How GN handles cross-compiling

About 2 min

How GN handles cross-compiling

As a GN user

GN has robust support for doing cross compiles and building things for multiple architectures in a single build (e.g., to build some things to run locally and some things to run on an embedded device). In fact, there is no limit on the number of different architectures you can build at once; the Chromium build uses at least four in some configurations.

To start, GN has the concepts of a host and a target. The host is the platform that the build is run on, and the target is the platform where the code will actually run (This is different from autotools' terminology, but uses the more common terminology for cross compiling).

(Confusingly, GN also refers to each build artifact -- an executable, library, etc. -- as a target. On this page, we will use "target" only to refer to the system you want to run your code on, and use "rule" or some other synonym to refer to a specific build artifact).

When GN starts up, the host_os and host_cpu variables are set automatically to match the operating system (they can be overridden in args files, which can be useful in weird corner cases). The user can specify that they want to do a cross-compile by setting either or both of target_os and target_cpu; if they are not set, the build config files will usually set them to the host's values, though the Chromium build will set target_cpu to "arm" if target_os is set to "android").

So, for example, running on an x64 Linux machine:

gn gen out/Default
1

is equivalent to:

gn gen out/Default --args='target_os="linux" target_cpu="x64"'
1

To do an 32-bit ARM Android cross-compile, do:

gn gen out/Default --args='target_os="android"'
1

(We don't have to specify target_cpu because of the conditionals mentioned above).

And, to do a 64-bit MIPS Chrome OS cross-compile:

gn gen out/Default --args='target_os="chromeos" target_cpu="mips64el"'
1

As a BUILD.gn author

If you are editing build files outside of the //build directory (i.e., not directly working on toolchains, compiler configs, etc.), generally you only need to worry about a few things:

The current_toolchain, current_cpu, and current_os variables reflect the settings that are currently in effect in a given rule. The is_linux, is_win etc. variables are updated to reflect the current settings, and changes to cflags, ldflags and so forth also only apply to the current toolchain and the current thing being built.

You can also refer to the target_cpu and target_os variables. This is useful if you need to do something different on the host depending on which target_arch is requested; the values are constant across all toolchains. You can do similar things for the host_cpu and host_os variables, but should generally never need to.

For the default toolchain, target_cpu and current_cpu are the same. For a secondary toolchain, current_cpu is set based on the toolchain definition and target_cpu remains the same. When writing rules, current_cpu should be used rather than target_cpu most of the time.

By default, dependencies listed in the deps variable of a rule use the same (currently active) toolchain. You may specify a different toolchain using the foo(bar) label notation as described in the label section of the reference doc.

Here's an example of when to use target_cpu vs current_cpu:

declare_args() {
  # Applies only to toolchains targeting target_cpu.
  sysroot = ""
}

config("my_config") {
  # Uses current_cpu because compile flags are toolchain-dependent.
  if (current_cpu == "arm") {
    defines = [ "CPU_IS_32_BIT" ]
  } else {
    defines = [ "CPU_IS_64_BIT" ]
  }
  # Compares current_cpu with target_cpu to see whether current_toolchain
  # has the same architecture as target_toolchain.
  if (sysroot != "" && current_cpu == target_cpu) {
    cflags = [
      "-isysroot",
      sysroot,
    ]
  }
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

As a //build/config or //build/toolchain author

The default_toolchain is declared in the //build/config/BUILDCONFIG.gn file. Usually the default_toolchain should be the toolchain for the target_os and target_cpu. The current_toolchain reflects the toolchain that is currently in effect for a rule.

Be sure you understand the differences between host_cpu, target_cpu, current_cpu, and toolchain_cpu (and the os equivalents). The first two are set as described above. You are responsible for making sure that current_cpu is set appropriately in your toolchain definitions; if you are using the stock templates like gcc_toolchain and msvc_toolchain, that means you are responsible for making sure that toolchain_cpu and toolchain_os are set as appropriate in the template invocations.