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