Don't barrier() when we're unwinding an exception

See libMesh/libmesh#4351 for more context.

Author: roystgnr

src/utilities/src/timpi_init.C

@@ -25,6 +25,9 @@
 #include "timpi/communicator.h"
 #include "timpi/timpi_assert.h"
 
+#ifdef TIMPI_ENABLE_EXCEPTIONS
+#include <exception>  // For std::uncaught_exceptions
+#endif
 
 #ifdef TIMPI_HAVE_MPI
 void TIMPI_MPI_Handler (MPI_Comm *, int *, ...)
@@ -119,14 +122,21 @@ TIMPIInit::TIMPIInit (int /* argc */, const char * const * /* argv */,
 TIMPIInit::~TIMPIInit()
 {
   // Every processor had better be ready to exit at the same time.
-  // This would be a timpi_parallel_only() function, except that
-  // timpi_parallel_only() uses timpi_assert() which throws an
-  // exception which causes compilers to scream about exceptions
-  // inside destructors.
-
   // Even if we're not doing parallel_only debugging, we don't want
   // one processor to try to exit until all others are done working.
-  this->comm().barrier();
+
+  // We could be destructing here because we're unwinding the stack
+  // due to a thrown exception, though.  It's possible that an
+  // application is catching exceptions outside of the LibMeshInit
+  // scope, or that we're using a C++ compiler that does unwinding
+  // for uncaught exceptions (the standard says whether to go straight
+  // to terminate() or unwind first is "implementation-defined").  If
+  // *that* is the case then we can't safely communicate with other
+  // processors that might not all be unwinding too.
+#ifdef TIMPI_ENABLE_EXCEPTIONS
+  if (!std::uncaught_exceptions())
+#endif
+    this->comm().barrier();
 
   // Trigger any SemiPermanent cleanup before potentially finalizing MPI
   _ref.reset();