Templates, friends, operators .. what more do we need to have fun?

[Micha]

Hello there,

I think I've run into some classic c++ pitfall and maybe some of
you guys can help me out.
For my project I will need to use matrices and vectors and so I
decided to implement them by myself. I know there are already
tons of vector and matrix implementations, but I wanted to have
one taylored for my needs and without debugging someones
else code. Also is's become somewhat personal meanwhile ;-).

The plan :
- define a Matrix<T> and a Vector<T> class
- make them friends to perform some operations directly - quick
and dirty
- define the operators processing instances of more than one type
as friends

After a while and learning a lot about c++ syntax subtleness I made
the first two steps.
Now the linker (I'm using g++ 3.3.4 without any switches refering
templates) is nagging with complaints about not knowing the friend
operators I've defined.

Lets come to the details :

I've defined both classes separated into a header and the
implementation
and for all my template classes I'm using a single .cpp file which
includes
the class template implementations to explicitly instantiate the
specialized classes.

Both header files (matrix and vector) include each other. They contain
- the forward declarations of the class templates
- the operators prototypes
- the class template itself
- including the references to friend classes and friend operators

The operators are defined in the according class implementation files.

Some more details - I will show it exemplarily :

Vector.h
===================================================================
....

#include "Grid2D.h"
#include "Matrix.h"

namespace calcapp {

//a template method
template <class T> void doWhateveryouwant(const T& v);

//forward declaration of class template
template <class T> class Vector;
template <class T> class Matrix;

//operator prototypes
template <class T> Matrix<T> operator/ (const Vector<T>& v1, const
Vector<T>& v2);
....

//prototype of a test-method
template <class T> Vector<T> doNothing(const Vector<T>& v);

template <class T>
class Vector : public calcapp::Grid2D<T> {
public:
template <class TF> friend class Matrix;

Vector(int count_dim);
Vector(const Vector & other);

....
//tensor product
friend Matrix<T> operator/<T> (const Vector<T>& v1, const Vector<T>&
v2);
//dummy
friend Vector<T> doNothingInNS<T>(const Vector<T>& v);
protected:
....
}; //class

} //namespace

Vector.cpp
===================================================================
#include "Vector.h"

....
namespace calcapp {

template <class T>
Matrix<T> calcapp:perator/ (const Vector<T>& v1, const Vector<T>& v2)
{

Matrix<T> result(v1.size_x, v2.size_x);
....
return result;
} //tensorproduct
....
} //namespace


TemplateInstances.cpp
===================================================================
#include "global.h"
#include "Vector.cpp"

namespace calcapp {

template class Grid2D<fptype>;

template class Matrix<fptype>;
template class Matrix<int32>;
template class Matrix<int64>;

template class Vector<fptype>;
template class Vector<int32>;
template class Vector<int64>;

}

test.cpp
===================================================================
....

Vector<fptype) v1(3), v2(3);
....

and here the compiler states :
../RunTest/test.cpp:310: undefined reference to `calcapp::Matrix<double> >calcapp:perator/<double>(calcapp::Vector<double> const&, calcapp::Vector<double> const&)'

....
===================================================================
So what the #!@* is wrong? That was obviously a linker error and for my
understanding the compiler didn't generate code for operator/<fptype>.
Correct?

--------------------------
The problem doesn't seem to be about operators because

....
v2 = calcapp::doNothing(v1);
....

would leed to an equivalent reaction.
--------------------------
And even

....
int i = 10;
doWhateveryouwant(i);
....

won't compile (link) :

undefined reference to `void calcapp::doWhateveryouwant<int>(int const&)'

So it seems it has nothing to do with classes or friends either.
But still somehow the linker figured out that the function was defined
in
the namespace calcapp.
--------------------------

Do I have to explicitly instantiate template methods too? I've
read nothing about that so far. How??
Operators?

Ok guys, what's the simple trick? ;-)


bye, Micha

 

[John Harrison]

Hello there,

I think I've run into some classic c++ pitfall and maybe some of
you guys can help me out.
For my project I will need to use matrices and vectors and so I
decided to implement them by myself. I know there are already
tons of vector and matrix implementations, but I wanted to have
one taylored for my needs and without debugging someones
else code. Also is's become somewhat personal meanwhile ;-).

The plan :
- define a Matrix<T> and a Vector<T> class
- make them friends to perform some operations directly - quick
and dirty
- define the operators processing instances of more than one type
as friends

After a while and learning a lot about c++ syntax subtleness I made
the first two steps.
Now the linker (I'm using g++ 3.3.4 without any switches refering
templates) is nagging with complaints about not knowing the friend
operators I've defined.

To be perfectly honest I haven't read all your post (bit too long for
me) but you might care to check the FAQ. It does at least describe a
problem with similar symptoms to yours.

http://www.parashift.com/c++-faq-lite/templates.html#faq-35.10

john

 

[Greg]

Hello there,

I think I've run into some classic c++ pitfall and maybe some of
you guys can help me out.
For my project I will need to use matrices and vectors and so I
decided to implement them by myself. I know there are already
tons of vector and matrix implementations, but I wanted to have
one taylored for my needs and without debugging someones
else code. Also is's become somewhat personal meanwhile ;-).

The plan :
- define a Matrix<T> and a Vector<T> class
- make them friends to perform some operations directly - quick
and dirty
- define the operators processing instances of more than one type
as friends

After a while and learning a lot about c++ syntax subtleness I made
the first two steps.
Now the linker (I'm using g++ 3.3.4 without any switches refering
templates) is nagging with complaints about not knowing the friend
operators I've defined.
....

Some more details - I will show it exemplarily :

Vector.h

First it's a not a good idea to have a header called "vector.h" since
there is an (obsolete) header with the same name in the system include
paths. It seems safer to minimize confusio by renaming it with a less
generic name.

Vector.cpp
===================================================================
#include "Vector.h"

...
namespace calcapp {

template <class T>
Matrix<T> calcapp:perator/ (const Vector<T>& v1, const Vector<T>& v2)
{

Matrix<T> result(v1.size_x, v2.size_x);
...
return result;
} //tensorproduct
...
} //namespace


TemplateInstances.cpp

It looks like this template function is being defined in a source file.
Template definitions need to be placed in header files so the compiler
will have seen the template defnition before it compiles source code
that instantiates the template. With the template definition in a
source file, it is effectively available only to other routines in this
file which come after it.

#include "global.h"
#include "Vector.cpp"

namespace calcapp {

template class Grid2D<fptype>;

template class Matrix<fptype>;
template class Matrix<int32>;
template class Matrix<int64>;

template class Vector<fptype>;
template class Vector<int32>;
template class Vector<int64>;

}

test.cpp
===================================================================
...

Vector<fptype) v1(3), v2(3);
...



...
===================================================================
So what the #!@* is wrong? That was obviously a linker error and for my
understanding the compiler didn't generate code for operator/<fptype>.
Correct?

When compiling test.cpp, the compiler did not see the definition for
the operator/ by the time it reached line 310 that references it.
Without a definition on hand, the compiler does not know what code to
generate. Nor does the compiler "remember" template definitions from
one source file to the next. Every source file that instantiates a
template must ensure that the the template definition is included
before its own code when it is compiled.

The solution is to move the operator/ template and any other template
definitions into a header file that test.cpp and other clients of these
template classes will then include.

Greg

 

[John Harrison]

It looks like this template function is being defined in a source file.
Template definitions need to be placed in header files so the compiler
will have seen the template defnition before it compiles source code
that instantiates the template. With the template definition in a
source file, it is effectively available only to other routines in this
file which come after it.

Greg, you missed that he wrote #include "Vector.cpp". Why some people
insist on putting template code in cpp files and then pulling all sorts
of tricks to compensate for this is beyond me.

john

 

[Greg]

Greg, you missed that he wrote #include "Vector.cpp". Why some people
insist on putting template code in cpp files and then pulling all sorts
of tricks to compensate for this is beyond me.

john

You're right. Though actually I did notice that #include at one point,
but I lost track of it in that ocean of code.

But it is worth mentioning nonetheless that source files #including
other source files is bad form. Such nested source listings can confuse
the debugger, the programmer, the compiler - and programming is
confusing enough as it is, we need not invent ways to make it more so.

Greg

 

[Micha]

As far as I got it, the idea is to have the compiler
generate the code for the specialized template classes
once only in the resulting TemplateInstances.o object
file.

Wouldn't implementing the whole template class in the
header file, force the compiler to instantiate the same
MyTemplateClass<MyType> class in every object file
generated from any .cpp file in which
MyTemplateClass<MyType> is used?

Btw thanks for your replies but I didn't get any further. My
classes work just fine until I try to use those friend operators.

The template classes which where instantiated in that .cpp
file which includes the .cpp files of the template classes
(together with their header files) get compiled and linked. I can
use their member operators.

As far as I figured it out until now, the problem is that the compiler
won't generate code for the operators despite seeing instantiations of
the template classes they are friends with.

Maybe I can solve my problem if I get the following example to work.

foo_templatestuff.h
====================================
#ifndef FOO_TEMPLATESTUFF_
#define FOO_TEMPLATESTUFF_

template <class T>
void doSomething(T x);
#endif


foo_templatestuff.cpp
====================================
#include "foo_templatestuff.h"
#include <iostream>

template <class T>
void doSomething(T x) {
std::cout << x;
}

main.cpp
====================================
#include "foo_templatestuff.h"

int main(int argc, char *argv[]) {

int i = 23;
doSomething(i);

return EXIT_SUCCESS;
}

Again the compiler figures out that it is
void doSomething<int>(int) that should be called while
processing main.cpp, but it didn't see a reason to instantiate that
function while processing foo_templatestuff.cpp.

Is there no other way but to define the whole function template
in the header file instead of the prototype only?

Ok, there is simple but dirty solution by adding a function
void __dummy__() {
doSomething<int>(5);
}
to foo_templatestuff.cpp.

It works but I don't like it. I don't want to write some dummy function
that touches all my operator templates for any specialization just to
make that *#!!$ compiler generate the accordant code.

There must be a better way ;-) ! Am I right?

bye, Micha

 

[Greg]

As far as I got it, the idea is to have the compiler
generate the code for the specialized template classes
once only in the resulting TemplateInstances.o object
file.

Wouldn't implementing the whole template class in the
header file, force the compiler to instantiate the same
MyTemplateClass<MyType> class in every object file
generated from any .cpp file in which
MyTemplateClass<MyType> is used?

Yes. In fact all of the STL container classes being templates are
entirely implemented in header files <vector>, <list> and so on have to
be included by any code that uses a std::vector, std::list and so
forth. The compiler will instantiate only the methods and templates
that a source file needs. And the linker discards duplicate template
instantiations across files when the program is linked.

In theory, instantiating the templates in a source file would speed up
compile times. But the program would then have to anticipate all of the
instantiations needed in other files. In the end, the same routines end
up in the app anyway, so it's easier to let compiler and linker handle
it.

Btw thanks for your replies but I didn't get any further. My
classes work just fine until I try to use those friend operators.

The template classes which where instantiated in that .cpp
file which includes the .cpp files of the template classes
(together with their header files) get compiled and linked. I can
use their member operators.

As far as I figured it out until now, the problem is that the compiler
won't generate code for the operators despite seeing instantiations of
the template classes they are friends with.

Declaring a template class or function a friend is not enough to
instantiate it, so this behavior is as expected.

Maybe I can solve my problem if I get the following example to work.

foo_templatestuff.h
====================================
#ifndef FOO_TEMPLATESTUFF_
#define FOO_TEMPLATESTUFF_

template <class T>
void doSomething(T x);
#endif


foo_templatestuff.cpp
====================================
#include "foo_templatestuff.h"
#include <iostream>

template <class T>
void doSomething(T x) {
std::cout << x;
}

main.cpp
====================================
#include "foo_templatestuff.h"

int main(int argc, char *argv[]) {

int i = 23;
doSomething(i);

return EXIT_SUCCESS;
}

Again the compiler figures out that it is
void doSomething<int>(int) that should be called while
processing main.cpp, but it didn't see a reason to instantiate that
function while processing foo_templatestuff.cpp.

Is there no other way but to define the whole function template
in the header file instead of the prototype only?

Yes, just put the whole function template in the header file. Unlike a
regular function definition, doing so will not cause link errors. Note
also that the function template is not inlined just because it is in
header file. The "inline" keyword would still have to be used if
inlining is intended.

Ok, there is simple but dirty solution by adding a function
void __dummy__() {
doSomething<int>(5);
}
to foo_templatestuff.cpp.

It works but I don't like it. I don't want to write some dummy function
that touches all my operator templates for any specialization just to
make that *#!!$ compiler generate the accordant code.

There must be a better way ;-) ! Am I right?

bye, Micha

Depending on the compiler, it should be possible to explicitly
instantiate the function template. Of course the definition for the
template will be needed at the point of instantiation. Specializing the
template should also define it.

Moving the definitions into a header file will spare you these
headaches. It would match the way that everybody else, include the
standard library, does it. The compiler may end up compiling a little
more than it needs to. But the program after it is linked is the same
program whether you do have placed the templates in headers, or have
succeeded after much effort in putting them somewhere else.

Greg

 

[John Harrison]

As far as I got it, the idea is to have the compiler
generate the code for the specialized template classes
once only in the resulting TemplateInstances.o object
file.

Wouldn't implementing the whole template class in the
header file, force the compiler to instantiate the same
MyTemplateClass<MyType> class in every object file
generated from any .cpp file in which
MyTemplateClass<MyType> is used?

Yes and this is sometimes known as code bloat. But these days the linker
can eliminate duplicate template code so this is not a problem.

Put all template code in header files, it feels strange at first but it
is the right way to do it.

john

 

[Micha]

Thank you guys, it works and I can have my sleep ;-)

bye, Micha

 

[ma740988]

Greg, you missed that he wrote #include "Vector.cpp". Why some people
insist on putting template code in cpp files and then pulling all sorts
of tricks to compensate for this is beyond me.

john

John, my current approach when dealing with a LOT of template code is
to to put the implementation in a .hh file. Include the .hh file in
the .h file. How does this sound?

I'm unsure why it's necessary to make the function definitions inline
but I suspect I could peruse the standard to find out why.

I too place all my template code in a .h to make life easy. This
reminds me of a conversation I had with an individual who insists on
using void within his function arguments for member functions that take
no arguments. It drives me crazy and if I'm not mistaken they (see
below) don't mean the same thing in a C++ world but I didn't have proof
so I couldn't convice him.

bool test::some_func(void) <- the use of void here is not the same in
C++ correct?
{ return 1;}

 

[Greg]

John, my current approach when dealing with a LOT of template code is
to to put the implementation in a .hh file. Include the .hh file in
the .h file. How does this sound?

I'm unsure why it's necessary to make the function definitions inline
but I suspect I could peruse the standard to find out why.

Placing the function templates in a header file does not implicitly
"inline" them. The functions still need to be declared with the
"inline" keyword for the compiler to treat them as inline functions.
The reason for putting the templates in the header file is to ensure
that the compiler will have seen the template definition should it need
to instantiate it when compiling a source file.

I too place all my template code in a .h to make life easy. This
reminds me of a conversation I had with an individual who insists on
using void within his function arguments for member functions that take
no arguments. It drives me crazy and if I'm not mistaken they (see
below) don't mean the same thing in a C++ world but I didn't have proof
so I couldn't convice him.

bool test::some_func(void) <- the use of void here is not the same in
C++ correct?
{ return 1;}

I think you have specifics reversed - that is, the declaration without
"void":

int some_func();

is not the same in C as it is in C++. In C++ some_func accepts no
parameters. In C, some_func may accept an int parameter (since the
parameter list has been omitted, an int parameter is assumed).

To declare a a function that accepts no parameters in C it is therefore
necessary to use the "void" parameter list in order to distinguish it
from the declaration without a paramter list - and for which an int
parameter is assumed. C++ never needs a void parameter list, since the
parameter list is always present in a function's declaration; if the
parameter list is empty, then the function accepts no parameters.

So although the void is unnecessary and even annoying in C++, it is not
actually changing the meaning of the declaration.

Greg

 

[ma740988]

Placing the function templates in a header file does not implicitly
"inline" them. The functions still need to be declared with the
"inline" keyword for the compiler to treat them as inline functions.
The reason for putting the templates in the header file is to ensure
that the compiler will have seen the template definition should it need
to instantiate it when compiling a source file.


I think you have specifics reversed - that is, the declaration without
"void":

int some_func();

is not the same in C as it is in C++. In C++ some_func accepts no
parameters. In C, some_func may accept an int parameter (since the
parameter list has been omitted, an int parameter is assumed).

To declare a a function that accepts no parameters in C it is therefore
necessary to use the "void" parameter list in order to distinguish it
from the declaration without a paramter list - and for which an int
parameter is assumed. C++ never needs a void parameter list, since the
parameter list is always present in a function's declaration; if the
parameter list is empty, then the function accepts no parameters.

So although the void is unnecessary and even annoying in C++, it is not
actually changing the meaning of the declaration.

Greg

Appreaciate it Greg!! You're right I had my specifics backwords. I'll
get over my annoyance - I suppose