About LLVM-embedded-toolchain-for-Arm atomic

[W-M-R]

ld.lld: error: undefined symbol: atomic_load
>>> referenced by spinlock.h:639 (nuttx/include/nuttx/spinlock.h:639)
>>>               serial.c.o:(rspin_lock) in archive drivers/libdrivers.a
>>> referenced by spinlock.h:639 (nuttx/include/nuttx/spinlock.h:639)
>>>               spinlock.c.o:(rspin_lock) in archive apps/testing/ostest/libapps_ostest.a
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)

I encountered a compilation error when using clang18.1.3. Is it because the compiler does not support it? The link is bin/../lib/clang-runtimes/arm-none-eabi/armv8m.main_hard_fp_exn_rtti/lib/libclang_rt.builtins.a

[W-M-R]

I compiled a C function, not C++

[smithp35]

It is difficult to tell what has happened from the information provided.

The error message is a linker error saying that the symbol atomic_load is not found. Typically when using <libatomic.h> atomic_load is translated to a compiler builtin that either gets inlined when the implementation can be performed using the CPUs atomic instructions, or a function call to a helper function defined in "libatomic" which would use __atomic_load with a prefix of two underscores.

In summary I wouldn't expect an undefined symbol for atomic_load unless the source code that you are using has provided its own implementation of atomics. There looks to be something that could be that in https://github.com/apache/nuttx/blob/71f558765c7d53d0dad001cd1617d7f6ecf23c8c/include/nuttx/atomic.h#L77 so you may want to look there first.

For LLVM embedded Toolchain we don't provide an implementation of libatomic so if the code to handle the operation can't be inlined then you will get an undefined symbol error, although I would expect it to be of the form __atomic_*. For a Cortex-M I would recommend not using atomics larger than 4-bytes in size to limit the chances that compiler calls out to libatomic.

If you want to experiment with rebuilding the libclang_rt.builtins.a there is an implementation that can be enabled with a CMake parameter. However we've not even tried to build it (could have dependencies that can't be satisfied by bare metal toolchain), so your mileage may vary.

https://github.com/llvm/llvm-project/blob/release/18.x/compiler-rt/lib/builtins/CMakeLists.txt#L226C8-L226C42

Just to make sure you're aware, the arm-toolchain project has replaced LLVM-ET (https://github.com/arm/arm-toolchain/) for LLVM releases >= 20.