We have succeeded in building embedded Linux with Yocto for a quad-core NXP i.MX6 (ARM Cortex-A9). Next, we want to cross-compile our own Qt application. As we use CMake for building our Qt application, we must create a CMake toolchain file. I am going to give a line-line by line explanation of the CMake toolchain file. I used Yocto Morty and CMake v3.5.1 (as it comes with Ubuntu 16.04 LTS).
We built the embedded Linux system with Yocto. We installed the SDK in /opt/imx6/sdk
. The GCC toolchain, which run on the x86_64-based host computer, is located in /opt/imx6/sdk/sysroots/x86_64-pokysdk-linux
. The root file system and files needed for cross-builds like headers are located in /opt/imx6/sdk/sysroots/cortexa9hf-neon-poky-linux-gnueabi
. We must set an environment variable SYSROOTS
to /opt/imx6/sdk/sysroots
such that CMake knows where to find the host and target sysroot.
In addition to the C++ compiler, we need Qt utilities like qmake, moc and rcc to build a Qt application. These utilities run on the Intel-powered host – as the whole the cross-build does. Unfortunately, Yocto builds these utilies for the ARM-powered host. There is an easy workaround. On the host and build computer, we install the same Qt version as is used on the target device. Typically, this Qt version is installed on the host already, because we develop the application with this Qt version on the host. Then, we must tell CMake where to find the utilities. If we installed Qt in /opt/Qt5.7.1
, for example, we set the environment variable QT_BIN_PATH to /opt/Qt5.7.1/5.7/gcc_64/bin
.
In summary: We must set two environment variables in the Linux shell, in which we want to run CMake and the cross-build later.
export SYSROOTS=/opt/imx6/sdk/sysroots
export QT_BIN_PATH=/opt/Qt5.7.1/5.7/gcc_64/bin
These two environment variables are all we need for cross-compiling with CMake. Especially, we need not run the script environment-setup-cortexa9hf-neon-poky-linux-gnueabi
to initialise the Yocto-SDK and set two dozen environment variables.
We are now ready to write the CMake toolchain file. The first two lines set the operating system and architecture of the target system.
set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_PROCESSOR armv7-a)
Our target system runs on Linux
powered by an NXP i.MX6, which uses the armv7-a
architecture. The variable CMAKE_SYSTEM_NAME
sets the variable CMAKE_CROSSCOMPILING
to true
. We use CMAKE_CROSSCOMPILING
in CMake files to set variables differently in native and cross builds.
We specify where CMake can find the C and C++ compilers.
set(TOOLCHAIN_PREFIX $ENV{SYSROOTS}/x86_64-pokysdk-linux/usr/bin/arm-poky-linux-gnueabi/arm-poky-linux-gnueabi-)
set(CMAKE_C_COMPILER ${TOOLCHAIN_PREFIX}gcc)
set(CMAKE_CXX_COMPILER ${TOOLCHAIN_PREFIX}g++)
If we use an older version of CMake or a different setup, CMake’s compiler check may fail. The reason is that the compiler and linker require special flags like -mfloat-abi=hard
. CMake’s compiler check calls the compiler executables without any flags and there is no way to provide flags. If, for example, our toolchain takes advantage of an on-chip floating-point unit (GCC option -mfloat-abi=hard
), the compiler check produces object files for floating-point emulation (the default GCC option is -mfloat-abi=soft
) and fails to link against the “hard” floating-point library.
We have two options to get out this mess. The first option is to set the CMake variable CMAKE_TRY_COMPILE_TARGET_TYPE
. This skips running the linker, which was exactly the problem in the compiler check. This variable is available from version 3.6.
set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
The second option is to switch off the compiler check with the macros from the module CMakeForceCompiler
. This option is deprecated starting with CMake v3.5, but works for older versions.
include(CMakeForceCompiler)
CMAKE_FORCE_C_COMPILER(${TOOLCHAIN_PREFIX}gcc GNU)
CMAKE_FORCE_CXX_COMPILER(${TOOLCHAIN_PREFIX}g++ GNU)
We can now return to the normal toolchain file. We set the target sysroot CMAKE_SYSROOT
as follows.
set(CMAKE_SYSROOT $ENV{SYSROOTS}/cortexa9hf-neon-poky-linux-gnueabi)
CMake adds the option --sysroot=/opt/imx6/sdk/sysroots/cortexa9hf-neon-poky-linux-gnueabi
to every compiler and linker call.
The next two lines work around the problem that the Yocto build produces ARM executables for Qt utilities like moc and rcc – instead of Intel ones.
set(CMAKE_PREFIX_PATH ${CMAKE_SYSROOT}/usr/lib/cmake)
set(OE_QMAKE_PATH_EXTERNAL_HOST_BINS $ENV{QT_BIN_PATH})
The first line tells CMake’s find_*()
functions where to look for its Qt configuration files like Qt5CoreConfig.cmake
. These configuration files use the Yocto variable OE_QMAKE_PATH_EXTERNAL_HOST_BINS
to call Qt utilities like moc and rcc. It is a bit of a hack, but it is simple and works.
The rest of the toolchain file is pretty smooth sailing.
set(CMAKE_FIND_ROOT_PATH ${CMAKE_SYSROOT})
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
The CMAKE_FIND_ROOT_PATH tells the CMake find_*()
functions to search for programs, libraries and headers in the target sysroot directory ${CMAKE_SYSROOT}
. The following three settings restrict the search. find_program()
is told not to search in the root path. This makes sense, because ARM executables don’t run on the Intel host. find_library()
may only search in the root path and below containing files for the target system. The same is true for find_file
and find_path
. If we use packages, we should add another line restricting find_package()
to the given root path:
set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
The next lines specify the flags for the C and C++ compilers.
set(COMPILER_FLAGS " -march=armv7-a -marm -mfpu=neon -mfloat-abi=hard -mcpu=cortex-a9")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${COMPILER_FLAGS}" CACHE STRING "" FORCE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${COMPILER_FLAGS}" CACHE STRING "" FORCE)
It is important that we force the change of the variables CMAKE_C_FLAGS
and CMAKE_CXX_FLAGS
in CMake’s cache. Otherwise, we’ll see the failing compiler checks described above. Of course, you can adapt the COMPILER_FLAGS
to your own needs.
If the linker needs special flags, we can set them through the variables CMAKE_EXE_LINKER_FLAGS
, CMAKE_SHARED_LINKER_FLAGS
and CMAKE_STATIC_LINKER_FLAGS
. We trust CMake to come up with the right default options for the linker.
Here is the complete toolchain file toolchain-imx6.cmake
under MIT license. Feel free to adapt the toolchain file to your needs and use it in your projects – but keep the license notice.
We run CMake for our Qt application and cross-build it with the following commands:
$ cmake '-GCodeBlocks - Unix Makefiles' -DCMAKE_BUILD_TYPE:STRING=Debug -DCMAKE_CXX_COMPILER:STRING=$SYSROOTS/x86_64-pokysdk-linux/usr/bin/arm-poky-linux-gnueabi/arm-poky-linux-gnueabi-g++ -DCMAKE_TOOLCHAIN_FILE:FILEPATH=/path/to/toolchain-imx6.cmake /path/to/project
$ cmake --build . --target project
Of course, /path/to/toolchain-imx6.cmake
, /path/to/project
and project
are only placeholders for the toolchain file path, the directory path containing the project’s top-level CMakeLists.txt and the project’s name, respectively.
References
- Official CMake documentation of “Cross-Compiling for Linux“.
- Wiki page “CMake Cross Compiling” by Kitware. Slightly dated, but still helpful.