This layer provides clang/llvm as alternative to system C/C++ compiler for OpenEmbedded/Yocto Project based distributions. This can cohabit with GNU compiler and can be used for specific recipes or full system compiler.
git clone https://github.com/openembedded/openembedded-core.git
cd openembedded-core
git clone https://github.com/openembedded/bitbake.git
git clone https://github.com/kraj/meta-clang.git
$ . ./oe-init-build-env
If using poky ( Yocto project reference Distribution )
git clone https://git.yoctoproject.org/git/poky
cd poky
git clone https://github.com/kraj/meta-clang.git
$ . ./oe-init-build-env
Add meta-clang overlay
bitbake-layers add-layer ../meta-clang
Check conf/bblayers.conf
to see that meta-clang is added to layer mix e.g.
Note that by default gcc will remain the system compiler, however if you wish clang to be the default compiler then set
TOOLCHAIN ?= "clang"
in local.conf
, this would now switch default cross-compiler to be clang
you can select clang per recipe too by writing bbappends for them containing
TOOLCHAIN = "clang"
also look at conf/nonclangable.conf
for list of recipes which do not yet fully
build with clang.
clang recipes can provide llvm and related packages too, it might be worth using single provider for llvm and clang to save some compile time and space, select the knobs to point to clang, default is to use the version provided by core layer.
PREFERRED_PROVIDER_llvm = "clang"
PREFERRED_PROVIDER_llvm-native = "clang-native"
PREFERRED_PROVIDER_nativesdk-llvm = "nativesdk-clang"
PROVIDES:pn-clang = "llvm"
PROVIDES:pn-clang-native = "llvm-native"
PROVIDES:pn-nativesdk-clang = "nativesdk-llvm"
Default is to use GNU runtime TC_CXX_RUNTIME = "gnu"
which consists of libgcc, libstdc++ to provide C/C++
runtime support. However it's possible to use LLVM runtime to replace it where
compile-rt, llvm libunwind, and libc++ are used to provide C/C++ runtime, while
GNU runtime works with both GCC and Clang, LLVM runtime is only tested with Clang
compiler, switching to use LLVM runtime is done via a config metadata knob
TC_CXX_RUNTIME = "llvm"
TC_CXX_RUNTIME variable influences individual runtime elements and can be set explicitly as well
e.g. LIBCPLUSPLUS
COMPILER_RT
and UNWINDLIB
.
Please note that this will still use crt files from GNU compiler always, while llvm now do provide crt files, they have not been yet integrated into the toolchain.
Using TC_CXX_RUNTIME variable will select which C++ runtime is used, however it can be overridden
if needed to by modifying LIBCPLUSPLUS
variable, usually defaults used by TC_CXX_RUNTIME
are
best fit. e.g. below we select LLVM C++ as default C++ runtime.
LIBCPLUSPLUS = "-stdlib=libc++"
in local.conf
.
You can select libstdc++ per package too by writing bbappends for them containing
LIBCPLUSPLUS:toolchain-clang:pn-<recipe> = "-stdlibc=libc++"
Defaults are chosen to be GNU for maximum compatibility with existing GNU systems. It's always good to use single runtime on a system, mixing runtimes can cause complications during compilation as well as runtime. However, it's up to distribution policies to decide which runtime to use.
Clang based cross compiler is not included into the generated SDK using bitbake meta-toolchain
or
bitbake -cpopulate_sdk <image>
if clang is expected to be part of SDK, add CLANGSDK = "1"
in local.conf
CLANGSDK = "1"
Newer kernels and Android kernels support clang compilation, and even support LTO, The following takes rockchip's kernel as an example to configure clang compilation. x86, arm and arm64 kernel supported full LLVM toolchain, other arch only support clang. more info https://docs.kernel.org/kbuild/llvm.html
- linux-rockchip_%bbappend
TOOLCHAIN:forcevariable = "clang"
DEPENDS:append:toolchain-clang = " clang-cross-${TARGET_ARCH}"
KERNEL_CC:toolchain-clang = "${CCACHE}clang ${HOST_CC_KERNEL_ARCH} -fuse-ld=lld ${DEBUG_PREFIX_MAP} -fdebug-prefix-map=${STAGING_KERNEL_DIR}=${KERNEL_SRC_PATH}"
KERNEL_LD:toolchain-clang = "${CCACHE}ld.lld"
KERNEL_AR:toolchain-clang = "${CCACHE}llvm-ar"
if you want use LLVM integrated assembler for some older kernel, newer vesion is enabled default.
do_compile:prepend:toolchain-clang() {
export LLVM_IAS=1
}
do_compile_kernelmodules:prepend:toolchain-clang() {
export LLVM_IAS=1
}
if you want enable LTO, append follow content.
FILESEXTRAPATHS:prepend := "${THISDIR}/${PN}:"
SRC_URI:append:toolchain-clang = "\
file://lto.cfg \
"
CONFIG_LTO_CLANG
is need for some android based kernel, mainline kernel will do auto detect.
- lto.cfg
CONFIG_LTO_CLANG=y
CONFIG_LTO=y
CONFIG_LTO_CLANG_THIN=y
Below we build for qemuarm machine as an example
$ MACHINE=qemuarm bitbake core-image-full-cmdline
$ runqemu nographic
Few components do not build with clang, if you have a component to add to that list
simply add it to conf/nonclangable.inc
e.g.
TOOLCHAIN:pn-<recipe> = "gcc"
and OE will start using gcc to cross compile that recipe.
If a component does not build with libc++, you can add it to conf/nonclangable.inc
e.g.
CXX:remove:pn-<recipe>:toolchain-clang = " -stdlib=libc++ "
If your DISTRO sets its own value of TARGET_VENDOR, then it needs to be added in CLANG_EXTRA_OE_VENDORS, it should be done automatically, but if compiler-rt fails like bellow, then check the end of work-shared/llvm-project-source-12.0.0-r0/temp/log.do_patch it should have a line like: NOTE: Adding support following TARGET_VENDOR values: foo in /OE/build/oe-core/tmp-glibc/work-shared/llvm-project-source-12.0.0-r0/git/llvm/lib/Support/Triple.cpp and /OE/build/oe-core/tmp-glibc/work-shared/llvm-project-source-12.0.0-r0/git/clang/lib/Driver/ToolChains/Gnu.cpp and check these files if //CLANG_EXTRA_OE_VENDORS* strings were replaced correctly. Read add_distro_vendor function in recipes-devtools/clang/llvm-project-source.inc for more details.
http://errors.yoctoproject.org/Errors/Details/574365/
-- Found assembler: TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/recipe-sysroot-native/usr/bin/x86_64-foo-linux/x86_64-foo-linux-clang
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - failed
-- Check for working C compiler: TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/recipe-sysroot-native/usr/bin/x86_64-foo-linux/x86_64-foo-linux-clang
-- Check for working C compiler: TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/recipe-sysroot-native/usr/bin/x86_64-foo-linux/x86_64-foo-linux-clang - broken
CMake Error at TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/recipe-sysroot-native/usr/share/cmake-3.19/Modules/CMakeTestCCompiler.cmake:66 (message):
The C compiler
"TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/recipe-sysroot-native/usr/bin/x86_64-foo-linux/x86_64-foo-linux-clang"
is not able to compile a simple test program.
It fails with the following output:
Change Dir: TOPDIR/tmp-glibc/work/core2-64-foo-linux/compiler-rt/12.0.0-r0/build/CMakeFiles/CMakeTmp
Run Build Command(s):ninja cmTC_928f4 && [1/2] Building C object CMakeFiles/cmTC_928f4.dir/testCCompiler.c.o
[2/2] Linking C executable cmTC_928f4
URI: https://github.com/openembedded/openembedded-core.git
branch: master
revision: HEAD
URI: ghttps://github.com/openembedded/bitbake.git
branch: master
revision: HEAD
All LLVM based recipes use single llvm source directory, As a LLVM
developer, you might want to work on your own repository to build
related recipes, devtool can we useful in establishing such a workflow
there is a script provided scripts/devtool-clang.sh
which can assist
in setting up all the recipes to use custom LLVM repository, in order
to setup repository make sure that it has all the needed patches applied
or else it will fail to build. Such a tree is already prepared and kept
in sync at
https://github.com/kraj/llvm-project
There are branches under oe
namespace which carry the needed OE patches
cd $TOPDIR/workspace/sources
git clone https://github.com/kraj/llvm-project -b oe/main llvm-project
Once project is setup and meta-clang is added, run devtool-clang.sh
script which will do the needed for setting up external sources for the
yocto recipes, now yocto will use custom LLVM tree for its needs.
You are encouraged to follow Github Pull request workflow to share changes and following commit message guidelines are recommended: OE patch guidelines.
Layer Maintainer: Khem Raj