clang-format does not recognize full path header, e.g. <.../MainModule.h>, as corresponding header when sorting

[Quuxplusone]

Bugzilla Link	PR26634
Status	CONFIRMED
Importance	P normal
Reported by	Tobias Markmann (tm@ayena.de)
Reported on	2016-02-16 07:54:48 -0800
Last modified on	2020-06-03 06:48:23 -0700
Version	unspecified
Hardware	All All
CC	ababos@bloomberg.net, ddomenichelli@drdanz.it, djasper@google.com, klimek@google.com, llvm-bugs@lists.llvm.org, marejde@gmail.com, mydeveloperday@gmail.com
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also

A source file located under Name/Folder/MainModule.cpp and the corresponding
header file at Name/Folder/MainModule.h. The source file starts with including
the corresponding header like

#include <Name/Folder/MainModule.h>
#include <cstdio>
#include <Name/Folder/SomeOtherHeader.h>
.

When running "clang-format -i Name/Folder/MainModule.cpp" with sorting in the
.clang-format configuration, it does not put the corresponding header, i.e.
<Name/Folder/MainModule.h>, first but sorts it all alphabetically. Changing it
to a local include like #include "MainModule.h" it is sorted first. Why does it
not recognize the corresponding header file when it's prefixed with a subpath?

[Quuxplusone]

This is because clang::tooling::IncludeCategoryManager::isMainHeader ignores
any header that's not included with quotes:
https://github.com/llvm-mirror/clang/blob/c104c07a624f422cb229a8cc94c9433f144029dc/lib/Tooling/Inclusions/HeaderIncludes.cpp#L178

This should be a very simple 2 line fix, just remove that quotes check.

[Quuxplusone]

But isn't the <> in #include <XXX> supposed to be reserved for system headers?

if we removed that check and imagine you were developing a class call list.cxx

#include "list.h"

#include <vector>
#include <list>

which would be the main header?  #include "list.h" or #include <list>

[Quuxplusone]

1. The meaning of <> vs. "" is implementation defined:
https://en.cppreference.com/w/cpp/preprocessor/include. In practice a lot of
C++ shops use <> for includes coming from the include path, and "" for relative
paths.
2. Since the main module is typically included first, an acceptable heuristic
in case of a conflict would be to pick the first suitable one.
3. We could also define a filter for standard headers to further avoid that
conflict.
4. How else would you suggest resolving the issue Tobias outlined? Assuming
changing the include style is not an option.

[Quuxplusone]

I still feel reading cppreference is that the intent is that <> are for those
supported by -isystem and " is for those by -i certainly I know I don't use "
for just relative includes, but thats probably projecting my own opinion ;-)

However I think as you say the issue is easily avoided by allow both, I don't
think we could do that without making it an option in the .clang-format file or
previously sorted includes could potentially start moving around

If someone would like to propose a name for a suitable option (In my view the
hardest part of this issue), I'l be happy to propose a potential change

Here's my input, but I'm sure someone could think of a better name

AllowSytemHeadersAsMainHeader: true?
AllowChevronHeadersAsMainHeader: true?

[Quuxplusone]

Either works... Although Chevron might be a bit more obscure for non-native speakers. An alternative suggestion could be to name it after the current behavior, which would be:

ExcludeSytemHeadersAsMainHeader

[Quuxplusone]

I like ExcludeSytemHeadersAsMainHeader much better