Tools we use: installing GDB for ARM

In this mini article, I’ll be going on a few strategies and nuances around getting a copy of GDB installed for debugging ARM chips.


GDB, the GNU Project Debugger, is by default typically compiled to target the same architecture as the host system. When we talk about “host” and “target architectures, what we mean is:

  1. the host architecture: where the GDB program itself is run
  2. the target architecture: where the program being debugged is run

Debugging a program that’s running on the same machine as the host might look like this:


In embedded engineering, we often want to target a foreign architecture (eg, the embedded device, connected to some debug probe). That setup might look like this:


We need a version of GDB that supports the target architecture of the program being debugged.

In this article, we’re interested in running GDB on an x86-64 or arm64 host and debugging an arm-v7m (32-bit) target. Getting the right toolchain setup can be tricky, especially if we want advanced features to be available. I’ll go over different approaches people use to download and setup GDB and explain the pros and cons of each of them.

Summary of strategies

I’m going to go through the various ways to install GDB with ARM support, but first here’s a table summarizing the approaches:

Strategy Binary or source install? Pinned to version Python support 3rd party package manager
Binaries from ARM 📦binary ✅ yes ⚠️ varies per version  
xPack GNU Arm Embedded GCC 📦binary ✅ yes ✅ yes  
System package manager 📦binary ⚠️ varies ⚠️ varies  
Conda 📦binary ✅ yes ✅ yes ⚠️ yes
Docker 📦binary ✅ yes ✅ yes ⚠️ yes
SDK-specific tools 📦binary ⚠️ varies ✅ (usually yes)  
Build from source 📁source ✅ yes ✅ (configurable)  
crosstool-NG 📁source ✅ yes ✅ (configurable)  

Details on each strategy

Binaries from ARM

ARM ships full prebuilt GCC + Binutils toolchains for Linux, Windows, and Mac. These were originally hosted on, but were moved here around 2016:

In 2022, the downloads relocated again to this page:

These packages are very handy- they include the following components:

  • GCC compiler
  • all binutils
    • GDB
    • LD
    • objdump, etc
  • prebuilt Newlib C library
  • prebuilt “Newlib-Nano” reduced size C library
  • various example linker scripts and startup files

Another benefit- you can have developers on Windows/Linux/Mac all using the same version of the compiler (doesn’t necessarily guarantee bit-for-bit reproducible builds, but chances are pretty good).

The same tools can also easily be used in whatever CI system you are using.

Note that these packages are the officially supported toolchain from ARM itself.

There is one gotcha with these packages, though- the Python support is whatever was configured at build time by ARM. This means if you’re using GDB’s Python API, you’ll need to make sure your Python scripts/dependencies/interpreter matches the interpreter embedded into GDB.

And in the latest release, Python support was dropped in the Mac release:

# check which dynamic libraries the mac arm-none-eabi-gdb binary depends on

# v11.2-2022.02 (no Python support 😔. note the *gdb-py binary is absent from
# the archive)
❯ llvm-otool-14 -L ~/Downloads/gcc-arm-11.2-2022.02-darwin-x86_64-arm-none-eabi/bin/arm-none-eabi-gdb
        /usr/lib/libncurses.5.4.dylib (compatibility version 5.4.0, current version 5.4.0)
        /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1281.100.1)
        /usr/lib/libiconv.2.dylib (compatibility version 7.0.0, current version 7.0.0)
        /usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 902.1.0)

# v10.3-2021.10 (python2.7! in 2021!)
❯ llvm-otool-14 -L ~/Downloads/gcc-arm-none-eabi-10.3-2021.10-mac/bin/arm-none-eabi-gdb-py
        /usr/lib/libncurses.5.4.dylib (compatibility version 5.4.0, current version 5.4.0)
        /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1252.250.1)
        /System/Library/Frameworks/Python.framework/Versions/2.7/Python (compatibility version 2.7.0, current version 2.7.10)
        /System/Library/Frameworks/CoreFoundation.framework/Versions/A/CoreFoundation (compatibility version 150.0.0, current version 1575.17.0)
        /usr/lib/libiconv.2.dylib (compatibility version 7.0.0, current version 7.0.0)
        /usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 400.9.4)

# still present in the linux version of v11.2-2022.02 though. But it's a
# somewhat venerable python3.6, which I don't have installed right now 😅
❯ ldd ~/Downloads/gcc-arm-11.2-2022.02-x86_64-arm-none-eabi/bin/arm-none-eabi-gdb (0x00007ffd3fd4d000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c717000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c6e8000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c6e3000) => not found => /lib/x86_64-linux-gnu/ (0x00007f9f6c6de000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c6d7000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c5f0000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c5bf000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c393000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c373000) => /lib/x86_64-linux-gnu/ (0x00007f9f6c14b000)
        /lib64/ (0x00007f9f6c764000)

This could potentially be a pretty serious problem if there is some tooling or scripting that uses the Python API and the release for your host system doesn’t have Python enabled.

xPack GNU Arm Embedded GCC

The xPack Project provides a prebuilt gcc-arm toolchain:

They tend to be pretty up-to-date, and more reliably include Python support.

There’s also build documentation, if you’d like to tweak the package:

Overall pretty simple and easy to use, and make a very good alternative to the official ARM GCC toolchains! I’ve used it as a drop-in replacement for the ARM GCC toolchains, and it’s working very well.


  • wider host support (darwin-arm64) than the official releases
  • more consistent Python support across host platform versions
  • documented + reproducible build instructions


  • only a single Python version supported per release
  • similarly non-customizable as the official ARM GCC packages. you get what you get and you don’t get upset!

System package manager

Depending on your host platform (Linux, Mac, Windows, *BSD, etc), you may or may not have a built-in or defacto package manager. For example, on Linux, you may have one of:

  • apt-get / apt (Debian, Ubuntu)
  • pacman (Arch)
  • yum (Red Hat)

Mac users will often install homebrew, since there is no built-in package manager.

Windows users may use something like chocolately or the recent winget tool.

It’s certainly convenient to be able to do something as simple as this to acquire a version of gdb compatible with arm targets:

sudo apt install gdb-multiarch

On Linux, those gdb-multiarch packages usually have a lot of supported targets. On my Ubuntu 22.04 machine, it’s 214 supported targets!

(the sed one-liner is from below)

# using the one liner from below to list supported targets
❯ gdb-multiarch --batch -nx --ex 'set architecture' 2>&1 | tail -n 1 | sed 's/auto./\n/g' | sed 's/,/\n/g' | sed 's/Requires an argument. Valid arguments are/\n/g' | sed '/^[ ]*$/d' | sort | uniq | wc -l


Of course, it’s only important that the targets we’re interested in are supported.

There are a few downsides to using a package manager to install the tools:

  • you are at the mercy of the available packages (or you need to create and maintain a package, which is probably more difficult than a normal from-source build!)
  • available versions could be pulled from registries for no reason, leaving your environment stranded
  • bugs in package versions might be difficult to deal with, since packages may not be updated frequently
  • package managers are usually host-specific (homebrew does support some Linux packages though!), which means if your development team is using different host machines, it may be difficult to standardize on one set of tools


Conda is a cross-platform package manager, which we’ve written about before on Interrupt:

It behaves quite similarly to system package managers, but can be used on different hosts (Linux/Mac/Windows), so a development team can standardize tooling without requiring a single host platform!

This is great! But there are a few downsides:

  • same as with system package managers, we are relying on externally maintained tooling packages
  • complex conda environments can get into dependency issues when updating packages
  • not all packages will provide versions for all of Linux + Mac + Windows
  • creating + maintaining conda packages can be pretty difficult, especially since cross-building conda packages is not yet fully supported

Plug (note: the author is an employee at Memfault), Memfault maintains conda packages for multi-arch GDB:


Docker is a well-known technology for freezing dependencies, by using virtualization.

For development tooling, you can set up a Docker image that contains all the tooling you need to build + debug an embedded project:

# select a specific SHA, to strictly pin the base image
FROM ubuntu:22.04@sha256:bace9fb0d5923a675c894d5c815da75ffe35e24970166a48a4460a48ae6e0d19

ARG DEBIAN_FRONTEND=noninteractive

# install GDB + GCC for ARM
RUN apt-get update && apt-get install -y --no-install-recommends \
    gcc-arm-none-eabi \

This approach has a few advantages:

  • anyone using the same Docker image should be using exactly the same version of the tools
  • the same Docker image can be used in your test / CI infrastructure, which simplifies things greatly
  • Docker images can be used on Linux/Mac/Windows without too much host setup

And a few downsides:

  • running Docker images on Mac or Windows is likely going to be lower performance than on Linux, especially for file IO, since on those platforms Docker containers are running in a virtualization environment
  • accessing host data (files) from within the Docker container can be tricky
  • accessing host hardware (USB or Serial ports) can be easy or impossible, depending on host platform
  • maintaining the Docker image itself requires some knowledge of Docker (though there are a lot of good tutorials online)

Plug- Memfault also provides a Docker image for running gdb-multiarch:

SDK-specific tools

SDK Vendors (and IDE vendors too!) will often package copies of GDB, along with other toolchain components, to make it easy to get started with their platform.

For example, Zephyr RTOS provides pre-built toolchains, that include the appropriate GDB versions:

ESP-IDF from Espressif provides similar tool packages:


  • these packages are often well-tested by the vendors, and you may even be able to get support if you encounter issues!
  • generally support all Linux/Mac/Windows these days


  • often these packages are very tailored to the particular environment (which usually is just fine), including custom patches for example
  • Python support can be pretty variable- some packages won’t have Python support at all, or it will only be for a specific version.
  • external dependencies may not be provided (eg libpython), and you’ll have to bring your own
  • can be based on pretty old GDB/binutils releases, so there might be long-resolved bugs present, or missing newer optimizations + features

Build from source

Building a custom copy of GDB is one way to enable support for the necessary architecture.

You might for example build GDB enabling all targets:

❯ ./configure
❯ make -j $(nproc)

This results in a nicely tuned copy of GDB- you can select only the features needed, enable different optimizations, select the correct Python version to integrate, etc., and are in full control of your software supply chain!

The downsides of from-source builds (in general):

  • can be challenging to build for Windows or other non-Unix platforms
  • need to maintain build tooling + binary repositories
  • using custom builds, you might encounter edge cases or warts that have been solved in more widely used copies of the same software

Quick note- if you want to see which architectures a given gdb binary supports, here’s a shell “one-liner” that can be useful (taken from here):

gdb --batch -nx --ex 'set architecture' 2>&1 | \
  tail -n 1 | \
  sed 's/auto./\n/g' | \
  sed 's/,/\n/g' | \
  sed 's/Requires an argument. Valid arguments are/\n/g' | \
  sed '/^[ ]*$/d' | sort | uniq

Warning, this won’t work with the built-in sed on macos, it’s too old and doesn’t have the necessary GNU extensions.


From the website:

Crosstool-NG is a versatile (cross) toolchain generator. It supports many architectures and components and has a simple yet powerful menuconfig-style interface.

It’s quite easy to get started with crosstool-ng, and there’s a good number of example configurations provided. A quickstart might run like this:

# 1. build crosstool-ng itself

# fetch the crosstool-ng repo
❯ git clone
❯ cd crosstool-ng
❯ git checkout crosstool-ng-1.25.0

# build crosstool-ng
❯ ./bootstrap && ./configure --enable-local && make -j$(nproc)

# 2. now to build a toolchain

# use one of the example configs
❯ ./ct-ng arm-nano-eabi
# you can use './ct-ng menuconfig' to adjust the config, eg, add CT_DEBUG_GDB=y
# to enable building gdb

# build the toolchain. this can take a while- on my middle-range laptop, it
# took about 12 minutes
❯ ./ct-ng build

# now to test it (the toolchain output needs to be inserted to path)export PATH=$PWD/.build/arm-nano-eabi/buildtools/bin:$PATHexport PATH=$PWD/.build/arm-nano-eabi/build/build-gdb-cross/gdb:$PATH
# confirm gdb is the freshly built one
❯ which gdb
# run the check-architecture script from above
❯ gdb --batch -nx --ex 'set architecture' 2>&1 | tail -n 1 | sed 's/auto./\n/g' | sed 's/,/\n/g' | sed 's/Requires an argument. Valid arguments are/\n/g' | sed '/^[ ]*$/d' | sort | uniq


# build an uninteresting programecho "int main(){return 0;}" > main.c
❯ arm-nano-eabi-gcc -mcpu=cortex-m0plus -ggdb3 -nostdlib -Wl,--entry=main -o arm.elf main.c
# check the elf it built with binutils
❯ gdb arm.elf -batch --ex 'disassemble /s main'
threads module disabled
Dump of assembler code for function main:
1       int main(){return 0;}
   0x00008000 <+0>:     push    {r7, lr}
   0x00008002 <+2>:     add     r7, sp, #0
   0x00008004 <+4>:     movs    r3, #0
   0x00008006 <+6>:     movs    r0, r3
   0x00008008 <+8>:     mov     sp, r7
   0x0000800a <+10>:    pop     {r7, pc}
End of assembler dump.

❯ arm-nano-eabi-objdump --disassemble=main --source arm.elf

arm.elf:     file format elf32-littlearm

Disassembly of section .text:

00008000 <main>:
int main(){return 0;}
    8000:       b580            push    {r7, lr}
    8002:       af00            add     r7, sp, #0
    8004:       2300            movs    r3, #0
    8006:       0018            movs    r0, r3
    8008:       46bd            mov     sp, r7
    800a:       bd80            pop     {r7, pc}

❯ arm-nano-eabi-size -Ax arm.elf
arm.elf  :
section             size      addr
.text                0xc    0x8000
.persistent          0x0   0x1800c
.noinit              0x0   0x1800c
.comment            0x22       0x0
.debug_aranges      0x20       0x0
.debug_info         0x48       0x0
.debug_abbrev       0x3a       0x0
.debug_line         0x3c       0x0
.debug_frame        0x2c       0x0
.debug_str        0x2c12       0x0
.debug_macro       0xa7d       0x0
.ARM.attributes     0x2c       0x0
Total             0x37f3

You end up with a fully-baked toolchain set up to your specifications, without a ton of work manually configuring and building various packages.

The Zephyr project currently bases its toolchain builds on crosstool-NG, with some custom sources:

And the famous Compiler Explorer project also uses crosstool-NG for cross toolchain building:


I learned a lot wading through the different varieties of GDB floating around! I was particularly surprised at how much Python support varied across all these options. That’s certainly a spot that’s caused me headaches in the past.

Overall I think the choice of toolchain/GDB depends on these criteria:

  • does your SDK/platform already provide a nicely working copy?
  • do you have specific needs that are not met by existing options?

The biggest difficulty overall seems to be sourcing a copy with

  1. the version of Python your tools are relying on
  2. supports Linux/Mac/Windows (depending on what developers on your team are using)

Though it might seem like an option of last resort, I was pleasantly surprised at how straightforward it was to build GDB from source, and crosstool-NG gets a special shoutout for really simplifying building a complete cross toolchain 🙌!

The great thing about these tools being open source: we have lots of options!

(also could be considered a downfall 😅)

Thanks so much for reading!

See anything you'd like to change? Submit a pull request or open an issue at GitHub

Noah Pendleton is an embedded software engineer at Memfault. Noah previously worked on embedded software teams at Fitbit and Markforged