[This is the second of a pair of articles I posted in 1997 on the difference between #include "" and #include <>, and strategies for using them. This article starts out by replying to John Winters, who had commented on the first one. I have altered the text somewhat for this web page.]

From: scs@eskimo.com (Steve Summit)
Subject: Re: Quotes vs. angle brackets in #includes
Date: Thu, 5 Jun 1997 14:54:42 GMT
Message-ID: <EBB5F7.BEG@eskimo.com>
Newsgroups: comp.lang.c,comp.unix.programmer

Several weeks ago, in <5kve70$t89@polo.demon.co.uk>, John Winters wrote:
> I find it hard to conceive of a justification for using angle
> brackets for your own header files...

Well, there was one described in the earlier article of mine (part of which you quote below, and which I'll say more about). But that involved a somewhat obscure situation, and I think I might even be prepared to justify the use of angle brackets for other than standard headers on more prosaic grounds, such as that they simply look better or send a different message. Double quotes say ``These are local headers.'' Angle brackets say, ``These are system headers.'' Where do library and third-party headers fall? Squarely in the middle. Two kinds of header inclusion aren't really enough, but they're all C gives us. If you've got a shared or third-party header to #include, and you don't want to make it look local, it's very tempting to put it in angle brackets (perhaps with a blank line separating it from the really standard headers up above). At least, I've been tempted to do so; in fact that's the way I tend to do it.

Now, it's true that you can certainly decide to use double quotes for all library headers, and it's true that if you focus on the Standard, angle brackets don't just say ``These are system headers'', they say, ``These are Standard headers.'' So there are definitely arguments in favor of leaning strongly towards double quotes for all but Standard headers, but there are also (I claim) some reasons -- which are at least decent, if not necessarily compelling -- to lean towards angle brackets for some of your less than purely local headers.

...

>> On the other hand, the problem as I've stated it only comes up if
>> you use double quotes. If you use angle brackets for the central
>> project header files (in particular, for nested inclusions of
>> central project header files), the preprocessor never starts at
>> the wrong ``current'' directory, but always traverses the search
>> path from the start again, and finds the one in your private
>> include directory, as you wanted, regardless.
>
> Presumably you had some particular implementation in mind, but in
> general there is no reason to expect the behaviour which you describe.

The ``particular implementation'' I had in mind is none other than Unix, along with at least half of the non-Unix compilers out there which have followed the old Unix preprocessor's lead in handling double quotes. The question is, if double quotes in an #include directive mean ``use the current directory'', which current directory is it? If the directory we were in when we invoked the compiler and the directory containing the .c file are the same, then there's little question. But what if the .c file is not in what we would otherwise think of as ``the current directory''? Worse, what if a .c file (in the current directory) includes a header file, which for whatever reason is not in the current directory, and that header file contains an #include directive using double quotes? Many preprocessors, including the traditional Unix ones, take the position that ``the current directory'' which is first searched by #include "" in that case is the directory containing the file containing the #include directive. (I know you know about this possibility, John, but evidently you haven't had to deal with some of its even more surprising consequences.)

Why would you want this rule at all? Why not make ``the current directory'' be unilaterally the current directory? The answer is that, at least some of the time, the Unix behavior is exactly what you want. Suppose I'm trying to share some code, but in an informal way, neither as a library nor even as a precompiled object file; suppose that I call (in my Makefile, project definition, or other build procedure) for some .c file in some other directory to be compiled and the resulting object linked into my program. (Why am I compiling it myself, instead of using a precompiled object or library? Perhaps I need to compile it with my own set of macro definitions to pick a certain set of #ifdef groupings.) Suppose further that this source file I'm trying to compile has, off in its own directory where I'm compiling it from, its own header file which I don't otherwise need. The Unix behavior will then ``do the right thing'': the file's header, which it includes using double quotes, will be found over in its directory.

Now, to be sure, there would have been other ways to skin this cat. If the rule were that the current directory is always the current directory, such that the preprocessor was unable to find the remote source file's header in this case, I could have added the remote file's directory to the list of directories to be searched for header files. But I assume that this is the sort of case that prompted the adoption of what I'm calling the ``Unix'' preprocessor behavior, and programmers on large projects need to understand it, because many compilers have followed Unix's lead, but the rule does have, as I've said, some even more surprising consequences.

Suppose I'm working with several other people on a large project. Suppose we've got a central directory, /usr/project, containing the project source code, and a subdirectory of central directory, /usr/project/include, containing all the shared project source files. (Perhaps we chose to place the header files in their own directory because they are shared between many of the libraries or modules which make up the project and which are distributed among several other subdirectories of /usr/project.) Suppose further that I have my own directory, /usr/scs/project, in which I'm working on my own parts of the project prior to releasing them to the central directory, and suppose that I've got a subdirectory /usr/scs/project/include containing any project header files which I'm modifying. Naturally, whenever I compile part of the project, I'll arrange (using -I switches, if under Unix) that the directories /usr/scs/project/include and /usr/project/include be additionally searched for header files, and in that order.

Finally, suppose that there are two files project.h and types.h in /usr/project/include, and that project.h contains the line

	#include "types.h"

One day, I have occasion to make a change to types.h. I make a copy in /usr/scs/project/include, edit it, and then arrange to recompile (in my own directory) those parts of the project which will be affected by the change, which will probably be nearly all of them. When I come back from lunch, to try the recompiled project out, I find that my changes to types.h have not taken effect! Why not? Because whenever any of the project's .c files included "project.h", they found it in /usr/project/include, because I don't have my own local copy of project.h. But by the rules of ``the current directory is the includer's directory,'' when /usr/project/include/project.h asks for "types.h", the first place to look is in /usr/project/include, so it finds /usr/project/include/types.h, not /usr/scs/project/include/types.h. To fix the problem (assuming I can even figure out what the heck is going on), I need to make a copy of project.h in /usr/scs/project/include, even though I have no intention of modifying it, just so that all compiles will find it first and so that its #include of "types.h" will find my modified copy of types.h. (Then, I go off for a long afternoon snack break while I recompile the entire project, again.)

Now, this may sound like such a complicated, contrived, and obscure situation that it's not worth making implementation strategy decisions on. I'd even be inclined to agree with you, except that Larry Weiss knew immediately and exactly what I was talking about, so he's obviously run into this problem, too. (And I suspect that he and I are not alone.) When you find yourself in this situation, you have exactly 4 choices that I can think of:

  1. Use a different compiler with a different set of rules for resolving #include "" lines.

  2. Disallow ``nested #include files''; don't have project.h do an #include on types.h at all.

  3. Always check out project.h (into your own, private work directory, that is) whenever you want to work on types.h, and accept some wasted compilation time when you forget.

  4. Have project.h use
    	#include <types.h>
    
    instead, so that it always unambiguously finds the first one in the header file search path, without tripping up on the question of what ``the current directory'' is.

The last time this happened to me (for yes, as you may have guessed, not all of these ``suppose'''s have been hypothetical) it occurred to me that this is a much stronger argument for a ban on nested header files (i.e. choice #2) than the more usual but much weaker argument of ``it makes it harder to find where things are defined,'' an argument which is truly convincing only to those who are still unclear on the concept of ``grep''. But if you're not willing to ban nested #include files, and if you're not in a position to switch compilers, and if you're tired of massive recompilations (or, worse, of searching in vain for some other cause of a bug after your fix to a central header file didn't seem to work), then #4 may be your best or only option, implementation-defined though it may be.

Heretical advice? Perhaps...

Steve Summit
scs@eskimo.com