1 /* 2 * Copyright (c) 2005, 2018, 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 "precompiled.hpp" 26 #include "ci/bcEscapeAnalyzer.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "libadt/vectset.hpp" 31 #include "memory/allocation.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/c2compiler.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/callnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/compile.hpp" 38 #include "opto/escape.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/movenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "utilities/macros.hpp" 43 44 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 45 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL), 46 _in_worklist(C->comp_arena()), 47 _next_pidx(0), 48 _collecting(true), 49 _verify(false), 50 _compile(C), 51 _igvn(igvn), 52 _node_map(C->comp_arena()) { 53 // Add unknown java object. 54 add_java_object(C->top(), PointsToNode::GlobalEscape); 55 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 56 // Add ConP(#NULL) and ConN(#NULL) nodes. 57 Node* oop_null = igvn->zerocon(T_OBJECT); 58 assert(oop_null->_idx < nodes_size(), "should be created already"); 59 add_java_object(oop_null, PointsToNode::NoEscape); 60 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 61 if (UseCompressedOops) { 62 Node* noop_null = igvn->zerocon(T_NARROWOOP); 63 assert(noop_null->_idx < nodes_size(), "should be created already"); 64 map_ideal_node(noop_null, null_obj); 65 } 66 _pcmp_neq = NULL; // Should be initialized 67 _pcmp_eq = NULL; 68 } 69 70 bool ConnectionGraph::has_candidates(Compile *C) { 71 // EA brings benefits only when the code has allocations and/or locks which 72 // are represented by ideal Macro nodes. 73 int cnt = C->macro_count(); 74 for (int i = 0; i < cnt; i++) { 75 Node *n = C->macro_node(i); 76 if (n->is_Allocate()) 77 return true; 78 if (n->is_Lock()) { 79 Node* obj = n->as_Lock()->obj_node()->uncast(); 80 if (!(obj->is_Parm() || obj->is_Con())) 81 return true; 82 } 83 if (n->is_CallStaticJava() && 84 n->as_CallStaticJava()->is_boxing_method()) { 85 return true; 86 } 87 } 88 return false; 89 } 90 91 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 92 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]); 93 ResourceMark rm; 94 95 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 96 // to create space for them in ConnectionGraph::_nodes[]. 97 Node* oop_null = igvn->zerocon(T_OBJECT); 98 Node* noop_null = igvn->zerocon(T_NARROWOOP); 99 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 100 // Perform escape analysis 101 if (congraph->compute_escape()) { 102 // There are non escaping objects. 103 C->set_congraph(congraph); 104 } 105 // Cleanup. 106 if (oop_null->outcnt() == 0) 107 igvn->hash_delete(oop_null); 108 if (noop_null->outcnt() == 0) 109 igvn->hash_delete(noop_null); 110 } 111 112 bool ConnectionGraph::compute_escape() { 113 Compile* C = _compile; 114 PhaseGVN* igvn = _igvn; 115 116 // Worklists used by EA. 117 Unique_Node_List delayed_worklist; 118 GrowableArray<Node*> alloc_worklist; 119 GrowableArray<Node*> ptr_cmp_worklist; 120 GrowableArray<Node*> storestore_worklist; 121 GrowableArray<ArrayCopyNode*> arraycopy_worklist; 122 GrowableArray<PointsToNode*> ptnodes_worklist; 123 GrowableArray<JavaObjectNode*> java_objects_worklist; 124 GrowableArray<JavaObjectNode*> non_escaped_worklist; 125 GrowableArray<FieldNode*> oop_fields_worklist; 126 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 127 128 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]); 129 130 // 1. Populate Connection Graph (CG) with PointsTo nodes. 131 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 132 // Initialize worklist 133 if (C->root() != NULL) { 134 ideal_nodes.push(C->root()); 135 } 136 // Processed ideal nodes are unique on ideal_nodes list 137 // but several ideal nodes are mapped to the phantom_obj. 138 // To avoid duplicated entries on the following worklists 139 // add the phantom_obj only once to them. 140 ptnodes_worklist.append(phantom_obj); 141 java_objects_worklist.append(phantom_obj); 142 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 143 Node* n = ideal_nodes.at(next); 144 if ((n->Opcode() == Op_LoadX || n->Opcode() == Op_StoreX) && 145 !n->in(MemNode::Address)->is_AddP() && 146 _igvn->type(n->in(MemNode::Address))->isa_oopptr()) { 147 // Load/Store at mark work address is at offset 0 so has no AddP which confuses EA 148 Node* addp = new AddPNode(n->in(MemNode::Address), n->in(MemNode::Address), _igvn->MakeConX(0)); 149 _igvn->register_new_node_with_optimizer(addp); 150 _igvn->replace_input_of(n, MemNode::Address, addp); 151 ideal_nodes.push(addp); 152 _nodes.at_put_grow(addp->_idx, NULL, NULL); 153 } 154 // Create PointsTo nodes and add them to Connection Graph. Called 155 // only once per ideal node since ideal_nodes is Unique_Node list. 156 add_node_to_connection_graph(n, &delayed_worklist); 157 PointsToNode* ptn = ptnode_adr(n->_idx); 158 if (ptn != NULL && ptn != phantom_obj) { 159 ptnodes_worklist.append(ptn); 160 if (ptn->is_JavaObject()) { 161 java_objects_worklist.append(ptn->as_JavaObject()); 162 if ((n->is_Allocate() || n->is_CallStaticJava()) && 163 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 164 // Only allocations and java static calls results are interesting. 165 non_escaped_worklist.append(ptn->as_JavaObject()); 166 } 167 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 168 oop_fields_worklist.append(ptn->as_Field()); 169 } 170 } 171 if (n->is_MergeMem()) { 172 // Collect all MergeMem nodes to add memory slices for 173 // scalar replaceable objects in split_unique_types(). 174 _mergemem_worklist.append(n->as_MergeMem()); 175 } else if (OptimizePtrCompare && n->is_Cmp() && 176 ((n->Opcode() == Op_CmpP && !(((CmpPNode*)n)->has_perturbed_operand() != NULL)) || 177 n->Opcode() == Op_CmpN)) { 178 // Collect compare pointers nodes. 179 ptr_cmp_worklist.append(n); 180 } else if (n->is_MemBarStoreStore()) { 181 // Collect all MemBarStoreStore nodes so that depending on the 182 // escape status of the associated Allocate node some of them 183 // may be eliminated. 184 storestore_worklist.append(n); 185 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 186 (n->req() > MemBarNode::Precedent)) { 187 record_for_optimizer(n); 188 #ifdef ASSERT 189 } else if (n->is_AddP()) { 190 // Collect address nodes for graph verification. 191 addp_worklist.append(n); 192 #endif 193 } else if (n->is_ArrayCopy()) { 194 // Keep a list of ArrayCopy nodes so if one of its input is non 195 // escaping, we can record a unique type 196 arraycopy_worklist.append(n->as_ArrayCopy()); 197 } 198 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 199 Node* m = n->fast_out(i); // Get user 200 ideal_nodes.push(m); 201 } 202 } 203 if (non_escaped_worklist.length() == 0) { 204 _collecting = false; 205 return false; // Nothing to do. 206 } 207 // Add final simple edges to graph. 208 while(delayed_worklist.size() > 0) { 209 Node* n = delayed_worklist.pop(); 210 add_final_edges(n); 211 } 212 int ptnodes_length = ptnodes_worklist.length(); 213 214 #ifdef ASSERT 215 if (VerifyConnectionGraph) { 216 // Verify that no new simple edges could be created and all 217 // local vars has edges. 218 _verify = true; 219 for (int next = 0; next < ptnodes_length; ++next) { 220 PointsToNode* ptn = ptnodes_worklist.at(next); 221 add_final_edges(ptn->ideal_node()); 222 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 223 ptn->dump(); 224 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 225 } 226 } 227 _verify = false; 228 } 229 #endif 230 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes 231 // processing, calls to CI to resolve symbols (types, fields, methods) 232 // referenced in bytecode. During symbol resolution VM may throw 233 // an exception which CI cleans and converts to compilation failure. 234 if (C->failing()) return false; 235 236 // 2. Finish Graph construction by propagating references to all 237 // java objects through graph. 238 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 239 java_objects_worklist, oop_fields_worklist)) { 240 // All objects escaped or hit time or iterations limits. 241 _collecting = false; 242 return false; 243 } 244 245 // 3. Adjust scalar_replaceable state of nonescaping objects and push 246 // scalar replaceable allocations on alloc_worklist for processing 247 // in split_unique_types(). 248 int non_escaped_length = non_escaped_worklist.length(); 249 for (int next = 0; next < non_escaped_length; next++) { 250 JavaObjectNode* ptn = non_escaped_worklist.at(next); 251 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 252 Node* n = ptn->ideal_node(); 253 if (n->is_Allocate()) { 254 n->as_Allocate()->_is_non_escaping = noescape; 255 } 256 if (n->is_CallStaticJava()) { 257 n->as_CallStaticJava()->_is_non_escaping = noescape; 258 } 259 if (noescape && ptn->scalar_replaceable()) { 260 adjust_scalar_replaceable_state(ptn); 261 if (ptn->scalar_replaceable()) { 262 alloc_worklist.append(ptn->ideal_node()); 263 } 264 } 265 } 266 267 #ifdef ASSERT 268 if (VerifyConnectionGraph) { 269 // Verify that graph is complete - no new edges could be added or needed. 270 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 271 java_objects_worklist, addp_worklist); 272 } 273 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 274 assert(null_obj->escape_state() == PointsToNode::NoEscape && 275 null_obj->edge_count() == 0 && 276 !null_obj->arraycopy_src() && 277 !null_obj->arraycopy_dst(), "sanity"); 278 #endif 279 280 _collecting = false; 281 282 } // TracePhase t3("connectionGraph") 283 284 // 4. Optimize ideal graph based on EA information. 285 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 286 if (has_non_escaping_obj) { 287 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 288 } 289 290 #ifndef PRODUCT 291 if (PrintEscapeAnalysis) { 292 dump(ptnodes_worklist); // Dump ConnectionGraph 293 } 294 #endif 295 296 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 297 #ifdef ASSERT 298 if (VerifyConnectionGraph) { 299 int alloc_length = alloc_worklist.length(); 300 for (int next = 0; next < alloc_length; ++next) { 301 Node* n = alloc_worklist.at(next); 302 PointsToNode* ptn = ptnode_adr(n->_idx); 303 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 304 } 305 } 306 #endif 307 308 // 5. Separate memory graph for scalar replaceable allcations. 309 if (has_scalar_replaceable_candidates && 310 C->AliasLevel() >= 3 && EliminateAllocations) { 311 // Now use the escape information to create unique types for 312 // scalar replaceable objects. 313 split_unique_types(alloc_worklist, arraycopy_worklist); 314 if (C->failing()) return false; 315 C->print_method(PHASE_AFTER_EA, 2); 316 317 #ifdef ASSERT 318 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 319 tty->print("=== No allocations eliminated for "); 320 C->method()->print_short_name(); 321 if(!EliminateAllocations) { 322 tty->print(" since EliminateAllocations is off ==="); 323 } else if(!has_scalar_replaceable_candidates) { 324 tty->print(" since there are no scalar replaceable candidates ==="); 325 } else if(C->AliasLevel() < 3) { 326 tty->print(" since AliasLevel < 3 ==="); 327 } 328 tty->cr(); 329 #endif 330 } 331 return has_non_escaping_obj; 332 } 333 334 // Utility function for nodes that load an object 335 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 336 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 337 // ThreadLocal has RawPtr type. 338 const Type* t = _igvn->type(n); 339 if (t->make_ptr() != NULL) { 340 Node* adr = n->in(MemNode::Address); 341 #ifdef ASSERT 342 if (!adr->is_AddP()) { 343 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 344 } else { 345 assert((ptnode_adr(adr->_idx) == NULL || 346 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 347 } 348 #endif 349 add_local_var_and_edge(n, PointsToNode::NoEscape, 350 adr, delayed_worklist); 351 } 352 } 353 354 // Populate Connection Graph with PointsTo nodes and create simple 355 // connection graph edges. 356 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 357 assert(!_verify, "this method should not be called for verification"); 358 PhaseGVN* igvn = _igvn; 359 uint n_idx = n->_idx; 360 PointsToNode* n_ptn = ptnode_adr(n_idx); 361 if (n_ptn != NULL) 362 return; // No need to redefine PointsTo node during first iteration. 363 364 if (n->is_Call()) { 365 // Arguments to allocation and locking don't escape. 366 if (n->is_AbstractLock()) { 367 // Put Lock and Unlock nodes on IGVN worklist to process them during 368 // first IGVN optimization when escape information is still available. 369 record_for_optimizer(n); 370 } else if (n->is_Allocate()) { 371 add_call_node(n->as_Call()); 372 record_for_optimizer(n); 373 } else { 374 if (n->is_CallStaticJava()) { 375 const char* name = n->as_CallStaticJava()->_name; 376 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 377 return; // Skip uncommon traps 378 } 379 // Don't mark as processed since call's arguments have to be processed. 380 delayed_worklist->push(n); 381 // Check if a call returns an object. 382 if ((n->as_Call()->returns_pointer() && 383 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) || 384 (n->is_CallStaticJava() && 385 n->as_CallStaticJava()->is_boxing_method())) { 386 add_call_node(n->as_Call()); 387 } else if (n->as_Call()->tf()->returns_value_type_as_fields()) { 388 bool returns_oop = false; 389 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) { 390 ProjNode* pn = n->fast_out(i)->as_Proj(); 391 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) { 392 returns_oop = true; 393 } 394 } 395 if (returns_oop) { 396 add_call_node(n->as_Call()); 397 } 398 } 399 } 400 return; 401 } 402 // Put this check here to process call arguments since some call nodes 403 // point to phantom_obj. 404 if (n_ptn == phantom_obj || n_ptn == null_obj) 405 return; // Skip predefined nodes. 406 407 int opcode = n->Opcode(); 408 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode); 409 if (gc_handled) { 410 return; // Ignore node if already handled by GC. 411 } 412 switch (opcode) { 413 case Op_AddP: { 414 Node* base = get_addp_base(n); 415 PointsToNode* ptn_base = ptnode_adr(base->_idx); 416 // Field nodes are created for all field types. They are used in 417 // adjust_scalar_replaceable_state() and split_unique_types(). 418 // Note, non-oop fields will have only base edges in Connection 419 // Graph because such fields are not used for oop loads and stores. 420 int offset = address_offset(n, igvn); 421 add_field(n, PointsToNode::NoEscape, offset); 422 if (ptn_base == NULL) { 423 delayed_worklist->push(n); // Process it later. 424 } else { 425 n_ptn = ptnode_adr(n_idx); 426 add_base(n_ptn->as_Field(), ptn_base); 427 } 428 break; 429 } 430 case Op_CastX2P: { 431 map_ideal_node(n, phantom_obj); 432 break; 433 } 434 case Op_CastPP: 435 case Op_CheckCastPP: 436 case Op_EncodeP: 437 case Op_DecodeN: 438 case Op_EncodePKlass: 439 case Op_DecodeNKlass: { 440 add_local_var_and_edge(n, PointsToNode::NoEscape, 441 n->in(1), delayed_worklist); 442 break; 443 } 444 case Op_CMoveP: { 445 add_local_var(n, PointsToNode::NoEscape); 446 // Do not add edges during first iteration because some could be 447 // not defined yet. 448 delayed_worklist->push(n); 449 break; 450 } 451 case Op_ConP: 452 case Op_ConN: 453 case Op_ConNKlass: { 454 // assume all oop constants globally escape except for null 455 PointsToNode::EscapeState es; 456 const Type* t = igvn->type(n); 457 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 458 es = PointsToNode::NoEscape; 459 } else { 460 es = PointsToNode::GlobalEscape; 461 } 462 add_java_object(n, es); 463 break; 464 } 465 case Op_CreateEx: { 466 // assume that all exception objects globally escape 467 map_ideal_node(n, phantom_obj); 468 break; 469 } 470 case Op_LoadKlass: 471 case Op_LoadNKlass: { 472 // Unknown class is loaded 473 map_ideal_node(n, phantom_obj); 474 break; 475 } 476 case Op_LoadP: 477 case Op_LoadN: 478 case Op_LoadPLocked: { 479 add_objload_to_connection_graph(n, delayed_worklist); 480 break; 481 } 482 case Op_Parm: { 483 map_ideal_node(n, phantom_obj); 484 break; 485 } 486 case Op_PartialSubtypeCheck: { 487 // Produces Null or notNull and is used in only in CmpP so 488 // phantom_obj could be used. 489 map_ideal_node(n, phantom_obj); // Result is unknown 490 break; 491 } 492 case Op_Phi: { 493 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 494 // ThreadLocal has RawPtr type. 495 const Type* t = n->as_Phi()->type(); 496 if (t->make_ptr() != NULL) { 497 add_local_var(n, PointsToNode::NoEscape); 498 // Do not add edges during first iteration because some could be 499 // not defined yet. 500 delayed_worklist->push(n); 501 } 502 break; 503 } 504 case Op_Proj: { 505 // we are only interested in the oop result projection from a call 506 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() && 507 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) { 508 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) || 509 n->in(0)->as_Call()->tf()->returns_value_type_as_fields(), "what kind of oop return is it?"); 510 add_local_var_and_edge(n, PointsToNode::NoEscape, 511 n->in(0), delayed_worklist); 512 } 513 break; 514 } 515 case Op_Rethrow: // Exception object escapes 516 case Op_Return: { 517 if (n->req() > TypeFunc::Parms && 518 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 519 // Treat Return value as LocalVar with GlobalEscape escape state. 520 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 521 n->in(TypeFunc::Parms), delayed_worklist); 522 } 523 break; 524 } 525 case Op_CompareAndExchangeP: 526 case Op_CompareAndExchangeN: 527 case Op_GetAndSetP: 528 case Op_GetAndSetN: { 529 add_objload_to_connection_graph(n, delayed_worklist); 530 // fallthrough 531 } 532 case Op_StoreP: 533 case Op_StoreN: 534 case Op_StoreNKlass: 535 case Op_StorePConditional: 536 case Op_WeakCompareAndSwapP: 537 case Op_WeakCompareAndSwapN: 538 case Op_CompareAndSwapP: 539 case Op_CompareAndSwapN: { 540 add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 541 break; 542 } 543 case Op_AryEq: 544 case Op_HasNegatives: 545 case Op_StrComp: 546 case Op_StrEquals: 547 case Op_StrIndexOf: 548 case Op_StrIndexOfChar: 549 case Op_StrInflatedCopy: 550 case Op_StrCompressedCopy: 551 case Op_EncodeISOArray: { 552 add_local_var(n, PointsToNode::ArgEscape); 553 delayed_worklist->push(n); // Process it later. 554 break; 555 } 556 case Op_ThreadLocal: { 557 add_java_object(n, PointsToNode::ArgEscape); 558 break; 559 } 560 default: 561 ; // Do nothing for nodes not related to EA. 562 } 563 return; 564 } 565 566 #ifdef ASSERT 567 #define ELSE_FAIL(name) \ 568 /* Should not be called for not pointer type. */ \ 569 n->dump(1); \ 570 assert(false, name); \ 571 break; 572 #else 573 #define ELSE_FAIL(name) \ 574 break; 575 #endif 576 577 // Add final simple edges to graph. 578 void ConnectionGraph::add_final_edges(Node *n) { 579 PointsToNode* n_ptn = ptnode_adr(n->_idx); 580 #ifdef ASSERT 581 if (_verify && n_ptn->is_JavaObject()) 582 return; // This method does not change graph for JavaObject. 583 #endif 584 585 if (n->is_Call()) { 586 process_call_arguments(n->as_Call()); 587 return; 588 } 589 assert(n->is_Store() || n->is_LoadStore() || 590 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 591 "node should be registered already"); 592 int opcode = n->Opcode(); 593 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode); 594 if (gc_handled) { 595 return; // Ignore node if already handled by GC. 596 } 597 switch (opcode) { 598 case Op_AddP: { 599 Node* base = get_addp_base(n); 600 PointsToNode* ptn_base = ptnode_adr(base->_idx); 601 assert(ptn_base != NULL, "field's base should be registered"); 602 add_base(n_ptn->as_Field(), ptn_base); 603 break; 604 } 605 case Op_CastPP: 606 case Op_CheckCastPP: 607 case Op_EncodeP: 608 case Op_DecodeN: 609 case Op_EncodePKlass: 610 case Op_DecodeNKlass: { 611 add_local_var_and_edge(n, PointsToNode::NoEscape, 612 n->in(1), NULL); 613 break; 614 } 615 case Op_CMoveP: { 616 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 617 Node* in = n->in(i); 618 if (in == NULL) 619 continue; // ignore NULL 620 Node* uncast_in = in->uncast(); 621 if (uncast_in->is_top() || uncast_in == n) 622 continue; // ignore top or inputs which go back this node 623 PointsToNode* ptn = ptnode_adr(in->_idx); 624 assert(ptn != NULL, "node should be registered"); 625 add_edge(n_ptn, ptn); 626 } 627 break; 628 } 629 case Op_LoadP: 630 case Op_LoadN: 631 case Op_LoadPLocked: { 632 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 633 // ThreadLocal has RawPtr type. 634 const Type* t = _igvn->type(n); 635 if (t->make_ptr() != NULL) { 636 Node* adr = n->in(MemNode::Address); 637 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 638 break; 639 } 640 ELSE_FAIL("Op_LoadP"); 641 } 642 case Op_Phi: { 643 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 644 // ThreadLocal has RawPtr type. 645 const Type* t = n->as_Phi()->type(); 646 if (t->make_ptr() != NULL) { 647 for (uint i = 1; i < n->req(); i++) { 648 Node* in = n->in(i); 649 if (in == NULL) 650 continue; // ignore NULL 651 Node* uncast_in = in->uncast(); 652 if (uncast_in->is_top() || uncast_in == n) 653 continue; // ignore top or inputs which go back this node 654 PointsToNode* ptn = ptnode_adr(in->_idx); 655 assert(ptn != NULL, "node should be registered"); 656 add_edge(n_ptn, ptn); 657 } 658 break; 659 } 660 ELSE_FAIL("Op_Phi"); 661 } 662 case Op_Proj: { 663 // we are only interested in the oop result projection from a call 664 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() && 665 (n->in(0)->as_Call()->returns_pointer()|| n->bottom_type()->isa_ptr())) { 666 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) || 667 n->in(0)->as_Call()->tf()->returns_value_type_as_fields(), "what kind of oop return is it?"); 668 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 669 break; 670 } 671 ELSE_FAIL("Op_Proj"); 672 } 673 case Op_Rethrow: // Exception object escapes 674 case Op_Return: { 675 if (n->req() > TypeFunc::Parms && 676 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 677 // Treat Return value as LocalVar with GlobalEscape escape state. 678 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 679 n->in(TypeFunc::Parms), NULL); 680 break; 681 } 682 ELSE_FAIL("Op_Return"); 683 } 684 case Op_StoreP: 685 case Op_StoreN: 686 case Op_StoreNKlass: 687 case Op_StorePConditional: 688 case Op_CompareAndExchangeP: 689 case Op_CompareAndExchangeN: 690 case Op_CompareAndSwapP: 691 case Op_CompareAndSwapN: 692 case Op_WeakCompareAndSwapP: 693 case Op_WeakCompareAndSwapN: 694 case Op_GetAndSetP: 695 case Op_GetAndSetN: { 696 if (add_final_edges_unsafe_access(n, opcode)) { 697 break; 698 } 699 ELSE_FAIL("Op_StoreP"); 700 } 701 case Op_AryEq: 702 case Op_HasNegatives: 703 case Op_StrComp: 704 case Op_StrEquals: 705 case Op_StrIndexOf: 706 case Op_StrIndexOfChar: 707 case Op_StrInflatedCopy: 708 case Op_StrCompressedCopy: 709 case Op_EncodeISOArray: { 710 // char[]/byte[] arrays passed to string intrinsic do not escape but 711 // they are not scalar replaceable. Adjust escape state for them. 712 // Start from in(2) edge since in(1) is memory edge. 713 for (uint i = 2; i < n->req(); i++) { 714 Node* adr = n->in(i); 715 const Type* at = _igvn->type(adr); 716 if (!adr->is_top() && at->isa_ptr()) { 717 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 718 at->isa_ptr() != NULL, "expecting a pointer"); 719 if (adr->is_AddP()) { 720 adr = get_addp_base(adr); 721 } 722 PointsToNode* ptn = ptnode_adr(adr->_idx); 723 assert(ptn != NULL, "node should be registered"); 724 add_edge(n_ptn, ptn); 725 } 726 } 727 break; 728 } 729 default: { 730 // This method should be called only for EA specific nodes which may 731 // miss some edges when they were created. 732 #ifdef ASSERT 733 n->dump(1); 734 #endif 735 guarantee(false, "unknown node"); 736 } 737 } 738 return; 739 } 740 741 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) { 742 Node* adr = n->in(MemNode::Address); 743 const Type* adr_type = _igvn->type(adr); 744 adr_type = adr_type->make_ptr(); 745 if (adr_type == NULL) { 746 return; // skip dead nodes 747 } 748 if (adr_type->isa_oopptr() 749 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 750 && adr_type == TypeRawPtr::NOTNULL 751 && adr->in(AddPNode::Address)->is_Proj() 752 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 753 delayed_worklist->push(n); // Process it later. 754 #ifdef ASSERT 755 assert (adr->is_AddP(), "expecting an AddP"); 756 if (adr_type == TypeRawPtr::NOTNULL) { 757 // Verify a raw address for a store captured by Initialize node. 758 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 759 assert(offs != Type::OffsetBot, "offset must be a constant"); 760 } 761 #endif 762 } else { 763 // Ignore copy the displaced header to the BoxNode (OSR compilation). 764 if (adr->is_BoxLock()) { 765 return; 766 } 767 // Stored value escapes in unsafe access. 768 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 769 delayed_worklist->push(n); // Process unsafe access later. 770 return; 771 } 772 #ifdef ASSERT 773 n->dump(1); 774 assert(false, "not unsafe"); 775 #endif 776 } 777 } 778 779 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) { 780 Node* adr = n->in(MemNode::Address); 781 const Type *adr_type = _igvn->type(adr); 782 adr_type = adr_type->make_ptr(); 783 #ifdef ASSERT 784 if (adr_type == NULL) { 785 n->dump(1); 786 assert(adr_type != NULL, "dead node should not be on list"); 787 return true; 788 } 789 #endif 790 791 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 792 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 793 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 794 } 795 796 if (adr_type->isa_oopptr() 797 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 798 && adr_type == TypeRawPtr::NOTNULL 799 && adr->in(AddPNode::Address)->is_Proj() 800 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 801 // Point Address to Value 802 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 803 assert(adr_ptn != NULL && 804 adr_ptn->as_Field()->is_oop(), "node should be registered"); 805 Node* val = n->in(MemNode::ValueIn); 806 PointsToNode* ptn = ptnode_adr(val->_idx); 807 assert(ptn != NULL, "node should be registered"); 808 add_edge(adr_ptn, ptn); 809 return true; 810 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 811 // Stored value escapes in unsafe access. 812 Node* val = n->in(MemNode::ValueIn); 813 PointsToNode* ptn = ptnode_adr(val->_idx); 814 assert(ptn != NULL, "node should be registered"); 815 set_escape_state(ptn, PointsToNode::GlobalEscape); 816 // Add edge to object for unsafe access with offset. 817 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 818 assert(adr_ptn != NULL, "node should be registered"); 819 if (adr_ptn->is_Field()) { 820 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 821 add_edge(adr_ptn, ptn); 822 } 823 return true; 824 } 825 return false; 826 } 827 828 void ConnectionGraph::add_call_node(CallNode* call) { 829 assert(call->returns_pointer() || call->tf()->returns_value_type_as_fields(), "only for call which returns pointer"); 830 uint call_idx = call->_idx; 831 if (call->is_Allocate()) { 832 Node* k = call->in(AllocateNode::KlassNode); 833 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 834 assert(kt != NULL, "TypeKlassPtr required."); 835 ciKlass* cik = kt->klass(); 836 PointsToNode::EscapeState es = PointsToNode::NoEscape; 837 bool scalar_replaceable = true; 838 if (call->is_AllocateArray()) { 839 if (!cik->is_array_klass()) { // StressReflectiveCode 840 es = PointsToNode::GlobalEscape; 841 } else { 842 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 843 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 844 // Not scalar replaceable if the length is not constant or too big. 845 scalar_replaceable = false; 846 } 847 } 848 } else { // Allocate instance 849 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 850 cik->is_subclass_of(_compile->env()->Reference_klass()) || 851 !cik->is_instance_klass() || // StressReflectiveCode 852 !cik->as_instance_klass()->can_be_instantiated() || 853 cik->as_instance_klass()->has_finalizer()) { 854 es = PointsToNode::GlobalEscape; 855 } 856 } 857 add_java_object(call, es); 858 PointsToNode* ptn = ptnode_adr(call_idx); 859 if (!scalar_replaceable && ptn->scalar_replaceable()) { 860 ptn->set_scalar_replaceable(false); 861 } 862 } else if (call->is_CallStaticJava()) { 863 // Call nodes could be different types: 864 // 865 // 1. CallDynamicJavaNode (what happened during call is unknown): 866 // 867 // - mapped to GlobalEscape JavaObject node if oop is returned; 868 // 869 // - all oop arguments are escaping globally; 870 // 871 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 872 // 873 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 874 // 875 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 876 // - mapped to NoEscape JavaObject node if non-escaping object allocated 877 // during call is returned; 878 // - mapped to ArgEscape LocalVar node pointed to object arguments 879 // which are returned and does not escape during call; 880 // 881 // - oop arguments escaping status is defined by bytecode analysis; 882 // 883 // For a static call, we know exactly what method is being called. 884 // Use bytecode estimator to record whether the call's return value escapes. 885 ciMethod* meth = call->as_CallJava()->method(); 886 if (meth == NULL) { 887 const char* name = call->as_CallStaticJava()->_name; 888 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 889 // Returns a newly allocated unescaped object. 890 add_java_object(call, PointsToNode::NoEscape); 891 ptnode_adr(call_idx)->set_scalar_replaceable(false); 892 } else if (meth->is_boxing_method()) { 893 // Returns boxing object 894 PointsToNode::EscapeState es; 895 vmIntrinsics::ID intr = meth->intrinsic_id(); 896 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 897 // It does not escape if object is always allocated. 898 es = PointsToNode::NoEscape; 899 } else { 900 // It escapes globally if object could be loaded from cache. 901 es = PointsToNode::GlobalEscape; 902 } 903 add_java_object(call, es); 904 } else { 905 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 906 call_analyzer->copy_dependencies(_compile->dependencies()); 907 if (call_analyzer->is_return_allocated()) { 908 // Returns a newly allocated unescaped object, simply 909 // update dependency information. 910 // Mark it as NoEscape so that objects referenced by 911 // it's fields will be marked as NoEscape at least. 912 add_java_object(call, PointsToNode::NoEscape); 913 ptnode_adr(call_idx)->set_scalar_replaceable(false); 914 } else { 915 // Determine whether any arguments are returned. 916 const TypeTuple* d = call->tf()->domain_cc(); 917 bool ret_arg = false; 918 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 919 if (d->field_at(i)->isa_ptr() != NULL && 920 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 921 ret_arg = true; 922 break; 923 } 924 } 925 if (ret_arg) { 926 add_local_var(call, PointsToNode::ArgEscape); 927 } else { 928 // Returns unknown object. 929 map_ideal_node(call, phantom_obj); 930 } 931 } 932 } 933 } else { 934 // An other type of call, assume the worst case: 935 // returned value is unknown and globally escapes. 936 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 937 map_ideal_node(call, phantom_obj); 938 } 939 } 940 941 void ConnectionGraph::process_call_arguments(CallNode *call) { 942 bool is_arraycopy = false; 943 switch (call->Opcode()) { 944 #ifdef ASSERT 945 case Op_Allocate: 946 case Op_AllocateArray: 947 case Op_Lock: 948 case Op_Unlock: 949 assert(false, "should be done already"); 950 break; 951 #endif 952 case Op_ArrayCopy: 953 case Op_CallLeafNoFP: 954 // Most array copies are ArrayCopy nodes at this point but there 955 // are still a few direct calls to the copy subroutines (See 956 // PhaseStringOpts::copy_string()) 957 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 958 call->as_CallLeaf()->is_call_to_arraycopystub(); 959 // fall through 960 case Op_CallLeaf: { 961 // Stub calls, objects do not escape but they are not scale replaceable. 962 // Adjust escape state for outgoing arguments. 963 const TypeTuple * d = call->tf()->domain_sig(); 964 bool src_has_oops = false; 965 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 966 const Type* at = d->field_at(i); 967 Node *arg = call->in(i); 968 if (arg == NULL) { 969 continue; 970 } 971 const Type *aat = _igvn->type(arg); 972 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 973 continue; 974 if (arg->is_AddP()) { 975 // 976 // The inline_native_clone() case when the arraycopy stub is called 977 // after the allocation before Initialize and CheckCastPP nodes. 978 // Or normal arraycopy for object arrays case. 979 // 980 // Set AddP's base (Allocate) as not scalar replaceable since 981 // pointer to the base (with offset) is passed as argument. 982 // 983 arg = get_addp_base(arg); 984 } 985 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 986 assert(arg_ptn != NULL, "should be registered"); 987 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 988 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 989 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 990 aat->isa_ptr() != NULL, "expecting an Ptr"); 991 bool arg_has_oops = aat->isa_oopptr() && 992 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 993 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()) || 994 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != NULL && 995 aat->isa_aryptr()->elem()->isa_valuetype() && 996 aat->isa_aryptr()->elem()->value_klass()->contains_oops())); 997 if (i == TypeFunc::Parms) { 998 src_has_oops = arg_has_oops; 999 } 1000 // 1001 // src or dst could be j.l.Object when other is basic type array: 1002 // 1003 // arraycopy(char[],0,Object*,0,size); 1004 // arraycopy(Object*,0,char[],0,size); 1005 // 1006 // Don't add edges in such cases. 1007 // 1008 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 1009 arg_has_oops && (i > TypeFunc::Parms); 1010 #ifdef ASSERT 1011 if (!(is_arraycopy || 1012 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) || 1013 (call->as_CallLeaf()->_name != NULL && 1014 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 1015 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 1016 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 1017 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 1018 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 1019 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 1020 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 1021 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 || 1022 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 1023 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 || 1024 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 1025 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 1026 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 1027 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 1028 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 1029 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 1030 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 1031 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 1032 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 1033 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 1034 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 || 1035 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 || 1036 strcmp(call->as_CallLeaf()->_name, "load_unknown_value") == 0 || 1037 strcmp(call->as_CallLeaf()->_name, "store_unknown_value") == 0) 1038 ))) { 1039 call->dump(); 1040 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name); 1041 } 1042 #endif 1043 // Always process arraycopy's destination object since 1044 // we need to add all possible edges to references in 1045 // source object. 1046 if (arg_esc >= PointsToNode::ArgEscape && 1047 !arg_is_arraycopy_dest) { 1048 continue; 1049 } 1050 PointsToNode::EscapeState es = PointsToNode::ArgEscape; 1051 if (call->is_ArrayCopy()) { 1052 ArrayCopyNode* ac = call->as_ArrayCopy(); 1053 if (ac->is_clonebasic() || 1054 ac->is_arraycopy_validated() || 1055 ac->is_copyof_validated() || 1056 ac->is_copyofrange_validated()) { 1057 es = PointsToNode::NoEscape; 1058 } 1059 } 1060 set_escape_state(arg_ptn, es); 1061 if (arg_is_arraycopy_dest) { 1062 Node* src = call->in(TypeFunc::Parms); 1063 if (src->is_AddP()) { 1064 src = get_addp_base(src); 1065 } 1066 PointsToNode* src_ptn = ptnode_adr(src->_idx); 1067 assert(src_ptn != NULL, "should be registered"); 1068 if (arg_ptn != src_ptn) { 1069 // Special arraycopy edge: 1070 // A destination object's field can't have the source object 1071 // as base since objects escape states are not related. 1072 // Only escape state of destination object's fields affects 1073 // escape state of fields in source object. 1074 add_arraycopy(call, es, src_ptn, arg_ptn); 1075 } 1076 } 1077 } 1078 } 1079 break; 1080 } 1081 case Op_CallStaticJava: { 1082 // For a static call, we know exactly what method is being called. 1083 // Use bytecode estimator to record the call's escape affects 1084 #ifdef ASSERT 1085 const char* name = call->as_CallStaticJava()->_name; 1086 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1087 #endif 1088 ciMethod* meth = call->as_CallJava()->method(); 1089 if ((meth != NULL) && meth->is_boxing_method()) { 1090 break; // Boxing methods do not modify any oops. 1091 } 1092 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1093 // fall-through if not a Java method or no analyzer information 1094 if (call_analyzer != NULL) { 1095 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1096 const TypeTuple* d = call->tf()->domain_cc(); 1097 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1098 const Type* at = d->field_at(i); 1099 int k = i - TypeFunc::Parms; 1100 Node* arg = call->in(i); 1101 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1102 if (at->isa_ptr() != NULL && 1103 call_analyzer->is_arg_returned(k)) { 1104 // The call returns arguments. 1105 if (call_ptn != NULL) { // Is call's result used? 1106 assert(call_ptn->is_LocalVar(), "node should be registered"); 1107 assert(arg_ptn != NULL, "node should be registered"); 1108 add_edge(call_ptn, arg_ptn); 1109 } 1110 } 1111 if (at->isa_oopptr() != NULL && 1112 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1113 if (!call_analyzer->is_arg_stack(k)) { 1114 // The argument global escapes 1115 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1116 } else { 1117 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1118 if (!call_analyzer->is_arg_local(k)) { 1119 // The argument itself doesn't escape, but any fields might 1120 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1121 } 1122 } 1123 } 1124 } 1125 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1126 // The call returns arguments. 1127 assert(call_ptn->edge_count() > 0, "sanity"); 1128 if (!call_analyzer->is_return_local()) { 1129 // Returns also unknown object. 1130 add_edge(call_ptn, phantom_obj); 1131 } 1132 } 1133 break; 1134 } 1135 } 1136 default: { 1137 // Fall-through here if not a Java method or no analyzer information 1138 // or some other type of call, assume the worst case: all arguments 1139 // globally escape. 1140 const TypeTuple* d = call->tf()->domain_cc(); 1141 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1142 const Type* at = d->field_at(i); 1143 if (at->isa_oopptr() != NULL) { 1144 Node* arg = call->in(i); 1145 if (arg->is_AddP()) { 1146 arg = get_addp_base(arg); 1147 } 1148 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1149 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1150 } 1151 } 1152 } 1153 } 1154 } 1155 1156 1157 // Finish Graph construction. 1158 bool ConnectionGraph::complete_connection_graph( 1159 GrowableArray<PointsToNode*>& ptnodes_worklist, 1160 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1161 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1162 GrowableArray<FieldNode*>& oop_fields_worklist) { 1163 // Normally only 1-3 passes needed to build Connection Graph depending 1164 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1165 // Set limit to 20 to catch situation when something did go wrong and 1166 // bailout Escape Analysis. 1167 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1168 #define CG_BUILD_ITER_LIMIT 20 1169 1170 // Propagate GlobalEscape and ArgEscape escape states and check that 1171 // we still have non-escaping objects. The method pushs on _worklist 1172 // Field nodes which reference phantom_object. 1173 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1174 return false; // Nothing to do. 1175 } 1176 // Now propagate references to all JavaObject nodes. 1177 int java_objects_length = java_objects_worklist.length(); 1178 elapsedTimer time; 1179 bool timeout = false; 1180 int new_edges = 1; 1181 int iterations = 0; 1182 do { 1183 while ((new_edges > 0) && 1184 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1185 double start_time = time.seconds(); 1186 time.start(); 1187 new_edges = 0; 1188 // Propagate references to phantom_object for nodes pushed on _worklist 1189 // by find_non_escaped_objects() and find_field_value(). 1190 new_edges += add_java_object_edges(phantom_obj, false); 1191 for (int next = 0; next < java_objects_length; ++next) { 1192 JavaObjectNode* ptn = java_objects_worklist.at(next); 1193 new_edges += add_java_object_edges(ptn, true); 1194 1195 #define SAMPLE_SIZE 4 1196 if ((next % SAMPLE_SIZE) == 0) { 1197 // Each 4 iterations calculate how much time it will take 1198 // to complete graph construction. 1199 time.stop(); 1200 // Poll for requests from shutdown mechanism to quiesce compiler 1201 // because Connection graph construction may take long time. 1202 CompileBroker::maybe_block(); 1203 double stop_time = time.seconds(); 1204 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1205 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1206 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1207 timeout = true; 1208 break; // Timeout 1209 } 1210 start_time = stop_time; 1211 time.start(); 1212 } 1213 #undef SAMPLE_SIZE 1214 1215 } 1216 if (timeout) break; 1217 if (new_edges > 0) { 1218 // Update escape states on each iteration if graph was updated. 1219 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1220 return false; // Nothing to do. 1221 } 1222 } 1223 time.stop(); 1224 if (time.seconds() >= EscapeAnalysisTimeout) { 1225 timeout = true; 1226 break; 1227 } 1228 } 1229 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1230 time.start(); 1231 // Find fields which have unknown value. 1232 int fields_length = oop_fields_worklist.length(); 1233 for (int next = 0; next < fields_length; next++) { 1234 FieldNode* field = oop_fields_worklist.at(next); 1235 if (field->edge_count() == 0) { 1236 new_edges += find_field_value(field); 1237 // This code may added new edges to phantom_object. 1238 // Need an other cycle to propagate references to phantom_object. 1239 } 1240 } 1241 time.stop(); 1242 if (time.seconds() >= EscapeAnalysisTimeout) { 1243 timeout = true; 1244 break; 1245 } 1246 } else { 1247 new_edges = 0; // Bailout 1248 } 1249 } while (new_edges > 0); 1250 1251 // Bailout if passed limits. 1252 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1253 Compile* C = _compile; 1254 if (C->log() != NULL) { 1255 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1256 C->log()->text("%s", timeout ? "time" : "iterations"); 1257 C->log()->end_elem(" limit'"); 1258 } 1259 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1260 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()); 1261 // Possible infinite build_connection_graph loop, 1262 // bailout (no changes to ideal graph were made). 1263 return false; 1264 } 1265 #ifdef ASSERT 1266 if (Verbose && PrintEscapeAnalysis) { 1267 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1268 iterations, nodes_size(), ptnodes_worklist.length()); 1269 } 1270 #endif 1271 1272 #undef CG_BUILD_ITER_LIMIT 1273 1274 // Find fields initialized by NULL for non-escaping Allocations. 1275 int non_escaped_length = non_escaped_worklist.length(); 1276 for (int next = 0; next < non_escaped_length; next++) { 1277 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1278 PointsToNode::EscapeState es = ptn->escape_state(); 1279 assert(es <= PointsToNode::ArgEscape, "sanity"); 1280 if (es == PointsToNode::NoEscape) { 1281 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1282 // Adding references to NULL object does not change escape states 1283 // since it does not escape. Also no fields are added to NULL object. 1284 add_java_object_edges(null_obj, false); 1285 } 1286 } 1287 Node* n = ptn->ideal_node(); 1288 if (n->is_Allocate()) { 1289 // The object allocated by this Allocate node will never be 1290 // seen by an other thread. Mark it so that when it is 1291 // expanded no MemBarStoreStore is added. 1292 InitializeNode* ini = n->as_Allocate()->initialization(); 1293 if (ini != NULL) 1294 ini->set_does_not_escape(); 1295 } 1296 } 1297 return true; // Finished graph construction. 1298 } 1299 1300 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1301 // and check that we still have non-escaping java objects. 1302 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1303 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1304 GrowableArray<PointsToNode*> escape_worklist; 1305 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1306 int ptnodes_length = ptnodes_worklist.length(); 1307 for (int next = 0; next < ptnodes_length; ++next) { 1308 PointsToNode* ptn = ptnodes_worklist.at(next); 1309 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1310 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1311 escape_worklist.push(ptn); 1312 } 1313 } 1314 // Set escape states to referenced nodes (edges list). 1315 while (escape_worklist.length() > 0) { 1316 PointsToNode* ptn = escape_worklist.pop(); 1317 PointsToNode::EscapeState es = ptn->escape_state(); 1318 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1319 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1320 es >= PointsToNode::ArgEscape) { 1321 // GlobalEscape or ArgEscape state of field means it has unknown value. 1322 if (add_edge(ptn, phantom_obj)) { 1323 // New edge was added 1324 add_field_uses_to_worklist(ptn->as_Field()); 1325 } 1326 } 1327 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1328 PointsToNode* e = i.get(); 1329 if (e->is_Arraycopy()) { 1330 assert(ptn->arraycopy_dst(), "sanity"); 1331 // Propagate only fields escape state through arraycopy edge. 1332 if (e->fields_escape_state() < field_es) { 1333 set_fields_escape_state(e, field_es); 1334 escape_worklist.push(e); 1335 } 1336 } else if (es >= field_es) { 1337 // fields_escape_state is also set to 'es' if it is less than 'es'. 1338 if (e->escape_state() < es) { 1339 set_escape_state(e, es); 1340 escape_worklist.push(e); 1341 } 1342 } else { 1343 // Propagate field escape state. 1344 bool es_changed = false; 1345 if (e->fields_escape_state() < field_es) { 1346 set_fields_escape_state(e, field_es); 1347 es_changed = true; 1348 } 1349 if ((e->escape_state() < field_es) && 1350 e->is_Field() && ptn->is_JavaObject() && 1351 e->as_Field()->is_oop()) { 1352 // Change escape state of referenced fields. 1353 set_escape_state(e, field_es); 1354 es_changed = true; 1355 } else if (e->escape_state() < es) { 1356 set_escape_state(e, es); 1357 es_changed = true; 1358 } 1359 if (es_changed) { 1360 escape_worklist.push(e); 1361 } 1362 } 1363 } 1364 } 1365 // Remove escaped objects from non_escaped list. 1366 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1367 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1368 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1369 non_escaped_worklist.delete_at(next); 1370 } 1371 if (ptn->escape_state() == PointsToNode::NoEscape) { 1372 // Find fields in non-escaped allocations which have unknown value. 1373 find_init_values(ptn, phantom_obj, NULL); 1374 } 1375 } 1376 return (non_escaped_worklist.length() > 0); 1377 } 1378 1379 // Add all references to JavaObject node by walking over all uses. 1380 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1381 int new_edges = 0; 1382 if (populate_worklist) { 1383 // Populate _worklist by uses of jobj's uses. 1384 for (UseIterator i(jobj); i.has_next(); i.next()) { 1385 PointsToNode* use = i.get(); 1386 if (use->is_Arraycopy()) 1387 continue; 1388 add_uses_to_worklist(use); 1389 if (use->is_Field() && use->as_Field()->is_oop()) { 1390 // Put on worklist all field's uses (loads) and 1391 // related field nodes (same base and offset). 1392 add_field_uses_to_worklist(use->as_Field()); 1393 } 1394 } 1395 } 1396 for (int l = 0; l < _worklist.length(); l++) { 1397 PointsToNode* use = _worklist.at(l); 1398 if (PointsToNode::is_base_use(use)) { 1399 // Add reference from jobj to field and from field to jobj (field's base). 1400 use = PointsToNode::get_use_node(use)->as_Field(); 1401 if (add_base(use->as_Field(), jobj)) { 1402 new_edges++; 1403 } 1404 continue; 1405 } 1406 assert(!use->is_JavaObject(), "sanity"); 1407 if (use->is_Arraycopy()) { 1408 if (jobj == null_obj) // NULL object does not have field edges 1409 continue; 1410 // Added edge from Arraycopy node to arraycopy's source java object 1411 if (add_edge(use, jobj)) { 1412 jobj->set_arraycopy_src(); 1413 new_edges++; 1414 } 1415 // and stop here. 1416 continue; 1417 } 1418 if (!add_edge(use, jobj)) 1419 continue; // No new edge added, there was such edge already. 1420 new_edges++; 1421 if (use->is_LocalVar()) { 1422 add_uses_to_worklist(use); 1423 if (use->arraycopy_dst()) { 1424 for (EdgeIterator i(use); i.has_next(); i.next()) { 1425 PointsToNode* e = i.get(); 1426 if (e->is_Arraycopy()) { 1427 if (jobj == null_obj) // NULL object does not have field edges 1428 continue; 1429 // Add edge from arraycopy's destination java object to Arraycopy node. 1430 if (add_edge(jobj, e)) { 1431 new_edges++; 1432 jobj->set_arraycopy_dst(); 1433 } 1434 } 1435 } 1436 } 1437 } else { 1438 // Added new edge to stored in field values. 1439 // Put on worklist all field's uses (loads) and 1440 // related field nodes (same base and offset). 1441 add_field_uses_to_worklist(use->as_Field()); 1442 } 1443 } 1444 _worklist.clear(); 1445 _in_worklist.Reset(); 1446 return new_edges; 1447 } 1448 1449 // Put on worklist all related field nodes. 1450 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1451 assert(field->is_oop(), "sanity"); 1452 int offset = field->offset(); 1453 add_uses_to_worklist(field); 1454 // Loop over all bases of this field and push on worklist Field nodes 1455 // with the same offset and base (since they may reference the same field). 1456 for (BaseIterator i(field); i.has_next(); i.next()) { 1457 PointsToNode* base = i.get(); 1458 add_fields_to_worklist(field, base); 1459 // Check if the base was source object of arraycopy and go over arraycopy's 1460 // destination objects since values stored to a field of source object are 1461 // accessable by uses (loads) of fields of destination objects. 1462 if (base->arraycopy_src()) { 1463 for (UseIterator j(base); j.has_next(); j.next()) { 1464 PointsToNode* arycp = j.get(); 1465 if (arycp->is_Arraycopy()) { 1466 for (UseIterator k(arycp); k.has_next(); k.next()) { 1467 PointsToNode* abase = k.get(); 1468 if (abase->arraycopy_dst() && abase != base) { 1469 // Look for the same arraycopy reference. 1470 add_fields_to_worklist(field, abase); 1471 } 1472 } 1473 } 1474 } 1475 } 1476 } 1477 } 1478 1479 // Put on worklist all related field nodes. 1480 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1481 int offset = field->offset(); 1482 if (base->is_LocalVar()) { 1483 for (UseIterator j(base); j.has_next(); j.next()) { 1484 PointsToNode* f = j.get(); 1485 if (PointsToNode::is_base_use(f)) { // Field 1486 f = PointsToNode::get_use_node(f); 1487 if (f == field || !f->as_Field()->is_oop()) 1488 continue; 1489 int offs = f->as_Field()->offset(); 1490 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1491 add_to_worklist(f); 1492 } 1493 } 1494 } 1495 } else { 1496 assert(base->is_JavaObject(), "sanity"); 1497 if (// Skip phantom_object since it is only used to indicate that 1498 // this field's content globally escapes. 1499 (base != phantom_obj) && 1500 // NULL object node does not have fields. 1501 (base != null_obj)) { 1502 for (EdgeIterator i(base); i.has_next(); i.next()) { 1503 PointsToNode* f = i.get(); 1504 // Skip arraycopy edge since store to destination object field 1505 // does not update value in source object field. 1506 if (f->is_Arraycopy()) { 1507 assert(base->arraycopy_dst(), "sanity"); 1508 continue; 1509 } 1510 if (f == field || !f->as_Field()->is_oop()) 1511 continue; 1512 int offs = f->as_Field()->offset(); 1513 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1514 add_to_worklist(f); 1515 } 1516 } 1517 } 1518 } 1519 } 1520 1521 // Find fields which have unknown value. 1522 int ConnectionGraph::find_field_value(FieldNode* field) { 1523 // Escaped fields should have init value already. 1524 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1525 int new_edges = 0; 1526 for (BaseIterator i(field); i.has_next(); i.next()) { 1527 PointsToNode* base = i.get(); 1528 if (base->is_JavaObject()) { 1529 // Skip Allocate's fields which will be processed later. 1530 if (base->ideal_node()->is_Allocate()) 1531 return 0; 1532 assert(base == null_obj, "only NULL ptr base expected here"); 1533 } 1534 } 1535 if (add_edge(field, phantom_obj)) { 1536 // New edge was added 1537 new_edges++; 1538 add_field_uses_to_worklist(field); 1539 } 1540 return new_edges; 1541 } 1542 1543 // Find fields initializing values for allocations. 1544 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1545 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1546 int new_edges = 0; 1547 Node* alloc = pta->ideal_node(); 1548 if (init_val == phantom_obj) { 1549 // Do nothing for Allocate nodes since its fields values are 1550 // "known" unless they are initialized by arraycopy/clone. 1551 if (alloc->is_Allocate() && !pta->arraycopy_dst()) 1552 return 0; 1553 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity"); 1554 #ifdef ASSERT 1555 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) { 1556 const char* name = alloc->as_CallStaticJava()->_name; 1557 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1558 } 1559 #endif 1560 // Non-escaped allocation returned from Java or runtime call have 1561 // unknown values in fields. 1562 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1563 PointsToNode* field = i.get(); 1564 if (field->is_Field() && field->as_Field()->is_oop()) { 1565 if (add_edge(field, phantom_obj)) { 1566 // New edge was added 1567 new_edges++; 1568 add_field_uses_to_worklist(field->as_Field()); 1569 } 1570 } 1571 } 1572 return new_edges; 1573 } 1574 assert(init_val == null_obj, "sanity"); 1575 // Do nothing for Call nodes since its fields values are unknown. 1576 if (!alloc->is_Allocate()) 1577 return 0; 1578 1579 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1580 bool visited_bottom_offset = false; 1581 GrowableArray<int> offsets_worklist; 1582 1583 // Check if an oop field's initializing value is recorded and add 1584 // a corresponding NULL if field's value if it is not recorded. 1585 // Connection Graph does not record a default initialization by NULL 1586 // captured by Initialize node. 1587 // 1588 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1589 PointsToNode* field = i.get(); // Field (AddP) 1590 if (!field->is_Field() || !field->as_Field()->is_oop()) 1591 continue; // Not oop field 1592 int offset = field->as_Field()->offset(); 1593 if (offset == Type::OffsetBot) { 1594 if (!visited_bottom_offset) { 1595 // OffsetBot is used to reference array's element, 1596 // always add reference to NULL to all Field nodes since we don't 1597 // known which element is referenced. 1598 if (add_edge(field, null_obj)) { 1599 // New edge was added 1600 new_edges++; 1601 add_field_uses_to_worklist(field->as_Field()); 1602 visited_bottom_offset = true; 1603 } 1604 } 1605 } else { 1606 // Check only oop fields. 1607 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1608 if (adr_type->isa_rawptr()) { 1609 #ifdef ASSERT 1610 // Raw pointers are used for initializing stores so skip it 1611 // since it should be recorded already 1612 Node* base = get_addp_base(field->ideal_node()); 1613 assert(adr_type->isa_rawptr() && base->is_Proj() && 1614 (base->in(0) == alloc),"unexpected pointer type"); 1615 #endif 1616 continue; 1617 } 1618 if (!offsets_worklist.contains(offset)) { 1619 offsets_worklist.append(offset); 1620 Node* value = NULL; 1621 if (ini != NULL) { 1622 // StoreP::memory_type() == T_ADDRESS 1623 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1624 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1625 // Make sure initializing store has the same type as this AddP. 1626 // This AddP may reference non existing field because it is on a 1627 // dead branch of bimorphic call which is not eliminated yet. 1628 if (store != NULL && store->is_Store() && 1629 store->as_Store()->memory_type() == ft) { 1630 value = store->in(MemNode::ValueIn); 1631 #ifdef ASSERT 1632 if (VerifyConnectionGraph) { 1633 // Verify that AddP already points to all objects the value points to. 1634 PointsToNode* val = ptnode_adr(value->_idx); 1635 assert((val != NULL), "should be processed already"); 1636 PointsToNode* missed_obj = NULL; 1637 if (val->is_JavaObject()) { 1638 if (!field->points_to(val->as_JavaObject())) { 1639 missed_obj = val; 1640 } 1641 } else { 1642 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1643 tty->print_cr("----------init store has invalid value -----"); 1644 store->dump(); 1645 val->dump(); 1646 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1647 } 1648 for (EdgeIterator j(val); j.has_next(); j.next()) { 1649 PointsToNode* obj = j.get(); 1650 if (obj->is_JavaObject()) { 1651 if (!field->points_to(obj->as_JavaObject())) { 1652 missed_obj = obj; 1653 break; 1654 } 1655 } 1656 } 1657 } 1658 if (missed_obj != NULL) { 1659 tty->print_cr("----------field---------------------------------"); 1660 field->dump(); 1661 tty->print_cr("----------missed reference to object------------"); 1662 missed_obj->dump(); 1663 tty->print_cr("----------object referenced by init store-------"); 1664 store->dump(); 1665 val->dump(); 1666 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1667 } 1668 } 1669 #endif 1670 } else { 1671 // There could be initializing stores which follow allocation. 1672 // For example, a volatile field store is not collected 1673 // by Initialize node. 1674 // 1675 // Need to check for dependent loads to separate such stores from 1676 // stores which follow loads. For now, add initial value NULL so 1677 // that compare pointers optimization works correctly. 1678 } 1679 } 1680 if (value == NULL) { 1681 // A field's initializing value was not recorded. Add NULL. 1682 if (add_edge(field, null_obj)) { 1683 // New edge was added 1684 new_edges++; 1685 add_field_uses_to_worklist(field->as_Field()); 1686 } 1687 } 1688 } 1689 } 1690 } 1691 return new_edges; 1692 } 1693 1694 // Adjust scalar_replaceable state after Connection Graph is built. 1695 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1696 // Search for non-escaping objects which are not scalar replaceable 1697 // and mark them to propagate the state to referenced objects. 1698 1699 // 1. An object is not scalar replaceable if the field into which it is 1700 // stored has unknown offset (stored into unknown element of an array). 1701 // 1702 for (UseIterator i(jobj); i.has_next(); i.next()) { 1703 PointsToNode* use = i.get(); 1704 if (use->is_Arraycopy()) { 1705 continue; 1706 } 1707 if (use->is_Field()) { 1708 FieldNode* field = use->as_Field(); 1709 assert(field->is_oop() && field->scalar_replaceable(), "sanity"); 1710 if (field->offset() == Type::OffsetBot) { 1711 jobj->set_scalar_replaceable(false); 1712 return; 1713 } 1714 // 2. An object is not scalar replaceable if the field into which it is 1715 // stored has multiple bases one of which is null. 1716 if (field->base_count() > 1) { 1717 for (BaseIterator i(field); i.has_next(); i.next()) { 1718 PointsToNode* base = i.get(); 1719 if (base == null_obj) { 1720 jobj->set_scalar_replaceable(false); 1721 return; 1722 } 1723 } 1724 } 1725 } 1726 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1727 // 3. An object is not scalar replaceable if it is merged with other objects. 1728 for (EdgeIterator j(use); j.has_next(); j.next()) { 1729 PointsToNode* ptn = j.get(); 1730 if (ptn->is_JavaObject() && ptn != jobj) { 1731 // Mark all objects. 1732 jobj->set_scalar_replaceable(false); 1733 ptn->set_scalar_replaceable(false); 1734 } 1735 } 1736 if (!jobj->scalar_replaceable()) { 1737 return; 1738 } 1739 } 1740 1741 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1742 if (j.get()->is_Arraycopy()) { 1743 continue; 1744 } 1745 1746 // Non-escaping object node should point only to field nodes. 1747 FieldNode* field = j.get()->as_Field(); 1748 int offset = field->as_Field()->offset(); 1749 1750 // 4. An object is not scalar replaceable if it has a field with unknown 1751 // offset (array's element is accessed in loop). 1752 if (offset == Type::OffsetBot) { 1753 jobj->set_scalar_replaceable(false); 1754 return; 1755 } 1756 // 5. Currently an object is not scalar replaceable if a LoadStore node 1757 // access its field since the field value is unknown after it. 1758 // 1759 Node* n = field->ideal_node(); 1760 1761 // Test for an unsafe access that was parsed as maybe off heap 1762 // (with a CheckCastPP to raw memory). 1763 assert(n->is_AddP(), "expect an address computation"); 1764 if (n->in(AddPNode::Base)->is_top() && 1765 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) { 1766 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected"); 1767 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected"); 1768 jobj->set_scalar_replaceable(false); 1769 return; 1770 } 1771 1772 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1773 Node* u = n->fast_out(i); 1774 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) { 1775 jobj->set_scalar_replaceable(false); 1776 return; 1777 } 1778 } 1779 1780 // 6. Or the address may point to more then one object. This may produce 1781 // the false positive result (set not scalar replaceable) 1782 // since the flow-insensitive escape analysis can't separate 1783 // the case when stores overwrite the field's value from the case 1784 // when stores happened on different control branches. 1785 // 1786 // Note: it will disable scalar replacement in some cases: 1787 // 1788 // Point p[] = new Point[1]; 1789 // p[0] = new Point(); // Will be not scalar replaced 1790 // 1791 // but it will save us from incorrect optimizations in next cases: 1792 // 1793 // Point p[] = new Point[1]; 1794 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1795 // 1796 if (field->base_count() > 1) { 1797 for (BaseIterator i(field); i.has_next(); i.next()) { 1798 PointsToNode* base = i.get(); 1799 // Don't take into account LocalVar nodes which 1800 // may point to only one object which should be also 1801 // this field's base by now. 1802 if (base->is_JavaObject() && base != jobj) { 1803 // Mark all bases. 1804 jobj->set_scalar_replaceable(false); 1805 base->set_scalar_replaceable(false); 1806 } 1807 } 1808 } 1809 } 1810 } 1811 1812 #ifdef ASSERT 1813 void ConnectionGraph::verify_connection_graph( 1814 GrowableArray<PointsToNode*>& ptnodes_worklist, 1815 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1816 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1817 GrowableArray<Node*>& addp_worklist) { 1818 // Verify that graph is complete - no new edges could be added. 1819 int java_objects_length = java_objects_worklist.length(); 1820 int non_escaped_length = non_escaped_worklist.length(); 1821 int new_edges = 0; 1822 for (int next = 0; next < java_objects_length; ++next) { 1823 JavaObjectNode* ptn = java_objects_worklist.at(next); 1824 new_edges += add_java_object_edges(ptn, true); 1825 } 1826 assert(new_edges == 0, "graph was not complete"); 1827 // Verify that escape state is final. 1828 int length = non_escaped_worklist.length(); 1829 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1830 assert((non_escaped_length == non_escaped_worklist.length()) && 1831 (non_escaped_length == length) && 1832 (_worklist.length() == 0), "escape state was not final"); 1833 1834 // Verify fields information. 1835 int addp_length = addp_worklist.length(); 1836 for (int next = 0; next < addp_length; ++next ) { 1837 Node* n = addp_worklist.at(next); 1838 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1839 if (field->is_oop()) { 1840 // Verify that field has all bases 1841 Node* base = get_addp_base(n); 1842 PointsToNode* ptn = ptnode_adr(base->_idx); 1843 if (ptn->is_JavaObject()) { 1844 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1845 } else { 1846 assert(ptn->is_LocalVar(), "sanity"); 1847 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1848 PointsToNode* e = i.get(); 1849 if (e->is_JavaObject()) { 1850 assert(field->has_base(e->as_JavaObject()), "sanity"); 1851 } 1852 } 1853 } 1854 // Verify that all fields have initializing values. 1855 if (field->edge_count() == 0) { 1856 tty->print_cr("----------field does not have references----------"); 1857 field->dump(); 1858 for (BaseIterator i(field); i.has_next(); i.next()) { 1859 PointsToNode* base = i.get(); 1860 tty->print_cr("----------field has next base---------------------"); 1861 base->dump(); 1862 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1863 tty->print_cr("----------base has fields-------------------------"); 1864 for (EdgeIterator j(base); j.has_next(); j.next()) { 1865 j.get()->dump(); 1866 } 1867 tty->print_cr("----------base has references---------------------"); 1868 for (UseIterator j(base); j.has_next(); j.next()) { 1869 j.get()->dump(); 1870 } 1871 } 1872 } 1873 for (UseIterator i(field); i.has_next(); i.next()) { 1874 i.get()->dump(); 1875 } 1876 assert(field->edge_count() > 0, "sanity"); 1877 } 1878 } 1879 } 1880 } 1881 #endif 1882 1883 // Optimize ideal graph. 1884 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1885 GrowableArray<Node*>& storestore_worklist) { 1886 Compile* C = _compile; 1887 PhaseIterGVN* igvn = _igvn; 1888 if (EliminateLocks) { 1889 // Mark locks before changing ideal graph. 1890 int cnt = C->macro_count(); 1891 for( int i=0; i < cnt; i++ ) { 1892 Node *n = C->macro_node(i); 1893 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1894 AbstractLockNode* alock = n->as_AbstractLock(); 1895 if (!alock->is_non_esc_obj()) { 1896 if (not_global_escape(alock->obj_node())) { 1897 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1898 // The lock could be marked eliminated by lock coarsening 1899 // code during first IGVN before EA. Replace coarsened flag 1900 // to eliminate all associated locks/unlocks. 1901 #ifdef ASSERT 1902 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1903 #endif 1904 alock->set_non_esc_obj(); 1905 } 1906 } 1907 } 1908 } 1909 } 1910 1911 if (OptimizePtrCompare) { 1912 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1913 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1914 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1915 // Optimize objects compare. 1916 while (ptr_cmp_worklist.length() != 0) { 1917 Node *n = ptr_cmp_worklist.pop(); 1918 Node *res = optimize_ptr_compare(n); 1919 if (res != NULL) { 1920 #ifndef PRODUCT 1921 if (PrintOptimizePtrCompare) { 1922 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, (res == _pcmp_eq ? "EQ" : "NotEQ")); 1923 if (Verbose) { 1924 n->dump(1); 1925 } 1926 } 1927 #endif 1928 igvn->replace_node(n, res); 1929 } 1930 } 1931 // cleanup 1932 if (_pcmp_neq->outcnt() == 0) 1933 igvn->hash_delete(_pcmp_neq); 1934 if (_pcmp_eq->outcnt() == 0) 1935 igvn->hash_delete(_pcmp_eq); 1936 } 1937 1938 // For MemBarStoreStore nodes added in library_call.cpp, check 1939 // escape status of associated AllocateNode and optimize out 1940 // MemBarStoreStore node if the allocated object never escapes. 1941 while (storestore_worklist.length() != 0) { 1942 Node *n = storestore_worklist.pop(); 1943 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1944 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1945 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1946 if (not_global_escape(alloc)) { 1947 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1948 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1949 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1950 igvn->register_new_node_with_optimizer(mb); 1951 igvn->replace_node(storestore, mb); 1952 } 1953 } 1954 } 1955 1956 // Optimize objects compare. 1957 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1958 assert(OptimizePtrCompare, "sanity"); 1959 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1960 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1961 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1962 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1963 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1964 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1965 1966 // Check simple cases first. 1967 if (jobj1 != NULL) { 1968 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1969 if (jobj1 == jobj2) { 1970 // Comparing the same not escaping object. 1971 return _pcmp_eq; 1972 } 1973 Node* obj = jobj1->ideal_node(); 1974 // Comparing not escaping allocation. 1975 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1976 !ptn2->points_to(jobj1)) { 1977 return _pcmp_neq; // This includes nullness check. 1978 } 1979 } 1980 } 1981 if (jobj2 != NULL) { 1982 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1983 Node* obj = jobj2->ideal_node(); 1984 // Comparing not escaping allocation. 1985 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1986 !ptn1->points_to(jobj2)) { 1987 return _pcmp_neq; // This includes nullness check. 1988 } 1989 } 1990 } 1991 if (jobj1 != NULL && jobj1 != phantom_obj && 1992 jobj2 != NULL && jobj2 != phantom_obj && 1993 jobj1->ideal_node()->is_Con() && 1994 jobj2->ideal_node()->is_Con()) { 1995 // Klass or String constants compare. Need to be careful with 1996 // compressed pointers - compare types of ConN and ConP instead of nodes. 1997 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1998 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1999 if (t1->make_ptr() == t2->make_ptr()) { 2000 return _pcmp_eq; 2001 } else { 2002 return _pcmp_neq; 2003 } 2004 } 2005 if (ptn1->meet(ptn2)) { 2006 return NULL; // Sets are not disjoint 2007 } 2008 2009 // Sets are disjoint. 2010 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 2011 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 2012 bool set1_has_null_ptr = ptn1->points_to(null_obj); 2013 bool set2_has_null_ptr = ptn2->points_to(null_obj); 2014 if ((set1_has_unknown_ptr && set2_has_null_ptr) || 2015 (set2_has_unknown_ptr && set1_has_null_ptr)) { 2016 // Check nullness of unknown object. 2017 return NULL; 2018 } 2019 2020 // Disjointness by itself is not sufficient since 2021 // alias analysis is not complete for escaped objects. 2022 // Disjoint sets are definitely unrelated only when 2023 // at least one set has only not escaping allocations. 2024 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 2025 if (ptn1->non_escaping_allocation()) { 2026 return _pcmp_neq; 2027 } 2028 } 2029 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 2030 if (ptn2->non_escaping_allocation()) { 2031 return _pcmp_neq; 2032 } 2033 } 2034 return NULL; 2035 } 2036 2037 // Connection Graph constuction functions. 2038 2039 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 2040 PointsToNode* ptadr = _nodes.at(n->_idx); 2041 if (ptadr != NULL) { 2042 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 2043 return; 2044 } 2045 Compile* C = _compile; 2046 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 2047 _nodes.at_put(n->_idx, ptadr); 2048 } 2049 2050 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 2051 PointsToNode* ptadr = _nodes.at(n->_idx); 2052 if (ptadr != NULL) { 2053 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2054 return; 2055 } 2056 Compile* C = _compile; 2057 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2058 _nodes.at_put(n->_idx, ptadr); 2059 } 2060 2061 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2062 PointsToNode* ptadr = _nodes.at(n->_idx); 2063 if (ptadr != NULL) { 2064 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2065 return; 2066 } 2067 bool unsafe = false; 2068 bool is_oop = is_oop_field(n, offset, &unsafe); 2069 if (unsafe) { 2070 es = PointsToNode::GlobalEscape; 2071 } 2072 Compile* C = _compile; 2073 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2074 _nodes.at_put(n->_idx, field); 2075 } 2076 2077 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2078 PointsToNode* src, PointsToNode* dst) { 2079 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2080 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2081 PointsToNode* ptadr = _nodes.at(n->_idx); 2082 if (ptadr != NULL) { 2083 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2084 return; 2085 } 2086 Compile* C = _compile; 2087 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2088 _nodes.at_put(n->_idx, ptadr); 2089 // Add edge from arraycopy node to source object. 2090 (void)add_edge(ptadr, src); 2091 src->set_arraycopy_src(); 2092 // Add edge from destination object to arraycopy node. 2093 (void)add_edge(dst, ptadr); 2094 dst->set_arraycopy_dst(); 2095 } 2096 2097 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2098 const Type* adr_type = n->as_AddP()->bottom_type(); 2099 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot; 2100 BasicType bt = T_INT; 2101 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) { 2102 // Check only oop fields. 2103 if (!adr_type->isa_aryptr() || 2104 (adr_type->isa_aryptr()->klass() == NULL) || 2105 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2106 // OffsetBot is used to reference array's element. Ignore first AddP. 2107 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2108 bt = T_OBJECT; 2109 } 2110 } 2111 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2112 if (adr_type->isa_instptr()) { 2113 ciField* field = _compile->alias_type(adr_type->is_ptr())->field(); 2114 if (field != NULL) { 2115 bt = field->layout_type(); 2116 } else { 2117 // Check for unsafe oop field access 2118 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2119 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2120 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2121 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2122 bt = T_OBJECT; 2123 (*unsafe) = true; 2124 } 2125 } 2126 } else if (adr_type->isa_aryptr()) { 2127 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2128 // Ignore array length load. 2129 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2130 // Ignore first AddP. 2131 } else { 2132 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2133 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 2134 ciValueKlass* vk = elemtype->value_klass(); 2135 field_offset += vk->first_field_offset(); 2136 bt = vk->get_field_by_offset(field_offset, false)->layout_type(); 2137 } else { 2138 bt = elemtype->array_element_basic_type(); 2139 } 2140 } 2141 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2142 // Allocation initialization, ThreadLocal field access, unsafe access 2143 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2144 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2145 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2146 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2147 bt = T_OBJECT; 2148 } 2149 } 2150 } 2151 return (bt == T_OBJECT || bt == T_VALUETYPE || bt == T_NARROWOOP || bt == T_ARRAY); 2152 } 2153 2154 // Returns unique pointed java object or NULL. 2155 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2156 assert(!_collecting, "should not call when constructed graph"); 2157 // If the node was created after the escape computation we can't answer. 2158 uint idx = n->_idx; 2159 if (idx >= nodes_size()) { 2160 return NULL; 2161 } 2162 PointsToNode* ptn = ptnode_adr(idx); 2163 if (ptn->is_JavaObject()) { 2164 return ptn->as_JavaObject(); 2165 } 2166 assert(ptn->is_LocalVar(), "sanity"); 2167 // Check all java objects it points to. 2168 JavaObjectNode* jobj = NULL; 2169 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2170 PointsToNode* e = i.get(); 2171 if (e->is_JavaObject()) { 2172 if (jobj == NULL) { 2173 jobj = e->as_JavaObject(); 2174 } else if (jobj != e) { 2175 return NULL; 2176 } 2177 } 2178 } 2179 return jobj; 2180 } 2181 2182 // Return true if this node points only to non-escaping allocations. 2183 bool PointsToNode::non_escaping_allocation() { 2184 if (is_JavaObject()) { 2185 Node* n = ideal_node(); 2186 if (n->is_Allocate() || n->is_CallStaticJava()) { 2187 return (escape_state() == PointsToNode::NoEscape); 2188 } else { 2189 return false; 2190 } 2191 } 2192 assert(is_LocalVar(), "sanity"); 2193 // Check all java objects it points to. 2194 for (EdgeIterator i(this); i.has_next(); i.next()) { 2195 PointsToNode* e = i.get(); 2196 if (e->is_JavaObject()) { 2197 Node* n = e->ideal_node(); 2198 if ((e->escape_state() != PointsToNode::NoEscape) || 2199 !(n->is_Allocate() || n->is_CallStaticJava())) { 2200 return false; 2201 } 2202 } 2203 } 2204 return true; 2205 } 2206 2207 // Return true if we know the node does not escape globally. 2208 bool ConnectionGraph::not_global_escape(Node *n) { 2209 assert(!_collecting, "should not call during graph construction"); 2210 // If the node was created after the escape computation we can't answer. 2211 uint idx = n->_idx; 2212 if (idx >= nodes_size()) { 2213 return false; 2214 } 2215 PointsToNode* ptn = ptnode_adr(idx); 2216 PointsToNode::EscapeState es = ptn->escape_state(); 2217 // If we have already computed a value, return it. 2218 if (es >= PointsToNode::GlobalEscape) 2219 return false; 2220 if (ptn->is_JavaObject()) { 2221 return true; // (es < PointsToNode::GlobalEscape); 2222 } 2223 assert(ptn->is_LocalVar(), "sanity"); 2224 // Check all java objects it points to. 2225 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2226 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2227 return false; 2228 } 2229 return true; 2230 } 2231 2232 2233 // Helper functions 2234 2235 // Return true if this node points to specified node or nodes it points to. 2236 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2237 if (is_JavaObject()) { 2238 return (this == ptn); 2239 } 2240 assert(is_LocalVar() || is_Field(), "sanity"); 2241 for (EdgeIterator i(this); i.has_next(); i.next()) { 2242 if (i.get() == ptn) 2243 return true; 2244 } 2245 return false; 2246 } 2247 2248 // Return true if one node points to an other. 2249 bool PointsToNode::meet(PointsToNode* ptn) { 2250 if (this == ptn) { 2251 return true; 2252 } else if (ptn->is_JavaObject()) { 2253 return this->points_to(ptn->as_JavaObject()); 2254 } else if (this->is_JavaObject()) { 2255 return ptn->points_to(this->as_JavaObject()); 2256 } 2257 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2258 int ptn_count = ptn->edge_count(); 2259 for (EdgeIterator i(this); i.has_next(); i.next()) { 2260 PointsToNode* this_e = i.get(); 2261 for (int j = 0; j < ptn_count; j++) { 2262 if (this_e == ptn->edge(j)) 2263 return true; 2264 } 2265 } 2266 return false; 2267 } 2268 2269 #ifdef ASSERT 2270 // Return true if bases point to this java object. 2271 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2272 for (BaseIterator i(this); i.has_next(); i.next()) { 2273 if (i.get() == jobj) 2274 return true; 2275 } 2276 return false; 2277 } 2278 #endif 2279 2280 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2281 const Type *adr_type = phase->type(adr); 2282 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2283 adr->in(AddPNode::Address)->is_Proj() && 2284 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2285 // We are computing a raw address for a store captured by an Initialize 2286 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2287 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2288 assert(offs != Type::OffsetBot || 2289 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2290 "offset must be a constant or it is initialization of array"); 2291 return offs; 2292 } 2293 return adr_type->is_ptr()->flattened_offset(); 2294 } 2295 2296 Node* ConnectionGraph::get_addp_base(Node *addp) { 2297 assert(addp->is_AddP(), "must be AddP"); 2298 // 2299 // AddP cases for Base and Address inputs: 2300 // case #1. Direct object's field reference: 2301 // Allocate 2302 // | 2303 // Proj #5 ( oop result ) 2304 // | 2305 // CheckCastPP (cast to instance type) 2306 // | | 2307 // AddP ( base == address ) 2308 // 2309 // case #2. Indirect object's field reference: 2310 // Phi 2311 // | 2312 // CastPP (cast to instance type) 2313 // | | 2314 // AddP ( base == address ) 2315 // 2316 // case #3. Raw object's field reference for Initialize node: 2317 // Allocate 2318 // | 2319 // Proj #5 ( oop result ) 2320 // top | 2321 // \ | 2322 // AddP ( base == top ) 2323 // 2324 // case #4. Array's element reference: 2325 // {CheckCastPP | CastPP} 2326 // | | | 2327 // | AddP ( array's element offset ) 2328 // | | 2329 // AddP ( array's offset ) 2330 // 2331 // case #5. Raw object's field reference for arraycopy stub call: 2332 // The inline_native_clone() case when the arraycopy stub is called 2333 // after the allocation before Initialize and CheckCastPP nodes. 2334 // Allocate 2335 // | 2336 // Proj #5 ( oop result ) 2337 // | | 2338 // AddP ( base == address ) 2339 // 2340 // case #6. Constant Pool, ThreadLocal, CastX2P or 2341 // Raw object's field reference: 2342 // {ConP, ThreadLocal, CastX2P, raw Load} 2343 // top | 2344 // \ | 2345 // AddP ( base == top ) 2346 // 2347 // case #7. Klass's field reference. 2348 // LoadKlass 2349 // | | 2350 // AddP ( base == address ) 2351 // 2352 // case #8. narrow Klass's field reference. 2353 // LoadNKlass 2354 // | 2355 // DecodeN 2356 // | | 2357 // AddP ( base == address ) 2358 // 2359 // case #9. Mixed unsafe access 2360 // {instance} 2361 // | 2362 // CheckCastPP (raw) 2363 // top | 2364 // \ | 2365 // AddP ( base == top ) 2366 // 2367 Node *base = addp->in(AddPNode::Base); 2368 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9. 2369 base = addp->in(AddPNode::Address); 2370 while (base->is_AddP()) { 2371 // Case #6 (unsafe access) may have several chained AddP nodes. 2372 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2373 base = base->in(AddPNode::Address); 2374 } 2375 if (base->Opcode() == Op_CheckCastPP && 2376 base->bottom_type()->isa_rawptr() && 2377 _igvn->type(base->in(1))->isa_oopptr()) { 2378 base = base->in(1); // Case #9 2379 } else { 2380 Node* uncast_base = base->uncast(); 2381 int opcode = uncast_base->Opcode(); 2382 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2383 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2384 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2385 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()) || 2386 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(uncast_base), "sanity"); 2387 } 2388 } 2389 return base; 2390 } 2391 2392 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2393 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2394 Node* addp2 = addp->raw_out(0); 2395 if (addp->outcnt() == 1 && addp2->is_AddP() && 2396 addp2->in(AddPNode::Base) == n && 2397 addp2->in(AddPNode::Address) == addp) { 2398 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2399 // 2400 // Find array's offset to push it on worklist first and 2401 // as result process an array's element offset first (pushed second) 2402 // to avoid CastPP for the array's offset. 2403 // Otherwise the inserted CastPP (LocalVar) will point to what 2404 // the AddP (Field) points to. Which would be wrong since 2405 // the algorithm expects the CastPP has the same point as 2406 // as AddP's base CheckCastPP (LocalVar). 2407 // 2408 // ArrayAllocation 2409 // | 2410 // CheckCastPP 2411 // | 2412 // memProj (from ArrayAllocation CheckCastPP) 2413 // | || 2414 // | || Int (element index) 2415 // | || | ConI (log(element size)) 2416 // | || | / 2417 // | || LShift 2418 // | || / 2419 // | AddP (array's element offset) 2420 // | | 2421 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2422 // | / / 2423 // AddP (array's offset) 2424 // | 2425 // Load/Store (memory operation on array's element) 2426 // 2427 return addp2; 2428 } 2429 return NULL; 2430 } 2431 2432 // 2433 // Adjust the type and inputs of an AddP which computes the 2434 // address of a field of an instance 2435 // 2436 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2437 PhaseGVN* igvn = _igvn; 2438 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2439 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2440 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2441 if (t == NULL) { 2442 // We are computing a raw address for a store captured by an Initialize 2443 // compute an appropriate address type (cases #3 and #5). 2444 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2445 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2446 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2447 assert(offs != Type::OffsetBot, "offset must be a constant"); 2448 if (base_t->isa_aryptr() != NULL) { 2449 // In the case of a flattened value type array, each field has its 2450 // own slice so we need to extract the field being accessed from 2451 // the address computation 2452 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr(); 2453 } else { 2454 t = base_t->add_offset(offs)->is_oopptr(); 2455 } 2456 } 2457 int inst_id = base_t->instance_id(); 2458 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2459 "old type must be non-instance or match new type"); 2460 2461 // The type 't' could be subclass of 'base_t'. 2462 // As result t->offset() could be large then base_t's size and it will 2463 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2464 // constructor verifies correctness of the offset. 2465 // 2466 // It could happened on subclass's branch (from the type profiling 2467 // inlining) which was not eliminated during parsing since the exactness 2468 // of the allocation type was not propagated to the subclass type check. 2469 // 2470 // Or the type 't' could be not related to 'base_t' at all. 2471 // It could happen when CHA type is different from MDO type on a dead path 2472 // (for example, from instanceof check) which is not collapsed during parsing. 2473 // 2474 // Do nothing for such AddP node and don't process its users since 2475 // this code branch will go away. 2476 // 2477 if (!t->is_known_instance() && 2478 !base_t->klass()->is_subtype_of(t->klass())) { 2479 return false; // bail out 2480 } 2481 const TypePtr* tinst = base_t->add_offset(t->offset()); 2482 if (tinst->isa_aryptr() && t->isa_aryptr()) { 2483 // In the case of a flattened value type array, each field has its 2484 // own slice so we need to keep track of the field being accessed. 2485 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get()); 2486 } 2487 2488 // Do NOT remove the next line: ensure a new alias index is allocated 2489 // for the instance type. Note: C++ will not remove it since the call 2490 // has side effect. 2491 int alias_idx = _compile->get_alias_index(tinst); 2492 igvn->set_type(addp, tinst); 2493 // record the allocation in the node map 2494 set_map(addp, get_map(base->_idx)); 2495 // Set addp's Base and Address to 'base'. 2496 Node *abase = addp->in(AddPNode::Base); 2497 Node *adr = addp->in(AddPNode::Address); 2498 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2499 adr->in(0)->_idx == (uint)inst_id) { 2500 // Skip AddP cases #3 and #5. 2501 } else { 2502 assert(!abase->is_top(), "sanity"); // AddP case #3 2503 if (abase != base) { 2504 igvn->hash_delete(addp); 2505 addp->set_req(AddPNode::Base, base); 2506 if (abase == adr) { 2507 addp->set_req(AddPNode::Address, base); 2508 } else { 2509 // AddP case #4 (adr is array's element offset AddP node) 2510 #ifdef ASSERT 2511 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2512 assert(adr->is_AddP() && atype != NULL && 2513 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2514 #endif 2515 } 2516 igvn->hash_insert(addp); 2517 } 2518 } 2519 // Put on IGVN worklist since at least addp's type was changed above. 2520 record_for_optimizer(addp); 2521 return true; 2522 } 2523 2524 // 2525 // Create a new version of orig_phi if necessary. Returns either the newly 2526 // created phi or an existing phi. Sets create_new to indicate whether a new 2527 // phi was created. Cache the last newly created phi in the node map. 2528 // 2529 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2530 Compile *C = _compile; 2531 PhaseGVN* igvn = _igvn; 2532 new_created = false; 2533 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2534 // nothing to do if orig_phi is bottom memory or matches alias_idx 2535 if (phi_alias_idx == alias_idx) { 2536 return orig_phi; 2537 } 2538 // Have we recently created a Phi for this alias index? 2539 PhiNode *result = get_map_phi(orig_phi->_idx); 2540 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2541 return result; 2542 } 2543 // Previous check may fail when the same wide memory Phi was split into Phis 2544 // for different memory slices. Search all Phis for this region. 2545 if (result != NULL) { 2546 Node* region = orig_phi->in(0); 2547 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2548 Node* phi = region->fast_out(i); 2549 if (phi->is_Phi() && 2550 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2551 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2552 return phi->as_Phi(); 2553 } 2554 } 2555 } 2556 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2557 if (C->do_escape_analysis() == true && !C->failing()) { 2558 // Retry compilation without escape analysis. 2559 // If this is the first failure, the sentinel string will "stick" 2560 // to the Compile object, and the C2Compiler will see it and retry. 2561 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2562 } 2563 return NULL; 2564 } 2565 orig_phi_worklist.append_if_missing(orig_phi); 2566 const TypePtr *atype = C->get_adr_type(alias_idx); 2567 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2568 C->copy_node_notes_to(result, orig_phi); 2569 igvn->set_type(result, result->bottom_type()); 2570 record_for_optimizer(result); 2571 set_map(orig_phi, result); 2572 new_created = true; 2573 return result; 2574 } 2575 2576 // 2577 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2578 // specified alias index. 2579 // 2580 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2581 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2582 Compile *C = _compile; 2583 PhaseGVN* igvn = _igvn; 2584 bool new_phi_created; 2585 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2586 if (!new_phi_created) { 2587 return result; 2588 } 2589 GrowableArray<PhiNode *> phi_list; 2590 GrowableArray<uint> cur_input; 2591 PhiNode *phi = orig_phi; 2592 uint idx = 1; 2593 bool finished = false; 2594 while(!finished) { 2595 while (idx < phi->req()) { 2596 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2597 if (mem != NULL && mem->is_Phi()) { 2598 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2599 if (new_phi_created) { 2600 // found an phi for which we created a new split, push current one on worklist and begin 2601 // processing new one 2602 phi_list.push(phi); 2603 cur_input.push(idx); 2604 phi = mem->as_Phi(); 2605 result = newphi; 2606 idx = 1; 2607 continue; 2608 } else { 2609 mem = newphi; 2610 } 2611 } 2612 if (C->failing()) { 2613 return NULL; 2614 } 2615 result->set_req(idx++, mem); 2616 } 2617 #ifdef ASSERT 2618 // verify that the new Phi has an input for each input of the original 2619 assert( phi->req() == result->req(), "must have same number of inputs."); 2620 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2621 #endif 2622 // Check if all new phi's inputs have specified alias index. 2623 // Otherwise use old phi. 2624 for (uint i = 1; i < phi->req(); i++) { 2625 Node* in = result->in(i); 2626 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2627 } 2628 // we have finished processing a Phi, see if there are any more to do 2629 finished = (phi_list.length() == 0 ); 2630 if (!finished) { 2631 phi = phi_list.pop(); 2632 idx = cur_input.pop(); 2633 PhiNode *prev_result = get_map_phi(phi->_idx); 2634 prev_result->set_req(idx++, result); 2635 result = prev_result; 2636 } 2637 } 2638 return result; 2639 } 2640 2641 // 2642 // The next methods are derived from methods in MemNode. 2643 // 2644 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2645 Node *mem = mmem; 2646 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2647 // means an array I have not precisely typed yet. Do not do any 2648 // alias stuff with it any time soon. 2649 if (toop->base() != Type::AnyPtr && 2650 !(toop->klass() != NULL && 2651 toop->klass()->is_java_lang_Object() && 2652 toop->offset() == Type::OffsetBot)) { 2653 mem = mmem->memory_at(alias_idx); 2654 // Update input if it is progress over what we have now 2655 } 2656 return mem; 2657 } 2658 2659 // 2660 // Move memory users to their memory slices. 2661 // 2662 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2663 Compile* C = _compile; 2664 PhaseGVN* igvn = _igvn; 2665 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2666 assert(tp != NULL, "ptr type"); 2667 int alias_idx = C->get_alias_index(tp); 2668 int general_idx = C->get_general_index(alias_idx); 2669 2670 // Move users first 2671 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2672 Node* use = n->fast_out(i); 2673 if (use->is_MergeMem()) { 2674 MergeMemNode* mmem = use->as_MergeMem(); 2675 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2676 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2677 continue; // Nothing to do 2678 } 2679 // Replace previous general reference to mem node. 2680 uint orig_uniq = C->unique(); 2681 Node* m = find_inst_mem(n, general_idx, orig_phis); 2682 assert(orig_uniq == C->unique(), "no new nodes"); 2683 mmem->set_memory_at(general_idx, m); 2684 --imax; 2685 --i; 2686 } else if (use->is_MemBar()) { 2687 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2688 if (use->req() > MemBarNode::Precedent && 2689 use->in(MemBarNode::Precedent) == n) { 2690 // Don't move related membars. 2691 record_for_optimizer(use); 2692 continue; 2693 } 2694 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2695 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) || 2696 alias_idx == general_idx) { 2697 continue; // Nothing to do 2698 } 2699 // Move to general memory slice. 2700 uint orig_uniq = C->unique(); 2701 Node* m = find_inst_mem(n, general_idx, orig_phis); 2702 assert(orig_uniq == C->unique(), "no new nodes"); 2703 igvn->hash_delete(use); 2704 imax -= use->replace_edge(n, m); 2705 igvn->hash_insert(use); 2706 record_for_optimizer(use); 2707 --i; 2708 #ifdef ASSERT 2709 } else if (use->is_Mem()) { 2710 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2711 // Don't move related cardmark. 2712 continue; 2713 } 2714 // Memory nodes should have new memory input. 2715 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2716 assert(tp != NULL, "ptr type"); 2717 int idx = C->get_alias_index(tp); 2718 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2719 "Following memory nodes should have new memory input or be on the same memory slice"); 2720 } else if (use->is_Phi()) { 2721 // Phi nodes should be split and moved already. 2722 tp = use->as_Phi()->adr_type()->isa_ptr(); 2723 assert(tp != NULL, "ptr type"); 2724 int idx = C->get_alias_index(tp); 2725 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2726 } else { 2727 use->dump(); 2728 assert(false, "should not be here"); 2729 #endif 2730 } 2731 } 2732 } 2733 2734 // 2735 // Search memory chain of "mem" to find a MemNode whose address 2736 // is the specified alias index. 2737 // 2738 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2739 if (orig_mem == NULL) 2740 return orig_mem; 2741 Compile* C = _compile; 2742 PhaseGVN* igvn = _igvn; 2743 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2744 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2745 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 2746 Node *prev = NULL; 2747 Node *result = orig_mem; 2748 while (prev != result) { 2749 prev = result; 2750 if (result == start_mem) 2751 break; // hit one of our sentinels 2752 if (result->is_Mem()) { 2753 const Type *at = igvn->type(result->in(MemNode::Address)); 2754 if (at == Type::TOP) 2755 break; // Dead 2756 assert (at->isa_ptr() != NULL, "pointer type required."); 2757 int idx = C->get_alias_index(at->is_ptr()); 2758 if (idx == alias_idx) 2759 break; // Found 2760 if (!is_instance && (at->isa_oopptr() == NULL || 2761 !at->is_oopptr()->is_known_instance())) { 2762 break; // Do not skip store to general memory slice. 2763 } 2764 result = result->in(MemNode::Memory); 2765 } 2766 if (!is_instance) 2767 continue; // don't search further for non-instance types 2768 // skip over a call which does not affect this memory slice 2769 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2770 Node *proj_in = result->in(0); 2771 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2772 break; // hit one of our sentinels 2773 } else if (proj_in->is_Call()) { 2774 // ArrayCopy node processed here as well 2775 CallNode *call = proj_in->as_Call(); 2776 if (!call->may_modify(toop, igvn)) { 2777 result = call->in(TypeFunc::Memory); 2778 } 2779 } else if (proj_in->is_Initialize()) { 2780 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2781 // Stop if this is the initialization for the object instance which 2782 // which contains this memory slice, otherwise skip over it. 2783 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2784 result = proj_in->in(TypeFunc::Memory); 2785 } 2786 } else if (proj_in->is_MemBar()) { 2787 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() && 2788 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() && 2789 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) { 2790 // clone 2791 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy(); 2792 if (ac->may_modify(toop, igvn)) { 2793 break; 2794 } 2795 } 2796 result = proj_in->in(TypeFunc::Memory); 2797 } 2798 } else if (result->is_MergeMem()) { 2799 MergeMemNode *mmem = result->as_MergeMem(); 2800 result = step_through_mergemem(mmem, alias_idx, toop); 2801 if (result == mmem->base_memory()) { 2802 // Didn't find instance memory, search through general slice recursively. 2803 result = mmem->memory_at(C->get_general_index(alias_idx)); 2804 result = find_inst_mem(result, alias_idx, orig_phis); 2805 if (C->failing()) { 2806 return NULL; 2807 } 2808 mmem->set_memory_at(alias_idx, result); 2809 } 2810 } else if (result->is_Phi() && 2811 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2812 Node *un = result->as_Phi()->unique_input(igvn); 2813 if (un != NULL) { 2814 orig_phis.append_if_missing(result->as_Phi()); 2815 result = un; 2816 } else { 2817 break; 2818 } 2819 } else if (result->is_ClearArray()) { 2820 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2821 // Can not bypass initialization of the instance 2822 // we are looking for. 2823 break; 2824 } 2825 // Otherwise skip it (the call updated 'result' value). 2826 } else if (result->Opcode() == Op_SCMemProj) { 2827 Node* mem = result->in(0); 2828 Node* adr = NULL; 2829 if (mem->is_LoadStore()) { 2830 adr = mem->in(MemNode::Address); 2831 } else { 2832 assert(mem->Opcode() == Op_EncodeISOArray || 2833 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 2834 adr = mem->in(3); // Memory edge corresponds to destination array 2835 } 2836 const Type *at = igvn->type(adr); 2837 if (at != Type::TOP) { 2838 assert(at->isa_ptr() != NULL, "pointer type required."); 2839 int idx = C->get_alias_index(at->is_ptr()); 2840 if (idx == alias_idx) { 2841 // Assert in debug mode 2842 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 2843 break; // In product mode return SCMemProj node 2844 } 2845 } 2846 result = mem->in(MemNode::Memory); 2847 } else if (result->Opcode() == Op_StrInflatedCopy) { 2848 Node* adr = result->in(3); // Memory edge corresponds to destination array 2849 const Type *at = igvn->type(adr); 2850 if (at != Type::TOP) { 2851 assert(at->isa_ptr() != NULL, "pointer type required."); 2852 int idx = C->get_alias_index(at->is_ptr()); 2853 if (idx == alias_idx) { 2854 // Assert in debug mode 2855 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 2856 break; // In product mode return SCMemProj node 2857 } 2858 } 2859 result = result->in(MemNode::Memory); 2860 } 2861 } 2862 if (result->is_Phi()) { 2863 PhiNode *mphi = result->as_Phi(); 2864 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2865 const TypePtr *t = mphi->adr_type(); 2866 if (!is_instance) { 2867 // Push all non-instance Phis on the orig_phis worklist to update inputs 2868 // during Phase 4 if needed. 2869 orig_phis.append_if_missing(mphi); 2870 } else if (C->get_alias_index(t) != alias_idx) { 2871 // Create a new Phi with the specified alias index type. 2872 result = split_memory_phi(mphi, alias_idx, orig_phis); 2873 } 2874 } 2875 // the result is either MemNode, PhiNode, InitializeNode. 2876 return result; 2877 } 2878 2879 // 2880 // Convert the types of unescaped object to instance types where possible, 2881 // propagate the new type information through the graph, and update memory 2882 // edges and MergeMem inputs to reflect the new type. 2883 // 2884 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2885 // The processing is done in 4 phases: 2886 // 2887 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2888 // types for the CheckCastPP for allocations where possible. 2889 // Propagate the new types through users as follows: 2890 // casts and Phi: push users on alloc_worklist 2891 // AddP: cast Base and Address inputs to the instance type 2892 // push any AddP users on alloc_worklist and push any memnode 2893 // users onto memnode_worklist. 2894 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2895 // search the Memory chain for a store with the appropriate type 2896 // address type. If a Phi is found, create a new version with 2897 // the appropriate memory slices from each of the Phi inputs. 2898 // For stores, process the users as follows: 2899 // MemNode: push on memnode_worklist 2900 // MergeMem: push on mergemem_worklist 2901 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2902 // moving the first node encountered of each instance type to the 2903 // the input corresponding to its alias index. 2904 // appropriate memory slice. 2905 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2906 // 2907 // In the following example, the CheckCastPP nodes are the cast of allocation 2908 // results and the allocation of node 29 is unescaped and eligible to be an 2909 // instance type. 2910 // 2911 // We start with: 2912 // 2913 // 7 Parm #memory 2914 // 10 ConI "12" 2915 // 19 CheckCastPP "Foo" 2916 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2917 // 29 CheckCastPP "Foo" 2918 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2919 // 2920 // 40 StoreP 25 7 20 ... alias_index=4 2921 // 50 StoreP 35 40 30 ... alias_index=4 2922 // 60 StoreP 45 50 20 ... alias_index=4 2923 // 70 LoadP _ 60 30 ... alias_index=4 2924 // 80 Phi 75 50 60 Memory alias_index=4 2925 // 90 LoadP _ 80 30 ... alias_index=4 2926 // 100 LoadP _ 80 20 ... alias_index=4 2927 // 2928 // 2929 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2930 // and creating a new alias index for node 30. This gives: 2931 // 2932 // 7 Parm #memory 2933 // 10 ConI "12" 2934 // 19 CheckCastPP "Foo" 2935 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2936 // 29 CheckCastPP "Foo" iid=24 2937 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2938 // 2939 // 40 StoreP 25 7 20 ... alias_index=4 2940 // 50 StoreP 35 40 30 ... alias_index=6 2941 // 60 StoreP 45 50 20 ... alias_index=4 2942 // 70 LoadP _ 60 30 ... alias_index=6 2943 // 80 Phi 75 50 60 Memory alias_index=4 2944 // 90 LoadP _ 80 30 ... alias_index=6 2945 // 100 LoadP _ 80 20 ... alias_index=4 2946 // 2947 // In phase 2, new memory inputs are computed for the loads and stores, 2948 // And a new version of the phi is created. In phase 4, the inputs to 2949 // node 80 are updated and then the memory nodes are updated with the 2950 // values computed in phase 2. This results in: 2951 // 2952 // 7 Parm #memory 2953 // 10 ConI "12" 2954 // 19 CheckCastPP "Foo" 2955 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2956 // 29 CheckCastPP "Foo" iid=24 2957 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2958 // 2959 // 40 StoreP 25 7 20 ... alias_index=4 2960 // 50 StoreP 35 7 30 ... alias_index=6 2961 // 60 StoreP 45 40 20 ... alias_index=4 2962 // 70 LoadP _ 50 30 ... alias_index=6 2963 // 80 Phi 75 40 60 Memory alias_index=4 2964 // 120 Phi 75 50 50 Memory alias_index=6 2965 // 90 LoadP _ 120 30 ... alias_index=6 2966 // 100 LoadP _ 80 20 ... alias_index=4 2967 // 2968 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 2969 GrowableArray<Node *> memnode_worklist; 2970 GrowableArray<PhiNode *> orig_phis; 2971 PhaseIterGVN *igvn = _igvn; 2972 uint new_index_start = (uint) _compile->num_alias_types(); 2973 Arena* arena = Thread::current()->resource_area(); 2974 VectorSet visited(arena); 2975 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2976 uint unique_old = _compile->unique(); 2977 2978 // Phase 1: Process possible allocations from alloc_worklist. 2979 // Create instance types for the CheckCastPP for allocations where possible. 2980 // 2981 // (Note: don't forget to change the order of the second AddP node on 2982 // the alloc_worklist if the order of the worklist processing is changed, 2983 // see the comment in find_second_addp().) 2984 // 2985 while (alloc_worklist.length() != 0) { 2986 Node *n = alloc_worklist.pop(); 2987 uint ni = n->_idx; 2988 if (n->is_Call()) { 2989 CallNode *alloc = n->as_Call(); 2990 // copy escape information to call node 2991 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2992 PointsToNode::EscapeState es = ptn->escape_state(); 2993 // We have an allocation or call which returns a Java object, 2994 // see if it is unescaped. 2995 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2996 continue; 2997 // Find CheckCastPP for the allocate or for the return value of a call 2998 n = alloc->result_cast(); 2999 if (n == NULL) { // No uses except Initialize node 3000 if (alloc->is_Allocate()) { 3001 // Set the scalar_replaceable flag for allocation 3002 // so it could be eliminated if it has no uses. 3003 alloc->as_Allocate()->_is_scalar_replaceable = true; 3004 } 3005 if (alloc->is_CallStaticJava()) { 3006 // Set the scalar_replaceable flag for boxing method 3007 // so it could be eliminated if it has no uses. 3008 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3009 } 3010 continue; 3011 } 3012 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 3013 assert(!alloc->is_Allocate(), "allocation should have unique type"); 3014 continue; 3015 } 3016 3017 // The inline code for Object.clone() casts the allocation result to 3018 // java.lang.Object and then to the actual type of the allocated 3019 // object. Detect this case and use the second cast. 3020 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 3021 // the allocation result is cast to java.lang.Object and then 3022 // to the actual Array type. 3023 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 3024 && (alloc->is_AllocateArray() || 3025 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 3026 Node *cast2 = NULL; 3027 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3028 Node *use = n->fast_out(i); 3029 if (use->is_CheckCastPP()) { 3030 cast2 = use; 3031 break; 3032 } 3033 } 3034 if (cast2 != NULL) { 3035 n = cast2; 3036 } else { 3037 // Non-scalar replaceable if the allocation type is unknown statically 3038 // (reflection allocation), the object can't be restored during 3039 // deoptimization without precise type. 3040 continue; 3041 } 3042 } 3043 3044 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 3045 if (t == NULL) 3046 continue; // not a TypeOopPtr 3047 if (!t->klass_is_exact()) 3048 continue; // not an unique type 3049 3050 if (alloc->is_Allocate()) { 3051 // Set the scalar_replaceable flag for allocation 3052 // so it could be eliminated. 3053 alloc->as_Allocate()->_is_scalar_replaceable = true; 3054 } 3055 if (alloc->is_CallStaticJava()) { 3056 // Set the scalar_replaceable flag for boxing method 3057 // so it could be eliminated. 3058 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3059 } 3060 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 3061 // in order for an object to be scalar-replaceable, it must be: 3062 // - a direct allocation (not a call returning an object) 3063 // - non-escaping 3064 // - eligible to be a unique type 3065 // - not determined to be ineligible by escape analysis 3066 set_map(alloc, n); 3067 set_map(n, alloc); 3068 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 3069 igvn->hash_delete(n); 3070 igvn->set_type(n, tinst); 3071 n->raise_bottom_type(tinst); 3072 igvn->hash_insert(n); 3073 record_for_optimizer(n); 3074 // Allocate an alias index for the header fields. Accesses to 3075 // the header emitted during macro expansion wouldn't have 3076 // correct memory state otherwise. 3077 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes())); 3078 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes())); 3079 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 3080 3081 // First, put on the worklist all Field edges from Connection Graph 3082 // which is more accurate than putting immediate users from Ideal Graph. 3083 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 3084 PointsToNode* tgt = e.get(); 3085 if (tgt->is_Arraycopy()) { 3086 continue; 3087 } 3088 Node* use = tgt->ideal_node(); 3089 assert(tgt->is_Field() && use->is_AddP(), 3090 "only AddP nodes are Field edges in CG"); 3091 if (use->outcnt() > 0) { // Don't process dead nodes 3092 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 3093 if (addp2 != NULL) { 3094 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3095 alloc_worklist.append_if_missing(addp2); 3096 } 3097 alloc_worklist.append_if_missing(use); 3098 } 3099 } 3100 3101 // An allocation may have an Initialize which has raw stores. Scan 3102 // the users of the raw allocation result and push AddP users 3103 // on alloc_worklist. 3104 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms); 3105 assert (raw_result != NULL, "must have an allocation result"); 3106 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 3107 Node *use = raw_result->fast_out(i); 3108 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3109 Node* addp2 = find_second_addp(use, raw_result); 3110 if (addp2 != NULL) { 3111 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3112 alloc_worklist.append_if_missing(addp2); 3113 } 3114 alloc_worklist.append_if_missing(use); 3115 } else if (use->is_MemBar()) { 3116 memnode_worklist.append_if_missing(use); 3117 } 3118 } 3119 } 3120 } else if (n->is_AddP()) { 3121 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3122 if (jobj == NULL || jobj == phantom_obj) { 3123 #ifdef ASSERT 3124 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3125 ptnode_adr(n->_idx)->dump(); 3126 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3127 #endif 3128 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3129 return; 3130 } 3131 Node *base = get_map(jobj->idx()); // CheckCastPP node 3132 if (!split_AddP(n, base)) continue; // wrong type from dead path 3133 } else if (n->is_Phi() || 3134 n->is_CheckCastPP() || 3135 n->is_EncodeP() || 3136 n->is_DecodeN() || 3137 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(n) || 3138 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3139 if (visited.test_set(n->_idx)) { 3140 assert(n->is_Phi(), "loops only through Phi's"); 3141 continue; // already processed 3142 } 3143 JavaObjectNode* jobj = unique_java_object(n); 3144 if (jobj == NULL || jobj == phantom_obj) { 3145 #ifdef ASSERT 3146 ptnode_adr(n->_idx)->dump(); 3147 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3148 #endif 3149 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3150 return; 3151 } else { 3152 Node *val = get_map(jobj->idx()); // CheckCastPP node 3153 TypeNode *tn = n->as_Type(); 3154 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3155 assert(tinst != NULL && tinst->is_known_instance() && 3156 tinst->instance_id() == jobj->idx() , "instance type expected."); 3157 3158 const Type *tn_type = igvn->type(tn); 3159 const TypeOopPtr *tn_t; 3160 if (tn_type->isa_narrowoop()) { 3161 tn_t = tn_type->make_ptr()->isa_oopptr(); 3162 } else { 3163 tn_t = tn_type->isa_oopptr(); 3164 } 3165 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3166 if (tn_type->isa_narrowoop()) { 3167 tn_type = tinst->make_narrowoop(); 3168 } else { 3169 tn_type = tinst; 3170 } 3171 igvn->hash_delete(tn); 3172 igvn->set_type(tn, tn_type); 3173 tn->set_type(tn_type); 3174 igvn->hash_insert(tn); 3175 record_for_optimizer(n); 3176 } else { 3177 assert(tn_type == TypePtr::NULL_PTR || 3178 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3179 "unexpected type"); 3180 continue; // Skip dead path with different type 3181 } 3182 } 3183 } else { 3184 debug_only(n->dump();) 3185 assert(false, "EA: unexpected node"); 3186 continue; 3187 } 3188 // push allocation's users on appropriate worklist 3189 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3190 Node *use = n->fast_out(i); 3191 if (use->is_Mem() && use->in(MemNode::Address) == n) { 3192 // Load/store to instance's field 3193 memnode_worklist.append_if_missing(use); 3194 } else if (use->is_MemBar()) { 3195 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3196 memnode_worklist.append_if_missing(use); 3197 } 3198 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3199 Node* addp2 = find_second_addp(use, n); 3200 if (addp2 != NULL) { 3201 alloc_worklist.append_if_missing(addp2); 3202 } 3203 alloc_worklist.append_if_missing(use); 3204 } else if (use->is_Phi() || 3205 use->is_CheckCastPP() || 3206 use->is_EncodeNarrowPtr() || 3207 use->is_DecodeNarrowPtr() || 3208 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(use) || 3209 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3210 alloc_worklist.append_if_missing(use); 3211 #ifdef ASSERT 3212 } else if (use->is_Mem()) { 3213 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3214 } else if (use->is_MergeMem()) { 3215 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3216 } else if (use->is_SafePoint()) { 3217 // Look for MergeMem nodes for calls which reference unique allocation 3218 // (through CheckCastPP nodes) even for debug info. 3219 Node* m = use->in(TypeFunc::Memory); 3220 if (m->is_MergeMem()) { 3221 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3222 } 3223 } else if (use->Opcode() == Op_EncodeISOArray) { 3224 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3225 // EncodeISOArray overwrites destination array 3226 memnode_worklist.append_if_missing(use); 3227 } 3228 } else if (use->Opcode() == Op_Return) { 3229 assert(_compile->tf()->returns_value_type_as_fields(), "must return a value type"); 3230 // Get ValueKlass by removing the tag bit from the metadata pointer 3231 Node* klass = use->in(TypeFunc::Parms); 3232 intptr_t ptr = igvn->type(klass)->isa_rawptr()->get_con(); 3233 clear_nth_bit(ptr, 0); 3234 assert(Metaspace::contains((void*)ptr), "should be klass"); 3235 assert(((ValueKlass*)ptr)->contains_oops(), "returned value type must contain a reference field"); 3236 } else { 3237 uint op = use->Opcode(); 3238 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) && 3239 (use->in(MemNode::Memory) == n)) { 3240 // They overwrite memory edge corresponding to destination array, 3241 memnode_worklist.append_if_missing(use); 3242 } else if (!(op == Op_CmpP || op == Op_Conv2B || 3243 op == Op_CastP2X || op == Op_StoreCM || 3244 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3245 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3246 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || 3247 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) || 3248 op == Op_ValueType)) { 3249 n->dump(); 3250 use->dump(); 3251 assert(false, "EA: missing allocation reference path"); 3252 } 3253 #endif 3254 } 3255 } 3256 3257 } 3258 3259 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3260 // type, record it in the ArrayCopy node so we know what memory this 3261 // node uses/modified. 3262 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3263 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3264 Node* dest = ac->in(ArrayCopyNode::Dest); 3265 if (dest->is_AddP()) { 3266 dest = get_addp_base(dest); 3267 } 3268 JavaObjectNode* jobj = unique_java_object(dest); 3269 if (jobj != NULL) { 3270 Node *base = get_map(jobj->idx()); 3271 if (base != NULL) { 3272 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3273 ac->_dest_type = base_t; 3274 } 3275 } 3276 Node* src = ac->in(ArrayCopyNode::Src); 3277 if (src->is_AddP()) { 3278 src = get_addp_base(src); 3279 } 3280 jobj = unique_java_object(src); 3281 if (jobj != NULL) { 3282 Node* base = get_map(jobj->idx()); 3283 if (base != NULL) { 3284 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3285 ac->_src_type = base_t; 3286 } 3287 } 3288 } 3289 3290 // New alias types were created in split_AddP(). 3291 uint new_index_end = (uint) _compile->num_alias_types(); 3292 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3293 3294 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3295 // compute new values for Memory inputs (the Memory inputs are not 3296 // actually updated until phase 4.) 3297 if (memnode_worklist.length() == 0) 3298 return; // nothing to do 3299 while (memnode_worklist.length() != 0) { 3300 Node *n = memnode_worklist.pop(); 3301 if (visited.test_set(n->_idx)) 3302 continue; 3303 if (n->is_Phi() || n->is_ClearArray()) { 3304 // we don't need to do anything, but the users must be pushed 3305 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3306 // we don't need to do anything, but the users must be pushed 3307 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory); 3308 if (n == NULL) 3309 continue; 3310 } else if (n->Opcode() == Op_StrCompressedCopy || 3311 n->Opcode() == Op_EncodeISOArray) { 3312 // get the memory projection 3313 n = n->find_out_with(Op_SCMemProj); 3314 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3315 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != NULL && 3316 strcmp(n->as_CallLeaf()->_name, "store_unknown_value") == 0) { 3317 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory); 3318 } else { 3319 assert(n->is_Mem(), "memory node required."); 3320 Node *addr = n->in(MemNode::Address); 3321 const Type *addr_t = igvn->type(addr); 3322 if (addr_t == Type::TOP) 3323 continue; 3324 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3325 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3326 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3327 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3328 if (_compile->failing()) { 3329 return; 3330 } 3331 if (mem != n->in(MemNode::Memory)) { 3332 // We delay the memory edge update since we need old one in 3333 // MergeMem code below when instances memory slices are separated. 3334 set_map(n, mem); 3335 } 3336 if (n->is_Load()) { 3337 continue; // don't push users 3338 } else if (n->is_LoadStore()) { 3339 // get the memory projection 3340 n = n->find_out_with(Op_SCMemProj); 3341 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3342 } 3343 } 3344 // push user on appropriate worklist 3345 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3346 Node *use = n->fast_out(i); 3347 if (use->is_Phi() || use->is_ClearArray()) { 3348 memnode_worklist.append_if_missing(use); 3349 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3350 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3351 continue; 3352 memnode_worklist.append_if_missing(use); 3353 } else if (use->is_MemBar()) { 3354 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3355 memnode_worklist.append_if_missing(use); 3356 } 3357 #ifdef ASSERT 3358 } else if (use->is_Mem()) { 3359 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3360 } else if (use->is_MergeMem()) { 3361 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3362 } else if (use->Opcode() == Op_EncodeISOArray) { 3363 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3364 // EncodeISOArray overwrites destination array 3365 memnode_worklist.append_if_missing(use); 3366 } 3367 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != NULL && 3368 strcmp(use->as_CallLeaf()->_name, "store_unknown_value") == 0) { 3369 // store_unknown_value overwrites destination array 3370 memnode_worklist.append_if_missing(use); 3371 } else { 3372 uint op = use->Opcode(); 3373 if ((use->in(MemNode::Memory) == n) && 3374 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) { 3375 // They overwrite memory edge corresponding to destination array, 3376 memnode_worklist.append_if_missing(use); 3377 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) || 3378 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3379 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3380 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3381 n->dump(); 3382 use->dump(); 3383 assert(false, "EA: missing memory path"); 3384 } 3385 #endif 3386 } 3387 } 3388 } 3389 3390 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3391 // Walk each memory slice moving the first node encountered of each 3392 // instance type to the input corresponding to its alias index. 3393 uint length = _mergemem_worklist.length(); 3394 for( uint next = 0; next < length; ++next ) { 3395 MergeMemNode* nmm = _mergemem_worklist.at(next); 3396 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3397 // Note: we don't want to use MergeMemStream here because we only want to 3398 // scan inputs which exist at the start, not ones we add during processing. 3399 // Note 2: MergeMem may already contains instance memory slices added 3400 // during find_inst_mem() call when memory nodes were processed above. 3401 igvn->hash_delete(nmm); 3402 uint nslices = MIN2(nmm->req(), new_index_start); 3403 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3404 Node* mem = nmm->in(i); 3405 Node* cur = NULL; 3406 if (mem == NULL || mem->is_top()) 3407 continue; 3408 // First, update mergemem by moving memory nodes to corresponding slices 3409 // if their type became more precise since this mergemem was created. 3410 while (mem->is_Mem()) { 3411 const Type *at = igvn->type(mem->in(MemNode::Address)); 3412 if (at != Type::TOP) { 3413 assert (at->isa_ptr() != NULL, "pointer type required."); 3414 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3415 if (idx == i) { 3416 if (cur == NULL) 3417 cur = mem; 3418 } else { 3419 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3420 nmm->set_memory_at(idx, mem); 3421 } 3422 } 3423 } 3424 mem = mem->in(MemNode::Memory); 3425 } 3426 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3427 // Find any instance of the current type if we haven't encountered 3428 // already a memory slice of the instance along the memory chain. 3429 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3430 if((uint)_compile->get_general_index(ni) == i) { 3431 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3432 if (nmm->is_empty_memory(m)) { 3433 Node* result = find_inst_mem(mem, ni, orig_phis); 3434 if (_compile->failing()) { 3435 return; 3436 } 3437 nmm->set_memory_at(ni, result); 3438 } 3439 } 3440 } 3441 } 3442 // Find the rest of instances values 3443 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3444 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3445 Node* result = step_through_mergemem(nmm, ni, tinst); 3446 if (result == nmm->base_memory()) { 3447 // Didn't find instance memory, search through general slice recursively. 3448 result = nmm->memory_at(_compile->get_general_index(ni)); 3449 result = find_inst_mem(result, ni, orig_phis); 3450 if (_compile->failing()) { 3451 return; 3452 } 3453 nmm->set_memory_at(ni, result); 3454 } 3455 } 3456 igvn->hash_insert(nmm); 3457 record_for_optimizer(nmm); 3458 } 3459 3460 // Phase 4: Update the inputs of non-instance memory Phis and 3461 // the Memory input of memnodes 3462 // First update the inputs of any non-instance Phi's from 3463 // which we split out an instance Phi. Note we don't have 3464 // to recursively process Phi's encountered on the input memory 3465 // chains as is done in split_memory_phi() since they will 3466 // also be processed here. 3467 for (int j = 0; j < orig_phis.length(); j++) { 3468 PhiNode *phi = orig_phis.at(j); 3469 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3470 igvn->hash_delete(phi); 3471 for (uint i = 1; i < phi->req(); i++) { 3472 Node *mem = phi->in(i); 3473 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3474 if (_compile->failing()) { 3475 return; 3476 } 3477 if (mem != new_mem) { 3478 phi->set_req(i, new_mem); 3479 } 3480 } 3481 igvn->hash_insert(phi); 3482 record_for_optimizer(phi); 3483 } 3484 3485 // Update the memory inputs of MemNodes with the value we computed 3486 // in Phase 2 and move stores memory users to corresponding memory slices. 3487 // Disable memory split verification code until the fix for 6984348. 3488 // Currently it produces false negative results since it does not cover all cases. 3489 #if 0 // ifdef ASSERT 3490 visited.Reset(); 3491 Node_Stack old_mems(arena, _compile->unique() >> 2); 3492 #endif 3493 for (uint i = 0; i < ideal_nodes.size(); i++) { 3494 Node* n = ideal_nodes.at(i); 3495 Node* nmem = get_map(n->_idx); 3496 assert(nmem != NULL, "sanity"); 3497 if (n->is_Mem()) { 3498 #if 0 // ifdef ASSERT 3499 Node* old_mem = n->in(MemNode::Memory); 3500 if (!visited.test_set(old_mem->_idx)) { 3501 old_mems.push(old_mem, old_mem->outcnt()); 3502 } 3503 #endif 3504 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3505 if (!n->is_Load()) { 3506 // Move memory users of a store first. 3507 move_inst_mem(n, orig_phis); 3508 } 3509 // Now update memory input 3510 igvn->hash_delete(n); 3511 n->set_req(MemNode::Memory, nmem); 3512 igvn->hash_insert(n); 3513 record_for_optimizer(n); 3514 } else { 3515 assert(n->is_Allocate() || n->is_CheckCastPP() || 3516 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3517 } 3518 } 3519 #if 0 // ifdef ASSERT 3520 // Verify that memory was split correctly 3521 while (old_mems.is_nonempty()) { 3522 Node* old_mem = old_mems.node(); 3523 uint old_cnt = old_mems.index(); 3524 old_mems.pop(); 3525 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3526 } 3527 #endif 3528 } 3529 3530 #ifndef PRODUCT 3531 static const char *node_type_names[] = { 3532 "UnknownType", 3533 "JavaObject", 3534 "LocalVar", 3535 "Field", 3536 "Arraycopy" 3537 }; 3538 3539 static const char *esc_names[] = { 3540 "UnknownEscape", 3541 "NoEscape", 3542 "ArgEscape", 3543 "GlobalEscape" 3544 }; 3545 3546 void PointsToNode::dump(bool print_state) const { 3547 NodeType nt = node_type(); 3548 tty->print("%s ", node_type_names[(int) nt]); 3549 if (print_state) { 3550 EscapeState es = escape_state(); 3551 EscapeState fields_es = fields_escape_state(); 3552 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3553 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3554 tty->print("NSR "); 3555 } 3556 if (is_Field()) { 3557 FieldNode* f = (FieldNode*)this; 3558 if (f->is_oop()) 3559 tty->print("oop "); 3560 if (f->offset() > 0) 3561 tty->print("+%d ", f->offset()); 3562 tty->print("("); 3563 for (BaseIterator i(f); i.has_next(); i.next()) { 3564 PointsToNode* b = i.get(); 3565 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3566 } 3567 tty->print(" )"); 3568 } 3569 tty->print("["); 3570 for (EdgeIterator i(this); i.has_next(); i.next()) { 3571 PointsToNode* e = i.get(); 3572 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3573 } 3574 tty->print(" ["); 3575 for (UseIterator i(this); i.has_next(); i.next()) { 3576 PointsToNode* u = i.get(); 3577 bool is_base = false; 3578 if (PointsToNode::is_base_use(u)) { 3579 is_base = true; 3580 u = PointsToNode::get_use_node(u)->as_Field(); 3581 } 3582 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3583 } 3584 tty->print(" ]] "); 3585 if (_node == NULL) 3586 tty->print_cr("<null>"); 3587 else 3588 _node->dump(); 3589 } 3590 3591 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3592 bool first = true; 3593 int ptnodes_length = ptnodes_worklist.length(); 3594 for (int i = 0; i < ptnodes_length; i++) { 3595 PointsToNode *ptn = ptnodes_worklist.at(i); 3596 if (ptn == NULL || !ptn->is_JavaObject()) 3597 continue; 3598 PointsToNode::EscapeState es = ptn->escape_state(); 3599 if ((es != PointsToNode::NoEscape) && !Verbose) { 3600 continue; 3601 } 3602 Node* n = ptn->ideal_node(); 3603 if (n->is_Allocate() || (n->is_CallStaticJava() && 3604 n->as_CallStaticJava()->is_boxing_method())) { 3605 if (first) { 3606 tty->cr(); 3607 tty->print("======== Connection graph for "); 3608 _compile->method()->print_short_name(); 3609 tty->cr(); 3610 first = false; 3611 } 3612 ptn->dump(); 3613 // Print all locals and fields which reference this allocation 3614 for (UseIterator j(ptn); j.has_next(); j.next()) { 3615 PointsToNode* use = j.get(); 3616 if (use->is_LocalVar()) { 3617 use->dump(Verbose); 3618 } else if (Verbose) { 3619 use->dump(); 3620 } 3621 } 3622 tty->cr(); 3623 } 3624 } 3625 } 3626 #endif 3627 3628 void ConnectionGraph::record_for_optimizer(Node *n) { 3629 _igvn->_worklist.push(n); 3630 _igvn->add_users_to_worklist(n); 3631 }