Automatically separate class definitions from decl

问题:

I am using a library that consists almost entirely of templated classes and functions in header files, like this:

// foo.h
template<class T>
class Foo {
  Foo(){}
  void computeXYZ() { /* heavy code */ }
};
template<class T>
void processFoo(const Foo<T>& foo) { /* more heavy code */ }

Now this is bad because compile times are unbearable whenever I include one of those header files (and actually I include many of them in each of my compilation units).

Since as a template parameter I only use one or two types anyway I am planning to create, for each library header file, a file that contains only declarations, without the heavy code, like this:

// NEW: fwd-foo.h
template<class T>
class Foo {
  Foo();
  void computeXYZ();
};
template<class T>
void processFoo(const Foo<T>& foo);

And then one file that creates all the instantiations that I'll need. That file can be compiled separately once and for all:

// NEW: foo.cpp
#include "foo.h"
template class Foo<int>;
template class Foo<double>;
template void processFoo(const Foo<int>& foo);
template void processFoo(const Foo<double>& foo);

Now I can just include fwd-foo.h in my code and have short compile times. I'll link against foo.o at the end.

The downside, of course, is that I have to create these new fwd-foo.h and foo.cpp files myself. And of course it's a maintenance problem: When a new library version is released I have to adapt them to that new version. Are there any other downsides?

And my main question is:

Is there any chance I can create these new files, especially fwd-foo.h, automatically from the original foo.h? I have to do this for many library header files (maybe 20 or so), and an automatic solution would be best especially in case a new library version is released and I have to do this again with the new version. Are any tools available for this task?

EDIT:

Additional question: How can the newly supported extern keyword help me in this case?

回答1:

We use lzz which splits out a single file into a separate header and translation unit. By default, it would normally put the template definitions into the header too, however, you can specify that you don't want this to happen.

To show you how you might use it consider the following:

// t.cc
#include "b.h"
#include "c.h"

template <typename T> 
class A {
  void foo () {
    C c;
    c.foo ();
    b.foo ();
  }
  B b;
}

Take the above file and copy it to 't.lzz' file. Place any #include directives into separate $hdr and $src blocks as necessary:

// t.lzz
$hdr
#include "b.h"
$end

$src
#include "c.h"
$end

template <typename T> 
class A {
  void foo () {
    C c;
    c.foo ();
    b.foo ();
  }
  B b;
}

Now finally, run lzz over the file specifying that it places the template definitions into the source file. You can either do this using a $pragma in the source file, or you can use the command line option "-ts":

This will result in the following files being generated:

// t.h
//

#ifndef LZZ_t_h
#define LZZ_t_h
#include "b.h"
#undef LZZ_INLINE
#ifdef LZZ_ENABLE_INLINE
#define LZZ_INLINE inline
#else
#define LZZ_INLINE       
#endif
template <typename T>
class A
{
  void foo ();
  B b;
};
#undef LZZ_INLINE
#endif

And:

// t.cpp
//

#include "t.h"
#include "c.h"
#define LZZ_INLINE inline
template <typename T>
void A <T>::foo ()
          {
    C c;
    c.foo ();
    b.foo ();
  }
#undef LZZ_INLINE

You can then run these through some grep/sed commands to remove the LZZ helper macros.

回答2:

Try using precompiled headers. I know GCC and MSVC support this feature. Usage is vender-specific, though.

回答3:

I have been working on the very same issue for quite a while now. In the solution you're proposing, you are defining your template classes twice. It will be ok if it defines the same stuff (in the same order), but you're bound to have problems sooner or later.

What i have come up with is to consider the problem the other way around. As long as you are not specializing your implementation, it works gracefully.

It uses two macros, which avois having to update template arguments in implementation file (be careful, though, if you want to add default template arguments to the class).

// foo.h
#define FOO_TEMPLATE template<typename T>
#define FOO_CLASS Foo<T>

FOO_TEMPLATE
class Foo {
  Foo();
  void computeXYZ();
};

// foo_impl.h
#include "foo.h"
FOO_TEMPLATE
FOO_CLASS::Foo(){}

FOO_TEMPLATE
void FOO_CLASS::computeXYZ() { /* heavy code */ }

By doing this, you essentially work the same way you do with non-template classes (you can do the same thing with template functions, of course).

EDIT : about the extern keyword in c++0x

I believe the extern keyword in c++0x will help, but it won't solve everything magically !

From this article,

Extern Templates

Every module that instantiates a template essentially creates a copy of it in the object code. Then, it's up to the linker to dispose of all of the redundant object code at the very last stage, thus slowing the critical edit-compile-link cycle that makes up a programmer's day (or sometimes daydreams). To short-circuit this object code garbage collection, a number of compiler vendors have already implemented an extern keyword that can go in front of templates. This is a case of standardization codifying existing industry practice (pun intended). In practice, this is implemented by sending a notice to the compiler basically to "not instantiate this here":

extern template class std::vector;

回答4:

C++0x will fix your compile time issues with extern templates. I don't know an automatic way to do what you ask, though.