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/bcEscapeAnalyzer.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc/shared/barrierSet.hpp"
  28 #include "gc/shared/c2/barrierSetC2.hpp"
  29 #include "libadt/vectset.hpp"
  30 #include "memory/allocation.hpp"
  31 #include "memory/metaspace.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "opto/arraycopynode.hpp"
  34 #include "opto/c2compiler.hpp"
  35 #include "opto/callnode.hpp"
  36 #include "opto/castnode.hpp"
  37 #include "opto/cfgnode.hpp"
  38 #include "opto/compile.hpp"
  39 #include "opto/escape.hpp"
  40 #include "opto/inlinetypenode.hpp"
  41 #include "opto/locknode.hpp"
  42 #include "opto/macro.hpp"
  43 #include "opto/movenode.hpp"
  44 #include "opto/narrowptrnode.hpp"
  45 #include "opto/phaseX.hpp"
  46 #include "opto/rootnode.hpp"
  47 #include "utilities/macros.hpp"
  48 
  49 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
  50   // If ReduceAllocationMerges is enabled we might call split_through_phi during
  51   // split_unique_types and that will create additional nodes that need to be
  52   // pushed to the ConnectionGraph. The code below bumps the initial capacity of
  53   // _nodes by 10% to account for these additional nodes. If capacity is exceeded
  54   // the array will be reallocated.
  55   _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
  56   _in_worklist(C->comp_arena()),
  57   _next_pidx(0),
  58   _collecting(true),
  59   _verify(false),
  60   _compile(C),
  61   _igvn(igvn),
  62   _invocation(invocation),
  63   _build_iterations(0),
  64   _build_time(0.),
  65   _node_map(C->comp_arena()) {
  66   // Add unknown java object.
  67   add_java_object(C->top(), PointsToNode::GlobalEscape);
  68   phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
  69   set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
  70   // Add ConP and ConN null oop nodes
  71   Node* oop_null = igvn->zerocon(T_OBJECT);
  72   assert(oop_null->_idx < nodes_size(), "should be created already");
  73   add_java_object(oop_null, PointsToNode::NoEscape);
  74   null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
  75   set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
  76   if (UseCompressedOops) {
  77     Node* noop_null = igvn->zerocon(T_NARROWOOP);
  78     assert(noop_null->_idx < nodes_size(), "should be created already");
  79     map_ideal_node(noop_null, null_obj);
  80   }
  81 }
  82 
  83 bool ConnectionGraph::has_candidates(Compile *C) {
  84   // EA brings benefits only when the code has allocations and/or locks which
  85   // are represented by ideal Macro nodes.
  86   int cnt = C->macro_count();
  87   for (int i = 0; i < cnt; i++) {
  88     Node *n = C->macro_node(i);
  89     if (n->is_Allocate()) {
  90       return true;
  91     }
  92     if (n->is_Lock()) {
  93       Node* obj = n->as_Lock()->obj_node()->uncast();
  94       if (!(obj->is_Parm() || obj->is_Con())) {
  95         return true;
  96       }
  97     }
  98     if (n->is_CallStaticJava() &&
  99         n->as_CallStaticJava()->is_boxing_method()) {
 100       return true;
 101     }
 102   }
 103   return false;
 104 }
 105 
 106 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
 107   Compile::TracePhase tp(Phase::_t_escapeAnalysis);
 108   ResourceMark rm;
 109 
 110   // Add ConP and ConN null oop nodes before ConnectionGraph construction
 111   // to create space for them in ConnectionGraph::_nodes[].
 112   Node* oop_null = igvn->zerocon(T_OBJECT);
 113   Node* noop_null = igvn->zerocon(T_NARROWOOP);
 114   int invocation = 0;
 115   if (C->congraph() != nullptr) {
 116     invocation = C->congraph()->_invocation + 1;
 117   }
 118   ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
 119   NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(congraph);)
 120   // Perform escape analysis
 121   if (congraph->compute_escape()) {
 122     // There are non escaping objects.
 123     C->set_congraph(congraph);
 124   }
 125   NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(nullptr);)
 126   // Cleanup.
 127   if (oop_null->outcnt() == 0) {
 128     igvn->hash_delete(oop_null);
 129   }
 130   if (noop_null->outcnt() == 0) {
 131     igvn->hash_delete(noop_null);
 132   }
 133 
 134   C->print_method(PHASE_AFTER_EA, 2);
 135 }
 136 
 137 bool ConnectionGraph::compute_escape() {
 138   Compile* C = _compile;
 139   PhaseGVN* igvn = _igvn;
 140 
 141   // Worklists used by EA.
 142   Unique_Node_List delayed_worklist;
 143   Unique_Node_List reducible_merges;
 144   GrowableArray<Node*> alloc_worklist;
 145   GrowableArray<Node*> ptr_cmp_worklist;
 146   GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
 147   GrowableArray<ArrayCopyNode*>  arraycopy_worklist;
 148   GrowableArray<PointsToNode*>   ptnodes_worklist;
 149   GrowableArray<JavaObjectNode*> java_objects_worklist;
 150   GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
 151   GrowableArray<FieldNode*>      oop_fields_worklist;
 152   GrowableArray<SafePointNode*>  sfn_worklist;
 153   GrowableArray<MergeMemNode*>   mergemem_worklist;
 154   DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
 155 
 156   { Compile::TracePhase tp(Phase::_t_connectionGraph);
 157 
 158   // 1. Populate Connection Graph (CG) with PointsTo nodes.
 159   ideal_nodes.map(C->live_nodes(), nullptr);  // preallocate space
 160   // Initialize worklist
 161   if (C->root() != nullptr) {
 162     ideal_nodes.push(C->root());
 163   }
 164   // Processed ideal nodes are unique on ideal_nodes list
 165   // but several ideal nodes are mapped to the phantom_obj.
 166   // To avoid duplicated entries on the following worklists
 167   // add the phantom_obj only once to them.
 168   ptnodes_worklist.append(phantom_obj);
 169   java_objects_worklist.append(phantom_obj);
 170   for( uint next = 0; next < ideal_nodes.size(); ++next ) {
 171     Node* n = ideal_nodes.at(next);
 172     if ((n->Opcode() == Op_LoadX || n->Opcode() == Op_StoreX) &&
 173         !n->in(MemNode::Address)->is_AddP() &&
 174         _igvn->type(n->in(MemNode::Address))->isa_oopptr()) {
 175       // Load/Store at mark work address is at offset 0 so has no AddP which confuses EA
 176       Node* addp = AddPNode::make_with_base(n->in(MemNode::Address), n->in(MemNode::Address), _igvn->MakeConX(0));
 177       _igvn->register_new_node_with_optimizer(addp);
 178       _igvn->replace_input_of(n, MemNode::Address, addp);
 179       ideal_nodes.push(addp);
 180       _nodes.at_put_grow(addp->_idx, nullptr, nullptr);
 181     }
 182     // Create PointsTo nodes and add them to Connection Graph. Called
 183     // only once per ideal node since ideal_nodes is Unique_Node list.
 184     add_node_to_connection_graph(n, &delayed_worklist);
 185     PointsToNode* ptn = ptnode_adr(n->_idx);
 186     if (ptn != nullptr && ptn != phantom_obj) {
 187       ptnodes_worklist.append(ptn);
 188       if (ptn->is_JavaObject()) {
 189         java_objects_worklist.append(ptn->as_JavaObject());
 190         if ((n->is_Allocate() || n->is_CallStaticJava()) &&
 191             (ptn->escape_state() < PointsToNode::GlobalEscape)) {
 192           // Only allocations and java static calls results are interesting.
 193           non_escaped_allocs_worklist.append(ptn->as_JavaObject());
 194         }
 195       } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
 196         oop_fields_worklist.append(ptn->as_Field());
 197       }
 198     }
 199     // Collect some interesting nodes for further use.
 200     switch (n->Opcode()) {
 201       case Op_MergeMem:
 202         // Collect all MergeMem nodes to add memory slices for
 203         // scalar replaceable objects in split_unique_types().
 204         mergemem_worklist.append(n->as_MergeMem());
 205         break;
 206       case Op_CmpP:
 207       case Op_CmpN:
 208         // Collect compare pointers nodes.
 209         if (OptimizePtrCompare) {
 210           ptr_cmp_worklist.append(n);
 211         }
 212         break;
 213       case Op_MemBarStoreStore:
 214         // Collect all MemBarStoreStore nodes so that depending on the
 215         // escape status of the associated Allocate node some of them
 216         // may be eliminated.
 217         if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
 218           storestore_worklist.append(n->as_MemBarStoreStore());
 219         }
 220         // If MemBarStoreStore has a precedent edge add it to the worklist (like MemBarRelease)
 221       case Op_MemBarRelease:
 222         if (n->req() > MemBarNode::Precedent) {
 223           record_for_optimizer(n);
 224         }
 225         break;
 226 #ifdef ASSERT
 227       case Op_AddP:
 228         // Collect address nodes for graph verification.
 229         addp_worklist.append(n);
 230         break;
 231 #endif
 232       case Op_ArrayCopy:
 233         // Keep a list of ArrayCopy nodes so if one of its input is non
 234         // escaping, we can record a unique type
 235         arraycopy_worklist.append(n->as_ArrayCopy());
 236         break;
 237       default:
 238         // not interested now, ignore...
 239         break;
 240     }
 241     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 242       Node* m = n->fast_out(i);   // Get user
 243       ideal_nodes.push(m);
 244     }
 245     if (n->is_SafePoint()) {
 246       sfn_worklist.append(n->as_SafePoint());
 247     }
 248   }
 249 
 250 #ifndef PRODUCT
 251   if (_compile->directive()->TraceEscapeAnalysisOption) {
 252     tty->print("+++++ Initial worklist for ");
 253     _compile->method()->print_name();
 254     tty->print_cr(" (ea_inv=%d)", _invocation);
 255     for (int i = 0; i < ptnodes_worklist.length(); i++) {
 256       PointsToNode* ptn = ptnodes_worklist.at(i);
 257       ptn->dump();
 258     }
 259     tty->print_cr("+++++ Calculating escape states and scalar replaceability");
 260   }
 261 #endif
 262 
 263   if (non_escaped_allocs_worklist.length() == 0) {
 264     _collecting = false;
 265     NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 266     return false; // Nothing to do.
 267   }
 268   // Add final simple edges to graph.
 269   while(delayed_worklist.size() > 0) {
 270     Node* n = delayed_worklist.pop();
 271     add_final_edges(n);
 272   }
 273 
 274 #ifdef ASSERT
 275   if (VerifyConnectionGraph) {
 276     // Verify that no new simple edges could be created and all
 277     // local vars has edges.
 278     _verify = true;
 279     int ptnodes_length = ptnodes_worklist.length();
 280     for (int next = 0; next < ptnodes_length; ++next) {
 281       PointsToNode* ptn = ptnodes_worklist.at(next);
 282       add_final_edges(ptn->ideal_node());
 283       if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
 284         ptn->dump();
 285         assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
 286       }
 287     }
 288     _verify = false;
 289   }
 290 #endif
 291   // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
 292   // processing, calls to CI to resolve symbols (types, fields, methods)
 293   // referenced in bytecode. During symbol resolution VM may throw
 294   // an exception which CI cleans and converts to compilation failure.
 295   if (C->failing()) {
 296     NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 297     return false;
 298   }
 299 
 300   _compile->print_method(PHASE_EA_AFTER_INITIAL_CONGRAPH, 4);
 301 
 302   // 2. Finish Graph construction by propagating references to all
 303   //    java objects through graph.
 304   if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
 305                                  java_objects_worklist, oop_fields_worklist)) {
 306     // All objects escaped or hit time or iterations limits.
 307     _collecting = false;
 308     NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 309     return false;
 310   }
 311 
 312   _compile->print_method(PHASE_EA_AFTER_COMPLETE_CONGRAPH, 4);
 313 
 314   // 3. Adjust scalar_replaceable state of nonescaping objects and push
 315   //    scalar replaceable allocations on alloc_worklist for processing
 316   //    in split_unique_types().
 317   GrowableArray<JavaObjectNode*> jobj_worklist;
 318   int non_escaped_length = non_escaped_allocs_worklist.length();
 319   bool found_nsr_alloc = false;
 320   for (int next = 0; next < non_escaped_length; next++) {
 321     JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
 322     bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
 323     Node* n = ptn->ideal_node();
 324     if (n->is_Allocate()) {
 325       n->as_Allocate()->_is_non_escaping = noescape;
 326     }
 327     if (noescape && ptn->scalar_replaceable()) {
 328       adjust_scalar_replaceable_state(ptn, reducible_merges);
 329       if (ptn->scalar_replaceable()) {
 330         jobj_worklist.push(ptn);
 331       } else {
 332         found_nsr_alloc = true;
 333       }
 334     }
 335     _compile->print_method(PHASE_EA_ADJUST_SCALAR_REPLACEABLE_ITER, 6, n);
 336   }
 337 
 338   // Propagate NSR (Not Scalar Replaceable) state.
 339   if (found_nsr_alloc) {
 340     find_scalar_replaceable_allocs(jobj_worklist, reducible_merges);
 341   }
 342 
 343   // alloc_worklist will be processed in reverse push order.
 344   // Therefore the reducible Phis will be processed for last and that's what we
 345   // want because by then the scalarizable inputs of the merge will already have
 346   // an unique instance type.
 347   for (uint i = 0; i < reducible_merges.size(); i++ ) {
 348     Node* n = reducible_merges.at(i);
 349     alloc_worklist.append(n);
 350   }
 351 
 352   for (int next = 0; next < jobj_worklist.length(); ++next) {
 353     JavaObjectNode* jobj = jobj_worklist.at(next);
 354     if (jobj->scalar_replaceable()) {
 355       alloc_worklist.append(jobj->ideal_node());
 356     }
 357   }
 358 
 359 #ifdef ASSERT
 360   if (VerifyConnectionGraph) {
 361     // Verify that graph is complete - no new edges could be added or needed.
 362     verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
 363                             java_objects_worklist, addp_worklist);
 364   }
 365   assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
 366   assert(null_obj->escape_state() == PointsToNode::NoEscape &&
 367          null_obj->edge_count() == 0 &&
 368          !null_obj->arraycopy_src() &&
 369          !null_obj->arraycopy_dst(), "sanity");
 370 #endif
 371 
 372   _collecting = false;
 373 
 374   _compile->print_method(PHASE_EA_AFTER_PROPAGATE_NSR, 4);
 375   } // TracePhase t3("connectionGraph")
 376 
 377   // 4. Optimize ideal graph based on EA information.
 378   bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
 379   if (has_non_escaping_obj) {
 380     optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
 381   }
 382 
 383 #ifndef PRODUCT
 384   if (PrintEscapeAnalysis) {
 385     dump(ptnodes_worklist); // Dump ConnectionGraph
 386   }
 387 #endif
 388 
 389 #ifdef ASSERT
 390   if (VerifyConnectionGraph) {
 391     int alloc_length = alloc_worklist.length();
 392     for (int next = 0; next < alloc_length; ++next) {
 393       Node* n = alloc_worklist.at(next);
 394       PointsToNode* ptn = ptnode_adr(n->_idx);
 395       assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
 396     }
 397   }
 398 
 399   if (VerifyReduceAllocationMerges) {
 400     for (uint i = 0; i < reducible_merges.size(); i++ ) {
 401       Node* n = reducible_merges.at(i);
 402       if (!can_reduce_phi(n->as_Phi())) {
 403         TraceReduceAllocationMerges = true;
 404         n->dump(2);
 405         n->dump(-2);
 406         assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
 407       }
 408     }
 409   }
 410 #endif
 411 
 412   _compile->print_method(PHASE_EA_AFTER_GRAPH_OPTIMIZATION, 4);
 413 
 414   // 5. Separate memory graph for scalar replaceable allcations.
 415   bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
 416   if (has_scalar_replaceable_candidates && EliminateAllocations) {
 417     assert(C->do_aliasing(), "Aliasing should be enabled");
 418     // Now use the escape information to create unique types for
 419     // scalar replaceable objects.
 420     split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
 421     if (C->failing()) {
 422       NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 423       return false;
 424     }
 425 
 426 #ifdef ASSERT
 427   } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
 428     tty->print("=== No allocations eliminated for ");
 429     C->method()->print_short_name();
 430     if (!EliminateAllocations) {
 431       tty->print(" since EliminateAllocations is off ===");
 432     } else if(!has_scalar_replaceable_candidates) {
 433       tty->print(" since there are no scalar replaceable candidates ===");
 434     }
 435     tty->cr();
 436 #endif
 437   }
 438 
 439   // 6. Expand flat accesses if the object does not escape. This adds nodes to
 440   // the graph, so it has to be after split_unique_types. This expands atomic
 441   // mismatched accesses (though encapsulated in LoadFlats and StoreFlats) into
 442   // non-mismatched accesses, so it is better before reduce allocation merges.
 443   if (has_non_escaping_obj) {
 444     optimize_flat_accesses(sfn_worklist);
 445   }
 446 
 447   _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES, 4);
 448 
 449   // 7. Reduce allocation merges used as debug information. This is done after
 450   // split_unique_types because the methods used to create SafePointScalarObject
 451   // need to traverse the memory graph to find values for object fields. We also
 452   // set to null the scalarized inputs of reducible Phis so that the Allocate
 453   // that they point can be later scalar replaced.
 454   bool delay = _igvn->delay_transform();
 455   _igvn->set_delay_transform(true);
 456   for (uint i = 0; i < reducible_merges.size(); i++) {
 457     Node* n = reducible_merges.at(i);
 458     if (n->outcnt() > 0) {
 459       if (!reduce_phi_on_safepoints(n->as_Phi())) {
 460         NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 461         C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
 462         return false;
 463       }
 464 
 465       // Now we set the scalar replaceable inputs of ophi to null, which is
 466       // the last piece that would prevent it from being scalar replaceable.
 467       reset_scalar_replaceable_entries(n->as_Phi());
 468     }
 469   }
 470   _igvn->set_delay_transform(delay);
 471 
 472   // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
 473   // java calls if they pass ArgEscape objects as parameters.
 474   if (has_non_escaping_obj &&
 475       (C->env()->should_retain_local_variables() ||
 476        C->env()->jvmti_can_get_owned_monitor_info() ||
 477        C->env()->jvmti_can_walk_any_space() ||
 478        DeoptimizeObjectsALot)) {
 479     int sfn_length = sfn_worklist.length();
 480     for (int next = 0; next < sfn_length; next++) {
 481       SafePointNode* sfn = sfn_worklist.at(next);
 482       sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
 483       if (sfn->is_CallJava()) {
 484         CallJavaNode* call = sfn->as_CallJava();
 485         call->set_arg_escape(has_arg_escape(call));
 486       }
 487     }
 488   }
 489 
 490   _compile->print_method(PHASE_EA_AFTER_REDUCE_PHI_ON_SAFEPOINTS, 4);
 491 
 492   NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
 493   return has_non_escaping_obj;
 494 }
 495 
 496 // Check if it's profitable to reduce the Phi passed as parameter.  Returns true
 497 // if at least one scalar replaceable allocation participates in the merge.
 498 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
 499   bool found_sr_allocate = false;
 500 
 501   for (uint i = 1; i < ophi->req(); i++) {
 502     JavaObjectNode* ptn = unique_java_object(ophi->in(i));
 503     if (ptn != nullptr && ptn->scalar_replaceable()) {
 504       AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
 505 
 506       // Don't handle arrays.
 507       if (alloc->Opcode() != Op_Allocate) {
 508         assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
 509         continue;
 510       }
 511 
 512       if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
 513         found_sr_allocate = true;
 514       } else {
 515         NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
 516         ptn->set_scalar_replaceable(false);
 517       }
 518     }
 519   }
 520 
 521   NOT_PRODUCT(if (TraceReduceAllocationMerges && !found_sr_allocate) tty->print_cr("Can NOT reduce Phi %d on invocation %d. No SR Allocate as input.", ophi->_idx, _invocation);)
 522   return found_sr_allocate;
 523 }
 524 
 525 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
 526 // I require the 'other' input to be a constant so that I can move the Cmp
 527 // around safely.
 528 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
 529   assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
 530   Node* left = cmp->in(1);
 531   Node* right = cmp->in(2);
 532 
 533   return (left == n || right == n) &&
 534          (left->is_Con() || right->is_Con()) &&
 535          cmp->outcnt() == 1;
 536 }
 537 
 538 // We are going to check if any of the SafePointScalarMerge entries
 539 // in the SafePoint reference the Phi that we are checking.
 540 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
 541   JVMState *jvms = sfpt->jvms();
 542 
 543   for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
 544     Node* sfpt_in = sfpt->in(i);
 545     if (sfpt_in->is_SafePointScalarMerge()) {
 546       SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
 547       Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
 548       if (nsr_ptr == n) {
 549         return true;
 550       }
 551     }
 552   }
 553 
 554   return false;
 555 }
 556 
 557 // Check if we are able to untangle the merge. The following patterns are
 558 // supported:
 559 //  - Phi -> SafePoints
 560 //  - Phi -> CmpP/N
 561 //  - Phi -> AddP -> Load
 562 //  - Phi -> CastPP -> SafePoints
 563 //  - Phi -> CastPP -> AddP -> Load
 564 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
 565   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 566     Node* use = n->fast_out(i);
 567 
 568     if (use->is_SafePoint()) {
 569       if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
 570         NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
 571         return false;
 572       } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
 573         NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
 574         return false;
 575       }
 576     } else if (use->is_AddP()) {
 577       Node* addp = use;
 578       for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
 579         Node* use_use = addp->fast_out(j);
 580         const Type* load_type = _igvn->type(use_use);
 581 
 582         if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
 583           NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. AddP user isn't a [splittable] Load(): %s", n->_idx, _invocation, use_use->Name());)
 584           return false;
 585         } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
 586           NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. [Narrow] Klass Load: %s", n->_idx, _invocation, use_use->Name());)
 587           return false;
 588         }
 589       }
 590     } else if (nesting > 0) {
 591       NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Unsupported user %s at nesting level %d.", n->_idx, _invocation, use->Name(), nesting);)
 592       return false;
 593     } else if (use->is_CastPP()) {
 594       const Type* cast_t = _igvn->type(use);
 595       if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
 596 #ifndef PRODUCT
 597         if (TraceReduceAllocationMerges) {
 598           tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
 599           use->dump();
 600         }
 601 #endif
 602         return false;
 603       }
 604 
 605       bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
 606       if (!is_trivial_control) {
 607         // If it's not a trivial control then we check if we can reduce the
 608         // CmpP/N used by the If controlling the cast.
 609         if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
 610           Node* iff = use->in(0)->in(0);
 611           // We may have an OpaqueConstantBool node between If and Bool nodes. But we could also have a sub class of IfNode,
 612           // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
 613           bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
 614           if (can_reduce) {
 615             Node* iff_cmp = iff->in(1)->in(1);
 616             int opc = iff_cmp->Opcode();
 617             can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
 618           }
 619           if (!can_reduce) {
 620 #ifndef PRODUCT
 621             if (TraceReduceAllocationMerges) {
 622               tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
 623               n->dump(5);
 624             }
 625 #endif
 626             return false;
 627           }
 628         }
 629       }
 630 
 631       if (!can_reduce_check_users(use, nesting+1)) {
 632         return false;
 633       }
 634     } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
 635       if (!can_reduce_cmp(n, use)) {
 636         NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. CmpP/N %d isn't reducible.", n->_idx, _invocation, use->_idx);)
 637         return false;
 638       }
 639     } else {
 640       NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. One of the uses is: %d %s", n->_idx, _invocation, use->_idx, use->Name());)
 641       return false;
 642     }
 643   }
 644 
 645   return true;
 646 }
 647 
 648 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
 649 // only used in some certain code shapes. Check comments in
 650 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
 651 // details.
 652 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
 653   // If there was an error attempting to reduce allocation merges for this
 654   // method we might have disabled the compilation and be retrying with RAM
 655   // disabled.
 656   if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
 657     return false;
 658   }
 659 
 660   const Type* phi_t = _igvn->type(ophi);
 661   if (phi_t == nullptr ||
 662       phi_t->make_ptr() == nullptr ||
 663       phi_t->make_ptr()->isa_aryptr() != nullptr) {
 664     return false;
 665   }
 666 
 667   if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
 668     return false;
 669   }
 670 
 671   NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
 672   return true;
 673 }
 674 
 675 // This method will return a CmpP/N that we need to use on the If controlling a
 676 // CastPP after it was split. This method is only called on bases that are
 677 // nullable therefore we always need a controlling if for the splitted CastPP.
 678 //
 679 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
 680 // If the CastPP currently doesn't have a control then the CmpP/N will be
 681 // against the null constant, otherwise it will be against the constant input of
 682 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
 683 // case because we have constraints on it and because the CastPP has a control
 684 // input.
 685 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
 686   const Type* t = base->bottom_type();
 687   Node* con = nullptr;
 688 
 689   if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
 690     con = _igvn->zerocon(t->basic_type());
 691   } else {
 692     // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
 693     assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
 694     Node* bol = curr_ctrl->in(0)->in(1);
 695     assert(bol->is_Bool(), "unexpected node %s", bol->Name());
 696     Node* curr_cmp = bol->in(1);
 697     assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
 698     con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
 699   }
 700 
 701   return CmpNode::make(base, con, t->basic_type());
 702 }
 703 
 704 // This method 'specializes' the CastPP passed as parameter to the base passed
 705 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
 706 // means that the CastPP now will be specific for a given base instead of a Phi.
 707 // An If-Then-Else-Region block is inserted to control the CastPP. The control
 708 // of the CastPP is a copy of the current one (if there is one) or a check
 709 // against null.
 710 //
 711 // Before:
 712 //
 713 //    C1     C2  ... Cn
 714 //     \      |      /
 715 //      \     |     /
 716 //       \    |    /
 717 //        \   |   /
 718 //         \  |  /
 719 //          \ | /
 720 //           \|/
 721 //          Region     B1      B2  ... Bn
 722 //            |          \      |      /
 723 //            |           \     |     /
 724 //            |            \    |    /
 725 //            |             \   |   /
 726 //            |              \  |  /
 727 //            |               \ | /
 728 //            ---------------> Phi
 729 //                              |
 730 //                      X       |
 731 //                      |       |
 732 //                      |       |
 733 //                      ------> CastPP
 734 //
 735 // After (only partial illustration; base = B2, current_control = C2):
 736 //
 737 //                      C2
 738 //                      |
 739 //                      If
 740 //                     / \
 741 //                    /   \
 742 //                   T     F
 743 //                  /\     /
 744 //                 /  \   /
 745 //                /    \ /
 746 //      C1    CastPP   Reg        Cn
 747 //       |              |          |
 748 //       |              |          |
 749 //       |              |          |
 750 //       -------------- | ----------
 751 //                    | | |
 752 //                    Region
 753 //
 754 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
 755   Node* control_successor  = current_control->unique_ctrl_out();
 756   Node* cmp                = _igvn->transform(specialize_cmp(base, castpp->in(0)));
 757   Node* bol                = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
 758   IfNode* if_ne            = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
 759   Node* not_eq_control     = _igvn->transform(new IfTrueNode(if_ne));
 760   Node* yes_eq_control     = _igvn->transform(new IfFalseNode(if_ne));
 761   Node* end_region         = _igvn->transform(new RegionNode(3));
 762 
 763   // Insert the new if-else-region block into the graph
 764   end_region->set_req(1, not_eq_control);
 765   end_region->set_req(2, yes_eq_control);
 766   control_successor->replace_edge(current_control, end_region, _igvn);
 767 
 768   _igvn->_worklist.push(current_control);
 769   _igvn->_worklist.push(control_successor);
 770 
 771   return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr));
 772 }
 773 
 774 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *>  &alloc_worklist) {
 775   const Type* load_type = _igvn->type(curr_load);
 776   Node* nsr_value = _igvn->zerocon(load_type->basic_type());
 777   Node* memory = curr_load->in(MemNode::Memory);
 778 
 779   // The data_phi merging the loads needs to be nullable if
 780   // we are loading pointers.
 781   if (load_type->make_ptr() != nullptr) {
 782     if (load_type->isa_narrowoop()) {
 783       load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
 784     } else if (load_type->isa_ptr()) {
 785       load_type = load_type->meet(TypePtr::NULL_PTR);
 786     } else {
 787       assert(false, "Unexpected load ptr type.");
 788     }
 789   }
 790 
 791   Node* data_phi = PhiNode::make(region, nsr_value, load_type);
 792 
 793   for (int i = 1; i < bases_for_loads->length(); i++) {
 794     Node* base = bases_for_loads->at(i);
 795     Node* cmp_region = nullptr;
 796     if (base != nullptr) {
 797       if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
 798         cmp_region = base->unique_ctrl_out_or_null();
 799         assert(cmp_region != nullptr, "There should be.");
 800         base = base->find_out_with(Op_CastPP);
 801       }
 802 
 803       Node* addr = _igvn->transform(AddPNode::make_with_base(base, curr_addp->in(AddPNode::Offset)));
 804       Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
 805       Node* load = curr_load->clone();
 806       load->set_req(0, nullptr);
 807       load->set_req(1, mem);
 808       load->set_req(2, addr);
 809 
 810       if (cmp_region != nullptr) { // see comment on previous if
 811         Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
 812         intermediate_phi->set_req(1, _igvn->transform(load));
 813         load = intermediate_phi;
 814       }
 815 
 816       data_phi->set_req(i, _igvn->transform(load));
 817     } else {
 818       // Just use the default, which is already in phi
 819     }
 820   }
 821 
 822   // Takes care of updating CG and split_unique_types worklists due
 823   // to cloned AddP->Load.
 824   updates_after_load_split(data_phi, curr_load, alloc_worklist);
 825 
 826   return _igvn->transform(data_phi);
 827 }
 828 
 829 // This method only reduces CastPP fields loads; SafePoints are handled
 830 // separately. The idea here is basically to clone the CastPP and place copies
 831 // on each input of the Phi, including non-scalar replaceable inputs.
 832 // Experimentation shows that the resulting IR graph is simpler that way than if
 833 // we just split the cast through scalar-replaceable inputs.
 834 //
 835 // The reduction process requires that CastPP's control be one of:
 836 //  1) no control,
 837 //  2) the same region as Ophi, or
 838 //  3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
 839 //
 840 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
 841 // flow. If the CastPP originally had an IfTrue/False control input then we'll
 842 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
 843 // juse use a CmpP/N against the null constant.
 844 //
 845 // The If-Then-Else-Region isn't always needed. For instance, if input to
 846 // splitted cast was not nullable (or if it was the null constant) then we don't
 847 // need (shouldn't) use a CastPP at all.
 848 //
 849 // After the casts are splitted we'll split the AddP->Loads through the Phi and
 850 // connect them to the just split CastPPs.
 851 //
 852 // Before (CastPP control is same as Phi):
 853 //
 854 //          Region     Allocate   Null    Call
 855 //            |             \      |      /
 856 //            |              \     |     /
 857 //            |               \    |    /
 858 //            |                \   |   /
 859 //            |                 \  |  /
 860 //            |                  \ | /
 861 //            ------------------> Phi            # Oop Phi
 862 //            |                    |
 863 //            |                    |
 864 //            |                    |
 865 //            |                    |
 866 //            ----------------> CastPP
 867 //                                 |
 868 //                               AddP
 869 //                                 |
 870 //                               Load
 871 //
 872 // After (Very much simplified):
 873 //
 874 //                         Call  Null
 875 //                            \  /
 876 //                            CmpP
 877 //                             |
 878 //                           Bool#NE
 879 //                             |
 880 //                             If
 881 //                            / \
 882 //                           T   F
 883 //                          / \ /
 884 //                         /   R
 885 //                     CastPP  |
 886 //                       |     |
 887 //                     AddP    |
 888 //                       |     |
 889 //                     Load    |
 890 //                         \   |   0
 891 //            Allocate      \  |  /
 892 //                \          \ | /
 893 //               AddP         Phi
 894 //                  \         /
 895 //                 Load      /
 896 //                    \  0  /
 897 //                     \ | /
 898 //                      \|/
 899 //                      Phi        # "Field" Phi
 900 //
 901 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node*> &alloc_worklist) {
 902   Node* ophi = curr_castpp->in(1);
 903   assert(ophi->is_Phi(), "Expected this to be a Phi node.");
 904 
 905   // Identify which base should be used for AddP->Load later when spliting the
 906   // CastPP->Loads through ophi. Three kind of values may be stored in this
 907   // array, depending on the nullability status of the corresponding input in
 908   // ophi.
 909   //
 910   //  - nullptr:    Meaning that the base is actually the null constant and therefore
 911   //                we won't try to load from it.
 912   //
 913   //  - CFG Node:   Meaning that the base is a CastPP that was specialized for
 914   //                this input of Ophi. I.e., we added an If->Then->Else-Region
 915   //                that will 'activate' the CastPp only when the input is not Null.
 916   //
 917   //  - Other Node: Meaning that the base is not nullable and therefore we'll try
 918   //                to load directly from it.
 919   GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
 920 
 921   for (uint i = 1; i < ophi->req(); i++) {
 922     Node* base = ophi->in(i);
 923     const Type* base_t = _igvn->type(base);
 924 
 925     if (base_t->maybe_null()) {
 926       if (base->is_Con()) {
 927         // Nothing todo as bases_for_loads[i] is already null
 928       } else {
 929         Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
 930         bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
 931       }
 932     } else {
 933       bases_for_loads.at_put(i, base);
 934     }
 935   }
 936 
 937   // Now let's split the CastPP->Loads through the Phi
 938   for (int i = curr_castpp->outcnt()-1; i >= 0;) {
 939     Node* use = curr_castpp->raw_out(i);
 940     if (use->is_AddP()) {
 941       for (int j = use->outcnt()-1; j >= 0;) {
 942         Node* use_use = use->raw_out(j);
 943         assert(use_use->is_Load(), "Expected this to be a Load node.");
 944 
 945         // We can't make an unconditional load from a nullable input. The
 946         // 'split_castpp_load_through_phi` method will add an
 947         // 'If-Then-Else-Region` around nullable bases and only load from them
 948         // when the input is not null.
 949         Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
 950         _igvn->replace_node(use_use, phi);
 951 
 952         --j;
 953         j = MIN2(j, (int)use->outcnt()-1);
 954       }
 955 
 956       _igvn->remove_dead_node(use, PhaseIterGVN::NodeOrigin::Graph);
 957     }
 958     --i;
 959     i = MIN2(i, (int)curr_castpp->outcnt()-1);
 960   }
 961 }
 962 
 963 // This method split a given CmpP/N through the Phi used in one of its inputs.
 964 // As a result we convert a comparison with a pointer to a comparison with an
 965 // integer.
 966 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
 967 // while the other must be a constant.
 968 // The splitting process is basically just cloning the CmpP/N above the input
 969 // Phi.  However, some (most) of the cloned CmpP/Ns won't be requred because we
 970 // can prove at compile time the result of the comparison.
 971 //
 972 // Before:
 973 //
 974 //             in1    in2 ... inN
 975 //              \      |      /
 976 //               \     |     /
 977 //                \    |    /
 978 //                 \   |   /
 979 //                  \  |  /
 980 //                   \ | /
 981 //                    Phi
 982 //                     |   Other
 983 //                     |    /
 984 //                     |   /
 985 //                     |  /
 986 //                    CmpP/N
 987 //
 988 // After:
 989 //
 990 //        in1  Other   in2 Other  inN  Other
 991 //         |    |      |   |      |    |
 992 //         \    |      |   |      |    |
 993 //          \  /       |   /      |    /
 994 //          CmpP/N    CmpP/N     CmpP/N
 995 //          Bool      Bool       Bool
 996 //            \        |        /
 997 //             \       |       /
 998 //              \      |      /
 999 //               \     |     /
1000 //                \    |    /
1001 //                 \   |   /
1002 //                  \  |  /
1003 //                   \ | /
1004 //                    Phi
1005 //                     |
1006 //                     |   Zero
1007 //                     |    /
1008 //                     |   /
1009 //                     |  /
1010 //                     CmpI
1011 //
1012 //
1013 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
1014   Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
1015   assert(ophi->is_Phi(), "Expected this to be a Phi node.");
1016 
1017   Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
1018   Node* zero = _igvn->intcon(0);
1019   Node* one = _igvn->intcon(1);
1020   BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
1021 
1022   // This Phi will merge the result of the Cmps split through the Phi
1023   Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
1024 
1025   for (uint i=1; i<ophi->req(); i++) {
1026     Node* ophi_input = ophi->in(i);
1027     Node* res_phi_input = nullptr;
1028 
1029     const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1030     if (tcmp->singleton()) {
1031       if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
1032           (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
1033         res_phi_input = one;
1034       } else {
1035         res_phi_input = zero;
1036       }
1037     } else {
1038       Node* ncmp = _igvn->transform(cmp->clone());
1039       ncmp->set_req(1, ophi_input);
1040       ncmp->set_req(2, other);
1041       Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1042       res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1043     }
1044 
1045     res_phi->set_req(i, res_phi_input);
1046   }
1047 
1048   // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1049   Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1050   _igvn->replace_node(cmp, new_cmp);
1051 }
1052 
1053 // Push the newly created AddP on alloc_worklist and patch
1054 // the connection graph. Note that the changes in the CG below
1055 // won't affect the ES of objects since the new nodes have the
1056 // same status as the old ones.
1057 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *>  &alloc_worklist) {
1058   assert(data_phi != nullptr, "Output of split_through_phi is null.");
1059   assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1060   assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1061 
1062   if (data_phi == nullptr || !data_phi->is_Phi()) {
1063     // Make this a retry?
1064     return ;
1065   }
1066 
1067   Node* previous_addp = previous_load->in(MemNode::Address);
1068   FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1069   for (uint i = 1; i < data_phi->req(); i++) {
1070     Node* new_load = data_phi->in(i);
1071 
1072     if (new_load->is_Phi()) {
1073       // new_load is currently the "intermediate_phi" from an specialized
1074       // CastPP.
1075       new_load = new_load->in(1);
1076     }
1077 
1078     // "new_load" might actually be a constant, parameter, etc.
1079     if (new_load->is_Load()) {
1080       Node* new_addp = new_load->in(MemNode::Address);
1081 
1082       // If new_load is a Load but not from an AddP, it means that the load is folded into another
1083       // load. And since this load is not from a field, we cannot create a unique type for it.
1084       // For example:
1085       //
1086       //   if (b) {
1087       //       Holder h1 = new Holder();
1088       //       Object o = ...;
1089       //       h.o = o.getClass();
1090       //   } else {
1091       //       Holder h2 = ...;
1092       //   }
1093       //   Holder h = Phi(h1, h2);
1094       //   Object r = h.o;
1095       //
1096       // Then, splitting r through the merge point results in:
1097       //
1098       //   if (b) {
1099       //       Holder h1 = new Holder();
1100       //       Object o = ...;
1101       //       h.o = o.getClass();
1102       //       Object o1 = h.o;
1103       //   } else {
1104       //       Holder h2 = ...;
1105       //       Object o2 = h2.o;
1106       //   }
1107       //   Object r = Phi(o1, o2);
1108       //
1109       // In this case, o1 is folded to o.getClass() which is a Load but not from an AddP, but from
1110       // an OopHandle that is loaded from the Klass of o.
1111       if (!new_addp->is_AddP()) {
1112         continue;
1113       }
1114       Node* base = get_addp_base(new_addp);
1115 
1116       // The base might not be something that we can create an unique
1117       // type for. If that's the case we are done with that input.
1118       PointsToNode* jobj_ptn = unique_java_object(base);
1119       if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1120         continue;
1121       }
1122 
1123       // Push to alloc_worklist since the base has an unique_type
1124       alloc_worklist.append_if_missing(new_addp);
1125 
1126       // Now let's add the node to the connection graph
1127       _nodes.at_grow(new_addp->_idx, nullptr);
1128       add_field(new_addp, fn->escape_state(), fn->offset());
1129       add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1130 
1131       // If the load doesn't load an object then it won't be
1132       // part of the connection graph
1133       PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1134       if (curr_load_ptn != nullptr) {
1135         _nodes.at_grow(new_load->_idx, nullptr);
1136         add_local_var(new_load, curr_load_ptn->escape_state());
1137         add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1138       }
1139     }
1140   }
1141 }
1142 
1143 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *>  &alloc_worklist) {
1144   // We'll pass this to 'split_through_phi' so that it'll do the split even
1145   // though the load doesn't have an unique instance type.
1146   bool ignore_missing_instance_id = true;
1147 
1148   // All AddPs are present in the connection graph
1149   FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1150 
1151   // Iterate over AddP looking for a Load
1152   for (int k = previous_addp->outcnt()-1; k >= 0;) {
1153     Node* previous_load = previous_addp->raw_out(k);
1154     if (previous_load->is_Load()) {
1155       Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1156 
1157       // Takes care of updating CG and split_unique_types worklists due to cloned
1158       // AddP->Load.
1159       updates_after_load_split(data_phi, previous_load, alloc_worklist);
1160 
1161       _igvn->replace_node(previous_load, data_phi);
1162     }
1163     --k;
1164     k = MIN2(k, (int)previous_addp->outcnt()-1);
1165   }
1166 
1167   // Remove the old AddP from the processing list because it's dead now
1168   assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1169   alloc_worklist.remove_if_existing(previous_addp);
1170 }
1171 
1172 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1173 // selector is:
1174 //    -> a '-1' constant, the i'th input of the original Phi is NSR.
1175 //    -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1176 //       the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1177 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1178   Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1179   Node* selector  = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1180   uint number_of_sr_objects = 0;
1181   for (uint i = 1; i < ophi->req(); i++) {
1182     Node* base = ophi->in(i);
1183     JavaObjectNode* ptn = unique_java_object(base);
1184 
1185     if (ptn != nullptr && ptn->scalar_replaceable()) {
1186       Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1187       selector->set_req(i, sr_obj_idx);
1188       number_of_sr_objects++;
1189     }
1190   }
1191 
1192   return selector->as_Phi();
1193 }
1194 
1195 // Returns true if the AddP node 'n' has at least one base that is a reducible
1196 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1197 // checked instead.
1198 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1199   PointsToNode* ptn = ptnode_adr(n->_idx);
1200   if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1201     return false;
1202   }
1203 
1204   for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1205     Node* base = i.get()->ideal_node();
1206 
1207     if (reducible_merges.member(base)) {
1208       return true;
1209     }
1210 
1211     if (base->is_CastPP() || base->is_CheckCastPP()) {
1212       base = base->in(1);
1213       if (reducible_merges.member(base)) {
1214         return true;
1215       }
1216     }
1217   }
1218 
1219   return false;
1220 }
1221 
1222 // This method will call its helper method to reduce SafePoint nodes that use
1223 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1224 // "version" of Phi use the same debug information (regarding the Phi).
1225 // Therefore, I collect all safepoints and patch them all at once.
1226 //
1227 // The safepoints using the Phi node have to be processed before safepoints of
1228 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1229 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1230 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1231 // algorithm that process Phi's safepoints will think that the added Phi
1232 // reference is a regular reference.
1233 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1234   PhiNode* selector = create_selector(ophi);
1235   Unique_Node_List safepoints;
1236   Unique_Node_List casts;
1237 
1238   // Just collect the users of the Phis for later processing
1239   // in the needed order.
1240   for (uint i = 0; i < ophi->outcnt(); i++) {
1241     Node* use = ophi->raw_out(i);
1242     if (use->is_SafePoint()) {
1243       safepoints.push(use);
1244     } else if (use->is_CastPP()) {
1245       casts.push(use);
1246     } else {
1247       assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1248     }
1249   }
1250 
1251   // Need to process safepoints using the Phi first
1252   if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1253     return false;
1254   }
1255 
1256   // Now process CastPP->safepoints
1257   for (uint i = 0; i < casts.size(); i++) {
1258     Node* cast = casts.at(i);
1259     Unique_Node_List cast_sfpts;
1260 
1261     for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1262       Node* use_use = cast->fast_out(j);
1263       if (use_use->is_SafePoint()) {
1264         cast_sfpts.push(use_use);
1265       } else {
1266         assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1267       }
1268     }
1269 
1270     if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1271       return false;
1272     }
1273   }
1274 
1275   return true;
1276 }
1277 
1278 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1279 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1280 // SafePointScalarObjectNode for each scalar replaceable input. Each
1281 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1282 // check detailed description in SafePointScalarMergeNode class header.
1283 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1284   PhaseMacroExpand mexp(*_igvn);
1285   Node* original_sfpt_parent =  cast != nullptr ? cast : ophi;
1286   const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1287 
1288   Node* nsr_merge_pointer = ophi;
1289   if (cast != nullptr) {
1290     const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1291     nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::DependencyType::FloatingNarrowing, nullptr));
1292   }
1293 
1294   for (uint spi = 0; spi < safepoints.size(); spi++) {
1295     SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1296 
1297     SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(*_igvn);
1298 
1299     // All sfpt inputs are implicitly included into debug info during the scalarization process below.
1300     // Keep non-debug inputs separately, so they stay non-debug.
1301     sfpt->remove_non_debug_edges(non_debug_edges_worklist);
1302 
1303     JVMState* jvms  = sfpt->jvms();
1304     uint merge_idx  = (sfpt->req() - jvms->scloff());
1305     int debug_start = jvms->debug_start();
1306 
1307     SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1308     smerge->init_req(0, _compile->root());
1309     _igvn->register_new_node_with_optimizer(smerge);
1310 
1311     assert(sfpt->jvms()->endoff() == sfpt->req(), "no extra edges past debug info allowed");
1312 
1313     // The next two inputs are:
1314     //  (1) A copy of the original pointer to NSR objects.
1315     //  (2) A selector, used to decide if we need to rematerialize an object
1316     //      or use the pointer to a NSR object.
1317     // See more details of these fields in the declaration of SafePointScalarMergeNode.
1318     // It is safe to include them into debug info straight away since create_scalarized_object_description()
1319     // will include all newly added inputs into debug info anyway.
1320     sfpt->add_req(nsr_merge_pointer);
1321     sfpt->add_req(selector);
1322     sfpt->jvms()->set_endoff(sfpt->req());
1323 
1324     for (uint i = 1; i < ophi->req(); i++) {
1325       Node* base = ophi->in(i);
1326       JavaObjectNode* ptn = unique_java_object(base);
1327 
1328       // If the base is not scalar replaceable we don't need to register information about
1329       // it at this time.
1330       if (ptn == nullptr || !ptn->scalar_replaceable()) {
1331         continue;
1332       }
1333 
1334       AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1335       Unique_Node_List value_worklist;
1336 #ifdef ASSERT
1337       const Type* res_type = alloc->result_cast()->bottom_type();
1338       if (res_type->is_inlinetypeptr() && !Compile::current()->has_circular_inline_type()) {
1339         assert(!ophi->as_Phi()->can_push_inline_types_down(_igvn), "missed earlier scalarization opportunity");
1340       }
1341 #endif
1342       SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt, &value_worklist);
1343       if (sobj == nullptr) {
1344         _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1345         sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1346         return false; // non-recoverable failure; recompile
1347       }
1348 
1349       // Now make a pass over the debug information replacing any references
1350       // to the allocated object with "sobj"
1351       Node* ccpp = alloc->result_cast();
1352       sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1353       non_debug_edges_worklist.remove_edge_if_present(ccpp); // drop scalarized input from non-debug info
1354 
1355       // Register the scalarized object as a candidate for reallocation
1356       smerge->add_req(sobj);
1357 
1358       // Scalarize inline types that were added to the safepoint.
1359       // Don't allow linking a constant oop (if available) for flat array elements
1360       // because Deoptimization::reassign_flat_array_elements needs field values.
1361       const bool allow_oop = !merge_t->is_flat();
1362       for (uint j = 0; j < value_worklist.size(); ++j) {
1363         InlineTypeNode* vt = value_worklist.at(j)->as_InlineType();
1364         vt->make_scalar_in_safepoints(_igvn, allow_oop);
1365       }
1366     }
1367 
1368     // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1369     sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1370     non_debug_edges_worklist.remove_edge_if_present(original_sfpt_parent); // drop scalarized input from non-debug info
1371 
1372     // The call to 'replace_edges_in_range' above might have removed the
1373     // reference to ophi that we need at _merge_pointer_idx. The line below make
1374     // sure the reference is maintained.
1375     sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1376 
1377     sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1378 
1379     _igvn->_worklist.push(sfpt);
1380   }
1381 
1382   return true;
1383 }
1384 
1385 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1386   bool delay = _igvn->delay_transform();
1387   _igvn->set_delay_transform(true);
1388   _igvn->hash_delete(ophi);
1389 
1390   // Copying all users first because some will be removed and others won't.
1391   // Ophi also may acquire some new users as part of Cast reduction.
1392   // CastPPs also need to be processed before CmpPs.
1393   Unique_Node_List castpps;
1394   Unique_Node_List others;
1395   for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1396     Node* use = ophi->fast_out(i);
1397 
1398     if (use->is_CastPP()) {
1399       castpps.push(use);
1400     } else if (use->is_AddP() || use->is_Cmp()) {
1401       others.push(use);
1402     } else {
1403       // Safepoints to be processed later; other users aren't expected here
1404       assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1405     }
1406   }
1407 
1408   _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1409 
1410   // CastPPs need to be processed before Cmps because during the process of
1411   // splitting CastPPs we make reference to the inputs of the Cmp that is used
1412   // by the If controlling the CastPP.
1413   for (uint i = 0; i < castpps.size(); i++) {
1414     reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1415     _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1416   }
1417 
1418   for (uint i = 0; i < others.size(); i++) {
1419     Node* use = others.at(i);
1420 
1421     if (use->is_AddP()) {
1422       reduce_phi_on_field_access(use, alloc_worklist);
1423       _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1424     } else if(use->is_Cmp()) {
1425       reduce_phi_on_cmp(use);
1426       _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1427     }
1428   }
1429 
1430   _igvn->set_delay_transform(delay);
1431 }
1432 
1433 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1434   Node* null_ptr            = _igvn->makecon(TypePtr::NULL_PTR);
1435   const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1436   const Type* new_t         = merge_t->meet(TypePtr::NULL_PTR);
1437   Node* new_phi             = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1438 
1439   for (uint i = 1; i < ophi->req(); i++) {
1440     Node* base          = ophi->in(i);
1441     JavaObjectNode* ptn = unique_java_object(base);
1442 
1443     if (ptn != nullptr && ptn->scalar_replaceable()) {
1444       new_phi->set_req(i, null_ptr);
1445     } else {
1446       new_phi->set_req(i, ophi->in(i));
1447     }
1448   }
1449 
1450   for (int i = ophi->outcnt()-1; i >= 0;) {
1451     Node* out = ophi->raw_out(i);
1452 
1453     if (out->is_ConstraintCast()) {
1454       const Type* out_t = _igvn->type(out)->make_ptr();
1455       const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1456       bool change = out_new_t != out_t;
1457 
1458       for (int j = out->outcnt()-1; change && j >= 0; --j) {
1459         Node* out2 = out->raw_out(j);
1460         if (!out2->is_SafePoint()) {
1461           change = false;
1462           break;
1463         }
1464       }
1465 
1466       if (change) {
1467         Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::DependencyType::NonFloatingNarrowing, nullptr);
1468         _igvn->replace_node(out, new_cast);
1469         _igvn->register_new_node_with_optimizer(new_cast);
1470       }
1471     }
1472 
1473     --i;
1474     i = MIN2(i, (int)ophi->outcnt()-1);
1475   }
1476 
1477   _igvn->replace_node(ophi, new_phi);
1478 }
1479 
1480 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1481   if (!C->do_reduce_allocation_merges()) return;
1482 
1483   Unique_Node_List ideal_nodes;
1484   ideal_nodes.map(C->live_nodes(), nullptr);  // preallocate space
1485   ideal_nodes.push(root);
1486 
1487   for (uint next = 0; next < ideal_nodes.size(); ++next) {
1488     Node* n = ideal_nodes.at(next);
1489 
1490     if (n->is_SafePointScalarMerge()) {
1491       SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1492 
1493       // Validate inputs of merge
1494       for (uint i = 1; i < merge->req(); i++) {
1495         if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1496           assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1497           C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1498         }
1499       }
1500 
1501       // Validate users of merge
1502       for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1503         Node* sfpt = merge->fast_out(i);
1504         if (sfpt->is_SafePoint()) {
1505           int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1506 
1507           if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1508             assert(false, "SafePointScalarMerge nodes can't be nested.");
1509             C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1510           }
1511         } else {
1512           assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1513           C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1514         }
1515       }
1516     }
1517 
1518     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1519       Node* m = n->fast_out(i);
1520       ideal_nodes.push(m);
1521     }
1522   }
1523 }
1524 
1525 // Returns true if there is an object in the scope of sfn that does not escape globally.
1526 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1527   Compile* C = _compile;
1528   for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1529     if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1530         DeoptimizeObjectsALot) {
1531       // Jvmti agents can access locals. Must provide info about local objects at runtime.
1532       int num_locs = jvms->loc_size();
1533       for (int idx = 0; idx < num_locs; idx++) {
1534         Node* l = sfn->local(jvms, idx);
1535         if (not_global_escape(l)) {
1536           return true;
1537         }
1538       }
1539     }
1540     if (C->env()->jvmti_can_get_owned_monitor_info() ||
1541         C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1542       // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1543       int num_mon = jvms->nof_monitors();
1544       for (int idx = 0; idx < num_mon; idx++) {
1545         Node* m = sfn->monitor_obj(jvms, idx);
1546         if (m != nullptr && not_global_escape(m)) {
1547           return true;
1548         }
1549       }
1550     }
1551   }
1552   return false;
1553 }
1554 
1555 // Returns true if at least one of the arguments to the call is an object
1556 // that does not escape globally.
1557 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1558   if (call->method() != nullptr) {
1559     uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1560     for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1561       Node* p = call->in(idx);
1562       if (not_global_escape(p)) {
1563         return true;
1564       }
1565     }
1566   } else {
1567     const char* name = call->as_CallStaticJava()->_name;
1568     assert(name != nullptr, "no name");
1569     // no arg escapes through uncommon traps
1570     if (strcmp(name, "uncommon_trap") != 0) {
1571       // process_call_arguments() assumes that all arguments escape globally
1572       const TypeTuple* d = call->tf()->domain_sig();
1573       for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1574         const Type* at = d->field_at(i);
1575         if (at->isa_oopptr() != nullptr) {
1576           return true;
1577         }
1578       }
1579     }
1580   }
1581   return false;
1582 }
1583 
1584 
1585 
1586 // Utility function for nodes that load an object
1587 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1588   // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1589   // ThreadLocal has RawPtr type.
1590   const Type* t = _igvn->type(n);
1591   if (t->make_ptr() != nullptr) {
1592     Node* adr = n->in(MemNode::Address);
1593 #ifdef ASSERT
1594     if (!adr->is_AddP()) {
1595       assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1596     } else {
1597       assert((ptnode_adr(adr->_idx) == nullptr ||
1598               ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1599     }
1600 #endif
1601     add_local_var_and_edge(n, PointsToNode::NoEscape,
1602                            adr, delayed_worklist);
1603   }
1604 }
1605 
1606 void ConnectionGraph::add_proj(Node* n, Unique_Node_List* delayed_worklist) {
1607   if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && n->in(0)->as_Call()->returns_pointer()) {
1608     add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1609   } else if (n->in(0)->is_LoadFlat()) {
1610     // Treat LoadFlat outputs similar to a call return value
1611     add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1612   } else if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() && n->bottom_type()->isa_ptr()) {
1613     CallNode* call = n->in(0)->as_Call();
1614     assert(call->tf()->returns_inline_type_as_fields(), "");
1615     if (n->as_Proj()->_con == TypeFunc::Parms || !returns_an_argument(call)) {
1616       // either:
1617       // - not an argument returned
1618       // - the returned buffer for a returned scalarized argument
1619       add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1620     } else {
1621       add_local_var(n, PointsToNode::NoEscape);
1622     }
1623   }
1624 }
1625 
1626 // Populate Connection Graph with PointsTo nodes and create simple
1627 // connection graph edges.
1628 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1629   assert(!_verify, "this method should not be called for verification");
1630   PhaseGVN* igvn = _igvn;
1631   uint n_idx = n->_idx;
1632   PointsToNode* n_ptn = ptnode_adr(n_idx);
1633   if (n_ptn != nullptr) {
1634     return; // No need to redefine PointsTo node during first iteration.
1635   }
1636   int opcode = n->Opcode();
1637   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1638   if (gc_handled) {
1639     return; // Ignore node if already handled by GC.
1640   }
1641 
1642   if (n->is_Call()) {
1643     // Arguments to allocation and locking don't escape.
1644     if (n->is_AbstractLock()) {
1645       // Put Lock and Unlock nodes on IGVN worklist to process them during
1646       // first IGVN optimization when escape information is still available.
1647       record_for_optimizer(n);
1648     } else if (n->is_Allocate()) {
1649       add_call_node(n->as_Call());
1650       record_for_optimizer(n);
1651     } else {
1652       if (n->is_CallStaticJava()) {
1653         const char* name = n->as_CallStaticJava()->_name;
1654         if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1655           return; // Skip uncommon traps
1656         }
1657       }
1658       // Don't mark as processed since call's arguments have to be processed.
1659       delayed_worklist->push(n);
1660       // Check if a call returns an object.
1661       if ((n->as_Call()->returns_pointer() &&
1662            n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1663           (n->is_CallStaticJava() &&
1664            n->as_CallStaticJava()->is_boxing_method())) {
1665         add_call_node(n->as_Call());
1666       } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1667         bool returns_oop = false;
1668         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1669           ProjNode* pn = n->fast_out(i)->as_Proj();
1670           if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1671             returns_oop = true;
1672           }
1673         }
1674         if (returns_oop) {
1675           add_call_node(n->as_Call());
1676         }
1677       }
1678     }
1679     return;
1680   }
1681   // Put this check here to process call arguments since some call nodes
1682   // point to phantom_obj.
1683   if (n_ptn == phantom_obj || n_ptn == null_obj) {
1684     return; // Skip predefined nodes.
1685   }
1686   switch (opcode) {
1687     case Op_AddP: {
1688       Node* base = get_addp_base(n);
1689       PointsToNode* ptn_base = ptnode_adr(base->_idx);
1690       // Field nodes are created for all field types. They are used in
1691       // adjust_scalar_replaceable_state() and split_unique_types().
1692       // Note, non-oop fields will have only base edges in Connection
1693       // Graph because such fields are not used for oop loads and stores.
1694       int offset = address_offset(n, igvn);
1695       add_field(n, PointsToNode::NoEscape, offset);
1696       if (ptn_base == nullptr) {
1697         delayed_worklist->push(n); // Process it later.
1698       } else {
1699         n_ptn = ptnode_adr(n_idx);
1700         add_base(n_ptn->as_Field(), ptn_base);
1701       }
1702       break;
1703     }
1704     case Op_CastX2P:
1705     case Op_CastI2N: {
1706       map_ideal_node(n, phantom_obj);
1707       break;
1708     }
1709     case Op_InlineType:
1710     case Op_CastPP:
1711     case Op_CheckCastPP:
1712     case Op_EncodeP:
1713     case Op_DecodeN:
1714     case Op_EncodePKlass:
1715     case Op_DecodeNKlass: {
1716       add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1717       break;
1718     }
1719     case Op_CMoveP: {
1720       add_local_var(n, PointsToNode::NoEscape);
1721       // Do not add edges during first iteration because some could be
1722       // not defined yet.
1723       delayed_worklist->push(n);
1724       break;
1725     }
1726     case Op_ConP:
1727     case Op_ConN:
1728     case Op_ConNKlass: {
1729       // assume all oop constants globally escape except for null
1730       PointsToNode::EscapeState es;
1731       const Type* t = igvn->type(n);
1732       if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1733         es = PointsToNode::NoEscape;
1734       } else {
1735         es = PointsToNode::GlobalEscape;
1736       }
1737       PointsToNode* ptn_con = add_java_object(n, es);
1738       set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1739       break;
1740     }
1741     case Op_CreateEx: {
1742       // assume that all exception objects globally escape
1743       map_ideal_node(n, phantom_obj);
1744       break;
1745     }
1746     case Op_LoadKlass:
1747     case Op_LoadNKlass: {
1748       // Unknown class is loaded
1749       map_ideal_node(n, phantom_obj);
1750       break;
1751     }
1752     case Op_LoadP:
1753     case Op_LoadN: {
1754       add_objload_to_connection_graph(n, delayed_worklist);
1755       break;
1756     }
1757     case Op_Parm: {
1758       map_ideal_node(n, phantom_obj);
1759       break;
1760     }
1761     case Op_PartialSubtypeCheck: {
1762       // Produces Null or notNull and is used in only in CmpP so
1763       // phantom_obj could be used.
1764       map_ideal_node(n, phantom_obj); // Result is unknown
1765       break;
1766     }
1767     case Op_Phi: {
1768       // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1769       // ThreadLocal has RawPtr type.
1770       const Type* t = n->as_Phi()->type();
1771       if (t->make_ptr() != nullptr) {
1772         add_local_var(n, PointsToNode::NoEscape);
1773         // Do not add edges during first iteration because some could be
1774         // not defined yet.
1775         delayed_worklist->push(n);
1776       }
1777       break;
1778     }
1779     case Op_LoadFlat:
1780       // Treat LoadFlat similar to an unknown call that receives nothing and produces its results
1781       map_ideal_node(n, phantom_obj);
1782       break;
1783     case Op_StoreFlat:
1784       // Treat StoreFlat similar to a call that escapes the stored flattened fields
1785       delayed_worklist->push(n);
1786       break;
1787     case Op_Proj: {
1788       // we are only interested in the oop result projection from a call
1789       add_proj(n, delayed_worklist);
1790       break;
1791     }
1792     case Op_Rethrow: // Exception object escapes
1793     case Op_Return: {
1794       if (n->req() > TypeFunc::Parms &&
1795           igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1796         // Treat Return value as LocalVar with GlobalEscape escape state.
1797         add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1798       }
1799       break;
1800     }
1801     case Op_CompareAndExchangeP:
1802     case Op_CompareAndExchangeN:
1803     case Op_GetAndSetP:
1804     case Op_GetAndSetN: {
1805       add_objload_to_connection_graph(n, delayed_worklist);
1806       // fall-through
1807     }
1808     case Op_StoreP:
1809     case Op_StoreN:
1810     case Op_StoreNKlass:
1811     case Op_WeakCompareAndSwapP:
1812     case Op_WeakCompareAndSwapN:
1813     case Op_CompareAndSwapP:
1814     case Op_CompareAndSwapN: {
1815       add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1816       break;
1817     }
1818     case Op_AryEq:
1819     case Op_CountPositives:
1820     case Op_StrComp:
1821     case Op_StrEquals:
1822     case Op_StrIndexOf:
1823     case Op_StrIndexOfChar:
1824     case Op_StrInflatedCopy:
1825     case Op_StrCompressedCopy:
1826     case Op_VectorizedHashCode:
1827     case Op_EncodeISOArray: {
1828       add_local_var(n, PointsToNode::ArgEscape);
1829       delayed_worklist->push(n); // Process it later.
1830       break;
1831     }
1832     case Op_ThreadLocal: {
1833       PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1834       set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1835       break;
1836     }
1837     case Op_Blackhole: {
1838       // All blackhole pointer arguments are globally escaping.
1839       // Only do this if there is at least one pointer argument.
1840       // Do not add edges during first iteration because some could be
1841       // not defined yet, defer to final step.
1842       for (uint i = 0; i < n->req(); i++) {
1843         Node* in = n->in(i);
1844         if (in != nullptr) {
1845           const Type* at = _igvn->type(in);
1846           if (!at->isa_ptr()) continue;
1847 
1848           add_local_var(n, PointsToNode::GlobalEscape);
1849           delayed_worklist->push(n);
1850           break;
1851         }
1852       }
1853       break;
1854     }
1855     default:
1856       ; // Do nothing for nodes not related to EA.
1857   }
1858   return;
1859 }
1860 
1861 // Add final simple edges to graph.
1862 void ConnectionGraph::add_final_edges(Node *n) {
1863   PointsToNode* n_ptn = ptnode_adr(n->_idx);
1864 #ifdef ASSERT
1865   if (_verify && n_ptn->is_JavaObject())
1866     return; // This method does not change graph for JavaObject.
1867 #endif
1868 
1869   if (n->is_Call()) {
1870     process_call_arguments(n->as_Call());
1871     return;
1872   }
1873   assert(n->is_Store() || n->is_LoadStore() || n->is_StoreFlat() ||
1874          ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1875          "node should be registered already");
1876   int opcode = n->Opcode();
1877   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1878   if (gc_handled) {
1879     return; // Ignore node if already handled by GC.
1880   }
1881   switch (opcode) {
1882     case Op_AddP: {
1883       Node* base = get_addp_base(n);
1884       PointsToNode* ptn_base = ptnode_adr(base->_idx);
1885       assert(ptn_base != nullptr, "field's base should be registered");
1886       add_base(n_ptn->as_Field(), ptn_base);
1887       break;
1888     }
1889     case Op_InlineType:
1890     case Op_CastPP:
1891     case Op_CheckCastPP:
1892     case Op_EncodeP:
1893     case Op_DecodeN:
1894     case Op_EncodePKlass:
1895     case Op_DecodeNKlass: {
1896       add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1897       break;
1898     }
1899     case Op_CMoveP: {
1900       for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1901         Node* in = n->in(i);
1902         if (in == nullptr) {
1903           continue;  // ignore null
1904         }
1905         Node* uncast_in = in->uncast();
1906         if (uncast_in->is_top() || uncast_in == n) {
1907           continue;  // ignore top or inputs which go back this node
1908         }
1909         PointsToNode* ptn = ptnode_adr(in->_idx);
1910         assert(ptn != nullptr, "node should be registered");
1911         add_edge(n_ptn, ptn);
1912       }
1913       break;
1914     }
1915     case Op_LoadP:
1916     case Op_LoadN: {
1917       // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1918       // ThreadLocal has RawPtr type.
1919       assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1920       add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1921       break;
1922     }
1923     case Op_Phi: {
1924       // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1925       // ThreadLocal has RawPtr type.
1926       assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1927       for (uint i = 1; i < n->req(); i++) {
1928         Node* in = n->in(i);
1929         if (in == nullptr) {
1930           continue;  // ignore null
1931         }
1932         Node* uncast_in = in->uncast();
1933         if (uncast_in->is_top() || uncast_in == n) {
1934           continue;  // ignore top or inputs which go back this node
1935         }
1936         PointsToNode* ptn = ptnode_adr(in->_idx);
1937         assert(ptn != nullptr, "node should be registered");
1938         add_edge(n_ptn, ptn);
1939       }
1940       break;
1941     }
1942     case Op_StoreFlat: {
1943       // StoreFlat globally escapes its stored flattened fields
1944       InlineTypeNode* value = n->as_StoreFlat()->value();
1945       ciInlineKlass* vk = _igvn->type(value)->inline_klass();
1946       for (int i = 0; i < vk->nof_nonstatic_fields(); i++) {
1947         ciField* field = vk->nonstatic_field_at(i);
1948         if (field->type()->is_primitive_type()) {
1949           continue;
1950         }
1951 
1952         Node* field_value = value->field_value_by_offset(field->offset_in_bytes(), true);
1953         PointsToNode* field_value_ptn = ptnode_adr(field_value->_idx);
1954         set_escape_state(field_value_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "store into a flat field"));
1955       }
1956       break;
1957     }
1958     case Op_Proj: {
1959       add_proj(n, nullptr);
1960       break;
1961     }
1962     case Op_Rethrow: // Exception object escapes
1963     case Op_Return: {
1964       assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1965              "Unexpected node type");
1966       // Treat Return value as LocalVar with GlobalEscape escape state.
1967       add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1968       break;
1969     }
1970     case Op_CompareAndExchangeP:
1971     case Op_CompareAndExchangeN:
1972     case Op_GetAndSetP:
1973     case Op_GetAndSetN:{
1974       assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1975       add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1976       // fall-through
1977     }
1978     case Op_CompareAndSwapP:
1979     case Op_CompareAndSwapN:
1980     case Op_WeakCompareAndSwapP:
1981     case Op_WeakCompareAndSwapN:
1982     case Op_StoreP:
1983     case Op_StoreN:
1984     case Op_StoreNKlass:{
1985       add_final_edges_unsafe_access(n, opcode);
1986       break;
1987     }
1988     case Op_VectorizedHashCode:
1989     case Op_AryEq:
1990     case Op_CountPositives:
1991     case Op_StrComp:
1992     case Op_StrEquals:
1993     case Op_StrIndexOf:
1994     case Op_StrIndexOfChar:
1995     case Op_StrInflatedCopy:
1996     case Op_StrCompressedCopy:
1997     case Op_EncodeISOArray: {
1998       // char[]/byte[] arrays passed to string intrinsic do not escape but
1999       // they are not scalar replaceable. Adjust escape state for them.
2000       // Start from in(2) edge since in(1) is memory edge.
2001       for (uint i = 2; i < n->req(); i++) {
2002         Node* adr = n->in(i);
2003         const Type* at = _igvn->type(adr);
2004         if (!adr->is_top() && at->isa_ptr()) {
2005           assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2006                  at->isa_ptr() != nullptr, "expecting a pointer");
2007           if (adr->is_AddP()) {
2008             adr = get_addp_base(adr);
2009           }
2010           PointsToNode* ptn = ptnode_adr(adr->_idx);
2011           assert(ptn != nullptr, "node should be registered");
2012           add_edge(n_ptn, ptn);
2013         }
2014       }
2015       break;
2016     }
2017     case Op_Blackhole: {
2018       // All blackhole pointer arguments are globally escaping.
2019       for (uint i = 0; i < n->req(); i++) {
2020         Node* in = n->in(i);
2021         if (in != nullptr) {
2022           const Type* at = _igvn->type(in);
2023           if (!at->isa_ptr()) continue;
2024 
2025           if (in->is_AddP()) {
2026             in = get_addp_base(in);
2027           }
2028 
2029           PointsToNode* ptn = ptnode_adr(in->_idx);
2030           assert(ptn != nullptr, "should be defined already");
2031           set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
2032           add_edge(n_ptn, ptn);
2033         }
2034       }
2035       break;
2036     }
2037     default: {
2038       // This method should be called only for EA specific nodes which may
2039       // miss some edges when they were created.
2040 #ifdef ASSERT
2041       n->dump(1);
2042 #endif
2043       guarantee(false, "unknown node");
2044     }
2045   }
2046   return;
2047 }
2048 
2049 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
2050   Node* adr = n->in(MemNode::Address);
2051   const Type* adr_type = _igvn->type(adr);
2052   adr_type = adr_type->make_ptr();
2053   if (adr_type == nullptr) {
2054     return; // skip dead nodes
2055   }
2056   if (adr_type->isa_oopptr()
2057       || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2058           && adr_type == TypeRawPtr::NOTNULL
2059           && is_captured_store_address(adr))) {
2060     delayed_worklist->push(n); // Process it later.
2061 #ifdef ASSERT
2062     assert (adr->is_AddP(), "expecting an AddP");
2063     if (adr_type == TypeRawPtr::NOTNULL) {
2064       // Verify a raw address for a store captured by Initialize node.
2065       int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2066       assert(offs != Type::OffsetBot, "offset must be a constant");
2067     }
2068 #endif
2069   } else {
2070     // Ignore copy the displaced header to the BoxNode (OSR compilation).
2071     if (adr->is_BoxLock()) {
2072       return;
2073     }
2074     // Stored value escapes in unsafe access.
2075     if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2076       delayed_worklist->push(n); // Process unsafe access later.
2077       return;
2078     }
2079 #ifdef ASSERT
2080     n->dump(1);
2081     assert(false, "not unsafe");
2082 #endif
2083   }
2084 }
2085 
2086 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
2087   Node* adr = n->in(MemNode::Address);
2088   const Type *adr_type = _igvn->type(adr);
2089   adr_type = adr_type->make_ptr();
2090 #ifdef ASSERT
2091   if (adr_type == nullptr) {
2092     n->dump(1);
2093     assert(adr_type != nullptr, "dead node should not be on list");
2094     return true;
2095   }
2096 #endif
2097 
2098   if (adr_type->isa_oopptr()
2099       || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2100            && adr_type == TypeRawPtr::NOTNULL
2101            && is_captured_store_address(adr))) {
2102     // Point Address to Value
2103     PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2104     assert(adr_ptn != nullptr &&
2105            adr_ptn->as_Field()->is_oop(), "node should be registered");
2106     Node* val = n->in(MemNode::ValueIn);
2107     PointsToNode* ptn = ptnode_adr(val->_idx);
2108     assert(ptn != nullptr, "node should be registered");
2109     add_edge(adr_ptn, ptn);
2110     return true;
2111   } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2112     // Stored value escapes in unsafe access.
2113     Node* val = n->in(MemNode::ValueIn);
2114     PointsToNode* ptn = ptnode_adr(val->_idx);
2115     assert(ptn != nullptr, "node should be registered");
2116     set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2117     // Add edge to object for unsafe access with offset.
2118     PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2119     assert(adr_ptn != nullptr, "node should be registered");
2120     if (adr_ptn->is_Field()) {
2121       assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2122       add_edge(adr_ptn, ptn);
2123     }
2124     return true;
2125   }
2126 #ifdef ASSERT
2127   n->dump(1);
2128   assert(false, "not unsafe");
2129 #endif
2130   return false;
2131 }
2132 
2133 // Iterate over the domains for the scalarized and non scalarized calling conventions: Only move to the next element
2134 // in the non scalarized calling convention once all elements of the scalarized calling convention for that parameter
2135 // have been iterated over. So (ignoring hidden arguments such as the null marker) iterating over:
2136 // value class MyValue {
2137 //   int f1;
2138 //   float f2;
2139 // }
2140 // void m(Object o, MyValue v, int i)
2141 // produces the pairs:
2142 // (Object, Object), (Myvalue, int), (MyValue, float), (int, int)
2143 class DomainIterator : public StackObj {
2144 private:
2145   const TypeTuple* _domain;
2146   const TypeTuple* _domain_cc;
2147   const GrowableArray<SigEntry>* _sig_cc;
2148 
2149   uint _i_domain;
2150   uint _i_domain_cc;
2151   int _i_sig_cc;
2152   uint _depth;
2153   uint _first_field_pos;
2154   const bool _is_static;
2155 
2156   void next_helper() {
2157     if (_sig_cc == nullptr) {
2158       return;
2159     }
2160     BasicType prev_bt = _i_sig_cc > 0 ? _sig_cc->at(_i_sig_cc-1)._bt : T_ILLEGAL;
2161     BasicType prev_prev_bt = _i_sig_cc > 1 ? _sig_cc->at(_i_sig_cc-2)._bt : T_ILLEGAL;
2162     while (_i_sig_cc < _sig_cc->length()) {
2163       BasicType bt = _sig_cc->at(_i_sig_cc)._bt;
2164       assert(bt != T_VOID || _sig_cc->at(_i_sig_cc-1)._bt == prev_bt, "incorrect prev bt");
2165       if (bt == T_METADATA) {
2166         if (_depth == 0) {
2167           _first_field_pos = _i_domain_cc;
2168         }
2169         _depth++;
2170       } else if (bt == T_VOID && (prev_bt != T_LONG && prev_bt != T_DOUBLE)) {
2171         _depth--;
2172         if (_depth == 0) {
2173           _i_domain++;
2174         }
2175       } else if (bt == T_OBJECT && prev_bt == T_METADATA && (_is_static || _i_domain > 0) && _sig_cc->at(_i_sig_cc)._offset == 0) {
2176         assert(_sig_cc->at(_i_sig_cc)._vt_oop, "buffer expected right after T_METADATA");
2177         assert(_depth == 1, "only root value has buffer");
2178         _i_domain_cc++;
2179         _first_field_pos = _i_domain_cc;
2180       } else if (bt == T_BOOLEAN && prev_prev_bt == T_METADATA && (_is_static || _i_domain > 0) && _sig_cc->at(_i_sig_cc)._offset == -1) {
2181         assert(_sig_cc->at(_i_sig_cc)._null_marker, "null marker expected right after T_METADATA");
2182         assert(_depth == 1, "only root value null marker");
2183         _i_domain_cc++;
2184         _first_field_pos = _i_domain_cc;
2185       } else {
2186         return;
2187       }
2188       prev_prev_bt = prev_bt;
2189       prev_bt = bt;
2190       _i_sig_cc++;
2191     }
2192   }
2193 
2194 public:
2195 
2196   DomainIterator(CallJavaNode* call) :
2197     _domain(call->tf()->domain_sig()),
2198     _domain_cc(call->tf()->domain_cc()),
2199     _sig_cc(call->method()->get_sig_cc()),
2200     _i_domain(TypeFunc::Parms),
2201     _i_domain_cc(TypeFunc::Parms),
2202     _i_sig_cc(0),
2203     _depth(0),
2204     _first_field_pos(0),
2205     _is_static(call->method()->is_static()) {
2206     next_helper();
2207   }
2208 
2209   bool has_next() const {
2210     assert(_sig_cc == nullptr || (_i_sig_cc < _sig_cc->length()) == (_i_domain < _domain->cnt()), "should reach end in sync");
2211     assert((_i_domain < _domain->cnt()) == (_i_domain_cc < _domain_cc->cnt()), "should reach end in sync");
2212     return _i_domain < _domain->cnt();
2213   }
2214 
2215   void next() {
2216     assert(_depth != 0 || _domain->field_at(_i_domain) == _domain_cc->field_at(_i_domain_cc), "should produce same non scalarized elements");
2217     _i_sig_cc++;
2218     if (_depth == 0) {
2219       _i_domain++;
2220     }
2221     _i_domain_cc++;
2222     next_helper();
2223   }
2224 
2225   uint i_domain() const {
2226     return _i_domain;
2227   }
2228 
2229   uint i_domain_cc() const {
2230     return _i_domain_cc;
2231   }
2232 
2233   const Type* current_domain() const {
2234     return _domain->field_at(_i_domain);
2235   }
2236 
2237   const Type* current_domain_cc() const {
2238     return _domain_cc->field_at(_i_domain_cc);
2239   }
2240 
2241   uint first_field_pos() const {
2242     assert(_first_field_pos >= TypeFunc::Parms, "not yet updated?");
2243     return _first_field_pos;
2244   }
2245 };
2246 
2247 // Determine whether any arguments are returned.
2248 bool ConnectionGraph::returns_an_argument(CallNode* call) {
2249   ciMethod* meth = call->as_CallJava()->method();
2250   BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2251   if (call_analyzer == nullptr) {
2252     return false;
2253   }
2254 
2255   const TypeTuple* d = call->tf()->domain_sig();
2256   bool ret_arg = false;
2257   for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2258     if (d->field_at(i)->isa_ptr() != nullptr &&
2259         call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2260       if (meth->is_scalarized_arg(i - TypeFunc::Parms) && !compatible_return(call->as_CallJava(), i)) {
2261         return false;
2262       }
2263       if (call->tf()->returns_inline_type_as_fields() != meth->is_scalarized_arg(i - TypeFunc::Parms)) {
2264         return false;
2265       }
2266       ret_arg = true;
2267     }
2268   }
2269   return ret_arg;
2270 }
2271 
2272 void ConnectionGraph::add_call_node(CallNode* call) {
2273   assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2274   uint call_idx = call->_idx;
2275   if (call->is_Allocate()) {
2276     Node* k = call->in(AllocateNode::KlassNode);
2277     const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2278     assert(kt != nullptr, "TypeKlassPtr  required.");
2279     PointsToNode::EscapeState es = PointsToNode::NoEscape;
2280     bool scalar_replaceable = true;
2281     NOT_PRODUCT(const char* nsr_reason = "");
2282     if (call->is_AllocateArray()) {
2283       if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2284         es = PointsToNode::GlobalEscape;
2285       } else {
2286         int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2287         if (length < 0) {
2288           // Not scalar replaceable if the length is not constant.
2289           scalar_replaceable = false;
2290           NOT_PRODUCT(nsr_reason = "has a non-constant length");
2291         } else if (length > EliminateAllocationArraySizeLimit) {
2292           // Not scalar replaceable if the length is too big.
2293           scalar_replaceable = false;
2294           NOT_PRODUCT(nsr_reason = "has a length that is too big");
2295         }
2296       }
2297     } else {  // Allocate instance
2298       if (!kt->isa_instklassptr()) { // StressReflectiveCode
2299         es = PointsToNode::GlobalEscape;
2300       } else {
2301         const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2302         ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2303         if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2304             ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2305             !ik->can_be_instantiated() ||
2306             ik->has_finalizer()) {
2307           es = PointsToNode::GlobalEscape;
2308         } else {
2309           int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2310           if (nfields > EliminateAllocationFieldsLimit) {
2311             // Not scalar replaceable if there are too many fields.
2312             scalar_replaceable = false;
2313             NOT_PRODUCT(nsr_reason = "has too many fields");
2314           }
2315         }
2316       }
2317     }
2318     add_java_object(call, es);
2319     PointsToNode* ptn = ptnode_adr(call_idx);
2320     if (!scalar_replaceable && ptn->scalar_replaceable()) {
2321       set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2322     }
2323   } else if (call->is_CallStaticJava()) {
2324     // Call nodes could be different types:
2325     //
2326     // 1. CallDynamicJavaNode (what happened during call is unknown):
2327     //
2328     //    - mapped to GlobalEscape JavaObject node if oop is returned;
2329     //
2330     //    - all oop arguments are escaping globally;
2331     //
2332     // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2333     //
2334     //    - the same as CallDynamicJavaNode if can't do bytecode analysis;
2335     //
2336     //    - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2337     //    - mapped to NoEscape JavaObject node if non-escaping object allocated
2338     //      during call is returned;
2339     //    - mapped to ArgEscape LocalVar node pointed to object arguments
2340     //      which are returned and does not escape during call;
2341     //
2342     //    - oop arguments escaping status is defined by bytecode analysis;
2343     //
2344     // For a static call, we know exactly what method is being called.
2345     // Use bytecode estimator to record whether the call's return value escapes.
2346     ciMethod* meth = call->as_CallJava()->method();
2347     if (meth == nullptr) {
2348       const char* name = call->as_CallStaticJava()->_name;
2349       assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
2350              strncmp(name, "load_unknown_inline", 19) == 0 ||
2351              strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "TODO: add failed case check");
2352       // Returns a newly allocated non-escaped object.
2353       add_java_object(call, PointsToNode::NoEscape);
2354       set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2355     } else if (meth->is_boxing_method()) {
2356       // Returns boxing object
2357       PointsToNode::EscapeState es;
2358       vmIntrinsics::ID intr = meth->intrinsic_id();
2359       if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2360         // It does not escape if object is always allocated.
2361         es = PointsToNode::NoEscape;
2362       } else {
2363         // It escapes globally if object could be loaded from cache.
2364         es = PointsToNode::GlobalEscape;
2365       }
2366       add_java_object(call, es);
2367       if (es == PointsToNode::GlobalEscape) {
2368         set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2369       }
2370     } else {
2371       BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2372       call_analyzer->copy_dependencies(_compile->dependencies());
2373       if (call_analyzer->is_return_allocated()) {
2374         // Returns a newly allocated non-escaped object, simply
2375         // update dependency information.
2376         // Mark it as NoEscape so that objects referenced by
2377         // it's fields will be marked as NoEscape at least.
2378         add_java_object(call, PointsToNode::NoEscape);
2379         set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2380       } else {
2381         // For non scalarized argument/return: add_proj() adds an edge between the return projection and the call,
2382         // process_call_arguments() adds an edge between the call and the argument
2383         // For scalarized argument/return: process_call_arguments() adds an edge between a call projection for a field
2384         // and the argument input to the call for that field. An edge is added between the projection for the returned
2385         // buffer and the call.
2386         if (returns_an_argument(call) && !call->tf()->returns_inline_type_as_fields()) {
2387           // returns non scalarized argument
2388           add_local_var(call, PointsToNode::ArgEscape);
2389         } else {
2390           // Returns unknown object or scalarized argument being returned
2391           map_ideal_node(call, phantom_obj);
2392         }
2393       }
2394     }
2395   } else {
2396     // An other type of call, assume the worst case:
2397     // returned value is unknown and globally escapes.
2398     assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2399     map_ideal_node(call, phantom_obj);
2400   }
2401 }
2402 
2403 // Check that the return type is compatible with the type of the argument being returned i.e. that there's no cast that
2404 // fails in the method
2405 bool ConnectionGraph::compatible_return(CallJavaNode* call, uint k) {
2406   return call->tf()->domain_sig()->field_at(k)->is_instptr()->instance_klass() == call->tf()->range_sig()->field_at(TypeFunc::Parms)->is_instptr()->instance_klass();
2407 }
2408 
2409 void ConnectionGraph::process_call_arguments(CallNode *call) {
2410     bool is_arraycopy = false;
2411     switch (call->Opcode()) {
2412 #ifdef ASSERT
2413     case Op_Allocate:
2414     case Op_AllocateArray:
2415     case Op_Lock:
2416     case Op_Unlock:
2417       assert(false, "should be done already");
2418       break;
2419 #endif
2420     case Op_ArrayCopy:
2421     case Op_CallLeafNoFP:
2422       // Most array copies are ArrayCopy nodes at this point but there
2423       // are still a few direct calls to the copy subroutines (See
2424       // PhaseStringOpts::copy_string())
2425       is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2426         call->as_CallLeaf()->is_call_to_arraycopystub();
2427       // fall through
2428     case Op_CallLeafVector:
2429     case Op_CallLeaf: {
2430       // Stub calls, objects do not escape but they are not scale replaceable.
2431       // Adjust escape state for outgoing arguments.
2432       const TypeTuple * d = call->tf()->domain_sig();
2433       bool src_has_oops = false;
2434       for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2435         const Type* at = d->field_at(i);
2436         Node *arg = call->in(i);
2437         if (arg == nullptr) {
2438           continue;
2439         }
2440         const Type *aat = _igvn->type(arg);
2441         if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2442           continue;
2443         }
2444         if (arg->is_AddP()) {
2445           //
2446           // The inline_native_clone() case when the arraycopy stub is called
2447           // after the allocation before Initialize and CheckCastPP nodes.
2448           // Or normal arraycopy for object arrays case.
2449           //
2450           // Set AddP's base (Allocate) as not scalar replaceable since
2451           // pointer to the base (with offset) is passed as argument.
2452           //
2453           arg = get_addp_base(arg);
2454         }
2455         PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2456         assert(arg_ptn != nullptr, "should be registered");
2457         PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2458         if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2459           assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2460                  aat->isa_ptr() != nullptr, "expecting an Ptr");
2461           bool arg_has_oops = aat->isa_oopptr() &&
2462                               (aat->isa_instptr() ||
2463                                (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2464                                (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2465                                                                aat->isa_aryptr()->is_flat() &&
2466                                                                aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2467           if (i == TypeFunc::Parms) {
2468             src_has_oops = arg_has_oops;
2469           }
2470           //
2471           // src or dst could be j.l.Object when other is basic type array:
2472           //
2473           //   arraycopy(char[],0,Object*,0,size);
2474           //   arraycopy(Object*,0,char[],0,size);
2475           //
2476           // Don't add edges in such cases.
2477           //
2478           bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2479                                        arg_has_oops && (i > TypeFunc::Parms);
2480 #ifdef ASSERT
2481           if (!(is_arraycopy ||
2482                 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2483                 (call->as_CallLeaf()->_name != nullptr &&
2484                  (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2485                   strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2486                   strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2487                   strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2488                   strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2489                   strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2490                   strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2491                   strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2492                   strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2493                   strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2494                   strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2495                   strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2496                   strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2497                   strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2498                   strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2499                   strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2500                   strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2501                   strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2502                   strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2503                   strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2504                   strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2505                   strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2506                   strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2507                   strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2508                   strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2509                   strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2510                   strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2511                   strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2512                   strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2513                   strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2514                   strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2515                   strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2516                   strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2517                   strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2518                   strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2519                   strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2520                   strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2521                   strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2522                   strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2523                   strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2524                   strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2525                   strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2526                   strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2527                   strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2528                   strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2529                   strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2530                   strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2531                   strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2532                   strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2533                   strcmp(call->as_CallLeaf()->_name, "store_inline_type_fields_to_buf") == 0 ||
2534                   strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2535                   strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2536                   strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2537                   strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2538                   strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2539                   strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2540                   strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2541                   strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2542                  ))) {
2543             call->dump();
2544             fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2545           }
2546 #endif
2547           // Always process arraycopy's destination object since
2548           // we need to add all possible edges to references in
2549           // source object.
2550           if (arg_esc >= PointsToNode::ArgEscape &&
2551               !arg_is_arraycopy_dest) {
2552             continue;
2553           }
2554           PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2555           if (call->is_ArrayCopy()) {
2556             ArrayCopyNode* ac = call->as_ArrayCopy();
2557             if (ac->is_clonebasic() ||
2558                 ac->is_arraycopy_validated() ||
2559                 ac->is_copyof_validated() ||
2560                 ac->is_copyofrange_validated()) {
2561               es = PointsToNode::NoEscape;
2562             }
2563           }
2564           set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2565           if (arg_is_arraycopy_dest) {
2566             Node* src = call->in(TypeFunc::Parms);
2567             if (src->is_AddP()) {
2568               src = get_addp_base(src);
2569             }
2570             PointsToNode* src_ptn = ptnode_adr(src->_idx);
2571             assert(src_ptn != nullptr, "should be registered");
2572             // Special arraycopy edge:
2573             // Only escape state of destination object's fields affects
2574             // escape state of fields in source object.
2575             add_arraycopy(call, es, src_ptn, arg_ptn);
2576           }
2577         }
2578       }
2579       break;
2580     }
2581     case Op_CallStaticJava: {
2582       // For a static call, we know exactly what method is being called.
2583       // Use bytecode estimator to record the call's escape affects
2584 #ifdef ASSERT
2585       const char* name = call->as_CallStaticJava()->_name;
2586       assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2587 #endif
2588       ciMethod* meth = call->as_CallJava()->method();
2589       if ((meth != nullptr) && meth->is_boxing_method()) {
2590         break; // Boxing methods do not modify any oops.
2591       }
2592       BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2593       // fall-through if not a Java method or no analyzer information
2594       if (call_analyzer != nullptr) {
2595         PointsToNode* call_ptn = ptnode_adr(call->_idx);
2596         bool ret_arg = returns_an_argument(call);
2597         for (DomainIterator di(call->as_CallJava()); di.has_next(); di.next()) {
2598           int k = di.i_domain() - TypeFunc::Parms;
2599           const Type* at = di.current_domain_cc();
2600           Node* arg = call->in(di.i_domain_cc());
2601           PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2602           assert(!call_analyzer->is_arg_returned(k) || !meth->is_scalarized_arg(k) ||
2603                  !compatible_return(call->as_CallJava(), di.i_domain()) ||
2604                  call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1) == nullptr ||
2605                  _igvn->type(call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1)) == at,
2606                  "scalarized return and scalarized argument should match");
2607           if (at->isa_ptr() != nullptr && call_analyzer->is_arg_returned(k) && ret_arg) {
2608             // The call returns arguments.
2609             if (meth->is_scalarized_arg(k)) {
2610               ProjNode* res_proj = call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1);
2611               if (res_proj != nullptr) {
2612                 assert(_igvn->type(res_proj)->isa_ptr(), "scalarized return and scalarized argument should match");
2613                 if (res_proj->_con != TypeFunc::Parms) {
2614                   // add an edge between the result projection for a field and the argument projection for the same argument field
2615                   PointsToNode* proj_ptn = ptnode_adr(res_proj->_idx);
2616                   add_edge(proj_ptn, arg_ptn);
2617                   if (!call_analyzer->is_return_local()) {
2618                     add_edge(proj_ptn, phantom_obj);
2619                   }
2620                 }
2621               }
2622             } else if (call_ptn != nullptr) { // Is call's result used?
2623               assert(call_ptn->is_LocalVar(), "node should be registered");
2624               assert(arg_ptn != nullptr, "node should be registered");
2625               add_edge(call_ptn, arg_ptn);
2626             }
2627           }
2628           if (at->isa_oopptr() != nullptr &&
2629               arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2630             if (!call_analyzer->is_arg_stack(k)) {
2631               // The argument global escapes
2632               set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2633             } else {
2634               set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2635               if (!call_analyzer->is_arg_local(k)) {
2636                 // The argument itself doesn't escape, but any fields might
2637                 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2638               }
2639             }
2640           }
2641         }
2642         if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2643           // The call returns arguments.
2644           assert(call_ptn->edge_count() > 0, "sanity");
2645           if (!call_analyzer->is_return_local()) {
2646             // Returns also unknown object.
2647             add_edge(call_ptn, phantom_obj);
2648           }
2649         }
2650         break;
2651       }
2652     }
2653     default: {
2654       // Fall-through here if not a Java method or no analyzer information
2655       // or some other type of call, assume the worst case: all arguments
2656       // globally escape.
2657       const TypeTuple* d = call->tf()->domain_cc();
2658       for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2659         const Type* at = d->field_at(i);
2660         if (at->isa_oopptr() != nullptr) {
2661           Node* arg = call->in(i);
2662           if (arg->is_AddP()) {
2663             arg = get_addp_base(arg);
2664           }
2665           assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2666           set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2667         }
2668       }
2669     }
2670   }
2671 }
2672 
2673 
2674 // Finish Graph construction.
2675 bool ConnectionGraph::complete_connection_graph(
2676                          GrowableArray<PointsToNode*>&   ptnodes_worklist,
2677                          GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2678                          GrowableArray<JavaObjectNode*>& java_objects_worklist,
2679                          GrowableArray<FieldNode*>&      oop_fields_worklist) {
2680   // Normally only 1-3 passes needed to build Connection Graph depending
2681   // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2682   // Set limit to 20 to catch situation when something did go wrong and
2683   // bailout Escape Analysis.
2684   // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2685 #define GRAPH_BUILD_ITER_LIMIT 20
2686 
2687   // Propagate GlobalEscape and ArgEscape escape states and check that
2688   // we still have non-escaping objects. The method pushs on _worklist
2689   // Field nodes which reference phantom_object.
2690   if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2691     return false; // Nothing to do.
2692   }
2693   // Now propagate references to all JavaObject nodes.
2694   int java_objects_length = java_objects_worklist.length();
2695   elapsedTimer build_time;
2696   build_time.start();
2697   elapsedTimer time;
2698   bool timeout = false;
2699   int new_edges = 1;
2700   int iterations = 0;
2701   do {
2702     while ((new_edges > 0) &&
2703            (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2704       double start_time = time.seconds();
2705       time.start();
2706       new_edges = 0;
2707       // Propagate references to phantom_object for nodes pushed on _worklist
2708       // by find_non_escaped_objects() and find_field_value().
2709       new_edges += add_java_object_edges(phantom_obj, false);
2710       for (int next = 0; next < java_objects_length; ++next) {
2711         JavaObjectNode* ptn = java_objects_worklist.at(next);
2712         new_edges += add_java_object_edges(ptn, true);
2713 
2714 #define SAMPLE_SIZE 4
2715         if ((next % SAMPLE_SIZE) == 0) {
2716           // Each 4 iterations calculate how much time it will take
2717           // to complete graph construction.
2718           time.stop();
2719           // Poll for requests from shutdown mechanism to quiesce compiler
2720           // because Connection graph construction may take long time.
2721           CompileBroker::maybe_block();
2722           double stop_time = time.seconds();
2723           double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2724           double time_until_end = time_per_iter * (double)(java_objects_length - next);
2725           if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2726             timeout = true;
2727             break; // Timeout
2728           }
2729           start_time = stop_time;
2730           time.start();
2731         }
2732 #undef SAMPLE_SIZE
2733 
2734       }
2735       if (timeout) break;
2736       if (new_edges > 0) {
2737         // Update escape states on each iteration if graph was updated.
2738         if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2739           return false; // Nothing to do.
2740         }
2741       }
2742       time.stop();
2743       if (time.seconds() >= EscapeAnalysisTimeout) {
2744         timeout = true;
2745         break;
2746       }
2747       _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2748     }
2749     if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2750       time.start();
2751       // Find fields which have unknown value.
2752       int fields_length = oop_fields_worklist.length();
2753       for (int next = 0; next < fields_length; next++) {
2754         FieldNode* field = oop_fields_worklist.at(next);
2755         if (field->edge_count() == 0) {
2756           new_edges += find_field_value(field);
2757           // This code may added new edges to phantom_object.
2758           // Need an other cycle to propagate references to phantom_object.
2759         }
2760       }
2761       time.stop();
2762       if (time.seconds() >= EscapeAnalysisTimeout) {
2763         timeout = true;
2764         break;
2765       }
2766     } else {
2767       new_edges = 0; // Bailout
2768     }
2769   } while (new_edges > 0);
2770 
2771   build_time.stop();
2772   _build_time = build_time.seconds();
2773   _build_iterations = iterations;
2774 
2775   // Bailout if passed limits.
2776   if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2777     Compile* C = _compile;
2778     if (C->log() != nullptr) {
2779       C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2780       C->log()->text("%s", timeout ? "time" : "iterations");
2781       C->log()->end_elem(" limit'");
2782     }
2783     assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2784            _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2785     // Possible infinite build_connection_graph loop,
2786     // bailout (no changes to ideal graph were made).
2787     return false;
2788   }
2789 
2790 #undef GRAPH_BUILD_ITER_LIMIT
2791 
2792   // Find fields initialized by null for non-escaping Allocations.
2793   int non_escaped_length = non_escaped_allocs_worklist.length();
2794   for (int next = 0; next < non_escaped_length; next++) {
2795     JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2796     PointsToNode::EscapeState es = ptn->escape_state();
2797     assert(es <= PointsToNode::ArgEscape, "sanity");
2798     if (es == PointsToNode::NoEscape) {
2799       if (find_init_values_null(ptn, _igvn) > 0) {
2800         // Adding references to null object does not change escape states
2801         // since it does not escape. Also no fields are added to null object.
2802         add_java_object_edges(null_obj, false);
2803       }
2804     }
2805     Node* n = ptn->ideal_node();
2806     if (n->is_Allocate()) {
2807       // The object allocated by this Allocate node will never be
2808       // seen by an other thread. Mark it so that when it is
2809       // expanded no MemBarStoreStore is added.
2810       InitializeNode* ini = n->as_Allocate()->initialization();
2811       if (ini != nullptr)
2812         ini->set_does_not_escape();
2813     }
2814   }
2815   return true; // Finished graph construction.
2816 }
2817 
2818 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2819 // and check that we still have non-escaping java objects.
2820 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2821                                                GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2822                                                bool print_method) {
2823   GrowableArray<PointsToNode*> escape_worklist;
2824   // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2825   int ptnodes_length = ptnodes_worklist.length();
2826   for (int next = 0; next < ptnodes_length; ++next) {
2827     PointsToNode* ptn = ptnodes_worklist.at(next);
2828     if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2829         ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2830       escape_worklist.push(ptn);
2831     }
2832   }
2833   // Set escape states to referenced nodes (edges list).
2834   while (escape_worklist.length() > 0) {
2835     PointsToNode* ptn = escape_worklist.pop();
2836     PointsToNode::EscapeState es  = ptn->escape_state();
2837     PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2838     if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2839         es >= PointsToNode::ArgEscape) {
2840       // GlobalEscape or ArgEscape state of field means it has unknown value.
2841       if (add_edge(ptn, phantom_obj)) {
2842         // New edge was added
2843         add_field_uses_to_worklist(ptn->as_Field());
2844       }
2845     }
2846     for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2847       PointsToNode* e = i.get();
2848       if (e->is_Arraycopy()) {
2849         assert(ptn->arraycopy_dst(), "sanity");
2850         // Propagate only fields escape state through arraycopy edge.
2851         if (e->fields_escape_state() < field_es) {
2852           set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2853           escape_worklist.push(e);
2854         }
2855       } else if (es >= field_es) {
2856         // fields_escape_state is also set to 'es' if it is less than 'es'.
2857         if (e->escape_state() < es) {
2858           set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2859           escape_worklist.push(e);
2860         }
2861       } else {
2862         // Propagate field escape state.
2863         bool es_changed = false;
2864         if (e->fields_escape_state() < field_es) {
2865           set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2866           es_changed = true;
2867         }
2868         if ((e->escape_state() < field_es) &&
2869             e->is_Field() && ptn->is_JavaObject() &&
2870             e->as_Field()->is_oop()) {
2871           // Change escape state of referenced fields.
2872           set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2873           es_changed = true;
2874         } else if (e->escape_state() < es) {
2875           set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2876           es_changed = true;
2877         }
2878         if (es_changed) {
2879           escape_worklist.push(e);
2880         }
2881       }
2882       if (print_method) {
2883         _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2884       }
2885     }
2886   }
2887   // Remove escaped objects from non_escaped list.
2888   for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2889     JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2890     if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2891       non_escaped_allocs_worklist.delete_at(next);
2892     }
2893     if (ptn->escape_state() == PointsToNode::NoEscape) {
2894       // Find fields in non-escaped allocations which have unknown value.
2895       find_init_values_phantom(ptn);
2896     }
2897   }
2898   return (non_escaped_allocs_worklist.length() > 0);
2899 }
2900 
2901 // Add all references to JavaObject node by walking over all uses.
2902 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2903   int new_edges = 0;
2904   if (populate_worklist) {
2905     // Populate _worklist by uses of jobj's uses.
2906     for (UseIterator i(jobj); i.has_next(); i.next()) {
2907       PointsToNode* use = i.get();
2908       if (use->is_Arraycopy()) {
2909         continue;
2910       }
2911       add_uses_to_worklist(use);
2912       if (use->is_Field() && use->as_Field()->is_oop()) {
2913         // Put on worklist all field's uses (loads) and
2914         // related field nodes (same base and offset).
2915         add_field_uses_to_worklist(use->as_Field());
2916       }
2917     }
2918   }
2919   for (int l = 0; l < _worklist.length(); l++) {
2920     PointsToNode* use = _worklist.at(l);
2921     if (PointsToNode::is_base_use(use)) {
2922       // Add reference from jobj to field and from field to jobj (field's base).
2923       use = PointsToNode::get_use_node(use)->as_Field();
2924       if (add_base(use->as_Field(), jobj)) {
2925         new_edges++;
2926       }
2927       continue;
2928     }
2929     assert(!use->is_JavaObject(), "sanity");
2930     if (use->is_Arraycopy()) {
2931       if (jobj == null_obj) { // null object does not have field edges
2932         continue;
2933       }
2934       // Added edge from Arraycopy node to arraycopy's source java object
2935       if (add_edge(use, jobj)) {
2936         jobj->set_arraycopy_src();
2937         new_edges++;
2938       }
2939       // and stop here.
2940       continue;
2941     }
2942     if (!add_edge(use, jobj)) {
2943       continue; // No new edge added, there was such edge already.
2944     }
2945     new_edges++;
2946     if (use->is_LocalVar()) {
2947       add_uses_to_worklist(use);
2948       if (use->arraycopy_dst()) {
2949         for (EdgeIterator i(use); i.has_next(); i.next()) {
2950           PointsToNode* e = i.get();
2951           if (e->is_Arraycopy()) {
2952             if (jobj == null_obj) { // null object does not have field edges
2953               continue;
2954             }
2955             // Add edge from arraycopy's destination java object to Arraycopy node.
2956             if (add_edge(jobj, e)) {
2957               new_edges++;
2958               jobj->set_arraycopy_dst();
2959             }
2960           }
2961         }
2962       }
2963     } else {
2964       // Added new edge to stored in field values.
2965       // Put on worklist all field's uses (loads) and
2966       // related field nodes (same base and offset).
2967       add_field_uses_to_worklist(use->as_Field());
2968     }
2969   }
2970   _worklist.clear();
2971   _in_worklist.reset();
2972   return new_edges;
2973 }
2974 
2975 // Put on worklist all related field nodes.
2976 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2977   assert(field->is_oop(), "sanity");
2978   int offset = field->offset();
2979   add_uses_to_worklist(field);
2980   // Loop over all bases of this field and push on worklist Field nodes
2981   // with the same offset and base (since they may reference the same field).
2982   for (BaseIterator i(field); i.has_next(); i.next()) {
2983     PointsToNode* base = i.get();
2984     add_fields_to_worklist(field, base);
2985     // Check if the base was source object of arraycopy and go over arraycopy's
2986     // destination objects since values stored to a field of source object are
2987     // accessible by uses (loads) of fields of destination objects.
2988     if (base->arraycopy_src()) {
2989       for (UseIterator j(base); j.has_next(); j.next()) {
2990         PointsToNode* arycp = j.get();
2991         if (arycp->is_Arraycopy()) {
2992           for (UseIterator k(arycp); k.has_next(); k.next()) {
2993             PointsToNode* abase = k.get();
2994             if (abase->arraycopy_dst() && abase != base) {
2995               // Look for the same arraycopy reference.
2996               add_fields_to_worklist(field, abase);
2997             }
2998           }
2999         }
3000       }
3001     }
3002   }
3003 }
3004 
3005 // Put on worklist all related field nodes.
3006 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
3007   int offset = field->offset();
3008   if (base->is_LocalVar()) {
3009     for (UseIterator j(base); j.has_next(); j.next()) {
3010       PointsToNode* f = j.get();
3011       if (PointsToNode::is_base_use(f)) { // Field
3012         f = PointsToNode::get_use_node(f);
3013         if (f == field || !f->as_Field()->is_oop()) {
3014           continue;
3015         }
3016         int offs = f->as_Field()->offset();
3017         if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
3018           add_to_worklist(f);
3019         }
3020       }
3021     }
3022   } else {
3023     assert(base->is_JavaObject(), "sanity");
3024     if (// Skip phantom_object since it is only used to indicate that
3025         // this field's content globally escapes.
3026         (base != phantom_obj) &&
3027         // null object node does not have fields.
3028         (base != null_obj)) {
3029       for (EdgeIterator i(base); i.has_next(); i.next()) {
3030         PointsToNode* f = i.get();
3031         // Skip arraycopy edge since store to destination object field
3032         // does not update value in source object field.
3033         if (f->is_Arraycopy()) {
3034           assert(base->arraycopy_dst(), "sanity");
3035           continue;
3036         }
3037         if (f == field || !f->as_Field()->is_oop()) {
3038           continue;
3039         }
3040         int offs = f->as_Field()->offset();
3041         if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
3042           add_to_worklist(f);
3043         }
3044       }
3045     }
3046   }
3047 }
3048 
3049 // Find fields which have unknown value.
3050 int ConnectionGraph::find_field_value(FieldNode* field) {
3051   // Escaped fields should have init value already.
3052   assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
3053   int new_edges = 0;
3054   for (BaseIterator i(field); i.has_next(); i.next()) {
3055     PointsToNode* base = i.get();
3056     if (base->is_JavaObject()) {
3057       // Skip Allocate's fields which will be processed later.
3058       if (base->ideal_node()->is_Allocate()) {
3059         return 0;
3060       }
3061       assert(base == null_obj, "only null ptr base expected here");
3062     }
3063   }
3064   if (add_edge(field, phantom_obj)) {
3065     // New edge was added
3066     new_edges++;
3067     add_field_uses_to_worklist(field);
3068   }
3069   return new_edges;
3070 }
3071 
3072 // Find fields initializing values for allocations.
3073 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
3074   assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
3075   PointsToNode* init_val = phantom_obj;
3076   Node* alloc = pta->ideal_node();
3077 
3078   // Do nothing for Allocate nodes since its fields values are
3079   // "known" unless they are initialized by arraycopy/clone.
3080   if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
3081     if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
3082       // Null-free inline type arrays are initialized with an init value instead of null
3083       init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
3084       assert(init_val != nullptr, "init value should be registered");
3085     } else {
3086       return 0;
3087     }
3088   }
3089   // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
3090   assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
3091 #ifdef ASSERT
3092   if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
3093     const char* name = alloc->as_CallStaticJava()->_name;
3094     assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
3095            strncmp(name, "load_unknown_inline", 19) == 0 ||
3096            strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "sanity");
3097   }
3098 #endif
3099   // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
3100   int new_edges = 0;
3101   for (EdgeIterator i(pta); i.has_next(); i.next()) {
3102     PointsToNode* field = i.get();
3103     if (field->is_Field() && field->as_Field()->is_oop()) {
3104       if (add_edge(field, init_val)) {
3105         // New edge was added
3106         new_edges++;
3107         add_field_uses_to_worklist(field->as_Field());
3108       }
3109     }
3110   }
3111   return new_edges;
3112 }
3113 
3114 // Find fields initializing values for allocations.
3115 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
3116   assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
3117   Node* alloc = pta->ideal_node();
3118   // Do nothing for Call nodes since its fields values are unknown.
3119   if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
3120     return 0;
3121   }
3122   InitializeNode* ini = alloc->as_Allocate()->initialization();
3123   bool visited_bottom_offset = false;
3124   GrowableArray<int> offsets_worklist;
3125   int new_edges = 0;
3126 
3127   // Check if an oop field's initializing value is recorded and add
3128   // a corresponding null if field's value if it is not recorded.
3129   // Connection Graph does not record a default initialization by null
3130   // captured by Initialize node.
3131   //
3132   for (EdgeIterator i(pta); i.has_next(); i.next()) {
3133     PointsToNode* field = i.get(); // Field (AddP)
3134     if (!field->is_Field() || !field->as_Field()->is_oop()) {
3135       continue; // Not oop field
3136     }
3137     int offset = field->as_Field()->offset();
3138     if (offset == Type::OffsetBot) {
3139       if (!visited_bottom_offset) {
3140         // OffsetBot is used to reference array's element,
3141         // always add reference to null to all Field nodes since we don't
3142         // known which element is referenced.
3143         if (add_edge(field, null_obj)) {
3144           // New edge was added
3145           new_edges++;
3146           add_field_uses_to_worklist(field->as_Field());
3147           visited_bottom_offset = true;
3148         }
3149       }
3150     } else {
3151       // Check only oop fields.
3152       const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
3153       if (adr_type->isa_rawptr()) {
3154 #ifdef ASSERT
3155         // Raw pointers are used for initializing stores so skip it
3156         // since it should be recorded already
3157         Node* base = get_addp_base(field->ideal_node());
3158         assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
3159 #endif
3160         continue;
3161       }
3162       if (!offsets_worklist.contains(offset)) {
3163         offsets_worklist.append(offset);
3164         Node* value = nullptr;
3165         if (ini != nullptr) {
3166           // StoreP::value_basic_type() == T_ADDRESS
3167           BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
3168           Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
3169           // Make sure initializing store has the same type as this AddP.
3170           // This AddP may reference non existing field because it is on a
3171           // dead branch of bimorphic call which is not eliminated yet.
3172           if (store != nullptr && store->is_Store() &&
3173               store->as_Store()->value_basic_type() == ft) {
3174             value = store->in(MemNode::ValueIn);
3175 #ifdef ASSERT
3176             if (VerifyConnectionGraph) {
3177               // Verify that AddP already points to all objects the value points to.
3178               PointsToNode* val = ptnode_adr(value->_idx);
3179               assert((val != nullptr), "should be processed already");
3180               PointsToNode* missed_obj = nullptr;
3181               if (val->is_JavaObject()) {
3182                 if (!field->points_to(val->as_JavaObject())) {
3183                   missed_obj = val;
3184                 }
3185               } else {
3186                 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
3187                   tty->print_cr("----------init store has invalid value -----");
3188                   store->dump();
3189                   val->dump();
3190                   assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
3191                 }
3192                 for (EdgeIterator j(val); j.has_next(); j.next()) {
3193                   PointsToNode* obj = j.get();
3194                   if (obj->is_JavaObject()) {
3195                     if (!field->points_to(obj->as_JavaObject())) {
3196                       missed_obj = obj;
3197                       break;
3198                     }
3199                   }
3200                 }
3201               }
3202               if (missed_obj != nullptr) {
3203                 tty->print_cr("----------field---------------------------------");
3204                 field->dump();
3205                 tty->print_cr("----------missed reference to object------------");
3206                 missed_obj->dump();
3207                 tty->print_cr("----------object referenced by init store-------");
3208                 store->dump();
3209                 val->dump();
3210                 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
3211               }
3212             }
3213 #endif
3214           } else {
3215             // There could be initializing stores which follow allocation.
3216             // For example, a volatile field store is not collected
3217             // by Initialize node.
3218             //
3219             // Need to check for dependent loads to separate such stores from
3220             // stores which follow loads. For now, add initial value null so
3221             // that compare pointers optimization works correctly.
3222           }
3223         }
3224         if (value == nullptr) {
3225           // A field's initializing value was not recorded. Add null.
3226           if (add_edge(field, null_obj)) {
3227             // New edge was added
3228             new_edges++;
3229             add_field_uses_to_worklist(field->as_Field());
3230           }
3231         }
3232       }
3233     }
3234   }
3235   return new_edges;
3236 }
3237 
3238 // Adjust scalar_replaceable state after Connection Graph is built.
3239 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
3240   // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
3241   // returns true. If one of the constraints in this method set 'jobj' to NSR
3242   // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
3243   // input, 'adjust_scalar_replaceable_state' will eventually be called with
3244   // that other object and the Phi will become a reducible Phi.
3245   // There could be multiple merges involving the same jobj.
3246   Unique_Node_List candidates;
3247 
3248   // Search for non-escaping objects which are not scalar replaceable
3249   // and mark them to propagate the state to referenced objects.
3250 
3251   for (UseIterator i(jobj); i.has_next(); i.next()) {
3252     PointsToNode* use = i.get();
3253     if (use->is_Arraycopy()) {
3254       continue;
3255     }
3256     if (use->is_Field()) {
3257       FieldNode* field = use->as_Field();
3258       assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3259       // 1. An object is not scalar replaceable if the field into which it is
3260       // stored has unknown offset (stored into unknown element of an array).
3261       if (field->offset() == Type::OffsetBot) {
3262         set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
3263         return;
3264       }
3265       for (BaseIterator i(field); i.has_next(); i.next()) {
3266         PointsToNode* base = i.get();
3267         // 2. An object is not scalar replaceable if the field into which it is
3268         // stored has multiple bases one of which is null.
3269         if ((base == null_obj) && (field->base_count() > 1)) {
3270           set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3271           return;
3272         }
3273         // 2.5. An object is not scalar replaceable if the field into which it is
3274         // stored has NSR base.
3275         if (!base->scalar_replaceable()) {
3276           set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3277           return;
3278         }
3279       }
3280     }
3281     assert(use->is_Field() || use->is_LocalVar(), "sanity");
3282     // 3. An object is not scalar replaceable if it is merged with other objects
3283     // and we can't remove the merge
3284     for (EdgeIterator j(use); j.has_next(); j.next()) {
3285       PointsToNode* ptn = j.get();
3286       if (ptn->is_JavaObject() && ptn != jobj) {
3287         Node* use_n = use->ideal_node();
3288 
3289         // These other local vars may point to multiple objects through a Phi
3290         // In this case we skip them and see if we can reduce the Phi.
3291         if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3292           use_n = use_n->in(1);
3293         }
3294 
3295         // If it's already a candidate or confirmed reducible merge we can skip verification
3296         if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3297           continue;
3298         }
3299 
3300         if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3301           candidates.push(use_n);
3302         } else {
3303           // Mark all objects as NSR if we can't remove the merge
3304           set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3305           set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3306         }
3307       }
3308     }
3309     if (!jobj->scalar_replaceable()) {
3310       return;
3311     }
3312   }
3313 
3314   for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3315     if (j.get()->is_Arraycopy()) {
3316       continue;
3317     }
3318 
3319     // Non-escaping object node should point only to field nodes.
3320     FieldNode* field = j.get()->as_Field();
3321     int offset = field->as_Field()->offset();
3322 
3323     // 4. An object is not scalar replaceable if it has a field with unknown
3324     // offset (array's element is accessed in loop).
3325     if (offset == Type::OffsetBot) {
3326       set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3327       return;
3328     }
3329     // 5. Currently an object is not scalar replaceable if a LoadStore node
3330     // access its field since the field value is unknown after it.
3331     //
3332     Node* n = field->ideal_node();
3333 
3334     // Test for an unsafe access that was parsed as maybe off heap
3335     // (with a CheckCastPP to raw memory).
3336     assert(n->is_AddP(), "expect an address computation");
3337     if (n->in(AddPNode::Base)->is_top() &&
3338         n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3339       assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3340       assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3341       set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3342       return;
3343     }
3344 
3345     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3346       Node* u = n->fast_out(i);
3347       if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3348         set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3349         return;
3350       }
3351     }
3352 
3353     // 6. Or the address may point to more then one object. This may produce
3354     // the false positive result (set not scalar replaceable)
3355     // since the flow-insensitive escape analysis can't separate
3356     // the case when stores overwrite the field's value from the case
3357     // when stores happened on different control branches.
3358     //
3359     // Note: it will disable scalar replacement in some cases:
3360     //
3361     //    Point p[] = new Point[1];
3362     //    p[0] = new Point(); // Will be not scalar replaced
3363     //
3364     // but it will save us from incorrect optimizations in next cases:
3365     //
3366     //    Point p[] = new Point[1];
3367     //    if ( x ) p[0] = new Point(); // Will be not scalar replaced
3368     //
3369     if (field->base_count() > 1 && candidates.size() == 0) {
3370       if (has_non_reducible_merge(field, reducible_merges)) {
3371         for (BaseIterator i(field); i.has_next(); i.next()) {
3372           PointsToNode* base = i.get();
3373           // Don't take into account LocalVar nodes which
3374           // may point to only one object which should be also
3375           // this field's base by now.
3376           if (base->is_JavaObject() && base != jobj) {
3377             // Mark all bases.
3378             set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3379             set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3380           }
3381         }
3382 
3383         if (!jobj->scalar_replaceable()) {
3384           return;
3385         }
3386       }
3387     }
3388   }
3389 
3390   // The candidate is truly a reducible merge only if none of the other
3391   // constraints ruled it as NSR. There could be multiple merges involving the
3392   // same jobj.
3393   assert(jobj->scalar_replaceable(), "sanity");
3394   for (uint i = 0; i < candidates.size(); i++ ) {
3395     Node* candidate = candidates.at(i);
3396     reducible_merges.push(candidate);
3397   }
3398 }
3399 
3400 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3401   for (BaseIterator i(field); i.has_next(); i.next()) {
3402     Node* base = i.get()->ideal_node();
3403     if (base->is_Phi() && !reducible_merges.member(base)) {
3404       return true;
3405     }
3406   }
3407   return false;
3408 }
3409 
3410 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3411   assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3412 
3413   // Look for 'phis' that refer to 'jobj' as the last
3414   // remaining scalar replaceable input.
3415   uint reducible_merges_cnt = reducible_merges.size();
3416   for (uint i = 0; i < reducible_merges_cnt; i++) {
3417     Node* phi = reducible_merges.at(i);
3418 
3419     // This 'Phi' will be a 'good' if it still points to
3420     // at least one scalar replaceable object. Note that 'obj'
3421     // was/should be marked as NSR before calling this function.
3422     bool good_phi = false;
3423 
3424     for (uint j = 1; j < phi->req(); j++) {
3425       JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3426       if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3427         good_phi = true;
3428         break;
3429       }
3430     }
3431 
3432     if (!good_phi) {
3433       NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3434       reducible_merges.remove(i);
3435 
3436       // Decrement the index because the 'remove' call above actually
3437       // moves the last entry of the list to position 'i'.
3438       i--;
3439 
3440       reducible_merges_cnt--;
3441     }
3442   }
3443 }
3444 
3445 // Propagate NSR (Not scalar replaceable) state.
3446 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3447   int jobj_length = jobj_worklist.length();
3448   bool found_nsr_alloc = true;
3449   while (found_nsr_alloc) {
3450     found_nsr_alloc = false;
3451     for (int next = 0; next < jobj_length; ++next) {
3452       JavaObjectNode* jobj = jobj_worklist.at(next);
3453       for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3454         PointsToNode* use = i.get();
3455         if (use->is_Field()) {
3456           FieldNode* field = use->as_Field();
3457           assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3458           assert(field->offset() != Type::OffsetBot, "sanity");
3459           for (BaseIterator i(field); i.has_next(); i.next()) {
3460             PointsToNode* base = i.get();
3461             // An object is not scalar replaceable if the field into which
3462             // it is stored has NSR base.
3463             if ((base != null_obj) && !base->scalar_replaceable()) {
3464               set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3465               // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3466               // because there is no point in reducing a Phi that won't improve the number of SR
3467               // objects.
3468               revisit_reducible_phi_status(jobj, reducible_merges);
3469               found_nsr_alloc = true;
3470               break;
3471             }
3472           }
3473         } else if (use->is_LocalVar()) {
3474           Node* phi = use->ideal_node();
3475           if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3476             set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3477             reducible_merges.yank(phi);
3478             found_nsr_alloc = true;
3479             break;
3480           }
3481         }
3482         _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3483       }
3484     }
3485   }
3486 }
3487 
3488 #ifdef ASSERT
3489 void ConnectionGraph::verify_connection_graph(
3490                          GrowableArray<PointsToNode*>&   ptnodes_worklist,
3491                          GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3492                          GrowableArray<JavaObjectNode*>& java_objects_worklist,
3493                          GrowableArray<Node*>& addp_worklist) {
3494   // Verify that graph is complete - no new edges could be added.
3495   int java_objects_length = java_objects_worklist.length();
3496   int non_escaped_length  = non_escaped_allocs_worklist.length();
3497   int new_edges = 0;
3498   for (int next = 0; next < java_objects_length; ++next) {
3499     JavaObjectNode* ptn = java_objects_worklist.at(next);
3500     new_edges += add_java_object_edges(ptn, true);
3501   }
3502   assert(new_edges == 0, "graph was not complete");
3503   // Verify that escape state is final.
3504   int length = non_escaped_allocs_worklist.length();
3505   find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3506   assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3507          (non_escaped_length == length) &&
3508          (_worklist.length() == 0), "escape state was not final");
3509 
3510   // Verify fields information.
3511   int addp_length = addp_worklist.length();
3512   for (int next = 0; next < addp_length; ++next ) {
3513     Node* n = addp_worklist.at(next);
3514     FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3515     if (field->is_oop()) {
3516       // Verify that field has all bases
3517       Node* base = get_addp_base(n);
3518       PointsToNode* ptn = ptnode_adr(base->_idx);
3519       if (ptn->is_JavaObject()) {
3520         assert(field->has_base(ptn->as_JavaObject()), "sanity");
3521       } else {
3522         assert(ptn->is_LocalVar(), "sanity");
3523         for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3524           PointsToNode* e = i.get();
3525           if (e->is_JavaObject()) {
3526             assert(field->has_base(e->as_JavaObject()), "sanity");
3527           }
3528         }
3529       }
3530       // Verify that all fields have initializing values.
3531       if (field->edge_count() == 0) {
3532         tty->print_cr("----------field does not have references----------");
3533         field->dump();
3534         for (BaseIterator i(field); i.has_next(); i.next()) {
3535           PointsToNode* base = i.get();
3536           tty->print_cr("----------field has next base---------------------");
3537           base->dump();
3538           if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3539             tty->print_cr("----------base has fields-------------------------");
3540             for (EdgeIterator j(base); j.has_next(); j.next()) {
3541               j.get()->dump();
3542             }
3543             tty->print_cr("----------base has references---------------------");
3544             for (UseIterator j(base); j.has_next(); j.next()) {
3545               j.get()->dump();
3546             }
3547           }
3548         }
3549         for (UseIterator i(field); i.has_next(); i.next()) {
3550           i.get()->dump();
3551         }
3552         assert(field->edge_count() > 0, "sanity");
3553       }
3554     }
3555   }
3556 }
3557 #endif
3558 
3559 // Optimize ideal graph.
3560 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3561                                            GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3562   Compile* C = _compile;
3563   PhaseIterGVN* igvn = _igvn;
3564   if (EliminateLocks) {
3565     // Mark locks before changing ideal graph.
3566     int cnt = C->macro_count();
3567     for (int i = 0; i < cnt; i++) {
3568       Node *n = C->macro_node(i);
3569       if (n->is_AbstractLock()) { // Lock and Unlock nodes
3570         AbstractLockNode* alock = n->as_AbstractLock();
3571         if (!alock->is_non_esc_obj()) {
3572           const Type* obj_type = igvn->type(alock->obj_node());
3573           if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3574             assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3575             // The lock could be marked eliminated by lock coarsening
3576             // code during first IGVN before EA. Replace coarsened flag
3577             // to eliminate all associated locks/unlocks.
3578 #ifdef ASSERT
3579             alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3580 #endif
3581             alock->set_non_esc_obj();
3582           }
3583         }
3584       }
3585     }
3586   }
3587 
3588   if (OptimizePtrCompare) {
3589     for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3590       Node *n = ptr_cmp_worklist.at(i);
3591       assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3592       const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3593       if (tcmp->singleton()) {
3594         Node* cmp = igvn->makecon(tcmp);
3595 #ifndef PRODUCT
3596         if (PrintOptimizePtrCompare) {
3597           tty->print_cr("++++ Replaced: %d %s(%d,%d) --> %s", n->_idx, (n->Opcode() == Op_CmpP ? "CmpP" : "CmpN"), n->in(1)->_idx, n->in(2)->_idx, (tcmp == TypeInt::CC_EQ ? "EQ" : "NotEQ"));
3598           if (Verbose) {
3599             n->dump(1);
3600           }
3601         }
3602 #endif
3603         igvn->replace_node(n, cmp);
3604       }
3605     }
3606   }
3607 
3608   // For MemBarStoreStore nodes added in library_call.cpp, check
3609   // escape status of associated AllocateNode and optimize out
3610   // MemBarStoreStore node if the allocated object never escapes.
3611   for (int i = 0; i < storestore_worklist.length(); i++) {
3612     Node* storestore = storestore_worklist.at(i);
3613     Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3614     if (alloc->is_Allocate() && not_global_escape(alloc)) {
3615       if (alloc->in(AllocateNode::InlineType) != nullptr) {
3616         // Non-escaping inline type buffer allocations don't require a membar
3617         storestore->as_MemBar()->remove(_igvn);
3618       } else {
3619         MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3620         mb->init_req(TypeFunc::Memory,  storestore->in(TypeFunc::Memory));
3621         mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3622         igvn->register_new_node_with_optimizer(mb);
3623         igvn->replace_node(storestore, mb);
3624       }
3625     }
3626   }
3627 }
3628 
3629 // Atomic flat accesses on non-escaping objects can be optimized to non-atomic accesses
3630 void ConnectionGraph::optimize_flat_accesses(GrowableArray<SafePointNode*>& sfn_worklist) {
3631   PhaseIterGVN& igvn = *_igvn;
3632   bool delay = igvn.delay_transform();
3633   igvn.set_delay_transform(true);
3634   igvn.C->for_each_flat_access([&](Node* n) {
3635     Node* base = n->is_LoadFlat() ? n->as_LoadFlat()->base() : n->as_StoreFlat()->base();
3636     if (!not_global_escape(base)) {
3637       return;
3638     }
3639 
3640     bool expanded;
3641     if (n->is_LoadFlat()) {
3642       expanded = n->as_LoadFlat()->expand_non_atomic(igvn);
3643     } else {
3644       expanded = n->as_StoreFlat()->expand_non_atomic(igvn);
3645     }
3646     if (expanded) {
3647       sfn_worklist.remove(n->as_SafePoint());
3648       igvn.C->remove_flat_access(n);
3649     }
3650   });
3651   igvn.set_delay_transform(delay);
3652 }
3653 
3654 // Optimize objects compare.
3655 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3656   const TypeInt* UNKNOWN = TypeInt::CC;    // [-1, 0,1]
3657   if (!OptimizePtrCompare) {
3658     return UNKNOWN;
3659   }
3660   const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3661   const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3662 
3663   PointsToNode* ptn1 = ptnode_adr(left->_idx);
3664   PointsToNode* ptn2 = ptnode_adr(right->_idx);
3665   JavaObjectNode* jobj1 = unique_java_object(left);
3666   JavaObjectNode* jobj2 = unique_java_object(right);
3667 
3668   // The use of this method during allocation merge reduction may cause 'left'
3669   // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3670   // that doesn't reference an unique java object.
3671   if (ptn1 == nullptr || ptn2 == nullptr ||
3672       jobj1 == nullptr || jobj2 == nullptr) {
3673     return UNKNOWN;
3674   }
3675 
3676   assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3677   assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3678 
3679   // Check simple cases first.
3680   if (jobj1 != nullptr) {
3681     if (jobj1->escape_state() == PointsToNode::NoEscape) {
3682       if (jobj1 == jobj2) {
3683         // Comparing the same not escaping object.
3684         return EQ;
3685       }
3686       Node* obj = jobj1->ideal_node();
3687       // Comparing not escaping allocation.
3688       if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3689           !ptn2->points_to(jobj1)) {
3690         return NE; // This includes nullness check.
3691       }
3692     }
3693   }
3694   if (jobj2 != nullptr) {
3695     if (jobj2->escape_state() == PointsToNode::NoEscape) {
3696       Node* obj = jobj2->ideal_node();
3697       // Comparing not escaping allocation.
3698       if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3699           !ptn1->points_to(jobj2)) {
3700         return NE; // This includes nullness check.
3701       }
3702     }
3703   }
3704   if (jobj1 != nullptr && jobj1 != phantom_obj &&
3705       jobj2 != nullptr && jobj2 != phantom_obj &&
3706       jobj1->ideal_node()->is_Con() &&
3707       jobj2->ideal_node()->is_Con()) {
3708     // Klass or String constants compare. Need to be careful with
3709     // compressed pointers - compare types of ConN and ConP instead of nodes.
3710     const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3711     const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3712     if (t1->make_ptr() == t2->make_ptr()) {
3713       return EQ;
3714     } else {
3715       return NE;
3716     }
3717   }
3718   if (ptn1->meet(ptn2)) {
3719     return UNKNOWN; // Sets are not disjoint
3720   }
3721 
3722   // Sets are disjoint.
3723   bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3724   bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3725   bool set1_has_null_ptr    = ptn1->points_to(null_obj);
3726   bool set2_has_null_ptr    = ptn2->points_to(null_obj);
3727   if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3728       (set2_has_unknown_ptr && set1_has_null_ptr)) {
3729     // Check nullness of unknown object.
3730     return UNKNOWN;
3731   }
3732 
3733   // Disjointness by itself is not sufficient since
3734   // alias analysis is not complete for escaped objects.
3735   // Disjoint sets are definitely unrelated only when
3736   // at least one set has only not escaping allocations.
3737   if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3738     if (ptn1->non_escaping_allocation()) {
3739       return NE;
3740     }
3741   }
3742   if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3743     if (ptn2->non_escaping_allocation()) {
3744       return NE;
3745     }
3746   }
3747   return UNKNOWN;
3748 }
3749 
3750 // Connection Graph construction functions.
3751 
3752 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3753   PointsToNode* ptadr = _nodes.at(n->_idx);
3754   if (ptadr != nullptr) {
3755     assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3756     return;
3757   }
3758   Compile* C = _compile;
3759   ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3760   map_ideal_node(n, ptadr);
3761 }
3762 
3763 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3764   PointsToNode* ptadr = _nodes.at(n->_idx);
3765   if (ptadr != nullptr) {
3766     assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3767     return ptadr;
3768   }
3769   Compile* C = _compile;
3770   ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3771   map_ideal_node(n, ptadr);
3772   return ptadr;
3773 }
3774 
3775 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3776   PointsToNode* ptadr = _nodes.at(n->_idx);
3777   if (ptadr != nullptr) {
3778     assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3779     return;
3780   }
3781   bool unsafe = false;
3782   bool is_oop = is_oop_field(n, offset, &unsafe);
3783   if (unsafe) {
3784     es = PointsToNode::GlobalEscape;
3785   }
3786   Compile* C = _compile;
3787   FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3788   map_ideal_node(n, field);
3789 }
3790 
3791 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3792                                     PointsToNode* src, PointsToNode* dst) {
3793   assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3794   assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3795   PointsToNode* ptadr = _nodes.at(n->_idx);
3796   if (ptadr != nullptr) {
3797     assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3798     return;
3799   }
3800   Compile* C = _compile;
3801   ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3802   map_ideal_node(n, ptadr);
3803   // Add edge from arraycopy node to source object.
3804   (void)add_edge(ptadr, src);
3805   src->set_arraycopy_src();
3806   // Add edge from destination object to arraycopy node.
3807   (void)add_edge(dst, ptadr);
3808   dst->set_arraycopy_dst();
3809 }
3810 
3811 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3812   const Type* adr_type = n->as_AddP()->bottom_type();
3813   int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3814   BasicType bt = T_INT;
3815   if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3816     // Check only oop fields.
3817     if (!adr_type->isa_aryptr() ||
3818         adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3819         adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3820       // OffsetBot is used to reference array's element. Ignore first AddP.
3821       if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3822         bt = T_OBJECT;
3823       }
3824     }
3825   } else if (offset != oopDesc::klass_offset_in_bytes()) {
3826     if (adr_type->isa_instptr()) {
3827       ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3828       if (field != nullptr) {
3829         bt = field->layout_type();
3830       } else {
3831         // Check for unsafe oop field access
3832         if (has_oop_node_outs(n)) {
3833           bt = T_OBJECT;
3834           (*unsafe) = true;
3835         }
3836       }
3837     } else if (adr_type->isa_aryptr()) {
3838       if (offset == arrayOopDesc::length_offset_in_bytes()) {
3839         // Ignore array length load.
3840       } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3841         // Ignore first AddP.
3842       } else {
3843         const Type* elemtype = adr_type->is_aryptr()->elem();
3844         if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3845           ciInlineKlass* vk = elemtype->inline_klass();
3846           field_offset += vk->payload_offset();
3847           ciField* field = vk->get_field_by_offset(field_offset, false);
3848           if (field != nullptr) {
3849             bt = field->layout_type();
3850           } else {
3851             assert(field_offset == vk->payload_offset() + vk->null_marker_offset_in_payload(), "no field or null marker of %s at offset %d", vk->name()->as_utf8(), field_offset);
3852             bt = T_BOOLEAN;
3853           }
3854         } else {
3855           bt = elemtype->array_element_basic_type();
3856         }
3857       }
3858     } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3859       // Allocation initialization, ThreadLocal field access, unsafe access
3860       if (has_oop_node_outs(n)) {
3861         bt = T_OBJECT;
3862       }
3863     }
3864   }
3865   // Note: T_NARROWOOP is not classed as a real reference type
3866   bool res = (is_reference_type(bt) || bt == T_NARROWOOP);
3867   assert(!has_oop_node_outs(n) || res, "sanity: AddP has oop outs, needs to be treated as oop field");
3868   return res;
3869 }
3870 
3871 bool ConnectionGraph::has_oop_node_outs(Node* n) {
3872   return n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3873          n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3874          n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3875          BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n);
3876 }
3877 
3878 // Returns unique pointed java object or null.
3879 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3880   // If the node was created after the escape computation we can't answer.
3881   uint idx = n->_idx;
3882   if (idx >= nodes_size()) {
3883     return nullptr;
3884   }
3885   PointsToNode* ptn = ptnode_adr(idx);
3886   if (ptn == nullptr) {
3887     return nullptr;
3888   }
3889   if (ptn->is_JavaObject()) {
3890     return ptn->as_JavaObject();
3891   }
3892   assert(ptn->is_LocalVar(), "sanity");
3893   // Check all java objects it points to.
3894   JavaObjectNode* jobj = nullptr;
3895   for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3896     PointsToNode* e = i.get();
3897     if (e->is_JavaObject()) {
3898       if (jobj == nullptr) {
3899         jobj = e->as_JavaObject();
3900       } else if (jobj != e) {
3901         return nullptr;
3902       }
3903     }
3904   }
3905   return jobj;
3906 }
3907 
3908 // Return true if this node points only to non-escaping allocations.
3909 bool PointsToNode::non_escaping_allocation() {
3910   if (is_JavaObject()) {
3911     Node* n = ideal_node();
3912     if (n->is_Allocate() || n->is_CallStaticJava()) {
3913       return (escape_state() == PointsToNode::NoEscape);
3914     } else {
3915       return false;
3916     }
3917   }
3918   assert(is_LocalVar(), "sanity");
3919   // Check all java objects it points to.
3920   for (EdgeIterator i(this); i.has_next(); i.next()) {
3921     PointsToNode* e = i.get();
3922     if (e->is_JavaObject()) {
3923       Node* n = e->ideal_node();
3924       if ((e->escape_state() != PointsToNode::NoEscape) ||
3925           !(n->is_Allocate() || n->is_CallStaticJava())) {
3926         return false;
3927       }
3928     }
3929   }
3930   return true;
3931 }
3932 
3933 // Return true if we know the node does not escape globally.
3934 bool ConnectionGraph::not_global_escape(Node *n) {
3935   assert(!_collecting, "should not call during graph construction");
3936   // If the node was created after the escape computation we can't answer.
3937   uint idx = n->_idx;
3938   if (idx >= nodes_size()) {
3939     return false;
3940   }
3941   PointsToNode* ptn = ptnode_adr(idx);
3942   if (ptn == nullptr) {
3943     return false; // not in congraph (e.g. ConI)
3944   }
3945   PointsToNode::EscapeState es = ptn->escape_state();
3946   // If we have already computed a value, return it.
3947   if (es >= PointsToNode::GlobalEscape) {
3948     return false;
3949   }
3950   if (ptn->is_JavaObject()) {
3951     return true; // (es < PointsToNode::GlobalEscape);
3952   }
3953   assert(ptn->is_LocalVar(), "sanity");
3954   // Check all java objects it points to.
3955   for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3956     if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3957       return false;
3958     }
3959   }
3960   return true;
3961 }
3962 
3963 // Return true if locked object does not escape globally
3964 // and locked code region (identified by BoxLockNode) is balanced:
3965 // all compiled code paths have corresponding Lock/Unlock pairs.
3966 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3967   if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3968     if (EliminateNestedLocks) {
3969       // We can mark whole locking region as Local only when only
3970       // one object is used for locking.
3971       alock->box_node()->as_BoxLock()->set_local();
3972     }
3973     return true;
3974   }
3975   return false;
3976 }
3977 
3978 // Helper functions
3979 
3980 // Return true if this node points to specified node or nodes it points to.
3981 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3982   if (is_JavaObject()) {
3983     return (this == ptn);
3984   }
3985   assert(is_LocalVar() || is_Field(), "sanity");
3986   for (EdgeIterator i(this); i.has_next(); i.next()) {
3987     if (i.get() == ptn) {
3988       return true;
3989     }
3990   }
3991   return false;
3992 }
3993 
3994 // Return true if one node points to an other.
3995 bool PointsToNode::meet(PointsToNode* ptn) {
3996   if (this == ptn) {
3997     return true;
3998   } else if (ptn->is_JavaObject()) {
3999     return this->points_to(ptn->as_JavaObject());
4000   } else if (this->is_JavaObject()) {
4001     return ptn->points_to(this->as_JavaObject());
4002   }
4003   assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
4004   int ptn_count =  ptn->edge_count();
4005   for (EdgeIterator i(this); i.has_next(); i.next()) {
4006     PointsToNode* this_e = i.get();
4007     for (int j = 0; j < ptn_count; j++) {
4008       if (this_e == ptn->edge(j)) {
4009         return true;
4010       }
4011     }
4012   }
4013   return false;
4014 }
4015 
4016 #ifdef ASSERT
4017 // Return true if bases point to this java object.
4018 bool FieldNode::has_base(JavaObjectNode* jobj) const {
4019   for (BaseIterator i(this); i.has_next(); i.next()) {
4020     if (i.get() == jobj) {
4021       return true;
4022     }
4023   }
4024   return false;
4025 }
4026 #endif
4027 
4028 bool ConnectionGraph::is_captured_store_address(Node* addp) {
4029   // Handle simple case first.
4030   assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
4031   if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
4032     return true;
4033   } else if (addp->in(AddPNode::Address)->is_Phi()) {
4034     for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
4035       Node* addp_use = addp->fast_out(i);
4036       if (addp_use->is_Store()) {
4037         for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
4038           if (addp_use->fast_out(j)->is_Initialize()) {
4039             return true;
4040           }
4041         }
4042       }
4043     }
4044   }
4045   return false;
4046 }
4047 
4048 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
4049   const Type *adr_type = phase->type(adr);
4050   if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
4051     // We are computing a raw address for a store captured by an Initialize
4052     // compute an appropriate address type. AddP cases #3 and #5 (see below).
4053     int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
4054     assert(offs != Type::OffsetBot ||
4055            adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
4056            "offset must be a constant or it is initialization of array");
4057     return offs;
4058   }
4059   return adr_type->is_ptr()->flat_offset();
4060 }
4061 
4062 Node* ConnectionGraph::get_addp_base(Node *addp) {
4063   assert(addp->is_AddP(), "must be AddP");
4064   //
4065   // AddP cases for Base and Address inputs:
4066   // case #1. Direct object's field reference:
4067   //     Allocate
4068   //       |
4069   //     Proj #5 ( oop result )
4070   //       |
4071   //     CheckCastPP (cast to instance type)
4072   //      | |
4073   //     AddP  ( base == address )
4074   //
4075   // case #2. Indirect object's field reference:
4076   //      Phi
4077   //       |
4078   //     CastPP (cast to instance type)
4079   //      | |
4080   //     AddP  ( base == address )
4081   //
4082   // case #3. Raw object's field reference for Initialize node.
4083   //          Could have an additional Phi merging multiple allocations.
4084   //      Allocate
4085   //        |
4086   //      Proj #5 ( oop result )
4087   //  top   |
4088   //     \  |
4089   //     AddP  ( base == top )
4090   //
4091   // case #4. Array's element reference:
4092   //   {CheckCastPP | CastPP}
4093   //     |  | |
4094   //     |  AddP ( array's element offset )
4095   //     |  |
4096   //     AddP ( array's offset )
4097   //
4098   // case #5. Raw object's field reference for arraycopy stub call:
4099   //          The inline_native_clone() case when the arraycopy stub is called
4100   //          after the allocation before Initialize and CheckCastPP nodes.
4101   //      Allocate
4102   //        |
4103   //      Proj #5 ( oop result )
4104   //       | |
4105   //       AddP  ( base == address )
4106   //
4107   // case #6. Constant Pool, ThreadLocal, CastX2P, Klass, OSR buffer buf or
4108   //          Raw object's field reference:
4109   //      {ConP, ThreadLocal, CastX2P, raw Load, Parm0}
4110   //  top   |
4111   //     \  |
4112   //     AddP  ( base == top )
4113   //
4114   // case #7. Klass's field reference.
4115   //      LoadKlass
4116   //       | |
4117   //       AddP  ( base == address )
4118   //
4119   // case #8. narrow Klass's field reference.
4120   //      LoadNKlass
4121   //       |
4122   //      DecodeN
4123   //       | |
4124   //       AddP  ( base == address )
4125   //
4126   // case #9. Mixed unsafe access
4127   //    {instance}
4128   //        |
4129   //      CheckCastPP (raw)
4130   //  top   |
4131   //     \  |
4132   //     AddP  ( base == top )
4133   //
4134   // case #10. Klass fetched with
4135   //           LibraryCallKit::load_*_refined_array_klass()
4136   //           which has en extra Phi.
4137   //  LoadKlass   LoadKlass
4138   //       |          |
4139   //     CastPP    CastPP
4140   //          \   /
4141   //           Phi
4142   //      top   |
4143   //         \  |
4144   //         AddP  ( base == top )
4145   //
4146   Node *base = addp->in(AddPNode::Base);
4147   if (base->uncast()->is_top()) { // The AddP case #3, #6, #9, and #10.
4148     base = addp->in(AddPNode::Address);
4149     while (base->is_AddP()) {
4150       // Case #6 (unsafe access) may have several chained AddP nodes.
4151       assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
4152       base = base->in(AddPNode::Address);
4153     }
4154     if (base->Opcode() == Op_CheckCastPP &&
4155         base->bottom_type()->isa_rawptr() &&
4156         _igvn->type(base->in(1))->isa_oopptr()) {
4157       base = base->in(1); // Case #9
4158     } else {
4159       // Case #3, #6, and #10
4160       Node* uncast_base = base->uncast();
4161       int opcode = uncast_base->Opcode();
4162       assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
4163              opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
4164              (_igvn->C->is_osr_compilation() && uncast_base->is_Parm() && uncast_base->as_Parm()->_con == TypeFunc::Parms)||
4165              (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
4166              (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_klassptr() != nullptr)) ||
4167              is_captured_store_address(addp) ||
4168              is_load_array_klass_related(uncast_base), "sanity");
4169     }
4170   }
4171   return base;
4172 }
4173 
4174 #ifdef ASSERT
4175 // Case #10
4176 bool ConnectionGraph::is_load_array_klass_related(const Node* uncast_base) {
4177   if (!uncast_base->is_Phi() || uncast_base->req() != 3) {
4178     return false;
4179   }
4180   Node* in1 = uncast_base->in(1);
4181   Node* in2 = uncast_base->in(2);
4182   return in1->uncast()->Opcode() == Op_LoadKlass &&
4183          in2->uncast()->Opcode() == Op_LoadKlass;
4184 }
4185 #endif
4186 
4187 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
4188   assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
4189   Node* addp2 = addp->raw_out(0);
4190   if (addp->outcnt() == 1 && addp2->is_AddP() &&
4191       addp2->in(AddPNode::Base) == n &&
4192       addp2->in(AddPNode::Address) == addp) {
4193     assert(addp->in(AddPNode::Base) == n, "expecting the same base");
4194     //
4195     // Find array's offset to push it on worklist first and
4196     // as result process an array's element offset first (pushed second)
4197     // to avoid CastPP for the array's offset.
4198     // Otherwise the inserted CastPP (LocalVar) will point to what
4199     // the AddP (Field) points to. Which would be wrong since
4200     // the algorithm expects the CastPP has the same point as
4201     // as AddP's base CheckCastPP (LocalVar).
4202     //
4203     //    ArrayAllocation
4204     //     |
4205     //    CheckCastPP
4206     //     |
4207     //    memProj (from ArrayAllocation CheckCastPP)
4208     //     |  ||
4209     //     |  ||   Int (element index)
4210     //     |  ||    |   ConI (log(element size))
4211     //     |  ||    |   /
4212     //     |  ||   LShift
4213     //     |  ||  /
4214     //     |  AddP (array's element offset)
4215     //     |  |
4216     //     |  | ConI (array's offset: #12(32-bits) or #24(64-bits))
4217     //     | / /
4218     //     AddP (array's offset)
4219     //      |
4220     //     Load/Store (memory operation on array's element)
4221     //
4222     return addp2;
4223   }
4224   return nullptr;
4225 }
4226 
4227 //
4228 // Adjust the type and inputs of an AddP which computes the
4229 // address of a field of an instance
4230 //
4231 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
4232   PhaseGVN* igvn = _igvn;
4233   const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
4234   assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
4235   const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
4236   if (t == nullptr) {
4237     // We are computing a raw address for a store captured by an Initialize
4238     // compute an appropriate address type (cases #3 and #5).
4239     assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
4240     assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
4241     intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
4242     assert(offs != Type::OffsetBot, "offset must be a constant");
4243     if (base_t->isa_aryptr() != nullptr) {
4244       // In the case of a flat inline type array, each field has its
4245       // own slice so we need to extract the field being accessed from
4246       // the address computation
4247       t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
4248     } else {
4249       t = base_t->add_offset(offs)->is_oopptr();
4250     }
4251   }
4252   int inst_id = base_t->instance_id();
4253   assert(!t->is_known_instance() || t->instance_id() == inst_id,
4254                              "old type must be non-instance or match new type");
4255 
4256   // The type 't' could be subclass of 'base_t'.
4257   // As result t->offset() could be large then base_t's size and it will
4258   // cause the failure in add_offset() with narrow oops since TypeOopPtr()
4259   // constructor verifies correctness of the offset.
4260   //
4261   // It could happened on subclass's branch (from the type profiling
4262   // inlining) which was not eliminated during parsing since the exactness
4263   // of the allocation type was not propagated to the subclass type check.
4264   //
4265   // Or the type 't' could be not related to 'base_t' at all.
4266   // It could happen when CHA type is different from MDO type on a dead path
4267   // (for example, from instanceof check) which is not collapsed during parsing.
4268   //
4269   // Do nothing for such AddP node and don't process its users since
4270   // this code branch will go away.
4271   //
4272   if (!t->is_known_instance() &&
4273       !base_t->maybe_java_subtype_of(t)) {
4274      return false; // bail out
4275   }
4276   const TypePtr* tinst = base_t->add_offset(t->offset());
4277   if (tinst->isa_aryptr() && t->isa_aryptr()) {
4278     // In the case of a flat inline type array, each field has its
4279     // own slice so we need to keep track of the field being accessed.
4280     tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
4281     // Keep array properties (not flat/null-free)
4282     tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
4283     if (tinst == nullptr) {
4284       return false; // Skip dead path with inconsistent properties
4285     }
4286   }
4287 
4288   // Do NOT remove the next line: ensure a new alias index is allocated
4289   // for the instance type. Note: C++ will not remove it since the call
4290   // has side effect.
4291   int alias_idx = _compile->get_alias_index(tinst);
4292   igvn->set_type(addp, tinst);
4293   // record the allocation in the node map
4294   set_map(addp, get_map(base->_idx));
4295   // Set addp's Base and Address to 'base'.
4296   Node *abase = addp->in(AddPNode::Base);
4297   Node *adr   = addp->in(AddPNode::Address);
4298   if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
4299       adr->in(0)->_idx == (uint)inst_id) {
4300     // Skip AddP cases #3 and #5.
4301   } else {
4302     assert(!abase->is_top(), "sanity"); // AddP case #3
4303     if (abase != base) {
4304       igvn->hash_delete(addp);
4305       addp->set_req(AddPNode::Base, base);
4306       if (abase == adr) {
4307         addp->set_req(AddPNode::Address, base);
4308       } else {
4309         // AddP case #4 (adr is array's element offset AddP node)
4310 #ifdef ASSERT
4311         const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
4312         assert(adr->is_AddP() && atype != nullptr &&
4313                atype->instance_id() == inst_id, "array's element offset should be processed first");
4314 #endif
4315       }
4316       igvn->hash_insert(addp);
4317     }
4318   }
4319   // Put on IGVN worklist since at least addp's type was changed above.
4320   record_for_optimizer(addp);
4321   return true;
4322 }
4323 
4324 //
4325 // Create a new version of orig_phi if necessary. Returns either the newly
4326 // created phi or an existing phi.  Sets create_new to indicate whether a new
4327 // phi was created.  Cache the last newly created phi in the node map.
4328 //
4329 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *>  &orig_phi_worklist, bool &new_created) {
4330   Compile *C = _compile;
4331   PhaseGVN* igvn = _igvn;
4332   new_created = false;
4333   int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
4334   // nothing to do if orig_phi is bottom memory or matches alias_idx
4335   if (phi_alias_idx == alias_idx) {
4336     return orig_phi;
4337   }
4338   // Have we recently created a Phi for this alias index?
4339   PhiNode *result = get_map_phi(orig_phi->_idx);
4340   if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
4341     return result;
4342   }
4343   // Previous check may fail when the same wide memory Phi was split into Phis
4344   // for different memory slices. Search all Phis for this region.
4345   if (result != nullptr) {
4346     Node* region = orig_phi->in(0);
4347     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4348       Node* phi = region->fast_out(i);
4349       if (phi->is_Phi() &&
4350           C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4351         assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4352         return phi->as_Phi();
4353       }
4354     }
4355   }
4356   if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4357     if (C->do_escape_analysis() == true && !C->failing()) {
4358       // Retry compilation without escape analysis.
4359       // If this is the first failure, the sentinel string will "stick"
4360       // to the Compile object, and the C2Compiler will see it and retry.
4361       C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4362     }
4363     return nullptr;
4364   }
4365   orig_phi_worklist.append_if_missing(orig_phi);
4366   const TypePtr *atype = C->get_adr_type(alias_idx);
4367   result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4368   C->copy_node_notes_to(result, orig_phi);
4369   igvn->set_type(result, result->bottom_type());
4370   record_for_optimizer(result);
4371   set_map(orig_phi, result);
4372   new_created = true;
4373   return result;
4374 }
4375 
4376 //
4377 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4378 // specified alias index.
4379 //
4380 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4381   assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4382   Compile *C = _compile;
4383   PhaseGVN* igvn = _igvn;
4384   bool new_phi_created;
4385   PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4386   if (!new_phi_created) {
4387     return result;
4388   }
4389   GrowableArray<PhiNode *>  phi_list;
4390   GrowableArray<uint>  cur_input;
4391   PhiNode *phi = orig_phi;
4392   uint idx = 1;
4393   bool finished = false;
4394   while(!finished) {
4395     while (idx < phi->req()) {
4396       Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4397       if (mem != nullptr && mem->is_Phi()) {
4398         PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4399         if (new_phi_created) {
4400           // found an phi for which we created a new split, push current one on worklist and begin
4401           // processing new one
4402           phi_list.push(phi);
4403           cur_input.push(idx);
4404           phi = mem->as_Phi();
4405           result = newphi;
4406           idx = 1;
4407           continue;
4408         } else {
4409           mem = newphi;
4410         }
4411       }
4412       if (C->failing()) {
4413         return nullptr;
4414       }
4415       result->set_req(idx++, mem);
4416     }
4417 #ifdef ASSERT
4418     // verify that the new Phi has an input for each input of the original
4419     assert( phi->req() == result->req(), "must have same number of inputs.");
4420     assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4421 #endif
4422     // Check if all new phi's inputs have specified alias index.
4423     // Otherwise use old phi.
4424     for (uint i = 1; i < phi->req(); i++) {
4425       Node* in = result->in(i);
4426       assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4427     }
4428     // we have finished processing a Phi, see if there are any more to do
4429     finished = (phi_list.length() == 0 );
4430     if (!finished) {
4431       phi = phi_list.pop();
4432       idx = cur_input.pop();
4433       PhiNode *prev_result = get_map_phi(phi->_idx);
4434       prev_result->set_req(idx++, result);
4435       result = prev_result;
4436     }
4437   }
4438   return result;
4439 }
4440 
4441 //
4442 // The next methods are derived from methods in MemNode.
4443 //
4444 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4445   Node *mem = mmem;
4446   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4447   // means an array I have not precisely typed yet.  Do not do any
4448   // alias stuff with it any time soon.
4449   if (toop->base() != Type::AnyPtr &&
4450       !(toop->isa_instptr() &&
4451         toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4452         toop->offset() == Type::OffsetBot)) {
4453     mem = mmem->memory_at(alias_idx);
4454     // Update input if it is progress over what we have now
4455   }
4456   return mem;
4457 }
4458 
4459 //
4460 // Move memory users to their memory slices.
4461 //
4462 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *>  &orig_phis) {
4463   Compile* C = _compile;
4464   PhaseGVN* igvn = _igvn;
4465   const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4466   assert(tp != nullptr, "ptr type");
4467   int alias_idx = C->get_alias_index(tp);
4468   int general_idx = C->get_general_index(alias_idx);
4469 
4470   // Move users first
4471   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4472     Node* use = n->fast_out(i);
4473     if (use->is_MergeMem()) {
4474       MergeMemNode* mmem = use->as_MergeMem();
4475       assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4476       if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4477         continue; // Nothing to do
4478       }
4479       // Replace previous general reference to mem node.
4480       uint orig_uniq = C->unique();
4481       Node* m = find_inst_mem(n, general_idx, orig_phis);
4482       assert(orig_uniq == C->unique(), "no new nodes");
4483       mmem->set_memory_at(general_idx, m);
4484       --imax;
4485       --i;
4486     } else if (use->is_MemBar()) {
4487       assert(!use->is_Initialize(), "initializing stores should not be moved");
4488       if (use->req() > MemBarNode::Precedent &&
4489           use->in(MemBarNode::Precedent) == n) {
4490         // Don't move related membars.
4491         record_for_optimizer(use);
4492         continue;
4493       }
4494       tp = use->as_MemBar()->adr_type()->isa_ptr();
4495       if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4496           alias_idx == general_idx) {
4497         continue; // Nothing to do
4498       }
4499       // Move to general memory slice.
4500       uint orig_uniq = C->unique();
4501       Node* m = find_inst_mem(n, general_idx, orig_phis);
4502       assert(orig_uniq == C->unique(), "no new nodes");
4503       igvn->hash_delete(use);
4504       imax -= use->replace_edge(n, m, igvn);
4505       igvn->hash_insert(use);
4506       record_for_optimizer(use);
4507       --i;
4508 #ifdef ASSERT
4509     } else if (use->is_Mem()) {
4510       // Memory nodes should have new memory input.
4511       tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4512       assert(tp != nullptr, "ptr type");
4513       int idx = C->get_alias_index(tp);
4514       assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4515              "Following memory nodes should have new memory input or be on the same memory slice");
4516     } else if (use->is_Phi()) {
4517       // Phi nodes should be split and moved already.
4518       tp = use->as_Phi()->adr_type()->isa_ptr();
4519       assert(tp != nullptr, "ptr type");
4520       int idx = C->get_alias_index(tp);
4521       assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4522     } else {
4523       use->dump();
4524       assert(false, "should not be here");
4525 #endif
4526     }
4527   }
4528 }
4529 
4530 //
4531 // Search memory chain of "mem" to find a MemNode whose address
4532 // is the specified alias index.
4533 //
4534 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4535 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *>  &orig_phis, uint rec_depth) {
4536   if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4537     _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4538     return nullptr;
4539   }
4540   if (orig_mem == nullptr) {
4541     return orig_mem;
4542   }
4543   Compile* C = _compile;
4544   PhaseGVN* igvn = _igvn;
4545   const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4546   bool is_instance = (toop != nullptr) && toop->is_known_instance();
4547   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4548   Node *prev = nullptr;
4549   Node *result = orig_mem;
4550   while (prev != result) {
4551     prev = result;
4552     if (result == start_mem) {
4553       break;  // hit one of our sentinels
4554     }
4555     if (result->is_Mem()) {
4556       const Type *at = igvn->type(result->in(MemNode::Address));
4557       if (at == Type::TOP) {
4558         break; // Dead
4559       }
4560       assert (at->isa_ptr() != nullptr, "pointer type required.");
4561       int idx = C->get_alias_index(at->is_ptr());
4562       if (idx == alias_idx) {
4563         break; // Found
4564       }
4565       if (!is_instance && (at->isa_oopptr() == nullptr ||
4566                            !at->is_oopptr()->is_known_instance())) {
4567         break; // Do not skip store to general memory slice.
4568       }
4569       result = result->in(MemNode::Memory);
4570     }
4571     if (!is_instance) {
4572       continue;  // don't search further for non-instance types
4573     }
4574     // skip over a call which does not affect this memory slice
4575     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4576       Node *proj_in = result->in(0);
4577       if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4578         break;  // hit one of our sentinels
4579       } else if (proj_in->is_Call()) {
4580         // ArrayCopy node processed here as well
4581         CallNode *call = proj_in->as_Call();
4582         if (!call->may_modify(toop, igvn)) {
4583           result = call->in(TypeFunc::Memory);
4584         }
4585       } else if (proj_in->is_Initialize()) {
4586         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4587         // Stop if this is the initialization for the object instance which
4588         // which contains this memory slice, otherwise skip over it.
4589         if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4590           result = proj_in->in(TypeFunc::Memory);
4591         } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4592           assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4593           result = get_map(result->_idx);
4594           assert(result != nullptr, "new projection should have been allocated");
4595           break;
4596         }
4597       } else if (proj_in->is_MemBar()) {
4598         // Check if there is an array copy for a clone
4599         // Step over GC barrier when ReduceInitialCardMarks is disabled
4600         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4601         Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4602 
4603         if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4604           // Stop if it is a clone
4605           ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4606           if (ac->may_modify(toop, igvn)) {
4607             break;
4608           }
4609         }
4610         result = proj_in->in(TypeFunc::Memory);
4611       }
4612     } else if (result->is_MergeMem()) {
4613       MergeMemNode *mmem = result->as_MergeMem();
4614       result = step_through_mergemem(mmem, alias_idx, toop);
4615       if (result == mmem->base_memory()) {
4616         // Didn't find instance memory, search through general slice recursively.
4617         result = mmem->memory_at(C->get_general_index(alias_idx));
4618         result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4619         if (C->failing()) {
4620           return nullptr;
4621         }
4622         mmem->set_memory_at(alias_idx, result);
4623       }
4624     } else if (result->is_Phi() &&
4625                C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4626       Node *un = result->as_Phi()->unique_input(igvn);
4627       if (un != nullptr) {
4628         orig_phis.append_if_missing(result->as_Phi());
4629         result = un;
4630       } else {
4631         break;
4632       }
4633     } else if (result->is_ClearArray()) {
4634       if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4635         // Can not bypass initialization of the instance
4636         // we are looking for.
4637         break;
4638       }
4639       // Otherwise skip it (the call updated 'result' value).
4640     } else if (result->Opcode() == Op_SCMemProj) {
4641       Node* mem = result->in(0);
4642       Node* adr = nullptr;
4643       if (mem->is_LoadStore()) {
4644         adr = mem->in(MemNode::Address);
4645       } else {
4646         assert(mem->Opcode() == Op_EncodeISOArray ||
4647                mem->Opcode() == Op_StrCompressedCopy, "sanity");
4648         adr = mem->in(3); // Memory edge corresponds to destination array
4649       }
4650       const Type *at = igvn->type(adr);
4651       if (at != Type::TOP) {
4652         assert(at->isa_ptr() != nullptr, "pointer type required.");
4653         int idx = C->get_alias_index(at->is_ptr());
4654         if (idx == alias_idx) {
4655           // Assert in debug mode
4656           assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4657           break; // In product mode return SCMemProj node
4658         }
4659       }
4660       result = mem->in(MemNode::Memory);
4661     } else if (result->Opcode() == Op_StrInflatedCopy) {
4662       Node* adr = result->in(3); // Memory edge corresponds to destination array
4663       const Type *at = igvn->type(adr);
4664       if (at != Type::TOP) {
4665         assert(at->isa_ptr() != nullptr, "pointer type required.");
4666         int idx = C->get_alias_index(at->is_ptr());
4667         if (idx == alias_idx) {
4668           // Assert in debug mode
4669           assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4670           break; // In product mode return SCMemProj node
4671         }
4672       }
4673       result = result->in(MemNode::Memory);
4674     }
4675   }
4676   if (result->is_Phi()) {
4677     PhiNode *mphi = result->as_Phi();
4678     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4679     const TypePtr *t = mphi->adr_type();
4680     if (!is_instance) {
4681       // Push all non-instance Phis on the orig_phis worklist to update inputs
4682       // during Phase 4 if needed.
4683       orig_phis.append_if_missing(mphi);
4684     } else if (C->get_alias_index(t) != alias_idx) {
4685       // Create a new Phi with the specified alias index type.
4686       result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4687     }
4688   }
4689   // the result is either MemNode, PhiNode, InitializeNode.
4690   return result;
4691 }
4692 
4693 //
4694 //  Convert the types of non-escaped object to instance types where possible,
4695 //  propagate the new type information through the graph, and update memory
4696 //  edges and MergeMem inputs to reflect the new type.
4697 //
4698 //  We start with allocations (and calls which may be allocations)  on alloc_worklist.
4699 //  The processing is done in 4 phases:
4700 //
4701 //  Phase 1:  Process possible allocations from alloc_worklist.  Create instance
4702 //            types for the CheckCastPP for allocations where possible.
4703 //            Propagate the new types through users as follows:
4704 //               casts and Phi:  push users on alloc_worklist
4705 //               AddP:  cast Base and Address inputs to the instance type
4706 //                      push any AddP users on alloc_worklist and push any memnode
4707 //                      users onto memnode_worklist.
4708 //  Phase 2:  Process MemNode's from memnode_worklist. compute new address type and
4709 //            search the Memory chain for a store with the appropriate type
4710 //            address type.  If a Phi is found, create a new version with
4711 //            the appropriate memory slices from each of the Phi inputs.
4712 //            For stores, process the users as follows:
4713 //               MemNode:  push on memnode_worklist
4714 //               MergeMem: push on mergemem_worklist
4715 //  Phase 3:  Process MergeMem nodes from mergemem_worklist.  Walk each memory slice
4716 //            moving the first node encountered of each  instance type to the
4717 //            the input corresponding to its alias index.
4718 //            appropriate memory slice.
4719 //  Phase 4:  Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4720 //
4721 // In the following example, the CheckCastPP nodes are the cast of allocation
4722 // results and the allocation of node 29 is non-escaped and eligible to be an
4723 // instance type.
4724 //
4725 // We start with:
4726 //
4727 //     7 Parm #memory
4728 //    10  ConI  "12"
4729 //    19  CheckCastPP   "Foo"
4730 //    20  AddP  _ 19 19 10  Foo+12  alias_index=4
4731 //    29  CheckCastPP   "Foo"
4732 //    30  AddP  _ 29 29 10  Foo+12  alias_index=4
4733 //
4734 //    40  StoreP  25   7  20   ... alias_index=4
4735 //    50  StoreP  35  40  30   ... alias_index=4
4736 //    60  StoreP  45  50  20   ... alias_index=4
4737 //    70  LoadP    _  60  30   ... alias_index=4
4738 //    80  Phi     75  50  60   Memory alias_index=4
4739 //    90  LoadP    _  80  30   ... alias_index=4
4740 //   100  LoadP    _  80  20   ... alias_index=4
4741 //
4742 //
4743 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4744 // and creating a new alias index for node 30.  This gives:
4745 //
4746 //     7 Parm #memory
4747 //    10  ConI  "12"
4748 //    19  CheckCastPP   "Foo"
4749 //    20  AddP  _ 19 19 10  Foo+12  alias_index=4
4750 //    29  CheckCastPP   "Foo"  iid=24
4751 //    30  AddP  _ 29 29 10  Foo+12  alias_index=6  iid=24
4752 //
4753 //    40  StoreP  25   7  20   ... alias_index=4
4754 //    50  StoreP  35  40  30   ... alias_index=6
4755 //    60  StoreP  45  50  20   ... alias_index=4
4756 //    70  LoadP    _  60  30   ... alias_index=6
4757 //    80  Phi     75  50  60   Memory alias_index=4
4758 //    90  LoadP    _  80  30   ... alias_index=6
4759 //   100  LoadP    _  80  20   ... alias_index=4
4760 //
4761 // In phase 2, new memory inputs are computed for the loads and stores,
4762 // And a new version of the phi is created.  In phase 4, the inputs to
4763 // node 80 are updated and then the memory nodes are updated with the
4764 // values computed in phase 2.  This results in:
4765 //
4766 //     7 Parm #memory
4767 //    10  ConI  "12"
4768 //    19  CheckCastPP   "Foo"
4769 //    20  AddP  _ 19 19 10  Foo+12  alias_index=4
4770 //    29  CheckCastPP   "Foo"  iid=24
4771 //    30  AddP  _ 29 29 10  Foo+12  alias_index=6  iid=24
4772 //
4773 //    40  StoreP  25  7   20   ... alias_index=4
4774 //    50  StoreP  35  7   30   ... alias_index=6
4775 //    60  StoreP  45  40  20   ... alias_index=4
4776 //    70  LoadP    _  50  30   ... alias_index=6
4777 //    80  Phi     75  40  60   Memory alias_index=4
4778 //   120  Phi     75  50  50   Memory alias_index=6
4779 //    90  LoadP    _ 120  30   ... alias_index=6
4780 //   100  LoadP    _  80  20   ... alias_index=4
4781 //
4782 void ConnectionGraph::split_unique_types(GrowableArray<Node *>  &alloc_worklist,
4783                                          GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4784                                          GrowableArray<MergeMemNode*> &mergemem_worklist,
4785                                          Unique_Node_List &reducible_merges) {
4786   DEBUG_ONLY(Unique_Node_List reduced_merges;)
4787   GrowableArray<Node *>  memnode_worklist;
4788   GrowableArray<PhiNode *>  orig_phis;
4789   PhaseIterGVN  *igvn = _igvn;
4790   uint new_index_start = (uint) _compile->num_alias_types();
4791   VectorSet visited;
4792   ideal_nodes.clear(); // Reset for use with set_map/get_map.
4793 
4794   //  Phase 1:  Process possible allocations from alloc_worklist.
4795   //  Create instance types for the CheckCastPP for allocations where possible.
4796   //
4797   // (Note: don't forget to change the order of the second AddP node on
4798   //  the alloc_worklist if the order of the worklist processing is changed,
4799   //  see the comment in find_second_addp().)
4800   //
4801   while (alloc_worklist.length() != 0) {
4802     Node *n = alloc_worklist.pop();
4803     uint ni = n->_idx;
4804     if (n->is_Call()) {
4805       CallNode *alloc = n->as_Call();
4806       // copy escape information to call node
4807       PointsToNode* ptn = ptnode_adr(alloc->_idx);
4808       PointsToNode::EscapeState es = ptn->escape_state();
4809       // We have an allocation or call which returns a Java object,
4810       // see if it is non-escaped.
4811       if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4812         continue;
4813       }
4814       // Find CheckCastPP for the allocate or for the return value of a call
4815       n = alloc->result_cast();
4816       if (n == nullptr) {            // No uses except Initialize node
4817         if (alloc->is_Allocate()) {
4818           // Set the scalar_replaceable flag for allocation
4819           // so it could be eliminated if it has no uses.
4820           alloc->as_Allocate()->_is_scalar_replaceable = true;
4821         }
4822         continue;
4823       }
4824       if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4825         // we could reach here for allocate case if one init is associated with many allocs.
4826         if (alloc->is_Allocate()) {
4827           alloc->as_Allocate()->_is_scalar_replaceable = false;
4828         }
4829         continue;
4830       }
4831 
4832       // The inline code for Object.clone() casts the allocation result to
4833       // java.lang.Object and then to the actual type of the allocated
4834       // object. Detect this case and use the second cast.
4835       // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4836       // the allocation result is cast to java.lang.Object and then
4837       // to the actual Array type.
4838       if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4839           && (alloc->is_AllocateArray() ||
4840               igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4841         Node *cast2 = nullptr;
4842         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4843           Node *use = n->fast_out(i);
4844           if (use->is_CheckCastPP()) {
4845             cast2 = use;
4846             break;
4847           }
4848         }
4849         if (cast2 != nullptr) {
4850           n = cast2;
4851         } else {
4852           // Non-scalar replaceable if the allocation type is unknown statically
4853           // (reflection allocation), the object can't be restored during
4854           // deoptimization without precise type.
4855           continue;
4856         }
4857       }
4858 
4859       const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4860       if (t == nullptr) {
4861         continue;  // not a TypeOopPtr
4862       }
4863       if (!t->klass_is_exact()) {
4864         continue; // not an unique type
4865       }
4866       if (alloc->is_Allocate()) {
4867         // Set the scalar_replaceable flag for allocation
4868         // so it could be eliminated.
4869         alloc->as_Allocate()->_is_scalar_replaceable = true;
4870       }
4871       set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4872       // in order for an object to be scalar-replaceable, it must be:
4873       //   - a direct allocation (not a call returning an object)
4874       //   - non-escaping
4875       //   - eligible to be a unique type
4876       //   - not determined to be ineligible by escape analysis
4877       set_map(alloc, n);
4878       set_map(n, alloc);
4879       const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4880       igvn->hash_delete(n);
4881       igvn->set_type(n,  tinst);
4882       n->raise_bottom_type(tinst);
4883       igvn->hash_insert(n);
4884       record_for_optimizer(n);
4885       // Allocate an alias index for the header fields. Accesses to
4886       // the header emitted during macro expansion wouldn't have
4887       // correct memory state otherwise.
4888       _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4889       _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4890       if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4891         // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4892         InitializeNode* init = alloc->as_Allocate()->initialization();
4893         assert(init != nullptr, "can't find Initialization node for this Allocate node");
4894         auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4895           const TypePtr* adr_type = proj->adr_type();
4896           const TypePtr* new_adr_type = tinst->with_offset(adr_type->offset());
4897           if (adr_type->isa_aryptr()) {
4898             // In the case of a flat inline type array, each field has its own slice so we need a
4899             // NarrowMemProj for each field of the flat array elements
4900             new_adr_type = new_adr_type->is_aryptr()->with_field_offset(adr_type->is_aryptr()->field_offset().get());
4901           }
4902           if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4903             // Do NOT remove the next line: ensure a new alias index is allocated for the instance type.
4904             uint alias_idx = _compile->get_alias_index(new_adr_type);
4905             assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4906             NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4907             igvn->set_type(new_proj, new_proj->bottom_type());
4908             record_for_optimizer(new_proj);
4909             set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4910           }
4911         };
4912         init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4913 
4914         // First, put on the worklist all Field edges from Connection Graph
4915         // which is more accurate than putting immediate users from Ideal Graph.
4916         for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4917           PointsToNode* tgt = e.get();
4918           if (tgt->is_Arraycopy()) {
4919             continue;
4920           }
4921           Node* use = tgt->ideal_node();
4922           assert(tgt->is_Field() && use->is_AddP(),
4923                  "only AddP nodes are Field edges in CG");
4924           if (use->outcnt() > 0) { // Don't process dead nodes
4925             Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4926             if (addp2 != nullptr) {
4927               assert(alloc->is_AllocateArray(),"array allocation was expected");
4928               alloc_worklist.append_if_missing(addp2);
4929             }
4930             alloc_worklist.append_if_missing(use);
4931           }
4932         }
4933 
4934         // An allocation may have an Initialize which has raw stores. Scan
4935         // the users of the raw allocation result and push AddP users
4936         // on alloc_worklist.
4937         Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4938         assert (raw_result != nullptr, "must have an allocation result");
4939         for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4940           Node *use = raw_result->fast_out(i);
4941           if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4942             Node* addp2 = find_second_addp(use, raw_result);
4943             if (addp2 != nullptr) {
4944               assert(alloc->is_AllocateArray(),"array allocation was expected");
4945               alloc_worklist.append_if_missing(addp2);
4946             }
4947             alloc_worklist.append_if_missing(use);
4948           } else if (use->is_MemBar()) {
4949             memnode_worklist.append_if_missing(use);
4950           }
4951         }
4952       }
4953     } else if (n->is_AddP()) {
4954       if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4955         // This AddP will go away when we reduce the Phi
4956         continue;
4957       }
4958       Node* addp_base = get_addp_base(n);
4959       JavaObjectNode* jobj = unique_java_object(addp_base);
4960       if (jobj == nullptr || jobj == phantom_obj) {
4961 #ifdef ASSERT
4962         ptnode_adr(get_addp_base(n)->_idx)->dump();
4963         ptnode_adr(n->_idx)->dump();
4964         assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4965 #endif
4966         _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4967         return;
4968       }
4969       Node *base = get_map(jobj->idx());  // CheckCastPP node
4970       if (!split_AddP(n, base)) continue; // wrong type from dead path
4971     } else if (n->is_Phi() ||
4972                n->is_CheckCastPP() ||
4973                n->is_EncodeP() ||
4974                n->is_DecodeN() ||
4975                (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4976       if (visited.test_set(n->_idx)) {
4977         assert(n->is_Phi(), "loops only through Phi's");
4978         continue;  // already processed
4979       }
4980       // Reducible Phi's will be removed from the graph after split_unique_types
4981       // finishes. For now we just try to split out the SR inputs of the merge.
4982       Node* parent = n->in(1);
4983       if (reducible_merges.member(n)) {
4984         reduce_phi(n->as_Phi(), alloc_worklist);
4985 #ifdef ASSERT
4986         if (VerifyReduceAllocationMerges) {
4987           reduced_merges.push(n);
4988         }
4989 #endif
4990         continue;
4991       } else if (reducible_merges.member(parent)) {
4992         // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4993         // part of reduce_merge.
4994         continue;
4995       }
4996       JavaObjectNode* jobj = unique_java_object(n);
4997       if (jobj == nullptr || jobj == phantom_obj) {
4998 #ifdef ASSERT
4999         ptnode_adr(n->_idx)->dump();
5000         assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
5001 #endif
5002         _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
5003         return;
5004       } else {
5005         Node *val = get_map(jobj->idx());   // CheckCastPP node
5006         TypeNode *tn = n->as_Type();
5007         const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
5008         assert(tinst != nullptr && tinst->is_known_instance() &&
5009                tinst->instance_id() == jobj->idx() , "instance type expected.");
5010 
5011         const Type *tn_type = igvn->type(tn);
5012         const TypeOopPtr *tn_t;
5013         if (tn_type->isa_narrowoop()) {
5014           tn_t = tn_type->make_ptr()->isa_oopptr();
5015         } else {
5016           tn_t = tn_type->isa_oopptr();
5017         }
5018         if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
5019           if (tn_t->isa_aryptr()) {
5020             // Keep array properties (not flat/null-free)
5021             tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
5022             if (tinst == nullptr) {
5023               continue; // Skip dead path with inconsistent properties
5024             }
5025           }
5026           if (tn_type->isa_narrowoop()) {
5027             tn_type = tinst->make_narrowoop();
5028           } else {
5029             tn_type = tinst;
5030           }
5031           igvn->hash_delete(tn);
5032           igvn->set_type(tn, tn_type);
5033           tn->set_type(tn_type);
5034           igvn->hash_insert(tn);
5035           record_for_optimizer(n);
5036         } else {
5037           assert(tn_type == TypePtr::NULL_PTR ||
5038                  (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
5039                  "unexpected type");
5040           continue; // Skip dead path with different type
5041         }
5042       }
5043     } else {
5044       DEBUG_ONLY(n->dump();)
5045       assert(false, "EA: unexpected node");
5046       continue;
5047     }
5048     // push allocation's users on appropriate worklist
5049     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5050       Node *use = n->fast_out(i);
5051       if (use->is_Mem() && use->in(MemNode::Address) == n) {
5052         // Load/store to instance's field
5053         memnode_worklist.append_if_missing(use);
5054       } else if (use->is_MemBar()) {
5055         if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5056           memnode_worklist.append_if_missing(use);
5057         }
5058       } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
5059         Node* addp2 = find_second_addp(use, n);
5060         if (addp2 != nullptr) {
5061           alloc_worklist.append_if_missing(addp2);
5062         }
5063         alloc_worklist.append_if_missing(use);
5064       } else if (use->is_Phi() ||
5065                  use->is_CheckCastPP() ||
5066                  use->is_EncodeNarrowPtr() ||
5067                  use->is_DecodeNarrowPtr() ||
5068                  (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
5069         alloc_worklist.append_if_missing(use);
5070 #ifdef ASSERT
5071       } else if (use->is_Mem()) {
5072         assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
5073       } else if (use->is_MergeMem()) {
5074         assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5075       } else if (use->is_SafePoint()) {
5076         // Look for MergeMem nodes for calls which reference unique allocation
5077         // (through CheckCastPP nodes) even for debug info.
5078         Node* m = use->in(TypeFunc::Memory);
5079         if (m->is_MergeMem()) {
5080           assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5081         }
5082       } else if (use->Opcode() == Op_EncodeISOArray) {
5083         if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5084           // EncodeISOArray overwrites destination array
5085           memnode_worklist.append_if_missing(use);
5086         }
5087       } else if (use->Opcode() == Op_Return) {
5088         // Allocation is referenced by field of returned inline type
5089         assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
5090       } else {
5091         uint op = use->Opcode();
5092         if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
5093             (use->in(MemNode::Memory) == n)) {
5094           // They overwrite memory edge corresponding to destination array,
5095           memnode_worklist.append_if_missing(use);
5096         } else if (!(op == Op_CmpP || op == Op_Conv2B ||
5097               op == Op_CastP2X ||
5098               op == Op_FastLock || op == Op_AryEq ||
5099               op == Op_StrComp || op == Op_CountPositives ||
5100               op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
5101               op == Op_StrEquals || op == Op_VectorizedHashCode ||
5102               op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
5103               op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
5104               op == Op_ReinterpretS2HF ||
5105               op == Op_ReachabilityFence ||
5106               BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
5107           n->dump();
5108           use->dump();
5109           assert(false, "EA: missing allocation reference path");
5110         }
5111 #endif
5112       }
5113     }
5114 
5115   }
5116 
5117 #ifdef ASSERT
5118   if (VerifyReduceAllocationMerges) {
5119     for (uint i = 0; i < reducible_merges.size(); i++) {
5120       Node* phi = reducible_merges.at(i);
5121 
5122       if (!reduced_merges.member(phi)) {
5123         phi->dump(2);
5124         phi->dump(-2);
5125         assert(false, "This reducible merge wasn't reduced.");
5126       }
5127 
5128       // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
5129       for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
5130         Node* use = phi->fast_out(j);
5131         if (!use->is_SafePoint() && !use->is_CastPP()) {
5132           phi->dump(2);
5133           phi->dump(-2);
5134           assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
5135         }
5136       }
5137     }
5138   }
5139 #endif
5140 
5141   // Go over all ArrayCopy nodes and if one of the inputs has a unique
5142   // type, record it in the ArrayCopy node so we know what memory this
5143   // node uses/modified.
5144   for (int next = 0; next < arraycopy_worklist.length(); next++) {
5145     ArrayCopyNode* ac = arraycopy_worklist.at(next);
5146     Node* dest = ac->in(ArrayCopyNode::Dest);
5147     if (dest->is_AddP()) {
5148       dest = get_addp_base(dest);
5149     }
5150     JavaObjectNode* jobj = unique_java_object(dest);
5151     if (jobj != nullptr) {
5152       Node *base = get_map(jobj->idx());
5153       if (base != nullptr) {
5154         const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5155         ac->_dest_type = base_t;
5156       }
5157     }
5158     Node* src = ac->in(ArrayCopyNode::Src);
5159     if (src->is_AddP()) {
5160       src = get_addp_base(src);
5161     }
5162     jobj = unique_java_object(src);
5163     if (jobj != nullptr) {
5164       Node* base = get_map(jobj->idx());
5165       if (base != nullptr) {
5166         const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5167         ac->_src_type = base_t;
5168       }
5169     }
5170   }
5171 
5172   // New alias types were created in split_AddP().
5173   uint new_index_end = (uint) _compile->num_alias_types();
5174 
5175   _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
5176 
5177   //  Phase 2:  Process MemNode's from memnode_worklist. compute new address type and
5178   //            compute new values for Memory inputs  (the Memory inputs are not
5179   //            actually updated until phase 4.)
5180   if (memnode_worklist.length() == 0)
5181     return;  // nothing to do
5182   while (memnode_worklist.length() != 0) {
5183     Node *n = memnode_worklist.pop();
5184     if (visited.test_set(n->_idx)) {
5185       continue;
5186     }
5187     if (n->is_Phi()) {
5188       if ((uint) _compile->get_alias_index(n->as_Phi()->adr_type()) < new_index_start) {
5189         // Push memory phis on the orig_phis worklist to update
5190         // during Phase 4 if needed.
5191         orig_phis.append_if_missing(n->as_Phi());
5192       }
5193     } else if (n->is_ClearArray()) {
5194      // we don't need to do anything, but the users must be pushed
5195     } else if (n->is_MemBar()) { // MemBar nodes
5196       if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
5197         // we don't need to do anything, but the users must be pushed
5198         n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
5199         if (n == nullptr) {
5200           continue;
5201         }
5202       }
5203     } else if (n->is_CallLeaf()) {
5204       // Runtime calls with narrow memory input (no MergeMem node)
5205       // get the memory projection
5206       n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
5207       if (n == nullptr) {
5208         continue;
5209       }
5210     } else if (n->Opcode() == Op_StrInflatedCopy) {
5211       // Check direct uses of StrInflatedCopy.
5212       // It is memory type Node - no special SCMemProj node.
5213     } else if (n->Opcode() == Op_StrCompressedCopy ||
5214                n->Opcode() == Op_EncodeISOArray) {
5215       // get the memory projection
5216       n = n->find_out_with(Op_SCMemProj);
5217       assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5218     } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
5219                strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5220       n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
5221     } else if (n->is_Proj()) {
5222       assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
5223     } else {
5224 #ifdef ASSERT
5225       if (!n->is_Mem()) {
5226         n->dump();
5227       }
5228       assert(n->is_Mem(), "memory node required.");
5229 #endif
5230       Node *addr = n->in(MemNode::Address);
5231       const Type *addr_t = igvn->type(addr);
5232       if (addr_t == Type::TOP) {
5233         continue;
5234       }
5235       assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
5236       int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
5237       assert ((uint)alias_idx < new_index_end, "wrong alias index");
5238       Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
5239       if (_compile->failing()) {
5240         return;
5241       }
5242       if (mem != n->in(MemNode::Memory)) {
5243         // We delay the memory edge update since we need old one in
5244         // MergeMem code below when instances memory slices are separated.
5245         set_map(n, mem);
5246       }
5247       if (n->is_Load()) {
5248         continue;  // don't push users
5249       } else if (n->is_LoadStore()) {
5250         // get the memory projection
5251         n = n->find_out_with(Op_SCMemProj);
5252         assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5253       }
5254     }
5255     // push user on appropriate worklist
5256     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5257       Node *use = n->fast_out(i);
5258       if (use->is_Phi() || use->is_ClearArray()) {
5259         memnode_worklist.append_if_missing(use);
5260       } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
5261         memnode_worklist.append_if_missing(use);
5262       } else if (use->is_MemBar() || use->is_CallLeaf()) {
5263         if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5264           memnode_worklist.append_if_missing(use);
5265         }
5266       } else if (use->is_Proj()) {
5267         assert(n->is_Initialize(), "We only push projections of Initialize");
5268         if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
5269           memnode_worklist.append_if_missing(use);
5270         }
5271 #ifdef ASSERT
5272       } else if (use->is_Mem()) {
5273         assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
5274       } else if (use->is_MergeMem()) {
5275         assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5276       } else if (use->Opcode() == Op_EncodeISOArray) {
5277         if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5278           // EncodeISOArray overwrites destination array
5279           memnode_worklist.append_if_missing(use);
5280         }
5281       } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
5282                  strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5283         // store_unknown_inline overwrites destination array
5284         memnode_worklist.append_if_missing(use);
5285       } else {
5286         uint op = use->Opcode();
5287         if ((use->in(MemNode::Memory) == n) &&
5288             (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
5289           // They overwrite memory edge corresponding to destination array,
5290           memnode_worklist.append_if_missing(use);
5291         } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
5292               op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
5293               op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
5294               op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
5295           n->dump();
5296           use->dump();
5297           assert(false, "EA: missing memory path");
5298         }
5299 #endif
5300       }
5301     }
5302   }
5303 
5304   //  Phase 3:  Process MergeMem nodes from mergemem_worklist.
5305   //            Walk each memory slice moving the first node encountered of each
5306   //            instance type to the input corresponding to its alias index.
5307   uint length = mergemem_worklist.length();
5308   for( uint next = 0; next < length; ++next ) {
5309     MergeMemNode* nmm = mergemem_worklist.at(next);
5310     assert(!visited.test_set(nmm->_idx), "should not be visited before");
5311     // Note: we don't want to use MergeMemStream here because we only want to
5312     // scan inputs which exist at the start, not ones we add during processing.
5313     // Note 2: MergeMem may already contains instance memory slices added
5314     // during find_inst_mem() call when memory nodes were processed above.
5315     igvn->hash_delete(nmm);
5316     uint nslices = MIN2(nmm->req(), new_index_start);
5317     for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
5318       Node* mem = nmm->in(i);
5319       Node* cur = nullptr;
5320       if (mem == nullptr || mem->is_top()) {
5321         continue;
5322       }
5323       // First, update mergemem by moving memory nodes to corresponding slices
5324       // if their type became more precise since this mergemem was created.
5325       while (mem->is_Mem()) {
5326         const Type* at = igvn->type(mem->in(MemNode::Address));
5327         if (at != Type::TOP) {
5328           assert (at->isa_ptr() != nullptr, "pointer type required.");
5329           uint idx = (uint)_compile->get_alias_index(at->is_ptr());
5330           if (idx == i) {
5331             if (cur == nullptr) {
5332               cur = mem;
5333             }
5334           } else {
5335             if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
5336               nmm->set_memory_at(idx, mem);
5337             }
5338           }
5339         }
5340         mem = mem->in(MemNode::Memory);
5341       }
5342       nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
5343       // Find any instance of the current type if we haven't encountered
5344       // already a memory slice of the instance along the memory chain.
5345       for (uint ni = new_index_start; ni < new_index_end; ni++) {
5346         if((uint)_compile->get_general_index(ni) == i) {
5347           Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
5348           if (nmm->is_empty_memory(m)) {
5349             Node* result = find_inst_mem(mem, ni, orig_phis);
5350             if (_compile->failing()) {
5351               return;
5352             }
5353             nmm->set_memory_at(ni, result);
5354           }
5355         }
5356       }
5357     }
5358     // Find the rest of instances values
5359     for (uint ni = new_index_start; ni < new_index_end; ni++) {
5360       const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
5361       Node* result = step_through_mergemem(nmm, ni, tinst);
5362       if (result == nmm->base_memory()) {
5363         // Didn't find instance memory, search through general slice recursively.
5364         result = nmm->memory_at(_compile->get_general_index(ni));
5365         result = find_inst_mem(result, ni, orig_phis);
5366         if (_compile->failing()) {
5367           return;
5368         }
5369         nmm->set_memory_at(ni, result);
5370       }
5371     }
5372 
5373     // If we have crossed the 3/4 point of max node limit it's too risky
5374     // to continue with EA/SR because we might hit the max node limit.
5375     if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
5376       if (_compile->do_reduce_allocation_merges()) {
5377         _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
5378       } else if (_invocation > 0) {
5379         _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
5380       } else {
5381         _compile->record_failure(C2Compiler::retry_no_escape_analysis());
5382       }
5383       return;
5384     }
5385 
5386     igvn->hash_insert(nmm);
5387     record_for_optimizer(nmm);
5388   }
5389 
5390   _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5391 
5392   //  Phase 4:  Update the inputs of non-instance memory Phis and
5393   //            the Memory input of memnodes
5394   // First update the inputs of any non-instance Phi's from
5395   // which we split out an instance Phi.  Note we don't have
5396   // to recursively process Phi's encountered on the input memory
5397   // chains as is done in split_memory_phi() since they will
5398   // also be processed here.
5399   for (int j = 0; j < orig_phis.length(); j++) {
5400     PhiNode *phi = orig_phis.at(j);
5401     int alias_idx = _compile->get_alias_index(phi->adr_type());
5402     igvn->hash_delete(phi);
5403     for (uint i = 1; i < phi->req(); i++) {
5404       Node *mem = phi->in(i);
5405       Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5406       if (_compile->failing()) {
5407         return;
5408       }
5409       if (mem != new_mem) {
5410         phi->set_req(i, new_mem);
5411       }
5412     }
5413     igvn->hash_insert(phi);
5414     record_for_optimizer(phi);
5415   }
5416 
5417   // Update the memory inputs of MemNodes with the value we computed
5418   // in Phase 2 and move stores memory users to corresponding memory slices.
5419   // Disable memory split verification code until the fix for 6984348.
5420   // Currently it produces false negative results since it does not cover all cases.
5421 #if 0 // ifdef ASSERT
5422   visited.Reset();
5423   Node_Stack old_mems(arena, _compile->unique() >> 2);
5424 #endif
5425   for (uint i = 0; i < ideal_nodes.size(); i++) {
5426     Node*    n = ideal_nodes.at(i);
5427     Node* nmem = get_map(n->_idx);
5428     assert(nmem != nullptr, "sanity");
5429     if (n->is_Mem()) {
5430 #if 0 // ifdef ASSERT
5431       Node* old_mem = n->in(MemNode::Memory);
5432       if (!visited.test_set(old_mem->_idx)) {
5433         old_mems.push(old_mem, old_mem->outcnt());
5434       }
5435 #endif
5436       assert(n->in(MemNode::Memory) != nmem, "sanity");
5437       if (!n->is_Load()) {
5438         // Move memory users of a store first.
5439         move_inst_mem(n, orig_phis);
5440       }
5441       // Now update memory input
5442       igvn->hash_delete(n);
5443       n->set_req(MemNode::Memory, nmem);
5444       igvn->hash_insert(n);
5445       record_for_optimizer(n);
5446     } else {
5447       assert(n->is_Allocate() || n->is_CheckCastPP() ||
5448              n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5449     }
5450   }
5451 #if 0 // ifdef ASSERT
5452   // Verify that memory was split correctly
5453   while (old_mems.is_nonempty()) {
5454     Node* old_mem = old_mems.node();
5455     uint  old_cnt = old_mems.index();
5456     old_mems.pop();
5457     assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5458   }
5459 #endif
5460   _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5461 }
5462 
5463 #ifndef PRODUCT
5464 int ConnectionGraph::_no_escape_counter = 0;
5465 int ConnectionGraph::_arg_escape_counter = 0;
5466 int ConnectionGraph::_global_escape_counter = 0;
5467 
5468 static const char *node_type_names[] = {
5469   "UnknownType",
5470   "JavaObject",
5471   "LocalVar",
5472   "Field",
5473   "Arraycopy"
5474 };
5475 
5476 static const char *esc_names[] = {
5477   "UnknownEscape",
5478   "NoEscape",
5479   "ArgEscape",
5480   "GlobalEscape"
5481 };
5482 
5483 const char* PointsToNode::esc_name() const {
5484   return esc_names[(int)escape_state()];
5485 }
5486 
5487 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5488   NodeType nt = node_type();
5489   out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5490   if (print_state) {
5491     EscapeState es = escape_state();
5492     EscapeState fields_es = fields_escape_state();
5493     out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5494     if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5495       out->print("NSR ");
5496     }
5497   }
5498 }
5499 
5500 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5501   dump_header(print_state, out);
5502   if (is_Field()) {
5503     FieldNode* f = (FieldNode*)this;
5504     if (f->is_oop()) {
5505       out->print("oop ");
5506     }
5507     if (f->offset() > 0) {
5508       out->print("+%d ", f->offset());
5509     }
5510     out->print("(");
5511     for (BaseIterator i(f); i.has_next(); i.next()) {
5512       PointsToNode* b = i.get();
5513       out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5514     }
5515     out->print(" )");
5516   }
5517   out->print("[");
5518   for (EdgeIterator i(this); i.has_next(); i.next()) {
5519     PointsToNode* e = i.get();
5520     out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5521   }
5522   out->print(" [");
5523   for (UseIterator i(this); i.has_next(); i.next()) {
5524     PointsToNode* u = i.get();
5525     bool is_base = false;
5526     if (PointsToNode::is_base_use(u)) {
5527       is_base = true;
5528       u = PointsToNode::get_use_node(u)->as_Field();
5529     }
5530     out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5531   }
5532   out->print(" ]]  ");
5533   if (_node == nullptr) {
5534     out->print("<null>%s", newline ? "\n" : "");
5535   } else {
5536     _node->dump(newline ? "\n" : "", false, out);
5537   }
5538 }
5539 
5540 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5541   bool first = true;
5542   int ptnodes_length = ptnodes_worklist.length();
5543   for (int i = 0; i < ptnodes_length; i++) {
5544     PointsToNode *ptn = ptnodes_worklist.at(i);
5545     if (ptn == nullptr || !ptn->is_JavaObject()) {
5546       continue;
5547     }
5548     PointsToNode::EscapeState es = ptn->escape_state();
5549     if ((es != PointsToNode::NoEscape) && !Verbose) {
5550       continue;
5551     }
5552     Node* n = ptn->ideal_node();
5553     if (n->is_Allocate() || (n->is_CallStaticJava() &&
5554                              n->as_CallStaticJava()->is_boxing_method())) {
5555       if (first) {
5556         tty->cr();
5557         tty->print("======== Connection graph for ");
5558         _compile->method()->print_short_name();
5559         tty->cr();
5560         tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5561                       _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5562         tty->cr();
5563         first = false;
5564       }
5565       ptn->dump();
5566       // Print all locals and fields which reference this allocation
5567       for (UseIterator j(ptn); j.has_next(); j.next()) {
5568         PointsToNode* use = j.get();
5569         if (use->is_LocalVar()) {
5570           use->dump(Verbose);
5571         } else if (Verbose) {
5572           use->dump();
5573         }
5574       }
5575       tty->cr();
5576     }
5577   }
5578 }
5579 
5580 void ConnectionGraph::print_statistics() {
5581   tty->print_cr("No escape = %d, Arg escape = %d, Global escape = %d", AtomicAccess::load(&_no_escape_counter), AtomicAccess::load(&_arg_escape_counter), AtomicAccess::load(&_global_escape_counter));
5582 }
5583 
5584 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5585   if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5586     return;
5587   }
5588   for (int next = 0; next < java_objects_worklist.length(); ++next) {
5589     JavaObjectNode* ptn = java_objects_worklist.at(next);
5590     if (ptn->ideal_node()->is_Allocate()) {
5591       if (ptn->escape_state() == PointsToNode::NoEscape) {
5592         AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5593       } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5594         AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5595       } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5596         AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5597       } else {
5598         assert(false, "Unexpected Escape State");
5599       }
5600     }
5601   }
5602 }
5603 
5604 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5605   if (_compile->directive()->TraceEscapeAnalysisOption) {
5606     assert(ptn != nullptr, "should not be null");
5607     assert(reason != nullptr, "should not be null");
5608     ptn->dump_header(true);
5609     PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5610     PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5611     tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5612   }
5613 }
5614 
5615 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5616   if (_compile->directive()->TraceEscapeAnalysisOption) {
5617     stringStream ss;
5618     ss.print("propagated from: ");
5619     from->dump(true, &ss, false);
5620     return ss.as_string();
5621   } else {
5622     return nullptr;
5623   }
5624 }
5625 
5626 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5627   if (_compile->directive()->TraceEscapeAnalysisOption) {
5628     stringStream ss;
5629     ss.print("escapes as arg to:");
5630     call->dump("", false, &ss);
5631     return ss.as_string();
5632   } else {
5633     return nullptr;
5634   }
5635 }
5636 
5637 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5638   if (_compile->directive()->TraceEscapeAnalysisOption) {
5639     stringStream ss;
5640     ss.print("is merged with other object: ");
5641     other->dump_header(true, &ss);
5642     return ss.as_string();
5643   } else {
5644     return nullptr;
5645   }
5646 }
5647 
5648 #endif
5649 
5650 void ConnectionGraph::record_for_optimizer(Node *n) {
5651   _igvn->_worklist.push(n);
5652   _igvn->add_users_to_worklist(n);
5653 }