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