Related
What order should include files be specified, i.e. what are the reasons for including one header before another?
For example, do the system files, STL, and Boost go before or after the local include files?
I don't think there's a recommended order, as long as it compiles! What's annoying is when some headers require other headers to be included first... That's a problem with the headers themselves, not with the order of includes.
My personal preference is to go from local to global, each subsection in alphabetical order, i.e.:
h file corresponding to this cpp file (if applicable)
headers from the same component,
headers from other components,
system headers.
My rationale for 1. is that it should prove that each header (for which there is a cpp) can be #included without prerequisites (terminus technicus: header is "self-contained"). And the rest just seems to flow logically from there.
The big thing to keep in mind is that your headers should not be dependent upon other headers being included first. One way to insure this is to include your headers before any other headers.
"Thinking in C++" in particular mentions this, referencing Lakos' "Large Scale C++ Software Design":
Latent usage errors can be avoided by ensuring that the .h file of a component parses by itself – without externally-provided declarations or definitions... Including the .h file as the very first line of the .c file ensures that no critical piece of information intrinsic to the physical interface of the component is missing from the .h file (or, if there is, that you will find out about it as soon as you try to compile the .c file).
That is to say, include in the following order:
The prototype/interface header for this implementation (ie, the .h/.hh file that corresponds to this .cpp/.cc file).
Other headers from the same project, as needed.
Headers from other non-standard, non-system libraries (for example, Qt, Eigen, etc).
Headers from other "almost-standard" libraries (for example, Boost)
Standard C++ headers (for example, iostream, functional, etc.)
Standard C headers (for example, cstdint, dirent.h, etc.)
If any of the headers have an issue with being included in this order, either fix them (if yours) or don't use them. Boycott libraries that don't write clean headers.
Google's C++ style guide argues almost the reverse, with really no justification at all; I personally tend to favor the Lakos approach.
I follow two simple rules that avoid the vast majority of problems:
All headers (and indeed any source files) should include what they need. They should not rely on their users including things.
As an adjunct, all headers should have include guards so that they don't get included multiple times by over-ambitious application of rule 1 above.
I also follow the guidelines of:
Include system headers first (stdio.h, etc) with a dividing line.
Group them logically.
In other words:
#include <stdio.h>
#include <string.h>
#include "btree.h"
#include "collect_hash.h"
#include "collect_arraylist.h"
#include "globals.h"
Although, being guidelines, that's a subjective thing. The rules on the other hand, I enforce rigidly, even to the point of providing 'wrapper' header files with include guards and grouped includes if some obnoxious third-party developer doesn't subscribe to my vision :-)
To add my own brick to the wall.
Each header needs to be self-sufficient, which can only be tested if it's included first at least once
One should not mistakenly modify the meaning of a third-party header by introducing symbols (macro, types, etc.)
So I usually go like this:
// myproject/src/example.cpp
#include "myproject/example.h"
#include <algorithm>
#include <set>
#include <vector>
#include <3rdparty/foo.h>
#include <3rdparty/bar.h>
#include "myproject/another.h"
#include "myproject/specific/bla.h"
#include "detail/impl.h"
Each group separated by a blank line from the next one:
Header corresponding to this cpp file first (sanity check)
System headers
Third-party headers, organized by dependency order
Project headers
Project private headers
Also note that, apart from system headers, each file is in a folder with the name of its namespace, just because it's easier to track them down this way.
I recommend:
The header for the .cc module you're building. (Helps ensure each header in your project doesn't have implicit dependencies on other headers in your project.)
C system files.
C++ system files.
Platform / OS / other header files (e.g. win32, gtk, openGL).
Other header files from your project.
And of course, alphabetical order within each section, where possible.
Always use forward declarations to avoid unnecessary #includes in your header files.
I'm pretty sure this isn't a recommended practice anywhere in the sane world, but I like to line system includes up by filename length, sorted lexically within the same length. Like so:
#include <set>
#include <vector>
#include <algorithm>
#include <functional>
I think it's a good idea to include your own headers before other peoples, to avoid the shame of include-order dependency.
This is not subjective. Make sure your headers don't rely on being #included in specific order. You can be sure it doesn't matter what order you include STL or Boost headers.
First include the header corresponding to the .cpp... in other words, source1.cpp should include source1.h before including anything else. The only exception I can think of is when using MSVC with pre-compiled headers in which case, you are forced to include stdafx.h before anything else.
Reasoning: Including the source1.h before any other files ensures that it can stand alone without it's dependencies. If source1.h takes on a dependency on a later date, the compiler will immediately alert you to add the required forward declarations to source1.h. This in turn ensures that headers can be included in any order by their dependants.
Example:
source1.h
class Class1 {
Class2 c2; // a dependency which has not been forward declared
};
source1.cpp
#include "source1.h" // now compiler will alert you saying that Class2 is undefined
// so you can forward declare Class2 within source1.h
...
MSVC users: I strongly recommend using pre-compiled headers. So, move all #include directives for standard headers (and other headers which are never going to change) to stdafx.h.
Include from the most specific to the least specific, starting with the corresponding .hpp for the .cpp, if one such exists. That way, any hidden dependencies in header files that are not self-sufficient will be revealed.
This is complicated by the use of pre-compiled headers. One way around this is, without making your project compiler-specific, is to use one of the project headers as the precompiled header include file.
Several separate considerations are conflated when deciding for a particular include order. Let try to me untangle.
1. check for self-containedness
Many answers suggest that the include order should act as a check that your headers are self-contained. That mixes up the consideration of testing and compilation
You can check separately whether your headers are self-included. That "static analysis" is independent of any compilation process. For example, run
gcc headerfile.h -fsyntax-only
Testing whether your header files are self-contained can easily be scripted/automated. Even your makefile can do that.
No offense but Lakos' book is from 1996 and putting those different concerns together sounds like 90s-style programming to me. That being said, there are ecosystems (Windows today or in the 90s?) which lack the tools for scripted/automated tests.
2. Readability
Another consideration is readability. When you look up your source file, you just want to easily see what stuff has been included. For that your personal tastes and preferences matter most, though typically you either order them from most specific to least specific or the other way around (I prefer the latter).
Within each group, I usually just include them alphabetically.
3. Does the include order matter?
If your header files are self-contained, then the include order technically shouldn't matter at all for the compilation result.
That is, unless you have (questionable?) specific design choices for your code, such as necessary macro definitions that are not automatically included. In that case, you should reconsider your program design, though it might work perfectly well for you of course.
It is a hard question in the C/C++ world, with so many elements beyond the standard.
I think header file order is not a serious problem as long as it compiles, like squelart said.
My ideas is: If there is no conflict of symbols in all those headers, any order is OK, and the header dependency issue can be fixed later by adding #include lines to the flawed .h.
The real hassle arises when some header changes its action (by checking #if conditions) according to what headers are above.
For example, in stddef.h in VS2005, there is:
#ifdef _WIN64
#define offsetof(s,m) (size_t)( (ptrdiff_t)&(((s *)0)->m) )
#else
#define offsetof(s,m) (size_t)&(((s *)0)->m)
#endif
Now the problem: If I have a custom header ("custom.h") that needs to be used with many compilers, including some older ones that don't provide offsetof in their system headers, I should write in my header:
#ifndef offsetof
#define offsetof(s,m) (size_t)&(((s *)0)->m)
#endif
And be sure to tell the user to #include "custom.h" after all system headers, otherwise, the line of offsetof in stddef.h will assert a macro redefinition error.
We pray not to meet any more of such cases in our career.
First of all i want to say that I read about precompiled headers and I understand that this is an optimization that saves me the time of compiling headers over and over on every built.
I'm reading the documentation of boost and I see that in the instructions they say:
In Configuration Properties > C/C++ > Precompiled Headers, change Use Precompiled Header (/Yu) to Not Using Precompiled Headers
And then they explain it:
There's no problem using Boost with precompiled headers; these instructions merely avoid precompiled headers because it would require Visual Studio-specific changes to the source code used in the examples.
Can some explain me the sentence I marked in bold? which visual studio specific changes they are talking about ? (Here is the link to the documentation I'm reading: http://www.boost.org/doc/libs/1_55_0/more/getting_started/windows.html#pch)
Why and when I would want to turn off the precompiled headers?
what is the difference between "Create" and "Use" in the precompiled header options.
Originally a comment, but I may as well post it. Note: this is specific to VC++:
The bold sentence is their way of saying the samples don't follow the mantra of a unified use-this-lead-in-header-for-pch-generation model. IOW, their samples aren't PCH-friendly, but you can still use pch with boost in your projects if properly configured.
You would turn them off for a variety of reasons. Some source modules, particularly ones from 3rd-parties, don't follow the PCH model of including "the" pch-through-header at their outset. Their samples are such code (and thus the advise to turn them off for their samples). Sometimes source files require different preprocessor configurations only for this files and not all files int he project; another reason to disable PCH for those files.
You typically use a source/header pair to generate "the One"; the precompiled header image. This header file typically includes:
Any system standard lib headers used by your project
3rd-party SDK headers
Just about everything else that is NOT in active development for your project.
The single source file tagged as Create typically includes one line of code : #include "YourHeaderFile.h", where YourHeaderFile.h is the header you filled with stuff from the list above. Tagging it as "Create" through header YourHeaderFile.h tells VC it is the file needed for rebuilding the PCH through that header when compiling other source files. All other source files are tagged as Use (except the ones where PCH is turned off) and should include, as their first line of code, the same #include "TheHeaderFile.h".
In short (hard to believe), <boost> is telling you their samples aren't setup like described above, and as such you should turn PCH off when building them.
When you use pre-compiled headers, you need to do something like:
#include <foo>
#include <bar>
#include <baz>
#pragma hdrstop
// other code here
Everything before the #pragma goes into the precompiled header. Everything after it depends on the precompiled header. The VC++ specific "magic" to make pre-compiled header work is that #pragma.
There's a little more to the story than just that though. To make pre-compiled headers work well, you want to include exactly the same set of headers in exactly the same order in every source file.
That leads to (typically) creating one header that includes all the other common headers and has the #pragma hdrstop right at its end, then including that in all the other source files.
Then, when the compiler does its thing, there are two phases: first you need to create a pre-compiled header. This means running the compiler with one switch. The compiler only looks at what comes before the #pragma hdrstop, builds a symbol table (and such) and puts the data into a .pch file.
Then comes the phase when you do a build using the pre-compiled header. In this phase, the compiler simply ignores everything in the the file up to the #pragma hdrstop. When it gets to that, it reads the compiler's internal state from the .pch file, and then starts compiling that individual file.
This means each source file typically includes a lot of headers it doesn't actually need. That, in turn, means that if you don't use pre-compiled headers, you end up with compilation that's much slower than if you hadn't done anything to support pre-compiled headers at all.
In other words, although the only part that's absolutely required is the #pragma hdrstop, which is fairly innocuous, a great deal more file re-structuring is needed to get much benefit from them--and those changes are likely to actively harmful to compilation time if you're using anything that doesn't support pre-compiled headers (and in the same way VC++ does them at that).
When precompiled headers is on every cpp source file must start with #include "stdafx.h"
So you would turn it off if you do not want to edit all the boost source files.
When precompiled headers is on stdafx.cpp "creates" the precompiled header. All other files "use" the precompiled header.
I have a precompiled header that contains includes for various 3rd party libraries, e.g.:
#ifndef PRECOMPILED_H
#define PRECOMPILED_H
#include "booststuff.h"
#include "luastuff.h"
#endif
Where booststuff.h and luastuff.h are header files in my project that just include various boost / lua related things and set up some typedefs / usings / namespace aliases.
I set up the precompiled header in the usual way inside visual studio (2012), and use the force include option to include it in every cpp file.
In the cpp files, I've also been fairly careful to #include "booststuff.h" where I actually use it as well (I sometimes disable precompiled headers to test this). However, I've been wondering lately whether that's a good idea. So:
Does anything bad happen if I include a file again that's already included in the precompiled header (I don't see why it would, but I've seen things about headers having to be included "in the same order", and not really understood what they were on about)?
Does it affect Intellisense (unusably slow with a fairly small project)? I'd be happy to give up some portability for better Intellisense since I currently have no desire to switch platforms.
If each include file has #pragma once in it, the compiler will completely skip reading the file on the second and subsequent attempts to include it. It isn't stated explicitly but I assume the precompiled header tracks this information as well.
What order should include files be specified, i.e. what are the reasons for including one header before another?
For example, do the system files, STL, and Boost go before or after the local include files?
I don't think there's a recommended order, as long as it compiles! What's annoying is when some headers require other headers to be included first... That's a problem with the headers themselves, not with the order of includes.
My personal preference is to go from local to global, each subsection in alphabetical order, i.e.:
h file corresponding to this cpp file (if applicable)
headers from the same component,
headers from other components,
system headers.
My rationale for 1. is that it should prove that each header (for which there is a cpp) can be #included without prerequisites (terminus technicus: header is "self-contained"). And the rest just seems to flow logically from there.
The big thing to keep in mind is that your headers should not be dependent upon other headers being included first. One way to insure this is to include your headers before any other headers.
"Thinking in C++" in particular mentions this, referencing Lakos' "Large Scale C++ Software Design":
Latent usage errors can be avoided by ensuring that the .h file of a component parses by itself – without externally-provided declarations or definitions... Including the .h file as the very first line of the .c file ensures that no critical piece of information intrinsic to the physical interface of the component is missing from the .h file (or, if there is, that you will find out about it as soon as you try to compile the .c file).
That is to say, include in the following order:
The prototype/interface header for this implementation (ie, the .h/.hh file that corresponds to this .cpp/.cc file).
Other headers from the same project, as needed.
Headers from other non-standard, non-system libraries (for example, Qt, Eigen, etc).
Headers from other "almost-standard" libraries (for example, Boost)
Standard C++ headers (for example, iostream, functional, etc.)
Standard C headers (for example, cstdint, dirent.h, etc.)
If any of the headers have an issue with being included in this order, either fix them (if yours) or don't use them. Boycott libraries that don't write clean headers.
Google's C++ style guide argues almost the reverse, with really no justification at all; I personally tend to favor the Lakos approach.
I follow two simple rules that avoid the vast majority of problems:
All headers (and indeed any source files) should include what they need. They should not rely on their users including things.
As an adjunct, all headers should have include guards so that they don't get included multiple times by over-ambitious application of rule 1 above.
I also follow the guidelines of:
Include system headers first (stdio.h, etc) with a dividing line.
Group them logically.
In other words:
#include <stdio.h>
#include <string.h>
#include "btree.h"
#include "collect_hash.h"
#include "collect_arraylist.h"
#include "globals.h"
Although, being guidelines, that's a subjective thing. The rules on the other hand, I enforce rigidly, even to the point of providing 'wrapper' header files with include guards and grouped includes if some obnoxious third-party developer doesn't subscribe to my vision :-)
To add my own brick to the wall.
Each header needs to be self-sufficient, which can only be tested if it's included first at least once
One should not mistakenly modify the meaning of a third-party header by introducing symbols (macro, types, etc.)
So I usually go like this:
// myproject/src/example.cpp
#include "myproject/example.h"
#include <algorithm>
#include <set>
#include <vector>
#include <3rdparty/foo.h>
#include <3rdparty/bar.h>
#include "myproject/another.h"
#include "myproject/specific/bla.h"
#include "detail/impl.h"
Each group separated by a blank line from the next one:
Header corresponding to this cpp file first (sanity check)
System headers
Third-party headers, organized by dependency order
Project headers
Project private headers
Also note that, apart from system headers, each file is in a folder with the name of its namespace, just because it's easier to track them down this way.
I recommend:
The header for the .cc module you're building. (Helps ensure each header in your project doesn't have implicit dependencies on other headers in your project.)
C system files.
C++ system files.
Platform / OS / other header files (e.g. win32, gtk, openGL).
Other header files from your project.
And of course, alphabetical order within each section, where possible.
Always use forward declarations to avoid unnecessary #includes in your header files.
I'm pretty sure this isn't a recommended practice anywhere in the sane world, but I like to line system includes up by filename length, sorted lexically within the same length. Like so:
#include <set>
#include <vector>
#include <algorithm>
#include <functional>
I think it's a good idea to include your own headers before other peoples, to avoid the shame of include-order dependency.
This is not subjective. Make sure your headers don't rely on being #included in specific order. You can be sure it doesn't matter what order you include STL or Boost headers.
First include the header corresponding to the .cpp... in other words, source1.cpp should include source1.h before including anything else. The only exception I can think of is when using MSVC with pre-compiled headers in which case, you are forced to include stdafx.h before anything else.
Reasoning: Including the source1.h before any other files ensures that it can stand alone without it's dependencies. If source1.h takes on a dependency on a later date, the compiler will immediately alert you to add the required forward declarations to source1.h. This in turn ensures that headers can be included in any order by their dependants.
Example:
source1.h
class Class1 {
Class2 c2; // a dependency which has not been forward declared
};
source1.cpp
#include "source1.h" // now compiler will alert you saying that Class2 is undefined
// so you can forward declare Class2 within source1.h
...
MSVC users: I strongly recommend using pre-compiled headers. So, move all #include directives for standard headers (and other headers which are never going to change) to stdafx.h.
Include from the most specific to the least specific, starting with the corresponding .hpp for the .cpp, if one such exists. That way, any hidden dependencies in header files that are not self-sufficient will be revealed.
This is complicated by the use of pre-compiled headers. One way around this is, without making your project compiler-specific, is to use one of the project headers as the precompiled header include file.
Several separate considerations are conflated when deciding for a particular include order. Let try to me untangle.
1. check for self-containedness
Many answers suggest that the include order should act as a check that your headers are self-contained. That mixes up the consideration of testing and compilation
You can check separately whether your headers are self-included. That "static analysis" is independent of any compilation process. For example, run
gcc headerfile.h -fsyntax-only
Testing whether your header files are self-contained can easily be scripted/automated. Even your makefile can do that.
No offense but Lakos' book is from 1996 and putting those different concerns together sounds like 90s-style programming to me. That being said, there are ecosystems (Windows today or in the 90s?) which lack the tools for scripted/automated tests.
2. Readability
Another consideration is readability. When you look up your source file, you just want to easily see what stuff has been included. For that your personal tastes and preferences matter most, though typically you either order them from most specific to least specific or the other way around (I prefer the latter).
Within each group, I usually just include them alphabetically.
3. Does the include order matter?
If your header files are self-contained, then the include order technically shouldn't matter at all for the compilation result.
That is, unless you have (questionable?) specific design choices for your code, such as necessary macro definitions that are not automatically included. In that case, you should reconsider your program design, though it might work perfectly well for you of course.
It is a hard question in the C/C++ world, with so many elements beyond the standard.
I think header file order is not a serious problem as long as it compiles, like squelart said.
My ideas is: If there is no conflict of symbols in all those headers, any order is OK, and the header dependency issue can be fixed later by adding #include lines to the flawed .h.
The real hassle arises when some header changes its action (by checking #if conditions) according to what headers are above.
For example, in stddef.h in VS2005, there is:
#ifdef _WIN64
#define offsetof(s,m) (size_t)( (ptrdiff_t)&(((s *)0)->m) )
#else
#define offsetof(s,m) (size_t)&(((s *)0)->m)
#endif
Now the problem: If I have a custom header ("custom.h") that needs to be used with many compilers, including some older ones that don't provide offsetof in their system headers, I should write in my header:
#ifndef offsetof
#define offsetof(s,m) (size_t)&(((s *)0)->m)
#endif
And be sure to tell the user to #include "custom.h" after all system headers, otherwise, the line of offsetof in stddef.h will assert a macro redefinition error.
We pray not to meet any more of such cases in our career.
I am working on a large C++ project in Visual Studio 2008, and there are a lot of files with unnecessary #include directives. Sometimes the #includes are just artifacts and everything will compile fine with them removed, and in other cases classes could be forward declared and the #include could be moved to the .cpp file. Are there any good tools for detecting both of these cases?
While it won't reveal unneeded include files, Visual studio has a setting /showIncludes (right click on a .cpp file, Properties->C/C++->Advanced) that will output a tree of all included files at compile time. This can help in identifying files that shouldn't need to be included.
You can also take a look at the pimpl idiom to let you get away with fewer header file dependencies to make it easier to see the cruft that you can remove.
PC Lint works quite well for this, and it finds all sorts of other goofy problems for you too. It has command line options that can be used to create External Tools in Visual Studio, but I've found that the Visual Lint addin is easier to work with. Even the free version of Visual Lint helps. But give PC-Lint a shot. Configuring it so it doesn't give you too many warnings takes a bit of time, but you'll be amazed at what it turns up.
There's a new Clang-based tool, include-what-you-use, that aims to do this.
!!DISCLAIMER!! I work on a commercial static analysis tool (not PC Lint). !!DISCLAIMER!!
There are several issues with a simple non parsing approach:
1) Overload Sets:
It's possible that an overloaded function has declarations that come from different files. It might be that removing one header file results in a different overload being chosen rather than a compile error! The result will be a silent change in semantics that may be very difficult to track down afterwards.
2) Template specializations:
Similar to the overload example, if you have partial or explicit specializations for a template you want them all to be visible when the template is used. It might be that specializations for the primary template are in different header files. Removing the header with the specialization will not cause a compile error, but may result in undefined behaviour if that specialization would have been selected. (See: Visibility of template specialization of C++ function)
As pointed out by 'msalters', performing a full analysis of the code also allows for analysis of class usage. By checking how a class is used though a specific path of files, it is possible that the definition of the class (and therefore all of its dependnecies) can be removed completely or at least moved to a level closer to the main source in the include tree.
I don't know of any such tools, and I have thought about writing one in the past, but it turns out that this is a difficult problem to solve.
Say your source file includes a.h and b.h; a.h contains #define USE_FEATURE_X and b.h uses #ifdef USE_FEATURE_X. If #include "a.h" is commented out, your file may still compile, but may not do what you expect. Detecting this programatically is non-trivial.
Whatever tool does this would need to know your build environment as well. If a.h looks like:
#if defined( WINNT )
#define USE_FEATURE_X
#endif
Then USE_FEATURE_X is only defined if WINNT is defined, so the tool would need to know what directives are generated by the compiler itself as well as which ones are specified in the compile command rather than in a header file.
Like Timmermans, I'm not familiar with any tools for this. But I have known programmers who wrote a Perl (or Python) script to try commenting out each include line one at a time and then compile each file.
It appears that now Eric Raymond has a tool for this.
Google's cpplint.py has an "include what you use" rule (among many others), but as far as I can tell, no "include only what you use." Even so, it can be useful.
If you're interested in this topic in general, you might want to check out Lakos' Large Scale C++ Software Design. It's a bit dated, but goes into lots of "physical design" issues like finding the absolute minimum of headers that need to be included. I haven't really seen this sort of thing discussed anywhere else.
Give Include Manager a try. It integrates easily in Visual Studio and visualizes your include paths which helps you to find unnecessary stuff.
Internally it uses Graphviz but there are many more cool features. And although it is a commercial product it has a very low price.
You can build an include graph using C/C++ Include File Dependencies Watcher, and find unneeded includes visually.
If your header files generally start with
#ifndef __SOMEHEADER_H__
#define __SOMEHEADER_H__
// header contents
#endif
(as opposed to using #pragma once) you could change that to:
#ifndef __SOMEHEADER_H__
#define __SOMEHEADER_H__
// header contents
#else
#pragma message("Someheader.h superfluously included")
#endif
And since the compiler outputs the name of the cpp file being compiled, that would let you know at least which cpp file is causing the header to be brought in multiple times.
PC-Lint can indeed do this. One easy way to do this is to configure it to detect just unused include files and ignore all other issues. This is pretty straightforward - to enable just message 766 ("Header file not used in module"), just include the options -w0 +e766 on the command line.
The same approach can also be used with related messages such as 964 ("Header file not directly used in module") and 966 ("Indirectly included header file not used in module").
FWIW I wrote about this in more detail in a blog post last week at http://www.riverblade.co.uk/blog.php?archive=2008_09_01_archive.xml#3575027665614976318.
Adding one or both of the following #defines
will exclude often unnecessary header files and
may substantially improve
compile times especially if the code that is not using Windows API functions.
#define WIN32_LEAN_AND_MEAN
#define VC_EXTRALEAN
See http://support.microsoft.com/kb/166474
If you are looking to remove unnecessary #include files in order to decrease build times, your time and money might be better spent parallelizing your build process using cl.exe /MP, make -j, Xoreax IncrediBuild, distcc/icecream, etc.
Of course, if you already have a parallel build process and you're still trying to speed it up, then by all means clean up your #include directives and remove those unnecessary dependencies.
Start with each include file, and ensure that each include file only includes what is necessary to compile itself. Any include files that are then missing for the C++ files, can be added to the C++ files themselves.
For each include and source file, comment out each include file one at a time and see if it compiles.
It is also a good idea to sort the include files alphabetically, and where this is not possible, add a comment.
If you aren't already, using a precompiled header to include everything that you're not going to change (platform headers, external SDK headers, or static already completed pieces of your project) will make a huge difference in build times.
http://msdn.microsoft.com/en-us/library/szfdksca(VS.71).aspx
Also, although it may be too late for your project, organizing your project into sections and not lumping all local headers to one big main header is a good practice, although it takes a little extra work.
If you would work with Eclipse CDT you could try out http://includator.com to optimize your include structure. However, Includator might not know enough about VC++'s predefined includes and setting up CDT to use VC++ with correct includes is not built into CDT yet.
The latest Jetbrains IDE, CLion, automatically shows (in gray) the includes that are not used in the current file.
It is also possible to have the list of all the unused includes (and also functions, methods, etc...) from the IDE.
Some of the existing answers state that it's hard. That's indeed true, because you need a full compiler to detect the cases in which a forward declaration would be appropriate. You cant parse C++ without knowing what the symbols mean; the grammar is simply too ambiguous for that. You must know whether a certain name names a class (could be forward-declared) or a variable (can't). Also, you need to be namespace-aware.
Maybe a little late, but I once found a WebKit perl script that did just what you wanted. It'll need some adapting I believe (I'm not well versed in perl), but it should do the trick:
http://trac.webkit.org/browser/branches/old/safari-3-2-branch/WebKitTools/Scripts/find-extra-includes
(this is an old branch because trunk doesn't have the file anymore)
If there's a particular header that you think isn't needed anymore (say
string.h), you can comment out that include then put this below all the
includes:
#ifdef _STRING_H_
# error string.h is included indirectly
#endif
Of course your interface headers might use a different #define convention
to record their inclusion in CPP memory. Or no convention, in which case
this approach won't work.
Then rebuild. There are three possibilities:
It builds ok. string.h wasn't compile-critical, and the include for it
can be removed.
The #error trips. string.g was included indirectly somehow
You still don't know if string.h is required. If it is required, you
should directly #include it (see below).
You get some other compilation error. string.h was needed and isn't being
included indirectly, so the include was correct to begin with.
Note that depending on indirect inclusion when your .h or .c directly uses
another .h is almost certainly a bug: you are in effect promising that your
code will only require that header as long as some other header you're using
requires it, which probably isn't what you meant.
The caveats mentioned in other answers about headers that modify behavior
rather that declaring things which cause build failures apply here as well.