PATCH: PR objc/18408 (but really a gimplifier fix)

[Ziemowit Laski]

This is for

    objc/18408  [4.0 Regression] ICE compiling code that involves 
casting classes

but really is a fix to the gimplifier, so I'd like GIMPLE-literate folk 
to take a look at this.

The gimplification process involves stripping useless (to the back-end 
-- ObjC does need them!) type conversions, such as NOP_EXPRs.  This can 
be a problem if the NOP_EXPR in question is the argument of an 
INDIRECT_REF (e.g., '(*(List_linked*)collection).firstLink'); once the 
NOP_EXPR is removed, the INDIRECT_REF now has a type that is 
inconsistent with that of its (new) operand, leading to an assert 
failure later on.

The following patch fixes this, causing the attached test case to 
compile correctly.  Does this look OK?  I'll do full before-and-after 
testing to ensure nothing regresses, of course.  Alternatively, is 
there a better way of accomplishing the same goal?

Thanks,

--Zem

[gcc/ChangeLog]
2004-11-11  Ziemowit Laski  <zlaski@apple.com>

         PR objc/18408
         * gimplify.c (gimplify_expr): Adjust the type of
         INDIRECT_REFs after their argument is gimplified.

[gcc/testsuite/ChangeLog]
2004-11-11  Ziemowit Laski  <zlaski@apple.com>

         PR objc/18408
         * objc.dg/comp-types-7.m: New test.

Index: gcc/gimplify.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/gimplify.c,v
retrieving revision 2.84
diff -u -3 -p -r2.84 gimplify.c
--- gcc/gimplify.c      23 Oct 2004 09:16:46 -0000      2.84
+++ gcc/gimplify.c      12 Nov 2004 01:05:27 -0000
@@ -3659,6 +3659,14 @@ gimplify_expr (tree *expr_p, tree *pre_p
         case INDIRECT_REF:
           ret = gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p, 
post_p,
                                is_gimple_reg, fb_rvalue);
+
+         /* Useless type conversions may have been discarded, so ensure
+            that the INDIRECT_REF has a type corresponding to the 
pointee
+            type of its operand.  */
+         if (TREE_CODE (*expr_p) == INDIRECT_REF)
+           TREE_TYPE (*expr_p)
+             = TREE_TYPE (TREE_TYPE (TREE_OPERAND (*expr_p, 0)));
+
           recalculate_side_effects (*expr_p);
           break;

Index: gcc/testsuite/objc.dg/comp-types-7.m
===================================================================
RCS file: gcc/testsuite/objc.dg/comp-types-7.m
diff -N gcc/testsuite/objc.dg/comp-types-7.m
--- /dev/null   1 Jan 1970 00:00:00 -0000
+++ gcc/testsuite/objc.dg/comp-types-7.m        12 Nov 2004 01:05:37 
-0000
@@ -0,0 +1,29 @@
+/* { dg-do compile } */
+@interface List
+{
+@public
+  int firstLink;
+}
+@end
+
+@interface ListIndex_linked
+{
+@public
+  List *collection;
+  int link;
+}
+@end
+
+@interface List_linked: List
+@end
+
+@implementation List
+@end
+
+@implementation ListIndex_linked
+- next
+{
+   /* The gimplifier should be able to eliminate the cast in the rhs 
below.  */
+   link = ((List_linked*)collection)->firstLink;
+}
+@end