[PATCH] Fix PR42730

[Richard Guenther]

This fixes PR42730 - maybe_fold_offset_to_reference ends up
choosing the wrong union field.  The easy fix is to simply
avoid doing the extra work for zero offsets.

Bootstrapped and tested on x86-64_unknown-linux-gnu, applied to trunk.

Richard.

2010-01-13  Richard Guenther  <rguenther@suse.de>

	PR tree-optimization/42730
	* tree-ssa-ccp.c (maybe_fold_stmt_indirect): Add shortcut for
	offset zero.

	* gcc.c-torture/compile/pr42730.c: New testcase.

Index: gcc/tree-ssa-ccp.c
===================================================================
*** gcc/tree-ssa-ccp.c	(revision 155867)
--- gcc/tree-ssa-ccp.c	(working copy)
*************** maybe_fold_stmt_indirect (tree expr, tre
*** 2120,2125 ****
--- 2120,2129 ----
  	  && is_gimple_min_invariant (DECL_INITIAL (base)))
  	return DECL_INITIAL (base);
  
+       /* If there is no offset involved simply return the folded base.  */
+       if (integer_zerop (offset))
+ 	return base;
+ 
        /* Try folding *(&B+O) to B.X.  */
        t = maybe_fold_offset_to_reference (loc, base_addr, offset,
  					  TREE_TYPE (expr));
Index: gcc/testsuite/gcc.c-torture/compile/pr42730.c
===================================================================
*** gcc/testsuite/gcc.c-torture/compile/pr42730.c	(revision 0)
--- gcc/testsuite/gcc.c-torture/compile/pr42730.c	(revision 0)
***************
*** 0 ****
--- 1,14 ----
+ union bzz
+ {
+   unsigned *pa;
+   void *pv;
+ };
+ 
+ void foo (void)
+ {
+   union bzz u;
+   void **x;
+   void *y = 0;
+   x = &u.pv;
+   *x = y;
+ }