Clean up your #include statements?

Question

How do you maintain the #include statements in your C or C++ project? It seems almost inevitable that eventually the set of include statements in a file is either insufficient (but happens to work because of the current state of the project) or includes stuff that is no longer needed.

Have you created any tools to spot or rectify problems? Any suggestions?

I've been thinking about writing something that compiles each non-header file individually many times, each time removing an #include statement. Continue doing this until a minimal set of includes is achieved.

To verify that header files are including everything they need, I would create a source file that all it does is include a header file and try to compile it. If the compile fails, then the header file itself is missing an include.

Before I create something though, I thought I should ask here. This seems like a somewhat universal problem.

Answer 1

Yup. We have a preprocessor of our own which gives us access to our own macro language. It also checks that header files are only included one time. Creating a simple preprocessor checking for multiple includes should be fairly easy.

Answer 2

To verify that header files are including everything they need, I would creating a source file that all it does is include a header file and try to compile it. If the compile fails, then the header file itself is missing an include.

You get the same effect by making the following rule: that the first header file which foo.c or foo.cpp must include should be the correspondingly-named foo.h. Doing this ensures that foo.h includes whatever it needs to compile.

Furthermore, Lakos' book Large-Scale C++ Software Design (for example) lists many, many techniques for moving implementation details out of a header and into the corresponding CPP file. If you take that to its extreme, using techniques like Cheshire Cat (which hides all implementation details) and Factory (which hides the existence of subclasses) then many headers would be able to stand alone without including other headers, and instead make do with just forward declaration to opaque types instead ... except perhaps for template classes.

In the end, each header file might need to include:

• No header files for types which are data members (instead, data members are defined/hidden in the CPP file using the "cheshire cat" a.k.a. "pimpl" technique)

• No header files for types which are parameters to or return types from methods (instead, these are predefined types like int; or, if they're user-defined types, then they're references in which case a forward-declared, opaque type declaration like merely class Foo; instead of #include "foo.h" in the header file is sufficient).

What you need then is the header file for:

• The superclass, if this is a subclass

• Possibly any templated types which are used as method parameters and/or return types: apparently you're supposed to be able to forward-declare template classes too, but some compiler implementations may have a problem with that (though you could also encapsulate any templates e.g. List<X> as implementation details of a user-defined type e.g. ListX).

In practice, I might make a "standard.h" which includes all the system files (e.g. STL headers, O/S-specific types and/or any #defines, etc) that are used by any/all header files in the project, and include that as the first header in every application header file (and tell the compiler to treat this "standard.h" as the 'precompiled header file').

---

//contents of foo.h
#ifndef INC_FOO_H //or #pragma once
#define INC_FOO_H

#include "standard.h"
class Foo
{
public: //methods
  ... Foo-specific methods here ...
private: //data
  struct Impl;
  Impl* m_impl;
};
#endif//INC_FOO_H

---

//contents of foo.cpp
#include "foo.h"
#include "bar.h"
Foo::Foo()
{
  m_impl = new Impl();
}
struct Foo::Impl
{
  Bar m_bar;
  ... etc ...
};
... etc ...

Answer 3

I usually create one source file (main.c, for example) and one header file for that source file (main.h). In the source file, I put all the main sort of "interface" functions that I use in that file (in main, it'd be main()), and then whatever functions I get after refactoring those functions (implementation details), go below. In the header file, I declare some extern functions which are defined in other source files, but used in the source file which uses that header. Then I declare whatever structs or other data types that I use in that source file.

Then I just compile and link them all together. It stays nice and clean. A typical include ... section, in my current project looks like this

#include<windows.h>
#include<windowsx.h>
#include<stdio.h>

#include"interface.h"
#include"thissourcefile.h"

//function prototypes

//source

there's an interface header which keeps track of the data structures I use in all of the forms in the project, and then thissourcefile.h which does exactly what I just explained (declares externs, etc).

Also, I never define anything in my headers, I only put declarations there. That way they can be included by different source files and still link successfully. Function prototypes (extern, static, or otherwise) and declarations go in the header, that way they can be used lots of times--definitions go in the source, because they only need to be in one place.

This would obviously be different if you were creating a library, or something like that. But just for internal project linking, I find this keeps everything nice and clean. Plus, if you write a makefile, (or you just use an IDE), then compiling is really simple and efficient.

Answer 4

As far as tools go, I've used Imagix (this was about 6 years ago) on windows to identify includes that are unneeded as well as includes which are needed but are indirectly included thru another include.

Answer 5

I have the habit of ordering my includes from high abstraction level to low abstraction level. This has as a consequence that headers must be self-sufficient, and that any lack of self-sufficiency is quickly revealed by compiler errors.

For example a class 'Tetris' has a Tetris.h and Tetris.cpp file. The include order for Tetris.cpp would be

#include "Tetris.h"     // corresponding header first
#include "Block.h"      // ..then application level includes
#include "Utils/Grid.h" // ..then library dependencies
#include <vector>       // ..then stl
#include <windows.h>    // ..then system includes

This system helps avoiding hidden dependencies.

And now I realize this doesn't really answer your question since this system does not really help to clean up unneeded includes. Ah well..

Answer 6

Depending on the size of your project, looking at the include graphs created by Doxygen (with the INCLUDE_GRAPH option on ) can be helpful.

Answer 7

I've been thinking about writing something that compiles each non-header file individually many times, each time removing an #include statement. Continue doing this until a minimal set of includes is achieved.

I think this is misguided, and will lead to "insufficient but just happens to work" include sets.

Suppose your source file uses numeric_limits, but also includes some header file, that for reasons of its own includes <limits>. That doesn't mean that your source file shouldn't include <limits>. That other header file probably isn't documented to define everything defined in <limits>, it just so happens to do so. Some day it might stop: maybe it only uses one value as a default parameter of some function, and maybe that default value changes from std::numeric_limits<T>::min() to 0. And now your source file doesn't compile any more, and the maintainer of that header file didn't even know that your file existed until it broke his build.

Unless you have crippling build problems right this minute, I think the best way to remove redundant includes is just to get into the habit of looking over the list whenever you touch a file for maintenance. If you find that you have dozens of includes, and having reviewed the file you still can't figure out what each one is for, consider breaking down into smaller files.

Answer 8

One big problem with the remove a header and recompile technique is that it can lead to still-compiling, but wrong or inefficient code.

1. Template specialization: If you have a template specialization for a specific type that is in one header and the more general template in another, removing the specialization may leave the code in a compilable state, but with undesired results.

2. Overload resolution: A similar issue - if you have two overloads of one function in different headers, but that take somewhat compatible types, you can end up removing the version that is the better fit in one case, but still have the code compile. This is probably less likely than the template specialization version, but it is possible.