1 /*
   2  * Copyright (c) 1999, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shared/c2/barrierSetC2.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/connode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/loopnode.hpp"
  35 #include "opto/matcher.hpp"
  36 #include "opto/movenode.hpp"
  37 #include "opto/mulnode.hpp"
  38 #include "opto/opaquenode.hpp"
  39 #include "opto/rootnode.hpp"
  40 #include "opto/subnode.hpp"
  41 #include "opto/subtypenode.hpp"
  42 #include "opto/superword.hpp"
  43 #include "opto/vectornode.hpp"
  44 #include "utilities/checkedCast.hpp"
  45 #include "utilities/macros.hpp"
  46 
  47 //=============================================================================
  48 //------------------------------split_thru_phi---------------------------------
  49 // Split Node 'n' through merge point if there is enough win.
  50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
  51   if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
  52       (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
  53     // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
  54     // so disable this for now
  55     return nullptr;
  56   }
  57 
  58   // Splitting range check CastIIs through a loop induction Phi can
  59   // cause new Phis to be created that are left unrelated to the loop
  60   // induction Phi and prevent optimizations (vectorization)
  61   if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
  62       n->in(1) == region->as_CountedLoop()->phi()) {
  63     return nullptr;
  64   }
  65 
  66   if (cannot_split_division(n, region)) {
  67     return nullptr;
  68   }
  69 
  70   SplitThruPhiWins wins(region);
  71   assert(!n->is_CFG(), "");
  72   assert(region->is_Region(), "");
  73 
  74   const Type* type = n->bottom_type();
  75   const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
  76   Node* phi;
  77   if (t_oop != nullptr && t_oop->is_known_instance_field()) {
  78     int iid    = t_oop->instance_id();
  79     int index  = C->get_alias_index(t_oop);
  80     int offset = t_oop->offset();
  81     phi = new PhiNode(region, type, nullptr, iid, index, offset);
  82   } else {
  83     phi = PhiNode::make_blank(region, n);
  84   }
  85   uint old_unique = C->unique();
  86   for (uint i = 1; i < region->req(); i++) {
  87     Node* x;
  88     Node* the_clone = nullptr;
  89     if (region->in(i) == C->top()) {
  90       x = C->top();             // Dead path?  Use a dead data op
  91     } else {
  92       x = n->clone();           // Else clone up the data op
  93       the_clone = x;            // Remember for possible deletion.
  94       // Alter data node to use pre-phi inputs
  95       if (n->in(0) == region)
  96         x->set_req( 0, region->in(i) );
  97       for (uint j = 1; j < n->req(); j++) {
  98         Node* in = n->in(j);
  99         if (in->is_Phi() && in->in(0) == region)
 100           x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
 101       }
 102     }
 103     // Check for a 'win' on some paths
 104     const Type* t = x->Value(&_igvn);
 105 
 106     bool singleton = t->singleton();
 107 
 108     // A TOP singleton indicates that there are no possible values incoming
 109     // along a particular edge. In most cases, this is OK, and the Phi will
 110     // be eliminated later in an Ideal call. However, we can't allow this to
 111     // happen if the singleton occurs on loop entry, as the elimination of
 112     // the PhiNode may cause the resulting node to migrate back to a previous
 113     // loop iteration.
 114     if (singleton && t == Type::TOP) {
 115       // Is_Loop() == false does not confirm the absence of a loop (e.g., an
 116       // irreducible loop may not be indicated by an affirmative is_Loop());
 117       // therefore, the only top we can split thru a phi is on a backedge of
 118       // a loop.
 119       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
 120     }
 121 
 122     if (singleton) {
 123       wins.add_win(i);
 124       x = makecon(t);
 125     } else {
 126       // We now call Identity to try to simplify the cloned node.
 127       // Note that some Identity methods call phase->type(this).
 128       // Make sure that the type array is big enough for
 129       // our new node, even though we may throw the node away.
 130       // (Note: This tweaking with igvn only works because x is a new node.)
 131       _igvn.set_type(x, t);
 132       // If x is a TypeNode, capture any more-precise type permanently into Node
 133       // otherwise it will be not updated during igvn->transform since
 134       // igvn->type(x) is set to x->Value() already.
 135       x->raise_bottom_type(t);
 136       Node* y = x->Identity(&_igvn);
 137       if (y != x) {
 138         wins.add_win(i);
 139         x = y;
 140       } else {
 141         y = _igvn.hash_find(x);
 142         if (y == nullptr) {
 143           y = similar_subtype_check(x, region->in(i));
 144         }
 145         if (y) {
 146           wins.add_win(i);
 147           x = y;
 148         } else {
 149           // Else x is a new node we are keeping
 150           // We do not need register_new_node_with_optimizer
 151           // because set_type has already been called.
 152           _igvn._worklist.push(x);
 153         }
 154       }
 155     }
 156 
 157     phi->set_req( i, x );
 158 
 159     if (the_clone == nullptr) {
 160       continue;
 161     }
 162 
 163     if (the_clone != x) {
 164       _igvn.remove_dead_node(the_clone);
 165     } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
 166                n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
 167       // it is not a win if 'x' moved from an outer to an inner loop
 168       // this edge case can only happen for Load nodes
 169       wins.reset();
 170       break;
 171     }
 172   }
 173   // Too few wins?
 174   if (!wins.profitable(policy)) {
 175     _igvn.remove_dead_node(phi);
 176     return nullptr;
 177   }
 178 
 179   // Record Phi
 180   register_new_node( phi, region );
 181 
 182   for (uint i2 = 1; i2 < phi->req(); i2++) {
 183     Node *x = phi->in(i2);
 184     // If we commoned up the cloned 'x' with another existing Node,
 185     // the existing Node picks up a new use.  We need to make the
 186     // existing Node occur higher up so it dominates its uses.
 187     Node *old_ctrl;
 188     IdealLoopTree *old_loop;
 189 
 190     if (x->is_Con()) {
 191       assert(get_ctrl(x) == C->root(), "constant control is not root");
 192       continue;
 193     }
 194     // The occasional new node
 195     if (x->_idx >= old_unique) {     // Found a new, unplaced node?
 196       old_ctrl = nullptr;
 197       old_loop = nullptr;               // Not in any prior loop
 198     } else {
 199       old_ctrl = get_ctrl(x);
 200       old_loop = get_loop(old_ctrl); // Get prior loop
 201     }
 202     // New late point must dominate new use
 203     Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
 204     if (new_ctrl == old_ctrl) // Nothing is changed
 205       continue;
 206 
 207     IdealLoopTree *new_loop = get_loop(new_ctrl);
 208 
 209     // Don't move x into a loop if its uses are
 210     // outside of loop. Otherwise x will be cloned
 211     // for each use outside of this loop.
 212     IdealLoopTree *use_loop = get_loop(region);
 213     if (!new_loop->is_member(use_loop) &&
 214         (old_loop == nullptr || !new_loop->is_member(old_loop))) {
 215       // Take early control, later control will be recalculated
 216       // during next iteration of loop optimizations.
 217       new_ctrl = get_early_ctrl(x);
 218       new_loop = get_loop(new_ctrl);
 219     }
 220     // Set new location
 221     set_ctrl(x, new_ctrl);
 222     // If changing loop bodies, see if we need to collect into new body
 223     if (old_loop != new_loop) {
 224       if (old_loop && !old_loop->_child)
 225         old_loop->_body.yank(x);
 226       if (!new_loop->_child)
 227         new_loop->_body.push(x);  // Collect body info
 228     }
 229   }
 230 
 231   split_thru_phi_yank_old_nodes(n, region);
 232   _igvn.replace_node(n, phi);
 233 
 234 #ifndef PRODUCT
 235   if (TraceLoopOpts) {
 236     tty->print_cr("Split %d %s through %d Phi in %d %s",
 237                   n->_idx, n->Name(), phi->_idx, region->_idx, region->Name());
 238   }
 239 #endif // !PRODUCT
 240 
 241   return phi;
 242 }
 243 
 244 // If the region is a Loop, we are removing the old n,
 245 // and need to yank it from the _body. If any phi we
 246 // just split through now has no use any more, it also
 247 // has to be removed.
 248 void PhaseIdealLoop::split_thru_phi_yank_old_nodes(Node* n, Node* region) {
 249   IdealLoopTree* region_loop = get_loop(region);
 250   if (region->is_Loop() && region_loop->is_innermost()) {
 251     region_loop->_body.yank(n);
 252     for (uint j = 1; j < n->req(); j++) {
 253       PhiNode* phi = n->in(j)->isa_Phi();
 254       // Check that phi belongs to the region and only has n as a use.
 255       if (phi != nullptr &&
 256           phi->in(0) == region &&
 257           phi->unique_multiple_edges_out_or_null() == n) {
 258         assert(get_ctrl(phi) == region, "sanity");
 259         assert(get_ctrl(n) == region, "sanity");
 260         region_loop->_body.yank(phi);
 261       }
 262     }
 263   }
 264 }
 265 
 266 // Test whether node 'x' can move into an inner loop relative to node 'n'.
 267 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
 268 // BUT it can also return true and 'x' is in the outer loop
 269 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
 270   IdealLoopTree* n_loop_tree = get_loop(n_loop);
 271   IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
 272   // x_loop_tree should be outer or same loop as n_loop_tree
 273   return !x_loop_tree->is_member(n_loop_tree);
 274 }
 275 
 276 // Subtype checks that carry profile data don't common so look for a replacement by following edges
 277 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
 278   if (x->is_SubTypeCheck()) {
 279     Node* in1 = x->in(1);
 280     for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
 281       Node* u = in1->fast_out(i);
 282       if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
 283         for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
 284           Node* bol = u->fast_out(j);
 285           for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
 286             Node* iff = bol->fast_out(k);
 287             // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
 288             // unrelated profile
 289             if (iff->is_If() && is_dominator(iff, r_in)) {
 290               return u;
 291             }
 292           }
 293         }
 294       }
 295     }
 296   }
 297   return nullptr;
 298 }
 299 
 300 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
 301 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
 302 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
 303 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
 304 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
 305 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
 306 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
 307 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
 308   const Type* zero;
 309   switch (n->Opcode()) {
 310     case Op_DivI:
 311     case Op_ModI:
 312     case Op_UDivI:
 313     case Op_UModI:
 314       zero = TypeInt::ZERO;
 315       break;
 316     case Op_DivL:
 317     case Op_ModL:
 318     case Op_UDivL:
 319     case Op_UModL:
 320       zero = TypeLong::ZERO;
 321       break;
 322     default:
 323       return false;
 324   }
 325 
 326   if (n->in(0) != nullptr) {
 327     // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
 328     return true;
 329   }
 330 
 331   Node* divisor = n->in(2);
 332   return is_divisor_loop_phi(divisor, region) &&
 333          loop_phi_backedge_type_contains_zero(divisor, zero);
 334 }
 335 
 336 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
 337   return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
 338 }
 339 
 340 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
 341     return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
 342 }
 343 
 344 //------------------------------dominated_by------------------------------------
 345 // Replace the dominated test with an obvious true or false.  Place it on the
 346 // IGVN worklist for later cleanup.  Move control-dependent data Nodes on the
 347 // live path up to the dominating control.
 348 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool prevdom_not_imply_this) {
 349   if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
 350 
 351   // prevdom is the dominating projection of the dominating test.
 352   assert(iff->Opcode() == Op_If ||
 353          iff->Opcode() == Op_CountedLoopEnd ||
 354          iff->Opcode() == Op_LongCountedLoopEnd ||
 355          iff->Opcode() == Op_RangeCheck ||
 356          iff->Opcode() == Op_ParsePredicate,
 357         "Check this code when new subtype is added");
 358 
 359   int pop = prevdom->Opcode();
 360   assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
 361   if (flip) {
 362     if (pop == Op_IfTrue)
 363       pop = Op_IfFalse;
 364     else
 365       pop = Op_IfTrue;
 366   }
 367   // 'con' is set to true or false to kill the dominated test.
 368   Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
 369   // Hack the dominated test
 370   _igvn.replace_input_of(iff, 1, con);
 371 
 372   // If I don't have a reachable TRUE and FALSE path following the IfNode then
 373   // I can assume this path reaches an infinite loop.  In this case it's not
 374   // important to optimize the data Nodes - either the whole compilation will
 375   // be tossed or this path (and all data Nodes) will go dead.
 376   if (iff->outcnt() != 2) {
 377     return;
 378   }
 379 
 380   // Make control-dependent data Nodes on the live path (path that will remain
 381   // once the dominated IF is removed) become control-dependent on the
 382   // dominating projection.
 383   Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
 384 
 385   if (dp == nullptr) {
 386     return;
 387   }
 388 
 389   rewire_safe_outputs_to_dominator(dp, prevdom, prevdom_not_imply_this);
 390 }
 391 
 392 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool dominator_not_imply_source) {
 393   IdealLoopTree* old_loop = get_loop(source);
 394 
 395   for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
 396     Node* out = source->fast_out(i); // Control-dependent node
 397     if (out->depends_only_on_test()) {
 398       assert(out->in(0) == source, "must be control dependent on source");
 399       _igvn.replace_input_of(out, 0, dominator);
 400       if (dominator_not_imply_source) {
 401         // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
 402         // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
 403         // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
 404         // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
 405         // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
 406         // dominating check.
 407         Node* clone = out->pin_node_under_control();
 408         if (clone != nullptr) {
 409           clone = _igvn.register_new_node_with_optimizer(clone, out);
 410           _igvn.replace_node(out, clone);
 411           out = clone;
 412         }
 413       }
 414       set_early_ctrl(out, false);
 415       IdealLoopTree* new_loop = get_loop(get_ctrl(out));
 416       if (old_loop != new_loop) {
 417         if (!old_loop->_child) {
 418           old_loop->_body.yank(out);
 419         }
 420         if (!new_loop->_child) {
 421           new_loop->_body.push(out);
 422         }
 423       }
 424       --i;
 425       --imax;
 426     }
 427   }
 428 }
 429 
 430 //------------------------------has_local_phi_input----------------------------
 431 // Return TRUE if 'n' has Phi inputs from its local block and no other
 432 // block-local inputs (all non-local-phi inputs come from earlier blocks)
 433 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
 434   Node *n_ctrl = get_ctrl(n);
 435   // See if some inputs come from a Phi in this block, or from before
 436   // this block.
 437   uint i;
 438   for( i = 1; i < n->req(); i++ ) {
 439     Node *phi = n->in(i);
 440     if( phi->is_Phi() && phi->in(0) == n_ctrl )
 441       break;
 442   }
 443   if( i >= n->req() )
 444     return nullptr;                // No Phi inputs; nowhere to clone thru
 445 
 446   // Check for inputs created between 'n' and the Phi input.  These
 447   // must split as well; they have already been given the chance
 448   // (courtesy of a post-order visit) and since they did not we must
 449   // recover the 'cost' of splitting them by being very profitable
 450   // when splitting 'n'.  Since this is unlikely we simply give up.
 451   for( i = 1; i < n->req(); i++ ) {
 452     Node *m = n->in(i);
 453     if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
 454       // We allow the special case of AddP's with no local inputs.
 455       // This allows us to split-up address expressions.
 456       if (m->is_AddP() &&
 457           get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
 458           get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
 459           get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
 460         // Move the AddP up to the dominating point. That's fine because control of m's inputs
 461         // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
 462         Node* c = find_non_split_ctrl(idom(n_ctrl));
 463         if (c->is_OuterStripMinedLoop()) {
 464           c->as_Loop()->verify_strip_mined(1);
 465           c = c->in(LoopNode::EntryControl);
 466         }
 467         set_ctrl_and_loop(m, c);
 468         continue;
 469       }
 470       return nullptr;
 471     }
 472     assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
 473   }
 474 
 475   return n_ctrl;
 476 }
 477 
 478 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
 479 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
 480   assert(bt == T_INT || bt == T_LONG, "only for integers");
 481   int n_op = n->Opcode();
 482 
 483   if (n_op == Op_LShift(bt)) {
 484     // Scale is loop invariant
 485     Node* scale = n->in(2);
 486     Node* scale_ctrl = get_ctrl(scale);
 487     IdealLoopTree* scale_loop = get_loop(scale_ctrl);
 488     if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
 489       return nullptr;
 490     }
 491     const TypeInt* scale_t = scale->bottom_type()->isa_int();
 492     if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
 493       return nullptr;              // Dont bother with byte/short masking
 494     }
 495     // Add must vary with loop (else shift would be loop-invariant)
 496     Node* add = n->in(1);
 497     Node* add_ctrl = get_ctrl(add);
 498     IdealLoopTree* add_loop = get_loop(add_ctrl);
 499     if (n_loop != add_loop) {
 500       return nullptr;  // happens w/ evil ZKM loops
 501     }
 502 
 503     // Convert I-V into I+ (0-V); same for V-I
 504     if (add->Opcode() == Op_Sub(bt) &&
 505         _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
 506       assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
 507       Node* zero = integercon(0, bt);
 508       Node* neg = SubNode::make(zero, add->in(2), bt);
 509       register_new_node_with_ctrl_of(neg, add->in(2));
 510       add = AddNode::make(add->in(1), neg, bt);
 511       register_new_node(add, add_ctrl);
 512     }
 513     if (add->Opcode() != Op_Add(bt)) return nullptr;
 514     assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
 515     // See if one add input is loop invariant
 516     Node* add_var = add->in(1);
 517     Node* add_var_ctrl = get_ctrl(add_var);
 518     IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
 519     Node* add_invar = add->in(2);
 520     Node* add_invar_ctrl = get_ctrl(add_invar);
 521     IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
 522     if (add_invar_loop == n_loop) {
 523       // Swap to find the invariant part
 524       add_invar = add_var;
 525       add_invar_ctrl = add_var_ctrl;
 526       add_invar_loop = add_var_loop;
 527       add_var = add->in(2);
 528     } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
 529       return nullptr;
 530     }
 531     if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
 532       return nullptr;              // No invariant part of the add?
 533     }
 534 
 535     // Yes!  Reshape address expression!
 536     Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
 537     Node* inv_scale_ctrl =
 538             dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
 539             add_invar_ctrl : scale_ctrl;
 540     register_new_node(inv_scale, inv_scale_ctrl);
 541     Node* var_scale = LShiftNode::make(add_var, scale, bt);
 542     register_new_node(var_scale, n_ctrl);
 543     Node* var_add = AddNode::make(var_scale, inv_scale, bt);
 544     register_new_node(var_add, n_ctrl);
 545     _igvn.replace_node(n, var_add);
 546     return var_add;
 547   }
 548   return nullptr;
 549 }
 550 
 551 //------------------------------remix_address_expressions----------------------
 552 // Rework addressing expressions to get the most loop-invariant stuff
 553 // moved out.  We'd like to do all associative operators, but it's especially
 554 // important (common) to do address expressions.
 555 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
 556   if (!has_ctrl(n))  return nullptr;
 557   Node* n_ctrl = get_ctrl(n);
 558   IdealLoopTree* n_loop = get_loop(n_ctrl);
 559 
 560   // See if 'n' mixes loop-varying and loop-invariant inputs and
 561   // itself is loop-varying.
 562 
 563   // Only interested in binary ops (and AddP)
 564   if (n->req() < 3 || n->req() > 4) return nullptr;
 565 
 566   Node* n1_ctrl = get_ctrl(n->in(                    1));
 567   Node* n2_ctrl = get_ctrl(n->in(                    2));
 568   Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
 569   IdealLoopTree* n1_loop = get_loop(n1_ctrl);
 570   IdealLoopTree* n2_loop = get_loop(n2_ctrl);
 571   IdealLoopTree* n3_loop = get_loop(n3_ctrl);
 572 
 573   // Does one of my inputs spin in a tighter loop than self?
 574   if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
 575       (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
 576       (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
 577     return nullptr;                // Leave well enough alone
 578   }
 579 
 580   // Is at least one of my inputs loop-invariant?
 581   if (n1_loop == n_loop &&
 582       n2_loop == n_loop &&
 583       n3_loop == n_loop) {
 584     return nullptr;                // No loop-invariant inputs
 585   }
 586 
 587   Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
 588   if (res != nullptr) {
 589     return res;
 590   }
 591   res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
 592   if (res != nullptr) {
 593     return res;
 594   }
 595 
 596   int n_op = n->Opcode();
 597   // Replace (I+V) with (V+I)
 598   if (n_op == Op_AddI ||
 599       n_op == Op_AddL ||
 600       n_op == Op_AddF ||
 601       n_op == Op_AddD ||
 602       n_op == Op_MulI ||
 603       n_op == Op_MulL ||
 604       n_op == Op_MulF ||
 605       n_op == Op_MulD) {
 606     if (n2_loop == n_loop) {
 607       assert(n1_loop != n_loop, "");
 608       n->swap_edges(1, 2);
 609     }
 610   }
 611 
 612   // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
 613   // but not if I2 is a constant. Skip for irreducible loops.
 614   if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
 615     if (n2_loop == n_loop && n3_loop != n_loop) {
 616       if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
 617         Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
 618         Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
 619         IdealLoopTree* n22loop = get_loop(n22_ctrl);
 620         IdealLoopTree* n23_loop = get_loop(n23_ctrl);
 621         if (n22loop != n_loop && n22loop->is_member(n_loop) &&
 622             n23_loop == n_loop) {
 623           Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
 624           // Stuff new AddP in the loop preheader
 625           register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
 626           Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
 627           register_new_node(add2, n_ctrl);
 628           _igvn.replace_node(n, add2);
 629           return add2;
 630         }
 631       }
 632     }
 633 
 634     // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
 635     if (n2_loop != n_loop && n3_loop == n_loop) {
 636       if (n->in(3)->Opcode() == Op_AddX) {
 637         Node* V = n->in(3)->in(1);
 638         Node* I = n->in(3)->in(2);
 639         if (ctrl_is_member(n_loop, V)) {
 640         } else {
 641           Node *tmp = V; V = I; I = tmp;
 642         }
 643         if (!ctrl_is_member(n_loop, I)) {
 644           Node* add1 = new AddPNode(n->in(1), n->in(2), I);
 645           // Stuff new AddP in the loop preheader
 646           register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
 647           Node* add2 = new AddPNode(n->in(1), add1, V);
 648           register_new_node(add2, n_ctrl);
 649           _igvn.replace_node(n, add2);
 650           return add2;
 651         }
 652       }
 653     }
 654   }
 655 
 656   return nullptr;
 657 }
 658 
 659 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
 660 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
 661   assert(n->Opcode() == Op_AddI, "sanity");
 662   Node * nn = nullptr;
 663   Node * in1 = n->in(1);
 664   Node * in2 = n->in(2);
 665   if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
 666     IdealLoopTree* loop_n = get_loop(get_ctrl(n));
 667     if (loop_n->is_counted() &&
 668         loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
 669         Matcher::match_rule_supported(Op_MulAddVS2VI) &&
 670         Matcher::match_rule_supported(Op_MulAddS2I)) {
 671       Node* mul_in1 = in1->in(1);
 672       Node* mul_in2 = in1->in(2);
 673       Node* mul_in3 = in2->in(1);
 674       Node* mul_in4 = in2->in(2);
 675       if (mul_in1->Opcode() == Op_LoadS &&
 676           mul_in2->Opcode() == Op_LoadS &&
 677           mul_in3->Opcode() == Op_LoadS &&
 678           mul_in4->Opcode() == Op_LoadS) {
 679         IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
 680         IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
 681         IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
 682         IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
 683         IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
 684         IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
 685         // All nodes should be in the same counted loop.
 686         if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
 687             loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
 688           Node* adr1 = mul_in1->in(MemNode::Address);
 689           Node* adr2 = mul_in2->in(MemNode::Address);
 690           Node* adr3 = mul_in3->in(MemNode::Address);
 691           Node* adr4 = mul_in4->in(MemNode::Address);
 692           if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
 693             if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
 694                 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
 695               nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
 696               register_new_node_with_ctrl_of(nn, n);
 697               _igvn.replace_node(n, nn);
 698               return nn;
 699             } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
 700                        (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
 701               nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
 702               register_new_node_with_ctrl_of(nn, n);
 703               _igvn.replace_node(n, nn);
 704               return nn;
 705             }
 706           }
 707         }
 708       }
 709     }
 710   }
 711   return nn;
 712 }
 713 
 714 //------------------------------conditional_move-------------------------------
 715 // Attempt to replace a Phi with a conditional move.  We have some pretty
 716 // strict profitability requirements.  All Phis at the merge point must
 717 // be converted, so we can remove the control flow.  We need to limit the
 718 // number of c-moves to a small handful.  All code that was in the side-arms
 719 // of the CFG diamond is now speculatively executed.  This code has to be
 720 // "cheap enough".  We are pretty much limited to CFG diamonds that merge
 721 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
 722 Node *PhaseIdealLoop::conditional_move( Node *region ) {
 723 
 724   assert(region->is_Region(), "sanity check");
 725   if (region->req() != 3) return nullptr;
 726 
 727   // Check for CFG diamond
 728   Node *lp = region->in(1);
 729   Node *rp = region->in(2);
 730   if (!lp || !rp) return nullptr;
 731   Node *lp_c = lp->in(0);
 732   if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
 733   IfNode *iff = lp_c->as_If();
 734 
 735   // Check for ops pinned in an arm of the diamond.
 736   // Can't remove the control flow in this case
 737   if (lp->outcnt() > 1) return nullptr;
 738   if (rp->outcnt() > 1) return nullptr;
 739 
 740   IdealLoopTree* r_loop = get_loop(region);
 741   assert(r_loop == get_loop(iff), "sanity");
 742   // Always convert to CMOVE if all results are used only outside this loop.
 743   bool used_inside_loop = (r_loop == _ltree_root);
 744 
 745   // Check profitability
 746   int cost = 0;
 747   int phis = 0;
 748   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 749     Node *out = region->fast_out(i);
 750     if (!out->is_Phi()) continue; // Ignore other control edges, etc
 751     phis++;
 752     PhiNode* phi = out->as_Phi();
 753     BasicType bt = phi->type()->basic_type();
 754     switch (bt) {
 755     case T_DOUBLE:
 756     case T_FLOAT:
 757       if (C->use_cmove()) {
 758         continue; //TODO: maybe we want to add some cost
 759       }
 760       cost += Matcher::float_cmove_cost(); // Could be very expensive
 761       break;
 762     case T_LONG: {
 763       cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
 764     }
 765     case T_INT:                 // These all CMOV fine
 766     case T_ADDRESS: {           // (RawPtr)
 767       cost++;
 768       break;
 769     }
 770     case T_NARROWOOP: // Fall through
 771     case T_OBJECT: {            // Base oops are OK, but not derived oops
 772       const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
 773       // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
 774       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 775       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 776       // have a Phi for the base here that we convert to a CMOVE all is well
 777       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 778       // the allocator will have to produce a base by creating a CMOVE of the
 779       // relevant bases.  This puts the allocator in the business of
 780       // manufacturing expensive instructions, generally a bad plan.
 781       // Just Say No to Conditionally-Moved Derived Pointers.
 782       if (tp && tp->offset() != 0)
 783         return nullptr;
 784       cost++;
 785       break;
 786     }
 787     default:
 788       return nullptr;              // In particular, can't do memory or I/O
 789     }
 790     // Add in cost any speculative ops
 791     for (uint j = 1; j < region->req(); j++) {
 792       Node *proj = region->in(j);
 793       Node *inp = phi->in(j);
 794       if (get_ctrl(inp) == proj) { // Found local op
 795         cost++;
 796         // Check for a chain of dependent ops; these will all become
 797         // speculative in a CMOV.
 798         for (uint k = 1; k < inp->req(); k++)
 799           if (get_ctrl(inp->in(k)) == proj)
 800             cost += ConditionalMoveLimit; // Too much speculative goo
 801       }
 802     }
 803     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 804     // This will likely Split-If, a higher-payoff operation.
 805     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 806       Node* use = phi->fast_out(k);
 807       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 808         cost += ConditionalMoveLimit;
 809       // Is there a use inside the loop?
 810       // Note: check only basic types since CMoveP is pinned.
 811       if (!used_inside_loop && is_java_primitive(bt)) {
 812         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 813         if (r_loop == u_loop || r_loop->is_member(u_loop)) {
 814           used_inside_loop = true;
 815         }
 816       }
 817     }
 818   }//for
 819   Node* bol = iff->in(1);
 820   assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
 821   if (bol->is_OpaqueTemplateAssertionPredicate()) {
 822     // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
 823     return nullptr;
 824   }
 825   if (bol->is_OpaqueMultiversioning()) {
 826     assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
 827     // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
 828     // and will make the multiversion_if constant fold in the next IGVN round.
 829     return nullptr;
 830   }
 831   if (!bol->is_Bool()) {
 832     assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
 833     return nullptr;
 834   }
 835   int cmp_op = bol->in(1)->Opcode();
 836   if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
 837     return nullptr;
 838   }
 839   // It is expensive to generate flags from a float compare.
 840   // Avoid duplicated float compare.
 841   if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
 842 
 843   float infrequent_prob = PROB_UNLIKELY_MAG(3);
 844   // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
 845   if (used_inside_loop) {
 846     if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
 847 
 848     // BlockLayoutByFrequency optimization moves infrequent branch
 849     // from hot path. No point in CMOV'ing in such case (110 is used
 850     // instead of 100 to take into account not exactness of float value).
 851     if (BlockLayoutByFrequency) {
 852       infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
 853     }
 854   }
 855   // Check for highly predictable branch.  No point in CMOV'ing if
 856   // we are going to predict accurately all the time.
 857   if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
 858     //keep going
 859   } else if (iff->_prob < infrequent_prob ||
 860       iff->_prob > (1.0f - infrequent_prob))
 861     return nullptr;
 862 
 863   // --------------
 864   // Now replace all Phis with CMOV's
 865   Node *cmov_ctrl = iff->in(0);
 866   uint flip = (lp->Opcode() == Op_IfTrue);
 867   Node_List wq;
 868   while (1) {
 869     PhiNode* phi = nullptr;
 870     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 871       Node *out = region->fast_out(i);
 872       if (out->is_Phi()) {
 873         phi = out->as_Phi();
 874         break;
 875       }
 876     }
 877     if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
 878       break;
 879     }
 880     // Move speculative ops
 881     wq.push(phi);
 882     while (wq.size() > 0) {
 883       Node *n = wq.pop();
 884       for (uint j = 1; j < n->req(); j++) {
 885         Node* m = n->in(j);
 886         if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
 887           set_ctrl(m, cmov_ctrl);
 888           wq.push(m);
 889         }
 890       }
 891     }
 892     Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
 893     register_new_node(cmov, cmov_ctrl);
 894     _igvn.replace_node(phi, cmov);
 895 #ifndef PRODUCT
 896     if (TraceLoopOpts) {
 897       tty->print("CMOV  ");
 898       r_loop->dump_head();
 899       if (Verbose) {
 900         bol->in(1)->dump(1);
 901         cmov->dump(1);
 902       }
 903     }
 904     DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
 905 #endif
 906   }
 907 
 908   // The useless CFG diamond will fold up later; see the optimization in
 909   // RegionNode::Ideal.
 910   _igvn._worklist.push(region);
 911 
 912   return iff->in(1);
 913 }
 914 
 915 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
 916   for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
 917     Node* u = m->fast_out(i);
 918     if (u->is_CFG()) {
 919       if (u->is_NeverBranch()) {
 920         u = u->as_NeverBranch()->proj_out(0);
 921         enqueue_cfg_uses(u, wq);
 922       } else {
 923         wq.push(u);
 924       }
 925     }
 926   }
 927 }
 928 
 929 // Try moving a store out of a loop, right before the loop
 930 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
 931   // Store has to be first in the loop body
 932   IdealLoopTree *n_loop = get_loop(n_ctrl);
 933   if (n->is_Store() && n_loop != _ltree_root &&
 934       n_loop->is_loop() && n_loop->_head->is_Loop() &&
 935       n->in(0) != nullptr) {
 936     Node* address = n->in(MemNode::Address);
 937     Node* value = n->in(MemNode::ValueIn);
 938     Node* mem = n->in(MemNode::Memory);
 939 
 940     // - address and value must be loop invariant
 941     // - memory must be a memory Phi for the loop
 942     // - Store must be the only store on this memory slice in the
 943     // loop: if there's another store following this one then value
 944     // written at iteration i by the second store could be overwritten
 945     // at iteration i+n by the first store: it's not safe to move the
 946     // first store out of the loop
 947     // - nothing must observe the memory Phi: it guarantees no read
 948     // before the store, we are also guaranteed the store post
 949     // dominates the loop head (ignoring a possible early
 950     // exit). Otherwise there would be extra Phi involved between the
 951     // loop's Phi and the store.
 952     // - there must be no early exit from the loop before the Store
 953     // (such an exit most of the time would be an extra use of the
 954     // memory Phi but sometimes is a bottom memory Phi that takes the
 955     // store as input).
 956 
 957     if (!ctrl_is_member(n_loop, address) &&
 958         !ctrl_is_member(n_loop, value) &&
 959         mem->is_Phi() && mem->in(0) == n_loop->_head &&
 960         mem->outcnt() == 1 &&
 961         mem->in(LoopNode::LoopBackControl) == n) {
 962 
 963       assert(n_loop->_tail != nullptr, "need a tail");
 964       assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
 965 
 966       // Verify that there's no early exit of the loop before the store.
 967       bool ctrl_ok = false;
 968       {
 969         // Follow control from loop head until n, we exit the loop or
 970         // we reach the tail
 971         ResourceMark rm;
 972         Unique_Node_List wq;
 973         wq.push(n_loop->_head);
 974 
 975         for (uint next = 0; next < wq.size(); ++next) {
 976           Node *m = wq.at(next);
 977           if (m == n->in(0)) {
 978             ctrl_ok = true;
 979             continue;
 980           }
 981           assert(!has_ctrl(m), "should be CFG");
 982           if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
 983             ctrl_ok = false;
 984             break;
 985           }
 986           enqueue_cfg_uses(m, wq);
 987           if (wq.size() > 10) {
 988             ctrl_ok = false;
 989             break;
 990           }
 991         }
 992       }
 993       if (ctrl_ok) {
 994         // move the Store
 995         _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
 996         _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
 997         _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
 998         // Disconnect the phi now. An empty phi can confuse other
 999         // optimizations in this pass of loop opts.
1000         _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
1001         n_loop->_body.yank(mem);
1002 
1003         set_ctrl_and_loop(n, n->in(0));
1004 
1005         return n;
1006       }
1007     }
1008   }
1009   return nullptr;
1010 }
1011 
1012 // Try moving a store out of a loop, right after the loop
1013 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
1014   if (n->is_Store() && n->in(0) != nullptr) {
1015     Node *n_ctrl = get_ctrl(n);
1016     IdealLoopTree *n_loop = get_loop(n_ctrl);
1017     // Store must be in a loop
1018     if (n_loop != _ltree_root && !n_loop->_irreducible) {
1019       Node* address = n->in(MemNode::Address);
1020       Node* value = n->in(MemNode::ValueIn);
1021       // address must be loop invariant
1022       if (!ctrl_is_member(n_loop, address)) {
1023         // Store must be last on this memory slice in the loop and
1024         // nothing in the loop must observe it
1025         Node* phi = nullptr;
1026         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1027           Node* u = n->fast_out(i);
1028           if (has_ctrl(u)) { // control use?
1029             if (!ctrl_is_member(n_loop, u)) {
1030               continue;
1031             }
1032             if (u->is_Phi() && u->in(0) == n_loop->_head) {
1033               assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1034               // multiple phis on the same slice are possible
1035               if (phi != nullptr) {
1036                 return;
1037               }
1038               phi = u;
1039               continue;
1040             }
1041           }
1042           return;
1043         }
1044         if (phi != nullptr) {
1045           // Nothing in the loop before the store (next iteration)
1046           // must observe the stored value
1047           bool mem_ok = true;
1048           {
1049             ResourceMark rm;
1050             Unique_Node_List wq;
1051             wq.push(phi);
1052             for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1053               Node *m = wq.at(next);
1054               for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1055                 Node* u = m->fast_out(i);
1056                 if (u->is_Store() || u->is_Phi()) {
1057                   if (u != n) {
1058                     wq.push(u);
1059                     mem_ok = (wq.size() <= 10);
1060                   }
1061                 } else {
1062                   mem_ok = false;
1063                   break;
1064                 }
1065               }
1066             }
1067           }
1068           if (mem_ok) {
1069             // Move the store out of the loop if the LCA of all
1070             // users (except for the phi) is outside the loop.
1071             Node* hook = new Node(1);
1072             hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1073             _igvn.rehash_node_delayed(phi);
1074             int count = phi->replace_edge(n, hook, &_igvn);
1075             assert(count > 0, "inconsistent phi");
1076 
1077             // Compute latest point this store can go
1078             Node* lca = get_late_ctrl(n, get_ctrl(n));
1079             if (lca->is_OuterStripMinedLoop()) {
1080               lca = lca->in(LoopNode::EntryControl);
1081             }
1082             if (n_loop->is_member(get_loop(lca))) {
1083               // LCA is in the loop - bail out
1084               _igvn.replace_node(hook, n);
1085               return;
1086             }
1087 #ifdef ASSERT
1088             if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1089               assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1090               n_loop->_head->as_Loop()->verify_strip_mined(1);
1091               Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1092               IdealLoopTree* outer_loop = get_loop(outer);
1093               assert(n_loop->_parent == outer_loop, "broken loop tree");
1094               assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1095             }
1096 #endif
1097             lca = place_outside_loop(lca, n_loop);
1098             assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1099             assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1100 
1101             // Move store out of the loop
1102             _igvn.replace_node(hook, n->in(MemNode::Memory));
1103             _igvn.replace_input_of(n, 0, lca);
1104             set_ctrl_and_loop(n, lca);
1105 
1106             // Disconnect the phi now. An empty phi can confuse other
1107             // optimizations in this pass of loop opts..
1108             if (phi->in(LoopNode::LoopBackControl) == phi) {
1109               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1110               n_loop->_body.yank(phi);
1111             }
1112           }
1113         }
1114       }
1115     }
1116   }
1117 }
1118 
1119 //------------------------------split_if_with_blocks_pre-----------------------
1120 // Do the real work in a non-recursive function.  Data nodes want to be
1121 // cloned in the pre-order so they can feed each other nicely.
1122 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1123   // Cloning these guys is unlikely to win
1124   int n_op = n->Opcode();
1125   if (n_op == Op_MergeMem) {
1126     return n;
1127   }
1128   if (n->is_Proj()) {
1129     return n;
1130   }
1131   // Do not clone-up CmpFXXX variations, as these are always
1132   // followed by a CmpI
1133   if (n->is_Cmp()) {
1134     return n;
1135   }
1136   // Attempt to use a conditional move instead of a phi/branch
1137   if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1138     Node *cmov = conditional_move( n );
1139     if (cmov) {
1140       return cmov;
1141     }
1142   }
1143   if (n->is_CFG() || n->is_LoadStore()) {
1144     return n;
1145   }
1146   if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1147     if (!C->major_progress()) {   // If chance of no more loop opts...
1148       _igvn._worklist.push(n);  // maybe we'll remove them
1149     }
1150     return n;
1151   }
1152 
1153   if (n->is_Con()) {
1154     return n;   // No cloning for Con nodes
1155   }
1156 
1157   Node *n_ctrl = get_ctrl(n);
1158   if (!n_ctrl) {
1159     return n;       // Dead node
1160   }
1161 
1162   Node* res = try_move_store_before_loop(n, n_ctrl);
1163   if (res != nullptr) {
1164     return n;
1165   }
1166 
1167   // Attempt to remix address expressions for loop invariants
1168   Node *m = remix_address_expressions( n );
1169   if( m ) return m;
1170 
1171   if (n_op == Op_AddI) {
1172     Node *nn = convert_add_to_muladd( n );
1173     if ( nn ) return nn;
1174   }
1175 
1176   if (n->is_ConstraintCast() && n->as_ConstraintCast()->dependency().narrows_type()) {
1177     Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1178     // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1179     // Node control inputs don't necessarily agree with loop control info (due to
1180     // transformations happened in between), thus additional dominance check is needed
1181     // to keep loop info valid.
1182     if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1183       _igvn.replace_node(n, dom_cast);
1184       return dom_cast;
1185     }
1186   }
1187 
1188   // Determine if the Node has inputs from some local Phi.
1189   // Returns the block to clone thru.
1190   Node *n_blk = has_local_phi_input( n );
1191   if( !n_blk ) return n;
1192 
1193   // Do not clone the trip counter through on a CountedLoop
1194   // (messes up the canonical shape).
1195   if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1196       (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1197     return n;
1198   }
1199   // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1200   if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1201       n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1202     return n;
1203   }
1204 
1205   // Check for having no control input; not pinned.  Allow
1206   // dominating control.
1207   if (n->in(0)) {
1208     Node *dom = idom(n_blk);
1209     if (dom_lca(n->in(0), dom) != n->in(0)) {
1210       return n;
1211     }
1212   }
1213   // Policy: when is it profitable.  You must get more wins than
1214   // policy before it is considered profitable.  Policy is usually 0,
1215   // so 1 win is considered profitable.  Big merges will require big
1216   // cloning, so get a larger policy.
1217   int policy = n_blk->req() >> 2;
1218 
1219   // If the loop is a candidate for range check elimination,
1220   // delay splitting through it's phi until a later loop optimization
1221   if (n_blk->is_BaseCountedLoop()) {
1222     IdealLoopTree *lp = get_loop(n_blk);
1223     if (lp && lp->_rce_candidate) {
1224       return n;
1225     }
1226   }
1227 
1228   if (must_throttle_split_if()) return n;
1229 
1230   // Split 'n' through the merge point if it is profitable, replacing it with a new phi.
1231   Node* phi = split_thru_phi(n, n_blk, policy);
1232   if (phi == nullptr) { return n; }
1233 
1234   // Moved a load around the loop, 'en-registering' something.
1235   if (n_blk->is_Loop() && n->is_Load() &&
1236       !phi->in(LoopNode::LoopBackControl)->is_Load())
1237     C->set_major_progress();
1238 
1239   return phi;
1240 }
1241 
1242 static bool merge_point_too_heavy(Compile* C, Node* region) {
1243   // Bail out if the region and its phis have too many users.
1244   int weight = 0;
1245   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1246     weight += region->fast_out(i)->outcnt();
1247   }
1248   int nodes_left = C->max_node_limit() - C->live_nodes();
1249   if (weight * 8 > nodes_left) {
1250     if (PrintOpto) {
1251       tty->print_cr("*** Split-if bails out:  %d nodes, region weight %d", C->unique(), weight);
1252     }
1253     return true;
1254   } else {
1255     return false;
1256   }
1257 }
1258 
1259 static bool merge_point_safe(Node* region) {
1260   // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1261   // having a PhiNode input. This sidesteps the dangerous case where the split
1262   // ConvI2LNode may become TOP if the input Value() does not
1263   // overlap the ConvI2L range, leaving a node which may not dominate its
1264   // uses.
1265   // A better fix for this problem can be found in the BugTraq entry, but
1266   // expediency for Mantis demands this hack.
1267 #ifdef _LP64
1268   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1269     Node* n = region->fast_out(i);
1270     if (n->is_Phi()) {
1271       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1272         Node* m = n->fast_out(j);
1273         if (m->Opcode() == Op_ConvI2L)
1274           return false;
1275         if (m->is_CastII()) {
1276           return false;
1277         }
1278       }
1279     }
1280   }
1281 #endif
1282   return true;
1283 }
1284 
1285 
1286 //------------------------------place_outside_loop---------------------------------
1287 // Place some computation outside of this loop on the path to the use passed as argument
1288 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1289   Node* head = loop->_head;
1290   assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1291   if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1292     loop = loop->_parent;
1293     assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1294   }
1295 
1296   // Pick control right outside the loop
1297   for (;;) {
1298     Node* dom = idom(useblock);
1299     if (loop->is_member(get_loop(dom))) {
1300       break;
1301     }
1302     useblock = dom;
1303   }
1304   assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1305   return useblock;
1306 }
1307 
1308 
1309 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1310   if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1311     return false;
1312   }
1313   if (!n->in(0)->is_Region()) {
1314     return false;
1315   }
1316 
1317   Node* region = n->in(0);
1318   Node* dom = idom(region);
1319   if (!dom->is_If() ||  !n->as_If()->same_condition(dom, &_igvn)) {
1320     return false;
1321   }
1322   IfNode* dom_if = dom->as_If();
1323   IfTrueNode* proj_true = dom_if->true_proj();
1324   IfFalseNode* proj_false = dom_if->false_proj();
1325 
1326   for (uint i = 1; i < region->req(); i++) {
1327     if (is_dominator(proj_true, region->in(i))) {
1328       continue;
1329     }
1330     if (is_dominator(proj_false, region->in(i))) {
1331       continue;
1332     }
1333     return false;
1334   }
1335 
1336   return true;
1337 }
1338 
1339 
1340 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1341   if (must_throttle_split_if()) {
1342     return false;
1343   }
1344 
1345   // Do not do 'split-if' if irreducible loops are present.
1346   if (_has_irreducible_loops) {
1347     return false;
1348   }
1349 
1350   if (merge_point_too_heavy(C, n_ctrl)) {
1351     return false;
1352   }
1353 
1354   // Do not do 'split-if' if some paths are dead.  First do dead code
1355   // elimination and then see if its still profitable.
1356   for (uint i = 1; i < n_ctrl->req(); i++) {
1357     if (n_ctrl->in(i) == C->top()) {
1358       return false;
1359     }
1360   }
1361 
1362   // If trying to do a 'Split-If' at the loop head, it is only
1363   // profitable if the cmp folds up on BOTH paths.  Otherwise we
1364   // risk peeling a loop forever.
1365 
1366   // CNC - Disabled for now.  Requires careful handling of loop
1367   // body selection for the cloned code.  Also, make sure we check
1368   // for any input path not being in the same loop as n_ctrl.  For
1369   // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1370   // because the alternative loop entry points won't be converted
1371   // into LoopNodes.
1372   IdealLoopTree *n_loop = get_loop(n_ctrl);
1373   for (uint j = 1; j < n_ctrl->req(); j++) {
1374     if (get_loop(n_ctrl->in(j)) != n_loop) {
1375       return false;
1376     }
1377   }
1378 
1379   // Check for safety of the merge point.
1380   if (!merge_point_safe(n_ctrl)) {
1381     return false;
1382   }
1383 
1384   return true;
1385 }
1386 
1387 // Detect if the node is the inner strip-mined loop
1388 // Return: null if it's not the case, or the exit of outer strip-mined loop
1389 static Node* is_inner_of_stripmined_loop(const Node* out) {
1390   Node* out_le = nullptr;
1391 
1392   if (out->is_CountedLoopEnd()) {
1393       const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1394 
1395       if (loop != nullptr && loop->is_strip_mined()) {
1396         out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1397       }
1398   }
1399 
1400   return out_le;
1401 }
1402 
1403 //------------------------------split_if_with_blocks_post----------------------
1404 // Do the real work in a non-recursive function.  CFG hackery wants to be
1405 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1406 // info.
1407 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1408 
1409   // Cloning Cmp through Phi's involves the split-if transform.
1410   // FastLock is not used by an If
1411   if (n->is_Cmp() && !n->is_FastLock()) {
1412     Node *n_ctrl = get_ctrl(n);
1413     // Determine if the Node has inputs from some local Phi.
1414     // Returns the block to clone thru.
1415     Node *n_blk = has_local_phi_input(n);
1416     if (n_blk != n_ctrl) {
1417       return;
1418     }
1419 
1420     if (!can_split_if(n_ctrl)) {
1421       return;
1422     }
1423 
1424     if (n->outcnt() != 1) {
1425       return; // Multiple bool's from 1 compare?
1426     }
1427     Node *bol = n->unique_out();
1428     assert(bol->is_Bool(), "expect a bool here");
1429     if (bol->outcnt() != 1) {
1430       return;// Multiple branches from 1 compare?
1431     }
1432     Node *iff = bol->unique_out();
1433 
1434     // Check some safety conditions
1435     if (iff->is_If()) {        // Classic split-if?
1436       if (iff->in(0) != n_ctrl) {
1437         return; // Compare must be in same blk as if
1438       }
1439     } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1440       // Can't split CMove with different control.
1441       if (get_ctrl(iff) != n_ctrl) {
1442         return;
1443       }
1444       if (get_ctrl(iff->in(2)) == n_ctrl ||
1445           get_ctrl(iff->in(3)) == n_ctrl) {
1446         return;                 // Inputs not yet split-up
1447       }
1448       if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1449         return;                 // Loop-invar test gates loop-varying CMOVE
1450       }
1451     } else {
1452       return;  // some other kind of node, such as an Allocate
1453     }
1454 
1455     // When is split-if profitable?  Every 'win' on means some control flow
1456     // goes dead, so it's almost always a win.
1457     int policy = 0;
1458     // Split compare 'n' through the merge point if it is profitable
1459     Node *phi = split_thru_phi( n, n_ctrl, policy);
1460     if (!phi) {
1461       return;
1462     }
1463 
1464     // Now split the bool up thru the phi
1465     Node* bolphi = split_thru_phi(bol, n_ctrl, -1);
1466     guarantee(bolphi != nullptr, "null boolean phi node");
1467     assert(iff->in(1) == bolphi, "");
1468 
1469     if (bolphi->Value(&_igvn)->singleton()) {
1470       return;
1471     }
1472 
1473     // Conditional-move?  Must split up now
1474     if (!iff->is_If()) {
1475       Node* cmovphi = split_thru_phi(iff, n_ctrl, -1);
1476       return;
1477     }
1478 
1479     // Now split the IF
1480     C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1481     if (TraceLoopOpts) {
1482       tty->print_cr("Split-If");
1483     }
1484     do_split_if(iff);
1485     C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1486     return;
1487   }
1488 
1489   // Two identical ifs back to back can be merged
1490   if (try_merge_identical_ifs(n)) {
1491     return;
1492   }
1493 
1494   // Check for an IF ready to split; one that has its
1495   // condition codes input coming from a Phi at the block start.
1496   int n_op = n->Opcode();
1497 
1498   // Check for an IF being dominated by another IF same test
1499   if (n_op == Op_If ||
1500       n_op == Op_RangeCheck) {
1501     Node *bol = n->in(1);
1502     uint max = bol->outcnt();
1503     // Check for same test used more than once?
1504     if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1505       // Search up IDOMs to see if this IF is dominated.
1506       Node* cmp = bol->in(1);
1507       Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1508 
1509       // Now search up IDOMs till cutoff, looking for a dominating test
1510       Node *prevdom = n;
1511       Node *dom = idom(prevdom);
1512       while (dom != cutoff) {
1513         if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1514             safe_for_if_replacement(dom)) {
1515           // It's invalid to move control dependent data nodes in the inner
1516           // strip-mined loop, because:
1517           //  1) break validation of LoopNode::verify_strip_mined()
1518           //  2) move code with side-effect in strip-mined loop
1519           // Move to the exit of outer strip-mined loop in that case.
1520           Node* out_le = is_inner_of_stripmined_loop(dom);
1521           if (out_le != nullptr) {
1522             prevdom = out_le;
1523           }
1524           // Replace the dominated test with an obvious true or false.
1525           // Place it on the IGVN worklist for later cleanup.
1526           C->set_major_progress();
1527           // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1528           // to prevent an array load from floating above its range check. There are three cases:
1529           // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1530           //    all its array accesses at that point.
1531           // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1532           //    accesses would start to float, since we don't pin at that point.
1533           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1534           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1535                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1536           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1537           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1538           return;
1539         }
1540         prevdom = dom;
1541         dom = idom(prevdom);
1542       }
1543     }
1544   }
1545 
1546   try_sink_out_of_loop(n);
1547   if (C->failing()) {
1548     return;
1549   }
1550 
1551   try_move_store_after_loop(n);
1552 }
1553 
1554 // Transform:
1555 //
1556 // if (some_condition) {
1557 //   // body 1
1558 // } else {
1559 //   // body 2
1560 // }
1561 // if (some_condition) {
1562 //   // body 3
1563 // } else {
1564 //   // body 4
1565 // }
1566 //
1567 // into:
1568 //
1569 //
1570 // if (some_condition) {
1571 //   // body 1
1572 //   // body 3
1573 // } else {
1574 //   // body 2
1575 //   // body 4
1576 // }
1577 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1578   if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1579     Node *n_ctrl = n->in(0);
1580     IfNode* dom_if = idom(n_ctrl)->as_If();
1581     if (n->in(1) != dom_if->in(1)) {
1582       assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1583              (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1584               dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1585       _igvn.replace_input_of(n, 1, dom_if->in(1));
1586     }
1587     IfTrueNode* dom_proj_true = dom_if->true_proj();
1588     IfFalseNode* dom_proj_false = dom_if->false_proj();
1589 
1590     // Now split the IF
1591     RegionNode* new_false_region;
1592     RegionNode* new_true_region;
1593     do_split_if(n, &new_false_region, &new_true_region);
1594     assert(new_false_region->req() == new_true_region->req(), "");
1595 #ifdef ASSERT
1596     for (uint i = 1; i < new_false_region->req(); ++i) {
1597       assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1598       assert(i == new_false_region->req() - 1 || new_false_region->in(i)->in(0)->in(1) == new_false_region->in(i + 1)->in(0)->in(1), "unexpected shape following split if");
1599     }
1600 #endif
1601     assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1602 
1603     // We now have:
1604     // if (some_condition) {
1605     //   // body 1
1606     //   if (some_condition) {
1607     //     body3: // new_true_region
1608     //     // body3
1609     //   } else {
1610     //     goto body4;
1611     //   }
1612     // } else {
1613     //   // body 2
1614     //  if (some_condition) {
1615     //     goto body3;
1616     //   } else {
1617     //     body4:   // new_false_region
1618     //     // body4;
1619     //   }
1620     // }
1621     //
1622 
1623     // clone pinned nodes thru the resulting regions
1624     push_pinned_nodes_thru_region(dom_if, new_true_region);
1625     push_pinned_nodes_thru_region(dom_if, new_false_region);
1626 
1627     // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1628     // on a projection of n to have the dom_if as a control dependency. We don't want the CastPP to end up with an
1629     // unrelated control dependency.
1630     for (uint i = 1; i < new_false_region->req(); i++) {
1631       if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1632         dominated_by(dom_proj_true, new_false_region->in(i)->in(0)->as_If());
1633       } else {
1634         assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1635         dominated_by(dom_proj_false, new_false_region->in(i)->in(0)->as_If());
1636       }
1637     }
1638     return true;
1639   }
1640   return false;
1641 }
1642 
1643 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1644   for (DUIterator i = region->outs(); region->has_out(i); i++) {
1645     Node* u = region->out(i);
1646     if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test()) {
1647       continue;
1648     }
1649     assert(u->in(0) == region, "not a control dependent node?");
1650     uint j = 1;
1651     for (; j < u->req(); ++j) {
1652       Node* in = u->in(j);
1653       if (!is_dominator(ctrl_or_self(in), dom_if)) {
1654         break;
1655       }
1656     }
1657     if (j == u->req()) {
1658       Node *phi = PhiNode::make_blank(region, u);
1659       for (uint k = 1; k < region->req(); ++k) {
1660         Node* clone = u->clone();
1661         clone->set_req(0, region->in(k));
1662         register_new_node(clone, region->in(k));
1663         phi->init_req(k, clone);
1664       }
1665       register_new_node(phi, region);
1666       _igvn.replace_node(u, phi);
1667       --i;
1668     }
1669   }
1670 }
1671 
1672 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1673   if (!dom->is_CountedLoopEnd()) {
1674     return true;
1675   }
1676   CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1677   CountedLoopNode* cl = le->loopnode();
1678   if (cl == nullptr) {
1679     return true;
1680   }
1681   if (!cl->is_main_loop()) {
1682     return true;
1683   }
1684   if (cl->is_canonical_loop_entry() == nullptr) {
1685     return true;
1686   }
1687   // Further unrolling is possible so loop exit condition might change
1688   return false;
1689 }
1690 
1691 // See if a shared loop-varying computation has no loop-varying uses.
1692 // Happens if something is only used for JVM state in uncommon trap exits,
1693 // like various versions of induction variable+offset.  Clone the
1694 // computation per usage to allow it to sink out of the loop.
1695 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1696   bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1697                             n->in(1)->bottom_type()->isa_rawptr() &&
1698                             !n->bottom_type()->isa_rawptr();
1699 
1700   if (has_ctrl(n) &&
1701       !n->is_Phi() &&
1702       !n->is_Bool() &&
1703       !n->is_Proj() &&
1704       !n->is_MergeMem() &&
1705       !n->is_CMove() &&
1706       !n->is_OpaqueConstantBool() &&
1707       !n->is_OpaqueInitializedAssertionPredicate() &&
1708       !n->is_OpaqueTemplateAssertionPredicate() &&
1709       !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1710       n->Opcode() != Op_CreateEx &&
1711       (KillPathsReachableByDeadTypeNode || !n->is_Type())
1712       ) {
1713     Node *n_ctrl = get_ctrl(n);
1714     IdealLoopTree *n_loop = get_loop(n_ctrl);
1715 
1716     if (n->in(0) != nullptr) {
1717       IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1718       if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1719         // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1720         // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1721         // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1722         Node* maybe_pinned_n = n;
1723         Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1724         if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) {
1725           if (n->depends_only_on_test()) {
1726             // If this node depends_only_on_test, it will be rewired to a control input that is not
1727             // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1728             // rewired to a dominating test equivalent to the new control.
1729             Node* pinned_clone = n->pin_node_under_control();
1730             if (pinned_clone != nullptr) {
1731               register_new_node(pinned_clone, n_ctrl);
1732               maybe_pinned_n = pinned_clone;
1733               _igvn.replace_node(n, pinned_clone);
1734             }
1735           }
1736           _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1737         }
1738       }
1739     }
1740     if (n_loop != _ltree_root && n->outcnt() > 1) {
1741       // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1742       // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1743       Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1744       if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1745           ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1746         if (n->is_Store() || n->is_LoadStore()) {
1747             assert(false, "no node with a side effect");
1748             C->record_failure("no node with a side effect");
1749             return;
1750         }
1751         Node* outer_loop_clone = nullptr;
1752         for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1753           Node* u = n->last_out(j); // Clone private computation per use
1754           _igvn.rehash_node_delayed(u);
1755           Node* x = nullptr;
1756           if (n->in(0) != nullptr && n->depends_only_on_test()) {
1757             // If this node depends_only_on_test, it will be rewired to a control input that is not
1758             // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1759             // rewired to a dominating test equivalent to the new control.
1760             x = n->pin_node_under_control();
1761           }
1762           if (x == nullptr) {
1763             x = n->clone();
1764           }
1765           Node* x_ctrl = nullptr;
1766           if (u->is_Phi()) {
1767             // Replace all uses of normal nodes.  Replace Phi uses
1768             // individually, so the separate Nodes can sink down
1769             // different paths.
1770             uint k = 1;
1771             while (u->in(k) != n) k++;
1772             u->set_req(k, x);
1773             // x goes next to Phi input path
1774             x_ctrl = u->in(0)->in(k);
1775             // Find control for 'x' next to use but not inside inner loops.
1776             x_ctrl = place_outside_loop(x_ctrl, n_loop);
1777             --j;
1778           } else {              // Normal use
1779             if (has_ctrl(u)) {
1780               x_ctrl = get_ctrl(u);
1781             } else {
1782               x_ctrl = u->in(0);
1783             }
1784             // Find control for 'x' next to use but not inside inner loops.
1785             x_ctrl = place_outside_loop(x_ctrl, n_loop);
1786             // Replace all uses
1787             if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1788               // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1789               // anymore now that we're going to pin n as well
1790               _igvn.replace_node(u, x);
1791               --j;
1792             } else {
1793               int nb = u->replace_edge(n, x, &_igvn);
1794               j -= nb;
1795             }
1796           }
1797 
1798           if (n->is_Load()) {
1799             // For loads, add a control edge to a CFG node outside of the loop
1800             // to force them to not combine and return back inside the loop
1801             // during GVN optimization (4641526).
1802             assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1803 
1804             IdealLoopTree* x_loop = get_loop(x_ctrl);
1805             Node* x_head = x_loop->_head;
1806             if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1807               // Do not add duplicate LoadNodes to the outer strip mined loop
1808               if (outer_loop_clone != nullptr) {
1809                 _igvn.replace_node(x, outer_loop_clone);
1810                 continue;
1811               }
1812               outer_loop_clone = x;
1813             }
1814             x->set_req(0, x_ctrl);
1815           } else if (n->in(0) != nullptr){
1816             x->set_req(0, x_ctrl);
1817           }
1818           assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1819           assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1820           register_new_node(x, x_ctrl);
1821 
1822           // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1823           // All AddP nodes must keep the same base after sinking so:
1824           // 1- We don't add a CastPP here until the last one of the chain is sunk: if part of the chain is not sunk,
1825           // their bases remain the same.
1826           // (see 2- below)
1827           assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1828                  x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1829                  !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1830           if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1831               !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1832             assert(!x->is_Load(), "load should be pinned");
1833             // Use a cast node to pin clone out of loop
1834             Node* cast = nullptr;
1835             for (uint k = 0; k < x->req(); k++) {
1836               Node* in = x->in(k);
1837               if (in != nullptr && ctrl_is_member(n_loop, in)) {
1838                 const Type* in_t = _igvn.type(in);
1839                 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1840                                                               ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr);
1841               }
1842               if (cast != nullptr) {
1843                 Node* prev = _igvn.hash_find_insert(cast);
1844                 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1845                   cast->destruct(&_igvn);
1846                   cast = prev;
1847                 } else {
1848                   register_new_node(cast, x_ctrl);
1849                 }
1850                 x->replace_edge(in, cast);
1851                 // Chain of AddP nodes:
1852                 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1853                 if (x->is_AddP() && k == AddPNode::Base) {
1854                   update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1855                 }
1856                 break;
1857               }
1858             }
1859             assert(cast != nullptr, "must have added a cast to pin the node");
1860           }
1861         }
1862         _igvn.remove_dead_node(n);
1863       }
1864       _dom_lca_tags_round = 0;
1865     }
1866   }
1867 }
1868 
1869 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1870   ResourceMark rm;
1871   Node_List wq;
1872   wq.push(x);
1873   while (wq.size() != 0) {
1874     Node* n = wq.pop();
1875     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1876       Node* u = n->fast_out(i);
1877       if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1878         _igvn.replace_input_of(u, AddPNode::Base, new_base);
1879         wq.push(u);
1880       }
1881     }
1882   }
1883 }
1884 
1885 // Compute the early control of a node by following its inputs until we reach
1886 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1887 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1888   Node* early_ctrl = nullptr;
1889   ResourceMark rm;
1890   Unique_Node_List wq;
1891   wq.push(n);
1892   for (uint i = 0; i < wq.size(); i++) {
1893     Node* m = wq.at(i);
1894     Node* c = nullptr;
1895     if (m->is_CFG()) {
1896       c = m;
1897     } else if (m->pinned()) {
1898       c = m->in(0);
1899     } else {
1900       for (uint j = 0; j < m->req(); j++) {
1901         Node* in = m->in(j);
1902         if (in != nullptr) {
1903           wq.push(in);
1904         }
1905       }
1906     }
1907     if (c != nullptr) {
1908       assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1909       if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1910         early_ctrl = c;
1911       }
1912     }
1913   }
1914   assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1915   return early_ctrl;
1916 }
1917 
1918 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1919   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1920     Node* u = n->fast_out(i);
1921     if (u->is_Opaque1()) {
1922       return false;  // Found loop limit, bugfix for 4677003
1923     }
1924     if (u->is_Phi()) {
1925       for (uint j = 1; j < u->req(); ++j) {
1926         if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1927           return false;
1928         }
1929       }
1930     } else {
1931       Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1932       if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1933         return false;
1934       }
1935     }
1936   }
1937   return true;
1938 }
1939 
1940 // Sinking a node from a pre loop to its main loop pins the node between the pre and main loops. If that node is input
1941 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1942 // test of the pre loop above the point in the graph where it's pinned. This results in a broken graph. One way to avoid
1943 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code
1944 // is generated for the main loop.
1945 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) {
1946   if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1947     CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1948     if (is_dominator(pre_loop->loopexit(), ctrl)) {
1949       return true;
1950     }
1951   }
1952   return false;
1953 }
1954 
1955 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1956   if (n->is_Load()) {
1957     // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure each call to
1958     // get_late_ctrl_with_anti_dep() uses its own tag
1959     _dom_lca_tags_round++;
1960     assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1961 
1962     ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1963   }
1964   IdealLoopTree *u_loop = get_loop(ctrl);
1965   if (u_loop == n_loop) {
1966     return false; // Found loop-varying use
1967   }
1968   if (n_loop->is_member(u_loop)) {
1969     return false; // Found use in inner loop
1970   }
1971   if (would_sink_below_pre_loop_exit(n_loop, ctrl)) {
1972     return false;
1973   }
1974   return true;
1975 }
1976 
1977 //------------------------------split_if_with_blocks---------------------------
1978 // Check for aggressive application of 'split-if' optimization,
1979 // using basic block level info.
1980 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1981   Node* root = C->root();
1982   visited.set(root->_idx); // first, mark root as visited
1983   // Do pre-visit work for root
1984   Node* n   = split_if_with_blocks_pre(root);
1985   uint  cnt = n->outcnt();
1986   uint  i   = 0;
1987 
1988   while (true) {
1989     // Visit all children
1990     if (i < cnt) {
1991       Node* use = n->raw_out(i);
1992       ++i;
1993       if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1994         // Now do pre-visit work for this use
1995         use = split_if_with_blocks_pre(use);
1996         nstack.push(n, i); // Save parent and next use's index.
1997         n   = use;         // Process all children of current use.
1998         cnt = use->outcnt();
1999         i   = 0;
2000       }
2001     }
2002     else {
2003       // All of n's children have been processed, complete post-processing.
2004       if (cnt != 0 && !n->is_Con()) {
2005         assert(has_node(n), "no dead nodes");
2006         split_if_with_blocks_post(n);
2007         if (C->failing()) {
2008           return;
2009         }
2010       }
2011       if (must_throttle_split_if()) {
2012         nstack.clear();
2013       }
2014       if (nstack.is_empty()) {
2015         // Finished all nodes on stack.
2016         break;
2017       }
2018       // Get saved parent node and next use's index. Visit the rest of uses.
2019       n   = nstack.node();
2020       cnt = n->outcnt();
2021       i   = nstack.index();
2022       nstack.pop();
2023     }
2024   }
2025 }
2026 
2027 
2028 //=============================================================================
2029 //
2030 //                   C L O N E   A   L O O P   B O D Y
2031 //
2032 
2033 //------------------------------clone_iff--------------------------------------
2034 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2035 // "Nearly" because all Nodes have been cloned from the original in the loop,
2036 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
2037 // through the Phi recursively, and return a Bool.
2038 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2039 
2040   // Convert this Phi into a Phi merging Bools
2041   uint i;
2042   for (i = 1; i < phi->req(); i++) {
2043     Node* b = phi->in(i);
2044     if (b->is_Phi()) {
2045       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2046     } else {
2047       assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(),
2048              "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node");
2049     }
2050   }
2051   Node* n = phi->in(1);
2052   Node* sample_opaque = nullptr;
2053   Node *sample_bool = nullptr;
2054   if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) {
2055     sample_opaque = n;
2056     sample_bool = n->in(1);
2057     assert(sample_bool->is_Bool(), "wrong type");
2058   } else {
2059     sample_bool = n;
2060   }
2061   Node *sample_cmp = sample_bool->in(1);
2062 
2063   // Make Phis to merge the Cmp's inputs.
2064   PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2065   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2066   for (i = 1; i < phi->req(); i++) {
2067     Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2068     Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2069     phi1->set_req(i, n1);
2070     phi2->set_req(i, n2);
2071     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2072     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2073   }
2074   // See if these Phis have been made before.
2075   // Register with optimizer
2076   Node *hit1 = _igvn.hash_find_insert(phi1);
2077   if (hit1) {                   // Hit, toss just made Phi
2078     _igvn.remove_dead_node(phi1); // Remove new phi
2079     assert(hit1->is_Phi(), "" );
2080     phi1 = (PhiNode*)hit1;      // Use existing phi
2081   } else {                      // Miss
2082     _igvn.register_new_node_with_optimizer(phi1);
2083   }
2084   Node *hit2 = _igvn.hash_find_insert(phi2);
2085   if (hit2) {                   // Hit, toss just made Phi
2086     _igvn.remove_dead_node(phi2); // Remove new phi
2087     assert(hit2->is_Phi(), "" );
2088     phi2 = (PhiNode*)hit2;      // Use existing phi
2089   } else {                      // Miss
2090     _igvn.register_new_node_with_optimizer(phi2);
2091   }
2092   // Register Phis with loop/block info
2093   set_ctrl(phi1, phi->in(0));
2094   set_ctrl(phi2, phi->in(0));
2095   // Make a new Cmp
2096   Node *cmp = sample_cmp->clone();
2097   cmp->set_req(1, phi1);
2098   cmp->set_req(2, phi2);
2099   _igvn.register_new_node_with_optimizer(cmp);
2100   set_ctrl(cmp, phi->in(0));
2101 
2102   // Make a new Bool
2103   Node *b = sample_bool->clone();
2104   b->set_req(1,cmp);
2105   _igvn.register_new_node_with_optimizer(b);
2106   set_ctrl(b, phi->in(0));
2107 
2108   if (sample_opaque != nullptr) {
2109     Node* opaque = sample_opaque->clone();
2110     opaque->set_req(1, b);
2111     _igvn.register_new_node_with_optimizer(opaque);
2112     set_ctrl(opaque, phi->in(0));
2113     return opaque;
2114   }
2115 
2116   assert(b->is_Bool(), "");
2117   return b;
2118 }
2119 
2120 //------------------------------clone_bool-------------------------------------
2121 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2122 // "Nearly" because all Nodes have been cloned from the original in the loop,
2123 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
2124 // through the Phi recursively, and return a Bool.
2125 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2126   uint i;
2127   // Convert this Phi into a Phi merging Bools
2128   for( i = 1; i < phi->req(); i++ ) {
2129     Node *b = phi->in(i);
2130     if( b->is_Phi() ) {
2131       _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2132     } else {
2133       assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2134     }
2135   }
2136 
2137   Node *sample_cmp = phi->in(1);
2138 
2139   // Make Phis to merge the Cmp's inputs.
2140   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2141   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2142   for( uint j = 1; j < phi->req(); j++ ) {
2143     Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2144     Node *n1, *n2;
2145     if( cmp_top->is_Cmp() ) {
2146       n1 = cmp_top->in(1);
2147       n2 = cmp_top->in(2);
2148     } else {
2149       n1 = n2 = cmp_top;
2150     }
2151     phi1->set_req( j, n1 );
2152     phi2->set_req( j, n2 );
2153     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2154     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2155   }
2156 
2157   // See if these Phis have been made before.
2158   // Register with optimizer
2159   Node *hit1 = _igvn.hash_find_insert(phi1);
2160   if( hit1 ) {                  // Hit, toss just made Phi
2161     _igvn.remove_dead_node(phi1); // Remove new phi
2162     assert( hit1->is_Phi(), "" );
2163     phi1 = (PhiNode*)hit1;      // Use existing phi
2164   } else {                      // Miss
2165     _igvn.register_new_node_with_optimizer(phi1);
2166   }
2167   Node *hit2 = _igvn.hash_find_insert(phi2);
2168   if( hit2 ) {                  // Hit, toss just made Phi
2169     _igvn.remove_dead_node(phi2); // Remove new phi
2170     assert( hit2->is_Phi(), "" );
2171     phi2 = (PhiNode*)hit2;      // Use existing phi
2172   } else {                      // Miss
2173     _igvn.register_new_node_with_optimizer(phi2);
2174   }
2175   // Register Phis with loop/block info
2176   set_ctrl(phi1, phi->in(0));
2177   set_ctrl(phi2, phi->in(0));
2178   // Make a new Cmp
2179   Node *cmp = sample_cmp->clone();
2180   cmp->set_req( 1, phi1 );
2181   cmp->set_req( 2, phi2 );
2182   _igvn.register_new_node_with_optimizer(cmp);
2183   set_ctrl(cmp, phi->in(0));
2184 
2185   assert( cmp->is_Cmp(), "" );
2186   return (CmpNode*)cmp;
2187 }
2188 
2189 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2190                                                  IdealLoopTree* loop, IdealLoopTree* outer_loop,
2191                                                  Node_List*& split_if_set, Node_List*& split_bool_set,
2192                                                  Node_List*& split_cex_set, Node_List& worklist,
2193                                                  uint new_counter, CloneLoopMode mode) {
2194   Node* nnn = old_new[old->_idx];
2195   // Copy uses to a worklist, so I can munge the def-use info
2196   // with impunity.
2197   for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2198     worklist.push(old->fast_out(j));
2199 
2200   while( worklist.size() ) {
2201     Node *use = worklist.pop();
2202     if (!has_node(use))  continue; // Ignore dead nodes
2203     if (use->in(0) == C->top())  continue;
2204     IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2205     // Check for data-use outside of loop - at least one of OLD or USE
2206     // must not be a CFG node.
2207 #ifdef ASSERT
2208     if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2209       Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2210       assert(mode != IgnoreStripMined, "incorrect cloning mode");
2211       assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2212     }
2213 #endif
2214     if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2215 
2216       // If the Data use is an IF, that means we have an IF outside the
2217       // loop that is switching on a condition that is set inside the
2218       // loop.  Happens if people set a loop-exit flag; then test the flag
2219       // in the loop to break the loop, then test is again outside the
2220       // loop to determine which way the loop exited.
2221       //
2222       // For several uses we need to make sure that there is no phi between,
2223       // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2224       // to avoid such a phi in between.
2225       // For example, it is unexpected that there is a Phi between an
2226       // AllocateArray node and its ValidLengthTest input that could cause
2227       // split if to break.
2228       assert(!use->is_OpaqueTemplateAssertionPredicate(),
2229              "should not clone a Template Assertion Predicate which should be removed once it's useless");
2230       if (use->is_If() || use->is_CMove() || use->is_OpaqueConstantBool() || use->is_OpaqueInitializedAssertionPredicate() ||
2231           (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2232         // Since this code is highly unlikely, we lazily build the worklist
2233         // of such Nodes to go split.
2234         if (!split_if_set) {
2235           split_if_set = new Node_List();
2236         }
2237         split_if_set->push(use);
2238       }
2239       if (use->is_Bool()) {
2240         if (!split_bool_set) {
2241           split_bool_set = new Node_List();
2242         }
2243         split_bool_set->push(use);
2244       }
2245       if (use->Opcode() == Op_CreateEx) {
2246         if (!split_cex_set) {
2247           split_cex_set = new Node_List();
2248         }
2249         split_cex_set->push(use);
2250       }
2251 
2252 
2253       // Get "block" use is in
2254       uint idx = 0;
2255       while( use->in(idx) != old ) idx++;
2256       Node *prev = use->is_CFG() ? use : get_ctrl(use);
2257       assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2258       Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2259         ? prev->in(2)
2260         : idom(prev);
2261       if( use->is_Phi() )     // Phi use is in prior block
2262         cfg = prev->in(idx);  // NOT in block of Phi itself
2263       if (cfg->is_top()) {    // Use is dead?
2264         _igvn.replace_input_of(use, idx, C->top());
2265         continue;
2266       }
2267 
2268       // If use is referenced through control edge... (idx == 0)
2269       if (mode == IgnoreStripMined && idx == 0) {
2270         LoopNode *head = loop->_head->as_Loop();
2271         if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2272           // That node is outside the inner loop, leave it outside the
2273           // outer loop as well to not confuse verification code.
2274           assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2275           _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2276           continue;
2277         }
2278       }
2279 
2280       while(!outer_loop->is_member(get_loop(cfg))) {
2281         prev = cfg;
2282         cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2283       }
2284       // If the use occurs after merging several exits from the loop, then
2285       // old value must have dominated all those exits.  Since the same old
2286       // value was used on all those exits we did not need a Phi at this
2287       // merge point.  NOW we do need a Phi here.  Each loop exit value
2288       // is now merged with the peeled body exit; each exit gets its own
2289       // private Phi and those Phis need to be merged here.
2290       Node *phi;
2291       if( prev->is_Region() ) {
2292         if( idx == 0 ) {      // Updating control edge?
2293           phi = prev;         // Just use existing control
2294         } else {              // Else need a new Phi
2295           phi = PhiNode::make( prev, old );
2296           // Now recursively fix up the new uses of old!
2297           for( uint i = 1; i < prev->req(); i++ ) {
2298             worklist.push(phi); // Onto worklist once for each 'old' input
2299           }
2300         }
2301       } else {
2302         // Get new RegionNode merging old and new loop exits
2303         prev = old_new[prev->_idx];
2304         assert( prev, "just made this in step 7" );
2305         if( idx == 0) {      // Updating control edge?
2306           phi = prev;         // Just use existing control
2307         } else {              // Else need a new Phi
2308           // Make a new Phi merging data values properly
2309           phi = PhiNode::make( prev, old );
2310           phi->set_req( 1, nnn );
2311         }
2312       }
2313       // If inserting a new Phi, check for prior hits
2314       if( idx != 0 ) {
2315         Node *hit = _igvn.hash_find_insert(phi);
2316         if( hit == nullptr ) {
2317           _igvn.register_new_node_with_optimizer(phi); // Register new phi
2318         } else {                                      // or
2319           // Remove the new phi from the graph and use the hit
2320           _igvn.remove_dead_node(phi);
2321           phi = hit;                                  // Use existing phi
2322         }
2323         set_ctrl(phi, prev);
2324       }
2325       // Make 'use' use the Phi instead of the old loop body exit value
2326       assert(use->in(idx) == old, "old is still input of use");
2327       // We notify all uses of old, including use, and the indirect uses,
2328       // that may now be optimized because we have replaced old with phi.
2329       _igvn.add_users_to_worklist(old);
2330       if (idx == 0 && use->depends_only_on_test()) {
2331         // If this node depends_only_on_test, it will be rewired to a control input that is not the
2332         // correct test. As a result, it must be pinned otherwise it can be incorrectly rewired to
2333         // a dominating test equivalent to the new control.
2334         Node* pinned_clone = use->pin_node_under_control();
2335         if (pinned_clone != nullptr) {
2336           pinned_clone->set_req(0, phi);
2337           register_new_node_with_ctrl_of(pinned_clone, use);
2338           _igvn.replace_node(use, pinned_clone);
2339           continue;
2340         }
2341       }
2342       _igvn.replace_input_of(use, idx, phi);
2343       if( use->_idx >= new_counter ) { // If updating new phis
2344         // Not needed for correctness, but prevents a weak assert
2345         // in AddPNode from tripping (when we end up with different
2346         // base & derived Phis that will become the same after
2347         // IGVN does CSE).
2348         Node *hit = _igvn.hash_find_insert(use);
2349         if( hit )             // Go ahead and re-hash for hits.
2350           _igvn.replace_node( use, hit );
2351       }
2352     }
2353   }
2354 }
2355 
2356 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2357                                                                 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2358                                                                 bool check_old_new) {
2359   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2360     Node* u = n->fast_out(j);
2361     assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2362     if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2363       assert(!phase->ctrl_is_member(loop, u) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2364       if (!phase->ctrl_is_member(loop, u)) {
2365         if (phase->ctrl_is_member(outer_loop, u)) {
2366           wq.push(u);
2367         } else {
2368           // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2369           // the outer loop too
2370           Node* u_c = u->in(0);
2371           if (u_c != nullptr) {
2372             IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2373             if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2374               wq.push(u);
2375             }
2376           }
2377         }
2378       }
2379     }
2380   }
2381 }
2382 
2383 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2384                                       IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2385                                       Node_List& extra_data_nodes) {
2386   if (head->is_strip_mined() && mode != IgnoreStripMined) {
2387     CountedLoopNode* cl = head->as_CountedLoop();
2388     Node* l = cl->outer_loop();
2389     Node* tail = cl->outer_loop_tail();
2390     IfNode* le = cl->outer_loop_end();
2391     Node* sfpt = cl->outer_safepoint();
2392     CountedLoopEndNode* cle = cl->loopexit();
2393     CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2394     CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2395     IfFalseNode* cle_out = cle->false_proj();
2396 
2397     Node* new_sfpt = nullptr;
2398     Node* new_cle_out = cle_out->clone();
2399     old_new.map(cle_out->_idx, new_cle_out);
2400     if (mode == CloneIncludesStripMined) {
2401       // clone outer loop body
2402       Node* new_l = l->clone();
2403       Node* new_tail = tail->clone();
2404       IfNode* new_le = le->clone()->as_If();
2405       new_sfpt = sfpt->clone();
2406 
2407       set_loop(new_l, outer_loop->_parent);
2408       set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2409       set_loop(new_cle_out, outer_loop->_parent);
2410       set_idom(new_cle_out, new_cle, dd);
2411       set_loop(new_sfpt, outer_loop->_parent);
2412       set_idom(new_sfpt, new_cle_out, dd);
2413       set_loop(new_le, outer_loop->_parent);
2414       set_idom(new_le, new_sfpt, dd);
2415       set_loop(new_tail, outer_loop->_parent);
2416       set_idom(new_tail, new_le, dd);
2417       set_idom(new_cl, new_l, dd);
2418 
2419       old_new.map(l->_idx, new_l);
2420       old_new.map(tail->_idx, new_tail);
2421       old_new.map(le->_idx, new_le);
2422       old_new.map(sfpt->_idx, new_sfpt);
2423 
2424       new_l->set_req(LoopNode::LoopBackControl, new_tail);
2425       new_l->set_req(0, new_l);
2426       new_tail->set_req(0, new_le);
2427       new_le->set_req(0, new_sfpt);
2428       new_sfpt->set_req(0, new_cle_out);
2429       new_cle_out->set_req(0, new_cle);
2430       new_cl->set_req(LoopNode::EntryControl, new_l);
2431 
2432       _igvn.register_new_node_with_optimizer(new_l);
2433       _igvn.register_new_node_with_optimizer(new_tail);
2434       _igvn.register_new_node_with_optimizer(new_le);
2435     } else {
2436       Node *newhead = old_new[loop->_head->_idx];
2437       newhead->as_Loop()->clear_strip_mined();
2438       _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2439       set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2440     }
2441     // Look at data node that were assigned a control in the outer
2442     // loop: they are kept in the outer loop by the safepoint so start
2443     // from the safepoint node's inputs.
2444     IdealLoopTree* outer_loop = get_loop(l);
2445     Node_Stack stack(2);
2446     stack.push(sfpt, 1);
2447     uint new_counter = C->unique();
2448     while (stack.size() > 0) {
2449       Node* n = stack.node();
2450       uint i = stack.index();
2451       while (i < n->req() &&
2452              (n->in(i) == nullptr ||
2453               !has_ctrl(n->in(i)) ||
2454               get_loop(get_ctrl(n->in(i))) != outer_loop ||
2455               (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2456         i++;
2457       }
2458       if (i < n->req()) {
2459         stack.set_index(i+1);
2460         stack.push(n->in(i), 0);
2461       } else {
2462         assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2463         Node* m = n == sfpt ? new_sfpt : n->clone();
2464         if (m != nullptr) {
2465           for (uint i = 0; i < n->req(); i++) {
2466             if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2467               m->set_req(i, old_new[m->in(i)->_idx]);
2468             }
2469           }
2470         } else {
2471           assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2472         }
2473         if (n != sfpt) {
2474           extra_data_nodes.push(n);
2475           _igvn.register_new_node_with_optimizer(m);
2476           assert(get_ctrl(n) == cle_out, "what other control?");
2477           set_ctrl(m, new_cle_out);
2478           old_new.map(n->_idx, m);
2479         }
2480         stack.pop();
2481       }
2482     }
2483     if (mode == CloneIncludesStripMined) {
2484       _igvn.register_new_node_with_optimizer(new_sfpt);
2485       _igvn.register_new_node_with_optimizer(new_cle_out);
2486     }
2487     // Some other transformation may have pessimistically assigned some
2488     // data nodes to the outer loop. Set their control so they are out
2489     // of the outer loop.
2490     ResourceMark rm;
2491     Unique_Node_List wq;
2492     for (uint i = 0; i < extra_data_nodes.size(); i++) {
2493       Node* old = extra_data_nodes.at(i);
2494       collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2495     }
2496 
2497     for (uint i = 0; i < loop->_body.size(); i++) {
2498       Node* old = loop->_body.at(i);
2499       collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2500     }
2501 
2502     Node* inner_out = sfpt->in(0);
2503     if (inner_out->outcnt() > 1) {
2504       collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2505     }
2506 
2507     Node* new_ctrl = cl->outer_loop_exit();
2508     assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2509     for (uint i = 0; i < wq.size(); i++) {
2510       Node* n = wq.at(i);
2511       set_ctrl(n, new_ctrl);
2512       if (n->in(0) != nullptr) {
2513         _igvn.replace_input_of(n, 0, new_ctrl);
2514       }
2515       collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2516     }
2517   } else {
2518     Node *newhead = old_new[loop->_head->_idx];
2519     set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2520   }
2521 }
2522 
2523 //------------------------------clone_loop-------------------------------------
2524 //
2525 //                   C L O N E   A   L O O P   B O D Y
2526 //
2527 // This is the basic building block of the loop optimizations.  It clones an
2528 // entire loop body.  It makes an old_new loop body mapping; with this mapping
2529 // you can find the new-loop equivalent to an old-loop node.  All new-loop
2530 // nodes are exactly equal to their old-loop counterparts, all edges are the
2531 // same.  All exits from the old-loop now have a RegionNode that merges the
2532 // equivalent new-loop path.  This is true even for the normal "loop-exit"
2533 // condition.  All uses of loop-invariant old-loop values now come from (one
2534 // or more) Phis that merge their new-loop equivalents.
2535 //
2536 // This operation leaves the graph in an illegal state: there are two valid
2537 // control edges coming from the loop pre-header to both loop bodies.  I'll
2538 // definitely have to hack the graph after running this transform.
2539 //
2540 // From this building block I will further edit edges to perform loop peeling
2541 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2542 //
2543 // Parameter side_by_size_idom:
2544 //   When side_by_size_idom is null, the dominator tree is constructed for
2545 //      the clone loop to dominate the original.  Used in construction of
2546 //      pre-main-post loop sequence.
2547 //   When nonnull, the clone and original are side-by-side, both are
2548 //      dominated by the side_by_side_idom node.  Used in construction of
2549 //      unswitched loops.
2550 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2551                                 CloneLoopMode mode, Node* side_by_side_idom) {
2552 
2553   LoopNode* head = loop->_head->as_Loop();
2554   head->verify_strip_mined(1);
2555 
2556   if (C->do_vector_loop() && PrintOpto) {
2557     const char* mname = C->method()->name()->as_quoted_ascii();
2558     if (mname != nullptr) {
2559       tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2560     }
2561   }
2562 
2563   CloneMap& cm = C->clone_map();
2564   if (C->do_vector_loop()) {
2565     cm.set_clone_idx(cm.max_gen()+1);
2566 #ifndef PRODUCT
2567     if (PrintOpto) {
2568       tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2569       loop->dump_head();
2570     }
2571 #endif
2572   }
2573 
2574   // Step 1: Clone the loop body.  Make the old->new mapping.
2575   clone_loop_body(loop->_body, old_new, &cm);
2576 
2577   IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2578 
2579   // Step 2: Fix the edges in the new body.  If the old input is outside the
2580   // loop use it.  If the old input is INside the loop, use the corresponding
2581   // new node instead.
2582   fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2583 
2584   Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2585   clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2586 
2587   // Step 3: Now fix control uses.  Loop varying control uses have already
2588   // been fixed up (as part of all input edges in Step 2).  Loop invariant
2589   // control uses must be either an IfFalse or an IfTrue.  Make a merge
2590   // point to merge the old and new IfFalse/IfTrue nodes; make the use
2591   // refer to this.
2592   Node_List worklist;
2593   uint new_counter = C->unique();
2594   fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2595 
2596   // Step 4: If loop-invariant use is not control, it must be dominated by a
2597   // loop exit IfFalse/IfTrue.  Find "proper" loop exit.  Make a Region
2598   // there if needed.  Make a Phi there merging old and new used values.
2599   Node_List *split_if_set = nullptr;
2600   Node_List *split_bool_set = nullptr;
2601   Node_List *split_cex_set = nullptr;
2602   fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2603 
2604   for (uint i = 0; i < extra_data_nodes.size(); i++) {
2605     Node* old = extra_data_nodes.at(i);
2606     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2607                                 split_bool_set, split_cex_set, worklist, new_counter,
2608                                 mode);
2609   }
2610 
2611   // Check for IFs that need splitting/cloning.  Happens if an IF outside of
2612   // the loop uses a condition set in the loop.  The original IF probably
2613   // takes control from one or more OLD Regions (which in turn get from NEW
2614   // Regions).  In any case, there will be a set of Phis for each merge point
2615   // from the IF up to where the original BOOL def exists the loop.
2616   finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2617 
2618 }
2619 
2620 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2621   if (split_if_set) {
2622     while (split_if_set->size()) {
2623       Node *iff = split_if_set->pop();
2624       uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2625       if (iff->in(input)->is_Phi()) {
2626         Node *b = clone_iff(iff->in(input)->as_Phi());
2627         _igvn.replace_input_of(iff, input, b);
2628       }
2629     }
2630   }
2631   if (split_bool_set) {
2632     while (split_bool_set->size()) {
2633       Node *b = split_bool_set->pop();
2634       Node *phi = b->in(1);
2635       assert(phi->is_Phi(), "");
2636       CmpNode *cmp = clone_bool((PhiNode*) phi);
2637       _igvn.replace_input_of(b, 1, cmp);
2638     }
2639   }
2640   if (split_cex_set) {
2641     while (split_cex_set->size()) {
2642       Node *b = split_cex_set->pop();
2643       assert(b->in(0)->is_Region(), "");
2644       assert(b->in(1)->is_Phi(), "");
2645       assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2646       split_up(b, b->in(0), nullptr);
2647     }
2648   }
2649 }
2650 
2651 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2652                                    uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2653                                    Node_List*& split_bool_set, Node_List*& split_cex_set) {
2654   for(uint i = 0; i < body.size(); i++ ) {
2655     Node* old = body.at(i);
2656     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2657                                 split_bool_set, split_cex_set, worklist, new_counter,
2658                                 mode);
2659   }
2660 }
2661 
2662 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2663                                    Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2664   LoopNode* head = loop->_head->as_Loop();
2665   for(uint i = 0; i < body.size(); i++ ) {
2666     Node* old = body.at(i);
2667     if( !old->is_CFG() ) continue;
2668 
2669     // Copy uses to a worklist, so I can munge the def-use info
2670     // with impunity.
2671     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2672       worklist.push(old->fast_out(j));
2673     }
2674 
2675     while (worklist.size()) {  // Visit all uses
2676       Node *use = worklist.pop();
2677       if (!has_node(use))  continue; // Ignore dead nodes
2678       IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2679       if (!loop->is_member(use_loop) && use->is_CFG()) {
2680         // Both OLD and USE are CFG nodes here.
2681         assert(use->is_Proj(), "" );
2682         Node* nnn = old_new[old->_idx];
2683 
2684         Node* newuse = nullptr;
2685         if (head->is_strip_mined() && mode != IgnoreStripMined) {
2686           CountedLoopNode* cl = head->as_CountedLoop();
2687           CountedLoopEndNode* cle = cl->loopexit();
2688           // is use the projection that exits the loop from the CountedLoopEndNode?
2689           if (use->in(0) == cle) {
2690             IfFalseNode* cle_out = use->as_IfFalse();
2691             IfNode* le = cl->outer_loop_end();
2692             use = le->false_proj();
2693             use_loop = get_loop(use);
2694             if (mode == CloneIncludesStripMined) {
2695               nnn = old_new[le->_idx];
2696             } else {
2697               newuse = old_new[cle_out->_idx];
2698             }
2699           }
2700         }
2701         if (newuse == nullptr) {
2702           newuse = use->clone();
2703         }
2704 
2705         // Clone the loop exit control projection
2706         if (C->do_vector_loop() && cm != nullptr) {
2707           cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2708         }
2709         newuse->set_req(0,nnn);
2710         _igvn.register_new_node_with_optimizer(newuse);
2711         set_loop(newuse, use_loop);
2712         set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2713 
2714         // We need a Region to merge the exit from the peeled body and the
2715         // exit from the old loop body.
2716         RegionNode *r = new RegionNode(3);
2717         uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2718         assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2719 
2720         // The original user of 'use' uses 'r' instead.
2721         for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2722           Node* useuse = use->last_out(l);
2723           _igvn.rehash_node_delayed(useuse);
2724           uint uses_found = 0;
2725           if (useuse->in(0) == use) {
2726             useuse->set_req(0, r);
2727             uses_found++;
2728             if (useuse->is_CFG()) {
2729               // This is not a dom_depth > dd_r because when new
2730               // control flow is constructed by a loop opt, a node and
2731               // its dominator can end up at the same dom_depth
2732               assert(dom_depth(useuse) >= dd_r, "");
2733               set_idom(useuse, r, dom_depth(useuse));
2734             }
2735           }
2736           for (uint k = 1; k < useuse->req(); k++) {
2737             if( useuse->in(k) == use ) {
2738               useuse->set_req(k, r);
2739               uses_found++;
2740               if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2741                 // This is not a dom_depth > dd_r because when new
2742                 // control flow is constructed by a loop opt, a node
2743                 // and its dominator can end up at the same dom_depth
2744                 assert(dom_depth(useuse) >= dd_r , "");
2745                 set_idom(useuse, r, dom_depth(useuse));
2746               }
2747             }
2748           }
2749           l -= uses_found;    // we deleted 1 or more copies of this edge
2750         }
2751 
2752         assert(use->is_Proj(), "loop exit should be projection");
2753         // replace_node_and_forward_ctrl() below moves all nodes that are:
2754         // - control dependent on the loop exit or
2755         // - have control set to the loop exit
2756         // below the post-loop merge point.
2757         // replace_node_and_forward_ctrl() takes a dead control as first input.
2758         // To make it possible to use it, the loop exit projection is cloned and becomes the
2759         // new exit projection. The initial one becomes dead and is "replaced" by the region.
2760         Node* use_clone = use->clone();
2761         register_control(use_clone, use_loop, idom(use), dom_depth(use));
2762         // Now finish up 'r'
2763         r->set_req(1, newuse);
2764         r->set_req(2, use_clone);
2765         _igvn.register_new_node_with_optimizer(r);
2766         set_loop(r, use_loop);
2767         set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2768         replace_node_and_forward_ctrl(use, r);
2769         // Map the (cloned) old use to the new merge point
2770         old_new.map(use_clone->_idx, r);
2771       } // End of if a loop-exit test
2772     }
2773   }
2774 }
2775 
2776 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2777                                     IdealLoopTree* parent, bool partial) {
2778   for(uint i = 0; i < body.size(); i++ ) {
2779     Node *old = body.at(i);
2780     Node *nnn = old_new[old->_idx];
2781     // Fix CFG/Loop controlling the new node
2782     if (has_ctrl(old)) {
2783       set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2784     } else {
2785       set_loop(nnn, parent);
2786       if (old->outcnt() > 0) {
2787         Node* dom = idom(old);
2788         if (old_new[dom->_idx] != nullptr) {
2789           dom = old_new[dom->_idx];
2790           set_idom(nnn, dom, dd );
2791         }
2792       }
2793     }
2794     // Correct edges to the new node
2795     for (uint j = 0; j < nnn->req(); j++) {
2796         Node *n = nnn->in(j);
2797         if (n != nullptr) {
2798           IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2799           if (loop->is_member(old_in_loop)) {
2800             if (old_new[n->_idx] != nullptr) {
2801               nnn->set_req(j, old_new[n->_idx]);
2802             } else {
2803               assert(!body.contains(n), "");
2804               assert(partial, "node not cloned");
2805             }
2806           }
2807         }
2808     }
2809     _igvn.hash_find_insert(nnn);
2810   }
2811 }
2812 
2813 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2814   for (uint i = 0; i < body.size(); i++) {
2815     Node* old = body.at(i);
2816     Node* nnn = old->clone();
2817     old_new.map(old->_idx, nnn);
2818     if (C->do_vector_loop() && cm != nullptr) {
2819       cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2820     }
2821     _igvn.register_new_node_with_optimizer(nnn);
2822   }
2823 }
2824 
2825 
2826 //---------------------- stride_of_possible_iv -------------------------------------
2827 // Looks for an iff/bool/comp with one operand of the compare
2828 // being a cycle involving an add and a phi,
2829 // with an optional truncation (left-shift followed by a right-shift)
2830 // of the add. Returns zero if not an iv.
2831 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2832   Node* trunc1 = nullptr;
2833   Node* trunc2 = nullptr;
2834   const TypeInteger* ttype = nullptr;
2835   if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2836     return 0;
2837   }
2838   BoolNode* bl = iff->in(1)->as_Bool();
2839   Node* cmp = bl->in(1);
2840   if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2841     return 0;
2842   }
2843   // Must have an invariant operand
2844   if (ctrl_is_member(get_loop(iff), cmp->in(2))) {
2845     return 0;
2846   }
2847   Node* add2 = nullptr;
2848   Node* cmp1 = cmp->in(1);
2849   if (cmp1->is_Phi()) {
2850     // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2851     Node* phi = cmp1;
2852     for (uint i = 1; i < phi->req(); i++) {
2853       Node* in = phi->in(i);
2854       Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2855                                 &trunc1, &trunc2, &ttype, T_INT);
2856       if (add && add->in(1) == phi) {
2857         add2 = add->in(2);
2858         break;
2859       }
2860     }
2861   } else {
2862     // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2863     Node* addtrunc = cmp1;
2864     Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2865                                 &trunc1, &trunc2, &ttype, T_INT);
2866     if (add && add->in(1)->is_Phi()) {
2867       Node* phi = add->in(1);
2868       for (uint i = 1; i < phi->req(); i++) {
2869         if (phi->in(i) == addtrunc) {
2870           add2 = add->in(2);
2871           break;
2872         }
2873       }
2874     }
2875   }
2876   if (add2 != nullptr) {
2877     const TypeInt* add2t = _igvn.type(add2)->is_int();
2878     if (add2t->is_con()) {
2879       return add2t->get_con();
2880     }
2881   }
2882   return 0;
2883 }
2884 
2885 
2886 //---------------------- stay_in_loop -------------------------------------
2887 // Return the (unique) control output node that's in the loop (if it exists.)
2888 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2889   Node* unique = nullptr;
2890   if (!n) return nullptr;
2891   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2892     Node* use = n->fast_out(i);
2893     if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2894       if (unique != nullptr) {
2895         return nullptr;
2896       }
2897       unique = use;
2898     }
2899   }
2900   return unique;
2901 }
2902 
2903 //------------------------------ register_node -------------------------------------
2904 // Utility to register node "n" with PhaseIdealLoop
2905 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2906   _igvn.register_new_node_with_optimizer(n);
2907   loop->_body.push(n);
2908   if (n->is_CFG()) {
2909     set_loop(n, loop);
2910     set_idom(n, pred, ddepth);
2911   } else {
2912     set_ctrl(n, pred);
2913   }
2914 }
2915 
2916 //------------------------------ proj_clone -------------------------------------
2917 // Utility to create an if-projection
2918 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2919   ProjNode* c = p->clone()->as_Proj();
2920   c->set_req(0, iff);
2921   return c;
2922 }
2923 
2924 //------------------------------ short_circuit_if -------------------------------------
2925 // Force the iff control output to be the live_proj
2926 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2927   guarantee(live_proj != nullptr, "null projection");
2928   int proj_con = live_proj->_con;
2929   assert(proj_con == 0 || proj_con == 1, "false or true projection");
2930   Node* con = intcon(proj_con);
2931   if (iff) {
2932     iff->set_req(1, con);
2933   }
2934   return con;
2935 }
2936 
2937 //------------------------------ insert_if_before_proj -------------------------------------
2938 // Insert a new if before an if projection (* - new node)
2939 //
2940 // before
2941 //           if(test)
2942 //           /     \
2943 //          v       v
2944 //    other-proj   proj (arg)
2945 //
2946 // after
2947 //           if(test)
2948 //           /     \
2949 //          /       v
2950 //         |      * proj-clone
2951 //         v          |
2952 //    other-proj      v
2953 //                * new_if(relop(cmp[IU](left,right)))
2954 //                  /  \
2955 //                 v    v
2956 //         * new-proj  proj
2957 //         (returned)
2958 //
2959 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2960   IfNode* iff = proj->in(0)->as_If();
2961   IdealLoopTree *loop = get_loop(proj);
2962   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2963   uint ddepth = dom_depth(proj);
2964 
2965   _igvn.rehash_node_delayed(iff);
2966   _igvn.rehash_node_delayed(proj);
2967 
2968   proj->set_req(0, nullptr);  // temporary disconnect
2969   ProjNode* proj2 = proj_clone(proj, iff);
2970   register_node(proj2, loop, iff, ddepth);
2971 
2972   Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2973   register_node(cmp, loop, proj2, ddepth);
2974 
2975   BoolNode* bol = new BoolNode(cmp, relop);
2976   register_node(bol, loop, proj2, ddepth);
2977 
2978   int opcode = iff->Opcode();
2979   assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2980   IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2981   register_node(new_if, loop, proj2, ddepth);
2982 
2983   proj->set_req(0, new_if); // reattach
2984   set_idom(proj, new_if, ddepth);
2985 
2986   ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2987   guarantee(new_exit != nullptr, "null exit node");
2988   register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2989 
2990   return new_exit;
2991 }
2992 
2993 //------------------------------ insert_region_before_proj -------------------------------------
2994 // Insert a region before an if projection (* - new node)
2995 //
2996 // before
2997 //           if(test)
2998 //          /      |
2999 //         v       |
3000 //       proj      v
3001 //               other-proj
3002 //
3003 // after
3004 //           if(test)
3005 //          /      |
3006 //         v       |
3007 // * proj-clone    v
3008 //         |     other-proj
3009 //         v
3010 // * new-region
3011 //         |
3012 //         v
3013 // *      dum_if
3014 //       /     \
3015 //      v       \
3016 // * dum-proj    v
3017 //              proj
3018 //
3019 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3020   IfNode* iff = proj->in(0)->as_If();
3021   IdealLoopTree *loop = get_loop(proj);
3022   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3023   uint ddepth = dom_depth(proj);
3024 
3025   _igvn.rehash_node_delayed(iff);
3026   _igvn.rehash_node_delayed(proj);
3027 
3028   proj->set_req(0, nullptr);  // temporary disconnect
3029   ProjNode* proj2 = proj_clone(proj, iff);
3030   register_node(proj2, loop, iff, ddepth);
3031 
3032   RegionNode* reg = new RegionNode(2);
3033   reg->set_req(1, proj2);
3034   register_node(reg, loop, iff, ddepth);
3035 
3036   IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3037   register_node(dum_if, loop, reg, ddepth);
3038 
3039   proj->set_req(0, dum_if); // reattach
3040   set_idom(proj, dum_if, ddepth);
3041 
3042   ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3043   register_node(dum_proj, loop, dum_if, ddepth);
3044 
3045   return reg;
3046 }
3047 
3048 // Idea
3049 // ----
3050 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3051 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3052 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3053 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3054 // test alone instead:
3055 //
3056 // Before Partial Peeling:
3057 //   Loop:
3058 //     <peeled section>
3059 //     Split off signed loop exit test
3060 //     <-- CUT HERE -->
3061 //     Unchanged unsigned loop exit test
3062 //     <rest of unpeeled section>
3063 //     goto Loop
3064 //
3065 // After Partial Peeling:
3066 //   <cloned peeled section>
3067 //   Cloned split off signed loop exit test
3068 //   Loop:
3069 //     Unchanged unsigned loop exit test
3070 //     <rest of unpeeled section>
3071 //     <peeled section>
3072 //     Split off signed loop exit test
3073 //     goto Loop
3074 //
3075 // Details
3076 // -------
3077 // Before:
3078 //          if (i <u limit)    Unsigned loop exit condition
3079 //         /       |
3080 //        v        v
3081 //   exit-proj   stay-in-loop-proj
3082 //
3083 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3084 // before the CmpU on the stay-in-loop path and keep both tests:
3085 //
3086 //          if (i <u limit)    Signed loop exit test
3087 //        /        |
3088 //       /  if (i <u limit)    Unsigned loop exit test
3089 //      /  /       |
3090 //     v  v        v
3091 //  exit-region  stay-in-loop-proj
3092 //
3093 // Implementation
3094 // --------------
3095 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3096 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3097 // exit tests is preserved, and their loop nesting is correct.
3098 //
3099 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3100 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3101 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3102 //
3103 //           if (stay-in-loop-const)  Killed original unsigned loop exit test
3104 //          /       |
3105 //         /        v
3106 //        /  if (i <  limit)          Split off signed loop exit test
3107 //       /  /       |
3108 //      /  /        v
3109 //     /  /  if (i <u limit)          Cloned unsigned loop exit test
3110 //    /  /   /      |
3111 //   v  v  v        |
3112 //  exit-region     |
3113 //        |         |
3114 //    dummy-if      |
3115 //     /  |         |
3116 // dead   |         |
3117 //        v         v
3118 //   exit-proj   stay-in-loop-proj
3119 //
3120 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3121 //       loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3122 //
3123 // Requirements
3124 // ------------
3125 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3126 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3127 // Otherwise, we need to bail out (see comments in the code below).
3128 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3129   const bool Signed   = true;
3130   const bool Unsigned = false;
3131 
3132   BoolNode* bol = if_cmpu->in(1)->as_Bool();
3133   if (bol->_test._test != BoolTest::lt) {
3134     return nullptr;
3135   }
3136   CmpNode* cmpu = bol->in(1)->as_Cmp();
3137   assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3138 
3139   int stride = stride_of_possible_iv(if_cmpu);
3140   if (stride == 0) {
3141     return nullptr;
3142   }
3143 
3144   Node* lp_proj = stay_in_loop(if_cmpu, loop);
3145   guarantee(lp_proj != nullptr, "null loop node");
3146 
3147   ProjNode* lp_continue = lp_proj->as_Proj();
3148   ProjNode* lp_exit     = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3149   if (!lp_exit->is_IfFalse()) {
3150     // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3151     // We therefore can't add a single exit condition.
3152     return nullptr;
3153   }
3154   // The unsigned loop exit condition is
3155   //   !(i <u  limit)
3156   // =   i >=u limit
3157   //
3158   // First, we note that for any x for which
3159   //   0 <= x <= INT_MAX
3160   // we can convert x to an unsigned int and still get the same guarantee:
3161   //   0 <=  (uint) x <=  INT_MAX = (uint) INT_MAX
3162   //   0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX   (LEMMA)
3163   //
3164   // With that in mind, if
3165   //   limit >= 0             (COND)
3166   // then the unsigned loop exit condition
3167   //   i >=u limit            (ULE)
3168   // is equivalent to
3169   //   i < 0 || i >= limit    (SLE-full)
3170   // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3171   //   (uint) i >=u MAX_INT >= limit >= 0
3172   // or otherwise
3173   //   i >= limit >= 0
3174   // holds due to (LEMMA).
3175   //
3176   // For completeness, a counterexample with limit < 0:
3177   // Assume i = -3 and limit = -2:
3178   //   i  < 0
3179   //   -2 < 0
3180   // is true and thus also "i < 0 || i >= limit". But
3181   //   i  >=u limit
3182   //   -3 >=u -2
3183   // is false.
3184   Node* limit = cmpu->in(2);
3185   const TypeInt* type_limit = _igvn.type(limit)->is_int();
3186   if (type_limit->_lo < 0) {
3187     return nullptr;
3188   }
3189 
3190   // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3191   //   stride < 0:
3192   //     i < 0        (SLE = SLE-negative)
3193   //   stride > 0:
3194   //     i >= limit   (SLE = SLE-positive)
3195   // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3196   //
3197   // Loop:
3198   //   <peeled section>
3199   //   i >= limit    (SLE-positive)
3200   //   <-- CUT HERE -->
3201   //   i >=u limit   (ULE)
3202   //   <rest of unpeeled section>
3203   //   goto Loop
3204   //
3205   // We exit the loop if:
3206   //   (SLE) is true OR (ULE) is true
3207   // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3208   // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3209   //   (SLE) IMPLIES (ULE)
3210   // This indeed holds when (COND) is given:
3211   // - stride > 0:
3212   //       i >=  limit             // (SLE = SLE-positive)
3213   //       i >=  limit >= 0        // (COND)
3214   //       i >=u limit >= 0        // (LEMMA)
3215   //     which is the unsigned loop exit condition (ULE).
3216   // - stride < 0:
3217   //       i        <  0           // (SLE = SLE-negative)
3218   //       (uint) i >u MAX_INT     // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3219   //       MAX_INT >= limit >= 0   // (COND)
3220   //       MAX_INT >=u limit >= 0  // (LEMMA)
3221   //     and thus from (NEG) and (LEMMA):
3222   //       i >=u limit
3223   //     which is the unsigned loop exit condition (ULE).
3224   //
3225   //
3226   // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3227   //   <cloned peeled section>
3228   //   i >= limit (SLE-positive)
3229   //   Loop:
3230   //     i >=u limit (ULE)
3231   //     <rest of unpeeled section>
3232   //     <peeled section>
3233   //     i >= limit (SLE-positive)
3234   //     goto Loop
3235   Node* rhs_cmpi;
3236   if (stride > 0) {
3237     rhs_cmpi = limit; // For i >= limit
3238   } else {
3239     rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3240   }
3241   // Create a new region on the exit path
3242   RegionNode* reg = insert_region_before_proj(lp_exit);
3243   guarantee(reg != nullptr, "null region node");
3244 
3245   // Clone the if-cmpu-true-false using a signed compare
3246   BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3247   ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3248   reg->add_req(cmpi_exit);
3249 
3250   // Clone the if-cmpu-true-false
3251   BoolTest::mask rel_u = bol->_test._test;
3252   ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3253   reg->add_req(cmpu_exit);
3254 
3255   // Force original if to stay in loop.
3256   short_circuit_if(if_cmpu, lp_continue);
3257 
3258   return cmpi_exit->in(0)->as_If();
3259 }
3260 
3261 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3262 // Remove a previously inserted signed compare loop exit.
3263 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3264   Node* lp_proj = stay_in_loop(if_cmp, loop);
3265   assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3266          stay_in_loop(lp_proj, loop)->is_If() &&
3267          stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3268   Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3269   if_cmp->set_req(1, con);
3270 }
3271 
3272 //------------------------------ scheduled_nodelist -------------------------------------
3273 // Create a post order schedule of nodes that are in the
3274 // "member" set.  The list is returned in "sched".
3275 // The first node in "sched" is the loop head, followed by
3276 // nodes which have no inputs in the "member" set, and then
3277 // followed by the nodes that have an immediate input dependence
3278 // on a node in "sched".
3279 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3280 
3281   assert(member.test(loop->_head->_idx), "loop head must be in member set");
3282   VectorSet visited;
3283   Node_Stack nstack(loop->_body.size());
3284 
3285   Node* n  = loop->_head;  // top of stack is cached in "n"
3286   uint idx = 0;
3287   visited.set(n->_idx);
3288 
3289   // Initially push all with no inputs from within member set
3290   for(uint i = 0; i < loop->_body.size(); i++ ) {
3291     Node *elt = loop->_body.at(i);
3292     if (member.test(elt->_idx)) {
3293       bool found = false;
3294       for (uint j = 0; j < elt->req(); j++) {
3295         Node* def = elt->in(j);
3296         if (def && member.test(def->_idx) && def != elt) {
3297           found = true;
3298           break;
3299         }
3300       }
3301       if (!found && elt != loop->_head) {
3302         nstack.push(n, idx);
3303         n = elt;
3304         assert(!visited.test(n->_idx), "not seen yet");
3305         visited.set(n->_idx);
3306       }
3307     }
3308   }
3309 
3310   // traverse out's that are in the member set
3311   while (true) {
3312     if (idx < n->outcnt()) {
3313       Node* use = n->raw_out(idx);
3314       idx++;
3315       if (!visited.test_set(use->_idx)) {
3316         if (member.test(use->_idx)) {
3317           nstack.push(n, idx);
3318           n = use;
3319           idx = 0;
3320         }
3321       }
3322     } else {
3323       // All outputs processed
3324       sched.push(n);
3325       if (nstack.is_empty()) break;
3326       n   = nstack.node();
3327       idx = nstack.index();
3328       nstack.pop();
3329     }
3330   }
3331 }
3332 
3333 
3334 //------------------------------ has_use_in_set -------------------------------------
3335 // Has a use in the vector set
3336 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3337   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3338     Node* use = n->fast_out(j);
3339     if (vset.test(use->_idx)) {
3340       return true;
3341     }
3342   }
3343   return false;
3344 }
3345 
3346 
3347 //------------------------------ has_use_internal_to_set -------------------------------------
3348 // Has use internal to the vector set (ie. not in a phi at the loop head)
3349 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3350   Node* head  = loop->_head;
3351   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3352     Node* use = n->fast_out(j);
3353     if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3354       return true;
3355     }
3356   }
3357   return false;
3358 }
3359 
3360 
3361 //------------------------------ clone_for_use_outside_loop -------------------------------------
3362 // clone "n" for uses that are outside of loop
3363 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3364   int cloned = 0;
3365   assert(worklist.size() == 0, "should be empty");
3366   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3367     Node* use = n->fast_out(j);
3368     if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3369       worklist.push(use);
3370     }
3371   }
3372 
3373   if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3374                           "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3375     return -1;
3376   }
3377 
3378   while( worklist.size() ) {
3379     Node *use = worklist.pop();
3380     if (!has_node(use) || use->in(0) == C->top()) continue;
3381     uint j;
3382     for (j = 0; j < use->req(); j++) {
3383       if (use->in(j) == n) break;
3384     }
3385     assert(j < use->req(), "must be there");
3386 
3387     // clone "n" and insert it between the inputs of "n" and the use outside the loop
3388     Node* n_clone = n->clone();
3389     _igvn.replace_input_of(use, j, n_clone);
3390     cloned++;
3391     Node* use_c;
3392     if (!use->is_Phi()) {
3393       use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3394     } else {
3395       // Use in a phi is considered a use in the associated predecessor block
3396       use_c = use->in(0)->in(j);
3397     }
3398     set_ctrl(n_clone, use_c);
3399     assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3400     get_loop(use_c)->_body.push(n_clone);
3401     _igvn.register_new_node_with_optimizer(n_clone);
3402 #ifndef PRODUCT
3403     if (TracePartialPeeling) {
3404       tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3405     }
3406 #endif
3407   }
3408   return cloned;
3409 }
3410 
3411 
3412 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3413 // clone "n" for special uses that are in the not_peeled region.
3414 // If these def-uses occur in separate blocks, the code generator
3415 // marks the method as not compilable.  For example, if a "BoolNode"
3416 // is in a different basic block than the "IfNode" that uses it, then
3417 // the compilation is aborted in the code generator.
3418 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3419                                                         VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3420   if (n->is_Phi() || n->is_Load()) {
3421     return;
3422   }
3423   assert(worklist.size() == 0, "should be empty");
3424   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3425     Node* use = n->fast_out(j);
3426     if ( not_peel.test(use->_idx) &&
3427          (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3428          use->in(1) == n)  {
3429       worklist.push(use);
3430     }
3431   }
3432   if (worklist.size() > 0) {
3433     // clone "n" and insert it between inputs of "n" and the use
3434     Node* n_clone = n->clone();
3435     loop->_body.push(n_clone);
3436     _igvn.register_new_node_with_optimizer(n_clone);
3437     set_ctrl(n_clone, get_ctrl(n));
3438     sink_list.push(n_clone);
3439     not_peel.set(n_clone->_idx);
3440 #ifndef PRODUCT
3441     if (TracePartialPeeling) {
3442       tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3443     }
3444 #endif
3445     while( worklist.size() ) {
3446       Node *use = worklist.pop();
3447       _igvn.rehash_node_delayed(use);
3448       for (uint j = 1; j < use->req(); j++) {
3449         if (use->in(j) == n) {
3450           use->set_req(j, n_clone);
3451         }
3452       }
3453     }
3454   }
3455 }
3456 
3457 
3458 //------------------------------ insert_phi_for_loop -------------------------------------
3459 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3460 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3461   Node *phi = PhiNode::make(lp, back_edge_val);
3462   phi->set_req(LoopNode::EntryControl, lp_entry_val);
3463   // Use existing phi if it already exists
3464   Node *hit = _igvn.hash_find_insert(phi);
3465   if( hit == nullptr ) {
3466     _igvn.register_new_node_with_optimizer(phi);
3467     set_ctrl(phi, lp);
3468   } else {
3469     // Remove the new phi from the graph and use the hit
3470     _igvn.remove_dead_node(phi);
3471     phi = hit;
3472   }
3473   _igvn.replace_input_of(use, idx, phi);
3474 }
3475 
3476 #ifdef ASSERT
3477 //------------------------------ is_valid_loop_partition -------------------------------------
3478 // Validate the loop partition sets: peel and not_peel
3479 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3480                                               VectorSet& not_peel ) {
3481   uint i;
3482   // Check that peel_list entries are in the peel set
3483   for (i = 0; i < peel_list.size(); i++) {
3484     if (!peel.test(peel_list.at(i)->_idx)) {
3485       return false;
3486     }
3487   }
3488   // Check at loop members are in one of peel set or not_peel set
3489   for (i = 0; i < loop->_body.size(); i++ ) {
3490     Node *def  = loop->_body.at(i);
3491     uint di = def->_idx;
3492     // Check that peel set elements are in peel_list
3493     if (peel.test(di)) {
3494       if (not_peel.test(di)) {
3495         return false;
3496       }
3497       // Must be in peel_list also
3498       bool found = false;
3499       for (uint j = 0; j < peel_list.size(); j++) {
3500         if (peel_list.at(j)->_idx == di) {
3501           found = true;
3502           break;
3503         }
3504       }
3505       if (!found) {
3506         return false;
3507       }
3508     } else if (not_peel.test(di)) {
3509       if (peel.test(di)) {
3510         return false;
3511       }
3512     } else {
3513       return false;
3514     }
3515   }
3516   return true;
3517 }
3518 
3519 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3520 // Ensure a use outside of loop is of the right form
3521 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3522   Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3523   return (use->is_Phi() &&
3524           use_c->is_Region() && use_c->req() == 3 &&
3525           (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3526            use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3527            use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3528           loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3529 }
3530 
3531 //------------------------------ is_valid_clone_loop_form -------------------------------------
3532 // Ensure that all uses outside of loop are of the right form
3533 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3534                                                uint orig_exit_idx, uint clone_exit_idx) {
3535   uint len = peel_list.size();
3536   for (uint i = 0; i < len; i++) {
3537     Node *def = peel_list.at(i);
3538 
3539     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3540       Node *use = def->fast_out(j);
3541       Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3542       if (!loop->is_member(get_loop(use_c))) {
3543         // use is not in the loop, check for correct structure
3544         if (use->in(0) == def) {
3545           // Okay
3546         } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3547           return false;
3548         }
3549       }
3550     }
3551   }
3552   return true;
3553 }
3554 #endif
3555 
3556 //------------------------------ partial_peel -------------------------------------
3557 // Partially peel (aka loop rotation) the top portion of a loop (called
3558 // the peel section below) by cloning it and placing one copy just before
3559 // the new loop head and the other copy at the bottom of the new loop.
3560 //
3561 //    before                       after                where it came from
3562 //
3563 //    stmt1                        stmt1
3564 //  loop:                          stmt2                     clone
3565 //    stmt2                        if condA goto exitA       clone
3566 //    if condA goto exitA        new_loop:                   new
3567 //    stmt3                        stmt3                     clone
3568 //    if !condB goto loop          if condB goto exitB       clone
3569 //  exitB:                         stmt2                     orig
3570 //    stmt4                        if !condA goto new_loop   orig
3571 //  exitA:                         goto exitA
3572 //                               exitB:
3573 //                                 stmt4
3574 //                               exitA:
3575 //
3576 // Step 1: find the cut point: an exit test on probable
3577 //         induction variable.
3578 // Step 2: schedule (with cloning) operations in the peel
3579 //         section that can be executed after the cut into
3580 //         the section that is not peeled.  This may need
3581 //         to clone operations into exit blocks.  For
3582 //         instance, a reference to A[i] in the not-peel
3583 //         section and a reference to B[i] in an exit block
3584 //         may cause a left-shift of i by 2 to be placed
3585 //         in the peel block.  This step will clone the left
3586 //         shift into the exit block and sink the left shift
3587 //         from the peel to the not-peel section.
3588 // Step 3: clone the loop, retarget the control, and insert
3589 //         phis for values that are live across the new loop
3590 //         head.  This is very dependent on the graph structure
3591 //         from clone_loop.  It creates region nodes for
3592 //         exit control and associated phi nodes for values
3593 //         flow out of the loop through that exit.  The region
3594 //         node is dominated by the clone's control projection.
3595 //         So the clone's peel section is placed before the
3596 //         new loop head, and the clone's not-peel section is
3597 //         forms the top part of the new loop.  The original
3598 //         peel section forms the tail of the new loop.
3599 // Step 4: update the dominator tree and recompute the
3600 //         dominator depth.
3601 //
3602 //                   orig
3603 //
3604 //                   stmt1
3605 //                     |
3606 //                     v
3607 //                 predicates
3608 //                     |
3609 //                     v
3610 //                   loop<----+
3611 //                     |      |
3612 //                   stmt2    |
3613 //                     |      |
3614 //                     v      |
3615 //                    ifA     |
3616 //                   / |      |
3617 //                  v  v      |
3618 //               false true   ^  <-- last_peel
3619 //               /     |      |
3620 //              /   ===|==cut |
3621 //             /     stmt3    |  <-- first_not_peel
3622 //            /        |      |
3623 //            |        v      |
3624 //            v       ifB     |
3625 //          exitA:   / \      |
3626 //                  /   \     |
3627 //                 v     v    |
3628 //               false true   |
3629 //               /       \    |
3630 //              /         ----+
3631 //             |
3632 //             v
3633 //           exitB:
3634 //           stmt4
3635 //
3636 //
3637 //            after clone loop
3638 //
3639 //                   stmt1
3640 //                     |
3641 //                     v
3642 //                predicates
3643 //                 /       \
3644 //        clone   /         \   orig
3645 //               /           \
3646 //              /             \
3647 //             v               v
3648 //   +---->loop                loop<----+
3649 //   |      |                    |      |
3650 //   |    stmt2                stmt2    |
3651 //   |      |                    |      |
3652 //   |      v                    v      |
3653 //   |      ifA                 ifA     |
3654 //   |      | \                / |      |
3655 //   |      v  v              v  v      |
3656 //   ^    true  false      false true   ^  <-- last_peel
3657 //   |      |   ^   \       /    |      |
3658 //   | cut==|==  \   \     /  ===|==cut |
3659 //   |    stmt3   \   \   /    stmt3    |  <-- first_not_peel
3660 //   |      |    dom   | |       |      |
3661 //   |      v      \  1v v2      v      |
3662 //   |      ifB     regionA     ifB     |
3663 //   |      / \        |       / \      |
3664 //   |     /   \       v      /   \     |
3665 //   |    v     v    exitA:  v     v    |
3666 //   |    true  false      false true   |
3667 //   |    /     ^   \      /       \    |
3668 //   +----       \   \    /         ----+
3669 //               dom  \  /
3670 //                 \  1v v2
3671 //                  regionB
3672 //                     |
3673 //                     v
3674 //                   exitB:
3675 //                   stmt4
3676 //
3677 //
3678 //           after partial peel
3679 //
3680 //                  stmt1
3681 //                     |
3682 //                     v
3683 //                predicates
3684 //                 /
3685 //        clone   /             orig
3686 //               /          TOP
3687 //              /             \
3688 //             v               v
3689 //    TOP->loop                loop----+
3690 //          |                    |      |
3691 //        stmt2                stmt2    |
3692 //          |                    |      |
3693 //          v                    v      |
3694 //          ifA                 ifA     |
3695 //          | \                / |      |
3696 //          v  v              v  v      |
3697 //        true  false      false true   |     <-- last_peel
3698 //          |   ^   \       /    +------|---+
3699 //  +->newloop   \   \     /  === ==cut |   |
3700 //  |     stmt3   \   \   /     TOP     |   |
3701 //  |       |    dom   | |      stmt3   |   | <-- first_not_peel
3702 //  |       v      \  1v v2      v      |   |
3703 //  |       ifB     regionA     ifB     ^   v
3704 //  |       / \        |       / \      |   |
3705 //  |      /   \       v      /   \     |   |
3706 //  |     v     v    exitA:  v     v    |   |
3707 //  |     true  false      false true   |   |
3708 //  |     /     ^   \      /       \    |   |
3709 //  |    |       \   \    /         v   |   |
3710 //  |    |       dom  \  /         TOP  |   |
3711 //  |    |         \  1v v2             |   |
3712 //  ^    v          regionB             |   |
3713 //  |    |             |                |   |
3714 //  |    |             v                ^   v
3715 //  |    |           exitB:             |   |
3716 //  |    |           stmt4              |   |
3717 //  |    +------------>-----------------+   |
3718 //  |                                       |
3719 //  +-----------------<---------------------+
3720 //
3721 //
3722 //              final graph
3723 //
3724 //                  stmt1
3725 //                    |
3726 //                    v
3727 //                predicates
3728 //                    |
3729 //                    v
3730 //                  stmt2 clone
3731 //                    |
3732 //                    v
3733 //         ........> ifA clone
3734 //         :        / |
3735 //        dom      /  |
3736 //         :      v   v
3737 //         :  false   true
3738 //         :  |       |
3739 //         :  |       v
3740 //         :  |    newloop<-----+
3741 //         :  |        |        |
3742 //         :  |     stmt3 clone |
3743 //         :  |        |        |
3744 //         :  |        v        |
3745 //         :  |       ifB       |
3746 //         :  |      / \        |
3747 //         :  |     v   v       |
3748 //         :  |  false true     |
3749 //         :  |   |     |       |
3750 //         :  |   v    stmt2    |
3751 //         :  | exitB:  |       |
3752 //         :  | stmt4   v       |
3753 //         :  |       ifA orig  |
3754 //         :  |      /  \       |
3755 //         :  |     /    \      |
3756 //         :  |    v     v      |
3757 //         :  |  false  true    |
3758 //         :  |  /        \     |
3759 //         :  v  v         -----+
3760 //          RegionA
3761 //             |
3762 //             v
3763 //           exitA
3764 //
3765 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3766 
3767   assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3768   if (!loop->_head->is_Loop()) {
3769     return false;
3770   }
3771   LoopNode *head = loop->_head->as_Loop();
3772 
3773   if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3774     return false;
3775   }
3776 
3777   // Check for complex exit control
3778   for (uint ii = 0; ii < loop->_body.size(); ii++) {
3779     Node *n = loop->_body.at(ii);
3780     int opc = n->Opcode();
3781     if (n->is_Call()        ||
3782         opc == Op_Catch     ||
3783         opc == Op_CatchProj ||
3784         opc == Op_Jump      ||
3785         opc == Op_JumpProj) {
3786 #ifndef PRODUCT
3787       if (TracePartialPeeling) {
3788         tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3789       }
3790 #endif
3791       return false;
3792     }
3793   }
3794 
3795   int dd = dom_depth(head);
3796 
3797   // Step 1: find cut point
3798 
3799   // Walk up dominators to loop head looking for first loop exit
3800   // which is executed on every path thru loop.
3801   IfNode *peel_if = nullptr;
3802   IfNode *peel_if_cmpu = nullptr;
3803 
3804   Node *iff = loop->tail();
3805   while (iff != head) {
3806     if (iff->is_If()) {
3807       Node *ctrl = get_ctrl(iff->in(1));
3808       if (ctrl->is_top()) return false; // Dead test on live IF.
3809       // If loop-varying exit-test, check for induction variable
3810       if (loop->is_member(get_loop(ctrl)) &&
3811           loop->is_loop_exit(iff) &&
3812           is_possible_iv_test(iff)) {
3813         Node* cmp = iff->in(1)->in(1);
3814         if (cmp->Opcode() == Op_CmpI) {
3815           peel_if = iff->as_If();
3816         } else {
3817           assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3818           peel_if_cmpu = iff->as_If();
3819         }
3820       }
3821     }
3822     iff = idom(iff);
3823   }
3824 
3825   // Prefer signed compare over unsigned compare.
3826   IfNode* new_peel_if = nullptr;
3827   if (peel_if == nullptr) {
3828     if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3829       return false;   // No peel point found
3830     }
3831     new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3832     if (new_peel_if == nullptr) {
3833       return false;   // No peel point found
3834     }
3835     peel_if = new_peel_if;
3836   }
3837   Node* last_peel        = stay_in_loop(peel_if, loop);
3838   Node* first_not_peeled = stay_in_loop(last_peel, loop);
3839   if (first_not_peeled == nullptr || first_not_peeled == head) {
3840     return false;
3841   }
3842 
3843 #ifndef PRODUCT
3844   if (TraceLoopOpts) {
3845     tty->print("PartialPeel  ");
3846     loop->dump_head();
3847   }
3848 
3849   if (TracePartialPeeling) {
3850     tty->print_cr("before partial peel one iteration");
3851     Node_List wl;
3852     Node* t = head->in(2);
3853     while (true) {
3854       wl.push(t);
3855       if (t == head) break;
3856       t = idom(t);
3857     }
3858     while (wl.size() > 0) {
3859       Node* tt = wl.pop();
3860       tt->dump();
3861       if (tt == last_peel) tty->print_cr("-- cut --");
3862     }
3863   }
3864 #endif
3865 
3866   C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3867 
3868   VectorSet peel;
3869   VectorSet not_peel;
3870   Node_List peel_list;
3871   Node_List worklist;
3872   Node_List sink_list;
3873 
3874   uint estimate = loop->est_loop_clone_sz(1);
3875   if (exceeding_node_budget(estimate)) {
3876     return false;
3877   }
3878 
3879   // Set of cfg nodes to peel are those that are executable from
3880   // the head through last_peel.
3881   assert(worklist.size() == 0, "should be empty");
3882   worklist.push(head);
3883   peel.set(head->_idx);
3884   while (worklist.size() > 0) {
3885     Node *n = worklist.pop();
3886     if (n != last_peel) {
3887       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3888         Node* use = n->fast_out(j);
3889         if (use->is_CFG() &&
3890             loop->is_member(get_loop(use)) &&
3891             !peel.test_set(use->_idx)) {
3892           worklist.push(use);
3893         }
3894       }
3895     }
3896   }
3897 
3898   // Set of non-cfg nodes to peel are those that are control
3899   // dependent on the cfg nodes.
3900   for (uint i = 0; i < loop->_body.size(); i++) {
3901     Node *n = loop->_body.at(i);
3902     Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3903     if (peel.test(n_c->_idx)) {
3904       peel.set(n->_idx);
3905     } else {
3906       not_peel.set(n->_idx);
3907     }
3908   }
3909 
3910   // Step 2: move operations from the peeled section down into the
3911   //         not-peeled section
3912 
3913   // Get a post order schedule of nodes in the peel region
3914   // Result in right-most operand.
3915   scheduled_nodelist(loop, peel, peel_list);
3916 
3917   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3918 
3919   // For future check for too many new phis
3920   uint old_phi_cnt = 0;
3921   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3922     Node* use = head->fast_out(j);
3923     if (use->is_Phi()) old_phi_cnt++;
3924   }
3925 
3926 #ifndef PRODUCT
3927   if (TracePartialPeeling) {
3928     tty->print_cr("\npeeled list");
3929   }
3930 #endif
3931 
3932   // Evacuate nodes in peel region into the not_peeled region if possible
3933   bool too_many_clones = false;
3934   uint new_phi_cnt = 0;
3935   uint cloned_for_outside_use = 0;
3936   for (uint i = 0; i < peel_list.size();) {
3937     Node* n = peel_list.at(i);
3938 #ifndef PRODUCT
3939   if (TracePartialPeeling) n->dump();
3940 #endif
3941     bool incr = true;
3942     if (!n->is_CFG()) {
3943       if (has_use_in_set(n, not_peel)) {
3944         // If not used internal to the peeled region,
3945         // move "n" from peeled to not_peeled region.
3946         if (!has_use_internal_to_set(n, peel, loop)) {
3947           // if not pinned and not a load (which maybe anti-dependent on a store)
3948           // and not a CMove (Matcher expects only bool->cmove).
3949           if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3950             int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3951             if (C->failing()) return false;
3952             if (new_clones == -1) {
3953               too_many_clones = true;
3954               break;
3955             }
3956             cloned_for_outside_use += new_clones;
3957             sink_list.push(n);
3958             peel.remove(n->_idx);
3959             not_peel.set(n->_idx);
3960             peel_list.remove(i);
3961             incr = false;
3962 #ifndef PRODUCT
3963             if (TracePartialPeeling) {
3964               tty->print_cr("sink to not_peeled region: %d newbb: %d",
3965                             n->_idx, get_ctrl(n)->_idx);
3966             }
3967 #endif
3968           }
3969         } else {
3970           // Otherwise check for special def-use cases that span
3971           // the peel/not_peel boundary such as bool->if
3972           clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3973           new_phi_cnt++;
3974         }
3975       }
3976     }
3977     if (incr) i++;
3978   }
3979 
3980   estimate += cloned_for_outside_use + new_phi_cnt;
3981   bool exceed_node_budget = !may_require_nodes(estimate);
3982   bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3983 
3984   if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3985 #ifndef PRODUCT
3986     if (TracePartialPeeling && exceed_phi_limit) {
3987       tty->print_cr("\nToo many new phis: %d  old %d new cmpi: %c",
3988                     new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3989     }
3990 #endif
3991     if (new_peel_if != nullptr) {
3992       remove_cmpi_loop_exit(new_peel_if, loop);
3993     }
3994     // Inhibit more partial peeling on this loop
3995     assert(!head->is_partial_peel_loop(), "not partial peeled");
3996     head->mark_partial_peel_failed();
3997     if (cloned_for_outside_use > 0) {
3998       // Terminate this round of loop opts because
3999       // the graph outside this loop was changed.
4000       C->set_major_progress();
4001       return true;
4002     }
4003     return false;
4004   }
4005 
4006   // Step 3: clone loop, retarget control, and insert new phis
4007 
4008   // Create new loop head for new phis and to hang
4009   // the nodes being moved (sinked) from the peel region.
4010   LoopNode* new_head = new LoopNode(last_peel, last_peel);
4011   new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4012   _igvn.register_new_node_with_optimizer(new_head);
4013   assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4014   _igvn.replace_input_of(first_not_peeled, 0, new_head);
4015   set_loop(new_head, loop);
4016   loop->_body.push(new_head);
4017   not_peel.set(new_head->_idx);
4018   set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4019   set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4020 
4021   while (sink_list.size() > 0) {
4022     Node* n = sink_list.pop();
4023     set_ctrl(n, new_head);
4024   }
4025 
4026   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4027 
4028   clone_loop(loop, old_new, dd, IgnoreStripMined);
4029 
4030   const uint clone_exit_idx = 1;
4031   const uint orig_exit_idx  = 2;
4032   assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4033 
4034   Node* head_clone             = old_new[head->_idx];
4035   LoopNode* new_head_clone     = old_new[new_head->_idx]->as_Loop();
4036   Node* orig_tail_clone        = head_clone->in(2);
4037 
4038   // Add phi if "def" node is in peel set and "use" is not
4039 
4040   for (uint i = 0; i < peel_list.size(); i++) {
4041     Node *def  = peel_list.at(i);
4042     if (!def->is_CFG()) {
4043       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4044         Node *use = def->fast_out(j);
4045         if (has_node(use) && use->in(0) != C->top() &&
4046             (!peel.test(use->_idx) ||
4047              (use->is_Phi() && use->in(0) == head)) ) {
4048           worklist.push(use);
4049         }
4050       }
4051       while( worklist.size() ) {
4052         Node *use = worklist.pop();
4053         for (uint j = 1; j < use->req(); j++) {
4054           Node* n = use->in(j);
4055           if (n == def) {
4056 
4057             // "def" is in peel set, "use" is not in peel set
4058             // or "use" is in the entry boundary (a phi) of the peel set
4059 
4060             Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4061 
4062             if ( loop->is_member(get_loop( use_c )) ) {
4063               // use is in loop
4064               if (old_new[use->_idx] != nullptr) { // null for dead code
4065                 Node* use_clone = old_new[use->_idx];
4066                 _igvn.replace_input_of(use, j, C->top());
4067                 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4068               }
4069             } else {
4070               assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4071               // use is not in the loop, check if the live range includes the cut
4072               Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4073               if (not_peel.test(lp_if->_idx)) {
4074                 assert(j == orig_exit_idx, "use from original loop");
4075                 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4076               }
4077             }
4078           }
4079         }
4080       }
4081     }
4082   }
4083 
4084   // Step 3b: retarget control
4085 
4086   // Redirect control to the new loop head if a cloned node in
4087   // the not_peeled region has control that points into the peeled region.
4088   // This necessary because the cloned peeled region will be outside
4089   // the loop.
4090   //                            from    to
4091   //          cloned-peeled    <---+
4092   //    new_head_clone:            |    <--+
4093   //          cloned-not_peeled  in(0)    in(0)
4094   //          orig-peeled
4095 
4096   for (uint i = 0; i < loop->_body.size(); i++) {
4097     Node *n = loop->_body.at(i);
4098     if (!n->is_CFG()           && n->in(0) != nullptr        &&
4099         not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4100       Node* n_clone = old_new[n->_idx];
4101       if (n_clone->depends_only_on_test()) {
4102         // If this node depends_only_on_test, it will be rewire to the loop head, which is not the
4103         // correct test
4104         Node* pinned_clone = n_clone->pin_node_under_control();
4105         if (pinned_clone != nullptr) {
4106           register_new_node_with_ctrl_of(pinned_clone, n_clone);
4107           old_new.map(n->_idx, pinned_clone);
4108           _igvn.replace_node(n_clone, pinned_clone);
4109           n_clone = pinned_clone;
4110         }
4111       }
4112       _igvn.replace_input_of(n_clone, 0, new_head_clone);
4113     }
4114   }
4115 
4116   // Backedge of the surviving new_head (the clone) is original last_peel
4117   _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4118 
4119   // Cut first node in original not_peel set
4120   _igvn.rehash_node_delayed(new_head);                     // Multiple edge updates:
4121   new_head->set_req(LoopNode::EntryControl,    C->top());  //   use rehash_node_delayed / set_req instead of
4122   new_head->set_req(LoopNode::LoopBackControl, C->top());  //   multiple replace_input_of calls
4123 
4124   // Copy head_clone back-branch info to original head
4125   // and remove original head's loop entry and
4126   // clone head's back-branch
4127   _igvn.rehash_node_delayed(head); // Multiple edge updates
4128   head->set_req(LoopNode::EntryControl,    head_clone->in(LoopNode::LoopBackControl));
4129   head->set_req(LoopNode::LoopBackControl, C->top());
4130   _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4131 
4132   // Similarly modify the phis
4133   for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4134     Node* use = head->fast_out(k);
4135     if (use->is_Phi() && use->outcnt() > 0) {
4136       Node* use_clone = old_new[use->_idx];
4137       _igvn.rehash_node_delayed(use); // Multiple edge updates
4138       use->set_req(LoopNode::EntryControl,    use_clone->in(LoopNode::LoopBackControl));
4139       use->set_req(LoopNode::LoopBackControl, C->top());
4140       _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4141     }
4142   }
4143 
4144   // Step 4: update dominator tree and dominator depth
4145 
4146   set_idom(head, orig_tail_clone, dd);
4147   recompute_dom_depth();
4148 
4149   // Inhibit more partial peeling on this loop
4150   new_head_clone->set_partial_peel_loop();
4151   C->set_major_progress();
4152   loop->record_for_igvn();
4153 
4154 #ifndef PRODUCT
4155   if (TracePartialPeeling) {
4156     tty->print_cr("\nafter partial peel one iteration");
4157     Node_List wl;
4158     Node* t = last_peel;
4159     while (true) {
4160       wl.push(t);
4161       if (t == head_clone) break;
4162       t = idom(t);
4163     }
4164     while (wl.size() > 0) {
4165       Node* tt = wl.pop();
4166       if (tt == head) tty->print_cr("orig head");
4167       else if (tt == new_head_clone) tty->print_cr("new head");
4168       else if (tt == head_clone) tty->print_cr("clone head");
4169       tt->dump();
4170     }
4171   }
4172 #endif
4173 
4174   C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4175 
4176   return true;
4177 }
4178 
4179 #ifdef ASSERT
4180 
4181 // Moves Template Assertion Predicates to a target loop by cloning and killing the old ones. The target loop is the
4182 // original, not-cloned loop. This is currently only used with StressLoopBackedge which is a develop flag only and
4183 // false with product builds. We can therefore guard it with an ifdef. More details can be found at the use-site.
4184 class MoveAssertionPredicatesVisitor : public PredicateVisitor {
4185   ClonePredicateToTargetLoop _clone_predicate_to_loop;
4186   PhaseIdealLoop* const _phase;
4187 
4188 public:
4189   MoveAssertionPredicatesVisitor(LoopNode* target_loop_head,
4190                                  const NodeInSingleLoopBody &node_in_loop_body,
4191                                  PhaseIdealLoop* phase)
4192     : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
4193       _phase(phase) {
4194   }
4195   NONCOPYABLE(MoveAssertionPredicatesVisitor);
4196 
4197   using PredicateVisitor::visit;
4198 
4199   void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
4200     _clone_predicate_to_loop.clone_template_assertion_predicate(template_assertion_predicate);
4201     template_assertion_predicate.kill(_phase->igvn());
4202   }
4203 };
4204 #endif // ASSERT
4205 
4206 // Transform:
4207 //
4208 // loop<-----------------+
4209 //  |                    |
4210 // stmt1 stmt2 .. stmtn  |
4211 //  |     |        |     |
4212 //  \     |       /      |
4213 //    v   v     v        |
4214 //       region          |
4215 //         |             |
4216 //     shared_stmt       |
4217 //         |             |
4218 //         v             |
4219 //         if            |
4220 //         / \           |
4221 //        |   -----------+
4222 //        v
4223 //
4224 // into:
4225 //
4226 //    loop<-------------------+
4227 //     |                      |
4228 //     v                      |
4229 // +->loop                    |
4230 // |   |                      |
4231 // |  stmt1 stmt2 .. stmtn    |
4232 // |   |     |        |       |
4233 // |   |      \       /       |
4234 // |   |       v     v        |
4235 // |   |        region1       |
4236 // |   |           |          |
4237 // |  shared_stmt shared_stmt |
4238 // |   |           |          |
4239 // |   v           v          |
4240 // |   if          if         |
4241 // |   /\          / \        |
4242 // +--   |         |   -------+
4243 //       \         /
4244 //        v       v
4245 //         region2
4246 //
4247 // (region2 is shown to merge mirrored projections of the loop exit
4248 // ifs to make the diagram clearer but they really merge the same
4249 // projection)
4250 //
4251 // Conditions for this transformation to trigger:
4252 // - the path through stmt1 is frequent enough
4253 // - the inner loop will be turned into a counted loop after transformation
4254 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4255   if (!DuplicateBackedge) {
4256     return false;
4257   }
4258   assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4259   if (!loop->_head->is_Loop()) {
4260     return false;
4261   }
4262 
4263   uint estimate = loop->est_loop_clone_sz(1);
4264   if (exceeding_node_budget(estimate)) {
4265     return false;
4266   }
4267 
4268   LoopNode *head = loop->_head->as_Loop();
4269 
4270   Node* region = nullptr;
4271   IfNode* exit_test = nullptr;
4272   uint inner;
4273   float f;
4274 #ifdef ASSERT
4275   if (StressDuplicateBackedge) {
4276     if (head->is_strip_mined()) {
4277       return false;
4278     }
4279     Node* c = head->in(LoopNode::LoopBackControl);
4280 
4281     while (c != head) {
4282       if (c->is_Region()) {
4283         region = c;
4284       }
4285       c = idom(c);
4286     }
4287 
4288     if (region == nullptr) {
4289       return false;
4290     }
4291 
4292     inner = 1;
4293   } else
4294 #endif //ASSERT
4295   {
4296     // Is the shape of the loop that of a counted loop...
4297     Node* back_control = loop_exit_control(head, loop);
4298     if (back_control == nullptr) {
4299       return false;
4300     }
4301 
4302     BoolTest::mask bt = BoolTest::illegal;
4303     float cl_prob = 0;
4304     Node* incr = nullptr;
4305     Node* limit = nullptr;
4306     Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4307     if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4308       return false;
4309     }
4310 
4311     // With an extra phi for the candidate iv?
4312     // Or the region node is the loop head
4313     if (!incr->is_Phi() || incr->in(0) == head) {
4314       return false;
4315     }
4316 
4317     PathFrequency pf(head, this);
4318     region = incr->in(0);
4319 
4320     // Go over all paths for the extra phi's region and see if that
4321     // path is frequent enough and would match the expected iv shape
4322     // if the extra phi is removed
4323     inner = 0;
4324     for (uint i = 1; i < incr->req(); ++i) {
4325       Node* in = incr->in(i);
4326       Node* trunc1 = nullptr;
4327       Node* trunc2 = nullptr;
4328       const TypeInteger* iv_trunc_t = nullptr;
4329       Node* orig_in = in;
4330       if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4331         continue;
4332       }
4333       assert(in->Opcode() == Op_AddI, "wrong increment code");
4334       Node* xphi = nullptr;
4335       Node* stride = loop_iv_stride(in, xphi);
4336 
4337       if (stride == nullptr) {
4338         continue;
4339       }
4340 
4341       PhiNode* phi = loop_iv_phi(xphi, nullptr, head);
4342       if (phi == nullptr ||
4343           (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4344           (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4345         return false;
4346       }
4347 
4348       f = pf.to(region->in(i));
4349       if (f > 0.5) {
4350         inner = i;
4351         break;
4352       }
4353     }
4354 
4355     if (inner == 0) {
4356       return false;
4357     }
4358 
4359     exit_test = back_control->in(0)->as_If();
4360   }
4361 
4362   if (idom(region)->is_Catch()) {
4363     return false;
4364   }
4365 
4366   // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4367   Unique_Node_List wq;
4368   wq.push(head->in(LoopNode::LoopBackControl));
4369   for (uint i = 0; i < wq.size(); i++) {
4370     Node* c = wq.at(i);
4371     assert(get_loop(c) == loop, "not in the right loop?");
4372     if (c->is_Region()) {
4373       if (c != region) {
4374         for (uint j = 1; j < c->req(); ++j) {
4375           wq.push(c->in(j));
4376         }
4377       }
4378     } else {
4379       wq.push(c->in(0));
4380     }
4381     assert(!is_strict_dominator(c, region), "shouldn't go above region");
4382   }
4383 
4384   Node* region_dom = idom(region);
4385 
4386   // Can't do the transformation if this would cause a membar pair to
4387   // be split
4388   for (uint i = 0; i < wq.size(); i++) {
4389     Node* c = wq.at(i);
4390     if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4391       assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4392       if (!wq.member(c->as_MemBar()->leading_membar())) {
4393         return false;
4394       }
4395     }
4396   }
4397   C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head);
4398 
4399   // Collect data nodes that need to be clones as well
4400   int dd = dom_depth(head);
4401 
4402   for (uint i = 0; i < loop->_body.size(); ++i) {
4403     Node* n = loop->_body.at(i);
4404     if (has_ctrl(n)) {
4405       Node* c = get_ctrl(n);
4406       if (wq.member(c)) {
4407         wq.push(n);
4408       }
4409     } else {
4410       set_idom(n, idom(n), dd);
4411     }
4412   }
4413 
4414   // clone shared_stmt
4415   clone_loop_body(wq, old_new, nullptr);
4416 
4417   Node* region_clone = old_new[region->_idx];
4418   region_clone->set_req(inner, C->top());
4419   set_idom(region, region->in(inner), dd);
4420 
4421   // Prepare the outer loop
4422   Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4423   register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4424   _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4425   set_idom(head, outer_head, dd);
4426 
4427   fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4428 
4429   // Make one of the shared_stmt copies only reachable from stmt1, the
4430   // other only from stmt2..stmtn.
4431   Node* dom = nullptr;
4432   for (uint i = 1; i < region->req(); ++i) {
4433     if (i != inner) {
4434       _igvn.replace_input_of(region, i, C->top());
4435     }
4436     Node* in = region_clone->in(i);
4437     if (in->is_top()) {
4438       continue;
4439     }
4440     if (dom == nullptr) {
4441       dom = in;
4442     } else {
4443       dom = dom_lca(dom, in);
4444     }
4445   }
4446 
4447   set_idom(region_clone, dom, dd);
4448 
4449   // Set up the outer loop
4450   for (uint i = 0; i < head->outcnt(); i++) {
4451     Node* u = head->raw_out(i);
4452     if (u->is_Phi()) {
4453       Node* outer_phi = u->clone();
4454       outer_phi->set_req(0, outer_head);
4455       Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4456       if (backedge == nullptr) {
4457         backedge = u->in(LoopNode::LoopBackControl);
4458       }
4459       outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4460       register_new_node(outer_phi, outer_head);
4461       _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4462     }
4463   }
4464 
4465   // create control and data nodes for out of loop uses (including region2)
4466   Node_List worklist;
4467   uint new_counter = C->unique();
4468   fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4469 
4470   Node_List *split_if_set = nullptr;
4471   Node_List *split_bool_set = nullptr;
4472   Node_List *split_cex_set = nullptr;
4473   fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4474                 split_if_set, split_bool_set, split_cex_set);
4475 
4476   finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4477 
4478   if (exit_test != nullptr) {
4479     float cnt = exit_test->_fcnt;
4480     if (cnt != COUNT_UNKNOWN) {
4481       exit_test->_fcnt = cnt * f;
4482       old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4483     }
4484   }
4485 
4486 #ifdef ASSERT
4487   if (StressDuplicateBackedge && head->is_CountedLoop()) {
4488     // The Template Assertion Predicates from the old counted loop are now at the new outer loop - clone them to
4489     // the inner counted loop and kill the old ones. We only need to do this with debug builds because
4490     // StressDuplicateBackedge is a devlop flag and false by default. Without StressDuplicateBackedge 'head' will be a
4491     // non-counted loop, and thus we have no Template Assertion Predicates above the old loop to move down.
4492     PredicateIterator predicate_iterator(outer_head->in(LoopNode::EntryControl));
4493     NodeInSingleLoopBody node_in_body(this, loop);
4494     MoveAssertionPredicatesVisitor move_assertion_predicates_visitor(head, node_in_body, this);
4495     predicate_iterator.for_each(move_assertion_predicates_visitor);
4496   }
4497 #endif // ASSERT
4498 
4499   C->set_major_progress();
4500 
4501   C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head);
4502 
4503   return true;
4504 }
4505 
4506 // AutoVectorize the loop: replace scalar ops with vector ops.
4507 PhaseIdealLoop::AutoVectorizeStatus
4508 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4509   // Counted loop only
4510   if (!lpt->is_counted()) {
4511     return AutoVectorizeStatus::Impossible;
4512   }
4513 
4514   // Main-loop only
4515   CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4516   if (!cl->is_main_loop()) {
4517     return AutoVectorizeStatus::Impossible;
4518   }
4519 
4520   VLoop vloop(lpt, false);
4521   if (!vloop.check_preconditions()) {
4522     return AutoVectorizeStatus::TriedAndFailed;
4523   }
4524 
4525   // Ensure the shared data is cleared before each use
4526   vshared.clear();
4527 
4528   const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4529   if (!vloop_analyzer.success()) {
4530     return AutoVectorizeStatus::TriedAndFailed;
4531   }
4532 
4533   SuperWord sw(vloop_analyzer);
4534   if (!sw.transform_loop()) {
4535     return AutoVectorizeStatus::TriedAndFailed;
4536   }
4537 
4538   return AutoVectorizeStatus::Success;
4539 }
4540 
4541 // Just before insert_pre_post_loops, we can multiversion the loop:
4542 //
4543 //              multiversion_if
4544 //               |       |
4545 //         fast_loop   slow_loop
4546 //
4547 // In the fast_loop we can make speculative assumptions, and put the
4548 // conditions into the multiversion_if. If the conditions hold at runtime,
4549 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4550 // instead which does not make any of the speculative assumptions.
4551 //
4552 // Note: we only multiversion the loop if the loop does not have any
4553 //       auto vectorization check Predicate. If we have that predicate,
4554 //       then we can simply add the speculative assumption checks to
4555 //       that Predicate. This means we do not need to duplicate the
4556 //       loop - we have a smaller graph and save compile time. Should
4557 //       the conditions ever fail, then we deopt / trap at the Predicate
4558 //       and recompile without that Predicate. At that point we will
4559 //       multiversion the loop, so that we can still have speculative
4560 //       runtime checks.
4561 //
4562 // We perform the multiversioning when the loop is still in its single
4563 // iteration form, even before we insert pre and post loops. This makes
4564 // the cloning much simpler. However, this means that both the fast
4565 // and the slow loop have to be optimized independently (adding pre
4566 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4567 // we may end up not needing any speculative assumptions in the fast_loop
4568 // and then rejecting the slow_loop by constant folding the multiversion_if.
4569 //
4570 // Therefore, we "delay" the optimization of the slow_loop until we add
4571 // at least one speculative assumption for the fast_loop. If we never
4572 // add such a speculative runtime check, the OpaqueMultiversioningNode
4573 // of the multiversion_if constant folds to true after loop opts, and the
4574 // multiversion_if folds away the "delayed" slow_loop. If we add any
4575 // speculative assumption, then we notify the OpaqueMultiversioningNode
4576 // with "notify_slow_loop_that_it_can_resume_optimizations".
4577 //
4578 // Note: new runtime checks can be added to the multiversion_if with
4579 //       PhaseIdealLoop::create_new_if_for_multiversion
4580 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4581   CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4582   LoopNode* outer_loop = cl->skip_strip_mined();
4583   Node* entry = outer_loop->in(LoopNode::EntryControl);
4584 
4585   // Check we have multiversioning enabled, and are not already multiversioned.
4586   if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4587 
4588   // Check that we do not have a parse-predicate where we can add the runtime checks
4589   // during auto-vectorization.
4590   const Predicates predicates(entry);
4591   const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4592   if (predicate_block->has_parse_predicate()) { return; }
4593 
4594   // Check node budget.
4595   uint estimate = lpt->est_loop_clone_sz(2);
4596   if (!may_require_nodes(estimate)) { return; }
4597 
4598   do_multiversioning(lpt, old_new);
4599 }
4600 
4601 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4602   for (uint i = 0; i < _data_nodes.size(); i++) {
4603     clone(_data_nodes[i], new_ctrl);
4604   }
4605 }
4606 
4607 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4608 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4609   Node* clone = node->clone();
4610   _phase->igvn().register_new_node_with_optimizer(clone);
4611   _orig_to_new.put(node, clone);
4612   _phase->set_ctrl(clone, new_ctrl);
4613   if (node->is_CastII()) {
4614     clone->set_req(0, new_ctrl);
4615   }
4616 }
4617 
4618 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4619 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4620 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4621   _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4622     for (uint i = 1; i < node->req(); i++) {
4623       Node** cloned_input = _orig_to_new.get(node->in(i));
4624       if (cloned_input != nullptr) {
4625         // Input was also cloned -> rewire clone to the cloned input.
4626         _phase->igvn().replace_input_of(clone, i, *cloned_input);
4627       }
4628     }
4629   });
4630 }
4631 
4632 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4633 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4634 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4635     const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4636     Node* new_ctrl) {
4637   for (uint i = 0; i < _data_nodes.size(); i++) {
4638     Node* data_node = _data_nodes[i];
4639     if (data_node->is_Opaque1()) {
4640       transform_opaque_node(transform_strategy, data_node);
4641     } else {
4642       clone(data_node, new_ctrl);
4643     }
4644   }
4645 }
4646 
4647 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4648   Node* transformed_node;
4649   if (node->is_OpaqueLoopInit()) {
4650     transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4651   } else {
4652     assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4653     transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4654   }
4655   // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4656   _orig_to_new.put(node, transformed_node);
4657 }