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