RFA: Compare Elimination Pass: Fix use of dataflow info

2011-10-07 Thread Nick Clifton 
Hi Richard, Hi Paulo,

  Following Paulo's suggestion, here is a patch to remove the calls to
  df_analyze from the compare-elimination pass, and to use
  df_get_live_[in/out].  I can confirm that with this patch applied PR
  49801 is fixed.

  OK to apply ?  For mainline and the 4.6 branch ?

Cheers
  Nick

gcc/ChangeLog
2011-10-07  Nick Clifton  ni...@redhat.com

PR middle-end/49801
* compare-elim.c (find_comparisons_in_bb): Use df_get_live_in and
df_get_live_out instead of accessing the bitmaps directly.
(execute_compare_elim_after_reload): Remove calls to df_set_flags,
df_live_add_problem and df_analyze.

Index: gcc/compare-elim.c
===
--- gcc/compare-elim.c  (revision 179647)
+++ gcc/compare-elim.c  (working copy)
@@ -356,7 +356,7 @@
 
   /* Look to see if the flags register is live outgoing here, and
 incoming to any successor not part of the extended basic block.  */
-  if (bitmap_bit_p (DF_LIVE_BB_INFO (bb)-out, targetm.flags_regnum))
+  if (bitmap_bit_p (df_get_live_out (bb), targetm.flags_regnum))
{
  edge e;
  edge_iterator ei;
@@ -364,7 +364,7 @@
  FOR_EACH_EDGE (e, ei, bb-succs)
{
  basic_block dest = e-dest;
- if (bitmap_bit_p (DF_LIVE_BB_INFO (dest)-in,
+ if (bitmap_bit_p (df_get_live_in (bb),
targetm.flags_regnum)
   !single_pred_p (dest))
{
@@ -580,10 +580,6 @@
 static unsigned int
 execute_compare_elim_after_reload (void)
 {
-  df_set_flags (DF_DEFER_INSN_RESCAN);
-  df_live_add_problem ();
-  df_analyze ();
-
   gcc_checking_assert (all_compares == NULL);
 
   /* Locate all comparisons and their uses, and eliminate duplicates.  */
@@ -602,8 +598,6 @@
 
   VEC_free (comparison_struct_p, heap, all_compares);
   all_compares = NULL;
-
-  df_analyze ();
 }
 
   return 0;

Re: RFA: Compare Elimination Pass: Fix use of dataflow info

2011-10-07 Thread Paolo Bonzini 

On 10/07/2011 12:06 PM, Nick Clifton wrote:

-  df_analyze ();
-
gcc_checking_assert (all_compares == NULL);



Ok except that this particular df_analyze should stay.  If it doesn't 
fix the bug, please try the parts of rth's patch that add the gate and 
the TODO_df_finish flag.


Paolo