Skip to content

tonistiigi/xx

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

xx - Dockerfile cross-compilation helpers

GitHub release CI Status Go Report Card Docker Pulls

xx provides tools to support cross-compilation from Dockerfiles that understand the --platform flag passed in from docker build or docker buildx build. These helpers allow you to build multi-platform images from any architecture into any architecture supported by your compiler with native performance. Adding xx to your Dockerfile should only need minimal updates and should not require custom conditions for specific architectures.


Dockerfile cross-compilation primer

Cross-compilation can be achieved in Dockerfiles by using multi-stage builds and defining some of the stages to always run on the native architecture used by the builder and execute the cross-compiling compiler. By default, a Dockerfile stage started with FROM keyword default to the target architecture, but this can be overridden with a FROM --platform flag. Using automatic platform ARGs in global scope, the platform of the cross-compiler stage can be set to $BUILDPLATFORM while the value of $TARGETPLATFORM can be passed to the compiler with an environment variable.

After compilation, the resulting assets can be copied into another stage that will become the result of the build. Usually, this stage does not use FROM --platform so that every stage is based on the expected target architecture.

FROM --platform=$BUILDPLATFORM alpine AS xbuild
ARG TARGETPLATFORM
RUN ./compile --target=$TARGETPLATFORM -o /out/myapp

FROM alpine
COPY --from=xbuild /out/myapp /bin

Installation

xx is distributed with a Docker image tonistiigi/xx that contains a collection of helper scripts that read TARGET* environment variables to automatically configure the compilation targets. The scripts are based on Posix Shell, so they should work on top of any image but currently xx is expected to work on Alpine and Debian/Ubuntu based distros. In order to avoid unexpected changes, you may want to pin the image using an immutable digest. Although xx only contains shell scripts that are identical for every platform it is recommended to also import xx with FROM --platform=$BUILDPLATFORM, so that import commands are shared for all compilation targets.

FROM --platform=$BUILDPLATFORM tonistiigi/xx AS xx

FROM --platform=$BUILDPLATFORM alpine
# copy xx scripts to your build stage
COPY --from=xx / /
# export TARGETPLATFORM (or other TARGET*)
ARG TARGETPLATFORM
# you can now call xx-* commands
RUN xx-info env

xx currently contains xx-info, xx-apk, xx-apt-get, xx-cc, xx-c++, xx-clang, xx-clang++, xx-go, xx-cargo, xx-verify. xx-clang (and its aliases) creates additional aliases, eg. ${triple}-clang, ${triple}-pkg-config, on first invocation or on xx-clang --setup-target-triple call.

Supported targets

xx supports building from and into Linux amd64, arm64, arm/v7, s390x, ppc64le and 386, and Alpine, Debian and Ubuntu. Risc-V is supported for Go and Rust builds and for newer distros that provide Risc-V packages like alpine:edge or debian:sid.

Go builds that don't depend on system packages can additionally target MacOS and Windows on all architectures. C/C++/CGo/Rust builds are supported for MacOS targets when an external SDK image is provided.

xx-info command also works on RHEL-style distros but no support is provided for package manager wrappers(eg. yum, dnf) there.

xx-info - Information about the build context

xx-info command returns normalized information about the current build context. It allows you to get various information about your build target and configuration and avoid the need for converting from one format to another in your own code. Invoking xx-info without any additional arguments will invoke xx-info triple.

Parsing current target

  • xx-info os - prints operating system component of TARGETPLATFORM (linux,darwin,windows,wasi)
  • xx-info arch - architecture component of TARGETPLATFORM
  • xx-info variant - variant component of TARGETPLATFORM if architecture is arm (eg. v7

Architecture formats

These commands return architecture names as used by specific tools to avoid conversion and tracking exceptions in your own code. E.g. arm64 repositories are called aarch64 in Alpine, but arm64 in Debian. uname -m returns aarch64 in Linux, but arm64 in Darwin etc.

  • xx-info march - Target machine architecture that is expected to match value of uname -m
  • xx-info alpine-arch - Target architecture for Alpine package repositories
  • xx-info debian-arch - Target architecture for Debian package repositories
  • xx-info rhel-arch - Target architecture for RPM package repositories
  • xx-info pkg-arch - Either alpine-arch, debian-arch or rhel-arch depending on the context

Target triple

Target triple is the target format taken as input in various gcc and llvm based compilers.

  • xx-info triple - Target triple in arch[-vendor]-os-abi form. This command is the default.
  • xx-info vendor - Vendor component of target triple
  • xx-info libc - Used libc (musl or gnu)

Build context

  • xx-info is-cross - Exit cleanly if target is not native architecture
  • xx-info env - Print XX_* variables defining target environment
$ xx-info env
XX_OS=linux
XX_MARCH=x86_64
XX_VENDOR=alpine
XX_PKG_ARCH=x86_64
XX_TRIPLE=x86_64-alpine-linux-musl
XX_LIBC=musl
TARGETOS=linux
TARGETARCH=amd64
TARGETVARIANT=

xx-apk, xx-apt, xx-apt-get - Installing packages for target architecture

These scripts allow managing packages (most commonly installing new packages) from either Alpine or Debian repositories. They can be invoked with any arguments regular apk or apt/apt-get commands accept. If cross-compiling for non-native architectures, the repositories for the target architecture are added automatically, and packages are installed from there. On Alpine, installing packages for a different architecture under the same root is not allowed, so xx-apk installs packages under a secondary root /${triple}. These scripts are meant for installing headers and libraries that compilers may need. To avoid unnecessary garbage, the non-native binaries under */bin are skipped on installation.

# alpine
ARG TARGETPLATFORM
RUN xx-apk add --no-cache musl-dev zlib-dev
# debian
ARG TARGETPLATFORM
RUN xx-apt-get install -y libc6-dev zlib1g-dev

Note

xx-apt --print-source-file can be used to print the path of the main Apt sources configuration file

Installing two meta-libraries, xx-c-essentials, xx-cxx-essentials is also allowed that expand the minimum necessary packages for either base image.

xx-verify - Verifying compilation results

xx-verify allows verifying that the cross-compile toolchain was correctly configured and outputted binaries for the expected target platform. xx-verify works by calling file utility and comparing the expected output. Optionally --static option can be passed to verify that the compiler produced a static binary that can be safely copied to another Dockerfile stage without runtime libraries. If the binary does not match the expected value, xx-verify returns with a non-zero exit code and error message.

ARG TARGETPLATFORM
RUN xx-clang --static -o /out/myapp app.c && \
    xx-verify --static /out/myapp

Note

XX_VERIFY_STATIC=1 environment variable can be defined to make xx-verify always verify that the compiler produced a static binary.

C/C++

The recommended method for C-based build is to use clang via xx-clang wrapper. Clang is natively a cross-compiler, but in order to use it, you also need a linker, compiler-rt or libgcc, and a C library(musl or glibc). All these are available as packages in Alpine and Debian based distros. Clang and linker are binaries and should be installed for your build architecture, while libgcc and C library should be installed for your target architecture.

The recommended linker is lld, but there are some caveats. lld is not supported on S390x, and based on our experience, sometimes has issues with preparing static binaries for Ppc64le. In these cases, ld from binutils is required. As separate ld binary needs to be built for each architecture, distros often do not provide it as a package. Therefore xx loads prebuilt ld binaries when needed. XX_CC_PREFER_LINKER=ld can be defined if you want to always use ld, even when lld is available on the system. Building MacOS binaries happens through a prebuilt ld64 linker that also adds ad-hoc code-signature to the resulting binary.

xx-clang can be called with any arguments clang binary accepts and will internally call the native clang binary with additional configuration for correct cross-compilation. On first invocation, xx-clang will also set up alias commands for the current target triple that can be later called directly. This helps with tooling that looks for programs with a target triple prefix from your PATH. This setup phase can be manually invoked by calling xx-clang --setup-target-triple that is a special flag that clang itself does not implement.

Alias commands include:

  • triple-clang, triple-clang++ if clang is installed
  • triple-ld if ld is used as linker
  • triple-pkg-config if pkg-config is installed
  • triple-addr2line, triple-ar, triple-as, triple-ranlib, triple-nm, triple-dlltool, triple-strip if cross-compilation capable tools are available though llvm package
  • triple-windres if llvm-rc is installed and compiling for Windows

Alias commands can be called directly and always build the configuration specified by their name, even if TARGETPLATFORM value has changed.

Building on Alpine

On Alpine, there is no special package for libgcc so you need to install gcc package with xx-apk even though the build happens through clang. To use compiler-rt instead of libgcc --rtlib needs to be passed manually. We will probably add default detection/loading for compiler-rt in the future to simplify this part. Default libc used in Alpine is Musl that can be installed with musl-dev package.

# ...
RUN apk add clang lld
# copy source
ARG TARGETPLATFORM
RUN xx-apk add gcc musl-dev
RUN xx-clang -o hello hello.c && \
    xx-verify hello

Clang binary can also be called directly with --target flag if you want to avoid xx- prefixes. --print-target-triple is a built-in flag in clang that can be used to query to correct default value.

# ...
RUN xx-apk add g++
RUN clang++ --target=$(xx-clang --print-target-triple) -o hello hello.cc

On the first invocation, aliases with triple- prefix are set up so the following also works:

# ...
RUN $(xx-clang --print-target-triple)-clang -o hello hello.c

If you prefer aliases to be created as a separate step on a separate layer, you can use --setup-target-triple.

# ...
RUN xx-clang --setup-target-triple
RUN $(xx-info)-clang -o hello hello.c

Building on Debian

Building on Debian/Ubuntu is very similar. The only required dependency that needs to be installed with xx-apt is libc6-dev or libstdc++-N-dev for C++.

# ...
RUN apt-get update && apt-get install -y clang lld
# copy source
ARG TARGETPLATFORM
RUN xx-apt install -y libc6-dev
RUN xx-clang -o hello hello.c

Refer to the previous section for other variants.

If you wish to build with GCC instead of Clang you need to install gcc and binutils packages additionally with xx-apt-get. xx-apt-get will automatically install the packages that generate binaries for the current target architecture. You can then call GCC directly with the correct target triple. Note that Debian currently only provides GCC cross-compilation packages if your native platform is amd64 or arm64.

# ...
# copy source
ARG TARGETPLATFORM
RUN xx-apt-get install -y binutils gcc libc6-dev
RUN $(xx-info)-gcc -o hello hello.c

Wrapping as default

Special flags xx-clang --wrap and xx-clang --unwrap can be used to override the default behavior of clang with xx-clang in the extreme cases where your build scripts have no way to point to alternative compiler names.

# export TARGETPLATFORM=linux/amd64
# xx-clang --print-target-triple
x86_64-alpine-linux-musl
# clang --print-target-triple
x86_64-alpine-linux-musl
# 
# xx-clang --wrap
# clang --print-target-triple
x86_64-alpine-linux-musl
# xx-clang --unwrap
# clang --print-target-triple
aarch64-alpine-linux-musl

Autotools

Autotools has built-in support for cross-compilation that works by passing --host, --build, and --target flags to the configure script. --host defines the target architecture of the build result, --build defines compilers native architecture(used for compiling helper tools etc.), and --target defines an architecture that the binary returns if it is running as a compiler of other binaries. Usually, only --host is needed.

# ...
ARG TARGETPLATFORM
RUN ./configure --host=$(xx-clang --print-target-triple) && make

If you need to pass --build, you can temporarily reset the TARGETPLATFORM variable to get the system value.

ARG TARGETPLATFORM
RUN ./configure --host=$(xx-clang --print-target-triple) --build=$(TARGETPLATFORM= xx-clang --print-target-triple) && make

Sometimes configure scripts misbehave and don't work correctly unless the name of the C compiler is passed directly. In these cases, you can use overrides like:

RUN CC=xx-clang ./configure ...
RUN ./configure --with-cc=xx-clang ...
RUN ./configure --with-cc=$(xx-clang --print-target-triple)-clang ...

CMake

In order to make cross-compiling with CMake easier, xx-clang has a special flag xx-clang --print-cmake-defines. Running that command returns the following Cmake definitions:

-DCMAKE_C_COMPILER=clang
-DCMAKE_CXX_COMPILER=clang++
-DCMAKE_ASM_COMPILER=clang
-DPKG_CONFIG_EXECUTABLE="$(xx-clang --print-prog-name=pkg-config)"
-DCMAKE_C_COMPILER_TARGET="$(xx-clang --print-target-triple)"
-DCMAKE_CXX_COMPILER_TARGET="$(xx-clang++ --print-target-triple)"
-DCMAKE_ASM_COMPILER_TARGET="$(xx-clang --print-target-triple)"

Usually, this should be enough to pick up the correct configuration.

RUN apk add cmake clang lld
ARG TARGETPLATFORM
RUN xx-apk musl-dev gcc
RUN mkdir build && cd build && \
    cmake $(xx-clang --print-cmake-defines) ..

Go / Cgo

Building Go can be achieved with the xx-go wrapper that automatically sets up values for GOOS, GOARCH, GOARM, GOAMD64 etc. It also sets up pkg-config and C compiler if building with CGo. Note that by default, CGo is enabled in Go when compiling for native architecture and disabled when cross-compiling. This can easily produce unexpected results; therefore, you should always define either CGO_ENABLED=1 or CGO_ENABLED=0 depending on if you expect your compilation to use CGo or not.

FROM --platform=$BUILDPLATFORM golang:alpine
# ...
ARG TARGETPLATFORM
ENV CGO_ENABLED=0
RUN xx-go build -o hello ./hello.go && \
    xx-verify hello
FROM --platform=$BUILDPLATFORM golang:alpine
RUN apk add clang lld
# ...
ARG TARGETPLATFORM
RUN xx-apk add musl-dev gcc
ENV CGO_ENABLED=1
RUN xx-go build -o hello ./hello.go && \
    xx-verify hello

If you want to make go compiler cross-compile by default, you can use xx-go --wrap and xx-go --unwrap

# ...
RUN xx-go --wrap
RUN go build -o hello hello.go && \
    xx-verify hello

Rust

Building Rust can be achieved with the xx-cargo wrapper that automatically sets up the target triple and also pkg-config and C compiler.

The wrapper supports rust installed via rustup (alpine/debian), distribution packages (alpine/debian) and the official rust image.

Building on Alpine

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM rust:alpine
RUN apk add clang lld
# ...
ARG TARGETPLATFORM
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

Cargo binary can also be called directly with --target flag if you don't want to use the wrapper. --print-target-triple is a built-in flag that can be used to set the correct target:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM rust:alpine
RUN apk add clang lld
# ...
ARG TARGETPLATFORM
RUN cargo build --target=$(xx-cargo --print-target-triple) --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

Note

xx-cargo --print-target-triple does not always have the same value as xx-clang --print-target-triple. This is because prebuilt Rust and C libraries sometimes use a different value.

The first invocation of xx-cargo will install the standard library for Rust matching the target if not already installed.

To fetch dependencies from crates.io you can use cargo fetch before building:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM rust:alpine
RUN apk add clang lld
# ...
RUN --mount=type=cache,target=/root/.cargo/git/db \
    --mount=type=cache,target=/root/.cargo/registry/cache \
    --mount=type=cache,target=/root/.cargo/registry/index \
    cargo fetch
ARG TARGETPLATFORM
RUN --mount=type=cache,target=/root/.cargo/git/db \
    --mount=type=cache,target=/root/.cargo/registry/cache \
    --mount=type=cache,target=/root/.cargo/registry/index \
    xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

Note

By calling cargo fetch before ARG TARGETPLATFORM your packages are fetched only once for the whole build while the building happens separately for each target architecture.

To avoid redownloading dependencies on every build, you can use cache mounts to store Git sources with packages and metadata of crate registries.

If you don't want to use the official Rust image, you can install rustup manually:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM alpine AS rustup
RUN apk add curl
RUN curl https://sh.rustup.rs -sSf | sh -s -- -y --default-toolchain stable --no-modify-path --profile minimal
ENV PATH="/root/.cargo/bin:$PATH"

FROM rustup
RUN apk add clang lld
# ...
ARG TARGETPLATFORM
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

If you install rust using distribution packages, rustup will not be available:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM alpine
RUN apk add clang lld rust cargo
# ...
ARG TARGETPLATFORM
RUN xx-apk add xx-c-essentials
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

In this case, you need to also install minimum necessary packages using xx-apk.

Building on Debian

Building on Debian/Ubuntu is very similar. If you are using rustup:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM rust:bookworm
RUN apt-get update && apt-get install -y clang lld
ARG TARGETPLATFORM
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo
# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM debian:bookworm AS rustup
RUN apt-get update && apt-get install -y curl ca-certificates
RUN curl https://sh.rustup.rs -sSf | sh -s -- -y --default-toolchain stable --no-modify-path --profile minimal
ENV PATH="/root/.cargo/bin:$PATH"

FROM rustup
RUN apt-get update && apt-get install -y clang lld
# ...
ARG TARGETPLATFORM
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

Or distribution packages:

# syntax=docker/dockerfile:1
FROM --platform=$BUILDPLATFORM debian:bookworm
RUN apt-get update && apt-get install -y clang lld cargo
ARG TARGETPLATFORM
RUN xx-apt-get install xx-c-essentials
RUN xx-cargo build --release --target-dir ./build && \
    xx-verify ./build/$(xx-cargo --print-target-triple)/release/hello_cargo

External SDK support

In addition to Linux targets, xx can also build binaries for MacOS and Windows. When building MacOS binaries from C, external MacOS SDK is needed in /xx-sdk directory. Such SDK can be built, for example, with gen_sdk_package script in osxcross project. Please consult XCode license terms when making such an image. RUN --mount syntax can be used in Dockerfile in order to avoid copying SDK files. No special tooling such as ld64 linker is required in the image itself.

Building Windows binaries from C/CGo is currently a work in progress and not functional.

# syntax=docker/dockerfile:1.2
# ...
RUN apk add clang lld
ARG TARGETPLATFORM
RUN --mount=from=my/sdk-image,target=/xx-sdk,src=/xx-sdk \
    xx-clang -o /hello hello.c && \
    xx-verify /hello

FROM scratch
COPY --from=build /hello /
docker buildx build --platform=darwin/amd64,darwin/arm64 -o bin .

-o/--output flag can be used to export binaries out from the builder without creating a container image.

Used by

These projects, as well as xx Dockerfile can be used for reference.

Issues

xx project welcomes contributions if you notice any issues or want to extend the capabilities with new features. We are also interested in cases where a popular project does not compile easily with xx so it can be improved, and tests can be added that try building these projects when xx gets updated. If you want to add support for a new architecture or language, please open an issue first to verify that the proposal matches the scope or xx.