naiveproxy/tools/gn/docs/cross_compiles.md
2018-01-29 00:30:36 +08:00

126 lines
4.8 KiB
Markdown

# 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](http://www.gnu.org/software/automake/manual/html_node/Cross_002dCompilation.html)'
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
```
is equivalent to:
```
gn gen out/Default --args='target_os="linux" target_cpu="x64"'
```
To do an 32-bit ARM Android cross-compile, do:
```
gn gen out/Default --args='target_os="android"'
```
(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"'
```
## 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](reference.md#Toolchains).
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,
]
}
}
```
## 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.