March 25, 2022 (toolchain,docker)
In a previous post I used docker to containerize building for a RISC-V target.
In this post I’ll simulate that target file in a containerized RISC-V ISA Simulator.
Spike ISA Sim
The official reference ISA simulator for RISC-V is spike. There are other more functional and performant open source and commercial simulators , however the aim of this post is to describe a solution for quick and easy debugging and simulation of the low level examples on this blog.
Spike implements a functional model of the RISC-V hart(s) and the debug interface. GDB must connect to OpenOCD, which will in turn connect to Spike’s model of the RISC-V serial debug interface. Spike’s Readme file describes how to do this, but it’s pretty tedious to load an example program.
The container example downloads and builds the appropriate source code and manages running the independent tools.
The docker files are in https://github.com/five-embeddev/build-and-verify. Assume it’s checked out to
Firstly, the docker images in these folders should be built:
cd build-and-verify/ make -C docker/riscv-openocd make -C docker/riscv-spike make -C docker/riscv-xpack-gcc make -C examples/debug-sim
Running a RISC-V Target Elf File on Desktop OS Host
Let’s assume we’ve already compiled an elf file
The Docker image is tagged as
five_embeddev/examples/riscv_spike_debug_env_sim_gdb, built in the folder
examples/debug-sim. In the command line below, docker arguments are placed before the image name and spike simulator arguments are after the image name.
The docker arguments needs to configure the local
build directory to be mounted as a volume inside the container, In this case to
/project/build. It also needs to configure an interactive TTY mode for debugging.
The spike arguments needs to configure the ISA, memory map and target image.
docker \ run \ -i --rm --tty \ -v `pwd`:/project \ five_embeddev/examples/riscv_spike_debug_env_sim_gdb \ --priv=m \ --isa=rv32imac \ -m0x8000000:0x2000,0x80000000:0x4000,0x20010000:0x6a120 \ /project/build/main.elf
It’s pretty easy to step through the code.
tui enable b main c next next
compose.yml file captures this configuration the mapping of local files to
/project in the container, and the simulator arguments.
services: spike_openocd_gdb: image: five_embeddev/examples/riscv_spike_debug_env_sim_gdb command: [ "--priv=m" ,"--isa=rv32imac", "-m0x8000000:0x2000,0x80000000:0x4000,0x20010000:0x6a120", "/project/build/main.elf"] tty: true volumes: - .:/project
docker-compose run spike_openocd_gdb
I’d like to customize spike for development flow and architecture experimentation, so having an automated from-source to single run command is convenient.
A few general advantages are:
- Containerized tools can be used to integrate with modern development environments, like VSCode.
- The latest tools can be downloaded and utilized using automated scripts.
- This can be built on for automated testing and cloud based CI.
March 01, 2022 (toolchain,docker)
Modern development is moving towards packaging tools in containers. There are several benefits to containerizing development tools:
- A simpler tool deployment process, the binary image with all dependencies can simply be pulled from a server and run.
- An automated and consistent deployment process that can be shared between development machines and cloud based build servers.
- A consistent set of tools for all team members.
- The ability to build with General Purpose IDE’s rather than whatever tools a device vendor may provide.
- Isolation of dependencies used for different devices and build targets.
For RISC-V there are a few more benefits:
- A “recipe” to consistently build the vendor neutral tools at, such as https://github.com/riscv-collab/riscv-gnu-toolchain.
- New versions of the tools with support for new extensions etc can be easily deployed.
I’ve put together a set of Docker images and Docker Compose build and run configurations with the aim of deploying them with GitHub’s workflows.
They are located here: https://github.com/five-embeddev/build-and-verify
Each image creates a default
/project that is
owned by the docker user. The local development directory can be bind mounted to that path.
The path to the tools is always added to the
PATH environment variable.
The container runs everything as user
docker_user with UID of 1000 by default.
To build a target binary call
docker directly or setup a
compose.yaml with the container
environment. The command line to be run in the container is passed as the arguments following the container
run command and it’s arguments.
test.c in the current directory with the
and delete the container after running:
docker \ run \ --rm \ -v `pwd`:/project \ five_embeddev/riscv_gnu_toolchain_dev_env \ riscv32-unknown-elf-gcc test.c -o test.elf
docker-compose with all of that captured in a
version: "3" services: build_gnu_toolchain: image: five_embeddev/riscv_gnu_toolchain_dev_env volumes: - .:/project command: ["riscv32-unknown-elf-gcc", "test.c", "-o", "test.elf" ]
docker-compose can simply run the build.
docker-compose run build_gnu_toolchain
The images have been setup with a docker user (by default UID 1000),
and the tools are installed to
/opt or to the docker user’s home
directory and added to the
PATH environment variable. There are
Makefiles to build the images, if those are used the current user
UID is used by the docker image default user. That allows the user’s
file system to be mounted within the container and file ownership
handled correctly. (By default docker containers runs as
can result in root owned target binaries.)
When tools are compiled from source the default method is to shallow
clone the source code within the container and build there. This
captures the build process in the
Dockerfile, but is inefficient due
to the docker caching mechanism. The
script does this but requires a large amount of resources.
Due to the resource requirements of building the entire toolchain the
riscv-gnu-toolchain-2 build script stores the source and build files
in host file system and mounts into the container via docker
volumes. Those can’t be used by a docker build process, so the build
process is more complex and relies on the
Makefile. This build
script should be preferred.
November 05, 2021 (toolchain,baremetal,C)
Information about extending cross compilation with PlatformIO has been added.
September 10, 2021 (registers,spec,interrupts,opcodes)
As RISC-V is a new architecture so there will be new development at all layers of the software and hardware stack. Rather than write code based on human language specifications from scratch, a smarter way to work can be to translate a machine readable specification to code.
“Machine Readable” does not need to be an all encompassing formal model of the architecture, there are many convenient formats such as csv, yaml, xml and json that can be parsed and transformed using the packages available in most scripting languages.
March 20, 2021 (toolchain,baremetal)
An example of cross compiling a baremetal program to RISC-V with CMake.
February 09, 2021 (gcc,base_isa,extensions,abi)
Note to self: When compiling the riscv-toolchain for embedded systems, set the configure options!
The toolchain can be cloned from the RISC-V official github. Once the dependencies are installed it’s straight forward to compile.
October 29, 2019 (toolchain,quickref)
An initial toolchain quick reference.
June 26, 2019 (gcc,base_isa,extensions,abi)
NOTE: Since this was written the riscv-toolchain-conventions document has been released.
Getting started with RISC-V. Compiling for the RISC-V target. This post covers the GCC machine architecture (-march), ABI (-mabi) options and how they relate to the RISC-V base ISA and extensions. It also looks at the multilib configuration for GCC.