[4.3] Fix wrong code for loop at -O

[Eric Botcazou <ebotcazou at adacore dot com>]

This is a wrong code regression present on the 4.3 branch.  It is actually 
caught if checking is enabled, for example in the stage 1 compiler:

loop_address2.adb: In function 'Loop_Address2':
loop_address2.adb:4: error: expected an SSA_NAME object
loop_address2.adb:4: error: in statement
# A.17 = VDEF <A.17> { A.17 }
A.17 = D.212_12;
+===========================GNAT BUG DETECTED==============================+
| 4.3.1 20080308 (prerelease) [gcc-4_3-branch revision 133036] 
(i586-suse-linux-gnu) GCC error:|
| verify_ssa failed          

This is IVOPTS turning

<bb 3>:
  # ivtmp.8_1 = PHI <ivtmp.8_13(6), 3(2)>
  # i_19 = PHI <i_7(6), 0(2)>
  b.1_4 = (const system__address) &b;
  C2b_9 = i_19 * 4;
  C9b_10 = VIEW_CONVERT_EXPR<system__address>(C2b_9);
  D.197_11 = C9b_10 + b.1_4;
  p_6 = (integer * {ref-all}) D.197_11;
  if (p_6 == 0B)
    goto <bb 4>;
  else
    goto <bb 5>;

into

<bb 2>:
  D.215_12 = 0;
  A.17 = D.215_12;
  A.18_5 = &A.17;
  D.218_22 = (system__address *) A.18_5;
  ivtmp.14_23 = (unsigned int) D.218_22;

<bb 3>:
  # ivtmp.14_14 = PHI <ivtmp.14_8(6), ivtmp.14_23(2)>
  b.1_4 = (const system__address) &b;
  D.219_24 = (long int) ivtmp.14_14;
  C9b_10 ={v} MEM[index: D.219_24];
  D.197_11 = C9b_10 + b.1_4;
  p_6 = (integer * {ref-all}) D.197_11;
  if (p_6 == 0B)
    goto <bb 4>;
  else
    goto <bb 5>;

The problem is that may_be_nonaddressable_p deems VIEW_CONVERT_EXPR<0> 
addressable.  This is already fixed on the mainline because of

    case VIEW_CONVERT_EXPR:
      /* This kind of view-conversions may wrap non-addressable objects
	 and make them look addressable.  After some processing the
	 non-addressability may be uncovered again, causing ADDR_EXPRs
	 of inappropriate objects to be built.  */
      if (is_gimple_reg (TREE_OPERAND (expr, 0))
	  || is_gimple_min_invariant (TREE_OPERAND (expr, 0)))
	return true;

so I've just backported the minimal change to make this work on the branch.

Tested on i586-suse-linux, applied on the branch as obvious and the testcase 
on the mainline as well.


2008-04-14  Eric Botcazou  <ebotcazou@adacore.com>

	* tree-ssa-loop-ivopts.c (may_be_nonaddressable_p) <VIEW_CONVERT_EXPR>:
	Return true if the operand is an INTEGER_CST.


2008-04-14  Eric Botcazou  <ebotcazou@adacore.com>

	* gnat.dg/loop_address2.adb: New test.


-- 
Eric Botcazou

Index: tree-ssa-loop-ivopts.c
===================================================================
--- tree-ssa-loop-ivopts.c	(revision 133036)
+++ tree-ssa-loop-ivopts.c	(working copy)
@@ -1451,6 +1451,8 @@ may_be_nonaddressable_p (tree expr)
       if (AGGREGATE_TYPE_P (TREE_TYPE (expr))
 	  && !AGGREGATE_TYPE_P (TREE_TYPE (TREE_OPERAND (expr, 0))))
 	return true;
+      if (TREE_CODE (TREE_OPERAND (expr, 0)) == INTEGER_CST)
+	return true;
 
       /* ... fall through ... */
 

-- { dg-do compile }
-- { dg-options "-O" }

with System, Ada.Unchecked_Conversion;
with System.Storage_Elements; use System.Storage_Elements;

procedure Loop_Address2 is

  type Ptr is access all Integer;

  function To_Ptr is new Ada.Unchecked_Conversion (System.Address, Ptr);

  function F (BM : System.Address; I : Integer) return System.Address is
  begin
    return BM + Storage_Offset (4*I);
  end;

  B : Integer;
  P : Ptr;

begin
  for I in 0 .. 2 loop
    P := To_Ptr (F (B'Address, I));
    P.all := 0;
  end loop;
end ;