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