Pocket article: Debug vs. Release Builds Considered Harmful

Separate “debug” and “release” builds are very common in embedded development. Typically the notion is improved debug capabilities (less aggressive compiler optimizations, more debugging information like logs) vs. highly optimized and hardened production release builds. I’m here to describe disadvantages to this practice, and why it might make sense to consolidate to a single build!

🌶️🔥 Warning! this article is taking a strong position on debug-vs-release builds. I’m hoping to highlight a few pitfalls you may run into, but it’s not intended to be a complete discounting of the reasons we end up with bifurcated builds (despite the clickbait title)! Please enjoy and as always comments are greatly appreciated 🙏

Debug vs. Release Builds for Embedded Applications

Many embedded development IDEs provide separate “Debug” and “Release” build configurations out of the box. Generally the differences between those can be summarized as:

Build config Debug info enabled? Compiler optimization? NDEBUG preprocessor definition?
Debug ✅ enabled (-g or similar) low (-O0 or -Og) DEBUG is set, or #define NDEBUG 0, etc
Release ❌ disabled high (-O3 or -Os/-Oz) NDEBUG is set

Let’s go one-by-one through these items!

Debug info

Debug info (“debugging information”) is metadata that’s included in the output symbol file (typically project.elf/project.out/project.axf, but almost always it’s an ELF file).

This debugging information is placed into the symbol file in “non-loadable” sections, which means:

  • When flashing the symbol file to a device, the debug info is not loaded (it’s completely ignored)
  • When converting the symbol file to a .bin or .hex format for factory programming, the debug info is omitted

Since enabling debug info has no impact on the final loaded executable, the only downsides to enabling it are:

  • Slightly longer build times (can be ~10%, but difficult to measure, and negligible if you’re using build caching!)
  • Much larger symbol file size (can go from approximately the on-target size of the program, to many megabytes). This is generally not a problem, except for some C++ projects, where the type information can grow dramatically (for example, building the Chrome browser with debug info enabled can exceed 4GiB ❗)

A quick example demonstrating that compiling with debug info has zero impact on the final program:

# a simple hello world programecho > hello.c <<EOF
#include <stdio.h>
int main(void) {
  printf("hello!\n");
  return 0;
}
EOF

# Compile with debug info disabled
❯ gcc -o hello-no-debug hello.c

# Compile it again with debug info enabled
❯ gcc -ggdb3 -o hello-with-debug hello.c

# Run the 'strip' tool to remove all data not part of the program contents.
# Note that we're also removing the (non-loaded) gnu build ID section, because
# the gnu build ID is computed over the program contents AND debug information;
# 'strip' doesn't remove it by default.
# See https://interrupt.memfault.com/blog/reproducible-firmware-builds for
# details!
❯ strip --remove-section=.note.gnu.build-id hello-no-debug hello-with-debug

# compare the two files; they're bit-for-bit identical
❯ diff hello-no-debug hello-with-debug && echo 'identical!'
identical!

Summary: debug info should always be enabled!

Compiler optimization level

Setting a lower (or zero) compiler optimization level can help in some step-debugging scenarios. It prevents the compiler from optimizing away variables when they’re no longer relevant, or rearranging/combining fragments of code during optimization passes.

Optimization can make step-debugging confusing, but the meaningful information, such as data structures, variables, and the call stack, is still present.

Anecdotally, I’ve personally never encountered a roadblock due to compiler optimization in step-debugging. In the worst case, I need to peek at a few assembly instructions to see what’s going on, but that’s very rare.

I’ve also used the compiler #pramga GCC optimize ("O0") or __attribute__((optimize("O0"))) to disable optimizations temporarily while debugging a particular chunk of code or a single file, instead of disabling it across an entire build.

There’s a HUGE downside to having different optimization levels for debug vs. release builds: we end up with a different executable:

  • Sizes will likely be different: stack or heap usage (affecting program stability!), code space (affecting flash memory allocation, OTA)
  • Different machine code executed by the CPU: this can impact timing (often where the nastiest bugs show up) and actual program validity, e.g. assumptions you are making about certain spots in your program may no longer be valid with different optimization levels (the classic one is instruction reordering causing concurrent access bugs!)

Because your debug (testing) build is not running the same code, you may see these bugs only show up during field test, rather than when the device is on your desk and connected to a debugger!

It’s also generally not desirable to adjust compiler optimization level when a project is nearing time to ship, to decrease risk. If the debug + release builds are using the same optimization settings, it’s less likely to hit a surprise out-of-space issue that can stop a project in its tracks 😱.

Summary: use the same optimization settings for all builds to decrease risk

NDEBUG preprocessor definition

Often these Debug/Release builds will have some preprocessor-defined symbols such as:

  • DEBUG=1 (debug enabled)
  • NDEBUG=0 (indicating this is NOT a non-debug build)

Usually these flags will be used to adjust some of the following features for production builds:

  1. Changing log verbosity (disabling DEBUG or INFO levels in production, to prevent unnecessary flash wear)
    • I’d recommend instead, sticking with the production level, and in exceptional cases (problematic device), upgrade to a more verbose level temporarily
    • Depending on the particular system, it may be preferable to adjust verbosity at runtime and keep the build the same. For example, leave WARNING and ERROR enabled by default, but allow increasing verbosity via a persistent on-device setting, or a temporary runtime flag.
    • Be especially wary of builds with different log verbosity- it can impact timing and memory usage dramatically!
  2. Disabling an “engineering interface” (eg ssh or local serial console backdoor)
    • This can be a security requirement. An alternative is to build authentication into the engineering console, but this can be error prone so proceed cautiously.
    • Having the interface available (in an authenticated way) on production units can be extremely useful when diagnosing faulty customer units that have been returned for analysis!
  3. Enabling signed/encrypted builds
    • Also very tricky! Best practice is to verify the signing/encryption implementation works throughout the development and test cycle, instead of causing surprises at the end (execute-in-place encryption too slow, key management not actually production ready, etc).
    • It might make the most sense to have a separate “development” key for signing/encrypting development builds. A key that is separate and has different security concerns than the production key.
  4. Disabling runtime asserts
    • If your system can tolerate a reset due to an ASSERT condition, these are stupendously valuable in production builds- see this post for more details

I’ve seen the DEBUG flag used pretty aggressively to cut out entire subsystems that aren’t strictly required for production. This can cause a lot of confusion, for example the following scenario:

  1. Production test build encounters a problem
  2. Engineer goes to retrieve logs, realizes the logging subsystem was entirely disabled in the “release” build
  3. 😩

Summary: it can be quite hazardous to allow splitting features based on preprocessor flags, and should only be done if there’s a strict requirement. Recommendation is to only do it when generating builds from a CI server, if absolutely necessary.

I prefer feature-specific flags when necessary. This is less confusing than a global DEBUG flag, and less likely to leak to other places:

#if !defined(LOG_LEVEL)
#define LOG_LEVEL kLogLevel_Debug
#endif

A crucial thing to be mindful of is turning off watchdogs in your debug version. Instead, it’s strongly advised to insert __asm("bkpt"); into the watchdog handler. This will enable you to promptly debug any instances where a watchdog is triggered during debugging.

Depending on the particular device, it may be necessary to add some GDB hooks for halt + continue to temporarily disable/reset a watchdog peripheral, so it doesn’t trigger when single-stepping. Some chips (like STM32’s) even have a feature that pauses a watchdog for you when the chip is debugger halted, which is great!

Summary

🙅🙅🙅🙅

Avoid separate “Debug” and “Release” builds at all costs!

More seriously, there are many downsides(complexity, wasted time, etc) if a different build is used during development vs. production. There are few cases where this is actually required!

My advice is to consider very carefully if you need a separate Release build, and make sure you’re ready to commit to supporting it.

Often it’s sufficient to enable or disable a few small features via a compilation flag for production builds. This can be done, for example, in your CI server, when preparing a tag for release. That build can also be where production keys/signing is performed, so it’s all in one place and can be audited later. Avoid “Noah’s laptop” builds making it to manufacturing 😅!

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

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