[PATCH] Fix up passing long long in ubsan with -m32 (PR sanitizer/59333)

[Jakub Jelinek]

On Thu, Dec 05, 2013 at 04:31:20PM +0100, Marek Polacek wrote:
> > > +	    }
> > > +	  t = build_fold_addr_expr (var);
> > > +	  return build2 (COMPOUND_EXPR, TREE_TYPE (t), tem, t);
> > 
> > I would expect you want to use this too even if in_expand_p...
> 
> Unfortunately, this won't be sufficient -- in expand_expr_real_1 we
> have 
>     case COMPOUND_EXPR:
>       /* Lowered by gimplify.c.  */
>       gcc_unreachable ();
> and with the code above we hit that.

Ah, ok, then perhaps you need to do expand_assignment (var, t, false);
there and then just build_fold_addr_expr.  The ugly thing is that those
assignments will be expanded there right away, so you want to move
the
  fn = ubsan_build_overflow_builtin (code, gimple_location (stmt),
                                     TREE_TYPE (arg0), arg0, arg1);
lines down right above expand_normal (fn);
Also, you want push_temp_slots (); and pop_temp_slots (); around it.
So
  push_temp_slots ();
  fn = ubsan_build_overflow_builtin (code, gimple_location (stmt),
                                     TREE_TYPE (arg0), arg0, arg1);
  expand_normal (fn);
  pop_temp_slots ();
That way if you have say multiple assign_stack_temp_for_type slots
in one spot for one diagnostics, those slots can be reused again
in another call.

	Jakub