1 /*
   2  * Copyright (c) 2005, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "ci/ciFlatArrayKlass.hpp"
  26 #include "ci/ciInlineKlass.hpp"
  27 #include "ci/ciInstanceKlass.hpp"
  28 #include "compiler/compileLog.hpp"
  29 #include "gc/shared/collectedHeap.inline.hpp"
  30 #include "gc/shared/tlab_globals.hpp"
  31 #include "libadt/vectset.hpp"
  32 #include "memory/universe.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/arraycopynode.hpp"
  35 #include "opto/callnode.hpp"
  36 #include "opto/castnode.hpp"
  37 #include "opto/cfgnode.hpp"
  38 #include "opto/compile.hpp"
  39 #include "opto/convertnode.hpp"
  40 #include "opto/graphKit.hpp"
  41 #include "opto/inlinetypenode.hpp"
  42 #include "opto/intrinsicnode.hpp"
  43 #include "opto/locknode.hpp"
  44 #include "opto/loopnode.hpp"
  45 #include "opto/macro.hpp"
  46 #include "opto/memnode.hpp"
  47 #include "opto/narrowptrnode.hpp"
  48 #include "opto/node.hpp"
  49 #include "opto/opaquenode.hpp"
  50 #include "opto/opcodes.hpp"
  51 #include "opto/phaseX.hpp"
  52 #include "opto/reachability.hpp"
  53 #include "opto/rootnode.hpp"
  54 #include "opto/runtime.hpp"
  55 #include "opto/subnode.hpp"
  56 #include "opto/subtypenode.hpp"
  57 #include "opto/type.hpp"
  58 #include "prims/jvmtiExport.hpp"
  59 #include "runtime/continuation.hpp"
  60 #include "runtime/sharedRuntime.hpp"
  61 #include "runtime/stubRoutines.hpp"
  62 #include "utilities/globalDefinitions.hpp"
  63 #include "utilities/macros.hpp"
  64 #include "utilities/powerOfTwo.hpp"
  65 #if INCLUDE_G1GC
  66 #include "gc/g1/g1ThreadLocalData.hpp"
  67 #endif // INCLUDE_G1GC
  68 
  69 
  70 //
  71 // Replace any references to "oldref" in inputs to "use" with "newref".
  72 // Returns the number of replacements made.
  73 //
  74 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  75   int nreplacements = 0;
  76   uint req = use->req();
  77   for (uint j = 0; j < use->len(); j++) {
  78     Node *uin = use->in(j);
  79     if (uin == oldref) {
  80       if (j < req)
  81         use->set_req(j, newref);
  82       else
  83         use->set_prec(j, newref);
  84       nreplacements++;
  85     } else if (j >= req && uin == nullptr) {
  86       break;
  87     }
  88   }
  89   return nreplacements;
  90 }
  91 
  92 
  93 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
  94   Node* cmp = word;
  95   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
  96   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
  97   transform_later(iff);
  98 
  99   // Fast path taken.
 100   Node *fast_taken = transform_later(new IfFalseNode(iff));
 101 
 102   // Fast path not-taken, i.e. slow path
 103   Node *slow_taken = transform_later(new IfTrueNode(iff));
 104 
 105     region->init_req(edge, fast_taken); // Capture fast-control
 106     return slow_taken;
 107 }
 108 
 109 //--------------------copy_predefined_input_for_runtime_call--------------------
 110 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
 111   // Set fixed predefined input arguments
 112   call->init_req( TypeFunc::Control, ctrl );
 113   call->init_req( TypeFunc::I_O    , oldcall->in( TypeFunc::I_O) );
 114   call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
 115   call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
 116   call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
 117 }
 118 
 119 //------------------------------make_slow_call---------------------------------
 120 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
 121                                            address slow_call, const char* leaf_name, Node* slow_path,
 122                                            Node* parm0, Node* parm1, Node* parm2) {
 123 
 124   // Slow-path call
 125  CallNode *call = leaf_name
 126    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 127    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
 128 
 129   // Slow path call has no side-effects, uses few values
 130   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 131   if (parm0 != nullptr)  call->init_req(TypeFunc::Parms+0, parm0);
 132   if (parm1 != nullptr)  call->init_req(TypeFunc::Parms+1, parm1);
 133   if (parm2 != nullptr)  call->init_req(TypeFunc::Parms+2, parm2);
 134   call->copy_call_debug_info(&_igvn, oldcall);
 135   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 136   _igvn.replace_node(oldcall, call);
 137   transform_later(call);
 138 
 139   return call;
 140 }
 141 
 142 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
 143   BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 144   bs->eliminate_gc_barrier(&_igvn, p2x);
 145 #ifndef PRODUCT
 146   if (PrintOptoStatistics) {
 147     AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
 148   }
 149 #endif
 150 }
 151 
 152 // Search for a memory operation for the specified memory slice.
 153 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 154   Node *orig_mem = mem;
 155   Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 156   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 157   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 158   while (true) {
 159     if (mem == alloc_mem || mem == start_mem ) {
 160       return mem;  // hit one of our sentinels
 161     } else if (mem->is_MergeMem()) {
 162       mem = mem->as_MergeMem()->memory_at(alias_idx);
 163     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 164       Node *in = mem->in(0);
 165       // we can safely skip over safepoints, calls, locks and membars because we
 166       // already know that the object is safe to eliminate.
 167       if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
 168         return in;
 169       } else if (in->is_Call()) {
 170         CallNode *call = in->as_Call();
 171         if (call->may_modify(tinst, phase)) {
 172           assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
 173           if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
 174             return in;
 175           }
 176         }
 177         mem = in->in(TypeFunc::Memory);
 178       } else if (in->is_MemBar()) {
 179         ArrayCopyNode* ac = nullptr;
 180         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
 181           if (ac != nullptr) {
 182             assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 183             return ac;
 184           }
 185         }
 186         mem = in->in(TypeFunc::Memory);
 187       } else if (in->is_LoadFlat() || in->is_StoreFlat()) {
 188         mem = in->in(TypeFunc::Memory);
 189       } else {
 190 #ifdef ASSERT
 191         in->dump();
 192         mem->dump();
 193         assert(false, "unexpected projection");
 194 #endif
 195       }
 196     } else if (mem->is_Store()) {
 197       const TypePtr* atype = mem->as_Store()->adr_type();
 198       int adr_idx = phase->C->get_alias_index(atype);
 199       if (adr_idx == alias_idx) {
 200         assert(atype->isa_oopptr(), "address type must be oopptr");
 201         int adr_offset = atype->flat_offset();
 202         uint adr_iid = atype->is_oopptr()->instance_id();
 203         // Array elements references have the same alias_idx
 204         // but different offset and different instance_id.
 205         if (adr_offset == offset && adr_iid == alloc->_idx) {
 206           return mem;
 207         }
 208       } else {
 209         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 210       }
 211       mem = mem->in(MemNode::Memory);
 212     } else if (mem->is_ClearArray()) {
 213       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 214         // Can not bypass initialization of the instance
 215         // we are looking.
 216         DEBUG_ONLY(intptr_t offset;)
 217         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 218         InitializeNode* init = alloc->as_Allocate()->initialization();
 219         // We are looking for stored value, return Initialize node
 220         // or memory edge from Allocate node.
 221         if (init != nullptr) {
 222           return init;
 223         } else {
 224           return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
 225         }
 226       }
 227       // Otherwise skip it (the call updated 'mem' value).
 228     } else if (mem->Opcode() == Op_SCMemProj) {
 229       mem = mem->in(0);
 230       Node* adr = nullptr;
 231       if (mem->is_LoadStore()) {
 232         adr = mem->in(MemNode::Address);
 233       } else {
 234         assert(mem->Opcode() == Op_EncodeISOArray ||
 235                mem->Opcode() == Op_StrCompressedCopy, "sanity");
 236         adr = mem->in(3); // Destination array
 237       }
 238       const TypePtr* atype = adr->bottom_type()->is_ptr();
 239       int adr_idx = phase->C->get_alias_index(atype);
 240       if (adr_idx == alias_idx) {
 241         DEBUG_ONLY(mem->dump();)
 242         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 243         return nullptr;
 244       }
 245       mem = mem->in(MemNode::Memory);
 246     } else if (mem->Opcode() == Op_StrInflatedCopy) {
 247       Node* adr = mem->in(3); // Destination array
 248       const TypePtr* atype = adr->bottom_type()->is_ptr();
 249       int adr_idx = phase->C->get_alias_index(atype);
 250       if (adr_idx == alias_idx) {
 251         DEBUG_ONLY(mem->dump();)
 252         assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
 253         return nullptr;
 254       }
 255       mem = mem->in(MemNode::Memory);
 256     } else {
 257       return mem;
 258     }
 259     assert(mem != orig_mem, "dead memory loop");
 260   }
 261 }
 262 
 263 // Determine if there is an interfering store between a rematerialization load and an arraycopy that is in the process
 264 // of being elided. Starting from the given rematerialization load this method starts a BFS traversal upwards through
 265 // the memory graph towards the provided ArrayCopyNode. For every node encountered on the traversal, check that it is
 266 // independent from the provided rematerialization. Returns false if every node on the traversal is independent and
 267 // true otherwise.
 268 bool has_interfering_store(const ArrayCopyNode* ac, LoadNode* load, PhaseGVN* phase) {
 269   assert(ac != nullptr && load != nullptr, "sanity");
 270   AccessAnalyzer acc(phase, load);
 271   ResourceMark rm;
 272   Unique_Node_List to_visit;
 273   to_visit.push(load->in(MemNode::Memory));
 274 
 275   for (uint worklist_idx = 0; worklist_idx < to_visit.size(); worklist_idx++) {
 276     Node* mem = to_visit.at(worklist_idx);
 277 
 278     if (mem->is_Proj() && mem->in(0) == ac) {
 279       // Reached the target, so visit what is left on the worklist.
 280       continue;
 281     }
 282 
 283     if (mem->is_Phi()) {
 284       assert(mem->bottom_type() == Type::MEMORY, "do not leave memory graph");
 285       // Add all non-control inputs of phis to be visited.
 286       for (uint phi_in = 1; phi_in < mem->len(); phi_in++) {
 287         Node* input = mem->in(phi_in);
 288         if (input != nullptr) {
 289           to_visit.push(input);
 290         }
 291       }
 292       continue;
 293     }
 294 
 295     AccessAnalyzer::AccessIndependence ind = acc.detect_access_independence(mem);
 296     if (ind.independent) {
 297       to_visit.push(ind.mem);
 298     } else {
 299       return true;
 300     }
 301   }
 302   // Did not find modification of source element in memory graph.
 303   return false;
 304 }
 305 
 306 // Generate loads from source of the arraycopy for fields of destination needed at a deoptimization point.
 307 // Returns nullptr if the load cannot be created because the arraycopy is not suitable for elimination
 308 // (e.g. copy inside the array with non-constant offsets) or the inputs do not match our assumptions (e.g.
 309 // the arraycopy does not actually write something at the provided offset).
 310 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type* ftype, AllocateNode* alloc) {
 311   assert((ctl == ac->control() && mem == ac->memory()) != (mem != ac->memory() && ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj()),
 312     "Either the control and memory are the same as for the arraycopy or they are pinned in an uncommon trap.");
 313   BasicType bt = ft;
 314   const Type *type = ftype;
 315   if (ft == T_NARROWOOP) {
 316     bt = T_OBJECT;
 317     type = ftype->make_oopptr();
 318   }
 319   Node* base = ac->in(ArrayCopyNode::Src);
 320   Node* adr = nullptr;
 321   const TypePtr* adr_type = nullptr;
 322 
 323   if (ac->is_clonebasic()) {
 324     assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
 325     adr = _igvn.transform(AddPNode::make_with_base(base, _igvn.MakeConX(offset)));
 326     adr_type = _igvn.type(base)->is_ptr();
 327     if (adr_type->isa_aryptr()) {
 328       adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset);
 329     } else {
 330       adr_type = adr_type->add_offset(offset);
 331     }
 332   } else {
 333     if (!ac->modifies(offset, offset, &_igvn, true)) {
 334       // If the arraycopy does not copy to this offset, we cannot generate a rematerialization load for it.
 335       return nullptr;
 336     }
 337     assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 338     uint shift = exact_log2(type2aelembytes(bt));
 339     Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
 340     Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
 341     const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
 342     const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
 343 
 344     adr_type = _igvn.type(base)->is_aryptr();
 345     if (((const TypeAryPtr*)adr_type)->is_flat()) {
 346       shift = ((const TypeAryPtr*)adr_type)->flat_log_elem_size();
 347     }
 348     if (src_pos_t->is_con() && dest_pos_t->is_con()) {
 349       intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
 350       adr = _igvn.transform(AddPNode::make_with_base(base, base, _igvn.MakeConX(off)));
 351       adr_type = _igvn.type(adr)->is_aryptr();
 352       assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
 353       if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 354         // Don't emit a new load from src if src == dst but try to get the value from memory instead
 355         return value_from_mem(ac, ctl, ft, ftype, (const TypeAryPtr*)adr_type, alloc);
 356       }
 357     } else {
 358       if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 359         // Non constant offset in the array: we can't statically
 360         // determine the value
 361         return nullptr;
 362       }
 363       Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 364 #ifdef _LP64
 365       diff = _igvn.transform(new ConvI2LNode(diff));
 366 #endif
 367       diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
 368 
 369       Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
 370       adr = _igvn.transform(AddPNode::make_with_base(base, base, off));
 371       // In the case of a flat inline type array, each field has its
 372       // own slice so we need to extract the field being accessed from
 373       // the address computation
 374       adr_type = ((TypeAryPtr*)adr_type)->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
 375       adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
 376     }
 377   }
 378   assert(adr != nullptr && adr_type != nullptr, "sanity");
 379 
 380   // Create the rematerialization load ...
 381   MergeMemNode* mergemem = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
 382   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 383   Node* res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemem, adr, adr_type, type, bt);
 384   assert(res != nullptr, "load should have been created");
 385 
 386   // ... and ensure that pinning the rematerialization load inside the uncommon path is safe.
 387   if (mem != ac->memory() && ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj() && res->is_Load() &&
 388       has_interfering_store(ac, res->as_Load(), &_igvn)) {
 389     // Not safe: use control and memory from the arraycopy to ensure correct memory state.
 390     _igvn.remove_dead_node(res, PhaseIterGVN::NodeOrigin::Graph); // Clean up the unusable rematerialization load.
 391     return make_arraycopy_load(ac, offset, ac->control(), ac->memory(), ft, ftype, alloc);
 392   }
 393 
 394   if (ftype->isa_narrowoop()) {
 395     // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
 396     res = _igvn.transform(new EncodePNode(res, ftype));
 397   }
 398   return res;
 399 }
 400 
 401 //
 402 // Given a Memory Phi, compute a value Phi containing the values from stores
 403 // on the input paths.
 404 // Note: this function is recursive, its depth is limited by the "level" argument
 405 // Returns the computed Phi, or null if it cannot compute it.
 406 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
 407   assert(mem->is_Phi(), "sanity");
 408   int alias_idx = C->get_alias_index(adr_t);
 409   int offset = adr_t->flat_offset();
 410   int instance_id = adr_t->instance_id();
 411 
 412   // Check if an appropriate value phi already exists.
 413   Node* region = mem->in(0);
 414   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 415     Node* phi = region->fast_out(k);
 416     if (phi->is_Phi() && phi != mem &&
 417         phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
 418       return phi;
 419     }
 420   }
 421   // Check if an appropriate new value phi already exists.
 422   Node* new_phi = value_phis->find(mem->_idx);
 423   if (new_phi != nullptr)
 424     return new_phi;
 425 
 426   if (level <= 0) {
 427     return nullptr; // Give up: phi tree too deep
 428   }
 429   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 430   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 431   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 432 
 433   uint length = mem->req();
 434   GrowableArray <Node *> values(length, length, nullptr);
 435 
 436   // create a new Phi for the value
 437   PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
 438   transform_later(phi);
 439   value_phis->push(phi, mem->_idx);
 440 
 441   for (uint j = 1; j < length; j++) {
 442     Node *in = mem->in(j);
 443     if (in == nullptr || in->is_top()) {
 444       values.at_put(j, in);
 445     } else {
 446       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 447       if (val == start_mem || val == alloc_mem) {
 448         // hit a sentinel, return appropriate value
 449         Node* init_value = value_from_alloc(ft, adr_t, alloc);
 450         if (init_value == nullptr) {
 451           return nullptr;
 452         } else {
 453           values.at_put(j, init_value);
 454           continue;
 455         }
 456       }
 457       if (val->is_Initialize()) {
 458         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 459       }
 460       if (val == nullptr) {
 461         return nullptr;  // can't find a value on this path
 462       }
 463       if (val == mem) {
 464         values.at_put(j, mem);
 465       } else if (val->is_Store()) {
 466         Node* n = val->in(MemNode::ValueIn);
 467         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 468         n = bs->step_over_gc_barrier(n);
 469         if (is_subword_type(ft)) {
 470           n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
 471         }
 472         values.at_put(j, n);
 473       } else if (val->is_Proj() && val->in(0) == alloc) {
 474         Node* init_value = value_from_alloc(ft, adr_t, alloc);
 475         if (init_value == nullptr) {
 476           return nullptr;
 477         } else {
 478           values.at_put(j, init_value);
 479         }
 480       } else if (val->is_Phi()) {
 481         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 482         if (val == nullptr) {
 483           return nullptr;
 484         }
 485         values.at_put(j, val);
 486       } else if (val->Opcode() == Op_SCMemProj) {
 487         assert(val->in(0)->is_LoadStore() ||
 488                val->in(0)->Opcode() == Op_EncodeISOArray ||
 489                val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
 490         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 491         return nullptr;
 492       } else if (val->is_ArrayCopy()) {
 493         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
 494         if (res == nullptr) {
 495           return nullptr;
 496         }
 497         values.at_put(j, res);
 498       } else if (val->is_top()) {
 499         // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
 500         // IGVN will clean this up later.
 501         values.at_put(j, val);
 502       } else {
 503         DEBUG_ONLY( val->dump(); )
 504         assert(false, "unknown node on this path");
 505         return nullptr;  // unknown node on this path
 506       }
 507     }
 508   }
 509   // Set Phi's inputs
 510   for (uint j = 1; j < length; j++) {
 511     if (values.at(j) == mem) {
 512       phi->init_req(j, phi);
 513     } else {
 514       phi->init_req(j, values.at(j));
 515     }
 516   }
 517   return phi;
 518 }
 519 
 520 // Extract the initial value of a field in an allocation
 521 Node* PhaseMacroExpand::value_from_alloc(BasicType ft, const TypeOopPtr* adr_t, AllocateNode* alloc) {
 522   Node* init_value = alloc->in(AllocateNode::InitValue);
 523   if (init_value == nullptr) {
 524     assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "conflicting InitValue and RawInitValue");
 525     return _igvn.zerocon(ft);
 526   }
 527 
 528   const TypeAryPtr* ary_t = adr_t->isa_aryptr();
 529   assert(ary_t != nullptr, "must be a pointer into an array");
 530 
 531   // If this is not a flat array, then it must be an oop array with elements being init_value
 532   if (ary_t->is_not_flat()) {
 533 #ifdef ASSERT
 534     BasicType init_bt = init_value->bottom_type()->basic_type();
 535     assert(ft == init_bt ||
 536            (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
 537            (is_subword_type(ft) && init_bt == T_INT),
 538            "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
 539 #endif // ASSERT
 540     return init_value;
 541   }
 542 
 543   assert(ary_t->klass_is_exact() && ary_t->is_flat(), "must be an exact flat array");
 544   assert(ary_t->field_offset().get() != Type::OffsetBot, "unknown offset");
 545   if (init_value->is_EncodeP()) {
 546     init_value = init_value->in(1);
 547   }
 548   // Cannot look through init_value if it is an oop
 549   if (!init_value->is_InlineType()) {
 550     return nullptr;
 551   }
 552 
 553   ciInlineKlass* vk = init_value->bottom_type()->inline_klass();
 554   if (ary_t->field_offset().get() == vk->null_marker_offset_in_payload()) {
 555     init_value = init_value->as_InlineType()->get_null_marker();
 556   } else {
 557     init_value = init_value->as_InlineType()->field_value_by_offset(ary_t->field_offset().get() + vk->payload_offset(), true);
 558   }
 559 
 560   if (ft == T_NARROWOOP) {
 561     assert(init_value->bottom_type()->isa_ptr(), "must be a pointer");
 562     init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_narrowoop()));
 563   }
 564 
 565 #ifdef ASSERT
 566   BasicType init_bt = init_value->bottom_type()->basic_type();
 567   assert(ft == init_bt ||
 568          (!is_java_primitive(ft) && !is_java_primitive(init_bt) && type2aelembytes(ft, true) == type2aelembytes(init_bt, true)) ||
 569          (is_subword_type(ft) && init_bt == T_INT),
 570          "invalid init_value of type %s for field of type %s", type2name(init_bt), type2name(ft));
 571 #endif // ASSERT
 572 
 573   return init_value;
 574 }
 575 
 576 // Search the last value stored into the object's field.
 577 Node* PhaseMacroExpand::value_from_mem(Node* origin, Node* ctl, BasicType ft, const Type* ftype, const TypeOopPtr* adr_t, AllocateNode* alloc) {
 578   assert(adr_t->is_known_instance_field(), "instance required");
 579   int instance_id = adr_t->instance_id();
 580   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 581 
 582   int alias_idx = C->get_alias_index(adr_t);
 583   int offset = adr_t->flat_offset();
 584   Node* orig_mem = origin->in(TypeFunc::Memory);
 585   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 586   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 587   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 588   VectorSet visited;
 589 
 590   bool done = orig_mem == alloc_mem;
 591   Node *mem = orig_mem;
 592   while (!done) {
 593     if (visited.test_set(mem->_idx)) {
 594       return nullptr;  // found a loop, give up
 595     }
 596     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 597     if (mem == start_mem || mem == alloc_mem) {
 598       done = true;  // hit a sentinel, return appropriate 0 value
 599     } else if (mem->is_Initialize()) {
 600       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 601       if (mem == nullptr) {
 602         done = true; // Something went wrong.
 603       } else if (mem->is_Store()) {
 604         const TypePtr* atype = mem->as_Store()->adr_type();
 605         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 606         done = true;
 607       }
 608     } else if (mem->is_Store()) {
 609       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 610       assert(atype != nullptr, "address type must be oopptr");
 611       assert(C->get_alias_index(atype) == alias_idx &&
 612              atype->is_known_instance_field() && atype->flat_offset() == offset &&
 613              atype->instance_id() == instance_id, "store is correct memory slice");
 614       done = true;
 615     } else if (mem->is_Phi()) {
 616       // try to find a phi's unique input
 617       Node *unique_input = nullptr;
 618       Node *top = C->top();
 619       for (uint i = 1; i < mem->req(); i++) {
 620         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 621         if (n == nullptr || n == top || n == mem) {
 622           continue;
 623         } else if (unique_input == nullptr) {
 624           unique_input = n;
 625         } else if (unique_input != n) {
 626           unique_input = top;
 627           break;
 628         }
 629       }
 630       if (unique_input != nullptr && unique_input != top) {
 631         mem = unique_input;
 632       } else {
 633         done = true;
 634       }
 635     } else if (mem->is_ArrayCopy()) {
 636       done = true;
 637     } else if (mem->is_top()) {
 638       // The slice is on a dead path. Returning nullptr would lead to elimination
 639       // bailout, but we want to prevent that. Just forwarding the top is also legal,
 640       // and IGVN can just clean things up, and remove whatever receives top.
 641       return mem;
 642     } else {
 643       DEBUG_ONLY( mem->dump(); )
 644       assert(false, "unexpected node");
 645     }
 646   }
 647   if (mem != nullptr) {
 648     if (mem == start_mem || mem == alloc_mem) {
 649       // hit a sentinel, return appropriate value
 650       return value_from_alloc(ft, adr_t, alloc);
 651     } else if (mem->is_Store()) {
 652       Node* n = mem->in(MemNode::ValueIn);
 653       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 654       n = bs->step_over_gc_barrier(n);
 655       return n;
 656     } else if (mem->is_Phi()) {
 657       // attempt to produce a Phi reflecting the values on the input paths of the Phi
 658       Node_Stack value_phis(8);
 659       Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 660       if (phi != nullptr) {
 661         return phi;
 662       } else {
 663         // Kill all new Phis
 664         while(value_phis.is_nonempty()) {
 665           Node* n = value_phis.node();
 666           _igvn.replace_node(n, C->top());
 667           value_phis.pop();
 668         }
 669       }
 670     } else if (mem->is_ArrayCopy()) {
 671       // Rematerialize the scalar-replaced array. If possible, pin the loads to the uncommon path of the uncommon trap.
 672       // Check for each element of the source array, whether it was modified. If not, pin both memory and control to
 673       // the uncommon path. Otherwise, use the control and memory state of the arraycopy. Control and memory state must
 674       // come from the same source to prevent anti-dependence problems in the backend.
 675       ArrayCopyNode* ac = mem->as_ArrayCopy();
 676       Node* ac_ctl = ac->control();
 677       Node* ac_mem = ac->memory();
 678       if (ctl->is_Proj() && ctl->as_Proj()->is_uncommon_trap_proj()) {
 679         // pin the loads in the uncommon trap path
 680         ac_ctl = ctl;
 681         ac_mem = orig_mem;
 682       }
 683       return make_arraycopy_load(ac, offset, ac_ctl, ac_mem, ft, ftype, alloc);
 684     }
 685   }
 686   // Something went wrong.
 687   return nullptr;
 688 }
 689 
 690 // Search the last value stored into the inline type's fields (for flat arrays).
 691 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
 692   auto report_failure = [&](int field_offset_in_element, bool is_forced_failure) {
 693 #ifndef PRODUCT
 694     if (PrintEliminateAllocations) {
 695       ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
 696       int offset = field_offset_in_element + elem_klass->payload_offset();
 697       ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
 698       assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
 699 
 700       tty->print("=== At SafePoint node %d ", sfpt->_idx);
 701       if (is_forced_failure) {
 702         tty->print_raw("forcibly abort elimination");
 703       } else {
 704         tty->print("can't find value of field [%s] of array element [%d]", flattened_field->name()->as_utf8(), elem_idx);
 705       }
 706       tty->print(", which prevents elimination of: ");
 707       alloc->dump();
 708     }
 709 #endif // PRODUCT
 710   };
 711 
 712   // Create a new InlineTypeNode and retrieve the field values from memory
 713   InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, null_free);
 714   transform_later(vt);
 715   if (null_free) {
 716     vt->set_null_marker(_igvn);
 717   } else {
 718     int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
 719     const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
 720     Node* nm_value = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
 721     bool force_scalarization_failure = StressEliminateAllocations &&
 722                                        (C->random() % StressEliminateAllocationsMean == 0);
 723     if (nm_value != nullptr && !force_scalarization_failure) {
 724       vt->set_null_marker(_igvn, nm_value);
 725     } else {
 726       report_failure(nm_offset_in_element, nm_value != nullptr);
 727       return nullptr;
 728     }
 729   }
 730 
 731   for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
 732     ciField* field = vt->field(i);
 733     ciType* field_type = field->type();
 734     int field_offset_in_element = offset_in_element + field->offset_in_bytes() - vk->payload_offset();
 735     Node* field_value = nullptr;
 736     assert(!field->is_flat() || field->type()->is_inlinetype(), "must be an inline type");
 737     if (field->is_flat()) {
 738       field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, field->is_null_free(), alloc, sfpt);
 739     } else {
 740       const Type* ft = Type::get_const_type(field_type);
 741       BasicType bt = type2field[field_type->basic_type()];
 742       if (UseCompressedOops && !is_java_primitive(bt)) {
 743         ft = ft->make_narrowoop();
 744         bt = T_NARROWOOP;
 745       }
 746       // Each inline type field has its own memory slice
 747       const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
 748       field_value = value_from_mem(sfpt, sfpt->control(), bt, ft, field_adr_type, alloc);
 749       bool force_scalarization_failure = StressEliminateAllocations &&
 750                                          (C->random() % StressEliminateAllocationsMean == 0);
 751       if (field_value == nullptr || force_scalarization_failure) {
 752         report_failure(field_offset_in_element, field_value != nullptr);
 753         return nullptr;
 754       } else if (ft->isa_narrowoop()) {
 755         assert(UseCompressedOops, "unexpected narrow oop");
 756         if (field_value->is_EncodeP()) {
 757           field_value = field_value->in(1);
 758         } else if (!field_value->is_InlineType()) {
 759           field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
 760         }
 761       }
 762     }
 763     assert(field_value != nullptr, "");
 764     vt->set_field_value(i, field_value);
 765   }
 766   return vt;
 767 }
 768 
 769 // Check the possibility of scalar replacement.
 770 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode* alloc, Unique_Node_List* safepoints) {
 771   //  Scan the uses of the allocation to check for anything that would
 772   //  prevent us from eliminating it.
 773   NOT_PRODUCT( const char* fail_eliminate = nullptr; )
 774   DEBUG_ONLY( Node* disq_node = nullptr; )
 775   bool can_eliminate = true;
 776   bool reduce_merge_precheck = (safepoints == nullptr);
 777 
 778   Unique_Node_List worklist;
 779   Node* res = alloc->result_cast();
 780   const TypeOopPtr* res_type = nullptr;
 781   if (res == nullptr) {
 782     // All users were eliminated.
 783   } else if (!res->is_CheckCastPP()) {
 784     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 785     can_eliminate = false;
 786   } else {
 787     worklist.push(res);
 788     res_type = igvn->type(res)->isa_oopptr();
 789     if (res_type == nullptr) {
 790       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 791       can_eliminate = false;
 792     } else if (!res_type->klass_is_exact()) {
 793       NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
 794       can_eliminate = false;
 795     } else if (res_type->isa_aryptr()) {
 796       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 797       if (length < 0) {
 798         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 799         can_eliminate = false;
 800       }
 801     }
 802   }
 803 
 804   while (can_eliminate && worklist.size() > 0) {
 805     BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 806     res = worklist.pop();
 807     for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
 808       Node* use = res->fast_out(j);
 809 
 810       if (use->is_AddP()) {
 811         const TypePtr* addp_type = igvn->type(use)->is_ptr();
 812         int offset = addp_type->offset();
 813 
 814         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 815           NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
 816           can_eliminate = false;
 817           break;
 818         }
 819         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 820                                    k < kmax && can_eliminate; k++) {
 821           Node* n = use->fast_out(k);
 822           if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
 823             DEBUG_ONLY(disq_node = n);
 824             NOT_PRODUCT(fail_eliminate = "Mismatched access");
 825             can_eliminate = false;
 826           }
 827           if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
 828             DEBUG_ONLY(disq_node = n;)
 829             if (n->is_Load() || n->is_LoadStore()) {
 830               NOT_PRODUCT(fail_eliminate = "Field load";)
 831             } else {
 832               NOT_PRODUCT(fail_eliminate = "Not store field reference";)
 833             }
 834             can_eliminate = false;
 835           }
 836         }
 837       } else if (use->is_ArrayCopy() &&
 838                  (use->as_ArrayCopy()->is_clonebasic() ||
 839                   use->as_ArrayCopy()->is_arraycopy_validated() ||
 840                   use->as_ArrayCopy()->is_copyof_validated() ||
 841                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 842                  use->in(ArrayCopyNode::Dest) == res) {
 843         // ok to eliminate
 844       } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
 845         // ok to eliminate
 846       } else if (use->is_SafePoint()) {
 847         SafePointNode* sfpt = use->as_SafePoint();
 848         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 849           // Object is passed as argument.
 850           DEBUG_ONLY(disq_node = use;)
 851           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 852           can_eliminate = false;
 853         }
 854         Node* sfptMem = sfpt->memory();
 855         if (sfptMem == nullptr || sfptMem->is_top()) {
 856           DEBUG_ONLY(disq_node = use;)
 857           NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
 858           can_eliminate = false;
 859         } else if (!reduce_merge_precheck) {
 860           assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
 861           safepoints->push(sfpt);
 862         }
 863       } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
 864         // Look at uses
 865         for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
 866           Node* u = use->fast_out(k);
 867           if (u->is_InlineType()) {
 868             // Use in flat field can be eliminated
 869             InlineTypeNode* vt = u->as_InlineType();
 870             for (uint i = 0; i < vt->field_count(); ++i) {
 871               if (vt->field_value(i) == use && !vt->field(i)->is_flat()) {
 872                 can_eliminate = false; // Use in non-flat field
 873                 break;
 874               }
 875             }
 876           } else {
 877             // Add other uses to the worklist to process individually
 878             worklist.push(use);
 879           }
 880         }
 881       } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
 882         // Store to mark word of inline type larval buffer
 883         assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
 884       } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
 885         // Inline type buffer allocations are followed by a membar
 886       } else if (reduce_merge_precheck &&
 887                  (use->is_Phi() || use->is_EncodeP() ||
 888                   use->Opcode() == Op_MemBarRelease ||
 889                   (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
 890         // Nothing to do
 891       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 892         if (use->is_Phi()) {
 893           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 894             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 895           } else {
 896             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 897           }
 898           DEBUG_ONLY(disq_node = use;)
 899         } else {
 900           if (use->Opcode() == Op_Return) {
 901             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 902           } else {
 903             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 904           }
 905           DEBUG_ONLY(disq_node = use;)
 906         }
 907         can_eliminate = false;
 908       } else {
 909         assert(use->Opcode() == Op_CastP2X, "should be");
 910         assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
 911       }
 912     }
 913   }
 914 
 915 #ifndef PRODUCT
 916   if (PrintEliminateAllocations && safepoints != nullptr) {
 917     if (can_eliminate) {
 918       tty->print("Scalar ");
 919       if (res == nullptr)
 920         alloc->dump();
 921       else
 922         res->dump();
 923     } else {
 924       tty->print("NotScalar (%s)", fail_eliminate);
 925       if (res == nullptr)
 926         alloc->dump();
 927       else
 928         res->dump();
 929 #ifdef ASSERT
 930       if (disq_node != nullptr) {
 931           tty->print("  >>>> ");
 932           disq_node->dump();
 933       }
 934 #endif /*ASSERT*/
 935     }
 936   }
 937 
 938   if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
 939     tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
 940     DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
 941   }
 942 #endif
 943   return can_eliminate;
 944 }
 945 
 946 void PhaseMacroExpand::undo_previous_scalarizations(Unique_Node_List& safepoints_done, AllocateNode* alloc) {
 947   Node* res = alloc->result_cast();
 948   int nfields = 0;
 949   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
 950 
 951   if (res != nullptr) {
 952     const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
 953 
 954     if (res_type->isa_instptr()) {
 955       // find the fields of the class which will be needed for safepoint debug information
 956       ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
 957       nfields = iklass->nof_nonstatic_fields();
 958     } else {
 959       // find the array's elements which will be needed for safepoint debug information
 960       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 961       assert(nfields >= 0, "must be an array klass.");
 962     }
 963   }
 964 
 965   // rollback processed safepoints
 966   while (safepoints_done.size() > 0) {
 967     SafePointNode* sfpt_done = safepoints_done.pop()->as_SafePoint();
 968 
 969     SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
 970 
 971     sfpt_done->remove_non_debug_edges(non_debug_edges_worklist);
 972 
 973     // remove any extra entries we added to the safepoint
 974     assert(sfpt_done->jvms()->endoff() == sfpt_done->req(), "no extra edges past debug info allowed");
 975     uint last = sfpt_done->req() - 1;
 976     for (int k = 0;  k < nfields; k++) {
 977       sfpt_done->del_req(last--);
 978     }
 979     JVMState *jvms = sfpt_done->jvms();
 980     jvms->set_endoff(sfpt_done->req());
 981     // Now make a pass over the debug information replacing any references
 982     // to SafePointScalarObjectNode with the allocated object.
 983     int start = jvms->debug_start();
 984     int end   = jvms->debug_end();
 985     for (int i = start; i < end; i++) {
 986       if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 987         SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 988         if (scobj->first_index(jvms) == sfpt_done->req() &&
 989             scobj->n_fields() == (uint)nfields) {
 990           assert(scobj->alloc() == alloc, "sanity");
 991           sfpt_done->set_req(i, res);
 992         }
 993       }
 994     }
 995 
 996     sfpt_done->restore_non_debug_edges(non_debug_edges_worklist);
 997 
 998     _igvn._worklist.push(sfpt_done);
 999   }
1000 }
1001 
1002 #ifdef ASSERT
1003   // Verify if a value can be written into a field.
1004   void verify_type_compatability(const Type* value_type, const Type* field_type) {
1005     BasicType value_bt = value_type->basic_type();
1006     BasicType field_bt = field_type->basic_type();
1007 
1008     // Primitive types must match.
1009     if (is_java_primitive(value_bt) && value_bt == field_bt) { return; }
1010 
1011     // I have been struggling to make a similar assert for non-primitive
1012     // types. I we can add one in the future. For now, I just let them
1013     // pass without checks.
1014     // In particular, I was struggling with a value that came from a call,
1015     // and had only a non-null check CastPP. There was also a checkcast
1016     // in the graph to verify the interface, but the corresponding
1017     // CheckCastPP result was not updated in the stack slot, and so
1018     // we ended up using the CastPP. That means that the field knows
1019     // that it should get an oop from an interface, but the value lost
1020     // that information, and so it is not a subtype.
1021     // There may be other issues, feel free to investigate further!
1022     if (!is_java_primitive(value_bt)) { return; }
1023 
1024     tty->print_cr("value not compatible for field: %s vs %s",
1025                   type2name(value_bt),
1026                   type2name(field_bt));
1027     tty->print("value_type: ");
1028     value_type->dump();
1029     tty->cr();
1030     tty->print("field_type: ");
1031     field_type->dump();
1032     tty->cr();
1033     assert(false, "value_type does not fit field_type");
1034   }
1035 #endif
1036 
1037 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1038   if (UseCompressedOops && field_type->isa_narrowoop()) {
1039     // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
1040     // to be able scalar replace the allocation.
1041     if (field_val->is_EncodeP()) {
1042       field_val = field_val->in(1);
1043     } else if (!field_val->is_InlineType()) {
1044       field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1045     }
1046   }
1047 
1048   // Keep track of inline types to scalarize them later
1049   if (field_val->is_InlineType()) {
1050     value_worklist->push(field_val);
1051   } else if (field_val->is_Phi()) {
1052     PhiNode* phi = field_val->as_Phi();
1053     // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1054     // scalarize all its fields in safepoints.
1055     field_val = phi->try_push_inline_types_down(&_igvn, true);
1056     if (field_val->is_InlineType()) {
1057       value_worklist->push(field_val);
1058     }
1059   }
1060   DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
1061   sfpt->add_req(field_val);
1062 }
1063 
1064 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1065   const Type* elem_type = array_type->elem();
1066   BasicType basic_elem_type = elem_type->array_element_basic_type();
1067 
1068   intptr_t elem_size;
1069   uint header_size;
1070   if (array_type->is_flat()) {
1071     elem_size = array_type->flat_elem_size();
1072     header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1073   } else {
1074     elem_size = type2aelembytes(basic_elem_type);
1075     header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1076   }
1077 
1078   int n_elems = alloc->in(AllocateNode::ALength)->get_int();
1079   for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
1080     intptr_t elem_offset = header_size + elem_idx * elem_size;
1081     const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
1082     Node* elem_val;
1083     if (array_type->is_flat()) {
1084       ciInlineKlass* elem_klass = elem_type->inline_klass();
1085       assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
1086       elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
1087     } else {
1088       elem_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
1089     }
1090     bool force_scalarization_failure = StressEliminateAllocations &&
1091                                        (C->random() % StressEliminateAllocationsMean == 0);
1092     if (elem_val == nullptr || force_scalarization_failure) {
1093 #ifndef PRODUCT
1094       if (PrintEliminateAllocations) {
1095         tty->print("=== At SafePoint node %d ", sfpt->_idx);
1096         if (elem_val == nullptr) {
1097           tty->print("can't find value of array element [%d]", elem_idx);
1098         } else {
1099           assert(force_scalarization_failure, "sanity");
1100           tty->print_raw("forcibly abort elimination");
1101         }
1102         tty->print(", which prevents elimination of: ");
1103         alloc->dump();
1104       }
1105 #endif // PRODUCT
1106       return false;
1107     }
1108 
1109     process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
1110   }
1111 
1112   return true;
1113 }
1114 
1115 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
1116 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
1117 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
1118 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
1119   const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
1120   auto report_failure = [&](int offset, bool is_forced_failure) {
1121 #ifndef PRODUCT
1122     if (PrintEliminateAllocations) {
1123       ciInstanceKlass* base_klass = base_type->instance_klass();
1124       ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
1125       assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
1126       tty->print("=== At SafePoint node %d ", sfpt->_idx);
1127       if (is_forced_failure) {
1128         tty->print_raw("forcibly abort elimination");
1129       } else {
1130         tty->print_raw("can't find value of field: ");
1131         flattened_field->print();
1132         int field_idx = C->alias_type(flattened_field)->index();
1133         tty->print(" (alias_idx=%d)", field_idx);
1134       }
1135       tty->print(", which prevents elimination of: ");
1136       base->dump();
1137     }
1138 #endif // PRODUCT
1139   };
1140 
1141   for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
1142     ciField* field = iklass->declared_nonstatic_field_at(i);
1143     if (field->is_flat()) {
1144       ciInlineKlass* fvk = field->type()->as_inline_klass();
1145       int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
1146       bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
1147       if (!success) {
1148         return false;
1149       }
1150 
1151       // The null marker of a field is added right after we scalarize that field
1152       if (!field->is_null_free()) {
1153         int nm_offset = offset_minus_header + field->null_marker_offset();
1154         Node* null_marker = value_from_mem(sfpt, sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
1155         bool force_scalarization_failure = StressEliminateAllocations &&
1156                                            (C->random() % StressEliminateAllocationsMean == 0);
1157         if (null_marker == nullptr || force_scalarization_failure) {
1158           report_failure(nm_offset, null_marker != nullptr);
1159           return false;
1160         }
1161         process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
1162       }
1163 
1164       continue;
1165     }
1166 
1167     int offset = offset_minus_header + field->offset_in_bytes();
1168     ciType* elem_type = field->type();
1169     BasicType basic_elem_type = field->layout_type();
1170 
1171     const Type* field_type;
1172     if (is_reference_type(basic_elem_type)) {
1173       if (!elem_type->is_loaded()) {
1174         field_type = TypeInstPtr::BOTTOM;
1175       } else {
1176         field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1177       }
1178       if (UseCompressedOops) {
1179         field_type = field_type->make_narrowoop();
1180         basic_elem_type = T_NARROWOOP;
1181       }
1182     } else {
1183       field_type = Type::get_const_basic_type(basic_elem_type);
1184     }
1185 
1186     const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1187     Node* field_val = value_from_mem(sfpt, sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1188     bool force_scalarization_failure = StressEliminateAllocations &&
1189                                        (C->random() % StressEliminateAllocationsMean == 0);
1190     if (field_val == nullptr || force_scalarization_failure) {
1191       report_failure(offset, field_val != nullptr);
1192       return false;
1193     }
1194     process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1195   }
1196 
1197   return true;
1198 }
1199 
1200 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1201                                                                                   Unique_Node_List* value_worklist) {
1202   assert(sfpt->jvms()->endoff() == sfpt->req(), "no extra edges past debug info allowed");
1203 
1204   // Fields of scalar objs are referenced only at the end
1205   // of regular debuginfo at the last (youngest) JVMS.
1206   // Record relative start index.
1207   ciInstanceKlass* iklass    = nullptr;
1208   const TypeOopPtr* res_type = nullptr;
1209   int nfields                = 0;
1210   uint first_ind             = (sfpt->req() - sfpt->jvms()->scloff());
1211   Node* res                  = alloc->result_cast();
1212 
1213   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1214   assert(sfpt->jvms() != nullptr, "missed JVMS");
1215   uint before_sfpt_req = sfpt->req();
1216 
1217   if (res != nullptr) { // Could be null when there are no users
1218     res_type = _igvn.type(res)->isa_oopptr();
1219 
1220     if (res_type->isa_instptr()) {
1221       // find the fields of the class which will be needed for safepoint debug information
1222       iklass = res_type->is_instptr()->instance_klass();
1223       nfields = iklass->nof_nonstatic_fields();
1224     } else {
1225       // find the array's elements which will be needed for safepoint debug information
1226       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1227       assert(nfields >= 0, "must be an array klass.");
1228     }
1229 
1230     if (res->bottom_type()->is_inlinetypeptr()) {
1231       // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1232       // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1233       // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1234       // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1235       // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1236       // later in PhaseOutput::filLocArray() for inline types.
1237       sfpt->add_req(_igvn.intcon(1));
1238     }
1239   }
1240 
1241   SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1242   sobj->init_req(0, C->root());
1243   transform_later(sobj);
1244 
1245   if (res == nullptr) {
1246     sfpt->jvms()->set_endoff(sfpt->req());
1247     return sobj;
1248   }
1249 
1250   bool success;
1251   if (iklass == nullptr) {
1252     success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1253   } else {
1254     success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1255   }
1256 
1257   // We weren't able to find a value for this field, remove all the fields added to the safepoint
1258   if (!success) {
1259     for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1260       sfpt->del_req(i);
1261     }
1262     _igvn._worklist.push(sfpt);
1263     return nullptr;
1264   }
1265 
1266   sfpt->jvms()->set_endoff(sfpt->req());
1267   return sobj;
1268 }
1269 
1270 // Do scalar replacement.
1271 bool PhaseMacroExpand::scalar_replacement(AllocateNode* alloc, Unique_Node_List& safepoints) {
1272   Unique_Node_List safepoints_done;
1273   Node* res = alloc->result_cast();
1274   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1275   const TypeOopPtr* res_type = nullptr;
1276   if (res != nullptr) { // Could be null when there are no users
1277     res_type = _igvn.type(res)->isa_oopptr();
1278   }
1279 
1280   // Process the safepoint uses
1281   Unique_Node_List value_worklist;
1282   while (safepoints.size() > 0) {
1283     SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
1284 
1285     SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(igvn());
1286 
1287     // All sfpt inputs are implicitly included into debug info during the scalarization process below.
1288     // Keep non-debug inputs separately, so they stay non-debug.
1289     sfpt->remove_non_debug_edges(non_debug_edges_worklist);
1290 
1291     SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1292 
1293     if (sobj == nullptr) {
1294       sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1295       undo_previous_scalarizations(safepoints_done, alloc);
1296       return false;
1297     }
1298 
1299     // Now make a pass over the debug information replacing any references
1300     // to the allocated object with "sobj"
1301     JVMState *jvms = sfpt->jvms();
1302     sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1303     non_debug_edges_worklist.remove_edge_if_present(res); // drop scalarized input from non-debug info
1304     sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1305     _igvn._worklist.push(sfpt);
1306 
1307     // keep it for rollback
1308     safepoints_done.push(sfpt);
1309   }
1310   // Scalarize inline types that were added to the safepoint.
1311   // Don't allow linking a constant oop (if available) for flat array elements
1312   // because Deoptimization::reassign_flat_array_elements needs field values.
1313   bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1314   for (uint i = 0; i < value_worklist.size(); ++i) {
1315     InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1316     vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1317   }
1318   return true;
1319 }
1320 
1321 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1322   Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1323   Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1324   if (ctl_proj != nullptr) {
1325     igvn.replace_node(ctl_proj, n->in(0));
1326   }
1327   if (mem_proj != nullptr) {
1328     igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1329   }
1330 }
1331 
1332 // Process users of eliminated allocation.
1333 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1334   Unique_Node_List worklist;
1335   Node* res = alloc->result_cast();
1336   if (res != nullptr) {
1337     worklist.push(res);
1338   }
1339   while (worklist.size() > 0) {
1340     res = worklist.pop();
1341     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1342       Node *use = res->last_out(j);
1343       uint oc1 = res->outcnt();
1344 
1345       if (use->is_AddP()) {
1346         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1347           Node *n = use->last_out(k);
1348           uint oc2 = use->outcnt();
1349           if (n->is_Store()) {
1350             for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1351               MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1352               if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1353                 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1354                 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1355                 mb->remove(&_igvn);
1356               }
1357             }
1358             _igvn.replace_node(n, n->in(MemNode::Memory));
1359           } else {
1360             eliminate_gc_barrier(n);
1361           }
1362           k -= (oc2 - use->outcnt());
1363         }
1364         _igvn.remove_dead_node(use, PhaseIterGVN::NodeOrigin::Graph);
1365       } else if (use->is_ArrayCopy()) {
1366         // Disconnect ArrayCopy node
1367         ArrayCopyNode* ac = use->as_ArrayCopy();
1368         if (ac->is_clonebasic()) {
1369           Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1370           disconnect_projections(ac, _igvn);
1371           assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1372           Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1373           disconnect_projections(membar_before->as_MemBar(), _igvn);
1374           if (membar_after->is_MemBar()) {
1375             disconnect_projections(membar_after->as_MemBar(), _igvn);
1376           }
1377         } else {
1378           assert(ac->is_arraycopy_validated() ||
1379                  ac->is_copyof_validated() ||
1380                  ac->is_copyofrange_validated(), "unsupported");
1381           CallProjections* callprojs = ac->extract_projections(true);
1382 
1383           _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1384           _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1385           _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1386 
1387           // Set control to top. IGVN will remove the remaining projections
1388           ac->set_req(0, top());
1389           ac->replace_edge(res, top(), &_igvn);
1390 
1391           // Disconnect src right away: it can help find new
1392           // opportunities for allocation elimination
1393           Node* src = ac->in(ArrayCopyNode::Src);
1394           ac->replace_edge(src, top(), &_igvn);
1395           // src can be top at this point if src and dest of the
1396           // arraycopy were the same
1397           if (src->outcnt() == 0 && !src->is_top()) {
1398             _igvn.remove_dead_node(src, PhaseIterGVN::NodeOrigin::Graph);
1399           }
1400         }
1401         _igvn._worklist.push(ac);
1402       } else if (use->is_InlineType()) {
1403         assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1404         // Cut off oop input and remove known instance id from type
1405         _igvn.rehash_node_delayed(use);
1406         use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1407         use->as_InlineType()->set_is_buffered(_igvn, false);
1408         const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1409         _igvn.set_type(use, toop);
1410         use->as_InlineType()->set_type(toop);
1411         // Process users
1412         for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1413           Node* u = use->fast_out(k);
1414           if (!u->is_InlineType() && !u->is_StoreFlat()) {
1415             worklist.push(u);
1416           }
1417         }
1418       } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1419         // Store to mark word of inline type larval buffer
1420         assert(inline_alloc, "Unexpected store to mark word");
1421         _igvn.replace_node(use, use->in(MemNode::Memory));
1422       } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1423         // Inline type buffer allocations are followed by a membar
1424         assert(inline_alloc, "Unexpected MemBarRelease");
1425         use->as_MemBar()->remove(&_igvn);
1426       } else if (use->is_ReachabilityFence() && OptimizeReachabilityFences) {
1427         use->as_ReachabilityFence()->clear_referent(_igvn); // redundant fence; will be removed during IGVN
1428       } else {
1429         eliminate_gc_barrier(use);
1430       }
1431       j -= (oc1 - res->outcnt());
1432     }
1433     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1434     _igvn.remove_dead_node(res, PhaseIterGVN::NodeOrigin::Graph);
1435   }
1436 
1437   //
1438   // Process other users of allocation's projections
1439   //
1440   if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1441     // First disconnect stores captured by Initialize node.
1442     // If Initialize node is eliminated first in the following code,
1443     // it will kill such stores and DUIterator_Last will assert.
1444     for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax);  j < jmax; j++) {
1445       Node* use = _callprojs->resproj[0]->fast_out(j);
1446       if (use->is_AddP()) {
1447         // raw memory addresses used only by the initialization
1448         _igvn.replace_node(use, C->top());
1449         --j; --jmax;
1450       }
1451     }
1452     for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1453       Node* use = _callprojs->resproj[0]->last_out(j);
1454       uint oc1 = _callprojs->resproj[0]->outcnt();
1455       if (use->is_Initialize()) {
1456         // Eliminate Initialize node.
1457         InitializeNode *init = use->as_Initialize();
1458         Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1459         if (ctrl_proj != nullptr) {
1460           _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1461 #ifdef ASSERT
1462           // If the InitializeNode has no memory out, it will die, and tmp will become null
1463           Node* tmp = init->in(TypeFunc::Control);
1464           assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1465 #endif
1466         }
1467         Node* mem = init->in(TypeFunc::Memory);
1468 #ifdef ASSERT
1469         if (init->number_of_projs(TypeFunc::Memory) > 0) {
1470           if (mem->is_MergeMem()) {
1471             assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs->fallthrough_memproj, "allocation memory projection");
1472           } else {
1473             assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1474           }
1475         }
1476 #endif
1477         init->replace_mem_projs_by(mem, &_igvn);
1478         assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1479       } else if (use->Opcode() == Op_MemBarStoreStore) {
1480         // Inline type buffer allocations are followed by a membar
1481         assert(inline_alloc, "Unexpected MemBarStoreStore");
1482         use->as_MemBar()->remove(&_igvn);
1483       } else  {
1484         assert(false, "only Initialize or AddP expected");
1485       }
1486       j -= (oc1 - _callprojs->resproj[0]->outcnt());
1487     }
1488   }
1489   if (_callprojs->fallthrough_catchproj != nullptr) {
1490     _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1491   }
1492   if (_callprojs->fallthrough_memproj != nullptr) {
1493     _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1494   }
1495   if (_callprojs->catchall_memproj != nullptr) {
1496     _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1497   }
1498   if (_callprojs->fallthrough_ioproj != nullptr) {
1499     _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1500   }
1501   if (_callprojs->catchall_ioproj != nullptr) {
1502     _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1503   }
1504   if (_callprojs->catchall_catchproj != nullptr) {
1505     _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1506   }
1507 }
1508 
1509 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1510   // If reallocation fails during deoptimization we'll pop all
1511   // interpreter frames for this compiled frame and that won't play
1512   // nice with JVMTI popframe.
1513   // We avoid this issue by eager reallocation when the popframe request
1514   // is received.
1515   if (!EliminateAllocations) {
1516     return false;
1517   }
1518   Node* klass = alloc->in(AllocateNode::KlassNode);
1519   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1520 
1521   // Attempt to eliminate inline type buffer allocations
1522   // regardless of usage and escape/replaceable status.
1523   bool inline_alloc = tklass->isa_instklassptr() &&
1524                       tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1525   if (!alloc->_is_non_escaping && !inline_alloc) {
1526     return false;
1527   }
1528   // Eliminate boxing allocations which are not used
1529   // regardless scalar replaceable status.
1530   Node* res = alloc->result_cast();
1531   bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1532                       tklass->isa_instklassptr() &&
1533                       tklass->is_instklassptr()->instance_klass()->is_box_klass();
1534   if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1535     return false;
1536   }
1537 
1538   _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1539 
1540   Unique_Node_List safepoints;
1541   if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1542     return false;
1543   }
1544 
1545   if (!alloc->_is_scalar_replaceable) {
1546     assert(res == nullptr || inline_alloc, "sanity");
1547     // We can only eliminate allocation if all debug info references
1548     // are already replaced with SafePointScalarObject because
1549     // we can't search for a fields value without instance_id.
1550     if (safepoints.size() > 0) {
1551       return false;
1552     }
1553   }
1554 
1555   if (!scalar_replacement(alloc, safepoints)) {
1556     return false;
1557   }
1558 
1559   CompileLog* log = C->log();
1560   if (log != nullptr) {
1561     log->head("eliminate_allocation type='%d'",
1562               log->identify(tklass->exact_klass()));
1563     JVMState* p = alloc->jvms();
1564     while (p != nullptr) {
1565       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1566       p = p->caller();
1567     }
1568     log->tail("eliminate_allocation");
1569   }
1570 
1571   process_users_of_allocation(alloc, inline_alloc);
1572 
1573 #ifndef PRODUCT
1574   if (PrintEliminateAllocations) {
1575     if (alloc->is_AllocateArray())
1576       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1577     else
1578       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1579   }
1580 #endif
1581 
1582   return true;
1583 }
1584 
1585 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode* call) {
1586   // EA should remove all uses of non-escaping boxing node.
1587   if (!C->eliminate_boxing()) {
1588     return false;
1589   }
1590   for (uint i = TypeFunc::Parms; i < call->tf()->range_cc()->cnt(); ++i) {
1591     if (call->proj_out_or_null(i) != nullptr) {
1592       return false;
1593     }
1594   }
1595 
1596   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1597 
1598   process_users_of_allocation(call);
1599 
1600   CompileLog* log = C->log();
1601   if (log != nullptr) {
1602     const TypeInstPtr* t = nullptr;
1603     if (call->is_boxing_method()) {
1604       const TypeTuple* range = call->tf()->range_sig();
1605       t = range->field_at(TypeFunc::Parms)->isa_instptr();
1606     } else {
1607       assert(call->is_unboxing_method(), "Unexpected call");
1608       const TypeTuple* domain = call->tf()->domain_sig();
1609       t = domain->field_at(TypeFunc::Parms)->isa_instptr();
1610       assert(!t->maybe_null(), "missing receiver null check?");
1611     }
1612     log->head("eliminate_boxing type='%d'",
1613               log->identify(t->instance_klass()));
1614     JVMState* p = call->jvms();
1615     while (p != nullptr) {
1616       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1617       p = p->caller();
1618     }
1619     log->tail("eliminate_boxing");
1620   }
1621 
1622 #ifndef PRODUCT
1623   if (PrintEliminateAllocations) {
1624     tty->print("++++ Eliminated: %d ", call->_idx);
1625     call->method()->print_short_name(tty);
1626     tty->cr();
1627   }
1628 #endif
1629 
1630   return true;
1631 }
1632 
1633 Node* PhaseMacroExpand::make_load_raw(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1634   Node* adr = off_heap_plus_addr(base, offset);
1635   const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1636   Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1637   transform_later(value);
1638   return value;
1639 }
1640 
1641 
1642 Node* PhaseMacroExpand::make_store_raw(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1643   Node* adr = off_heap_plus_addr(base, offset);
1644   mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1645   transform_later(mem);
1646   return mem;
1647 }
1648 
1649 //=============================================================================
1650 //
1651 //                              A L L O C A T I O N
1652 //
1653 // Allocation attempts to be fast in the case of frequent small objects.
1654 // It breaks down like this:
1655 //
1656 // 1) Size in doublewords is computed.  This is a constant for objects and
1657 // variable for most arrays.  Doubleword units are used to avoid size
1658 // overflow of huge doubleword arrays.  We need doublewords in the end for
1659 // rounding.
1660 //
1661 // 2) Size is checked for being 'too large'.  Too-large allocations will go
1662 // the slow path into the VM.  The slow path can throw any required
1663 // exceptions, and does all the special checks for very large arrays.  The
1664 // size test can constant-fold away for objects.  For objects with
1665 // finalizers it constant-folds the otherway: you always go slow with
1666 // finalizers.
1667 //
1668 // 3) If NOT using TLABs, this is the contended loop-back point.
1669 // Load-Locked the heap top.  If using TLABs normal-load the heap top.
1670 //
1671 // 4) Check that heap top + size*8 < max.  If we fail go the slow ` route.
1672 // NOTE: "top+size*8" cannot wrap the 4Gig line!  Here's why: for largish
1673 // "size*8" we always enter the VM, where "largish" is a constant picked small
1674 // enough that there's always space between the eden max and 4Gig (old space is
1675 // there so it's quite large) and large enough that the cost of entering the VM
1676 // is dwarfed by the cost to initialize the space.
1677 //
1678 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1679 // down.  If contended, repeat at step 3.  If using TLABs normal-store
1680 // adjusted heap top back down; there is no contention.
1681 //
1682 // 6) If !ZeroTLAB then Bulk-clear the object/array.  Fill in klass & mark
1683 // fields.
1684 //
1685 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1686 // oop flavor.
1687 //
1688 //=============================================================================
1689 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1690 // Allocations bigger than this always go the slow route.
1691 // This value must be small enough that allocation attempts that need to
1692 // trigger exceptions go the slow route.  Also, it must be small enough so
1693 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1694 //=============================================================================j//
1695 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1696 // The allocator will coalesce int->oop copies away.  See comment in
1697 // coalesce.cpp about how this works.  It depends critically on the exact
1698 // code shape produced here, so if you are changing this code shape
1699 // make sure the GC info for the heap-top is correct in and around the
1700 // slow-path call.
1701 //
1702 
1703 void PhaseMacroExpand::expand_allocate_common(
1704             AllocateNode* alloc, // allocation node to be expanded
1705             Node* length,  // array length for an array allocation
1706             Node* init_val, // value to initialize the array with
1707             const TypeFunc* slow_call_type, // Type of slow call
1708             address slow_call_address,  // Address of slow call
1709             Node* valid_length_test // whether length is valid or not
1710     )
1711 {
1712   Node* ctrl = alloc->in(TypeFunc::Control);
1713   Node* mem  = alloc->in(TypeFunc::Memory);
1714   Node* i_o  = alloc->in(TypeFunc::I_O);
1715   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1716   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1717   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1718   assert(ctrl != nullptr, "must have control");
1719 
1720   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1721   // they will not be used if "always_slow" is set
1722   enum { slow_result_path = 1, fast_result_path = 2 };
1723   Node *result_region = nullptr;
1724   Node *result_phi_rawmem = nullptr;
1725   Node *result_phi_rawoop = nullptr;
1726   Node *result_phi_i_o = nullptr;
1727 
1728   // The initial slow comparison is a size check, the comparison
1729   // we want to do is a BoolTest::gt
1730   bool expand_fast_path = true;
1731   int tv = _igvn.find_int_con(initial_slow_test, -1);
1732   if (tv >= 0) {
1733     // InitialTest has constant result
1734     //   0 - can fit in TLAB
1735     //   1 - always too big or negative
1736     assert(tv <= 1, "0 or 1 if a constant");
1737     expand_fast_path = (tv == 0);
1738     initial_slow_test = nullptr;
1739   } else {
1740     initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1741   }
1742 
1743   if (!UseTLAB) {
1744     // Force slow-path allocation
1745     expand_fast_path = false;
1746     initial_slow_test = nullptr;
1747   }
1748 
1749   // ArrayCopyNode right after an allocation operates on the raw result projection for the Allocate node so it's not
1750   // safe to remove such an allocation even if it has no result cast.
1751   bool allocation_has_use = (alloc->result_cast() != nullptr) || (alloc->initialization() != nullptr && alloc->initialization()->is_complete_with_arraycopy());
1752   if (!allocation_has_use) {
1753     InitializeNode* init = alloc->initialization();
1754     if (init != nullptr) {
1755       init->remove(&_igvn);
1756     }
1757     if (expand_fast_path && (initial_slow_test == nullptr)) {
1758       // Remove allocation node and return.
1759       // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1760       // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1761 #ifndef PRODUCT
1762       if (PrintEliminateAllocations) {
1763         tty->print("NotUsed ");
1764         Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1765         if (res != nullptr) {
1766           res->dump();
1767         } else {
1768           alloc->dump();
1769         }
1770       }
1771 #endif
1772       yank_alloc_node(alloc);
1773       return;
1774     }
1775   }
1776 
1777   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1778   Node *slow_region = nullptr;
1779   Node *toobig_false = ctrl;
1780 
1781   // generate the initial test if necessary
1782   if (initial_slow_test != nullptr ) {
1783     assert (expand_fast_path, "Only need test if there is a fast path");
1784     slow_region = new RegionNode(3);
1785 
1786     // Now make the initial failure test.  Usually a too-big test but
1787     // might be a TRUE for finalizers.
1788     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1789     transform_later(toobig_iff);
1790     // Plug the failing-too-big test into the slow-path region
1791     Node* toobig_true = new IfTrueNode(toobig_iff);
1792     transform_later(toobig_true);
1793     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1794     toobig_false = new IfFalseNode(toobig_iff);
1795     transform_later(toobig_false);
1796   } else {
1797     // No initial test, just fall into next case
1798     assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1799     toobig_false = ctrl;
1800     DEBUG_ONLY(slow_region = NodeSentinel);
1801   }
1802 
1803   // If we are here there are several possibilities
1804   // - expand_fast_path is false - then only a slow path is expanded. That's it.
1805   // no_initial_check means a constant allocation.
1806   // - If check always evaluates to false -> expand_fast_path is false (see above)
1807   // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1808   // if !allocation_has_use the fast path is empty
1809   // if !allocation_has_use && no_initial_check
1810   // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1811   //   removed by yank_alloc_node above.
1812 
1813   Node *slow_mem = mem;  // save the current memory state for slow path
1814   // generate the fast allocation code unless we know that the initial test will always go slow
1815   if (expand_fast_path) {
1816     // Fast path modifies only raw memory.
1817     if (mem->is_MergeMem()) {
1818       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1819     }
1820 
1821     // allocate the Region and Phi nodes for the result
1822     result_region = new RegionNode(3);
1823     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1824     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1825 
1826     // Grab regular I/O before optional prefetch may change it.
1827     // Slow-path does no I/O so just set it to the original I/O.
1828     result_phi_i_o->init_req(slow_result_path, i_o);
1829 
1830     // Name successful fast-path variables
1831     Node* fast_oop_ctrl;
1832     Node* fast_oop_rawmem;
1833 
1834     if (allocation_has_use) {
1835       Node* needgc_ctrl = nullptr;
1836       result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1837 
1838       intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1839       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1840       Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1841                                         fast_oop_ctrl, fast_oop_rawmem,
1842                                         prefetch_lines);
1843 
1844       if (initial_slow_test != nullptr) {
1845         // This completes all paths into the slow merge point
1846         slow_region->init_req(need_gc_path, needgc_ctrl);
1847         transform_later(slow_region);
1848       } else {
1849         // No initial slow path needed!
1850         // Just fall from the need-GC path straight into the VM call.
1851         slow_region = needgc_ctrl;
1852       }
1853 
1854       InitializeNode* init = alloc->initialization();
1855       fast_oop_rawmem = initialize_object(alloc,
1856                                           fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1857                                           klass_node, length, size_in_bytes);
1858       expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1859       expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1860 
1861       result_phi_rawoop->init_req(fast_result_path, fast_oop);
1862     } else {
1863       assert (initial_slow_test != nullptr, "sanity");
1864       fast_oop_ctrl   = toobig_false;
1865       fast_oop_rawmem = mem;
1866       transform_later(slow_region);
1867     }
1868 
1869     // Plug in the successful fast-path into the result merge point
1870     result_region    ->init_req(fast_result_path, fast_oop_ctrl);
1871     result_phi_i_o   ->init_req(fast_result_path, i_o);
1872     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1873   } else {
1874     slow_region = ctrl;
1875     result_phi_i_o = i_o; // Rename it to use in the following code.
1876   }
1877 
1878   // Generate slow-path call
1879   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1880                                OptoRuntime::stub_name(slow_call_address),
1881                                TypePtr::BOTTOM);
1882   call->init_req(TypeFunc::Control,   slow_region);
1883   call->init_req(TypeFunc::I_O,       top());    // does no i/o
1884   call->init_req(TypeFunc::Memory,    slow_mem); // may gc ptrs
1885   call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1886   call->init_req(TypeFunc::FramePtr,  alloc->in(TypeFunc::FramePtr));
1887 
1888   call->init_req(TypeFunc::Parms+0, klass_node);
1889   if (length != nullptr) {
1890     call->init_req(TypeFunc::Parms+1, length);
1891     if (init_val != nullptr) {
1892       call->init_req(TypeFunc::Parms+2, init_val);
1893     }
1894   }
1895 
1896   // Copy debug information and adjust JVMState information, then replace
1897   // allocate node with the call
1898   call->copy_call_debug_info(&_igvn, alloc);
1899   // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1900   // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1901   // path dies).
1902   if (valid_length_test != nullptr) {
1903     call->add_req(valid_length_test);
1904   }
1905   if (expand_fast_path) {
1906     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1907   } else {
1908     // Hook i_o projection to avoid its elimination during allocation
1909     // replacement (when only a slow call is generated).
1910     call->set_req(TypeFunc::I_O, result_phi_i_o);
1911   }
1912   _igvn.replace_node(alloc, call);
1913   transform_later(call);
1914 
1915   // Identify the output projections from the allocate node and
1916   // adjust any references to them.
1917   // The control and io projections look like:
1918   //
1919   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1920   //  Allocate                   Catch
1921   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1922   //
1923   //  We are interested in the CatchProj nodes.
1924   //
1925   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1926 
1927   // An allocate node has separate memory projections for the uses on
1928   // the control and i_o paths. Replace the control memory projection with
1929   // result_phi_rawmem (unless we are only generating a slow call when
1930   // both memory projections are combined)
1931   if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1932     _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1933   }
1934   // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1935   // catchall_memproj so we end up with a call that has only 1 memory projection.
1936   if (_callprojs->catchall_memproj != nullptr) {
1937     if (_callprojs->fallthrough_memproj == nullptr) {
1938       _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1939       transform_later(_callprojs->fallthrough_memproj);
1940     }
1941     _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1942     _igvn.remove_dead_node(_callprojs->catchall_memproj, PhaseIterGVN::NodeOrigin::Graph);
1943   }
1944 
1945   // An allocate node has separate i_o projections for the uses on the control
1946   // and i_o paths. Always replace the control i_o projection with result i_o
1947   // otherwise incoming i_o become dead when only a slow call is generated
1948   // (it is different from memory projections where both projections are
1949   // combined in such case).
1950   if (_callprojs->fallthrough_ioproj != nullptr) {
1951     _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1952   }
1953   // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1954   // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1955   if (_callprojs->catchall_ioproj != nullptr) {
1956     if (_callprojs->fallthrough_ioproj == nullptr) {
1957       _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1958       transform_later(_callprojs->fallthrough_ioproj);
1959     }
1960     _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1961     _igvn.remove_dead_node(_callprojs->catchall_ioproj, PhaseIterGVN::NodeOrigin::Graph);
1962   }
1963 
1964   // if we generated only a slow call, we are done
1965   if (!expand_fast_path) {
1966     // Now we can unhook i_o.
1967     if (result_phi_i_o->outcnt() > 1) {
1968       call->set_req(TypeFunc::I_O, top());
1969     } else {
1970       assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1971       // Case of new array with negative size known during compilation.
1972       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1973       // following code since call to runtime will throw exception.
1974       // As result there will be no users of i_o after the call.
1975       // Leave i_o attached to this call to avoid problems in preceding graph.
1976     }
1977     return;
1978   }
1979 
1980   if (_callprojs->fallthrough_catchproj != nullptr) {
1981     ctrl = _callprojs->fallthrough_catchproj->clone();
1982     transform_later(ctrl);
1983     _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1984   } else {
1985     ctrl = top();
1986   }
1987   Node *slow_result;
1988   if (_callprojs->resproj[0] == nullptr) {
1989     // no uses of the allocation result
1990     slow_result = top();
1991   } else {
1992     slow_result = _callprojs->resproj[0]->clone();
1993     transform_later(slow_result);
1994     _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1995   }
1996 
1997   // Plug slow-path into result merge point
1998   result_region->init_req( slow_result_path, ctrl);
1999   transform_later(result_region);
2000   if (allocation_has_use) {
2001     result_phi_rawoop->init_req(slow_result_path, slow_result);
2002     transform_later(result_phi_rawoop);
2003   }
2004   result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
2005   transform_later(result_phi_rawmem);
2006   transform_later(result_phi_i_o);
2007   // This completes all paths into the result merge point
2008 }
2009 
2010 // Remove alloc node that has no uses.
2011 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
2012   Node* ctrl = alloc->in(TypeFunc::Control);
2013   Node* mem  = alloc->in(TypeFunc::Memory);
2014   Node* i_o  = alloc->in(TypeFunc::I_O);
2015 
2016   _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2017   if (_callprojs->resproj[0] != nullptr) {
2018     for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
2019       Node* use = _callprojs->resproj[0]->fast_out(i);
2020       use->isa_MemBar()->remove(&_igvn);
2021       --imax;
2022       --i; // back up iterator
2023     }
2024     assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
2025     _igvn.remove_dead_node(_callprojs->resproj[0], PhaseIterGVN::NodeOrigin::Graph);
2026   }
2027   if (_callprojs->fallthrough_catchproj != nullptr) {
2028     _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
2029     _igvn.remove_dead_node(_callprojs->fallthrough_catchproj, PhaseIterGVN::NodeOrigin::Graph);
2030   }
2031   if (_callprojs->catchall_catchproj != nullptr) {
2032     _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
2033     _callprojs->catchall_catchproj->set_req(0, top());
2034   }
2035   if (_callprojs->fallthrough_proj != nullptr) {
2036     Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
2037     _igvn.remove_dead_node(catchnode, PhaseIterGVN::NodeOrigin::Graph);
2038     _igvn.remove_dead_node(_callprojs->fallthrough_proj, PhaseIterGVN::NodeOrigin::Graph);
2039   }
2040   if (_callprojs->fallthrough_memproj != nullptr) {
2041     _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
2042     _igvn.remove_dead_node(_callprojs->fallthrough_memproj, PhaseIterGVN::NodeOrigin::Graph);
2043   }
2044   if (_callprojs->fallthrough_ioproj != nullptr) {
2045     _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
2046     _igvn.remove_dead_node(_callprojs->fallthrough_ioproj, PhaseIterGVN::NodeOrigin::Graph);
2047   }
2048   if (_callprojs->catchall_memproj != nullptr) {
2049     _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
2050     _callprojs->catchall_memproj->set_req(0, top());
2051   }
2052   if (_callprojs->catchall_ioproj != nullptr) {
2053     _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
2054     _callprojs->catchall_ioproj->set_req(0, top());
2055   }
2056 #ifndef PRODUCT
2057   if (PrintEliminateAllocations) {
2058     if (alloc->is_AllocateArray()) {
2059       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
2060     } else {
2061       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
2062     }
2063   }
2064 #endif
2065   _igvn.remove_dead_node(alloc, PhaseIterGVN::NodeOrigin::Graph);
2066 }
2067 
2068 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
2069                                                 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
2070   // If initialization is performed by an array copy, any required
2071   // MemBarStoreStore was already added. If the object does not
2072   // escape no need for a MemBarStoreStore. If the object does not
2073   // escape in its initializer and memory barrier (MemBarStoreStore or
2074   // stronger) is already added at exit of initializer, also no need
2075   // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
2076   // so that stores that initialize this object can't be reordered
2077   // with a subsequent store that makes this object accessible by
2078   // other threads.
2079   // Other threads include java threads and JVM internal threads
2080   // (for example concurrent GC threads). Current concurrent GC
2081   // implementation: G1 will not scan newly created object,
2082   // so it's safe to skip storestore barrier when allocation does
2083   // not escape.
2084   if (!alloc->does_not_escape_thread() &&
2085     !alloc->is_allocation_MemBar_redundant() &&
2086     (init == nullptr || !init->is_complete_with_arraycopy())) {
2087     if (init == nullptr || init->req() < InitializeNode::RawStores) {
2088       // No InitializeNode or no stores captured by zeroing
2089       // elimination. Simply add the MemBarStoreStore after object
2090       // initialization.
2091       // What we want is to prevent the compiler and the CPU from re-ordering the stores that initialize this object
2092       // with subsequent stores to any slice. As a consequence, this MemBar should capture the entire memory state at
2093       // this point in the IR and produce a new memory state that should cover all slices. However, the Initialize node
2094       // only captures/produces a partial memory state making it complicated to insert such a MemBar. Because
2095       // re-ordering by the compiler can't happen by construction (a later Store that publishes the just allocated
2096       // object reference is indirectly control dependent on the Initialize node), preventing reordering by the CPU is
2097       // sufficient. For that a MemBar on the raw memory slice is good enough.
2098       // If init is null, this allocation does have an InitializeNode but this logic can't locate it (see comment in
2099       // PhaseMacroExpand::initialize_object()).
2100       MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2101       transform_later(mb);
2102 
2103       mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
2104       mb->init_req(TypeFunc::Control, fast_oop_ctrl);
2105       fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
2106       transform_later(fast_oop_ctrl);
2107       fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
2108       transform_later(fast_oop_rawmem);
2109     } else {
2110       // Add the MemBarStoreStore after the InitializeNode so that
2111       // all stores performing the initialization that were moved
2112       // before the InitializeNode happen before the storestore
2113       // barrier.
2114 
2115       Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
2116 
2117       // See comment above that explains why a raw memory MemBar is good enough.
2118       MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
2119       transform_later(mb);
2120 
2121       Node* ctrl = new ProjNode(init, TypeFunc::Control);
2122       transform_later(ctrl);
2123       Node* old_raw_mem_proj = nullptr;
2124       auto find_raw_mem = [&](ProjNode* proj) {
2125         if (C->get_alias_index(proj->adr_type()) == Compile::AliasIdxRaw) {
2126           assert(old_raw_mem_proj == nullptr, "only one expected");
2127           old_raw_mem_proj = proj;
2128         }
2129       };
2130       init->for_each_proj(find_raw_mem, TypeFunc::Memory);
2131       assert(old_raw_mem_proj != nullptr, "should have found raw mem Proj");
2132       Node* raw_mem_proj = new ProjNode(init, TypeFunc::Memory);
2133       transform_later(raw_mem_proj);
2134 
2135       // The MemBarStoreStore depends on control and memory coming
2136       // from the InitializeNode
2137       mb->init_req(TypeFunc::Memory, raw_mem_proj);
2138       mb->init_req(TypeFunc::Control, ctrl);
2139 
2140       ctrl = new ProjNode(mb, TypeFunc::Control);
2141       transform_later(ctrl);
2142       Node* mem = new ProjNode(mb, TypeFunc::Memory);
2143       transform_later(mem);
2144 
2145       // All nodes that depended on the InitializeNode for control
2146       // and memory must now depend on the MemBarNode that itself
2147       // depends on the InitializeNode
2148       if (init_ctrl != nullptr) {
2149         _igvn.replace_node(init_ctrl, ctrl);
2150       }
2151       _igvn.replace_node(old_raw_mem_proj, mem);
2152     }
2153   }
2154 }
2155 
2156 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
2157                                                 Node*& ctrl, Node*& rawmem) {
2158   if (C->env()->dtrace_alloc_probes()) {
2159     // Slow-path call
2160     int size = TypeFunc::Parms + 2;
2161     CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
2162                                           CAST_FROM_FN_PTR(address,
2163                                           static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
2164                                           "dtrace_object_alloc",
2165                                           TypeRawPtr::BOTTOM);
2166 
2167     // Get base of thread-local storage area
2168     Node* thread = new ThreadLocalNode();
2169     transform_later(thread);
2170 
2171     call->init_req(TypeFunc::Parms + 0, thread);
2172     call->init_req(TypeFunc::Parms + 1, oop);
2173     call->init_req(TypeFunc::Control, ctrl);
2174     call->init_req(TypeFunc::I_O    , top()); // does no i/o
2175     call->init_req(TypeFunc::Memory , rawmem);
2176     call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
2177     call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
2178     transform_later(call);
2179     ctrl = new ProjNode(call, TypeFunc::Control);
2180     transform_later(ctrl);
2181     rawmem = new ProjNode(call, TypeFunc::Memory);
2182     transform_later(rawmem);
2183   }
2184 }
2185 
2186 // Helper for PhaseMacroExpand::expand_allocate_common.
2187 // Initializes the newly-allocated storage.
2188 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
2189                                           Node* control, Node* rawmem, Node* object,
2190                                           Node* klass_node, Node* length,
2191                                           Node* size_in_bytes) {
2192   InitializeNode* init = alloc->initialization();
2193   // Store the klass & mark bits
2194   Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
2195   if (!mark_node->is_Con()) {
2196     transform_later(mark_node);
2197   }
2198   rawmem = make_store_raw(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
2199 
2200   if (!UseCompactObjectHeaders) {
2201     rawmem = make_store_raw(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2202   }
2203   int header_size = alloc->minimum_header_size();  // conservatively small
2204 
2205   // Array length
2206   if (length != nullptr) {         // Arrays need length field
2207     rawmem = make_store_raw(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
2208     // conservatively small header size:
2209     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2210     if (_igvn.type(klass_node)->isa_aryklassptr()) {   // we know the exact header size in most cases:
2211       BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_exact_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2212       if (is_reference_type(elem, true)) {
2213         elem = T_OBJECT;
2214       }
2215       header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2216     }
2217   }
2218 
2219   // Clear the object body, if necessary.
2220   if (init == nullptr) {
2221     // The init has somehow disappeared; be cautious and clear everything.
2222     //
2223     // This can happen if a node is allocated but an uncommon trap occurs
2224     // immediately.  In this case, the Initialize gets associated with the
2225     // trap, and may be placed in a different (outer) loop, if the Allocate
2226     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
2227     // there can be two Allocates to one Initialize.  The answer in all these
2228     // edge cases is safety first.  It is always safe to clear immediately
2229     // within an Allocate, and then (maybe or maybe not) clear some more later.
2230     if (!(UseTLAB && ZeroTLAB)) {
2231       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2232                                             alloc->in(AllocateNode::InitValue),
2233                                             alloc->in(AllocateNode::RawInitValue),
2234                                             header_size, size_in_bytes,
2235                                             true,
2236                                             &_igvn);
2237     }
2238   } else {
2239     if (!init->is_complete()) {
2240       // Try to win by zeroing only what the init does not store.
2241       // We can also try to do some peephole optimizations,
2242       // such as combining some adjacent subword stores.
2243       rawmem = init->complete_stores(control, rawmem, object,
2244                                      header_size, size_in_bytes, &_igvn);
2245     }
2246     // We have no more use for this link, since the AllocateNode goes away:
2247     init->set_req(InitializeNode::RawAddress, top());
2248     // (If we keep the link, it just confuses the register allocator,
2249     // who thinks he sees a real use of the address by the membar.)
2250   }
2251 
2252   return rawmem;
2253 }
2254 
2255 // Generate prefetch instructions for next allocations.
2256 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2257                                         Node*& contended_phi_rawmem,
2258                                         Node* old_eden_top, Node* new_eden_top,
2259                                         intx lines) {
2260    enum { fall_in_path = 1, pf_path = 2 };
2261    if (UseTLAB && AllocatePrefetchStyle == 2) {
2262       // Generate prefetch allocation with watermark check.
2263       // As an allocation hits the watermark, we will prefetch starting
2264       // at a "distance" away from watermark.
2265 
2266       Node* pf_region = new RegionNode(3);
2267       Node* pf_phi_rawmem = new PhiNode(pf_region, Type::MEMORY,
2268                                                 TypeRawPtr::BOTTOM);
2269       // I/O is used for Prefetch
2270       Node* pf_phi_abio = new PhiNode(pf_region, Type::ABIO);
2271 
2272       Node* thread = new ThreadLocalNode();
2273       transform_later(thread);
2274 
2275       Node* eden_pf_adr = AddPNode::make_off_heap(thread,
2276                    _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())));
2277       transform_later(eden_pf_adr);
2278 
2279       Node* old_pf_wm = new LoadPNode(needgc_false,
2280                                    contended_phi_rawmem, eden_pf_adr,
2281                                    TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2282                                    MemNode::unordered);
2283       transform_later(old_pf_wm);
2284 
2285       // check against new_eden_top
2286       Node* need_pf_cmp = new CmpPNode(new_eden_top, old_pf_wm);
2287       transform_later(need_pf_cmp);
2288       Node* need_pf_bol = new BoolNode(need_pf_cmp, BoolTest::ge);
2289       transform_later(need_pf_bol);
2290       IfNode* need_pf_iff = new IfNode(needgc_false, need_pf_bol,
2291                                        PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2292       transform_later(need_pf_iff);
2293 
2294       // true node, add prefetchdistance
2295       Node* need_pf_true = new IfTrueNode(need_pf_iff);
2296       transform_later(need_pf_true);
2297 
2298       Node* need_pf_false = new IfFalseNode(need_pf_iff);
2299       transform_later(need_pf_false);
2300 
2301       Node* new_pf_wmt = AddPNode::make_off_heap(old_pf_wm,
2302                                                  _igvn.MakeConX(AllocatePrefetchDistance));
2303       transform_later(new_pf_wmt);
2304       new_pf_wmt->set_req(0, need_pf_true);
2305 
2306       Node* store_new_wmt = new StorePNode(need_pf_true,
2307                                        contended_phi_rawmem, eden_pf_adr,
2308                                        TypeRawPtr::BOTTOM, new_pf_wmt,
2309                                        MemNode::unordered);
2310       transform_later(store_new_wmt);
2311 
2312       // adding prefetches
2313       pf_phi_abio->init_req(fall_in_path, i_o);
2314 
2315       Node* prefetch_adr;
2316       Node* prefetch;
2317       uint step_size = AllocatePrefetchStepSize;
2318       uint distance = 0;
2319 
2320       for (intx i = 0; i < lines; i++) {
2321         prefetch_adr = AddPNode::make_off_heap(new_pf_wmt,
2322                                                _igvn.MakeConX(distance));
2323         transform_later(prefetch_adr);
2324         prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2325         transform_later(prefetch);
2326         distance += step_size;
2327         i_o = prefetch;
2328       }
2329       pf_phi_abio->set_req(pf_path, i_o);
2330 
2331       pf_region->init_req(fall_in_path, need_pf_false);
2332       pf_region->init_req(pf_path, need_pf_true);
2333 
2334       pf_phi_rawmem->init_req(fall_in_path, contended_phi_rawmem);
2335       pf_phi_rawmem->init_req(pf_path, store_new_wmt);
2336 
2337       transform_later(pf_region);
2338       transform_later(pf_phi_rawmem);
2339       transform_later(pf_phi_abio);
2340 
2341       needgc_false = pf_region;
2342       contended_phi_rawmem = pf_phi_rawmem;
2343       i_o = pf_phi_abio;
2344    } else if (UseTLAB && AllocatePrefetchStyle == 3) {
2345       // Insert a prefetch instruction for each allocation.
2346       // This code is used to generate 1 prefetch instruction per cache line.
2347 
2348       // Generate several prefetch instructions.
2349       uint step_size = AllocatePrefetchStepSize;
2350       uint distance = AllocatePrefetchDistance;
2351 
2352       // Next cache address.
2353       Node* cache_adr = AddPNode::make_off_heap(old_eden_top,
2354                                                 _igvn.MakeConX(step_size + distance));
2355       transform_later(cache_adr);
2356       cache_adr = new CastP2XNode(needgc_false, cache_adr);
2357       transform_later(cache_adr);
2358       // Address is aligned to execute prefetch to the beginning of cache line size.
2359       Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2360       cache_adr = new AndXNode(cache_adr, mask);
2361       transform_later(cache_adr);
2362       cache_adr = new CastX2PNode(cache_adr);
2363       transform_later(cache_adr);
2364 
2365       // Prefetch
2366       Node* prefetch = new PrefetchAllocationNode(contended_phi_rawmem, cache_adr);
2367       prefetch->set_req(0, needgc_false);
2368       transform_later(prefetch);
2369       contended_phi_rawmem = prefetch;
2370       Node* prefetch_adr;
2371       distance = step_size;
2372       for (intx i = 1; i < lines; i++) {
2373         prefetch_adr = AddPNode::make_off_heap(cache_adr,
2374                                                _igvn.MakeConX(distance));
2375         transform_later(prefetch_adr);
2376         prefetch = new PrefetchAllocationNode(contended_phi_rawmem, prefetch_adr);
2377         transform_later(prefetch);
2378         distance += step_size;
2379         contended_phi_rawmem = prefetch;
2380       }
2381    } else if (AllocatePrefetchStyle > 0) {
2382       // Insert a prefetch for each allocation only on the fast-path
2383       Node* prefetch_adr;
2384       Node* prefetch;
2385       // Generate several prefetch instructions.
2386       uint step_size = AllocatePrefetchStepSize;
2387       uint distance = AllocatePrefetchDistance;
2388       for (intx i = 0; i < lines; i++) {
2389         prefetch_adr = AddPNode::make_off_heap(new_eden_top,
2390                                                _igvn.MakeConX(distance));
2391         transform_later(prefetch_adr);
2392         prefetch = new PrefetchAllocationNode(i_o, prefetch_adr);
2393         // Do not let it float too high, since if eden_top == eden_end,
2394         // both might be null.
2395         if (i == 0) { // Set control for first prefetch, next follows it
2396           prefetch->init_req(0, needgc_false);
2397         }
2398         transform_later(prefetch);
2399         distance += step_size;
2400         i_o = prefetch;
2401       }
2402    }
2403    return i_o;
2404 }
2405 
2406 
2407 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2408   expand_allocate_common(alloc, nullptr, nullptr,
2409                          OptoRuntime::new_instance_Type(),
2410                          OptoRuntime::new_instance_Java(), nullptr);
2411 }
2412 
2413 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2414   Node* length = alloc->in(AllocateNode::ALength);
2415   Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2416   InitializeNode* init = alloc->initialization();
2417   Node* klass_node = alloc->in(AllocateNode::KlassNode);
2418   Node* init_value = alloc->in(AllocateNode::InitValue);
2419   const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2420   assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2421          ary_klass_t->is_refined_type(), "Must be a refined array klass");
2422   const TypeFunc* slow_call_type;
2423   address slow_call_address;  // Address of slow call
2424   if (init != nullptr && init->is_complete_with_arraycopy() &&
2425       ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2426     // Don't zero type array during slow allocation in VM since
2427     // it will be initialized later by arraycopy in compiled code.
2428     slow_call_address = OptoRuntime::new_array_nozero_Java();
2429     slow_call_type = OptoRuntime::new_array_nozero_Type();
2430   } else {
2431     slow_call_address = OptoRuntime::new_array_Java();
2432     slow_call_type = OptoRuntime::new_array_Type();
2433 
2434     if (init_value == nullptr) {
2435       init_value = _igvn.zerocon(T_OBJECT);
2436     } else if (UseCompressedOops) {
2437       init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2438     }
2439   }
2440   expand_allocate_common(alloc, length, init_value,
2441                          slow_call_type,
2442                          slow_call_address, valid_length_test);
2443 }
2444 
2445 //-------------------mark_eliminated_box----------------------------------
2446 //
2447 // During EA obj may point to several objects but after few ideal graph
2448 // transformations (CCP) it may point to only one non escaping object
2449 // (but still using phi), corresponding locks and unlocks will be marked
2450 // for elimination. Later obj could be replaced with a new node (new phi)
2451 // and which does not have escape information. And later after some graph
2452 // reshape other locks and unlocks (which were not marked for elimination
2453 // before) are connected to this new obj (phi) but they still will not be
2454 // marked for elimination since new obj has no escape information.
2455 // Mark all associated (same box and obj) lock and unlock nodes for
2456 // elimination if some of them marked already.
2457 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2458   BoxLockNode* oldbox = box->as_BoxLock();
2459   if (oldbox->is_eliminated()) {
2460     return; // This BoxLock node was processed already.
2461   }
2462   assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2463   // New implementation (EliminateNestedLocks) has separate BoxLock
2464   // node for each locked region so mark all associated locks/unlocks as
2465   // eliminated even if different objects are referenced in one locked region
2466   // (for example, OSR compilation of nested loop inside locked scope).
2467   if (EliminateNestedLocks ||
2468       oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2469     // Box is used only in one lock region. Mark this box as eliminated.
2470     oldbox->set_local();      // This verifies correct state of BoxLock
2471     _igvn.hash_delete(oldbox);
2472     oldbox->set_eliminated(); // This changes box's hash value
2473      _igvn.hash_insert(oldbox);
2474 
2475     for (uint i = 0; i < oldbox->outcnt(); i++) {
2476       Node* u = oldbox->raw_out(i);
2477       if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2478         AbstractLockNode* alock = u->as_AbstractLock();
2479         // Check lock's box since box could be referenced by Lock's debug info.
2480         if (alock->box_node() == oldbox) {
2481           // Mark eliminated all related locks and unlocks.
2482 #ifdef ASSERT
2483           alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2484 #endif
2485           alock->set_non_esc_obj();
2486         }
2487       }
2488     }
2489     return;
2490   }
2491 
2492   // Create new "eliminated" BoxLock node and use it in monitor debug info
2493   // instead of oldbox for the same object.
2494   BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2495 
2496   // Note: BoxLock node is marked eliminated only here and it is used
2497   // to indicate that all associated lock and unlock nodes are marked
2498   // for elimination.
2499   newbox->set_local(); // This verifies correct state of BoxLock
2500   newbox->set_eliminated();
2501   transform_later(newbox);
2502 
2503   // Replace old box node with new box for all users of the same object.
2504   for (uint i = 0; i < oldbox->outcnt();) {
2505     bool next_edge = true;
2506 
2507     Node* u = oldbox->raw_out(i);
2508     if (u->is_AbstractLock()) {
2509       AbstractLockNode* alock = u->as_AbstractLock();
2510       if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2511         // Replace Box and mark eliminated all related locks and unlocks.
2512 #ifdef ASSERT
2513         alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2514 #endif
2515         alock->set_non_esc_obj();
2516         _igvn.rehash_node_delayed(alock);
2517         alock->set_box_node(newbox);
2518         next_edge = false;
2519       }
2520     }
2521     if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2522       FastLockNode* flock = u->as_FastLock();
2523       assert(flock->box_node() == oldbox, "sanity");
2524       _igvn.rehash_node_delayed(flock);
2525       flock->set_box_node(newbox);
2526       next_edge = false;
2527     }
2528 
2529     // Replace old box in monitor debug info.
2530     if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2531       SafePointNode* sfn = u->as_SafePoint();
2532       JVMState* youngest_jvms = sfn->jvms();
2533       int max_depth = youngest_jvms->depth();
2534       for (int depth = 1; depth <= max_depth; depth++) {
2535         JVMState* jvms = youngest_jvms->of_depth(depth);
2536         int num_mon  = jvms->nof_monitors();
2537         // Loop over monitors
2538         for (int idx = 0; idx < num_mon; idx++) {
2539           Node* obj_node = sfn->monitor_obj(jvms, idx);
2540           Node* box_node = sfn->monitor_box(jvms, idx);
2541           if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2542             int j = jvms->monitor_box_offset(idx);
2543             _igvn.replace_input_of(u, j, newbox);
2544             next_edge = false;
2545           }
2546         }
2547       }
2548     }
2549     if (next_edge) i++;
2550   }
2551 }
2552 
2553 //-----------------------mark_eliminated_locking_nodes-----------------------
2554 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2555   if (!alock->is_balanced()) {
2556     return; // Can't do any more elimination for this locking region
2557   }
2558   if (EliminateNestedLocks) {
2559     if (alock->is_nested()) {
2560        assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2561        return;
2562     } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2563       // Only Lock node has JVMState needed here.
2564       // Not that preceding claim is documented anywhere else.
2565       if (alock->jvms() != nullptr) {
2566         if (alock->as_Lock()->is_nested_lock_region()) {
2567           // Mark eliminated related nested locks and unlocks.
2568           Node* obj = alock->obj_node();
2569           BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2570           assert(!box_node->is_eliminated(), "should not be marked yet");
2571           // Note: BoxLock node is marked eliminated only here
2572           // and it is used to indicate that all associated lock
2573           // and unlock nodes are marked for elimination.
2574           box_node->set_eliminated(); // Box's hash is always NO_HASH here
2575           for (uint i = 0; i < box_node->outcnt(); i++) {
2576             Node* u = box_node->raw_out(i);
2577             if (u->is_AbstractLock()) {
2578               alock = u->as_AbstractLock();
2579               if (alock->box_node() == box_node) {
2580                 // Verify that this Box is referenced only by related locks.
2581                 assert(alock->obj_node()->eqv_uncast(obj), "");
2582                 // Mark all related locks and unlocks.
2583 #ifdef ASSERT
2584                 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2585 #endif
2586                 alock->set_nested();
2587               }
2588             }
2589           }
2590         } else {
2591 #ifdef ASSERT
2592           alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2593           if (C->log() != nullptr)
2594             alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2595 #endif
2596         }
2597       }
2598       return;
2599     }
2600     // Process locks for non escaping object
2601     assert(alock->is_non_esc_obj(), "");
2602   } // EliminateNestedLocks
2603 
2604   if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2605     // Look for all locks of this object and mark them and
2606     // corresponding BoxLock nodes as eliminated.
2607     Node* obj = alock->obj_node();
2608     for (uint j = 0; j < obj->outcnt(); j++) {
2609       Node* o = obj->raw_out(j);
2610       if (o->is_AbstractLock() &&
2611           o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2612         alock = o->as_AbstractLock();
2613         Node* box = alock->box_node();
2614         // Replace old box node with new eliminated box for all users
2615         // of the same object and mark related locks as eliminated.
2616         mark_eliminated_box(box, obj);
2617       }
2618     }
2619   }
2620 }
2621 
2622 // we have determined that this lock/unlock can be eliminated, we simply
2623 // eliminate the node without expanding it.
2624 //
2625 // Note:  The membar's associated with the lock/unlock are currently not
2626 //        eliminated.  This should be investigated as a future enhancement.
2627 //
2628 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2629 
2630   if (!alock->is_eliminated()) {
2631     return false;
2632   }
2633 #ifdef ASSERT
2634   if (!alock->is_coarsened()) {
2635     // Check that new "eliminated" BoxLock node is created.
2636     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2637     assert(oldbox->is_eliminated(), "should be done already");
2638   }
2639 #endif
2640 
2641   alock->log_lock_optimization(C, "eliminate_lock");
2642 
2643 #ifndef PRODUCT
2644   if (PrintEliminateLocks) {
2645     tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2646   }
2647 #endif
2648 
2649   Node* mem  = alock->in(TypeFunc::Memory);
2650   Node* ctrl = alock->in(TypeFunc::Control);
2651   guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2652 
2653   _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2654   // There are 2 projections from the lock.  The lock node will
2655   // be deleted when its last use is subsumed below.
2656   assert(alock->outcnt() == 2 &&
2657          _callprojs->fallthrough_proj != nullptr &&
2658          _callprojs->fallthrough_memproj != nullptr,
2659          "Unexpected projections from Lock/Unlock");
2660 
2661   Node* fallthroughproj = _callprojs->fallthrough_proj;
2662   Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2663 
2664   // The memory projection from a lock/unlock is RawMem
2665   // The input to a Lock is merged memory, so extract its RawMem input
2666   // (unless the MergeMem has been optimized away.)
2667   if (alock->is_Lock()) {
2668     // Search for MemBarAcquireLock node and delete it also.
2669     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2670     assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2671     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2672     Node* memproj = membar->proj_out(TypeFunc::Memory);
2673     _igvn.replace_node(ctrlproj, fallthroughproj);
2674     _igvn.replace_node(memproj, memproj_fallthrough);
2675 
2676     // Delete FastLock node also if this Lock node is unique user
2677     // (a loop peeling may clone a Lock node).
2678     Node* flock = alock->as_Lock()->fastlock_node();
2679     if (flock->outcnt() == 1) {
2680       assert(flock->unique_out() == alock, "sanity");
2681       _igvn.replace_node(flock, top());
2682     }
2683   }
2684 
2685   // Search for MemBarReleaseLock node and delete it also.
2686   if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2687     MemBarNode* membar = ctrl->in(0)->as_MemBar();
2688     assert(membar->Opcode() == Op_MemBarReleaseLock &&
2689            mem->is_Proj() && membar == mem->in(0), "");
2690     _igvn.replace_node(fallthroughproj, ctrl);
2691     _igvn.replace_node(memproj_fallthrough, mem);
2692     fallthroughproj = ctrl;
2693     memproj_fallthrough = mem;
2694     ctrl = membar->in(TypeFunc::Control);
2695     mem  = membar->in(TypeFunc::Memory);
2696   }
2697 
2698   _igvn.replace_node(fallthroughproj, ctrl);
2699   _igvn.replace_node(memproj_fallthrough, mem);
2700   return true;
2701 }
2702 
2703 
2704 //------------------------------expand_lock_node----------------------
2705 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2706 
2707   Node* ctrl = lock->in(TypeFunc::Control);
2708   Node* mem = lock->in(TypeFunc::Memory);
2709   Node* obj = lock->obj_node();
2710   Node* box = lock->box_node();
2711   Node* flock = lock->fastlock_node();
2712 
2713   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2714 
2715   // Make the merge point
2716   Node *region;
2717   Node *mem_phi;
2718   Node *slow_path;
2719 
2720   region  = new RegionNode(3);
2721   // create a Phi for the memory state
2722   mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2723 
2724   // Optimize test; set region slot 2
2725   slow_path = opt_bits_test(ctrl, region, 2, flock);
2726   mem_phi->init_req(2, mem);
2727 
2728   // Make slow path call
2729   CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2730                                   OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2731                                   obj, box, nullptr);
2732 
2733   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2734 
2735   // Slow path can only throw asynchronous exceptions, which are always
2736   // de-opted.  So the compiler thinks the slow-call can never throw an
2737   // exception.  If it DOES throw an exception we would need the debug
2738   // info removed first (since if it throws there is no monitor).
2739   assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2740          _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2741 
2742   // Capture slow path
2743   // disconnect fall-through projection from call and create a new one
2744   // hook up users of fall-through projection to region
2745   Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2746   transform_later(slow_ctrl);
2747   _igvn.hash_delete(_callprojs->fallthrough_proj);
2748   _callprojs->fallthrough_proj->disconnect_inputs(C);
2749   region->init_req(1, slow_ctrl);
2750   // region inputs are now complete
2751   transform_later(region);
2752   _igvn.replace_node(_callprojs->fallthrough_proj, region);
2753 
2754   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2755 
2756   mem_phi->init_req(1, memproj);
2757 
2758   transform_later(mem_phi);
2759 
2760   _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2761 }
2762 
2763 //------------------------------expand_unlock_node----------------------
2764 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2765 
2766   Node* ctrl = unlock->in(TypeFunc::Control);
2767   Node* mem = unlock->in(TypeFunc::Memory);
2768   Node* obj = unlock->obj_node();
2769   Node* box = unlock->box_node();
2770 
2771   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2772 
2773   // No need for a null check on unlock
2774 
2775   // Make the merge point
2776   Node* region = new RegionNode(3);
2777 
2778   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2779   funlock = transform_later( funlock )->as_FastUnlock();
2780   // Optimize test; set region slot 2
2781   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2782   Node *thread = transform_later(new ThreadLocalNode());
2783 
2784   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2785                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2786                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2787 
2788   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2789   assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2790          _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2791 
2792   // No exceptions for unlocking
2793   // Capture slow path
2794   // disconnect fall-through projection from call and create a new one
2795   // hook up users of fall-through projection to region
2796   Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2797   transform_later(slow_ctrl);
2798   _igvn.hash_delete(_callprojs->fallthrough_proj);
2799   _callprojs->fallthrough_proj->disconnect_inputs(C);
2800   region->init_req(1, slow_ctrl);
2801   // region inputs are now complete
2802   transform_later(region);
2803   _igvn.replace_node(_callprojs->fallthrough_proj, region);
2804 
2805   if (_callprojs->fallthrough_memproj != nullptr) {
2806     // create a Phi for the memory state
2807     Node* mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2808     Node* memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2809     mem_phi->init_req(1, memproj);
2810     mem_phi->init_req(2, mem);
2811     transform_later(mem_phi);
2812     _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2813   }
2814 }
2815 
2816 // An inline type might be returned from the call but we don't know its
2817 // type. Either we get a buffered inline type (and nothing needs to be done)
2818 // or one of the values being returned is the klass of the inline type
2819 // and we need to allocate an inline type instance of that type and
2820 // initialize it with other values being returned. In that case, we
2821 // first try a fast path allocation and initialize the value with the
2822 // inline klass's pack handler or we fall back to a runtime call.
2823 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2824   assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2825   Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2826   if (ret == nullptr) {
2827     return;
2828   }
2829   const TypeFunc* tf = call->_tf;
2830   const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2831   const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2832   call->_tf = new_tf;
2833   // Make sure the change of type is applied before projections are processed by igvn
2834   _igvn.set_type(call, call->Value(&_igvn));
2835   _igvn.set_type(ret, ret->Value(&_igvn));
2836 
2837   // Before any new projection is added:
2838   CallProjections* projs = call->extract_projections(true, true);
2839 
2840   // Create temporary hook nodes that will be replaced below.
2841   // Add an input to prevent hook nodes from being dead.
2842   Node* ctl = new Node(call);
2843   Node* mem = new Node(ctl);
2844   Node* io = new Node(ctl);
2845   Node* ex_ctl = new Node(ctl);
2846   Node* ex_mem = new Node(ctl);
2847   Node* ex_io = new Node(ctl);
2848   Node* res = new Node(ctl);
2849 
2850   // Allocate a new buffered inline type only if a new one is not returned
2851   Node* cast = transform_later(new CastP2XNode(ctl, res));
2852   Node* mask = MakeConX(0x1);
2853   Node* masked = transform_later(new AndXNode(cast, mask));
2854   Node* cmp = transform_later(new CmpXNode(masked, mask));
2855   Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2856   IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2857   transform_later(allocation_iff);
2858   Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2859   Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2860   Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2861 
2862   // Try to allocate a new buffered inline instance either from TLAB or eden space
2863   Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2864   CallLeafNoFPNode* handler_call;
2865   const bool alloc_in_place = UseTLAB;
2866   if (alloc_in_place) {
2867     Node* fast_oop_ctrl = nullptr;
2868     Node* fast_oop_rawmem = nullptr;
2869     Node* mask2 = MakeConX(-2);
2870     Node* masked2 = transform_later(new AndXNode(cast, mask2));
2871     Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2872     Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2873     Node* layout_val = make_load_raw(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2874     Node* size_in_bytes = ConvI2X(layout_val);
2875     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2876     Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2877                                       fast_oop_ctrl, fast_oop_rawmem,
2878                                       AllocateInstancePrefetchLines);
2879     // Allocation succeed, initialize buffered inline instance header firstly,
2880     // and then initialize its fields with an inline class specific handler
2881     Node* mark_word_node;
2882     if (UseCompactObjectHeaders) {
2883       // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2884       mark_word_node = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2885     } else {
2886       // Otherwise, use the static prototype.
2887       mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2888     }
2889 
2890     fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2891     if (!UseCompactObjectHeaders) {
2892       // COH: Everything is encoded in the mark word, so nothing left to do.
2893       fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2894       fast_oop_rawmem = make_store_raw(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2895     }
2896     Node* members  = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InlineKlass::adr_members_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2897     Node* pack_handler = make_load_raw(fast_oop_ctrl, fast_oop_rawmem, members, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2898     handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2899                                         nullptr,
2900                                         "pack handler",
2901                                         TypeRawPtr::BOTTOM);
2902     handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2903     handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2904     handler_call->init_req(TypeFunc::I_O, top());
2905     handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2906     handler_call->init_req(TypeFunc::ReturnAdr, top());
2907     handler_call->init_req(TypeFunc::Parms, pack_handler);
2908     handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2909   } else {
2910     needgc_ctrl = allocation_ctl;
2911   }
2912 
2913   // Allocation failed, fall back to a runtime call
2914   CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2915                                                          SharedRuntime::store_inline_type_fields_to_buf_entry(),
2916                                                          "store_inline_type_fields",
2917                                                          TypePtr::BOTTOM);
2918   slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2919   slow_call->init_req(TypeFunc::Memory, mem);
2920   slow_call->init_req(TypeFunc::I_O, io);
2921   slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2922   slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2923   slow_call->init_req(TypeFunc::Parms, res);
2924 
2925   Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2926   Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2927   Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2928   Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2929   Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2930   Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2931   Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci));
2932 
2933   Node* ex_r = new RegionNode(3);
2934   Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2935   Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2936   ex_r->init_req(1, slow_excp);
2937   ex_mem_phi->init_req(1, slow_mem);
2938   ex_io_phi->init_req(1, slow_io);
2939   ex_r->init_req(2, ex_ctl);
2940   ex_mem_phi->init_req(2, ex_mem);
2941   ex_io_phi->init_req(2, ex_io);
2942   transform_later(ex_r);
2943   transform_later(ex_mem_phi);
2944   transform_later(ex_io_phi);
2945 
2946   // We don't know how many values are returned. This assumes the
2947   // worst case, that all available registers are used.
2948   for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2949     if (domain->field_at(i) == Type::HALF) {
2950       slow_call->init_req(i, top());
2951       if (alloc_in_place) {
2952         handler_call->init_req(i+1, top());
2953       }
2954       continue;
2955     }
2956     Node* proj = transform_later(new ProjNode(call, i));
2957     slow_call->init_req(i, proj);
2958     if (alloc_in_place) {
2959       handler_call->init_req(i+1, proj);
2960     }
2961   }
2962   // We can safepoint at that new call
2963   slow_call->copy_call_debug_info(&_igvn, call);
2964   transform_later(slow_call);
2965   if (alloc_in_place) {
2966     transform_later(handler_call);
2967   }
2968 
2969   Node* fast_ctl = nullptr;
2970   Node* fast_res = nullptr;
2971   MergeMemNode* fast_mem = nullptr;
2972   if (alloc_in_place) {
2973     fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2974     Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2975     fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2976     fast_mem = MergeMemNode::make(mem);
2977     fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2978     transform_later(fast_mem);
2979   }
2980 
2981   Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2982   Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2983   Node* io_phi = new PhiNode(r, Type::ABIO);
2984   Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2985   r->init_req(1, no_allocation_ctl);
2986   mem_phi->init_req(1, mem);
2987   io_phi->init_req(1, io);
2988   res_phi->init_req(1, no_allocation_res);
2989   r->init_req(2, slow_norm);
2990   mem_phi->init_req(2, slow_mem);
2991   io_phi->init_req(2, slow_io);
2992   res_phi->init_req(2, slow_res);
2993   if (alloc_in_place) {
2994     r->init_req(3, fast_ctl);
2995     mem_phi->init_req(3, fast_mem);
2996     io_phi->init_req(3, io);
2997     res_phi->init_req(3, fast_res);
2998   }
2999   transform_later(r);
3000   transform_later(mem_phi);
3001   transform_later(io_phi);
3002   transform_later(res_phi);
3003 
3004   // Do not let stores that initialize this buffer be reordered with a subsequent
3005   // store that would make this buffer accessible by other threads.
3006   MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
3007   transform_later(mb);
3008   mb->init_req(TypeFunc::Memory, mem_phi);
3009   mb->init_req(TypeFunc::Control, r);
3010   r = new ProjNode(mb, TypeFunc::Control);
3011   transform_later(r);
3012   mem_phi = new ProjNode(mb, TypeFunc::Memory);
3013   transform_later(mem_phi);
3014 
3015   assert(projs->nb_resproj == 1, "unexpected number of results");
3016   _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
3017   _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
3018   _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
3019   _igvn.replace_in_uses(projs->resproj[0], res_phi);
3020   _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
3021   _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
3022   _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
3023   // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
3024   Node* cn = projs->fallthrough_catchproj->in(0);
3025   _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
3026 
3027   _igvn.replace_node(ctl, projs->fallthrough_catchproj);
3028   _igvn.replace_node(mem, projs->fallthrough_memproj);
3029   _igvn.replace_node(io, projs->fallthrough_ioproj);
3030   _igvn.replace_node(res, projs->resproj[0]);
3031   _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
3032   _igvn.replace_node(ex_mem, projs->catchall_memproj);
3033   _igvn.replace_node(ex_io, projs->catchall_ioproj);
3034  }
3035 
3036 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
3037   assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
3038   Node* bol = check->unique_out();
3039   Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
3040   Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
3041   assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
3042 
3043   for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
3044     Node* iff = bol->last_out(i);
3045     assert(iff->is_If(), "where's the if?");
3046 
3047     if (iff->in(0)->is_top()) {
3048       _igvn.replace_input_of(iff, 1, C->top());
3049       continue;
3050     }
3051 
3052     IfTrueNode* iftrue = iff->as_If()->true_proj();
3053     IfFalseNode* iffalse = iff->as_If()->false_proj();
3054     Node* ctrl = iff->in(0);
3055 
3056     Node* subklass = nullptr;
3057     if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
3058       subklass = obj_or_subklass;
3059     } else {
3060       Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
3061       subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3062     }
3063 
3064     Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
3065 
3066     _igvn.replace_input_of(iff, 0, C->top());
3067     _igvn.replace_node(iftrue, not_subtype_ctrl);
3068     _igvn.replace_node(iffalse, ctrl);
3069   }
3070   _igvn.replace_node(check, C->top());
3071 }
3072 
3073 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
3074 //
3075 // long mark = array1.mark | array2.mark | ...;
3076 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
3077 // if (locked_bit == 0) {
3078 //   // One array is locked, load prototype header from the klass
3079 //   mark = array1.klass.proto | array2.klass.proto | ...
3080 // }
3081 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
3082 //    ...
3083 // }
3084 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
3085   bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
3086   if (array_inputs) {
3087     Node* mark = MakeConX(0);
3088     Node* locked_bit = MakeConX(markWord::unlocked_value);
3089     Node* mem = check->in(FlatArrayCheckNode::Memory);
3090     for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3091       Node* ary = check->in(i);
3092       const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
3093       assert(t != nullptr, "Mixing array and klass inputs");
3094       assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3095       Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
3096       Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3097       mark = _igvn.transform(new OrXNode(mark, mark_load));
3098       locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
3099     }
3100     assert(!mark->is_Con(), "Should have been optimized out");
3101     Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3102     Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
3103 
3104     // BoolNode might be shared, replace each if user
3105     Node* old_bol = check->unique_out();
3106     assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
3107     for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3108       IfNode* old_iff = old_bol->last_out(i)->as_If();
3109       Node* ctrl = old_iff->in(0);
3110       RegionNode* region = new RegionNode(3);
3111       Node* mark_phi = new PhiNode(region, TypeX_X);
3112 
3113       // Check if array is unlocked
3114       IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
3115 
3116       // Unlocked: Use bits from mark word
3117       region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
3118       mark_phi->init_req(1, mark);
3119 
3120       // Locked: Load prototype header from klass
3121       ctrl = _igvn.transform(new IfFalseNode(iff));
3122       Node* proto = MakeConX(0);
3123       for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3124         Node* ary = check->in(i);
3125         // Make loads control dependent to make sure they are only executed if array is locked
3126         Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3127         Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3128         Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3129         Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3130         proto = _igvn.transform(new OrXNode(proto, proto_load));
3131       }
3132       region->init_req(2, ctrl);
3133       mark_phi->init_req(2, proto);
3134 
3135       // Check if flat array bits are set
3136       Node* mask = MakeConX(markWord::flat_array_bit_in_place);
3137       Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
3138       cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
3139       Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
3140 
3141       ctrl = _igvn.transform(region);
3142       iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
3143       _igvn.replace_node(old_iff, iff);
3144     }
3145     _igvn.replace_node(check, C->top());
3146   } else {
3147     // Fall back to layout helper check
3148     Node* lhs = intcon(0);
3149     for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
3150       Node* array_or_klass = check->in(i);
3151       Node* klass = nullptr;
3152       const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
3153       assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
3154       if (t->isa_oopptr() != nullptr) {
3155         Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
3156         klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3157       } else {
3158         assert(t->isa_klassptr(), "Unexpected input type");
3159         klass = array_or_klass;
3160       }
3161       Node* lh_addr = basic_plus_adr(top(), klass, in_bytes(Klass::layout_helper_offset()));
3162       Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3163       lhs = _igvn.transform(new OrINode(lhs, lh_val));
3164     }
3165     Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
3166     Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
3167     Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
3168     Node* m2b = transform_later(new Conv2BNode(masked));
3169     // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
3170     // pattern, but the input to other potential users (e.g. Phi) to be some
3171     // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
3172     Node* old_bol = check->unique_out();
3173     for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
3174       Node* user = old_bol->last_out(i);
3175       for (uint j = 0; j < user->req(); j++) {
3176         Node* n = user->in(j);
3177         if (n == old_bol) {
3178           _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
3179         }
3180       }
3181     }
3182     _igvn.replace_node(check, C->top());
3183   }
3184 }
3185 
3186 // Perform refining of strip mined loop nodes in the macro nodes list.
3187 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
3188    for (int i = C->macro_count(); i > 0; i--) {
3189     Node* n = C->macro_node(i - 1);
3190     if (n->is_OuterStripMinedLoop()) {
3191       n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3192     }
3193   }
3194 }
3195 
3196 //---------------------------eliminate_macro_nodes----------------------
3197 // Eliminate scalar replaced allocations and associated locks.
3198 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
3199   if (C->macro_count() == 0) {
3200     return;
3201   }
3202 
3203   if (StressMacroElimination) {
3204     C->shuffle_macro_nodes();
3205   }
3206   NOT_PRODUCT(int membar_before = count_MemBar(C);)
3207 
3208   int iteration = 0;
3209   while (C->macro_count() > 0) {
3210     if (iteration++ > 100) {
3211       assert(false, "Too slow convergence of macro elimination");
3212       break;
3213     }
3214 
3215     // Postpone lock elimination to after EA when most allocations are eliminated
3216     // because they might block lock elimination if their escape state isn't
3217     // determined yet and we only got one chance at eliminating the lock.
3218     if (eliminate_locks) {
3219       // Before elimination may re-mark (change to Nested or NonEscObj)
3220       // all associated (same box and obj) lock and unlock nodes.
3221       int cnt = C->macro_count();
3222       for (int i=0; i < cnt; i++) {
3223         Node *n = C->macro_node(i);
3224         if (n->is_AbstractLock()) { // Lock and Unlock nodes
3225           mark_eliminated_locking_nodes(n->as_AbstractLock());
3226         }
3227       }
3228       // Re-marking may break consistency of Coarsened locks.
3229       if (!C->coarsened_locks_consistent()) {
3230         return; // recompile without Coarsened locks if broken
3231       } else {
3232         // After coarsened locks are eliminated locking regions
3233         // become unbalanced. We should not execute any more
3234         // locks elimination optimizations on them.
3235         C->mark_unbalanced_boxes();
3236       }
3237     }
3238 
3239     bool progress = false;
3240     for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3241       Node* n = C->macro_node(i - 1);
3242       bool success = false;
3243       DEBUG_ONLY(int old_macro_count = C->macro_count();)
3244       switch (n->class_id()) {
3245       case Node::Class_Allocate:
3246       case Node::Class_AllocateArray:
3247         success = eliminate_allocate_node(n->as_Allocate());
3248 #ifndef PRODUCT
3249         if (success && PrintOptoStatistics) {
3250           AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3251         }
3252 #endif
3253         break;
3254       case Node::Class_CallStaticJava: {
3255         CallStaticJavaNode* call = n->as_CallStaticJava();
3256         if (!call->method()->is_method_handle_intrinsic()) {
3257           success = eliminate_boxing_node(n->as_CallStaticJava());
3258         }
3259         break;
3260       }
3261       case Node::Class_Lock:
3262       case Node::Class_Unlock:
3263         if (eliminate_locks) {
3264           success = eliminate_locking_node(n->as_AbstractLock());
3265 #ifndef PRODUCT
3266           if (success && PrintOptoStatistics) {
3267             AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3268           }
3269 #endif
3270         }
3271         break;
3272       case Node::Class_ArrayCopy:
3273         break;
3274       case Node::Class_OuterStripMinedLoop:
3275         break;
3276       case Node::Class_SubTypeCheck:
3277         break;
3278       case Node::Class_Opaque1:
3279         break;
3280       case Node::Class_FlatArrayCheck:
3281         break;
3282       default:
3283         assert(n->Opcode() == Op_LoopLimit ||
3284                n->Opcode() == Op_ModD ||
3285                n->Opcode() == Op_ModF ||
3286                n->Opcode() == Op_PowD ||
3287                n->is_OpaqueConstantBool()    ||
3288                n->is_OpaqueInitializedAssertionPredicate() ||
3289                n->Opcode() == Op_MaxL      ||
3290                n->Opcode() == Op_MinL      ||
3291                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3292                "unknown node type in macro list");
3293       }
3294       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3295       progress = progress || success;
3296       if (success) {
3297         C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3298       }
3299     }
3300 
3301     // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3302     // transformation is pending). If an allocation is used only in safepoints, elimination of
3303     // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3304     // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3305     // iteration.
3306     _igvn.set_delay_transform(false);
3307     _igvn.optimize();
3308     if (C->failing()) {
3309       return;
3310     }
3311     _igvn.set_delay_transform(true);
3312 
3313     if (!progress) {
3314       break;
3315     }
3316   }
3317 #ifndef PRODUCT
3318   if (PrintOptoStatistics) {
3319     int membar_after = count_MemBar(C);
3320     AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3321   }
3322 #endif
3323 }
3324 
3325 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3326   if (C->macro_count() == 0) {
3327     return;
3328   }
3329   refine_strip_mined_loop_macro_nodes();
3330   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3331   bool progress = true;
3332   while (progress) {
3333     progress = false;
3334     for (int i = C->macro_count(); i > 0; i--) {
3335       Node* n = C->macro_node(i-1);
3336       bool success = false;
3337       DEBUG_ONLY(int old_macro_count = C->macro_count();)
3338       if (n->Opcode() == Op_LoopLimit) {
3339         // Remove it from macro list and put on IGVN worklist to optimize.
3340         C->remove_macro_node(n);
3341         _igvn._worklist.push(n);
3342         success = true;
3343       } else if (n->Opcode() == Op_CallStaticJava) {
3344         CallStaticJavaNode* call = n->as_CallStaticJava();
3345         if (!call->method()->is_method_handle_intrinsic()) {
3346           // Remove it from macro list and put on IGVN worklist to optimize.
3347           C->remove_macro_node(n);
3348           _igvn._worklist.push(n);
3349           success = true;
3350         }
3351       } else if (n->is_Opaque1()) {
3352         _igvn.replace_node(n, n->in(1));
3353         success = true;
3354       } else if (n->is_OpaqueConstantBool()) {
3355         // Tests with OpaqueConstantBool nodes are implicitly known. Replace the node with true/false. In debug builds,
3356         // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3357         // we will execute a Halt node.
3358 #ifdef ASSERT
3359         _igvn.replace_node(n, n->in(1));
3360 #else
3361         _igvn.replace_node(n, _igvn.intcon(n->as_OpaqueConstantBool()->constant()));
3362 #endif
3363         success = true;
3364       } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3365           // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3366           // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3367           // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3368           // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3369           // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3370 #ifdef ASSERT
3371         _igvn.replace_node(n, n->in(1));
3372 #else
3373         _igvn.replace_node(n, _igvn.intcon(1));
3374 #endif // ASSERT
3375       } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3376         C->remove_macro_node(n);
3377         success = true;
3378       } else if (n->Opcode() == Op_MaxL) {
3379         // Since MaxL and MinL are not implemented in the backend, we expand them to
3380         // a CMoveL construct now. At least until here, the type could be computed
3381         // precisely. CMoveL is not so smart, but we can give it at least the best
3382         // type we know abouot n now.
3383         Node* repl = MinMaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3384         _igvn.replace_node(n, repl);
3385         success = true;
3386       } else if (n->Opcode() == Op_MinL) {
3387         Node* repl = MinMaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3388         _igvn.replace_node(n, repl);
3389         success = true;
3390       }
3391       assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3392       progress = progress || success;
3393       if (success) {
3394         C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3395       }
3396     }
3397   }
3398 }
3399 
3400 //------------------------------expand_macro_nodes----------------------
3401 //  Returns true if a failure occurred.
3402 bool PhaseMacroExpand::expand_macro_nodes() {
3403   if (StressMacroExpansion) {
3404     C->shuffle_macro_nodes();
3405   }
3406 
3407   // Clean up the graph so we're less likely to hit the maximum node
3408   // limit
3409   _igvn.set_delay_transform(false);
3410   _igvn.optimize();
3411   if (C->failing())  return true;
3412   _igvn.set_delay_transform(true);
3413 
3414 
3415   // Because we run IGVN after each expansion, some macro nodes may go
3416   // dead and be removed from the list as we iterate over it. Move
3417   // Allocate nodes (processed in a second pass) at the beginning of
3418   // the list and then iterate from the last element of the list until
3419   // an Allocate node is seen. This is robust to random deletion in
3420   // the list due to nodes going dead.
3421   C->sort_macro_nodes();
3422 
3423   // expand arraycopy "macro" nodes first
3424   // For ReduceBulkZeroing, we must first process all arraycopy nodes
3425   // before the allocate nodes are expanded.
3426   while (C->macro_count() > 0) {
3427     int macro_count = C->macro_count();
3428     Node * n = C->macro_node(macro_count-1);
3429     assert(n->is_macro(), "only macro nodes expected here");
3430     if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3431       // node is unreachable, so don't try to expand it
3432       C->remove_macro_node(n);
3433       continue;
3434     }
3435     if (n->is_Allocate()) {
3436       break;
3437     }
3438     // Make sure expansion will not cause node limit to be exceeded.
3439     // Worst case is a macro node gets expanded into about 200 nodes.
3440     // Allow 50% more for optimization.
3441     if (C->check_node_count(300, "out of nodes before macro expansion")) {
3442       return true;
3443     }
3444 
3445     DEBUG_ONLY(int old_macro_count = C->macro_count();)
3446     switch (n->class_id()) {
3447     case Node::Class_Lock:
3448       expand_lock_node(n->as_Lock());
3449       break;
3450     case Node::Class_Unlock:
3451       expand_unlock_node(n->as_Unlock());
3452       break;
3453     case Node::Class_ArrayCopy:
3454       expand_arraycopy_node(n->as_ArrayCopy());
3455       break;
3456     case Node::Class_SubTypeCheck:
3457       expand_subtypecheck_node(n->as_SubTypeCheck());
3458       break;
3459     case Node::Class_CallStaticJava:
3460       expand_mh_intrinsic_return(n->as_CallStaticJava());
3461       C->remove_macro_node(n);
3462       break;
3463     case Node::Class_FlatArrayCheck:
3464       expand_flatarraycheck_node(n->as_FlatArrayCheck());
3465       break;
3466     default:
3467       switch (n->Opcode()) {
3468       case Op_ModD:
3469       case Op_ModF:
3470       case Op_PowD: {
3471         CallLeafPureNode* call_macro = n->as_CallLeafPure();
3472         CallLeafPureNode* call = call_macro->inline_call_leaf_pure_node();
3473         _igvn.replace_node(call_macro, call);
3474         transform_later(call);
3475         break;
3476       }
3477       default:
3478         assert(false, "unknown node type in macro list");
3479       }
3480     }
3481     assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3482     if (C->failing())  return true;
3483     C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3484 
3485     // Clean up the graph so we're less likely to hit the maximum node
3486     // limit
3487     _igvn.set_delay_transform(false);
3488     _igvn.optimize();
3489     if (C->failing())  return true;
3490     _igvn.set_delay_transform(true);
3491   }
3492 
3493   // All nodes except Allocate nodes are expanded now. There could be
3494   // new optimization opportunities (such as folding newly created
3495   // load from a just allocated object). Run IGVN.
3496 
3497   // expand "macro" nodes
3498   // nodes are removed from the macro list as they are processed
3499   while (C->macro_count() > 0) {
3500     int macro_count = C->macro_count();
3501     Node * n = C->macro_node(macro_count-1);
3502     assert(n->is_macro(), "only macro nodes expected here");
3503     if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3504       // node is unreachable, so don't try to expand it
3505       C->remove_macro_node(n);
3506       continue;
3507     }
3508     // Make sure expansion will not cause node limit to be exceeded.
3509     // Worst case is a macro node gets expanded into about 200 nodes.
3510     // Allow 50% more for optimization.
3511     if (C->check_node_count(300, "out of nodes before macro expansion")) {
3512       return true;
3513     }
3514     switch (n->class_id()) {
3515     case Node::Class_Allocate:
3516       expand_allocate(n->as_Allocate());
3517       break;
3518     case Node::Class_AllocateArray:
3519       expand_allocate_array(n->as_AllocateArray());
3520       break;
3521     default:
3522       assert(false, "unknown node type in macro list");
3523     }
3524     assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3525     if (C->failing())  return true;
3526     C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3527 
3528     // Clean up the graph so we're less likely to hit the maximum node
3529     // limit
3530     _igvn.set_delay_transform(false);
3531     _igvn.optimize();
3532     if (C->failing())  return true;
3533     _igvn.set_delay_transform(true);
3534   }
3535 
3536   _igvn.set_delay_transform(false);
3537   return false;
3538 }
3539 
3540 #ifndef PRODUCT
3541 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3542 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3543 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3544 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3545 
3546 void PhaseMacroExpand::print_statistics() {
3547   tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
3548   tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
3549   tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
3550   tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
3551 }
3552 
3553 int PhaseMacroExpand::count_MemBar(Compile *C) {
3554   if (!PrintOptoStatistics) {
3555     return 0;
3556   }
3557   Unique_Node_List ideal_nodes;
3558   int total = 0;
3559   ideal_nodes.map(C->live_nodes(), nullptr);
3560   ideal_nodes.push(C->root());
3561   for (uint next = 0; next < ideal_nodes.size(); ++next) {
3562     Node* n = ideal_nodes.at(next);
3563     if (n->is_MemBar()) {
3564       total++;
3565     }
3566     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3567       Node* m = n->fast_out(i);
3568       ideal_nodes.push(m);
3569     }
3570   }
3571   return total;
3572 }
3573 #endif