Technical Documentation: riscv-collab/riscv-gnu-toolchain
โน๏ธ Provenance: Multi-Tier Ingestion ยท CodeWiki Reference
Overview (`README.md`)
RISC-V GNU Compiler Toolchain
This is the RISC-V C and C++ cross-compiler. It can build a bare-metal
ELF toolchain (using Newlib or Picolibc) or a Linux-ELF toolchain
(using glibc, musl, or uClibc).
Getting the sources
This repository uses submodules, but submodules will fetch automatically on demand,
so --recursive or git submodule update --init --recursive is not needed.
$ git clone https://github.com/riscv/riscv-gnu-toolchainWarning: git clone takes around 6.65 GB of disk and download size
Prerequisites
Several standard packages are needed to build the toolchain.
On Ubuntu, executing the following command should suffice:
$ sudo apt-get install autoconf automake autotools-dev curl python3 python3-pip python3-tomli libmpc-dev libmpfr-dev libgmp-dev gawk build-essential bison flex texinfo gperf libtool patchutils bc zlib1g-dev libexpat-dev meson ninja-build git cmake libglib2.0-dev expect device-tree-compiler libslirp-dev libzstd-dev libncurses-devOn Fedora/CentOS/RHEL OS, executing the following command should suffice:
$ sudo yum install autoconf automake python3 libmpc-devel mpfr-devel gmp-devel gawk bison flex texinfo patchutils gcc gcc-c++ zlib-devel expat-devel libslirp-devel ncurses-devel meson ninja-build cmake dtcOn Arch Linux, executing the following command should suffice:
$ sudo pacman -Syu curl python3 libmpc mpfr gmp base-devel texinfo gperf patchutils bc zlib expat libslirp ncurses meson ninja cmakeAlso available for Arch users on the AUR: https://aur.archlinux.org/packages/riscv-gnu-toolchain-bin
On macOS, you can use Homebrew to install the dependencies:
$ brew install python3 gawk gnu-sed make gmp mpfr libmpc isl zlib expat texinfo flock libslirp ncurses meson ninja cmake bison m4 wgetWhen executing the instructions in this README, please use gmake instead of make to use the newly installed version of make.
To build the glibc (Linux) on macOS, you will need to build within a case-sensitive file
system. The simplest approach is to create and mount a new disk image with
a case sensitive format. Make sure that the mount point does not contain spaces. This is not necessary to build newlib or gcc itself on macOS.
This process will start by downloading about 200 MiB of upstream sources, then
will patch, build, and install the toolchain. If a local cache of the
upstream sources exists in $(DISTDIR), it will be used; the default location
is /var/cache/distfiles. Your computer will need about 8 GiB of disk space to
complete the process.
Installation
Pick an install path that is writeable โ say /opt/riscv โ and add/opt/riscv/bin to your PATH. Then configure:
./configure --prefix=/opt/riscvand build the C library you want:
| C library | build |
|---|---|
| Newlib | make |
| Picolibc | make picolibc |
| Linux (glibc) | make linux |
| Linux (musl) | make musl |
| Linux (uClibc) | make uclibc |
You should now be able to use, e.g., riscv64-unknown-elf-gcc and its
cousins. The bare-metal (Newlib, Picolibc) tools are prefixedriscv64-unknown-elf-, the Linux tools riscv64-unknown-linux-{gnu,musl,uclibc}-.
A plain make builds the default target, which is Newlib. To make a
different libc the default โ so that make on its own builds it โ configure
with --enable-newlib, --enable-linux, --enable-musl, --enable-uclibc
or --enable-picolibc.
The build defaults to targeting RV64GC (64-bit), even on a 32-bit build
environment. To build a 32-bit toolchain, pass the arch and ABI, e.g.:
./configure --prefix=/opt/riscv --with-arch=rv32gc --with-abi=ilp32dSupported architectures are rv32i or rv64i plus standard extensions (a)tomics,
(m)ultiplication and division, (f)loat, (d)ouble, or (g)eneral for MAFD.
Supported ABIs are ilp32 (32-bit soft-float), ilp32d (32-bit hard-float),
ilp32f (32-bit with single-precision in registers and double in memory, niche
use only), lp64 lp64f lp64d (same but with 64-bit long and pointers).
Note: If you're planning to use an external library that replaces part of newlib (for example libgloss-htif), read the FAQ.
Multilib
Add --enable-multilib to build runtime libraries for both 32-bit and
64-bit:
./configure --prefix=/opt/riscv --enable-multilibThe resulting compiler can target both 32-bit and 64-bit systems and
supports the most common -march/-mabi options, which can be seen with
the --print-multi-lib flag.
Multilib is only available for the Newlib and Linux/glibc toolchains; the
musl, uClibc and Picolibc toolchains build a single variant, and configuring
them with --enable-multilib is rejected. Building them explicitly (e.g.make musl) in a tree configured for multilib still yields a single
variant.
Linux toolchain has an additional option --enable-default-pie to control the
default PIE enablement for GCC, which is disable by default.
To customize the enabled languages, use option --with-languages=. For example,
if you want to enable c,c++,fortran, use ./configure --with-languages=c,c++,fortran.
This option only takes effect for the GNU toolchain.
The toolchain has an option --enable-strip to control strip of host binaries,
strip is disabled by default.
Installation (MacOS ARM)
First, ensure you have cloned the toolchain repository in a case-sensitive volume.
Now source macos.zsh to setup the PATH variable so that the build scripts can use the tools from homebrew, which are needed to build the GNU toolchain.
Then, run configure with your desired flags - For example:
./configure --prefix=/Volumes/case-sensitive/opt/riscv --with-arch=rv64gc_zifencei --with-abi=lp64d --enable-linux --disable-gdbThen, raise the limit of open files. Run: ulimit -n 65536
Builds on MacOS are highly specific to OS versions and the versions of the developer tools installed. We recommend running make check-binutils first, which will help surface a some of the more frequent build errors we've seen without having to start a full build.
If make check-binutils errors, check the following documentation for a list of common errors when building on MacOS and their solutions.
When make check-binutils finishes successfully, you run the build normally with make or make linux.
Big Endian configuration
To build either cross-compiler with support for big endian targets, run the following command:
./configure --prefix=/opt/riscv --with-endian=big
make (target)Troubleshooting Build Problems
Builds work best if installing into an empty directory. If you build a
hard-float toolchain and then try to build a soft-float toolchain with
the same --prefix directory, then the build scripts may get confused
and exit with a linker error complaining that hard float code can't be
linked with soft float code. Removing the existing toolchain first, or
using a different prefix for the second build, avoids the problem. It
is OK to build one newlib and one linux toolchain with the same prefix.
But you should avoid building two newlib or two linux toolchains with
the same prefix.
If building a linux toolchain on a MacOS system, or on a Windows system
using the Linux subsystem or cygwin, you must ensure that the filesystem
is case-sensitive. A build on a case-insensitive filesystem will fail when
building glibc because *.os and *.oS files will clobber each other during
the build eventually resulting in confusing link errors.
CentOS (and RHEL) provide old GNU tools versions that may be too old to build
a RISC-V toolchain. There is an alternate toolset provided that includes
current versions of the GNU tools. This is the devtoolset provided as part
of the Software Collection service. For more info, see the
devtoolset-7
URL. There are various versions of the devtoolset that are available, so you
can also try other versions of it, but we have at least one report that
devtoolset-7 works.
Advanced Options
There are a number of additional options that may be passed to
configure. See './configure --help' for more details.
Also you can define extra flags to pass to specific projects: BINUTILS_NATIVE_FLAGS_EXTRA, BINUTILS_TARGET_FLAGS_EXTRA, GCC_EXTRA_CONFIGURE_FLAGS, GDB_NATIVE_FLAGS_EXTRA, GDB_TARGET_FLAGS_EXTRA, GLIBC_TARGET_FLAGS_EXTRA, NEWLIB_TARGET_FLAGS_EXTRA, LLVM_EXTRA_CONFIGURE_FLAGS, LLVM_OPENMP_EXTRA_CONFIGURE_FLAGS, QEMU_EXTRA_CONFIGURE_FLAGS.
Example: GCC_EXTRA_CONFIGURE_FLAGS=--with-gmp=/opt/gmp make linux
Set default ISA spec version
--with-isa-spec= can specify the default version of the RISC-V Unprivileged
(formerly User-Level) ISA specification.
Possible options are: 2.2, 20190608 and 20191213.
The default version is 20191213.
More details about this option you can refer this post RISC-V GNU toolchain bumping default ISA spec to 20191213.
Build with customized multi-lib configure.
--with-multilib-generator= can specify what multilibs to build. The argument
is a semicolon separated list of values, possibly consisting of a single value.
Currently only supported for riscv*--elf. The accepted values and meanings
are given below.
Every config is constructed with four components: architecture string, ABI,
reuse rule with architecture string and reuse rule with sub-extension.
Re-use part support expansion operator (*) to simplify the combination of
different sub-extensions, example 4 demonstrate how it uses and works.
Example 1: Add multi-lib support for rv32i with ilp32.
./configure --with-multilib-generator="rv32i-ilp32--"Example 2: Add multi-lib support for rv32i with ilp32 and rv32imafd with ilp32.
./configure --with-multilib-generator="rv32i-ilp32--;rv32imafd-ilp32--"Example 3: Add multi-lib support for rv32i with ilp32; rv32im with ilp32 and
rv32ic with ilp32 will reuse this multi-lib set.
./configure --with-multilib-generator="rv32i-ilp32-rv32im-c"Example 4: Add multi-lib support for rv64ima with lp64; rv64imaf with lp64,
rv64imac with lp64 and rv64imafc with lp64 will reuse this multi-lib set.
./configure --with-multilib-generator="rv64ima-lp64--f*c"Enabling QEMU System Targets
The --enable-qemu-system configuration flag allows you to include QEMU system emulation targets in addition to the default user-mode emulation.
Enabled targets:
riscv64-linux-userriscv32-linux-userriscv64-softmmuriscv32-softmmu
Default targets (without this flag):
riscv64-linux-userriscv32-linux-user
Use this option if you need full system emulation for RISC-V. Example configuration:
./configure --enable-qemu-system --prefix=/opt/riscv
make build-sim SIM=qemuThis flag is particularly useful for developers testing and emulating full RISC-V systems rather than just user-space applications.
Test Suite
The Dejagnu test suite has been ported to RISC-V. This can be run with a
simulator for the elf and linux toolchains. The simulator can be selected
by the SIM variable in the Makefile, e.g. SIM=qemu, SIM=gdb, or SIM=spike
(experimental).In addition, the simulator can also be selected with the
configure time option --with-sim=.However, the testsuite allowlist is
only maintained for qemu.Other simulators might get extra failures.
The make check and make report targets are wired only for the Newlib
and Linux/glibc toolchains.
Additional Prerequisite
A helper script to setup testing environment requires
pyelftools.
On newer versions of Ubuntu, executing the following command
should suffice:
$ sudo apt-get install python3-pyelftoolsOn newer versions of Fedora and CentOS/RHEL OS (9 or later), executing
the following command should suffice:
$ sudo yum install python3-pyelftoolsOn Arch Linux, executing the following command should suffice:
$ sudo pacman -Syyu python-pyelftools python-sphinx python-sphinx_rtd_theme ninjaIf your distribution/OS does not have pyelftools package, you can install
it using PIP.
# Assuming that PIP is installed
$ pip3 install --user pyelftoolsTesting GCC
To test GCC, run the following commands:
./configure --prefix=$RISCV --disable-linux --with-arch=rv64ima # or --with-arch=rv32ima
make newlib
make report-newlib SIM=gdb # Run with gdb simulator
./configure --prefix=$RISCV
make linux
make report-linux SIM=qemu # Run with qemu
./configure --prefix=$RISCV --with-sim=spike
make linux
make report # Run with spikeNote:
- spike only support rv64* bare-metal/elf toolchain.
- gdb simulator only support bare-metal/elf toolchain.
Selecting the tests to run in GCC's regression test suite
By default GCC will execute all tests of its regression test suite.
While running them in parallel (e.g. make -j$(nproc) report) will
significantly speed up the execution time on multi-processor systems,
the required time for executing all tests is usually too high for
typical development cycles. Therefore GCC allows to select the tests
that are being executed using the environment variable RUNTESTFLAGS.
To restrict a test run to only RISC-V specific tests
the following command can be used:
RUNTESTFLAGS="riscv.exp" make reportTo restrict a test run to only RISC-V specific tests with match the
pattern "zb*.c" and "sm*.c" the following command can be used:
RUNTESTFLAGS="riscv.exp=zb*.c\ sm*.c" make reportTesting GCC, Binutils, and glibc of a Linux toolchain
The default Makefile target to run toolchain tests is report.
This will run all tests of the GCC regression test suite.
Alternatively, the following command can be used to do the same:
make check-gccThe following command can be used to run the Binutils tests:
make check-binutilsThe command below can be used to run the glibc tests:
make check-glibc-linuxAdding more arch/abi combination for testing without introducing multilib
--with-extra-multilib-test can be used when you want to test more combination
of arch/ABI, for example: built a linux toolchain with multilib withrv64gc/lp64d and rv64imac/lp64, but you want to test more configuration likerv64gcv/lp64d or rv64gcv_zba/lp64d, then you can use --with-extra-multilib-test
to specify that via --with-extra-multilib-test="rv64gcv-lp64d;rv64gcv_zba-lp64d",
then the testing will run for rv64gc/lp64d, rv64imac/lp64, rv64gcv/lp64d
and rv64gcv_zba/lp64d.
--with-extra-multilib-test support bare-metal and linux toolchain and support
even multilib is disable, but the user must ensure extra multilib test
configuration can be work with existing lib/multilib, e.g. rv32gcv/ilp32 test
can't work if multilib didn't have any rv32 multilib.
--with-extra-multilib-test also support more complicated format to fit the
requirements of end-users. First of all, the argument is a list of test
configurations. Each test configuration are separated by ;. For example:
rv64gcv-lp64d;rv64_zvl256b_zvfh-lp64d
For each test configuration, it has two parts, aka required arch-abi part and
optional build flags. We leverage : to separate them with some restrictions.
- arch-abi should be required and there must be only one at the begining of
the test configuration. - build flags is a array-like flags after the arch-abi, there will be two
ways to arrange them, aka AND, OR operation. - If you would like the flags in build flags array acts on arch-abi
simultaneously, you can use:to separate them. For example:
rv64gcv-lp64d:--param=riscv-autovec-lmul=dynamic:--param=riscv-autovec-preference=fixed-vlmaxwill be consider as one target board same as below:
riscv-sim/-march=rv64gcv/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-lmul=dynamic/--param=riscv-autovec-preference=fixed-vlmax- If you would like the flags in build flags array acts on arch-abi
respectively, you can use ',' to separate them. For example:
rv64gcv-lp64d:--param=riscv-autovec-lmul=dynamic,--param=riscv-autovec-preference=fixed-vlmaxwill be consider as two target boards same as below:
riscv-sim/-march=rv64gcv/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-preference=fixed-vlmax
riscv-sim/-march=rv64gcv/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-lmul=dynamic- However, you can also leverage AND(
:), OR(,) operator together but the
OR(,) will always have the higher priority. For example:
rv64gcv-lp64d:--param=riscv-autovec-lmul=dynamic:--param=riscv-autovec-preference=fixed-vlmax,--param=riscv-autovec-lmul=m2will be consider as tow target boars same as below:
riscv-sim/-march=rv64gcv/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-lmul=dynamic/--param=riscv-autovec-preference=fixed-vlmax
riscv-sim/-march=rv64gcv/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-lmul=m2LLVM / clang
LLVM can be used in combination with the RISC-V GNU Compiler Toolchain
to build RISC-V applications. To build LLVM with C and C++ support the
configure flag --enable-llvm can be used.
E.g. to build LLVM on top of a RV64 Linux toolchain the following commands
can be used:
./configure --prefix=$RISCV --enable-llvm --enable-linux
makeNote, that a combination of --enable-llvm and multilib configuration flags
is not supported.
The LLVM builds use the Ninja CMake generator by default. SetLLVM_GENERATOR="Unix Makefiles" to use make instead:
make -j$(nproc) LLVM_GENERATOR="Unix Makefiles" allBelow are examples how to build a rv64gc Linux/newlib toolchain with LLVM support,
how to use it to build a C and a C++ application using clang, and how to
execute the generated binaries using QEMU.
Build Linux toolchain and run examples:
# Build rv64gc toolchain with LLVM
./configure --prefix=$RISCV --enable-llvm --enable-linux --with-arch=rv64gc --with-abi=lp64d
make -j$(nproc) all build-sim SIM=qemu
# Build C application with clang
$RISCV/bin/clang -march=rv64imafdc -o hello_world hello_world.c
$RISCV/bin/qemu-riscv64 -L $RISCV/sysroot ./hello_world
# Build C++ application with clang
$RISCV/bin/clang++ -march=rv64imafdc -stdlib=libc++ -o hello_world_cpp hello_world_cpp.cxx
$RISCV/bin/qemu-riscv64 -L $RISCV/sysroot ./hello_world_cppBuild newlib toolchain and run examples (don't work with --with-multilib-generator=):
# Build rv64gc bare-metal toolchain with LLVM
./configure --prefix=$RISCV --enable-llvm --disable-linux --with-arch=rv64gc --with-abi=lp64d
make -j$(nproc) all build-sim SIM=qemu
# Build C application with clang
$RISCV/bin/clang -march=rv64imafdc -o hello_world hello_world.c
$RISCV/bin/qemu-riscv64 -L $RISCV/sysroot ./hello_world
# Build C++ application with clang using static link
$RISCV/bin/clang++ -march=rv64imafdc -static -o hello_world_cpp hello_world_cpp.cxx
$RISCV/bin/qemu-riscv64 -L $RISCV/sysroot ./hello_world_cppDevelopment
This section is only for developer or advanced user, or you want to build
toolchain with your own source tree.
Update Source Tree
riscv-gnu-toolchain contain stable but not latest source for each submodule,
in case you want to using latest development tree, you can use following command
to upgrade all submodule.
git submodule update --remoteOr you can upgrade specific submodule only.
git submodule update --remote <component>For example, upgrade gcc only, you can using following command:
git submodule update --remote gccHow to Check Which Branch are Used for Specific submodule
The branch info has recorded in .gitmodules file, which can set or update viagit submodule add -b or git submodule set-branch.
However the only way to check which branch are using is to check .gitmodules
file, here is the example for gcc, it's using releases/gcc-12 branch, so
it will has a section named gcc and has a field branch isreleases/gcc-12.
[submodule "gcc"]
path = gcc
url = ../gcc.git
branch = releases/gcc-12Use Source Tree Other Than `riscv-gnu-toolchain`
riscv-gnu-toolchain also supports using out-of-tree source to build the toolchain.
There are several configure options for specifying the source tree of each
submodule/component.
For example, if you have GCC sources in $HOME/gcc, use --with-gcc-src to build the toolchain using those sources:
./configure ... --with-gcc-src=$HOME/gccHere is the list of configure options for specifying alternative sources for the various submodules/components:
--with-binutils-src
--with-dejagnu-src
--with-gcc-src
--with-gdb-src
--with-glibc-src
--with-linux-headers-src
--with-llvm-src
--with-musl-src
--with-newlib-src
--with-picolibc-src
--with-pk-src
--with-qemu-src
--with-spike-src
--with-uclibc-srcBuild host GCC to check for compiler warnings
GCC contributions have to meet several requirements to qualify for upstream
inclusion. Warning free compilation with a compiler build from the same
sources is among them. The flag --enable-host-gcc does exaclty that:
- Initially a host GCC will be built
- This host GCC is then used to build the cross compiler
- The cross compiler will be built with
-Werrorto identify code issues
FAQ
Ensuring Code Model Consistency
If parts of newlib are going to be replaced with an external library (such as with libgloss-htif for Berkeley Host-Target Interface),
you should take care to ensure that both newlib and the external library are built using the same code model. For more information about RISC-V code models,
read this SiFive blog article.
Errors that indicate a code model mismatch include "relocation overflow" or "relocation truncated" errors from the linker being unable to successfully relocate symbols in the executable.
By default, riscv-gnu-toolchain builds newlib with -mcmodel=medlow. You can use the alternative medany code model (as used in libgloss-htif) by passing --with-cmodel=medany to the configure script.
Overview (`test/allowlist/README.md`)
This folder contain all allowlist files for testsuite result,
it used for riscv-gnu-toolchain/scripts/testsuite-filter,
naming rule of allowlist file as below:
<toolname>/common.log
<toolname>/[<lib>.][rv(32|64|128).][<ext>.][<abi>.]logtoolnamecan begcc,binutilsorgdb.<toolname>/common.log: Every target/library combination for the<toolname>
will use this allowlist file.<toolname>/[<lib>.][rv(32|64|128).][<ext>.][<abi>.]log:testsuite-filter
will according the target/library combination to match corresponding allowlist
files.For example, rv32im,ilp32/newlib will match following 24 files, and ignored if
file not exist:- common.log
- newlib.log
- rv32.log
- ilp32.log
- rv32.ilp32.log
- newlib.rv32.log
- newlib.ilp32.log
- newlib.rv32.ilp32.log
- i.log
- rv32.i.log
- i.ilp32.log
- rv32.i.ilp32.log
- newlib.i.log
- newlib.rv32.i.log
- newlib.i.ilp32.log
- newlib.rv32.i.ilp32.log
- m.log
- rv32.m.log
- m.ilp32.log
- rv32.m.ilp32.log
- newlib.m.log
- newlib.rv32.m.log
- newlib.m.ilp32.log
- newlib.rv32.m.ilp32.log
--- METRICS ---
- Files Extracted: 2
- Estimated Token Budget: ~5990 tokens
- Canonical Reference: https://codewiki.google/github.com/riscv-collab/riscv-gnu-toolchain