RFA: [PR objc/27377] Fix warnings about conditional expressions for compatible ObjC types

[David Ayers <ayers at fsfe dot org>]

This patch fixes the regression of pr27377.  The issue is that for
conditional expressions like:

 expr ? (id)val1 : (Object *)val2

the types of val1 and val2 are detected as incompatible even though they
are compatible according to the Objective-C type system.

This fix for build_conditional_expr is modeled after build_binary_op in
that it skips the diagnostic if the objc_compare_types verifies that the
types are compatible.

I can follow up with a little optimization for the C front end for both
build_conditional_expr and build_binary_op in that objc_compare_types
will only be called for c_dialect_objc.

Bootstrapped and regression tested on i686-pc-linux-gnu.

I know were in stage 4 and this regression is only P5... but I'll ask
anyway:
OK to commit trunk?
or should I wait for 4.5?

Cheers,
David

PS:
[This applies for the C front end as it's not specific to Objective-C]:
Actually it seems like an unfortunate workaround that
build_conditional_expr is comparing val1 and val2 to create an
artificial type combining the two.  I'm wondering  if one could
- avoid comparing val1 with val2 in build_conditional_expr
- avoid merging them to an artificial type
- have the built expression actually maintain both types
- have any usage diagnose both types independently and warn accordingly
but I'll need to dig into working with trees a bit more to understand if
this is actually feasible...

gcc/
2009-03-00  David Ayers  <ayers@fsfe.org>

	PR objc/27377
	* gcc/c-typeck.c (build_conditional_expr): Emit ObjC warnings
	by calling objc_compare_types and surpress warnings about
	incompatible C pointers that are compatible ObjC pointers.
	
testsuite/
2009-03-00  David Ayers  <ayers@fsfe.org>

	PR objc/27377
	* gcc/testsuite/objc.dg/conditional-1.m: New tests.
	
Index: gcc/c-typeck.c
===================================================================
--- gcc/c-typeck.c	(Revision 144857)
+++ gcc/c-typeck.c	(Arbeitskopie)
@@ -3438,6 +3438,9 @@
   tree result_type = NULL;
   tree orig_op1 = op1, orig_op2 = op2;
 
+  /* True means types are compatible as far as ObjC is concerned.  */
+  bool objc_ok;
+
   /* Promote both alternatives.  */
 
   if (TREE_CODE (TREE_TYPE (op1)) != VOID_TYPE)
@@ -3463,6 +3466,8 @@
       return error_mark_node;
     }
 
+  objc_ok = objc_compare_types (type1, type2, -3, NULL_TREE);
+
   /* Quickly detect the usual case where op1 and op2 have the same type
      after promotion.  */
   if (TYPE_MAIN_VARIANT (type1) == TYPE_MAIN_VARIANT (type2))
@@ -3546,8 +3551,9 @@
 	}
       else
 	{
-	  pedwarn (input_location, 0, 
-		   "pointer type mismatch in conditional expression");
+	  if (!objc_ok)
+	    pedwarn (input_location, 0, 
+		     "pointer type mismatch in conditional expression");
 	  result_type = build_pointer_type (void_type_node);
 	}
     }
Index: gcc/testsuite/objc.dg/conditional-1.m
===================================================================
--- gcc/testsuite/objc.dg/conditional-1.m	(Revision 0)
+++ gcc/testsuite/objc.dg/conditional-1.m	(Revision 0)
@@ -0,0 +1,45 @@
+/* Testing conditional warnings (without headers).  */
+/* Author: David Ayers */
+
+/* { dg-do compile } */
+
+#define nil ((id)0)
+@interface MyObject
+@end
+
+@protocol MyProtocol
+@end
+
+@interface MyProtoObject <MyProtocol>
+@end
+
+
+int
+main (int argc, char *argv[])
+{
+  id var_id = nil;
+  id <MyProtocol> var_id_p = nil;
+  MyObject *var_obj = nil;
+  MyProtoObject *var_obj_p = nil;
+
+  var_id = (var_id == var_obj) ? var_id : var_obj;
+  var_id = (var_id == var_obj) ? var_id : var_obj_p;
+
+  /* Ayers: Currently, the following test case passes for
+     technically the wrong reason (see below).
+  */
+  var_obj_p = (var_id == var_obj) ? var_obj_p : var_obj; /* { dg-warning "distinct Objective-C types" } */
+  var_obj_p = (var_id == var_obj) ? var_obj_p : var_id_p;
+
+  /* Ayers: The first of the following test cases
+     should probably warn for var_obj_p = var_obj,
+     yet that would require extensive changes to
+     build_conditional_expr to create a tree with
+     multiple types that the assignment would have
+     to evaluate both versions for correct diagnostics.
+  */
+  var_obj_p = (var_id == var_obj) ? var_id : var_obj;  
+  var_obj_p = (var_id == var_obj) ? var_id : var_obj_p;
+
+  return 0;
+}