--- old/src/hotspot/share/opto/loopnode.cpp	2018-05-16 09:40:20.964249170 +0200
+++ new/src/hotspot/share/opto/loopnode.cpp	2018-05-16 09:40:14.778267124 +0200
@@ -1252,9 +1252,7 @@
   return l->outer_safepoint();
 }
 
-Node* CountedLoopNode::skip_predicates() {
-  if (is_main_loop()) {
-    Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
+Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
     while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
@@ -1263,6 +1261,13 @@
 
     return ctrl;
   }
+
+Node* CountedLoopNode::skip_predicates() {
+  if (is_main_loop()) {
+    Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
+
+    return skip_predicates_from_entry(ctrl);
+  }
   return in(LoopNode::EntryControl);
 }
 
@@ -2371,6 +2376,13 @@
     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
     if (entry != NULL) {
       tty->print(" predicated");
+      entry = PhaseIdealLoop::skip_loop_predicates(entry);
+    }
+  }
+  if (UseProfiledLoopPredicate) {
+    entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
+    if (entry != NULL) {
+      tty->print(" profile_predicated");
     }
   }
   if (_head->is_CountedLoop()) {
@@ -2478,11 +2490,18 @@
     if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
-      entry = entry->in(0)->in(0);
+      entry = skip_loop_predicates(entry);
     }
     predicate_proj = find_predicate(entry); // Predicate
     if (predicate_proj != NULL ) {
       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
+      entry = skip_loop_predicates(entry);
+    }
+    if (UseProfiledLoopPredicate) {
+      predicate_proj = find_predicate(entry); // Predicate
+      if (predicate_proj != NULL ) {
+        useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
+      }
     }
   }
 
@@ -4165,12 +4184,33 @@
   // which can inhibit range check elimination.
   if (least != early) {
     Node* ctrl_out = least->unique_ctrl_out();
-    if (ctrl_out && ctrl_out->is_Loop() &&
-        least == ctrl_out->in(LoopNode::EntryControl) &&
-        (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
-      Node* least_dom = idom(least);
-      if (get_loop(least_dom)->is_member(get_loop(least))) {
-        least = least_dom;
+    if (ctrl_out && ctrl_out->is_CountedLoop() &&
+        least == ctrl_out->in(LoopNode::EntryControl)) {
+      Node* new_ctrl = least;
+      // Move the node above predicates so a following pass of loop
+      // predication doesn't hoist a predicate that depends on it
+      // above that node.
+      if (find_predicate_insertion_point(new_ctrl, Deoptimization::Reason_loop_limit_check) != NULL) {
+        new_ctrl = new_ctrl->in(0)->in(0);
+        assert(is_dominator(early, new_ctrl), "least != early so we can move up the dominator tree");
+      }
+      if (find_predicate_insertion_point(new_ctrl, Deoptimization::Reason_profile_predicate) != NULL) {
+        Node* c = new_ctrl->in(0)->in(0);
+        assert(is_dominator(early, c), "least != early so we can move up the dominator tree");
+        new_ctrl = c;
+      }
+      if (find_predicate_insertion_point(new_ctrl, Deoptimization::Reason_predicate) != NULL) {
+        Node* c = new_ctrl->in(0)->in(0);
+        assert(is_dominator(early, c), "least != early so we can move up the dominator tree");
+        new_ctrl = c;
+      }
+      if (new_ctrl != ctrl_out) {
+        least = new_ctrl;
+      } else if (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop()) {
+        Node* least_dom = idom(least);
+        if (get_loop(least_dom)->is_member(get_loop(least))) {
+          least = least_dom;
+        }
       }
     }
   }