1 /*
   2  * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "ci/ciFlatArrayKlass.hpp"
  27 #include "ci/ciInlineKlass.hpp"
  28 #include "ci/ciInstanceKlass.hpp"
  29 #include "classfile/javaClasses.hpp"
  30 #include "classfile/systemDictionary.hpp"
  31 #include "classfile/vmIntrinsics.hpp"
  32 #include "compiler/compileLog.hpp"
  33 #include "gc/shared/barrierSet.hpp"
  34 #include "gc/shared/c2/barrierSetC2.hpp"
  35 #include "gc/shared/tlab_globals.hpp"
  36 #include "memory/allocation.inline.hpp"
  37 #include "memory/resourceArea.hpp"
  38 #include "oops/flatArrayKlass.hpp"
  39 #include "oops/objArrayKlass.hpp"
  40 #include "opto/addnode.hpp"
  41 #include "opto/arraycopynode.hpp"
  42 #include "opto/callnode.hpp"
  43 #include "opto/cfgnode.hpp"
  44 #include "opto/compile.hpp"
  45 #include "opto/connode.hpp"
  46 #include "opto/convertnode.hpp"
  47 #include "opto/inlinetypenode.hpp"
  48 #include "opto/loopnode.hpp"
  49 #include "opto/machnode.hpp"
  50 #include "opto/matcher.hpp"
  51 #include "opto/memnode.hpp"
  52 #include "opto/mempointer.hpp"
  53 #include "opto/mulnode.hpp"
  54 #include "opto/narrowptrnode.hpp"
  55 #include "opto/opcodes.hpp"
  56 #include "opto/phaseX.hpp"
  57 #include "opto/regalloc.hpp"
  58 #include "opto/regmask.hpp"
  59 #include "opto/rootnode.hpp"
  60 #include "opto/traceMergeStoresTag.hpp"
  61 #include "opto/vectornode.hpp"
  62 #include "runtime/arguments.hpp"
  63 #include "utilities/align.hpp"
  64 #include "utilities/copy.hpp"
  65 #include "utilities/globalDefinitions.hpp"
  66 #include "utilities/macros.hpp"
  67 #include "utilities/powerOfTwo.hpp"
  68 #include "utilities/vmError.hpp"
  69 
  70 // Portions of code courtesy of Clifford Click
  71 
  72 // Optimization - Graph Style
  73 
  74 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  75 
  76 //=============================================================================
  77 uint MemNode::size_of() const { return sizeof(*this); }
  78 
  79 const TypePtr *MemNode::adr_type() const {
  80   Node* adr = in(Address);
  81   if (adr == nullptr)  return nullptr; // node is dead
  82   const TypePtr* cross_check = nullptr;
  83   DEBUG_ONLY(cross_check = _adr_type);
  84   return calculate_adr_type(adr->bottom_type(), cross_check);
  85 }
  86 
  87 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
  88   if (adr != nullptr) {
  89     if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
  90       return true;
  91     }
  92   }
  93   return false;
  94 }
  95 
  96 #ifndef PRODUCT
  97 void MemNode::dump_spec(outputStream *st) const {
  98   if (in(Address) == nullptr)  return; // node is dead
  99 #ifndef ASSERT
 100   // fake the missing field
 101   const TypePtr* _adr_type = nullptr;
 102   if (in(Address) != nullptr)
 103     _adr_type = in(Address)->bottom_type()->isa_ptr();
 104 #endif
 105   dump_adr_type(_adr_type, st);
 106 
 107   Compile* C = Compile::current();
 108   if (C->alias_type(_adr_type)->is_volatile()) {
 109     st->print(" Volatile!");
 110   }
 111   if (_unaligned_access) {
 112     st->print(" unaligned");
 113   }
 114   if (_mismatched_access) {
 115     st->print(" mismatched");
 116   }
 117   if (_unsafe_access) {
 118     st->print(" unsafe");
 119   }
 120 }
 121 
 122 void MemNode::dump_adr_type(const TypePtr* adr_type, outputStream* st) {
 123   st->print(" @");
 124   if (adr_type == nullptr) {
 125     st->print("null");
 126   } else {
 127     adr_type->dump_on(st);
 128     Compile* C = Compile::current();
 129     Compile::AliasType* atp = nullptr;
 130     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
 131     if (atp == nullptr)
 132       st->print(", idx=?\?;");
 133     else if (atp->index() == Compile::AliasIdxBot)
 134       st->print(", idx=Bot;");
 135     else if (atp->index() == Compile::AliasIdxTop)
 136       st->print(", idx=Top;");
 137     else if (atp->index() == Compile::AliasIdxRaw)
 138       st->print(", idx=Raw;");
 139     else {
 140       ciField* field = atp->field();
 141       if (field) {
 142         st->print(", name=");
 143         field->print_name_on(st);
 144       }
 145       st->print(", idx=%d;", atp->index());
 146     }
 147   }
 148 }
 149 
 150 extern void print_alias_types();
 151 
 152 #endif
 153 
 154 // Find the memory output corresponding to the fall-through path of a call
 155 static Node* find_call_fallthrough_mem_output(CallNode* call) {
 156   ResourceMark rm;
 157   CallProjections* projs = call->extract_projections(false, false);
 158   Node* res = projs->fallthrough_memproj;
 159   assert(res != nullptr, "must have a fallthrough mem output");
 160   return res;
 161 }
 162 
 163 // Try to find a better memory input for a load from a strict final field
 164 static Node* try_optimize_strict_final_load_memory(PhaseGVN* phase, Node* adr, ProjNode*& base_local) {
 165   intptr_t offset = 0;
 166   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 167   if (base == nullptr) {
 168     return nullptr;
 169   }
 170 
 171   Node* base_uncasted = base->uncast();
 172   if (base_uncasted->is_Proj()) {
 173     Node* multi = base_uncasted->in(0);
 174     if (multi->is_top()) {
 175       // The pointer dies, make the memory die, too
 176       return multi;
 177     } else if (multi->is_Allocate()) {
 178       base_local = base_uncasted->as_Proj();
 179       return nullptr;
 180     } else if (multi->is_Call()) {
 181       if (!multi->is_CallJava() || multi->as_CallJava()->method() == nullptr || !multi->as_CallJava()->method()->return_value_is_larval()) {
 182         // The oop is returned from a call, the memory can be the fallthrough output of the call
 183         return find_call_fallthrough_mem_output(multi->as_Call());
 184       }
 185     } else if (multi->is_Start()) {
 186       // The oop is a parameter
 187       if (base_uncasted->as_Proj()->_con == TypeFunc::Parms && phase->C->method()->receiver_maybe_larval()) {
 188         // The receiver of a constructor is similar to the result of an AllocateNode
 189         base_local = base_uncasted->as_Proj();
 190         return nullptr;
 191       } else {
 192         // Use the start memory otherwise
 193         return multi->as_Start()->proj_out(TypeFunc::Memory);
 194       }
 195     }
 196   }
 197 
 198   return nullptr;
 199 }
 200 
 201 // Whether a call can modify a strict final field, given that the object is allocated inside the
 202 // current compilation unit, or is the first parameter when the compilation root is a constructor.
 203 // This is equivalent to asking whether 'call' is a constructor invocation and the class declaring
 204 // the target method is a subclass of the class declaring 'field'.
 205 static bool call_can_modify_local_object(ciField* field, CallNode* call) {
 206   if (!call->is_CallJava()) {
 207     return false;
 208   }
 209 
 210   ciMethod* target = call->as_CallJava()->method();
 211   if (target == nullptr) {
 212     return false;
 213   } else if (target->intrinsic_id() == vmIntrinsicID::_linkToSpecial) {
 214     // linkToSpecial can be used to call a constructor, used in the construction of objects in the
 215     // reflection API
 216     return true;
 217   } else if (!target->is_object_constructor()) {
 218     return false;
 219   }
 220 
 221   // If 'field' is declared in a class that is a subclass of the one declaring the constructor,
 222   // then the field is set inside the constructor, else the field must be set before the
 223   // constructor invocation. E.g. A field Super.x will be set during the execution of Sub::<init>,
 224   // while a field Sub.y must be set before Super::<init> is invoked.
 225   // We can try to be more heroic and decide if the receiver of the constructor invocation is the
 226   // object from which we are loading from. This, however, may be problematic as deciding if 2
 227   // nodes are definitely different may not be trivial, especially if the graph is not canonical.
 228   // As a result, it is made more conservative for now.
 229   assert(call->req() > TypeFunc::Parms, "constructor must have at least 1 argument");
 230   return target->holder()->is_subclass_of(field->holder());
 231 }
 232 
 233 Node* MemNode::optimize_simple_memory_chain(Node* mchain, const TypeOopPtr* t_oop, Node* load, PhaseGVN* phase) {
 234   assert(t_oop != nullptr, "sanity");
 235   bool is_known_instance = t_oop->is_known_instance_field();
 236   bool is_strict_final_load = false;
 237 
 238   // After macro expansion, an allocation may become a call, changing the memory input to the
 239   // memory output of that call would be illegal. As a result, disallow this transformation after
 240   // macro expansion.
 241   if (phase->is_IterGVN() && phase->C->allow_macro_nodes() && load != nullptr && load->is_Load() && !load->as_Load()->is_mismatched_access()) {
 242     is_strict_final_load = t_oop->is_ptr_to_strict_final_field();
 243 #ifdef ASSERT
 244     if ((t_oop->is_inlinetypeptr() && t_oop->inline_klass()->contains_field_offset(t_oop->offset())) || t_oop->is_ptr_to_boxed_value()) {
 245       assert(is_strict_final_load, "sanity check for basic cases");
 246     }
 247 #endif // ASSERT
 248   }
 249 
 250   if (!is_known_instance && !is_strict_final_load) {
 251     return mchain;
 252   }
 253 
 254   Node* result = mchain;
 255   ProjNode* base_local = nullptr;
 256 
 257   ciField* field = nullptr;
 258   if (is_strict_final_load) {
 259     field = phase->C->alias_type(t_oop)->field();
 260     assert(field != nullptr, "must point to a field");
 261 
 262     Node* adr = load->in(MemNode::Address);
 263     assert(phase->type(adr) == t_oop, "inconsistent type");
 264     Node* tmp = try_optimize_strict_final_load_memory(phase, adr, base_local);
 265     if (tmp != nullptr) {
 266       result = tmp;
 267     }
 268   }
 269 
 270   uint instance_id = t_oop->instance_id();
 271   Node* start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
 272   Node* prev = nullptr;
 273   while (prev != result) {
 274     prev = result;
 275     if (result == start_mem) {
 276       // start_mem is the earliest memory possible
 277       break;
 278     }
 279 
 280     // skip over a call which does not affect this memory slice
 281     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 282       Node* proj_in = result->in(0);
 283       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 284         // This is the allocation that creates the object from which we are loading from
 285         break;
 286       } else if (proj_in->is_Call()) {
 287         // ArrayCopyNodes processed here as well
 288         CallNode* call = proj_in->as_Call();
 289         if (!call->may_modify(t_oop, phase)) {
 290           result = call->in(TypeFunc::Memory);
 291         } else if (is_strict_final_load && base_local != nullptr && !call_can_modify_local_object(field, call)) {
 292           result = call->in(TypeFunc::Memory);
 293         }
 294       } else if (proj_in->Opcode() == Op_Tuple) {
 295         // The call will be folded, skip over it.
 296         break;
 297       } else if (proj_in->is_Initialize()) {
 298         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 299         // Stop if this is the initialization for the object instance which
 300         // contains this memory slice, otherwise skip over it.
 301         if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
 302           break;
 303         }
 304         if (is_known_instance) {
 305           result = proj_in->in(TypeFunc::Memory);
 306         } else if (is_strict_final_load) {
 307           Node* klass = alloc->in(AllocateNode::KlassNode);
 308           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 309           if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
 310             // Allocation of another type, must be another object
 311             result = proj_in->in(TypeFunc::Memory);
 312           } else if (base_local != nullptr && (base_local->is_Parm() || base_local->in(0) != alloc)) {
 313             // Allocation of another object
 314             result = proj_in->in(TypeFunc::Memory);
 315           }
 316         }
 317       } else if (proj_in->is_MemBar()) {
 318         ArrayCopyNode* ac = nullptr;
 319         if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
 320           break;
 321         }
 322         result = proj_in->in(TypeFunc::Memory);
 323       } else if (proj_in->is_LoadFlat() || proj_in->is_StoreFlat()) {
 324         if (is_strict_final_load) {
 325           // LoadFlat and StoreFlat cannot happen to strict final fields
 326           result = proj_in->in(TypeFunc::Memory);
 327         }
 328       } else if (proj_in->is_top()) {
 329         break; // dead code
 330       } else {
 331         assert(false, "unexpected projection of %s", proj_in->Name());
 332       }
 333     } else if (result->is_ClearArray()) {
 334       if (!is_known_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 335         // Can not bypass initialization of the instance
 336         // we are looking for.
 337         break;
 338       }
 339       // Otherwise skip it (the call updated 'result' value).
 340     } else if (result->is_MergeMem()) {
 341       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
 342     }
 343   }
 344   return result;
 345 }
 346 
 347 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 348   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 349   if (t_oop == nullptr)
 350     return mchain;  // don't try to optimize non-oop types
 351   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 352   bool is_instance = t_oop->is_known_instance_field();
 353   PhaseIterGVN *igvn = phase->is_IterGVN();
 354   if (is_instance && igvn != nullptr && result->is_Phi()) {
 355     PhiNode *mphi = result->as_Phi();
 356     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 357     const TypePtr *t = mphi->adr_type();
 358     bool do_split = false;
 359     // In the following cases, Load memory input can be further optimized based on
 360     // its precise address type
 361     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
 362       do_split = true;
 363     } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
 364       const TypeOopPtr* mem_t =
 365         t->is_oopptr()->cast_to_exactness(true)
 366         ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 367         ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
 368       if (t_oop->isa_aryptr()) {
 369         mem_t = mem_t->is_aryptr()
 370                      ->cast_to_stable(t_oop->is_aryptr()->is_stable())
 371                      ->cast_to_size(t_oop->is_aryptr()->size())
 372                      ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
 373                      ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
 374                      ->with_offset(t_oop->is_aryptr()->offset())
 375                      ->is_aryptr();
 376       }
 377       do_split = mem_t == t_oop;
 378     }
 379     if (do_split) {
 380       // clone the Phi with our address type
 381       result = mphi->split_out_instance(t_adr, igvn);
 382     } else {
 383       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 384     }
 385   }
 386   return result;
 387 }
 388 
 389 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 390   uint alias_idx = phase->C->get_alias_index(tp);
 391   Node *mem = mmem;
 392 #ifdef ASSERT
 393   {
 394     // Check that current type is consistent with the alias index used during graph construction
 395     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 396     bool consistent =  adr_check == nullptr || adr_check->empty() ||
 397                        phase->C->must_alias(adr_check, alias_idx );
 398     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 399     if( !consistent && adr_check != nullptr && !adr_check->empty() &&
 400         tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 401         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 402         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 403           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 404           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 405       // don't assert if it is dead code.
 406       consistent = true;
 407     }
 408     if( !consistent ) {
 409       st->print("alias_idx==%d, adr_check==", alias_idx);
 410       if( adr_check == nullptr ) {
 411         st->print("null");
 412       } else {
 413         adr_check->dump();
 414       }
 415       st->cr();
 416       print_alias_types();
 417       assert(consistent, "adr_check must match alias idx");
 418     }
 419   }
 420 #endif
 421   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 422   // means an array I have not precisely typed yet.  Do not do any
 423   // alias stuff with it any time soon.
 424   const TypeOopPtr *toop = tp->isa_oopptr();
 425   if (tp->base() != Type::AnyPtr &&
 426       !(toop &&
 427         toop->isa_instptr() &&
 428         toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
 429         toop->offset() == Type::OffsetBot)) {
 430     // IGVN _delay_transform may be set to true and if that is the case and mmem
 431     // is already a registered node then the validation inside transform will
 432     // complain.
 433     Node* m = mmem;
 434     PhaseIterGVN* igvn = phase->is_IterGVN();
 435     if (igvn == nullptr || !igvn->delay_transform()) {
 436       // compress paths and change unreachable cycles to TOP
 437       // If not, we can update the input infinitely along a MergeMem cycle
 438       // Equivalent code in PhiNode::Ideal
 439       m = phase->transform(mmem);
 440     }
 441     // If transformed to a MergeMem, get the desired slice
 442     // Otherwise the returned node represents memory for every slice
 443     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 444     // Update input if it is progress over what we have now
 445   }
 446   return mem;
 447 }
 448 
 449 //--------------------------Ideal_common---------------------------------------
 450 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 451 // Unhook non-raw memories from complete (macro-expanded) initializations.
 452 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 453   // If our control input is a dead region, kill all below the region
 454   Node *ctl = in(MemNode::Control);
 455   if (ctl && remove_dead_region(phase, can_reshape))
 456     return this;
 457   ctl = in(MemNode::Control);
 458   // Don't bother trying to transform a dead node
 459   if (ctl && ctl->is_top())  return NodeSentinel;
 460 
 461   PhaseIterGVN *igvn = phase->is_IterGVN();
 462   // Wait if control on the worklist.
 463   if (ctl && can_reshape && igvn != nullptr) {
 464     Node* bol = nullptr;
 465     Node* cmp = nullptr;
 466     if (ctl->in(0)->is_If()) {
 467       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 468       bol = ctl->in(0)->in(1);
 469       if (bol->is_Bool())
 470         cmp = ctl->in(0)->in(1)->in(1);
 471     }
 472     if (igvn->_worklist.member(ctl) ||
 473         (bol != nullptr && igvn->_worklist.member(bol)) ||
 474         (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
 475       // This control path may be dead.
 476       // Delay this memory node transformation until the control is processed.
 477       igvn->_worklist.push(this);
 478       return NodeSentinel; // caller will return null
 479     }
 480   }
 481   // Ignore if memory is dead, or self-loop
 482   Node *mem = in(MemNode::Memory);
 483   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
 484   assert(mem != this, "dead loop in MemNode::Ideal");
 485 
 486   if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
 487     // This memory slice may be dead.
 488     // Delay this mem node transformation until the memory is processed.
 489     igvn->_worklist.push(this);
 490     return NodeSentinel; // caller will return null
 491   }
 492 
 493   Node *address = in(MemNode::Address);
 494   const Type *t_adr = phase->type(address);
 495   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return null
 496 
 497   if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
 498     // Unsafe off-heap access with zero address. Remove access and other control users
 499     // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
 500     assert(ctl != nullptr, "unsafe accesses should be control dependent");
 501     for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
 502       Node* u = ctl->fast_out(i);
 503       if (u != ctl) {
 504         igvn->rehash_node_delayed(u);
 505         int nb = u->replace_edge(ctl, phase->C->top(), igvn);
 506         --i, imax -= nb;
 507       }
 508     }
 509     Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
 510     Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
 511     phase->C->root()->add_req(halt);
 512     return this;
 513   }
 514 
 515   if (can_reshape && igvn != nullptr &&
 516       (igvn->_worklist.member(address) ||
 517        (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
 518     // The address's base and type may change when the address is processed.
 519     // Delay this mem node transformation until the address is processed.
 520     igvn->_worklist.push(this);
 521     return NodeSentinel; // caller will return null
 522   }
 523 
 524   // Do NOT remove or optimize the next lines: ensure a new alias index
 525   // is allocated for an oop pointer type before Escape Analysis.
 526   // Note: C++ will not remove it since the call has side effect.
 527   if (t_adr->isa_oopptr()) {
 528     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 529   }
 530 
 531   Node* base = nullptr;
 532   if (address->is_AddP()) {
 533     base = address->in(AddPNode::Base);
 534   }
 535   if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 536       !t_adr->isa_rawptr()) {
 537     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 538     // Skip this node optimization if its address has TOP base.
 539     return NodeSentinel; // caller will return null
 540   }
 541 
 542   // Avoid independent memory operations
 543   Node* old_mem = mem;
 544 
 545   // The code which unhooks non-raw memories from complete (macro-expanded)
 546   // initializations was removed. After macro-expansion all stores caught
 547   // by Initialize node became raw stores and there is no information
 548   // which memory slices they modify. So it is unsafe to move any memory
 549   // operation above these stores. Also in most cases hooked non-raw memories
 550   // were already unhooked by using information from detect_ptr_independence()
 551   // and find_previous_store().
 552 
 553   if (mem->is_MergeMem()) {
 554     MergeMemNode* mmem = mem->as_MergeMem();
 555     const TypePtr *tp = t_adr->is_ptr();
 556 
 557     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 558   }
 559 
 560   if (mem != old_mem) {
 561     set_req_X(MemNode::Memory, mem, phase);
 562     if (phase->type(mem) == Type::TOP) return NodeSentinel;
 563     return this;
 564   }
 565 
 566   // let the subclass continue analyzing...
 567   return nullptr;
 568 }
 569 
 570 // Helper function for proving some simple control dominations.
 571 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 572 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 573 // is not a constant (dominated by the method's StartNode).
 574 // Used by MemNode::find_previous_store to prove that the
 575 // control input of a memory operation predates (dominates)
 576 // an allocation it wants to look past.
 577 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
 578 // dominate 'sub', 'DomResult::NotDominate' if not,
 579 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
 580 // dead code, but at the end of IGVN, we know the definite result
 581 // once the dead code is cleaned up.
 582 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub, PhaseGVN* phase) {
 583   if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
 584     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 585   }
 586 
 587   // Check 'dom'. Skip Proj and CatchProj nodes.
 588   dom = dom->find_exact_control(dom);
 589   if (dom == nullptr || dom->is_top()) {
 590     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 591   }
 592 
 593   if (dom == sub) {
 594     // For the case when, for example, 'sub' is Initialize and the original
 595     // 'dom' is Proj node of the 'sub'.
 596     return DomResult::NotDominate;
 597   }
 598 
 599   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
 600     return DomResult::Dominate;
 601   }
 602 
 603   // 'dom' dominates 'sub' if its control edge and control edges
 604   // of all its inputs dominate or equal to sub's control edge.
 605 
 606   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 607   // Or Region for the check in LoadNode::Ideal();
 608   // 'sub' should have sub->in(0) != nullptr.
 609   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 610          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 611 
 612   // Get control edge of 'sub'.
 613   Node* orig_sub = sub;
 614   sub = sub->find_exact_control(sub->in(0));
 615   if (sub == nullptr || sub->is_top()) {
 616     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 617   }
 618 
 619   assert(sub->is_CFG(), "expecting control");
 620 
 621   if (sub == dom) {
 622     return DomResult::Dominate;
 623   }
 624 
 625   if (sub->is_Start() || sub->is_Root()) {
 626     return DomResult::NotDominate;
 627   }
 628 
 629   {
 630     // Check all control edges of 'dom'.
 631 
 632     ResourceMark rm;
 633     Node_List nlist;
 634     Unique_Node_List dom_list;
 635 
 636     dom_list.push(dom);
 637     bool only_dominating_controls = false;
 638 
 639     for (uint next = 0; next < dom_list.size(); next++) {
 640       Node* n = dom_list.at(next);
 641       if (n == orig_sub) {
 642         return DomResult::NotDominate; // One of dom's inputs dominated by sub.
 643       }
 644       if (!n->is_CFG() && n->pinned()) {
 645         // Check only own control edge for pinned non-control nodes.
 646         n = n->find_exact_control(n->in(0));
 647         if (n == nullptr || n->is_top()) {
 648           return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 649         }
 650         assert(n->is_CFG(), "expecting control");
 651         dom_list.push(n);
 652       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 653         only_dominating_controls = true;
 654       } else if (n->is_CFG()) {
 655         DomResult dom_result = n->dominates(sub, nlist);
 656         if (dom_result == DomResult::Dominate) {
 657           only_dominating_controls = true;
 658         } else {
 659           return dom_result;
 660         }
 661       } else {
 662         if (n->Value(phase) == Type::TOP) {
 663           return DomResult::EncounteredDeadCode;
 664         }
 665         // First, own control edge.
 666         Node* m = n->find_exact_control(n->in(0));
 667         if (m != nullptr) {
 668           if (m->is_top()) {
 669             return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 670           }
 671           dom_list.push(m);
 672         }
 673         // Now, the rest of edges.
 674         uint cnt = n->req();
 675         for (uint i = 1; i < cnt; i++) {
 676           m = n->find_exact_control(n->in(i));
 677           if (m == nullptr || m->is_top()) {
 678             continue;
 679           }
 680           dom_list.push(m);
 681         }
 682       }
 683     }
 684     return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
 685   }
 686 }
 687 
 688 //---------------------detect_ptr_independence---------------------------------
 689 // Used by MemNode::find_previous_store to prove that two base
 690 // pointers are never equal.
 691 // The pointers are accompanied by their associated allocations,
 692 // if any, which have been previously discovered by the caller.
 693 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 694                                       Node* p2, AllocateNode* a2,
 695                                       PhaseGVN* phase) {
 696   // Trivial case: Non-overlapping values. Be careful, we can cast a raw pointer to an oop (e.g. in
 697   // the allocation pattern) so joining the types only works if both are oops. join may also give
 698   // an incorrect result when both pointers are nullable and the result is supposed to be
 699   // TypePtr::NULL_PTR, so we exclude that case.
 700   const Type* p1_type = p1->bottom_type();
 701   const Type* p2_type = p2->bottom_type();
 702   if (p1_type != p2_type && p1_type->isa_oopptr() && p2_type->isa_oopptr() &&
 703       (!p1_type->maybe_null() || !p2_type->maybe_null()) &&
 704       p1_type->join(p2_type)->empty()) {
 705     return true;
 706   }
 707 
 708   // Attempt to prove that these two pointers cannot be aliased.
 709   // They may both manifestly be allocations, and they should differ.
 710   // Or, if they are not both allocations, they can be distinct constants.
 711   // Otherwise, one is an allocation and the other a pre-existing value.
 712   if (a1 == nullptr && a2 == nullptr) {           // neither an allocation
 713     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 714   } else if (a1 != nullptr && a2 != nullptr) {    // both allocations
 715     return (a1 != a2);
 716   } else if (a1 != nullptr) {                  // one allocation a1
 717     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 718     return all_controls_dominate(p2->uncast(), a1, phase);
 719   } else { //(a2 != null)                   // one allocation a2
 720     return all_controls_dominate(p1->uncast(), a2, phase);
 721   }
 722   return false;
 723 }
 724 
 725 // Find an arraycopy ac that produces the memory state represented by parameter mem.
 726 // Return ac if
 727 // (a) can_see_stored_value=true  and ac must have set the value for this load or if
 728 // (b) can_see_stored_value=false and ac could have set the value for this load or if
 729 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
 730 // In case (c) change the parameter mem to the memory input of ac to skip it
 731 // when searching stored value.
 732 // Otherwise return null.
 733 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 734   ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
 735   if (ac != nullptr) {
 736     Node* ld_addp = in(MemNode::Address);
 737     Node* src = ac->in(ArrayCopyNode::Src);
 738     const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
 739 
 740     // This is a load from a cloned array. The corresponding arraycopy ac must
 741     // have set the value for the load and we can return ac but only if the load
 742     // is known to be within bounds. This is checked below.
 743     if (ary_t != nullptr && ld_addp->is_AddP()) {
 744       Node* ld_offs = ld_addp->in(AddPNode::Offset);
 745       BasicType ary_elem = ary_t->elem()->array_element_basic_type();
 746       jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 747       jlong elemsize = type2aelembytes(ary_elem);
 748 
 749       const TypeX*   ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 750       const TypeInt* sizetype  = ary_t->size();
 751 
 752       if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
 753         // The load is known to be within bounds. It receives its value from ac.
 754         return ac;
 755       }
 756       // The load is known to be out-of-bounds.
 757     }
 758     // The load could be out-of-bounds. It must not be hoisted but must remain
 759     // dependent on the runtime range check. This is achieved by returning null.
 760   } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
 761     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 762 
 763     if (ac->is_arraycopy_validated() ||
 764         ac->is_copyof_validated() ||
 765         ac->is_copyofrange_validated()) {
 766       Node* ld_addp = in(MemNode::Address);
 767       if (ld_addp->is_AddP()) {
 768         Node* ld_base = ld_addp->in(AddPNode::Address);
 769         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 770 
 771         Node* dest = ac->in(ArrayCopyNode::Dest);
 772 
 773         if (dest == ld_base) {
 774           const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 775           assert(!ld_offs_t->empty(), "dead reference should be checked already");
 776           // Take into account vector or unsafe access size
 777           jlong ld_size_in_bytes = (jlong)memory_size();
 778           jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
 779           offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
 780           if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
 781             return ac;
 782           }
 783           if (!can_see_stored_value) {
 784             mem = ac->in(TypeFunc::Memory);
 785             return ac;
 786           }
 787         }
 788       }
 789     }
 790   }
 791   return nullptr;
 792 }
 793 
 794 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
 795   if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 796                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 797     if (ld_alloc != nullptr) {
 798       // Check if there is an array copy for a clone
 799       Node* mb = mem->in(0);
 800       ArrayCopyNode* ac = nullptr;
 801       if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
 802           mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
 803         ac = mb->in(0)->in(0)->as_ArrayCopy();
 804       } else {
 805         // Step over GC barrier when ReduceInitialCardMarks is disabled
 806         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 807         Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
 808 
 809         if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
 810           ac = control_proj_ac->in(0)->as_ArrayCopy();
 811         }
 812       }
 813 
 814       if (ac != nullptr && ac->is_clonebasic()) {
 815         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
 816         if (alloc != nullptr && alloc == ld_alloc) {
 817           return ac;
 818         }
 819       }
 820     }
 821   }
 822   return nullptr;
 823 }
 824 
 825 // The logic for reordering loads and stores uses four steps:
 826 // (a) Walk carefully past stores and initializations which we
 827 //     can prove are independent of this load.
 828 // (b) Observe that the next memory state makes an exact match
 829 //     with self (load or store), and locate the relevant store.
 830 // (c) Ensure that, if we were to wire self directly to the store,
 831 //     the optimizer would fold it up somehow.
 832 // (d) Do the rewiring, and return, depending on some other part of
 833 //     the optimizer to fold up the load.
 834 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 835 // specific to loads and stores, so they are handled by the callers.
 836 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 837 //
 838 Node* MemNode::find_previous_store(PhaseGVN* phase) {
 839   AccessAnalyzer analyzer(phase, this);
 840 
 841   Node* mem = in(MemNode::Memory); // start searching here...
 842   int cnt = 50;                    // Cycle limiter
 843   for (;; cnt--) {
 844     // While we can dance past unrelated stores...
 845     if (phase->type(mem) == Type::TOP) {
 846       // Encounter a dead node
 847       return phase->C->top();
 848     } else if (cnt <= 0) {
 849       // Caught in cycle or a complicated dance?
 850       return nullptr;
 851     } else if (mem->is_Phi()) {
 852       return nullptr;
 853     }
 854 
 855     AccessAnalyzer::AccessIndependence independence = analyzer.detect_access_independence(mem);
 856     if (independence.independent) {
 857       // (a) advance through the independent store
 858       mem = independence.mem;
 859       assert(mem != nullptr, "must not be nullptr");
 860     } else {
 861       // (b) found the store that this access observes if this is not null
 862       // Otherwise, give up if it is null
 863       return independence.mem;
 864     }
 865   }
 866 
 867   return nullptr;              // bail out
 868 }
 869 
 870 //----------------------calculate_adr_type-------------------------------------
 871 // Helper function.  Notices when the given type of address hits top or bottom.
 872 // Also, asserts a cross-check of the type against the expected address type.
 873 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 874   if (t == Type::TOP)  return nullptr; // does not touch memory any more?
 875   #ifdef ASSERT
 876   if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump())  cross_check = nullptr;
 877   #endif
 878   const TypePtr* tp = t->isa_ptr();
 879   if (tp == nullptr) {
 880     assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 881     return TypePtr::BOTTOM;           // touches lots of memory
 882   } else {
 883     #ifdef ASSERT
 884     // %%%% [phh] We don't check the alias index if cross_check is
 885     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 886     if (cross_check != nullptr &&
 887         cross_check != TypePtr::BOTTOM &&
 888         cross_check != TypeRawPtr::BOTTOM) {
 889       // Recheck the alias index, to see if it has changed (due to a bug).
 890       Compile* C = Compile::current();
 891       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 892              "must stay in the original alias category");
 893       // The type of the address must be contained in the adr_type,
 894       // disregarding "null"-ness.
 895       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 896       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 897       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 898              "real address must not escape from expected memory type");
 899     }
 900     #endif
 901     return tp;
 902   }
 903 }
 904 
 905 uint8_t MemNode::barrier_data(const Node* n) {
 906   if (n->is_LoadStore()) {
 907     return n->as_LoadStore()->barrier_data();
 908   } else if (n->is_Mem()) {
 909     return n->as_Mem()->barrier_data();
 910   }
 911   return 0;
 912 }
 913 
 914 AccessAnalyzer::AccessAnalyzer(PhaseGVN* phase, MemNode* n)
 915   : _phase(phase), _n(n), _memory_size(n->memory_size()), _alias_idx(-1) {
 916   Node* adr  = _n->in(MemNode::Address);
 917   _offset    = 0;
 918   _base      = AddPNode::Ideal_base_and_offset(adr, _phase, _offset);
 919   _maybe_raw = MemNode::check_if_adr_maybe_raw(adr);
 920   _alloc     = AllocateNode::Ideal_allocation(_base);
 921   _adr_type = _n->adr_type();
 922 
 923   if (_adr_type != nullptr && _adr_type->base() != TypePtr::AnyPtr) {
 924     // Avoid the cases that will upset Compile::get_alias_index
 925     _alias_idx = _phase->C->get_alias_index(_adr_type);
 926     assert(_alias_idx != Compile::AliasIdxTop, "must not be a dead node");
 927     assert(_alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
 928   }
 929 }
 930 
 931 // Decide whether the memory accessed by '_n' and 'other' may overlap. This function may be used
 932 // when we want to walk the memory graph to fold a load, or when we want to hoist a load above a
 933 // loop when there are no stores that may overlap with the load inside the loop.
 934 AccessAnalyzer::AccessIndependence AccessAnalyzer::detect_access_independence(Node* other) const {
 935   assert(_phase->type(other) == Type::MEMORY, "must be a memory node %s", other->Name());
 936   assert(!other->is_Phi(), "caller must handle Phi");
 937 
 938   if (_adr_type == nullptr) {
 939     // This means the access is dead
 940     return {false, _phase->C->top()};
 941   } else if (_adr_type->base() == TypePtr::AnyPtr) {
 942     // An example for this case is an access into the memory address 0 performed using Unsafe
 943     assert(_adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
 944     return {false, nullptr};
 945   }
 946 
 947   if (_offset == Type::OffsetBot) {
 948     // cannot unalias unless there are precise offsets
 949     return {false, nullptr};
 950   }
 951 
 952   const TypeOopPtr* adr_oop_type = _adr_type->isa_oopptr();
 953   Node* prev = other;
 954   if (other->is_Store()) {
 955     Node* st_adr = other->in(MemNode::Address);
 956     intptr_t st_offset = 0;
 957     Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, _phase, st_offset);
 958     if (st_base == nullptr) {
 959       // inscrutable pointer
 960       return {false, nullptr};
 961     }
 962 
 963     // If the bases are the same and the offsets are the same, it seems that this is the exact
 964     // store we are looking for, the caller will check if the type of the store matches using
 965     // MemNode::can_see_stored_value
 966     if (st_base == _base && st_offset == _offset) {
 967       return {false, other};
 968     }
 969 
 970     // If it is provable that the memory accessed by 'other' does not overlap the memory accessed
 971     // by '_n', we may walk past 'other'.
 972     // For raw accesses, 2 accesses are independent if they have the same base and the offsets
 973     // say that they do not overlap.
 974     // For heap accesses, 2 accesses are independent if either the bases are provably different
 975     // at runtime or the offsets say that the accesses do not overlap.
 976     if ((_maybe_raw || MemNode::check_if_adr_maybe_raw(st_adr)) && st_base != _base) {
 977       // Raw accesses can only be provably independent if they have the same base
 978       return {false, nullptr};
 979     }
 980 
 981     // If the offsets say that the accesses do not overlap, then it is provable that 'other' and
 982     // '_n' do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
 983     // Object+12, no matter what the bases are.
 984     if (st_offset != _offset && st_offset != Type::OffsetBot) {
 985       const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
 986       assert(other->as_Store()->memory_size() <= MAX_STORE, "");
 987       if (st_offset >= _offset + _memory_size ||
 988           st_offset <= _offset - MAX_STORE ||
 989           st_offset <= _offset - other->as_Store()->memory_size()) {
 990         // Success:  The offsets are provably independent.
 991         // (You may ask, why not just test st_offset != offset and be done?
 992         // The answer is that stores of different sizes can co-exist
 993         // in the same sequence of RawMem effects.  We sometimes initialize
 994         // a whole 'tile' of array elements with a single jint or jlong.)
 995         return {true, other->in(MemNode::Memory)};
 996       }
 997     }
 998 
 999     // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
1000     if (st_base == _base) {
1001       return {false, nullptr};
1002     }
1003 
1004     // Try to prove that 2 different base nodes at compile time are different values at runtime
1005     bool known_independent = false;
1006     if (MemNode::detect_ptr_independence(_base, _alloc, st_base, AllocateNode::Ideal_allocation(st_base), _phase)) {
1007       known_independent = true;
1008     }
1009 
1010     if (known_independent) {
1011       return {true, other->in(MemNode::Memory)};
1012     }
1013   } else if (other->is_Proj() && other->in(0)->is_Initialize()) {
1014     InitializeNode* st_init = other->in(0)->as_Initialize();
1015     AllocateNode* st_alloc = st_init->allocation();
1016     if (st_alloc == nullptr) {
1017       // Something degenerated
1018       return {false, nullptr};
1019     }
1020     bool known_identical = false;
1021     bool known_independent = false;
1022     if (_alloc == st_alloc) {
1023       known_identical = true;
1024     } else if (_alloc != nullptr) {
1025       known_independent = true;
1026     } else if (MemNode::all_controls_dominate(_base->uncast(), st_alloc, _phase)) {
1027       known_independent = true;
1028     }
1029 
1030     if (known_independent) {
1031       // The bases are provably independent: Either they are
1032       // manifestly distinct allocations, or else the control
1033       // of _base dominates the store's allocation.
1034       if (_alias_idx == Compile::AliasIdxRaw) {
1035         other = st_alloc->in(TypeFunc::Memory);
1036       } else {
1037         other = st_init->memory(_alias_idx);
1038       }
1039       return {true, other};
1040     }
1041 
1042     // If we are not looking at a store initializing the same
1043     // allocation we are loading from, we lose.
1044     if (known_identical) {
1045       // From caller, can_see_stored_value will consult find_captured_store.
1046       return {false, other};
1047     }
1048 
1049   } else if (_n->find_previous_arraycopy(_phase, _alloc, other, false) != nullptr) {
1050     // Find an arraycopy that may or may not affect the MemNode
1051     return {prev != other, other};
1052   } else if (other->is_MergeMem()) {
1053     return {true, other->as_MergeMem()->memory_at(_alias_idx)};
1054   } else if (adr_oop_type != nullptr && adr_oop_type->is_known_instance_field()) {
1055     // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
1056     if (other->is_Proj() && other->in(0)->is_Call()) {
1057       // ArrayCopyNodes processed here as well.
1058       CallNode* call = other->in(0)->as_Call();
1059       if (!call->may_modify(adr_oop_type, _phase)) {
1060         return {true, call->in(TypeFunc::Memory)};
1061       }
1062     } else if (other->is_Proj() && other->in(0)->is_MemBar()) {
1063       ArrayCopyNode* ac = nullptr;
1064       if (!ArrayCopyNode::may_modify(adr_oop_type, other->in(0)->as_MemBar(), _phase, ac)) {
1065         return {true, other->in(0)->in(TypeFunc::Memory)};
1066       }
1067     } else if (other->is_ClearArray()) {
1068       if (ClearArrayNode::step_through(&other, (uint)adr_oop_type->instance_id(), _phase)) {
1069         // (the call updated 'other' value)
1070         return {true, other};
1071       } else {
1072         // Can not bypass initialization of the instance
1073         // we are looking for.
1074         return {false, other};
1075       }
1076     }
1077   }
1078 
1079   return {false, nullptr};
1080 }
1081 
1082 //=============================================================================
1083 // Should LoadNode::Ideal() attempt to remove control edges?
1084 bool LoadNode::can_remove_control() const {
1085   return !has_pinned_control_dependency();
1086 }
1087 uint LoadNode::size_of() const { return sizeof(*this); }
1088 bool LoadNode::cmp(const Node &n) const {
1089   LoadNode& load = (LoadNode &)n;
1090   return Type::equals(_type, load._type) &&
1091          _control_dependency == load._control_dependency &&
1092          _mo == load._mo;
1093 }
1094 const Type *LoadNode::bottom_type() const { return _type; }
1095 uint LoadNode::ideal_reg() const {
1096   return _type->ideal_reg();
1097 }
1098 
1099 #ifndef PRODUCT
1100 void LoadNode::dump_spec(outputStream *st) const {
1101   MemNode::dump_spec(st);
1102   if( !Verbose && !WizardMode ) {
1103     // standard dump does this in Verbose and WizardMode
1104     st->print(" #"); _type->dump_on(st);
1105   }
1106   if (in(0) != nullptr && !depends_only_on_test()) {
1107     st->print(" (does not depend only on test, ");
1108     if (control_dependency() == UnknownControl) {
1109       st->print("unknown control");
1110     } else if (control_dependency() == Pinned) {
1111       st->print("pinned");
1112     } else if (adr_type() == TypeRawPtr::BOTTOM) {
1113       st->print("raw access");
1114     } else {
1115       st->print("unknown reason");
1116     }
1117     st->print(")");
1118   }
1119 }
1120 #endif
1121 
1122 #ifdef ASSERT
1123 //----------------------------is_immutable_value-------------------------------
1124 // Helper function to allow a raw load without control edge for some cases
1125 bool LoadNode::is_immutable_value(Node* adr) {
1126   if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1127       adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
1128 
1129     jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1130     int offsets[] = {
1131       in_bytes(JavaThread::osthread_offset()),
1132       in_bytes(JavaThread::threadObj_offset()),
1133       in_bytes(JavaThread::vthread_offset()),
1134       in_bytes(JavaThread::scopedValueCache_offset()),
1135     };
1136 
1137     for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1138       if (offset == offsets[i]) {
1139         return true;
1140       }
1141     }
1142   }
1143 
1144   return false;
1145 }
1146 #endif
1147 
1148 //----------------------------LoadNode::make-----------------------------------
1149 // Polymorphic factory method:
1150 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1151                      ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1152   Compile* C = gvn.C;
1153   assert(adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr(), true) == C->get_alias_index(adr_type, true), "adr and adr_type must agree");
1154 
1155   // sanity check the alias category against the created node type
1156   assert(!(adr_type->isa_oopptr() &&
1157            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1158          "use LoadKlassNode instead");
1159   assert(!(adr_type->isa_aryptr() &&
1160            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1161          "use LoadRangeNode instead");
1162   // Check control edge of raw loads
1163   assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1164           // oop will be recorded in oop map if load crosses safepoint
1165           rt->isa_oopptr() || is_immutable_value(adr),
1166           "raw memory operations should have control edge");
1167   LoadNode* load = nullptr;
1168   switch (bt) {
1169   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1170   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1171   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1172   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1173   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1174   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1175   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
1176   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency, require_atomic_access); break;
1177   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
1178   case T_ARRAY:
1179   case T_OBJECT:
1180   case T_NARROWOOP:
1181 #ifdef _LP64
1182     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1183       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1184     } else
1185 #endif
1186     {
1187       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1188       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1189     }
1190     break;
1191   default:
1192     guarantee(false, "unexpected basic type %s", type2name(bt));
1193     break;
1194   }
1195   assert(load != nullptr, "LoadNode should have been created");
1196   if (unaligned) {
1197     load->set_unaligned_access();
1198   }
1199   if (mismatched) {
1200     load->set_mismatched_access();
1201   }
1202   if (unsafe) {
1203     load->set_unsafe_access();
1204   }
1205   load->set_barrier_data(barrier_data);
1206   if (load->Opcode() == Op_LoadN) {
1207     Node* ld = gvn.transform(load);
1208     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1209   }
1210 
1211   return load;
1212 }
1213 
1214 //------------------------------hash-------------------------------------------
1215 uint LoadNode::hash() const {
1216   // unroll addition of interesting fields
1217   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1218 }
1219 
1220 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1221   if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1222     bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1223     bool is_stable_ary = FoldStableValues &&
1224                          (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1225                          tp->isa_aryptr()->is_stable();
1226 
1227     return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1228   }
1229 
1230   return false;
1231 }
1232 
1233 // Is the value loaded previously stored by an arraycopy? If so return
1234 // a load node that reads from the source array so we may be able to
1235 // optimize out the ArrayCopy node later.
1236 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1237   Node* ld_adr = in(MemNode::Address);
1238   intptr_t ld_off = 0;
1239   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1240   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1241   if (ac != nullptr) {
1242     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1243 
1244     Node* mem = ac->in(TypeFunc::Memory);
1245     Node* ctl = ac->in(0);
1246     Node* src = ac->in(ArrayCopyNode::Src);
1247 
1248     if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1249       return nullptr;
1250     }
1251 
1252     // load depends on the tests that validate the arraycopy
1253     LoadNode* ld = clone_pinned();
1254     Node* addp = in(MemNode::Address)->clone();
1255     if (ac->as_ArrayCopy()->is_clonebasic()) {
1256       assert(ld_alloc != nullptr, "need an alloc");
1257       assert(addp->is_AddP(), "address must be addp");
1258       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1259       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1260       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1261       addp->set_req(AddPNode::Base, src);
1262       addp->set_req(AddPNode::Address, src);
1263     } else {
1264       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1265              ac->as_ArrayCopy()->is_copyof_validated() ||
1266              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1267       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1268       addp->set_req(AddPNode::Base, src);
1269       addp->set_req(AddPNode::Address, src);
1270 
1271       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1272       BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1273       if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1274 
1275       uint shift  = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1276 
1277       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1278 #ifdef _LP64
1279       diff = phase->transform(new ConvI2LNode(diff));
1280 #endif
1281       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1282 
1283       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1284       addp->set_req(AddPNode::Offset, offset);
1285     }
1286     addp = phase->transform(addp);
1287 #ifdef ASSERT
1288     const TypePtr* adr_type = phase->type(addp)->is_ptr();
1289     ld->_adr_type = adr_type;
1290 #endif
1291     ld->set_req(MemNode::Address, addp);
1292     ld->set_req(0, ctl);
1293     ld->set_req(MemNode::Memory, mem);
1294     return ld;
1295   }
1296   return nullptr;
1297 }
1298 
1299 static Node* see_through_inline_type(PhaseValues* phase, const LoadNode* load, Node* base, int offset) {
1300   if (load->is_mismatched_access() || base == nullptr) {
1301     return nullptr;
1302   }
1303 
1304   InlineTypeNode* vt = base->isa_InlineType();
1305   if (vt == nullptr || offset < vt->type()->inline_klass()->payload_offset()) {
1306     return nullptr;
1307   }
1308 
1309   Node* value = vt->field_value_by_offset(offset, true);
1310   assert(value != nullptr, "must see some value");
1311   return value;
1312 }
1313 
1314 // This routine exists to make sure this set of tests is done the same
1315 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
1316 // will change the graph shape in a way which makes memory alive twice at the
1317 // same time (uses the Oracle model of aliasing), then some
1318 // LoadXNode::Identity will fold things back to the equivalence-class model
1319 // of aliasing.
1320 // This method may find an unencoded node instead of the corresponding encoded one.
1321 Node* LoadNode::can_see_stored_value_through_membars(Node* st, PhaseValues* phase) const {
1322   Node* ld_adr = in(MemNode::Address);
1323   intptr_t ld_off = 0;
1324   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1325   // Try to see through an InlineTypeNode
1326   Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1327   if (value != nullptr) {
1328     return value;
1329   }
1330 
1331   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1332   Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1333 
1334   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1335     uint alias_idx = atp->index();
1336     Node* result = nullptr;
1337     Node* current = st;
1338     // Skip through chains of MemBarNodes checking the MergeMems for new states for the slice of
1339     // this load. Stop once any other kind of node is encountered.
1340     //
1341     // In principle, folding a load is moving it up until it meets a matching store.
1342     //
1343     // store(ptr, v);          store(ptr, v);          store(ptr, v);
1344     // membar1;          ->    membar1;          ->    load(ptr);
1345     // membar2;                load(ptr);              membar1;
1346     // load(ptr);              membar2;                membar2;
1347     //
1348     // So, we can decide which kinds of barriers we can walk past. It is not safe to step over
1349     // MemBarCPUOrder, even if the memory is not rewritable, because alias info above them may be
1350     // inaccurate (e.g., due to mixed/mismatched unsafe accesses).
1351     bool is_final_mem = !atp->is_rewritable();
1352     while (current->is_Proj()) {
1353       int opc = current->in(0)->Opcode();
1354       if ((is_final_mem && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock || opc == Op_LoadFence)) ||
1355           opc == Op_MemBarRelease ||
1356           opc == Op_StoreFence ||
1357           opc == Op_MemBarReleaseLock ||
1358           opc == Op_MemBarStoreStore ||
1359           opc == Op_StoreStoreFence) {
1360         Node* mem = current->in(0)->in(TypeFunc::Memory);
1361         if (mem->is_MergeMem()) {
1362           MergeMemNode* merge = mem->as_MergeMem();
1363           Node* new_st = merge->memory_at(alias_idx);
1364           if (new_st == merge->base_memory()) {
1365             // Keep searching
1366             current = new_st;
1367             continue;
1368           }
1369           // Save the new memory state for the slice and fall through
1370           // to exit.
1371           result = new_st;
1372         }
1373       }
1374       break;
1375     }
1376     if (result != nullptr) {
1377       st = result;
1378     }
1379   }
1380 
1381   return can_see_stored_value(st, phase);
1382 }
1383 
1384 // If st is a store to the same location as this, return the stored value
1385 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1386   Node* ld_adr = in(MemNode::Address);
1387   intptr_t ld_off = 0;
1388   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1389   Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1390   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1391 
1392   // Loop around twice in the case Load -> Initialize -> Store.
1393   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1394   for (int trip = 0; trip <= 1; trip++) {
1395 
1396     if (st->is_Store()) {
1397       Node* st_adr = st->in(MemNode::Address);
1398       if (st_adr != ld_adr) {
1399         // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1400         intptr_t st_off = 0;
1401         Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1402         if (ld_base == nullptr)                                return nullptr;
1403         if (st_base == nullptr)                                return nullptr;
1404         if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1405         if (ld_off != st_off)                                  return nullptr;
1406         if (ld_off == Type::OffsetBot)                         return nullptr;
1407         // Same base, same offset.
1408         // Possible improvement for arrays: check index value instead of absolute offset.
1409 
1410         // At this point we have proven something like this setup:
1411         //   B = << base >>
1412         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1413         //   S = StoreQ(AddP(             B , #Off), V)
1414         // (Actually, we haven't yet proven the Q's are the same.)
1415         // In other words, we are loading from a casted version of
1416         // the same pointer-and-offset that we stored to.
1417         // Casted version may carry a dependency and it is respected.
1418         // Thus, we are able to replace L by V.
1419       }
1420       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1421       if (store_Opcode() != st->Opcode()) {
1422         return nullptr;
1423       }
1424       // LoadVector/StoreVector needs additional check to ensure the types match.
1425       if (st->is_StoreVector()) {
1426         if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1427           // Some kind of masked access or gather/scatter
1428           return nullptr;
1429         }
1430 
1431         const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1432         const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1433         if (in_vt != out_vt) {
1434           return nullptr;
1435         }
1436       }
1437       return st->in(MemNode::ValueIn);
1438     }
1439 
1440     // A load from a freshly-created object always returns zero.
1441     // (This can happen after LoadNode::Ideal resets the load's memory input
1442     // to find_captured_store, which returned InitializeNode::zero_memory.)
1443     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1444         (st->in(0) == ld_alloc) &&
1445         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1446       // return a zero value for the load's basic type
1447       // (This is one of the few places where a generic PhaseTransform
1448       // can create new nodes.  Think of it as lazily manifesting
1449       // virtually pre-existing constants.)
1450       Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1451       if (init_value != nullptr) {
1452         const TypeAryPtr* ld_adr_type = phase->type(ld_adr)->isa_aryptr();
1453         if (ld_adr_type == nullptr) {
1454           return nullptr;
1455         }
1456 
1457         // We know that this is not a flat array, the load should return the whole oop
1458         if (ld_adr_type->is_not_flat()) {
1459           return init_value;
1460         }
1461 
1462         // If this is a flat array, try to see through init_value
1463         if (init_value->is_EncodeP()) {
1464           init_value = init_value->in(1);
1465         }
1466         if (!init_value->is_InlineType() || ld_adr_type->field_offset() == Type::Offset::bottom) {
1467           return nullptr;
1468         }
1469 
1470         ciInlineKlass* vk = phase->type(init_value)->inline_klass();
1471         int field_offset_in_payload = ld_adr_type->field_offset().get();
1472         if (field_offset_in_payload == vk->null_marker_offset_in_payload()) {
1473           return init_value->as_InlineType()->get_null_marker();
1474         } else {
1475           return init_value->as_InlineType()->field_value_by_offset(field_offset_in_payload + vk->payload_offset(), true);
1476         }
1477       }
1478       assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1479       if (value_basic_type() != T_VOID) {
1480         if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1481           // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1482           // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1483           // by the ArrayCopyNode.
1484           return phase->zerocon(value_basic_type());
1485         }
1486       } else {
1487         // TODO: materialize all-zero vector constant
1488         assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1489       }
1490     }
1491 
1492     // A load from an initialization barrier can match a captured store.
1493     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1494       InitializeNode* init = st->in(0)->as_Initialize();
1495       AllocateNode* alloc = init->allocation();
1496       if ((alloc != nullptr) && (alloc == ld_alloc)) {
1497         // examine a captured store value
1498         st = init->find_captured_store(ld_off, memory_size(), phase);
1499         if (st != nullptr) {
1500           continue;             // take one more trip around
1501         }
1502       }
1503     }
1504 
1505     // Load boxed value from result of valueOf() call is input parameter.
1506     if (this->is_Load() && ld_adr->is_AddP() &&
1507         (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1508       intptr_t ignore = 0;
1509       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1510       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1511       base = bs->step_over_gc_barrier(base);
1512       if (base != nullptr && base->is_Proj() &&
1513           base->as_Proj()->_con == TypeFunc::Parms &&
1514           base->in(0)->is_CallStaticJava() &&
1515           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1516         return base->in(0)->in(TypeFunc::Parms);
1517       }
1518     }
1519 
1520     break;
1521   }
1522 
1523   return nullptr;
1524 }
1525 
1526 //----------------------is_instance_field_load_with_local_phi------------------
1527 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1528   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1529       in(Address)->is_AddP() ) {
1530     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1531     // Only known instances and immutable fields
1532     if( t_oop != nullptr &&
1533         (t_oop->is_ptr_to_strict_final_field() ||
1534          t_oop->is_known_instance_field()) &&
1535         t_oop->offset() != Type::OffsetBot &&
1536         t_oop->offset() != Type::OffsetTop) {
1537       return true;
1538     }
1539   }
1540   return false;
1541 }
1542 
1543 //------------------------------Identity---------------------------------------
1544 // Loads are identity if previous store is to same address
1545 Node* LoadNode::Identity(PhaseGVN* phase) {
1546   // If the previous store-maker is the right kind of Store, and the store is
1547   // to the same address, then we are equal to the value stored.
1548   Node* mem = in(Memory);
1549   Node* value = can_see_stored_value_through_membars(mem, phase);
1550   if( value ) {
1551     // byte, short & char stores truncate naturally.
1552     // A load has to load the truncated value which requires
1553     // some sort of masking operation and that requires an
1554     // Ideal call instead of an Identity call.
1555     if (memory_size() < BytesPerInt) {
1556       // If the input to the store does not fit with the load's result type,
1557       // it must be truncated via an Ideal call.
1558       if (!phase->type(value)->higher_equal(phase->type(this)))
1559         return this;
1560     }
1561 
1562     if (phase->type(value)->isa_ptr() && phase->type(this)->isa_narrowoop()) {
1563       return this;
1564     }
1565     // (This works even when value is a Con, but LoadNode::Value
1566     // usually runs first, producing the singleton type of the Con.)
1567     if (!has_pinned_control_dependency() || value->is_Con()) {
1568       return value;
1569     } else {
1570       return this;
1571     }
1572   }
1573 
1574   if (has_pinned_control_dependency()) {
1575     return this;
1576   }
1577   // Search for an existing data phi which was generated before for the same
1578   // instance's field to avoid infinite generation of phis in a loop.
1579   Node *region = mem->in(0);
1580   if (is_instance_field_load_with_local_phi(region)) {
1581     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1582     int this_index  = phase->C->get_alias_index(addr_t);
1583     int this_offset = addr_t->offset();
1584     int this_iid    = addr_t->instance_id();
1585     if (!addr_t->is_known_instance() &&
1586          addr_t->is_ptr_to_strict_final_field()) {
1587       // Use _idx of address base (could be Phi node) for immutable fields in unknown instances
1588       intptr_t   ignore = 0;
1589       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1590       if (base == nullptr) {
1591         return this;
1592       }
1593       this_iid = base->_idx;
1594     }
1595     const Type* this_type = bottom_type();
1596     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1597       Node* phi = region->fast_out(i);
1598       if (phi->is_Phi() && phi != mem &&
1599           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1600         return phi;
1601       }
1602     }
1603   }
1604 
1605   return this;
1606 }
1607 
1608 // Construct an equivalent unsigned load.
1609 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1610   BasicType bt = T_ILLEGAL;
1611   const Type* rt = nullptr;
1612   switch (Opcode()) {
1613     case Op_LoadUB: return this;
1614     case Op_LoadUS: return this;
1615     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1616     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1617     default:
1618       assert(false, "no unsigned variant: %s", Name());
1619       return nullptr;
1620   }
1621   const Type* mem_t = gvn.type(in(MemNode::Address));
1622   if (mem_t == Type::TOP) {
1623     return gvn.C->top();
1624   }
1625   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1626                         mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1627                         false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1628 }
1629 
1630 // Construct an equivalent signed load.
1631 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1632   BasicType bt = T_ILLEGAL;
1633   const Type* rt = nullptr;
1634   switch (Opcode()) {
1635     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1636     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1637     case Op_LoadB: // fall through
1638     case Op_LoadS: // fall through
1639     case Op_LoadI: // fall through
1640     case Op_LoadL: return this;
1641     default:
1642       assert(false, "no signed variant: %s", Name());
1643       return nullptr;
1644   }
1645   const Type* mem_t = gvn.type(in(MemNode::Address));
1646   if (mem_t == Type::TOP) {
1647     return gvn.C->top();
1648   }
1649   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1650                         mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1651                         false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1652 }
1653 
1654 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1655   BasicType bt = rt->basic_type();
1656   switch (Opcode()) {
1657     case Op_LoadI: return (bt == T_FLOAT);
1658     case Op_LoadL: return (bt == T_DOUBLE);
1659     case Op_LoadF: return (bt == T_INT);
1660     case Op_LoadD: return (bt == T_LONG);
1661 
1662     default: return false;
1663   }
1664 }
1665 
1666 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1667   BasicType bt = rt->basic_type();
1668   assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1669   bool is_mismatched = is_mismatched_access();
1670   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1671   if (raw_type == nullptr) {
1672     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1673   }
1674   const int op = Opcode();
1675   bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1676                                (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1677   const Type* mem_t = gvn.type(in(MemNode::Address));
1678   if (mem_t == Type::TOP) {
1679     return gvn.C->top();
1680   }
1681   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1682                         mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1683                         require_atomic_access, is_unaligned_access(), is_mismatched);
1684 }
1685 
1686 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1687   BasicType bt = vt->basic_type();
1688   switch (Opcode()) {
1689     case Op_StoreI: return (bt == T_FLOAT);
1690     case Op_StoreL: return (bt == T_DOUBLE);
1691     case Op_StoreF: return (bt == T_INT);
1692     case Op_StoreD: return (bt == T_LONG);
1693 
1694     default: return false;
1695   }
1696 }
1697 
1698 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1699   BasicType bt = vt->basic_type();
1700   assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1701   const int op = Opcode();
1702   bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1703                                (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1704   const Type* mem_t = gvn.type(in(MemNode::Address));
1705   if (mem_t == Type::TOP) {
1706     return gvn.C->top();
1707   }
1708   StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1709                                   mem_t->is_ptr(), val, bt, _mo, require_atomic_access);
1710 
1711   bool is_mismatched = is_mismatched_access();
1712   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1713   if (raw_type == nullptr) {
1714     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1715   }
1716   if (is_mismatched) {
1717     st->set_mismatched_access();
1718   }
1719   return st;
1720 }
1721 
1722 // We're loading from an object which has autobox behaviour.
1723 // If this object is result of a valueOf call we'll have a phi
1724 // merging a newly allocated object and a load from the cache.
1725 // We want to replace this load with the original incoming
1726 // argument to the valueOf call.
1727 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1728   assert(igvn->C->eliminate_boxing(), "sanity");
1729   intptr_t ignore = 0;
1730   Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1731   if ((base == nullptr) || base->is_Phi()) {
1732     // Push the loads from the phi that comes from valueOf up
1733     // through it to allow elimination of the loads and the recovery
1734     // of the original value. It is done in split_through_phi().
1735     return nullptr;
1736   } else if (base->is_Load() ||
1737              (base->is_DecodeN() && base->in(1)->is_Load())) {
1738     // Eliminate the load of boxed value for integer types from the cache
1739     // array by deriving the value from the index into the array.
1740     // Capture the offset of the load and then reverse the computation.
1741 
1742     // Get LoadN node which loads a boxing object from 'cache' array.
1743     if (base->is_DecodeN()) {
1744       base = base->in(1);
1745     }
1746     if (!base->in(Address)->is_AddP()) {
1747       return nullptr; // Complex address
1748     }
1749     AddPNode* address = base->in(Address)->as_AddP();
1750     Node* cache_base = address->in(AddPNode::Base);
1751     if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1752       // Get ConP node which is static 'cache' field.
1753       cache_base = cache_base->in(1);
1754     }
1755     if ((cache_base != nullptr) && cache_base->is_Con()) {
1756       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1757       if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1758         Node* elements[4];
1759         int shift = exact_log2(type2aelembytes(T_OBJECT));
1760         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1761         if (count > 0 && elements[0]->is_Con() &&
1762             (count == 1 ||
1763              (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1764                             elements[1]->in(2) == igvn->intcon(shift)))) {
1765           ciObjArray* array = base_type->const_oop()->as_obj_array();
1766           // Fetch the box object cache[0] at the base of the array and get its value
1767           ciInstance* box = array->obj_at(0)->as_instance();
1768           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1769           assert(ik->is_box_klass(), "sanity");
1770           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1771           if (ik->nof_nonstatic_fields() == 1) {
1772             // This should be true nonstatic_field_at requires calling
1773             // nof_nonstatic_fields so check it anyway
1774             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1775             BasicType bt = c.basic_type();
1776             // Only integer types have boxing cache.
1777             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1778                    bt == T_BYTE    || bt == T_SHORT ||
1779                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1780             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1781             if (cache_low != (int)cache_low) {
1782               return nullptr; // should not happen since cache is array indexed by value
1783             }
1784             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1785             if (offset != (int)offset) {
1786               return nullptr; // should not happen since cache is array indexed by value
1787             }
1788            // Add up all the offsets making of the address of the load
1789             Node* result = elements[0];
1790             for (int i = 1; i < count; i++) {
1791               result = igvn->transform(new AddXNode(result, elements[i]));
1792             }
1793             // Remove the constant offset from the address and then
1794             result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1795             // remove the scaling of the offset to recover the original index.
1796             if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1797               // Peel the shift off directly but wrap it in a dummy node
1798               // since Ideal can't return existing nodes
1799               igvn->_worklist.push(result); // remove dead node later
1800               result = new RShiftXNode(result->in(1), igvn->intcon(0));
1801             } else if (result->is_Add() && result->in(2)->is_Con() &&
1802                        result->in(1)->Opcode() == Op_LShiftX &&
1803                        result->in(1)->in(2) == igvn->intcon(shift)) {
1804               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1805               // but for boxing cache access we know that X<<Z will not overflow
1806               // (there is range check) so we do this optimizatrion by hand here.
1807               igvn->_worklist.push(result); // remove dead node later
1808               Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1809               result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1810             } else {
1811               result = new RShiftXNode(result, igvn->intcon(shift));
1812             }
1813 #ifdef _LP64
1814             if (bt != T_LONG) {
1815               result = new ConvL2INode(igvn->transform(result));
1816             }
1817 #else
1818             if (bt == T_LONG) {
1819               result = new ConvI2LNode(igvn->transform(result));
1820             }
1821 #endif
1822             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1823             // Need to preserve unboxing load type if it is unsigned.
1824             switch(this->Opcode()) {
1825               case Op_LoadUB:
1826                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1827                 break;
1828               case Op_LoadUS:
1829                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1830                 break;
1831             }
1832             return result;
1833           }
1834         }
1835       }
1836     }
1837   }
1838   return nullptr;
1839 }
1840 
1841 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1842   Node* region = phi->in(0);
1843   if (region == nullptr) {
1844     return false; // Wait stable graph
1845   }
1846   uint cnt = phi->req();
1847   for (uint i = 1; i < cnt; i++) {
1848     Node* rc = region->in(i);
1849     if (rc == nullptr || phase->type(rc) == Type::TOP)
1850       return false; // Wait stable graph
1851     Node* in = phi->in(i);
1852     if (in == nullptr || phase->type(in) == Type::TOP)
1853       return false; // Wait stable graph
1854   }
1855   return true;
1856 }
1857 
1858 //------------------------------split_through_phi------------------------------
1859 // Check whether a call to 'split_through_phi' would split this load through the
1860 // Phi *base*. This method is essentially a copy of the validations performed
1861 // by 'split_through_phi'. The first use of this method was in EA code as part
1862 // of simplification of allocation merges.
1863 // Some differences from original method (split_through_phi):
1864 //  - If base->is_CastPP(): base = base->in(1)
1865 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1866   Node* mem        = in(Memory);
1867   Node* address    = in(Address);
1868   intptr_t ignore  = 0;
1869   Node*    base    = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1870 
1871   if (base == nullptr) {
1872     return false;
1873   }
1874 
1875   if (base->is_CastPP()) {
1876     base = base->in(1);
1877   }
1878 
1879   if (req() > 3 || !base->is_Phi()) {
1880     return false;
1881   }
1882 
1883   if (!mem->is_Phi()) {
1884     if (!MemNode::all_controls_dominate(mem, base->in(0), phase)) {
1885       return false;
1886     }
1887   } else if (base->in(0) != mem->in(0)) {
1888     if (!MemNode::all_controls_dominate(mem, base->in(0), phase)) {
1889       return false;
1890     }
1891   }
1892 
1893   return true;
1894 }
1895 
1896 //------------------------------split_through_phi------------------------------
1897 // Split instance or boxed field load through Phi.
1898 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1899   if (req() > 3) {
1900     assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1901     // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1902     return nullptr;
1903   }
1904   Node* mem     = in(Memory);
1905   Node* address = in(Address);
1906   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1907 
1908   assert((t_oop != nullptr) &&
1909          (ignore_missing_instance_id ||
1910           t_oop->is_known_instance_field() ||
1911           t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1912 
1913   Compile* C = phase->C;
1914   intptr_t ignore = 0;
1915   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1916   bool base_is_phi = (base != nullptr) && base->is_Phi();
1917   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1918                            (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1919                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1920 
1921   if (!((mem->is_Phi() || base_is_phi) &&
1922         (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1923     return nullptr; // Neither memory or base are Phi
1924   }
1925 
1926   if (mem->is_Phi()) {
1927     if (!stable_phi(mem->as_Phi(), phase)) {
1928       return nullptr; // Wait stable graph
1929     }
1930     uint cnt = mem->req();
1931     // Check for loop invariant memory.
1932     if (cnt == 3) {
1933       for (uint i = 1; i < cnt; i++) {
1934         Node* in = mem->in(i);
1935         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1936         if (m == mem) {
1937           if (i == 1) {
1938             // if the first edge was a loop, check second edge too.
1939             // If both are replaceable - we are in an infinite loop
1940             Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1941             if (n == mem) {
1942               break;
1943             }
1944           }
1945           set_req(Memory, mem->in(cnt - i));
1946           return this; // made change
1947         }
1948       }
1949     }
1950   }
1951   if (base_is_phi) {
1952     if (!stable_phi(base->as_Phi(), phase)) {
1953       return nullptr; // Wait stable graph
1954     }
1955     uint cnt = base->req();
1956     // Check for loop invariant memory.
1957     if (cnt == 3) {
1958       for (uint i = 1; i < cnt; i++) {
1959         if (base->in(i) == base) {
1960           return nullptr; // Wait stable graph
1961         }
1962       }
1963     }
1964   }
1965 
1966   // Split through Phi (see original code in loopopts.cpp).
1967   assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1968 
1969   // Do nothing here if Identity will find a value
1970   // (to avoid infinite chain of value phis generation).
1971   if (this != Identity(phase)) {
1972     return nullptr;
1973   }
1974 
1975   // Select Region to split through.
1976   Node* region;
1977   DomResult dom_result = DomResult::Dominate;
1978   if (!base_is_phi) {
1979     assert(mem->is_Phi(), "sanity");
1980     region = mem->in(0);
1981     // Skip if the region dominates some control edge of the address.
1982     // We will check `dom_result` later.
1983     dom_result = MemNode::maybe_all_controls_dominate(address, region, phase);
1984   } else if (!mem->is_Phi()) {
1985     assert(base_is_phi, "sanity");
1986     region = base->in(0);
1987     // Skip if the region dominates some control edge of the memory.
1988     // We will check `dom_result` later.
1989     dom_result = MemNode::maybe_all_controls_dominate(mem, region, phase);
1990   } else if (base->in(0) != mem->in(0)) {
1991     assert(base_is_phi && mem->is_Phi(), "sanity");
1992     dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0), phase);
1993     if (dom_result == DomResult::Dominate) {
1994       region = base->in(0);
1995     } else {
1996       dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0), phase);
1997       if (dom_result == DomResult::Dominate) {
1998         region = mem->in(0);
1999       }
2000       // Otherwise we encountered a complex graph.
2001     }
2002   } else {
2003     assert(base->in(0) == mem->in(0), "sanity");
2004     region = mem->in(0);
2005   }
2006 
2007   PhaseIterGVN* igvn = phase->is_IterGVN();
2008   if (dom_result != DomResult::Dominate) {
2009     if (dom_result == DomResult::EncounteredDeadCode) {
2010       // There is some dead code which eventually will be removed in IGVN.
2011       // Once this is the case, we get an unambiguous dominance result.
2012       // Push the node to the worklist again until the dead code is removed.
2013       igvn->_worklist.push(this);
2014     }
2015     return nullptr;
2016   }
2017 
2018   Node* phi = nullptr;
2019   const Type* this_type = this->bottom_type();
2020   if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
2021     int this_index = C->get_alias_index(t_oop);
2022     int this_offset = t_oop->offset();
2023     int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
2024     phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
2025   } else if (ignore_missing_instance_id) {
2026     phi = new PhiNode(region, this_type, nullptr, mem->_idx);
2027   } else {
2028     return nullptr;
2029   }
2030 
2031   for (uint i = 1; i < region->req(); i++) {
2032     Node* x;
2033     Node* the_clone = nullptr;
2034     Node* in = region->in(i);
2035     if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
2036         in != nullptr && in->is_OuterStripMinedLoop()) {
2037       // No node should go in the outer strip mined loop
2038       in = in->in(LoopNode::EntryControl);
2039     }
2040     if (in == nullptr || in == C->top()) {
2041       x = C->top();      // Dead path?  Use a dead data op
2042     } else {
2043       x = this->clone();        // Else clone up the data op
2044       the_clone = x;            // Remember for possible deletion.
2045       // Alter data node to use pre-phi inputs
2046       if (this->in(0) == region) {
2047         x->set_req(0, in);
2048       } else {
2049         x->set_req(0, nullptr);
2050       }
2051       if (mem->is_Phi() && (mem->in(0) == region)) {
2052         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
2053       }
2054       if (address->is_Phi() && address->in(0) == region) {
2055         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
2056       }
2057       if (base_is_phi && (base->in(0) == region)) {
2058         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
2059         Node* adr_x = phase->transform(AddPNode::make_with_base(base_x, address->in(AddPNode::Offset)));
2060         x->set_req(Address, adr_x);
2061       }
2062     }
2063     // Check for a 'win' on some paths
2064     const Type *t = x->Value(igvn);
2065 
2066     bool singleton = t->singleton();
2067 
2068     // See comments in PhaseIdealLoop::split_thru_phi().
2069     if (singleton && t == Type::TOP) {
2070       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
2071     }
2072 
2073     if (singleton) {
2074       x = igvn->makecon(t);
2075     } else {
2076       // We now call Identity to try to simplify the cloned node.
2077       // Note that some Identity methods call phase->type(this).
2078       // Make sure that the type array is big enough for
2079       // our new node, even though we may throw the node away.
2080       // (This tweaking with igvn only works because x is a new node.)
2081       igvn->set_type(x, t);
2082       // If x is a TypeNode, capture any more-precise type permanently into Node
2083       // otherwise it will be not updated during igvn->transform since
2084       // igvn->type(x) is set to x->Value() already.
2085       x->raise_bottom_type(t);
2086       Node* y = x->Identity(igvn);
2087       if (y != x) {
2088         x = y;
2089       } else {
2090         y = igvn->hash_find_insert(x);
2091         if (y) {
2092           x = y;
2093         } else {
2094           // Else x is a new node we are keeping
2095           // We do not need register_new_node_with_optimizer
2096           // because set_type has already been called.
2097           igvn->_worklist.push(x);
2098         }
2099       }
2100     }
2101     if (x != the_clone && the_clone != nullptr) {
2102       igvn->remove_dead_node(the_clone, PhaseIterGVN::NodeOrigin::Speculative);
2103     }
2104     phi->set_req(i, x);
2105   }
2106   // Record Phi
2107   igvn->register_new_node_with_optimizer(phi);
2108   return phi;
2109 }
2110 
2111 AllocateNode* LoadNode::is_new_object_mark_load() const {
2112   if (Opcode() == Op_LoadX) {
2113     Node* address = in(MemNode::Address);
2114     AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
2115     Node* mem = in(MemNode::Memory);
2116     if (alloc != nullptr && mem->is_Proj() &&
2117         mem->in(0) != nullptr &&
2118         mem->in(0) == alloc->initialization() &&
2119         alloc->initialization()->proj_out_or_null(0) != nullptr) {
2120       return alloc;
2121     }
2122   }
2123   return nullptr;
2124 }
2125 
2126 
2127 //------------------------------Ideal------------------------------------------
2128 // If the load is from Field memory and the pointer is non-null, it might be possible to
2129 // zero out the control input.
2130 // If the offset is constant and the base is an object allocation,
2131 // try to hook me up to the exact initializing store.
2132 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2133   if (has_pinned_control_dependency()) {
2134     return nullptr;
2135   }
2136   Node* p = MemNode::Ideal_common(phase, can_reshape);
2137   if (p)  return (p == NodeSentinel) ? nullptr : p;
2138 
2139   Node* ctrl    = in(MemNode::Control);
2140   Node* address = in(MemNode::Address);
2141 
2142   bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
2143          phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
2144 
2145   // Skip up past a SafePoint control.  Cannot do this for Stores because
2146   // pointer stores & cardmarks must stay on the same side of a SafePoint.
2147   if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
2148       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw  &&
2149       !addr_mark &&
2150       (depends_only_on_test() || has_unknown_control_dependency())) {
2151     ctrl = ctrl->in(0);
2152     set_req(MemNode::Control,ctrl);
2153     return this;
2154   }
2155 
2156   intptr_t ignore = 0;
2157   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
2158   if (base != nullptr
2159       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
2160     // Check for useless control edge in some common special cases
2161     if (in(MemNode::Control) != nullptr
2162         // TODO 8350865 Can we re-enable this?
2163         && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
2164         && can_remove_control()
2165         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
2166         && all_controls_dominate(base, phase->C->start(), phase)) {
2167       // A method-invariant, non-null address (constant or 'this' argument).
2168       set_req(MemNode::Control, nullptr);
2169       return this;
2170     }
2171   }
2172 
2173   Node* mem = in(MemNode::Memory);
2174   const TypePtr *addr_t = phase->type(address)->isa_ptr();
2175 
2176   if (can_reshape && (addr_t != nullptr)) {
2177     // try to optimize our memory input
2178     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
2179     if (opt_mem != mem) {
2180       set_req_X(MemNode::Memory, opt_mem, phase);
2181       if (phase->type( opt_mem ) == Type::TOP) return nullptr;
2182       return this;
2183     }
2184     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
2185     if ((t_oop != nullptr) &&
2186         (t_oop->is_known_instance_field() ||
2187          t_oop->is_ptr_to_boxed_value())) {
2188       PhaseIterGVN *igvn = phase->is_IterGVN();
2189       assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
2190       if (igvn->_worklist.member(opt_mem)) {
2191         // Delay this transformation until memory Phi is processed.
2192         igvn->_worklist.push(this);
2193         return nullptr;
2194       }
2195       // Split instance field load through Phi.
2196       Node* result = split_through_phi(phase);
2197       if (result != nullptr) return result;
2198 
2199       if (t_oop->is_ptr_to_boxed_value()) {
2200         Node* result = eliminate_autobox(igvn);
2201         if (result != nullptr) return result;
2202       }
2203     }
2204   }
2205 
2206   // Is there a dominating load that loads the same value?  Leave
2207   // anything that is not a load of a field/array element (like
2208   // barriers etc.) alone
2209   if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
2210     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2211       Node *use = mem->fast_out(i);
2212       if (use != this &&
2213           use->Opcode() == Opcode() &&
2214           use->in(0) != nullptr &&
2215           use->in(0) != in(0) &&
2216           use->in(Address) == in(Address)) {
2217         Node* ctl = in(0);
2218         for (int i = 0; i < 10 && ctl != nullptr; i++) {
2219           ctl = IfNode::up_one_dom(ctl);
2220           if (ctl == use->in(0)) {
2221             set_req(0, use->in(0));
2222             return this;
2223           }
2224         }
2225       }
2226     }
2227   }
2228 
2229   // Check for prior store with a different base or offset; make Load
2230   // independent.  Skip through any number of them.  Bail out if the stores
2231   // are in an endless dead cycle and report no progress.  This is a key
2232   // transform for Reflection.  However, if after skipping through the Stores
2233   // we can't then fold up against a prior store do NOT do the transform as
2234   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
2235   // array memory alive twice: once for the hoisted Load and again after the
2236   // bypassed Store.  This situation only works if EVERYBODY who does
2237   // anti-dependence work knows how to bypass.  I.e. we need all
2238   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
2239   // the alias index stuff.  So instead, peek through Stores and IFF we can
2240   // fold up, do so.
2241   Node* prev_mem = find_previous_store(phase);
2242   if (prev_mem != nullptr && prev_mem->is_top()) {
2243     // find_previous_store returns top when the access is dead
2244     return prev_mem;
2245   }
2246   if (prev_mem != nullptr) {
2247     Node* value = can_see_arraycopy_value(prev_mem, phase);
2248     if (value != nullptr) {
2249       return value;
2250     }
2251   }
2252   // Steps (a), (b):  Walk past independent stores to find an exact match.
2253   if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2254     // (c) See if we can fold up on the spot, but don't fold up here.
2255     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2256     // just return a prior value, which is done by Identity calls.
2257     if (can_see_stored_value_through_membars(prev_mem, phase)) {
2258       // Make ready for step (d):
2259       set_req_X(MemNode::Memory, prev_mem, phase);
2260       return this;
2261     }
2262   }
2263 
2264   if (!can_reshape) {
2265     phase->record_for_igvn(this);
2266   }
2267 
2268   return nullptr;
2269 }
2270 
2271 // Helper to recognize certain Klass fields which are invariant across
2272 // some group of array types (e.g., int[] or all T[] where T < Object).
2273 const Type*
2274 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2275                                  ciKlass* klass) const {
2276   assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2277          "must not happen");
2278 
2279   if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2280     // The field is InstanceKlass::_access_flags.  Return its (constant) value.
2281     assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2282     ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2283     return TypeInt::make(iklass->access_flags());
2284   }
2285   if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2286     // The field is Klass::_misc_flags.  Return its (constant) value.
2287     assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2288     return TypeInt::make(klass->misc_flags());
2289   }
2290   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2291     // The field is Klass::_layout_helper.  Return its constant value if known.
2292     assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2293     return TypeInt::make(klass->layout_helper());
2294   }
2295 
2296   // No match.
2297   return nullptr;
2298 }
2299 
2300 //------------------------------Value-----------------------------------------
2301 const Type* LoadNode::Value(PhaseGVN* phase) const {
2302   // Either input is TOP ==> the result is TOP
2303   Node* mem = in(MemNode::Memory);
2304   const Type *t1 = phase->type(mem);
2305   if (t1 == Type::TOP)  return Type::TOP;
2306   Node* adr = in(MemNode::Address);
2307   const TypePtr* tp = phase->type(adr)->isa_ptr();
2308   if (tp == nullptr || tp->empty())  return Type::TOP;
2309   int off = tp->offset();
2310   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2311   Compile* C = phase->C;
2312 
2313   // If load can see a previous constant store, use that.
2314   Node* value = can_see_stored_value_through_membars(mem, phase);
2315   if (value != nullptr && value->is_Con()) {
2316     if (phase->type(value)->isa_ptr() && _type->isa_narrowoop()) {
2317       return phase->type(value)->make_narrowoop();
2318     } else {
2319       assert(value->bottom_type()->higher_equal(_type), "sanity");
2320       return phase->type(value);
2321     }
2322   }
2323   // Try to guess loaded type from pointer type
2324   if (tp->isa_aryptr()) {
2325     const TypeAryPtr* ary = tp->is_aryptr();
2326     const Type* t = ary->elem();
2327 
2328     // Determine whether the reference is beyond the header or not, by comparing
2329     // the offset against the offset of the start of the array's data.
2330     // Different array types begin at slightly different offsets (12 vs. 16).
2331     // We choose T_BYTE as an example base type that is least restrictive
2332     // as to alignment, which will therefore produce the smallest
2333     // possible base offset.
2334     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2335     const bool off_beyond_header = (off >= min_base_off);
2336 
2337     // Try to constant-fold a stable array element.
2338     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2339       // Make sure the reference is not into the header and the offset is constant
2340       ciObject* aobj = ary->const_oop();
2341       if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2342         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2343         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, ary->field_offset().get(),
2344                                                                       stable_dimension,
2345                                                                       value_basic_type(), is_unsigned());
2346         if (con_type != nullptr) {
2347           return con_type;
2348         }
2349       }
2350     }
2351 
2352     // Don't do this for integer types. There is only potential profit if
2353     // the element type t is lower than _type; that is, for int types, if _type is
2354     // more restrictive than t.  This only happens here if one is short and the other
2355     // char (both 16 bits), and in those cases we've made an intentional decision
2356     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2357     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2358     //
2359     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2360     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
2361     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2362     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
2363     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2364     // In fact, that could have been the original type of p1, and p1 could have
2365     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2366     // expression (LShiftL quux 3) independently optimized to the constant 8.
2367     if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2368         && (_type->isa_vect() == nullptr)
2369         && !ary->is_flat()
2370         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2371       // t might actually be lower than _type, if _type is a unique
2372       // concrete subclass of abstract class t.
2373       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
2374         const Type* jt = t->join_speculative(_type);
2375         // In any case, do not allow the join, per se, to empty out the type.
2376         if (jt->empty() && !t->empty()) {
2377           // This can happen if a interface-typed array narrows to a class type.
2378           jt = _type;
2379         }
2380 #ifdef ASSERT
2381         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2382           // The pointers in the autobox arrays are always non-null
2383           Node* base = adr->in(AddPNode::Base);
2384           if ((base != nullptr) && base->is_DecodeN()) {
2385             // Get LoadN node which loads IntegerCache.cache field
2386             base = base->in(1);
2387           }
2388           if ((base != nullptr) && base->is_Con()) {
2389             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2390             if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2391               // It could be narrow oop
2392               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2393             }
2394           }
2395         }
2396 #endif
2397         return jt;
2398       }
2399     }
2400   } else if (tp->base() == Type::InstPtr) {
2401     assert( off != Type::OffsetBot ||
2402             // arrays can be cast to Objects
2403             !tp->isa_instptr() ||
2404             tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2405             // Default value load
2406             tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2407             // unsafe field access may not have a constant offset
2408             is_unsafe_access(),
2409             "Field accesses must be precise" );
2410     // For oop loads, we expect the _type to be precise.
2411 
2412     const TypeInstPtr* tinst = tp->is_instptr();
2413     BasicType bt = value_basic_type();
2414 
2415     // Fold loads of the field map
2416     if (tinst != nullptr) {
2417       ciInstanceKlass* ik = tinst->instance_klass();
2418       int offset = tinst->offset();
2419       if (ik == phase->C->env()->Class_klass()) {
2420         ciType* t = tinst->java_mirror_type();
2421         if (t != nullptr && t->is_inlinetype() && offset == t->as_inline_klass()->field_map_offset()) {
2422           ciConstant map = t->as_inline_klass()->get_field_map();
2423           bool is_narrow_oop = (bt == T_NARROWOOP);
2424           return Type::make_from_constant(map, true, 1, is_narrow_oop);
2425         }
2426       }
2427     }
2428 
2429     // Optimize loads from constant fields.
2430     ciObject* const_oop = tinst->const_oop();
2431     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2432       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2433       if (con_type != nullptr) {
2434         return con_type;
2435       }
2436     }
2437   } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2438     assert(off != Type::OffsetBot ||
2439             !tp->isa_instklassptr() ||
2440            // arrays can be cast to Objects
2441            tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2442            // also allow array-loading from the primary supertype
2443            // array during subtype checks
2444            Opcode() == Op_LoadKlass,
2445            "Field accesses must be precise");
2446     // For klass/static loads, we expect the _type to be precise
2447   } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2448     /* With mirrors being an indirect in the Klass*
2449      * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2450      * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2451      *
2452      * So check the type and klass of the node before the LoadP.
2453      */
2454     Node* adr2 = adr->in(MemNode::Address);
2455     const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2456     if (tkls != nullptr && !StressReflectiveCode) {
2457       if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2458         ciKlass* klass = tkls->exact_klass();
2459         assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2460         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2461         return TypeInstPtr::make(klass->java_mirror());
2462       }
2463     }
2464   }
2465 
2466   const TypeKlassPtr *tkls = tp->isa_klassptr();
2467   if (tkls != nullptr) {
2468     if (tkls->is_loaded() && tkls->klass_is_exact()) {
2469       ciKlass* klass = tkls->exact_klass();
2470       // We are loading a field from a Klass metaobject whose identity
2471       // is known at compile time (the type is "exact" or "precise").
2472       // Check for fields we know are maintained as constants by the VM.
2473       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2474         // The field is Klass::_super_check_offset.  Return its (constant) value.
2475         // (Folds up type checking code.)
2476         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2477         return TypeInt::make(klass->super_check_offset());
2478       }
2479       if (klass->is_inlinetype() && tkls->offset() == in_bytes(InstanceKlass::acmp_maps_offset_offset())) {
2480         return TypeInt::make(klass->as_inline_klass()->field_map_offset());
2481       }
2482       if (klass->is_obj_array_klass() && tkls->offset() == in_bytes(ObjArrayKlass::next_refined_array_klass_offset())) {
2483         // Fold loads from LibraryCallKit::load_default_refined_array_klass
2484         return tkls->is_aryklassptr()->cast_to_refined_array_klass_ptr();
2485       }
2486       if (klass->is_array_klass() && tkls->offset() == in_bytes(ObjArrayKlass::properties_offset())) {
2487         assert(klass->is_type_array_klass() || tkls->is_aryklassptr()->is_refined_type(), "Must be a refined array klass pointer");
2488         return TypeInt::make((jint)klass->as_array_klass()->properties().value());
2489       }
2490       if (klass->is_flat_array_klass() && tkls->offset() == in_bytes(FlatArrayKlass::layout_kind_offset())) {
2491         assert(Opcode() == Op_LoadI, "must load an int from _layout_kind");
2492         return TypeInt::make(static_cast<jint>(klass->as_flat_array_klass()->layout_kind()));
2493       }
2494       if (UseCompactObjectHeaders && tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2495         // The field is Klass::_prototype_header. Return its (constant) value.
2496         assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2497         return TypeX::make(klass->prototype_header());
2498       }
2499       // Compute index into primary_supers array
2500       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2501       // Check for overflowing; use unsigned compare to handle the negative case.
2502       if( depth < ciKlass::primary_super_limit() ) {
2503         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2504         // (Folds up type checking code.)
2505         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2506         ciKlass *ss = klass->super_of_depth(depth);
2507         return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2508       }
2509       const Type* aift = load_array_final_field(tkls, klass);
2510       if (aift != nullptr)  return aift;
2511     }
2512 
2513     // We can still check if we are loading from the primary_supers array at a
2514     // shallow enough depth.  Even though the klass is not exact, entries less
2515     // than or equal to its super depth are correct.
2516     if (tkls->is_loaded()) {
2517       ciKlass* klass = nullptr;
2518       if (tkls->isa_instklassptr()) {
2519         klass = tkls->is_instklassptr()->instance_klass();
2520       } else {
2521         int dims;
2522         const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2523         if (inner->isa_instklassptr()) {
2524           klass = inner->is_instklassptr()->instance_klass();
2525           klass = ciObjArrayKlass::make(klass, dims);
2526         }
2527       }
2528       if (klass != nullptr) {
2529         // Compute index into primary_supers array
2530         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2531         // Check for overflowing; use unsigned compare to handle the negative case.
2532         if (depth < ciKlass::primary_super_limit() &&
2533             depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2534           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2535           // (Folds up type checking code.)
2536           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2537           ciKlass *ss = klass->super_of_depth(depth);
2538           return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2539         }
2540       }
2541     }
2542 
2543     // If the type is enough to determine that the thing is not an array,
2544     // we can give the layout_helper a positive interval type.
2545     // This will help short-circuit some reflective code.
2546     if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2547         tkls->isa_instklassptr() && // not directly typed as an array
2548         !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable
2549     ) {
2550       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2551       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2552       // The key property of this type is that it folds up tests
2553       // for array-ness, since it proves that the layout_helper is positive.
2554       // Thus, a generic value like the basic object layout helper works fine.
2555       return TypeInt::make(min_size, max_jint, Type::WidenMin);
2556     }
2557   }
2558 
2559   // If we are loading from a freshly-allocated object/array, produce a zero.
2560   // Things to check:
2561   //   1. Load is beyond the header: headers are not guaranteed to be zero
2562   //   2. Load is not vectorized: vectors have no zero constant
2563   //   3. Load has no matching store, i.e. the input is the initial memory state
2564   const TypeOopPtr* tinst = tp->isa_oopptr();
2565   bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2566   bool is_not_vect = (_type->isa_vect() == nullptr);
2567   if (is_not_header && is_not_vect) {
2568     Node* mem = in(MemNode::Memory);
2569     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2570       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2571       // TODO 8350865 Scalar replacement does not work well for flat arrays.
2572       // Escape Analysis assumes that arrays are always zeroed during allocation which is not true for null-free arrays
2573       // ConnectionGraph::split_unique_types will re-wire the memory of loads from such arrays around the allocation
2574       // TestArrays::test6 and test152 and TestBasicFunctionality::test20 are affected by this.
2575       if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && tp->is_aryptr()->is_null_free()) {
2576         intptr_t offset = 0;
2577         Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2578         AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
2579         if (alloc != nullptr && alloc->is_AllocateArray() && alloc->in(AllocateNode::InitValue) != nullptr) {
2580           return _type;
2581         }
2582       }
2583       return Type::get_zero_type(_type->basic_type());
2584     }
2585   }
2586   if (!UseCompactObjectHeaders) {
2587     Node* alloc = is_new_object_mark_load();
2588     if (alloc != nullptr) {
2589       if (Arguments::is_valhalla_enabled()) {
2590         // The mark word may contain property bits (inline, flat, null-free)
2591         Node* klass_node = alloc->in(AllocateNode::KlassNode);
2592         const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2593         if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2594           return TypeX::make(tkls->exact_klass()->prototype_header());
2595         }
2596       } else {
2597         return TypeX::make(markWord::prototype().value());
2598       }
2599     }
2600   }
2601 
2602   return _type;
2603 }
2604 
2605 //------------------------------match_edge-------------------------------------
2606 // Do we Match on this edge index or not?  Match only the address.
2607 uint LoadNode::match_edge(uint idx) const {
2608   return idx == MemNode::Address;
2609 }
2610 
2611 //--------------------------LoadBNode::Ideal--------------------------------------
2612 //
2613 //  If the previous store is to the same address as this load,
2614 //  and the value stored was larger than a byte, replace this load
2615 //  with the value stored truncated to a byte.  If no truncation is
2616 //  needed, the replacement is done in LoadNode::Identity().
2617 //
2618 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2619   Node* mem = in(MemNode::Memory);
2620   Node* value = can_see_stored_value_through_membars(mem, phase);
2621   if (value != nullptr) {
2622     Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2623     if (narrow != value) {
2624       return narrow;
2625     }
2626   }
2627   // Identity call will handle the case where truncation is not needed.
2628   return LoadNode::Ideal(phase, can_reshape);
2629 }
2630 
2631 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2632   Node* mem = in(MemNode::Memory);
2633   Node* value = can_see_stored_value_through_membars(mem, phase);
2634   if (value != nullptr && value->is_Con() &&
2635       !value->bottom_type()->higher_equal(_type)) {
2636     // If the input to the store does not fit with the load's result type,
2637     // it must be truncated. We can't delay until Ideal call since
2638     // a singleton Value is needed for split_thru_phi optimization.
2639     int con = value->get_int();
2640     return TypeInt::make((con << 24) >> 24);
2641   }
2642   return LoadNode::Value(phase);
2643 }
2644 
2645 //--------------------------LoadUBNode::Ideal-------------------------------------
2646 //
2647 //  If the previous store is to the same address as this load,
2648 //  and the value stored was larger than a byte, replace this load
2649 //  with the value stored truncated to a byte.  If no truncation is
2650 //  needed, the replacement is done in LoadNode::Identity().
2651 //
2652 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2653   Node* mem = in(MemNode::Memory);
2654   Node* value = can_see_stored_value_through_membars(mem, phase);
2655   if (value != nullptr) {
2656     Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2657     if (narrow != value) {
2658       return narrow;
2659     }
2660   }
2661   // Identity call will handle the case where truncation is not needed.
2662   return LoadNode::Ideal(phase, can_reshape);
2663 }
2664 
2665 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2666   Node* mem = in(MemNode::Memory);
2667   Node* value = can_see_stored_value_through_membars(mem, phase);
2668   if (value != nullptr && value->is_Con() &&
2669       !value->bottom_type()->higher_equal(_type)) {
2670     // If the input to the store does not fit with the load's result type,
2671     // it must be truncated. We can't delay until Ideal call since
2672     // a singleton Value is needed for split_thru_phi optimization.
2673     int con = value->get_int();
2674     return TypeInt::make(con & 0xFF);
2675   }
2676   return LoadNode::Value(phase);
2677 }
2678 
2679 //--------------------------LoadUSNode::Ideal-------------------------------------
2680 //
2681 //  If the previous store is to the same address as this load,
2682 //  and the value stored was larger than a char, replace this load
2683 //  with the value stored truncated to a char.  If no truncation is
2684 //  needed, the replacement is done in LoadNode::Identity().
2685 //
2686 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2687   Node* mem = in(MemNode::Memory);
2688   Node* value = can_see_stored_value_through_membars(mem, phase);
2689   if (value != nullptr) {
2690     Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2691     if (narrow != value) {
2692       return narrow;
2693     }
2694   }
2695   // Identity call will handle the case where truncation is not needed.
2696   return LoadNode::Ideal(phase, can_reshape);
2697 }
2698 
2699 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2700   Node* mem = in(MemNode::Memory);
2701   Node* value = can_see_stored_value_through_membars(mem, phase);
2702   if (value != nullptr && value->is_Con() &&
2703       !value->bottom_type()->higher_equal(_type)) {
2704     // If the input to the store does not fit with the load's result type,
2705     // it must be truncated. We can't delay until Ideal call since
2706     // a singleton Value is needed for split_thru_phi optimization.
2707     int con = value->get_int();
2708     return TypeInt::make(con & 0xFFFF);
2709   }
2710   return LoadNode::Value(phase);
2711 }
2712 
2713 //--------------------------LoadSNode::Ideal--------------------------------------
2714 //
2715 //  If the previous store is to the same address as this load,
2716 //  and the value stored was larger than a short, replace this load
2717 //  with the value stored truncated to a short.  If no truncation is
2718 //  needed, the replacement is done in LoadNode::Identity().
2719 //
2720 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2721   Node* mem = in(MemNode::Memory);
2722   Node* value = can_see_stored_value_through_membars(mem, phase);
2723   if (value != nullptr) {
2724     Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2725     if (narrow != value) {
2726       return narrow;
2727     }
2728   }
2729   // Identity call will handle the case where truncation is not needed.
2730   return LoadNode::Ideal(phase, can_reshape);
2731 }
2732 
2733 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2734   Node* mem = in(MemNode::Memory);
2735   Node* value = can_see_stored_value_through_membars(mem, phase);
2736   if (value != nullptr && value->is_Con() &&
2737       !value->bottom_type()->higher_equal(_type)) {
2738     // If the input to the store does not fit with the load's result type,
2739     // it must be truncated. We can't delay until Ideal call since
2740     // a singleton Value is needed for split_thru_phi optimization.
2741     int con = value->get_int();
2742     return TypeInt::make((con << 16) >> 16);
2743   }
2744   return LoadNode::Value(phase);
2745 }
2746 
2747 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2748   // Can see the corresponding value, may need to add an EncodeP
2749   Node* value = can_see_stored_value_through_membars(in(Memory), phase);
2750   if (value != nullptr && phase->type(value)->isa_ptr() && type()->isa_narrowoop()) {
2751     return new EncodePNode(value, type());
2752   }
2753 
2754   // Identity call will handle the case where EncodeP is unnecessary
2755   return LoadNode::Ideal(phase, can_reshape);
2756 }
2757 
2758 //=============================================================================
2759 //----------------------------LoadKlassNode::make------------------------------
2760 // Polymorphic factory method:
2761 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2762   // sanity check the alias category against the created node type
2763   const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2764   assert(adr_type != nullptr, "expecting TypeKlassPtr");
2765 #ifdef _LP64
2766   if (adr_type->is_ptr_to_narrowklass()) {
2767     Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2768     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2769   }
2770 #endif
2771   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2772   return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2773 }
2774 
2775 //------------------------------Value------------------------------------------
2776 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2777   return klass_value_common(phase);
2778 }
2779 
2780 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2781   // Either input is TOP ==> the result is TOP
2782   const Type *t1 = phase->type( in(MemNode::Memory) );
2783   if (t1 == Type::TOP)  return Type::TOP;
2784   Node *adr = in(MemNode::Address);
2785   const Type *t2 = phase->type( adr );
2786   if (t2 == Type::TOP)  return Type::TOP;
2787   const TypePtr *tp = t2->is_ptr();
2788   if (TypePtr::above_centerline(tp->ptr()) ||
2789       tp->ptr() == TypePtr::Null)  return Type::TOP;
2790 
2791   // Return a more precise klass, if possible
2792   const TypeInstPtr *tinst = tp->isa_instptr();
2793   if (tinst != nullptr) {
2794     ciInstanceKlass* ik = tinst->instance_klass();
2795     int offset = tinst->offset();
2796     if (ik == phase->C->env()->Class_klass()
2797         && (offset == java_lang_Class::klass_offset() ||
2798             offset == java_lang_Class::array_klass_offset())) {
2799       // We are loading a special hidden field from a Class mirror object,
2800       // the field which points to the VM's Klass metaobject.
2801       ciType* t = tinst->java_mirror_type();
2802       // java_mirror_type returns non-null for compile-time Class constants.
2803       if (t != nullptr) {
2804         // constant oop => constant klass
2805         if (offset == java_lang_Class::array_klass_offset()) {
2806           if (t->is_void()) {
2807             // We cannot create a void array.  Since void is a primitive type return null
2808             // klass.  Users of this result need to do a null check on the returned klass.
2809             return TypePtr::NULL_PTR;
2810           }
2811           return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2812         }
2813         if (!t->is_klass()) {
2814           // a primitive Class (e.g., int.class) has null for a klass field
2815           return TypePtr::NULL_PTR;
2816         }
2817         // Fold up the load of the hidden field
2818         return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2819       }
2820       // non-constant mirror, so we can't tell what's going on
2821     }
2822     if (!tinst->is_loaded())
2823       return _type;             // Bail out if not loaded
2824     if (offset == oopDesc::klass_offset_in_bytes()) {
2825       return tinst->as_klass_type(true);
2826     }
2827   }
2828 
2829   // Check for loading klass from an array
2830   const TypeAryPtr* tary = tp->isa_aryptr();
2831   if (tary != nullptr &&
2832       tary->offset() == oopDesc::klass_offset_in_bytes()) {
2833     return tary->as_klass_type(true)->is_aryklassptr();
2834   }
2835 
2836   // Check for loading klass from an array klass
2837   const TypeKlassPtr *tkls = tp->isa_klassptr();
2838   if (tkls != nullptr && !StressReflectiveCode) {
2839     if (!tkls->is_loaded())
2840      return _type;             // Bail out if not loaded
2841     if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2842         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2843       // // Always returning precise element type is incorrect,
2844       // // e.g., element type could be object and array may contain strings
2845       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2846 
2847       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2848       // according to the element type's subclassing.
2849       return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2850     }
2851     if (tkls->isa_aryklassptr() != nullptr && tkls->klass_is_exact() &&
2852         !tkls->exact_klass()->is_type_array_klass() &&
2853         tkls->offset() == in_bytes(Klass::super_offset())) {
2854       // We are loading the super klass of a refined array klass, return the non-refined klass pointer
2855       assert(tkls->is_aryklassptr()->is_refined_type(), "Must be a refined array klass pointer");
2856       return tkls->is_aryklassptr()->with_offset(0)->cast_to_non_refined();
2857     }
2858     if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2859         tkls->offset() == in_bytes(Klass::super_offset())) {
2860       ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2861       // The field is Klass::_super.  Return its (constant) value.
2862       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2863       return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2864     }
2865   }
2866 
2867   if (tkls != nullptr && !UseSecondarySupersCache
2868       && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset()))  {
2869     // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2870     return TypePtr::NULL_PTR;
2871   }
2872 
2873   // Bailout case
2874   return LoadNode::Value(phase);
2875 }
2876 
2877 //------------------------------Identity---------------------------------------
2878 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2879 // Also feed through the klass in Allocate(...klass...)._klass.
2880 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2881   return klass_identity_common(phase);
2882 }
2883 
2884 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2885   Node* x = LoadNode::Identity(phase);
2886   if (x != this)  return x;
2887 
2888   // Take apart the address into an oop and offset.
2889   // Return 'this' if we cannot.
2890   Node*    adr    = in(MemNode::Address);
2891   intptr_t offset = 0;
2892   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2893   if (base == nullptr)     return this;
2894   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2895   if (toop == nullptr)     return this;
2896 
2897   // Step over potential GC barrier for OopHandle resolve
2898   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2899   if (bs->is_gc_barrier_node(base)) {
2900     base = bs->step_over_gc_barrier(base);
2901   }
2902 
2903   // We can fetch the klass directly through an AllocateNode.
2904   // This works even if the klass is not constant (clone or newArray).
2905   if (offset == oopDesc::klass_offset_in_bytes()) {
2906     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2907     if (allocated_klass != nullptr) {
2908       return allocated_klass;
2909     }
2910   }
2911 
2912   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2913   // See inline_native_Class_query for occurrences of these patterns.
2914   // Java Example:  x.getClass().isAssignableFrom(y)
2915   //
2916   // This improves reflective code, often making the Class
2917   // mirror go completely dead.  (Current exception:  Class
2918   // mirrors may appear in debug info, but we could clean them out by
2919   // introducing a new debug info operator for Klass.java_mirror).
2920   //
2921   // This optimization does not apply to arrays because if k is not a
2922   // constant, it was obtained via load_klass which returns the refined type
2923   // and '.java_mirror.as_klass' should return the Java type instead.
2924 
2925   if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2926       && offset == java_lang_Class::klass_offset()) {
2927     if (base->is_Load()) {
2928       Node* base2 = base->in(MemNode::Address);
2929       if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2930         Node* adr2 = base2->in(MemNode::Address);
2931         const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2932         if (tkls != nullptr && !tkls->empty()
2933             && ((tkls->isa_instklassptr() && !tkls->is_instklassptr()->might_be_an_array()))
2934             && adr2->is_AddP()) {
2935           int mirror_field = in_bytes(Klass::java_mirror_offset());
2936           if (tkls->offset() == mirror_field) {
2937 #ifdef ASSERT
2938             const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2939             assert(tkls2->offset() == 0, "not a load of java_mirror");
2940 #endif
2941             assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2942             assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2943             return adr2->in(AddPNode::Address);
2944           }
2945         }
2946       }
2947     }
2948   }
2949 
2950   return this;
2951 }
2952 
2953 LoadNode* LoadNode::clone_pinned() const {
2954   LoadNode* ld = clone()->as_Load();
2955   ld->_control_dependency = UnknownControl;
2956   return ld;
2957 }
2958 
2959 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2960 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2961 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2962 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2963 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2964 // those case, since there is not a dependency between the node and its control input, we do not
2965 // need to pin it.
2966 LoadNode* LoadNode::pin_node_under_control_impl() const {
2967   const TypePtr* adr_type = this->adr_type();
2968   if (adr_type != nullptr && adr_type->isa_aryptr()) {
2969     // Only array accesses have dependencies on their control input
2970     return clone_pinned();
2971   }
2972   return nullptr;
2973 }
2974 
2975 //------------------------------Value------------------------------------------
2976 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2977   const Type *t = klass_value_common(phase);
2978   if (t == Type::TOP)
2979     return t;
2980 
2981   return t->make_narrowklass();
2982 }
2983 
2984 //------------------------------Identity---------------------------------------
2985 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2986 // Also feed through the klass in Allocate(...klass...)._klass.
2987 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2988   Node *x = klass_identity_common(phase);
2989 
2990   const Type *t = phase->type( x );
2991   if( t == Type::TOP ) return x;
2992   if( t->isa_narrowklass()) return x;
2993   assert (!t->isa_narrowoop(), "no narrow oop here");
2994 
2995   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2996 }
2997 
2998 //------------------------------Value-----------------------------------------
2999 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
3000   // Either input is TOP ==> the result is TOP
3001   const Type *t1 = phase->type( in(MemNode::Memory) );
3002   if( t1 == Type::TOP ) return Type::TOP;
3003   Node *adr = in(MemNode::Address);
3004   const Type *t2 = phase->type( adr );
3005   if( t2 == Type::TOP ) return Type::TOP;
3006   const TypePtr *tp = t2->is_ptr();
3007   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
3008   const TypeAryPtr *tap = tp->isa_aryptr();
3009   if( !tap ) return _type;
3010   return tap->size();
3011 }
3012 
3013 //-------------------------------Ideal---------------------------------------
3014 // Feed through the length in AllocateArray(...length...)._length.
3015 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3016   Node* p = MemNode::Ideal_common(phase, can_reshape);
3017   if (p)  return (p == NodeSentinel) ? nullptr : p;
3018 
3019   // Take apart the address into an oop and offset.
3020   // Return 'this' if we cannot.
3021   Node*    adr    = in(MemNode::Address);
3022   intptr_t offset = 0;
3023   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
3024   if (base == nullptr)     return nullptr;
3025   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
3026   if (tary == nullptr)     return nullptr;
3027 
3028   // We can fetch the length directly through an AllocateArrayNode.
3029   // This works even if the length is not constant (clone or newArray).
3030   if (offset == arrayOopDesc::length_offset_in_bytes()) {
3031     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
3032     if (alloc != nullptr) {
3033       Node* allocated_length = alloc->Ideal_length();
3034       Node* len = alloc->make_ideal_length(tary, phase);
3035       if (allocated_length != len) {
3036         // New CastII improves on this.
3037         return len;
3038       }
3039     }
3040   }
3041 
3042   return nullptr;
3043 }
3044 
3045 //------------------------------Identity---------------------------------------
3046 // Feed through the length in AllocateArray(...length...)._length.
3047 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
3048   Node* x = LoadINode::Identity(phase);
3049   if (x != this)  return x;
3050 
3051   // Take apart the address into an oop and offset.
3052   // Return 'this' if we cannot.
3053   Node*    adr    = in(MemNode::Address);
3054   intptr_t offset = 0;
3055   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3056   if (base == nullptr)     return this;
3057   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
3058   if (tary == nullptr)     return this;
3059 
3060   // We can fetch the length directly through an AllocateArrayNode.
3061   // This works even if the length is not constant (clone or newArray).
3062   if (offset == arrayOopDesc::length_offset_in_bytes()) {
3063     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
3064     if (alloc != nullptr) {
3065       Node* allocated_length = alloc->Ideal_length();
3066       // Do not allow make_ideal_length to allocate a CastII node.
3067       Node* len = alloc->make_ideal_length(tary, phase, false);
3068       if (allocated_length == len) {
3069         // Return allocated_length only if it would not be improved by a CastII.
3070         return allocated_length;
3071       }
3072     }
3073   }
3074 
3075   return this;
3076 
3077 }
3078 
3079 //=============================================================================
3080 //---------------------------StoreNode::make-----------------------------------
3081 // Polymorphic factory method:
3082 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) {
3083   assert((mo == unordered || mo == release), "unexpected");
3084   Compile* C = gvn.C;
3085   assert(adr_type == nullptr || adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr()) == C->get_alias_index(adr_type), "adr and adr_type must agree");
3086   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
3087          ctl != nullptr, "raw memory operations should have control edge");
3088 
3089   switch (bt) {
3090   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
3091   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
3092   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
3093   case T_CHAR:
3094   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
3095   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
3096   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
3097   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
3098   case T_METADATA:
3099   case T_ADDRESS:
3100   case T_OBJECT:
3101   case T_ARRAY:
3102 #ifdef _LP64
3103     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3104       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
3105       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
3106     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
3107                (val->bottom_type()->isa_klassptr() && adr->bottom_type()->isa_rawptr())) {
3108       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
3109       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
3110     }
3111 #endif
3112     {
3113       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
3114     }
3115   default:
3116     guarantee(false, "unexpected basic type %s", type2name(bt));
3117     return (StoreNode*)nullptr;
3118   }
3119 }
3120 
3121 //--------------------------bottom_type----------------------------------------
3122 const Type *StoreNode::bottom_type() const {
3123   return Type::MEMORY;
3124 }
3125 
3126 //------------------------------hash-------------------------------------------
3127 uint StoreNode::hash() const {
3128   // unroll addition of interesting fields
3129   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
3130 
3131   // Since they are not commoned, do not hash them:
3132   return NO_HASH;
3133 }
3134 
3135 // Link together multiple stores (B/S/C/I) into a longer one.
3136 //
3137 // Example: _store = StoreB[i+3]
3138 //
3139 //   RangeCheck[i+0]           RangeCheck[i+0]
3140 //   StoreB[i+0]
3141 //   RangeCheck[i+3]           RangeCheck[i+3]
3142 //   StoreB[i+1]         -->   pass:             fail:
3143 //   StoreB[i+2]               StoreI[i+0]       StoreB[i+0]
3144 //   StoreB[i+3]
3145 //
3146 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
3147 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
3148 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
3149 // failing path.
3150 //
3151 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
3152 //       - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
3153 //                                       and possibly no RangeChecks remain between the stores.
3154 //       - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
3155 //                              and those later RangeChecks can be removed. Example:
3156 //
3157 //                              RangeCheck[i+0]                         RangeCheck[i+0] <- before first store
3158 //                              StoreB[i+0]                             StoreB[i+0]     <- first store
3159 //                              RangeCheck[i+1]     --> smeared -->     RangeCheck[i+3] <- only RC between first and last store
3160 //                              StoreB[i+1]                             StoreB[i+1]     <- second store
3161 //                              RangeCheck[i+2]     --> removed
3162 //                              StoreB[i+2]                             StoreB[i+2]
3163 //                              RangeCheck[i+3]     --> removed
3164 //                              StoreB[i+3]                             StoreB[i+3]     <- last store
3165 //
3166 //                              Thus, it is a common pattern that between the first and last store in a chain
3167 //                              of adjacent stores there remains exactly one RangeCheck, located between the
3168 //                              first and the second store (e.g. RangeCheck[i+3]).
3169 //
3170 class MergePrimitiveStores : public StackObj {
3171 private:
3172   PhaseGVN* const _phase;
3173   StoreNode* const _store;
3174   // State machine with initial state Unknown
3175   // Allowed transitions:
3176   //   Unknown     -> Const
3177   //   Unknown     -> Platform
3178   //   Unknown     -> Reverse
3179   //   Unknown     -> NotAdjacent
3180   //   Const       -> Const
3181   //   Const       -> NotAdjacent
3182   //   Platform    -> Platform
3183   //   Platform    -> NotAdjacent
3184   //   Reverse     -> Reverse
3185   //   Reverse     -> NotAdjacent
3186   //   NotAdjacent -> NotAdjacent
3187   enum ValueOrder : uint8_t {
3188     Unknown,     // Initial state
3189     Const,       // Input values are const
3190     Platform,    // Platform order
3191     Reverse,     // Reverse platform order
3192     NotAdjacent  // Not adjacent
3193   };
3194   ValueOrder  _value_order;
3195 
3196   NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
3197 
3198 public:
3199   MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
3200     _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
3201     NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
3202     {}
3203 
3204   StoreNode* run();
3205 
3206 private:
3207   bool is_compatible_store(const StoreNode* other_store) const;
3208   bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3209   bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3210   static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
3211   enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3212   static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3213 
3214   class Status {
3215   private:
3216     StoreNode* _found_store;
3217     bool       _found_range_check;
3218 
3219     Status(StoreNode* found_store, bool found_range_check)
3220       : _found_store(found_store), _found_range_check(found_range_check) {}
3221 
3222   public:
3223     StoreNode* found_store() const { return _found_store; }
3224     bool found_range_check() const { return _found_range_check; }
3225     static Status make_failure() { return Status(nullptr, false); }
3226 
3227     static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3228       if (cfg_status == CFGStatus::Failure) {
3229         return Status::make_failure();
3230       }
3231       return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3232     }
3233 
3234 #ifndef PRODUCT
3235     void print_on(outputStream* st) const {
3236       if (_found_store == nullptr) {
3237         st->print_cr("None");
3238       } else {
3239         st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
3240                                          _found_range_check ? "RC" : "no-RC");
3241       }
3242     }
3243 #endif
3244   };
3245 
3246   enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
3247   Status find_adjacent_use_store(const StoreNode* def_store) const;
3248   Status find_adjacent_def_store(const StoreNode* use_store) const;
3249   Status find_use_store(const StoreNode* def_store) const;
3250   Status find_def_store(const StoreNode* use_store) const;
3251   Status find_use_store_unidirectional(const StoreNode* def_store) const;
3252   Status find_def_store_unidirectional(const StoreNode* use_store) const;
3253 
3254   void collect_merge_list(Node_List& merge_list) const;
3255   Node* make_merged_input_value(const Node_List& merge_list);
3256   StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3257 
3258 #ifndef PRODUCT
3259   // Access to TraceMergeStores tags
3260   bool is_trace(TraceMergeStores::Tag tag) const {
3261     return _trace_tags.at(tag);
3262   }
3263 
3264   bool is_trace_basic() const {
3265     return is_trace(TraceMergeStores::Tag::BASIC);
3266   }
3267 
3268   bool is_trace_pointer_parsing() const {
3269     return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
3270   }
3271 
3272   bool is_trace_pointer_aliasing() const {
3273     return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
3274   }
3275 
3276   bool is_trace_pointer_adjacency() const {
3277     return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
3278   }
3279 
3280   bool is_trace_success() const {
3281     return is_trace(TraceMergeStores::Tag::SUCCESS);
3282   }
3283 #endif
3284 
3285   NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3286 };
3287 
3288 StoreNode* MergePrimitiveStores::run() {
3289   // Check for B/S/C/I
3290   int opc = _store->Opcode();
3291   if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3292     return nullptr;
3293   }
3294 
3295   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3296 
3297   // The _store must be the "last" store in a chain. If we find a use we could merge with
3298   // then that use or a store further down is the "last" store.
3299   Status status_use = find_adjacent_use_store(_store);
3300   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3301   if (status_use.found_store() != nullptr) {
3302     return nullptr;
3303   }
3304 
3305   // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3306   Status status_def = find_adjacent_def_store(_store);
3307   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3308   Node* def_store = status_def.found_store();
3309   if (def_store == nullptr) {
3310     return nullptr;
3311   }
3312 
3313   // Initialize value order
3314   _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3315                                                  _store->in(MemNode::ValueIn),
3316                                                  _store->memory_size());
3317   assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3318 
3319   ResourceMark rm;
3320   Node_List merge_list;
3321   collect_merge_list(merge_list);
3322 
3323   Node* merged_input_value = make_merged_input_value(merge_list);
3324   if (merged_input_value == nullptr) { return nullptr; }
3325 
3326   StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3327 
3328   NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3329 
3330   return merged_store;
3331 }
3332 
3333 // Check compatibility between _store and other_store.
3334 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3335   int opc = _store->Opcode();
3336   assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3337 
3338   if (other_store == nullptr ||
3339       _store->Opcode() != other_store->Opcode()) {
3340     return false;
3341   }
3342 
3343   return true;
3344 }
3345 
3346 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3347   if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3348                               use_store->in(MemNode::ValueIn),
3349                               def_store->memory_size())) {
3350     return false;
3351   }
3352 
3353   ResourceMark rm;
3354 #ifndef PRODUCT
3355   const TraceMemPointer trace(is_trace_pointer_parsing(),
3356                               is_trace_pointer_aliasing(),
3357                               is_trace_pointer_adjacency(),
3358                               true);
3359 #endif
3360   const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3361   const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3362   return pointer_def.is_adjacent_to_and_before(pointer_use);
3363 }
3364 
3365 // Check input values n1 and n2 can be merged and return the value order
3366 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3367                                                                                        const int memory_size) const {
3368   // Pattern: [n1 = ConI, n2 = ConI]
3369   if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3370     return ValueOrder::Const;
3371   }
3372 
3373   Node const *base_n2;
3374   jint shift_n2;
3375   if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3376     return ValueOrder::NotAdjacent;
3377   }
3378   Node const *base_n1;
3379   jint shift_n1;
3380   if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3381     return ValueOrder::NotAdjacent;
3382   }
3383 
3384   int bits_per_store = memory_size * 8;
3385   if (base_n1 != base_n2 ||
3386       abs(shift_n1 - shift_n2) != bits_per_store ||
3387       shift_n1 % bits_per_store != 0) {
3388     // Values are not adjacent
3389     return ValueOrder::NotAdjacent;
3390   }
3391 
3392   // Detect value order
3393 #ifdef VM_LITTLE_ENDIAN
3394   return shift_n1 < shift_n2 ? ValueOrder::Platform     // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3395                              : ValueOrder::Reverse;     // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3396 #else
3397   return shift_n1 > shift_n2 ? ValueOrder::Platform     // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3398                              : ValueOrder::Reverse;     // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3399 #endif
3400 }
3401 
3402 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3403   ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3404 
3405   switch (input_value_order) {
3406     case ValueOrder::NotAdjacent:
3407       return false;
3408     case ValueOrder::Reverse:
3409       if (memory_size != 1 ||
3410           !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3411           !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3412           !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3413         // ReverseBytes are not supported by platform
3414         return false;
3415       }
3416       // fall-through.
3417     case ValueOrder::Const:
3418     case ValueOrder::Platform:
3419       if (_value_order == ValueOrder::Unknown) {
3420         // Initial state is Unknown, and we find a valid input value order
3421         return true;
3422       }
3423       // The value order can not be changed
3424       return _value_order == input_value_order;
3425     case ValueOrder::Unknown:
3426     default:
3427       ShouldNotReachHere();
3428   }
3429   return false;
3430 }
3431 
3432 // Detect pattern: n = base_out >> shift_out
3433 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3434   assert(n != nullptr, "precondition");
3435 
3436   int opc = n->Opcode();
3437   if (opc == Op_ConvL2I) {
3438     n = n->in(1);
3439     opc = n->Opcode();
3440   }
3441 
3442   if ((opc == Op_RShiftI ||
3443        opc == Op_RShiftL ||
3444        opc == Op_URShiftI ||
3445        opc == Op_URShiftL) &&
3446       n->in(2)->is_ConI()) {
3447     base_out = n->in(1);
3448     shift_out = n->in(2)->get_int();
3449     // The shift must be positive:
3450     return shift_out >= 0;
3451   }
3452 
3453   if (phase->type(n)->isa_int()  != nullptr ||
3454       phase->type(n)->isa_long() != nullptr) {
3455     // (base >> 0)
3456     base_out = n;
3457     shift_out = 0;
3458     return true;
3459   }
3460   return false;
3461 }
3462 
3463 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3464 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3465   assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3466 
3467   Node* ctrl_use = use_store->in(MemNode::Control);
3468   Node* ctrl_def = def_store->in(MemNode::Control);
3469   if (ctrl_use == nullptr || ctrl_def == nullptr) {
3470     return CFGStatus::Failure;
3471   }
3472 
3473   if (ctrl_use == ctrl_def) {
3474     // Same ctrl -> no RangeCheck in between.
3475     // Check: use_store must be the only use of def_store.
3476     if (def_store->outcnt() > 1) {
3477       return CFGStatus::Failure;
3478     }
3479     return CFGStatus::SuccessNoRangeCheck;
3480   }
3481 
3482   // Different ctrl -> could have RangeCheck in between.
3483   // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3484   //        2. ctrl separated by RangeCheck.
3485   if (def_store->outcnt() != 2) {
3486     return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3487   }
3488   int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3489   Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3490   if (merge_mem == nullptr ||
3491       merge_mem->outcnt() != 1) {
3492     return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3493   }
3494   if (!ctrl_use->is_IfProj() ||
3495       !ctrl_use->in(0)->is_RangeCheck() ||
3496       ctrl_use->in(0)->outcnt() != 2) {
3497     return CFGStatus::Failure; // Not RangeCheck.
3498   }
3499   IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3500   Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3501   if (trap != merge_mem->unique_out() ||
3502       ctrl_use->in(0)->in(0) != ctrl_def) {
3503     return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3504   }
3505 
3506   return CFGStatus::SuccessWithRangeCheck;
3507 }
3508 
3509 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3510   Status status_use = find_use_store(def_store);
3511   StoreNode* use_store = status_use.found_store();
3512   if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3513     return Status::make_failure();
3514   }
3515   return status_use;
3516 }
3517 
3518 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3519   Status status_def = find_def_store(use_store);
3520   StoreNode* def_store = status_def.found_store();
3521   if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3522     return Status::make_failure();
3523   }
3524   return status_def;
3525 }
3526 
3527 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3528   Status status_use = find_use_store_unidirectional(def_store);
3529 
3530 #ifdef ASSERT
3531   StoreNode* use_store = status_use.found_store();
3532   if (use_store != nullptr) {
3533     Status status_def = find_def_store_unidirectional(use_store);
3534     assert(status_def.found_store() == def_store &&
3535            status_def.found_range_check() == status_use.found_range_check(),
3536            "find_use_store and find_def_store must be symmetric");
3537   }
3538 #endif
3539 
3540   return status_use;
3541 }
3542 
3543 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3544   Status status_def = find_def_store_unidirectional(use_store);
3545 
3546 #ifdef ASSERT
3547   StoreNode* def_store = status_def.found_store();
3548   if (def_store != nullptr) {
3549     Status status_use = find_use_store_unidirectional(def_store);
3550     assert(status_use.found_store() == use_store &&
3551            status_use.found_range_check() == status_def.found_range_check(),
3552            "find_use_store and find_def_store must be symmetric");
3553   }
3554 #endif
3555 
3556   return status_def;
3557 }
3558 
3559 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3560   assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3561 
3562   for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3563     StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3564     if (is_compatible_store(use_store)) {
3565       return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3566     }
3567   }
3568 
3569   return Status::make_failure();
3570 }
3571 
3572 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3573   assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3574 
3575   StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3576   if (!is_compatible_store(def_store)) {
3577     return Status::make_failure();
3578   }
3579 
3580   return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3581 }
3582 
3583 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3584   // The merged store can be at most 8 bytes.
3585   const uint merge_list_max_size = 8 / _store->memory_size();
3586   assert(merge_list_max_size >= 2 &&
3587          merge_list_max_size <= 8 &&
3588          is_power_of_2(merge_list_max_size),
3589          "must be 2, 4 or 8");
3590 
3591   // Traverse up the chain of adjacent def stores.
3592   StoreNode* current = _store;
3593   merge_list.push(current);
3594   while (current != nullptr && merge_list.size() < merge_list_max_size) {
3595     Status status = find_adjacent_def_store(current);
3596     NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3597 
3598     current = status.found_store();
3599     if (current != nullptr) {
3600       merge_list.push(current);
3601 
3602       // We can have at most one RangeCheck.
3603       if (status.found_range_check()) {
3604         NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3605         break;
3606       }
3607     }
3608   }
3609 
3610   NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3611 
3612   // Truncate the merge_list to a power of 2.
3613   const uint pow2size = round_down_power_of_2(merge_list.size());
3614   assert(pow2size >= 2, "must be merging at least 2 stores");
3615   while (merge_list.size() > pow2size) { merge_list.pop(); }
3616 
3617   NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3618 }
3619 
3620 // Merge the input values of the smaller stores to a single larger input value.
3621 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3622   int new_memory_size = _store->memory_size() * merge_list.size();
3623   Node* first = merge_list.at(merge_list.size()-1);
3624   Node* merged_input_value = nullptr;
3625   if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3626     assert(_value_order == ValueOrder::Const, "must match");
3627     // Pattern: [ConI, ConI, ...] -> new constant
3628     jlong con = 0;
3629     jlong bits_per_store = _store->memory_size() * 8;
3630     jlong mask = (((jlong)1) << bits_per_store) - 1;
3631     for (uint i = 0; i < merge_list.size(); i++) {
3632       jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3633 #ifdef VM_LITTLE_ENDIAN
3634       con = con << bits_per_store;
3635       con = con | (mask & con_i);
3636 #else // VM_LITTLE_ENDIAN
3637       con_i = (mask & con_i) << (i * bits_per_store);
3638       con = con | con_i;
3639 #endif // VM_LITTLE_ENDIAN
3640     }
3641     merged_input_value = _phase->longcon(con);
3642   } else {
3643     assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3644     // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3645     //             |                                  |
3646     //           _store                             first
3647     //
3648     Node* hi = _store->in(MemNode::ValueIn);
3649     Node* lo = first->in(MemNode::ValueIn);
3650 #ifndef VM_LITTLE_ENDIAN
3651     // `_store` and `first` are swapped in the diagram above
3652     swap(hi, lo);
3653 #endif // !VM_LITTLE_ENDIAN
3654     if (_value_order == ValueOrder::Reverse) {
3655       swap(hi, lo);
3656     }
3657     Node const* hi_base;
3658     jint hi_shift;
3659     merged_input_value = lo;
3660     bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3661     assert(is_true, "must detect con RShift");
3662     if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3663       // look through
3664       merged_input_value = merged_input_value->in(1);
3665     }
3666     if (merged_input_value != hi_base) {
3667       // merged_input_value is not the base
3668       return nullptr;
3669     }
3670   }
3671 
3672   if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3673     // Example:
3674     //
3675     //   long base = ...;
3676     //   a[0] = (byte)(base >> 0);
3677     //   a[1] = (byte)(base >> 8);
3678     //
3679     merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3680   }
3681 
3682   assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3683          (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3684          "merged_input_value is either int or long, and new_memory_size is small enough");
3685 
3686   if (_value_order == ValueOrder::Reverse) {
3687     assert(_store->memory_size() == 1, "only implemented for bytes");
3688     if (new_memory_size == 8) {
3689       merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3690     } else if (new_memory_size == 4) {
3691       merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3692     } else {
3693       assert(new_memory_size == 2, "sanity check");
3694       merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3695     }
3696   }
3697   return merged_input_value;
3698 }
3699 
3700 //                                                                                                          //
3701 // first_ctrl    first_mem   first_adr                first_ctrl    first_mem         first_adr             //
3702 //  |                |           |                     |                |                 |                 //
3703 //  |                |           |                     |                +---------------+ |                 //
3704 //  |                |           |                     |                |               | |                 //
3705 //  |                | +---------+                     |                | +---------------+                 //
3706 //  |                | |                               |                | |             | |                 //
3707 //  +--------------+ | |  v1                           +------------------------------+ | |  v1             //
3708 //  |              | | |  |                            |                | |           | | |  |              //
3709 // RangeCheck     first_store                         RangeCheck        | |          first_store            //
3710 //  |                |  |                              |                | |                |                //
3711 // last_ctrl         |  +----> unc_trap               last_ctrl         | |                +----> unc_trap  //
3712 //  |                |                       ===>      |                | |                                 //
3713 //  +--------------+ | a2 v2                           |                | |                                 //
3714 //  |              | | |  |                            |                | |                                 //
3715 //  |             second_store                         |                | |                                 //
3716 //  |                |                                 |                | | [v1 v2   ...   vn]              //
3717 // ...              ...                                |                | |         |                       //
3718 //  |                |                                 |                | |         v                       //
3719 //  +--------------+ | an vn                           +--------------+ | | merged_input_value              //
3720 //                 | | |  |                                           | | |  |                              //
3721 //                last_store (= _store)                              merged_store                           //
3722 //                                                                                                          //
3723 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3724   Node* first_store = merge_list.at(merge_list.size()-1);
3725   Node* last_ctrl   = _store->in(MemNode::Control); // after (optional) RangeCheck
3726   Node* first_mem   = first_store->in(MemNode::Memory);
3727   Node* first_adr   = first_store->in(MemNode::Address);
3728 
3729   const TypePtr* new_adr_type = _store->adr_type();
3730 
3731   int new_memory_size = _store->memory_size() * merge_list.size();
3732   BasicType bt = T_ILLEGAL;
3733   switch (new_memory_size) {
3734     case 2: bt = T_SHORT; break;
3735     case 4: bt = T_INT;   break;
3736     case 8: bt = T_LONG;  break;
3737   }
3738 
3739   StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3740                                             new_adr_type, merged_input_value, bt, MemNode::unordered);
3741 
3742   // Marking the store mismatched is sufficient to prevent reordering, since array stores
3743   // are all on the same slice. Hence, we need no barriers.
3744   merged_store->set_mismatched_access();
3745 
3746   // Constants above may now also be be packed -> put candidate on worklist
3747   _phase->is_IterGVN()->_worklist.push(first_mem);
3748 
3749   return merged_store;
3750 }
3751 
3752 #ifndef PRODUCT
3753 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3754   stringStream ss;
3755   ss.print_cr("[TraceMergeStores]: Replace");
3756   for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3757     merge_list.at(i)->dump("\n", false, &ss);
3758   }
3759   ss.print_cr("[TraceMergeStores]: with");
3760   merged_input_value->dump("\n", false, &ss);
3761   merged_store->dump("\n", false, &ss);
3762   tty->print("%s", ss.as_string());
3763 }
3764 #endif
3765 
3766 //------------------------------Ideal------------------------------------------
3767 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3768 // When a store immediately follows a relevant allocation/initialization,
3769 // try to capture it into the initialization, or hoist it above.
3770 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3771   Node* p = MemNode::Ideal_common(phase, can_reshape);
3772   if (p)  return (p == NodeSentinel) ? nullptr : p;
3773 
3774   Node* mem     = in(MemNode::Memory);
3775   Node* address = in(MemNode::Address);
3776   Node* value   = in(MemNode::ValueIn);
3777   // Back-to-back stores to same address?  Fold em up.  Generally
3778   // unsafe if I have intervening uses...
3779   if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3780     Node* st = mem;
3781     // If Store 'st' has more than one use, we cannot fold 'st' away.
3782     // For example, 'st' might be the final state at a conditional
3783     // return.  Or, 'st' might be used by some node which is live at
3784     // the same time 'st' is live, which might be unschedulable.  So,
3785     // require exactly ONE user until such time as we clone 'mem' for
3786     // each of 'mem's uses (thus making the exactly-1-user-rule hold
3787     // true).
3788     while (st->is_Store() && st->outcnt() == 1) {
3789       // Looking at a dead closed cycle of memory?
3790       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3791       assert(Opcode() == st->Opcode() ||
3792              st->Opcode() == Op_StoreVector ||
3793              Opcode() == Op_StoreVector ||
3794              st->Opcode() == Op_StoreVectorScatter ||
3795              Opcode() == Op_StoreVectorScatter ||
3796              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3797              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3798              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3799              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3800              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3801              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3802 
3803       if (st->in(MemNode::Address)->eqv_uncast(address) &&
3804           st->as_Store()->memory_size() <= this->memory_size()) {
3805         Node* use = st->raw_out(0);
3806         if (phase->is_IterGVN()) {
3807           phase->is_IterGVN()->rehash_node_delayed(use);
3808         }
3809         // It's OK to do this in the parser, since DU info is always accurate,
3810         // and the parser always refers to nodes via SafePointNode maps.
3811         use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3812         return this;
3813       }
3814       st = st->in(MemNode::Memory);
3815     }
3816   }
3817 
3818 
3819   // Capture an unaliased, unconditional, simple store into an initializer.
3820   // Or, if it is independent of the allocation, hoist it above the allocation.
3821   if (ReduceFieldZeroing && /*can_reshape &&*/
3822       mem->is_Proj() && mem->in(0)->is_Initialize()) {
3823     InitializeNode* init = mem->in(0)->as_Initialize();
3824     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3825     if (offset > 0) {
3826       Node* moved = init->capture_store(this, offset, phase, can_reshape);
3827       // If the InitializeNode captured me, it made a raw copy of me,
3828       // and I need to disappear.
3829       if (moved != nullptr) {
3830         // %%% hack to ensure that Ideal returns a new node:
3831         mem = MergeMemNode::make(mem);
3832         return mem;             // fold me away
3833       }
3834     }
3835   }
3836 
3837   // Fold reinterpret cast into memory operation:
3838   //    StoreX mem (MoveY2X v) => StoreY mem v
3839   if (value->is_Move()) {
3840     const Type* vt = value->in(1)->bottom_type();
3841     if (has_reinterpret_variant(vt)) {
3842       if (phase->C->post_loop_opts_phase()) {
3843         return convert_to_reinterpret_store(*phase, value->in(1), vt);
3844       } else {
3845         phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3846       }
3847     }
3848   }
3849 
3850   if (MergeStores && UseUnalignedAccesses) {
3851     if (phase->C->merge_stores_phase()) {
3852       MergePrimitiveStores merge(phase, this);
3853       Node* progress = merge.run();
3854       if (progress != nullptr) { return progress; }
3855     } else {
3856       // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3857       // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3858       // the stores, and we merge the wrong sequence of stores.
3859       // Example:
3860       //   StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3861       // Apply MergeStores:
3862       //   StoreI RangeCheck [   StoreL  ] RangeCheck StoreI
3863       // Remove more RangeChecks:
3864       //   StoreI            [   StoreL  ]            StoreI
3865       // But now it would have been better to do this instead:
3866       //   [         StoreL       ] [       StoreL         ]
3867       phase->C->record_for_merge_stores_igvn(this);
3868     }
3869   }
3870 
3871   return nullptr;                  // No further progress
3872 }
3873 
3874 //------------------------------Value-----------------------------------------
3875 const Type* StoreNode::Value(PhaseGVN* phase) const {
3876   // Either input is TOP ==> the result is TOP
3877   const Type *t1 = phase->type( in(MemNode::Memory) );
3878   if( t1 == Type::TOP ) return Type::TOP;
3879   const Type *t2 = phase->type( in(MemNode::Address) );
3880   if( t2 == Type::TOP ) return Type::TOP;
3881   const Type *t3 = phase->type( in(MemNode::ValueIn) );
3882   if( t3 == Type::TOP ) return Type::TOP;
3883   return Type::MEMORY;
3884 }
3885 
3886 //------------------------------Identity---------------------------------------
3887 // Remove redundant stores:
3888 //   Store(m, p, Load(m, p)) changes to m.
3889 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3890 Node* StoreNode::Identity(PhaseGVN* phase) {
3891   Node* mem = in(MemNode::Memory);
3892   Node* adr = in(MemNode::Address);
3893   Node* val = in(MemNode::ValueIn);
3894 
3895   Node* result = this;
3896 
3897   // Load then Store?  Then the Store is useless
3898   if (val->is_Load() &&
3899       val->in(MemNode::Address)->eqv_uncast(adr) &&
3900       val->in(MemNode::Memory )->eqv_uncast(mem) &&
3901       val->as_Load()->store_Opcode() == Opcode()) {
3902     if (!is_StoreVector()) {
3903       result = mem;
3904     } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3905                as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3906       // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3907       result = mem;
3908     }
3909   }
3910 
3911   // Two stores in a row of the same value?
3912   if (result == this &&
3913       mem->is_Store() &&
3914       mem->in(MemNode::Address)->eqv_uncast(adr) &&
3915       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3916       mem->Opcode() == Opcode()) {
3917     if (!is_StoreVector()) {
3918       result = mem;
3919     } else {
3920       const StoreVectorNode* store_vector = as_StoreVector();
3921       const StoreVectorNode* mem_vector = mem->as_StoreVector();
3922       const Node* store_indices = store_vector->indices();
3923       const Node* mem_indices = mem_vector->indices();
3924       const Node* store_mask = store_vector->mask();
3925       const Node* mem_mask = mem_vector->mask();
3926       // Ensure types, indices, and masks match
3927       if (store_vector->vect_type() == mem_vector->vect_type() &&
3928           ((store_indices == nullptr) == (mem_indices == nullptr) &&
3929            (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3930           ((store_mask == nullptr) == (mem_mask == nullptr) &&
3931            (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3932         result = mem;
3933       }
3934     }
3935   }
3936 
3937   // Store of zero anywhere into a freshly-allocated object?
3938   // Then the store is useless.
3939   // (It must already have been captured by the InitializeNode.)
3940   if (result == this && ReduceFieldZeroing) {
3941     // a newly allocated object is already all-zeroes everywhere
3942     if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3943         (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3944       result = mem;
3945     }
3946 
3947     if (result == this && phase->type(val)->is_zero_type()) {
3948       // the store may also apply to zero-bits in an earlier object
3949       Node* prev_mem = find_previous_store(phase);
3950       // Steps (a), (b):  Walk past independent stores to find an exact match.
3951       if (prev_mem != nullptr) {
3952         if (prev_mem->is_top()) {
3953           // find_previous_store returns top when the access is dead
3954           return prev_mem;
3955         }
3956         Node* prev_val = can_see_stored_value(prev_mem, phase);
3957         if (prev_val != nullptr && prev_val == val) {
3958           // prev_val and val might differ by a cast; it would be good
3959           // to keep the more informative of the two.
3960           result = mem;
3961         }
3962       }
3963     }
3964   }
3965 
3966   PhaseIterGVN* igvn = phase->is_IterGVN();
3967   if (result != this && igvn != nullptr) {
3968     MemBarNode* trailing = trailing_membar();
3969     if (trailing != nullptr) {
3970 #ifdef ASSERT
3971       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3972       assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3973 #endif
3974       trailing->remove(igvn);
3975     }
3976   }
3977 
3978   return result;
3979 }
3980 
3981 //------------------------------match_edge-------------------------------------
3982 // Do we Match on this edge index or not?  Match only memory & value
3983 uint StoreNode::match_edge(uint idx) const {
3984   return idx == MemNode::Address || idx == MemNode::ValueIn;
3985 }
3986 
3987 //------------------------------cmp--------------------------------------------
3988 // Do not common stores up together.  They generally have to be split
3989 // back up anyways, so do not bother.
3990 bool StoreNode::cmp( const Node &n ) const {
3991   return (&n == this);          // Always fail except on self
3992 }
3993 
3994 //------------------------------Ideal_masked_input-----------------------------
3995 // Check for a useless mask before a partial-word store
3996 // (StoreB ... (AndI valIn conIa) )
3997 // If (conIa & mask == mask) this simplifies to
3998 // (StoreB ... (valIn) )
3999 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
4000   Node *val = in(MemNode::ValueIn);
4001   if( val->Opcode() == Op_AndI ) {
4002     const TypeInt *t = phase->type( val->in(2) )->isa_int();
4003     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
4004       set_req_X(MemNode::ValueIn, val->in(1), phase);
4005       return this;
4006     }
4007   }
4008   return nullptr;
4009 }
4010 
4011 
4012 //------------------------------Ideal_sign_extended_input----------------------
4013 // Check for useless sign-extension before a partial-word store
4014 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
4015 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
4016 // (StoreB ... (v))
4017 // If (conIL > conIR) under some conditions, it can be simplified into
4018 // (StoreB ... (LShiftI _ v (conIL - conIR)))
4019 // This case happens when the value of the store was itself a left shift, that
4020 // gets merged into the inner left shift of the sign-extension. For instance,
4021 // if we have
4022 // array_of_shorts[0] = (short)(v << 2)
4023 // We get a structure such as:
4024 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
4025 // that is simplified into
4026 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
4027 // It is thus useful to handle cases where conIL > conIR. But this simplification
4028 // does not always hold. Let's see in which cases it's valid.
4029 //
4030 // Let's assume we have the following 32 bits integer v:
4031 // +----------------------------------+
4032 // |             v[0..31]             |
4033 // +----------------------------------+
4034 //  31                               0
4035 // that will be stuffed in 8 bits byte after a shift left and a shift right of
4036 // potentially different magnitudes.
4037 // We denote num_rejected_bits the number of bits of the discarded part. In this
4038 // case, num_rejected_bits == 24.
4039 //
4040 // Statement (proved further below in case analysis):
4041 //   Given:
4042 //   - 0 <= conIL < BitsPerJavaInteger   (no wrap in shift, enforced by maskShiftAmount)
4043 //   - 0 <= conIR < BitsPerJavaInteger   (no wrap in shift, enforced by maskShiftAmount)
4044 //   - conIL >= conIR
4045 //   - num_rejected_bits >= conIR
4046 //   Then this form:
4047 //      (RShiftI _ (LShiftI _ v conIL) conIR)
4048 //   can be replaced with this form:
4049 //      (LShiftI _ v (conIL-conIR))
4050 //
4051 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
4052 //       have to be correct, as the higher bits are rejected / truncated by the store.
4053 //
4054 // The hypotheses
4055 //   0 <= conIL < BitsPerJavaInteger
4056 //   0 <= conIR < BitsPerJavaInteger
4057 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
4058 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
4059 // to v << 1).
4060 //
4061 //
4062 // If you don't like case analysis, jump after the conclusion.
4063 // ### Case 1 : conIL == conIR
4064 // ###### Case 1.1: conIL == conIR == num_rejected_bits
4065 // If we do the shift left then right by 24 bits, we get:
4066 // after: v << 24
4067 // +---------+------------------------+
4068 // | v[0..7] |           0            |
4069 // +---------+------------------------+
4070 //  31     24 23                      0
4071 // after: (v << 24) >> 24
4072 // +------------------------+---------+
4073 // |        sign bit        | v[0..7] |
4074 // +------------------------+---------+
4075 //  31                     8 7        0
4076 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
4077 // result) are the same before and after, so, indeed, simplifying is correct.
4078 
4079 // ###### Case 1.2: conIL == conIR < num_rejected_bits
4080 // If we do the shift left then right by 22 bits, we get:
4081 // after: v << 22
4082 // +---------+------------------------+
4083 // | v[0..9] |           0            |
4084 // +---------+------------------------+
4085 //  31     22 21                      0
4086 // after: (v << 22) >> 22
4087 // +------------------------+---------+
4088 // |        sign bit        | v[0..9] |
4089 // +------------------------+---------+
4090 //  31                    10 9        0
4091 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
4092 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
4093 // still correct.
4094 
4095 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
4096 // If we do the shift left then right by 26 bits, we get:
4097 // after: v << 26
4098 // +---------+------------------------+
4099 // | v[0..5] |           0            |
4100 // +---------+------------------------+
4101 //  31     26 25                      0
4102 // after: (v << 26) >> 26
4103 // +------------------------+---------+
4104 // |        sign bit        | v[0..5] |
4105 // +------------------------+---------+
4106 //  31                     6 5        0
4107 // The non-rejected bits are made of
4108 // - 0-5 => the bits 0 to 5 of v
4109 // - 6-7 => the sign bit of v[0..5] (that is v[5])
4110 // Simplifying this as v is not correct.
4111 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
4112 //
4113 // ### Case 2: conIL > conIR
4114 // ###### Case 2.1: num_rejected_bits == conIR
4115 // We take conIL == 26 for this example.
4116 // after: v << 26
4117 // +---------+------------------------+
4118 // | v[0..5] |           0            |
4119 // +---------+------------------------+
4120 //  31     26 25                      0
4121 // after: (v << 26) >> 24
4122 // +------------------+---------+-----+
4123 // |     sign bit     | v[0..5] |  0  |
4124 // +------------------+---------+-----+
4125 //  31               8 7       2 1   0
4126 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
4127 // The bits 6 and 7 of v have been thrown away after the shift left.
4128 // The simplification is still correct.
4129 //
4130 // ###### Case 2.2: num_rejected_bits > conIR.
4131 // Let's say conIL == 26 and conIR == 22.
4132 // after: v << 26
4133 // +---------+------------------------+
4134 // | v[0..5] |           0            |
4135 // +---------+------------------------+
4136 //  31     26 25                      0
4137 // after: (v << 26) >> 22
4138 // +------------------+---------+-----+
4139 // |     sign bit     | v[0..5] |  0  |
4140 // +------------------+---------+-----+
4141 //  31              10 9       4 3   0
4142 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
4143 // - 0-3 => 0
4144 // - 4-7 => bits 0 to 3 of v
4145 // The simplification is still correct.
4146 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
4147 // matter as they are not in the 8 lower bits: they will be cut out by the store.
4148 //
4149 // ###### But! Case 2.3: num_rejected_bits < conIR.
4150 // Let's see that this case is not as easy to simplify.
4151 // Let's say conIL == 28 and conIR == 26.
4152 // after: v << 28
4153 // +---------+------------------------+
4154 // | v[0..3] |           0            |
4155 // +---------+------------------------+
4156 //  31     28 27                      0
4157 // after: (v << 28) >> 26
4158 // +------------------+---------+-----+
4159 // |     sign bit     | v[0..3] |  0  |
4160 // +------------------+---------+-----+
4161 //  31               6 5       2 1   0
4162 // The non-rejected bits are made of
4163 // - 0-1 => 0
4164 // - 2-5 => the bits 0 to 3 of v
4165 // - 6-7 => the sign bit of v[0..3] (that is v[3])
4166 // Simplifying this as (v << 2) is not correct.
4167 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
4168 //
4169 // ### Conclusion:
4170 // Our hypotheses are indeed sufficient:
4171 //   - 0 <= conIL < BitsPerJavaInteger
4172 //   - 0 <= conIR < BitsPerJavaInteger
4173 //   - conIL >= conIR
4174 //   - num_rejected_bits >= conIR
4175 //
4176 // ### A rationale without case analysis:
4177 // After the shift left, conIL upper  bits of v are discarded and conIL lower bit
4178 // zeroes are added. After the shift right, conIR lower bits of the previous result
4179 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
4180 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
4181 // the point of the simplification is to get an expression of the form
4182 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
4183 // simplification is not correct if conIL < conIR.
4184 //
4185 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
4186 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
4187 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
4188 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
4189 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
4190 //
4191 // We do not treat the case conIL < conIR here since the point of this function is
4192 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
4193 // of treating conIL > conIR comes from the cases where the sign-extended value is
4194 // also left-shift expression. Computing the sign-extension of a right-shift expression
4195 // doesn't yield a situation such as
4196 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
4197 // where conIL < conIR.
4198 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
4199   Node* shr = in(MemNode::ValueIn);
4200   if (shr->Opcode() == Op_RShiftI) {
4201     const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
4202     if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
4203       Node* shl = shr->in(1);
4204       if (shl->Opcode() == Op_LShiftI) {
4205         const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
4206         if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
4207           if (conIL->get_con() == conIR->get_con()) {
4208             set_req_X(MemNode::ValueIn, shl->in(1), phase);
4209             return this;
4210           }
4211           if (conIL->get_con() > conIR->get_con()) {
4212             Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
4213             set_req_X(MemNode::ValueIn, new_shl, phase);
4214             return this;
4215           }
4216         }
4217       }
4218     }
4219   }
4220   return nullptr;
4221 }
4222 
4223 //------------------------------value_never_loaded-----------------------------------
4224 // Determine whether there are any possible loads of the value stored.
4225 // For simplicity, we actually check if there are any loads from the
4226 // address stored to, not just for loads of the value stored by this node.
4227 //
4228 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
4229   Node *adr = in(Address);
4230   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
4231   if (adr_oop == nullptr)
4232     return false;
4233   if (!adr_oop->is_known_instance_field())
4234     return false; // if not a distinct instance, there may be aliases of the address
4235   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
4236     Node *use = adr->fast_out(i);
4237     if (use->is_Load() || use->is_LoadStore()) {
4238       return false;
4239     }
4240   }
4241   return true;
4242 }
4243 
4244 MemBarNode* StoreNode::trailing_membar() const {
4245   if (is_release()) {
4246     MemBarNode* trailing_mb = nullptr;
4247     for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4248       Node* u = fast_out(i);
4249       if (u->is_MemBar()) {
4250         if (u->as_MemBar()->trailing_store()) {
4251           assert(u->Opcode() == Op_MemBarVolatile, "");
4252           assert(trailing_mb == nullptr, "only one");
4253           trailing_mb = u->as_MemBar();
4254 #ifdef ASSERT
4255           Node* leading = u->as_MemBar()->leading_membar();
4256           assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4257           assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
4258           assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
4259 #endif
4260         } else {
4261           assert(u->as_MemBar()->standalone(), "");
4262         }
4263       }
4264     }
4265     return trailing_mb;
4266   }
4267   return nullptr;
4268 }
4269 
4270 
4271 //=============================================================================
4272 //------------------------------Ideal------------------------------------------
4273 // If the store is from an AND mask that leaves the low bits untouched, then
4274 // we can skip the AND operation.  If the store is from a sign-extension
4275 // (a left shift, then right shift) we can skip both.
4276 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
4277   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
4278   if( progress != nullptr ) return progress;
4279 
4280   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
4281   if( progress != nullptr ) return progress;
4282 
4283   // Finally check the default case
4284   return StoreNode::Ideal(phase, can_reshape);
4285 }
4286 
4287 //=============================================================================
4288 //------------------------------Ideal------------------------------------------
4289 // If the store is from an AND mask that leaves the low bits untouched, then
4290 // we can skip the AND operation
4291 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4292   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4293   if( progress != nullptr ) return progress;
4294 
4295   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4296   if( progress != nullptr ) return progress;
4297 
4298   // Finally check the default case
4299   return StoreNode::Ideal(phase, can_reshape);
4300 }
4301 
4302 //=============================================================================
4303 //----------------------------------SCMemProjNode------------------------------
4304 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4305 {
4306   if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4307     return Type::TOP;
4308   }
4309   return bottom_type();
4310 }
4311 
4312 //=============================================================================
4313 //----------------------------------LoadStoreNode------------------------------
4314 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4315   : Node(required),
4316     _type(rt),
4317     _barrier_data(0)
4318 {
4319   init_req(MemNode::Control, c  );
4320   init_req(MemNode::Memory , mem);
4321   init_req(MemNode::Address, adr);
4322   init_req(MemNode::ValueIn, val);
4323   init_class_id(Class_LoadStore);
4324   DEBUG_ONLY(_adr_type = at; adr_type();)
4325 }
4326 
4327 //------------------------------Value-----------------------------------------
4328 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4329   // Either input is TOP ==> the result is TOP
4330   if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4331     return Type::TOP;
4332   }
4333   const Type* t = phase->type(in(MemNode::Memory));
4334   if (t == Type::TOP) {
4335     return Type::TOP;
4336   }
4337   t = phase->type(in(MemNode::Address));
4338   if (t == Type::TOP) {
4339     return Type::TOP;
4340   }
4341   t = phase->type(in(MemNode::ValueIn));
4342   if (t == Type::TOP) {
4343     return Type::TOP;
4344   }
4345   return bottom_type();
4346 }
4347 
4348 const TypePtr* LoadStoreNode::adr_type() const {
4349   const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4350   return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4351 }
4352 
4353 uint LoadStoreNode::ideal_reg() const {
4354   return _type->ideal_reg();
4355 }
4356 
4357 // This method conservatively checks if the result of a LoadStoreNode is
4358 // used, that is, if it returns true, then it is definitely the case that
4359 // the result of the node is not needed.
4360 // For example, GetAndAdd can be matched into a lock_add instead of a
4361 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4362 bool LoadStoreNode::result_not_used() const {
4363   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4364     Node *x = fast_out(i);
4365     if (x->Opcode() == Op_SCMemProj || x->is_ReachabilityFence()) {
4366       continue;
4367     }
4368     if (x->bottom_type() == TypeTuple::MEMBAR &&
4369         !x->is_Call() &&
4370         x->Opcode() != Op_Blackhole) {
4371       continue;
4372     }
4373     return false;
4374   }
4375   return true;
4376 }
4377 
4378 MemBarNode* LoadStoreNode::trailing_membar() const {
4379   MemBarNode* trailing = nullptr;
4380   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4381     Node* u = fast_out(i);
4382     if (u->is_MemBar()) {
4383       if (u->as_MemBar()->trailing_load_store()) {
4384         assert(u->Opcode() == Op_MemBarAcquire, "");
4385         assert(trailing == nullptr, "only one");
4386         trailing = u->as_MemBar();
4387 #ifdef ASSERT
4388         Node* leading = trailing->leading_membar();
4389         assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4390         assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4391         assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4392 #endif
4393       } else {
4394         assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4395       }
4396     }
4397   }
4398 
4399   return trailing;
4400 }
4401 
4402 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4403 
4404 //=============================================================================
4405 //----------------------------------LoadStoreConditionalNode--------------------
4406 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4407   init_req(ExpectedIn, ex );
4408 }
4409 
4410 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4411   // Either input is TOP ==> the result is TOP
4412   const Type* t = phase->type(in(ExpectedIn));
4413   if (t == Type::TOP) {
4414     return Type::TOP;
4415   }
4416   return LoadStoreNode::Value(phase);
4417 }
4418 
4419 //=============================================================================
4420 //-------------------------------adr_type--------------------------------------
4421 const TypePtr* ClearArrayNode::adr_type() const {
4422   Node *adr = in(3);
4423   if (adr == nullptr)  return nullptr; // node is dead
4424   return MemNode::calculate_adr_type(adr->bottom_type());
4425 }
4426 
4427 //------------------------------match_edge-------------------------------------
4428 // Do we Match on this edge index or not?  Do not match memory
4429 uint ClearArrayNode::match_edge(uint idx) const {
4430   return idx > 1;
4431 }
4432 
4433 //------------------------------Identity---------------------------------------
4434 // Clearing a zero length array does nothing
4435 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4436   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
4437 }
4438 
4439 //------------------------------Idealize---------------------------------------
4440 // Clearing a short array is faster with stores
4441 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4442   // Already know this is a large node, do not try to ideal it
4443   if (_is_large) return nullptr;
4444 
4445   const int unit = BytesPerLong;
4446   const TypeX* t = phase->type(in(2))->isa_intptr_t();
4447   if (!t)  return nullptr;
4448   if (!t->is_con())  return nullptr;
4449   intptr_t raw_count = t->get_con();
4450   intptr_t size = raw_count;
4451   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4452   // Clearing nothing uses the Identity call.
4453   // Negative clears are possible on dead ClearArrays
4454   // (see jck test stmt114.stmt11402.val).
4455   if (size <= 0 || size % unit != 0)  return nullptr;
4456   intptr_t count = size / unit;
4457   // Length too long; communicate this to matchers and assemblers.
4458   // Assemblers are responsible to produce fast hardware clears for it.
4459   if (size > InitArrayShortSize) {
4460     return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4461   } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4462     return nullptr;
4463   }
4464   if (!IdealizeClearArrayNode) return nullptr;
4465   Node *mem = in(1);
4466   if( phase->type(mem)==Type::TOP ) return nullptr;
4467   Node *adr = in(3);
4468   const Type* at = phase->type(adr);
4469   if( at==Type::TOP ) return nullptr;
4470   const TypePtr* atp = at->isa_ptr();
4471   // adjust atp to be the correct array element address type
4472   if (atp == nullptr)  atp = TypePtr::BOTTOM;
4473   else              atp = atp->add_offset(Type::OffsetBot);
4474   // Get base for derived pointer purposes
4475   if( adr->Opcode() != Op_AddP ) Unimplemented();
4476   Node *base = adr->in(1);
4477 
4478   Node *val = in(4);
4479   Node *off  = phase->MakeConX(BytesPerLong);
4480   mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4481   count--;
4482   while (count--) {
4483     mem = phase->transform(mem);
4484     adr = phase->transform(AddPNode::make_with_base(base,adr,off));
4485     mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4486   }
4487   return mem;
4488 }
4489 
4490 //----------------------------step_through----------------------------------
4491 // Return allocation input memory edge if it is different instance
4492 // or itself if it is the one we are looking for.
4493 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4494   Node* n = *np;
4495   assert(n->is_ClearArray(), "sanity");
4496   intptr_t offset;
4497   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4498   // This method is called only before Allocate nodes are expanded
4499   // during macro nodes expansion. Before that ClearArray nodes are
4500   // only generated in PhaseMacroExpand::generate_arraycopy() (before
4501   // Allocate nodes are expanded) which follows allocations.
4502   assert(alloc != nullptr, "should have allocation");
4503   if (alloc->_idx == instance_id) {
4504     // Can not bypass initialization of the instance we are looking for.
4505     return false;
4506   }
4507   // Otherwise skip it.
4508   InitializeNode* init = alloc->initialization();
4509   if (init != nullptr)
4510     *np = init->in(TypeFunc::Memory);
4511   else
4512     *np = alloc->in(TypeFunc::Memory);
4513   return true;
4514 }
4515 
4516 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4517   Node* base = dest;
4518   if (raw_base) {
4519     // May be called as part of the initialization of a just allocated object
4520     base = phase->C->top();
4521   }
4522   return phase->transform(AddPNode::make_with_base(base, dest, offset));
4523 }
4524 
4525 //----------------------------clear_memory-------------------------------------
4526 // Generate code to initialize object storage to zero.
4527 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4528                                    Node* val,
4529                                    Node* raw_val,
4530                                    intptr_t start_offset,
4531                                    Node* end_offset,
4532                                    bool raw_base,
4533                                    PhaseGVN* phase) {
4534   intptr_t offset = start_offset;
4535 
4536   int unit = BytesPerLong;
4537   if ((offset % unit) != 0) {
4538     Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4539     const TypePtr* atp = TypeRawPtr::BOTTOM;
4540     if (val != nullptr) {
4541       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4542       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4543     } else {
4544       assert(raw_val == nullptr, "val may not be null");
4545       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4546     }
4547     mem = phase->transform(mem);
4548     offset += BytesPerInt;
4549   }
4550   assert((offset % unit) == 0, "");
4551 
4552   // Initialize the remaining stuff, if any, with a ClearArray.
4553   return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, raw_base, phase);
4554 }
4555 
4556 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4557                                    Node* raw_val,
4558                                    Node* start_offset,
4559                                    Node* end_offset,
4560                                    bool raw_base,
4561                                    PhaseGVN* phase) {
4562   if (start_offset == end_offset) {
4563     // nothing to do
4564     return mem;
4565   }
4566 
4567   int unit = BytesPerLong;
4568   Node* zbase = start_offset;
4569   Node* zend  = end_offset;
4570 
4571   // Scale to the unit required by the CPU:
4572   if (!Matcher::init_array_count_is_in_bytes) {
4573     Node* shift = phase->intcon(exact_log2(unit));
4574     zbase = phase->transform(new URShiftXNode(zbase, shift) );
4575     zend  = phase->transform(new URShiftXNode(zend,  shift) );
4576   }
4577 
4578   // Bulk clear double-words
4579   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4580   Node* adr = make_address(dest, start_offset, raw_base, phase);
4581   if (raw_val == nullptr) {
4582     raw_val = phase->MakeConX(0);
4583   }
4584   mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4585   return phase->transform(mem);
4586 }
4587 
4588 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4589                                    Node* val,
4590                                    Node* raw_val,
4591                                    intptr_t start_offset,
4592                                    intptr_t end_offset,
4593                                    bool raw_base,
4594                                    PhaseGVN* phase) {
4595   if (start_offset == end_offset) {
4596     // nothing to do
4597     return mem;
4598   }
4599 
4600   assert((end_offset % BytesPerInt) == 0, "odd end offset");
4601   intptr_t done_offset = end_offset;
4602   if ((done_offset % BytesPerLong) != 0) {
4603     done_offset -= BytesPerInt;
4604   }
4605   if (done_offset > start_offset) {
4606     mem = clear_memory(ctl, mem, dest, val, raw_val,
4607                        start_offset, phase->MakeConX(done_offset), raw_base, phase);
4608   }
4609   if (done_offset < end_offset) { // emit the final 32-bit store
4610     Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4611     const TypePtr* atp = TypeRawPtr::BOTTOM;
4612     if (val != nullptr) {
4613       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4614       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4615     } else {
4616       assert(raw_val == nullptr, "val may not be null");
4617       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4618     }
4619     mem = phase->transform(mem);
4620     done_offset += BytesPerInt;
4621   }
4622   assert(done_offset == end_offset, "");
4623   return mem;
4624 }
4625 
4626 //=============================================================================
4627 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4628   : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4629     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4630 #ifdef ASSERT
4631   , _pair_idx(0)
4632 #endif
4633 {
4634   init_class_id(Class_MemBar);
4635   Node* top = C->top();
4636   init_req(TypeFunc::I_O,top);
4637   init_req(TypeFunc::FramePtr,top);
4638   init_req(TypeFunc::ReturnAdr,top);
4639   if (precedent != nullptr)
4640     init_req(TypeFunc::Parms, precedent);
4641 }
4642 
4643 //------------------------------cmp--------------------------------------------
4644 uint MemBarNode::hash() const { return NO_HASH; }
4645 bool MemBarNode::cmp( const Node &n ) const {
4646   return (&n == this);          // Always fail except on self
4647 }
4648 
4649 //------------------------------make-------------------------------------------
4650 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4651   switch (opcode) {
4652   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
4653   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
4654   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
4655   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
4656   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
4657   case Op_StoreStoreFence:   return new StoreStoreFenceNode(C, atp, pn);
4658   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4659   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4660   case Op_MemBarStoreLoad:   return new MemBarStoreLoadNode(C, atp, pn);
4661   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
4662   case Op_MemBarFull:        return new MemBarFullNode(C, atp, pn);
4663   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
4664   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
4665   case Op_Initialize:        return new InitializeNode(C, atp, pn);
4666   default: ShouldNotReachHere(); return nullptr;
4667   }
4668 }
4669 
4670 void MemBarNode::remove(PhaseIterGVN *igvn) {
4671   assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4672   if (trailing_store() || trailing_load_store()) {
4673     MemBarNode* leading = leading_membar();
4674     if (leading != nullptr) {
4675       assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4676       leading->remove(igvn);
4677     }
4678   }
4679   if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4680     igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4681   }
4682   if (is_Initialize()) {
4683     as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4684   } else {
4685     if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4686       igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4687     }
4688   }
4689 }
4690 
4691 //------------------------------Ideal------------------------------------------
4692 // Return a node which is more "ideal" than the current node.  Strip out
4693 // control copies
4694 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4695   if (remove_dead_region(phase, can_reshape)) return this;
4696   // Don't bother trying to transform a dead node
4697   if (in(0) && in(0)->is_top()) {
4698     return nullptr;
4699   }
4700 
4701   bool progress = false;
4702   // Eliminate volatile MemBars for scalar replaced objects.
4703   if (can_reshape && req() == (Precedent+1)) {
4704     bool eliminate = false;
4705     int opc = Opcode();
4706     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4707       // Volatile field loads and stores.
4708       Node* my_mem = in(MemBarNode::Precedent);
4709       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4710       if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4711         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4712         // replace this Precedent (decodeN) with the Load instead.
4713         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
4714           Node* load_node = my_mem->in(1);
4715           set_req(MemBarNode::Precedent, load_node);
4716           phase->is_IterGVN()->_worklist.push(my_mem);
4717           my_mem = load_node;
4718         } else {
4719           assert(my_mem->unique_out() == this, "sanity");
4720           assert(!trailing_load_store(), "load store node can't be eliminated");
4721           del_req(Precedent);
4722           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4723           my_mem = nullptr;
4724         }
4725         progress = true;
4726       }
4727       if (my_mem != nullptr && my_mem->is_Mem()) {
4728         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4729         // Check for scalar replaced object reference.
4730         if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4731             t_oop->offset() != Type::OffsetBot &&
4732             t_oop->offset() != Type::OffsetTop) {
4733           eliminate = true;
4734         }
4735       }
4736     } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4737       // Final field stores.
4738       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4739       if ((alloc != nullptr) && alloc->is_Allocate() &&
4740           alloc->as_Allocate()->does_not_escape_thread()) {
4741         // The allocated object does not escape.
4742         eliminate = true;
4743       }
4744     }
4745     if (eliminate) {
4746       // Replace MemBar projections by its inputs.
4747       PhaseIterGVN* igvn = phase->is_IterGVN();
4748       remove(igvn);
4749       // Must return either the original node (now dead) or a new node
4750       // (Do not return a top here, since that would break the uniqueness of top.)
4751       return new ConINode(TypeInt::ZERO);
4752     }
4753   }
4754   return progress ? this : nullptr;
4755 }
4756 
4757 //------------------------------Value------------------------------------------
4758 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4759   if( !in(0) ) return Type::TOP;
4760   if( phase->type(in(0)) == Type::TOP )
4761     return Type::TOP;
4762   return TypeTuple::MEMBAR;
4763 }
4764 
4765 //------------------------------match------------------------------------------
4766 // Construct projections for memory.
4767 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4768   switch (proj->_con) {
4769   case TypeFunc::Control:
4770   case TypeFunc::Memory:
4771     return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4772   }
4773   ShouldNotReachHere();
4774   return nullptr;
4775 }
4776 
4777 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4778   trailing->_kind = TrailingStore;
4779   leading->_kind = LeadingStore;
4780 #ifdef ASSERT
4781   trailing->_pair_idx = leading->_idx;
4782   leading->_pair_idx = leading->_idx;
4783 #endif
4784 }
4785 
4786 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4787   trailing->_kind = TrailingLoadStore;
4788   leading->_kind = LeadingLoadStore;
4789 #ifdef ASSERT
4790   trailing->_pair_idx = leading->_idx;
4791   leading->_pair_idx = leading->_idx;
4792 #endif
4793 }
4794 
4795 MemBarNode* MemBarNode::trailing_membar() const {
4796   ResourceMark rm;
4797   Node* trailing = (Node*)this;
4798   VectorSet seen;
4799   Node_Stack multis(0);
4800   do {
4801     Node* c = trailing;
4802     uint i = 0;
4803     do {
4804       trailing = nullptr;
4805       for (; i < c->outcnt(); i++) {
4806         Node* next = c->raw_out(i);
4807         if (next != c && next->is_CFG()) {
4808           if (c->is_MultiBranch()) {
4809             if (multis.node() == c) {
4810               multis.set_index(i+1);
4811             } else {
4812               multis.push(c, i+1);
4813             }
4814           }
4815           trailing = next;
4816           break;
4817         }
4818       }
4819       if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4820         break;
4821       }
4822       while (multis.size() > 0) {
4823         c = multis.node();
4824         i = multis.index();
4825         if (i < c->req()) {
4826           break;
4827         }
4828         multis.pop();
4829       }
4830     } while (multis.size() > 0);
4831   } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4832 
4833   MemBarNode* mb = trailing->as_MemBar();
4834   assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4835          (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4836   assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4837   return mb;
4838 }
4839 
4840 MemBarNode* MemBarNode::leading_membar() const {
4841   ResourceMark rm;
4842   VectorSet seen;
4843   Node_Stack regions(0);
4844   Node* leading = in(0);
4845   while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4846     while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4847       leading = nullptr;
4848       while (regions.size() > 0 && leading == nullptr) {
4849         Node* r = regions.node();
4850         uint i = regions.index();
4851         if (i < r->req()) {
4852           leading = r->in(i);
4853           regions.set_index(i+1);
4854         } else {
4855           regions.pop();
4856         }
4857       }
4858       if (leading == nullptr) {
4859         assert(regions.size() == 0, "all paths should have been tried");
4860         return nullptr;
4861       }
4862     }
4863     if (leading->is_Region()) {
4864       regions.push(leading, 2);
4865       leading = leading->in(1);
4866     } else {
4867       leading = leading->in(0);
4868     }
4869   }
4870 #ifdef ASSERT
4871   Unique_Node_List wq;
4872   wq.push((Node*)this);
4873   uint found = 0;
4874   for (uint i = 0; i < wq.size(); i++) {
4875     Node* n = wq.at(i);
4876     if (n->is_Region()) {
4877       for (uint j = 1; j < n->req(); j++) {
4878         Node* in = n->in(j);
4879         if (in != nullptr && !in->is_top()) {
4880           wq.push(in);
4881         }
4882       }
4883     } else {
4884       if (n->is_MemBar() && n->as_MemBar()->leading()) {
4885         assert(n == leading, "consistency check failed");
4886         found++;
4887       } else {
4888         Node* in = n->in(0);
4889         if (in != nullptr && !in->is_top()) {
4890           wq.push(in);
4891         }
4892       }
4893     }
4894   }
4895   assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4896 #endif
4897   if (leading == nullptr) {
4898     return nullptr;
4899   }
4900   MemBarNode* mb = leading->as_MemBar();
4901   assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4902          (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4903   assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4904   return mb;
4905 }
4906 
4907 
4908 //===========================InitializeNode====================================
4909 // SUMMARY:
4910 // This node acts as a memory barrier on raw memory, after some raw stores.
4911 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4912 // The Initialize can 'capture' suitably constrained stores as raw inits.
4913 // It can coalesce related raw stores into larger units (called 'tiles').
4914 // It can avoid zeroing new storage for memory units which have raw inits.
4915 // At macro-expansion, it is marked 'complete', and does not optimize further.
4916 //
4917 // EXAMPLE:
4918 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4919 //   ctl = incoming control; mem* = incoming memory
4920 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
4921 // First allocate uninitialized memory and fill in the header:
4922 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
4923 //   ctl := alloc.Control; mem* := alloc.Memory*
4924 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
4925 // Then initialize to zero the non-header parts of the raw memory block:
4926 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4927 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4928 // After the initialize node executes, the object is ready for service:
4929 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4930 // Suppose its body is immediately initialized as {1,2}:
4931 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4932 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
4933 //   mem.SLICE(#short[*]) := store2
4934 //
4935 // DETAILS:
4936 // An InitializeNode collects and isolates object initialization after
4937 // an AllocateNode and before the next possible safepoint.  As a
4938 // memory barrier (MemBarNode), it keeps critical stores from drifting
4939 // down past any safepoint or any publication of the allocation.
4940 // Before this barrier, a newly-allocated object may have uninitialized bits.
4941 // After this barrier, it may be treated as a real oop, and GC is allowed.
4942 //
4943 // The semantics of the InitializeNode include an implicit zeroing of
4944 // the new object from object header to the end of the object.
4945 // (The object header and end are determined by the AllocateNode.)
4946 //
4947 // Certain stores may be added as direct inputs to the InitializeNode.
4948 // These stores must update raw memory, and they must be to addresses
4949 // derived from the raw address produced by AllocateNode, and with
4950 // a constant offset.  They must be ordered by increasing offset.
4951 // The first one is at in(RawStores), the last at in(req()-1).
4952 // Unlike most memory operations, they are not linked in a chain,
4953 // but are displayed in parallel as users of the rawmem output of
4954 // the allocation.
4955 //
4956 // (See comments in InitializeNode::capture_store, which continue
4957 // the example given above.)
4958 //
4959 // When the associated Allocate is macro-expanded, the InitializeNode
4960 // may be rewritten to optimize collected stores.  A ClearArrayNode
4961 // may also be created at that point to represent any required zeroing.
4962 // The InitializeNode is then marked 'complete', prohibiting further
4963 // capturing of nearby memory operations.
4964 //
4965 // During macro-expansion, all captured initializations which store
4966 // constant values of 32 bits or smaller are coalesced (if advantageous)
4967 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
4968 // initialized in fewer memory operations.  Memory words which are
4969 // covered by neither tiles nor non-constant stores are pre-zeroed
4970 // by explicit stores of zero.  (The code shape happens to do all
4971 // zeroing first, then all other stores, with both sequences occurring
4972 // in order of ascending offsets.)
4973 //
4974 // Alternatively, code may be inserted between an AllocateNode and its
4975 // InitializeNode, to perform arbitrary initialization of the new object.
4976 // E.g., the object copying intrinsics insert complex data transfers here.
4977 // The initialization must then be marked as 'complete' disable the
4978 // built-in zeroing semantics and the collection of initializing stores.
4979 //
4980 // While an InitializeNode is incomplete, reads from the memory state
4981 // produced by it are optimizable if they match the control edge and
4982 // new oop address associated with the allocation/initialization.
4983 // They return a stored value (if the offset matches) or else zero.
4984 // A write to the memory state, if it matches control and address,
4985 // and if it is to a constant offset, may be 'captured' by the
4986 // InitializeNode.  It is cloned as a raw memory operation and rewired
4987 // inside the initialization, to the raw oop produced by the allocation.
4988 // Operations on addresses which are provably distinct (e.g., to
4989 // other AllocateNodes) are allowed to bypass the initialization.
4990 //
4991 // The effect of all this is to consolidate object initialization
4992 // (both arrays and non-arrays, both piecewise and bulk) into a
4993 // single location, where it can be optimized as a unit.
4994 //
4995 // Only stores with an offset less than TrackedInitializationLimit words
4996 // will be considered for capture by an InitializeNode.  This puts a
4997 // reasonable limit on the complexity of optimized initializations.
4998 
4999 //---------------------------InitializeNode------------------------------------
5000 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
5001   : MemBarNode(C, adr_type, rawoop),
5002     _is_complete(Incomplete), _does_not_escape(false)
5003 {
5004   init_class_id(Class_Initialize);
5005 
5006   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
5007   assert(in(RawAddress) == rawoop, "proper init");
5008   // Note:  allocation() can be null, for secondary initialization barriers
5009 }
5010 
5011 // Since this node is not matched, it will be processed by the
5012 // register allocator.  Declare that there are no constraints
5013 // on the allocation of the RawAddress edge.
5014 const RegMask &InitializeNode::in_RegMask(uint idx) const {
5015   // This edge should be set to top, by the set_complete.  But be conservative.
5016   if (idx == InitializeNode::RawAddress)
5017     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
5018   return RegMask::EMPTY;
5019 }
5020 
5021 Node* InitializeNode::memory(uint alias_idx) {
5022   Node* mem = in(Memory);
5023   if (mem->is_MergeMem()) {
5024     return mem->as_MergeMem()->memory_at(alias_idx);
5025   } else {
5026     // incoming raw memory is not split
5027     return mem;
5028   }
5029 }
5030 
5031 bool InitializeNode::is_non_zero() {
5032   if (is_complete())  return false;
5033   remove_extra_zeroes();
5034   return (req() > RawStores);
5035 }
5036 
5037 void InitializeNode::set_complete(PhaseGVN* phase) {
5038   assert(!is_complete(), "caller responsibility");
5039   _is_complete = Complete;
5040 
5041   // After this node is complete, it contains a bunch of
5042   // raw-memory initializations.  There is no need for
5043   // it to have anything to do with non-raw memory effects.
5044   // Therefore, tell all non-raw users to re-optimize themselves,
5045   // after skipping the memory effects of this initialization.
5046   PhaseIterGVN* igvn = phase->is_IterGVN();
5047   if (igvn)  igvn->add_users_to_worklist(this);
5048 }
5049 
5050 // convenience function
5051 // return false if the init contains any stores already
5052 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
5053   InitializeNode* init = initialization();
5054   if (init == nullptr || init->is_complete()) {
5055     return false;
5056   }
5057   init->remove_extra_zeroes();
5058   // for now, if this allocation has already collected any inits, bail:
5059   if (init->is_non_zero())  return false;
5060   init->set_complete(phase);
5061   return true;
5062 }
5063 
5064 void InitializeNode::remove_extra_zeroes() {
5065   if (req() == RawStores)  return;
5066   Node* zmem = zero_memory();
5067   uint fill = RawStores;
5068   for (uint i = fill; i < req(); i++) {
5069     Node* n = in(i);
5070     if (n->is_top() || n == zmem)  continue;  // skip
5071     if (fill < i)  set_req(fill, n);          // compact
5072     ++fill;
5073   }
5074   // delete any empty spaces created:
5075   while (fill < req()) {
5076     del_req(fill);
5077   }
5078 }
5079 
5080 // Helper for remembering which stores go with which offsets.
5081 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
5082   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
5083   intptr_t offset = -1;
5084   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
5085                                                phase, offset);
5086   if (base == nullptr)  return -1;  // something is dead,
5087   if (offset < 0)       return -1;  //        dead, dead
5088   return offset;
5089 }
5090 
5091 // Helper for proving that an initialization expression is
5092 // "simple enough" to be folded into an object initialization.
5093 // Attempts to prove that a store's initial value 'n' can be captured
5094 // within the initialization without creating a vicious cycle, such as:
5095 //     { Foo p = new Foo(); p.next = p; }
5096 // True for constants and parameters and small combinations thereof.
5097 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
5098   ResourceMark rm;
5099   Unique_Node_List worklist;
5100   worklist.push(value);
5101 
5102   uint complexity_limit = 20;
5103   for (uint j = 0; j < worklist.size(); j++) {
5104     if (j >= complexity_limit) {
5105       return false;  // Bail out if processed too many nodes
5106     }
5107 
5108     Node* n = worklist.at(j);
5109     if (n == nullptr)   continue;   // (can this really happen?)
5110     if (n->is_Proj())   n = n->in(0);
5111     if (n == this)      return false;  // found a cycle
5112     if (n->is_Con())    continue;
5113     if (n->is_Start())  continue;   // params, etc., are OK
5114     if (n->is_Root())   continue;   // even better
5115 
5116     // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
5117     if (n->is_CFG() && phase->is_dominator(n, allocation())) {
5118       continue;
5119     }
5120 
5121     Node* ctl = n->in(0);
5122     if (ctl != nullptr && !ctl->is_top()) {
5123       if (ctl->is_Proj())  ctl = ctl->in(0);
5124       if (ctl == this)  return false;
5125 
5126       // If we already know that the enclosing memory op is pinned right after
5127       // the init, then any control flow that the store has picked up
5128       // must have preceded the init, or else be equal to the init.
5129       // Even after loop optimizations (which might change control edges)
5130       // a store is never pinned *before* the availability of its inputs.
5131       if (!MemNode::all_controls_dominate(n, this, phase)) {
5132         return false;                  // failed to prove a good control
5133       }
5134     }
5135 
5136     // Check data edges for possible dependencies on 'this'.
5137     for (uint i = 1; i < n->req(); i++) {
5138       Node* m = n->in(i);
5139       if (m == nullptr || m == n || m->is_top())  continue;
5140 
5141       // Only process data inputs once
5142       worklist.push(m);
5143     }
5144   }
5145 
5146   return true;
5147 }
5148 
5149 // Here are all the checks a Store must pass before it can be moved into
5150 // an initialization.  Returns zero if a check fails.
5151 // On success, returns the (constant) offset to which the store applies,
5152 // within the initialized memory.
5153 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
5154   const int FAIL = 0;
5155   if (st->req() != MemNode::ValueIn + 1)
5156     return FAIL;                // an inscrutable StoreNode (card mark?)
5157   Node* ctl = st->in(MemNode::Control);
5158   if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
5159     return FAIL;                // must be unconditional after the initialization
5160   Node* mem = st->in(MemNode::Memory);
5161   if (!(mem->is_Proj() && mem->in(0) == this))
5162     return FAIL;                // must not be preceded by other stores
5163   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5164   if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
5165       !bs->can_initialize_object(st)) {
5166     return FAIL;
5167   }
5168   Node* adr = st->in(MemNode::Address);
5169   intptr_t offset;
5170   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
5171   if (alloc == nullptr)
5172     return FAIL;                // inscrutable address
5173   if (alloc != allocation())
5174     return FAIL;                // wrong allocation!  (store needs to float up)
5175   int size_in_bytes = st->memory_size();
5176   if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
5177     return FAIL;                // mismatched access
5178   }
5179   Node* val = st->in(MemNode::ValueIn);
5180 
5181   if (!detect_init_independence(val, phase))
5182     return FAIL;                // stored value must be 'simple enough'
5183 
5184   // The Store can be captured only if nothing after the allocation
5185   // and before the Store is using the memory location that the store
5186   // overwrites.
5187   bool failed = false;
5188   // If is_complete_with_arraycopy() is true the shape of the graph is
5189   // well defined and is safe so no need for extra checks.
5190   if (!is_complete_with_arraycopy()) {
5191     // We are going to look at each use of the memory state following
5192     // the allocation to make sure nothing reads the memory that the
5193     // Store writes.
5194     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
5195     int alias_idx = phase->C->get_alias_index(t_adr);
5196     ResourceMark rm;
5197     Unique_Node_List mems;
5198     mems.push(mem);
5199     Node* unique_merge = nullptr;
5200     for (uint next = 0; next < mems.size(); ++next) {
5201       Node *m  = mems.at(next);
5202       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5203         Node *n = m->fast_out(j);
5204         if (n->outcnt() == 0) {
5205           continue;
5206         }
5207         if (n == st) {
5208           continue;
5209         } else if (n->in(0) != nullptr && n->in(0) != ctl) {
5210           // If the control of this use is different from the control
5211           // of the Store which is right after the InitializeNode then
5212           // this node cannot be between the InitializeNode and the
5213           // Store.
5214           continue;
5215         } else if (n->is_MergeMem()) {
5216           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
5217             // We can hit a MergeMemNode (that will likely go away
5218             // later) that is a direct use of the memory state
5219             // following the InitializeNode on the same slice as the
5220             // store node that we'd like to capture. We need to check
5221             // the uses of the MergeMemNode.
5222             mems.push(n);
5223           }
5224         } else if (n->is_Mem()) {
5225           Node* other_adr = n->in(MemNode::Address);
5226           if (other_adr == adr) {
5227             failed = true;
5228             break;
5229           } else {
5230             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
5231             if (other_t_adr != nullptr) {
5232               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
5233               if (other_alias_idx == alias_idx) {
5234                 // A load from the same memory slice as the store right
5235                 // after the InitializeNode. We check the control of the
5236                 // object/array that is loaded from. If it's the same as
5237                 // the store control then we cannot capture the store.
5238                 assert(!n->is_Store(), "2 stores to same slice on same control?");
5239                 Node* base = other_adr;
5240                 if (base->is_Phi()) {
5241                   // In rare case, base may be a PhiNode and it may read
5242                   // the same memory slice between InitializeNode and store.
5243                   failed = true;
5244                   break;
5245                 }
5246                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
5247                 base = base->in(AddPNode::Base);
5248                 if (base != nullptr) {
5249                   base = base->uncast();
5250                   if (base->is_Proj() && base->in(0) == alloc) {
5251                     failed = true;
5252                     break;
5253                   }
5254                 }
5255               }
5256             }
5257           }
5258         } else {
5259           failed = true;
5260           break;
5261         }
5262       }
5263     }
5264   }
5265   if (failed) {
5266     if (!can_reshape) {
5267       // We decided we couldn't capture the store during parsing. We
5268       // should try again during the next IGVN once the graph is
5269       // cleaner.
5270       phase->C->record_for_igvn(st);
5271     }
5272     return FAIL;
5273   }
5274 
5275   return offset;                // success
5276 }
5277 
5278 // Find the captured store in(i) which corresponds to the range
5279 // [start..start+size) in the initialized object.
5280 // If there is one, return its index i.  If there isn't, return the
5281 // negative of the index where it should be inserted.
5282 // Return 0 if the queried range overlaps an initialization boundary
5283 // or if dead code is encountered.
5284 // If size_in_bytes is zero, do not bother with overlap checks.
5285 int InitializeNode::captured_store_insertion_point(intptr_t start,
5286                                                    int size_in_bytes,
5287                                                    PhaseValues* phase) {
5288   const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
5289 
5290   if (is_complete())
5291     return FAIL;                // arraycopy got here first; punt
5292 
5293   assert(allocation() != nullptr, "must be present");
5294 
5295   // no negatives, no header fields:
5296   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
5297 
5298   // after a certain size, we bail out on tracking all the stores:
5299   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5300   if (start >= ti_limit)  return FAIL;
5301 
5302   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
5303     if (i >= limit)  return -(int)i; // not found; here is where to put it
5304 
5305     Node*    st     = in(i);
5306     intptr_t st_off = get_store_offset(st, phase);
5307     if (st_off < 0) {
5308       if (st != zero_memory()) {
5309         return FAIL;            // bail out if there is dead garbage
5310       }
5311     } else if (st_off > start) {
5312       // ...we are done, since stores are ordered
5313       if (st_off < start + size_in_bytes) {
5314         return FAIL;            // the next store overlaps
5315       }
5316       return -(int)i;           // not found; here is where to put it
5317     } else if (st_off < start) {
5318       assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5319       if (size_in_bytes != 0 &&
5320           start < st_off + MAX_STORE &&
5321           start < st_off + st->as_Store()->memory_size()) {
5322         return FAIL;            // the previous store overlaps
5323       }
5324     } else {
5325       if (size_in_bytes != 0 &&
5326           st->as_Store()->memory_size() != size_in_bytes) {
5327         return FAIL;            // mismatched store size
5328       }
5329       return i;
5330     }
5331 
5332     ++i;
5333   }
5334 }
5335 
5336 // Look for a captured store which initializes at the offset 'start'
5337 // with the given size.  If there is no such store, and no other
5338 // initialization interferes, then return zero_memory (the memory
5339 // projection of the AllocateNode).
5340 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5341                                           PhaseValues* phase) {
5342   assert(stores_are_sane(phase), "");
5343   int i = captured_store_insertion_point(start, size_in_bytes, phase);
5344   if (i == 0) {
5345     return nullptr;              // something is dead
5346   } else if (i < 0) {
5347     return zero_memory();       // just primordial zero bits here
5348   } else {
5349     Node* st = in(i);           // here is the store at this position
5350     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5351     return st;
5352   }
5353 }
5354 
5355 // Create, as a raw pointer, an address within my new object at 'offset'.
5356 Node* InitializeNode::make_raw_address(intptr_t offset,
5357                                        PhaseGVN* phase) {
5358   Node* addr = in(RawAddress);
5359   if (offset != 0) {
5360     Compile* C = phase->C;
5361     addr = phase->transform(AddPNode::make_off_heap(addr, phase->MakeConX(offset)));
5362   }
5363   return addr;
5364 }
5365 
5366 // Clone the given store, converting it into a raw store
5367 // initializing a field or element of my new object.
5368 // Caller is responsible for retiring the original store,
5369 // with subsume_node or the like.
5370 //
5371 // From the example above InitializeNode::InitializeNode,
5372 // here are the old stores to be captured:
5373 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5374 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
5375 //
5376 // Here is the changed code; note the extra edges on init:
5377 //   alloc = (Allocate ...)
5378 //   rawoop = alloc.RawAddress
5379 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5380 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5381 //   init = (Initialize alloc.Control alloc.Memory rawoop
5382 //                      rawstore1 rawstore2)
5383 //
5384 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5385                                     PhaseGVN* phase, bool can_reshape) {
5386   assert(stores_are_sane(phase), "");
5387 
5388   if (start < 0)  return nullptr;
5389   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5390 
5391   Compile* C = phase->C;
5392   int size_in_bytes = st->memory_size();
5393   int i = captured_store_insertion_point(start, size_in_bytes, phase);
5394   if (i == 0)  return nullptr;  // bail out
5395   Node* prev_mem = nullptr;     // raw memory for the captured store
5396   if (i > 0) {
5397     prev_mem = in(i);           // there is a pre-existing store under this one
5398     set_req(i, C->top());       // temporarily disconnect it
5399     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5400   } else {
5401     i = -i;                     // no pre-existing store
5402     prev_mem = zero_memory();   // a slice of the newly allocated object
5403     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5404       set_req(--i, C->top());   // reuse this edge; it has been folded away
5405     else
5406       ins_req(i, C->top());     // build a new edge
5407   }
5408   Node* new_st = st->clone_with_adr_type(TypeRawPtr::BOTTOM);
5409   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5410   new_st->set_req(MemNode::Control, in(Control));
5411   new_st->set_req(MemNode::Memory,  prev_mem);
5412   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5413   bs->eliminate_gc_barrier_data(new_st);
5414   new_st = phase->transform(new_st);
5415 
5416   // At this point, new_st might have swallowed a pre-existing store
5417   // at the same offset, or perhaps new_st might have disappeared,
5418   // if it redundantly stored the same value (or zero to fresh memory).
5419 
5420   // In any case, wire it in:
5421   PhaseIterGVN* igvn = phase->is_IterGVN();
5422   if (igvn) {
5423     igvn->rehash_node_delayed(this);
5424   }
5425   set_req(i, new_st);
5426 
5427   // The caller may now kill the old guy.
5428   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5429   assert(check_st == new_st || check_st == nullptr, "must be findable");
5430   assert(!is_complete(), "");
5431   return new_st;
5432 }
5433 
5434 static bool store_constant(jlong* tiles, int num_tiles,
5435                            intptr_t st_off, int st_size,
5436                            jlong con) {
5437   if ((st_off & (st_size-1)) != 0)
5438     return false;               // strange store offset (assume size==2**N)
5439   address addr = (address)tiles + st_off;
5440   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5441   switch (st_size) {
5442   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
5443   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
5444   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
5445   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
5446   default: return false;        // strange store size (detect size!=2**N here)
5447   }
5448   return true;                  // return success to caller
5449 }
5450 
5451 // Coalesce subword constants into int constants and possibly
5452 // into long constants.  The goal, if the CPU permits,
5453 // is to initialize the object with a small number of 64-bit tiles.
5454 // Also, convert floating-point constants to bit patterns.
5455 // Non-constants are not relevant to this pass.
5456 //
5457 // In terms of the running example on InitializeNode::InitializeNode
5458 // and InitializeNode::capture_store, here is the transformation
5459 // of rawstore1 and rawstore2 into rawstore12:
5460 //   alloc = (Allocate ...)
5461 //   rawoop = alloc.RawAddress
5462 //   tile12 = 0x00010002
5463 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5464 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5465 //
5466 void
5467 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5468                                         Node* size_in_bytes,
5469                                         PhaseGVN* phase) {
5470   Compile* C = phase->C;
5471 
5472   assert(stores_are_sane(phase), "");
5473   // Note:  After this pass, they are not completely sane,
5474   // since there may be some overlaps.
5475 
5476   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5477 
5478   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5479   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5480   size_limit = MIN2(size_limit, ti_limit);
5481   size_limit = align_up(size_limit, BytesPerLong);
5482   int num_tiles = size_limit / BytesPerLong;
5483 
5484   // allocate space for the tile map:
5485   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5486   jlong  tiles_buf[small_len];
5487   Node*  nodes_buf[small_len];
5488   jlong  inits_buf[small_len];
5489   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5490                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5491   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5492                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5493   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5494                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5495   // tiles: exact bitwise model of all primitive constants
5496   // nodes: last constant-storing node subsumed into the tiles model
5497   // inits: which bytes (in each tile) are touched by any initializations
5498 
5499   //// Pass A: Fill in the tile model with any relevant stores.
5500 
5501   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5502   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5503   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5504   Node* zmem = zero_memory(); // initially zero memory state
5505   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5506     Node* st = in(i);
5507     intptr_t st_off = get_store_offset(st, phase);
5508 
5509     // Figure out the store's offset and constant value:
5510     if (st_off < header_size)             continue; //skip (ignore header)
5511     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
5512     int st_size = st->as_Store()->memory_size();
5513     if (st_off + st_size > size_limit)    break;
5514 
5515     // Record which bytes are touched, whether by constant or not.
5516     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5517       continue;                 // skip (strange store size)
5518 
5519     const Type* val = phase->type(st->in(MemNode::ValueIn));
5520     if (!val->singleton())                continue; //skip (non-con store)
5521     BasicType type = val->basic_type();
5522 
5523     jlong con = 0;
5524     switch (type) {
5525     case T_INT:    con = val->is_int()->get_con();  break;
5526     case T_LONG:   con = val->is_long()->get_con(); break;
5527     case T_FLOAT:  con = jint_cast(val->getf());    break;
5528     case T_DOUBLE: con = jlong_cast(val->getd());   break;
5529     default:                              continue; //skip (odd store type)
5530     }
5531 
5532     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5533         st->Opcode() == Op_StoreL) {
5534       continue;                 // This StoreL is already optimal.
5535     }
5536 
5537     // Store down the constant.
5538     store_constant(tiles, num_tiles, st_off, st_size, con);
5539 
5540     intptr_t j = st_off >> LogBytesPerLong;
5541 
5542     if (type == T_INT && st_size == BytesPerInt
5543         && (st_off & BytesPerInt) == BytesPerInt) {
5544       jlong lcon = tiles[j];
5545       if (!Matcher::isSimpleConstant64(lcon) &&
5546           st->Opcode() == Op_StoreI) {
5547         // This StoreI is already optimal by itself.
5548         jint* intcon = (jint*) &tiles[j];
5549         intcon[1] = 0;  // undo the store_constant()
5550 
5551         // If the previous store is also optimal by itself, back up and
5552         // undo the action of the previous loop iteration... if we can.
5553         // But if we can't, just let the previous half take care of itself.
5554         st = nodes[j];
5555         st_off -= BytesPerInt;
5556         con = intcon[0];
5557         if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5558           assert(st_off >= header_size, "still ignoring header");
5559           assert(get_store_offset(st, phase) == st_off, "must be");
5560           assert(in(i-1) == zmem, "must be");
5561           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5562           assert(con == tcon->is_int()->get_con(), "must be");
5563           // Undo the effects of the previous loop trip, which swallowed st:
5564           intcon[0] = 0;        // undo store_constant()
5565           set_req(i-1, st);     // undo set_req(i, zmem)
5566           nodes[j] = nullptr;   // undo nodes[j] = st
5567           --old_subword;        // undo ++old_subword
5568         }
5569         continue;               // This StoreI is already optimal.
5570       }
5571     }
5572 
5573     // This store is not needed.
5574     set_req(i, zmem);
5575     nodes[j] = st;              // record for the moment
5576     if (st_size < BytesPerLong) // something has changed
5577           ++old_subword;        // includes int/float, but who's counting...
5578     else  ++old_long;
5579   }
5580 
5581   if ((old_subword + old_long) == 0)
5582     return;                     // nothing more to do
5583 
5584   //// Pass B: Convert any non-zero tiles into optimal constant stores.
5585   // Be sure to insert them before overlapping non-constant stores.
5586   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
5587   for (int j = 0; j < num_tiles; j++) {
5588     jlong con  = tiles[j];
5589     jlong init = inits[j];
5590     if (con == 0)  continue;
5591     jint con0,  con1;           // split the constant, address-wise
5592     jint init0, init1;          // split the init map, address-wise
5593     { union { jlong con; jint intcon[2]; } u;
5594       u.con = con;
5595       con0  = u.intcon[0];
5596       con1  = u.intcon[1];
5597       u.con = init;
5598       init0 = u.intcon[0];
5599       init1 = u.intcon[1];
5600     }
5601 
5602     Node* old = nodes[j];
5603     assert(old != nullptr, "need the prior store");
5604     intptr_t offset = (j * BytesPerLong);
5605 
5606     bool split = !Matcher::isSimpleConstant64(con);
5607 
5608     if (offset < header_size) {
5609       assert(offset + BytesPerInt >= header_size, "second int counts");
5610       assert(*(jint*)&tiles[j] == 0, "junk in header");
5611       split = true;             // only the second word counts
5612       // Example:  int a[] = { 42 ... }
5613     } else if (con0 == 0 && init0 == -1) {
5614       split = true;             // first word is covered by full inits
5615       // Example:  int a[] = { ... foo(), 42 ... }
5616     } else if (con1 == 0 && init1 == -1) {
5617       split = true;             // second word is covered by full inits
5618       // Example:  int a[] = { ... 42, foo() ... }
5619     }
5620 
5621     // Here's a case where init0 is neither 0 nor -1:
5622     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
5623     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5624     // In this case the tile is not split; it is (jlong)42.
5625     // The big tile is stored down, and then the foo() value is inserted.
5626     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5627 
5628     Node* ctl = old->in(MemNode::Control);
5629     Node* adr = make_raw_address(offset, phase);
5630     const TypePtr* atp = TypeRawPtr::BOTTOM;
5631 
5632     // One or two coalesced stores to plop down.
5633     Node*    st[2];
5634     intptr_t off[2];
5635     int  nst = 0;
5636     if (!split) {
5637       ++new_long;
5638       off[nst] = offset;
5639       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5640                                   phase->longcon(con), T_LONG, MemNode::unordered);
5641     } else {
5642       // Omit either if it is a zero.
5643       if (con0 != 0) {
5644         ++new_int;
5645         off[nst]  = offset;
5646         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5647                                     phase->intcon(con0), T_INT, MemNode::unordered);
5648       }
5649       if (con1 != 0) {
5650         ++new_int;
5651         offset += BytesPerInt;
5652         adr = make_raw_address(offset, phase);
5653         off[nst]  = offset;
5654         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5655                                     phase->intcon(con1), T_INT, MemNode::unordered);
5656       }
5657     }
5658 
5659     // Insert second store first, then the first before the second.
5660     // Insert each one just before any overlapping non-constant stores.
5661     while (nst > 0) {
5662       Node* st1 = st[--nst];
5663       C->copy_node_notes_to(st1, old);
5664       st1 = phase->transform(st1);
5665       offset = off[nst];
5666       assert(offset >= header_size, "do not smash header");
5667       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5668       guarantee(ins_idx != 0, "must re-insert constant store");
5669       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
5670       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5671         set_req(--ins_idx, st1);
5672       else
5673         ins_req(ins_idx, st1);
5674     }
5675   }
5676 
5677   if (PrintCompilation && WizardMode)
5678     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5679                   old_subword, old_long, new_int, new_long);
5680   if (C->log() != nullptr)
5681     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5682                    old_subword, old_long, new_int, new_long);
5683 
5684   // Clean up any remaining occurrences of zmem:
5685   remove_extra_zeroes();
5686 }
5687 
5688 // Explore forward from in(start) to find the first fully initialized
5689 // word, and return its offset.  Skip groups of subword stores which
5690 // together initialize full words.  If in(start) is itself part of a
5691 // fully initialized word, return the offset of in(start).  If there
5692 // are no following full-word stores, or if something is fishy, return
5693 // a negative value.
5694 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5695   int       int_map = 0;
5696   intptr_t  int_map_off = 0;
5697   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
5698 
5699   for (uint i = start, limit = req(); i < limit; i++) {
5700     Node* st = in(i);
5701 
5702     intptr_t st_off = get_store_offset(st, phase);
5703     if (st_off < 0)  break;  // return conservative answer
5704 
5705     int st_size = st->as_Store()->memory_size();
5706     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5707       return st_off;            // we found a complete word init
5708     }
5709 
5710     // update the map:
5711 
5712     intptr_t this_int_off = align_down(st_off, BytesPerInt);
5713     if (this_int_off != int_map_off) {
5714       // reset the map:
5715       int_map = 0;
5716       int_map_off = this_int_off;
5717     }
5718 
5719     int subword_off = st_off - this_int_off;
5720     int_map |= right_n_bits(st_size) << subword_off;
5721     if ((int_map & FULL_MAP) == FULL_MAP) {
5722       return this_int_off;      // we found a complete word init
5723     }
5724 
5725     // Did this store hit or cross the word boundary?
5726     intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5727     if (next_int_off == this_int_off + BytesPerInt) {
5728       // We passed the current int, without fully initializing it.
5729       int_map_off = next_int_off;
5730       int_map >>= BytesPerInt;
5731     } else if (next_int_off > this_int_off + BytesPerInt) {
5732       // We passed the current and next int.
5733       return this_int_off + BytesPerInt;
5734     }
5735   }
5736 
5737   return -1;
5738 }
5739 
5740 
5741 // Called when the associated AllocateNode is expanded into CFG.
5742 // At this point, we may perform additional optimizations.
5743 // Linearize the stores by ascending offset, to make memory
5744 // activity as coherent as possible.
5745 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5746                                       intptr_t header_size,
5747                                       Node* size_in_bytes,
5748                                       PhaseIterGVN* phase) {
5749   assert(!is_complete(), "not already complete");
5750   assert(stores_are_sane(phase), "");
5751   assert(allocation() != nullptr, "must be present");
5752 
5753   remove_extra_zeroes();
5754 
5755   if (ReduceFieldZeroing || ReduceBulkZeroing)
5756     // reduce instruction count for common initialization patterns
5757     coalesce_subword_stores(header_size, size_in_bytes, phase);
5758 
5759   Node* zmem = zero_memory();   // initially zero memory state
5760   Node* inits = zmem;           // accumulating a linearized chain of inits
5761   #ifdef ASSERT
5762   intptr_t first_offset = allocation()->minimum_header_size();
5763   intptr_t last_init_off = first_offset;  // previous init offset
5764   intptr_t last_init_end = first_offset;  // previous init offset+size
5765   intptr_t last_tile_end = first_offset;  // previous tile offset+size
5766   #endif
5767   intptr_t zeroes_done = header_size;
5768 
5769   bool do_zeroing = true;       // we might give up if inits are very sparse
5770   int  big_init_gaps = 0;       // how many large gaps have we seen?
5771 
5772   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
5773   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
5774 
5775   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5776     Node* st = in(i);
5777     intptr_t st_off = get_store_offset(st, phase);
5778     if (st_off < 0)
5779       break;                    // unknown junk in the inits
5780     if (st->in(MemNode::Memory) != zmem)
5781       break;                    // complicated store chains somehow in list
5782 
5783     int st_size = st->as_Store()->memory_size();
5784     intptr_t next_init_off = st_off + st_size;
5785 
5786     if (do_zeroing && zeroes_done < next_init_off) {
5787       // See if this store needs a zero before it or under it.
5788       intptr_t zeroes_needed = st_off;
5789 
5790       if (st_size < BytesPerInt) {
5791         // Look for subword stores which only partially initialize words.
5792         // If we find some, we must lay down some word-level zeroes first,
5793         // underneath the subword stores.
5794         //
5795         // Examples:
5796         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
5797         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
5798         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
5799         //
5800         // Note:  coalesce_subword_stores may have already done this,
5801         // if it was prompted by constant non-zero subword initializers.
5802         // But this case can still arise with non-constant stores.
5803 
5804         intptr_t next_full_store = find_next_fullword_store(i, phase);
5805 
5806         // In the examples above:
5807         //   in(i)          p   q   r   s     x   y     z
5808         //   st_off        12  13  14  15    12  13    14
5809         //   st_size        1   1   1   1     1   1     1
5810         //   next_full_s.  12  16  16  16    16  16    16
5811         //   z's_done      12  16  16  16    12  16    12
5812         //   z's_needed    12  16  16  16    16  16    16
5813         //   zsize          0   0   0   0     4   0     4
5814         if (next_full_store < 0) {
5815           // Conservative tack:  Zero to end of current word.
5816           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5817         } else {
5818           // Zero to beginning of next fully initialized word.
5819           // Or, don't zero at all, if we are already in that word.
5820           assert(next_full_store >= zeroes_needed, "must go forward");
5821           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5822           zeroes_needed = next_full_store;
5823         }
5824       }
5825 
5826       if (zeroes_needed > zeroes_done) {
5827         intptr_t zsize = zeroes_needed - zeroes_done;
5828         // Do some incremental zeroing on rawmem, in parallel with inits.
5829         zeroes_done = align_down(zeroes_done, BytesPerInt);
5830         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5831                                               allocation()->in(AllocateNode::InitValue),
5832                                               allocation()->in(AllocateNode::RawInitValue),
5833                                               zeroes_done, zeroes_needed,
5834                                               true,
5835                                               phase);
5836         zeroes_done = zeroes_needed;
5837         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5838           do_zeroing = false;   // leave the hole, next time
5839       }
5840     }
5841 
5842     // Collect the store and move on:
5843     phase->replace_input_of(st, MemNode::Memory, inits);
5844     inits = st;                 // put it on the linearized chain
5845     set_req(i, zmem);           // unhook from previous position
5846 
5847     if (zeroes_done == st_off)
5848       zeroes_done = next_init_off;
5849 
5850     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5851 
5852     #ifdef ASSERT
5853     // Various order invariants.  Weaker than stores_are_sane because
5854     // a large constant tile can be filled in by smaller non-constant stores.
5855     assert(st_off >= last_init_off, "inits do not reverse");
5856     last_init_off = st_off;
5857     const Type* val = nullptr;
5858     if (st_size >= BytesPerInt &&
5859         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5860         (int)val->basic_type() < (int)T_OBJECT) {
5861       assert(st_off >= last_tile_end, "tiles do not overlap");
5862       assert(st_off >= last_init_end, "tiles do not overwrite inits");
5863       last_tile_end = MAX2(last_tile_end, next_init_off);
5864     } else {
5865       intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5866       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5867       assert(st_off      >= last_init_end, "inits do not overlap");
5868       last_init_end = next_init_off;  // it's a non-tile
5869     }
5870     #endif //ASSERT
5871   }
5872 
5873   remove_extra_zeroes();        // clear out all the zmems left over
5874   add_req(inits);
5875 
5876   if (!(UseTLAB && ZeroTLAB)) {
5877     // If anything remains to be zeroed, zero it all now.
5878     zeroes_done = align_down(zeroes_done, BytesPerInt);
5879     // if it is the last unused 4 bytes of an instance, forget about it
5880     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5881     if (zeroes_done + BytesPerLong >= size_limit) {
5882       AllocateNode* alloc = allocation();
5883       assert(alloc != nullptr, "must be present");
5884       if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5885         Node* klass_node = alloc->in(AllocateNode::KlassNode);
5886         ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5887         if (zeroes_done == k->layout_helper())
5888           zeroes_done = size_limit;
5889       }
5890     }
5891     if (zeroes_done < size_limit) {
5892       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5893                                             allocation()->in(AllocateNode::InitValue),
5894                                             allocation()->in(AllocateNode::RawInitValue),
5895                                             zeroes_done, size_in_bytes, true, phase);
5896     }
5897   }
5898 
5899   set_complete(phase);
5900   return rawmem;
5901 }
5902 
5903 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5904   auto replace_proj = [&](ProjNode* proj) {
5905     C->gvn_replace_by(proj, mem);
5906     return CONTINUE;
5907   };
5908   apply_to_projs(replace_proj, TypeFunc::Memory);
5909 }
5910 
5911 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5912   DUIterator_Fast imax, i = fast_outs(imax);
5913   auto replace_proj = [&](ProjNode* proj) {
5914     igvn->replace_node(proj, mem);
5915     --i; --imax;
5916     return CONTINUE;
5917   };
5918   apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5919 }
5920 
5921 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5922   auto find_proj = [&](ProjNode* proj) {
5923     if (proj->adr_type() == adr_type) {
5924       return BREAK_AND_RETURN_CURRENT_PROJ;
5925     }
5926     return CONTINUE;
5927   };
5928   DUIterator_Fast imax, i = fast_outs(imax);
5929   return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5930 }
5931 
5932 MachProjNode* InitializeNode::mem_mach_proj() const {
5933   auto find_proj = [](ProjNode* proj) {
5934     if (proj->is_MachProj()) {
5935       return BREAK_AND_RETURN_CURRENT_PROJ;
5936     }
5937     return CONTINUE;
5938   };
5939   ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5940   if (proj == nullptr) {
5941     return nullptr;
5942   }
5943   return proj->as_MachProj();
5944 }
5945 
5946 #ifdef ASSERT
5947 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5948   if (is_complete())
5949     return true;                // stores could be anything at this point
5950   assert(allocation() != nullptr, "must be present");
5951   intptr_t last_off = allocation()->minimum_header_size();
5952   for (uint i = InitializeNode::RawStores; i < req(); i++) {
5953     Node* st = in(i);
5954     intptr_t st_off = get_store_offset(st, phase);
5955     if (st_off < 0)  continue;  // ignore dead garbage
5956     if (last_off > st_off) {
5957       tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5958       this->dump(2);
5959       assert(false, "ascending store offsets");
5960       return false;
5961     }
5962     last_off = st_off + st->as_Store()->memory_size();
5963   }
5964   return true;
5965 }
5966 #endif //ASSERT
5967 
5968 
5969 
5970 
5971 //============================MergeMemNode=====================================
5972 //
5973 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
5974 // contributing store or call operations.  Each contributor provides the memory
5975 // state for a particular "alias type" (see Compile::alias_type).  For example,
5976 // if a MergeMem has an input X for alias category #6, then any memory reference
5977 // to alias category #6 may use X as its memory state input, as an exact equivalent
5978 // to using the MergeMem as a whole.
5979 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5980 //
5981 // (Here, the <N> notation gives the index of the relevant adr_type.)
5982 //
5983 // In one special case (and more cases in the future), alias categories overlap.
5984 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5985 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
5986 // it is exactly equivalent to that state W:
5987 //   MergeMem(<Bot>: W) <==> W
5988 //
5989 // Usually, the merge has more than one input.  In that case, where inputs
5990 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5991 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5992 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5993 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5994 //
5995 // A merge can take a "wide" memory state as one of its narrow inputs.
5996 // This simply means that the merge observes out only the relevant parts of
5997 // the wide input.  That is, wide memory states arriving at narrow merge inputs
5998 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
5999 //
6000 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
6001 // and that memory slices "leak through":
6002 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
6003 //
6004 // But, in such a cascade, repeated memory slices can "block the leak":
6005 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
6006 //
6007 // In the last example, Y is not part of the combined memory state of the
6008 // outermost MergeMem.  The system must, of course, prevent unschedulable
6009 // memory states from arising, so you can be sure that the state Y is somehow
6010 // a precursor to state Y'.
6011 //
6012 //
6013 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
6014 // of each MergeMemNode array are exactly the numerical alias indexes, including
6015 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
6016 // Compile::alias_type (and kin) produce and manage these indexes.
6017 //
6018 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
6019 // (Note that this provides quick access to the top node inside MergeMem methods,
6020 // without the need to reach out via TLS to Compile::current.)
6021 //
6022 // As a consequence of what was just described, a MergeMem that represents a full
6023 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
6024 // containing all alias categories.
6025 //
6026 // MergeMem nodes never (?) have control inputs, so in(0) is null.
6027 //
6028 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
6029 // a memory state for the alias type <N>, or else the top node, meaning that
6030 // there is no particular input for that alias type.  Note that the length of
6031 // a MergeMem is variable, and may be extended at any time to accommodate new
6032 // memory states at larger alias indexes.  When merges grow, they are of course
6033 // filled with "top" in the unused in() positions.
6034 //
6035 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
6036 // (Top was chosen because it works smoothly with passes like GCM.)
6037 //
6038 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
6039 // the type of random VM bits like TLS references.)  Since it is always the
6040 // first non-Bot memory slice, some low-level loops use it to initialize an
6041 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
6042 //
6043 //
6044 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
6045 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
6046 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
6047 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
6048 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
6049 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
6050 //
6051 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
6052 // really that different from the other memory inputs.  An abbreviation called
6053 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
6054 //
6055 //
6056 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
6057 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
6058 // that "emerges though" the base memory will be marked as excluding the alias types
6059 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
6060 //
6061 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
6062 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
6063 //
6064 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
6065 // (It is currently unimplemented.)  As you can see, the resulting merge is
6066 // actually a disjoint union of memory states, rather than an overlay.
6067 //
6068 
6069 //------------------------------MergeMemNode-----------------------------------
6070 Node* MergeMemNode::make_empty_memory() {
6071   Node* empty_memory = (Node*) Compile::current()->top();
6072   assert(empty_memory->is_top(), "correct sentinel identity");
6073   return empty_memory;
6074 }
6075 
6076 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
6077   init_class_id(Class_MergeMem);
6078   // all inputs are nullified in Node::Node(int)
6079   // set_input(0, nullptr);  // no control input
6080 
6081   // Initialize the edges uniformly to top, for starters.
6082   Node* empty_mem = make_empty_memory();
6083   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
6084     init_req(i,empty_mem);
6085   }
6086   assert(empty_memory() == empty_mem, "");
6087 
6088   if( new_base != nullptr && new_base->is_MergeMem() ) {
6089     MergeMemNode* mdef = new_base->as_MergeMem();
6090     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
6091     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
6092       mms.set_memory(mms.memory2());
6093     }
6094     assert(base_memory() == mdef->base_memory(), "");
6095   } else {
6096     set_base_memory(new_base);
6097   }
6098 }
6099 
6100 // Make a new, untransformed MergeMem with the same base as 'mem'.
6101 // If mem is itself a MergeMem, populate the result with the same edges.
6102 MergeMemNode* MergeMemNode::make(Node* mem) {
6103   return new MergeMemNode(mem);
6104 }
6105 
6106 //------------------------------cmp--------------------------------------------
6107 uint MergeMemNode::hash() const { return NO_HASH; }
6108 bool MergeMemNode::cmp( const Node &n ) const {
6109   return (&n == this);          // Always fail except on self
6110 }
6111 
6112 //------------------------------Identity---------------------------------------
6113 Node* MergeMemNode::Identity(PhaseGVN* phase) {
6114   // Identity if this merge point does not record any interesting memory
6115   // disambiguations.
6116   Node* base_mem = base_memory();
6117   Node* empty_mem = empty_memory();
6118   if (base_mem != empty_mem) {  // Memory path is not dead?
6119     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6120       Node* mem = in(i);
6121       if (mem != empty_mem && mem != base_mem) {
6122         return this;            // Many memory splits; no change
6123       }
6124     }
6125   }
6126   return base_mem;              // No memory splits; ID on the one true input
6127 }
6128 
6129 //------------------------------Ideal------------------------------------------
6130 // This method is invoked recursively on chains of MergeMem nodes
6131 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
6132   if (Identity(phase) != this) {
6133     // Let Identity handle this case
6134     return nullptr;
6135   }
6136 
6137   // Remove chain'd MergeMems
6138   //
6139   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
6140   // relative to the "in(Bot)".  Since we are patching both at the same time,
6141   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
6142   // but rewrite each "in(i)" relative to the new "in(Bot)".
6143   Node *progress = nullptr;
6144 
6145 
6146   Node* old_base = base_memory();
6147   Node* empty_mem = empty_memory();
6148   if (old_base == empty_mem)
6149     return nullptr; // Dead memory path.
6150 
6151   MergeMemNode* old_mbase;
6152   if (old_base != nullptr && old_base->is_MergeMem())
6153     old_mbase = old_base->as_MergeMem();
6154   else
6155     old_mbase = nullptr;
6156   Node* new_base = old_base;
6157 
6158   // simplify stacked MergeMems in base memory
6159   if (old_mbase)  new_base = old_mbase->base_memory();
6160 
6161   // the base memory might contribute new slices beyond my req()
6162   if (old_mbase)  grow_to_match(old_mbase);
6163 
6164   // Note:  We do not call verify_sparse on entry, because inputs
6165   // can normalize to the base_memory via subsume_node or similar
6166   // mechanisms.  This method repairs that damage.
6167 
6168   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
6169 
6170   // Look at each slice.
6171   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6172     Node* old_in = in(i);
6173     // calculate the old memory value
6174     Node* old_mem = old_in;
6175     if (old_mem == empty_mem)  old_mem = old_base;
6176     assert(old_mem == memory_at(i), "");
6177 
6178     // maybe update (reslice) the old memory value
6179 
6180     // simplify stacked MergeMems
6181     Node* new_mem = old_mem;
6182     MergeMemNode* old_mmem;
6183     if (old_mem != nullptr && old_mem->is_MergeMem())
6184       old_mmem = old_mem->as_MergeMem();
6185     else
6186       old_mmem = nullptr;
6187     if (old_mmem == this) {
6188       // This can happen if loops break up and safepoints disappear.
6189       // A merge of BotPtr (default) with a RawPtr memory derived from a
6190       // safepoint can be rewritten to a merge of the same BotPtr with
6191       // the BotPtr phi coming into the loop.  If that phi disappears
6192       // also, we can end up with a self-loop of the mergemem.
6193       // In general, if loops degenerate and memory effects disappear,
6194       // a mergemem can be left looking at itself.  This simply means
6195       // that the mergemem's default should be used, since there is
6196       // no longer any apparent effect on this slice.
6197       // Note: If a memory slice is a MergeMem cycle, it is unreachable
6198       //       from start.  Update the input to TOP.
6199       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
6200     }
6201     else if (old_mmem != nullptr) {
6202       new_mem = old_mmem->memory_at(i);
6203     }
6204     // else preceding memory was not a MergeMem
6205 
6206     // maybe store down a new value
6207     Node* new_in = new_mem;
6208     if (new_in == new_base)  new_in = empty_mem;
6209 
6210     if (new_in != old_in) {
6211       // Warning:  Do not combine this "if" with the previous "if"
6212       // A memory slice might have be be rewritten even if it is semantically
6213       // unchanged, if the base_memory value has changed.
6214       set_req_X(i, new_in, phase);
6215       progress = this;          // Report progress
6216     }
6217   }
6218 
6219   if (new_base != old_base) {
6220     set_req_X(Compile::AliasIdxBot, new_base, phase);
6221     // Don't use set_base_memory(new_base), because we need to update du.
6222     assert(base_memory() == new_base, "");
6223     progress = this;
6224   }
6225 
6226   if( base_memory() == this ) {
6227     // a self cycle indicates this memory path is dead
6228     set_req(Compile::AliasIdxBot, empty_mem);
6229   }
6230 
6231   // Resolve external cycles by calling Ideal on a MergeMem base_memory
6232   // Recursion must occur after the self cycle check above
6233   if( base_memory()->is_MergeMem() ) {
6234     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
6235     Node *m = phase->transform(new_mbase);  // Rollup any cycles
6236     if( m != nullptr &&
6237         (m->is_top() ||
6238          (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
6239       // propagate rollup of dead cycle to self
6240       set_req(Compile::AliasIdxBot, empty_mem);
6241     }
6242   }
6243 
6244   if( base_memory() == empty_mem ) {
6245     progress = this;
6246     // Cut inputs during Parse phase only.
6247     // During Optimize phase a dead MergeMem node will be subsumed by Top.
6248     if( !can_reshape ) {
6249       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6250         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
6251       }
6252     }
6253   }
6254 
6255   if( !progress && base_memory()->is_Phi() && can_reshape ) {
6256     // Check if PhiNode::Ideal's "Split phis through memory merges"
6257     // transform should be attempted. Look for this->phi->this cycle.
6258     uint merge_width = req();
6259     if (merge_width > Compile::AliasIdxRaw) {
6260       PhiNode* phi = base_memory()->as_Phi();
6261       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
6262         if (phi->in(i) == this) {
6263           phase->is_IterGVN()->_worklist.push(phi);
6264           break;
6265         }
6266       }
6267     }
6268   }
6269 
6270   assert(progress || verify_sparse(), "please, no dups of base");
6271   return progress;
6272 }
6273 
6274 //-------------------------set_base_memory-------------------------------------
6275 void MergeMemNode::set_base_memory(Node *new_base) {
6276   Node* empty_mem = empty_memory();
6277   set_req(Compile::AliasIdxBot, new_base);
6278   assert(memory_at(req()) == new_base, "must set default memory");
6279   // Clear out other occurrences of new_base:
6280   if (new_base != empty_mem) {
6281     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6282       if (in(i) == new_base)  set_req(i, empty_mem);
6283     }
6284   }
6285 }
6286 
6287 //------------------------------out_RegMask------------------------------------
6288 const RegMask &MergeMemNode::out_RegMask() const {
6289   return RegMask::EMPTY;
6290 }
6291 
6292 //------------------------------dump_spec--------------------------------------
6293 #ifndef PRODUCT
6294 void MergeMemNode::dump_spec(outputStream *st) const {
6295   st->print(" {");
6296   Node* base_mem = base_memory();
6297   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
6298     Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
6299     if (mem == base_mem) { st->print(" -"); continue; }
6300     st->print( " N%d:", mem->_idx );
6301     Compile::current()->get_adr_type(i)->dump_on(st);
6302   }
6303   st->print(" }");
6304 }
6305 #endif // !PRODUCT
6306 
6307 
6308 #ifdef ASSERT
6309 static bool might_be_same(Node* a, Node* b) {
6310   if (a == b)  return true;
6311   if (!(a->is_Phi() || b->is_Phi()))  return false;
6312   // phis shift around during optimization
6313   return true;  // pretty stupid...
6314 }
6315 
6316 // verify a narrow slice (either incoming or outgoing)
6317 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
6318   if (!VerifyAliases)                return;  // don't bother to verify unless requested
6319   if (VMError::is_error_reported())  return;  // muzzle asserts when debugging an error
6320   if (Node::in_dump())               return;  // muzzle asserts when printing
6321   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
6322   assert(n != nullptr, "");
6323   // Elide intervening MergeMem's
6324   while (n->is_MergeMem()) {
6325     n = n->as_MergeMem()->memory_at(alias_idx);
6326   }
6327   Compile* C = Compile::current();
6328   const TypePtr* n_adr_type = n->adr_type();
6329   if (n == m->empty_memory()) {
6330     // Implicit copy of base_memory()
6331   } else if (n_adr_type != TypePtr::BOTTOM) {
6332     assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6333     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6334   } else {
6335     // A few places like make_runtime_call "know" that VM calls are narrow,
6336     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6337     bool expected_wide_mem = false;
6338     if (n == m->base_memory()) {
6339       expected_wide_mem = true;
6340     } else if (alias_idx == Compile::AliasIdxRaw ||
6341                n == m->memory_at(Compile::AliasIdxRaw)) {
6342       expected_wide_mem = true;
6343     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6344       // memory can "leak through" calls on channels that
6345       // are write-once.  Allow this also.
6346       expected_wide_mem = true;
6347     }
6348     assert(expected_wide_mem, "expected narrow slice replacement");
6349   }
6350 }
6351 #else // !ASSERT
6352 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
6353 #endif
6354 
6355 
6356 //-----------------------------memory_at---------------------------------------
6357 Node* MergeMemNode::memory_at(uint alias_idx) const {
6358   assert(alias_idx >= Compile::AliasIdxRaw ||
6359          (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6360          "must avoid base_memory and AliasIdxTop");
6361 
6362   // Otherwise, it is a narrow slice.
6363   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6364   if (is_empty_memory(n)) {
6365     // the array is sparse; empty slots are the "top" node
6366     n = base_memory();
6367     assert(Node::in_dump()
6368            || n == nullptr || n->bottom_type() == Type::TOP
6369            || n->adr_type() == nullptr // address is TOP
6370            || n->adr_type() == TypePtr::BOTTOM
6371            || n->adr_type() == TypeRawPtr::BOTTOM
6372            || n->is_NarrowMemProj()
6373            || !Compile::current()->do_aliasing(),
6374            "must be a wide memory");
6375     // do_aliasing == false if we are organizing the memory states manually.
6376     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6377   } else {
6378     // make sure the stored slice is sane
6379     #ifdef ASSERT
6380     if (VMError::is_error_reported() || Node::in_dump()) {
6381     } else if (might_be_same(n, base_memory())) {
6382       // Give it a pass:  It is a mostly harmless repetition of the base.
6383       // This can arise normally from node subsumption during optimization.
6384     } else {
6385       verify_memory_slice(this, alias_idx, n);
6386     }
6387     #endif
6388   }
6389   return n;
6390 }
6391 
6392 //---------------------------set_memory_at-------------------------------------
6393 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6394   verify_memory_slice(this, alias_idx, n);
6395   Node* empty_mem = empty_memory();
6396   if (n == base_memory())  n = empty_mem;  // collapse default
6397   uint need_req = alias_idx+1;
6398   if (req() < need_req) {
6399     if (n == empty_mem)  return;  // already the default, so do not grow me
6400     // grow the sparse array
6401     do {
6402       add_req(empty_mem);
6403     } while (req() < need_req);
6404   }
6405   set_req( alias_idx, n );
6406 }
6407 
6408 
6409 
6410 //--------------------------iteration_setup------------------------------------
6411 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6412   if (other != nullptr) {
6413     grow_to_match(other);
6414     // invariant:  the finite support of mm2 is within mm->req()
6415     #ifdef ASSERT
6416     for (uint i = req(); i < other->req(); i++) {
6417       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6418     }
6419     #endif
6420   }
6421   // Replace spurious copies of base_memory by top.
6422   Node* base_mem = base_memory();
6423   if (base_mem != nullptr && !base_mem->is_top()) {
6424     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6425       if (in(i) == base_mem)
6426         set_req(i, empty_memory());
6427     }
6428   }
6429 }
6430 
6431 //---------------------------grow_to_match-------------------------------------
6432 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6433   Node* empty_mem = empty_memory();
6434   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6435   // look for the finite support of the other memory
6436   for (uint i = other->req(); --i >= req(); ) {
6437     if (other->in(i) != empty_mem) {
6438       uint new_len = i+1;
6439       while (req() < new_len)  add_req(empty_mem);
6440       break;
6441     }
6442   }
6443 }
6444 
6445 //---------------------------verify_sparse-------------------------------------
6446 #ifndef PRODUCT
6447 bool MergeMemNode::verify_sparse() const {
6448   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6449   Node* base_mem = base_memory();
6450   // The following can happen in degenerate cases, since empty==top.
6451   if (is_empty_memory(base_mem))  return true;
6452   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6453     assert(in(i) != nullptr, "sane slice");
6454     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
6455   }
6456   return true;
6457 }
6458 
6459 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6460   Node* n;
6461   n = mm->in(idx);
6462   if (mem == n)  return true;  // might be empty_memory()
6463   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6464   if (mem == n)  return true;
6465   return false;
6466 }
6467 #endif // !PRODUCT