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, PhaseIterGVN::NodeOrigin::Speculative);
 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, PhaseIterGVN::NodeOrigin::Speculative);
 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 = AddPNode::make_with_base(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 = AddPNode::make_with_base(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 = AddPNode::make_with_base(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 = AddPNode::make_with_base(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 #ifndef PRODUCT
1482     if (TraceLoopOpts || TraceSplitIf) {
1483       tty->print_cr("Split-If: %d %s", iff->_idx, iff->Name());
1484     }
1485 #endif
1486     do_split_if(iff);
1487     C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1488     return;
1489   }
1490 
1491   // Two identical ifs back to back can be merged
1492   if (try_merge_identical_ifs(n)) {
1493     return;
1494   }
1495 
1496   // Check for an IF ready to split; one that has its
1497   // condition codes input coming from a Phi at the block start.
1498   int n_op = n->Opcode();
1499 
1500   // Check for an IF being dominated by another IF same test
1501   if (n_op == Op_If ||
1502       n_op == Op_RangeCheck) {
1503     Node *bol = n->in(1);
1504     uint max = bol->outcnt();
1505     // Check for same test used more than once?
1506     if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1507       // Search up IDOMs to see if this IF is dominated.
1508       Node* cmp = bol->in(1);
1509       Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1510 
1511       // Now search up IDOMs till cutoff, looking for a dominating test
1512       Node *prevdom = n;
1513       Node *dom = idom(prevdom);
1514       while (dom != cutoff) {
1515         if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1516             safe_for_if_replacement(dom)) {
1517           // It's invalid to move control dependent data nodes in the inner
1518           // strip-mined loop, because:
1519           //  1) break validation of LoopNode::verify_strip_mined()
1520           //  2) move code with side-effect in strip-mined loop
1521           // Move to the exit of outer strip-mined loop in that case.
1522           Node* out_le = is_inner_of_stripmined_loop(dom);
1523           if (out_le != nullptr) {
1524             prevdom = out_le;
1525           }
1526           // Replace the dominated test with an obvious true or false.
1527           // Place it on the IGVN worklist for later cleanup.
1528           C->set_major_progress();
1529           // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1530           // to prevent an array load from floating above its range check. There are three cases:
1531           // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1532           //    all its array accesses at that point.
1533           // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1534           //    accesses would start to float, since we don't pin at that point.
1535           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1536           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1537                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1538           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1539           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1540           return;
1541         }
1542         prevdom = dom;
1543         dom = idom(prevdom);
1544       }
1545     }
1546   }
1547 
1548   try_sink_out_of_loop(n);
1549   if (C->failing()) {
1550     return;
1551   }
1552 
1553   try_move_store_after_loop(n);
1554 }
1555 
1556 // Transform:
1557 //
1558 // if (some_condition) {
1559 //   // body 1
1560 // } else {
1561 //   // body 2
1562 // }
1563 // if (some_condition) {
1564 //   // body 3
1565 // } else {
1566 //   // body 4
1567 // }
1568 //
1569 // into:
1570 //
1571 //
1572 // if (some_condition) {
1573 //   // body 1
1574 //   // body 3
1575 // } else {
1576 //   // body 2
1577 //   // body 4
1578 // }
1579 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1580   if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1581     Node *n_ctrl = n->in(0);
1582     IfNode* dom_if = idom(n_ctrl)->as_If();
1583     if (n->in(1) != dom_if->in(1)) {
1584       assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1585              (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1586               dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1587       _igvn.replace_input_of(n, 1, dom_if->in(1));
1588     }
1589     IfTrueNode* dom_proj_true = dom_if->true_proj();
1590     IfFalseNode* dom_proj_false = dom_if->false_proj();
1591 
1592     // Now split the IF
1593     RegionNode* new_false_region;
1594     RegionNode* new_true_region;
1595 #ifndef PRODUCT
1596     if (TraceLoopOpts || TraceSplitIf) {
1597       tty->print_cr("Split-If Merging Identical Ifs: Dom-If: %d %s, If: %d %s", dom_if->_idx, dom_if->Name(), n->_idx, n->Name());
1598     }
1599 #endif
1600     do_split_if(n, &new_false_region, &new_true_region);
1601     assert(new_false_region->req() == new_true_region->req(), "");
1602 #ifdef ASSERT
1603     for (uint i = 1; i < new_false_region->req(); ++i) {
1604       assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1605       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");
1606     }
1607 #endif
1608     assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1609 
1610     // We now have:
1611     // if (some_condition) {
1612     //   // body 1
1613     //   if (some_condition) {
1614     //     body3: // new_true_region
1615     //     // body3
1616     //   } else {
1617     //     goto body4;
1618     //   }
1619     // } else {
1620     //   // body 2
1621     //  if (some_condition) {
1622     //     goto body3;
1623     //   } else {
1624     //     body4:   // new_false_region
1625     //     // body4;
1626     //   }
1627     // }
1628     //
1629 
1630     // clone pinned nodes thru the resulting regions
1631     push_pinned_nodes_thru_region(dom_if, new_true_region);
1632     push_pinned_nodes_thru_region(dom_if, new_false_region);
1633 
1634     // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1635     // 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
1636     // unrelated control dependency.
1637     for (uint i = 1; i < new_false_region->req(); i++) {
1638       if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1639         dominated_by(dom_proj_true, new_false_region->in(i)->in(0)->as_If());
1640       } else {
1641         assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1642         dominated_by(dom_proj_false, new_false_region->in(i)->in(0)->as_If());
1643       }
1644     }
1645     return true;
1646   }
1647   return false;
1648 }
1649 
1650 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1651   for (DUIterator i = region->outs(); region->has_out(i); i++) {
1652     Node* u = region->out(i);
1653     if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test()) {
1654       continue;
1655     }
1656     assert(u->in(0) == region, "not a control dependent node?");
1657     uint j = 1;
1658     for (; j < u->req(); ++j) {
1659       Node* in = u->in(j);
1660       if (!is_dominator(ctrl_or_self(in), dom_if)) {
1661         break;
1662       }
1663     }
1664     if (j == u->req()) {
1665       Node *phi = PhiNode::make_blank(region, u);
1666       for (uint k = 1; k < region->req(); ++k) {
1667         Node* clone = u->clone();
1668         clone->set_req(0, region->in(k));
1669         register_new_node(clone, region->in(k));
1670         phi->init_req(k, clone);
1671       }
1672       register_new_node(phi, region);
1673       _igvn.replace_node(u, phi);
1674       --i;
1675     }
1676   }
1677 }
1678 
1679 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1680   if (!dom->is_CountedLoopEnd()) {
1681     return true;
1682   }
1683   CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1684   CountedLoopNode* cl = le->loopnode();
1685   if (cl == nullptr) {
1686     return true;
1687   }
1688   if (!cl->is_main_loop()) {
1689     return true;
1690   }
1691   if (cl->is_canonical_loop_entry() == nullptr) {
1692     return true;
1693   }
1694   // Further unrolling is possible so loop exit condition might change
1695   return false;
1696 }
1697 
1698 // See if a shared loop-varying computation has no loop-varying uses.
1699 // Happens if something is only used for JVM state in uncommon trap exits,
1700 // like various versions of induction variable+offset.  Clone the
1701 // computation per usage to allow it to sink out of the loop.
1702 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1703   bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1704                             n->in(1)->bottom_type()->isa_rawptr() &&
1705                             !n->bottom_type()->isa_rawptr();
1706 
1707   if (has_ctrl(n) &&
1708       !n->is_Phi() &&
1709       !n->is_Bool() &&
1710       !n->is_Proj() &&
1711       !n->is_MergeMem() &&
1712       !n->is_CMove() &&
1713       !n->is_OpaqueConstantBool() &&
1714       !n->is_OpaqueInitializedAssertionPredicate() &&
1715       !n->is_OpaqueTemplateAssertionPredicate() &&
1716       !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1717       n->Opcode() != Op_CreateEx &&
1718       (KillPathsReachableByDeadTypeNode || !n->is_Type())
1719       ) {
1720     Node *n_ctrl = get_ctrl(n);
1721     IdealLoopTree *n_loop = get_loop(n_ctrl);
1722 
1723     if (n->in(0) != nullptr) {
1724       IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1725       if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1726         // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1727         // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1728         // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1729         Node* maybe_pinned_n = n;
1730         Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1731         if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) {
1732           if (n->depends_only_on_test()) {
1733             // If this node depends_only_on_test, it will be rewired to a control input that is not
1734             // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1735             // rewired to a dominating test equivalent to the new control.
1736             Node* pinned_clone = n->pin_node_under_control();
1737             if (pinned_clone != nullptr) {
1738               register_new_node(pinned_clone, n_ctrl);
1739               maybe_pinned_n = pinned_clone;
1740               _igvn.replace_node(n, pinned_clone);
1741             }
1742           }
1743           _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1744         }
1745       }
1746     }
1747     if (n_loop != _ltree_root && n->outcnt() > 1) {
1748       // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1749       // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1750       Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1751       if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1752           ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1753         if (n->is_Store() || n->is_LoadStore()) {
1754             assert(false, "no node with a side effect");
1755             C->record_failure("no node with a side effect");
1756             return;
1757         }
1758         Node* outer_loop_clone = nullptr;
1759         for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1760           Node* u = n->last_out(j); // Clone private computation per use
1761           _igvn.rehash_node_delayed(u);
1762           Node* x = nullptr;
1763           if (n->in(0) != nullptr && n->depends_only_on_test()) {
1764             // If this node depends_only_on_test, it will be rewired to a control input that is not
1765             // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1766             // rewired to a dominating test equivalent to the new control.
1767             x = n->pin_node_under_control();
1768           }
1769           if (x == nullptr) {
1770             x = n->clone();
1771           }
1772           Node* x_ctrl = nullptr;
1773           if (u->is_Phi()) {
1774             // Replace all uses of normal nodes.  Replace Phi uses
1775             // individually, so the separate Nodes can sink down
1776             // different paths.
1777             uint k = 1;
1778             while (u->in(k) != n) k++;
1779             u->set_req(k, x);
1780             // x goes next to Phi input path
1781             x_ctrl = u->in(0)->in(k);
1782             // Find control for 'x' next to use but not inside inner loops.
1783             x_ctrl = place_outside_loop(x_ctrl, n_loop);
1784             --j;
1785           } else {              // Normal use
1786             if (has_ctrl(u)) {
1787               x_ctrl = get_ctrl(u);
1788             } else {
1789               x_ctrl = u->in(0);
1790             }
1791             // Find control for 'x' next to use but not inside inner loops.
1792             x_ctrl = place_outside_loop(x_ctrl, n_loop);
1793             // Replace all uses
1794             if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1795               // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1796               // anymore now that we're going to pin n as well
1797               _igvn.replace_node(u, x);
1798               --j;
1799             } else {
1800               int nb = u->replace_edge(n, x, &_igvn);
1801               j -= nb;
1802             }
1803           }
1804 
1805           if (n->is_Load()) {
1806             // For loads, add a control edge to a CFG node outside of the loop
1807             // to force them to not combine and return back inside the loop
1808             // during GVN optimization (4641526).
1809             assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1810 
1811             IdealLoopTree* x_loop = get_loop(x_ctrl);
1812             Node* x_head = x_loop->_head;
1813             if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1814               // Do not add duplicate LoadNodes to the outer strip mined loop
1815               if (outer_loop_clone != nullptr) {
1816                 _igvn.replace_node(x, outer_loop_clone);
1817                 continue;
1818               }
1819               outer_loop_clone = x;
1820             }
1821             x->set_req(0, x_ctrl);
1822           } else if (n->in(0) != nullptr){
1823             x->set_req(0, x_ctrl);
1824           }
1825           assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1826           assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1827           register_new_node(x, x_ctrl);
1828 
1829           // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1830           // All AddP nodes must keep the same base after sinking so:
1831           // 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,
1832           // their bases remain the same.
1833           // (see 2- below)
1834           assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1835                  x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1836                  !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1837           if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1838               !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1839             assert(!x->is_Load(), "load should be pinned");
1840             // Use a cast node to pin clone out of loop
1841             Node* cast = nullptr;
1842             for (uint k = 0; k < x->req(); k++) {
1843               Node* in = x->in(k);
1844               if (in != nullptr && ctrl_is_member(n_loop, in)) {
1845                 const Type* in_t = _igvn.type(in);
1846                 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1847                                                               ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr);
1848               }
1849               if (cast != nullptr) {
1850                 Node* prev = _igvn.hash_find_insert(cast);
1851                 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1852                   cast->destruct(&_igvn);
1853                   cast = prev;
1854                 } else {
1855                   register_new_node(cast, x_ctrl);
1856                 }
1857                 x->replace_edge(in, cast);
1858                 // Chain of AddP nodes:
1859                 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1860                 if (x->is_AddP() && k == AddPNode::Base) {
1861                   update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1862                 }
1863                 break;
1864               }
1865             }
1866             assert(cast != nullptr, "must have added a cast to pin the node");
1867           }
1868         }
1869         _igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
1870       }
1871       _dom_lca_tags_round = 0;
1872     }
1873   }
1874 }
1875 
1876 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1877   ResourceMark rm;
1878   Node_List wq;
1879   wq.push(x);
1880   while (wq.size() != 0) {
1881     Node* n = wq.pop();
1882     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1883       Node* u = n->fast_out(i);
1884       if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1885         _igvn.replace_input_of(u, AddPNode::Base, new_base);
1886         wq.push(u);
1887       }
1888     }
1889   }
1890 }
1891 
1892 // Compute the early control of a node by following its inputs until we reach
1893 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1894 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1895   Node* early_ctrl = nullptr;
1896   ResourceMark rm;
1897   Unique_Node_List wq;
1898   wq.push(n);
1899   for (uint i = 0; i < wq.size(); i++) {
1900     Node* m = wq.at(i);
1901     Node* c = nullptr;
1902     if (m->is_CFG()) {
1903       c = m;
1904     } else if (m->pinned()) {
1905       c = m->in(0);
1906     } else {
1907       for (uint j = 0; j < m->req(); j++) {
1908         Node* in = m->in(j);
1909         if (in != nullptr) {
1910           wq.push(in);
1911         }
1912       }
1913     }
1914     if (c != nullptr) {
1915       assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1916       if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1917         early_ctrl = c;
1918       }
1919     }
1920   }
1921   assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1922   return early_ctrl;
1923 }
1924 
1925 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1926   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1927     Node* u = n->fast_out(i);
1928     if (u->is_Opaque1()) {
1929       return false;  // Found loop limit, bugfix for 4677003
1930     }
1931     if (u->is_Phi()) {
1932       for (uint j = 1; j < u->req(); ++j) {
1933         if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1934           return false;
1935         }
1936       }
1937     } else {
1938       Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1939       if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1940         return false;
1941       }
1942     }
1943   }
1944   return true;
1945 }
1946 
1947 // 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
1948 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1949 // 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
1950 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code
1951 // is generated for the main loop.
1952 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) {
1953   if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1954     CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1955     if (is_dominator(pre_loop->loopexit(), ctrl)) {
1956       return true;
1957     }
1958   }
1959   return false;
1960 }
1961 
1962 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1963   if (n->is_Load()) {
1964     // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure each call to
1965     // get_late_ctrl_with_anti_dep() uses its own tag
1966     _dom_lca_tags_round++;
1967     assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1968 
1969     ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1970   }
1971   IdealLoopTree *u_loop = get_loop(ctrl);
1972   if (u_loop == n_loop) {
1973     return false; // Found loop-varying use
1974   }
1975   if (n_loop->is_member(u_loop)) {
1976     return false; // Found use in inner loop
1977   }
1978   if (would_sink_below_pre_loop_exit(n_loop, ctrl)) {
1979     return false;
1980   }
1981   return true;
1982 }
1983 
1984 //------------------------------split_if_with_blocks---------------------------
1985 // Check for aggressive application of 'split-if' optimization,
1986 // using basic block level info.
1987 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1988   Node* root = C->root();
1989   visited.set(root->_idx); // first, mark root as visited
1990   // Do pre-visit work for root
1991   Node* n   = split_if_with_blocks_pre(root);
1992   uint  cnt = n->outcnt();
1993   uint  i   = 0;
1994 
1995   while (true) {
1996     // Visit all children
1997     if (i < cnt) {
1998       Node* use = n->raw_out(i);
1999       ++i;
2000       if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
2001         // Now do pre-visit work for this use
2002         use = split_if_with_blocks_pre(use);
2003         nstack.push(n, i); // Save parent and next use's index.
2004         n   = use;         // Process all children of current use.
2005         cnt = use->outcnt();
2006         i   = 0;
2007       }
2008     }
2009     else {
2010       // All of n's children have been processed, complete post-processing.
2011       if (cnt != 0 && !n->is_Con()) {
2012         assert(has_node(n), "no dead nodes");
2013         split_if_with_blocks_post(n);
2014         if (C->failing()) {
2015           return;
2016         }
2017       }
2018       if (must_throttle_split_if()) {
2019         nstack.clear();
2020       }
2021       if (nstack.is_empty()) {
2022         // Finished all nodes on stack.
2023         break;
2024       }
2025       // Get saved parent node and next use's index. Visit the rest of uses.
2026       n   = nstack.node();
2027       cnt = n->outcnt();
2028       i   = nstack.index();
2029       nstack.pop();
2030     }
2031   }
2032 }
2033 
2034 
2035 //=============================================================================
2036 //
2037 //                   C L O N E   A   L O O P   B O D Y
2038 //
2039 
2040 //------------------------------clone_iff--------------------------------------
2041 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2042 // "Nearly" because all Nodes have been cloned from the original in the loop,
2043 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
2044 // through the Phi recursively, and return a Bool.
2045 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2046 
2047   // Convert this Phi into a Phi merging Bools
2048   uint i;
2049   for (i = 1; i < phi->req(); i++) {
2050     Node* b = phi->in(i);
2051     if (b->is_Phi()) {
2052       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2053     } else {
2054       assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(),
2055              "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node");
2056     }
2057   }
2058   Node* n = phi->in(1);
2059   Node* sample_opaque = nullptr;
2060   Node *sample_bool = nullptr;
2061   if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) {
2062     sample_opaque = n;
2063     sample_bool = n->in(1);
2064     assert(sample_bool->is_Bool(), "wrong type");
2065   } else {
2066     sample_bool = n;
2067   }
2068   Node *sample_cmp = sample_bool->in(1);
2069 
2070   // Make Phis to merge the Cmp's inputs.
2071   PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2072   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2073   for (i = 1; i < phi->req(); i++) {
2074     Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2075     Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2076     phi1->set_req(i, n1);
2077     phi2->set_req(i, n2);
2078     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2079     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2080   }
2081   // See if these Phis have been made before.
2082   // Register with optimizer
2083   Node *hit1 = _igvn.hash_find_insert(phi1);
2084   if (hit1) {                   // Hit, toss just made Phi
2085     _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2086     assert(hit1->is_Phi(), "" );
2087     phi1 = (PhiNode*)hit1;      // Use existing phi
2088   } else {                      // Miss
2089     _igvn.register_new_node_with_optimizer(phi1);
2090   }
2091   Node *hit2 = _igvn.hash_find_insert(phi2);
2092   if (hit2) {                   // Hit, toss just made Phi
2093     _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2094     assert(hit2->is_Phi(), "" );
2095     phi2 = (PhiNode*)hit2;      // Use existing phi
2096   } else {                      // Miss
2097     _igvn.register_new_node_with_optimizer(phi2);
2098   }
2099   // Register Phis with loop/block info
2100   set_ctrl(phi1, phi->in(0));
2101   set_ctrl(phi2, phi->in(0));
2102   // Make a new Cmp
2103   Node *cmp = sample_cmp->clone();
2104   cmp->set_req(1, phi1);
2105   cmp->set_req(2, phi2);
2106   _igvn.register_new_node_with_optimizer(cmp);
2107   set_ctrl(cmp, phi->in(0));
2108 
2109   // Make a new Bool
2110   Node *b = sample_bool->clone();
2111   b->set_req(1,cmp);
2112   _igvn.register_new_node_with_optimizer(b);
2113   set_ctrl(b, phi->in(0));
2114 
2115   if (sample_opaque != nullptr) {
2116     Node* opaque = sample_opaque->clone();
2117     opaque->set_req(1, b);
2118     _igvn.register_new_node_with_optimizer(opaque);
2119     set_ctrl(opaque, phi->in(0));
2120     return opaque;
2121   }
2122 
2123   assert(b->is_Bool(), "");
2124   return b;
2125 }
2126 
2127 //------------------------------clone_bool-------------------------------------
2128 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2129 // "Nearly" because all Nodes have been cloned from the original in the loop,
2130 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
2131 // through the Phi recursively, and return a Bool.
2132 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2133   uint i;
2134   // Convert this Phi into a Phi merging Bools
2135   for( i = 1; i < phi->req(); i++ ) {
2136     Node *b = phi->in(i);
2137     if( b->is_Phi() ) {
2138       _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2139     } else {
2140       assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2141     }
2142   }
2143 
2144   Node *sample_cmp = phi->in(1);
2145 
2146   // Make Phis to merge the Cmp's inputs.
2147   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2148   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2149   for( uint j = 1; j < phi->req(); j++ ) {
2150     Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2151     Node *n1, *n2;
2152     if( cmp_top->is_Cmp() ) {
2153       n1 = cmp_top->in(1);
2154       n2 = cmp_top->in(2);
2155     } else {
2156       n1 = n2 = cmp_top;
2157     }
2158     phi1->set_req( j, n1 );
2159     phi2->set_req( j, n2 );
2160     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2161     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2162   }
2163 
2164   // See if these Phis have been made before.
2165   // Register with optimizer
2166   Node *hit1 = _igvn.hash_find_insert(phi1);
2167   if( hit1 ) {                  // Hit, toss just made Phi
2168     _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2169     assert( hit1->is_Phi(), "" );
2170     phi1 = (PhiNode*)hit1;      // Use existing phi
2171   } else {                      // Miss
2172     _igvn.register_new_node_with_optimizer(phi1);
2173   }
2174   Node *hit2 = _igvn.hash_find_insert(phi2);
2175   if( hit2 ) {                  // Hit, toss just made Phi
2176     _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2177     assert( hit2->is_Phi(), "" );
2178     phi2 = (PhiNode*)hit2;      // Use existing phi
2179   } else {                      // Miss
2180     _igvn.register_new_node_with_optimizer(phi2);
2181   }
2182   // Register Phis with loop/block info
2183   set_ctrl(phi1, phi->in(0));
2184   set_ctrl(phi2, phi->in(0));
2185   // Make a new Cmp
2186   Node *cmp = sample_cmp->clone();
2187   cmp->set_req( 1, phi1 );
2188   cmp->set_req( 2, phi2 );
2189   _igvn.register_new_node_with_optimizer(cmp);
2190   set_ctrl(cmp, phi->in(0));
2191 
2192   assert( cmp->is_Cmp(), "" );
2193   return (CmpNode*)cmp;
2194 }
2195 
2196 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2197                                                  IdealLoopTree* loop, IdealLoopTree* outer_loop,
2198                                                  Node_List*& split_if_set, Node_List*& split_bool_set,
2199                                                  Node_List*& split_cex_set, Node_List& worklist,
2200                                                  uint new_counter, CloneLoopMode mode) {
2201   Node* nnn = old_new[old->_idx];
2202   // Copy uses to a worklist, so I can munge the def-use info
2203   // with impunity.
2204   for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2205     worklist.push(old->fast_out(j));
2206 
2207   while( worklist.size() ) {
2208     Node *use = worklist.pop();
2209     if (!has_node(use))  continue; // Ignore dead nodes
2210     if (use->in(0) == C->top())  continue;
2211     IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2212     // Check for data-use outside of loop - at least one of OLD or USE
2213     // must not be a CFG node.
2214 #ifdef ASSERT
2215     if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2216       Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2217       assert(mode != IgnoreStripMined, "incorrect cloning mode");
2218       assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2219     }
2220 #endif
2221     if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2222 
2223       // If the Data use is an IF, that means we have an IF outside the
2224       // loop that is switching on a condition that is set inside the
2225       // loop.  Happens if people set a loop-exit flag; then test the flag
2226       // in the loop to break the loop, then test is again outside the
2227       // loop to determine which way the loop exited.
2228       //
2229       // For several uses we need to make sure that there is no phi between,
2230       // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2231       // to avoid such a phi in between.
2232       // For example, it is unexpected that there is a Phi between an
2233       // AllocateArray node and its ValidLengthTest input that could cause
2234       // split if to break.
2235       assert(!use->is_OpaqueTemplateAssertionPredicate(),
2236              "should not clone a Template Assertion Predicate which should be removed once it's useless");
2237       if (use->is_If() || use->is_CMove() || use->is_OpaqueConstantBool() || use->is_OpaqueInitializedAssertionPredicate() ||
2238           (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2239         // Since this code is highly unlikely, we lazily build the worklist
2240         // of such Nodes to go split.
2241         if (!split_if_set) {
2242           split_if_set = new Node_List();
2243         }
2244         split_if_set->push(use);
2245       }
2246       if (use->is_Bool()) {
2247         if (!split_bool_set) {
2248           split_bool_set = new Node_List();
2249         }
2250         split_bool_set->push(use);
2251       }
2252       if (use->Opcode() == Op_CreateEx) {
2253         if (!split_cex_set) {
2254           split_cex_set = new Node_List();
2255         }
2256         split_cex_set->push(use);
2257       }
2258 
2259 
2260       // Get "block" use is in
2261       uint idx = 0;
2262       while( use->in(idx) != old ) idx++;
2263       Node *prev = use->is_CFG() ? use : get_ctrl(use);
2264       assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2265       Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2266         ? prev->in(2)
2267         : idom(prev);
2268       if( use->is_Phi() )     // Phi use is in prior block
2269         cfg = prev->in(idx);  // NOT in block of Phi itself
2270       if (cfg->is_top()) {    // Use is dead?
2271         _igvn.replace_input_of(use, idx, C->top());
2272         continue;
2273       }
2274 
2275       // If use is referenced through control edge... (idx == 0)
2276       if (mode == IgnoreStripMined && idx == 0) {
2277         LoopNode *head = loop->_head->as_Loop();
2278         if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2279           // That node is outside the inner loop, leave it outside the
2280           // outer loop as well to not confuse verification code.
2281           assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2282           _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2283           continue;
2284         }
2285       }
2286 
2287       while(!outer_loop->is_member(get_loop(cfg))) {
2288         prev = cfg;
2289         cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2290       }
2291       // If the use occurs after merging several exits from the loop, then
2292       // old value must have dominated all those exits.  Since the same old
2293       // value was used on all those exits we did not need a Phi at this
2294       // merge point.  NOW we do need a Phi here.  Each loop exit value
2295       // is now merged with the peeled body exit; each exit gets its own
2296       // private Phi and those Phis need to be merged here.
2297       Node *phi;
2298       if( prev->is_Region() ) {
2299         if( idx == 0 ) {      // Updating control edge?
2300           phi = prev;         // Just use existing control
2301         } else {              // Else need a new Phi
2302           phi = PhiNode::make( prev, old );
2303           // Now recursively fix up the new uses of old!
2304           for( uint i = 1; i < prev->req(); i++ ) {
2305             worklist.push(phi); // Onto worklist once for each 'old' input
2306           }
2307         }
2308       } else {
2309         // Get new RegionNode merging old and new loop exits
2310         prev = old_new[prev->_idx];
2311         assert( prev, "just made this in step 7" );
2312         if( idx == 0) {      // Updating control edge?
2313           phi = prev;         // Just use existing control
2314         } else {              // Else need a new Phi
2315           // Make a new Phi merging data values properly
2316           phi = PhiNode::make( prev, old );
2317           phi->set_req( 1, nnn );
2318         }
2319       }
2320       // If inserting a new Phi, check for prior hits
2321       if( idx != 0 ) {
2322         Node *hit = _igvn.hash_find_insert(phi);
2323         if( hit == nullptr ) {
2324           _igvn.register_new_node_with_optimizer(phi); // Register new phi
2325         } else {                                      // or
2326           // Remove the new phi from the graph and use the hit
2327           _igvn.remove_dead_node(phi, phi == prev ? PhaseIterGVN::NodeOrigin::Graph : PhaseIterGVN::NodeOrigin::Speculative);
2328           phi = hit;                                  // Use existing phi
2329         }
2330         set_ctrl(phi, prev);
2331       }
2332       // Make 'use' use the Phi instead of the old loop body exit value
2333       assert(use->in(idx) == old, "old is still input of use");
2334       // We notify all uses of old, including use, and the indirect uses,
2335       // that may now be optimized because we have replaced old with phi.
2336       _igvn.add_users_to_worklist(old);
2337       if (idx == 0 && use->depends_only_on_test()) {
2338         // If this node depends_only_on_test, it will be rewired to a control input that is not the
2339         // correct test. As a result, it must be pinned otherwise it can be incorrectly rewired to
2340         // a dominating test equivalent to the new control.
2341         Node* pinned_clone = use->pin_node_under_control();
2342         if (pinned_clone != nullptr) {
2343           pinned_clone->set_req(0, phi);
2344           register_new_node_with_ctrl_of(pinned_clone, use);
2345           _igvn.replace_node(use, pinned_clone);
2346           continue;
2347         }
2348       }
2349       _igvn.replace_input_of(use, idx, phi);
2350       if( use->_idx >= new_counter ) { // If updating new phis
2351         // Not needed for correctness, but prevents a weak assert
2352         // in AddPNode from tripping (when we end up with different
2353         // base & derived Phis that will become the same after
2354         // IGVN does CSE).
2355         Node *hit = _igvn.hash_find_insert(use);
2356         if( hit )             // Go ahead and re-hash for hits.
2357           _igvn.replace_node( use, hit );
2358       }
2359     }
2360   }
2361 }
2362 
2363 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2364                                                                 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2365                                                                 bool check_old_new) {
2366   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2367     Node* u = n->fast_out(j);
2368     assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2369     if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2370       assert(!phase->ctrl_is_member(loop, u) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2371       if (!phase->ctrl_is_member(loop, u)) {
2372         if (phase->ctrl_is_member(outer_loop, u)) {
2373           wq.push(u);
2374         } else {
2375           // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2376           // the outer loop too
2377           Node* u_c = u->in(0);
2378           if (u_c != nullptr) {
2379             IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2380             if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2381               wq.push(u);
2382             }
2383           }
2384         }
2385       }
2386     }
2387   }
2388 }
2389 
2390 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2391                                       IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2392                                       Node_List& extra_data_nodes) {
2393   if (head->is_strip_mined() && mode != IgnoreStripMined) {
2394     CountedLoopNode* cl = head->as_CountedLoop();
2395     Node* l = cl->outer_loop();
2396     Node* tail = cl->outer_loop_tail();
2397     IfNode* le = cl->outer_loop_end();
2398     Node* sfpt = cl->outer_safepoint();
2399     CountedLoopEndNode* cle = cl->loopexit();
2400     CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2401     CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2402     IfFalseNode* cle_out = cle->false_proj();
2403 
2404     Node* new_sfpt = nullptr;
2405     Node* new_cle_out = cle_out->clone();
2406     old_new.map(cle_out->_idx, new_cle_out);
2407     if (mode == CloneIncludesStripMined) {
2408       // clone outer loop body
2409       Node* new_l = l->clone();
2410       Node* new_tail = tail->clone();
2411       IfNode* new_le = le->clone()->as_If();
2412       new_sfpt = sfpt->clone();
2413 
2414       set_loop(new_l, outer_loop->_parent);
2415       set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2416       set_loop(new_cle_out, outer_loop->_parent);
2417       set_idom(new_cle_out, new_cle, dd);
2418       set_loop(new_sfpt, outer_loop->_parent);
2419       set_idom(new_sfpt, new_cle_out, dd);
2420       set_loop(new_le, outer_loop->_parent);
2421       set_idom(new_le, new_sfpt, dd);
2422       set_loop(new_tail, outer_loop->_parent);
2423       set_idom(new_tail, new_le, dd);
2424       set_idom(new_cl, new_l, dd);
2425 
2426       old_new.map(l->_idx, new_l);
2427       old_new.map(tail->_idx, new_tail);
2428       old_new.map(le->_idx, new_le);
2429       old_new.map(sfpt->_idx, new_sfpt);
2430 
2431       new_l->set_req(LoopNode::LoopBackControl, new_tail);
2432       new_l->set_req(0, new_l);
2433       new_tail->set_req(0, new_le);
2434       new_le->set_req(0, new_sfpt);
2435       new_sfpt->set_req(0, new_cle_out);
2436       new_cle_out->set_req(0, new_cle);
2437       new_cl->set_req(LoopNode::EntryControl, new_l);
2438 
2439       _igvn.register_new_node_with_optimizer(new_l);
2440       _igvn.register_new_node_with_optimizer(new_tail);
2441       _igvn.register_new_node_with_optimizer(new_le);
2442     } else {
2443       Node *newhead = old_new[loop->_head->_idx];
2444       newhead->as_Loop()->clear_strip_mined();
2445       _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2446       set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2447     }
2448     // Look at data node that were assigned a control in the outer
2449     // loop: they are kept in the outer loop by the safepoint so start
2450     // from the safepoint node's inputs.
2451     IdealLoopTree* outer_loop = get_loop(l);
2452     Node_Stack stack(2);
2453     stack.push(sfpt, 1);
2454     uint new_counter = C->unique();
2455     while (stack.size() > 0) {
2456       Node* n = stack.node();
2457       uint i = stack.index();
2458       while (i < n->req() &&
2459              (n->in(i) == nullptr ||
2460               !has_ctrl(n->in(i)) ||
2461               get_loop(get_ctrl(n->in(i))) != outer_loop ||
2462               (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2463         i++;
2464       }
2465       if (i < n->req()) {
2466         stack.set_index(i+1);
2467         stack.push(n->in(i), 0);
2468       } else {
2469         assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2470         Node* m = n == sfpt ? new_sfpt : n->clone();
2471         if (m != nullptr) {
2472           for (uint i = 0; i < n->req(); i++) {
2473             if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2474               m->set_req(i, old_new[m->in(i)->_idx]);
2475             }
2476           }
2477         } else {
2478           assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2479         }
2480         if (n != sfpt) {
2481           extra_data_nodes.push(n);
2482           _igvn.register_new_node_with_optimizer(m);
2483           assert(get_ctrl(n) == cle_out, "what other control?");
2484           set_ctrl(m, new_cle_out);
2485           old_new.map(n->_idx, m);
2486         }
2487         stack.pop();
2488       }
2489     }
2490     if (mode == CloneIncludesStripMined) {
2491       _igvn.register_new_node_with_optimizer(new_sfpt);
2492       _igvn.register_new_node_with_optimizer(new_cle_out);
2493     }
2494     // Some other transformation may have pessimistically assigned some
2495     // data nodes to the outer loop. Set their control so they are out
2496     // of the outer loop.
2497     ResourceMark rm;
2498     Unique_Node_List wq;
2499     for (uint i = 0; i < extra_data_nodes.size(); i++) {
2500       Node* old = extra_data_nodes.at(i);
2501       collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2502     }
2503 
2504     for (uint i = 0; i < loop->_body.size(); i++) {
2505       Node* old = loop->_body.at(i);
2506       collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2507     }
2508 
2509     Node* inner_out = sfpt->in(0);
2510     if (inner_out->outcnt() > 1) {
2511       collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2512     }
2513 
2514     Node* new_ctrl = cl->outer_loop_exit();
2515     assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2516     for (uint i = 0; i < wq.size(); i++) {
2517       Node* n = wq.at(i);
2518       set_ctrl(n, new_ctrl);
2519       if (n->in(0) != nullptr) {
2520         _igvn.replace_input_of(n, 0, new_ctrl);
2521       }
2522       collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2523     }
2524   } else {
2525     Node *newhead = old_new[loop->_head->_idx];
2526     set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2527   }
2528 }
2529 
2530 //------------------------------clone_loop-------------------------------------
2531 //
2532 //                   C L O N E   A   L O O P   B O D Y
2533 //
2534 // This is the basic building block of the loop optimizations.  It clones an
2535 // entire loop body.  It makes an old_new loop body mapping; with this mapping
2536 // you can find the new-loop equivalent to an old-loop node.  All new-loop
2537 // nodes are exactly equal to their old-loop counterparts, all edges are the
2538 // same.  All exits from the old-loop now have a RegionNode that merges the
2539 // equivalent new-loop path.  This is true even for the normal "loop-exit"
2540 // condition.  All uses of loop-invariant old-loop values now come from (one
2541 // or more) Phis that merge their new-loop equivalents.
2542 //
2543 // This operation leaves the graph in an illegal state: there are two valid
2544 // control edges coming from the loop pre-header to both loop bodies.  I'll
2545 // definitely have to hack the graph after running this transform.
2546 //
2547 // From this building block I will further edit edges to perform loop peeling
2548 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2549 //
2550 // Parameter side_by_size_idom:
2551 //   When side_by_size_idom is null, the dominator tree is constructed for
2552 //      the clone loop to dominate the original.  Used in construction of
2553 //      pre-main-post loop sequence.
2554 //   When nonnull, the clone and original are side-by-side, both are
2555 //      dominated by the side_by_side_idom node.  Used in construction of
2556 //      unswitched loops.
2557 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2558                                 CloneLoopMode mode, Node* side_by_side_idom) {
2559 
2560   LoopNode* head = loop->_head->as_Loop();
2561   head->verify_strip_mined(1);
2562 
2563   if (C->do_vector_loop() && PrintOpto) {
2564     const char* mname = C->method()->name()->as_quoted_ascii();
2565     if (mname != nullptr) {
2566       tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2567     }
2568   }
2569 
2570   CloneMap& cm = C->clone_map();
2571   if (C->do_vector_loop()) {
2572     cm.set_clone_idx(cm.max_gen()+1);
2573 #ifndef PRODUCT
2574     if (PrintOpto) {
2575       tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2576       loop->dump_head();
2577     }
2578 #endif
2579   }
2580 
2581   // Step 1: Clone the loop body.  Make the old->new mapping.
2582   clone_loop_body(loop->_body, old_new, &cm);
2583 
2584   IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2585 
2586   // Step 2: Fix the edges in the new body.  If the old input is outside the
2587   // loop use it.  If the old input is INside the loop, use the corresponding
2588   // new node instead.
2589   fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2590 
2591   Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2592   clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2593 
2594   // Step 3: Now fix control uses.  Loop varying control uses have already
2595   // been fixed up (as part of all input edges in Step 2).  Loop invariant
2596   // control uses must be either an IfFalse or an IfTrue.  Make a merge
2597   // point to merge the old and new IfFalse/IfTrue nodes; make the use
2598   // refer to this.
2599   Node_List worklist;
2600   uint new_counter = C->unique();
2601   fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2602 
2603   // Step 4: If loop-invariant use is not control, it must be dominated by a
2604   // loop exit IfFalse/IfTrue.  Find "proper" loop exit.  Make a Region
2605   // there if needed.  Make a Phi there merging old and new used values.
2606   Node_List *split_if_set = nullptr;
2607   Node_List *split_bool_set = nullptr;
2608   Node_List *split_cex_set = nullptr;
2609   fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2610 
2611   for (uint i = 0; i < extra_data_nodes.size(); i++) {
2612     Node* old = extra_data_nodes.at(i);
2613     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2614                                 split_bool_set, split_cex_set, worklist, new_counter,
2615                                 mode);
2616   }
2617 
2618   // Check for IFs that need splitting/cloning.  Happens if an IF outside of
2619   // the loop uses a condition set in the loop.  The original IF probably
2620   // takes control from one or more OLD Regions (which in turn get from NEW
2621   // Regions).  In any case, there will be a set of Phis for each merge point
2622   // from the IF up to where the original BOOL def exists the loop.
2623   finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2624 
2625 }
2626 
2627 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2628   if (split_if_set) {
2629     while (split_if_set->size()) {
2630       Node *iff = split_if_set->pop();
2631       uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2632       if (iff->in(input)->is_Phi()) {
2633         Node *b = clone_iff(iff->in(input)->as_Phi());
2634         _igvn.replace_input_of(iff, input, b);
2635       }
2636     }
2637   }
2638   if (split_bool_set) {
2639     while (split_bool_set->size()) {
2640       Node *b = split_bool_set->pop();
2641       Node *phi = b->in(1);
2642       assert(phi->is_Phi(), "");
2643       CmpNode *cmp = clone_bool((PhiNode*) phi);
2644       _igvn.replace_input_of(b, 1, cmp);
2645     }
2646   }
2647   if (split_cex_set) {
2648     while (split_cex_set->size()) {
2649       Node *b = split_cex_set->pop();
2650       assert(b->in(0)->is_Region(), "");
2651       assert(b->in(1)->is_Phi(), "");
2652       assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2653       split_up(b, b->in(0), nullptr);
2654     }
2655   }
2656 }
2657 
2658 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2659                                    uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2660                                    Node_List*& split_bool_set, Node_List*& split_cex_set) {
2661   for(uint i = 0; i < body.size(); i++ ) {
2662     Node* old = body.at(i);
2663     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2664                                 split_bool_set, split_cex_set, worklist, new_counter,
2665                                 mode);
2666   }
2667 }
2668 
2669 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2670                                    Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2671   LoopNode* head = loop->_head->as_Loop();
2672   for(uint i = 0; i < body.size(); i++ ) {
2673     Node* old = body.at(i);
2674     if( !old->is_CFG() ) continue;
2675 
2676     // Copy uses to a worklist, so I can munge the def-use info
2677     // with impunity.
2678     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2679       worklist.push(old->fast_out(j));
2680     }
2681 
2682     while (worklist.size()) {  // Visit all uses
2683       Node *use = worklist.pop();
2684       if (!has_node(use))  continue; // Ignore dead nodes
2685       IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2686       if (!loop->is_member(use_loop) && use->is_CFG()) {
2687         // Both OLD and USE are CFG nodes here.
2688         assert(use->is_Proj(), "" );
2689         Node* nnn = old_new[old->_idx];
2690 
2691         Node* newuse = nullptr;
2692         if (head->is_strip_mined() && mode != IgnoreStripMined) {
2693           CountedLoopNode* cl = head->as_CountedLoop();
2694           CountedLoopEndNode* cle = cl->loopexit();
2695           // is use the projection that exits the loop from the CountedLoopEndNode?
2696           if (use->in(0) == cle) {
2697             IfFalseNode* cle_out = use->as_IfFalse();
2698             IfNode* le = cl->outer_loop_end();
2699             use = le->false_proj();
2700             use_loop = get_loop(use);
2701             if (mode == CloneIncludesStripMined) {
2702               nnn = old_new[le->_idx];
2703             } else {
2704               newuse = old_new[cle_out->_idx];
2705             }
2706           }
2707         }
2708         if (newuse == nullptr) {
2709           newuse = use->clone();
2710         }
2711 
2712         // Clone the loop exit control projection
2713         if (C->do_vector_loop() && cm != nullptr) {
2714           cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2715         }
2716         newuse->set_req(0,nnn);
2717         _igvn.register_new_node_with_optimizer(newuse);
2718         set_loop(newuse, use_loop);
2719         set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2720 
2721         // We need a Region to merge the exit from the peeled body and the
2722         // exit from the old loop body.
2723         RegionNode *r = new RegionNode(3);
2724         uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2725         assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2726 
2727         // The original user of 'use' uses 'r' instead.
2728         for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2729           Node* useuse = use->last_out(l);
2730           _igvn.rehash_node_delayed(useuse);
2731           uint uses_found = 0;
2732           if (useuse->in(0) == use) {
2733             useuse->set_req(0, r);
2734             uses_found++;
2735             if (useuse->is_CFG()) {
2736               // This is not a dom_depth > dd_r because when new
2737               // control flow is constructed by a loop opt, a node and
2738               // its dominator can end up at the same dom_depth
2739               assert(dom_depth(useuse) >= dd_r, "");
2740               set_idom(useuse, r, dom_depth(useuse));
2741             }
2742           }
2743           for (uint k = 1; k < useuse->req(); k++) {
2744             if( useuse->in(k) == use ) {
2745               useuse->set_req(k, r);
2746               uses_found++;
2747               if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2748                 // This is not a dom_depth > dd_r because when new
2749                 // control flow is constructed by a loop opt, a node
2750                 // and its dominator can end up at the same dom_depth
2751                 assert(dom_depth(useuse) >= dd_r , "");
2752                 set_idom(useuse, r, dom_depth(useuse));
2753               }
2754             }
2755           }
2756           l -= uses_found;    // we deleted 1 or more copies of this edge
2757         }
2758 
2759         assert(use->is_Proj(), "loop exit should be projection");
2760         // replace_node_and_forward_ctrl() below moves all nodes that are:
2761         // - control dependent on the loop exit or
2762         // - have control set to the loop exit
2763         // below the post-loop merge point.
2764         // replace_node_and_forward_ctrl() takes a dead control as first input.
2765         // To make it possible to use it, the loop exit projection is cloned and becomes the
2766         // new exit projection. The initial one becomes dead and is "replaced" by the region.
2767         Node* use_clone = use->clone();
2768         register_control(use_clone, use_loop, idom(use), dom_depth(use));
2769         // Now finish up 'r'
2770         r->set_req(1, newuse);
2771         r->set_req(2, use_clone);
2772         _igvn.register_new_node_with_optimizer(r);
2773         set_loop(r, use_loop);
2774         set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2775         replace_node_and_forward_ctrl(use, r);
2776         // Map the (cloned) old use to the new merge point
2777         old_new.map(use_clone->_idx, r);
2778       } // End of if a loop-exit test
2779     }
2780   }
2781 }
2782 
2783 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2784                                     IdealLoopTree* parent, bool partial) {
2785   for(uint i = 0; i < body.size(); i++ ) {
2786     Node *old = body.at(i);
2787     Node *nnn = old_new[old->_idx];
2788     // Fix CFG/Loop controlling the new node
2789     if (has_ctrl(old)) {
2790       set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2791     } else {
2792       set_loop(nnn, parent);
2793       if (old->outcnt() > 0) {
2794         Node* dom = idom(old);
2795         if (old_new[dom->_idx] != nullptr) {
2796           dom = old_new[dom->_idx];
2797           set_idom(nnn, dom, dd );
2798         }
2799       }
2800     }
2801     // Correct edges to the new node
2802     for (uint j = 0; j < nnn->req(); j++) {
2803         Node *n = nnn->in(j);
2804         if (n != nullptr) {
2805           IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2806           if (loop->is_member(old_in_loop)) {
2807             if (old_new[n->_idx] != nullptr) {
2808               nnn->set_req(j, old_new[n->_idx]);
2809             } else {
2810               assert(!body.contains(n), "");
2811               assert(partial, "node not cloned");
2812             }
2813           }
2814         }
2815     }
2816     _igvn.hash_find_insert(nnn);
2817   }
2818 }
2819 
2820 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2821   for (uint i = 0; i < body.size(); i++) {
2822     Node* old = body.at(i);
2823     Node* nnn = old->clone();
2824     old_new.map(old->_idx, nnn);
2825     if (C->do_vector_loop() && cm != nullptr) {
2826       cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2827     }
2828     _igvn.register_new_node_with_optimizer(nnn);
2829   }
2830 }
2831 
2832 
2833 //---------------------- stride_of_possible_iv -------------------------------------
2834 // Looks for an iff/bool/comp with one operand of the compare
2835 // being a cycle involving an add and a phi,
2836 // with an optional truncation (left-shift followed by a right-shift)
2837 // of the add. Returns zero if not an iv.
2838 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2839   const TypeInteger* ttype = nullptr;
2840   if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2841     return 0;
2842   }
2843   BoolNode* bl = iff->in(1)->as_Bool();
2844   Node* cmp = bl->in(1);
2845   if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2846     return 0;
2847   }
2848   // Must have an invariant operand
2849   if (ctrl_is_member(get_loop(iff), cmp->in(2))) {
2850     return 0;
2851   }
2852   Node* add2 = nullptr;
2853   Node* cmp1 = cmp->in(1);
2854   if (cmp1->is_Phi()) {
2855     // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2856     Node* phi = cmp1;
2857     for (uint i = 1; i < phi->req(); i++) {
2858       Node* in = phi->in(i);
2859       CountedLoopConverter::TruncatedIncrement add(T_INT);
2860       add.build(in);
2861       if (add.is_valid() && add.incr()->in(1) == phi) {
2862         add2 = add.incr()->in(2);
2863         break;
2864       }
2865     }
2866   } else {
2867     // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2868     Node* addtrunc = cmp1;
2869     CountedLoopConverter::TruncatedIncrement add(T_INT);
2870     add.build(addtrunc);
2871     if (add.is_valid() && add.incr()->in(1)->is_Phi()) {
2872       Node* phi = add.incr()->in(1);
2873       for (uint i = 1; i < phi->req(); i++) {
2874         if (phi->in(i) == addtrunc) {
2875           add2 = add.incr()->in(2);
2876           break;
2877         }
2878       }
2879     }
2880   }
2881   if (add2 != nullptr) {
2882     const TypeInt* add2t = _igvn.type(add2)->is_int();
2883     if (add2t->is_con()) {
2884       return add2t->get_con();
2885     }
2886   }
2887   return 0;
2888 }
2889 
2890 
2891 //---------------------- stay_in_loop -------------------------------------
2892 // Return the (unique) control output node that's in the loop (if it exists.)
2893 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2894   Node* unique = nullptr;
2895   if (!n) return nullptr;
2896   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2897     Node* use = n->fast_out(i);
2898     if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2899       if (unique != nullptr) {
2900         return nullptr;
2901       }
2902       unique = use;
2903     }
2904   }
2905   return unique;
2906 }
2907 
2908 //------------------------------ register_node -------------------------------------
2909 // Utility to register node "n" with PhaseIdealLoop
2910 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2911   _igvn.register_new_node_with_optimizer(n);
2912   loop->_body.push(n);
2913   if (n->is_CFG()) {
2914     set_loop(n, loop);
2915     set_idom(n, pred, ddepth);
2916   } else {
2917     set_ctrl(n, pred);
2918   }
2919 }
2920 
2921 //------------------------------ proj_clone -------------------------------------
2922 // Utility to create an if-projection
2923 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2924   ProjNode* c = p->clone()->as_Proj();
2925   c->set_req(0, iff);
2926   return c;
2927 }
2928 
2929 //------------------------------ short_circuit_if -------------------------------------
2930 // Force the iff control output to be the live_proj
2931 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2932   guarantee(live_proj != nullptr, "null projection");
2933   int proj_con = live_proj->_con;
2934   assert(proj_con == 0 || proj_con == 1, "false or true projection");
2935   Node* con = intcon(proj_con);
2936   if (iff) {
2937     iff->set_req(1, con);
2938   }
2939   return con;
2940 }
2941 
2942 //------------------------------ insert_if_before_proj -------------------------------------
2943 // Insert a new if before an if projection (* - new node)
2944 //
2945 // before
2946 //           if(test)
2947 //           /     \
2948 //          v       v
2949 //    other-proj   proj (arg)
2950 //
2951 // after
2952 //           if(test)
2953 //           /     \
2954 //          /       v
2955 //         |      * proj-clone
2956 //         v          |
2957 //    other-proj      v
2958 //                * new_if(relop(cmp[IU](left,right)))
2959 //                  /  \
2960 //                 v    v
2961 //         * new-proj  proj
2962 //         (returned)
2963 //
2964 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2965   IfNode* iff = proj->in(0)->as_If();
2966   IdealLoopTree *loop = get_loop(proj);
2967   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2968   uint ddepth = dom_depth(proj);
2969 
2970   _igvn.rehash_node_delayed(iff);
2971   _igvn.rehash_node_delayed(proj);
2972 
2973   proj->set_req(0, nullptr);  // temporary disconnect
2974   ProjNode* proj2 = proj_clone(proj, iff);
2975   register_node(proj2, loop, iff, ddepth);
2976 
2977   Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2978   register_node(cmp, loop, proj2, ddepth);
2979 
2980   BoolNode* bol = new BoolNode(cmp, relop);
2981   register_node(bol, loop, proj2, ddepth);
2982 
2983   int opcode = iff->Opcode();
2984   assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2985   IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2986   register_node(new_if, loop, proj2, ddepth);
2987 
2988   proj->set_req(0, new_if); // reattach
2989   set_idom(proj, new_if, ddepth);
2990 
2991   ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2992   guarantee(new_exit != nullptr, "null exit node");
2993   register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2994 
2995   return new_exit;
2996 }
2997 
2998 //------------------------------ insert_region_before_proj -------------------------------------
2999 // Insert a region before an if projection (* - new node)
3000 //
3001 // before
3002 //           if(test)
3003 //          /      |
3004 //         v       |
3005 //       proj      v
3006 //               other-proj
3007 //
3008 // after
3009 //           if(test)
3010 //          /      |
3011 //         v       |
3012 // * proj-clone    v
3013 //         |     other-proj
3014 //         v
3015 // * new-region
3016 //         |
3017 //         v
3018 // *      dum_if
3019 //       /     \
3020 //      v       \
3021 // * dum-proj    v
3022 //              proj
3023 //
3024 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3025   IfNode* iff = proj->in(0)->as_If();
3026   IdealLoopTree *loop = get_loop(proj);
3027   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3028   uint ddepth = dom_depth(proj);
3029 
3030   _igvn.rehash_node_delayed(iff);
3031   _igvn.rehash_node_delayed(proj);
3032 
3033   proj->set_req(0, nullptr);  // temporary disconnect
3034   ProjNode* proj2 = proj_clone(proj, iff);
3035   register_node(proj2, loop, iff, ddepth);
3036 
3037   RegionNode* reg = new RegionNode(2);
3038   reg->set_req(1, proj2);
3039   register_node(reg, loop, iff, ddepth);
3040 
3041   IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3042   register_node(dum_if, loop, reg, ddepth);
3043 
3044   proj->set_req(0, dum_if); // reattach
3045   set_idom(proj, dum_if, ddepth);
3046 
3047   ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3048   register_node(dum_proj, loop, dum_if, ddepth);
3049 
3050   return reg;
3051 }
3052 
3053 // Idea
3054 // ----
3055 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3056 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3057 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3058 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3059 // test alone instead:
3060 //
3061 // Before Partial Peeling:
3062 //   Loop:
3063 //     <peeled section>
3064 //     Split off signed loop exit test
3065 //     <-- CUT HERE -->
3066 //     Unchanged unsigned loop exit test
3067 //     <rest of unpeeled section>
3068 //     goto Loop
3069 //
3070 // After Partial Peeling:
3071 //   <cloned peeled section>
3072 //   Cloned split off signed loop exit test
3073 //   Loop:
3074 //     Unchanged unsigned loop exit test
3075 //     <rest of unpeeled section>
3076 //     <peeled section>
3077 //     Split off signed loop exit test
3078 //     goto Loop
3079 //
3080 // Details
3081 // -------
3082 // Before:
3083 //          if (i <u limit)    Unsigned loop exit condition
3084 //         /       |
3085 //        v        v
3086 //   exit-proj   stay-in-loop-proj
3087 //
3088 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3089 // before the CmpU on the stay-in-loop path and keep both tests:
3090 //
3091 //          if (i <u limit)    Signed loop exit test
3092 //        /        |
3093 //       /  if (i <u limit)    Unsigned loop exit test
3094 //      /  /       |
3095 //     v  v        v
3096 //  exit-region  stay-in-loop-proj
3097 //
3098 // Implementation
3099 // --------------
3100 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3101 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3102 // exit tests is preserved, and their loop nesting is correct.
3103 //
3104 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3105 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3106 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3107 //
3108 //           if (stay-in-loop-const)  Killed original unsigned loop exit test
3109 //          /       |
3110 //         /        v
3111 //        /  if (i <  limit)          Split off signed loop exit test
3112 //       /  /       |
3113 //      /  /        v
3114 //     /  /  if (i <u limit)          Cloned unsigned loop exit test
3115 //    /  /   /      |
3116 //   v  v  v        |
3117 //  exit-region     |
3118 //        |         |
3119 //    dummy-if      |
3120 //     /  |         |
3121 // dead   |         |
3122 //        v         v
3123 //   exit-proj   stay-in-loop-proj
3124 //
3125 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3126 //       loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3127 //
3128 // Requirements
3129 // ------------
3130 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3131 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3132 // Otherwise, we need to bail out (see comments in the code below).
3133 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3134   const bool Signed   = true;
3135   const bool Unsigned = false;
3136 
3137   BoolNode* bol = if_cmpu->in(1)->as_Bool();
3138   if (bol->_test._test != BoolTest::lt) {
3139     return nullptr;
3140   }
3141   CmpNode* cmpu = bol->in(1)->as_Cmp();
3142   assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3143 
3144   int stride = stride_of_possible_iv(if_cmpu);
3145   if (stride == 0) {
3146     return nullptr;
3147   }
3148 
3149   Node* lp_proj = stay_in_loop(if_cmpu, loop);
3150   guarantee(lp_proj != nullptr, "null loop node");
3151 
3152   ProjNode* lp_continue = lp_proj->as_Proj();
3153   ProjNode* lp_exit     = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3154   if (!lp_exit->is_IfFalse()) {
3155     // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3156     // We therefore can't add a single exit condition.
3157     return nullptr;
3158   }
3159   // The unsigned loop exit condition is
3160   //   !(i <u  limit)
3161   // =   i >=u limit
3162   //
3163   // First, we note that for any x for which
3164   //   0 <= x <= INT_MAX
3165   // we can convert x to an unsigned int and still get the same guarantee:
3166   //   0 <=  (uint) x <=  INT_MAX = (uint) INT_MAX
3167   //   0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX   (LEMMA)
3168   //
3169   // With that in mind, if
3170   //   limit >= 0             (COND)
3171   // then the unsigned loop exit condition
3172   //   i >=u limit            (ULE)
3173   // is equivalent to
3174   //   i < 0 || i >= limit    (SLE-full)
3175   // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3176   //   (uint) i >=u MAX_INT >= limit >= 0
3177   // or otherwise
3178   //   i >= limit >= 0
3179   // holds due to (LEMMA).
3180   //
3181   // For completeness, a counterexample with limit < 0:
3182   // Assume i = -3 and limit = -2:
3183   //   i  < 0
3184   //   -2 < 0
3185   // is true and thus also "i < 0 || i >= limit". But
3186   //   i  >=u limit
3187   //   -3 >=u -2
3188   // is false.
3189   Node* limit = cmpu->in(2);
3190   const TypeInt* type_limit = _igvn.type(limit)->is_int();
3191   if (type_limit->_lo < 0) {
3192     return nullptr;
3193   }
3194 
3195   // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3196   //   stride < 0:
3197   //     i < 0        (SLE = SLE-negative)
3198   //   stride > 0:
3199   //     i >= limit   (SLE = SLE-positive)
3200   // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3201   //
3202   // Loop:
3203   //   <peeled section>
3204   //   i >= limit    (SLE-positive)
3205   //   <-- CUT HERE -->
3206   //   i >=u limit   (ULE)
3207   //   <rest of unpeeled section>
3208   //   goto Loop
3209   //
3210   // We exit the loop if:
3211   //   (SLE) is true OR (ULE) is true
3212   // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3213   // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3214   //   (SLE) IMPLIES (ULE)
3215   // This indeed holds when (COND) is given:
3216   // - stride > 0:
3217   //       i >=  limit             // (SLE = SLE-positive)
3218   //       i >=  limit >= 0        // (COND)
3219   //       i >=u limit >= 0        // (LEMMA)
3220   //     which is the unsigned loop exit condition (ULE).
3221   // - stride < 0:
3222   //       i        <  0           // (SLE = SLE-negative)
3223   //       (uint) i >u MAX_INT     // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3224   //       MAX_INT >= limit >= 0   // (COND)
3225   //       MAX_INT >=u limit >= 0  // (LEMMA)
3226   //     and thus from (NEG) and (LEMMA):
3227   //       i >=u limit
3228   //     which is the unsigned loop exit condition (ULE).
3229   //
3230   //
3231   // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3232   //   <cloned peeled section>
3233   //   i >= limit (SLE-positive)
3234   //   Loop:
3235   //     i >=u limit (ULE)
3236   //     <rest of unpeeled section>
3237   //     <peeled section>
3238   //     i >= limit (SLE-positive)
3239   //     goto Loop
3240   Node* rhs_cmpi;
3241   if (stride > 0) {
3242     rhs_cmpi = limit; // For i >= limit
3243   } else {
3244     rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3245   }
3246   // Create a new region on the exit path
3247   RegionNode* reg = insert_region_before_proj(lp_exit);
3248   guarantee(reg != nullptr, "null region node");
3249 
3250   // Clone the if-cmpu-true-false using a signed compare
3251   BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3252   ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3253   reg->add_req(cmpi_exit);
3254 
3255   // Clone the if-cmpu-true-false
3256   BoolTest::mask rel_u = bol->_test._test;
3257   ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3258   reg->add_req(cmpu_exit);
3259 
3260   // Force original if to stay in loop.
3261   short_circuit_if(if_cmpu, lp_continue);
3262 
3263   return cmpi_exit->in(0)->as_If();
3264 }
3265 
3266 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3267 // Remove a previously inserted signed compare loop exit.
3268 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3269   Node* lp_proj = stay_in_loop(if_cmp, loop);
3270   assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3271          stay_in_loop(lp_proj, loop)->is_If() &&
3272          stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3273   Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3274   if_cmp->set_req(1, con);
3275 }
3276 
3277 //------------------------------ scheduled_nodelist -------------------------------------
3278 // Create a post order schedule of nodes that are in the
3279 // "member" set.  The list is returned in "sched".
3280 // The first node in "sched" is the loop head, followed by
3281 // nodes which have no inputs in the "member" set, and then
3282 // followed by the nodes that have an immediate input dependence
3283 // on a node in "sched".
3284 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3285 
3286   assert(member.test(loop->_head->_idx), "loop head must be in member set");
3287   VectorSet visited;
3288   Node_Stack nstack(loop->_body.size());
3289 
3290   Node* n  = loop->_head;  // top of stack is cached in "n"
3291   uint idx = 0;
3292   visited.set(n->_idx);
3293 
3294   // Initially push all with no inputs from within member set
3295   for(uint i = 0; i < loop->_body.size(); i++ ) {
3296     Node *elt = loop->_body.at(i);
3297     if (member.test(elt->_idx)) {
3298       bool found = false;
3299       for (uint j = 0; j < elt->req(); j++) {
3300         Node* def = elt->in(j);
3301         if (def && member.test(def->_idx) && def != elt) {
3302           found = true;
3303           break;
3304         }
3305       }
3306       if (!found && elt != loop->_head) {
3307         nstack.push(n, idx);
3308         n = elt;
3309         assert(!visited.test(n->_idx), "not seen yet");
3310         visited.set(n->_idx);
3311       }
3312     }
3313   }
3314 
3315   // traverse out's that are in the member set
3316   while (true) {
3317     if (idx < n->outcnt()) {
3318       Node* use = n->raw_out(idx);
3319       idx++;
3320       if (!visited.test_set(use->_idx)) {
3321         if (member.test(use->_idx)) {
3322           nstack.push(n, idx);
3323           n = use;
3324           idx = 0;
3325         }
3326       }
3327     } else {
3328       // All outputs processed
3329       sched.push(n);
3330       if (nstack.is_empty()) break;
3331       n   = nstack.node();
3332       idx = nstack.index();
3333       nstack.pop();
3334     }
3335   }
3336 }
3337 
3338 
3339 //------------------------------ has_use_in_set -------------------------------------
3340 // Has a use in the vector set
3341 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3342   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3343     Node* use = n->fast_out(j);
3344     if (vset.test(use->_idx)) {
3345       return true;
3346     }
3347   }
3348   return false;
3349 }
3350 
3351 
3352 //------------------------------ has_use_internal_to_set -------------------------------------
3353 // Has use internal to the vector set (ie. not in a phi at the loop head)
3354 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3355   Node* head  = loop->_head;
3356   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3357     Node* use = n->fast_out(j);
3358     if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3359       return true;
3360     }
3361   }
3362   return false;
3363 }
3364 
3365 
3366 //------------------------------ clone_for_use_outside_loop -------------------------------------
3367 // clone "n" for uses that are outside of loop
3368 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3369   int cloned = 0;
3370   assert(worklist.size() == 0, "should be empty");
3371   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3372     Node* use = n->fast_out(j);
3373     if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3374       worklist.push(use);
3375     }
3376   }
3377 
3378   if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3379                           "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3380     return -1;
3381   }
3382 
3383   while( worklist.size() ) {
3384     Node *use = worklist.pop();
3385     if (!has_node(use) || use->in(0) == C->top()) continue;
3386     uint j;
3387     for (j = 0; j < use->req(); j++) {
3388       if (use->in(j) == n) break;
3389     }
3390     assert(j < use->req(), "must be there");
3391 
3392     // clone "n" and insert it between the inputs of "n" and the use outside the loop
3393     Node* n_clone = n->clone();
3394     _igvn.replace_input_of(use, j, n_clone);
3395     cloned++;
3396     Node* use_c;
3397     if (!use->is_Phi()) {
3398       use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3399     } else {
3400       // Use in a phi is considered a use in the associated predecessor block
3401       use_c = use->in(0)->in(j);
3402     }
3403     set_ctrl(n_clone, use_c);
3404     assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3405     get_loop(use_c)->_body.push(n_clone);
3406     _igvn.register_new_node_with_optimizer(n_clone);
3407 #ifndef PRODUCT
3408     if (TracePartialPeeling) {
3409       tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3410     }
3411 #endif
3412   }
3413   return cloned;
3414 }
3415 
3416 
3417 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3418 // clone "n" for special uses that are in the not_peeled region.
3419 // If these def-uses occur in separate blocks, the code generator
3420 // marks the method as not compilable.  For example, if a "BoolNode"
3421 // is in a different basic block than the "IfNode" that uses it, then
3422 // the compilation is aborted in the code generator.
3423 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3424                                                         VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3425   if (n->is_Phi() || n->is_Load()) {
3426     return;
3427   }
3428   assert(worklist.size() == 0, "should be empty");
3429   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3430     Node* use = n->fast_out(j);
3431     if ( not_peel.test(use->_idx) &&
3432          (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3433          use->in(1) == n)  {
3434       worklist.push(use);
3435     }
3436   }
3437   if (worklist.size() > 0) {
3438     // clone "n" and insert it between inputs of "n" and the use
3439     Node* n_clone = n->clone();
3440     loop->_body.push(n_clone);
3441     _igvn.register_new_node_with_optimizer(n_clone);
3442     set_ctrl(n_clone, get_ctrl(n));
3443     sink_list.push(n_clone);
3444     not_peel.set(n_clone->_idx);
3445 #ifndef PRODUCT
3446     if (TracePartialPeeling) {
3447       tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3448     }
3449 #endif
3450     while( worklist.size() ) {
3451       Node *use = worklist.pop();
3452       _igvn.rehash_node_delayed(use);
3453       for (uint j = 1; j < use->req(); j++) {
3454         if (use->in(j) == n) {
3455           use->set_req(j, n_clone);
3456         }
3457       }
3458     }
3459   }
3460 }
3461 
3462 
3463 //------------------------------ insert_phi_for_loop -------------------------------------
3464 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3465 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3466   Node *phi = PhiNode::make(lp, back_edge_val);
3467   phi->set_req(LoopNode::EntryControl, lp_entry_val);
3468   // Use existing phi if it already exists
3469   Node *hit = _igvn.hash_find_insert(phi);
3470   if( hit == nullptr ) {
3471     _igvn.register_new_node_with_optimizer(phi);
3472     set_ctrl(phi, lp);
3473   } else {
3474     // Remove the new phi from the graph and use the hit
3475     _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Speculative);
3476     phi = hit;
3477   }
3478   _igvn.replace_input_of(use, idx, phi);
3479 }
3480 
3481 #ifdef ASSERT
3482 //------------------------------ is_valid_loop_partition -------------------------------------
3483 // Validate the loop partition sets: peel and not_peel
3484 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3485                                               VectorSet& not_peel ) {
3486   uint i;
3487   // Check that peel_list entries are in the peel set
3488   for (i = 0; i < peel_list.size(); i++) {
3489     if (!peel.test(peel_list.at(i)->_idx)) {
3490       return false;
3491     }
3492   }
3493   // Check at loop members are in one of peel set or not_peel set
3494   for (i = 0; i < loop->_body.size(); i++ ) {
3495     Node *def  = loop->_body.at(i);
3496     uint di = def->_idx;
3497     // Check that peel set elements are in peel_list
3498     if (peel.test(di)) {
3499       if (not_peel.test(di)) {
3500         return false;
3501       }
3502       // Must be in peel_list also
3503       bool found = false;
3504       for (uint j = 0; j < peel_list.size(); j++) {
3505         if (peel_list.at(j)->_idx == di) {
3506           found = true;
3507           break;
3508         }
3509       }
3510       if (!found) {
3511         return false;
3512       }
3513     } else if (not_peel.test(di)) {
3514       if (peel.test(di)) {
3515         return false;
3516       }
3517     } else {
3518       return false;
3519     }
3520   }
3521   return true;
3522 }
3523 
3524 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3525 // Ensure a use outside of loop is of the right form
3526 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3527   Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3528   return (use->is_Phi() &&
3529           use_c->is_Region() && use_c->req() == 3 &&
3530           (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3531            use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3532            use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3533           loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3534 }
3535 
3536 //------------------------------ is_valid_clone_loop_form -------------------------------------
3537 // Ensure that all uses outside of loop are of the right form
3538 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3539                                                uint orig_exit_idx, uint clone_exit_idx) {
3540   uint len = peel_list.size();
3541   for (uint i = 0; i < len; i++) {
3542     Node *def = peel_list.at(i);
3543 
3544     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3545       Node *use = def->fast_out(j);
3546       Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3547       if (!loop->is_member(get_loop(use_c))) {
3548         // use is not in the loop, check for correct structure
3549         if (use->in(0) == def) {
3550           // Okay
3551         } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3552           return false;
3553         }
3554       }
3555     }
3556   }
3557   return true;
3558 }
3559 #endif
3560 
3561 //------------------------------ partial_peel -------------------------------------
3562 // Partially peel (aka loop rotation) the top portion of a loop (called
3563 // the peel section below) by cloning it and placing one copy just before
3564 // the new loop head and the other copy at the bottom of the new loop.
3565 //
3566 //    before                       after                where it came from
3567 //
3568 //    stmt1                        stmt1
3569 //  loop:                          stmt2                     clone
3570 //    stmt2                        if condA goto exitA       clone
3571 //    if condA goto exitA        new_loop:                   new
3572 //    stmt3                        stmt3                     clone
3573 //    if !condB goto loop          if condB goto exitB       clone
3574 //  exitB:                         stmt2                     orig
3575 //    stmt4                        if !condA goto new_loop   orig
3576 //  exitA:                         goto exitA
3577 //                               exitB:
3578 //                                 stmt4
3579 //                               exitA:
3580 //
3581 // Step 1: find the cut point: an exit test on probable
3582 //         induction variable.
3583 // Step 2: schedule (with cloning) operations in the peel
3584 //         section that can be executed after the cut into
3585 //         the section that is not peeled.  This may need
3586 //         to clone operations into exit blocks.  For
3587 //         instance, a reference to A[i] in the not-peel
3588 //         section and a reference to B[i] in an exit block
3589 //         may cause a left-shift of i by 2 to be placed
3590 //         in the peel block.  This step will clone the left
3591 //         shift into the exit block and sink the left shift
3592 //         from the peel to the not-peel section.
3593 // Step 3: clone the loop, retarget the control, and insert
3594 //         phis for values that are live across the new loop
3595 //         head.  This is very dependent on the graph structure
3596 //         from clone_loop.  It creates region nodes for
3597 //         exit control and associated phi nodes for values
3598 //         flow out of the loop through that exit.  The region
3599 //         node is dominated by the clone's control projection.
3600 //         So the clone's peel section is placed before the
3601 //         new loop head, and the clone's not-peel section is
3602 //         forms the top part of the new loop.  The original
3603 //         peel section forms the tail of the new loop.
3604 // Step 4: update the dominator tree and recompute the
3605 //         dominator depth.
3606 //
3607 //                   orig
3608 //
3609 //                   stmt1
3610 //                     |
3611 //                     v
3612 //                 predicates
3613 //                     |
3614 //                     v
3615 //                   loop<----+
3616 //                     |      |
3617 //                   stmt2    |
3618 //                     |      |
3619 //                     v      |
3620 //                    ifA     |
3621 //                   / |      |
3622 //                  v  v      |
3623 //               false true   ^  <-- last_peel
3624 //               /     |      |
3625 //              /   ===|==cut |
3626 //             /     stmt3    |  <-- first_not_peel
3627 //            /        |      |
3628 //            |        v      |
3629 //            v       ifB     |
3630 //          exitA:   / \      |
3631 //                  /   \     |
3632 //                 v     v    |
3633 //               false true   |
3634 //               /       \    |
3635 //              /         ----+
3636 //             |
3637 //             v
3638 //           exitB:
3639 //           stmt4
3640 //
3641 //
3642 //            after clone loop
3643 //
3644 //                   stmt1
3645 //                     |
3646 //                     v
3647 //                predicates
3648 //                 /       \
3649 //        clone   /         \   orig
3650 //               /           \
3651 //              /             \
3652 //             v               v
3653 //   +---->loop                loop<----+
3654 //   |      |                    |      |
3655 //   |    stmt2                stmt2    |
3656 //   |      |                    |      |
3657 //   |      v                    v      |
3658 //   |      ifA                 ifA     |
3659 //   |      | \                / |      |
3660 //   |      v  v              v  v      |
3661 //   ^    true  false      false true   ^  <-- last_peel
3662 //   |      |   ^   \       /    |      |
3663 //   | cut==|==  \   \     /  ===|==cut |
3664 //   |    stmt3   \   \   /    stmt3    |  <-- first_not_peel
3665 //   |      |    dom   | |       |      |
3666 //   |      v      \  1v v2      v      |
3667 //   |      ifB     regionA     ifB     |
3668 //   |      / \        |       / \      |
3669 //   |     /   \       v      /   \     |
3670 //   |    v     v    exitA:  v     v    |
3671 //   |    true  false      false true   |
3672 //   |    /     ^   \      /       \    |
3673 //   +----       \   \    /         ----+
3674 //               dom  \  /
3675 //                 \  1v v2
3676 //                  regionB
3677 //                     |
3678 //                     v
3679 //                   exitB:
3680 //                   stmt4
3681 //
3682 //
3683 //           after partial peel
3684 //
3685 //                  stmt1
3686 //                     |
3687 //                     v
3688 //                predicates
3689 //                 /
3690 //        clone   /             orig
3691 //               /          TOP
3692 //              /             \
3693 //             v               v
3694 //    TOP->loop                loop----+
3695 //          |                    |      |
3696 //        stmt2                stmt2    |
3697 //          |                    |      |
3698 //          v                    v      |
3699 //          ifA                 ifA     |
3700 //          | \                / |      |
3701 //          v  v              v  v      |
3702 //        true  false      false true   |     <-- last_peel
3703 //          |   ^   \       /    +------|---+
3704 //  +->newloop   \   \     /  === ==cut |   |
3705 //  |     stmt3   \   \   /     TOP     |   |
3706 //  |       |    dom   | |      stmt3   |   | <-- first_not_peel
3707 //  |       v      \  1v v2      v      |   |
3708 //  |       ifB     regionA     ifB     ^   v
3709 //  |       / \        |       / \      |   |
3710 //  |      /   \       v      /   \     |   |
3711 //  |     v     v    exitA:  v     v    |   |
3712 //  |     true  false      false true   |   |
3713 //  |     /     ^   \      /       \    |   |
3714 //  |    |       \   \    /         v   |   |
3715 //  |    |       dom  \  /         TOP  |   |
3716 //  |    |         \  1v v2             |   |
3717 //  ^    v          regionB             |   |
3718 //  |    |             |                |   |
3719 //  |    |             v                ^   v
3720 //  |    |           exitB:             |   |
3721 //  |    |           stmt4              |   |
3722 //  |    +------------>-----------------+   |
3723 //  |                                       |
3724 //  +-----------------<---------------------+
3725 //
3726 //
3727 //              final graph
3728 //
3729 //                  stmt1
3730 //                    |
3731 //                    v
3732 //                predicates
3733 //                    |
3734 //                    v
3735 //                  stmt2 clone
3736 //                    |
3737 //                    v
3738 //         ........> ifA clone
3739 //         :        / |
3740 //        dom      /  |
3741 //         :      v   v
3742 //         :  false   true
3743 //         :  |       |
3744 //         :  |       v
3745 //         :  |    newloop<-----+
3746 //         :  |        |        |
3747 //         :  |     stmt3 clone |
3748 //         :  |        |        |
3749 //         :  |        v        |
3750 //         :  |       ifB       |
3751 //         :  |      / \        |
3752 //         :  |     v   v       |
3753 //         :  |  false true     |
3754 //         :  |   |     |       |
3755 //         :  |   v    stmt2    |
3756 //         :  | exitB:  |       |
3757 //         :  | stmt4   v       |
3758 //         :  |       ifA orig  |
3759 //         :  |      /  \       |
3760 //         :  |     /    \      |
3761 //         :  |    v     v      |
3762 //         :  |  false  true    |
3763 //         :  |  /        \     |
3764 //         :  v  v         -----+
3765 //          RegionA
3766 //             |
3767 //             v
3768 //           exitA
3769 //
3770 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3771 
3772   assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3773   if (!loop->_head->is_Loop()) {
3774     return false;
3775   }
3776   LoopNode *head = loop->_head->as_Loop();
3777 
3778   if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3779     return false;
3780   }
3781 
3782   // Check for complex exit control
3783   for (uint ii = 0; ii < loop->_body.size(); ii++) {
3784     Node *n = loop->_body.at(ii);
3785     int opc = n->Opcode();
3786     if (n->is_Call()        ||
3787         opc == Op_Catch     ||
3788         opc == Op_CatchProj ||
3789         opc == Op_Jump      ||
3790         opc == Op_JumpProj) {
3791 #ifndef PRODUCT
3792       if (TracePartialPeeling) {
3793         tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3794       }
3795 #endif
3796       return false;
3797     }
3798   }
3799 
3800   int dd = dom_depth(head);
3801 
3802   // Step 1: find cut point
3803 
3804   // Walk up dominators to loop head looking for first loop exit
3805   // which is executed on every path thru loop.
3806   IfNode *peel_if = nullptr;
3807   IfNode *peel_if_cmpu = nullptr;
3808 
3809   Node *iff = loop->tail();
3810   while (iff != head) {
3811     if (iff->is_If()) {
3812       Node *ctrl = get_ctrl(iff->in(1));
3813       if (ctrl->is_top()) return false; // Dead test on live IF.
3814       // If loop-varying exit-test, check for induction variable
3815       if (loop->is_member(get_loop(ctrl)) &&
3816           loop->is_loop_exit(iff) &&
3817           is_possible_iv_test(iff)) {
3818         Node* cmp = iff->in(1)->in(1);
3819         if (cmp->Opcode() == Op_CmpI) {
3820           peel_if = iff->as_If();
3821         } else {
3822           assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3823           peel_if_cmpu = iff->as_If();
3824         }
3825       }
3826     }
3827     iff = idom(iff);
3828   }
3829 
3830   // Prefer signed compare over unsigned compare.
3831   IfNode* new_peel_if = nullptr;
3832   if (peel_if == nullptr) {
3833     if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3834       return false;   // No peel point found
3835     }
3836     new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3837     if (new_peel_if == nullptr) {
3838       return false;   // No peel point found
3839     }
3840     peel_if = new_peel_if;
3841   }
3842   Node* last_peel        = stay_in_loop(peel_if, loop);
3843   Node* first_not_peeled = stay_in_loop(last_peel, loop);
3844   if (first_not_peeled == nullptr || first_not_peeled == head) {
3845     return false;
3846   }
3847 
3848 #ifndef PRODUCT
3849   if (TraceLoopOpts) {
3850     tty->print("PartialPeel  ");
3851     loop->dump_head();
3852   }
3853 
3854   if (TracePartialPeeling) {
3855     tty->print_cr("before partial peel one iteration");
3856     Node_List wl;
3857     Node* t = head->in(2);
3858     while (true) {
3859       wl.push(t);
3860       if (t == head) break;
3861       t = idom(t);
3862     }
3863     while (wl.size() > 0) {
3864       Node* tt = wl.pop();
3865       tt->dump();
3866       if (tt == last_peel) tty->print_cr("-- cut --");
3867     }
3868   }
3869 #endif
3870 
3871   C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3872 
3873   VectorSet peel;
3874   VectorSet not_peel;
3875   Node_List peel_list;
3876   Node_List worklist;
3877   Node_List sink_list;
3878 
3879   uint estimate = loop->est_loop_clone_sz(1);
3880   if (exceeding_node_budget(estimate)) {
3881     return false;
3882   }
3883 
3884   // Set of cfg nodes to peel are those that are executable from
3885   // the head through last_peel.
3886   assert(worklist.size() == 0, "should be empty");
3887   worklist.push(head);
3888   peel.set(head->_idx);
3889   while (worklist.size() > 0) {
3890     Node *n = worklist.pop();
3891     if (n != last_peel) {
3892       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3893         Node* use = n->fast_out(j);
3894         if (use->is_CFG() &&
3895             loop->is_member(get_loop(use)) &&
3896             !peel.test_set(use->_idx)) {
3897           worklist.push(use);
3898         }
3899       }
3900     }
3901   }
3902 
3903   // Set of non-cfg nodes to peel are those that are control
3904   // dependent on the cfg nodes.
3905   for (uint i = 0; i < loop->_body.size(); i++) {
3906     Node *n = loop->_body.at(i);
3907     Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3908     if (peel.test(n_c->_idx)) {
3909       peel.set(n->_idx);
3910     } else {
3911       not_peel.set(n->_idx);
3912     }
3913   }
3914 
3915   // Step 2: move operations from the peeled section down into the
3916   //         not-peeled section
3917 
3918   // Get a post order schedule of nodes in the peel region
3919   // Result in right-most operand.
3920   scheduled_nodelist(loop, peel, peel_list);
3921 
3922   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3923 
3924   // For future check for too many new phis
3925   uint old_phi_cnt = 0;
3926   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3927     Node* use = head->fast_out(j);
3928     if (use->is_Phi()) old_phi_cnt++;
3929   }
3930 
3931 #ifndef PRODUCT
3932   if (TracePartialPeeling) {
3933     tty->print_cr("\npeeled list");
3934   }
3935 #endif
3936 
3937   // Evacuate nodes in peel region into the not_peeled region if possible
3938   bool too_many_clones = false;
3939   uint new_phi_cnt = 0;
3940   uint cloned_for_outside_use = 0;
3941   for (uint i = 0; i < peel_list.size();) {
3942     Node* n = peel_list.at(i);
3943 #ifndef PRODUCT
3944   if (TracePartialPeeling) n->dump();
3945 #endif
3946     bool incr = true;
3947     if (!n->is_CFG()) {
3948       if (has_use_in_set(n, not_peel)) {
3949         // If not used internal to the peeled region,
3950         // move "n" from peeled to not_peeled region.
3951         if (!has_use_internal_to_set(n, peel, loop)) {
3952           // if not pinned and not a load (which maybe anti-dependent on a store)
3953           // and not a CMove (Matcher expects only bool->cmove).
3954           if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3955             int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3956             if (C->failing()) return false;
3957             if (new_clones == -1) {
3958               too_many_clones = true;
3959               break;
3960             }
3961             cloned_for_outside_use += new_clones;
3962             sink_list.push(n);
3963             peel.remove(n->_idx);
3964             not_peel.set(n->_idx);
3965             peel_list.remove(i);
3966             incr = false;
3967 #ifndef PRODUCT
3968             if (TracePartialPeeling) {
3969               tty->print_cr("sink to not_peeled region: %d newbb: %d",
3970                             n->_idx, get_ctrl(n)->_idx);
3971             }
3972 #endif
3973           }
3974         } else {
3975           // Otherwise check for special def-use cases that span
3976           // the peel/not_peel boundary such as bool->if
3977           clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3978           new_phi_cnt++;
3979         }
3980       }
3981     }
3982     if (incr) i++;
3983   }
3984 
3985   estimate += cloned_for_outside_use + new_phi_cnt;
3986   bool exceed_node_budget = !may_require_nodes(estimate);
3987   bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3988 
3989   if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3990 #ifndef PRODUCT
3991     if (TracePartialPeeling && exceed_phi_limit) {
3992       tty->print_cr("\nToo many new phis: %d  old %d new cmpi: %c",
3993                     new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3994     }
3995 #endif
3996     if (new_peel_if != nullptr) {
3997       remove_cmpi_loop_exit(new_peel_if, loop);
3998     }
3999     // Inhibit more partial peeling on this loop
4000     assert(!head->is_partial_peel_loop(), "not partial peeled");
4001     head->mark_partial_peel_failed();
4002     if (cloned_for_outside_use > 0) {
4003       // Terminate this round of loop opts because
4004       // the graph outside this loop was changed.
4005       C->set_major_progress();
4006       return true;
4007     }
4008     return false;
4009   }
4010 
4011   // Step 3: clone loop, retarget control, and insert new phis
4012 
4013   // Create new loop head for new phis and to hang
4014   // the nodes being moved (sinked) from the peel region.
4015   LoopNode* new_head = new LoopNode(last_peel, last_peel);
4016   new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4017   _igvn.register_new_node_with_optimizer(new_head);
4018   assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4019   _igvn.replace_input_of(first_not_peeled, 0, new_head);
4020   set_loop(new_head, loop);
4021   loop->_body.push(new_head);
4022   not_peel.set(new_head->_idx);
4023   set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4024   set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4025 
4026   while (sink_list.size() > 0) {
4027     Node* n = sink_list.pop();
4028     set_ctrl(n, new_head);
4029   }
4030 
4031   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4032 
4033   clone_loop(loop, old_new, dd, IgnoreStripMined);
4034 
4035   const uint clone_exit_idx = 1;
4036   const uint orig_exit_idx  = 2;
4037   assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4038 
4039   Node* head_clone             = old_new[head->_idx];
4040   LoopNode* new_head_clone     = old_new[new_head->_idx]->as_Loop();
4041   Node* orig_tail_clone        = head_clone->in(2);
4042 
4043   // Add phi if "def" node is in peel set and "use" is not
4044 
4045   for (uint i = 0; i < peel_list.size(); i++) {
4046     Node *def  = peel_list.at(i);
4047     if (!def->is_CFG()) {
4048       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4049         Node *use = def->fast_out(j);
4050         if (has_node(use) && use->in(0) != C->top() &&
4051             (!peel.test(use->_idx) ||
4052              (use->is_Phi() && use->in(0) == head)) ) {
4053           worklist.push(use);
4054         }
4055       }
4056       while( worklist.size() ) {
4057         Node *use = worklist.pop();
4058         for (uint j = 1; j < use->req(); j++) {
4059           Node* n = use->in(j);
4060           if (n == def) {
4061 
4062             // "def" is in peel set, "use" is not in peel set
4063             // or "use" is in the entry boundary (a phi) of the peel set
4064 
4065             Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4066 
4067             if ( loop->is_member(get_loop( use_c )) ) {
4068               // use is in loop
4069               if (old_new[use->_idx] != nullptr) { // null for dead code
4070                 Node* use_clone = old_new[use->_idx];
4071                 _igvn.replace_input_of(use, j, C->top());
4072                 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4073               }
4074             } else {
4075               assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4076               // use is not in the loop, check if the live range includes the cut
4077               Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4078               if (not_peel.test(lp_if->_idx)) {
4079                 assert(j == orig_exit_idx, "use from original loop");
4080                 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4081               }
4082             }
4083           }
4084         }
4085       }
4086     }
4087   }
4088 
4089   // Step 3b: retarget control
4090 
4091   // Redirect control to the new loop head if a cloned node in
4092   // the not_peeled region has control that points into the peeled region.
4093   // This necessary because the cloned peeled region will be outside
4094   // the loop.
4095   //                            from    to
4096   //          cloned-peeled    <---+
4097   //    new_head_clone:            |    <--+
4098   //          cloned-not_peeled  in(0)    in(0)
4099   //          orig-peeled
4100 
4101   for (uint i = 0; i < loop->_body.size(); i++) {
4102     Node *n = loop->_body.at(i);
4103     if (!n->is_CFG()           && n->in(0) != nullptr        &&
4104         not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4105       Node* n_clone = old_new[n->_idx];
4106       if (n_clone->depends_only_on_test()) {
4107         // If this node depends_only_on_test, it will be rewire to the loop head, which is not the
4108         // correct test
4109         Node* pinned_clone = n_clone->pin_node_under_control();
4110         if (pinned_clone != nullptr) {
4111           register_new_node_with_ctrl_of(pinned_clone, n_clone);
4112           old_new.map(n->_idx, pinned_clone);
4113           _igvn.replace_node(n_clone, pinned_clone);
4114           n_clone = pinned_clone;
4115         }
4116       }
4117       _igvn.replace_input_of(n_clone, 0, new_head_clone);
4118     }
4119   }
4120 
4121   // Backedge of the surviving new_head (the clone) is original last_peel
4122   _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4123 
4124   // Cut first node in original not_peel set
4125   _igvn.rehash_node_delayed(new_head);                     // Multiple edge updates:
4126   new_head->set_req(LoopNode::EntryControl,    C->top());  //   use rehash_node_delayed / set_req instead of
4127   new_head->set_req(LoopNode::LoopBackControl, C->top());  //   multiple replace_input_of calls
4128 
4129   // Copy head_clone back-branch info to original head
4130   // and remove original head's loop entry and
4131   // clone head's back-branch
4132   _igvn.rehash_node_delayed(head); // Multiple edge updates
4133   head->set_req(LoopNode::EntryControl,    head_clone->in(LoopNode::LoopBackControl));
4134   head->set_req(LoopNode::LoopBackControl, C->top());
4135   _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4136 
4137   // Similarly modify the phis
4138   for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4139     Node* use = head->fast_out(k);
4140     if (use->is_Phi() && use->outcnt() > 0) {
4141       Node* use_clone = old_new[use->_idx];
4142       _igvn.rehash_node_delayed(use); // Multiple edge updates
4143       use->set_req(LoopNode::EntryControl,    use_clone->in(LoopNode::LoopBackControl));
4144       use->set_req(LoopNode::LoopBackControl, C->top());
4145       _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4146     }
4147   }
4148 
4149   // Step 4: update dominator tree and dominator depth
4150 
4151   set_idom(head, orig_tail_clone, dd);
4152   recompute_dom_depth();
4153 
4154   // Inhibit more partial peeling on this loop
4155   new_head_clone->set_partial_peel_loop();
4156   C->set_major_progress();
4157   loop->record_for_igvn();
4158 
4159 #ifndef PRODUCT
4160   if (TracePartialPeeling) {
4161     tty->print_cr("\nafter partial peel one iteration");
4162     Node_List wl;
4163     Node* t = last_peel;
4164     while (true) {
4165       wl.push(t);
4166       if (t == head_clone) break;
4167       t = idom(t);
4168     }
4169     while (wl.size() > 0) {
4170       Node* tt = wl.pop();
4171       if (tt == head) tty->print_cr("orig head");
4172       else if (tt == new_head_clone) tty->print_cr("new head");
4173       else if (tt == head_clone) tty->print_cr("clone head");
4174       tt->dump();
4175     }
4176   }
4177 #endif
4178 
4179   C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4180 
4181   return true;
4182 }
4183 
4184 #ifdef ASSERT
4185 
4186 // Moves Template Assertion Predicates to a target loop by cloning and killing the old ones. The target loop is the
4187 // original, not-cloned loop. This is currently only used with StressLoopBackedge which is a develop flag only and
4188 // false with product builds. We can therefore guard it with an ifdef. More details can be found at the use-site.
4189 class MoveAssertionPredicatesVisitor : public PredicateVisitor {
4190   ClonePredicateToTargetLoop _clone_predicate_to_loop;
4191   PhaseIdealLoop* const _phase;
4192 
4193 public:
4194   MoveAssertionPredicatesVisitor(LoopNode* target_loop_head,
4195                                  const NodeInSingleLoopBody &node_in_loop_body,
4196                                  PhaseIdealLoop* phase)
4197     : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
4198       _phase(phase) {
4199   }
4200   NONCOPYABLE(MoveAssertionPredicatesVisitor);
4201 
4202   using PredicateVisitor::visit;
4203 
4204   void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
4205     _clone_predicate_to_loop.clone_template_assertion_predicate(template_assertion_predicate);
4206     template_assertion_predicate.kill(_phase->igvn());
4207   }
4208 };
4209 #endif // ASSERT
4210 
4211 // Transform:
4212 //
4213 // loop<-----------------+
4214 //  |                    |
4215 // stmt1 stmt2 .. stmtn  |
4216 //  |     |        |     |
4217 //  \     |       /      |
4218 //    v   v     v        |
4219 //       region          |
4220 //         |             |
4221 //     shared_stmt       |
4222 //         |             |
4223 //         v             |
4224 //         if            |
4225 //         / \           |
4226 //        |   -----------+
4227 //        v
4228 //
4229 // into:
4230 //
4231 //    loop<-------------------+
4232 //     |                      |
4233 //     v                      |
4234 // +->loop                    |
4235 // |   |                      |
4236 // |  stmt1 stmt2 .. stmtn    |
4237 // |   |     |        |       |
4238 // |   |      \       /       |
4239 // |   |       v     v        |
4240 // |   |        region1       |
4241 // |   |           |          |
4242 // |  shared_stmt shared_stmt |
4243 // |   |           |          |
4244 // |   v           v          |
4245 // |   if          if         |
4246 // |   /\          / \        |
4247 // +--   |         |   -------+
4248 //       \         /
4249 //        v       v
4250 //         region2
4251 //
4252 // (region2 is shown to merge mirrored projections of the loop exit
4253 // ifs to make the diagram clearer but they really merge the same
4254 // projection)
4255 //
4256 // Conditions for this transformation to trigger:
4257 // - the path through stmt1 is frequent enough
4258 // - the inner loop will be turned into a counted loop after transformation
4259 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4260   if (!DuplicateBackedge) {
4261     return false;
4262   }
4263   assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4264   if (!loop->_head->is_Loop()) {
4265     return false;
4266   }
4267 
4268   uint estimate = loop->est_loop_clone_sz(1);
4269   if (exceeding_node_budget(estimate)) {
4270     return false;
4271   }
4272 
4273   LoopNode *head = loop->_head->as_Loop();
4274 
4275   Node* region = nullptr;
4276   IfNode* exit_test = nullptr;
4277   uint inner;
4278   float f;
4279 #ifdef ASSERT
4280   if (StressDuplicateBackedge) {
4281     if (head->is_strip_mined()) {
4282       return false;
4283     }
4284     Node* c = head->in(LoopNode::LoopBackControl);
4285 
4286     while (c != head) {
4287       if (c->is_Region()) {
4288         region = c;
4289       }
4290       c = idom(c);
4291     }
4292 
4293     if (region == nullptr) {
4294       return false;
4295     }
4296 
4297     inner = 1;
4298   } else
4299 #endif //ASSERT
4300   {
4301     // Is the shape of the loop that of a counted loop...
4302     Node* back_control = loop_exit_control(loop);
4303     if (back_control == nullptr) {
4304       return false;
4305     }
4306 
4307     LoopExitTest loop_exit(back_control, loop, this);
4308     loop_exit.build();
4309     if (!loop_exit.is_valid_with_bt(T_INT)) {
4310       return false;
4311     }
4312 
4313     const Node* loop_incr = loop_exit.incr();
4314 
4315     // With an extra phi for the candidate iv?
4316     // Or the region node is the loop head
4317     if (!loop_incr->is_Phi() || loop_incr->in(0) == head) {
4318       return false;
4319     }
4320 
4321     PathFrequency pf(head, this);
4322     region = loop_incr->in(0);
4323 
4324     // Go over all paths for the extra phi's region and see if that
4325     // path is frequent enough and would match the expected iv shape
4326     // if the extra phi is removed
4327     inner = 0;
4328     for (uint i = 1; i < loop_incr->req(); ++i) {
4329       CountedLoopConverter::TruncatedIncrement increment(T_INT);
4330       increment.build(loop_incr->in(i));
4331       if (!increment.is_valid()) {
4332         continue;
4333       }
4334       assert(increment.incr()->Opcode() == Op_AddI, "wrong increment code");
4335 
4336       LoopIVStride stride = LoopIVStride(T_INT);
4337       stride.build(increment.incr());
4338       if (!stride.is_valid()) {
4339         continue;
4340       }
4341 
4342       PhiNode* phi = loop_iv_phi(stride.xphi(), nullptr, head);
4343       if (phi == nullptr ||
4344           (increment.outer_trunc() == nullptr && phi->in(LoopNode::LoopBackControl) != loop_exit.incr()) ||
4345           (increment.outer_trunc() != nullptr && phi->in(LoopNode::LoopBackControl) != increment.outer_trunc())) {
4346         return false;
4347       }
4348 
4349       f = pf.to(region->in(i));
4350       if (f > 0.5) {
4351         inner = i;
4352         break;
4353       }
4354     }
4355 
4356     if (inner == 0) {
4357       return false;
4358     }
4359 
4360     exit_test = back_control->in(0)->as_If();
4361   }
4362 
4363   if (idom(region)->is_Catch()) {
4364     return false;
4365   }
4366 
4367   // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4368   Unique_Node_List wq;
4369   wq.push(head->in(LoopNode::LoopBackControl));
4370   for (uint i = 0; i < wq.size(); i++) {
4371     Node* c = wq.at(i);
4372     assert(get_loop(c) == loop, "not in the right loop?");
4373     if (c->is_Region()) {
4374       if (c != region) {
4375         for (uint j = 1; j < c->req(); ++j) {
4376           wq.push(c->in(j));
4377         }
4378       }
4379     } else {
4380       wq.push(c->in(0));
4381     }
4382     assert(!is_strict_dominator(c, region), "shouldn't go above region");
4383   }
4384 
4385   Node* region_dom = idom(region);
4386 
4387   // Can't do the transformation if this would cause a membar pair to
4388   // be split
4389   for (uint i = 0; i < wq.size(); i++) {
4390     Node* c = wq.at(i);
4391     if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4392       assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4393       if (!wq.member(c->as_MemBar()->leading_membar())) {
4394         return false;
4395       }
4396     }
4397   }
4398   C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head);
4399 
4400   // Collect data nodes that need to be clones as well
4401   int dd = dom_depth(head);
4402 
4403   for (uint i = 0; i < loop->_body.size(); ++i) {
4404     Node* n = loop->_body.at(i);
4405     if (has_ctrl(n)) {
4406       Node* c = get_ctrl(n);
4407       if (wq.member(c)) {
4408         wq.push(n);
4409       }
4410     } else {
4411       set_idom(n, idom(n), dd);
4412     }
4413   }
4414 
4415   // clone shared_stmt
4416   clone_loop_body(wq, old_new, nullptr);
4417 
4418   Node* region_clone = old_new[region->_idx];
4419   region_clone->set_req(inner, C->top());
4420   set_idom(region, region->in(inner), dd);
4421 
4422   // Prepare the outer loop
4423   Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4424   register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4425   _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4426   set_idom(head, outer_head, dd);
4427 
4428   fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4429 
4430   // Make one of the shared_stmt copies only reachable from stmt1, the
4431   // other only from stmt2..stmtn.
4432   Node* dom = nullptr;
4433   for (uint i = 1; i < region->req(); ++i) {
4434     if (i != inner) {
4435       _igvn.replace_input_of(region, i, C->top());
4436     }
4437     Node* in = region_clone->in(i);
4438     if (in->is_top()) {
4439       continue;
4440     }
4441     if (dom == nullptr) {
4442       dom = in;
4443     } else {
4444       dom = dom_lca(dom, in);
4445     }
4446   }
4447 
4448   set_idom(region_clone, dom, dd);
4449 
4450   // Set up the outer loop
4451   for (uint i = 0; i < head->outcnt(); i++) {
4452     Node* u = head->raw_out(i);
4453     if (u->is_Phi()) {
4454       Node* outer_phi = u->clone();
4455       outer_phi->set_req(0, outer_head);
4456       Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4457       if (backedge == nullptr) {
4458         backedge = u->in(LoopNode::LoopBackControl);
4459       }
4460       outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4461       register_new_node(outer_phi, outer_head);
4462       _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4463     }
4464   }
4465 
4466   // create control and data nodes for out of loop uses (including region2)
4467   Node_List worklist;
4468   uint new_counter = C->unique();
4469   fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4470 
4471   Node_List *split_if_set = nullptr;
4472   Node_List *split_bool_set = nullptr;
4473   Node_List *split_cex_set = nullptr;
4474   fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4475                 split_if_set, split_bool_set, split_cex_set);
4476 
4477   finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4478 
4479   if (exit_test != nullptr) {
4480     float cnt = exit_test->_fcnt;
4481     if (cnt != COUNT_UNKNOWN) {
4482       exit_test->_fcnt = cnt * f;
4483       old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4484     }
4485   }
4486 
4487 #ifdef ASSERT
4488   if (StressDuplicateBackedge && head->is_CountedLoop()) {
4489     // The Template Assertion Predicates from the old counted loop are now at the new outer loop - clone them to
4490     // the inner counted loop and kill the old ones. We only need to do this with debug builds because
4491     // StressDuplicateBackedge is a devlop flag and false by default. Without StressDuplicateBackedge 'head' will be a
4492     // non-counted loop, and thus we have no Template Assertion Predicates above the old loop to move down.
4493     PredicateIterator predicate_iterator(outer_head->in(LoopNode::EntryControl));
4494     NodeInSingleLoopBody node_in_body(this, loop);
4495     MoveAssertionPredicatesVisitor move_assertion_predicates_visitor(head, node_in_body, this);
4496     predicate_iterator.for_each(move_assertion_predicates_visitor);
4497   }
4498 #endif // ASSERT
4499 
4500   C->set_major_progress();
4501 
4502   C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head);
4503 
4504   return true;
4505 }
4506 
4507 // AutoVectorize the loop: replace scalar ops with vector ops.
4508 PhaseIdealLoop::AutoVectorizeStatus
4509 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4510   // Counted loop only
4511   if (!lpt->is_counted()) {
4512     return AutoVectorizeStatus::Impossible;
4513   }
4514 
4515   // Main-loop only
4516   CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4517   if (!cl->is_main_loop()) {
4518     return AutoVectorizeStatus::Impossible;
4519   }
4520 
4521   VLoop vloop(lpt, false);
4522   if (!vloop.check_preconditions()) {
4523     return AutoVectorizeStatus::TriedAndFailed;
4524   }
4525 
4526   // Ensure the shared data is cleared before each use
4527   vshared.clear();
4528 
4529   const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4530   if (!vloop_analyzer.success()) {
4531     return AutoVectorizeStatus::TriedAndFailed;
4532   }
4533 
4534   SuperWord sw(vloop_analyzer);
4535   if (!sw.transform_loop()) {
4536     return AutoVectorizeStatus::TriedAndFailed;
4537   }
4538 
4539   return AutoVectorizeStatus::Success;
4540 }
4541 
4542 // Just before insert_pre_post_loops, we can multiversion the loop:
4543 //
4544 //              multiversion_if
4545 //               |       |
4546 //         fast_loop   slow_loop
4547 //
4548 // In the fast_loop we can make speculative assumptions, and put the
4549 // conditions into the multiversion_if. If the conditions hold at runtime,
4550 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4551 // instead which does not make any of the speculative assumptions.
4552 //
4553 // Note: we only multiversion the loop if the loop does not have any
4554 //       auto vectorization check Predicate. If we have that predicate,
4555 //       then we can simply add the speculative assumption checks to
4556 //       that Predicate. This means we do not need to duplicate the
4557 //       loop - we have a smaller graph and save compile time. Should
4558 //       the conditions ever fail, then we deopt / trap at the Predicate
4559 //       and recompile without that Predicate. At that point we will
4560 //       multiversion the loop, so that we can still have speculative
4561 //       runtime checks.
4562 //
4563 // We perform the multiversioning when the loop is still in its single
4564 // iteration form, even before we insert pre and post loops. This makes
4565 // the cloning much simpler. However, this means that both the fast
4566 // and the slow loop have to be optimized independently (adding pre
4567 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4568 // we may end up not needing any speculative assumptions in the fast_loop
4569 // and then rejecting the slow_loop by constant folding the multiversion_if.
4570 //
4571 // Therefore, we "delay" the optimization of the slow_loop until we add
4572 // at least one speculative assumption for the fast_loop. If we never
4573 // add such a speculative runtime check, the OpaqueMultiversioningNode
4574 // of the multiversion_if constant folds to true after loop opts, and the
4575 // multiversion_if folds away the "delayed" slow_loop. If we add any
4576 // speculative assumption, then we notify the OpaqueMultiversioningNode
4577 // with "notify_slow_loop_that_it_can_resume_optimizations".
4578 //
4579 // Note: new runtime checks can be added to the multiversion_if with
4580 //       PhaseIdealLoop::create_new_if_for_multiversion
4581 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4582   CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4583   LoopNode* outer_loop = cl->skip_strip_mined();
4584   Node* entry = outer_loop->in(LoopNode::EntryControl);
4585 
4586   // Check we have multiversioning enabled, and are not already multiversioned.
4587   if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4588 
4589   // Check that we do not have a parse-predicate where we can add the runtime checks
4590   // during auto-vectorization.
4591   const Predicates predicates(entry);
4592   const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4593   if (predicate_block->has_parse_predicate()) { return; }
4594 
4595   // Check node budget.
4596   uint estimate = lpt->est_loop_clone_sz(2);
4597   if (!may_require_nodes(estimate)) { return; }
4598 
4599   do_multiversioning(lpt, old_new);
4600 }
4601 
4602 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4603   for (uint i = 0; i < _data_nodes.size(); i++) {
4604     clone(_data_nodes[i], new_ctrl);
4605   }
4606 }
4607 
4608 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4609 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4610   Node* clone = node->clone();
4611   _phase->igvn().register_new_node_with_optimizer(clone);
4612   _orig_to_new.put(node, clone);
4613   _phase->set_ctrl(clone, new_ctrl);
4614   if (node->is_CastII()) {
4615     clone->set_req(0, new_ctrl);
4616   }
4617 }
4618 
4619 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4620 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4621 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4622   _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4623     for (uint i = 1; i < node->req(); i++) {
4624       Node** cloned_input = _orig_to_new.get(node->in(i));
4625       if (cloned_input != nullptr) {
4626         // Input was also cloned -> rewire clone to the cloned input.
4627         _phase->igvn().replace_input_of(clone, i, *cloned_input);
4628       }
4629     }
4630   });
4631 }
4632 
4633 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4634 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4635 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4636     const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4637     Node* new_ctrl) {
4638   for (uint i = 0; i < _data_nodes.size(); i++) {
4639     Node* data_node = _data_nodes[i];
4640     if (data_node->is_Opaque1()) {
4641       transform_opaque_node(transform_strategy, data_node);
4642     } else {
4643       clone(data_node, new_ctrl);
4644     }
4645   }
4646 }
4647 
4648 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4649   Node* transformed_node;
4650   if (node->is_OpaqueLoopInit()) {
4651     transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4652   } else {
4653     assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4654     transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4655   }
4656   // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4657   _orig_to_new.put(node, transformed_node);
4658 }