicated that gcc doesn't do templates very well, such that I need to add template class Set<int> and so on, for each kind of Set class I want. That's certainly lame, but I was willing. That works for awhile, but as I add more member functions, I invariably end up with more errors.

One last thing: The prepackaged source code from the book (Accelerated C++), which has template classes, compiles with makefiles just fine. So it must be possible to get templates working, it's just that I'm stupid.

I don't know anything about creating makefiles, but should I go that route if I want things to work? Surely many others want to make template classes in OS X...

Mithras · Apr 26, 2002, 09:37 AM

Hard to believe no one else is doing this.

Anyway, I eventually found a message on the projectbuilder-users mailing list that indicated that gcc is unhappy unless all of your template function implemenations are in the header file with the class definition, instead of in a separate .cpp file.

So for anyone else out there: Just stick everything in your .h

kamprath · Apr 26, 2002, 11:19 AM

Originally posted by Mithras:
Hard to believe no one else is doing this.

Anyway, I eventually found a message on the projectbuilder-users mailing list that indicated that gcc is unhappy unless all of your template function implemenations are in the header file with the class definition, instead of in a separate .cpp file.

So for anyone else out there: Just stick everything in your .h

Requiring template method implementation in the local compile unit when used (effectively, it should be in the header) is not a "gcc thing," its a "C++ thing." Any C++ book would have covered this.

Michael

Mithras · Apr 26, 2002, 01:21 PM

Accelerated C++, p142:
The C++ standard says nothing about how implementations should manage template instantiation, so every implementation handles instantiation in its own particular way...

Most current implementations require that in order to instantiate a template, the definition of the template, not just the declaration, has to be accessible to the implementation. Generally, this requirement implies access to the source files that define the template, as well as the header file. How the implementation locates the source file differs from one implementation to another. Many implementations expect the header file for the template to include the source file, either directly or via a #include. The most certain way to know what your implementation expects is to check its documentation.

Definitely an implementation, not a specification, issue. But yes, I should have caught this bit before. Thanks for your illuminating, helpful, and kind post.

[ 04-26-2002: Message edited by: Mithras ]

[ 04-26-2002: Message edited by: Mithras ]

kamprath · Apr 26, 2002, 04:52 PM

Originally posted by Mithras:


Definitely an implementation, not a specification, issue. But yes, I should have caught this bit before. Thanks for your illuminating, helpful, and kind post.

[ 04-26-2002: Message edited by: Mithras ]

[ 04-26-2002: Message edited by: Mithras ]

Actually, we are both wrong. The C++ Standard does say something about defining a template's implementation outside of the local compile unit. As you suspected, you are allowed to do so, however, like inlined functions defined outside the local compile unit, the implementations must be exported from the compile unit they are in. However, it is noted by Stroustrup that compile/link time is greatly enhanced by always defining template definitions int he local compile unit (but that runs the risk of namespace collisions).

That means you implementation needed to be:

myclass.cpp

<BLOCKQUOTE>code:<HR><pre>
#include "myclass.h"

export template<class T>
Set<T>::Set()
{
//no-op
}
[/code]

Be wary of any programming book that has "Accelerated" in the title. Spend some money and buy Stroustrup's latest C++ definition (WELL worth the money, if you are thinking about being a proffessional developper, don't buy the cheap books). Good books can help you properly resolve issues like this, even if you have been programming for 15+ years. The information on my above comments can be found in P13.7 of "The C++ Programming Language."

Michael

[ 04-26-2002: Message edited by: kamprath ]

Mithras · Apr 26, 2002, 07:17 PM

Thanks a bunch. You took my cross, sarcastic response and responded sincerely and helpfully to it, darn you!

But yes, thanks for explicating it all. I've settled on includeing the source back into the header as a fair compromise, for now. I appreciate the thoughts.

I certainly agree with your comment about "Accelerated" - it's definitely a lousy book, intended for work-a-day programmers. (It hands you the STL and tells you to use that, without much discussion of algorithms or data structures or why STL authors made the choices they did...)

I just graduated from a good school with a degree in another field, and decided to take a couple of CS classes - but the cheapness of the school at which I'm taking them is evident in the syllabus, I'm afraid...