Cross building

Cross building (or cross compilation) is the process of generating binaries for a platform that is not the one where the compiling process is running.

Cross compilation is mostly used to build software for an alien device, such as an embedded device where you don’t have an operating system nor a compiler available. It’s also used to build software for slower devices, like an Android machine or a Raspberry Pi where running the native compilation will take too much time.

In order to cross build a codebase the right toolchain is needed, with a proper compiler (cross compiler), a linker and the set of libraries matching the host platform.

GNU triplet convention

According to the GNU convention, there are three platforms involved in the software building:

  • Build platform: The platform on which the compilation tools are being executed.

  • Host platform: The platform on which the generated binaries will run.

  • Target platform: Only when building a cross compiler, it is the platform it will generate binaries for.

Depending on the values of these platforms, there are different scenarios:

  • Native building: when the build and the host platforms are the same, it means that the platform where the compiler is running is the same one where the generated binaries will run. This is the most common scenario.

  • Cross building: when the build and the host platform are different, it requires a cross compiler running in the build platform that generates binaries for the host platform.

The target platform plays an important role when compiling a cross compiler, in that scenario the target is the platform the compiler will generate binaries for: in order to be a cross compiler the host platform (where the cross compiler will run) has to be different from the target platform. If the build platform is also different, it is called Canadian Cross.

Let’s illustrate these scenarios with some examples:

  • The Android NDK is a cross compiler to Android: it can be executed in Linux (the build platform) to generate binaries for Android (the host platform).

  • The Android NDK was once compiled, during that compilation a different compiler was used running in a build platform (maybe Windows) to generate the actual Android NDK that will run in the host platform Linux, and as we saw before, that Android NDK cross compiler will generate binaries for a target platform which is Android.

The values of the build , host and target platforms are not absolute, and they depend on the process we are talking about: the host when compiling a cross compiler turns into the build when using that same cross compiler, or the target of the cross compiler is the host platform when we are using it to build binaries.

See also

One way to avoid this complexity is to run the compilation in the host platform, so both build and host will take the same value and it will be a native compilation. Docker is a very successful tool that can help you with this, read more about it in this section.

Cross building with Conan

If you want to cross build a Conan package (for example using your Linux machine) to build the zlib Conan package for Windows, you need to tell Conan where to find your toolchain/cross compiler.

There are two approaches:

  • Using a profile: install the toolchain in your computer and use a profile to declare the settings and point to the needed tools/libraries in the toolchain using the [env] section to declare, at least, the CC and CXX environment variables.

  • Using build requires: package the toolchain as a Conan package and include it as a build_requires.

Using a profile

Using a Conan profile we can declare not only the settings that will identify our binary (host settings), but also all the environment variables needed to use a toolchain or cross compiler. The profile needs the following sections:

  • A [settings] section containing the regular settings: os, arch, compiler and build_type depending on your library. These settings will identify your binary.

  • An [env] section with a PATH variable pointing to your installed toolchain. Also any other variable that the toolchain expects (read the docs of your compiler). Some build systems need a variable SYSROOT to locate where the host system libraries and tools are.

For example, in the following profile we declare the host platform to be Windows x86_64 with the compiler, version and other settings we are using. And we add the [env] section with all the variables needed to use an installed toolchain:

toolchain=/usr/x86_64-w64-mingw32 # Adjust this path
target_host=x86_64-w64-mingw32
cc_compiler=gcc
cxx_compiler=g++

[env]
CONAN_CMAKE_FIND_ROOT_PATH=$toolchain  # Optional, for CMake to find things in that folder
CONAN_CMAKE_SYSROOT=$toolchain  # Optional, if we want to define sysroot
CHOST=$target_host
AR=$target_host-ar
AS=$target_host-as
RANLIB=$target_host-ranlib
CC=$target_host-$cc_compiler
CXX=$target_host-$cxx_compiler
STRIP=$target_host-strip
RC=$target_host-windres

[settings]
# We are cross-building to Windows
os=Windows
arch=x86_64
compiler=gcc

# Adjust to the gcc version of your MinGW package
compiler.version=7.3
compiler.libcxx=libstdc++11
build_type=Release

You can find working examples at the bottom of this section.

Using build requires

Warning

This section refers to the experimental feature that is activated when using --profile:build and --profile:host in the command-line. It is currently under development, features can be added or removed in the following versions.

Instead of manually downloading the toolchain and creating a profile, you can create a Conan package with it. Starting with Conan v1.24 and the command line arguments --profile:host and --profile:build this should be a regular recipe, for older versions some more work is needed.

Conan v1.24 and newer

A recipe with a toolchain is like any other recipe with a binary executable:

import os
from conans import ConanFile

class MyToolchainXXXConan(ConanFile):
    name = "my_toolchain"
    version = "0.1"
    settings = "os", "arch", "compiler", "build_type"

    # Implement source() and build() as usual

    def package(self):
        # Copy all the required files for your toolchain
        self.copy("*", dst="", src="toolchain")

    def package_info(self):
        bin_folder = os.path.join(self.package_folder, "bin")
        self.env_info.CC = os.path.join(bin_folder, "mycompiler-cc")
        self.env_info.CXX = os.path.join(bin_folder, "mycompiler-cxx")
        self.env_info.SYSROOT = self.package_folder

The Conan package with the toolchain needs to fill the env_info object in the package_info() method with the same variables we’ve specified in the examples above in the [env] section of profiles.

Then you will need to consume this recipe as any regular build requires that belongs to the build context: you need to use the --profile:build argument in the command line while creating your library:

conan create path/to/conanfile.py --profile:build=profile_build --profile:host=profile_host
../_images/conan-my_toolchain.png

The profile profile_build will contain just the settings related to your build platform, where you are running the command, and the profile_host will list the settings for the host platform (and eventually the my_toolchain/0.1 as build_requires if it is not listed in the recipe itself).

Conan will apply the appropiate profile to each recipe, and will inject the environment of all the build requirements that belong to the build context before running the build() method of the libraries being compiled. That way, the environment variables CC, CXX and SYSROOT from my_toolchain/0.1 will be available and also the path to the bindirs directory from that package.

The above means that Conan is able to compile the full graph in a single execution, it will compile the build requires using the profile_build and then it will compile the libraries using the host_profile settings applying the environment of the former ones.

Starting with Conan v1.25 (if the user provides the --profile:build) it is possible to get the relative context where a recipe is running during a Conan invocation. The object instatiated from the recipe contains the following attributes:

  • self.settings will always contain the settings corresponding to the binary to build/retrieve. It will contain the settings from the profile profile_host when this recipe appears in the host context and the settings from the profile profile:build if this object belongs to the build context.

  • self.settings_build will always contain the settings provided in the profile profile_build, even if the recipe appears in the build context, the build requirements of the build requirements are expected to run in the build machine too.

  • self.settings_target: for recipes in the host context this attribute will be equal to None, for those in the build context, if will depend on the level of anidation:

    • for recipes that are build requirements of packages in the host context, this attribute will contain the settings from the profile profile_host, while

    • for recipes that are build requirements of other build requirements the self.settings_target will contain the values of the profile_build.

With previous attributes, a draft for a recipe that packages a cross compiler could follow this pattern:

class CrossCompiler(ConanFile):
    name = "my_compiler"

    settings = "os", "arch", "compiler", "build_type"
    options = {"target": "ANY"}
    default_options = {"shared": False, "target": None}

    def configure(self):
        settings_target = getattr(self, 'settings_target', None)
        if settings_target is None:
            # It is running in 'host', so Conan is compiling this package
            if not self.options.target:
                raise ConanInvalidConfiguration("A value for option 'target' has to be provided")
        else:
            # It is running in 'build' and it is being used as a BR, 'target' can be inferred from settings
            if self.options.target:
                raise ConanInvalidConfiguration("Value for the option 'target' will be computed from settings_target")
            self.options.target = "<target-value>"  # Use 'self.settings_target' to get this value

Conan older than v1.24

Warning

We ask you to use the previous approach for Conan 1.24 and newer, and avoid any specific modification of your recipes to make them work as build requirements in a cross building scenario.

With this approach, only one profile is provided in the command line (the --profile:host or just --profile) and it has to define the os_build and arch_build settings too. The recipe of this build requires has to be modified to take into account these settings and the compiler and build_type settings have to be removed because their values for the build platform are not defined in the profile:

from conans import ConanFile
import os


class MyToolchainXXXConan(ConanFile):
    name = "my_toolchain"
    version = "0.1"
    settings = "os_build", "arch_build"

    # As typically, this recipe doesn't declare 'compiler' and 'build_type',
    #   the source() and build() methods need a custom implementation
    def build(self):
        # Typically download the toolchain for the 'build' platform
        url = "http://fake_url.com/installers/%s/%s/toolchain.tgz" % (os_build, os_arch)
        tools.download(url, "toolchain.tgz")
        tools.unzip("toolchain.tgz")

    def package(self):
        # Copy all the required files for your toolchain
        self.copy("*", dst="", src="toolchain")

    def package_info(self):
        bin_folder = os.path.join(self.package_folder, "bin")
        self.env_info.PATH.append(bin_folder)
        self.env_info.CC = os.path.join(bin_folder, "mycompiler-cc")
        self.env_info.CXX = os.path.join(bin_folder, "mycompiler-cxx")
        self.env_info.SYSROOT = self.package_folder

With this approach we also need to add the path to the binaries to the PATH environment variable. The one and only profile has to include a [build_requires] section with the reference to our new packaged toolchain and it will also contain a [settings] section with the regular settings plus the os_build and arch_build ones.

This approach requires a special profile, and it needs a modified recipe without the compiler and build_type settings, Conan can still compile it from sources but it won’t be able to identify the binary properly and it can be really to tackle if the build requirements has other Conan dependencies.

Host settings os_build, arch_build, os_target and arch_target

Warning

These settings are being reviewed and might be deprecated in the future, we encourage you to try not to use them. If you need help with your use case, please open an issue in the Conan repository and we will help you.

Before Conan v1.24 the recommended way to deal with cross building was to use some extra settings like os_build, arch_build and os_target and arch_target. These settings have a special meaning for some Conan tools and build helpers, but they also need to be listed in the recipes themselves creating a dedicated set of recipes for installers and tools in general. This approach should be superseeded with the introduction in Conan 1.24 of the command line arguments --profile:host and --profile:build that allow to declare two different profiles with all the information needed for the corresponding platforms (see section above this one).

The meaning of those settings is the following:

  • The settings os_build and arch_build identify the build platform according to the GNU convention triplet. These settings are detected the first time you run Conan with the same values than the host settings, so by default, we are doing native building. You will probably never need to change the value of this setting because they describe where are you running Conan.

  • The settings os_target and arch_target identify the target platform. If you are building a cross compiler, these settings specify where the compiled code will run.

The rest of settings, as we already know, identify the host platform.

ARM architecture reference

Remember that the Conan settings are intended to unify the different names for operating systems, compilers, architectures etc.

Conan has different architecture settings for ARM: armv6, armv7, armv7hf, armv8. The “problem” with ARM architecture is that it’s frequently named in different ways, so maybe you are wondering what setting do you need to specify in your case.

Here is a table with some typical ARM platforms:

Platform

Conan setting

Raspberry PI 1

armv6

Raspberry PI 2

armv7 or armv7hf if we want to use the float point hard support

Raspberry PI 3

armv8 also known as armv64-v8a

Visual Studio

armv7 currently Visual Studio builds armv7 binaries when you select ARM.

Android armbeabi-v7a

armv7

Android armv64-v8a

armv8

Android armeabi

armv6 (as a minimal compatible, will be compatible with v7 too)

Examples

Examples using profiles

Linux to Windows

  • Install the needed toolchain, in Ubuntu:

    sudo apt-get install g++-mingw-w64 gcc-mingw-w64
    
  • Create a file named linux_to_win64 with the contents:

    toolchain=/usr/x86_64-w64-mingw32 # Adjust this path
    target_host=x86_64-w64-mingw32
    cc_compiler=gcc
    cxx_compiler=g++
    
    [env]
    CONAN_CMAKE_FIND_ROOT_PATH=$toolchain  # Optional, for CMake to find things in that folder
    CONAN_CMAKE_SYSROOT=$toolchain  # Optional, if we want to define sysroot
    CHOST=$target_host
    AR=$target_host-ar
    AS=$target_host-as
    RANLIB=$target_host-ranlib
    CC=$target_host-$cc_compiler
    CXX=$target_host-$cxx_compiler
    STRIP=$target_host-strip
    RC=$target_host-windres
    
    [settings]
    # We are cross-building to Windows
    os=Windows
    arch=x86_64
    compiler=gcc
    
    # Adjust to the gcc version of your MinGW package
    compiler.version=7.3
    compiler.libcxx=libstdc++11
    build_type=Release
    
  • Clone an example recipe or use your own recipe:

    git clone https://github.com/memsharded/conan-hello.git
    
  • Call conan create using the created linux_to_win64

    $ cd conan-hello && conan create . conan/testing --profile ../linux_to_win64
    ...
    [ 50%] Building CXX object CMakeFiles/example.dir/example.cpp.obj
    [100%] Linking CXX executable bin/example.exe
    [100%] Built target example
    

A bin/example.exe for Win64 platform has been built.

Windows to Raspberry Pi (Linux/ARM)

  • Install the toolchain: https://gnutoolchains.com/raspberry/ You can choose different versions of the GCC cross compiler. Choose one and adjust the following settings in the profile accordingly.

  • Create a file named win_to_rpi with the contents:

    target_host=arm-linux-gnueabihf
    standalone_toolchain=C:/sysgcc/raspberry
    cc_compiler=gcc
    cxx_compiler=g++
    
    [settings]
    os=Linux
    arch=armv7 # Change to armv6 if you are using Raspberry 1
    compiler=gcc
    compiler.version=6
    compiler.libcxx=libstdc++11
    build_type=Release
    
    [env]
    CONAN_CMAKE_FIND_ROOT_PATH=$standalone_toolchain/$target_host
    CONAN_CMAKE_SYSROOT=$standalone_toolchain/$target_host/sysroot
    PATH=[$standalone_toolchain/bin]
    CHOST=$target_host
    AR=$target_host-ar
    AS=$target_host-as
    RANLIB=$target_host-ranlib
    LD=$target_host-ld
    STRIP=$target_host-strip
    CC=$target_host-$cc_compiler
    CXX=$target_host-$cxx_compiler
    CXXFLAGS=-I"$standalone_toolchain/$target_host/lib/include"
    

The profiles to target Linux are all very similar. You probably just need to adjust the variables declared at the top of the profile:

  • target_host: All the executables in the toolchain starts with this prefix.

  • standalone_toolchain: Path to the toolchain installation.

  • cc_compiler/cxx_compiler: In this case gcc/g++, but could be clang/clang++.

  • Clone an example recipe or use your own recipe:

    git clone https://github.com/memsharded/conan-hello.git
    
  • Call conan create using the created profile.

    $ cd conan-hello && conan create . conan/testing --profile=../win_to_rpi
    ...
    [ 50%] Building CXX object CMakeFiles/example.dir/example.cpp.obj
    [100%] Linking CXX executable bin/example
    [100%] Built target example
    

A bin/example for Raspberry PI (Linux/armv7hf) platform has been built.

Windows to Windows CE

The Windows CE (WinCE) operating system is supported for CMake and MSBuild. Since WinCE depends on the MSVC compiler, Visual Studio and the according Windows CE platform SDK for the WinCE device have to be installed on the build host.

The os.platform defines the WinCE Platform SDK and is equal to the Platform in Visual Studio.

Some examples for Windows CE platforms:

  • SDK_AM335X_SK_WEC2013_V310

  • STANDARDSDK_500 (ARMV4I)

  • Windows Mobile 5.0 Pocket PC SDK (ARMV4I)

  • Toradex_CE800 (ARMV7)

The os.version defines the WinCE version and must be "5.0", "6.0" or "7.0".

CMake supports Visual Studio 2008 (compiler.version=9) and Visual Studio 2012 (compiler.version=11).

Example of an Windows CE conan profile:

[settings]
os=WindowsCE
os.version=8.0
os.platform=Toradex_CE800 (ARMV7)
arch=armv7
compiler=Visual Studio
compiler.version=11

# Release configuration
build_type=Release
compiler.runtime=MD

Note

Further information about CMake and WinCE can be found in the CMake documentation:

CMake - Cross Compiling for Windows CE

Linux/Windows/macOS to Android

Cross-building a library for Android is very similar to the previous examples, except the complexity of managing different architectures (armeabi, armeabi-v7a, x86, arm64-v8a) and the Android API levels.

Download the Android NDK here and unzip it.

Note

If you are in Windows the process will be almost the same, but unzip the file in the root folder of your hard disk (C:\) to avoid issues with path lengths.

Note

If you are using Android Studio, you may use already available Android NDK

To use the clang compiler, create a profile android_21_arm_clang. Once again, the profile is very similar to the RPI one:

include(default)
target_host=aarch64-linux-android
android_ndk=/Users/sse4/Library/Android/sdk/ndk-bundle  # Adjust this path
api_level=21
[settings]
arch=armv8
build_type=Release
compiler=clang
compiler.libcxx=libc++
compiler.version=9
os=Android
os.api_level=$api_level
[build_requires]
[options]
[env]
PATH=[$android_ndk/toolchains/llvm/prebuilt/darwin-x86_64/bin]  # Adjust this path
CHOST=$target_host
AR=$target_host-ar
AS=$target_host-as
RANLIB=$target_host-ranlib
CC=$target_host$api_level-clang
CXX=$target_host$api_level-clang++
LD=$target_host-ld
STRIP=$target_host-strip
CONAN_CMAKE_TOOLCHAIN_FILE=$android_ndk/build/cmake/android.toolchain.cmake
  • Clone, for example, the zlib library to try to build it to Android

git clone https://github.com/conan-io/conan-center-index.git
  • Call conan create using the created profile.

$ cd conan-center-index/recipes/zlib/1.2.11 && conan create . 1.2.11@ -pr:h ../android_21_arm_clang -pr:b default

...
-- Build files have been written to: /tmp/conan-zlib/test_package/build/ba0b9dbae0576b9a23ce7005180b00e4fdef1198
Scanning dependencies of target enough
[ 50%] Building C object CMakeFiles/enough.dir/enough.c.o
[100%] Linking C executable bin/enough
[100%] Built target enough
zlib/1.2.11 (test package): Running test()

A bin/enough for Android ARM platform has been built.

Examples using build requires

You can find one example on how to use build requires for cross-compiling to iOS in the iOS integration section in the documentation.

See also

See also