1 /*
   2  * Copyright (c) 1997, 2025, 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 "classfile/javaClasses.hpp"
  28 #include "classfile/systemDictionary.hpp"
  29 #include "compiler/compileLog.hpp"
  30 #include "gc/shared/barrierSet.hpp"
  31 #include "gc/shared/c2/barrierSetC2.hpp"
  32 #include "gc/shared/tlab_globals.hpp"
  33 #include "memory/allocation.inline.hpp"
  34 #include "memory/resourceArea.hpp"
  35 #include "oops/objArrayKlass.hpp"
  36 #include "opto/addnode.hpp"
  37 #include "opto/arraycopynode.hpp"
  38 #include "opto/cfgnode.hpp"
  39 #include "opto/compile.hpp"
  40 #include "opto/connode.hpp"
  41 #include "opto/convertnode.hpp"
  42 #include "opto/inlinetypenode.hpp"
  43 #include "opto/loopnode.hpp"
  44 #include "opto/machnode.hpp"
  45 #include "opto/matcher.hpp"
  46 #include "opto/memnode.hpp"
  47 #include "opto/mempointer.hpp"
  48 #include "opto/mulnode.hpp"
  49 #include "opto/narrowptrnode.hpp"
  50 #include "opto/phaseX.hpp"
  51 #include "opto/regalloc.hpp"
  52 #include "opto/regmask.hpp"
  53 #include "opto/rootnode.hpp"
  54 #include "opto/traceMergeStoresTag.hpp"
  55 #include "opto/vectornode.hpp"
  56 #include "utilities/align.hpp"
  57 #include "utilities/copy.hpp"
  58 #include "utilities/macros.hpp"
  59 #include "utilities/powerOfTwo.hpp"
  60 #include "utilities/vmError.hpp"
  61 
  62 // Portions of code courtesy of Clifford Click
  63 
  64 // Optimization - Graph Style
  65 
  66 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  67 
  68 //=============================================================================
  69 uint MemNode::size_of() const { return sizeof(*this); }
  70 
  71 const TypePtr *MemNode::adr_type() const {
  72   Node* adr = in(Address);
  73   if (adr == nullptr)  return nullptr; // node is dead
  74   const TypePtr* cross_check = nullptr;
  75   DEBUG_ONLY(cross_check = _adr_type);
  76   return calculate_adr_type(adr->bottom_type(), cross_check);
  77 }
  78 
  79 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
  80   if (adr != nullptr) {
  81     if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
  82       return true;
  83     }
  84   }
  85   return false;
  86 }
  87 
  88 #ifndef PRODUCT
  89 void MemNode::dump_spec(outputStream *st) const {
  90   if (in(Address) == nullptr)  return; // node is dead
  91 #ifndef ASSERT
  92   // fake the missing field
  93   const TypePtr* _adr_type = nullptr;
  94   if (in(Address) != nullptr)
  95     _adr_type = in(Address)->bottom_type()->isa_ptr();
  96 #endif
  97   dump_adr_type(this, _adr_type, st);
  98 
  99   Compile* C = Compile::current();
 100   if (C->alias_type(_adr_type)->is_volatile()) {
 101     st->print(" Volatile!");
 102   }
 103   if (_unaligned_access) {
 104     st->print(" unaligned");
 105   }
 106   if (_mismatched_access) {
 107     st->print(" mismatched");
 108   }
 109   if (_unsafe_access) {
 110     st->print(" unsafe");
 111   }
 112 }
 113 
 114 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
 115   st->print(" @");
 116   if (adr_type == nullptr) {
 117     st->print("null");
 118   } else {
 119     adr_type->dump_on(st);
 120     Compile* C = Compile::current();
 121     Compile::AliasType* atp = nullptr;
 122     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
 123     if (atp == nullptr)
 124       st->print(", idx=?\?;");
 125     else if (atp->index() == Compile::AliasIdxBot)
 126       st->print(", idx=Bot;");
 127     else if (atp->index() == Compile::AliasIdxTop)
 128       st->print(", idx=Top;");
 129     else if (atp->index() == Compile::AliasIdxRaw)
 130       st->print(", idx=Raw;");
 131     else {
 132       ciField* field = atp->field();
 133       if (field) {
 134         st->print(", name=");
 135         field->print_name_on(st);
 136       }
 137       st->print(", idx=%d;", atp->index());
 138     }
 139   }
 140 }
 141 
 142 extern void print_alias_types();
 143 
 144 #endif
 145 
 146 // Find the memory output corresponding to the fall-through path of a call
 147 static Node* find_call_fallthrough_mem_output(CallNode* call) {
 148   ResourceMark rm;
 149   CallProjections* projs = call->extract_projections(false, false);
 150   Node* res = projs->fallthrough_memproj;
 151   assert(res != nullptr, "must have a fallthrough mem output");
 152   return res;
 153 }
 154 
 155 // Try to find a better memory input for a load from a strict final field
 156 static Node* try_optimize_strict_final_load_memory(PhaseGVN* phase, Node* adr, ProjNode*& base_local) {
 157   intptr_t offset = 0;
 158   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 159   if (base == nullptr) {
 160     return nullptr;
 161   }
 162 
 163   Node* base_uncasted = base->uncast();
 164   if (base_uncasted->is_Proj()) {
 165     MultiNode* multi = base_uncasted->in(0)->as_Multi();
 166     if (multi->is_Allocate()) {
 167       base_local = base_uncasted->as_Proj();
 168       return nullptr;
 169     } else if (multi->is_Call()) {
 170       // The oop is returned from a call, the memory can be the fallthrough output of the call
 171       return find_call_fallthrough_mem_output(multi->as_Call());
 172     } else if (multi->is_Start()) {
 173       // The oop is a parameter
 174       if (phase->C->method()->is_object_constructor() && base_uncasted->as_Proj()->_con == TypeFunc::Parms) {
 175         // The receiver of a constructor is similar to the result of an AllocateNode
 176         base_local = base_uncasted->as_Proj();
 177         return nullptr;
 178       } else {
 179         // Use the start memory otherwise
 180         return multi->proj_out(TypeFunc::Memory);
 181       }
 182     }
 183   }
 184 
 185   return nullptr;
 186 }
 187 
 188 // Whether a call can modify a strict final field, given that the object is allocated inside the
 189 // current compilation unit, or is the first parameter when the compilation root is a constructor.
 190 // This is equivalent to asking whether 'call' is a constructor invocation and the class declaring
 191 // the target method is a subclass of the class declaring 'field'.
 192 static bool call_can_modify_local_object(ciField* field, CallNode* call) {
 193   if (!call->is_CallJava()) {
 194     return false;
 195   }
 196 
 197   ciMethod* target = call->as_CallJava()->method();
 198   if (target == nullptr || !target->is_object_constructor()) {
 199     return false;
 200   }
 201 
 202   // If 'field' is declared in a class that is a subclass of the one declaring the constructor,
 203   // then the field is set inside the constructor, else the field must be set before the
 204   // constructor invocation. E.g. A field Super.x will be set during the execution of Sub::<init>,
 205   // while a field Sub.y must be set before Super::<init> is invoked.
 206   // We can try to be more heroic and decide if the receiver of the constructor invocation is the
 207   // object from which we are loading from. This, however, may be problematic as deciding if 2
 208   // nodes are definitely different may not be trivial, especially if the graph is not canonical.
 209   // As a result, it is made more conservative for now.
 210   assert(call->req() > TypeFunc::Parms, "constructor must have at least 1 argument");
 211   return target->holder()->is_subclass_of(field->holder());
 212 }
 213 
 214 Node* MemNode::optimize_simple_memory_chain(Node* mchain, const TypeOopPtr* t_oop, Node* load, PhaseGVN* phase) {
 215   assert(t_oop != nullptr, "sanity");
 216   bool is_instance = t_oop->is_known_instance_field();
 217 
 218   ciField* field = phase->C->alias_type(t_oop)->field();
 219   bool is_strict_final_load = false;
 220 
 221   // After macro expansion, an allocation may become a call, changing the memory input to the
 222   // memory output of that call would be illegal. As a result, disallow this transformation after
 223   // macro expansion.
 224   if (phase->is_IterGVN() && phase->C->allow_macro_nodes() && load != nullptr && load->is_Load() && !load->as_Load()->is_mismatched_access()) {
 225     if (EnableValhalla) {
 226       if (field != nullptr && (field->holder()->is_inlinetype() || field->holder()->is_abstract_value_klass())) {
 227         is_strict_final_load = true;
 228       }
 229 #ifdef ASSERT
 230       if (t_oop->is_inlinetypeptr() && t_oop->inline_klass()->contains_field_offset(t_oop->offset())) {
 231         assert(is_strict_final_load, "sanity check for basic cases");
 232       }
 233 #endif
 234     } else {
 235       is_strict_final_load = field != nullptr && t_oop->is_ptr_to_boxed_value();
 236     }
 237   }
 238 
 239   if (!is_instance && !is_strict_final_load) {
 240     return mchain;
 241   }
 242 
 243   Node* result = mchain;
 244   ProjNode* base_local = nullptr;
 245 
 246   if (is_strict_final_load) {
 247     Node* adr = load->in(MemNode::Address);
 248     assert(phase->type(adr) == t_oop, "inconsistent type");
 249     Node* tmp = try_optimize_strict_final_load_memory(phase, adr, base_local);
 250     if (tmp != nullptr) {
 251       result = tmp;
 252     }
 253   }
 254 
 255   uint instance_id = t_oop->instance_id();
 256   Node* start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
 257   Node* prev = nullptr;
 258   while (prev != result) {
 259     prev = result;
 260     if (result == start_mem) {
 261       // start_mem is the earliest memory possible
 262       break;
 263     }
 264 
 265     // skip over a call which does not affect this memory slice
 266     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 267       Node* proj_in = result->in(0);
 268       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 269         // This is the allocation that creates the object from which we are loading from
 270         break;
 271       } else if (proj_in->is_Call()) {
 272         // ArrayCopyNodes processed here as well
 273         CallNode* call = proj_in->as_Call();
 274         if (!call->may_modify(t_oop, phase)) {
 275           result = call->in(TypeFunc::Memory);
 276         } else if (is_strict_final_load && base_local != nullptr && !call_can_modify_local_object(field, call)) {
 277           result = call->in(TypeFunc::Memory);
 278         }
 279       } else if (proj_in->is_Initialize()) {
 280         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 281         // Stop if this is the initialization for the object instance which
 282         // contains this memory slice, otherwise skip over it.
 283         if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
 284           break;
 285         }
 286         if (is_instance) {
 287           result = proj_in->in(TypeFunc::Memory);
 288         } else if (is_strict_final_load) {
 289           Node* klass = alloc->in(AllocateNode::KlassNode);
 290           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 291           if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
 292             // Allocation of another type, must be another object
 293             result = proj_in->in(TypeFunc::Memory);
 294           } else if (base_local != nullptr && (base_local->is_Parm() || base_local->in(0) != alloc)) {
 295             // Allocation of another object
 296             result = proj_in->in(TypeFunc::Memory);
 297           }
 298         }
 299       } else if (proj_in->is_MemBar()) {
 300         ArrayCopyNode* ac = nullptr;
 301         if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
 302           break;
 303         }
 304         result = proj_in->in(TypeFunc::Memory);
 305       } else if (proj_in->is_top()) {
 306         break; // dead code
 307       } else {
 308         assert(false, "unexpected projection");
 309       }
 310     } else if (result->is_ClearArray()) {
 311       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 312         // Can not bypass initialization of the instance
 313         // we are looking for.
 314         break;
 315       }
 316       // Otherwise skip it (the call updated 'result' value).
 317     } else if (result->is_MergeMem()) {
 318       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
 319     }
 320   }
 321   return result;
 322 }
 323 
 324 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 325   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 326   if (t_oop == nullptr)
 327     return mchain;  // don't try to optimize non-oop types
 328   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 329   bool is_instance = t_oop->is_known_instance_field();
 330   PhaseIterGVN *igvn = phase->is_IterGVN();
 331   if (is_instance && igvn != nullptr && result->is_Phi()) {
 332     PhiNode *mphi = result->as_Phi();
 333     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 334     const TypePtr *t = mphi->adr_type();
 335     bool do_split = false;
 336     // In the following cases, Load memory input can be further optimized based on
 337     // its precise address type
 338     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
 339       do_split = true;
 340     } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
 341       const TypeOopPtr* mem_t =
 342         t->is_oopptr()->cast_to_exactness(true)
 343         ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 344         ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
 345       if (t_oop->isa_aryptr()) {
 346         mem_t = mem_t->is_aryptr()
 347                      ->cast_to_stable(t_oop->is_aryptr()->is_stable())
 348                      ->cast_to_size(t_oop->is_aryptr()->size())
 349                      ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
 350                      ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
 351                      ->with_offset(t_oop->is_aryptr()->offset())
 352                      ->is_aryptr();
 353       }
 354       do_split = mem_t == t_oop;
 355     }
 356     if (do_split) {
 357       // clone the Phi with our address type
 358       result = mphi->split_out_instance(t_adr, igvn);
 359     } else {
 360       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 361     }
 362   }
 363   return result;
 364 }
 365 
 366 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 367   uint alias_idx = phase->C->get_alias_index(tp);
 368   Node *mem = mmem;
 369 #ifdef ASSERT
 370   {
 371     // Check that current type is consistent with the alias index used during graph construction
 372     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 373     bool consistent =  adr_check == nullptr || adr_check->empty() ||
 374                        phase->C->must_alias(adr_check, alias_idx );
 375     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 376     if( !consistent && adr_check != nullptr && !adr_check->empty() &&
 377         tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 378         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 379         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 380           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 381           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 382       // don't assert if it is dead code.
 383       consistent = true;
 384     }
 385     if( !consistent ) {
 386       st->print("alias_idx==%d, adr_check==", alias_idx);
 387       if( adr_check == nullptr ) {
 388         st->print("null");
 389       } else {
 390         adr_check->dump();
 391       }
 392       st->cr();
 393       print_alias_types();
 394       assert(consistent, "adr_check must match alias idx");
 395     }
 396   }
 397 #endif
 398   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 399   // means an array I have not precisely typed yet.  Do not do any
 400   // alias stuff with it any time soon.
 401   const TypeOopPtr *toop = tp->isa_oopptr();
 402   if (tp->base() != Type::AnyPtr &&
 403       !(toop &&
 404         toop->isa_instptr() &&
 405         toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
 406         toop->offset() == Type::OffsetBot)) {
 407     // IGVN _delay_transform may be set to true and if that is the case and mmem
 408     // is already a registered node then the validation inside transform will
 409     // complain.
 410     Node* m = mmem;
 411     PhaseIterGVN* igvn = phase->is_IterGVN();
 412     if (igvn == nullptr || !igvn->delay_transform()) {
 413       // compress paths and change unreachable cycles to TOP
 414       // If not, we can update the input infinitely along a MergeMem cycle
 415       // Equivalent code in PhiNode::Ideal
 416       m = phase->transform(mmem);
 417     }
 418     // If transformed to a MergeMem, get the desired slice
 419     // Otherwise the returned node represents memory for every slice
 420     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 421     // Update input if it is progress over what we have now
 422   }
 423   return mem;
 424 }
 425 
 426 //--------------------------Ideal_common---------------------------------------
 427 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 428 // Unhook non-raw memories from complete (macro-expanded) initializations.
 429 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 430   // If our control input is a dead region, kill all below the region
 431   Node *ctl = in(MemNode::Control);
 432   if (ctl && remove_dead_region(phase, can_reshape))
 433     return this;
 434   ctl = in(MemNode::Control);
 435   // Don't bother trying to transform a dead node
 436   if (ctl && ctl->is_top())  return NodeSentinel;
 437 
 438   PhaseIterGVN *igvn = phase->is_IterGVN();
 439   // Wait if control on the worklist.
 440   if (ctl && can_reshape && igvn != nullptr) {
 441     Node* bol = nullptr;
 442     Node* cmp = nullptr;
 443     if (ctl->in(0)->is_If()) {
 444       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 445       bol = ctl->in(0)->in(1);
 446       if (bol->is_Bool())
 447         cmp = ctl->in(0)->in(1)->in(1);
 448     }
 449     if (igvn->_worklist.member(ctl) ||
 450         (bol != nullptr && igvn->_worklist.member(bol)) ||
 451         (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
 452       // This control path may be dead.
 453       // Delay this memory node transformation until the control is processed.
 454       igvn->_worklist.push(this);
 455       return NodeSentinel; // caller will return null
 456     }
 457   }
 458   // Ignore if memory is dead, or self-loop
 459   Node *mem = in(MemNode::Memory);
 460   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
 461   assert(mem != this, "dead loop in MemNode::Ideal");
 462 
 463   if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
 464     // This memory slice may be dead.
 465     // Delay this mem node transformation until the memory is processed.
 466     igvn->_worklist.push(this);
 467     return NodeSentinel; // caller will return null
 468   }
 469 
 470   Node *address = in(MemNode::Address);
 471   const Type *t_adr = phase->type(address);
 472   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return null
 473 
 474   if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
 475     // Unsafe off-heap access with zero address. Remove access and other control users
 476     // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
 477     assert(ctl != nullptr, "unsafe accesses should be control dependent");
 478     for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
 479       Node* u = ctl->fast_out(i);
 480       if (u != ctl) {
 481         igvn->rehash_node_delayed(u);
 482         int nb = u->replace_edge(ctl, phase->C->top(), igvn);
 483         --i, imax -= nb;
 484       }
 485     }
 486     Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
 487     Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
 488     phase->C->root()->add_req(halt);
 489     return this;
 490   }
 491 
 492   if (can_reshape && igvn != nullptr &&
 493       (igvn->_worklist.member(address) ||
 494        (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
 495     // The address's base and type may change when the address is processed.
 496     // Delay this mem node transformation until the address is processed.
 497     igvn->_worklist.push(this);
 498     return NodeSentinel; // caller will return null
 499   }
 500 
 501   // Do NOT remove or optimize the next lines: ensure a new alias index
 502   // is allocated for an oop pointer type before Escape Analysis.
 503   // Note: C++ will not remove it since the call has side effect.
 504   if (t_adr->isa_oopptr()) {
 505     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 506   }
 507 
 508   Node* base = nullptr;
 509   if (address->is_AddP()) {
 510     base = address->in(AddPNode::Base);
 511   }
 512   if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 513       !t_adr->isa_rawptr()) {
 514     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 515     // Skip this node optimization if its address has TOP base.
 516     return NodeSentinel; // caller will return null
 517   }
 518 
 519   // Avoid independent memory operations
 520   Node* old_mem = mem;
 521 
 522   // The code which unhooks non-raw memories from complete (macro-expanded)
 523   // initializations was removed. After macro-expansion all stores caught
 524   // by Initialize node became raw stores and there is no information
 525   // which memory slices they modify. So it is unsafe to move any memory
 526   // operation above these stores. Also in most cases hooked non-raw memories
 527   // were already unhooked by using information from detect_ptr_independence()
 528   // and find_previous_store().
 529 
 530   if (mem->is_MergeMem()) {
 531     MergeMemNode* mmem = mem->as_MergeMem();
 532     const TypePtr *tp = t_adr->is_ptr();
 533 
 534     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 535   }
 536 
 537   if (mem != old_mem) {
 538     set_req_X(MemNode::Memory, mem, phase);
 539     if (phase->type(mem) == Type::TOP) return NodeSentinel;
 540     return this;
 541   }
 542 
 543   // let the subclass continue analyzing...
 544   return nullptr;
 545 }
 546 
 547 // Helper function for proving some simple control dominations.
 548 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 549 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 550 // is not a constant (dominated by the method's StartNode).
 551 // Used by MemNode::find_previous_store to prove that the
 552 // control input of a memory operation predates (dominates)
 553 // an allocation it wants to look past.
 554 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
 555 // dominate 'sub', 'DomResult::NotDominate' if not,
 556 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
 557 // dead code, but at the end of IGVN, we know the definite result
 558 // once the dead code is cleaned up.
 559 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
 560   if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
 561     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 562   }
 563 
 564   // Check 'dom'. Skip Proj and CatchProj nodes.
 565   dom = dom->find_exact_control(dom);
 566   if (dom == nullptr || dom->is_top()) {
 567     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 568   }
 569 
 570   if (dom == sub) {
 571     // For the case when, for example, 'sub' is Initialize and the original
 572     // 'dom' is Proj node of the 'sub'.
 573     return DomResult::NotDominate;
 574   }
 575 
 576   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
 577     return DomResult::Dominate;
 578   }
 579 
 580   // 'dom' dominates 'sub' if its control edge and control edges
 581   // of all its inputs dominate or equal to sub's control edge.
 582 
 583   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 584   // Or Region for the check in LoadNode::Ideal();
 585   // 'sub' should have sub->in(0) != nullptr.
 586   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 587          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 588 
 589   // Get control edge of 'sub'.
 590   Node* orig_sub = sub;
 591   sub = sub->find_exact_control(sub->in(0));
 592   if (sub == nullptr || sub->is_top()) {
 593     return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 594   }
 595 
 596   assert(sub->is_CFG(), "expecting control");
 597 
 598   if (sub == dom) {
 599     return DomResult::Dominate;
 600   }
 601 
 602   if (sub->is_Start() || sub->is_Root()) {
 603     return DomResult::NotDominate;
 604   }
 605 
 606   {
 607     // Check all control edges of 'dom'.
 608 
 609     ResourceMark rm;
 610     Node_List nlist;
 611     Unique_Node_List dom_list;
 612 
 613     dom_list.push(dom);
 614     bool only_dominating_controls = false;
 615 
 616     for (uint next = 0; next < dom_list.size(); next++) {
 617       Node* n = dom_list.at(next);
 618       if (n == orig_sub) {
 619         return DomResult::NotDominate; // One of dom's inputs dominated by sub.
 620       }
 621       if (!n->is_CFG() && n->pinned()) {
 622         // Check only own control edge for pinned non-control nodes.
 623         n = n->find_exact_control(n->in(0));
 624         if (n == nullptr || n->is_top()) {
 625           return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 626         }
 627         assert(n->is_CFG(), "expecting control");
 628         dom_list.push(n);
 629       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 630         only_dominating_controls = true;
 631       } else if (n->is_CFG()) {
 632         DomResult dom_result = n->dominates(sub, nlist);
 633         if (dom_result == DomResult::Dominate) {
 634           only_dominating_controls = true;
 635         } else {
 636           return dom_result;
 637         }
 638       } else {
 639         // First, own control edge.
 640         Node* m = n->find_exact_control(n->in(0));
 641         if (m != nullptr) {
 642           if (m->is_top()) {
 643             return DomResult::EncounteredDeadCode; // Conservative answer for dead code
 644           }
 645           dom_list.push(m);
 646         }
 647         // Now, the rest of edges.
 648         uint cnt = n->req();
 649         for (uint i = 1; i < cnt; i++) {
 650           m = n->find_exact_control(n->in(i));
 651           if (m == nullptr || m->is_top()) {
 652             continue;
 653           }
 654           dom_list.push(m);
 655         }
 656       }
 657     }
 658     return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
 659   }
 660 }
 661 
 662 //---------------------detect_ptr_independence---------------------------------
 663 // Used by MemNode::find_previous_store to prove that two base
 664 // pointers are never equal.
 665 // The pointers are accompanied by their associated allocations,
 666 // if any, which have been previously discovered by the caller.
 667 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 668                                       Node* p2, AllocateNode* a2,
 669                                       PhaseTransform* phase) {
 670   // Attempt to prove that these two pointers cannot be aliased.
 671   // They may both manifestly be allocations, and they should differ.
 672   // Or, if they are not both allocations, they can be distinct constants.
 673   // Otherwise, one is an allocation and the other a pre-existing value.
 674   if (a1 == nullptr && a2 == nullptr) {           // neither an allocation
 675     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 676   } else if (a1 != nullptr && a2 != nullptr) {    // both allocations
 677     return (a1 != a2);
 678   } else if (a1 != nullptr) {                  // one allocation a1
 679     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 680     return all_controls_dominate(p2, a1);
 681   } else { //(a2 != null)                   // one allocation a2
 682     return all_controls_dominate(p1, a2);
 683   }
 684   return false;
 685 }
 686 
 687 
 688 // Find an arraycopy ac that produces the memory state represented by parameter mem.
 689 // Return ac if
 690 // (a) can_see_stored_value=true  and ac must have set the value for this load or if
 691 // (b) can_see_stored_value=false and ac could have set the value for this load or if
 692 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
 693 // In case (c) change the parameter mem to the memory input of ac to skip it
 694 // when searching stored value.
 695 // Otherwise return null.
 696 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 697   ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
 698   if (ac != nullptr) {
 699     Node* ld_addp = in(MemNode::Address);
 700     Node* src = ac->in(ArrayCopyNode::Src);
 701     const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
 702 
 703     // This is a load from a cloned array. The corresponding arraycopy ac must
 704     // have set the value for the load and we can return ac but only if the load
 705     // is known to be within bounds. This is checked below.
 706     if (ary_t != nullptr && ld_addp->is_AddP()) {
 707       Node* ld_offs = ld_addp->in(AddPNode::Offset);
 708       BasicType ary_elem = ary_t->elem()->array_element_basic_type();
 709       jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 710       jlong elemsize = type2aelembytes(ary_elem);
 711 
 712       const TypeX*   ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 713       const TypeInt* sizetype  = ary_t->size();
 714 
 715       if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
 716         // The load is known to be within bounds. It receives its value from ac.
 717         return ac;
 718       }
 719       // The load is known to be out-of-bounds.
 720     }
 721     // The load could be out-of-bounds. It must not be hoisted but must remain
 722     // dependent on the runtime range check. This is achieved by returning null.
 723   } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
 724     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 725 
 726     if (ac->is_arraycopy_validated() ||
 727         ac->is_copyof_validated() ||
 728         ac->is_copyofrange_validated()) {
 729       Node* ld_addp = in(MemNode::Address);
 730       if (ld_addp->is_AddP()) {
 731         Node* ld_base = ld_addp->in(AddPNode::Address);
 732         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 733 
 734         Node* dest = ac->in(ArrayCopyNode::Dest);
 735 
 736         if (dest == ld_base) {
 737           const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 738           assert(!ld_offs_t->empty(), "dead reference should be checked already");
 739           // Take into account vector or unsafe access size
 740           jlong ld_size_in_bytes = (jlong)memory_size();
 741           jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
 742           offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
 743           if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
 744             return ac;
 745           }
 746           if (!can_see_stored_value) {
 747             mem = ac->in(TypeFunc::Memory);
 748             return ac;
 749           }
 750         }
 751       }
 752     }
 753   }
 754   return nullptr;
 755 }
 756 
 757 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
 758   if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 759                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 760     if (ld_alloc != nullptr) {
 761       // Check if there is an array copy for a clone
 762       Node* mb = mem->in(0);
 763       ArrayCopyNode* ac = nullptr;
 764       if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
 765           mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
 766         ac = mb->in(0)->in(0)->as_ArrayCopy();
 767       } else {
 768         // Step over GC barrier when ReduceInitialCardMarks is disabled
 769         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 770         Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
 771 
 772         if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
 773           ac = control_proj_ac->in(0)->as_ArrayCopy();
 774         }
 775       }
 776 
 777       if (ac != nullptr && ac->is_clonebasic()) {
 778         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
 779         if (alloc != nullptr && alloc == ld_alloc) {
 780           return ac;
 781         }
 782       }
 783     }
 784   }
 785   return nullptr;
 786 }
 787 
 788 // The logic for reordering loads and stores uses four steps:
 789 // (a) Walk carefully past stores and initializations which we
 790 //     can prove are independent of this load.
 791 // (b) Observe that the next memory state makes an exact match
 792 //     with self (load or store), and locate the relevant store.
 793 // (c) Ensure that, if we were to wire self directly to the store,
 794 //     the optimizer would fold it up somehow.
 795 // (d) Do the rewiring, and return, depending on some other part of
 796 //     the optimizer to fold up the load.
 797 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 798 // specific to loads and stores, so they are handled by the callers.
 799 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 800 //
 801 Node* MemNode::find_previous_store(PhaseValues* phase) {
 802   Node*         ctrl   = in(MemNode::Control);
 803   Node*         adr    = in(MemNode::Address);
 804   intptr_t      offset = 0;
 805   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 806   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base);
 807 
 808   if (offset == Type::OffsetBot)
 809     return nullptr;            // cannot unalias unless there are precise offsets
 810 
 811   const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
 812   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 813 
 814   intptr_t size_in_bytes = memory_size();
 815 
 816   Node* mem = in(MemNode::Memory);   // start searching here...
 817 
 818   int cnt = 50;             // Cycle limiter
 819   for (;;) {                // While we can dance past unrelated stores...
 820     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 821 
 822     Node* prev = mem;
 823     if (mem->is_Store()) {
 824       Node* st_adr = mem->in(MemNode::Address);
 825       intptr_t st_offset = 0;
 826       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 827       if (st_base == nullptr)
 828         break;              // inscrutable pointer
 829 
 830       // For raw accesses it's not enough to prove that constant offsets don't intersect.
 831       // We need the bases to be the equal in order for the offset check to make sense.
 832       if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
 833         break;
 834       }
 835 
 836       if (st_offset != offset && st_offset != Type::OffsetBot) {
 837         const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
 838         assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
 839         if (st_offset >= offset + size_in_bytes ||
 840             st_offset <= offset - MAX_STORE ||
 841             st_offset <= offset - mem->as_Store()->memory_size()) {
 842           // Success:  The offsets are provably independent.
 843           // (You may ask, why not just test st_offset != offset and be done?
 844           // The answer is that stores of different sizes can co-exist
 845           // in the same sequence of RawMem effects.  We sometimes initialize
 846           // a whole 'tile' of array elements with a single jint or jlong.)
 847           mem = mem->in(MemNode::Memory);
 848           continue;           // (a) advance through independent store memory
 849         }
 850       }
 851       if (st_base != base &&
 852           detect_ptr_independence(base, alloc,
 853                                   st_base,
 854                                   AllocateNode::Ideal_allocation(st_base),
 855                                   phase)) {
 856         // Success:  The bases are provably independent.
 857         mem = mem->in(MemNode::Memory);
 858         continue;           // (a) advance through independent store memory
 859       }
 860 
 861       // (b) At this point, if the bases or offsets do not agree, we lose,
 862       // since we have not managed to prove 'this' and 'mem' independent.
 863       if (st_base == base && st_offset == offset) {
 864         return mem;         // let caller handle steps (c), (d)
 865       }
 866 
 867     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 868       InitializeNode* st_init = mem->in(0)->as_Initialize();
 869       AllocateNode*  st_alloc = st_init->allocation();
 870       if (st_alloc == nullptr) {
 871         break;              // something degenerated
 872       }
 873       bool known_identical = false;
 874       bool known_independent = false;
 875       if (alloc == st_alloc) {
 876         known_identical = true;
 877       } else if (alloc != nullptr) {
 878         known_independent = true;
 879       } else if (all_controls_dominate(this, st_alloc)) {
 880         known_independent = true;
 881       }
 882 
 883       if (known_independent) {
 884         // The bases are provably independent: Either they are
 885         // manifestly distinct allocations, or else the control
 886         // of this load dominates the store's allocation.
 887         int alias_idx = phase->C->get_alias_index(adr_type());
 888         if (alias_idx == Compile::AliasIdxRaw) {
 889           mem = st_alloc->in(TypeFunc::Memory);
 890         } else {
 891           mem = st_init->memory(alias_idx);
 892         }
 893         continue;           // (a) advance through independent store memory
 894       }
 895 
 896       // (b) at this point, if we are not looking at a store initializing
 897       // the same allocation we are loading from, we lose.
 898       if (known_identical) {
 899         // From caller, can_see_stored_value will consult find_captured_store.
 900         return mem;         // let caller handle steps (c), (d)
 901       }
 902 
 903     } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
 904       if (prev != mem) {
 905         // Found an arraycopy but it doesn't affect that load
 906         continue;
 907       }
 908       // Found an arraycopy that may affect that load
 909       return mem;
 910     } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
 911       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 912       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 913         // ArrayCopyNodes processed here as well.
 914         CallNode *call = mem->in(0)->as_Call();
 915         if (!call->may_modify(addr_t, phase)) {
 916           mem = call->in(TypeFunc::Memory);
 917           continue;         // (a) advance through independent call memory
 918         }
 919       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 920         ArrayCopyNode* ac = nullptr;
 921         if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
 922           break;
 923         }
 924         mem = mem->in(0)->in(TypeFunc::Memory);
 925         continue;           // (a) advance through independent MemBar memory
 926       } else if (mem->is_ClearArray()) {
 927         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 928           // (the call updated 'mem' value)
 929           continue;         // (a) advance through independent allocation memory
 930         } else {
 931           // Can not bypass initialization of the instance
 932           // we are looking for.
 933           return mem;
 934         }
 935       } else if (mem->is_MergeMem()) {
 936         int alias_idx = phase->C->get_alias_index(adr_type());
 937         mem = mem->as_MergeMem()->memory_at(alias_idx);
 938         continue;           // (a) advance through independent MergeMem memory
 939       }
 940     }
 941 
 942     // Unless there is an explicit 'continue', we must bail out here,
 943     // because 'mem' is an inscrutable memory state (e.g., a call).
 944     break;
 945   }
 946 
 947   return nullptr;              // bail out
 948 }
 949 
 950 //----------------------calculate_adr_type-------------------------------------
 951 // Helper function.  Notices when the given type of address hits top or bottom.
 952 // Also, asserts a cross-check of the type against the expected address type.
 953 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 954   if (t == Type::TOP)  return nullptr; // does not touch memory any more?
 955   #ifdef ASSERT
 956   if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump())  cross_check = nullptr;
 957   #endif
 958   const TypePtr* tp = t->isa_ptr();
 959   if (tp == nullptr) {
 960     assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 961     return TypePtr::BOTTOM;           // touches lots of memory
 962   } else {
 963     #ifdef ASSERT
 964     // %%%% [phh] We don't check the alias index if cross_check is
 965     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 966     if (cross_check != nullptr &&
 967         cross_check != TypePtr::BOTTOM &&
 968         cross_check != TypeRawPtr::BOTTOM) {
 969       // Recheck the alias index, to see if it has changed (due to a bug).
 970       Compile* C = Compile::current();
 971       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 972              "must stay in the original alias category");
 973       // The type of the address must be contained in the adr_type,
 974       // disregarding "null"-ness.
 975       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 976       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 977       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 978              "real address must not escape from expected memory type");
 979     }
 980     #endif
 981     return tp;
 982   }
 983 }
 984 
 985 uint8_t MemNode::barrier_data(const Node* n) {
 986   if (n->is_LoadStore()) {
 987     return n->as_LoadStore()->barrier_data();
 988   } else if (n->is_Mem()) {
 989     return n->as_Mem()->barrier_data();
 990   }
 991   return 0;
 992 }
 993 
 994 //=============================================================================
 995 // Should LoadNode::Ideal() attempt to remove control edges?
 996 bool LoadNode::can_remove_control() const {
 997   return !has_pinned_control_dependency();
 998 }
 999 uint LoadNode::size_of() const { return sizeof(*this); }
1000 bool LoadNode::cmp(const Node &n) const {
1001   LoadNode& load = (LoadNode &)n;
1002   return Type::equals(_type, load._type) &&
1003          _control_dependency == load._control_dependency &&
1004          _mo == load._mo;
1005 }
1006 const Type *LoadNode::bottom_type() const { return _type; }
1007 uint LoadNode::ideal_reg() const {
1008   return _type->ideal_reg();
1009 }
1010 
1011 #ifndef PRODUCT
1012 void LoadNode::dump_spec(outputStream *st) const {
1013   MemNode::dump_spec(st);
1014   if( !Verbose && !WizardMode ) {
1015     // standard dump does this in Verbose and WizardMode
1016     st->print(" #"); _type->dump_on(st);
1017   }
1018   if (!depends_only_on_test()) {
1019     st->print(" (does not depend only on test, ");
1020     if (control_dependency() == UnknownControl) {
1021       st->print("unknown control");
1022     } else if (control_dependency() == Pinned) {
1023       st->print("pinned");
1024     } else if (adr_type() == TypeRawPtr::BOTTOM) {
1025       st->print("raw access");
1026     } else {
1027       st->print("unknown reason");
1028     }
1029     st->print(")");
1030   }
1031 }
1032 #endif
1033 
1034 #ifdef ASSERT
1035 //----------------------------is_immutable_value-------------------------------
1036 // Helper function to allow a raw load without control edge for some cases
1037 bool LoadNode::is_immutable_value(Node* adr) {
1038   if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1039       adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
1040 
1041     jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1042     int offsets[] = {
1043       in_bytes(JavaThread::osthread_offset()),
1044       in_bytes(JavaThread::threadObj_offset()),
1045       in_bytes(JavaThread::vthread_offset()),
1046       in_bytes(JavaThread::scopedValueCache_offset()),
1047     };
1048 
1049     for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1050       if (offset == offsets[i]) {
1051         return true;
1052       }
1053     }
1054   }
1055 
1056   return false;
1057 }
1058 #endif
1059 
1060 //----------------------------LoadNode::make-----------------------------------
1061 // Polymorphic factory method:
1062 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1063                      ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1064   Compile* C = gvn.C;
1065 
1066   // sanity check the alias category against the created node type
1067   assert(!(adr_type->isa_oopptr() &&
1068            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1069          "use LoadKlassNode instead");
1070   assert(!(adr_type->isa_aryptr() &&
1071            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1072          "use LoadRangeNode instead");
1073   // Check control edge of raw loads
1074   assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1075           // oop will be recorded in oop map if load crosses safepoint
1076           rt->isa_oopptr() || is_immutable_value(adr),
1077           "raw memory operations should have control edge");
1078   LoadNode* load = nullptr;
1079   switch (bt) {
1080   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1081   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1082   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1083   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1084   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
1085   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1086   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
1087   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency, require_atomic_access); break;
1088   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
1089   case T_OBJECT:
1090   case T_NARROWOOP:
1091 #ifdef _LP64
1092     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1093       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1094     } else
1095 #endif
1096     {
1097       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1098       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1099     }
1100     break;
1101   default:
1102     ShouldNotReachHere();
1103     break;
1104   }
1105   assert(load != nullptr, "LoadNode should have been created");
1106   if (unaligned) {
1107     load->set_unaligned_access();
1108   }
1109   if (mismatched) {
1110     load->set_mismatched_access();
1111   }
1112   if (unsafe) {
1113     load->set_unsafe_access();
1114   }
1115   load->set_barrier_data(barrier_data);
1116   if (load->Opcode() == Op_LoadN) {
1117     Node* ld = gvn.transform(load);
1118     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1119   }
1120 
1121   return load;
1122 }
1123 
1124 //------------------------------hash-------------------------------------------
1125 uint LoadNode::hash() const {
1126   // unroll addition of interesting fields
1127   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1128 }
1129 
1130 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1131   if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1132     bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1133     bool is_stable_ary = FoldStableValues &&
1134                          (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1135                          tp->isa_aryptr()->is_stable();
1136 
1137     return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1138   }
1139 
1140   return false;
1141 }
1142 
1143 LoadNode* LoadNode::pin_array_access_node() const {
1144   const TypePtr* adr_type = this->adr_type();
1145   if (adr_type != nullptr && adr_type->isa_aryptr()) {
1146     return clone_pinned();
1147   }
1148   return nullptr;
1149 }
1150 
1151 // Is the value loaded previously stored by an arraycopy? If so return
1152 // a load node that reads from the source array so we may be able to
1153 // optimize out the ArrayCopy node later.
1154 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1155   Node* ld_adr = in(MemNode::Address);
1156   intptr_t ld_off = 0;
1157   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1158   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1159   if (ac != nullptr) {
1160     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1161 
1162     Node* mem = ac->in(TypeFunc::Memory);
1163     Node* ctl = ac->in(0);
1164     Node* src = ac->in(ArrayCopyNode::Src);
1165 
1166     if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1167       return nullptr;
1168     }
1169 
1170     // load depends on the tests that validate the arraycopy
1171     LoadNode* ld = clone_pinned();
1172     Node* addp = in(MemNode::Address)->clone();
1173     if (ac->as_ArrayCopy()->is_clonebasic()) {
1174       assert(ld_alloc != nullptr, "need an alloc");
1175       assert(addp->is_AddP(), "address must be addp");
1176       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1177       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1178       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1179       addp->set_req(AddPNode::Base, src);
1180       addp->set_req(AddPNode::Address, src);
1181     } else {
1182       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1183              ac->as_ArrayCopy()->is_copyof_validated() ||
1184              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1185       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1186       addp->set_req(AddPNode::Base, src);
1187       addp->set_req(AddPNode::Address, src);
1188 
1189       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1190       BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1191       if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1192 
1193       uint shift  = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1194 
1195       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1196 #ifdef _LP64
1197       diff = phase->transform(new ConvI2LNode(diff));
1198 #endif
1199       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1200 
1201       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1202       addp->set_req(AddPNode::Offset, offset);
1203     }
1204     addp = phase->transform(addp);
1205 #ifdef ASSERT
1206     const TypePtr* adr_type = phase->type(addp)->is_ptr();
1207     ld->_adr_type = adr_type;
1208 #endif
1209     ld->set_req(MemNode::Address, addp);
1210     ld->set_req(0, ctl);
1211     ld->set_req(MemNode::Memory, mem);
1212     return ld;
1213   }
1214   return nullptr;
1215 }
1216 
1217 static Node* see_through_inline_type(PhaseValues* phase, const MemNode* load, Node* base, int offset) {
1218   if (!load->is_mismatched_access() && base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1219     InlineTypeNode* vt = base->as_InlineType();
1220     Node* value = vt->field_value_by_offset(offset, true);
1221     assert(value != nullptr, "must see some value");
1222     return value;
1223   }
1224 
1225   return nullptr;
1226 }
1227 
1228 //---------------------------can_see_stored_value------------------------------
1229 // This routine exists to make sure this set of tests is done the same
1230 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
1231 // will change the graph shape in a way which makes memory alive twice at the
1232 // same time (uses the Oracle model of aliasing), then some
1233 // LoadXNode::Identity will fold things back to the equivalence-class model
1234 // of aliasing.
1235 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1236   Node* ld_adr = in(MemNode::Address);
1237   intptr_t ld_off = 0;
1238   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1239   // Try to see through an InlineTypeNode
1240   // LoadN is special because the input is not compressed
1241   if (Opcode() != Op_LoadN) {
1242     Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1243     if (value != nullptr) {
1244       return value;
1245     }
1246   }
1247 
1248   Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1249   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1250   Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1251   // This is more general than load from boxing objects.
1252   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1253     uint alias_idx = atp->index();
1254     Node* result = nullptr;
1255     Node* current = st;
1256     // Skip through chains of MemBarNodes checking the MergeMems for
1257     // new states for the slice of this load.  Stop once any other
1258     // kind of node is encountered.  Loads from final memory can skip
1259     // through any kind of MemBar but normal loads shouldn't skip
1260     // through MemBarAcquire since the could allow them to move out of
1261     // a synchronized region. It is not safe to step over MemBarCPUOrder,
1262     // because alias info above them may be inaccurate (e.g., due to
1263     // mixed/mismatched unsafe accesses).
1264     bool is_final_mem = !atp->is_rewritable();
1265     while (current->is_Proj()) {
1266       int opc = current->in(0)->Opcode();
1267       if ((is_final_mem && (opc == Op_MemBarAcquire ||
1268                             opc == Op_MemBarAcquireLock ||
1269                             opc == Op_LoadFence)) ||
1270           opc == Op_MemBarRelease ||
1271           opc == Op_StoreFence ||
1272           opc == Op_MemBarReleaseLock ||
1273           opc == Op_MemBarStoreStore ||
1274           opc == Op_StoreStoreFence) {
1275         Node* mem = current->in(0)->in(TypeFunc::Memory);
1276         if (mem->is_MergeMem()) {
1277           MergeMemNode* merge = mem->as_MergeMem();
1278           Node* new_st = merge->memory_at(alias_idx);
1279           if (new_st == merge->base_memory()) {
1280             // Keep searching
1281             current = new_st;
1282             continue;
1283           }
1284           // Save the new memory state for the slice and fall through
1285           // to exit.
1286           result = new_st;
1287         }
1288       }
1289       break;
1290     }
1291     if (result != nullptr) {
1292       st = result;
1293     }
1294   }
1295 
1296   // Loop around twice in the case Load -> Initialize -> Store.
1297   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1298   for (int trip = 0; trip <= 1; trip++) {
1299 
1300     if (st->is_Store()) {
1301       Node* st_adr = st->in(MemNode::Address);
1302       if (st_adr != ld_adr) {
1303         // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1304         intptr_t st_off = 0;
1305         Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1306         if (ld_base == nullptr)                                return nullptr;
1307         if (st_base == nullptr)                                return nullptr;
1308         if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1309         if (ld_off != st_off)                                  return nullptr;
1310         if (ld_off == Type::OffsetBot)                         return nullptr;
1311         // Same base, same offset.
1312         // Possible improvement for arrays: check index value instead of absolute offset.
1313 
1314         // At this point we have proven something like this setup:
1315         //   B = << base >>
1316         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1317         //   S = StoreQ(AddP(             B , #Off), V)
1318         // (Actually, we haven't yet proven the Q's are the same.)
1319         // In other words, we are loading from a casted version of
1320         // the same pointer-and-offset that we stored to.
1321         // Casted version may carry a dependency and it is respected.
1322         // Thus, we are able to replace L by V.
1323       }
1324       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1325       if (store_Opcode() != st->Opcode()) {
1326         return nullptr;
1327       }
1328       // LoadVector/StoreVector needs additional check to ensure the types match.
1329       if (st->is_StoreVector()) {
1330         const TypeVect*  in_vt = st->as_StoreVector()->vect_type();
1331         const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1332         if (in_vt != out_vt) {
1333           return nullptr;
1334         }
1335       }
1336       return st->in(MemNode::ValueIn);
1337     }
1338 
1339     // A load from a freshly-created object always returns zero.
1340     // (This can happen after LoadNode::Ideal resets the load's memory input
1341     // to find_captured_store, which returned InitializeNode::zero_memory.)
1342     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1343         (st->in(0) == ld_alloc) &&
1344         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1345       // return a zero value for the load's basic type
1346       // (This is one of the few places where a generic PhaseTransform
1347       // can create new nodes.  Think of it as lazily manifesting
1348       // virtually pre-existing constants.)
1349       Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1350       if (init_value != nullptr) {
1351         // TODO 8350865 Scalar replacement does not work well for flat arrays.
1352         // Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
1353         return init_value;
1354       }
1355       assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1356       if (value_basic_type() != T_VOID) {
1357         if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1358           // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1359           // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1360           // by the ArrayCopyNode.
1361           return phase->zerocon(value_basic_type());
1362         }
1363       } else {
1364         // TODO: materialize all-zero vector constant
1365         assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1366       }
1367     }
1368 
1369     // A load from an initialization barrier can match a captured store.
1370     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1371       InitializeNode* init = st->in(0)->as_Initialize();
1372       AllocateNode* alloc = init->allocation();
1373       if ((alloc != nullptr) && (alloc == ld_alloc)) {
1374         // examine a captured store value
1375         st = init->find_captured_store(ld_off, memory_size(), phase);
1376         if (st != nullptr) {
1377           continue;             // take one more trip around
1378         }
1379       }
1380     }
1381 
1382     // Load boxed value from result of valueOf() call is input parameter.
1383     if (this->is_Load() && ld_adr->is_AddP() &&
1384         (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1385       intptr_t ignore = 0;
1386       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1387       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1388       base = bs->step_over_gc_barrier(base);
1389       if (base != nullptr && base->is_Proj() &&
1390           base->as_Proj()->_con == TypeFunc::Parms &&
1391           base->in(0)->is_CallStaticJava() &&
1392           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1393         return base->in(0)->in(TypeFunc::Parms);
1394       }
1395     }
1396 
1397     break;
1398   }
1399 
1400   return nullptr;
1401 }
1402 
1403 //----------------------is_instance_field_load_with_local_phi------------------
1404 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1405   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1406       in(Address)->is_AddP() ) {
1407     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1408     // Only instances and boxed values.
1409     if( t_oop != nullptr &&
1410         (t_oop->is_ptr_to_boxed_value() ||
1411          t_oop->is_known_instance_field()) &&
1412         t_oop->offset() != Type::OffsetBot &&
1413         t_oop->offset() != Type::OffsetTop) {
1414       return true;
1415     }
1416   }
1417   return false;
1418 }
1419 
1420 //------------------------------Identity---------------------------------------
1421 // Loads are identity if previous store is to same address
1422 Node* LoadNode::Identity(PhaseGVN* phase) {
1423   // If the previous store-maker is the right kind of Store, and the store is
1424   // to the same address, then we are equal to the value stored.
1425   Node* mem = in(Memory);
1426   Node* value = can_see_stored_value(mem, phase);
1427   if( value ) {
1428     // byte, short & char stores truncate naturally.
1429     // A load has to load the truncated value which requires
1430     // some sort of masking operation and that requires an
1431     // Ideal call instead of an Identity call.
1432     if (memory_size() < BytesPerInt) {
1433       // If the input to the store does not fit with the load's result type,
1434       // it must be truncated via an Ideal call.
1435       if (!phase->type(value)->higher_equal(phase->type(this)))
1436         return this;
1437     }
1438     // (This works even when value is a Con, but LoadNode::Value
1439     // usually runs first, producing the singleton type of the Con.)
1440     if (!has_pinned_control_dependency() || value->is_Con()) {
1441       return value;
1442     } else {
1443       return this;
1444     }
1445   }
1446 
1447   if (has_pinned_control_dependency()) {
1448     return this;
1449   }
1450   // Search for an existing data phi which was generated before for the same
1451   // instance's field to avoid infinite generation of phis in a loop.
1452   Node *region = mem->in(0);
1453   if (is_instance_field_load_with_local_phi(region)) {
1454     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1455     int this_index  = phase->C->get_alias_index(addr_t);
1456     int this_offset = addr_t->offset();
1457     int this_iid    = addr_t->instance_id();
1458     if (!addr_t->is_known_instance() &&
1459          addr_t->is_ptr_to_boxed_value()) {
1460       // Use _idx of address base (could be Phi node) for boxed values.
1461       intptr_t   ignore = 0;
1462       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1463       if (base == nullptr) {
1464         return this;
1465       }
1466       this_iid = base->_idx;
1467     }
1468     const Type* this_type = bottom_type();
1469     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1470       Node* phi = region->fast_out(i);
1471       if (phi->is_Phi() && phi != mem &&
1472           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1473         return phi;
1474       }
1475     }
1476   }
1477 
1478   return this;
1479 }
1480 
1481 // Construct an equivalent unsigned load.
1482 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1483   BasicType bt = T_ILLEGAL;
1484   const Type* rt = nullptr;
1485   switch (Opcode()) {
1486     case Op_LoadUB: return this;
1487     case Op_LoadUS: return this;
1488     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1489     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1490     default:
1491       assert(false, "no unsigned variant: %s", Name());
1492       return nullptr;
1493   }
1494   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1495                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1496                         false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1497 }
1498 
1499 // Construct an equivalent signed load.
1500 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1501   BasicType bt = T_ILLEGAL;
1502   const Type* rt = nullptr;
1503   switch (Opcode()) {
1504     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1505     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1506     case Op_LoadB: // fall through
1507     case Op_LoadS: // fall through
1508     case Op_LoadI: // fall through
1509     case Op_LoadL: return this;
1510     default:
1511       assert(false, "no signed variant: %s", Name());
1512       return nullptr;
1513   }
1514   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1515                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1516                         false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1517 }
1518 
1519 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1520   BasicType bt = rt->basic_type();
1521   switch (Opcode()) {
1522     case Op_LoadI: return (bt == T_FLOAT);
1523     case Op_LoadL: return (bt == T_DOUBLE);
1524     case Op_LoadF: return (bt == T_INT);
1525     case Op_LoadD: return (bt == T_LONG);
1526 
1527     default: return false;
1528   }
1529 }
1530 
1531 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1532   BasicType bt = rt->basic_type();
1533   assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1534   bool is_mismatched = is_mismatched_access();
1535   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1536   if (raw_type == nullptr) {
1537     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1538   }
1539   const int op = Opcode();
1540   bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1541                                (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1542   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1543                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1544                         require_atomic_access, is_unaligned_access(), is_mismatched);
1545 }
1546 
1547 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1548   BasicType bt = vt->basic_type();
1549   switch (Opcode()) {
1550     case Op_StoreI: return (bt == T_FLOAT);
1551     case Op_StoreL: return (bt == T_DOUBLE);
1552     case Op_StoreF: return (bt == T_INT);
1553     case Op_StoreD: return (bt == T_LONG);
1554 
1555     default: return false;
1556   }
1557 }
1558 
1559 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1560   BasicType bt = vt->basic_type();
1561   assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1562   const int op = Opcode();
1563   bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1564                                (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1565   StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1566                                   raw_adr_type(), val, bt, _mo, require_atomic_access);
1567 
1568   bool is_mismatched = is_mismatched_access();
1569   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1570   if (raw_type == nullptr) {
1571     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1572   }
1573   if (is_mismatched) {
1574     st->set_mismatched_access();
1575   }
1576   return st;
1577 }
1578 
1579 // We're loading from an object which has autobox behaviour.
1580 // If this object is result of a valueOf call we'll have a phi
1581 // merging a newly allocated object and a load from the cache.
1582 // We want to replace this load with the original incoming
1583 // argument to the valueOf call.
1584 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1585   assert(igvn->C->eliminate_boxing(), "sanity");
1586   intptr_t ignore = 0;
1587   Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1588   if ((base == nullptr) || base->is_Phi()) {
1589     // Push the loads from the phi that comes from valueOf up
1590     // through it to allow elimination of the loads and the recovery
1591     // of the original value. It is done in split_through_phi().
1592     return nullptr;
1593   } else if (base->is_Load() ||
1594              (base->is_DecodeN() && base->in(1)->is_Load())) {
1595     // Eliminate the load of boxed value for integer types from the cache
1596     // array by deriving the value from the index into the array.
1597     // Capture the offset of the load and then reverse the computation.
1598 
1599     // Get LoadN node which loads a boxing object from 'cache' array.
1600     if (base->is_DecodeN()) {
1601       base = base->in(1);
1602     }
1603     if (!base->in(Address)->is_AddP()) {
1604       return nullptr; // Complex address
1605     }
1606     AddPNode* address = base->in(Address)->as_AddP();
1607     Node* cache_base = address->in(AddPNode::Base);
1608     if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1609       // Get ConP node which is static 'cache' field.
1610       cache_base = cache_base->in(1);
1611     }
1612     if ((cache_base != nullptr) && cache_base->is_Con()) {
1613       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1614       if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1615         Node* elements[4];
1616         int shift = exact_log2(type2aelembytes(T_OBJECT));
1617         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1618         if (count > 0 && elements[0]->is_Con() &&
1619             (count == 1 ||
1620              (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1621                             elements[1]->in(2) == igvn->intcon(shift)))) {
1622           ciObjArray* array = base_type->const_oop()->as_obj_array();
1623           // Fetch the box object cache[0] at the base of the array and get its value
1624           ciInstance* box = array->obj_at(0)->as_instance();
1625           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1626           assert(ik->is_box_klass(), "sanity");
1627           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1628           if (ik->nof_nonstatic_fields() == 1) {
1629             // This should be true nonstatic_field_at requires calling
1630             // nof_nonstatic_fields so check it anyway
1631             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1632             BasicType bt = c.basic_type();
1633             // Only integer types have boxing cache.
1634             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1635                    bt == T_BYTE    || bt == T_SHORT ||
1636                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1637             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1638             if (cache_low != (int)cache_low) {
1639               return nullptr; // should not happen since cache is array indexed by value
1640             }
1641             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1642             if (offset != (int)offset) {
1643               return nullptr; // should not happen since cache is array indexed by value
1644             }
1645            // Add up all the offsets making of the address of the load
1646             Node* result = elements[0];
1647             for (int i = 1; i < count; i++) {
1648               result = igvn->transform(new AddXNode(result, elements[i]));
1649             }
1650             // Remove the constant offset from the address and then
1651             result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1652             // remove the scaling of the offset to recover the original index.
1653             if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1654               // Peel the shift off directly but wrap it in a dummy node
1655               // since Ideal can't return existing nodes
1656               igvn->_worklist.push(result); // remove dead node later
1657               result = new RShiftXNode(result->in(1), igvn->intcon(0));
1658             } else if (result->is_Add() && result->in(2)->is_Con() &&
1659                        result->in(1)->Opcode() == Op_LShiftX &&
1660                        result->in(1)->in(2) == igvn->intcon(shift)) {
1661               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1662               // but for boxing cache access we know that X<<Z will not overflow
1663               // (there is range check) so we do this optimizatrion by hand here.
1664               igvn->_worklist.push(result); // remove dead node later
1665               Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1666               result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1667             } else {
1668               result = new RShiftXNode(result, igvn->intcon(shift));
1669             }
1670 #ifdef _LP64
1671             if (bt != T_LONG) {
1672               result = new ConvL2INode(igvn->transform(result));
1673             }
1674 #else
1675             if (bt == T_LONG) {
1676               result = new ConvI2LNode(igvn->transform(result));
1677             }
1678 #endif
1679             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1680             // Need to preserve unboxing load type if it is unsigned.
1681             switch(this->Opcode()) {
1682               case Op_LoadUB:
1683                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1684                 break;
1685               case Op_LoadUS:
1686                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1687                 break;
1688             }
1689             return result;
1690           }
1691         }
1692       }
1693     }
1694   }
1695   return nullptr;
1696 }
1697 
1698 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1699   Node* region = phi->in(0);
1700   if (region == nullptr) {
1701     return false; // Wait stable graph
1702   }
1703   uint cnt = phi->req();
1704   for (uint i = 1; i < cnt; i++) {
1705     Node* rc = region->in(i);
1706     if (rc == nullptr || phase->type(rc) == Type::TOP)
1707       return false; // Wait stable graph
1708     Node* in = phi->in(i);
1709     if (in == nullptr || phase->type(in) == Type::TOP)
1710       return false; // Wait stable graph
1711   }
1712   return true;
1713 }
1714 
1715 //------------------------------split_through_phi------------------------------
1716 // Check whether a call to 'split_through_phi' would split this load through the
1717 // Phi *base*. This method is essentially a copy of the validations performed
1718 // by 'split_through_phi'. The first use of this method was in EA code as part
1719 // of simplification of allocation merges.
1720 // Some differences from original method (split_through_phi):
1721 //  - If base->is_CastPP(): base = base->in(1)
1722 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1723   Node* mem        = in(Memory);
1724   Node* address    = in(Address);
1725   intptr_t ignore  = 0;
1726   Node*    base    = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1727 
1728   if (base->is_CastPP()) {
1729     base = base->in(1);
1730   }
1731 
1732   if (req() > 3 || base == nullptr || !base->is_Phi()) {
1733     return false;
1734   }
1735 
1736   if (!mem->is_Phi()) {
1737     if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1738       return false;
1739     }
1740   } else if (base->in(0) != mem->in(0)) {
1741     if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1742       return false;
1743     }
1744   }
1745 
1746   return true;
1747 }
1748 
1749 //------------------------------split_through_phi------------------------------
1750 // Split instance or boxed field load through Phi.
1751 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1752   if (req() > 3) {
1753     assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1754     // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1755     return nullptr;
1756   }
1757   Node* mem     = in(Memory);
1758   Node* address = in(Address);
1759   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1760 
1761   assert((t_oop != nullptr) &&
1762          (ignore_missing_instance_id ||
1763           t_oop->is_known_instance_field() ||
1764           t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1765 
1766   Compile* C = phase->C;
1767   intptr_t ignore = 0;
1768   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1769   bool base_is_phi = (base != nullptr) && base->is_Phi();
1770   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1771                            (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1772                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1773 
1774   if (!((mem->is_Phi() || base_is_phi) &&
1775         (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1776     return nullptr; // Neither memory or base are Phi
1777   }
1778 
1779   if (mem->is_Phi()) {
1780     if (!stable_phi(mem->as_Phi(), phase)) {
1781       return nullptr; // Wait stable graph
1782     }
1783     uint cnt = mem->req();
1784     // Check for loop invariant memory.
1785     if (cnt == 3) {
1786       for (uint i = 1; i < cnt; i++) {
1787         Node* in = mem->in(i);
1788         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1789         if (m == mem) {
1790           if (i == 1) {
1791             // if the first edge was a loop, check second edge too.
1792             // If both are replaceable - we are in an infinite loop
1793             Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1794             if (n == mem) {
1795               break;
1796             }
1797           }
1798           set_req(Memory, mem->in(cnt - i));
1799           return this; // made change
1800         }
1801       }
1802     }
1803   }
1804   if (base_is_phi) {
1805     if (!stable_phi(base->as_Phi(), phase)) {
1806       return nullptr; // Wait stable graph
1807     }
1808     uint cnt = base->req();
1809     // Check for loop invariant memory.
1810     if (cnt == 3) {
1811       for (uint i = 1; i < cnt; i++) {
1812         if (base->in(i) == base) {
1813           return nullptr; // Wait stable graph
1814         }
1815       }
1816     }
1817   }
1818 
1819   // Split through Phi (see original code in loopopts.cpp).
1820   assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1821 
1822   // Do nothing here if Identity will find a value
1823   // (to avoid infinite chain of value phis generation).
1824   if (this != Identity(phase)) {
1825     return nullptr;
1826   }
1827 
1828   // Select Region to split through.
1829   Node* region;
1830   DomResult dom_result = DomResult::Dominate;
1831   if (!base_is_phi) {
1832     assert(mem->is_Phi(), "sanity");
1833     region = mem->in(0);
1834     // Skip if the region dominates some control edge of the address.
1835     // We will check `dom_result` later.
1836     dom_result = MemNode::maybe_all_controls_dominate(address, region);
1837   } else if (!mem->is_Phi()) {
1838     assert(base_is_phi, "sanity");
1839     region = base->in(0);
1840     // Skip if the region dominates some control edge of the memory.
1841     // We will check `dom_result` later.
1842     dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1843   } else if (base->in(0) != mem->in(0)) {
1844     assert(base_is_phi && mem->is_Phi(), "sanity");
1845     dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1846     if (dom_result == DomResult::Dominate) {
1847       region = base->in(0);
1848     } else {
1849       dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1850       if (dom_result == DomResult::Dominate) {
1851         region = mem->in(0);
1852       }
1853       // Otherwise we encountered a complex graph.
1854     }
1855   } else {
1856     assert(base->in(0) == mem->in(0), "sanity");
1857     region = mem->in(0);
1858   }
1859 
1860   PhaseIterGVN* igvn = phase->is_IterGVN();
1861   if (dom_result != DomResult::Dominate) {
1862     if (dom_result == DomResult::EncounteredDeadCode) {
1863       // There is some dead code which eventually will be removed in IGVN.
1864       // Once this is the case, we get an unambiguous dominance result.
1865       // Push the node to the worklist again until the dead code is removed.
1866       igvn->_worklist.push(this);
1867     }
1868     return nullptr;
1869   }
1870 
1871   Node* phi = nullptr;
1872   const Type* this_type = this->bottom_type();
1873   if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1874     int this_index = C->get_alias_index(t_oop);
1875     int this_offset = t_oop->offset();
1876     int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1877     phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1878   } else if (ignore_missing_instance_id) {
1879     phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1880   } else {
1881     return nullptr;
1882   }
1883 
1884   for (uint i = 1; i < region->req(); i++) {
1885     Node* x;
1886     Node* the_clone = nullptr;
1887     Node* in = region->in(i);
1888     if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1889         in != nullptr && in->is_OuterStripMinedLoop()) {
1890       // No node should go in the outer strip mined loop
1891       in = in->in(LoopNode::EntryControl);
1892     }
1893     if (in == nullptr || in == C->top()) {
1894       x = C->top();      // Dead path?  Use a dead data op
1895     } else {
1896       x = this->clone();        // Else clone up the data op
1897       the_clone = x;            // Remember for possible deletion.
1898       // Alter data node to use pre-phi inputs
1899       if (this->in(0) == region) {
1900         x->set_req(0, in);
1901       } else {
1902         x->set_req(0, nullptr);
1903       }
1904       if (mem->is_Phi() && (mem->in(0) == region)) {
1905         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1906       }
1907       if (address->is_Phi() && address->in(0) == region) {
1908         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1909       }
1910       if (base_is_phi && (base->in(0) == region)) {
1911         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1912         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1913         x->set_req(Address, adr_x);
1914       }
1915     }
1916     // Check for a 'win' on some paths
1917     const Type *t = x->Value(igvn);
1918 
1919     bool singleton = t->singleton();
1920 
1921     // See comments in PhaseIdealLoop::split_thru_phi().
1922     if (singleton && t == Type::TOP) {
1923       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1924     }
1925 
1926     if (singleton) {
1927       x = igvn->makecon(t);
1928     } else {
1929       // We now call Identity to try to simplify the cloned node.
1930       // Note that some Identity methods call phase->type(this).
1931       // Make sure that the type array is big enough for
1932       // our new node, even though we may throw the node away.
1933       // (This tweaking with igvn only works because x is a new node.)
1934       igvn->set_type(x, t);
1935       // If x is a TypeNode, capture any more-precise type permanently into Node
1936       // otherwise it will be not updated during igvn->transform since
1937       // igvn->type(x) is set to x->Value() already.
1938       x->raise_bottom_type(t);
1939       Node* y = x->Identity(igvn);
1940       if (y != x) {
1941         x = y;
1942       } else {
1943         y = igvn->hash_find_insert(x);
1944         if (y) {
1945           x = y;
1946         } else {
1947           // Else x is a new node we are keeping
1948           // We do not need register_new_node_with_optimizer
1949           // because set_type has already been called.
1950           igvn->_worklist.push(x);
1951         }
1952       }
1953     }
1954     if (x != the_clone && the_clone != nullptr) {
1955       igvn->remove_dead_node(the_clone);
1956     }
1957     phi->set_req(i, x);
1958   }
1959   // Record Phi
1960   igvn->register_new_node_with_optimizer(phi);
1961   return phi;
1962 }
1963 
1964 AllocateNode* LoadNode::is_new_object_mark_load() const {
1965   if (Opcode() == Op_LoadX) {
1966     Node* address = in(MemNode::Address);
1967     AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1968     Node* mem = in(MemNode::Memory);
1969     if (alloc != nullptr && mem->is_Proj() &&
1970         mem->in(0) != nullptr &&
1971         mem->in(0) == alloc->initialization() &&
1972         alloc->initialization()->proj_out_or_null(0) != nullptr) {
1973       return alloc;
1974     }
1975   }
1976   return nullptr;
1977 }
1978 
1979 
1980 //------------------------------Ideal------------------------------------------
1981 // If the load is from Field memory and the pointer is non-null, it might be possible to
1982 // zero out the control input.
1983 // If the offset is constant and the base is an object allocation,
1984 // try to hook me up to the exact initializing store.
1985 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1986   if (has_pinned_control_dependency()) {
1987     return nullptr;
1988   }
1989   Node* p = MemNode::Ideal_common(phase, can_reshape);
1990   if (p)  return (p == NodeSentinel) ? nullptr : p;
1991 
1992   Node* ctrl    = in(MemNode::Control);
1993   Node* address = in(MemNode::Address);
1994   bool progress = false;
1995 
1996   bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1997          phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1998 
1999   // Skip up past a SafePoint control.  Cannot do this for Stores because
2000   // pointer stores & cardmarks must stay on the same side of a SafePoint.
2001   if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
2002       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw  &&
2003       !addr_mark &&
2004       (depends_only_on_test() || has_unknown_control_dependency())) {
2005     ctrl = ctrl->in(0);
2006     set_req(MemNode::Control,ctrl);
2007     progress = true;
2008   }
2009 
2010   intptr_t ignore = 0;
2011   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
2012   if (base != nullptr
2013       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
2014     // Check for useless control edge in some common special cases
2015     if (in(MemNode::Control) != nullptr
2016         && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
2017         && can_remove_control()
2018         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
2019         && all_controls_dominate(base, phase->C->start())) {
2020       // A method-invariant, non-null address (constant or 'this' argument).
2021       set_req(MemNode::Control, nullptr);
2022       progress = true;
2023     }
2024   }
2025 
2026   Node* mem = in(MemNode::Memory);
2027   const TypePtr *addr_t = phase->type(address)->isa_ptr();
2028 
2029   if (can_reshape && (addr_t != nullptr)) {
2030     // try to optimize our memory input
2031     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
2032     if (opt_mem != mem) {
2033       set_req_X(MemNode::Memory, opt_mem, phase);
2034       if (phase->type( opt_mem ) == Type::TOP) return nullptr;
2035       return this;
2036     }
2037     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
2038     if ((t_oop != nullptr) &&
2039         (t_oop->is_known_instance_field() ||
2040          t_oop->is_ptr_to_boxed_value())) {
2041       PhaseIterGVN *igvn = phase->is_IterGVN();
2042       assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
2043       if (igvn->_worklist.member(opt_mem)) {
2044         // Delay this transformation until memory Phi is processed.
2045         igvn->_worklist.push(this);
2046         return nullptr;
2047       }
2048       // Split instance field load through Phi.
2049       Node* result = split_through_phi(phase);
2050       if (result != nullptr) return result;
2051 
2052       if (t_oop->is_ptr_to_boxed_value()) {
2053         Node* result = eliminate_autobox(igvn);
2054         if (result != nullptr) return result;
2055       }
2056     }
2057   }
2058 
2059   // Is there a dominating load that loads the same value?  Leave
2060   // anything that is not a load of a field/array element (like
2061   // barriers etc.) alone
2062   if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
2063     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2064       Node *use = mem->fast_out(i);
2065       if (use != this &&
2066           use->Opcode() == Opcode() &&
2067           use->in(0) != nullptr &&
2068           use->in(0) != in(0) &&
2069           use->in(Address) == in(Address)) {
2070         Node* ctl = in(0);
2071         for (int i = 0; i < 10 && ctl != nullptr; i++) {
2072           ctl = IfNode::up_one_dom(ctl);
2073           if (ctl == use->in(0)) {
2074             set_req(0, use->in(0));
2075             return this;
2076           }
2077         }
2078       }
2079     }
2080   }
2081 
2082   // Check for prior store with a different base or offset; make Load
2083   // independent.  Skip through any number of them.  Bail out if the stores
2084   // are in an endless dead cycle and report no progress.  This is a key
2085   // transform for Reflection.  However, if after skipping through the Stores
2086   // we can't then fold up against a prior store do NOT do the transform as
2087   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
2088   // array memory alive twice: once for the hoisted Load and again after the
2089   // bypassed Store.  This situation only works if EVERYBODY who does
2090   // anti-dependence work knows how to bypass.  I.e. we need all
2091   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
2092   // the alias index stuff.  So instead, peek through Stores and IFF we can
2093   // fold up, do so.
2094   Node* prev_mem = find_previous_store(phase);
2095   if (prev_mem != nullptr) {
2096     Node* value = can_see_arraycopy_value(prev_mem, phase);
2097     if (value != nullptr) {
2098       return value;
2099     }
2100   }
2101   // Steps (a), (b):  Walk past independent stores to find an exact match.
2102   if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2103     // (c) See if we can fold up on the spot, but don't fold up here.
2104     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2105     // just return a prior value, which is done by Identity calls.
2106     if (can_see_stored_value(prev_mem, phase)) {
2107       // Make ready for step (d):
2108       set_req_X(MemNode::Memory, prev_mem, phase);
2109       return this;
2110     }
2111   }
2112 
2113   if (progress) {
2114     return this;
2115   }
2116 
2117   if (!can_reshape) {
2118     phase->record_for_igvn(this);
2119   }
2120   return nullptr;
2121 }
2122 
2123 // Helper to recognize certain Klass fields which are invariant across
2124 // some group of array types (e.g., int[] or all T[] where T < Object).
2125 const Type*
2126 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2127                                  ciKlass* klass) const {
2128   assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2129          "must not happen");
2130   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
2131     // The field is Klass::_access_flags.  Return its (constant) value.
2132     assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2133     return TypeInt::make(klass->access_flags());
2134   }
2135   if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2136     // The field is Klass::_misc_flags.  Return its (constant) value.
2137     assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2138     return TypeInt::make(klass->misc_flags());
2139   }
2140   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2141     // The field is Klass::_layout_helper.  Return its constant value if known.
2142     assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2143     return TypeInt::make(klass->layout_helper());
2144   }
2145 
2146   // No match.
2147   return nullptr;
2148 }
2149 
2150 //------------------------------Value-----------------------------------------
2151 const Type* LoadNode::Value(PhaseGVN* phase) const {
2152   // Either input is TOP ==> the result is TOP
2153   Node* mem = in(MemNode::Memory);
2154   const Type *t1 = phase->type(mem);
2155   if (t1 == Type::TOP)  return Type::TOP;
2156   Node* adr = in(MemNode::Address);
2157   const TypePtr* tp = phase->type(adr)->isa_ptr();
2158   if (tp == nullptr || tp->empty())  return Type::TOP;
2159   int off = tp->offset();
2160   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2161   Compile* C = phase->C;
2162 
2163   // If we are loading from a freshly-allocated object, produce a zero,
2164   // if the load is provably beyond the header of the object.
2165   // (Also allow a variable load from a fresh array to produce zero.)
2166   const TypeOopPtr* tinst = tp->isa_oopptr();
2167   bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2168   Node* value = can_see_stored_value(mem, phase);
2169   if (value != nullptr && value->is_Con()) {
2170     assert(value->bottom_type()->higher_equal(_type), "sanity");
2171     return value->bottom_type();
2172   }
2173 
2174   // Try to guess loaded type from pointer type
2175   if (tp->isa_aryptr()) {
2176     const TypeAryPtr* ary = tp->is_aryptr();
2177     const Type* t = ary->elem();
2178 
2179     // Determine whether the reference is beyond the header or not, by comparing
2180     // the offset against the offset of the start of the array's data.
2181     // Different array types begin at slightly different offsets (12 vs. 16).
2182     // We choose T_BYTE as an example base type that is least restrictive
2183     // as to alignment, which will therefore produce the smallest
2184     // possible base offset.
2185     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2186     const bool off_beyond_header = (off >= min_base_off);
2187 
2188     // Try to constant-fold a stable array element.
2189     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2190       // Make sure the reference is not into the header and the offset is constant
2191       ciObject* aobj = ary->const_oop();
2192       if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2193         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2194         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2195                                                                       stable_dimension,
2196                                                                       value_basic_type(), is_unsigned());
2197         if (con_type != nullptr) {
2198           return con_type;
2199         }
2200       }
2201     }
2202 
2203     // Don't do this for integer types. There is only potential profit if
2204     // the element type t is lower than _type; that is, for int types, if _type is
2205     // more restrictive than t.  This only happens here if one is short and the other
2206     // char (both 16 bits), and in those cases we've made an intentional decision
2207     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2208     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2209     //
2210     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2211     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
2212     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2213     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
2214     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2215     // In fact, that could have been the original type of p1, and p1 could have
2216     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2217     // expression (LShiftL quux 3) independently optimized to the constant 8.
2218     if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2219         && (_type->isa_vect() == nullptr)
2220         && !ary->is_flat()
2221         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2222       // t might actually be lower than _type, if _type is a unique
2223       // concrete subclass of abstract class t.
2224       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
2225         const Type* jt = t->join_speculative(_type);
2226         // In any case, do not allow the join, per se, to empty out the type.
2227         if (jt->empty() && !t->empty()) {
2228           // This can happen if a interface-typed array narrows to a class type.
2229           jt = _type;
2230         }
2231 #ifdef ASSERT
2232         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2233           // The pointers in the autobox arrays are always non-null
2234           Node* base = adr->in(AddPNode::Base);
2235           if ((base != nullptr) && base->is_DecodeN()) {
2236             // Get LoadN node which loads IntegerCache.cache field
2237             base = base->in(1);
2238           }
2239           if ((base != nullptr) && base->is_Con()) {
2240             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2241             if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2242               // It could be narrow oop
2243               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2244             }
2245           }
2246         }
2247 #endif
2248         return jt;
2249       }
2250     }
2251   } else if (tp->base() == Type::InstPtr) {
2252     assert( off != Type::OffsetBot ||
2253             // arrays can be cast to Objects
2254             !tp->isa_instptr() ||
2255             tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2256             // Default value load
2257             tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2258             // unsafe field access may not have a constant offset
2259             C->has_unsafe_access(),
2260             "Field accesses must be precise" );
2261     // For oop loads, we expect the _type to be precise.
2262 
2263     const TypeInstPtr* tinst = tp->is_instptr();
2264     BasicType bt = value_basic_type();
2265 
2266     // Optimize loads from constant fields.
2267     ciObject* const_oop = tinst->const_oop();
2268     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2269       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2270       if (con_type != nullptr) {
2271         return con_type;
2272       }
2273     }
2274   } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2275     assert(off != Type::OffsetBot ||
2276             !tp->isa_instklassptr() ||
2277            // arrays can be cast to Objects
2278            tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2279            // also allow array-loading from the primary supertype
2280            // array during subtype checks
2281            Opcode() == Op_LoadKlass,
2282            "Field accesses must be precise");
2283     // For klass/static loads, we expect the _type to be precise
2284   } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2285     /* With mirrors being an indirect in the Klass*
2286      * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2287      * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2288      *
2289      * So check the type and klass of the node before the LoadP.
2290      */
2291     Node* adr2 = adr->in(MemNode::Address);
2292     const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2293     if (tkls != nullptr && !StressReflectiveCode) {
2294       if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2295         ciKlass* klass = tkls->exact_klass();
2296         assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2297         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2298         return TypeInstPtr::make(klass->java_mirror());
2299       }
2300     }
2301   }
2302 
2303   const TypeKlassPtr *tkls = tp->isa_klassptr();
2304   if (tkls != nullptr) {
2305     if (tkls->is_loaded() && tkls->klass_is_exact()) {
2306       ciKlass* klass = tkls->exact_klass();
2307       // We are loading a field from a Klass metaobject whose identity
2308       // is known at compile time (the type is "exact" or "precise").
2309       // Check for fields we know are maintained as constants by the VM.
2310       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2311         // The field is Klass::_super_check_offset.  Return its (constant) value.
2312         // (Folds up type checking code.)
2313         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2314         return TypeInt::make(klass->super_check_offset());
2315       }
2316       if (UseCompactObjectHeaders) { // TODO: Should EnableValhalla also take this path ?
2317         if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2318           // The field is Klass::_prototype_header. Return its (constant) value.
2319           assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2320           return TypeX::make(klass->prototype_header());
2321         }
2322       }
2323       // Compute index into primary_supers array
2324       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2325       // Check for overflowing; use unsigned compare to handle the negative case.
2326       if( depth < ciKlass::primary_super_limit() ) {
2327         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2328         // (Folds up type checking code.)
2329         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2330         ciKlass *ss = klass->super_of_depth(depth);
2331         return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2332       }
2333       const Type* aift = load_array_final_field(tkls, klass);
2334       if (aift != nullptr)  return aift;
2335     }
2336 
2337     // We can still check if we are loading from the primary_supers array at a
2338     // shallow enough depth.  Even though the klass is not exact, entries less
2339     // than or equal to its super depth are correct.
2340     if (tkls->is_loaded()) {
2341       ciKlass* klass = nullptr;
2342       if (tkls->isa_instklassptr()) {
2343         klass = tkls->is_instklassptr()->instance_klass();
2344       } else {
2345         int dims;
2346         const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2347         if (inner->isa_instklassptr()) {
2348           klass = inner->is_instklassptr()->instance_klass();
2349           klass = ciObjArrayKlass::make(klass, dims);
2350         }
2351       }
2352       if (klass != nullptr) {
2353         // Compute index into primary_supers array
2354         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2355         // Check for overflowing; use unsigned compare to handle the negative case.
2356         if (depth < ciKlass::primary_super_limit() &&
2357             depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2358           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2359           // (Folds up type checking code.)
2360           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2361           ciKlass *ss = klass->super_of_depth(depth);
2362           return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2363         }
2364       }
2365     }
2366 
2367     // If the type is enough to determine that the thing is not an array,
2368     // we can give the layout_helper a positive interval type.
2369     // This will help short-circuit some reflective code.
2370     if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2371         tkls->isa_instklassptr() && // not directly typed as an array
2372         !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
2373     ) {
2374       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2375       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2376       // The key property of this type is that it folds up tests
2377       // for array-ness, since it proves that the layout_helper is positive.
2378       // Thus, a generic value like the basic object layout helper works fine.
2379       return TypeInt::make(min_size, max_jint, Type::WidenMin);
2380     }
2381   }
2382 
2383   bool is_vect = (_type->isa_vect() != nullptr);
2384   if (is_instance && !is_vect) {
2385     // If we have an instance type and our memory input is the
2386     // programs's initial memory state, there is no matching store,
2387     // so just return a zero of the appropriate type -
2388     // except if it is vectorized - then we have no zero constant.
2389     Node *mem = in(MemNode::Memory);
2390     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2391       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2392       // TODO 8350865 Scalar replacement does not work well for flat arrays.
2393       // Escape Analysis assumes that arrays are always zeroed during allocation which is not true for null-free arrays
2394       // ConnectionGraph::split_unique_types will re-wire the memory of loads from such arrays around the allocation
2395       // TestArrays::test6 and test152 and TestBasicFunctionality::test20 are affected by this.
2396       if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && tp->is_aryptr()->is_null_free()) {
2397         intptr_t offset = 0;
2398         Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2399         AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
2400         if (alloc != nullptr && alloc->is_AllocateArray() && alloc->in(AllocateNode::InitValue) != nullptr) {
2401           return _type;
2402         }
2403       }
2404       return Type::get_zero_type(_type->basic_type());
2405     }
2406   }
2407   if (!UseCompactObjectHeaders) {
2408     Node* alloc = is_new_object_mark_load();
2409     if (alloc != nullptr) {
2410       if (EnableValhalla) {
2411         // The mark word may contain property bits (inline, flat, null-free)
2412         Node* klass_node = alloc->in(AllocateNode::KlassNode);
2413         const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2414         if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2415           return TypeX::make(tkls->exact_klass()->prototype_header());
2416         }
2417       } else {
2418         return TypeX::make(markWord::prototype().value());
2419       }
2420     }
2421   }
2422 
2423   return _type;
2424 }
2425 
2426 //------------------------------match_edge-------------------------------------
2427 // Do we Match on this edge index or not?  Match only the address.
2428 uint LoadNode::match_edge(uint idx) const {
2429   return idx == MemNode::Address;
2430 }
2431 
2432 //--------------------------LoadBNode::Ideal--------------------------------------
2433 //
2434 //  If the previous store is to the same address as this load,
2435 //  and the value stored was larger than a byte, replace this load
2436 //  with the value stored truncated to a byte.  If no truncation is
2437 //  needed, the replacement is done in LoadNode::Identity().
2438 //
2439 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2440   Node* mem = in(MemNode::Memory);
2441   Node* value = can_see_stored_value(mem,phase);
2442   if (value != nullptr) {
2443     Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2444     if (narrow != value) {
2445       return narrow;
2446     }
2447   }
2448   // Identity call will handle the case where truncation is not needed.
2449   return LoadNode::Ideal(phase, can_reshape);
2450 }
2451 
2452 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2453   Node* mem = in(MemNode::Memory);
2454   Node* value = can_see_stored_value(mem,phase);
2455   if (value != nullptr && value->is_Con() &&
2456       !value->bottom_type()->higher_equal(_type)) {
2457     // If the input to the store does not fit with the load's result type,
2458     // it must be truncated. We can't delay until Ideal call since
2459     // a singleton Value is needed for split_thru_phi optimization.
2460     int con = value->get_int();
2461     return TypeInt::make((con << 24) >> 24);
2462   }
2463   return LoadNode::Value(phase);
2464 }
2465 
2466 //--------------------------LoadUBNode::Ideal-------------------------------------
2467 //
2468 //  If the previous store is to the same address as this load,
2469 //  and the value stored was larger than a byte, replace this load
2470 //  with the value stored truncated to a byte.  If no truncation is
2471 //  needed, the replacement is done in LoadNode::Identity().
2472 //
2473 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2474   Node* mem = in(MemNode::Memory);
2475   Node* value = can_see_stored_value(mem, phase);
2476   if (value != nullptr) {
2477     Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2478     if (narrow != value) {
2479       return narrow;
2480     }
2481   }
2482   // Identity call will handle the case where truncation is not needed.
2483   return LoadNode::Ideal(phase, can_reshape);
2484 }
2485 
2486 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2487   Node* mem = in(MemNode::Memory);
2488   Node* value = can_see_stored_value(mem,phase);
2489   if (value != nullptr && value->is_Con() &&
2490       !value->bottom_type()->higher_equal(_type)) {
2491     // If the input to the store does not fit with the load's result type,
2492     // it must be truncated. We can't delay until Ideal call since
2493     // a singleton Value is needed for split_thru_phi optimization.
2494     int con = value->get_int();
2495     return TypeInt::make(con & 0xFF);
2496   }
2497   return LoadNode::Value(phase);
2498 }
2499 
2500 //--------------------------LoadUSNode::Ideal-------------------------------------
2501 //
2502 //  If the previous store is to the same address as this load,
2503 //  and the value stored was larger than a char, replace this load
2504 //  with the value stored truncated to a char.  If no truncation is
2505 //  needed, the replacement is done in LoadNode::Identity().
2506 //
2507 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2508   Node* mem = in(MemNode::Memory);
2509   Node* value = can_see_stored_value(mem,phase);
2510   if (value != nullptr) {
2511     Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2512     if (narrow != value) {
2513       return narrow;
2514     }
2515   }
2516   // Identity call will handle the case where truncation is not needed.
2517   return LoadNode::Ideal(phase, can_reshape);
2518 }
2519 
2520 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2521   Node* mem = in(MemNode::Memory);
2522   Node* value = can_see_stored_value(mem,phase);
2523   if (value != nullptr && value->is_Con() &&
2524       !value->bottom_type()->higher_equal(_type)) {
2525     // If the input to the store does not fit with the load's result type,
2526     // it must be truncated. We can't delay until Ideal call since
2527     // a singleton Value is needed for split_thru_phi optimization.
2528     int con = value->get_int();
2529     return TypeInt::make(con & 0xFFFF);
2530   }
2531   return LoadNode::Value(phase);
2532 }
2533 
2534 //--------------------------LoadSNode::Ideal--------------------------------------
2535 //
2536 //  If the previous store is to the same address as this load,
2537 //  and the value stored was larger than a short, replace this load
2538 //  with the value stored truncated to a short.  If no truncation is
2539 //  needed, the replacement is done in LoadNode::Identity().
2540 //
2541 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2542   Node* mem = in(MemNode::Memory);
2543   Node* value = can_see_stored_value(mem,phase);
2544   if (value != nullptr) {
2545     Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2546     if (narrow != value) {
2547       return narrow;
2548     }
2549   }
2550   // Identity call will handle the case where truncation is not needed.
2551   return LoadNode::Ideal(phase, can_reshape);
2552 }
2553 
2554 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2555   Node* mem = in(MemNode::Memory);
2556   Node* value = can_see_stored_value(mem,phase);
2557   if (value != nullptr && value->is_Con() &&
2558       !value->bottom_type()->higher_equal(_type)) {
2559     // If the input to the store does not fit with the load's result type,
2560     // it must be truncated. We can't delay until Ideal call since
2561     // a singleton Value is needed for split_thru_phi optimization.
2562     int con = value->get_int();
2563     return TypeInt::make((con << 16) >> 16);
2564   }
2565   return LoadNode::Value(phase);
2566 }
2567 
2568 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2569   // Loading from an InlineType, find the input and make an EncodeP
2570   Node* addr = in(Address);
2571   intptr_t offset;
2572   Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
2573   Node* value = see_through_inline_type(phase, this, base, offset);
2574   if (value != nullptr) {
2575     return new EncodePNode(value, type());
2576   }
2577 
2578   return LoadNode::Ideal(phase, can_reshape);
2579 }
2580 
2581 //=============================================================================
2582 //----------------------------LoadKlassNode::make------------------------------
2583 // Polymorphic factory method:
2584 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2585   // sanity check the alias category against the created node type
2586   const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2587   assert(adr_type != nullptr, "expecting TypeKlassPtr");
2588 #ifdef _LP64
2589   if (adr_type->is_ptr_to_narrowklass()) {
2590     assert(UseCompressedClassPointers, "no compressed klasses");
2591     Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2592     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2593   }
2594 #endif
2595   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2596   return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2597 }
2598 
2599 //------------------------------Value------------------------------------------
2600 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2601   return klass_value_common(phase);
2602 }
2603 
2604 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2605   // Either input is TOP ==> the result is TOP
2606   const Type *t1 = phase->type( in(MemNode::Memory) );
2607   if (t1 == Type::TOP)  return Type::TOP;
2608   Node *adr = in(MemNode::Address);
2609   const Type *t2 = phase->type( adr );
2610   if (t2 == Type::TOP)  return Type::TOP;
2611   const TypePtr *tp = t2->is_ptr();
2612   if (TypePtr::above_centerline(tp->ptr()) ||
2613       tp->ptr() == TypePtr::Null)  return Type::TOP;
2614 
2615   // Return a more precise klass, if possible
2616   const TypeInstPtr *tinst = tp->isa_instptr();
2617   if (tinst != nullptr) {
2618     ciInstanceKlass* ik = tinst->instance_klass();
2619     int offset = tinst->offset();
2620     if (ik == phase->C->env()->Class_klass()
2621         && (offset == java_lang_Class::klass_offset() ||
2622             offset == java_lang_Class::array_klass_offset())) {
2623       // We are loading a special hidden field from a Class mirror object,
2624       // the field which points to the VM's Klass metaobject.
2625       ciType* t = tinst->java_mirror_type();
2626       // java_mirror_type returns non-null for compile-time Class constants.
2627       if (t != nullptr) {
2628         // constant oop => constant klass
2629         if (offset == java_lang_Class::array_klass_offset()) {
2630           if (t->is_void()) {
2631             // We cannot create a void array.  Since void is a primitive type return null
2632             // klass.  Users of this result need to do a null check on the returned klass.
2633             return TypePtr::NULL_PTR;
2634           }
2635           return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2636         }
2637         if (!t->is_klass()) {
2638           // a primitive Class (e.g., int.class) has null for a klass field
2639           return TypePtr::NULL_PTR;
2640         }
2641         // Fold up the load of the hidden field
2642         return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2643       }
2644       // non-constant mirror, so we can't tell what's going on
2645     }
2646     if (!tinst->is_loaded())
2647       return _type;             // Bail out if not loaded
2648     if (offset == oopDesc::klass_offset_in_bytes()) {
2649       return tinst->as_klass_type(true);
2650     }
2651   }
2652 
2653   // Check for loading klass from an array
2654   const TypeAryPtr* tary = tp->isa_aryptr();
2655   if (tary != nullptr &&
2656       tary->offset() == oopDesc::klass_offset_in_bytes()) {
2657     return tary->as_klass_type(true);
2658   }
2659 
2660   // Check for loading klass from an array klass
2661   const TypeKlassPtr *tkls = tp->isa_klassptr();
2662   if (tkls != nullptr && !StressReflectiveCode) {
2663     if (!tkls->is_loaded())
2664      return _type;             // Bail out if not loaded
2665     if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2666         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2667       // // Always returning precise element type is incorrect,
2668       // // e.g., element type could be object and array may contain strings
2669       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2670 
2671       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2672       // according to the element type's subclassing.
2673       return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2674     }
2675     if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2676         tkls->offset() == in_bytes(Klass::super_offset())) {
2677       ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2678       // The field is Klass::_super.  Return its (constant) value.
2679       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2680       return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2681     }
2682   }
2683 
2684   if (tkls != nullptr && !UseSecondarySupersCache
2685       && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset()))  {
2686     // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2687     return TypePtr::NULL_PTR;
2688   }
2689 
2690   // Bailout case
2691   return LoadNode::Value(phase);
2692 }
2693 
2694 //------------------------------Identity---------------------------------------
2695 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2696 // Also feed through the klass in Allocate(...klass...)._klass.
2697 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2698   return klass_identity_common(phase);
2699 }
2700 
2701 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2702   Node* x = LoadNode::Identity(phase);
2703   if (x != this)  return x;
2704 
2705   // Take apart the address into an oop and offset.
2706   // Return 'this' if we cannot.
2707   Node*    adr    = in(MemNode::Address);
2708   intptr_t offset = 0;
2709   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2710   if (base == nullptr)     return this;
2711   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2712   if (toop == nullptr)     return this;
2713 
2714   // Step over potential GC barrier for OopHandle resolve
2715   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2716   if (bs->is_gc_barrier_node(base)) {
2717     base = bs->step_over_gc_barrier(base);
2718   }
2719 
2720   // We can fetch the klass directly through an AllocateNode.
2721   // This works even if the klass is not constant (clone or newArray).
2722   if (offset == oopDesc::klass_offset_in_bytes()) {
2723     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2724     if (allocated_klass != nullptr) {
2725       return allocated_klass;
2726     }
2727   }
2728 
2729   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2730   // See inline_native_Class_query for occurrences of these patterns.
2731   // Java Example:  x.getClass().isAssignableFrom(y)
2732   //
2733   // This improves reflective code, often making the Class
2734   // mirror go completely dead.  (Current exception:  Class
2735   // mirrors may appear in debug info, but we could clean them out by
2736   // introducing a new debug info operator for Klass.java_mirror).
2737 
2738   if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2739       && offset == java_lang_Class::klass_offset()) {
2740     if (base->is_Load()) {
2741       Node* base2 = base->in(MemNode::Address);
2742       if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2743         Node* adr2 = base2->in(MemNode::Address);
2744         const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2745         // TODO 8366668 Re-enable this for arrays
2746         if (tkls != nullptr && !tkls->empty()
2747             && ((tkls->isa_instklassptr() && !tkls->is_instklassptr()->might_be_an_array()) || (tkls->isa_aryklassptr() && false))
2748             && adr2->is_AddP()
2749            ) {
2750           int mirror_field = in_bytes(Klass::java_mirror_offset());
2751           if (tkls->offset() == mirror_field) {
2752             return adr2->in(AddPNode::Base);
2753           }
2754         }
2755       }
2756     }
2757   }
2758 
2759   return this;
2760 }
2761 
2762 LoadNode* LoadNode::clone_pinned() const {
2763   LoadNode* ld = clone()->as_Load();
2764   ld->_control_dependency = UnknownControl;
2765   return ld;
2766 }
2767 
2768 
2769 //------------------------------Value------------------------------------------
2770 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2771   const Type *t = klass_value_common(phase);
2772   if (t == Type::TOP)
2773     return t;
2774 
2775   return t->make_narrowklass();
2776 }
2777 
2778 //------------------------------Identity---------------------------------------
2779 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2780 // Also feed through the klass in Allocate(...klass...)._klass.
2781 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2782   Node *x = klass_identity_common(phase);
2783 
2784   const Type *t = phase->type( x );
2785   if( t == Type::TOP ) return x;
2786   if( t->isa_narrowklass()) return x;
2787   assert (!t->isa_narrowoop(), "no narrow oop here");
2788 
2789   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2790 }
2791 
2792 //------------------------------Value-----------------------------------------
2793 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2794   // Either input is TOP ==> the result is TOP
2795   const Type *t1 = phase->type( in(MemNode::Memory) );
2796   if( t1 == Type::TOP ) return Type::TOP;
2797   Node *adr = in(MemNode::Address);
2798   const Type *t2 = phase->type( adr );
2799   if( t2 == Type::TOP ) return Type::TOP;
2800   const TypePtr *tp = t2->is_ptr();
2801   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2802   const TypeAryPtr *tap = tp->isa_aryptr();
2803   if( !tap ) return _type;
2804   return tap->size();
2805 }
2806 
2807 //-------------------------------Ideal---------------------------------------
2808 // Feed through the length in AllocateArray(...length...)._length.
2809 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2810   Node* p = MemNode::Ideal_common(phase, can_reshape);
2811   if (p)  return (p == NodeSentinel) ? nullptr : p;
2812 
2813   // Take apart the address into an oop and offset.
2814   // Return 'this' if we cannot.
2815   Node*    adr    = in(MemNode::Address);
2816   intptr_t offset = 0;
2817   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2818   if (base == nullptr)     return nullptr;
2819   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2820   if (tary == nullptr)     return nullptr;
2821 
2822   // We can fetch the length directly through an AllocateArrayNode.
2823   // This works even if the length is not constant (clone or newArray).
2824   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2825     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2826     if (alloc != nullptr) {
2827       Node* allocated_length = alloc->Ideal_length();
2828       Node* len = alloc->make_ideal_length(tary, phase);
2829       if (allocated_length != len) {
2830         // New CastII improves on this.
2831         return len;
2832       }
2833     }
2834   }
2835 
2836   return nullptr;
2837 }
2838 
2839 //------------------------------Identity---------------------------------------
2840 // Feed through the length in AllocateArray(...length...)._length.
2841 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2842   Node* x = LoadINode::Identity(phase);
2843   if (x != this)  return x;
2844 
2845   // Take apart the address into an oop and offset.
2846   // Return 'this' if we cannot.
2847   Node*    adr    = in(MemNode::Address);
2848   intptr_t offset = 0;
2849   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2850   if (base == nullptr)     return this;
2851   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2852   if (tary == nullptr)     return this;
2853 
2854   // We can fetch the length directly through an AllocateArrayNode.
2855   // This works even if the length is not constant (clone or newArray).
2856   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2857     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2858     if (alloc != nullptr) {
2859       Node* allocated_length = alloc->Ideal_length();
2860       // Do not allow make_ideal_length to allocate a CastII node.
2861       Node* len = alloc->make_ideal_length(tary, phase, false);
2862       if (allocated_length == len) {
2863         // Return allocated_length only if it would not be improved by a CastII.
2864         return allocated_length;
2865       }
2866     }
2867   }
2868 
2869   return this;
2870 
2871 }
2872 
2873 //=============================================================================
2874 //---------------------------StoreNode::make-----------------------------------
2875 // Polymorphic factory method:
2876 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) {
2877   assert((mo == unordered || mo == release), "unexpected");
2878   Compile* C = gvn.C;
2879   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2880          ctl != nullptr, "raw memory operations should have control edge");
2881 
2882   switch (bt) {
2883   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2884   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2885   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2886   case T_CHAR:
2887   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2888   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2889   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2890   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2891   case T_METADATA:
2892   case T_ADDRESS:
2893   case T_OBJECT:
2894 #ifdef _LP64
2895     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2896       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2897       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2898     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2899                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2900                 adr->bottom_type()->isa_rawptr())) {
2901       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2902       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2903     }
2904 #endif
2905     {
2906       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2907     }
2908   default:
2909     ShouldNotReachHere();
2910     return (StoreNode*)nullptr;
2911   }
2912 }
2913 
2914 //--------------------------bottom_type----------------------------------------
2915 const Type *StoreNode::bottom_type() const {
2916   return Type::MEMORY;
2917 }
2918 
2919 //------------------------------hash-------------------------------------------
2920 uint StoreNode::hash() const {
2921   // unroll addition of interesting fields
2922   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2923 
2924   // Since they are not commoned, do not hash them:
2925   return NO_HASH;
2926 }
2927 
2928 // Link together multiple stores (B/S/C/I) into a longer one.
2929 //
2930 // Example: _store = StoreB[i+3]
2931 //
2932 //   RangeCheck[i+0]           RangeCheck[i+0]
2933 //   StoreB[i+0]
2934 //   RangeCheck[i+3]           RangeCheck[i+3]
2935 //   StoreB[i+1]         -->   pass:             fail:
2936 //   StoreB[i+2]               StoreI[i+0]       StoreB[i+0]
2937 //   StoreB[i+3]
2938 //
2939 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2940 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2941 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2942 // failing path.
2943 //
2944 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2945 //       - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2946 //                                       and possibly no RangeChecks remain between the stores.
2947 //       - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2948 //                              and those later RangeChecks can be removed. Example:
2949 //
2950 //                              RangeCheck[i+0]                         RangeCheck[i+0] <- before first store
2951 //                              StoreB[i+0]                             StoreB[i+0]     <- first store
2952 //                              RangeCheck[i+1]     --> smeared -->     RangeCheck[i+3] <- only RC between first and last store
2953 //                              StoreB[i+1]                             StoreB[i+1]     <- second store
2954 //                              RangeCheck[i+2]     --> removed
2955 //                              StoreB[i+2]                             StoreB[i+2]
2956 //                              RangeCheck[i+3]     --> removed
2957 //                              StoreB[i+3]                             StoreB[i+3]     <- last store
2958 //
2959 //                              Thus, it is a common pattern that between the first and last store in a chain
2960 //                              of adjacent stores there remains exactly one RangeCheck, located between the
2961 //                              first and the second store (e.g. RangeCheck[i+3]).
2962 //
2963 class MergePrimitiveStores : public StackObj {
2964 private:
2965   PhaseGVN* const _phase;
2966   StoreNode* const _store;
2967   // State machine with initial state Unknown
2968   // Allowed transitions:
2969   //   Unknown     -> Const
2970   //   Unknown     -> Platform
2971   //   Unknown     -> Reverse
2972   //   Unknown     -> NotAdjacent
2973   //   Const       -> Const
2974   //   Const       -> NotAdjacent
2975   //   Platform    -> Platform
2976   //   Platform    -> NotAdjacent
2977   //   Reverse     -> Reverse
2978   //   Reverse     -> NotAdjacent
2979   //   NotAdjacent -> NotAdjacent
2980   enum ValueOrder : uint8_t {
2981     Unknown,     // Initial state
2982     Const,       // Input values are const
2983     Platform,    // Platform order
2984     Reverse,     // Reverse platform order
2985     NotAdjacent  // Not adjacent
2986   };
2987   ValueOrder  _value_order;
2988 
2989   NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2990 
2991 public:
2992   MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2993     _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2994     NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2995     {}
2996 
2997   StoreNode* run();
2998 
2999 private:
3000   bool is_compatible_store(const StoreNode* other_store) const;
3001   bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3002   bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3003   static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
3004   enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3005   static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3006 
3007   class Status {
3008   private:
3009     StoreNode* _found_store;
3010     bool       _found_range_check;
3011 
3012     Status(StoreNode* found_store, bool found_range_check)
3013       : _found_store(found_store), _found_range_check(found_range_check) {}
3014 
3015   public:
3016     StoreNode* found_store() const { return _found_store; }
3017     bool found_range_check() const { return _found_range_check; }
3018     static Status make_failure() { return Status(nullptr, false); }
3019 
3020     static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3021       if (cfg_status == CFGStatus::Failure) {
3022         return Status::make_failure();
3023       }
3024       return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3025     }
3026 
3027 #ifndef PRODUCT
3028     void print_on(outputStream* st) const {
3029       if (_found_store == nullptr) {
3030         st->print_cr("None");
3031       } else {
3032         st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
3033                                          _found_range_check ? "RC" : "no-RC");
3034       }
3035     }
3036 #endif
3037   };
3038 
3039   enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
3040   Status find_adjacent_use_store(const StoreNode* def_store) const;
3041   Status find_adjacent_def_store(const StoreNode* use_store) const;
3042   Status find_use_store(const StoreNode* def_store) const;
3043   Status find_def_store(const StoreNode* use_store) const;
3044   Status find_use_store_unidirectional(const StoreNode* def_store) const;
3045   Status find_def_store_unidirectional(const StoreNode* use_store) const;
3046 
3047   void collect_merge_list(Node_List& merge_list) const;
3048   Node* make_merged_input_value(const Node_List& merge_list);
3049   StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3050 
3051 #ifndef PRODUCT
3052   // Access to TraceMergeStores tags
3053   bool is_trace(TraceMergeStores::Tag tag) const {
3054     return _trace_tags.at(tag);
3055   }
3056 
3057   bool is_trace_basic() const {
3058     return is_trace(TraceMergeStores::Tag::BASIC);
3059   }
3060 
3061   bool is_trace_pointer_parsing() const {
3062     return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
3063   }
3064 
3065   bool is_trace_pointer_aliasing() const {
3066     return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
3067   }
3068 
3069   bool is_trace_pointer_adjacency() const {
3070     return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
3071   }
3072 
3073   bool is_trace_success() const {
3074     return is_trace(TraceMergeStores::Tag::SUCCESS);
3075   }
3076 #endif
3077 
3078   NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3079 };
3080 
3081 StoreNode* MergePrimitiveStores::run() {
3082   // Check for B/S/C/I
3083   int opc = _store->Opcode();
3084   if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3085     return nullptr;
3086   }
3087 
3088   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3089 
3090   // The _store must be the "last" store in a chain. If we find a use we could merge with
3091   // then that use or a store further down is the "last" store.
3092   Status status_use = find_adjacent_use_store(_store);
3093   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3094   if (status_use.found_store() != nullptr) {
3095     return nullptr;
3096   }
3097 
3098   // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3099   Status status_def = find_adjacent_def_store(_store);
3100   NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3101   Node* def_store = status_def.found_store();
3102   if (def_store == nullptr) {
3103     return nullptr;
3104   }
3105 
3106   // Initialize value order
3107   _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3108                                                  _store->in(MemNode::ValueIn),
3109                                                  _store->memory_size());
3110   assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3111 
3112   ResourceMark rm;
3113   Node_List merge_list;
3114   collect_merge_list(merge_list);
3115 
3116   Node* merged_input_value = make_merged_input_value(merge_list);
3117   if (merged_input_value == nullptr) { return nullptr; }
3118 
3119   StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3120 
3121   NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3122 
3123   return merged_store;
3124 }
3125 
3126 // Check compatibility between _store and other_store.
3127 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3128   int opc = _store->Opcode();
3129   assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3130 
3131   if (other_store == nullptr ||
3132       _store->Opcode() != other_store->Opcode()) {
3133     return false;
3134   }
3135 
3136   return true;
3137 }
3138 
3139 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3140   if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3141                               use_store->in(MemNode::ValueIn),
3142                               def_store->memory_size())) {
3143     return false;
3144   }
3145 
3146   ResourceMark rm;
3147 #ifndef PRODUCT
3148   const TraceMemPointer trace(is_trace_pointer_parsing(),
3149                               is_trace_pointer_aliasing(),
3150                               is_trace_pointer_adjacency(),
3151                               true);
3152 #endif
3153   const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3154   const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3155   return pointer_def.is_adjacent_to_and_before(pointer_use);
3156 }
3157 
3158 // Check input values n1 and n2 can be merged and return the value order
3159 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3160                                                                                        const int memory_size) const {
3161   // Pattern: [n1 = ConI, n2 = ConI]
3162   if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3163     return ValueOrder::Const;
3164   }
3165 
3166   Node const *base_n2;
3167   jint shift_n2;
3168   if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3169     return ValueOrder::NotAdjacent;
3170   }
3171   Node const *base_n1;
3172   jint shift_n1;
3173   if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3174     return ValueOrder::NotAdjacent;
3175   }
3176 
3177   int bits_per_store = memory_size * 8;
3178   if (base_n1 != base_n2 ||
3179       abs(shift_n1 - shift_n2) != bits_per_store ||
3180       shift_n1 % bits_per_store != 0) {
3181     // Values are not adjacent
3182     return ValueOrder::NotAdjacent;
3183   }
3184 
3185   // Detect value order
3186 #ifdef VM_LITTLE_ENDIAN
3187   return shift_n1 < shift_n2 ? ValueOrder::Platform     // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3188                              : ValueOrder::Reverse;     // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3189 #else
3190   return shift_n1 > shift_n2 ? ValueOrder::Platform     // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3191                              : ValueOrder::Reverse;     // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3192 #endif
3193 }
3194 
3195 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3196   ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3197 
3198   switch (input_value_order) {
3199     case ValueOrder::NotAdjacent:
3200       return false;
3201     case ValueOrder::Reverse:
3202       if (memory_size != 1 ||
3203           !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3204           !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3205           !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3206         // ReverseBytes are not supported by platform
3207         return false;
3208       }
3209       // fall-through.
3210     case ValueOrder::Const:
3211     case ValueOrder::Platform:
3212       if (_value_order == ValueOrder::Unknown) {
3213         // Initial state is Unknown, and we find a valid input value order
3214         return true;
3215       }
3216       // The value order can not be changed
3217       return _value_order == input_value_order;
3218     case ValueOrder::Unknown:
3219     default:
3220       ShouldNotReachHere();
3221   }
3222   return false;
3223 }
3224 
3225 // Detect pattern: n = base_out >> shift_out
3226 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3227   assert(n != nullptr, "precondition");
3228 
3229   int opc = n->Opcode();
3230   if (opc == Op_ConvL2I) {
3231     n = n->in(1);
3232     opc = n->Opcode();
3233   }
3234 
3235   if ((opc == Op_RShiftI ||
3236        opc == Op_RShiftL ||
3237        opc == Op_URShiftI ||
3238        opc == Op_URShiftL) &&
3239       n->in(2)->is_ConI()) {
3240     base_out = n->in(1);
3241     shift_out = n->in(2)->get_int();
3242     // The shift must be positive:
3243     return shift_out >= 0;
3244   }
3245 
3246   if (phase->type(n)->isa_int()  != nullptr ||
3247       phase->type(n)->isa_long() != nullptr) {
3248     // (base >> 0)
3249     base_out = n;
3250     shift_out = 0;
3251     return true;
3252   }
3253   return false;
3254 }
3255 
3256 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3257 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3258   assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3259 
3260   Node* ctrl_use = use_store->in(MemNode::Control);
3261   Node* ctrl_def = def_store->in(MemNode::Control);
3262   if (ctrl_use == nullptr || ctrl_def == nullptr) {
3263     return CFGStatus::Failure;
3264   }
3265 
3266   if (ctrl_use == ctrl_def) {
3267     // Same ctrl -> no RangeCheck in between.
3268     // Check: use_store must be the only use of def_store.
3269     if (def_store->outcnt() > 1) {
3270       return CFGStatus::Failure;
3271     }
3272     return CFGStatus::SuccessNoRangeCheck;
3273   }
3274 
3275   // Different ctrl -> could have RangeCheck in between.
3276   // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3277   //        2. ctrl separated by RangeCheck.
3278   if (def_store->outcnt() != 2) {
3279     return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3280   }
3281   int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3282   Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3283   if (merge_mem == nullptr ||
3284       merge_mem->outcnt() != 1) {
3285     return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3286   }
3287   if (!ctrl_use->is_IfProj() ||
3288       !ctrl_use->in(0)->is_RangeCheck() ||
3289       ctrl_use->in(0)->outcnt() != 2) {
3290     return CFGStatus::Failure; // Not RangeCheck.
3291   }
3292   ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3293   Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3294   if (trap != merge_mem->unique_out() ||
3295       ctrl_use->in(0)->in(0) != ctrl_def) {
3296     return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3297   }
3298 
3299   return CFGStatus::SuccessWithRangeCheck;
3300 }
3301 
3302 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3303   Status status_use = find_use_store(def_store);
3304   StoreNode* use_store = status_use.found_store();
3305   if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3306     return Status::make_failure();
3307   }
3308   return status_use;
3309 }
3310 
3311 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3312   Status status_def = find_def_store(use_store);
3313   StoreNode* def_store = status_def.found_store();
3314   if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3315     return Status::make_failure();
3316   }
3317   return status_def;
3318 }
3319 
3320 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3321   Status status_use = find_use_store_unidirectional(def_store);
3322 
3323 #ifdef ASSERT
3324   StoreNode* use_store = status_use.found_store();
3325   if (use_store != nullptr) {
3326     Status status_def = find_def_store_unidirectional(use_store);
3327     assert(status_def.found_store() == def_store &&
3328            status_def.found_range_check() == status_use.found_range_check(),
3329            "find_use_store and find_def_store must be symmetric");
3330   }
3331 #endif
3332 
3333   return status_use;
3334 }
3335 
3336 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3337   Status status_def = find_def_store_unidirectional(use_store);
3338 
3339 #ifdef ASSERT
3340   StoreNode* def_store = status_def.found_store();
3341   if (def_store != nullptr) {
3342     Status status_use = find_use_store_unidirectional(def_store);
3343     assert(status_use.found_store() == use_store &&
3344            status_use.found_range_check() == status_def.found_range_check(),
3345            "find_use_store and find_def_store must be symmetric");
3346   }
3347 #endif
3348 
3349   return status_def;
3350 }
3351 
3352 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3353   assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3354 
3355   for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3356     StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3357     if (is_compatible_store(use_store)) {
3358       return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3359     }
3360   }
3361 
3362   return Status::make_failure();
3363 }
3364 
3365 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3366   assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3367 
3368   StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3369   if (!is_compatible_store(def_store)) {
3370     return Status::make_failure();
3371   }
3372 
3373   return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3374 }
3375 
3376 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3377   // The merged store can be at most 8 bytes.
3378   const uint merge_list_max_size = 8 / _store->memory_size();
3379   assert(merge_list_max_size >= 2 &&
3380          merge_list_max_size <= 8 &&
3381          is_power_of_2(merge_list_max_size),
3382          "must be 2, 4 or 8");
3383 
3384   // Traverse up the chain of adjacent def stores.
3385   StoreNode* current = _store;
3386   merge_list.push(current);
3387   while (current != nullptr && merge_list.size() < merge_list_max_size) {
3388     Status status = find_adjacent_def_store(current);
3389     NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3390 
3391     current = status.found_store();
3392     if (current != nullptr) {
3393       merge_list.push(current);
3394 
3395       // We can have at most one RangeCheck.
3396       if (status.found_range_check()) {
3397         NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3398         break;
3399       }
3400     }
3401   }
3402 
3403   NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3404 
3405   // Truncate the merge_list to a power of 2.
3406   const uint pow2size = round_down_power_of_2(merge_list.size());
3407   assert(pow2size >= 2, "must be merging at least 2 stores");
3408   while (merge_list.size() > pow2size) { merge_list.pop(); }
3409 
3410   NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3411 }
3412 
3413 // Merge the input values of the smaller stores to a single larger input value.
3414 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3415   int new_memory_size = _store->memory_size() * merge_list.size();
3416   Node* first = merge_list.at(merge_list.size()-1);
3417   Node* merged_input_value = nullptr;
3418   if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3419     assert(_value_order == ValueOrder::Const, "must match");
3420     // Pattern: [ConI, ConI, ...] -> new constant
3421     jlong con = 0;
3422     jlong bits_per_store = _store->memory_size() * 8;
3423     jlong mask = (((jlong)1) << bits_per_store) - 1;
3424     for (uint i = 0; i < merge_list.size(); i++) {
3425       jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3426 #ifdef VM_LITTLE_ENDIAN
3427       con = con << bits_per_store;
3428       con = con | (mask & con_i);
3429 #else // VM_LITTLE_ENDIAN
3430       con_i = (mask & con_i) << (i * bits_per_store);
3431       con = con | con_i;
3432 #endif // VM_LITTLE_ENDIAN
3433     }
3434     merged_input_value = _phase->longcon(con);
3435   } else {
3436     assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3437     // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3438     //             |                                  |
3439     //           _store                             first
3440     //
3441     Node* hi = _store->in(MemNode::ValueIn);
3442     Node* lo = first->in(MemNode::ValueIn);
3443 #ifndef VM_LITTLE_ENDIAN
3444     // `_store` and `first` are swapped in the diagram above
3445     swap(hi, lo);
3446 #endif // !VM_LITTLE_ENDIAN
3447     if (_value_order == ValueOrder::Reverse) {
3448       swap(hi, lo);
3449     }
3450     Node const* hi_base;
3451     jint hi_shift;
3452     merged_input_value = lo;
3453     bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3454     assert(is_true, "must detect con RShift");
3455     if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3456       // look through
3457       merged_input_value = merged_input_value->in(1);
3458     }
3459     if (merged_input_value != hi_base) {
3460       // merged_input_value is not the base
3461       return nullptr;
3462     }
3463   }
3464 
3465   if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3466     // Example:
3467     //
3468     //   long base = ...;
3469     //   a[0] = (byte)(base >> 0);
3470     //   a[1] = (byte)(base >> 8);
3471     //
3472     merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3473   }
3474 
3475   assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3476          (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3477          "merged_input_value is either int or long, and new_memory_size is small enough");
3478 
3479   if (_value_order == ValueOrder::Reverse) {
3480     assert(_store->memory_size() == 1, "only implemented for bytes");
3481     if (new_memory_size == 8) {
3482       merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3483     } else if (new_memory_size == 4) {
3484       merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3485     } else {
3486       assert(new_memory_size == 2, "sanity check");
3487       merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3488     }
3489   }
3490   return merged_input_value;
3491 }
3492 
3493 //                                                                                                          //
3494 // first_ctrl    first_mem   first_adr                first_ctrl    first_mem         first_adr             //
3495 //  |                |           |                     |                |                 |                 //
3496 //  |                |           |                     |                +---------------+ |                 //
3497 //  |                |           |                     |                |               | |                 //
3498 //  |                | +---------+                     |                | +---------------+                 //
3499 //  |                | |                               |                | |             | |                 //
3500 //  +--------------+ | |  v1                           +------------------------------+ | |  v1             //
3501 //  |              | | |  |                            |                | |           | | |  |              //
3502 // RangeCheck     first_store                         RangeCheck        | |          first_store            //
3503 //  |                |  |                              |                | |                |                //
3504 // last_ctrl         |  +----> unc_trap               last_ctrl         | |                +----> unc_trap  //
3505 //  |                |                       ===>      |                | |                                 //
3506 //  +--------------+ | a2 v2                           |                | |                                 //
3507 //  |              | | |  |                            |                | |                                 //
3508 //  |             second_store                         |                | |                                 //
3509 //  |                |                                 |                | | [v1 v2   ...   vn]              //
3510 // ...              ...                                |                | |         |                       //
3511 //  |                |                                 |                | |         v                       //
3512 //  +--------------+ | an vn                           +--------------+ | | merged_input_value              //
3513 //                 | | |  |                                           | | |  |                              //
3514 //                last_store (= _store)                              merged_store                           //
3515 //                                                                                                          //
3516 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3517   Node* first_store = merge_list.at(merge_list.size()-1);
3518   Node* last_ctrl   = _store->in(MemNode::Control); // after (optional) RangeCheck
3519   Node* first_mem   = first_store->in(MemNode::Memory);
3520   Node* first_adr   = first_store->in(MemNode::Address);
3521 
3522   const TypePtr* new_adr_type = _store->adr_type();
3523 
3524   int new_memory_size = _store->memory_size() * merge_list.size();
3525   BasicType bt = T_ILLEGAL;
3526   switch (new_memory_size) {
3527     case 2: bt = T_SHORT; break;
3528     case 4: bt = T_INT;   break;
3529     case 8: bt = T_LONG;  break;
3530   }
3531 
3532   StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3533                                             new_adr_type, merged_input_value, bt, MemNode::unordered);
3534 
3535   // Marking the store mismatched is sufficient to prevent reordering, since array stores
3536   // are all on the same slice. Hence, we need no barriers.
3537   merged_store->set_mismatched_access();
3538 
3539   // Constants above may now also be be packed -> put candidate on worklist
3540   _phase->is_IterGVN()->_worklist.push(first_mem);
3541 
3542   return merged_store;
3543 }
3544 
3545 #ifndef PRODUCT
3546 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3547   stringStream ss;
3548   ss.print_cr("[TraceMergeStores]: Replace");
3549   for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3550     merge_list.at(i)->dump("\n", false, &ss);
3551   }
3552   ss.print_cr("[TraceMergeStores]: with");
3553   merged_input_value->dump("\n", false, &ss);
3554   merged_store->dump("\n", false, &ss);
3555   tty->print("%s", ss.as_string());
3556 }
3557 #endif
3558 
3559 //------------------------------Ideal------------------------------------------
3560 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3561 // When a store immediately follows a relevant allocation/initialization,
3562 // try to capture it into the initialization, or hoist it above.
3563 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3564   Node* p = MemNode::Ideal_common(phase, can_reshape);
3565   if (p)  return (p == NodeSentinel) ? nullptr : p;
3566 
3567   Node* mem     = in(MemNode::Memory);
3568   Node* address = in(MemNode::Address);
3569   Node* value   = in(MemNode::ValueIn);
3570   // Back-to-back stores to same address?  Fold em up.  Generally
3571   // unsafe if I have intervening uses...
3572   if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3573     Node* st = mem;
3574     // If Store 'st' has more than one use, we cannot fold 'st' away.
3575     // For example, 'st' might be the final state at a conditional
3576     // return.  Or, 'st' might be used by some node which is live at
3577     // the same time 'st' is live, which might be unschedulable.  So,
3578     // require exactly ONE user until such time as we clone 'mem' for
3579     // each of 'mem's uses (thus making the exactly-1-user-rule hold
3580     // true).
3581     while (st->is_Store() && st->outcnt() == 1) {
3582       // Looking at a dead closed cycle of memory?
3583       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3584       assert(Opcode() == st->Opcode() ||
3585              st->Opcode() == Op_StoreVector ||
3586              Opcode() == Op_StoreVector ||
3587              st->Opcode() == Op_StoreVectorScatter ||
3588              Opcode() == Op_StoreVectorScatter ||
3589              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3590              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3591              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3592              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3593              (st->adr_type()->isa_aryptr() && st->adr_type()->is_aryptr()->is_flat()) || // TODO 8343835
3594              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3595              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3596 
3597       if (st->in(MemNode::Address)->eqv_uncast(address) &&
3598           st->as_Store()->memory_size() <= this->memory_size()) {
3599         Node* use = st->raw_out(0);
3600         if (phase->is_IterGVN()) {
3601           phase->is_IterGVN()->rehash_node_delayed(use);
3602         }
3603         // It's OK to do this in the parser, since DU info is always accurate,
3604         // and the parser always refers to nodes via SafePointNode maps.
3605         use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3606         return this;
3607       }
3608       st = st->in(MemNode::Memory);
3609     }
3610   }
3611 
3612 
3613   // Capture an unaliased, unconditional, simple store into an initializer.
3614   // Or, if it is independent of the allocation, hoist it above the allocation.
3615   if (ReduceFieldZeroing && /*can_reshape &&*/
3616       mem->is_Proj() && mem->in(0)->is_Initialize()) {
3617     InitializeNode* init = mem->in(0)->as_Initialize();
3618     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3619     if (offset > 0) {
3620       Node* moved = init->capture_store(this, offset, phase, can_reshape);
3621       // If the InitializeNode captured me, it made a raw copy of me,
3622       // and I need to disappear.
3623       if (moved != nullptr) {
3624         // %%% hack to ensure that Ideal returns a new node:
3625         mem = MergeMemNode::make(mem);
3626         return mem;             // fold me away
3627       }
3628     }
3629   }
3630 
3631   // Fold reinterpret cast into memory operation:
3632   //    StoreX mem (MoveY2X v) => StoreY mem v
3633   if (value->is_Move()) {
3634     const Type* vt = value->in(1)->bottom_type();
3635     if (has_reinterpret_variant(vt)) {
3636       if (phase->C->post_loop_opts_phase()) {
3637         return convert_to_reinterpret_store(*phase, value->in(1), vt);
3638       } else {
3639         phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3640       }
3641     }
3642   }
3643 
3644   if (MergeStores && UseUnalignedAccesses) {
3645     if (phase->C->merge_stores_phase()) {
3646       MergePrimitiveStores merge(phase, this);
3647       Node* progress = merge.run();
3648       if (progress != nullptr) { return progress; }
3649     } else {
3650       // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3651       // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3652       // the stores, and we merge the wrong sequence of stores.
3653       // Example:
3654       //   StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3655       // Apply MergeStores:
3656       //   StoreI RangeCheck [   StoreL  ] RangeCheck StoreI
3657       // Remove more RangeChecks:
3658       //   StoreI            [   StoreL  ]            StoreI
3659       // But now it would have been better to do this instead:
3660       //   [         StoreL       ] [       StoreL         ]
3661       phase->C->record_for_merge_stores_igvn(this);
3662     }
3663   }
3664 
3665   return nullptr;                  // No further progress
3666 }
3667 
3668 //------------------------------Value-----------------------------------------
3669 const Type* StoreNode::Value(PhaseGVN* phase) const {
3670   // Either input is TOP ==> the result is TOP
3671   const Type *t1 = phase->type( in(MemNode::Memory) );
3672   if( t1 == Type::TOP ) return Type::TOP;
3673   const Type *t2 = phase->type( in(MemNode::Address) );
3674   if( t2 == Type::TOP ) return Type::TOP;
3675   const Type *t3 = phase->type( in(MemNode::ValueIn) );
3676   if( t3 == Type::TOP ) return Type::TOP;
3677   return Type::MEMORY;
3678 }
3679 
3680 //------------------------------Identity---------------------------------------
3681 // Remove redundant stores:
3682 //   Store(m, p, Load(m, p)) changes to m.
3683 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3684 Node* StoreNode::Identity(PhaseGVN* phase) {
3685   Node* mem = in(MemNode::Memory);
3686   Node* adr = in(MemNode::Address);
3687   Node* val = in(MemNode::ValueIn);
3688 
3689   Node* result = this;
3690 
3691   // Load then Store?  Then the Store is useless
3692   if (val->is_Load() &&
3693       val->in(MemNode::Address)->eqv_uncast(adr) &&
3694       val->in(MemNode::Memory )->eqv_uncast(mem) &&
3695       val->as_Load()->store_Opcode() == Opcode()) {
3696     // Ensure vector type is the same
3697     if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3698       result = mem;
3699     }
3700   }
3701 
3702   // Two stores in a row of the same value?
3703   if (result == this &&
3704       mem->is_Store() &&
3705       mem->in(MemNode::Address)->eqv_uncast(adr) &&
3706       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3707       mem->Opcode() == Opcode()) {
3708     if (!is_StoreVector()) {
3709       result = mem;
3710     } else {
3711       const StoreVectorNode* store_vector = as_StoreVector();
3712       const StoreVectorNode* mem_vector = mem->as_StoreVector();
3713       const Node* store_indices = store_vector->indices();
3714       const Node* mem_indices = mem_vector->indices();
3715       const Node* store_mask = store_vector->mask();
3716       const Node* mem_mask = mem_vector->mask();
3717       // Ensure types, indices, and masks match
3718       if (store_vector->vect_type() == mem_vector->vect_type() &&
3719           ((store_indices == nullptr) == (mem_indices == nullptr) &&
3720            (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3721           ((store_mask == nullptr) == (mem_mask == nullptr) &&
3722            (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3723         result = mem;
3724       }
3725     }
3726   }
3727 
3728   // Store of zero anywhere into a freshly-allocated object?
3729   // Then the store is useless.
3730   // (It must already have been captured by the InitializeNode.)
3731   if (result == this && ReduceFieldZeroing) {
3732     // a newly allocated object is already all-zeroes everywhere
3733     if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3734         (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3735       result = mem;
3736     }
3737 
3738     if (result == this && phase->type(val)->is_zero_type()) {
3739       // the store may also apply to zero-bits in an earlier object
3740       Node* prev_mem = find_previous_store(phase);
3741       // Steps (a), (b):  Walk past independent stores to find an exact match.
3742       if (prev_mem != nullptr) {
3743         Node* prev_val = can_see_stored_value(prev_mem, phase);
3744         if (prev_val != nullptr && prev_val == val) {
3745           // prev_val and val might differ by a cast; it would be good
3746           // to keep the more informative of the two.
3747           result = mem;
3748         }
3749       }
3750     }
3751   }
3752 
3753   PhaseIterGVN* igvn = phase->is_IterGVN();
3754   if (result != this && igvn != nullptr) {
3755     MemBarNode* trailing = trailing_membar();
3756     if (trailing != nullptr) {
3757 #ifdef ASSERT
3758       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3759       assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3760 #endif
3761       trailing->remove(igvn);
3762     }
3763   }
3764 
3765   return result;
3766 }
3767 
3768 //------------------------------match_edge-------------------------------------
3769 // Do we Match on this edge index or not?  Match only memory & value
3770 uint StoreNode::match_edge(uint idx) const {
3771   return idx == MemNode::Address || idx == MemNode::ValueIn;
3772 }
3773 
3774 //------------------------------cmp--------------------------------------------
3775 // Do not common stores up together.  They generally have to be split
3776 // back up anyways, so do not bother.
3777 bool StoreNode::cmp( const Node &n ) const {
3778   return (&n == this);          // Always fail except on self
3779 }
3780 
3781 //------------------------------Ideal_masked_input-----------------------------
3782 // Check for a useless mask before a partial-word store
3783 // (StoreB ... (AndI valIn conIa) )
3784 // If (conIa & mask == mask) this simplifies to
3785 // (StoreB ... (valIn) )
3786 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3787   Node *val = in(MemNode::ValueIn);
3788   if( val->Opcode() == Op_AndI ) {
3789     const TypeInt *t = phase->type( val->in(2) )->isa_int();
3790     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3791       set_req_X(MemNode::ValueIn, val->in(1), phase);
3792       return this;
3793     }
3794   }
3795   return nullptr;
3796 }
3797 
3798 
3799 //------------------------------Ideal_sign_extended_input----------------------
3800 // Check for useless sign-extension before a partial-word store
3801 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3802 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3803 // (StoreB ... (v))
3804 // If (conIL > conIR) under some conditions, it can be simplified into
3805 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3806 // This case happens when the value of the store was itself a left shift, that
3807 // gets merged into the inner left shift of the sign-extension. For instance,
3808 // if we have
3809 // array_of_shorts[0] = (short)(v << 2)
3810 // We get a structure such as:
3811 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3812 // that is simplified into
3813 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3814 // It is thus useful to handle cases where conIL > conIR. But this simplification
3815 // does not always hold. Let's see in which cases it's valid.
3816 //
3817 // Let's assume we have the following 32 bits integer v:
3818 // +----------------------------------+
3819 // |             v[0..31]             |
3820 // +----------------------------------+
3821 //  31                               0
3822 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3823 // potentially different magnitudes.
3824 // We denote num_rejected_bits the number of bits of the discarded part. In this
3825 // case, num_rejected_bits == 24.
3826 //
3827 // Statement (proved further below in case analysis):
3828 //   Given:
3829 //   - 0 <= conIL < BitsPerJavaInteger   (no wrap in shift, enforced by maskShiftAmount)
3830 //   - 0 <= conIR < BitsPerJavaInteger   (no wrap in shift, enforced by maskShiftAmount)
3831 //   - conIL >= conIR
3832 //   - num_rejected_bits >= conIR
3833 //   Then this form:
3834 //      (RShiftI _ (LShiftI _ v conIL) conIR)
3835 //   can be replaced with this form:
3836 //      (LShiftI _ v (conIL-conIR))
3837 //
3838 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3839 //       have to be correct, as the higher bits are rejected / truncated by the store.
3840 //
3841 // The hypotheses
3842 //   0 <= conIL < BitsPerJavaInteger
3843 //   0 <= conIR < BitsPerJavaInteger
3844 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3845 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3846 // to v << 1).
3847 //
3848 //
3849 // If you don't like case analysis, jump after the conclusion.
3850 // ### Case 1 : conIL == conIR
3851 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3852 // If we do the shift left then right by 24 bits, we get:
3853 // after: v << 24
3854 // +---------+------------------------+
3855 // | v[0..7] |           0            |
3856 // +---------+------------------------+
3857 //  31     24 23                      0
3858 // after: (v << 24) >> 24
3859 // +------------------------+---------+
3860 // |        sign bit        | v[0..7] |
3861 // +------------------------+---------+
3862 //  31                     8 7        0
3863 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3864 // result) are the same before and after, so, indeed, simplifying is correct.
3865 
3866 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3867 // If we do the shift left then right by 22 bits, we get:
3868 // after: v << 22
3869 // +---------+------------------------+
3870 // | v[0..9] |           0            |
3871 // +---------+------------------------+
3872 //  31     22 21                      0
3873 // after: (v << 22) >> 22
3874 // +------------------------+---------+
3875 // |        sign bit        | v[0..9] |
3876 // +------------------------+---------+
3877 //  31                    10 9        0
3878 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3879 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3880 // still correct.
3881 
3882 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3883 // If we do the shift left then right by 26 bits, we get:
3884 // after: v << 26
3885 // +---------+------------------------+
3886 // | v[0..5] |           0            |
3887 // +---------+------------------------+
3888 //  31     26 25                      0
3889 // after: (v << 26) >> 26
3890 // +------------------------+---------+
3891 // |        sign bit        | v[0..5] |
3892 // +------------------------+---------+
3893 //  31                     6 5        0
3894 // The non-rejected bits are made of
3895 // - 0-5 => the bits 0 to 5 of v
3896 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3897 // Simplifying this as v is not correct.
3898 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3899 //
3900 // ### Case 2: conIL > conIR
3901 // ###### Case 2.1: num_rejected_bits == conIR
3902 // We take conIL == 26 for this example.
3903 // after: v << 26
3904 // +---------+------------------------+
3905 // | v[0..5] |           0            |
3906 // +---------+------------------------+
3907 //  31     26 25                      0
3908 // after: (v << 26) >> 24
3909 // +------------------+---------+-----+
3910 // |     sign bit     | v[0..5] |  0  |
3911 // +------------------+---------+-----+
3912 //  31               8 7       2 1   0
3913 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3914 // The bits 6 and 7 of v have been thrown away after the shift left.
3915 // The simplification is still correct.
3916 //
3917 // ###### Case 2.2: num_rejected_bits > conIR.
3918 // Let's say conIL == 26 and conIR == 22.
3919 // after: v << 26
3920 // +---------+------------------------+
3921 // | v[0..5] |           0            |
3922 // +---------+------------------------+
3923 //  31     26 25                      0
3924 // after: (v << 26) >> 22
3925 // +------------------+---------+-----+
3926 // |     sign bit     | v[0..5] |  0  |
3927 // +------------------+---------+-----+
3928 //  31              10 9       4 3   0
3929 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3930 // - 0-3 => 0
3931 // - 4-7 => bits 0 to 3 of v
3932 // The simplification is still correct.
3933 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3934 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3935 //
3936 // ###### But! Case 2.3: num_rejected_bits < conIR.
3937 // Let's see that this case is not as easy to simplify.
3938 // Let's say conIL == 28 and conIR == 26.
3939 // after: v << 28
3940 // +---------+------------------------+
3941 // | v[0..3] |           0            |
3942 // +---------+------------------------+
3943 //  31     28 27                      0
3944 // after: (v << 28) >> 26
3945 // +------------------+---------+-----+
3946 // |     sign bit     | v[0..3] |  0  |
3947 // +------------------+---------+-----+
3948 //  31               6 5       2 1   0
3949 // The non-rejected bits are made of
3950 // - 0-1 => 0
3951 // - 2-5 => the bits 0 to 3 of v
3952 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3953 // Simplifying this as (v << 2) is not correct.
3954 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3955 //
3956 // ### Conclusion:
3957 // Our hypotheses are indeed sufficient:
3958 //   - 0 <= conIL < BitsPerJavaInteger
3959 //   - 0 <= conIR < BitsPerJavaInteger
3960 //   - conIL >= conIR
3961 //   - num_rejected_bits >= conIR
3962 //
3963 // ### A rationale without case analysis:
3964 // After the shift left, conIL upper  bits of v are discarded and conIL lower bit
3965 // zeroes are added. After the shift right, conIR lower bits of the previous result
3966 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3967 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3968 // the point of the simplification is to get an expression of the form
3969 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3970 // simplification is not correct if conIL < conIR.
3971 //
3972 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3973 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3974 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3975 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3976 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3977 //
3978 // We do not treat the case conIL < conIR here since the point of this function is
3979 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3980 // of treating conIL > conIR comes from the cases where the sign-extended value is
3981 // also left-shift expression. Computing the sign-extension of a right-shift expression
3982 // doesn't yield a situation such as
3983 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3984 // where conIL < conIR.
3985 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3986   Node* shr = in(MemNode::ValueIn);
3987   if (shr->Opcode() == Op_RShiftI) {
3988     const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3989     if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3990       Node* shl = shr->in(1);
3991       if (shl->Opcode() == Op_LShiftI) {
3992         const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3993         if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3994           if (conIL->get_con() == conIR->get_con()) {
3995             set_req_X(MemNode::ValueIn, shl->in(1), phase);
3996             return this;
3997           }
3998           if (conIL->get_con() > conIR->get_con()) {
3999             Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
4000             set_req_X(MemNode::ValueIn, new_shl, phase);
4001             return this;
4002           }
4003         }
4004       }
4005     }
4006   }
4007   return nullptr;
4008 }
4009 
4010 //------------------------------value_never_loaded-----------------------------------
4011 // Determine whether there are any possible loads of the value stored.
4012 // For simplicity, we actually check if there are any loads from the
4013 // address stored to, not just for loads of the value stored by this node.
4014 //
4015 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
4016   Node *adr = in(Address);
4017   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
4018   if (adr_oop == nullptr)
4019     return false;
4020   if (!adr_oop->is_known_instance_field())
4021     return false; // if not a distinct instance, there may be aliases of the address
4022   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
4023     Node *use = adr->fast_out(i);
4024     if (use->is_Load() || use->is_LoadStore()) {
4025       return false;
4026     }
4027   }
4028   return true;
4029 }
4030 
4031 MemBarNode* StoreNode::trailing_membar() const {
4032   if (is_release()) {
4033     MemBarNode* trailing_mb = nullptr;
4034     for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4035       Node* u = fast_out(i);
4036       if (u->is_MemBar()) {
4037         if (u->as_MemBar()->trailing_store()) {
4038           assert(u->Opcode() == Op_MemBarVolatile, "");
4039           assert(trailing_mb == nullptr, "only one");
4040           trailing_mb = u->as_MemBar();
4041 #ifdef ASSERT
4042           Node* leading = u->as_MemBar()->leading_membar();
4043           assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4044           assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
4045           assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
4046 #endif
4047         } else {
4048           assert(u->as_MemBar()->standalone(), "");
4049         }
4050       }
4051     }
4052     return trailing_mb;
4053   }
4054   return nullptr;
4055 }
4056 
4057 
4058 //=============================================================================
4059 //------------------------------Ideal------------------------------------------
4060 // If the store is from an AND mask that leaves the low bits untouched, then
4061 // we can skip the AND operation.  If the store is from a sign-extension
4062 // (a left shift, then right shift) we can skip both.
4063 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
4064   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
4065   if( progress != nullptr ) return progress;
4066 
4067   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
4068   if( progress != nullptr ) return progress;
4069 
4070   // Finally check the default case
4071   return StoreNode::Ideal(phase, can_reshape);
4072 }
4073 
4074 //=============================================================================
4075 //------------------------------Ideal------------------------------------------
4076 // If the store is from an AND mask that leaves the low bits untouched, then
4077 // we can skip the AND operation
4078 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4079   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4080   if( progress != nullptr ) return progress;
4081 
4082   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4083   if( progress != nullptr ) return progress;
4084 
4085   // Finally check the default case
4086   return StoreNode::Ideal(phase, can_reshape);
4087 }
4088 
4089 //=============================================================================
4090 //----------------------------------SCMemProjNode------------------------------
4091 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4092 {
4093   if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4094     return Type::TOP;
4095   }
4096   return bottom_type();
4097 }
4098 
4099 //=============================================================================
4100 //----------------------------------LoadStoreNode------------------------------
4101 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4102   : Node(required),
4103     _type(rt),
4104     _adr_type(at),
4105     _barrier_data(0)
4106 {
4107   init_req(MemNode::Control, c  );
4108   init_req(MemNode::Memory , mem);
4109   init_req(MemNode::Address, adr);
4110   init_req(MemNode::ValueIn, val);
4111   init_class_id(Class_LoadStore);
4112 }
4113 
4114 //------------------------------Value-----------------------------------------
4115 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4116   // Either input is TOP ==> the result is TOP
4117   if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4118     return Type::TOP;
4119   }
4120   const Type* t = phase->type(in(MemNode::Memory));
4121   if (t == Type::TOP) {
4122     return Type::TOP;
4123   }
4124   t = phase->type(in(MemNode::Address));
4125   if (t == Type::TOP) {
4126     return Type::TOP;
4127   }
4128   t = phase->type(in(MemNode::ValueIn));
4129   if (t == Type::TOP) {
4130     return Type::TOP;
4131   }
4132   return bottom_type();
4133 }
4134 
4135 uint LoadStoreNode::ideal_reg() const {
4136   return _type->ideal_reg();
4137 }
4138 
4139 // This method conservatively checks if the result of a LoadStoreNode is
4140 // used, that is, if it returns true, then it is definitely the case that
4141 // the result of the node is not needed.
4142 // For example, GetAndAdd can be matched into a lock_add instead of a
4143 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4144 bool LoadStoreNode::result_not_used() const {
4145   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4146     Node *x = fast_out(i);
4147     if (x->Opcode() == Op_SCMemProj) {
4148       continue;
4149     }
4150     if (x->bottom_type() == TypeTuple::MEMBAR &&
4151         !x->is_Call() &&
4152         x->Opcode() != Op_Blackhole) {
4153       continue;
4154     }
4155     return false;
4156   }
4157   return true;
4158 }
4159 
4160 MemBarNode* LoadStoreNode::trailing_membar() const {
4161   MemBarNode* trailing = nullptr;
4162   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4163     Node* u = fast_out(i);
4164     if (u->is_MemBar()) {
4165       if (u->as_MemBar()->trailing_load_store()) {
4166         assert(u->Opcode() == Op_MemBarAcquire, "");
4167         assert(trailing == nullptr, "only one");
4168         trailing = u->as_MemBar();
4169 #ifdef ASSERT
4170         Node* leading = trailing->leading_membar();
4171         assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4172         assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4173         assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4174 #endif
4175       } else {
4176         assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4177       }
4178     }
4179   }
4180 
4181   return trailing;
4182 }
4183 
4184 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4185 
4186 //=============================================================================
4187 //----------------------------------LoadStoreConditionalNode--------------------
4188 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4189   init_req(ExpectedIn, ex );
4190 }
4191 
4192 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4193   // Either input is TOP ==> the result is TOP
4194   const Type* t = phase->type(in(ExpectedIn));
4195   if (t == Type::TOP) {
4196     return Type::TOP;
4197   }
4198   return LoadStoreNode::Value(phase);
4199 }
4200 
4201 //=============================================================================
4202 //-------------------------------adr_type--------------------------------------
4203 const TypePtr* ClearArrayNode::adr_type() const {
4204   Node *adr = in(3);
4205   if (adr == nullptr)  return nullptr; // node is dead
4206   return MemNode::calculate_adr_type(adr->bottom_type());
4207 }
4208 
4209 //------------------------------match_edge-------------------------------------
4210 // Do we Match on this edge index or not?  Do not match memory
4211 uint ClearArrayNode::match_edge(uint idx) const {
4212   return idx > 1;
4213 }
4214 
4215 //------------------------------Identity---------------------------------------
4216 // Clearing a zero length array does nothing
4217 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4218   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
4219 }
4220 
4221 //------------------------------Idealize---------------------------------------
4222 // Clearing a short array is faster with stores
4223 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4224   // Already know this is a large node, do not try to ideal it
4225   if (_is_large) return nullptr;
4226 
4227   const int unit = BytesPerLong;
4228   const TypeX* t = phase->type(in(2))->isa_intptr_t();
4229   if (!t)  return nullptr;
4230   if (!t->is_con())  return nullptr;
4231   intptr_t raw_count = t->get_con();
4232   intptr_t size = raw_count;
4233   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4234   // Clearing nothing uses the Identity call.
4235   // Negative clears are possible on dead ClearArrays
4236   // (see jck test stmt114.stmt11402.val).
4237   if (size <= 0 || size % unit != 0)  return nullptr;
4238   intptr_t count = size / unit;
4239   // Length too long; communicate this to matchers and assemblers.
4240   // Assemblers are responsible to produce fast hardware clears for it.
4241   if (size > InitArrayShortSize) {
4242     return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4243   } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4244     return nullptr;
4245   }
4246   if (!IdealizeClearArrayNode) return nullptr;
4247   Node *mem = in(1);
4248   if( phase->type(mem)==Type::TOP ) return nullptr;
4249   Node *adr = in(3);
4250   const Type* at = phase->type(adr);
4251   if( at==Type::TOP ) return nullptr;
4252   const TypePtr* atp = at->isa_ptr();
4253   // adjust atp to be the correct array element address type
4254   if (atp == nullptr)  atp = TypePtr::BOTTOM;
4255   else              atp = atp->add_offset(Type::OffsetBot);
4256   // Get base for derived pointer purposes
4257   if( adr->Opcode() != Op_AddP ) Unimplemented();
4258   Node *base = adr->in(1);
4259 
4260   Node *val = in(4);
4261   Node *off  = phase->MakeConX(BytesPerLong);
4262   mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4263   count--;
4264   while( count-- ) {
4265     mem = phase->transform(mem);
4266     adr = phase->transform(new AddPNode(base,adr,off));
4267     mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4268   }
4269   return mem;
4270 }
4271 
4272 //----------------------------step_through----------------------------------
4273 // Return allocation input memory edge if it is different instance
4274 // or itself if it is the one we are looking for.
4275 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4276   Node* n = *np;
4277   assert(n->is_ClearArray(), "sanity");
4278   intptr_t offset;
4279   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4280   // This method is called only before Allocate nodes are expanded
4281   // during macro nodes expansion. Before that ClearArray nodes are
4282   // only generated in PhaseMacroExpand::generate_arraycopy() (before
4283   // Allocate nodes are expanded) which follows allocations.
4284   assert(alloc != nullptr, "should have allocation");
4285   if (alloc->_idx == instance_id) {
4286     // Can not bypass initialization of the instance we are looking for.
4287     return false;
4288   }
4289   // Otherwise skip it.
4290   InitializeNode* init = alloc->initialization();
4291   if (init != nullptr)
4292     *np = init->in(TypeFunc::Memory);
4293   else
4294     *np = alloc->in(TypeFunc::Memory);
4295   return true;
4296 }
4297 
4298 //----------------------------clear_memory-------------------------------------
4299 // Generate code to initialize object storage to zero.
4300 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4301                                    Node* val,
4302                                    Node* raw_val,
4303                                    intptr_t start_offset,
4304                                    Node* end_offset,
4305                                    PhaseGVN* phase) {
4306   intptr_t offset = start_offset;
4307 
4308   int unit = BytesPerLong;
4309   if ((offset % unit) != 0) {
4310     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4311     adr = phase->transform(adr);
4312     const TypePtr* atp = TypeRawPtr::BOTTOM;
4313     if (val != nullptr) {
4314       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4315       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4316     } else {
4317       assert(raw_val == nullptr, "val may not be null");
4318       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4319     }
4320     mem = phase->transform(mem);
4321     offset += BytesPerInt;
4322   }
4323   assert((offset % unit) == 0, "");
4324 
4325   // Initialize the remaining stuff, if any, with a ClearArray.
4326   return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4327 }
4328 
4329 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4330                                    Node* raw_val,
4331                                    Node* start_offset,
4332                                    Node* end_offset,
4333                                    PhaseGVN* phase) {
4334   if (start_offset == end_offset) {
4335     // nothing to do
4336     return mem;
4337   }
4338 
4339   int unit = BytesPerLong;
4340   Node* zbase = start_offset;
4341   Node* zend  = end_offset;
4342 
4343   // Scale to the unit required by the CPU:
4344   if (!Matcher::init_array_count_is_in_bytes) {
4345     Node* shift = phase->intcon(exact_log2(unit));
4346     zbase = phase->transform(new URShiftXNode(zbase, shift) );
4347     zend  = phase->transform(new URShiftXNode(zend,  shift) );
4348   }
4349 
4350   // Bulk clear double-words
4351   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4352   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4353   if (raw_val == nullptr) {
4354     raw_val = phase->MakeConX(0);
4355   }
4356   mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4357   return phase->transform(mem);
4358 }
4359 
4360 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4361                                    Node* val,
4362                                    Node* raw_val,
4363                                    intptr_t start_offset,
4364                                    intptr_t end_offset,
4365                                    PhaseGVN* phase) {
4366   if (start_offset == end_offset) {
4367     // nothing to do
4368     return mem;
4369   }
4370 
4371   assert((end_offset % BytesPerInt) == 0, "odd end offset");
4372   intptr_t done_offset = end_offset;
4373   if ((done_offset % BytesPerLong) != 0) {
4374     done_offset -= BytesPerInt;
4375   }
4376   if (done_offset > start_offset) {
4377     mem = clear_memory(ctl, mem, dest, val, raw_val,
4378                        start_offset, phase->MakeConX(done_offset), phase);
4379   }
4380   if (done_offset < end_offset) { // emit the final 32-bit store
4381     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4382     adr = phase->transform(adr);
4383     const TypePtr* atp = TypeRawPtr::BOTTOM;
4384     if (val != nullptr) {
4385       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4386       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4387     } else {
4388       assert(raw_val == nullptr, "val may not be null");
4389       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4390     }
4391     mem = phase->transform(mem);
4392     done_offset += BytesPerInt;
4393   }
4394   assert(done_offset == end_offset, "");
4395   return mem;
4396 }
4397 
4398 //=============================================================================
4399 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4400   : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4401     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4402 #ifdef ASSERT
4403   , _pair_idx(0)
4404 #endif
4405 {
4406   init_class_id(Class_MemBar);
4407   Node* top = C->top();
4408   init_req(TypeFunc::I_O,top);
4409   init_req(TypeFunc::FramePtr,top);
4410   init_req(TypeFunc::ReturnAdr,top);
4411   if (precedent != nullptr)
4412     init_req(TypeFunc::Parms, precedent);
4413 }
4414 
4415 //------------------------------cmp--------------------------------------------
4416 uint MemBarNode::hash() const { return NO_HASH; }
4417 bool MemBarNode::cmp( const Node &n ) const {
4418   return (&n == this);          // Always fail except on self
4419 }
4420 
4421 //------------------------------make-------------------------------------------
4422 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4423   switch (opcode) {
4424   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
4425   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
4426   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
4427   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
4428   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
4429   case Op_StoreStoreFence:   return new StoreStoreFenceNode(C, atp, pn);
4430   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4431   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4432   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
4433   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
4434   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
4435   case Op_Initialize:        return new InitializeNode(C, atp, pn);
4436   default: ShouldNotReachHere(); return nullptr;
4437   }
4438 }
4439 
4440 void MemBarNode::remove(PhaseIterGVN *igvn) {
4441   if (outcnt() != 2) {
4442     assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4443     assert(outcnt() == 1, "Only control then");
4444   }
4445   if (trailing_store() || trailing_load_store()) {
4446     MemBarNode* leading = leading_membar();
4447     if (leading != nullptr) {
4448       assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4449       leading->remove(igvn);
4450     }
4451   }
4452   if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4453     igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4454   }
4455   if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4456     igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4457   }
4458 }
4459 
4460 //------------------------------Ideal------------------------------------------
4461 // Return a node which is more "ideal" than the current node.  Strip out
4462 // control copies
4463 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4464   if (remove_dead_region(phase, can_reshape)) return this;
4465   // Don't bother trying to transform a dead node
4466   if (in(0) && in(0)->is_top()) {
4467     return nullptr;
4468   }
4469 
4470   bool progress = false;
4471   // Eliminate volatile MemBars for scalar replaced objects.
4472   if (can_reshape && req() == (Precedent+1)) {
4473     bool eliminate = false;
4474     int opc = Opcode();
4475     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4476       // Volatile field loads and stores.
4477       Node* my_mem = in(MemBarNode::Precedent);
4478       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4479       if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4480         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4481         // replace this Precedent (decodeN) with the Load instead.
4482         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
4483           Node* load_node = my_mem->in(1);
4484           set_req(MemBarNode::Precedent, load_node);
4485           phase->is_IterGVN()->_worklist.push(my_mem);
4486           my_mem = load_node;
4487         } else {
4488           assert(my_mem->unique_out() == this, "sanity");
4489           assert(!trailing_load_store(), "load store node can't be eliminated");
4490           del_req(Precedent);
4491           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4492           my_mem = nullptr;
4493         }
4494         progress = true;
4495       }
4496       if (my_mem != nullptr && my_mem->is_Mem()) {
4497         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4498         // Check for scalar replaced object reference.
4499         if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4500             t_oop->offset() != Type::OffsetBot &&
4501             t_oop->offset() != Type::OffsetTop) {
4502           eliminate = true;
4503         }
4504       }
4505     } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4506       // Final field stores.
4507       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4508       if ((alloc != nullptr) && alloc->is_Allocate() &&
4509           alloc->as_Allocate()->does_not_escape_thread()) {
4510         // The allocated object does not escape.
4511         eliminate = true;
4512       }
4513     }
4514     if (eliminate) {
4515       // Replace MemBar projections by its inputs.
4516       PhaseIterGVN* igvn = phase->is_IterGVN();
4517       remove(igvn);
4518       // Must return either the original node (now dead) or a new node
4519       // (Do not return a top here, since that would break the uniqueness of top.)
4520       return new ConINode(TypeInt::ZERO);
4521     }
4522   }
4523   return progress ? this : nullptr;
4524 }
4525 
4526 //------------------------------Value------------------------------------------
4527 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4528   if( !in(0) ) return Type::TOP;
4529   if( phase->type(in(0)) == Type::TOP )
4530     return Type::TOP;
4531   return TypeTuple::MEMBAR;
4532 }
4533 
4534 //------------------------------match------------------------------------------
4535 // Construct projections for memory.
4536 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4537   switch (proj->_con) {
4538   case TypeFunc::Control:
4539   case TypeFunc::Memory:
4540     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4541   }
4542   ShouldNotReachHere();
4543   return nullptr;
4544 }
4545 
4546 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4547   trailing->_kind = TrailingStore;
4548   leading->_kind = LeadingStore;
4549 #ifdef ASSERT
4550   trailing->_pair_idx = leading->_idx;
4551   leading->_pair_idx = leading->_idx;
4552 #endif
4553 }
4554 
4555 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4556   trailing->_kind = TrailingLoadStore;
4557   leading->_kind = LeadingLoadStore;
4558 #ifdef ASSERT
4559   trailing->_pair_idx = leading->_idx;
4560   leading->_pair_idx = leading->_idx;
4561 #endif
4562 }
4563 
4564 MemBarNode* MemBarNode::trailing_membar() const {
4565   ResourceMark rm;
4566   Node* trailing = (Node*)this;
4567   VectorSet seen;
4568   Node_Stack multis(0);
4569   do {
4570     Node* c = trailing;
4571     uint i = 0;
4572     do {
4573       trailing = nullptr;
4574       for (; i < c->outcnt(); i++) {
4575         Node* next = c->raw_out(i);
4576         if (next != c && next->is_CFG()) {
4577           if (c->is_MultiBranch()) {
4578             if (multis.node() == c) {
4579               multis.set_index(i+1);
4580             } else {
4581               multis.push(c, i+1);
4582             }
4583           }
4584           trailing = next;
4585           break;
4586         }
4587       }
4588       if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4589         break;
4590       }
4591       while (multis.size() > 0) {
4592         c = multis.node();
4593         i = multis.index();
4594         if (i < c->req()) {
4595           break;
4596         }
4597         multis.pop();
4598       }
4599     } while (multis.size() > 0);
4600   } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4601 
4602   MemBarNode* mb = trailing->as_MemBar();
4603   assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4604          (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4605   assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4606   return mb;
4607 }
4608 
4609 MemBarNode* MemBarNode::leading_membar() const {
4610   ResourceMark rm;
4611   VectorSet seen;
4612   Node_Stack regions(0);
4613   Node* leading = in(0);
4614   while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4615     while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4616       leading = nullptr;
4617       while (regions.size() > 0 && leading == nullptr) {
4618         Node* r = regions.node();
4619         uint i = regions.index();
4620         if (i < r->req()) {
4621           leading = r->in(i);
4622           regions.set_index(i+1);
4623         } else {
4624           regions.pop();
4625         }
4626       }
4627       if (leading == nullptr) {
4628         assert(regions.size() == 0, "all paths should have been tried");
4629         return nullptr;
4630       }
4631     }
4632     if (leading->is_Region()) {
4633       regions.push(leading, 2);
4634       leading = leading->in(1);
4635     } else {
4636       leading = leading->in(0);
4637     }
4638   }
4639 #ifdef ASSERT
4640   Unique_Node_List wq;
4641   wq.push((Node*)this);
4642   uint found = 0;
4643   for (uint i = 0; i < wq.size(); i++) {
4644     Node* n = wq.at(i);
4645     if (n->is_Region()) {
4646       for (uint j = 1; j < n->req(); j++) {
4647         Node* in = n->in(j);
4648         if (in != nullptr && !in->is_top()) {
4649           wq.push(in);
4650         }
4651       }
4652     } else {
4653       if (n->is_MemBar() && n->as_MemBar()->leading()) {
4654         assert(n == leading, "consistency check failed");
4655         found++;
4656       } else {
4657         Node* in = n->in(0);
4658         if (in != nullptr && !in->is_top()) {
4659           wq.push(in);
4660         }
4661       }
4662     }
4663   }
4664   assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4665 #endif
4666   if (leading == nullptr) {
4667     return nullptr;
4668   }
4669   MemBarNode* mb = leading->as_MemBar();
4670   assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4671          (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4672   assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4673   return mb;
4674 }
4675 
4676 
4677 //===========================InitializeNode====================================
4678 // SUMMARY:
4679 // This node acts as a memory barrier on raw memory, after some raw stores.
4680 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4681 // The Initialize can 'capture' suitably constrained stores as raw inits.
4682 // It can coalesce related raw stores into larger units (called 'tiles').
4683 // It can avoid zeroing new storage for memory units which have raw inits.
4684 // At macro-expansion, it is marked 'complete', and does not optimize further.
4685 //
4686 // EXAMPLE:
4687 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4688 //   ctl = incoming control; mem* = incoming memory
4689 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
4690 // First allocate uninitialized memory and fill in the header:
4691 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
4692 //   ctl := alloc.Control; mem* := alloc.Memory*
4693 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
4694 // Then initialize to zero the non-header parts of the raw memory block:
4695 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4696 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4697 // After the initialize node executes, the object is ready for service:
4698 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4699 // Suppose its body is immediately initialized as {1,2}:
4700 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4701 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
4702 //   mem.SLICE(#short[*]) := store2
4703 //
4704 // DETAILS:
4705 // An InitializeNode collects and isolates object initialization after
4706 // an AllocateNode and before the next possible safepoint.  As a
4707 // memory barrier (MemBarNode), it keeps critical stores from drifting
4708 // down past any safepoint or any publication of the allocation.
4709 // Before this barrier, a newly-allocated object may have uninitialized bits.
4710 // After this barrier, it may be treated as a real oop, and GC is allowed.
4711 //
4712 // The semantics of the InitializeNode include an implicit zeroing of
4713 // the new object from object header to the end of the object.
4714 // (The object header and end are determined by the AllocateNode.)
4715 //
4716 // Certain stores may be added as direct inputs to the InitializeNode.
4717 // These stores must update raw memory, and they must be to addresses
4718 // derived from the raw address produced by AllocateNode, and with
4719 // a constant offset.  They must be ordered by increasing offset.
4720 // The first one is at in(RawStores), the last at in(req()-1).
4721 // Unlike most memory operations, they are not linked in a chain,
4722 // but are displayed in parallel as users of the rawmem output of
4723 // the allocation.
4724 //
4725 // (See comments in InitializeNode::capture_store, which continue
4726 // the example given above.)
4727 //
4728 // When the associated Allocate is macro-expanded, the InitializeNode
4729 // may be rewritten to optimize collected stores.  A ClearArrayNode
4730 // may also be created at that point to represent any required zeroing.
4731 // The InitializeNode is then marked 'complete', prohibiting further
4732 // capturing of nearby memory operations.
4733 //
4734 // During macro-expansion, all captured initializations which store
4735 // constant values of 32 bits or smaller are coalesced (if advantageous)
4736 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
4737 // initialized in fewer memory operations.  Memory words which are
4738 // covered by neither tiles nor non-constant stores are pre-zeroed
4739 // by explicit stores of zero.  (The code shape happens to do all
4740 // zeroing first, then all other stores, with both sequences occurring
4741 // in order of ascending offsets.)
4742 //
4743 // Alternatively, code may be inserted between an AllocateNode and its
4744 // InitializeNode, to perform arbitrary initialization of the new object.
4745 // E.g., the object copying intrinsics insert complex data transfers here.
4746 // The initialization must then be marked as 'complete' disable the
4747 // built-in zeroing semantics and the collection of initializing stores.
4748 //
4749 // While an InitializeNode is incomplete, reads from the memory state
4750 // produced by it are optimizable if they match the control edge and
4751 // new oop address associated with the allocation/initialization.
4752 // They return a stored value (if the offset matches) or else zero.
4753 // A write to the memory state, if it matches control and address,
4754 // and if it is to a constant offset, may be 'captured' by the
4755 // InitializeNode.  It is cloned as a raw memory operation and rewired
4756 // inside the initialization, to the raw oop produced by the allocation.
4757 // Operations on addresses which are provably distinct (e.g., to
4758 // other AllocateNodes) are allowed to bypass the initialization.
4759 //
4760 // The effect of all this is to consolidate object initialization
4761 // (both arrays and non-arrays, both piecewise and bulk) into a
4762 // single location, where it can be optimized as a unit.
4763 //
4764 // Only stores with an offset less than TrackedInitializationLimit words
4765 // will be considered for capture by an InitializeNode.  This puts a
4766 // reasonable limit on the complexity of optimized initializations.
4767 
4768 //---------------------------InitializeNode------------------------------------
4769 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4770   : MemBarNode(C, adr_type, rawoop),
4771     _is_complete(Incomplete), _does_not_escape(false)
4772 {
4773   init_class_id(Class_Initialize);
4774 
4775   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4776   assert(in(RawAddress) == rawoop, "proper init");
4777   // Note:  allocation() can be null, for secondary initialization barriers
4778 }
4779 
4780 // Since this node is not matched, it will be processed by the
4781 // register allocator.  Declare that there are no constraints
4782 // on the allocation of the RawAddress edge.
4783 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4784   // This edge should be set to top, by the set_complete.  But be conservative.
4785   if (idx == InitializeNode::RawAddress)
4786     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4787   return RegMask::Empty;
4788 }
4789 
4790 Node* InitializeNode::memory(uint alias_idx) {
4791   Node* mem = in(Memory);
4792   if (mem->is_MergeMem()) {
4793     return mem->as_MergeMem()->memory_at(alias_idx);
4794   } else {
4795     // incoming raw memory is not split
4796     return mem;
4797   }
4798 }
4799 
4800 bool InitializeNode::is_non_zero() {
4801   if (is_complete())  return false;
4802   remove_extra_zeroes();
4803   return (req() > RawStores);
4804 }
4805 
4806 void InitializeNode::set_complete(PhaseGVN* phase) {
4807   assert(!is_complete(), "caller responsibility");
4808   _is_complete = Complete;
4809 
4810   // After this node is complete, it contains a bunch of
4811   // raw-memory initializations.  There is no need for
4812   // it to have anything to do with non-raw memory effects.
4813   // Therefore, tell all non-raw users to re-optimize themselves,
4814   // after skipping the memory effects of this initialization.
4815   PhaseIterGVN* igvn = phase->is_IterGVN();
4816   if (igvn)  igvn->add_users_to_worklist(this);
4817 }
4818 
4819 // convenience function
4820 // return false if the init contains any stores already
4821 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4822   InitializeNode* init = initialization();
4823   if (init == nullptr || init->is_complete()) {
4824     return false;
4825   }
4826   init->remove_extra_zeroes();
4827   // for now, if this allocation has already collected any inits, bail:
4828   if (init->is_non_zero())  return false;
4829   init->set_complete(phase);
4830   return true;
4831 }
4832 
4833 void InitializeNode::remove_extra_zeroes() {
4834   if (req() == RawStores)  return;
4835   Node* zmem = zero_memory();
4836   uint fill = RawStores;
4837   for (uint i = fill; i < req(); i++) {
4838     Node* n = in(i);
4839     if (n->is_top() || n == zmem)  continue;  // skip
4840     if (fill < i)  set_req(fill, n);          // compact
4841     ++fill;
4842   }
4843   // delete any empty spaces created:
4844   while (fill < req()) {
4845     del_req(fill);
4846   }
4847 }
4848 
4849 // Helper for remembering which stores go with which offsets.
4850 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4851   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
4852   intptr_t offset = -1;
4853   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4854                                                phase, offset);
4855   if (base == nullptr)  return -1;  // something is dead,
4856   if (offset < 0)       return -1;  //        dead, dead
4857   return offset;
4858 }
4859 
4860 // Helper for proving that an initialization expression is
4861 // "simple enough" to be folded into an object initialization.
4862 // Attempts to prove that a store's initial value 'n' can be captured
4863 // within the initialization without creating a vicious cycle, such as:
4864 //     { Foo p = new Foo(); p.next = p; }
4865 // True for constants and parameters and small combinations thereof.
4866 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4867   ResourceMark rm;
4868   Unique_Node_List worklist;
4869   worklist.push(value);
4870 
4871   uint complexity_limit = 20;
4872   for (uint j = 0; j < worklist.size(); j++) {
4873     if (j >= complexity_limit) {
4874       return false;  // Bail out if processed too many nodes
4875     }
4876 
4877     Node* n = worklist.at(j);
4878     if (n == nullptr)   continue;   // (can this really happen?)
4879     if (n->is_Proj())   n = n->in(0);
4880     if (n == this)      return false;  // found a cycle
4881     if (n->is_Con())    continue;
4882     if (n->is_Start())  continue;   // params, etc., are OK
4883     if (n->is_Root())   continue;   // even better
4884 
4885     // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4886     if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4887       continue;
4888     }
4889 
4890     Node* ctl = n->in(0);
4891     if (ctl != nullptr && !ctl->is_top()) {
4892       if (ctl->is_Proj())  ctl = ctl->in(0);
4893       if (ctl == this)  return false;
4894 
4895       // If we already know that the enclosing memory op is pinned right after
4896       // the init, then any control flow that the store has picked up
4897       // must have preceded the init, or else be equal to the init.
4898       // Even after loop optimizations (which might change control edges)
4899       // a store is never pinned *before* the availability of its inputs.
4900       if (!MemNode::all_controls_dominate(n, this)) {
4901         return false;                  // failed to prove a good control
4902       }
4903     }
4904 
4905     // Check data edges for possible dependencies on 'this'.
4906     for (uint i = 1; i < n->req(); i++) {
4907       Node* m = n->in(i);
4908       if (m == nullptr || m == n || m->is_top())  continue;
4909 
4910       // Only process data inputs once
4911       worklist.push(m);
4912     }
4913   }
4914 
4915   return true;
4916 }
4917 
4918 // Here are all the checks a Store must pass before it can be moved into
4919 // an initialization.  Returns zero if a check fails.
4920 // On success, returns the (constant) offset to which the store applies,
4921 // within the initialized memory.
4922 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4923   const int FAIL = 0;
4924   if (st->req() != MemNode::ValueIn + 1)
4925     return FAIL;                // an inscrutable StoreNode (card mark?)
4926   Node* ctl = st->in(MemNode::Control);
4927   if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4928     return FAIL;                // must be unconditional after the initialization
4929   Node* mem = st->in(MemNode::Memory);
4930   if (!(mem->is_Proj() && mem->in(0) == this))
4931     return FAIL;                // must not be preceded by other stores
4932   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4933   if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4934       !bs->can_initialize_object(st)) {
4935     return FAIL;
4936   }
4937   Node* adr = st->in(MemNode::Address);
4938   intptr_t offset;
4939   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4940   if (alloc == nullptr)
4941     return FAIL;                // inscrutable address
4942   if (alloc != allocation())
4943     return FAIL;                // wrong allocation!  (store needs to float up)
4944   int size_in_bytes = st->memory_size();
4945   if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4946     return FAIL;                // mismatched access
4947   }
4948   Node* val = st->in(MemNode::ValueIn);
4949 
4950   if (!detect_init_independence(val, phase))
4951     return FAIL;                // stored value must be 'simple enough'
4952 
4953   // The Store can be captured only if nothing after the allocation
4954   // and before the Store is using the memory location that the store
4955   // overwrites.
4956   bool failed = false;
4957   // If is_complete_with_arraycopy() is true the shape of the graph is
4958   // well defined and is safe so no need for extra checks.
4959   if (!is_complete_with_arraycopy()) {
4960     // We are going to look at each use of the memory state following
4961     // the allocation to make sure nothing reads the memory that the
4962     // Store writes.
4963     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4964     int alias_idx = phase->C->get_alias_index(t_adr);
4965     ResourceMark rm;
4966     Unique_Node_List mems;
4967     mems.push(mem);
4968     Node* unique_merge = nullptr;
4969     for (uint next = 0; next < mems.size(); ++next) {
4970       Node *m  = mems.at(next);
4971       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4972         Node *n = m->fast_out(j);
4973         if (n->outcnt() == 0) {
4974           continue;
4975         }
4976         if (n == st) {
4977           continue;
4978         } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4979           // If the control of this use is different from the control
4980           // of the Store which is right after the InitializeNode then
4981           // this node cannot be between the InitializeNode and the
4982           // Store.
4983           continue;
4984         } else if (n->is_MergeMem()) {
4985           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4986             // We can hit a MergeMemNode (that will likely go away
4987             // later) that is a direct use of the memory state
4988             // following the InitializeNode on the same slice as the
4989             // store node that we'd like to capture. We need to check
4990             // the uses of the MergeMemNode.
4991             mems.push(n);
4992           }
4993         } else if (n->is_Mem()) {
4994           Node* other_adr = n->in(MemNode::Address);
4995           if (other_adr == adr) {
4996             failed = true;
4997             break;
4998           } else {
4999             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
5000             if (other_t_adr != nullptr) {
5001               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
5002               if (other_alias_idx == alias_idx) {
5003                 // A load from the same memory slice as the store right
5004                 // after the InitializeNode. We check the control of the
5005                 // object/array that is loaded from. If it's the same as
5006                 // the store control then we cannot capture the store.
5007                 assert(!n->is_Store(), "2 stores to same slice on same control?");
5008                 Node* base = other_adr;
5009                 if (base->is_Phi()) {
5010                   // In rare case, base may be a PhiNode and it may read
5011                   // the same memory slice between InitializeNode and store.
5012                   failed = true;
5013                   break;
5014                 }
5015                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
5016                 base = base->in(AddPNode::Base);
5017                 if (base != nullptr) {
5018                   base = base->uncast();
5019                   if (base->is_Proj() && base->in(0) == alloc) {
5020                     failed = true;
5021                     break;
5022                   }
5023                 }
5024               }
5025             }
5026           }
5027         } else {
5028           failed = true;
5029           break;
5030         }
5031       }
5032     }
5033   }
5034   if (failed) {
5035     if (!can_reshape) {
5036       // We decided we couldn't capture the store during parsing. We
5037       // should try again during the next IGVN once the graph is
5038       // cleaner.
5039       phase->C->record_for_igvn(st);
5040     }
5041     return FAIL;
5042   }
5043 
5044   return offset;                // success
5045 }
5046 
5047 // Find the captured store in(i) which corresponds to the range
5048 // [start..start+size) in the initialized object.
5049 // If there is one, return its index i.  If there isn't, return the
5050 // negative of the index where it should be inserted.
5051 // Return 0 if the queried range overlaps an initialization boundary
5052 // or if dead code is encountered.
5053 // If size_in_bytes is zero, do not bother with overlap checks.
5054 int InitializeNode::captured_store_insertion_point(intptr_t start,
5055                                                    int size_in_bytes,
5056                                                    PhaseValues* phase) {
5057   const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
5058 
5059   if (is_complete())
5060     return FAIL;                // arraycopy got here first; punt
5061 
5062   assert(allocation() != nullptr, "must be present");
5063 
5064   // no negatives, no header fields:
5065   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
5066 
5067   // after a certain size, we bail out on tracking all the stores:
5068   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5069   if (start >= ti_limit)  return FAIL;
5070 
5071   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
5072     if (i >= limit)  return -(int)i; // not found; here is where to put it
5073 
5074     Node*    st     = in(i);
5075     intptr_t st_off = get_store_offset(st, phase);
5076     if (st_off < 0) {
5077       if (st != zero_memory()) {
5078         return FAIL;            // bail out if there is dead garbage
5079       }
5080     } else if (st_off > start) {
5081       // ...we are done, since stores are ordered
5082       if (st_off < start + size_in_bytes) {
5083         return FAIL;            // the next store overlaps
5084       }
5085       return -(int)i;           // not found; here is where to put it
5086     } else if (st_off < start) {
5087       assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5088       if (size_in_bytes != 0 &&
5089           start < st_off + MAX_STORE &&
5090           start < st_off + st->as_Store()->memory_size()) {
5091         return FAIL;            // the previous store overlaps
5092       }
5093     } else {
5094       if (size_in_bytes != 0 &&
5095           st->as_Store()->memory_size() != size_in_bytes) {
5096         return FAIL;            // mismatched store size
5097       }
5098       return i;
5099     }
5100 
5101     ++i;
5102   }
5103 }
5104 
5105 // Look for a captured store which initializes at the offset 'start'
5106 // with the given size.  If there is no such store, and no other
5107 // initialization interferes, then return zero_memory (the memory
5108 // projection of the AllocateNode).
5109 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5110                                           PhaseValues* phase) {
5111   assert(stores_are_sane(phase), "");
5112   int i = captured_store_insertion_point(start, size_in_bytes, phase);
5113   if (i == 0) {
5114     return nullptr;              // something is dead
5115   } else if (i < 0) {
5116     return zero_memory();       // just primordial zero bits here
5117   } else {
5118     Node* st = in(i);           // here is the store at this position
5119     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5120     return st;
5121   }
5122 }
5123 
5124 // Create, as a raw pointer, an address within my new object at 'offset'.
5125 Node* InitializeNode::make_raw_address(intptr_t offset,
5126                                        PhaseGVN* phase) {
5127   Node* addr = in(RawAddress);
5128   if (offset != 0) {
5129     Compile* C = phase->C;
5130     addr = phase->transform( new AddPNode(C->top(), addr,
5131                                                  phase->MakeConX(offset)) );
5132   }
5133   return addr;
5134 }
5135 
5136 // Clone the given store, converting it into a raw store
5137 // initializing a field or element of my new object.
5138 // Caller is responsible for retiring the original store,
5139 // with subsume_node or the like.
5140 //
5141 // From the example above InitializeNode::InitializeNode,
5142 // here are the old stores to be captured:
5143 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5144 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
5145 //
5146 // Here is the changed code; note the extra edges on init:
5147 //   alloc = (Allocate ...)
5148 //   rawoop = alloc.RawAddress
5149 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5150 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5151 //   init = (Initialize alloc.Control alloc.Memory rawoop
5152 //                      rawstore1 rawstore2)
5153 //
5154 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5155                                     PhaseGVN* phase, bool can_reshape) {
5156   assert(stores_are_sane(phase), "");
5157 
5158   if (start < 0)  return nullptr;
5159   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5160 
5161   Compile* C = phase->C;
5162   int size_in_bytes = st->memory_size();
5163   int i = captured_store_insertion_point(start, size_in_bytes, phase);
5164   if (i == 0)  return nullptr;  // bail out
5165   Node* prev_mem = nullptr;     // raw memory for the captured store
5166   if (i > 0) {
5167     prev_mem = in(i);           // there is a pre-existing store under this one
5168     set_req(i, C->top());       // temporarily disconnect it
5169     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5170   } else {
5171     i = -i;                     // no pre-existing store
5172     prev_mem = zero_memory();   // a slice of the newly allocated object
5173     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5174       set_req(--i, C->top());   // reuse this edge; it has been folded away
5175     else
5176       ins_req(i, C->top());     // build a new edge
5177   }
5178   Node* new_st = st->clone();
5179   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5180   new_st->set_req(MemNode::Control, in(Control));
5181   new_st->set_req(MemNode::Memory,  prev_mem);
5182   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5183   bs->eliminate_gc_barrier_data(new_st);
5184   new_st = phase->transform(new_st);
5185 
5186   // At this point, new_st might have swallowed a pre-existing store
5187   // at the same offset, or perhaps new_st might have disappeared,
5188   // if it redundantly stored the same value (or zero to fresh memory).
5189 
5190   // In any case, wire it in:
5191   PhaseIterGVN* igvn = phase->is_IterGVN();
5192   if (igvn) {
5193     igvn->rehash_node_delayed(this);
5194   }
5195   set_req(i, new_st);
5196 
5197   // The caller may now kill the old guy.
5198   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5199   assert(check_st == new_st || check_st == nullptr, "must be findable");
5200   assert(!is_complete(), "");
5201   return new_st;
5202 }
5203 
5204 static bool store_constant(jlong* tiles, int num_tiles,
5205                            intptr_t st_off, int st_size,
5206                            jlong con) {
5207   if ((st_off & (st_size-1)) != 0)
5208     return false;               // strange store offset (assume size==2**N)
5209   address addr = (address)tiles + st_off;
5210   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5211   switch (st_size) {
5212   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
5213   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
5214   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
5215   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
5216   default: return false;        // strange store size (detect size!=2**N here)
5217   }
5218   return true;                  // return success to caller
5219 }
5220 
5221 // Coalesce subword constants into int constants and possibly
5222 // into long constants.  The goal, if the CPU permits,
5223 // is to initialize the object with a small number of 64-bit tiles.
5224 // Also, convert floating-point constants to bit patterns.
5225 // Non-constants are not relevant to this pass.
5226 //
5227 // In terms of the running example on InitializeNode::InitializeNode
5228 // and InitializeNode::capture_store, here is the transformation
5229 // of rawstore1 and rawstore2 into rawstore12:
5230 //   alloc = (Allocate ...)
5231 //   rawoop = alloc.RawAddress
5232 //   tile12 = 0x00010002
5233 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5234 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5235 //
5236 void
5237 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5238                                         Node* size_in_bytes,
5239                                         PhaseGVN* phase) {
5240   Compile* C = phase->C;
5241 
5242   assert(stores_are_sane(phase), "");
5243   // Note:  After this pass, they are not completely sane,
5244   // since there may be some overlaps.
5245 
5246   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5247 
5248   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5249   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5250   size_limit = MIN2(size_limit, ti_limit);
5251   size_limit = align_up(size_limit, BytesPerLong);
5252   int num_tiles = size_limit / BytesPerLong;
5253 
5254   // allocate space for the tile map:
5255   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5256   jlong  tiles_buf[small_len];
5257   Node*  nodes_buf[small_len];
5258   jlong  inits_buf[small_len];
5259   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5260                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5261   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5262                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5263   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5264                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5265   // tiles: exact bitwise model of all primitive constants
5266   // nodes: last constant-storing node subsumed into the tiles model
5267   // inits: which bytes (in each tile) are touched by any initializations
5268 
5269   //// Pass A: Fill in the tile model with any relevant stores.
5270 
5271   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5272   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5273   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5274   Node* zmem = zero_memory(); // initially zero memory state
5275   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5276     Node* st = in(i);
5277     intptr_t st_off = get_store_offset(st, phase);
5278 
5279     // Figure out the store's offset and constant value:
5280     if (st_off < header_size)             continue; //skip (ignore header)
5281     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
5282     int st_size = st->as_Store()->memory_size();
5283     if (st_off + st_size > size_limit)    break;
5284 
5285     // Record which bytes are touched, whether by constant or not.
5286     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5287       continue;                 // skip (strange store size)
5288 
5289     const Type* val = phase->type(st->in(MemNode::ValueIn));
5290     if (!val->singleton())                continue; //skip (non-con store)
5291     BasicType type = val->basic_type();
5292 
5293     jlong con = 0;
5294     switch (type) {
5295     case T_INT:    con = val->is_int()->get_con();  break;
5296     case T_LONG:   con = val->is_long()->get_con(); break;
5297     case T_FLOAT:  con = jint_cast(val->getf());    break;
5298     case T_DOUBLE: con = jlong_cast(val->getd());   break;
5299     default:                              continue; //skip (odd store type)
5300     }
5301 
5302     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5303         st->Opcode() == Op_StoreL) {
5304       continue;                 // This StoreL is already optimal.
5305     }
5306 
5307     // Store down the constant.
5308     store_constant(tiles, num_tiles, st_off, st_size, con);
5309 
5310     intptr_t j = st_off >> LogBytesPerLong;
5311 
5312     if (type == T_INT && st_size == BytesPerInt
5313         && (st_off & BytesPerInt) == BytesPerInt) {
5314       jlong lcon = tiles[j];
5315       if (!Matcher::isSimpleConstant64(lcon) &&
5316           st->Opcode() == Op_StoreI) {
5317         // This StoreI is already optimal by itself.
5318         jint* intcon = (jint*) &tiles[j];
5319         intcon[1] = 0;  // undo the store_constant()
5320 
5321         // If the previous store is also optimal by itself, back up and
5322         // undo the action of the previous loop iteration... if we can.
5323         // But if we can't, just let the previous half take care of itself.
5324         st = nodes[j];
5325         st_off -= BytesPerInt;
5326         con = intcon[0];
5327         if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5328           assert(st_off >= header_size, "still ignoring header");
5329           assert(get_store_offset(st, phase) == st_off, "must be");
5330           assert(in(i-1) == zmem, "must be");
5331           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5332           assert(con == tcon->is_int()->get_con(), "must be");
5333           // Undo the effects of the previous loop trip, which swallowed st:
5334           intcon[0] = 0;        // undo store_constant()
5335           set_req(i-1, st);     // undo set_req(i, zmem)
5336           nodes[j] = nullptr;   // undo nodes[j] = st
5337           --old_subword;        // undo ++old_subword
5338         }
5339         continue;               // This StoreI is already optimal.
5340       }
5341     }
5342 
5343     // This store is not needed.
5344     set_req(i, zmem);
5345     nodes[j] = st;              // record for the moment
5346     if (st_size < BytesPerLong) // something has changed
5347           ++old_subword;        // includes int/float, but who's counting...
5348     else  ++old_long;
5349   }
5350 
5351   if ((old_subword + old_long) == 0)
5352     return;                     // nothing more to do
5353 
5354   //// Pass B: Convert any non-zero tiles into optimal constant stores.
5355   // Be sure to insert them before overlapping non-constant stores.
5356   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
5357   for (int j = 0; j < num_tiles; j++) {
5358     jlong con  = tiles[j];
5359     jlong init = inits[j];
5360     if (con == 0)  continue;
5361     jint con0,  con1;           // split the constant, address-wise
5362     jint init0, init1;          // split the init map, address-wise
5363     { union { jlong con; jint intcon[2]; } u;
5364       u.con = con;
5365       con0  = u.intcon[0];
5366       con1  = u.intcon[1];
5367       u.con = init;
5368       init0 = u.intcon[0];
5369       init1 = u.intcon[1];
5370     }
5371 
5372     Node* old = nodes[j];
5373     assert(old != nullptr, "need the prior store");
5374     intptr_t offset = (j * BytesPerLong);
5375 
5376     bool split = !Matcher::isSimpleConstant64(con);
5377 
5378     if (offset < header_size) {
5379       assert(offset + BytesPerInt >= header_size, "second int counts");
5380       assert(*(jint*)&tiles[j] == 0, "junk in header");
5381       split = true;             // only the second word counts
5382       // Example:  int a[] = { 42 ... }
5383     } else if (con0 == 0 && init0 == -1) {
5384       split = true;             // first word is covered by full inits
5385       // Example:  int a[] = { ... foo(), 42 ... }
5386     } else if (con1 == 0 && init1 == -1) {
5387       split = true;             // second word is covered by full inits
5388       // Example:  int a[] = { ... 42, foo() ... }
5389     }
5390 
5391     // Here's a case where init0 is neither 0 nor -1:
5392     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
5393     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5394     // In this case the tile is not split; it is (jlong)42.
5395     // The big tile is stored down, and then the foo() value is inserted.
5396     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5397 
5398     Node* ctl = old->in(MemNode::Control);
5399     Node* adr = make_raw_address(offset, phase);
5400     const TypePtr* atp = TypeRawPtr::BOTTOM;
5401 
5402     // One or two coalesced stores to plop down.
5403     Node*    st[2];
5404     intptr_t off[2];
5405     int  nst = 0;
5406     if (!split) {
5407       ++new_long;
5408       off[nst] = offset;
5409       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5410                                   phase->longcon(con), T_LONG, MemNode::unordered);
5411     } else {
5412       // Omit either if it is a zero.
5413       if (con0 != 0) {
5414         ++new_int;
5415         off[nst]  = offset;
5416         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5417                                     phase->intcon(con0), T_INT, MemNode::unordered);
5418       }
5419       if (con1 != 0) {
5420         ++new_int;
5421         offset += BytesPerInt;
5422         adr = make_raw_address(offset, phase);
5423         off[nst]  = offset;
5424         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5425                                     phase->intcon(con1), T_INT, MemNode::unordered);
5426       }
5427     }
5428 
5429     // Insert second store first, then the first before the second.
5430     // Insert each one just before any overlapping non-constant stores.
5431     while (nst > 0) {
5432       Node* st1 = st[--nst];
5433       C->copy_node_notes_to(st1, old);
5434       st1 = phase->transform(st1);
5435       offset = off[nst];
5436       assert(offset >= header_size, "do not smash header");
5437       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5438       guarantee(ins_idx != 0, "must re-insert constant store");
5439       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
5440       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5441         set_req(--ins_idx, st1);
5442       else
5443         ins_req(ins_idx, st1);
5444     }
5445   }
5446 
5447   if (PrintCompilation && WizardMode)
5448     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5449                   old_subword, old_long, new_int, new_long);
5450   if (C->log() != nullptr)
5451     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5452                    old_subword, old_long, new_int, new_long);
5453 
5454   // Clean up any remaining occurrences of zmem:
5455   remove_extra_zeroes();
5456 }
5457 
5458 // Explore forward from in(start) to find the first fully initialized
5459 // word, and return its offset.  Skip groups of subword stores which
5460 // together initialize full words.  If in(start) is itself part of a
5461 // fully initialized word, return the offset of in(start).  If there
5462 // are no following full-word stores, or if something is fishy, return
5463 // a negative value.
5464 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5465   int       int_map = 0;
5466   intptr_t  int_map_off = 0;
5467   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
5468 
5469   for (uint i = start, limit = req(); i < limit; i++) {
5470     Node* st = in(i);
5471 
5472     intptr_t st_off = get_store_offset(st, phase);
5473     if (st_off < 0)  break;  // return conservative answer
5474 
5475     int st_size = st->as_Store()->memory_size();
5476     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5477       return st_off;            // we found a complete word init
5478     }
5479 
5480     // update the map:
5481 
5482     intptr_t this_int_off = align_down(st_off, BytesPerInt);
5483     if (this_int_off != int_map_off) {
5484       // reset the map:
5485       int_map = 0;
5486       int_map_off = this_int_off;
5487     }
5488 
5489     int subword_off = st_off - this_int_off;
5490     int_map |= right_n_bits(st_size) << subword_off;
5491     if ((int_map & FULL_MAP) == FULL_MAP) {
5492       return this_int_off;      // we found a complete word init
5493     }
5494 
5495     // Did this store hit or cross the word boundary?
5496     intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5497     if (next_int_off == this_int_off + BytesPerInt) {
5498       // We passed the current int, without fully initializing it.
5499       int_map_off = next_int_off;
5500       int_map >>= BytesPerInt;
5501     } else if (next_int_off > this_int_off + BytesPerInt) {
5502       // We passed the current and next int.
5503       return this_int_off + BytesPerInt;
5504     }
5505   }
5506 
5507   return -1;
5508 }
5509 
5510 
5511 // Called when the associated AllocateNode is expanded into CFG.
5512 // At this point, we may perform additional optimizations.
5513 // Linearize the stores by ascending offset, to make memory
5514 // activity as coherent as possible.
5515 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5516                                       intptr_t header_size,
5517                                       Node* size_in_bytes,
5518                                       PhaseIterGVN* phase) {
5519   assert(!is_complete(), "not already complete");
5520   assert(stores_are_sane(phase), "");
5521   assert(allocation() != nullptr, "must be present");
5522 
5523   remove_extra_zeroes();
5524 
5525   if (ReduceFieldZeroing || ReduceBulkZeroing)
5526     // reduce instruction count for common initialization patterns
5527     coalesce_subword_stores(header_size, size_in_bytes, phase);
5528 
5529   Node* zmem = zero_memory();   // initially zero memory state
5530   Node* inits = zmem;           // accumulating a linearized chain of inits
5531   #ifdef ASSERT
5532   intptr_t first_offset = allocation()->minimum_header_size();
5533   intptr_t last_init_off = first_offset;  // previous init offset
5534   intptr_t last_init_end = first_offset;  // previous init offset+size
5535   intptr_t last_tile_end = first_offset;  // previous tile offset+size
5536   #endif
5537   intptr_t zeroes_done = header_size;
5538 
5539   bool do_zeroing = true;       // we might give up if inits are very sparse
5540   int  big_init_gaps = 0;       // how many large gaps have we seen?
5541 
5542   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
5543   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
5544 
5545   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5546     Node* st = in(i);
5547     intptr_t st_off = get_store_offset(st, phase);
5548     if (st_off < 0)
5549       break;                    // unknown junk in the inits
5550     if (st->in(MemNode::Memory) != zmem)
5551       break;                    // complicated store chains somehow in list
5552 
5553     int st_size = st->as_Store()->memory_size();
5554     intptr_t next_init_off = st_off + st_size;
5555 
5556     if (do_zeroing && zeroes_done < next_init_off) {
5557       // See if this store needs a zero before it or under it.
5558       intptr_t zeroes_needed = st_off;
5559 
5560       if (st_size < BytesPerInt) {
5561         // Look for subword stores which only partially initialize words.
5562         // If we find some, we must lay down some word-level zeroes first,
5563         // underneath the subword stores.
5564         //
5565         // Examples:
5566         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
5567         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
5568         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
5569         //
5570         // Note:  coalesce_subword_stores may have already done this,
5571         // if it was prompted by constant non-zero subword initializers.
5572         // But this case can still arise with non-constant stores.
5573 
5574         intptr_t next_full_store = find_next_fullword_store(i, phase);
5575 
5576         // In the examples above:
5577         //   in(i)          p   q   r   s     x   y     z
5578         //   st_off        12  13  14  15    12  13    14
5579         //   st_size        1   1   1   1     1   1     1
5580         //   next_full_s.  12  16  16  16    16  16    16
5581         //   z's_done      12  16  16  16    12  16    12
5582         //   z's_needed    12  16  16  16    16  16    16
5583         //   zsize          0   0   0   0     4   0     4
5584         if (next_full_store < 0) {
5585           // Conservative tack:  Zero to end of current word.
5586           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5587         } else {
5588           // Zero to beginning of next fully initialized word.
5589           // Or, don't zero at all, if we are already in that word.
5590           assert(next_full_store >= zeroes_needed, "must go forward");
5591           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5592           zeroes_needed = next_full_store;
5593         }
5594       }
5595 
5596       if (zeroes_needed > zeroes_done) {
5597         intptr_t zsize = zeroes_needed - zeroes_done;
5598         // Do some incremental zeroing on rawmem, in parallel with inits.
5599         zeroes_done = align_down(zeroes_done, BytesPerInt);
5600         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5601                                               allocation()->in(AllocateNode::InitValue),
5602                                               allocation()->in(AllocateNode::RawInitValue),
5603                                               zeroes_done, zeroes_needed,
5604                                               phase);
5605         zeroes_done = zeroes_needed;
5606         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5607           do_zeroing = false;   // leave the hole, next time
5608       }
5609     }
5610 
5611     // Collect the store and move on:
5612     phase->replace_input_of(st, MemNode::Memory, inits);
5613     inits = st;                 // put it on the linearized chain
5614     set_req(i, zmem);           // unhook from previous position
5615 
5616     if (zeroes_done == st_off)
5617       zeroes_done = next_init_off;
5618 
5619     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5620 
5621     #ifdef ASSERT
5622     // Various order invariants.  Weaker than stores_are_sane because
5623     // a large constant tile can be filled in by smaller non-constant stores.
5624     assert(st_off >= last_init_off, "inits do not reverse");
5625     last_init_off = st_off;
5626     const Type* val = nullptr;
5627     if (st_size >= BytesPerInt &&
5628         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5629         (int)val->basic_type() < (int)T_OBJECT) {
5630       assert(st_off >= last_tile_end, "tiles do not overlap");
5631       assert(st_off >= last_init_end, "tiles do not overwrite inits");
5632       last_tile_end = MAX2(last_tile_end, next_init_off);
5633     } else {
5634       intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5635       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5636       assert(st_off      >= last_init_end, "inits do not overlap");
5637       last_init_end = next_init_off;  // it's a non-tile
5638     }
5639     #endif //ASSERT
5640   }
5641 
5642   remove_extra_zeroes();        // clear out all the zmems left over
5643   add_req(inits);
5644 
5645   if (!(UseTLAB && ZeroTLAB)) {
5646     // If anything remains to be zeroed, zero it all now.
5647     zeroes_done = align_down(zeroes_done, BytesPerInt);
5648     // if it is the last unused 4 bytes of an instance, forget about it
5649     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5650     if (zeroes_done + BytesPerLong >= size_limit) {
5651       AllocateNode* alloc = allocation();
5652       assert(alloc != nullptr, "must be present");
5653       if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5654         Node* klass_node = alloc->in(AllocateNode::KlassNode);
5655         ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5656         if (zeroes_done == k->layout_helper())
5657           zeroes_done = size_limit;
5658       }
5659     }
5660     if (zeroes_done < size_limit) {
5661       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5662                                             allocation()->in(AllocateNode::InitValue),
5663                                             allocation()->in(AllocateNode::RawInitValue),
5664                                             zeroes_done, size_in_bytes, phase);
5665     }
5666   }
5667 
5668   set_complete(phase);
5669   return rawmem;
5670 }
5671 
5672 
5673 #ifdef ASSERT
5674 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5675   if (is_complete())
5676     return true;                // stores could be anything at this point
5677   assert(allocation() != nullptr, "must be present");
5678   intptr_t last_off = allocation()->minimum_header_size();
5679   for (uint i = InitializeNode::RawStores; i < req(); i++) {
5680     Node* st = in(i);
5681     intptr_t st_off = get_store_offset(st, phase);
5682     if (st_off < 0)  continue;  // ignore dead garbage
5683     if (last_off > st_off) {
5684       tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5685       this->dump(2);
5686       assert(false, "ascending store offsets");
5687       return false;
5688     }
5689     last_off = st_off + st->as_Store()->memory_size();
5690   }
5691   return true;
5692 }
5693 #endif //ASSERT
5694 
5695 
5696 
5697 
5698 //============================MergeMemNode=====================================
5699 //
5700 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
5701 // contributing store or call operations.  Each contributor provides the memory
5702 // state for a particular "alias type" (see Compile::alias_type).  For example,
5703 // if a MergeMem has an input X for alias category #6, then any memory reference
5704 // to alias category #6 may use X as its memory state input, as an exact equivalent
5705 // to using the MergeMem as a whole.
5706 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5707 //
5708 // (Here, the <N> notation gives the index of the relevant adr_type.)
5709 //
5710 // In one special case (and more cases in the future), alias categories overlap.
5711 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5712 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
5713 // it is exactly equivalent to that state W:
5714 //   MergeMem(<Bot>: W) <==> W
5715 //
5716 // Usually, the merge has more than one input.  In that case, where inputs
5717 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5718 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5719 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5720 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5721 //
5722 // A merge can take a "wide" memory state as one of its narrow inputs.
5723 // This simply means that the merge observes out only the relevant parts of
5724 // the wide input.  That is, wide memory states arriving at narrow merge inputs
5725 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
5726 //
5727 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5728 // and that memory slices "leak through":
5729 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5730 //
5731 // But, in such a cascade, repeated memory slices can "block the leak":
5732 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5733 //
5734 // In the last example, Y is not part of the combined memory state of the
5735 // outermost MergeMem.  The system must, of course, prevent unschedulable
5736 // memory states from arising, so you can be sure that the state Y is somehow
5737 // a precursor to state Y'.
5738 //
5739 //
5740 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5741 // of each MergeMemNode array are exactly the numerical alias indexes, including
5742 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
5743 // Compile::alias_type (and kin) produce and manage these indexes.
5744 //
5745 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5746 // (Note that this provides quick access to the top node inside MergeMem methods,
5747 // without the need to reach out via TLS to Compile::current.)
5748 //
5749 // As a consequence of what was just described, a MergeMem that represents a full
5750 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5751 // containing all alias categories.
5752 //
5753 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5754 //
5755 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5756 // a memory state for the alias type <N>, or else the top node, meaning that
5757 // there is no particular input for that alias type.  Note that the length of
5758 // a MergeMem is variable, and may be extended at any time to accommodate new
5759 // memory states at larger alias indexes.  When merges grow, they are of course
5760 // filled with "top" in the unused in() positions.
5761 //
5762 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5763 // (Top was chosen because it works smoothly with passes like GCM.)
5764 //
5765 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
5766 // the type of random VM bits like TLS references.)  Since it is always the
5767 // first non-Bot memory slice, some low-level loops use it to initialize an
5768 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
5769 //
5770 //
5771 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
5772 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
5773 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5774 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
5775 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
5776 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5777 //
5778 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5779 // really that different from the other memory inputs.  An abbreviation called
5780 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5781 //
5782 //
5783 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
5784 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
5785 // that "emerges though" the base memory will be marked as excluding the alias types
5786 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5787 //
5788 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5789 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5790 //
5791 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5792 // (It is currently unimplemented.)  As you can see, the resulting merge is
5793 // actually a disjoint union of memory states, rather than an overlay.
5794 //
5795 
5796 //------------------------------MergeMemNode-----------------------------------
5797 Node* MergeMemNode::make_empty_memory() {
5798   Node* empty_memory = (Node*) Compile::current()->top();
5799   assert(empty_memory->is_top(), "correct sentinel identity");
5800   return empty_memory;
5801 }
5802 
5803 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5804   init_class_id(Class_MergeMem);
5805   // all inputs are nullified in Node::Node(int)
5806   // set_input(0, nullptr);  // no control input
5807 
5808   // Initialize the edges uniformly to top, for starters.
5809   Node* empty_mem = make_empty_memory();
5810   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5811     init_req(i,empty_mem);
5812   }
5813   assert(empty_memory() == empty_mem, "");
5814 
5815   if( new_base != nullptr && new_base->is_MergeMem() ) {
5816     MergeMemNode* mdef = new_base->as_MergeMem();
5817     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5818     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5819       mms.set_memory(mms.memory2());
5820     }
5821     assert(base_memory() == mdef->base_memory(), "");
5822   } else {
5823     set_base_memory(new_base);
5824   }
5825 }
5826 
5827 // Make a new, untransformed MergeMem with the same base as 'mem'.
5828 // If mem is itself a MergeMem, populate the result with the same edges.
5829 MergeMemNode* MergeMemNode::make(Node* mem) {
5830   return new MergeMemNode(mem);
5831 }
5832 
5833 //------------------------------cmp--------------------------------------------
5834 uint MergeMemNode::hash() const { return NO_HASH; }
5835 bool MergeMemNode::cmp( const Node &n ) const {
5836   return (&n == this);          // Always fail except on self
5837 }
5838 
5839 //------------------------------Identity---------------------------------------
5840 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5841   // Identity if this merge point does not record any interesting memory
5842   // disambiguations.
5843   Node* base_mem = base_memory();
5844   Node* empty_mem = empty_memory();
5845   if (base_mem != empty_mem) {  // Memory path is not dead?
5846     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5847       Node* mem = in(i);
5848       if (mem != empty_mem && mem != base_mem) {
5849         return this;            // Many memory splits; no change
5850       }
5851     }
5852   }
5853   return base_mem;              // No memory splits; ID on the one true input
5854 }
5855 
5856 //------------------------------Ideal------------------------------------------
5857 // This method is invoked recursively on chains of MergeMem nodes
5858 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5859   // Remove chain'd MergeMems
5860   //
5861   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5862   // relative to the "in(Bot)".  Since we are patching both at the same time,
5863   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5864   // but rewrite each "in(i)" relative to the new "in(Bot)".
5865   Node *progress = nullptr;
5866 
5867 
5868   Node* old_base = base_memory();
5869   Node* empty_mem = empty_memory();
5870   if (old_base == empty_mem)
5871     return nullptr; // Dead memory path.
5872 
5873   MergeMemNode* old_mbase;
5874   if (old_base != nullptr && old_base->is_MergeMem())
5875     old_mbase = old_base->as_MergeMem();
5876   else
5877     old_mbase = nullptr;
5878   Node* new_base = old_base;
5879 
5880   // simplify stacked MergeMems in base memory
5881   if (old_mbase)  new_base = old_mbase->base_memory();
5882 
5883   // the base memory might contribute new slices beyond my req()
5884   if (old_mbase)  grow_to_match(old_mbase);
5885 
5886   // Note:  We do not call verify_sparse on entry, because inputs
5887   // can normalize to the base_memory via subsume_node or similar
5888   // mechanisms.  This method repairs that damage.
5889 
5890   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5891 
5892   // Look at each slice.
5893   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5894     Node* old_in = in(i);
5895     // calculate the old memory value
5896     Node* old_mem = old_in;
5897     if (old_mem == empty_mem)  old_mem = old_base;
5898     assert(old_mem == memory_at(i), "");
5899 
5900     // maybe update (reslice) the old memory value
5901 
5902     // simplify stacked MergeMems
5903     Node* new_mem = old_mem;
5904     MergeMemNode* old_mmem;
5905     if (old_mem != nullptr && old_mem->is_MergeMem())
5906       old_mmem = old_mem->as_MergeMem();
5907     else
5908       old_mmem = nullptr;
5909     if (old_mmem == this) {
5910       // This can happen if loops break up and safepoints disappear.
5911       // A merge of BotPtr (default) with a RawPtr memory derived from a
5912       // safepoint can be rewritten to a merge of the same BotPtr with
5913       // the BotPtr phi coming into the loop.  If that phi disappears
5914       // also, we can end up with a self-loop of the mergemem.
5915       // In general, if loops degenerate and memory effects disappear,
5916       // a mergemem can be left looking at itself.  This simply means
5917       // that the mergemem's default should be used, since there is
5918       // no longer any apparent effect on this slice.
5919       // Note: If a memory slice is a MergeMem cycle, it is unreachable
5920       //       from start.  Update the input to TOP.
5921       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5922     }
5923     else if (old_mmem != nullptr) {
5924       new_mem = old_mmem->memory_at(i);
5925     }
5926     // else preceding memory was not a MergeMem
5927 
5928     // maybe store down a new value
5929     Node* new_in = new_mem;
5930     if (new_in == new_base)  new_in = empty_mem;
5931 
5932     if (new_in != old_in) {
5933       // Warning:  Do not combine this "if" with the previous "if"
5934       // A memory slice might have be be rewritten even if it is semantically
5935       // unchanged, if the base_memory value has changed.
5936       set_req_X(i, new_in, phase);
5937       progress = this;          // Report progress
5938     }
5939   }
5940 
5941   if (new_base != old_base) {
5942     set_req_X(Compile::AliasIdxBot, new_base, phase);
5943     // Don't use set_base_memory(new_base), because we need to update du.
5944     assert(base_memory() == new_base, "");
5945     progress = this;
5946   }
5947 
5948   if( base_memory() == this ) {
5949     // a self cycle indicates this memory path is dead
5950     set_req(Compile::AliasIdxBot, empty_mem);
5951   }
5952 
5953   // Resolve external cycles by calling Ideal on a MergeMem base_memory
5954   // Recursion must occur after the self cycle check above
5955   if( base_memory()->is_MergeMem() ) {
5956     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5957     Node *m = phase->transform(new_mbase);  // Rollup any cycles
5958     if( m != nullptr &&
5959         (m->is_top() ||
5960          (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5961       // propagate rollup of dead cycle to self
5962       set_req(Compile::AliasIdxBot, empty_mem);
5963     }
5964   }
5965 
5966   if( base_memory() == empty_mem ) {
5967     progress = this;
5968     // Cut inputs during Parse phase only.
5969     // During Optimize phase a dead MergeMem node will be subsumed by Top.
5970     if( !can_reshape ) {
5971       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5972         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5973       }
5974     }
5975   }
5976 
5977   if( !progress && base_memory()->is_Phi() && can_reshape ) {
5978     // Check if PhiNode::Ideal's "Split phis through memory merges"
5979     // transform should be attempted. Look for this->phi->this cycle.
5980     uint merge_width = req();
5981     if (merge_width > Compile::AliasIdxRaw) {
5982       PhiNode* phi = base_memory()->as_Phi();
5983       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5984         if (phi->in(i) == this) {
5985           phase->is_IterGVN()->_worklist.push(phi);
5986           break;
5987         }
5988       }
5989     }
5990   }
5991 
5992   assert(progress || verify_sparse(), "please, no dups of base");
5993   return progress;
5994 }
5995 
5996 //-------------------------set_base_memory-------------------------------------
5997 void MergeMemNode::set_base_memory(Node *new_base) {
5998   Node* empty_mem = empty_memory();
5999   set_req(Compile::AliasIdxBot, new_base);
6000   assert(memory_at(req()) == new_base, "must set default memory");
6001   // Clear out other occurrences of new_base:
6002   if (new_base != empty_mem) {
6003     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6004       if (in(i) == new_base)  set_req(i, empty_mem);
6005     }
6006   }
6007 }
6008 
6009 //------------------------------out_RegMask------------------------------------
6010 const RegMask &MergeMemNode::out_RegMask() const {
6011   return RegMask::Empty;
6012 }
6013 
6014 //------------------------------dump_spec--------------------------------------
6015 #ifndef PRODUCT
6016 void MergeMemNode::dump_spec(outputStream *st) const {
6017   st->print(" {");
6018   Node* base_mem = base_memory();
6019   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
6020     Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
6021     if (mem == base_mem) { st->print(" -"); continue; }
6022     st->print( " N%d:", mem->_idx );
6023     Compile::current()->get_adr_type(i)->dump_on(st);
6024   }
6025   st->print(" }");
6026 }
6027 #endif // !PRODUCT
6028 
6029 
6030 #ifdef ASSERT
6031 static bool might_be_same(Node* a, Node* b) {
6032   if (a == b)  return true;
6033   if (!(a->is_Phi() || b->is_Phi()))  return false;
6034   // phis shift around during optimization
6035   return true;  // pretty stupid...
6036 }
6037 
6038 // verify a narrow slice (either incoming or outgoing)
6039 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
6040   if (!VerifyAliases)                return;  // don't bother to verify unless requested
6041   if (VMError::is_error_reported())  return;  // muzzle asserts when debugging an error
6042   if (Node::in_dump())               return;  // muzzle asserts when printing
6043   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
6044   assert(n != nullptr, "");
6045   // Elide intervening MergeMem's
6046   while (n->is_MergeMem()) {
6047     n = n->as_MergeMem()->memory_at(alias_idx);
6048   }
6049   Compile* C = Compile::current();
6050   const TypePtr* n_adr_type = n->adr_type();
6051   if (n == m->empty_memory()) {
6052     // Implicit copy of base_memory()
6053   } else if (n_adr_type != TypePtr::BOTTOM) {
6054     assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6055     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6056   } else {
6057     // A few places like make_runtime_call "know" that VM calls are narrow,
6058     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6059     bool expected_wide_mem = false;
6060     if (n == m->base_memory()) {
6061       expected_wide_mem = true;
6062     } else if (alias_idx == Compile::AliasIdxRaw ||
6063                n == m->memory_at(Compile::AliasIdxRaw)) {
6064       expected_wide_mem = true;
6065     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6066       // memory can "leak through" calls on channels that
6067       // are write-once.  Allow this also.
6068       expected_wide_mem = true;
6069     }
6070     assert(expected_wide_mem, "expected narrow slice replacement");
6071   }
6072 }
6073 #else // !ASSERT
6074 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
6075 #endif
6076 
6077 
6078 //-----------------------------memory_at---------------------------------------
6079 Node* MergeMemNode::memory_at(uint alias_idx) const {
6080   assert(alias_idx >= Compile::AliasIdxRaw ||
6081          (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6082          "must avoid base_memory and AliasIdxTop");
6083 
6084   // Otherwise, it is a narrow slice.
6085   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6086   if (is_empty_memory(n)) {
6087     // the array is sparse; empty slots are the "top" node
6088     n = base_memory();
6089     assert(Node::in_dump()
6090            || n == nullptr || n->bottom_type() == Type::TOP
6091            || n->adr_type() == nullptr // address is TOP
6092            || n->adr_type() == TypePtr::BOTTOM
6093            || n->adr_type() == TypeRawPtr::BOTTOM
6094            || !Compile::current()->do_aliasing(),
6095            "must be a wide memory");
6096     // do_aliasing == false if we are organizing the memory states manually.
6097     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6098   } else {
6099     // make sure the stored slice is sane
6100     #ifdef ASSERT
6101     if (VMError::is_error_reported() || Node::in_dump()) {
6102     } else if (might_be_same(n, base_memory())) {
6103       // Give it a pass:  It is a mostly harmless repetition of the base.
6104       // This can arise normally from node subsumption during optimization.
6105     } else {
6106       verify_memory_slice(this, alias_idx, n);
6107     }
6108     #endif
6109   }
6110   return n;
6111 }
6112 
6113 //---------------------------set_memory_at-------------------------------------
6114 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6115   verify_memory_slice(this, alias_idx, n);
6116   Node* empty_mem = empty_memory();
6117   if (n == base_memory())  n = empty_mem;  // collapse default
6118   uint need_req = alias_idx+1;
6119   if (req() < need_req) {
6120     if (n == empty_mem)  return;  // already the default, so do not grow me
6121     // grow the sparse array
6122     do {
6123       add_req(empty_mem);
6124     } while (req() < need_req);
6125   }
6126   set_req( alias_idx, n );
6127 }
6128 
6129 
6130 
6131 //--------------------------iteration_setup------------------------------------
6132 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6133   if (other != nullptr) {
6134     grow_to_match(other);
6135     // invariant:  the finite support of mm2 is within mm->req()
6136     #ifdef ASSERT
6137     for (uint i = req(); i < other->req(); i++) {
6138       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6139     }
6140     #endif
6141   }
6142   // Replace spurious copies of base_memory by top.
6143   Node* base_mem = base_memory();
6144   if (base_mem != nullptr && !base_mem->is_top()) {
6145     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6146       if (in(i) == base_mem)
6147         set_req(i, empty_memory());
6148     }
6149   }
6150 }
6151 
6152 //---------------------------grow_to_match-------------------------------------
6153 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6154   Node* empty_mem = empty_memory();
6155   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6156   // look for the finite support of the other memory
6157   for (uint i = other->req(); --i >= req(); ) {
6158     if (other->in(i) != empty_mem) {
6159       uint new_len = i+1;
6160       while (req() < new_len)  add_req(empty_mem);
6161       break;
6162     }
6163   }
6164 }
6165 
6166 //---------------------------verify_sparse-------------------------------------
6167 #ifndef PRODUCT
6168 bool MergeMemNode::verify_sparse() const {
6169   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6170   Node* base_mem = base_memory();
6171   // The following can happen in degenerate cases, since empty==top.
6172   if (is_empty_memory(base_mem))  return true;
6173   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6174     assert(in(i) != nullptr, "sane slice");
6175     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
6176   }
6177   return true;
6178 }
6179 
6180 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6181   Node* n;
6182   n = mm->in(idx);
6183   if (mem == n)  return true;  // might be empty_memory()
6184   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6185   if (mem == n)  return true;
6186   return false;
6187 }
6188 #endif // !PRODUCT