Re: GCC EH unwinding bug and libjava calling std::terminate ()

Jan Hubicka wrote:

> current mainline is buggy in EH unwinding effectivly ignoring
> MUST_NOT_THROW regions when reached via RESX from local handlers.
> See http://gcc.gnu.org/ml/gcc-patches/2009-03/msg01285.html for details.
> 
> Unfortunately this patch causes bootstrap failure when building libjava,
> because std::terminate() is now called.  The call comes from
> run_proxy in natVMProxy.cc where we have cleanup code calling
> destructor of:
> 
> _Jv_InterpFrame frame_desc (self->self, thread, proxyClass,
>                               NULL, frame_proxy);
> 
> Now the desctuctor is pretty simple but because of:
> 'if a destructor called during stack unwinding exits with an exception,
> std::terminate is called'
> 
> and because we use -fnon-call-excpetion and destructor is accessing
> memory, we keep MUST_NOT_THROW terminate () call accessible
> because after inlining the destructor, cleanup might unwind up
> to that MUST_NOT_THROW.
> 
> Questio is how to fix this situation? Shall we link with C++ runtime,

Please don't.

> or use -fno-non-call-exceptions

No, because it might segfault, and we need to catch it.

> to build this file or somehow restruture
> code to avoid this case?

/* Gimplify a MUST_NOT_THROW_EXPR.  */

static enum gimplify_status
gimplify_must_not_throw_expr (tree *expr_p, gimple_seq *pre_p)
{
  tree stmt = *expr_p;
  tree temp = voidify_wrapper_expr (stmt, NULL);
  tree body = TREE_OPERAND (stmt, 0);
  tree termination;

  if (pragma_java_exceptions)
    termination = terminate_node;
  else
    termination = abort_node;

  stmt = build_gimple_eh_filter_tree (body, NULL_TREE,
				      build_call_n (termination, 0));



Andrew.