Go creates static binaries by default unless you use cgo to call C code, in which case it will create a dynamically linked library. Turns out that using cgo is more common than many people assume as the os/user and net packages use cgo, so importing either (directly or indirectly) will result in a dynamic binary.

The easiest way to check if your program is statically compiled is to run file on it:

$ file test.dynamic | tr , '\n'
test.dynamic: ELF 64-bit LSB executable
 x86-64
 version 1 (SYSV)
 dynamically linked
 interpreter /lib/ld-linux-x86-64.so.2
 Go BuildID=LxsDWU_fMQ9Cox6y4bSV/fdMBNuZAmOuPSIKb2RXJ/rcazy_d6AbaoNtes-qID/nRiDtV1fOY2eoEVlyqnu
 not stripped

$ file test.static | tr , '\n'
test.static:  ELF 64-bit LSB executable
 x86-64
 version 1 (SYSV)
 statically linked
 Go BuildID=hz56qplN20RU01EMBelb/58lm7IuCas399AWvpycN/BGETSDXvSFKK3BUjfgon/5xa5xLDJTC90556SUlNh
 not stripped

Notice the “dynamically linked” and “statically linked”. You can also run ldd, but note this only works if the binary matches your system’s architecture:

$ ldd test.dynamic
test.dynamic:
        linux-vdso.so.1 (0x00007ffe00302000)
        libpthread.so.0 => /usr/lib/libpthread.so.0 (0x00007f3f86f4a000)
        libc.so.6 => /usr/lib/libc.so.6 (0x00007f3f86d87000)
        /lib/ld-linux-x86-64.so.2 (0x00007f3f86f80000)

$ ldd test.static
test.static:
        not a dynamic executable

You can verify that a binary runs without external dependencies with chroot (this requires root on most platforms):

$ chroot . ./test.static
Hello, world!

$ chroot . ./test.dynamic
chroot: failed to run command './test.dynamic': No such file or directory

The “No such file or directory” error is a bit obscure, but it means that the dynamic linker (ld-linux) isn’t found. Unfortunately this is the exact same message as when the test.dynamic itself isn’t found, so make sure you didn’t typo it. I’m not sure if there’s any way to get Linux to emit a more useful message for this.[1]


There are two packages in the standard library that use cgo:

  • os/user contains cgo code to use the standard C library to map user and group ids to user and group names. There is also a Go implementation which parses /etc/passwd and /etc/group. The advantage of using the C library is that it can also get user information from LDAP or NIS. If you don’t use that – most people don’t – then there is no real difference.

    On Windows there is only a Go implementation, and this doesn’t apply.

  • net can use the C standard library to resolve domain names, but by default it uses the Go client. The C library has a few more features (e.g. you can configure getaddrinfo() with /etc/gai.conf) and some platforms don’t have a resolv.conf (e.g. Android), but for most cases the Go library should work well.

Your binary will no longer be statically linked if your program imports one of those two packages, either directly through a dependency. Especially the net one is quite common.

You can use the osusergo and netgo build tags to skip building the cgo parts:

$ go build -tags osusergo
$ go build -tags netgo
$ go build -tags osusergo,netgo

For simple cases where you don’t use any other cgo code it’s probably easier to just disable cgo, since the cgo code is protected with +build cgo:

$ CGO_ENABLED=0 go build

What if we want to use the cgo versions of the above? Or what if we want to use a cgo package such as SQLite? In those cases you can tell the C linker to statically link with -extldflags:

$ go build -ldflags="-extldflags=-static"

The nested -s look a bit confusing and are easy to forget, so be sure to pay attention (or maybe that’s just me… 🤦‍♂️).[2]

Some packages – such as SQLite – may produce warnings:

$ go build -ldflags="-extldflags=-static"
# test
/usr/bin/ld: /tmp/go-link-400285317/000010.o: in function `unixDlOpen':
/[..]/sqlite3-binding.c:39689: warning: Using 'dlopen' in statically linked
applications requires at runtime the shared libraries from the glibc version used
for linking

dlopen() loads shared libraries at runtime; looking at the SQLite source code it’s only used only for dynamically loading extensions. This is not a commonly used feature so this warning can be safely ignored for most programs (you can verify with the chroot mentioned earlier).

The go-sqlite3 package provides a build flag to disable this, if you want to make the warnings go away and ensure this feature isn’t used:

$ go build -ldflags="-extldflags=-static" -tags sqlite_omit_load_extension

The os/user and net packages will give you a similar warnings about the getpwnam_r() etc. and getaddrinfo() functions; which also depend on runtime configurations. You can use the tags mentioned earlier to make sure the Go code is used.

Other cgo packages may emit similar warnings; you’ll have to check the documentation or source code to see if they’re significant or not.


One of Go’s nicer features is that you can cross-compile to any system/architecture combination from any system by just setting setting GOOS and GOARCH. I can build Windows binaries on OpenBSD with GOOS=windows GOARCH=amd64 go build. Neat!

With cgo cross-compiling gets a bit trickier as cross-compiling C code is trickier.

The short version is that cross-compiling to different architectures (amd64, arm, etc.) for the same OS isn’t too hard, but cross-compiling to different operating systems is rather harder. It’s certainly doable, but you need the entire toolchain and libraries for the target OS. It’s a bit of a hassle and probably easier to just start a virtual machine.

You’ll need to install the toolchain for the target architecture (and OS, if you’re compiling to a different OS); if you’re on Linux your package manager will probably already include it, but they’re named different on different distros. Usually searching for -linux-gnu (or -linux-musl) should give you an overview.

I’m very 😎 so I use Void Linux, and for extra 😎 I want to use musl libc, so that’s what I’ll use in this example to cross-compile to arm and arm64; let me know if you have the commands for other systems and I’ll add them as well.[3]

# Replace musl with gnu if you want to use GNU libc.
$ xbps-install cross-aarch64-linux-musl cross-armv7l-linux-musleabihf

$ GOOS=linux GOARCH=arm64 CGO_ENABLED=1 CC=aarch64-linux-musl-gcc \
    go build -ldflags='-extldflags=-static' -o test.arm64 ./test.go

$ GOOS=linux GOARCH=arm CGO_ENABLED=1 CC=armv7l-linux-musleabihf-gcc \
    go build -ldflags='-extldflags=-static' -o test.arm ./test.go

aarch64 and arm64 are the same thing, just with a different name, just as x86_64 and amd64. To confirm that it works, you can use QEMU; for example with a simple programs which runs select date() on a :memory: database:

$ qemu-aarch64 ./test.arm64
<nil> 2020-04-11

$ qemu-arm ./test.arm
<nil> 2020-04-11

Huzzah!


Finally, to make releasing binaries a bit easier I wrote a small script: gogo-release. It’s just a glorified for loop (a lot of software is, really) to make the above a bit easier. For non-cgo projects the defaults settings should work without problems. This is what I use to build GoatCounter (which includes SQLite):

matrix="
linux amd64
linux arm   CC=armv7l-linux-musleabihf-gcc
linux arm64 CC=aarch64-linux-musl-gcc
"

build_flags="-trimpath -ldflags='-extldflags=-static -w -s -X main.version=$tag' -tags osusergo,netgo,sqlite_omit_load_extension ./cmd/goatcounter"

export CGO_ENABLED=1

And then just run gogo-release.

I’m not really aware of a good tool to make cross-compiling to different systems easier; the closest I know of is xgo, which installs the required build environment in a container. It’s not bad (although it is a bit messy), but it only supports Linux, macOS, and Windows. This covers most use cases but ideally I’d like a generic solution to cover all platforms. I may work on this in the future, time and enthusiasm permitting 😅

Footnotes
  1. It’s a common source of confusion when a hashbang (#!/bin/prog) is set to a program that doesn’t exist at that location. 

  2. The value of -ldflags is passed to go tool link, which passes the value of -extldflags to the C compiler. There’s an issue to make this easier

  3. Figuring them out just from the package index is a bit too much work/hassle, and also untested so it may contain errors or silly typos.