[patch, fortran] PR65996 [5/6 Regression] gfortran ICE with -dH

[Jerry DeLisle]

This patch follows the suggestion Jakub made in the PR and is very 
straightforward.

With the patch, an abort is given on actual errors, in agreement with the
documentation for -dH.
(Yes, not very useful, but we can clear this PR)

Buffered errors bypass this abort by saving and restoring the state.

Regression tested on x86-64-linux.

OK for trunk and then back port to 5 in about a week?

A test case will be included, similar to that given by Dominique in the PR.

Regards,

Jerry

2016-01-18  Jerry DeLisle  <jvdeli...@gcc.gnu.org>

        PR fortran/65996
        * error.c (gfc_error): Save the state of abort_on_error and set
        it to false for buffered errors to allow normal processing.
        Restore the state before leaving.


Index: error.c
===================================================================
--- error.c	(revision 232535)
+++ error.c	(working copy)
@@ -1226,6 +1226,7 @@ gfc_error (const char *gmsgid, va_list ap)
 {
   va_list argp;
   va_copy (argp, ap);
+  bool saved_abort_on_error = false;
 
   if (warnings_not_errors)
     {
@@ -1250,10 +1251,14 @@ gfc_error (const char *gmsgid, va_list ap)
 
   if (buffered_p)
     {
+      /* To prevent -dH from triggering an abort on a buffered error,
+	 save abort_on_error and restore it below.  */
+      saved_abort_on_error = global_dc->abort_on_error;
+      global_dc->abort_on_error = false;
       pp->buffer = pp_error_buffer;
       global_dc->fatal_errors = false;
       /* To prevent -fmax-errors= triggering, we decrease it before
-     report_diagnostic increases it.  */
+	 report_diagnostic increases it.  */
       --errorcount;
     }
 
@@ -1264,6 +1269,8 @@ gfc_error (const char *gmsgid, va_list ap)
     {
       pp->buffer = tmp_buffer;
       global_dc->fatal_errors = fatal_errors;
+      global_dc->abort_on_error = saved_abort_on_error;
+
     }
 
   va_end (argp);