System headers

This guide tries to explain how clangd finds system headers while providing its functionality. It aims to provide users with enough understanding to resolve any issues around these headers being missing.

Contents

What are system headers ?

In the context of this guide, any header a project depends on but doesn’t exist in the repository is considered a system header. These usually include:

These headers are usually provided either by a custom toolchain, which might be part of the repository, or directly via system installed libraries.

Clangd itself only ships with its own built-in headers, because they are tied to the version of clang embedded in clangd. The rest (including C++ STL) must be provided by your system.

How clangd finds those headers

Clangd comes with an embedded clang parser. Hence it makes use of all the mechanisms that exist in clang for lookups, while adding some extra spices to increase chances of discovery on Mac environments. Here follows some information about what clang does.

Search directories mentioned with compile flags

As most other compilers clang provides some command line flags to control system header search explicitly. Most important of these is -isystem, which adds a directory to system include search path.

Best way to ensure clangd can find your system includes is by putting the directories to be searched into your compile flags via -isystem. You can achieve this with compile_flags.txt, compile_commands.json or a clangd configuration file.

You might also want to take a look at -isysroot, -system-header-prefix and env variables respected by clang.

Heuristic search for system headers

Clang performs some toolchain specific searches to find suitable directories for system header search. The heuristics used by most of these search algorithms primarily rely on the directory containing the clang driver and the target triple.

You can investigate this search by invoking any clang with -v, for example clang -v -c -xc++ /dev/null (you can replace /dev/null with nul on windows). This prints out:

...
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/10
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/8
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/9
Selected GCC installation: /usr/lib/gcc/x86_64-linux-gnu/10
...
ignoring nonexistent directory "/usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include"
ignoring nonexistent directory "/include"
#include "..." search starts here:
#include <...> search starts here:
 /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10
 /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10
 /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward
 /usr/lib/clang/13.0.0/include
 /usr/local/include
 /usr/include/x86_64-linux-gnu
 /usr/include
End of search list.

Directories after #include <...> search starts here includes all the directories that will be used for system header search.

Directory of the driver

These heuristics often expect the standard library to be found near the compiler. Therefore clangd needs to know where the compiler is, especially when using a custom toolchain.

Clangd makes use of the first argument of the compile flags as the driver’s path. Ideally this argument should specify full path to the compiler.

For example, for a compile_commands.json entry like: { "directory": "/home/user/llvm/build", "command": "/usr/bin/clang++ -c -o file.o file.cc", "file": "file.cc" }, first argument is /usr/bin/clang++.

Note that, in case of a compile_flags.txt driver name defaults to clang-tool sitting next to clangd binary.

Target Triple

The second important factor is target triple, which specifies the architecture and OS to build for. It can be explicitly specified with --target compile flag or can be deduced implicitly from the driver name.

This enables clang to try to locate appropriate headers for the target platform for example with --target=x86_64-w64-mingw32 clang will look for mingw installed headers, which is one common toolchain for windows. You can see its effects on the header search dirs by executing clang --target=x86_64-w64-mingw32 -xc++ -v -c /dev/null (and without the target info).

This can also be achieved by implicitly including target information in the driver name, but is a lot more subtle and less convenient. So this guide doesn’t go into much details about it, but you can find more here.

Query-driver

Instead of trying to guess the header search paths, clangd can also try to query the actual compiler. For example if your compile flags has /custom/compiler as the driver name, clangd will run something similar to /custom/compiler -E -xc++ -v /dev/null and parse its output (this should work with variants of gcc, clang, and other compilers with compatible interfaces).

Note that this is a mechanism that solely exists in clangd and has nothing to do with clang.

It can be used as a last resort when clang’s heuristics are not enough to detect standard library locations being used by your custom toolcahin.

Since it implies executing arbitrary binaries, that might be checked-in with the project, clangd does not perform this inference automatically. You need to allowlist binaries you deem safe for execution using --query-driver clangd command line option. Please note that this option is just an allowlist and the real driver to be executed still comes from the compile command. It is a list of comma separated globs and a driver from a compile command needs to match at least one of these globs. For example to whitelist drivers in paths:

You can pass --query-driver="/path/to/my/project/arm-g*,/path/to/other/project/gcc" into clangd. You can find details about changing clangd arguments used by your editor in here, but it is always best to check your editor/LSP client’s documents directly.

Fixing missing system header issues

Since we’ve established some basic understanding of how system header search works for clang and clangd. Now let’s talk about how to fix missing system header issues.

Headers not present in the system at all

As mentioned above, clangd doesn’t ship with its own standard library. If you can build the project you are working on the same machine you are using clangd, you probably have the headers you need on your system but clangd is failing to find them, so you can just skip this section.

If you know your system lacks one, you should get it from some place suitable for your platform. Unfortunately this document is not the best place to talk about choices or how to get them but here are some choices:

After getting the headers clangd should hopefully be able to detect them, assuming they are not installed to a non-default location.

You can build your project but clangd is complaining about missing headers

In such a case you can start by checking your clangd logs to see compile flags being used by clangd. Easiest way to achieve this is by executing clangd --check=/path/to/a/file/in/your/project.cc. Outputs logs should contain something like:

I[17:11:57.203] Compile command from CDB is: ...
I[17:11:57.206] internal (cc1) args are: -cc1 ...

Note that --check is only available starting from clangd-12. For earlier versions you can open the file in your editor and access clangd logs through your LSP plugin.

If you are seeing a log line containing Generic fallback command is instead of the one above, it means clangd is not able to pick your compile commands. If you don’t have any compilation database, it is expected. But otherwise you should fix that first.

You should first try executing the command from CDB to see if it can compile your file. If it can’t, it means you have problems with your compile command again and probably there’s a discrepancy between what your build system uses and what’s being fed into clangd.

Compile command provided to clangd works on its own

There are usually two reasons for this failure. Either driver mentioned in the compile command is relative or it is employing custom heuristics unknown to clang.

Relative driver name

As mentioned above most of the clang’s heuristics rely on the location of the driver. If clangd cannot figure out the absolute path for your driver, then all those relative search heuristics will fail.

The best option here is changing the driver name in your CDB to use absolute path rather than relative paths. Other than that you can also try putting the directory containing your driver into $PATH so clangd can make it absolute itself.

Your driver has heuristics unknown to clang

This is the worst scenario to hit, and unfortunately is common for custom toolchains targetting embedded devices.

You can execute the driver with -v option to see all the search directories it has found and the target triple being used. Afterwards there are a couple options:

✏️