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