Merge

23 files changed, 1681 insertions, 941 deletions

diff --git a/src/share/vm/opto/callGenerator.cpp b/src/share/vm/opto/callGenerator.cpp
index 6d201139e..ed02629d6 100644
--- a/src/share/vm/opto/callGenerator.cpp
+++ b/src/share/vm/opto/callGenerator.cpp
@@ -978,31 +978,19 @@ WarmCallInfo* WarmCallInfo::remove_from(WarmCallInfo* head) {
   return head;
 }
 
-WarmCallInfo* WarmCallInfo::_always_hot  = NULL;
-WarmCallInfo* WarmCallInfo::_always_cold = NULL;
+WarmCallInfo WarmCallInfo::_always_hot(WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE(),
+                                       WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE());
+WarmCallInfo WarmCallInfo::_always_cold(WarmCallInfo::MIN_VALUE(), WarmCallInfo::MIN_VALUE(),
+                                        WarmCallInfo::MAX_VALUE(), WarmCallInfo::MAX_VALUE());
 
 WarmCallInfo* WarmCallInfo::always_hot() {
-  if (_always_hot == NULL) {
-    static double bits[sizeof(WarmCallInfo) / sizeof(double) + 1] = {0};
-    WarmCallInfo* ci = (WarmCallInfo*) bits;
-    ci->_profit = ci->_count = MAX_VALUE();
-    ci->_work   = ci->_size  = MIN_VALUE();
-    _always_hot = ci;
-  }
-  assert(_always_hot->is_hot(), "must always be hot");
-  return _always_hot;
+  assert(_always_hot.is_hot(), "must always be hot");
+  return &_always_hot;
 }
 
 WarmCallInfo* WarmCallInfo::always_cold() {
-  if (_always_cold == NULL) {
-    static double bits[sizeof(WarmCallInfo) / sizeof(double) + 1] = {0};
-    WarmCallInfo* ci = (WarmCallInfo*) bits;
-    ci->_profit = ci->_count = MIN_VALUE();
-    ci->_work   = ci->_size  = MAX_VALUE();
-    _always_cold = ci;
-  }
-  assert(_always_cold->is_cold(), "must always be cold");
-  return _always_cold;
+  assert(_always_cold.is_cold(), "must always be cold");
+  return &_always_cold;
 }
 
 
diff --git a/src/share/vm/opto/callGenerator.hpp b/src/share/vm/opto/callGenerator.hpp
index 75e8ba004..4a50f990e 100644
--- a/src/share/vm/opto/callGenerator.hpp
+++ b/src/share/vm/opto/callGenerator.hpp
@@ -215,8 +215,20 @@ class WarmCallInfo : public ResourceObj {
   WarmCallInfo*  next() const          { return _next; }
   void       set_next(WarmCallInfo* n) { _next = n; }
 
-  static WarmCallInfo* _always_hot;
-  static WarmCallInfo* _always_cold;
+  static WarmCallInfo _always_hot;
+  static WarmCallInfo _always_cold;
+
+  // Constructor intitialization of always_hot and always_cold
+  WarmCallInfo(float c, float p, float w, float s) {
+    _call = NULL;
+    _hot_cg = NULL;
+    _next = NULL;
+    _count = c;
+    _profit = p;
+    _work = w;
+    _size = s;
+    _heat = 0;
+  }
 
  public:
   // Because WarmInfo objects live over the entire lifetime of the
diff --git a/src/share/vm/opto/cfgnode.cpp b/src/share/vm/opto/cfgnode.cpp
index f6ca3ce7f..90f0234f0 100644
--- a/src/share/vm/opto/cfgnode.cpp
+++ b/src/share/vm/opto/cfgnode.cpp
@@ -1349,9 +1349,17 @@ static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) {
 static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) {
   igvn->hash_delete(n);         // Remove from hash before hacking edges
 
+  Node* predicate_proj = NULL;
   uint j = 1;
-  for( uint i = phi->req()-1; i > 0; i-- ) {
-    if( phi->in(i) == val ) {   // Found a path with val?
+  for (uint i = phi->req()-1; i > 0; i--) {
+    if (phi->in(i) == val) {   // Found a path with val?
+      if (n->is_Region()) {
+        Node* proj = PhaseIdealLoop::find_predicate(n->in(i));
+        if (proj != NULL) {
+          assert(predicate_proj == NULL, "only one predicate entry expected");
+          predicate_proj = proj;
+        }
+      }
       // Add to NEW Region/Phi, no DU info
       newn->set_req( j++, n->in(i) );
       // Remove from OLD Region/Phi
@@ -1362,6 +1370,12 @@ static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *
   // Register the new node but do not transform it.  Cannot transform until the
   // entire Region/Phi conglomerate has been hacked as a single huge transform.
   igvn->register_new_node_with_optimizer( newn );
+
+  // Clone loop predicates
+  if (predicate_proj != NULL) {
+    newn = igvn->clone_loop_predicates(predicate_proj, newn);
+  }
+
   // Now I can point to the new node.
   n->add_req(newn);
   igvn->_worklist.push(n);
diff --git a/src/share/vm/opto/compile.cpp b/src/share/vm/opto/compile.cpp
index 39bcf30bd..e6084be0e 100644
--- a/src/share/vm/opto/compile.cpp
+++ b/src/share/vm/opto/compile.cpp
@@ -1632,7 +1632,6 @@ void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
     igvn.replace_node(n, n->in(1));
   }
   assert(predicate_count()==0, "should be clean!");
-  igvn.optimize();
 }
 
 //------------------------------Optimize---------------------------------------
@@ -1689,7 +1688,7 @@ void Compile::Optimize() {
   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
     {
       TracePhase t2("idealLoop", &_t_idealLoop, true);
-      PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate);
+      PhaseIdealLoop ideal_loop( igvn, true );
       loop_opts_cnt--;
       if (major_progress()) print_method("PhaseIdealLoop 1", 2);
       if (failing())  return;
@@ -1697,7 +1696,7 @@ void Compile::Optimize() {
     // Loop opts pass if partial peeling occurred in previous pass
     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
       TracePhase t3("idealLoop", &_t_idealLoop, true);
-      PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
+      PhaseIdealLoop ideal_loop( igvn, false );
       loop_opts_cnt--;
       if (major_progress()) print_method("PhaseIdealLoop 2", 2);
       if (failing())  return;
@@ -1705,7 +1704,7 @@ void Compile::Optimize() {
     // Loop opts pass for loop-unrolling before CCP
     if(major_progress() && (loop_opts_cnt > 0)) {
       TracePhase t4("idealLoop", &_t_idealLoop, true);
-      PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
+      PhaseIdealLoop ideal_loop( igvn, false );
       loop_opts_cnt--;
       if (major_progress()) print_method("PhaseIdealLoop 3", 2);
     }
@@ -1743,21 +1742,13 @@ void Compile::Optimize() {
   // peeling, unrolling, etc.
   if(loop_opts_cnt > 0) {
     debug_only( int cnt = 0; );
-    bool loop_predication = UseLoopPredicate;
     while(major_progress() && (loop_opts_cnt > 0)) {
       TracePhase t2("idealLoop", &_t_idealLoop, true);
       assert( cnt++ < 40, "infinite cycle in loop optimization" );
-      PhaseIdealLoop ideal_loop( igvn, true, loop_predication);
+      PhaseIdealLoop ideal_loop( igvn, true);
       loop_opts_cnt--;
       if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
       if (failing())  return;
-      // Perform loop predication optimization during first iteration after CCP.
-      // After that switch it off and cleanup unused loop predicates.
-      if (loop_predication) {
-        loop_predication = false;
-        cleanup_loop_predicates(igvn);
-        if (failing())  return;
-      }
     }
   }
 
diff --git a/src/share/vm/opto/compile.hpp b/src/share/vm/opto/compile.hpp
index 5b562ecee..b2d81897a 100644
--- a/src/share/vm/opto/compile.hpp
+++ b/src/share/vm/opto/compile.hpp
@@ -489,6 +489,9 @@ class Compile : public Phase {
   // remove the opaque nodes that protect the predicates so that the unused checks and
   // uncommon traps will be eliminated from the graph.
   void cleanup_loop_predicates(PhaseIterGVN &igvn);
+  bool is_predicate_opaq(Node * n) {
+    return _predicate_opaqs->contains(n);
+  }
 
   // Compilation environment.
   Arena*            comp_arena()                { return &_comp_arena; }
diff --git a/src/share/vm/opto/doCall.cpp b/src/share/vm/opto/doCall.cpp
index e7cdaf47e..de066e992 100644
--- a/src/share/vm/opto/doCall.cpp
+++ b/src/share/vm/opto/doCall.cpp
@@ -63,6 +63,7 @@ CallGenerator* Compile::call_generator(ciMethod* call_method, int vtable_index,
                                        JVMState* jvms, bool allow_inline,
                                        float prof_factor) {
   CallGenerator* cg;
+  guarantee(call_method != NULL, "failed method resolution");
 
   // Dtrace currently doesn't work unless all calls are vanilla
   if (env()->dtrace_method_probes()) {
@@ -130,8 +131,9 @@ CallGenerator* Compile::call_generator(ciMethod* call_method, int vtable_index,
 
         // Get an adapter for the MethodHandle.
         ciMethod* target_method = method_handle->get_method_handle_adapter();
-
-        CallGenerator* hit_cg = this->call_generator(target_method, vtable_index, false, jvms, true, prof_factor);
+        CallGenerator* hit_cg = NULL;
+        if (target_method != NULL)
+          hit_cg = this->call_generator(target_method, vtable_index, false, jvms, true, prof_factor);
         if (hit_cg != NULL && hit_cg->is_inline())
           return hit_cg;
       }
@@ -152,8 +154,9 @@ CallGenerator* Compile::call_generator(ciMethod* call_method, int vtable_index,
 
       // Get an adapter for the MethodHandle.
       ciMethod* target_method = method_handle->get_invokedynamic_adapter();
-
-      CallGenerator* hit_cg = this->call_generator(target_method, vtable_index, false, jvms, true, prof_factor);
+      CallGenerator* hit_cg = NULL;
+      if (target_method != NULL)
+        hit_cg = this->call_generator(target_method, vtable_index, false, jvms, true, prof_factor);
       if (hit_cg != NULL && hit_cg->is_inline()) {
         CallGenerator* miss_cg = CallGenerator::for_dynamic_call(call_method);
         return CallGenerator::for_predicted_dynamic_call(method_handle, miss_cg, hit_cg, prof_factor);
diff --git a/src/share/vm/opto/escape.cpp b/src/share/vm/opto/escape.cpp
index 788a79785..a7b509e34 100644
--- a/src/share/vm/opto/escape.cpp
+++ b/src/share/vm/opto/escape.cpp
@@ -594,7 +594,7 @@ bool ConnectionGraph::split_AddP(Node *addp, Node *base,  PhaseGVN  *igvn) {
 
 //
 // Create a new version of orig_phi if necessary. Returns either the newly
-// created phi or an existing phi.  Sets create_new to indicate wheter  a new
+// created phi or an existing phi.  Sets create_new to indicate whether a new
 // phi was created.  Cache the last newly created phi in the node map.
 //
 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *>  &orig_phi_worklist, PhaseGVN  *igvn, bool &new_created) {
@@ -649,7 +649,7 @@ PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, Gro
 }
 
 //
-// Return a new version  of Memory Phi "orig_phi" with the inputs having the
+// Return a new version of Memory Phi "orig_phi" with the inputs having the
 // specified alias index.
 //
 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *>  &orig_phi_worklist, PhaseGVN  *igvn) {
@@ -828,11 +828,15 @@ Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArra
       break;  // hit one of our sentinels
     if (result->is_Mem()) {
       const Type *at = phase->type(result->in(MemNode::Address));
-      if (at != Type::TOP) {
-        assert (at->isa_ptr() != NULL, "pointer type required.");
-        int idx = C->get_alias_index(at->is_ptr());
-        if (idx == alias_idx)
-          break;
+      if (at == Type::TOP)
+        break; // Dead
+      assert (at->isa_ptr() != NULL, "pointer type required.");
+      int idx = C->get_alias_index(at->is_ptr());
+      if (idx == alias_idx)
+        break; // Found
+      if (!is_instance && (at->isa_oopptr() == NULL ||
+                           !at->is_oopptr()->is_known_instance())) {
+        break; // Do not skip store to general memory slice.
       }
       result = result->in(MemNode::Memory);
     }
@@ -902,13 +906,13 @@ Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArra
     PhiNode *mphi = result->as_Phi();
     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
     const TypePtr *t = mphi->adr_type();
-    if (C->get_alias_index(t) != alias_idx) {
-      // Create a new Phi with the specified alias index type.
-      result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
-    } else if (!is_instance) {
+    if (!is_instance) {
       // Push all non-instance Phis on the orig_phis worklist to update inputs
       // during Phase 4 if needed.
       orig_phis.append_if_missing(mphi);
+    } else if (C->get_alias_index(t) != alias_idx) {
+      // Create a new Phi with the specified alias index type.
+      result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
     }
   }
   // the result is either MemNode, PhiNode, InitializeNode.
diff --git a/src/share/vm/opto/graphKit.cpp b/src/share/vm/opto/graphKit.cpp
index 794a0f717..afa2efff9 100644
--- a/src/share/vm/opto/graphKit.cpp
+++ b/src/share/vm/opto/graphKit.cpp
@@ -3385,12 +3385,17 @@ void GraphKit::add_predicate(int nargs) {
 #define __ ideal.
 
 void GraphKit::sync_kit(IdealKit& ideal) {
-  // Final sync IdealKit and graphKit.
-  __ drain_delay_transform();
   set_all_memory(__ merged_memory());
+  set_i_o(__ i_o());
   set_control(__ ctrl());
 }
 
+void GraphKit::final_sync(IdealKit& ideal) {
+  // Final sync IdealKit and graphKit.
+  __ drain_delay_transform();
+  sync_kit(ideal);
+}
+
 // vanilla/CMS post barrier
 // Insert a write-barrier store.  This is to let generational GC work; we have
 // to flag all oop-stores before the next GC point.
@@ -3435,7 +3440,7 @@ void GraphKit::write_barrier_post(Node* oop_store,
   // (Else it's an array (or unknown), and we want more precise card marks.)
   assert(adr != NULL, "");
 
-  IdealKit ideal(gvn(), control(), merged_memory(), true);
+  IdealKit ideal(this, true);
 
   // Convert the pointer to an int prior to doing math on it
   Node* cast = __ CastPX(__ ctrl(), adr);
@@ -3461,7 +3466,7 @@ void GraphKit::write_barrier_post(Node* oop_store,
   }
 
   // Final sync IdealKit and GraphKit.
-  sync_kit(ideal);
+  final_sync(ideal);
 }
 
 // G1 pre/post barriers
@@ -3471,7 +3476,7 @@ void GraphKit::g1_write_barrier_pre(Node* obj,
                                     Node* val,
                                     const TypeOopPtr* val_type,
                                     BasicType bt) {
-  IdealKit ideal(gvn(), control(), merged_memory(), true);
+  IdealKit ideal(this, true);
 
   Node* tls = __ thread(); // ThreadLocalStorage
 
@@ -3548,7 +3553,7 @@ void GraphKit::g1_write_barrier_pre(Node* obj,
   } __ end_if();  // (!marking)
 
   // Final sync IdealKit and GraphKit.
-  sync_kit(ideal);
+  final_sync(ideal);
 }
 
 //
@@ -3614,7 +3619,7 @@ void GraphKit::g1_write_barrier_post(Node* oop_store,
   // (Else it's an array (or unknown), and we want more precise card marks.)
   assert(adr != NULL, "");
 
-  IdealKit ideal(gvn(), control(), merged_memory(), true);
+  IdealKit ideal(this, true);
 
   Node* tls = __ thread(); // ThreadLocalStorage
 
@@ -3688,6 +3693,6 @@ void GraphKit::g1_write_barrier_post(Node* oop_store,
   }
 
   // Final sync IdealKit and GraphKit.
-  sync_kit(ideal);
+  final_sync(ideal);
 }
 #undef __
diff --git a/src/share/vm/opto/graphKit.hpp b/src/share/vm/opto/graphKit.hpp
index 1302184fc..4523326db 100644
--- a/src/share/vm/opto/graphKit.hpp
+++ b/src/share/vm/opto/graphKit.hpp
@@ -662,7 +662,9 @@ class GraphKit : public Phase {
             && Universe::heap()->can_elide_tlab_store_barriers());
   }
 
+  // Sync Ideal and Graph kits.
   void sync_kit(IdealKit& ideal);
+  void final_sync(IdealKit& ideal);
 
   // vanilla/CMS post barrier
   void write_barrier_post(Node *store, Node* obj,
diff --git a/src/share/vm/opto/idealKit.cpp b/src/share/vm/opto/idealKit.cpp
index a143a6492..c9cb0f855 100644
--- a/src/share/vm/opto/idealKit.cpp
+++ b/src/share/vm/opto/idealKit.cpp
@@ -38,15 +38,16 @@
 const uint IdealKit::first_var = TypeFunc::Parms + 1;
 
 //----------------------------IdealKit-----------------------------------------
-IdealKit::IdealKit(PhaseGVN &gvn, Node* control, Node* mem, bool delay_all_transforms, bool has_declarations) :
-  _gvn(gvn), C(gvn.C) {
-  _initial_ctrl = control;
-  _initial_memory = mem;
+IdealKit::IdealKit(GraphKit* gkit, bool delay_all_transforms, bool has_declarations) :
+  _gvn(gkit->gvn()), C(gkit->C) {
+  _initial_ctrl = gkit->control();
+  _initial_memory = gkit->merged_memory();
+  _initial_i_o = gkit->i_o();
   _delay_all_transforms = delay_all_transforms;
   _var_ct = 0;
   _cvstate = NULL;
   // We can go memory state free or else we need the entire memory state
-  assert(mem == NULL || mem->Opcode() == Op_MergeMem, "memory must be pre-split");
+  assert(_initial_memory == NULL || _initial_memory->Opcode() == Op_MergeMem, "memory must be pre-split");
   int init_size = 5;
   _pending_cvstates = new (C->node_arena()) GrowableArray<Node*>(C->node_arena(), init_size, 0, 0);
   _delay_transform  = new (C->node_arena()) GrowableArray<Node*>(C->node_arena(), init_size, 0, 0);
@@ -56,6 +57,13 @@ IdealKit::IdealKit(PhaseGVN &gvn, Node* control, Node* mem, bool delay_all_trans
   }
 }
 
+//----------------------------sync_kit-----------------------------------------
+void IdealKit::sync_kit(GraphKit* gkit) {
+  set_all_memory(gkit->merged_memory());
+  set_i_o(gkit->i_o());
+  set_ctrl(gkit->control());
+}
+
 //-------------------------------if_then-------------------------------------
 // Create:  if(left relop right)
 //          /  \
@@ -156,16 +164,14 @@ void IdealKit::end_if() {
 // onto the stack.
 void IdealKit::loop(GraphKit* gkit, int nargs, IdealVariable& iv, Node* init, BoolTest::mask relop, Node* limit, float prob, float cnt) {
   assert((state() & (BlockS|LoopS|IfThenS|ElseS)), "bad state for new loop");
-
-  // Sync IdealKit and graphKit.
-  gkit->set_all_memory(this->merged_memory());
-  gkit->set_control(this->ctrl());
-  // Add loop predicate.
-  gkit->add_predicate(nargs);
-  // Update IdealKit memory.
-  this->set_all_memory(gkit->merged_memory());
-  this->set_ctrl(gkit->control());
-
+  if (UseLoopPredicate) {
+    // Sync IdealKit and graphKit.
+    gkit->sync_kit(*this);
+    // Add loop predicate.
+    gkit->add_predicate(nargs);
+    // Update IdealKit memory.
+    sync_kit(gkit);
+  }
   set(iv, init);
   Node* head = make_label(1);
   bind(head);
@@ -280,6 +286,7 @@ void IdealKit::declarations_done() {
   _cvstate = new_cvstate();   // initialize current cvstate
   set_ctrl(_initial_ctrl);    // initialize control in current cvstate
   set_all_memory(_initial_memory);// initialize memory in current cvstate
+  set_i_o(_initial_i_o);      // initialize i_o in current cvstate
   DEBUG_ONLY(_state->push(BlockS));
 }
 
@@ -421,6 +428,9 @@ void IdealKit::do_memory_merge(Node* merging, Node* join) {
   // Get the region for the join state
   Node* join_region = join->in(TypeFunc::Control);
   assert(join_region != NULL, "join region must exist");
+  if (join->in(TypeFunc::I_O) == NULL ) {
+    join->set_req(TypeFunc::I_O,  merging->in(TypeFunc::I_O));
+  }
   if (join->in(TypeFunc::Memory) == NULL ) {
     join->set_req(TypeFunc::Memory,  merging->in(TypeFunc::Memory));
     return;
@@ -467,6 +477,20 @@ void IdealKit::do_memory_merge(Node* merging, Node* join) {
       mms.set_memory(phi);
     }
   }
+
+  Node* join_io    = join->in(TypeFunc::I_O);
+  Node* merging_io = merging->in(TypeFunc::I_O);
+  if (join_io != merging_io) {
+    PhiNode* phi;
+    if (join_io->is_Phi() && join_io->as_Phi()->region() == join_region) {
+      phi = join_io->as_Phi();
+    } else {
+      phi = PhiNode::make(join_region, join_io, Type::ABIO);
+      phi = (PhiNode*) delay_transform(phi);
+      join->set_req(TypeFunc::I_O, phi);
+    }
+    phi->set_req(slot, merging_io);
+  }
 }
 
 
@@ -477,7 +501,8 @@ void IdealKit::make_leaf_call(const TypeFunc *slow_call_type,
                               const char *leaf_name,
                               Node* parm0,
                               Node* parm1,
-                              Node* parm2) {
+                              Node* parm2,
+                              Node* parm3) {
 
   // We only handle taking in RawMem and modifying RawMem
   const TypePtr* adr_type = TypeRawPtr::BOTTOM;
@@ -498,6 +523,55 @@ void IdealKit::make_leaf_call(const TypeFunc *slow_call_type,
   if (parm0 != NULL)  call->init_req(TypeFunc::Parms+0, parm0);
   if (parm1 != NULL)  call->init_req(TypeFunc::Parms+1, parm1);
   if (parm2 != NULL)  call->init_req(TypeFunc::Parms+2, parm2);
+  if (parm3 != NULL)  call->init_req(TypeFunc::Parms+3, parm3);
+
+  // Node *c = _gvn.transform(call);
+  call = (CallNode *) _gvn.transform(call);
+  Node *c = call; // dbx gets confused with call call->dump()
+
+  // Slow leaf call has no side-effects, sets few values
+
+  set_ctrl(transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
+
+  // Make memory for the call
+  Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
+
+  // Set the RawPtr memory state only.
+  set_memory(mem, adr_idx);
+
+  assert(C->alias_type(call->adr_type()) == C->alias_type(adr_type),
+         "call node must be constructed correctly");
+}
+
+
+void IdealKit::make_leaf_call_no_fp(const TypeFunc *slow_call_type,
+                              address slow_call,
+                              const char *leaf_name,
+                              const TypePtr* adr_type,
+                              Node* parm0,
+                              Node* parm1,
+                              Node* parm2,
+                              Node* parm3) {
+
+  // We only handle taking in RawMem and modifying RawMem
+  uint adr_idx = C->get_alias_index(adr_type);
+
+  // Slow-path leaf call
+  int size = slow_call_type->domain()->cnt();
+  CallNode *call =  (CallNode*)new (C, size) CallLeafNoFPNode( slow_call_type, slow_call, leaf_name, adr_type);
+
+  // Set fixed predefined input arguments
+  call->init_req( TypeFunc::Control, ctrl() );
+  call->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
+  // Narrow memory as only memory input
+  call->init_req( TypeFunc::Memory , memory(adr_idx));
+  call->init_req( TypeFunc::FramePtr, top() /* frameptr() */ );
+  call->init_req( TypeFunc::ReturnAdr, top() );
+
+  if (parm0 != NULL)  call->init_req(TypeFunc::Parms+0, parm0);
+  if (parm1 != NULL)  call->init_req(TypeFunc::Parms+1, parm1);
+  if (parm2 != NULL)  call->init_req(TypeFunc::Parms+2, parm2);
+  if (parm3 != NULL)  call->init_req(TypeFunc::Parms+3, parm3);
 
   // Node *c = _gvn.transform(call);
   call = (CallNode *) _gvn.transform(call);
diff --git a/src/share/vm/opto/idealKit.hpp b/src/share/vm/opto/idealKit.hpp
index 1ece14259..6787e533a 100644
--- a/src/share/vm/opto/idealKit.hpp
+++ b/src/share/vm/opto/idealKit.hpp
@@ -108,6 +108,7 @@ class IdealKit: public StackObj {
   bool _delay_all_transforms;              // flag forcing all transforms to be delayed
   Node* _initial_ctrl;                     // saves initial control until variables declared
   Node* _initial_memory;                   // saves initial memory  until variables declared
+  Node* _initial_i_o;                      // saves initial i_o  until variables declared
 
   PhaseGVN& gvn() const { return _gvn; }
   // Create a new cvstate filled with nulls
@@ -142,17 +143,21 @@ class IdealKit: public StackObj {
   Node* memory(uint alias_idx);
 
  public:
-  IdealKit(PhaseGVN &gvn, Node* control, Node* memory, bool delay_all_transforms = false, bool has_declarations = false);
+  IdealKit(GraphKit* gkit, bool delay_all_transforms = false, bool has_declarations = false);
   ~IdealKit() {
     stop();
     drain_delay_transform();
   }
+  void sync_kit(GraphKit* gkit);
+
   // Control
   Node* ctrl()                          { return _cvstate->in(TypeFunc::Control); }
   void set_ctrl(Node* ctrl)             { _cvstate->set_req(TypeFunc::Control, ctrl); }
   Node* top()                           { return C->top(); }
   MergeMemNode* merged_memory()         { return _cvstate->in(TypeFunc::Memory)->as_MergeMem(); }
   void set_all_memory(Node* mem)        { _cvstate->set_req(TypeFunc::Memory, mem); }
+  Node* i_o()                           { return _cvstate->in(TypeFunc::I_O); }
+  void set_i_o(Node* c)                 { _cvstate->set_req(TypeFunc::I_O, c); }
   void set(IdealVariable& v, Node* rhs) { _cvstate->set_req(first_var + v.id(), rhs); }
   Node* value(IdealVariable& v)         { return _cvstate->in(first_var + v.id()); }
   void dead(IdealVariable& v)           { set(v, (Node*)NULL); }
@@ -239,7 +244,18 @@ class IdealKit: public StackObj {
                       const char *leaf_name,
                       Node* parm0,
                       Node* parm1 = NULL,
-                      Node* parm2 = NULL);
+                      Node* parm2 = NULL,
+                      Node* parm3 = NULL);
+
+  void make_leaf_call_no_fp(const TypeFunc *slow_call_type,
+                            address slow_call,
+                            const char *leaf_name,
+                            const TypePtr* adr_type,
+                            Node* parm0,
+                            Node* parm1,
+                            Node* parm2,
+                            Node* parm3);
+
 };
 
 #endif // SHARE_VM_OPTO_IDEALKIT_HPP
diff --git a/src/share/vm/opto/ifnode.cpp b/src/share/vm/opto/ifnode.cpp
index 25f2e3151..bcdf5ad64 100644
--- a/src/share/vm/opto/ifnode.cpp
+++ b/src/share/vm/opto/ifnode.cpp
@@ -27,6 +27,7 @@
 #include "opto/addnode.hpp"
 #include "opto/cfgnode.hpp"
 #include "opto/connode.hpp"
+#include "opto/loopnode.hpp"
 #include "opto/phaseX.hpp"
 #include "opto/runtime.hpp"
 #include "opto/subnode.hpp"
@@ -222,22 +223,35 @@ static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) {
 
   // Make a region merging constants and a region merging the rest
   uint req_c = 0;
+  Node* predicate_proj = NULL;
   for (uint ii = 1; ii < r->req(); ii++) {
-    if( phi->in(ii) == con1 ) {
+    if (phi->in(ii) == con1) {
       req_c++;
     }
+    Node* proj = PhaseIdealLoop::find_predicate(r->in(ii));
+    if (proj != NULL) {
+      assert(predicate_proj == NULL, "only one predicate entry expected");
+      predicate_proj = proj;
+    }
   }
+  Node* predicate_c = NULL;
+  Node* predicate_x = NULL;
+
   Node *region_c = new (igvn->C, req_c + 1) RegionNode(req_c + 1);
   Node *phi_c    = con1;
   uint  len      = r->req();
-  Node *region_x = new (igvn->C, len - req_c + 1) RegionNode(len - req_c + 1);
+  Node *region_x = new (igvn->C, len - req_c) RegionNode(len - req_c);
   Node *phi_x    = PhiNode::make_blank(region_x, phi);
   for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) {
-    if( phi->in(i) == con1 ) {
+    if (phi->in(i) == con1) {
       region_c->init_req( i_c++, r  ->in(i) );
+      if (r->in(i) == predicate_proj)
+        predicate_c = predicate_proj;
     } else {
       region_x->init_req( i_x,   r  ->in(i) );
       phi_x   ->init_req( i_x++, phi->in(i) );
+      if (r->in(i) == predicate_proj)
+        predicate_x = predicate_proj;
     }
   }
 
@@ -277,8 +291,20 @@ static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) {
   // Make the true/false arms
   Node *iff_c_t = phase->transform(new (igvn->C, 1) IfTrueNode (iff_c));
   Node *iff_c_f = phase->transform(new (igvn->C, 1) IfFalseNode(iff_c));
+  if (predicate_c != NULL) {
+    assert(predicate_x == NULL, "only one predicate entry expected");
+    // Clone loop predicates to each path
+    iff_c_t = igvn->clone_loop_predicates(predicate_c, iff_c_t);
+    iff_c_f = igvn->clone_loop_predicates(predicate_c, iff_c_f);
+  }
   Node *iff_x_t = phase->transform(new (igvn->C, 1) IfTrueNode (iff_x));
   Node *iff_x_f = phase->transform(new (igvn->C, 1) IfFalseNode(iff_x));
+  if (predicate_x != NULL) {
+    assert(predicate_c == NULL, "only one predicate entry expected");
+    // Clone loop predicates to each path
+    iff_x_t = igvn->clone_loop_predicates(predicate_x, iff_x_t);
+    iff_x_f = igvn->clone_loop_predicates(predicate_x, iff_x_f);
+  }
 
   // Merge the TRUE paths
   Node *region_s = new (igvn->C, 3) RegionNode(3);
diff --git a/src/share/vm/opto/library_call.cpp b/src/share/vm/opto/library_call.cpp
index be8136bf5..0f73e144d 100644
--- a/src/share/vm/opto/library_call.cpp
+++ b/src/share/vm/opto/library_call.cpp
@@ -1120,7 +1120,7 @@ Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_ar
   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
 
-  IdealKit kit(gvn(), control(), merged_memory(), false, true);
+  IdealKit kit(this, false, true);
 #define __ kit.
   Node* zero             = __ ConI(0);
   Node* one              = __ ConI(1);
@@ -1171,7 +1171,7 @@ Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_ar
   __ bind(return_);
 
   // Final sync IdealKit and GraphKit.
-  sync_kit(kit);
+  final_sync(kit);
   Node* result = __ value(rtn);
 #undef __
   C->set_has_loops(true);
@@ -2318,22 +2318,20 @@ bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, Bas
         // of it. So we need to emit code to conditionally do the proper type of
         // store.
 
-        IdealKit ideal(gvn(), control(),  merged_memory());
+        IdealKit ideal(this);
 #define __ ideal.
         // QQQ who knows what probability is here??
         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
           // Sync IdealKit and graphKit.
-          set_all_memory( __ merged_memory());
-          set_control(__ ctrl());
+          sync_kit(ideal);
           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
           // Update IdealKit memory.
-          __ set_all_memory(merged_memory());
-          __ set_ctrl(control());
+          __ sync_kit(this);
         } __ else_(); {
           __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
         } __ end_if();
         // Final sync IdealKit and GraphKit.
-        sync_kit(ideal);
+        final_sync(ideal);
 #undef __
       }
     }
@@ -4294,81 +4292,6 @@ bool LibraryCallKit::inline_native_clone(bool is_virtual) {
   return true;
 }
 
-
-// constants for computing the copy function
-enum {
-  COPYFUNC_UNALIGNED = 0,
-  COPYFUNC_ALIGNED = 1,                 // src, dest aligned to HeapWordSize
-  COPYFUNC_CONJOINT = 0,
-  COPYFUNC_DISJOINT = 2                 // src != dest, or transfer can descend
-};
-
-// Note:  The condition "disjoint" applies also for overlapping copies
-// where an descending copy is permitted (i.e., dest_offset <= src_offset).
-static address
-select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
-  int selector =
-    (aligned  ? COPYFUNC_ALIGNED  : COPYFUNC_UNALIGNED) +
-    (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
-
-#define RETURN_STUB(xxx_arraycopy) { \
-  name = #xxx_arraycopy; \
-  return StubRoutines::xxx_arraycopy(); }
-
-#define RETURN_STUB_PARM(xxx_arraycopy, parm) {           \
-  name = #xxx_arraycopy; \
-  return StubRoutines::xxx_arraycopy(parm); }
-
-  switch (t) {
-  case T_BYTE:
-  case T_BOOLEAN:
-    switch (selector) {
-    case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jbyte_arraycopy);
-    case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jbyte_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jbyte_disjoint_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
-    }
-  case T_CHAR:
-  case T_SHORT:
-    switch (selector) {
-    case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jshort_arraycopy);
-    case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jshort_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jshort_disjoint_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
-    }
-  case T_INT:
-  case T_FLOAT:
-    switch (selector) {
-    case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jint_arraycopy);
-    case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jint_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jint_disjoint_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jint_disjoint_arraycopy);
-    }
-  case T_DOUBLE:
-  case T_LONG:
-    switch (selector) {
-    case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jlong_arraycopy);
-    case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jlong_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jlong_disjoint_arraycopy);
-    case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
-    }
-  case T_ARRAY:
-  case T_OBJECT:
-    switch (selector) {
-    case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
-    case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
-    case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
-    case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
-    }
-  default:
-    ShouldNotReachHere();
-    return NULL;
-  }
-
-#undef RETURN_STUB
-#undef RETURN_STUB_PARM
-}
-
 //------------------------------basictype2arraycopy----------------------------
 address LibraryCallKit::basictype2arraycopy(BasicType t,
                                             Node* src_offset,
@@ -4401,7 +4324,7 @@ address LibraryCallKit::basictype2arraycopy(BasicType t,
     disjoint = true;
   }
 
-  return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
+  return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 }
 
 
diff --git a/src/share/vm/opto/loopPredicate.cpp b/src/share/vm/opto/loopPredicate.cpp
new file mode 100644
index 000000000..c5ac2c682
--- /dev/null
+++ b/src/share/vm/opto/loopPredicate.cpp
@@ -0,0 +1,960 @@
+/*
+ * Copyright (c) 2011, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ *
+ */
+
+#include "precompiled.hpp"
+#include "opto/loopnode.hpp"
+#include "opto/addnode.hpp"
+#include "opto/callnode.hpp"
+#include "opto/connode.hpp"
+#include "opto/loopnode.hpp"
+#include "opto/mulnode.hpp"
+#include "opto/rootnode.hpp"
+#include "opto/subnode.hpp"
+
+/*
+ * The general idea of Loop Predication is to insert a predicate on the entry
+ * path to a loop, and raise a uncommon trap if the check of the condition fails.
+ * The condition checks are promoted from inside the loop body, and thus
+ * the checks inside the loop could be eliminated. Currently, loop predication
+ * optimization has been applied to remove array range check and loop invariant
+ * checks (such as null checks).
+*/
+
+//-------------------------------is_uncommon_trap_proj----------------------------
+// Return true if proj is the form of "proj->[region->..]call_uct"
+bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) {
+  int path_limit = 10;
+  assert(proj, "invalid argument");
+  Node* out = proj;
+  for (int ct = 0; ct < path_limit; ct++) {
+    out = out->unique_ctrl_out();
+    if (out == NULL)
+      return false;
+    if (out->is_CallStaticJava()) {
+      int req = out->as_CallStaticJava()->uncommon_trap_request();
+      if (req != 0) {
+        Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
+        if (trap_reason == reason || reason == Deoptimization::Reason_none) {
+           return true;
+        }
+      }
+      return false; // don't do further after call
+    }
+    if (out->Opcode() != Op_Region)
+      return false;
+  }
+  return false;
+}
+
+//-------------------------------is_uncommon_trap_if_pattern-------------------------
+// Return true  for "if(test)-> proj -> ...
+//                          |
+//                          V
+//                      other_proj->[region->..]call_uct"
+//
+// "must_reason_predicate" means the uct reason must be Reason_predicate
+bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) {
+  Node *in0 = proj->in(0);
+  if (!in0->is_If()) return false;
+  // Variation of a dead If node.
+  if (in0->outcnt() < 2)  return false;
+  IfNode* iff = in0->as_If();
+
+  // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
+  if (reason != Deoptimization::Reason_none) {
+    if (iff->in(1)->Opcode() != Op_Conv2B ||
+       iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
+      return false;
+    }
+  }
+
+  ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
+  if (is_uncommon_trap_proj(other_proj, reason)) {
+    assert(reason == Deoptimization::Reason_none ||
+           Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list");
+    return true;
+  }
+  return false;
+}
+
+//-------------------------------register_control-------------------------
+void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
+  assert(n->is_CFG(), "must be control node");
+  _igvn.register_new_node_with_optimizer(n);
+  loop->_body.push(n);
+  set_loop(n, loop);
+  // When called from beautify_loops() idom is not constructed yet.
+  if (_idom != NULL) {
+    set_idom(n, pred, dom_depth(pred));
+  }
+}
+
+//------------------------------create_new_if_for_predicate------------------------
+// create a new if above the uct_if_pattern for the predicate to be promoted.
+//
+//          before                                after
+//        ----------                           ----------
+//           ctrl                                 ctrl
+//            |                                     |
+//            |                                     |
+//            v                                     v
+//           iff                                 new_iff
+//          /    \                                /      \
+//         /      \                              /        \
+//        v        v                            v          v
+//  uncommon_proj cont_proj                   if_uct     if_cont
+// \      |        |                           |          |
+//  \     |        |                           |          |
+//   v    v        v                           |          v
+//     rgn       loop                          |         iff
+//      |                                      |        /     \
+//      |                                      |       /       \
+//      v                                      |      v         v
+// uncommon_trap                               | uncommon_proj cont_proj
+//                                           \  \    |           |
+//                                            \  \   |           |
+//                                             v  v  v           v
+//                                               rgn           loop
+//                                                |
+//                                                |
+//                                                v
+//                                           uncommon_trap
+//
+//
+// We will create a region to guard the uct call if there is no one there.
+// The true projecttion (if_cont) of the new_iff is returned.
+// This code is also used to clone predicates to clonned loops.
+ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
+                                                      Deoptimization::DeoptReason reason) {
+  assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
+  IfNode* iff = cont_proj->in(0)->as_If();
+
+  ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
+  Node     *rgn   = uncommon_proj->unique_ctrl_out();
+  assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
+
+  uint proj_index = 1; // region's edge corresponding to uncommon_proj
+  if (!rgn->is_Region()) { // create a region to guard the call
+    assert(rgn->is_Call(), "must be call uct");
+    CallNode* call = rgn->as_Call();
+    IdealLoopTree* loop = get_loop(call);
+    rgn = new (C, 1) RegionNode(1);
+    rgn->add_req(uncommon_proj);
+    register_control(rgn, loop, uncommon_proj);
+    _igvn.hash_delete(call);
+    call->set_req(0, rgn);
+    // When called from beautify_loops() idom is not constructed yet.
+    if (_idom != NULL) {
+      set_idom(call, rgn, dom_depth(rgn));
+    }
+  } else {
+    // Find region's edge corresponding to uncommon_proj
+    for (; proj_index < rgn->req(); proj_index++)
+      if (rgn->in(proj_index) == uncommon_proj) break;
+    assert(proj_index < rgn->req(), "sanity");
+  }
+
+  Node* entry = iff->in(0);
+  if (new_entry != NULL) {
+    // Clonning the predicate to new location.
+    entry = new_entry;
+  }
+  // Create new_iff
+  IdealLoopTree* lp = get_loop(entry);
+  IfNode *new_iff = iff->clone()->as_If();
+  new_iff->set_req(0, entry);
+  register_control(new_iff, lp, entry);
+  Node *if_cont = new (C, 1) IfTrueNode(new_iff);
+  Node *if_uct  = new (C, 1) IfFalseNode(new_iff);
+  if (cont_proj->is_IfFalse()) {
+    // Swap
+    Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
+  }
+  register_control(if_cont, lp, new_iff);
+  register_control(if_uct, get_loop(rgn), new_iff);
+
+  // if_uct to rgn
+  _igvn.hash_delete(rgn);
+  rgn->add_req(if_uct);
+  // When called from beautify_loops() idom is not constructed yet.
+  if (_idom != NULL) {
+    Node* ridom = idom(rgn);
+    Node* nrdom = dom_lca(ridom, new_iff);
+    set_idom(rgn, nrdom, dom_depth(rgn));
+  }
+
+  // If rgn has phis add new edges which has the same
+  // value as on original uncommon_proj pass.
+  assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
+  bool has_phi = false;
+  for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
+    Node* use = rgn->fast_out(i);
+    if (use->is_Phi() && use->outcnt() > 0) {
+      assert(use->in(0) == rgn, "");
+      _igvn.hash_delete(use);
+      use->add_req(use->in(proj_index));
+      _igvn._worklist.push(use);
+      has_phi = true;
+    }
+  }
+  assert(!has_phi || rgn->req() > 3, "no phis when region is created");
+
+  if (new_entry == NULL) {
+    // Attach if_cont to iff
+    _igvn.hash_delete(iff);
+    iff->set_req(0, if_cont);
+    if (_idom != NULL) {
+      set_idom(iff, if_cont, dom_depth(iff));
+    }
+  }
+  return if_cont->as_Proj();
+}
+
+//------------------------------create_new_if_for_predicate------------------------
+// Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
+ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
+                                                    Deoptimization::DeoptReason reason) {
+  assert(new_entry != 0, "only used for clone predicate");
+  assert(PhaseIdealLoop::is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
+  IfNode* iff = cont_proj->in(0)->as_If();
+
+  ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
+  Node     *rgn   = uncommon_proj->unique_ctrl_out();
+  assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
+
+  uint proj_index = 1; // region's edge corresponding to uncommon_proj
+  if (!rgn->is_Region()) { // create a region to guard the call
+    assert(rgn->is_Call(), "must be call uct");
+    CallNode* call = rgn->as_Call();
+    rgn = new (C, 1) RegionNode(1);
+    register_new_node_with_optimizer(rgn);
+    rgn->add_req(uncommon_proj);
+    hash_delete(call);
+    call->set_req(0, rgn);
+  } else {
+    // Find region's edge corresponding to uncommon_proj
+    for (; proj_index < rgn->req(); proj_index++)
+      if (rgn->in(proj_index) == uncommon_proj) break;
+    assert(proj_index < rgn->req(), "sanity");
+  }
+
+  // Create new_iff in new location.
+  IfNode *new_iff = iff->clone()->as_If();
+  new_iff->set_req(0, new_entry);
+
+  register_new_node_with_optimizer(new_iff);
+  Node *if_cont = new (C, 1) IfTrueNode(new_iff);
+  Node *if_uct  = new (C, 1) IfFalseNode(new_iff);
+  if (cont_proj->is_IfFalse()) {
+    // Swap
+    Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
+  }
+  register_new_node_with_optimizer(if_cont);
+  register_new_node_with_optimizer(if_uct);
+
+  // if_uct to rgn
+  hash_delete(rgn);
+  rgn->add_req(if_uct);
+
+  // If rgn has phis add corresponding new edges which has the same
+  // value as on original uncommon_proj pass.
+  assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
+  bool has_phi = false;
+  for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
+    Node* use = rgn->fast_out(i);
+    if (use->is_Phi() && use->outcnt() > 0) {
+      hash_delete(use);
+      use->add_req(use->in(proj_index));
+      _worklist.push(use);
+      has_phi = true;
+    }
+  }
+  assert(!has_phi || rgn->req() > 3, "no phis when region is created");
+
+  return if_cont->as_Proj();
+}
+
+//--------------------------clone_predicate-----------------------
+ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
+                                          Deoptimization::DeoptReason reason,
+                                          PhaseIdealLoop* loop_phase,
+                                          PhaseIterGVN* igvn) {
+  ProjNode* new_predicate_proj;
+  if (loop_phase != NULL) {
+    new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason);
+  } else {
+    new_predicate_proj =       igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason);
+  }
+  IfNode* iff = new_predicate_proj->in(0)->as_If();
+  Node* ctrl  = iff->in(0);
+
+  // Match original condition since predicate's projections could be swapped.
+  assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
+  Node* opq = new (igvn->C, 2) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
+  igvn->C->add_predicate_opaq(opq);
+
+  Node* bol = new (igvn->C, 2) Conv2BNode(opq);
+  if (loop_phase != NULL) {
+    loop_phase->register_new_node(opq, ctrl);
+    loop_phase->register_new_node(bol, ctrl);
+  } else {
+    igvn->register_new_node_with_optimizer(opq);
+    igvn->register_new_node_with_optimizer(bol);
+  }
+  igvn->hash_delete(iff);
+  iff->set_req(1, bol);
+  return new_predicate_proj;
+}
+
+//--------------------------move_predicate-----------------------
+// Cut predicate from old place and move it to new.
+ProjNode* PhaseIdealLoop::move_predicate(ProjNode* predicate_proj, Node* new_entry,
+                                         Deoptimization::DeoptReason reason,
+                                         PhaseIdealLoop* loop_phase,
+                                         PhaseIterGVN* igvn) {
+  assert(new_entry != NULL, "must be");
+  assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
+  IfNode* iff = predicate_proj->in(0)->as_If();
+  Node* old_entry = iff->in(0);
+
+  // Cut predicate from old place.
+  Node* old = predicate_proj;
+  igvn->_worklist.push(old);
+  for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
+    Node* use = old->last_out(i);  // for each use...
+    igvn->hash_delete(use);
+    igvn->_worklist.push(use);
+    // Update use-def info
+    uint uses_found = 0;
+    for (uint j = 0; j < use->req(); j++) {
+      if (use->in(j) == old) {
+        use->set_req(j, old_entry);
+        uses_found++;
+        if (loop_phase != NULL) {
+          if (use->is_CFG()) {
+            // When called from beautify_loops() idom is not constructed yet.
+            if (loop_phase->_idom != NULL)
+              loop_phase->set_idom(use, old_entry, loop_phase->dom_depth(use));
+          } else {
+            loop_phase->set_ctrl(use, old_entry);
+          }
+        }
+      }
+    }
+    i -= uses_found;    // we deleted 1 or more copies of this edge
+  }
+
+  // Move predicate.
+  igvn->hash_delete(iff);
+  iff->set_req(0, new_entry);
+  igvn->_worklist.push(iff);
+
+  if (loop_phase != NULL) {
+    // Fix up idom and ctrl.
+    loop_phase->set_ctrl(iff->in(1), new_entry);
+    loop_phase->set_ctrl(iff->in(1)->in(1), new_entry);
+    // When called from beautify_loops() idom is not constructed yet.
+    if (loop_phase->_idom != NULL)
+      loop_phase->set_idom(iff, new_entry, loop_phase->dom_depth(iff));
+  }
+
+  return predicate_proj;
+}
+
+//--------------------------clone_loop_predicates-----------------------
+// Interface from IGVN
+Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry) {
+  return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, false, NULL, this);
+}
+Node* PhaseIterGVN::move_loop_predicates(Node* old_entry, Node* new_entry) {
+  return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, true, NULL, this);
+}
+
+// Interface from PhaseIdealLoop
+Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry) {
+  return clone_loop_predicates(old_entry, new_entry, false, this, &this->_igvn);
+}
+Node* PhaseIdealLoop::move_loop_predicates(Node* old_entry, Node* new_entry) {
+  return clone_loop_predicates(old_entry, new_entry, true, this, &this->_igvn);
+}
+
+// Clone loop predicates to cloned loops (peeled, unswitched, split_if).
+Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
+                                                bool move_predicates,
+                                                PhaseIdealLoop* loop_phase,
+                                                PhaseIterGVN* igvn) {
+#ifdef ASSERT
+  if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
+    if (new_entry != NULL)
+      new_entry->dump();
+    assert(false, "not IfTrue, IfFalse, Region or SafePoint");
+  }
+#endif
+  // Search original predicates
+  Node* entry = old_entry;
+  if (UseLoopPredicate) {
+    ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
+    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");
+      if (move_predicates) {
+        new_entry =  move_predicate(predicate_proj, new_entry,
+                                    Deoptimization::Reason_predicate,
+                                    loop_phase, igvn);
+        assert(new_entry == predicate_proj, "old predicate fall through projection");
+      } else {
+        // clone predicate
+        new_entry = clone_predicate(predicate_proj, new_entry,
+                                    Deoptimization::Reason_predicate,
+                                    loop_phase, igvn);
+        assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
+      }
+      if (TraceLoopPredicate) {
+        tty->print_cr("Loop Predicate %s: ", move_predicates ? "moved" : "cloned");
+        debug_only( new_entry->in(0)->dump(); )
+      }
+    }
+  }
+  return new_entry;
+}
+
+//--------------------------eliminate_loop_predicates-----------------------
+void PhaseIdealLoop::eliminate_loop_predicates(Node* entry) {
+  if (UseLoopPredicate) {
+    ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
+    if (predicate_proj != NULL) { // right pattern that can be used by loop predication
+      Node* n = entry->in(0)->in(1)->in(1);
+      assert(n->Opcode()==Op_Opaque1, "must be");
+      // Remove Opaque1 node from predicates list.
+      // IGVN will remove this predicate check.
+      _igvn.replace_node(n, n->in(1));
+    }
+  }
+}
+
+//--------------------------skip_loop_predicates------------------------------
+// Skip related predicates.
+Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
+  Node* predicate = NULL;
+  if (UseLoopPredicate) {
+    predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
+    if (predicate != NULL) { // right pattern that can be used by loop predication
+      assert(entry->is_Proj() && entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
+      IfNode* iff = entry->in(0)->as_If();
+      ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
+      Node* rgn = uncommon_proj->unique_ctrl_out();
+      assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
+      entry = entry->in(0)->in(0);
+      while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
+        uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
+        if (uncommon_proj->unique_ctrl_out() != rgn)
+          break;
+        entry = entry->in(0)->in(0);
+      }
+    }
+  }
+  return entry;
+}
+
+//--------------------------find_predicate_insertion_point-------------------
+// Find a good location to insert a predicate
+ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
+  if (start_c == NULL || !start_c->is_Proj())
+    return NULL;
+  if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
+    return start_c->as_Proj();
+  }
+  return NULL;
+}
+
+//--------------------------find_predicate------------------------------------
+// Find a predicate
+Node* PhaseIdealLoop::find_predicate(Node* entry) {
+  Node* predicate = NULL;
+  if (UseLoopPredicate) {
+    predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
+    if (predicate != NULL) { // right pattern that can be used by loop predication
+      assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
+      return entry;
+    }
+  }
+  return NULL;
+}
+
+//------------------------------Invariance-----------------------------------
+// Helper class for loop_predication_impl to compute invariance on the fly and
+// clone invariants.
+class Invariance : public StackObj {
+  VectorSet _visited, _invariant;
+  Node_Stack _stack;
+  VectorSet _clone_visited;
+  Node_List _old_new; // map of old to new (clone)
+  IdealLoopTree* _lpt;
+  PhaseIdealLoop* _phase;
+
+  // Helper function to set up the invariance for invariance computation
+  // If n is a known invariant, set up directly. Otherwise, look up the
+  // the possibility to push n onto the stack for further processing.
+  void visit(Node* use, Node* n) {
+    if (_lpt->is_invariant(n)) { // known invariant
+      _invariant.set(n->_idx);
+    } else if (!n->is_CFG()) {
+      Node *n_ctrl = _phase->ctrl_or_self(n);
+      Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
+      if (_phase->is_dominator(n_ctrl, u_ctrl)) {
+        _stack.push(n, n->in(0) == NULL ? 1 : 0);
+      }
+    }
+  }
+
+  // Compute invariance for "the_node" and (possibly) all its inputs recursively
+  // on the fly
+  void compute_invariance(Node* n) {
+    assert(_visited.test(n->_idx), "must be");
+    visit(n, n);
+    while (_stack.is_nonempty()) {
+      Node*  n = _stack.node();
+      uint idx = _stack.index();
+      if (idx == n->req()) { // all inputs are processed
+        _stack.pop();
+        // n is invariant if it's inputs are all invariant
+        bool all_inputs_invariant = true;
+        for (uint i = 0; i < n->req(); i++) {
+          Node* in = n->in(i);
+          if (in == NULL) continue;
+          assert(_visited.test(in->_idx), "must have visited input");
+          if (!_invariant.test(in->_idx)) { // bad guy
+            all_inputs_invariant = false;
+            break;
+          }
+        }
+        if (all_inputs_invariant) {
+          _invariant.set(n->_idx); // I am a invariant too
+        }
+      } else { // process next input
+        _stack.set_index(idx + 1);
+        Node* m = n->in(idx);
+        if (m != NULL && !_visited.test_set(m->_idx)) {
+          visit(n, m);
+        }
+      }
+    }
+  }
+
+  // Helper function to set up _old_new map for clone_nodes.
+  // If n is a known invariant, set up directly ("clone" of n == n).
+  // Otherwise, push n onto the stack for real cloning.
+  void clone_visit(Node* n) {
+    assert(_invariant.test(n->_idx), "must be invariant");
+    if (_lpt->is_invariant(n)) { // known invariant
+      _old_new.map(n->_idx, n);
+    } else { // to be cloned
+      assert(!n->is_CFG(), "should not see CFG here");
+      _stack.push(n, n->in(0) == NULL ? 1 : 0);
+    }
+  }
+
+  // Clone "n" and (possibly) all its inputs recursively
+  void clone_nodes(Node* n, Node* ctrl) {
+    clone_visit(n);
+    while (_stack.is_nonempty()) {
+      Node*  n = _stack.node();
+      uint idx = _stack.index();
+      if (idx == n->req()) { // all inputs processed, clone n!
+        _stack.pop();
+        // clone invariant node
+        Node* n_cl = n->clone();
+        _old_new.map(n->_idx, n_cl);
+        _phase->register_new_node(n_cl, ctrl);
+        for (uint i = 0; i < n->req(); i++) {
+          Node* in = n_cl->in(i);
+          if (in == NULL) continue;
+          n_cl->set_req(i, _old_new[in->_idx]);
+        }
+      } else { // process next input
+        _stack.set_index(idx + 1);
+        Node* m = n->in(idx);
+        if (m != NULL && !_clone_visited.test_set(m->_idx)) {
+          clone_visit(m); // visit the input
+        }
+      }
+    }
+  }
+
+ public:
+  Invariance(Arena* area, IdealLoopTree* lpt) :
+    _lpt(lpt), _phase(lpt->_phase),
+    _visited(area), _invariant(area), _stack(area, 10 /* guess */),
+    _clone_visited(area), _old_new(area)
+  {}
+
+  // Map old to n for invariance computation and clone
+  void map_ctrl(Node* old, Node* n) {
+    assert(old->is_CFG() && n->is_CFG(), "must be");
+    _old_new.map(old->_idx, n); // "clone" of old is n
+    _invariant.set(old->_idx);  // old is invariant
+    _clone_visited.set(old->_idx);
+  }
+
+  // Driver function to compute invariance
+  bool is_invariant(Node* n) {
+    if (!_visited.test_set(n->_idx))
+      compute_invariance(n);
+    return (_invariant.test(n->_idx) != 0);
+  }
+
+  // Driver function to clone invariant
+  Node* clone(Node* n, Node* ctrl) {
+    assert(ctrl->is_CFG(), "must be");
+    assert(_invariant.test(n->_idx), "must be an invariant");
+    if (!_clone_visited.test(n->_idx))
+      clone_nodes(n, ctrl);
+    return _old_new[n->_idx];
+  }
+};
+
+//------------------------------is_range_check_if -----------------------------------
+// Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
+// Note: this function is particularly designed for loop predication. We require load_range
+//       and offset to be loop invariant computed on the fly by "invar"
+bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
+  if (!is_loop_exit(iff)) {
+    return false;
+  }
+  if (!iff->in(1)->is_Bool()) {
+    return false;
+  }
+  const BoolNode *bol = iff->in(1)->as_Bool();
+  if (bol->_test._test != BoolTest::lt) {
+    return false;
+  }
+  if (!bol->in(1)->is_Cmp()) {
+    return false;
+  }
+  const CmpNode *cmp = bol->in(1)->as_Cmp();
+  if (cmp->Opcode() != Op_CmpU) {
+    return false;
+  }
+  Node* range = cmp->in(2);
+  if (range->Opcode() != Op_LoadRange) {
+    const TypeInt* tint = phase->_igvn.type(range)->isa_int();
+    if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) {
+      // Allow predication on positive values that aren't LoadRanges.
+      // This allows optimization of loops where the length of the
+      // array is a known value and doesn't need to be loaded back
+      // from the array.
+      return false;
+    }
+  }
+  if (!invar.is_invariant(range)) {
+    return false;
+  }
+  Node *iv     = _head->as_CountedLoop()->phi();
+  int   scale  = 0;
+  Node *offset = NULL;
+  if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
+    return false;
+  }
+  if (offset && !invar.is_invariant(offset)) { // offset must be invariant
+    return false;
+  }
+  return true;
+}
+
+//------------------------------rc_predicate-----------------------------------
+// Create a range check predicate
+//
+// for (i = init; i < limit; i += stride) {
+//    a[scale*i+offset]
+// }
+//
+// Compute max(scale*i + offset) for init <= i < limit and build the predicate
+// as "max(scale*i + offset) u< a.length".
+//
+// There are two cases for max(scale*i + offset):
+// (1) stride*scale > 0
+//   max(scale*i + offset) = scale*(limit-stride) + offset
+// (2) stride*scale < 0
+//   max(scale*i + offset) = scale*init + offset
+BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl,
+                                       int scale, Node* offset,
+                                       Node* init, Node* limit, Node* stride,
+                                       Node* range, bool upper) {
+  DEBUG_ONLY(ttyLocker ttyl);
+  if (TraceLoopPredicate) tty->print("rc_predicate ");
+
+  Node* max_idx_expr  = init;
+  int stride_con = stride->get_int();
+  if ((stride_con > 0) == (scale > 0) == upper) {
+    max_idx_expr = new (C, 3) SubINode(limit, stride);
+    register_new_node(max_idx_expr, ctrl);
+    if (TraceLoopPredicate) tty->print("(limit - stride) ");
+  } else {
+    if (TraceLoopPredicate) tty->print("init ");
+  }
+
+  if (scale != 1) {
+    ConNode* con_scale = _igvn.intcon(scale);
+    max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
+    register_new_node(max_idx_expr, ctrl);
+    if (TraceLoopPredicate) tty->print("* %d ", scale);
+  }
+
+  if (offset && (!offset->is_Con() || offset->get_int() != 0)){
+    max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
+    register_new_node(max_idx_expr, ctrl);
+    if (TraceLoopPredicate)
+      if (offset->is_Con()) tty->print("+ %d ", offset->get_int());
+      else tty->print("+ offset ");
+  }
+
+  CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
+  register_new_node(cmp, ctrl);
+  BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
+  register_new_node(bol, ctrl);
+
+  if (TraceLoopPredicate) tty->print_cr("<u range");
+  return bol;
+}
+
+//------------------------------ loop_predication_impl--------------------------
+// Insert loop predicates for null checks and range checks
+bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
+  if (!UseLoopPredicate) return false;
+
+  if (!loop->_head->is_Loop()) {
+    // Could be a simple region when irreducible loops are present.
+    return false;
+  }
+
+  if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
+    // do nothing for infinite loops
+    return false;
+  }
+
+  CountedLoopNode *cl = NULL;
+  if (loop->_head->is_CountedLoop()) {
+    cl = loop->_head->as_CountedLoop();
+    // do nothing for iteration-splitted loops
+    if (!cl->is_normal_loop()) return false;
+  }
+
+  LoopNode *lpn  = loop->_head->as_Loop();
+  Node* entry = lpn->in(LoopNode::EntryControl);
+
+  ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
+  if (!predicate_proj) {
+#ifndef PRODUCT
+    if (TraceLoopPredicate) {
+      tty->print("missing predicate:");
+      loop->dump_head();
+      lpn->dump(1);
+    }
+#endif
+    return false;
+  }
+  ConNode* zero = _igvn.intcon(0);
+  set_ctrl(zero, C->root());
+
+  ResourceArea *area = Thread::current()->resource_area();
+  Invariance invar(area, loop);
+
+  // Create list of if-projs such that a newer proj dominates all older
+  // projs in the list, and they all dominate loop->tail()
+  Node_List if_proj_list(area);
+  LoopNode *head  = loop->_head->as_Loop();
+  Node *current_proj = loop->tail(); //start from tail
+  while (current_proj != head) {
+    if (loop == get_loop(current_proj) && // still in the loop ?
+        current_proj->is_Proj()        && // is a projection  ?
+        current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
+      if_proj_list.push(current_proj);
+    }
+    current_proj = idom(current_proj);
+  }
+
+  bool hoisted = false; // true if at least one proj is promoted
+  while (if_proj_list.size() > 0) {
+    // Following are changed to nonnull when a predicate can be hoisted
+    ProjNode* new_predicate_proj = NULL;
+
+    ProjNode* proj = if_proj_list.pop()->as_Proj();
+    IfNode*   iff  = proj->in(0)->as_If();
+
+    if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
+      if (loop->is_loop_exit(iff)) {
+        // stop processing the remaining projs in the list because the execution of them
+        // depends on the condition of "iff" (iff->in(1)).
+        break;
+      } else {
+        // Both arms are inside the loop. There are two cases:
+        // (1) there is one backward branch. In this case, any remaining proj
+        //     in the if_proj list post-dominates "iff". So, the condition of "iff"
+        //     does not determine the execution the remining projs directly, and we
+        //     can safely continue.
+        // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
+        //     does not dominate loop->tail(), so it can not be in the if_proj list.
+        continue;
+      }
+    }
+
+    Node*     test = iff->in(1);
+    if (!test->is_Bool()){ //Conv2B, ...
+      continue;
+    }
+    BoolNode* bol = test->as_Bool();
+    if (invar.is_invariant(bol)) {
+      // Invariant test
+      new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
+                                                       Deoptimization::Reason_predicate);
+      Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
+      BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
+
+      // Negate test if necessary
+      bool negated = false;
+      if (proj->_con != predicate_proj->_con) {
+        new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
+        register_new_node(new_predicate_bol, ctrl);
+        negated = true;
+      }
+      IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
+      _igvn.hash_delete(new_predicate_iff);
+      new_predicate_iff->set_req(1, new_predicate_bol);
+#ifndef PRODUCT
+      if (TraceLoopPredicate) {
+        tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
+        loop->dump_head();
+      } else if (TraceLoopOpts) {
+        tty->print("Predicate IC ");
+        loop->dump_head();
+      }
+#endif
+    } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
+      assert(proj->_con == predicate_proj->_con, "must match");
+
+      // Range check for counted loops
+      const Node*    cmp    = bol->in(1)->as_Cmp();
+      Node*          idx    = cmp->in(1);
+      assert(!invar.is_invariant(idx), "index is variant");
+      assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be");
+      Node* rng = cmp->in(2);
+      assert(invar.is_invariant(rng), "range must be invariant");
+      int scale    = 1;
+      Node* offset = zero;
+      bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
+      assert(ok, "must be index expression");
+
+      Node* init    = cl->init_trip();
+      Node* limit   = cl->limit();
+      Node* stride  = cl->stride();
+
+      // Build if's for the upper and lower bound tests.  The
+      // lower_bound test will dominate the upper bound test and all
+      // cloned or created nodes will use the lower bound test as
+      // their declared control.
+      ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
+      ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
+      assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
+      Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
+
+      // Perform cloning to keep Invariance state correct since the
+      // late schedule will place invariant things in the loop.
+      rng = invar.clone(rng, ctrl);
+      if (offset && offset != zero) {
+        assert(invar.is_invariant(offset), "offset must be loop invariant");
+        offset = invar.clone(offset, ctrl);
+      }
+
+      // Test the lower bound
+      Node*  lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false);
+      IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
+      _igvn.hash_delete(lower_bound_iff);
+      lower_bound_iff->set_req(1, lower_bound_bol);
+      if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
+
+      // Test the upper bound
+      Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true);
+      IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
+      _igvn.hash_delete(upper_bound_iff);
+      upper_bound_iff->set_req(1, upper_bound_bol);
+      if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
+
+      // Fall through into rest of the clean up code which will move
+      // any dependent nodes onto the upper bound test.
+      new_predicate_proj = upper_bound_proj;
+
+#ifndef PRODUCT
+      if (TraceLoopOpts && !TraceLoopPredicate) {
+        tty->print("Predicate RC ");
+        loop->dump_head();
+      }
+#endif
+    } else {
+      // Loop variant check (for example, range check in non-counted loop)
+      // with uncommon trap.
+      continue;
+    }
+    assert(new_predicate_proj != NULL, "sanity");
+    // Success - attach condition (new_predicate_bol) to predicate if
+    invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
+
+    // Eliminate the old If in the loop body
+    dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
+
+    hoisted = true;
+    C->set_major_progress();
+  } // end while
+
+#ifndef PRODUCT
+  // report that the loop predication has been actually performed
+  // for this loop
+  if (TraceLoopPredicate && hoisted) {
+    tty->print("Loop Predication Performed:");
+    loop->dump_head();
+  }
+#endif
+
+  return hoisted;
+}
+
+//------------------------------loop_predication--------------------------------
+// driver routine for loop predication optimization
+bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
+  bool hoisted = false;
+  // Recursively promote predicates
+  if (_child) {
+    hoisted = _child->loop_predication( phase);
+  }
+
+  // self
+  if (!_irreducible && !tail()->is_top()) {
+    hoisted |= phase->loop_predication_impl(this);
+  }
+
+  if (_next) { //sibling
+    hoisted |= _next->loop_predication( phase);
+  }
+
+  return hoisted;
+}
+
diff --git a/src/share/vm/opto/loopTransform.cpp b/src/share/vm/opto/loopTransform.cpp
index 9c5a7e034..b182f4e4f 100644
--- a/src/share/vm/opto/loopTransform.cpp
+++ b/src/share/vm/opto/loopTransform.cpp
@@ -63,6 +63,46 @@ void IdealLoopTree::record_for_igvn() {
   }
 }
 
+//------------------------------compute_exact_trip_count-----------------------
+// Compute loop exact trip count if possible. Do not recalculate trip count for
+// split loops (pre-main-post) which have their limits and inits behind Opaque node.
+void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
+  if (!_head->as_Loop()->is_valid_counted_loop()) {
+    return;
+  }
+  CountedLoopNode* cl = _head->as_CountedLoop();
+  // Trip count may become nonexact for iteration split loops since
+  // RCE modifies limits. Note, _trip_count value is not reset since
+  // it is used to limit unrolling of main loop.
+  cl->set_nonexact_trip_count();
+
+  // Loop's test should be part of loop.
+  if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
+    return; // Infinite loop
+
+#ifdef ASSERT
+  BoolTest::mask bt = cl->loopexit()->test_trip();
+  assert(bt == BoolTest::lt || bt == BoolTest::gt ||
+         bt == BoolTest::ne, "canonical test is expected");
+#endif
+
+  Node* init_n = cl->init_trip();
+  Node* limit_n = cl->limit();
+  if (init_n  != NULL &&  init_n->is_Con() &&
+      limit_n != NULL && limit_n->is_Con()) {
+    // Use longs to avoid integer overflow.
+    int stride_con  = cl->stride_con();
+    long init_con   = cl->init_trip()->get_int();
+    long limit_con  = cl->limit()->get_int();
+    int stride_m    = stride_con - (stride_con > 0 ? 1 : -1);
+    long trip_count = (limit_con - init_con + stride_m)/stride_con;
+    if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
+      // Set exact trip count.
+      cl->set_exact_trip_count((uint)trip_count);
+    }
+  }
+}
+
 //------------------------------compute_profile_trip_cnt----------------------------
 // Compute loop trip count from profile data as
 //    (backedge_count + loop_exit_count) / loop_exit_count
@@ -301,6 +341,132 @@ void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_n
 //         peeled-loop backedge has 2 users.
 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
 //         extra backedge user.
+//
+//                   orig
+//
+//                  stmt1
+//                    |
+//                    v
+//              loop predicate
+//                    |
+//                    v
+//                   loop<----+
+//                     |      |
+//                   stmt2    |
+//                     |      |
+//                     v      |
+//                    if      ^
+//                   / \      |
+//                  /   \     |
+//                 v     v    |
+//               false true   |
+//               /       \    |
+//              /         ----+
+//             |
+//             v
+//           exit
+//
+//
+//            after clone loop
+//
+//                   stmt1
+//                     |
+//                     v
+//               loop predicate
+//                 /       \
+//        clone   /         \   orig
+//               /           \
+//              /             \
+//             v               v
+//   +---->loop clone          loop<----+
+//   |      |                    |      |
+//   |    stmt2 clone          stmt2    |
+//   |      |                    |      |
+//   |      v                    v      |
+//   ^      if clone            If      ^
+//   |      / \                / \      |
+//   |     /   \              /   \     |
+//   |    v     v            v     v    |
+//   |    true  false      false true   |
+//   |    /         \      /       \    |
+//   +----           \    /         ----+
+//                    \  /
+//                    1v v2
+//                  region
+//                     |
+//                     v
+//                   exit
+//
+//
+//         after peel and predicate move
+//
+//                   stmt1
+//                    /
+//                   /
+//        clone     /            orig
+//                 /
+//                /              +----------+
+//               /               |          |
+//              /          loop predicate   |
+//             /                 |          |
+//            v                  v          |
+//   TOP-->loop clone          loop<----+   |
+//          |                    |      |   |
+//        stmt2 clone          stmt2    |   |
+//          |                    |      |   ^
+//          v                    v      |   |
+//          if clone            If      ^   |
+//          / \                / \      |   |
+//         /   \              /   \     |   |
+//        v     v            v     v    |   |
+//      true   false      false  true   |   |
+//        |         \      /       \    |   |
+//        |          \    /         ----+   ^
+//        |           \  /                  |
+//        |           1v v2                 |
+//        v         region                  |
+//        |            |                    |
+//        |            v                    |
+//        |          exit                   |
+//        |                                 |
+//        +--------------->-----------------+
+//
+//
+//              final graph
+//
+//                  stmt1
+//                    |
+//                    v
+//                  stmt2 clone
+//                    |
+//                    v
+//                   if clone
+//                  / |
+//                 /  |
+//                v   v
+//            false  true
+//             |      |
+//             |      v
+//             | loop predicate
+//             |      |
+//             |      v
+//             |     loop<----+
+//             |      |       |
+//             |    stmt2     |
+//             |      |       |
+//             |      v       |
+//             v      if      ^
+//             |     /  \     |
+//             |    /    \    |
+//             |   v     v    |
+//             | false  true  |
+//             |  |        \  |
+//             v  v         --+
+//            region
+//              |
+//              v
+//             exit
+//
 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 
   C->set_major_progress();
@@ -315,9 +481,10 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
     loop->dump_head();
   }
 #endif
-  Node *h = loop->_head;
-  if (h->is_CountedLoop()) {
-    CountedLoopNode *cl = h->as_CountedLoop();
+  Node* head = loop->_head;
+  bool counted_loop = head->is_CountedLoop();
+  if (counted_loop) {
+    CountedLoopNode *cl = head->as_CountedLoop();
     assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
     cl->set_trip_count(cl->trip_count() - 1);
     if (cl->is_main_loop()) {
@@ -330,11 +497,11 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 #endif
     }
   }
+  Node* entry = head->in(LoopNode::EntryControl);
 
   // Step 1: Clone the loop body.  The clone becomes the peeled iteration.
   //         The pre-loop illegally has 2 control users (old & new loops).
-  clone_loop( loop, old_new, dom_depth(loop->_head) );
-
+  clone_loop( loop, old_new, dom_depth(head) );
 
   // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
   //         Do this by making the old-loop fall-in edges act as if they came
@@ -342,12 +509,15 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
   //         backedges) and then map to the new peeled iteration.  This leaves
   //         the pre-loop with only 1 user (the new peeled iteration), but the
   //         peeled-loop backedge has 2 users.
-  for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) {
-    Node* old = loop->_head->fast_out(j);
-    if( old->in(0) == loop->_head && old->req() == 3 &&
-        (old->is_Loop() || old->is_Phi()) ) {
-      Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
-      if( !new_exit_value )     // Backedge value is ALSO loop invariant?
+  Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx];
+  new_exit_value = move_loop_predicates(entry, new_exit_value);
+  _igvn.hash_delete(head);
+  head->set_req(LoopNode::EntryControl, new_exit_value);
+  for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
+    Node* old = head->fast_out(j);
+    if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
+      new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
+      if (!new_exit_value )     // Backedge value is ALSO loop invariant?
         // Then loop body backedge value remains the same.
         new_exit_value = old->in(LoopNode::LoopBackControl);
       _igvn.hash_delete(old);
@@ -358,12 +528,12 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 
   // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
   //         extra backedge user.
-  Node *nnn = old_new[loop->_head->_idx];
-  _igvn.hash_delete(nnn);
-  nnn->set_req(LoopNode::LoopBackControl, C->top());
-  for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) {
-    Node* use = nnn->fast_out(j2);
-    if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) {
+  Node* new_head = old_new[head->_idx];
+  _igvn.hash_delete(new_head);
+  new_head->set_req(LoopNode::LoopBackControl, C->top());
+  for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
+    Node* use = new_head->fast_out(j2);
+    if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
       _igvn.hash_delete(use);
       use->set_req(LoopNode::LoopBackControl, C->top());
     }
@@ -371,15 +541,15 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 
 
   // Step 4: Correct dom-depth info.  Set to loop-head depth.
-  int dd = dom_depth(loop->_head);
-  set_idom(loop->_head, loop->_head->in(1), dd);
+  int dd = dom_depth(head);
+  set_idom(head, head->in(1), dd);
   for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
     Node *old = loop->_body.at(j3);
     Node *nnn = old_new[old->_idx];
     if (!has_ctrl(nnn))
       set_idom(nnn, idom(nnn), dd-1);
     // While we're at it, remove any SafePoints from the peeled code
-    if( old->Opcode() == Op_SafePoint ) {
+    if (old->Opcode() == Op_SafePoint) {
       Node *nnn = old_new[old->_idx];
       lazy_replace(nnn,nnn->in(TypeFunc::Control));
     }
@@ -392,34 +562,26 @@ void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
   loop->record_for_igvn();
 }
 
+#define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
+
 //------------------------------policy_maximally_unroll------------------------
-// Return exact loop trip count, or 0 if not maximally unrolling
+// Calculate exact loop trip count and return true if loop can be maximally
+// unrolled.
 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
   CountedLoopNode *cl = _head->as_CountedLoop();
   assert(cl->is_normal_loop(), "");
+  if (!cl->is_valid_counted_loop())
+    return false; // Malformed counted loop
 
-  Node *init_n = cl->init_trip();
-  Node *limit_n = cl->limit();
-
-  // Non-constant bounds
-  if (init_n   == NULL || !init_n->is_Con()  ||
-      limit_n  == NULL || !limit_n->is_Con() ||
-      // protect against stride not being a constant
-      !cl->stride_is_con()) {
+  if (!cl->has_exact_trip_count()) {
+    // Trip count is not exact.
     return false;
   }
-  int init   = init_n->get_int();
-  int limit  = limit_n->get_int();
-  int span   = limit - init;
-  int stride = cl->stride_con();
 
-  if (init >= limit || stride > span) {
-    // return a false (no maximally unroll) and the regular unroll/peel
-    // route will make a small mess which CCP will fold away.
-    return false;
-  }
-  uint trip_count = span/stride;   // trip_count can be greater than 2 Gig.
-  assert( (int)trip_count*stride == span, "must divide evenly" );
+  uint trip_count = cl->trip_count();
+  // Note, max_juint is used to indicate unknown trip count.
+  assert(trip_count > 1, "one iteration loop should be optimized out already");
+  assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
 
   // Real policy: if we maximally unroll, does it get too big?
   // Allow the unrolled mess to get larger than standard loop
@@ -427,15 +589,29 @@ bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
   uint body_size    = _body.size();
   uint unroll_limit = (uint)LoopUnrollLimit * 4;
   assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
-  cl->set_trip_count(trip_count);
   if (trip_count > unroll_limit || body_size > unroll_limit) {
     return false;
   }
 
+  // Take into account that after unroll conjoined heads and tails will fold,
+  // otherwise policy_unroll() may allow more unrolling than max unrolling.
+  uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
+  uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
+  if (body_size != tst_body_size) // Check for int overflow
+    return false;
+  if (new_body_size > unroll_limit ||
+      // Unrolling can result in a large amount of node construction
+      new_body_size >= MaxNodeLimit - phase->C->unique()) {
+    return false;
+  }
+
   // Currently we don't have policy to optimize one iteration loops.
   // Maximally unrolling transformation is used for that:
   // it is peeled and the original loop become non reachable (dead).
-  if (trip_count == 1)
+  // Also fully unroll a loop with few iterations regardless next
+  // conditions since following loop optimizations will split
+  // such loop anyway (pre-main-post).
+  if (trip_count <= 3)
     return true;
 
   // Do not unroll a loop with String intrinsics code.
@@ -452,17 +628,7 @@ bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
     } // switch
   }
 
-  if (body_size <= unroll_limit) {
-    uint new_body_size = body_size * trip_count;
-    if (new_body_size <= unroll_limit &&
-        body_size == new_body_size / trip_count &&
-        // Unrolling can result in a large amount of node construction
-        new_body_size < MaxNodeLimit - phase->C->unique()) {
-      return true;    // maximally unroll
-    }
-  }
-
-  return false;               // Do not maximally unroll
+  return true; // Do maximally unroll
 }
 
 
@@ -474,12 +640,15 @@ bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
   CountedLoopNode *cl = _head->as_CountedLoop();
   assert(cl->is_normal_loop() || cl->is_main_loop(), "");
 
-  // protect against stride not being a constant
-  if (!cl->stride_is_con()) return false;
+  if (!cl->is_valid_counted_loop())
+    return false; // Malformed counted loop
 
   // protect against over-unrolling
   if (cl->trip_count() <= 1) return false;
 
+  // Check for stride being a small enough constant
+  if (abs(cl->stride_con()) > (1<<3)) return false;
+
   int future_unroll_ct = cl->unrolled_count() * 2;
 
   // Don't unroll if the next round of unrolling would push us
@@ -560,9 +729,6 @@ bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
     return false;
   }
 
-  // Check for stride being a small enough constant
-  if (abs(cl->stride_con()) > (1<<3)) return false;
-
   // Unroll once!  (Each trip will soon do double iterations)
   return true;
 }
@@ -956,7 +1122,11 @@ void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool ad
     tty->print("Unrolling ");
     loop->dump_head();
   } else if (TraceLoopOpts) {
-    tty->print("Unroll     %d ", loop_head->unrolled_count()*2);
+    if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
+      tty->print("Unroll  %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
+    } else {
+      tty->print("Unroll  %d     ", loop_head->unrolled_count()*2);
+    }
     loop->dump_head();
   }
 #endif
@@ -1631,7 +1801,7 @@ void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
 // have on the last iteration.  This will break the loop.
 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
   // Minimum size must be empty loop
-  if (_body.size() > 7/*number of nodes in an empty loop*/)
+  if (_body.size() > EMPTY_LOOP_SIZE)
     return false;
 
   if (!_head->is_CountedLoop())
@@ -1658,8 +1828,19 @@ bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
   // main and post loops have explicitly created zero trip guard
   bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
   if (needs_guard) {
+    // Skip guard if values not overlap.
+    const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
+    const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
+    int  stride_con = cl->stride_con();
+    if (stride_con > 0) {
+      needs_guard = (init_t->_hi >= limit_t->_lo);
+    } else {
+      needs_guard = (init_t->_lo <= limit_t->_hi);
+    }
+  }
+  if (needs_guard) {
     // Check for an obvious zero trip guard.
-    Node* inctrl = cl->in(LoopNode::EntryControl);
+    Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
     if (inctrl->Opcode() == Op_IfTrue) {
       // The test should look like just the backedge of a CountedLoop
       Node* iff = inctrl->in(0);
@@ -1702,12 +1883,49 @@ bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
   return true;
 }
 
+//------------------------------policy_do_one_iteration_loop-------------------
+// Convert one iteration loop into normal code.
+bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
+  if (!_head->as_Loop()->is_valid_counted_loop())
+    return false; // Only for counted loop
+
+  CountedLoopNode *cl = _head->as_CountedLoop();
+  if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
+    return false;
+  }
+
+#ifndef PRODUCT
+  if(TraceLoopOpts) {
+    tty->print("OneIteration ");
+    this->dump_head();
+  }
+#endif
+
+  Node *init_n = cl->init_trip();
+#ifdef ASSERT
+  // Loop boundaries should be constant since trip count is exact.
+  assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
+#endif
+  // Replace the phi at loop head with the value of the init_trip.
+  // Then the CountedLoopEnd will collapse (backedge will not be taken)
+  // and all loop-invariant uses of the exit values will be correct.
+  phase->_igvn.replace_node(cl->phi(), cl->init_trip());
+  phase->C->set_major_progress();
+  return true;
+}
 
 //=============================================================================
 //------------------------------iteration_split_impl---------------------------
 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
+  // Compute exact loop trip count if possible.
+  compute_exact_trip_count(phase);
+
+  // Convert one iteration loop into normal code.
+  if (policy_do_one_iteration_loop(phase))
+    return true;
+
   // Check and remove empty loops (spam micro-benchmarks)
-  if( policy_do_remove_empty_loop(phase) )
+  if (policy_do_remove_empty_loop(phase))
     return true;  // Here we removed an empty loop
 
   bool should_peel = policy_peeling(phase); // Should we peel?
@@ -1716,40 +1934,40 @@ bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_
 
   // Non-counted loops may be peeled; exactly 1 iteration is peeled.
   // This removes loop-invariant tests (usually null checks).
-  if( !_head->is_CountedLoop() ) { // Non-counted loop
+  if (!_head->is_CountedLoop()) { // Non-counted loop
     if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
       // Partial peel succeeded so terminate this round of loop opts
       return false;
     }
-    if( should_peel ) {            // Should we peel?
+    if (should_peel) {            // Should we peel?
 #ifndef PRODUCT
       if (PrintOpto) tty->print_cr("should_peel");
 #endif
       phase->do_peeling(this,old_new);
-    } else if( should_unswitch ) {
+    } else if (should_unswitch) {
       phase->do_unswitching(this, old_new);
     }
     return true;
   }
   CountedLoopNode *cl = _head->as_CountedLoop();
 
-  if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
+  if (!cl->loopexit()) return true; // Ignore various kinds of broken loops
 
   // Do nothing special to pre- and post- loops
-  if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
+  if (cl->is_pre_loop() || cl->is_post_loop()) return true;
 
   // Compute loop trip count from profile data
   compute_profile_trip_cnt(phase);
 
   // Before attempting fancy unrolling, RCE or alignment, see if we want
   // to completely unroll this loop or do loop unswitching.
-  if( cl->is_normal_loop() ) {
+  if (cl->is_normal_loop()) {
     if (should_unswitch) {
       phase->do_unswitching(this, old_new);
       return true;
     }
     bool should_maximally_unroll =  policy_maximally_unroll(phase);
-    if( should_maximally_unroll ) {
+    if (should_maximally_unroll) {
       // Here we did some unrolling and peeling.  Eventually we will
       // completely unroll this loop and it will no longer be a loop.
       phase->do_maximally_unroll(this,old_new);
@@ -1757,6 +1975,12 @@ bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_
     }
   }
 
+  // Skip next optimizations if running low on nodes. Note that
+  // policy_unswitching and policy_maximally_unroll have this check.
+  uint nodes_left = MaxNodeLimit - phase->C->unique();
+  if ((2 * _body.size()) > nodes_left) {
+    return true;
+  }
 
   // Counted loops may be peeled, may need some iterations run up
   // front for RCE, and may want to align loop refs to a cache
@@ -1787,14 +2011,14 @@ bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_
   // If we have any of these conditions (RCE, alignment, unrolling) met, then
   // we switch to the pre-/main-/post-loop model.  This model also covers
   // peeling.
-  if( should_rce || should_align || should_unroll ) {
-    if( cl->is_normal_loop() )  // Convert to 'pre/main/post' loops
+  if (should_rce || should_align || should_unroll) {
+    if (cl->is_normal_loop())  // Convert to 'pre/main/post' loops
       phase->insert_pre_post_loops(this,old_new, !may_rce_align);
 
     // Adjust the pre- and main-loop limits to let the pre and post loops run
     // with full checks, but the main-loop with no checks.  Remove said
     // checks from the main body.
-    if( should_rce )
+    if (should_rce)
       phase->do_range_check(this,old_new);
 
     // Double loop body for unrolling.  Adjust the minimum-trip test (will do
@@ -1802,16 +2026,16 @@ bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_
     // an even number of trips).  If we are peeling, we might enable some RCE
     // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
     // peeling.
-      if( should_unroll && !should_peel )
-        phase->do_unroll(this,old_new, true);
+    if (should_unroll && !should_peel)
+      phase->do_unroll(this,old_new, true);
 
     // Adjust the pre-loop limits to align the main body
     // iterations.
-    if( should_align )
+    if (should_align)
       Unimplemented();
 
   } else {                      // Else we have an unchanged counted loop
-    if( should_peel )           // Might want to peel but do nothing else
+    if (should_peel)           // Might want to peel but do nothing else
       phase->do_peeling(this,old_new);
   }
   return true;
@@ -1861,651 +2085,8 @@ bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new )
   return true;
 }
 
-//-------------------------------is_uncommon_trap_proj----------------------------
-// Return true if proj is the form of "proj->[region->..]call_uct"
-bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) {
-  int path_limit = 10;
-  assert(proj, "invalid argument");
-  Node* out = proj;
-  for (int ct = 0; ct < path_limit; ct++) {
-    out = out->unique_ctrl_out();
-    if (out == NULL || out->is_Root() || out->is_Start())
-      return false;
-    if (out->is_CallStaticJava()) {
-      int req = out->as_CallStaticJava()->uncommon_trap_request();
-      if (req != 0) {
-        Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
-        if (trap_reason == reason || reason == Deoptimization::Reason_none) {
-           return true;
-        }
-      }
-      return false; // don't do further after call
-    }
-  }
-  return false;
-}
-
-//-------------------------------is_uncommon_trap_if_pattern-------------------------
-// Return true  for "if(test)-> proj -> ...
-//                          |
-//                          V
-//                      other_proj->[region->..]call_uct"
-//
-// "must_reason_predicate" means the uct reason must be Reason_predicate
-bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) {
-  Node *in0 = proj->in(0);
-  if (!in0->is_If()) return false;
-  // Variation of a dead If node.
-  if (in0->outcnt() < 2)  return false;
-  IfNode* iff = in0->as_If();
-
-  // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
-  if (reason != Deoptimization::Reason_none) {
-    if (iff->in(1)->Opcode() != Op_Conv2B ||
-       iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
-      return false;
-    }
-  }
-
-  ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
-  return is_uncommon_trap_proj(other_proj, reason);
-}
-
-//-------------------------------register_control-------------------------
-void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
-  assert(n->is_CFG(), "must be control node");
-  _igvn.register_new_node_with_optimizer(n);
-  loop->_body.push(n);
-  set_loop(n, loop);
-  // When called from beautify_loops() idom is not constructed yet.
-  if (_idom != NULL) {
-    set_idom(n, pred, dom_depth(pred));
-  }
-}
-
-//------------------------------create_new_if_for_predicate------------------------
-// create a new if above the uct_if_pattern for the predicate to be promoted.
-//
-//          before                                after
-//        ----------                           ----------
-//           ctrl                                 ctrl
-//            |                                     |
-//            |                                     |
-//            v                                     v
-//           iff                                 new_iff
-//          /    \                                /      \
-//         /      \                              /        \
-//        v        v                            v          v
-//  uncommon_proj cont_proj                   if_uct     if_cont
-// \      |        |                           |          |
-//  \     |        |                           |          |
-//   v    v        v                           |          v
-//     rgn       loop                          |         iff
-//      |                                      |        /     \
-//      |                                      |       /       \
-//      v                                      |      v         v
-// uncommon_trap                               | uncommon_proj cont_proj
-//                                           \  \    |           |
-//                                            \  \   |           |
-//                                             v  v  v           v
-//                                               rgn           loop
-//                                                |
-//                                                |
-//                                                v
-//                                           uncommon_trap
-//
-//
-// We will create a region to guard the uct call if there is no one there.
-// The true projecttion (if_cont) of the new_iff is returned.
-// This code is also used to clone predicates to clonned loops.
-ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
-                                                      Deoptimization::DeoptReason reason) {
-  assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
-  IfNode* iff = cont_proj->in(0)->as_If();
-
-  ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
-  Node     *rgn   = uncommon_proj->unique_ctrl_out();
-  assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
-
-  if (!rgn->is_Region()) { // create a region to guard the call
-    assert(rgn->is_Call(), "must be call uct");
-    CallNode* call = rgn->as_Call();
-    IdealLoopTree* loop = get_loop(call);
-    rgn = new (C, 1) RegionNode(1);
-    rgn->add_req(uncommon_proj);
-    register_control(rgn, loop, uncommon_proj);
-    _igvn.hash_delete(call);
-    call->set_req(0, rgn);
-    // When called from beautify_loops() idom is not constructed yet.
-    if (_idom != NULL) {
-      set_idom(call, rgn, dom_depth(rgn));
-    }
-  }
-
-  Node* entry = iff->in(0);
-  if (new_entry != NULL) {
-    // Clonning the predicate to new location.
-    entry = new_entry;
-  }
-  // Create new_iff
-  IdealLoopTree* lp = get_loop(entry);
-  IfNode *new_iff = new (C, 2) IfNode(entry, NULL, iff->_prob, iff->_fcnt);
-  register_control(new_iff, lp, entry);
-  Node *if_cont = new (C, 1) IfTrueNode(new_iff);
-  Node *if_uct  = new (C, 1) IfFalseNode(new_iff);
-  if (cont_proj->is_IfFalse()) {
-    // Swap
-    Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
-  }
-  register_control(if_cont, lp, new_iff);
-  register_control(if_uct, get_loop(rgn), new_iff);
-
-  // if_uct to rgn
-  _igvn.hash_delete(rgn);
-  rgn->add_req(if_uct);
-  // When called from beautify_loops() idom is not constructed yet.
-  if (_idom != NULL) {
-    Node* ridom = idom(rgn);
-    Node* nrdom = dom_lca(ridom, new_iff);
-    set_idom(rgn, nrdom, dom_depth(rgn));
-  }
-  // rgn must have no phis
-  assert(!rgn->as_Region()->has_phi(), "region must have no phis");
-
-  if (new_entry == NULL) {
-    // Attach if_cont to iff
-    _igvn.hash_delete(iff);
-    iff->set_req(0, if_cont);
-    if (_idom != NULL) {
-      set_idom(iff, if_cont, dom_depth(iff));
-    }
-  }
-  return if_cont->as_Proj();
-}
-
-//--------------------------find_predicate_insertion_point-------------------
-// Find a good location to insert a predicate
-ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
-  if (start_c == NULL || !start_c->is_Proj())
-    return NULL;
-  if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
-    return start_c->as_Proj();
-  }
-  return NULL;
-}
-
-//--------------------------find_predicate------------------------------------
-// Find a predicate
-Node* PhaseIdealLoop::find_predicate(Node* entry) {
-  Node* predicate = NULL;
-  if (UseLoopPredicate) {
-    predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
-    if (predicate != NULL) { // right pattern that can be used by loop predication
-      assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
-      return entry;
-    }
-  }
-  return NULL;
-}
-
-//------------------------------Invariance-----------------------------------
-// Helper class for loop_predication_impl to compute invariance on the fly and
-// clone invariants.
-class Invariance : public StackObj {
-  VectorSet _visited, _invariant;
-  Node_Stack _stack;
-  VectorSet _clone_visited;
-  Node_List _old_new; // map of old to new (clone)
-  IdealLoopTree* _lpt;
-  PhaseIdealLoop* _phase;
-
-  // Helper function to set up the invariance for invariance computation
-  // If n is a known invariant, set up directly. Otherwise, look up the
-  // the possibility to push n onto the stack for further processing.
-  void visit(Node* use, Node* n) {
-    if (_lpt->is_invariant(n)) { // known invariant
-      _invariant.set(n->_idx);
-    } else if (!n->is_CFG()) {
-      Node *n_ctrl = _phase->ctrl_or_self(n);
-      Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
-      if (_phase->is_dominator(n_ctrl, u_ctrl)) {
-        _stack.push(n, n->in(0) == NULL ? 1 : 0);
-      }
-    }
-  }
-
-  // Compute invariance for "the_node" and (possibly) all its inputs recursively
-  // on the fly
-  void compute_invariance(Node* n) {
-    assert(_visited.test(n->_idx), "must be");
-    visit(n, n);
-    while (_stack.is_nonempty()) {
-      Node*  n = _stack.node();
-      uint idx = _stack.index();
-      if (idx == n->req()) { // all inputs are processed
-        _stack.pop();
-        // n is invariant if it's inputs are all invariant
-        bool all_inputs_invariant = true;
-        for (uint i = 0; i < n->req(); i++) {
-          Node* in = n->in(i);
-          if (in == NULL) continue;
-          assert(_visited.test(in->_idx), "must have visited input");
-          if (!_invariant.test(in->_idx)) { // bad guy
-            all_inputs_invariant = false;
-            break;
-          }
-        }
-        if (all_inputs_invariant) {
-          _invariant.set(n->_idx); // I am a invariant too
-        }
-      } else { // process next input
-        _stack.set_index(idx + 1);
-        Node* m = n->in(idx);
-        if (m != NULL && !_visited.test_set(m->_idx)) {
-          visit(n, m);
-        }
-      }
-    }
-  }
-
-  // Helper function to set up _old_new map for clone_nodes.
-  // If n is a known invariant, set up directly ("clone" of n == n).
-  // Otherwise, push n onto the stack for real cloning.
-  void clone_visit(Node* n) {
-    assert(_invariant.test(n->_idx), "must be invariant");
-    if (_lpt->is_invariant(n)) { // known invariant
-      _old_new.map(n->_idx, n);
-    } else{ // to be cloned
-      assert (!n->is_CFG(), "should not see CFG here");
-      _stack.push(n, n->in(0) == NULL ? 1 : 0);
-    }
-  }
-
-  // Clone "n" and (possibly) all its inputs recursively
-  void clone_nodes(Node* n, Node* ctrl) {
-    clone_visit(n);
-    while (_stack.is_nonempty()) {
-      Node*  n = _stack.node();
-      uint idx = _stack.index();
-      if (idx == n->req()) { // all inputs processed, clone n!
-        _stack.pop();
-        // clone invariant node
-        Node* n_cl = n->clone();
-        _old_new.map(n->_idx, n_cl);
-        _phase->register_new_node(n_cl, ctrl);
-        for (uint i = 0; i < n->req(); i++) {
-          Node* in = n_cl->in(i);
-          if (in == NULL) continue;
-          n_cl->set_req(i, _old_new[in->_idx]);
-        }
-      } else { // process next input
-        _stack.set_index(idx + 1);
-        Node* m = n->in(idx);
-        if (m != NULL && !_clone_visited.test_set(m->_idx)) {
-          clone_visit(m); // visit the input
-        }
-      }
-    }
-  }
-
- public:
-  Invariance(Arena* area, IdealLoopTree* lpt) :
-    _lpt(lpt), _phase(lpt->_phase),
-    _visited(area), _invariant(area), _stack(area, 10 /* guess */),
-    _clone_visited(area), _old_new(area)
-  {}
-
-  // Map old to n for invariance computation and clone
-  void map_ctrl(Node* old, Node* n) {
-    assert(old->is_CFG() && n->is_CFG(), "must be");
-    _old_new.map(old->_idx, n); // "clone" of old is n
-    _invariant.set(old->_idx);  // old is invariant
-    _clone_visited.set(old->_idx);
-  }
-
-  // Driver function to compute invariance
-  bool is_invariant(Node* n) {
-    if (!_visited.test_set(n->_idx))
-      compute_invariance(n);
-    return (_invariant.test(n->_idx) != 0);
-  }
-
-  // Driver function to clone invariant
-  Node* clone(Node* n, Node* ctrl) {
-    assert(ctrl->is_CFG(), "must be");
-    assert(_invariant.test(n->_idx), "must be an invariant");
-    if (!_clone_visited.test(n->_idx))
-      clone_nodes(n, ctrl);
-    return _old_new[n->_idx];
-  }
-};
-
-//------------------------------is_range_check_if -----------------------------------
-// Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
-// Note: this function is particularly designed for loop predication. We require load_range
-//       and offset to be loop invariant computed on the fly by "invar"
-bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
-  if (!is_loop_exit(iff)) {
-    return false;
-  }
-  if (!iff->in(1)->is_Bool()) {
-    return false;
-  }
-  const BoolNode *bol = iff->in(1)->as_Bool();
-  if (bol->_test._test != BoolTest::lt) {
-    return false;
-  }
-  if (!bol->in(1)->is_Cmp()) {
-    return false;
-  }
-  const CmpNode *cmp = bol->in(1)->as_Cmp();
-  if (cmp->Opcode() != Op_CmpU ) {
-    return false;
-  }
-  Node* range = cmp->in(2);
-  if (range->Opcode() != Op_LoadRange) {
-    const TypeInt* tint = phase->_igvn.type(range)->isa_int();
-    if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) {
-      // Allow predication on positive values that aren't LoadRanges.
-      // This allows optimization of loops where the length of the
-      // array is a known value and doesn't need to be loaded back
-      // from the array.
-      return false;
-    }
-  }
-  if (!invar.is_invariant(range)) {
-    return false;
-  }
-  Node *iv     = _head->as_CountedLoop()->phi();
-  int   scale  = 0;
-  Node *offset = NULL;
-  if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
-    return false;
-  }
-  if(offset && !invar.is_invariant(offset)) { // offset must be invariant
-    return false;
-  }
-  return true;
-}
-
-//------------------------------rc_predicate-----------------------------------
-// Create a range check predicate
-//
-// for (i = init; i < limit; i += stride) {
-//    a[scale*i+offset]
-// }
-//
-// Compute max(scale*i + offset) for init <= i < limit and build the predicate
-// as "max(scale*i + offset) u< a.length".
-//
-// There are two cases for max(scale*i + offset):
-// (1) stride*scale > 0
-//   max(scale*i + offset) = scale*(limit-stride) + offset
-// (2) stride*scale < 0
-//   max(scale*i + offset) = scale*init + offset
-BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl,
-                                       int scale, Node* offset,
-                                       Node* init, Node* limit, Node* stride,
-                                       Node* range, bool upper) {
-  DEBUG_ONLY(ttyLocker ttyl);
-  if (TraceLoopPredicate) tty->print("rc_predicate ");
-
-  Node* max_idx_expr  = init;
-  int stride_con = stride->get_int();
-  if ((stride_con > 0) == (scale > 0) == upper) {
-    max_idx_expr = new (C, 3) SubINode(limit, stride);
-    register_new_node(max_idx_expr, ctrl);
-    if (TraceLoopPredicate) tty->print("(limit - stride) ");
-  } else {
-    if (TraceLoopPredicate) tty->print("init ");
-  }
-
-  if (scale != 1) {
-    ConNode* con_scale = _igvn.intcon(scale);
-    max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
-    register_new_node(max_idx_expr, ctrl);
-    if (TraceLoopPredicate) tty->print("* %d ", scale);
-  }
-
-  if (offset && (!offset->is_Con() || offset->get_int() != 0)){
-    max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
-    register_new_node(max_idx_expr, ctrl);
-    if (TraceLoopPredicate)
-      if (offset->is_Con()) tty->print("+ %d ", offset->get_int());
-      else tty->print("+ offset ");
-  }
-
-  CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
-  register_new_node(cmp, ctrl);
-  BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
-  register_new_node(bol, ctrl);
-
-  if (TraceLoopPredicate) tty->print_cr("<u range");
-  return bol;
-}
-
-//------------------------------ loop_predication_impl--------------------------
-// Insert loop predicates for null checks and range checks
-bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
-  if (!UseLoopPredicate) return false;
-
-  if (!loop->_head->is_Loop()) {
-    // Could be a simple region when irreducible loops are present.
-    return false;
-  }
-
-  if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
-    // do nothing for infinite loops
-    return false;
-  }
-
-  CountedLoopNode *cl = NULL;
-  if (loop->_head->is_CountedLoop()) {
-    cl = loop->_head->as_CountedLoop();
-    // do nothing for iteration-splitted loops
-    if (!cl->is_normal_loop()) return false;
-  }
-
-  LoopNode *lpn  = loop->_head->as_Loop();
-  Node* entry = lpn->in(LoopNode::EntryControl);
-
-  ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
-  if (!predicate_proj) {
-#ifndef PRODUCT
-    if (TraceLoopPredicate) {
-      tty->print("missing predicate:");
-      loop->dump_head();
-      lpn->dump(1);
-    }
-#endif
-    return false;
-  }
-  ConNode* zero = _igvn.intcon(0);
-  set_ctrl(zero, C->root());
-
-  ResourceArea *area = Thread::current()->resource_area();
-  Invariance invar(area, loop);
-
-  // Create list of if-projs such that a newer proj dominates all older
-  // projs in the list, and they all dominate loop->tail()
-  Node_List if_proj_list(area);
-  LoopNode *head  = loop->_head->as_Loop();
-  Node *current_proj = loop->tail(); //start from tail
-  while ( current_proj != head ) {
-    if (loop == get_loop(current_proj) && // still in the loop ?
-        current_proj->is_Proj()        && // is a projection  ?
-        current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
-      if_proj_list.push(current_proj);
-    }
-    current_proj = idom(current_proj);
-  }
-
-  bool hoisted = false; // true if at least one proj is promoted
-  while (if_proj_list.size() > 0) {
-    // Following are changed to nonnull when a predicate can be hoisted
-    ProjNode* new_predicate_proj = NULL;
-
-    ProjNode* proj = if_proj_list.pop()->as_Proj();
-    IfNode*   iff  = proj->in(0)->as_If();
-
-    if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
-      if (loop->is_loop_exit(iff)) {
-        // stop processing the remaining projs in the list because the execution of them
-        // depends on the condition of "iff" (iff->in(1)).
-        break;
-      } else {
-        // Both arms are inside the loop. There are two cases:
-        // (1) there is one backward branch. In this case, any remaining proj
-        //     in the if_proj list post-dominates "iff". So, the condition of "iff"
-        //     does not determine the execution the remining projs directly, and we
-        //     can safely continue.
-        // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
-        //     does not dominate loop->tail(), so it can not be in the if_proj list.
-        continue;
-      }
-    }
-
-    Node*     test = iff->in(1);
-    if (!test->is_Bool()){ //Conv2B, ...
-      continue;
-    }
-    BoolNode* bol = test->as_Bool();
-    if (invar.is_invariant(bol)) {
-      // Invariant test
-      new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
-                                                       Deoptimization::Reason_predicate);
-      Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
-      BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
-
-      // Negate test if necessary
-      bool negated = false;
-      if (proj->_con != predicate_proj->_con) {
-        new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
-        register_new_node(new_predicate_bol, ctrl);
-        negated = true;
-      }
-      IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
-      _igvn.hash_delete(new_predicate_iff);
-      new_predicate_iff->set_req(1, new_predicate_bol);
-#ifndef PRODUCT
-      if (TraceLoopPredicate) {
-        tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
-        loop->dump_head();
-      } else if (TraceLoopOpts) {
-        tty->print("Predicate IC ");
-        loop->dump_head();
-      }
-#endif
-    } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
-      assert(proj->_con == predicate_proj->_con, "must match");
-
-      // Range check for counted loops
-      const Node*    cmp    = bol->in(1)->as_Cmp();
-      Node*          idx    = cmp->in(1);
-      assert(!invar.is_invariant(idx), "index is variant");
-      assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be");
-      Node* rng = cmp->in(2);
-      assert(invar.is_invariant(rng), "range must be invariant");
-      int scale    = 1;
-      Node* offset = zero;
-      bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
-      assert(ok, "must be index expression");
-
-      Node* init    = cl->init_trip();
-      Node* limit   = cl->limit();
-      Node* stride  = cl->stride();
-
-      // Build if's for the upper and lower bound tests.  The
-      // lower_bound test will dominate the upper bound test and all
-      // cloned or created nodes will use the lower bound test as
-      // their declared control.
-      ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
-      ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
-      assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
-      Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
-
-      // Perform cloning to keep Invariance state correct since the
-      // late schedule will place invariant things in the loop.
-      rng = invar.clone(rng, ctrl);
-      if (offset && offset != zero) {
-        assert(invar.is_invariant(offset), "offset must be loop invariant");
-        offset = invar.clone(offset, ctrl);
-      }
-
-      // Test the lower bound
-      Node*  lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false);
-      IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
-      _igvn.hash_delete(lower_bound_iff);
-      lower_bound_iff->set_req(1, lower_bound_bol);
-      if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
-
-      // Test the upper bound
-      Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true);
-      IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
-      _igvn.hash_delete(upper_bound_iff);
-      upper_bound_iff->set_req(1, upper_bound_bol);
-      if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
-
-      // Fall through into rest of the clean up code which will move
-      // any dependent nodes onto the upper bound test.
-      new_predicate_proj = upper_bound_proj;
-
-#ifndef PRODUCT
-      if (TraceLoopOpts && !TraceLoopPredicate) {
-        tty->print("Predicate RC ");
-        loop->dump_head();
-      }
-#endif
-    } else {
-      // Loop variant check (for example, range check in non-counted loop)
-      // with uncommon trap.
-      continue;
-    }
-    assert(new_predicate_proj != NULL, "sanity");
-    // Success - attach condition (new_predicate_bol) to predicate if
-    invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
-
-    // Eliminate the old If in the loop body
-    dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
-
-    hoisted = true;
-    C->set_major_progress();
-  } // end while
-
-#ifndef PRODUCT
-  // report that the loop predication has been actually performed
-  // for this loop
-  if (TraceLoopPredicate && hoisted) {
-    tty->print("Loop Predication Performed:");
-    loop->dump_head();
-  }
-#endif
-
-  return hoisted;
-}
-
-//------------------------------loop_predication--------------------------------
-// driver routine for loop predication optimization
-bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
-  bool hoisted = false;
-  // Recursively promote predicates
-  if ( _child ) {
-    hoisted = _child->loop_predication( phase);
-  }
-
-  // self
-  if (!_irreducible && !tail()->is_top()) {
-    hoisted |= phase->loop_predication_impl(this);
-  }
-
-  if ( _next ) { //sibling
-    hoisted |= _next->loop_predication( phase);
-  }
-
-  return hoisted;
-}
-
 
+//=============================================================================
 // Process all the loops in the loop tree and replace any fill
 // patterns with an intrisc version.
 bool PhaseIdealLoop::do_intrinsify_fill() {
@@ -2625,9 +2206,12 @@ bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& st
       if (value != head->phi()) {
         msg = "unhandled shift in address";
       } else {
-        found_index = true;
-        shift = n;
-        assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match");
+        if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
+          msg = "scale doesn't match";
+        } else {
+          found_index = true;
+          shift = n;
+        }
       }
     } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
       if (n->in(1) == head->phi()) {
@@ -2762,6 +2346,13 @@ bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
     return false;
   }
 
+#ifndef PRODUCT
+  if (TraceLoopOpts) {
+    tty->print("ArrayFill    ");
+    lpt->dump_head();
+  }
+#endif
+
   // Now replace the whole loop body by a call to a fill routine that
   // covers the same region as the loop.
   Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
diff --git a/src/share/vm/opto/loopUnswitch.cpp b/src/share/vm/opto/loopUnswitch.cpp
index 544b845f8..8a103b228 100644
--- a/src/share/vm/opto/loopUnswitch.cpp
+++ b/src/share/vm/opto/loopUnswitch.cpp
@@ -32,15 +32,17 @@
 //
 // orig:                       transformed:
 //                               if (invariant-test) then
+//  predicate                      predicate
 //  loop                           loop
 //    stmt1                          stmt1
 //    if (invariant-test) then       stmt2
 //      stmt2                        stmt4
 //    else                         endloop
 //      stmt3                    else
-//    endif                        loop [clone]
-//    stmt4                          stmt1 [clone]
-//  endloop                          stmt3
+//    endif                        predicate [clone]
+//    stmt4                        loop [clone]
+//  endloop                          stmt1 [clone]
+//                                   stmt3
 //                                   stmt4 [clone]
 //                                 endloop
 //                               endif
@@ -124,8 +126,15 @@ void PhaseIdealLoop::do_unswitching (IdealLoopTree *loop, Node_List &old_new) {
 
   ProjNode* proj_true = create_slow_version_of_loop(loop, old_new);
 
-  assert(proj_true->is_IfTrue() && proj_true->unique_ctrl_out() == head, "by construction");
-
+#ifdef ASSERT
+  Node* uniqc = proj_true->unique_ctrl_out();
+  Node* entry = head->in(LoopNode::EntryControl);
+  Node* predicate = find_predicate(entry);
+  if (predicate != NULL) predicate = predicate->in(0);
+  assert(proj_true->is_IfTrue() &&
+         (predicate == NULL && uniqc == head ||
+          predicate != NULL && uniqc == predicate), "by construction");
+#endif
   // Increment unswitch count
   LoopNode* head_clone = old_new[head->_idx]->as_Loop();
   int nct = head->unswitch_count() + 1;
@@ -227,21 +236,24 @@ ProjNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
   register_node(ifslow, outer_loop, iff, dom_depth(iff));
 
   // Clone the loop body.  The clone becomes the fast loop.  The
-  // original pre-header will (illegally) have 2 control users (old & new loops).
+  // original pre-header will (illegally) have 3 control users
+  // (old & new loops & new if).
   clone_loop(loop, old_new, dom_depth(head), iff);
   assert(old_new[head->_idx]->is_Loop(), "" );
 
   // Fast (true) control
+  Node* iffast_pred = clone_loop_predicates(entry, iffast);
   _igvn.hash_delete(head);
-  head->set_req(LoopNode::EntryControl, iffast);
-  set_idom(head, iffast, dom_depth(head));
+  head->set_req(LoopNode::EntryControl, iffast_pred);
+  set_idom(head, iffast_pred, dom_depth(head));
   _igvn._worklist.push(head);
 
   // Slow (false) control
+  Node* ifslow_pred = move_loop_predicates(entry, ifslow);
   LoopNode* slow_head = old_new[head->_idx]->as_Loop();
   _igvn.hash_delete(slow_head);
-  slow_head->set_req(LoopNode::EntryControl, ifslow);
-  set_idom(slow_head, ifslow, dom_depth(slow_head));
+  slow_head->set_req(LoopNode::EntryControl, ifslow_pred);
+  set_idom(slow_head, ifslow_pred, dom_depth(slow_head));
   _igvn._worklist.push(slow_head);
 
   recompute_dom_depth();
diff --git a/src/share/vm/opto/loopnode.cpp b/src/share/vm/opto/loopnode.cpp
index 894a105e8..cd6948712 100644
--- a/src/share/vm/opto/loopnode.cpp
+++ b/src/share/vm/opto/loopnode.cpp
@@ -341,7 +341,12 @@ bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
   //
   assert(x->Opcode() == Op_Loop, "regular loops only");
   C->print_method("Before CountedLoop", 3);
-
+#ifndef PRODUCT
+  if (TraceLoopOpts) {
+    tty->print("Counted      ");
+    loop->dump_head();
+  }
+#endif
   // If compare points to incr, we are ok.  Otherwise the compare
   // can directly point to the phi; in this case adjust the compare so that
   // it points to the incr by adjusting the limit.
@@ -864,8 +869,10 @@ void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
   Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
   outer = igvn.register_new_node_with_optimizer(outer, _head);
   phase->set_created_loop_node();
+
+  Node* pred = phase->clone_loop_predicates(ctl, outer);
   // Outermost loop falls into '_head' loop
-  _head->set_req(LoopNode::EntryControl, outer);
+  _head->set_req(LoopNode::EntryControl, pred);
   _head->del_req(outer_idx);
   // Split all the Phis up between '_head' loop and 'outer' loop.
   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
@@ -1103,12 +1110,13 @@ bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
   // backedges into a private merge point and use the merge point as
   // the one true backedge.
   if( _head->req() > 3 ) {
-    // Merge the many backedges into a single backedge.
+    // Merge the many backedges into a single backedge but leave
+    // the hottest backedge as separate edge for the following peel.
     merge_many_backedges( phase );
     result = true;
   }
 
-  // If I am a shared header (multiple backedges), peel off myself loop.
+  // If I have one hot backedge, peel off myself loop.
   // I better be the outermost loop.
   if( _head->req() > 3 ) {
     split_outer_loop( phase );
@@ -1433,15 +1441,30 @@ void IdealLoopTree::dump_head( ) const {
   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
   if (_irreducible) tty->print(" IRREDUCIBLE");
   if (UseLoopPredicate) {
-    Node* entry = _head->in(LoopNode::EntryControl);
-    if (entry != NULL && entry->is_Proj() &&
-        PhaseIdealLoop::is_uncommon_trap_if_pattern(entry->as_Proj(), Deoptimization::Reason_predicate)) {
+    Node* entry = PhaseIdealLoop::find_predicate_insertion_point(_head->in(LoopNode::EntryControl),
+                                                                 Deoptimization::Reason_predicate);
+    if (entry != NULL) {
       tty->print(" predicated");
     }
   }
   if (_head->is_CountedLoop()) {
     CountedLoopNode *cl = _head->as_CountedLoop();
     tty->print(" counted");
+
+    Node* init_n = cl->init_trip();
+    if (init_n  != NULL &&  init_n->is_Con())
+      tty->print(" [%d,", cl->init_trip()->get_int());
+    else
+      tty->print(" [int,");
+    Node* limit_n = cl->limit();
+    if (limit_n  != NULL &&  limit_n->is_Con())
+      tty->print("%d),", cl->limit()->get_int());
+    else
+      tty->print("int),");
+    int stride_con  = cl->stride_con();
+    if (stride_con > 0) tty->print("+");
+    tty->print("%d", stride_con);
+
     if (cl->is_pre_loop ()) tty->print(" pre" );
     if (cl->is_main_loop()) tty->print(" main");
     if (cl->is_post_loop()) tty->print(" post");
@@ -1541,7 +1564,7 @@ void PhaseIdealLoop::eliminate_useless_predicates() {
 //----------------------------build_and_optimize-------------------------------
 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.
-void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
+void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
   ResourceMark rm;
 
   int old_progress = C->major_progress();
@@ -1573,6 +1596,13 @@ void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
   // Do not need a safepoint at the top level
   _ltree_root->_has_sfpt = 1;
 
+  // Initialize Dominators.
+  // Checked in clone_loop_predicate() during beautify_loops().
+  _idom_size = 0;
+  _idom      = NULL;
+  _dom_depth = NULL;
+  _dom_stk   = NULL;
+
   // Empty pre-order array
   allocate_preorders();
 
@@ -1698,8 +1728,9 @@ void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
     return;
   }
 
-  // some parser-inserted loop predicates could never be used by loop
-  // predication. Eliminate them before loop optimization
+  // Some parser-inserted loop predicates could never be used by loop
+  // predication or they were moved away from loop during some optimizations.
+  // For example, peeling. Eliminate them before next loop optimizations.
   if (UseLoopPredicate) {
     eliminate_useless_predicates();
   }
@@ -1750,7 +1781,7 @@ void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
   }
 
   // Perform loop predication before iteration splitting
-  if (do_loop_pred && C->has_loops() && !C->major_progress()) {
+  if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
     _ltree_root->_child->loop_predication(this);
   }
 
@@ -1793,8 +1824,20 @@ void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
     C->set_major_progress();
   }
 
-  // Convert scalar to superword operations
+  // Keep loop predicates and perform optimizations with them
+  // until no more loop optimizations could be done.
+  // After that switch predicates off and do more loop optimizations.
+  if (!C->major_progress() && (C->predicate_count() > 0)) {
+     C->cleanup_loop_predicates(_igvn);
+#ifndef PRODUCT
+     if (TraceLoopOpts) {
+       tty->print_cr("PredicatesOff");
+     }
+#endif
+     C->set_major_progress();
+  }
 
+  // Convert scalar to superword operations at the end of all loop opts.
   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
     // SuperWord transform
     SuperWord sw(this);
diff --git a/src/share/vm/opto/loopnode.hpp b/src/share/vm/opto/loopnode.hpp
index 9edcdf292..2b6ea369a 100644
--- a/src/share/vm/opto/loopnode.hpp
+++ b/src/share/vm/opto/loopnode.hpp
@@ -57,7 +57,12 @@ class LoopNode : public RegionNode {
 protected:
   short _loop_flags;
   // Names for flag bitfields
-  enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32  };
+  enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
+         MainHasNoPreLoop=4,
+         HasExactTripCount=8,
+         InnerLoop=16,
+         PartialPeelLoop=32,
+         PartialPeelFailed=64 };
   char _unswitch_count;
   enum { _unswitch_max=3 };
 
@@ -65,13 +70,13 @@ public:
   // Names for edge indices
   enum { Self=0, EntryControl, LoopBackControl };
 
-  int is_inner_loop() const { return _loop_flags & inner_loop; }
-  void set_inner_loop() { _loop_flags |= inner_loop; }
+  int is_inner_loop() const { return _loop_flags & InnerLoop; }
+  void set_inner_loop() { _loop_flags |= InnerLoop; }
 
-  int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
-  void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
-  int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
-  void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
+  int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
+  void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
+  int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
+  void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
 
   int unswitch_max() { return _unswitch_max; }
   int unswitch_count() { return _unswitch_count; }
@@ -137,8 +142,8 @@ class CountedLoopNode : public LoopNode {
   // the Main CountedLoop.  Used to assert that we understand the graph shape.
   node_idx_t _main_idx;
 
-  // Known trip count calculated by policy_maximally_unroll
-  int   _trip_count;
+  // Known trip count calculated by compute_exact_trip_count()
+  uint  _trip_count;
 
   // Expected trip count from profile data
   float _profile_trip_cnt;
@@ -152,7 +157,7 @@ class CountedLoopNode : public LoopNode {
 
 public:
   CountedLoopNode( Node *entry, Node *backedge )
-    : LoopNode(entry, backedge), _trip_count(max_jint),
+    : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
       _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
       _node_count_before_unroll(0) {
     init_class_id(Class_CountedLoop);
@@ -194,13 +199,12 @@ public:
 
   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
   // Aligned, may be missing it's pre-loop.
-  enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
-  int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
-  int is_pre_loop   () const { return (_loop_flags&PrePostFlagsMask) == Pre;    }
-  int is_main_loop  () const { return (_loop_flags&PrePostFlagsMask) == Main;   }
-  int is_post_loop  () const { return (_loop_flags&PrePostFlagsMask) == Post;   }
-  int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
-  void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
+  int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
+  int is_pre_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
+  int is_main_loop  () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
+  int is_post_loop  () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
+  int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
+  void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 
   int main_idx() const { return _main_idx; }
 
@@ -208,10 +212,19 @@ public:
   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
-  void set_normal_loop(                    ) { _loop_flags &= ~PrePostFlagsMask; }
+  void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 
-  void set_trip_count(int tc) { _trip_count = tc; }
-  int trip_count()            { return _trip_count; }
+  void set_trip_count(uint tc) { _trip_count = tc; }
+  uint trip_count()            { return _trip_count; }
+
+  bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
+  void set_exact_trip_count(uint tc) {
+    _trip_count = tc;
+    _loop_flags |= HasExactTripCount;
+  }
+  void set_nonexact_trip_count() {
+    _loop_flags &= ~HasExactTripCount;
+  }
 
   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
   float profile_trip_cnt()             { return _profile_trip_cnt; }
@@ -384,6 +397,9 @@ public:
   // Micro-benchmark spamming.  Remove empty loops.
   bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
 
+  // Convert one iteration loop into normal code.
+  bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
+
   // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
   // make some loop-invariant test (usually a null-check) happen before the
   // loop.
@@ -412,6 +428,9 @@ public:
   // Return TRUE if "iff" is a range check.
   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
 
+  // Compute loop exact trip count if possible
+  void compute_exact_trip_count( PhaseIdealLoop *phase );
+
   // Compute loop trip count from profile data
   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 
@@ -706,11 +725,11 @@ private:
     _dom_lca_tags(arena()), // Thread::resource_area
     _verify_me(NULL),
     _verify_only(true) {
-    build_and_optimize(false, false);
+    build_and_optimize(false);
   }
 
   // build the loop tree and perform any requested optimizations
-  void build_and_optimize(bool do_split_if, bool do_loop_pred);
+  void build_and_optimize(bool do_split_if);
 
 public:
   // Dominators for the sea of nodes
@@ -721,13 +740,13 @@ public:
   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
 
   // Compute the Ideal Node to Loop mapping
-  PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) :
+  PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs) :
     PhaseTransform(Ideal_Loop),
     _igvn(igvn),
     _dom_lca_tags(arena()), // Thread::resource_area
     _verify_me(NULL),
     _verify_only(false) {
-    build_and_optimize(do_split_ifs, do_loop_pred);
+    build_and_optimize(do_split_ifs);
   }
 
   // Verify that verify_me made the same decisions as a fresh run.
@@ -737,7 +756,7 @@ public:
     _dom_lca_tags(arena()), // Thread::resource_area
     _verify_me(verify_me),
     _verify_only(false) {
-    build_and_optimize(false, false);
+    build_and_optimize(false);
   }
 
   // Build and verify the loop tree without modifying the graph.  This
@@ -830,7 +849,26 @@ public:
                                         Deoptimization::DeoptReason reason);
   void register_control(Node* n, IdealLoopTree *loop, Node* pred);
 
-   // Find a good location to insert a predicate
+  // Clone loop predicates to cloned loops (peeled, unswitched)
+  static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
+                                   Deoptimization::DeoptReason reason,
+                                   PhaseIdealLoop* loop_phase,
+                                   PhaseIterGVN* igvn);
+  static ProjNode*  move_predicate(ProjNode* predicate_proj, Node* new_entry,
+                                   Deoptimization::DeoptReason reason,
+                                   PhaseIdealLoop* loop_phase,
+                                   PhaseIterGVN* igvn);
+  static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
+                                         bool move_predicates,
+                                         PhaseIdealLoop* loop_phase,
+                                         PhaseIterGVN* igvn);
+  Node* clone_loop_predicates(Node* old_entry, Node* new_entry);
+  Node*  move_loop_predicates(Node* old_entry, Node* new_entry);
+
+  void eliminate_loop_predicates(Node* entry);
+  static Node* skip_loop_predicates(Node* entry);
+
+  // Find a good location to insert a predicate
   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
   // Find a predicate
   static Node* find_predicate(Node* entry);
diff --git a/src/share/vm/opto/loopopts.cpp b/src/share/vm/opto/loopopts.cpp
index 831c6cb6e..1be374206 100644
--- a/src/share/vm/opto/loopopts.cpp
+++ b/src/share/vm/opto/loopopts.cpp
@@ -2139,9 +2139,12 @@ bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& p
 //
 //                   orig
 //
-//                  stmt1
-//                    |
-//                    v
+//                   stmt1
+//                     |
+//                     v
+//               loop predicate
+//                     |
+//                     v
 //                   loop<----+
 //                     |      |
 //                   stmt2    |
@@ -2172,6 +2175,9 @@ bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& p
 //            after clone loop
 //
 //                   stmt1
+//                     |
+//                     v
+//               loop predicate
 //                 /       \
 //        clone   /         \   orig
 //               /           \
@@ -2210,12 +2216,15 @@ bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& p
 //           after partial peel
 //
 //                  stmt1
+//                     |
+//                     v
+//               loop predicate
 //                 /
 //        clone   /             orig
 //               /          TOP
 //              /             \
 //             v               v
-//    TOP->region             region----+
+//    TOP->loop                loop----+
 //          |                    |      |
 //        stmt2                stmt2    |
 //          |                    |      |
@@ -2253,13 +2262,17 @@ bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& p
 //                  stmt1
 //                    |
 //                    v
+//                  stmt2 clone
+//                    |
+//                    v
 //         ........> ifA clone
 //         :        / |
 //        dom      /  |
 //         :      v   v
 //         :  false   true
 //         :  |       |
-//         :  |     stmt2 clone
+//         :  |       v
+//         :  | loop predicate
 //         :  |       |
 //         :  |       v
 //         :  |    newloop<-----+
@@ -2289,6 +2302,7 @@ bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& p
 //
 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
 
+  assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
   if (!loop->_head->is_Loop()) {
     return false;  }
 
@@ -2316,6 +2330,7 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
     }
   }
 
+  Node* entry = head->in(LoopNode::EntryControl);
   int dd = dom_depth(head);
 
   // Step 1: find cut point
@@ -2612,6 +2627,8 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
 
   // Backedge of the surviving new_head (the clone) is original last_peel
   _igvn.hash_delete(new_head_clone);
+  Node* new_entry = move_loop_predicates(entry, new_head_clone->in(LoopNode::EntryControl));
+  new_head_clone->set_req(LoopNode::EntryControl, new_entry);
   new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
   _igvn._worklist.push(new_head_clone);
 
diff --git a/src/share/vm/opto/phaseX.hpp b/src/share/vm/opto/phaseX.hpp
index 7f1060c05..4d19708ed 100644
--- a/src/share/vm/opto/phaseX.hpp
+++ b/src/share/vm/opto/phaseX.hpp
@@ -471,6 +471,13 @@ public:
     _delay_transform = delay;
   }
 
+  // Clone loop predicates. Defined in loopTransform.cpp.
+  Node* clone_loop_predicates(Node* old_entry, Node* new_entry);
+  Node*  move_loop_predicates(Node* old_entry, Node* new_entry);
+  // Create a new if below new_entry for the predicate to be cloned
+  ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
+                                        Deoptimization::DeoptReason reason);
+
 #ifndef PRODUCT
 protected:
   // Sub-quadratic implementation of VerifyIterativeGVN.
diff --git a/src/share/vm/opto/split_if.cpp b/src/share/vm/opto/split_if.cpp
index 91792368f..dba6ed1f2 100644
--- a/src/share/vm/opto/split_if.cpp
+++ b/src/share/vm/opto/split_if.cpp
@@ -399,6 +399,9 @@ void PhaseIdealLoop::do_split_if( Node *iff ) {
 #ifndef PRODUCT
   if( PrintOpto && VerifyLoopOptimizations )
     tty->print_cr("Split-if");
+  if (TraceLoopOpts) {
+    tty->print_cr("SplitIf");
+  }
 #endif
   C->set_major_progress();
   Node *region = iff->in(0);
diff --git a/src/share/vm/opto/superword.cpp b/src/share/vm/opto/superword.cpp
index 1230ebe51..6e1be24dd 100644
--- a/src/share/vm/opto/superword.cpp
+++ b/src/share/vm/opto/superword.cpp
@@ -1132,6 +1132,13 @@ void SuperWord::co_locate_pack(Node_List* pk) {
 void SuperWord::output() {
   if (_packset.length() == 0) return;
 
+#ifndef PRODUCT
+  if (TraceLoopOpts) {
+    tty->print("SuperWord    ");
+    lpt()->dump_head();
+  }
+#endif
+
   // MUST ENSURE main loop's initial value is properly aligned:
   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
 
diff --git a/src/share/vm/opto/vectornode.hpp b/src/share/vm/opto/vectornode.hpp
index b4915ddfd..4d28703c8 100644
--- a/src/share/vm/opto/vectornode.hpp
+++ b/src/share/vm/opto/vectornode.hpp
@@ -32,6 +32,7 @@
 //------------------------------VectorNode--------------------------------------
 // Vector Operation
 class VectorNode : public Node {
+  virtual uint size_of() const { return sizeof(*this); }
  protected:
   uint _length; // vector length
   virtual BasicType elt_basic_type() const = 0; // Vector element basic type