I found that the node weight updates on cloned nodes during ipa-cp were
leading to incorrect/insane weights. Both the original and new node weight
computations used truncating divides, leading to a loss of total node weight.
I have fixed this by making both rounding integer divides.

Bootstrapped and tested on x86-64-unknown-linux-gnu. Ok for trunk?

2013-03-27  Teresa Johnson  <tejohn...@google.com>

        * ipa-cp.c (update_profiling_info): Perform rounding integer
        division when updating weights instead of truncating.
        (update_specialized_profile): Ditto.

Index: ipa-cp.c
===================================================================
--- ipa-cp.c    (revision 197118)
+++ ipa-cp.c    (working copy)
@@ -2588,14 +2588,18 @@ update_profiling_info (struct cgraph_node *orig_no
 
   for (cs = new_node->callees; cs ; cs = cs->next_callee)
     if (cs->frequency)
-      cs->count = cs->count * (new_sum * REG_BR_PROB_BASE
-                              / orig_node_count) / REG_BR_PROB_BASE;
+      cs->count = (cs->count
+                   * ((new_sum * REG_BR_PROB_BASE + orig_node_count/2)
+                      / orig_node_count)
+                   + REG_BR_PROB_BASE/2) / REG_BR_PROB_BASE;
     else
       cs->count = 0;
 
   for (cs = orig_node->callees; cs ; cs = cs->next_callee)
-    cs->count = cs->count * (remainder * REG_BR_PROB_BASE
-                            / orig_node_count) / REG_BR_PROB_BASE;
+    cs->count = (cs->count
+                 * ((remainder * REG_BR_PROB_BASE + orig_node_count/2)
+                    / orig_node_count)
+                 + REG_BR_PROB_BASE/2) / REG_BR_PROB_BASE;
 
   if (dump_file)
     dump_profile_updates (orig_node, new_node);
@@ -2627,14 +2631,19 @@ update_specialized_profile (struct cgraph_node *ne
 
   for (cs = new_node->callees; cs ; cs = cs->next_callee)
     if (cs->frequency)
-      cs->count += cs->count * redirected_sum / new_node_count;
+      cs->count += (cs->count
+                    * ((redirected_sum * REG_BR_PROB_BASE
+                        + new_node_count/2) / new_node_count)
+                    + REG_BR_PROB_BASE/2) / REG_BR_PROB_BASE;
     else
       cs->count = 0;
 
   for (cs = orig_node->callees; cs ; cs = cs->next_callee)
     {
-      gcov_type dec = cs->count * (redirected_sum * REG_BR_PROB_BASE
-                                  / orig_node_count) / REG_BR_PROB_BASE;
+      gcov_type dec = (cs->count
+                       * ((redirected_sum * REG_BR_PROB_BASE
+                           + orig_node_count/2) / orig_node_count)
+                       + REG_BR_PROB_BASE/2) / REG_BR_PROB_BASE;
       if (dec < cs->count)
        cs->count -= dec;
       else

--
This patch is available for review at http://codereview.appspot.com/7812053

Reply via email to