7004535: Clone loop predicate during loop unswitch

Summary: Clone loop predicate for clonned loops
Reviewed-by: never

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);

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;
-      }
     }
   }
 

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; }

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);

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();

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 );
@@ -1433,9 +1441,9 @@ 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");
     }
   }
@@ -1541,7 +1549,7 @@ 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 +1581,13 @@ 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 +1713,9 @@ 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 +1766,7 @@ 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 +1809,20 @@ 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);

src/share/vm/opto/loopnode.hpp

@@ -706,11 +706,11 @@ 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 +721,13 @@ 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 +737,7 @@ 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 +830,26 @@ 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);

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);
 

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.

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);

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
 

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