[Bug rtl-optimization/51821] [4.5/4.6/4.7 Regression] 64bit > 32bit conversion produces incorrect results with optimizations

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51821

--- Comment #16 from Uros Bizjak <ubizjak at gmail dot com> 2012-01-12 17:00:48 UTC ---
(In reply to comment #15)
> Yes, this seems to be the correct approach.

Patch that fixes the failure:

Index: recog.c
===================================================================
--- recog.c     (revision 183053)
+++ recog.c     (working copy)
@@ -3038,6 +3038,7 @@ peep2_find_free_register (int from, int to, const
   static int search_ofs;
   enum reg_class cl;
   HARD_REG_SET live;
+  df_ref *def_rec;
   int i;

   gcc_assert (from < MAX_INSNS_PER_PEEP2 + 1);
@@ -3051,12 +3052,14 @@ peep2_find_free_register (int from, int to, const

   while (from != to)
     {
-      HARD_REG_SET this_live;
+      gcc_assert (peep2_insn_data[from].insn != NULL_RTX);

+      /* Don't use registers set or clobbered by the insn.  */
+      for (def_rec = DF_INSN_DEFS (peep2_insn_data[from].insn);
+          *def_rec; def_rec++)
+       SET_HARD_REG_BIT (live, DF_REF_REGNO (*def_rec));
+
       from = peep2_buf_position (from + 1);
-      gcc_assert (peep2_insn_data[from].insn != NULL_RTX);
-      REG_SET_TO_HARD_REG_SET (this_live, peep2_insn_data[from].live_before);
-      IOR_HARD_REG_SET (live, this_live);
     }

   cl = (class_str[0] == 'r' ? GENERAL_REGS