[PATCH] Fix IVOPTs induction variable choice for &a[0] and &a[4]

[Richard Guenther <rguenther at suse dot de>]

This fixes IVOPTs which was choosing two induction variables for uses
based on &a[0] and &a[4].  For the use based on &a[4] we add the bases
&a[4], 0 and (T *)&a as candidates while for &a[0] we only add the
bases &a[0] and 0.  So we end up choosing two different induction
variables.  (This happens in SPEC2k6 calculix in the hottest loop)

The following fixes this by adding a candidate also if the base
object differs after stripping constant offsets.

Bootstrap & regtest running on x86_64-unknown-linux-gnu, I'll apply
this to the trunk if that succeeds.

Richard.

2008-08-01  Richard Guenther  <rguenther@suse.de>

	* tree-ssa-loop-ivopts.c (add_iv_value_candidates): Also add
	the candidate with the stripped base if that base is different
	from the original base even for offset zero.

Index: gcc/tree-ssa-loop-ivopts.c
===================================================================
*** gcc/tree-ssa-loop-ivopts.c	(revision 138521)
--- gcc/tree-ssa-loop-ivopts.c	(working copy)
*************** add_iv_value_candidates (struct ivopts_d
*** 2275,2283 ****
    add_candidate (data, build_int_cst (basetype, 0),
  		 iv->step, true, use);
  
!   /* Third, try removing the constant offset.  */
    base = strip_offset (iv->base, &offset);
!   if (offset)
      add_candidate (data, base, iv->step, false, use);
  }
  
--- 2275,2285 ----
    add_candidate (data, build_int_cst (basetype, 0),
  		 iv->step, true, use);
  
!   /* Third, try removing the constant offset.  Make sure to even
!      add a candidate for &a[0] vs. (T *)&a.  */
    base = strip_offset (iv->base, &offset);
!   if (offset
!       || base != iv->base)
      add_candidate (data, base, iv->step, false, use);
  }