Test failures: gcc.dg/cleanup-{8,9,10,11}.c

[fxcoudert]

FAIL: gcc.dg/cleanup-8.c execution test
FAIL: gcc.dg/cleanup-9.c execution test
FAIL: gcc.dg/cleanup-10.c execution test
FAIL: gcc.dg/cleanup-11.c execution test

All four failures are identical, the simplest code is cleanup-8.c.

The abort that is triggered is this one:

static _Unwind_Reason_Code
force_unwind_stop (int version, _Unwind_Action actions,
                   _Unwind_Exception_Class exc_class,
                   struct _Unwind_Exception *exc_obj,
                   struct _Unwind_Context *context,
                   void *stop_parameter)
{
  if (actions & _UA_END_OF_STACK)
    abort ();
  return _URC_NO_REASON;
}

That function, force_unwind_stop(), is called 7 times:

• the first 6 times, the actions argument has value _UA_CLEANUP_PHASE | _UA_FORCE_UNWIND
• the last time, it is _UA_END_OF_STACK

Running the same code under Linux, the seventh time the argument is still _UA_CLEANUP_PHASE | _UA_FORCE_UNWIND.

PS: sorry @iains I couldn't reduce that one much further, I really don't know much about exceptions and this unwinding code.

[iains]

thanks. This looks exciting to debug ..

[iains]

Out of interest, does the same code build and/or function if the compiler is clang?

(since the unwinder we are using is libunwind [not libgcc_s] I wonder if the actions are supported).

[iains]

to answer my own question: clang ~/src-local/gcc-master/gcc/testsuite/gcc.dg/cleanup-8.c -I gcc/include -fexceptions -fnon-call-exceptions -o t ./t (no output)

$ ./gcc/xgcc -Bgcc ~/src-local/gcc-master/gcc/testsuite/gcc.dg/cleanup-8.c -fexceptions -fnon-call-exceptions -o t $ ./t Abort trap: 6

(I suppose at least we can compare the unwind frames between the two)

[iains]

these fails are triggered by a missing FDE record that updates the CFA to refer to the saved FP(we actually save the FP, but did not update the CFA to refer to it).

For the record, in principle, the unwinder should be able to use the existing CFA (stack pointer) info - but even if that is (a) a bug and (b) it gets fixed for some future OS release, that will not help the existing ones.

The fix was to force the FP on for Darwin (which is OK, since the ABI mandates a FP for non-leaf functions).

We will allow the user to override this with -fomit-frame-pointer on the assumption that such a user knows the intent.