From patchwork Wed Jul 28 09:53:26 2010
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: Fix PR tree-optimization/44885
Date: Tue, 27 Jul 2010 23:53:26 -0000
From: Eric Botcazou <ebotcazou@adacore.com>
X-Patchwork-Id: 60129
Message-Id: <201007281153.26503.ebotcazou@adacore.com>
To: gcc-patches@gcc.gnu.org

Hi,

this is the failure of gnat.dg/aliasing3.adb on 32-bit platforms exposed by 
the mem-ref2 merge.  But the root cause is a latent problem in IPA-SRA: the 
pass takes the address of a component of a TYPE_NONALIASED_COMPONENT array.

Proposed fix attached, tested on x86-64-suse-linux, OK for mainline?


2010-07-28  Eric Botcazou  <ebotcazou@adacore.com>

	PR tree-optimization/44885
	* tree-sra.c (find_param_candidates): Skip pointer types to arrays
	with non-aliased component.


Index: tree-sra.c
===================================================================
--- tree-sra.c	(revision 162618)
+++ tree-sra.c	(working copy)
@@ -3147,6 +3147,8 @@ find_param_candidates (void)
 
 	  if (TREE_CODE (type) == FUNCTION_TYPE
 	      || TYPE_VOLATILE (type)
+	      || (TREE_CODE (type) == ARRAY_TYPE
+		  && TYPE_NONALIASED_COMPONENT (type))
 	      || !is_gimple_reg (parm)
 	      || is_va_list_type (type)
 	      || ptr_parm_has_direct_uses (parm))