How to set up CMake to cross compile with clang for ARM embedded on Windows?
My question was answered through the reply to my bug report, but I add the answer here to have all information in one place for future reference.
In short: CMake currently does not support to use the clang/clang++ command line interface if you install Clang from llvm.org. If you want to use the clang/clang++ interface (which is necessary to cross-compile for ARM) you have to install Clang via msys2.
In detail
Clang on Windows has two different command line interfaces:
- clang/clang++ the default interface that attempts to be compatible with GCCs gcc/g++ and targets the GNU ABI
- clang-cl that attempts to be compatible with Microsofts Visual C++ compiler cl.exe and targets the MSVC ABI
In order to cross-compile for ARM you need the clang/clang++ interface. The Problem is CMake supports different interfaces depending on how you installed Clang (see the bug in the CMake issue tracker for more details):
- If you install Clang from llvm.org CMake only supports the clang-cl interface.
- If you install Clang via msys2 CMake supports the clang/clang++ interface.
So here is what I did:
- Install msys2
- Install Clang and CMake with pacman. There are two clang packages in msys2, a mingw32 and a mingw64 version. I used the mingw64 package (mingw-w64-x86_64-clang).
- Launch the mingw64 shell and run CMake and build from there.
Toolchain file
There were two problems with my original toolchain file that took my a long time to fix. So I hope this will save others some time:
- The target triple (e.g. arm-none-eabi) needs to match the prefix of the GCC binutils exactly. The prefix of my binutils was arm-none-eabi (e.g. arm-none-eabi-ar) so I had to change the target triple accordingly.
- CMAKE_TRY_COMPILE_TARGET_TYPE needs to be changed to STATIC_LIBRARY in order to prevent CMake from running the linker during the compile check.
Here is the final toolchain file I used (you can also find a good example for a toolchain file in this GitHub repo):
cmake_minimum_required(VERSION 3.13)
set(CMAKE_SYSTEM_NAME Generic)
set(CMAKE_SYSTEM_PROCESSOR ARM)
if(DEFINED ENV{GCC_ARM_TOOLCHAIN})
set(GCC_ARM_TOOLCHAIN $ENV{GCC_ARM_TOOLCHAIN})
else()
set(GCC_ARM_TOOLCHAIN "C:/Users/user/tools/gcc-arm-none-eabi-7-2018-q2-update-win32")
endif()
LIST(APPEND CMAKE_PROGRAM_PATH ${GCC_ARM_TOOLCHAIN})
# Specify the cross compiler
# The target triple needs to match the prefix of the binutils exactly
# (e.g. CMake looks for arm-none-eabi-ar)
set(CLANG_TARGET_TRIPLE arm-none-eabi)
set(GCC_ARM_TOOLCHAIN_PREFIX ${CLANG_CLANG_TARGET_TRIPLE})
set(CMAKE_C_COMPILER clang)
set(CMAKE_C_COMPILER_TARGET ${CLANG_TARGET_TRIPLE})
set(CMAKE_CXX_COMPILER clang++)
set(CMAKE_CXX_COMPILER_TARGET ${CLANG_TARGET_TRIPLE})
set(CMAKE_ASM_COMPILER clang)
set(CMAKE_ASM_COMPILER_TARGET ${CLANG_TARGET_TRIPLE})
# Don't run the linker on compiler check
set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
# Specify compiler flags
set(ARCH_FLAGS "-mcpu=cortex-a5 -mthumb -mfpu=neon-vfpv4 -mfloat-abi=hard -mno-unaligned-access")
set(CMAKE_C_FLAGS "-Wall ${ARCH_FLAGS}" CACHE STRING "Common flags for C compiler")
set(CMAKE_CXX_FLAGS "-Wall -std=c++17 -fno-exceptions -fno-rtti -fno-threadsafe-statics ${ARCH_FLAGS}" CACHE STRING "Common flags for C++ compiler")
set(CMAKE_ASM_FLAGS "-Wall ${ARCH_FLAGS} -x assembler-with-cpp" CACHE STRING "Common flags for assembler")
set(CMAKE_EXE_LINKER_FLAGS "-nostartfiles -Wl,-Map,kernel.map,--gc-sections -fuse-linker-plugin -Wl,--use-blx --specs=nano.specs --specs=nosys.specs" CACHE STRING "")
# C/C++ toolchain
set(GCC_ARM_SYSROOT "${GCC_ARM_TOOLCHAIN}/${GCC_ARM_TOOLCHAIN_PREFIX}")
# set(CMAKE_SYSROOT ${GCC_ARM_SYSROOT})
set(CMAKE_FIND_ROOT_PATH ${GCC_ARM_SYSROOT})
# Search for programs in the build host directories
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
# For libraries and headers in the target directories
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
Here is an alternative answer that uses CMake to cross compile for embedded ARM using the GCC toolchain, rather than clang, and also uses plain old make rather than Ninja. I know this does not directly answer the question, but it is a very reasonable alternative for embedded ARM, and many of the Eclipse-based vendor toolchains we are using now are based on GCC rather than LLVM/clang.
First install the latest "GNU MCU Eclipse ARM Embedded GCC" toolchain from https://github.com/gnu-mcu-eclipse/arm-none-eabi-gcc/releases (NOTE: no installer, just unzip to where you want it).
Next, install latest MSYS2/MinGW using the installer available at https://www.msys2.org/.
For some reason the default install of MSYS2 does not include make. So, after using pacman to update the default packages per the install instructions, execute 'pacman -S make' to install make.
There is a cmake build available within MSYS2. However, it does not include generators for MSYS or MinGW (no idea why), and also (or because of) seems to have problems finding the external GCC toolchain, even when full paths are explicitly provided on the command line or in a toolchain file. So, this solution uses the native Window build of CMake, available at https://cmake.org/download/, and NOT the MSYS2 build of cmake.
By default the MSYS shell does NOT inherit the Windows environment, so the native Windows CMake will not be on the MSYS shell path. The simple solution is to specify the full path to cmake on the command line. In my case, I have the GCC cross compiler toolchain at C:/GNU MCU Eclipse/ARM Embedded GCC/8.2.1-1.4-20190214-0604. So, a full command to cmake looks like this.
/c/Program\ Files/CMake/bin/cmake.exe -DCMAKE_BUILD_TYPE=Debug -DCMAKE_TOOLCHAIN_FILE=cmake/toolchain-arm-none-eabi.cmake "-DTOOLCHAIN_PREFIX=C:/GNU MCU Eclipse/ARM Embedded GCC/8.2.1-1.4-20190214-0604" -G "MSYS Makefiles" ../
Here the command is being executed from a build directory one level down from the project root, and a sibling cmake directory contains the specified toolchain file. Note, backslashes are necessary to escape any spaces in the path to cmake.exe, but are NOT used in the quoted paths passed as argument to cmake.
You can also run cmake and make right from a plain old Windows command line (cmd.exe) as long as both the cmake and msys bin directories on on the Windows path. In that case there is no need to specify the full path to cmake. No success from PowerShell, however.