[ira] patch fixing sigfault on sh

[Vladimir Makarov]

The patch solves the problem reported by Kaz Kojima in right way
(sigfault in init_fake_stack_mems).

For sh, MEMORY_MOVE_COST call in init_ira uses
memory_move_secondary_cost which needs top_of_stack initialized in
init_fake_stack_mems.  So init_ira was called in a wrong palce.

The patch does not change generated code for x86/x86_64, ppc64,
and itanium.

2008-04-05  Vladimir Makarov  <vmakarov@redhat.com>

    * toplev.c (backend_init_target): Move init_ira after
    init_fake_stack_mems.

    * config/sh/sh.c (sh_secondary_reload): Remove check of x on null.

Index: toplev.c
===================================================================
--- toplev.c    (revision 133840)
+++ toplev.c    (working copy)
@@ -1966,12 +1966,12 @@ backend_init_target (void)
      mode-dependent.  */
   init_regs ();
 
-  /* This invokes IRA to set up reg related data structures.  */
-  init_ira ();
-
   /* This depends on stack_pointer_rtx.  */
   init_fake_stack_mems ();
 
+  /* This invokes IRA to set up reg related data structures.  */
+  init_ira ();
+
   /* Sets static_base_value[HARD_FRAME_POINTER_REGNUM], which is
      mode-dependent.  */
   init_alias_target ();