C++ exception handler on gnu arm cortex m4 with freertos

Question

Update 2016-12 There is now also a minimal example for this behavior: https://community.nxp.com/message/862676

---

I'm using a ARM Cortex M4 with freertos using freescales freedom Kinetis IDE (gnu arm toolchain). Problem is that

try {
    throw 4; // old scenario also not working: throw std::runtime_error("wut");
} catch (...) {
}

results in a halted CPU and code after the try or (when some is added) in the catch handler is not executed.

And assembly can be found here: https://gist.github.com/Superlokkus/3c4201893b4c51e154e2a0afdf78fef0

I ASSUMED that this results in an SVC interrupt, I'm sorry I got that wrong, Freertos tricked me into this, because when I throw something it halts in DefaultISR.

The throw indeeds jump to __cxa_throw then from there to ___Unwind_RaiseException __gnu_Unwind_RaiseException __cxa_begin_catch> <_ZSt9terminatev> So it looks like std::terminate is called, but the catch all block should not allow this. Or is my assumption wrong and this behavior is because the gcc C++ runtime exception support is a stub which always calls terminate?!

Update 2016-09: Because I saw that rand() tries to use malloc(), I also defined a working malloc()/freeRTOS function and et voilà: __cxa_allocate_exception uses malloc (I wonder how the toolchain expects me to handle a bad_alloc case). So now, it still crashes, but after exception allocation (I think): The excecution path is :

(throwing function after exception allocation)
__cxa_throw
   ...                        //(some intructions in __cxa_throw)
   __cxa_begin_catch  //I guess something went wrong here
    _ZSt9terminatev // Immediately after __cxa_begin_catch
        _ZN10__cxxabiv111__terminateEPFvvE:
         00016dfc: push {r3, lr}
         00016dfe: blx r0  //Goes directly to WDOG_EWM_IRQHandler or hard fault handler
         00016e00: bl 0x194ac <abort>

If you wonder or it might help: My debuggers say its the WDOG_EWM_IRQHandler I crash into, if I not define the hard_fault handler and an own default handler.

So I guess something went wrong in the stack unwinding, because I go thru some symbols with "finished stack unwinding" in the name in _throw, but I didn't catched the break point I set in a destructor of an object which should have been cleaned up. And that seems to motivate __cxa_begin_catch to call abort or something.

( Kinetis Design Studio 3.2.0. with the GNU ARM C/C++ Cross Compiler Version: 1.12.1.201502281154 for our FRDM-KV31F)

Answer 1

From the RTOS side of things, C++ exceptions are just a glorified jump. As long as they're jumping from one bit of your code to another, they don't interfere with the RTOS. So you can write a try { } catch(std::exception) { }.

When there is no C++ handler, the RTOS indeed will have to step in as your C++ code stops running.

Answer 2

By fault, most of your exceptions will execute the default handler, so the first thing you need to do is determine which exception is actually executing. You can see the "Determining Which Exception Handler is Executing" section on the following page: http://www.freertos.org/Debugging-Hard-Faults-On-Cortex-M-Microcontrollers.html

I would guess, since you are not using a peripheral in your code, that it will be a fault handler, probably the hard fault. The same page (see link above) gives instructions on debugging that too.

Other than that - ensure you do the normal FreeRTOS debug things, like ensure you have configASSERT() defined, and that you have stack overflow checking on. Info on those topics is found on this page: http://www.freertos.org/FAQHelp.html

Answer 3

After successfully creating a blank project with default settings in freescales Kinetis, and asking the same problem on the nxp community, Alice_Yang , an NXP engineer (assuming by the NXP badge), told me the answer:

By default new projects link to newlib-nano which has it exception support disabled.

The libstdc++ built along with newlib-nano have exception handling disabled.

So the solution is to link simply to newlib. This can be done by simply removing the line "-specs=nano.specs" in "other linker flags" and also make sure the check box which adds the same option is also disabled. Then everything works as expected. Only the code increased by 27 kB in ROM/text size and 2kB in RAM/data.