1 /* 2 * Copyright (c) 2005, 2015, 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 "libadt/vectset.hpp" 29 #include "memory/allocation.hpp" 30 #include "opto/c2compiler.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/escape.hpp" 35 #include "opto/phaseX.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/rootnode.hpp" 38 39 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 40 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL), 41 _in_worklist(C->comp_arena()), 42 _next_pidx(0), 43 _collecting(true), 44 _verify(false), 45 _compile(C), 46 _igvn(igvn), 47 _node_map(C->comp_arena()) { 48 // Add unknown java object. 49 add_java_object(C->top(), PointsToNode::GlobalEscape); 50 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 51 // Add ConP(#NULL) and ConN(#NULL) nodes. 52 Node* oop_null = igvn->zerocon(T_OBJECT); 53 assert(oop_null->_idx < nodes_size(), "should be created already"); 54 add_java_object(oop_null, PointsToNode::NoEscape); 55 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 56 if (UseCompressedOops) { 57 Node* noop_null = igvn->zerocon(T_NARROWOOP); 58 assert(noop_null->_idx < nodes_size(), "should be created already"); 59 map_ideal_node(noop_null, null_obj); 60 } 61 _pcmp_neq = NULL; // Should be initialized 62 _pcmp_eq = NULL; 63 } 64 65 bool ConnectionGraph::has_candidates(Compile *C) { 66 // EA brings benefits only when the code has allocations and/or locks which 67 // are represented by ideal Macro nodes. 68 int cnt = C->macro_count(); 69 for (int i = 0; i < cnt; i++) { 70 Node *n = C->macro_node(i); 71 if (n->is_Allocate()) 72 return true; 73 if (n->is_Lock()) { 74 Node* obj = n->as_Lock()->obj_node()->uncast(); 75 if (!(obj->is_Parm() || obj->is_Con())) 76 return true; 77 } 78 if (n->is_CallStaticJava() && 79 n->as_CallStaticJava()->is_boxing_method()) { 80 return true; 81 } 82 } 83 return false; 84 } 85 86 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 87 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]); 88 ResourceMark rm; 89 90 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 91 // to create space for them in ConnectionGraph::_nodes[]. 92 Node* oop_null = igvn->zerocon(T_OBJECT); 93 Node* noop_null = igvn->zerocon(T_NARROWOOP); 94 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 95 // Perform escape analysis 96 if (congraph->compute_escape()) { 97 // There are non escaping objects. 98 C->set_congraph(congraph); 99 } 100 // Cleanup. 101 if (oop_null->outcnt() == 0) 102 igvn->hash_delete(oop_null); 103 if (noop_null->outcnt() == 0) 104 igvn->hash_delete(noop_null); 105 } 106 107 bool ConnectionGraph::compute_escape() { 108 Compile* C = _compile; 109 PhaseGVN* igvn = _igvn; 110 111 // Worklists used by EA. 112 Unique_Node_List delayed_worklist; 113 GrowableArray<Node*> alloc_worklist; 114 GrowableArray<Node*> ptr_cmp_worklist; 115 GrowableArray<Node*> storestore_worklist; 116 GrowableArray<PointsToNode*> ptnodes_worklist; 117 GrowableArray<JavaObjectNode*> java_objects_worklist; 118 GrowableArray<JavaObjectNode*> non_escaped_worklist; 119 GrowableArray<FieldNode*> oop_fields_worklist; 120 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 121 122 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]); 123 124 // 1. Populate Connection Graph (CG) with PointsTo nodes. 125 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 126 // Initialize worklist 127 if (C->root() != NULL) { 128 ideal_nodes.push(C->root()); 129 } 130 // Processed ideal nodes are unique on ideal_nodes list 131 // but several ideal nodes are mapped to the phantom_obj. 132 // To avoid duplicated entries on the following worklists 133 // add the phantom_obj only once to them. 134 ptnodes_worklist.append(phantom_obj); 135 java_objects_worklist.append(phantom_obj); 136 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 137 Node* n = ideal_nodes.at(next); 138 // Create PointsTo nodes and add them to Connection Graph. Called 139 // only once per ideal node since ideal_nodes is Unique_Node list. 140 add_node_to_connection_graph(n, &delayed_worklist); 141 PointsToNode* ptn = ptnode_adr(n->_idx); 142 if (ptn != NULL && ptn != phantom_obj) { 143 ptnodes_worklist.append(ptn); 144 if (ptn->is_JavaObject()) { 145 java_objects_worklist.append(ptn->as_JavaObject()); 146 if ((n->is_Allocate() || n->is_CallStaticJava()) && 147 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 148 // Only allocations and java static calls results are interesting. 149 non_escaped_worklist.append(ptn->as_JavaObject()); 150 } 151 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 152 oop_fields_worklist.append(ptn->as_Field()); 153 } 154 } 155 if (n->is_MergeMem()) { 156 // Collect all MergeMem nodes to add memory slices for 157 // scalar replaceable objects in split_unique_types(). 158 _mergemem_worklist.append(n->as_MergeMem()); 159 } else if (OptimizePtrCompare && n->is_Cmp() && 160 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) { 161 // Collect compare pointers nodes. 162 ptr_cmp_worklist.append(n); 163 } else if (n->is_MemBarStoreStore()) { 164 // Collect all MemBarStoreStore nodes so that depending on the 165 // escape status of the associated Allocate node some of them 166 // may be eliminated. 167 storestore_worklist.append(n); 168 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 169 (n->req() > MemBarNode::Precedent)) { 170 record_for_optimizer(n); 171 #ifdef ASSERT 172 } else if (n->is_AddP()) { 173 // Collect address nodes for graph verification. 174 addp_worklist.append(n); 175 #endif 176 } 177 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 178 Node* m = n->fast_out(i); // Get user 179 ideal_nodes.push(m); 180 } 181 } 182 if (non_escaped_worklist.length() == 0) { 183 _collecting = false; 184 return false; // Nothing to do. 185 } 186 // Add final simple edges to graph. 187 while(delayed_worklist.size() > 0) { 188 Node* n = delayed_worklist.pop(); 189 add_final_edges(n); 190 } 191 int ptnodes_length = ptnodes_worklist.length(); 192 193 #ifdef ASSERT 194 if (VerifyConnectionGraph) { 195 // Verify that no new simple edges could be created and all 196 // local vars has edges. 197 _verify = true; 198 for (int next = 0; next < ptnodes_length; ++next) { 199 PointsToNode* ptn = ptnodes_worklist.at(next); 200 add_final_edges(ptn->ideal_node()); 201 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 202 ptn->dump(); 203 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 204 } 205 } 206 _verify = false; 207 } 208 #endif 209 210 // 2. Finish Graph construction by propagating references to all 211 // java objects through graph. 212 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 213 java_objects_worklist, oop_fields_worklist)) { 214 // All objects escaped or hit time or iterations limits. 215 _collecting = false; 216 return false; 217 } 218 219 // 3. Adjust scalar_replaceable state of nonescaping objects and push 220 // scalar replaceable allocations on alloc_worklist for processing 221 // in split_unique_types(). 222 int non_escaped_length = non_escaped_worklist.length(); 223 for (int next = 0; next < non_escaped_length; next++) { 224 JavaObjectNode* ptn = non_escaped_worklist.at(next); 225 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 226 Node* n = ptn->ideal_node(); 227 if (n->is_Allocate()) { 228 n->as_Allocate()->_is_non_escaping = noescape; 229 } 230 if (n->is_CallStaticJava()) { 231 n->as_CallStaticJava()->_is_non_escaping = noescape; 232 } 233 if (noescape && ptn->scalar_replaceable()) { 234 adjust_scalar_replaceable_state(ptn); 235 if (ptn->scalar_replaceable()) { 236 alloc_worklist.append(ptn->ideal_node()); 237 } 238 } 239 } 240 241 #ifdef ASSERT 242 if (VerifyConnectionGraph) { 243 // Verify that graph is complete - no new edges could be added or needed. 244 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 245 java_objects_worklist, addp_worklist); 246 } 247 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 248 assert(null_obj->escape_state() == PointsToNode::NoEscape && 249 null_obj->edge_count() == 0 && 250 !null_obj->arraycopy_src() && 251 !null_obj->arraycopy_dst(), "sanity"); 252 #endif 253 254 _collecting = false; 255 256 } // TracePhase t3("connectionGraph") 257 258 // 4. Optimize ideal graph based on EA information. 259 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 260 if (has_non_escaping_obj) { 261 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 262 } 263 264 #ifndef PRODUCT 265 if (PrintEscapeAnalysis) { 266 dump(ptnodes_worklist); // Dump ConnectionGraph 267 } 268 #endif 269 270 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 271 #ifdef ASSERT 272 if (VerifyConnectionGraph) { 273 int alloc_length = alloc_worklist.length(); 274 for (int next = 0; next < alloc_length; ++next) { 275 Node* n = alloc_worklist.at(next); 276 PointsToNode* ptn = ptnode_adr(n->_idx); 277 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 278 } 279 } 280 #endif 281 282 // 5. Separate memory graph for scalar replaceable allcations. 283 if (has_scalar_replaceable_candidates && 284 C->AliasLevel() >= 3 && EliminateAllocations) { 285 // Now use the escape information to create unique types for 286 // scalar replaceable objects. 287 split_unique_types(alloc_worklist); 288 if (C->failing()) return false; 289 C->print_method(PHASE_AFTER_EA, 2); 290 291 #ifdef ASSERT 292 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 293 tty->print("=== No allocations eliminated for "); 294 C->method()->print_short_name(); 295 if(!EliminateAllocations) { 296 tty->print(" since EliminateAllocations is off ==="); 297 } else if(!has_scalar_replaceable_candidates) { 298 tty->print(" since there are no scalar replaceable candidates ==="); 299 } else if(C->AliasLevel() < 3) { 300 tty->print(" since AliasLevel < 3 ==="); 301 } 302 tty->cr(); 303 #endif 304 } 305 return has_non_escaping_obj; 306 } 307 308 // Utility function for nodes that load an object 309 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 310 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 311 // ThreadLocal has RawPtr type. 312 const Type* t = _igvn->type(n); 313 if (t->make_ptr() != NULL) { 314 Node* adr = n->in(MemNode::Address); 315 #ifdef ASSERT 316 if (!adr->is_AddP()) { 317 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 318 } else { 319 assert((ptnode_adr(adr->_idx) == NULL || 320 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 321 } 322 #endif 323 add_local_var_and_edge(n, PointsToNode::NoEscape, 324 adr, delayed_worklist); 325 } 326 } 327 328 // Populate Connection Graph with PointsTo nodes and create simple 329 // connection graph edges. 330 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 331 assert(!_verify, "this method sould not be called for verification"); 332 PhaseGVN* igvn = _igvn; 333 uint n_idx = n->_idx; 334 PointsToNode* n_ptn = ptnode_adr(n_idx); 335 if (n_ptn != NULL) 336 return; // No need to redefine PointsTo node during first iteration. 337 338 if (n->is_Call()) { 339 // Arguments to allocation and locking don't escape. 340 if (n->is_AbstractLock()) { 341 // Put Lock and Unlock nodes on IGVN worklist to process them during 342 // first IGVN optimization when escape information is still available. 343 record_for_optimizer(n); 344 } else if (n->is_Allocate()) { 345 add_call_node(n->as_Call()); 346 record_for_optimizer(n); 347 } else { 348 if (n->is_CallStaticJava()) { 349 const char* name = n->as_CallStaticJava()->_name; 350 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 351 return; // Skip uncommon traps 352 } 353 // Don't mark as processed since call's arguments have to be processed. 354 delayed_worklist->push(n); 355 // Check if a call returns an object. 356 if ((n->as_Call()->returns_pointer() && 357 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) || 358 (n->is_CallStaticJava() && 359 n->as_CallStaticJava()->is_boxing_method())) { 360 add_call_node(n->as_Call()); 361 } 362 } 363 return; 364 } 365 // Put this check here to process call arguments since some call nodes 366 // point to phantom_obj. 367 if (n_ptn == phantom_obj || n_ptn == null_obj) 368 return; // Skip predefined nodes. 369 370 int opcode = n->Opcode(); 371 switch (opcode) { 372 case Op_AddP: { 373 Node* base = get_addp_base(n); 374 PointsToNode* ptn_base = ptnode_adr(base->_idx); 375 // Field nodes are created for all field types. They are used in 376 // adjust_scalar_replaceable_state() and split_unique_types(). 377 // Note, non-oop fields will have only base edges in Connection 378 // Graph because such fields are not used for oop loads and stores. 379 int offset = address_offset(n, igvn); 380 add_field(n, PointsToNode::NoEscape, offset); 381 if (ptn_base == NULL) { 382 delayed_worklist->push(n); // Process it later. 383 } else { 384 n_ptn = ptnode_adr(n_idx); 385 add_base(n_ptn->as_Field(), ptn_base); 386 } 387 break; 388 } 389 case Op_CastX2P: { 390 map_ideal_node(n, phantom_obj); 391 break; 392 } 393 case Op_CastPP: 394 case Op_CheckCastPP: 395 case Op_EncodeP: 396 case Op_DecodeN: 397 case Op_EncodePKlass: 398 case Op_DecodeNKlass: { 399 add_local_var_and_edge(n, PointsToNode::NoEscape, 400 n->in(1), delayed_worklist); 401 break; 402 } 403 case Op_CMoveP: { 404 add_local_var(n, PointsToNode::NoEscape); 405 // Do not add edges during first iteration because some could be 406 // not defined yet. 407 delayed_worklist->push(n); 408 break; 409 } 410 case Op_ConP: 411 case Op_ConN: 412 case Op_ConNKlass: { 413 // assume all oop constants globally escape except for null 414 PointsToNode::EscapeState es; 415 const Type* t = igvn->type(n); 416 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 417 es = PointsToNode::NoEscape; 418 } else { 419 es = PointsToNode::GlobalEscape; 420 } 421 add_java_object(n, es); 422 break; 423 } 424 case Op_CreateEx: { 425 // assume that all exception objects globally escape 426 map_ideal_node(n, phantom_obj); 427 break; 428 } 429 case Op_LoadKlass: 430 case Op_LoadNKlass: { 431 // Unknown class is loaded 432 map_ideal_node(n, phantom_obj); 433 break; 434 } 435 case Op_LoadP: 436 case Op_LoadN: 437 case Op_LoadPLocked: { 438 add_objload_to_connection_graph(n, delayed_worklist); 439 break; 440 } 441 case Op_Parm: { 442 map_ideal_node(n, phantom_obj); 443 break; 444 } 445 case Op_PartialSubtypeCheck: { 446 // Produces Null or notNull and is used in only in CmpP so 447 // phantom_obj could be used. 448 map_ideal_node(n, phantom_obj); // Result is unknown 449 break; 450 } 451 case Op_Phi: { 452 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 453 // ThreadLocal has RawPtr type. 454 const Type* t = n->as_Phi()->type(); 455 if (t->make_ptr() != NULL) { 456 add_local_var(n, PointsToNode::NoEscape); 457 // Do not add edges during first iteration because some could be 458 // not defined yet. 459 delayed_worklist->push(n); 460 } 461 break; 462 } 463 case Op_Proj: { 464 // we are only interested in the oop result projection from a call 465 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 466 n->in(0)->as_Call()->returns_pointer()) { 467 add_local_var_and_edge(n, PointsToNode::NoEscape, 468 n->in(0), delayed_worklist); 469 } 470 break; 471 } 472 case Op_Rethrow: // Exception object escapes 473 case Op_Return: { 474 if (n->req() > TypeFunc::Parms && 475 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 476 // Treat Return value as LocalVar with GlobalEscape escape state. 477 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 478 n->in(TypeFunc::Parms), delayed_worklist); 479 } 480 break; 481 } 482 case Op_GetAndSetP: 483 case Op_GetAndSetN: { 484 add_objload_to_connection_graph(n, delayed_worklist); 485 // fallthrough 486 } 487 case Op_StoreP: 488 case Op_StoreN: 489 case Op_StoreNKlass: 490 case Op_StorePConditional: 491 case Op_CompareAndSwapP: 492 case Op_CompareAndSwapN: { 493 Node* adr = n->in(MemNode::Address); 494 const Type *adr_type = igvn->type(adr); 495 adr_type = adr_type->make_ptr(); 496 if (adr_type == NULL) { 497 break; // skip dead nodes 498 } 499 if (adr_type->isa_oopptr() || 500 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 501 (adr_type == TypeRawPtr::NOTNULL && 502 adr->in(AddPNode::Address)->is_Proj() && 503 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 504 delayed_worklist->push(n); // Process it later. 505 #ifdef ASSERT 506 assert(adr->is_AddP(), "expecting an AddP"); 507 if (adr_type == TypeRawPtr::NOTNULL) { 508 // Verify a raw address for a store captured by Initialize node. 509 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 510 assert(offs != Type::OffsetBot, "offset must be a constant"); 511 } 512 #endif 513 } else { 514 // Ignore copy the displaced header to the BoxNode (OSR compilation). 515 if (adr->is_BoxLock()) 516 break; 517 // Stored value escapes in unsafe access. 518 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) { 519 // Pointer stores in G1 barriers looks like unsafe access. 520 // Ignore such stores to be able scalar replace non-escaping 521 // allocations. 522 if (UseG1GC && adr->is_AddP()) { 523 Node* base = get_addp_base(adr); 524 if (base->Opcode() == Op_LoadP && 525 base->in(MemNode::Address)->is_AddP()) { 526 adr = base->in(MemNode::Address); 527 Node* tls = get_addp_base(adr); 528 if (tls->Opcode() == Op_ThreadLocal) { 529 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 530 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() + 531 PtrQueue::byte_offset_of_buf())) { 532 break; // G1 pre barier previous oop value store. 533 } 534 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() + 535 PtrQueue::byte_offset_of_buf())) { 536 break; // G1 post barier card address store. 537 } 538 } 539 } 540 } 541 delayed_worklist->push(n); // Process unsafe access later. 542 break; 543 } 544 #ifdef ASSERT 545 n->dump(1); 546 assert(false, "not unsafe or G1 barrier raw StoreP"); 547 #endif 548 } 549 break; 550 } 551 case Op_AryEq: 552 case Op_StrComp: 553 case Op_StrEquals: 554 case Op_StrIndexOf: 555 case Op_EncodeISOArray: { 556 add_local_var(n, PointsToNode::ArgEscape); 557 delayed_worklist->push(n); // Process it later. 558 break; 559 } 560 case Op_ThreadLocal: { 561 add_java_object(n, PointsToNode::ArgEscape); 562 break; 563 } 564 default: 565 ; // Do nothing for nodes not related to EA. 566 } 567 return; 568 } 569 570 #ifdef ASSERT 571 #define ELSE_FAIL(name) \ 572 /* Should not be called for not pointer type. */ \ 573 n->dump(1); \ 574 assert(false, name); \ 575 break; 576 #else 577 #define ELSE_FAIL(name) \ 578 break; 579 #endif 580 581 // Add final simple edges to graph. 582 void ConnectionGraph::add_final_edges(Node *n) { 583 PointsToNode* n_ptn = ptnode_adr(n->_idx); 584 #ifdef ASSERT 585 if (_verify && n_ptn->is_JavaObject()) 586 return; // This method does not change graph for JavaObject. 587 #endif 588 589 if (n->is_Call()) { 590 process_call_arguments(n->as_Call()); 591 return; 592 } 593 assert(n->is_Store() || n->is_LoadStore() || 594 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 595 "node should be registered already"); 596 int opcode = n->Opcode(); 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()) { 666 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 667 break; 668 } 669 ELSE_FAIL("Op_Proj"); 670 } 671 case Op_Rethrow: // Exception object escapes 672 case Op_Return: { 673 if (n->req() > TypeFunc::Parms && 674 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 675 // Treat Return value as LocalVar with GlobalEscape escape state. 676 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 677 n->in(TypeFunc::Parms), NULL); 678 break; 679 } 680 ELSE_FAIL("Op_Return"); 681 } 682 case Op_StoreP: 683 case Op_StoreN: 684 case Op_StoreNKlass: 685 case Op_StorePConditional: 686 case Op_CompareAndSwapP: 687 case Op_CompareAndSwapN: 688 case Op_GetAndSetP: 689 case Op_GetAndSetN: { 690 Node* adr = n->in(MemNode::Address); 691 const Type *adr_type = _igvn->type(adr); 692 adr_type = adr_type->make_ptr(); 693 #ifdef ASSERT 694 if (adr_type == NULL) { 695 n->dump(1); 696 assert(adr_type != NULL, "dead node should not be on list"); 697 break; 698 } 699 #endif 700 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) { 701 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 702 } 703 if (adr_type->isa_oopptr() || 704 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 705 (adr_type == TypeRawPtr::NOTNULL && 706 adr->in(AddPNode::Address)->is_Proj() && 707 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 708 // Point Address to Value 709 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 710 assert(adr_ptn != NULL && 711 adr_ptn->as_Field()->is_oop(), "node should be registered"); 712 Node *val = n->in(MemNode::ValueIn); 713 PointsToNode* ptn = ptnode_adr(val->_idx); 714 assert(ptn != NULL, "node should be registered"); 715 add_edge(adr_ptn, ptn); 716 break; 717 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) { 718 // Stored value escapes in unsafe access. 719 Node *val = n->in(MemNode::ValueIn); 720 PointsToNode* ptn = ptnode_adr(val->_idx); 721 assert(ptn != NULL, "node should be registered"); 722 set_escape_state(ptn, PointsToNode::GlobalEscape); 723 // Add edge to object for unsafe access with offset. 724 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 725 assert(adr_ptn != NULL, "node should be registered"); 726 if (adr_ptn->is_Field()) { 727 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 728 add_edge(adr_ptn, ptn); 729 } 730 break; 731 } 732 ELSE_FAIL("Op_StoreP"); 733 } 734 case Op_AryEq: 735 case Op_StrComp: 736 case Op_StrEquals: 737 case Op_StrIndexOf: 738 case Op_EncodeISOArray: { 739 // char[] arrays passed to string intrinsic do not escape but 740 // they are not scalar replaceable. Adjust escape state for them. 741 // Start from in(2) edge since in(1) is memory edge. 742 for (uint i = 2; i < n->req(); i++) { 743 Node* adr = n->in(i); 744 const Type* at = _igvn->type(adr); 745 if (!adr->is_top() && at->isa_ptr()) { 746 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 747 at->isa_ptr() != NULL, "expecting a pointer"); 748 if (adr->is_AddP()) { 749 adr = get_addp_base(adr); 750 } 751 PointsToNode* ptn = ptnode_adr(adr->_idx); 752 assert(ptn != NULL, "node should be registered"); 753 add_edge(n_ptn, ptn); 754 } 755 } 756 break; 757 } 758 default: { 759 // This method should be called only for EA specific nodes which may 760 // miss some edges when they were created. 761 #ifdef ASSERT 762 n->dump(1); 763 #endif 764 guarantee(false, "unknown node"); 765 } 766 } 767 return; 768 } 769 770 void ConnectionGraph::add_call_node(CallNode* call) { 771 assert(call->returns_pointer(), "only for call which returns pointer"); 772 uint call_idx = call->_idx; 773 if (call->is_Allocate()) { 774 Node* k = call->in(AllocateNode::KlassNode); 775 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 776 assert(kt != NULL, "TypeKlassPtr required."); 777 ciKlass* cik = kt->klass(); 778 PointsToNode::EscapeState es = PointsToNode::NoEscape; 779 bool scalar_replaceable = true; 780 if (call->is_AllocateArray()) { 781 if (!cik->is_array_klass()) { // StressReflectiveCode 782 es = PointsToNode::GlobalEscape; 783 } else { 784 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 785 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 786 // Not scalar replaceable if the length is not constant or too big. 787 scalar_replaceable = false; 788 } 789 } 790 } else { // Allocate instance 791 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 792 cik->is_subclass_of(_compile->env()->Reference_klass()) || 793 !cik->is_instance_klass() || // StressReflectiveCode 794 cik->as_instance_klass()->has_finalizer()) { 795 es = PointsToNode::GlobalEscape; 796 } 797 } 798 add_java_object(call, es); 799 PointsToNode* ptn = ptnode_adr(call_idx); 800 if (!scalar_replaceable && ptn->scalar_replaceable()) { 801 ptn->set_scalar_replaceable(false); 802 } 803 } else if (call->is_CallStaticJava()) { 804 // Call nodes could be different types: 805 // 806 // 1. CallDynamicJavaNode (what happened during call is unknown): 807 // 808 // - mapped to GlobalEscape JavaObject node if oop is returned; 809 // 810 // - all oop arguments are escaping globally; 811 // 812 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 813 // 814 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 815 // 816 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 817 // - mapped to NoEscape JavaObject node if non-escaping object allocated 818 // during call is returned; 819 // - mapped to ArgEscape LocalVar node pointed to object arguments 820 // which are returned and does not escape during call; 821 // 822 // - oop arguments escaping status is defined by bytecode analysis; 823 // 824 // For a static call, we know exactly what method is being called. 825 // Use bytecode estimator to record whether the call's return value escapes. 826 ciMethod* meth = call->as_CallJava()->method(); 827 if (meth == NULL) { 828 const char* name = call->as_CallStaticJava()->_name; 829 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 830 // Returns a newly allocated unescaped object. 831 add_java_object(call, PointsToNode::NoEscape); 832 ptnode_adr(call_idx)->set_scalar_replaceable(false); 833 } else if (meth->is_boxing_method()) { 834 // Returns boxing object 835 PointsToNode::EscapeState es; 836 vmIntrinsics::ID intr = meth->intrinsic_id(); 837 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 838 // It does not escape if object is always allocated. 839 es = PointsToNode::NoEscape; 840 } else { 841 // It escapes globally if object could be loaded from cache. 842 es = PointsToNode::GlobalEscape; 843 } 844 add_java_object(call, es); 845 } else { 846 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 847 call_analyzer->copy_dependencies(_compile->dependencies()); 848 if (call_analyzer->is_return_allocated()) { 849 // Returns a newly allocated unescaped object, simply 850 // update dependency information. 851 // Mark it as NoEscape so that objects referenced by 852 // it's fields will be marked as NoEscape at least. 853 add_java_object(call, PointsToNode::NoEscape); 854 ptnode_adr(call_idx)->set_scalar_replaceable(false); 855 } else { 856 // Determine whether any arguments are returned. 857 const TypeTuple* d = call->tf()->domain(); 858 bool ret_arg = false; 859 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 860 if (d->field_at(i)->isa_ptr() != NULL && 861 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 862 ret_arg = true; 863 break; 864 } 865 } 866 if (ret_arg) { 867 add_local_var(call, PointsToNode::ArgEscape); 868 } else { 869 // Returns unknown object. 870 map_ideal_node(call, phantom_obj); 871 } 872 } 873 } 874 } else { 875 // An other type of call, assume the worst case: 876 // returned value is unknown and globally escapes. 877 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 878 map_ideal_node(call, phantom_obj); 879 } 880 } 881 882 void ConnectionGraph::process_call_arguments(CallNode *call) { 883 bool is_arraycopy = false; 884 switch (call->Opcode()) { 885 #ifdef ASSERT 886 case Op_Allocate: 887 case Op_AllocateArray: 888 case Op_Lock: 889 case Op_Unlock: 890 assert(false, "should be done already"); 891 break; 892 #endif 893 case Op_ArrayCopy: 894 case Op_CallLeafNoFP: 895 // Most array copies are ArrayCopy nodes at this point but there 896 // are still a few direct calls to the copy subroutines (See 897 // PhaseStringOpts::copy_string()) 898 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 899 (call->as_CallLeaf()->_name != NULL && 900 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0); 901 // fall through 902 case Op_CallLeaf: { 903 // Stub calls, objects do not escape but they are not scale replaceable. 904 // Adjust escape state for outgoing arguments. 905 const TypeTuple * d = call->tf()->domain(); 906 bool src_has_oops = false; 907 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 908 const Type* at = d->field_at(i); 909 Node *arg = call->in(i); 910 if (arg == NULL) { 911 continue; 912 } 913 const Type *aat = _igvn->type(arg); 914 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 915 continue; 916 if (arg->is_AddP()) { 917 // 918 // The inline_native_clone() case when the arraycopy stub is called 919 // after the allocation before Initialize and CheckCastPP nodes. 920 // Or normal arraycopy for object arrays case. 921 // 922 // Set AddP's base (Allocate) as not scalar replaceable since 923 // pointer to the base (with offset) is passed as argument. 924 // 925 arg = get_addp_base(arg); 926 } 927 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 928 assert(arg_ptn != NULL, "should be registered"); 929 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 930 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 931 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 932 aat->isa_ptr() != NULL, "expecting an Ptr"); 933 bool arg_has_oops = aat->isa_oopptr() && 934 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 935 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 936 if (i == TypeFunc::Parms) { 937 src_has_oops = arg_has_oops; 938 } 939 // 940 // src or dst could be j.l.Object when other is basic type array: 941 // 942 // arraycopy(char[],0,Object*,0,size); 943 // arraycopy(Object*,0,char[],0,size); 944 // 945 // Don't add edges in such cases. 946 // 947 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 948 arg_has_oops && (i > TypeFunc::Parms); 949 #ifdef ASSERT 950 if (!(is_arraycopy || 951 (call->as_CallLeaf()->_name != NULL && 952 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 953 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 || 954 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 955 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 956 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 957 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 958 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 959 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 960 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 961 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 962 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 963 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 964 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 965 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 966 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0) 967 ))) { 968 call->dump(); 969 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name)); 970 } 971 #endif 972 // Always process arraycopy's destination object since 973 // we need to add all possible edges to references in 974 // source object. 975 if (arg_esc >= PointsToNode::ArgEscape && 976 !arg_is_arraycopy_dest) { 977 continue; 978 } 979 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 980 if (arg_is_arraycopy_dest) { 981 Node* src = call->in(TypeFunc::Parms); 982 if (src->is_AddP()) { 983 src = get_addp_base(src); 984 } 985 PointsToNode* src_ptn = ptnode_adr(src->_idx); 986 assert(src_ptn != NULL, "should be registered"); 987 if (arg_ptn != src_ptn) { 988 // Special arraycopy edge: 989 // A destination object's field can't have the source object 990 // as base since objects escape states are not related. 991 // Only escape state of destination object's fields affects 992 // escape state of fields in source object. 993 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn); 994 } 995 } 996 } 997 } 998 break; 999 } 1000 case Op_CallStaticJava: { 1001 // For a static call, we know exactly what method is being called. 1002 // Use bytecode estimator to record the call's escape affects 1003 #ifdef ASSERT 1004 const char* name = call->as_CallStaticJava()->_name; 1005 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1006 #endif 1007 ciMethod* meth = call->as_CallJava()->method(); 1008 if ((meth != NULL) && meth->is_boxing_method()) { 1009 break; // Boxing methods do not modify any oops. 1010 } 1011 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1012 // fall-through if not a Java method or no analyzer information 1013 if (call_analyzer != NULL) { 1014 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1015 const TypeTuple* d = call->tf()->domain(); 1016 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1017 const Type* at = d->field_at(i); 1018 int k = i - TypeFunc::Parms; 1019 Node* arg = call->in(i); 1020 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1021 if (at->isa_ptr() != NULL && 1022 call_analyzer->is_arg_returned(k)) { 1023 // The call returns arguments. 1024 if (call_ptn != NULL) { // Is call's result used? 1025 assert(call_ptn->is_LocalVar(), "node should be registered"); 1026 assert(arg_ptn != NULL, "node should be registered"); 1027 add_edge(call_ptn, arg_ptn); 1028 } 1029 } 1030 if (at->isa_oopptr() != NULL && 1031 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1032 if (!call_analyzer->is_arg_stack(k)) { 1033 // The argument global escapes 1034 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1035 } else { 1036 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1037 if (!call_analyzer->is_arg_local(k)) { 1038 // The argument itself doesn't escape, but any fields might 1039 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1040 } 1041 } 1042 } 1043 } 1044 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1045 // The call returns arguments. 1046 assert(call_ptn->edge_count() > 0, "sanity"); 1047 if (!call_analyzer->is_return_local()) { 1048 // Returns also unknown object. 1049 add_edge(call_ptn, phantom_obj); 1050 } 1051 } 1052 break; 1053 } 1054 } 1055 default: { 1056 // Fall-through here if not a Java method or no analyzer information 1057 // or some other type of call, assume the worst case: all arguments 1058 // globally escape. 1059 const TypeTuple* d = call->tf()->domain(); 1060 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1061 const Type* at = d->field_at(i); 1062 if (at->isa_oopptr() != NULL) { 1063 Node* arg = call->in(i); 1064 if (arg->is_AddP()) { 1065 arg = get_addp_base(arg); 1066 } 1067 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1068 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1069 } 1070 } 1071 } 1072 } 1073 } 1074 1075 1076 // Finish Graph construction. 1077 bool ConnectionGraph::complete_connection_graph( 1078 GrowableArray<PointsToNode*>& ptnodes_worklist, 1079 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1080 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1081 GrowableArray<FieldNode*>& oop_fields_worklist) { 1082 // Normally only 1-3 passes needed to build Connection Graph depending 1083 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1084 // Set limit to 20 to catch situation when something did go wrong and 1085 // bailout Escape Analysis. 1086 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1087 #define CG_BUILD_ITER_LIMIT 20 1088 1089 // Propagate GlobalEscape and ArgEscape escape states and check that 1090 // we still have non-escaping objects. The method pushs on _worklist 1091 // Field nodes which reference phantom_object. 1092 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1093 return false; // Nothing to do. 1094 } 1095 // Now propagate references to all JavaObject nodes. 1096 int java_objects_length = java_objects_worklist.length(); 1097 elapsedTimer time; 1098 bool timeout = false; 1099 int new_edges = 1; 1100 int iterations = 0; 1101 do { 1102 while ((new_edges > 0) && 1103 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1104 double start_time = time.seconds(); 1105 time.start(); 1106 new_edges = 0; 1107 // Propagate references to phantom_object for nodes pushed on _worklist 1108 // by find_non_escaped_objects() and find_field_value(). 1109 new_edges += add_java_object_edges(phantom_obj, false); 1110 for (int next = 0; next < java_objects_length; ++next) { 1111 JavaObjectNode* ptn = java_objects_worklist.at(next); 1112 new_edges += add_java_object_edges(ptn, true); 1113 1114 #define SAMPLE_SIZE 4 1115 if ((next % SAMPLE_SIZE) == 0) { 1116 // Each 4 iterations calculate how much time it will take 1117 // to complete graph construction. 1118 time.stop(); 1119 // Poll for requests from shutdown mechanism to quiesce compiler 1120 // because Connection graph construction may take long time. 1121 CompileBroker::maybe_block(); 1122 double stop_time = time.seconds(); 1123 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1124 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1125 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1126 timeout = true; 1127 break; // Timeout 1128 } 1129 start_time = stop_time; 1130 time.start(); 1131 } 1132 #undef SAMPLE_SIZE 1133 1134 } 1135 if (timeout) break; 1136 if (new_edges > 0) { 1137 // Update escape states on each iteration if graph was updated. 1138 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1139 return false; // Nothing to do. 1140 } 1141 } 1142 time.stop(); 1143 if (time.seconds() >= EscapeAnalysisTimeout) { 1144 timeout = true; 1145 break; 1146 } 1147 } 1148 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1149 time.start(); 1150 // Find fields which have unknown value. 1151 int fields_length = oop_fields_worklist.length(); 1152 for (int next = 0; next < fields_length; next++) { 1153 FieldNode* field = oop_fields_worklist.at(next); 1154 if (field->edge_count() == 0) { 1155 new_edges += find_field_value(field); 1156 // This code may added new edges to phantom_object. 1157 // Need an other cycle to propagate references to phantom_object. 1158 } 1159 } 1160 time.stop(); 1161 if (time.seconds() >= EscapeAnalysisTimeout) { 1162 timeout = true; 1163 break; 1164 } 1165 } else { 1166 new_edges = 0; // Bailout 1167 } 1168 } while (new_edges > 0); 1169 1170 // Bailout if passed limits. 1171 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1172 Compile* C = _compile; 1173 if (C->log() != NULL) { 1174 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1175 C->log()->text("%s", timeout ? "time" : "iterations"); 1176 C->log()->end_elem(" limit'"); 1177 } 1178 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1179 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length())); 1180 // Possible infinite build_connection_graph loop, 1181 // bailout (no changes to ideal graph were made). 1182 return false; 1183 } 1184 #ifdef ASSERT 1185 if (Verbose && PrintEscapeAnalysis) { 1186 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1187 iterations, nodes_size(), ptnodes_worklist.length()); 1188 } 1189 #endif 1190 1191 #undef CG_BUILD_ITER_LIMIT 1192 1193 // Find fields initialized by NULL for non-escaping Allocations. 1194 int non_escaped_length = non_escaped_worklist.length(); 1195 for (int next = 0; next < non_escaped_length; next++) { 1196 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1197 PointsToNode::EscapeState es = ptn->escape_state(); 1198 assert(es <= PointsToNode::ArgEscape, "sanity"); 1199 if (es == PointsToNode::NoEscape) { 1200 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1201 // Adding references to NULL object does not change escape states 1202 // since it does not escape. Also no fields are added to NULL object. 1203 add_java_object_edges(null_obj, false); 1204 } 1205 } 1206 Node* n = ptn->ideal_node(); 1207 if (n->is_Allocate()) { 1208 // The object allocated by this Allocate node will never be 1209 // seen by an other thread. Mark it so that when it is 1210 // expanded no MemBarStoreStore is added. 1211 InitializeNode* ini = n->as_Allocate()->initialization(); 1212 if (ini != NULL) 1213 ini->set_does_not_escape(); 1214 } 1215 } 1216 return true; // Finished graph construction. 1217 } 1218 1219 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1220 // and check that we still have non-escaping java objects. 1221 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1222 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1223 GrowableArray<PointsToNode*> escape_worklist; 1224 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1225 int ptnodes_length = ptnodes_worklist.length(); 1226 for (int next = 0; next < ptnodes_length; ++next) { 1227 PointsToNode* ptn = ptnodes_worklist.at(next); 1228 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1229 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1230 escape_worklist.push(ptn); 1231 } 1232 } 1233 // Set escape states to referenced nodes (edges list). 1234 while (escape_worklist.length() > 0) { 1235 PointsToNode* ptn = escape_worklist.pop(); 1236 PointsToNode::EscapeState es = ptn->escape_state(); 1237 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1238 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1239 es >= PointsToNode::ArgEscape) { 1240 // GlobalEscape or ArgEscape state of field means it has unknown value. 1241 if (add_edge(ptn, phantom_obj)) { 1242 // New edge was added 1243 add_field_uses_to_worklist(ptn->as_Field()); 1244 } 1245 } 1246 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1247 PointsToNode* e = i.get(); 1248 if (e->is_Arraycopy()) { 1249 assert(ptn->arraycopy_dst(), "sanity"); 1250 // Propagate only fields escape state through arraycopy edge. 1251 if (e->fields_escape_state() < field_es) { 1252 set_fields_escape_state(e, field_es); 1253 escape_worklist.push(e); 1254 } 1255 } else if (es >= field_es) { 1256 // fields_escape_state is also set to 'es' if it is less than 'es'. 1257 if (e->escape_state() < es) { 1258 set_escape_state(e, es); 1259 escape_worklist.push(e); 1260 } 1261 } else { 1262 // Propagate field escape state. 1263 bool es_changed = false; 1264 if (e->fields_escape_state() < field_es) { 1265 set_fields_escape_state(e, field_es); 1266 es_changed = true; 1267 } 1268 if ((e->escape_state() < field_es) && 1269 e->is_Field() && ptn->is_JavaObject() && 1270 e->as_Field()->is_oop()) { 1271 // Change escape state of referenced fileds. 1272 set_escape_state(e, field_es); 1273 es_changed = true;; 1274 } else if (e->escape_state() < es) { 1275 set_escape_state(e, es); 1276 es_changed = true;; 1277 } 1278 if (es_changed) { 1279 escape_worklist.push(e); 1280 } 1281 } 1282 } 1283 } 1284 // Remove escaped objects from non_escaped list. 1285 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1286 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1287 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1288 non_escaped_worklist.delete_at(next); 1289 } 1290 if (ptn->escape_state() == PointsToNode::NoEscape) { 1291 // Find fields in non-escaped allocations which have unknown value. 1292 find_init_values(ptn, phantom_obj, NULL); 1293 } 1294 } 1295 return (non_escaped_worklist.length() > 0); 1296 } 1297 1298 // Add all references to JavaObject node by walking over all uses. 1299 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1300 int new_edges = 0; 1301 if (populate_worklist) { 1302 // Populate _worklist by uses of jobj's uses. 1303 for (UseIterator i(jobj); i.has_next(); i.next()) { 1304 PointsToNode* use = i.get(); 1305 if (use->is_Arraycopy()) 1306 continue; 1307 add_uses_to_worklist(use); 1308 if (use->is_Field() && use->as_Field()->is_oop()) { 1309 // Put on worklist all field's uses (loads) and 1310 // related field nodes (same base and offset). 1311 add_field_uses_to_worklist(use->as_Field()); 1312 } 1313 } 1314 } 1315 for (int l = 0; l < _worklist.length(); l++) { 1316 PointsToNode* use = _worklist.at(l); 1317 if (PointsToNode::is_base_use(use)) { 1318 // Add reference from jobj to field and from field to jobj (field's base). 1319 use = PointsToNode::get_use_node(use)->as_Field(); 1320 if (add_base(use->as_Field(), jobj)) { 1321 new_edges++; 1322 } 1323 continue; 1324 } 1325 assert(!use->is_JavaObject(), "sanity"); 1326 if (use->is_Arraycopy()) { 1327 if (jobj == null_obj) // NULL object does not have field edges 1328 continue; 1329 // Added edge from Arraycopy node to arraycopy's source java object 1330 if (add_edge(use, jobj)) { 1331 jobj->set_arraycopy_src(); 1332 new_edges++; 1333 } 1334 // and stop here. 1335 continue; 1336 } 1337 if (!add_edge(use, jobj)) 1338 continue; // No new edge added, there was such edge already. 1339 new_edges++; 1340 if (use->is_LocalVar()) { 1341 add_uses_to_worklist(use); 1342 if (use->arraycopy_dst()) { 1343 for (EdgeIterator i(use); i.has_next(); i.next()) { 1344 PointsToNode* e = i.get(); 1345 if (e->is_Arraycopy()) { 1346 if (jobj == null_obj) // NULL object does not have field edges 1347 continue; 1348 // Add edge from arraycopy's destination java object to Arraycopy node. 1349 if (add_edge(jobj, e)) { 1350 new_edges++; 1351 jobj->set_arraycopy_dst(); 1352 } 1353 } 1354 } 1355 } 1356 } else { 1357 // Added new edge to stored in field values. 1358 // Put on worklist all field's uses (loads) and 1359 // related field nodes (same base and offset). 1360 add_field_uses_to_worklist(use->as_Field()); 1361 } 1362 } 1363 _worklist.clear(); 1364 _in_worklist.Reset(); 1365 return new_edges; 1366 } 1367 1368 // Put on worklist all related field nodes. 1369 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1370 assert(field->is_oop(), "sanity"); 1371 int offset = field->offset(); 1372 add_uses_to_worklist(field); 1373 // Loop over all bases of this field and push on worklist Field nodes 1374 // with the same offset and base (since they may reference the same field). 1375 for (BaseIterator i(field); i.has_next(); i.next()) { 1376 PointsToNode* base = i.get(); 1377 add_fields_to_worklist(field, base); 1378 // Check if the base was source object of arraycopy and go over arraycopy's 1379 // destination objects since values stored to a field of source object are 1380 // accessable by uses (loads) of fields of destination objects. 1381 if (base->arraycopy_src()) { 1382 for (UseIterator j(base); j.has_next(); j.next()) { 1383 PointsToNode* arycp = j.get(); 1384 if (arycp->is_Arraycopy()) { 1385 for (UseIterator k(arycp); k.has_next(); k.next()) { 1386 PointsToNode* abase = k.get(); 1387 if (abase->arraycopy_dst() && abase != base) { 1388 // Look for the same arracopy reference. 1389 add_fields_to_worklist(field, abase); 1390 } 1391 } 1392 } 1393 } 1394 } 1395 } 1396 } 1397 1398 // Put on worklist all related field nodes. 1399 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1400 int offset = field->offset(); 1401 if (base->is_LocalVar()) { 1402 for (UseIterator j(base); j.has_next(); j.next()) { 1403 PointsToNode* f = j.get(); 1404 if (PointsToNode::is_base_use(f)) { // Field 1405 f = PointsToNode::get_use_node(f); 1406 if (f == field || !f->as_Field()->is_oop()) 1407 continue; 1408 int offs = f->as_Field()->offset(); 1409 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1410 add_to_worklist(f); 1411 } 1412 } 1413 } 1414 } else { 1415 assert(base->is_JavaObject(), "sanity"); 1416 if (// Skip phantom_object since it is only used to indicate that 1417 // this field's content globally escapes. 1418 (base != phantom_obj) && 1419 // NULL object node does not have fields. 1420 (base != null_obj)) { 1421 for (EdgeIterator i(base); i.has_next(); i.next()) { 1422 PointsToNode* f = i.get(); 1423 // Skip arraycopy edge since store to destination object field 1424 // does not update value in source object field. 1425 if (f->is_Arraycopy()) { 1426 assert(base->arraycopy_dst(), "sanity"); 1427 continue; 1428 } 1429 if (f == field || !f->as_Field()->is_oop()) 1430 continue; 1431 int offs = f->as_Field()->offset(); 1432 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1433 add_to_worklist(f); 1434 } 1435 } 1436 } 1437 } 1438 } 1439 1440 // Find fields which have unknown value. 1441 int ConnectionGraph::find_field_value(FieldNode* field) { 1442 // Escaped fields should have init value already. 1443 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1444 int new_edges = 0; 1445 for (BaseIterator i(field); i.has_next(); i.next()) { 1446 PointsToNode* base = i.get(); 1447 if (base->is_JavaObject()) { 1448 // Skip Allocate's fields which will be processed later. 1449 if (base->ideal_node()->is_Allocate()) 1450 return 0; 1451 assert(base == null_obj, "only NULL ptr base expected here"); 1452 } 1453 } 1454 if (add_edge(field, phantom_obj)) { 1455 // New edge was added 1456 new_edges++; 1457 add_field_uses_to_worklist(field); 1458 } 1459 return new_edges; 1460 } 1461 1462 // Find fields initializing values for allocations. 1463 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1464 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1465 int new_edges = 0; 1466 Node* alloc = pta->ideal_node(); 1467 if (init_val == phantom_obj) { 1468 // Do nothing for Allocate nodes since its fields values are "known". 1469 if (alloc->is_Allocate()) 1470 return 0; 1471 assert(alloc->as_CallStaticJava(), "sanity"); 1472 #ifdef ASSERT 1473 if (alloc->as_CallStaticJava()->method() == NULL) { 1474 const char* name = alloc->as_CallStaticJava()->_name; 1475 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1476 } 1477 #endif 1478 // Non-escaped allocation returned from Java or runtime call have 1479 // unknown values in fields. 1480 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1481 PointsToNode* field = i.get(); 1482 if (field->is_Field() && field->as_Field()->is_oop()) { 1483 if (add_edge(field, phantom_obj)) { 1484 // New edge was added 1485 new_edges++; 1486 add_field_uses_to_worklist(field->as_Field()); 1487 } 1488 } 1489 } 1490 return new_edges; 1491 } 1492 assert(init_val == null_obj, "sanity"); 1493 // Do nothing for Call nodes since its fields values are unknown. 1494 if (!alloc->is_Allocate()) 1495 return 0; 1496 1497 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1498 bool visited_bottom_offset = false; 1499 GrowableArray<int> offsets_worklist; 1500 1501 // Check if an oop field's initializing value is recorded and add 1502 // a corresponding NULL if field's value if it is not recorded. 1503 // Connection Graph does not record a default initialization by NULL 1504 // captured by Initialize node. 1505 // 1506 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1507 PointsToNode* field = i.get(); // Field (AddP) 1508 if (!field->is_Field() || !field->as_Field()->is_oop()) 1509 continue; // Not oop field 1510 int offset = field->as_Field()->offset(); 1511 if (offset == Type::OffsetBot) { 1512 if (!visited_bottom_offset) { 1513 // OffsetBot is used to reference array's element, 1514 // always add reference to NULL to all Field nodes since we don't 1515 // known which element is referenced. 1516 if (add_edge(field, null_obj)) { 1517 // New edge was added 1518 new_edges++; 1519 add_field_uses_to_worklist(field->as_Field()); 1520 visited_bottom_offset = true; 1521 } 1522 } 1523 } else { 1524 // Check only oop fields. 1525 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1526 if (adr_type->isa_rawptr()) { 1527 #ifdef ASSERT 1528 // Raw pointers are used for initializing stores so skip it 1529 // since it should be recorded already 1530 Node* base = get_addp_base(field->ideal_node()); 1531 assert(adr_type->isa_rawptr() && base->is_Proj() && 1532 (base->in(0) == alloc),"unexpected pointer type"); 1533 #endif 1534 continue; 1535 } 1536 if (!offsets_worklist.contains(offset)) { 1537 offsets_worklist.append(offset); 1538 Node* value = NULL; 1539 if (ini != NULL) { 1540 // StoreP::memory_type() == T_ADDRESS 1541 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1542 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1543 // Make sure initializing store has the same type as this AddP. 1544 // This AddP may reference non existing field because it is on a 1545 // dead branch of bimorphic call which is not eliminated yet. 1546 if (store != NULL && store->is_Store() && 1547 store->as_Store()->memory_type() == ft) { 1548 value = store->in(MemNode::ValueIn); 1549 #ifdef ASSERT 1550 if (VerifyConnectionGraph) { 1551 // Verify that AddP already points to all objects the value points to. 1552 PointsToNode* val = ptnode_adr(value->_idx); 1553 assert((val != NULL), "should be processed already"); 1554 PointsToNode* missed_obj = NULL; 1555 if (val->is_JavaObject()) { 1556 if (!field->points_to(val->as_JavaObject())) { 1557 missed_obj = val; 1558 } 1559 } else { 1560 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1561 tty->print_cr("----------init store has invalid value -----"); 1562 store->dump(); 1563 val->dump(); 1564 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1565 } 1566 for (EdgeIterator j(val); j.has_next(); j.next()) { 1567 PointsToNode* obj = j.get(); 1568 if (obj->is_JavaObject()) { 1569 if (!field->points_to(obj->as_JavaObject())) { 1570 missed_obj = obj; 1571 break; 1572 } 1573 } 1574 } 1575 } 1576 if (missed_obj != NULL) { 1577 tty->print_cr("----------field---------------------------------"); 1578 field->dump(); 1579 tty->print_cr("----------missed referernce to object-----------"); 1580 missed_obj->dump(); 1581 tty->print_cr("----------object referernced by init store -----"); 1582 store->dump(); 1583 val->dump(); 1584 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1585 } 1586 } 1587 #endif 1588 } else { 1589 // There could be initializing stores which follow allocation. 1590 // For example, a volatile field store is not collected 1591 // by Initialize node. 1592 // 1593 // Need to check for dependent loads to separate such stores from 1594 // stores which follow loads. For now, add initial value NULL so 1595 // that compare pointers optimization works correctly. 1596 } 1597 } 1598 if (value == NULL) { 1599 // A field's initializing value was not recorded. Add NULL. 1600 if (add_edge(field, null_obj)) { 1601 // New edge was added 1602 new_edges++; 1603 add_field_uses_to_worklist(field->as_Field()); 1604 } 1605 } 1606 } 1607 } 1608 } 1609 return new_edges; 1610 } 1611 1612 // Adjust scalar_replaceable state after Connection Graph is built. 1613 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1614 // Search for non-escaping objects which are not scalar replaceable 1615 // and mark them to propagate the state to referenced objects. 1616 1617 // 1. An object is not scalar replaceable if the field into which it is 1618 // stored has unknown offset (stored into unknown element of an array). 1619 // 1620 for (UseIterator i(jobj); i.has_next(); i.next()) { 1621 PointsToNode* use = i.get(); 1622 assert(!use->is_Arraycopy(), "sanity"); 1623 if (use->is_Field()) { 1624 FieldNode* field = use->as_Field(); 1625 assert(field->is_oop() && field->scalar_replaceable() && 1626 field->fields_escape_state() == PointsToNode::NoEscape, "sanity"); 1627 if (field->offset() == Type::OffsetBot) { 1628 jobj->set_scalar_replaceable(false); 1629 return; 1630 } 1631 // 2. An object is not scalar replaceable if the field into which it is 1632 // stored has multiple bases one of which is null. 1633 if (field->base_count() > 1) { 1634 for (BaseIterator i(field); i.has_next(); i.next()) { 1635 PointsToNode* base = i.get(); 1636 if (base == null_obj) { 1637 jobj->set_scalar_replaceable(false); 1638 return; 1639 } 1640 } 1641 } 1642 } 1643 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1644 // 3. An object is not scalar replaceable if it is merged with other objects. 1645 for (EdgeIterator j(use); j.has_next(); j.next()) { 1646 PointsToNode* ptn = j.get(); 1647 if (ptn->is_JavaObject() && ptn != jobj) { 1648 // Mark all objects. 1649 jobj->set_scalar_replaceable(false); 1650 ptn->set_scalar_replaceable(false); 1651 } 1652 } 1653 if (!jobj->scalar_replaceable()) { 1654 return; 1655 } 1656 } 1657 1658 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1659 // Non-escaping object node should point only to field nodes. 1660 FieldNode* field = j.get()->as_Field(); 1661 int offset = field->as_Field()->offset(); 1662 1663 // 4. An object is not scalar replaceable if it has a field with unknown 1664 // offset (array's element is accessed in loop). 1665 if (offset == Type::OffsetBot) { 1666 jobj->set_scalar_replaceable(false); 1667 return; 1668 } 1669 // 5. Currently an object is not scalar replaceable if a LoadStore node 1670 // access its field since the field value is unknown after it. 1671 // 1672 Node* n = field->ideal_node(); 1673 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1674 if (n->fast_out(i)->is_LoadStore()) { 1675 jobj->set_scalar_replaceable(false); 1676 return; 1677 } 1678 } 1679 1680 // 6. Or the address may point to more then one object. This may produce 1681 // the false positive result (set not scalar replaceable) 1682 // since the flow-insensitive escape analysis can't separate 1683 // the case when stores overwrite the field's value from the case 1684 // when stores happened on different control branches. 1685 // 1686 // Note: it will disable scalar replacement in some cases: 1687 // 1688 // Point p[] = new Point[1]; 1689 // p[0] = new Point(); // Will be not scalar replaced 1690 // 1691 // but it will save us from incorrect optimizations in next cases: 1692 // 1693 // Point p[] = new Point[1]; 1694 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1695 // 1696 if (field->base_count() > 1) { 1697 for (BaseIterator i(field); i.has_next(); i.next()) { 1698 PointsToNode* base = i.get(); 1699 // Don't take into account LocalVar nodes which 1700 // may point to only one object which should be also 1701 // this field's base by now. 1702 if (base->is_JavaObject() && base != jobj) { 1703 // Mark all bases. 1704 jobj->set_scalar_replaceable(false); 1705 base->set_scalar_replaceable(false); 1706 } 1707 } 1708 } 1709 } 1710 } 1711 1712 #ifdef ASSERT 1713 void ConnectionGraph::verify_connection_graph( 1714 GrowableArray<PointsToNode*>& ptnodes_worklist, 1715 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1716 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1717 GrowableArray<Node*>& addp_worklist) { 1718 // Verify that graph is complete - no new edges could be added. 1719 int java_objects_length = java_objects_worklist.length(); 1720 int non_escaped_length = non_escaped_worklist.length(); 1721 int new_edges = 0; 1722 for (int next = 0; next < java_objects_length; ++next) { 1723 JavaObjectNode* ptn = java_objects_worklist.at(next); 1724 new_edges += add_java_object_edges(ptn, true); 1725 } 1726 assert(new_edges == 0, "graph was not complete"); 1727 // Verify that escape state is final. 1728 int length = non_escaped_worklist.length(); 1729 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1730 assert((non_escaped_length == non_escaped_worklist.length()) && 1731 (non_escaped_length == length) && 1732 (_worklist.length() == 0), "escape state was not final"); 1733 1734 // Verify fields information. 1735 int addp_length = addp_worklist.length(); 1736 for (int next = 0; next < addp_length; ++next ) { 1737 Node* n = addp_worklist.at(next); 1738 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1739 if (field->is_oop()) { 1740 // Verify that field has all bases 1741 Node* base = get_addp_base(n); 1742 PointsToNode* ptn = ptnode_adr(base->_idx); 1743 if (ptn->is_JavaObject()) { 1744 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1745 } else { 1746 assert(ptn->is_LocalVar(), "sanity"); 1747 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1748 PointsToNode* e = i.get(); 1749 if (e->is_JavaObject()) { 1750 assert(field->has_base(e->as_JavaObject()), "sanity"); 1751 } 1752 } 1753 } 1754 // Verify that all fields have initializing values. 1755 if (field->edge_count() == 0) { 1756 tty->print_cr("----------field does not have references----------"); 1757 field->dump(); 1758 for (BaseIterator i(field); i.has_next(); i.next()) { 1759 PointsToNode* base = i.get(); 1760 tty->print_cr("----------field has next base---------------------"); 1761 base->dump(); 1762 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1763 tty->print_cr("----------base has fields-------------------------"); 1764 for (EdgeIterator j(base); j.has_next(); j.next()) { 1765 j.get()->dump(); 1766 } 1767 tty->print_cr("----------base has references---------------------"); 1768 for (UseIterator j(base); j.has_next(); j.next()) { 1769 j.get()->dump(); 1770 } 1771 } 1772 } 1773 for (UseIterator i(field); i.has_next(); i.next()) { 1774 i.get()->dump(); 1775 } 1776 assert(field->edge_count() > 0, "sanity"); 1777 } 1778 } 1779 } 1780 } 1781 #endif 1782 1783 // Optimize ideal graph. 1784 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1785 GrowableArray<Node*>& storestore_worklist) { 1786 Compile* C = _compile; 1787 PhaseIterGVN* igvn = _igvn; 1788 if (EliminateLocks) { 1789 // Mark locks before changing ideal graph. 1790 int cnt = C->macro_count(); 1791 for( int i=0; i < cnt; i++ ) { 1792 Node *n = C->macro_node(i); 1793 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1794 AbstractLockNode* alock = n->as_AbstractLock(); 1795 if (!alock->is_non_esc_obj()) { 1796 if (not_global_escape(alock->obj_node())) { 1797 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1798 // The lock could be marked eliminated by lock coarsening 1799 // code during first IGVN before EA. Replace coarsened flag 1800 // to eliminate all associated locks/unlocks. 1801 alock->set_non_esc_obj(); 1802 } 1803 } 1804 } 1805 } 1806 } 1807 1808 if (OptimizePtrCompare) { 1809 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1810 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1811 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1812 // Optimize objects compare. 1813 while (ptr_cmp_worklist.length() != 0) { 1814 Node *n = ptr_cmp_worklist.pop(); 1815 Node *res = optimize_ptr_compare(n); 1816 if (res != NULL) { 1817 #ifndef PRODUCT 1818 if (PrintOptimizePtrCompare) { 1819 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")); 1820 if (Verbose) { 1821 n->dump(1); 1822 } 1823 } 1824 #endif 1825 igvn->replace_node(n, res); 1826 } 1827 } 1828 // cleanup 1829 if (_pcmp_neq->outcnt() == 0) 1830 igvn->hash_delete(_pcmp_neq); 1831 if (_pcmp_eq->outcnt() == 0) 1832 igvn->hash_delete(_pcmp_eq); 1833 } 1834 1835 // For MemBarStoreStore nodes added in library_call.cpp, check 1836 // escape status of associated AllocateNode and optimize out 1837 // MemBarStoreStore node if the allocated object never escapes. 1838 while (storestore_worklist.length() != 0) { 1839 Node *n = storestore_worklist.pop(); 1840 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1841 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1842 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1843 if (not_global_escape(alloc)) { 1844 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1845 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1846 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1847 igvn->register_new_node_with_optimizer(mb); 1848 igvn->replace_node(storestore, mb); 1849 } 1850 } 1851 } 1852 1853 // Optimize objects compare. 1854 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1855 assert(OptimizePtrCompare, "sanity"); 1856 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1857 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1858 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1859 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1860 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1861 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1862 1863 // Check simple cases first. 1864 if (jobj1 != NULL) { 1865 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1866 if (jobj1 == jobj2) { 1867 // Comparing the same not escaping object. 1868 return _pcmp_eq; 1869 } 1870 Node* obj = jobj1->ideal_node(); 1871 // Comparing not escaping allocation. 1872 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1873 !ptn2->points_to(jobj1)) { 1874 return _pcmp_neq; // This includes nullness check. 1875 } 1876 } 1877 } 1878 if (jobj2 != NULL) { 1879 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1880 Node* obj = jobj2->ideal_node(); 1881 // Comparing not escaping allocation. 1882 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1883 !ptn1->points_to(jobj2)) { 1884 return _pcmp_neq; // This includes nullness check. 1885 } 1886 } 1887 } 1888 if (jobj1 != NULL && jobj1 != phantom_obj && 1889 jobj2 != NULL && jobj2 != phantom_obj && 1890 jobj1->ideal_node()->is_Con() && 1891 jobj2->ideal_node()->is_Con()) { 1892 // Klass or String constants compare. Need to be careful with 1893 // compressed pointers - compare types of ConN and ConP instead of nodes. 1894 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1895 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1896 if (t1->make_ptr() == t2->make_ptr()) { 1897 return _pcmp_eq; 1898 } else { 1899 return _pcmp_neq; 1900 } 1901 } 1902 if (ptn1->meet(ptn2)) { 1903 return NULL; // Sets are not disjoint 1904 } 1905 1906 // Sets are disjoint. 1907 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1908 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1909 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1910 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1911 if (set1_has_unknown_ptr && set2_has_null_ptr || 1912 set2_has_unknown_ptr && set1_has_null_ptr) { 1913 // Check nullness of unknown object. 1914 return NULL; 1915 } 1916 1917 // Disjointness by itself is not sufficient since 1918 // alias analysis is not complete for escaped objects. 1919 // Disjoint sets are definitely unrelated only when 1920 // at least one set has only not escaping allocations. 1921 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1922 if (ptn1->non_escaping_allocation()) { 1923 return _pcmp_neq; 1924 } 1925 } 1926 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 1927 if (ptn2->non_escaping_allocation()) { 1928 return _pcmp_neq; 1929 } 1930 } 1931 return NULL; 1932 } 1933 1934 // Connection Graph constuction functions. 1935 1936 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 1937 PointsToNode* ptadr = _nodes.at(n->_idx); 1938 if (ptadr != NULL) { 1939 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 1940 return; 1941 } 1942 Compile* C = _compile; 1943 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 1944 _nodes.at_put(n->_idx, ptadr); 1945 } 1946 1947 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 1948 PointsToNode* ptadr = _nodes.at(n->_idx); 1949 if (ptadr != NULL) { 1950 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 1951 return; 1952 } 1953 Compile* C = _compile; 1954 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 1955 _nodes.at_put(n->_idx, ptadr); 1956 } 1957 1958 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 1959 PointsToNode* ptadr = _nodes.at(n->_idx); 1960 if (ptadr != NULL) { 1961 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 1962 return; 1963 } 1964 bool unsafe = false; 1965 bool is_oop = is_oop_field(n, offset, &unsafe); 1966 if (unsafe) { 1967 es = PointsToNode::GlobalEscape; 1968 } 1969 Compile* C = _compile; 1970 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 1971 _nodes.at_put(n->_idx, field); 1972 } 1973 1974 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 1975 PointsToNode* src, PointsToNode* dst) { 1976 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 1977 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 1978 PointsToNode* ptadr = _nodes.at(n->_idx); 1979 if (ptadr != NULL) { 1980 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 1981 return; 1982 } 1983 Compile* C = _compile; 1984 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 1985 _nodes.at_put(n->_idx, ptadr); 1986 // Add edge from arraycopy node to source object. 1987 (void)add_edge(ptadr, src); 1988 src->set_arraycopy_src(); 1989 // Add edge from destination object to arraycopy node. 1990 (void)add_edge(dst, ptadr); 1991 dst->set_arraycopy_dst(); 1992 } 1993 1994 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 1995 const Type* adr_type = n->as_AddP()->bottom_type(); 1996 BasicType bt = T_INT; 1997 if (offset == Type::OffsetBot) { 1998 // Check only oop fields. 1999 if (!adr_type->isa_aryptr() || 2000 (adr_type->isa_aryptr()->klass() == NULL) || 2001 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2002 // OffsetBot is used to reference array's element. Ignore first AddP. 2003 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2004 bt = T_OBJECT; 2005 } 2006 } 2007 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2008 if (adr_type->isa_instptr()) { 2009 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2010 if (field != NULL) { 2011 bt = field->layout_type(); 2012 } else { 2013 // Check for unsafe oop field access 2014 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN)) { 2015 bt = T_OBJECT; 2016 (*unsafe) = true; 2017 } 2018 } 2019 } else if (adr_type->isa_aryptr()) { 2020 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2021 // Ignore array length load. 2022 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2023 // Ignore first AddP. 2024 } else { 2025 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2026 bt = elemtype->array_element_basic_type(); 2027 } 2028 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2029 // Allocation initialization, ThreadLocal field access, unsafe access 2030 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN)) { 2031 bt = T_OBJECT; 2032 } 2033 } 2034 } 2035 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 2036 } 2037 2038 // Returns unique pointed java object or NULL. 2039 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2040 assert(!_collecting, "should not call when contructed graph"); 2041 // If the node was created after the escape computation we can't answer. 2042 uint idx = n->_idx; 2043 if (idx >= nodes_size()) { 2044 return NULL; 2045 } 2046 PointsToNode* ptn = ptnode_adr(idx); 2047 if (ptn->is_JavaObject()) { 2048 return ptn->as_JavaObject(); 2049 } 2050 assert(ptn->is_LocalVar(), "sanity"); 2051 // Check all java objects it points to. 2052 JavaObjectNode* jobj = NULL; 2053 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2054 PointsToNode* e = i.get(); 2055 if (e->is_JavaObject()) { 2056 if (jobj == NULL) { 2057 jobj = e->as_JavaObject(); 2058 } else if (jobj != e) { 2059 return NULL; 2060 } 2061 } 2062 } 2063 return jobj; 2064 } 2065 2066 // Return true if this node points only to non-escaping allocations. 2067 bool PointsToNode::non_escaping_allocation() { 2068 if (is_JavaObject()) { 2069 Node* n = ideal_node(); 2070 if (n->is_Allocate() || n->is_CallStaticJava()) { 2071 return (escape_state() == PointsToNode::NoEscape); 2072 } else { 2073 return false; 2074 } 2075 } 2076 assert(is_LocalVar(), "sanity"); 2077 // Check all java objects it points to. 2078 for (EdgeIterator i(this); i.has_next(); i.next()) { 2079 PointsToNode* e = i.get(); 2080 if (e->is_JavaObject()) { 2081 Node* n = e->ideal_node(); 2082 if ((e->escape_state() != PointsToNode::NoEscape) || 2083 !(n->is_Allocate() || n->is_CallStaticJava())) { 2084 return false; 2085 } 2086 } 2087 } 2088 return true; 2089 } 2090 2091 // Return true if we know the node does not escape globally. 2092 bool ConnectionGraph::not_global_escape(Node *n) { 2093 assert(!_collecting, "should not call during graph construction"); 2094 // If the node was created after the escape computation we can't answer. 2095 uint idx = n->_idx; 2096 if (idx >= nodes_size()) { 2097 return false; 2098 } 2099 PointsToNode* ptn = ptnode_adr(idx); 2100 PointsToNode::EscapeState es = ptn->escape_state(); 2101 // If we have already computed a value, return it. 2102 if (es >= PointsToNode::GlobalEscape) 2103 return false; 2104 if (ptn->is_JavaObject()) { 2105 return true; // (es < PointsToNode::GlobalEscape); 2106 } 2107 assert(ptn->is_LocalVar(), "sanity"); 2108 // Check all java objects it points to. 2109 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2110 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2111 return false; 2112 } 2113 return true; 2114 } 2115 2116 2117 // Helper functions 2118 2119 // Return true if this node points to specified node or nodes it points to. 2120 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2121 if (is_JavaObject()) { 2122 return (this == ptn); 2123 } 2124 assert(is_LocalVar() || is_Field(), "sanity"); 2125 for (EdgeIterator i(this); i.has_next(); i.next()) { 2126 if (i.get() == ptn) 2127 return true; 2128 } 2129 return false; 2130 } 2131 2132 // Return true if one node points to an other. 2133 bool PointsToNode::meet(PointsToNode* ptn) { 2134 if (this == ptn) { 2135 return true; 2136 } else if (ptn->is_JavaObject()) { 2137 return this->points_to(ptn->as_JavaObject()); 2138 } else if (this->is_JavaObject()) { 2139 return ptn->points_to(this->as_JavaObject()); 2140 } 2141 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2142 int ptn_count = ptn->edge_count(); 2143 for (EdgeIterator i(this); i.has_next(); i.next()) { 2144 PointsToNode* this_e = i.get(); 2145 for (int j = 0; j < ptn_count; j++) { 2146 if (this_e == ptn->edge(j)) 2147 return true; 2148 } 2149 } 2150 return false; 2151 } 2152 2153 #ifdef ASSERT 2154 // Return true if bases point to this java object. 2155 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2156 for (BaseIterator i(this); i.has_next(); i.next()) { 2157 if (i.get() == jobj) 2158 return true; 2159 } 2160 return false; 2161 } 2162 #endif 2163 2164 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2165 const Type *adr_type = phase->type(adr); 2166 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2167 adr->in(AddPNode::Address)->is_Proj() && 2168 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2169 // We are computing a raw address for a store captured by an Initialize 2170 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2171 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2172 assert(offs != Type::OffsetBot || 2173 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2174 "offset must be a constant or it is initialization of array"); 2175 return offs; 2176 } 2177 const TypePtr *t_ptr = adr_type->isa_ptr(); 2178 assert(t_ptr != NULL, "must be a pointer type"); 2179 return t_ptr->offset(); 2180 } 2181 2182 Node* ConnectionGraph::get_addp_base(Node *addp) { 2183 assert(addp->is_AddP(), "must be AddP"); 2184 // 2185 // AddP cases for Base and Address inputs: 2186 // case #1. Direct object's field reference: 2187 // Allocate 2188 // | 2189 // Proj #5 ( oop result ) 2190 // | 2191 // CheckCastPP (cast to instance type) 2192 // | | 2193 // AddP ( base == address ) 2194 // 2195 // case #2. Indirect object's field reference: 2196 // Phi 2197 // | 2198 // CastPP (cast to instance type) 2199 // | | 2200 // AddP ( base == address ) 2201 // 2202 // case #3. Raw object's field reference for Initialize node: 2203 // Allocate 2204 // | 2205 // Proj #5 ( oop result ) 2206 // top | 2207 // \ | 2208 // AddP ( base == top ) 2209 // 2210 // case #4. Array's element reference: 2211 // {CheckCastPP | CastPP} 2212 // | | | 2213 // | AddP ( array's element offset ) 2214 // | | 2215 // AddP ( array's offset ) 2216 // 2217 // case #5. Raw object's field reference for arraycopy stub call: 2218 // The inline_native_clone() case when the arraycopy stub is called 2219 // after the allocation before Initialize and CheckCastPP nodes. 2220 // Allocate 2221 // | 2222 // Proj #5 ( oop result ) 2223 // | | 2224 // AddP ( base == address ) 2225 // 2226 // case #6. Constant Pool, ThreadLocal, CastX2P or 2227 // Raw object's field reference: 2228 // {ConP, ThreadLocal, CastX2P, raw Load} 2229 // top | 2230 // \ | 2231 // AddP ( base == top ) 2232 // 2233 // case #7. Klass's field reference. 2234 // LoadKlass 2235 // | | 2236 // AddP ( base == address ) 2237 // 2238 // case #8. narrow Klass's field reference. 2239 // LoadNKlass 2240 // | 2241 // DecodeN 2242 // | | 2243 // AddP ( base == address ) 2244 // 2245 Node *base = addp->in(AddPNode::Base); 2246 if (base->uncast()->is_top()) { // The AddP case #3 and #6. 2247 base = addp->in(AddPNode::Address); 2248 while (base->is_AddP()) { 2249 // Case #6 (unsafe access) may have several chained AddP nodes. 2250 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2251 base = base->in(AddPNode::Address); 2252 } 2253 Node* uncast_base = base->uncast(); 2254 int opcode = uncast_base->Opcode(); 2255 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2256 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2257 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2258 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity"); 2259 } 2260 return base; 2261 } 2262 2263 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2264 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2265 Node* addp2 = addp->raw_out(0); 2266 if (addp->outcnt() == 1 && addp2->is_AddP() && 2267 addp2->in(AddPNode::Base) == n && 2268 addp2->in(AddPNode::Address) == addp) { 2269 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2270 // 2271 // Find array's offset to push it on worklist first and 2272 // as result process an array's element offset first (pushed second) 2273 // to avoid CastPP for the array's offset. 2274 // Otherwise the inserted CastPP (LocalVar) will point to what 2275 // the AddP (Field) points to. Which would be wrong since 2276 // the algorithm expects the CastPP has the same point as 2277 // as AddP's base CheckCastPP (LocalVar). 2278 // 2279 // ArrayAllocation 2280 // | 2281 // CheckCastPP 2282 // | 2283 // memProj (from ArrayAllocation CheckCastPP) 2284 // | || 2285 // | || Int (element index) 2286 // | || | ConI (log(element size)) 2287 // | || | / 2288 // | || LShift 2289 // | || / 2290 // | AddP (array's element offset) 2291 // | | 2292 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2293 // | / / 2294 // AddP (array's offset) 2295 // | 2296 // Load/Store (memory operation on array's element) 2297 // 2298 return addp2; 2299 } 2300 return NULL; 2301 } 2302 2303 // 2304 // Adjust the type and inputs of an AddP which computes the 2305 // address of a field of an instance 2306 // 2307 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2308 PhaseGVN* igvn = _igvn; 2309 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2310 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2311 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2312 if (t == NULL) { 2313 // We are computing a raw address for a store captured by an Initialize 2314 // compute an appropriate address type (cases #3 and #5). 2315 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2316 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2317 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2318 assert(offs != Type::OffsetBot, "offset must be a constant"); 2319 t = base_t->add_offset(offs)->is_oopptr(); 2320 } 2321 int inst_id = base_t->instance_id(); 2322 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2323 "old type must be non-instance or match new type"); 2324 2325 // The type 't' could be subclass of 'base_t'. 2326 // As result t->offset() could be large then base_t's size and it will 2327 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2328 // constructor verifies correctness of the offset. 2329 // 2330 // It could happened on subclass's branch (from the type profiling 2331 // inlining) which was not eliminated during parsing since the exactness 2332 // of the allocation type was not propagated to the subclass type check. 2333 // 2334 // Or the type 't' could be not related to 'base_t' at all. 2335 // It could happened when CHA type is different from MDO type on a dead path 2336 // (for example, from instanceof check) which is not collapsed during parsing. 2337 // 2338 // Do nothing for such AddP node and don't process its users since 2339 // this code branch will go away. 2340 // 2341 if (!t->is_known_instance() && 2342 !base_t->klass()->is_subtype_of(t->klass())) { 2343 return false; // bail out 2344 } 2345 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2346 // Do NOT remove the next line: ensure a new alias index is allocated 2347 // for the instance type. Note: C++ will not remove it since the call 2348 // has side effect. 2349 int alias_idx = _compile->get_alias_index(tinst); 2350 igvn->set_type(addp, tinst); 2351 // record the allocation in the node map 2352 set_map(addp, get_map(base->_idx)); 2353 // Set addp's Base and Address to 'base'. 2354 Node *abase = addp->in(AddPNode::Base); 2355 Node *adr = addp->in(AddPNode::Address); 2356 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2357 adr->in(0)->_idx == (uint)inst_id) { 2358 // Skip AddP cases #3 and #5. 2359 } else { 2360 assert(!abase->is_top(), "sanity"); // AddP case #3 2361 if (abase != base) { 2362 igvn->hash_delete(addp); 2363 addp->set_req(AddPNode::Base, base); 2364 if (abase == adr) { 2365 addp->set_req(AddPNode::Address, base); 2366 } else { 2367 // AddP case #4 (adr is array's element offset AddP node) 2368 #ifdef ASSERT 2369 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2370 assert(adr->is_AddP() && atype != NULL && 2371 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2372 #endif 2373 } 2374 igvn->hash_insert(addp); 2375 } 2376 } 2377 // Put on IGVN worklist since at least addp's type was changed above. 2378 record_for_optimizer(addp); 2379 return true; 2380 } 2381 2382 // 2383 // Create a new version of orig_phi if necessary. Returns either the newly 2384 // created phi or an existing phi. Sets create_new to indicate whether a new 2385 // phi was created. Cache the last newly created phi in the node map. 2386 // 2387 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2388 Compile *C = _compile; 2389 PhaseGVN* igvn = _igvn; 2390 new_created = false; 2391 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2392 // nothing to do if orig_phi is bottom memory or matches alias_idx 2393 if (phi_alias_idx == alias_idx) { 2394 return orig_phi; 2395 } 2396 // Have we recently created a Phi for this alias index? 2397 PhiNode *result = get_map_phi(orig_phi->_idx); 2398 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2399 return result; 2400 } 2401 // Previous check may fail when the same wide memory Phi was split into Phis 2402 // for different memory slices. Search all Phis for this region. 2403 if (result != NULL) { 2404 Node* region = orig_phi->in(0); 2405 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2406 Node* phi = region->fast_out(i); 2407 if (phi->is_Phi() && 2408 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2409 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2410 return phi->as_Phi(); 2411 } 2412 } 2413 } 2414 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2415 if (C->do_escape_analysis() == true && !C->failing()) { 2416 // Retry compilation without escape analysis. 2417 // If this is the first failure, the sentinel string will "stick" 2418 // to the Compile object, and the C2Compiler will see it and retry. 2419 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2420 } 2421 return NULL; 2422 } 2423 orig_phi_worklist.append_if_missing(orig_phi); 2424 const TypePtr *atype = C->get_adr_type(alias_idx); 2425 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2426 C->copy_node_notes_to(result, orig_phi); 2427 igvn->set_type(result, result->bottom_type()); 2428 record_for_optimizer(result); 2429 set_map(orig_phi, result); 2430 new_created = true; 2431 return result; 2432 } 2433 2434 // 2435 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2436 // specified alias index. 2437 // 2438 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2439 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2440 Compile *C = _compile; 2441 PhaseGVN* igvn = _igvn; 2442 bool new_phi_created; 2443 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2444 if (!new_phi_created) { 2445 return result; 2446 } 2447 GrowableArray<PhiNode *> phi_list; 2448 GrowableArray<uint> cur_input; 2449 PhiNode *phi = orig_phi; 2450 uint idx = 1; 2451 bool finished = false; 2452 while(!finished) { 2453 while (idx < phi->req()) { 2454 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2455 if (mem != NULL && mem->is_Phi()) { 2456 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2457 if (new_phi_created) { 2458 // found an phi for which we created a new split, push current one on worklist and begin 2459 // processing new one 2460 phi_list.push(phi); 2461 cur_input.push(idx); 2462 phi = mem->as_Phi(); 2463 result = newphi; 2464 idx = 1; 2465 continue; 2466 } else { 2467 mem = newphi; 2468 } 2469 } 2470 if (C->failing()) { 2471 return NULL; 2472 } 2473 result->set_req(idx++, mem); 2474 } 2475 #ifdef ASSERT 2476 // verify that the new Phi has an input for each input of the original 2477 assert( phi->req() == result->req(), "must have same number of inputs."); 2478 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2479 #endif 2480 // Check if all new phi's inputs have specified alias index. 2481 // Otherwise use old phi. 2482 for (uint i = 1; i < phi->req(); i++) { 2483 Node* in = result->in(i); 2484 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2485 } 2486 // we have finished processing a Phi, see if there are any more to do 2487 finished = (phi_list.length() == 0 ); 2488 if (!finished) { 2489 phi = phi_list.pop(); 2490 idx = cur_input.pop(); 2491 PhiNode *prev_result = get_map_phi(phi->_idx); 2492 prev_result->set_req(idx++, result); 2493 result = prev_result; 2494 } 2495 } 2496 return result; 2497 } 2498 2499 // 2500 // The next methods are derived from methods in MemNode. 2501 // 2502 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2503 Node *mem = mmem; 2504 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2505 // means an array I have not precisely typed yet. Do not do any 2506 // alias stuff with it any time soon. 2507 if (toop->base() != Type::AnyPtr && 2508 !(toop->klass() != NULL && 2509 toop->klass()->is_java_lang_Object() && 2510 toop->offset() == Type::OffsetBot)) { 2511 mem = mmem->memory_at(alias_idx); 2512 // Update input if it is progress over what we have now 2513 } 2514 return mem; 2515 } 2516 2517 // 2518 // Move memory users to their memory slices. 2519 // 2520 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2521 Compile* C = _compile; 2522 PhaseGVN* igvn = _igvn; 2523 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2524 assert(tp != NULL, "ptr type"); 2525 int alias_idx = C->get_alias_index(tp); 2526 int general_idx = C->get_general_index(alias_idx); 2527 2528 // Move users first 2529 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2530 Node* use = n->fast_out(i); 2531 if (use->is_MergeMem()) { 2532 MergeMemNode* mmem = use->as_MergeMem(); 2533 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2534 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2535 continue; // Nothing to do 2536 } 2537 // Replace previous general reference to mem node. 2538 uint orig_uniq = C->unique(); 2539 Node* m = find_inst_mem(n, general_idx, orig_phis); 2540 assert(orig_uniq == C->unique(), "no new nodes"); 2541 mmem->set_memory_at(general_idx, m); 2542 --imax; 2543 --i; 2544 } else if (use->is_MemBar()) { 2545 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2546 if (use->req() > MemBarNode::Precedent && 2547 use->in(MemBarNode::Precedent) == n) { 2548 // Don't move related membars. 2549 record_for_optimizer(use); 2550 continue; 2551 } 2552 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2553 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 2554 alias_idx == general_idx) { 2555 continue; // Nothing to do 2556 } 2557 // Move to general memory slice. 2558 uint orig_uniq = C->unique(); 2559 Node* m = find_inst_mem(n, general_idx, orig_phis); 2560 assert(orig_uniq == C->unique(), "no new nodes"); 2561 igvn->hash_delete(use); 2562 imax -= use->replace_edge(n, m); 2563 igvn->hash_insert(use); 2564 record_for_optimizer(use); 2565 --i; 2566 #ifdef ASSERT 2567 } else if (use->is_Mem()) { 2568 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2569 // Don't move related cardmark. 2570 continue; 2571 } 2572 // Memory nodes should have new memory input. 2573 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2574 assert(tp != NULL, "ptr type"); 2575 int idx = C->get_alias_index(tp); 2576 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2577 "Following memory nodes should have new memory input or be on the same memory slice"); 2578 } else if (use->is_Phi()) { 2579 // Phi nodes should be split and moved already. 2580 tp = use->as_Phi()->adr_type()->isa_ptr(); 2581 assert(tp != NULL, "ptr type"); 2582 int idx = C->get_alias_index(tp); 2583 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2584 } else { 2585 use->dump(); 2586 assert(false, "should not be here"); 2587 #endif 2588 } 2589 } 2590 } 2591 2592 // 2593 // Search memory chain of "mem" to find a MemNode whose address 2594 // is the specified alias index. 2595 // 2596 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2597 if (orig_mem == NULL) 2598 return orig_mem; 2599 Compile* C = _compile; 2600 PhaseGVN* igvn = _igvn; 2601 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2602 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2603 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 2604 Node *prev = NULL; 2605 Node *result = orig_mem; 2606 while (prev != result) { 2607 prev = result; 2608 if (result == start_mem) 2609 break; // hit one of our sentinels 2610 if (result->is_Mem()) { 2611 const Type *at = igvn->type(result->in(MemNode::Address)); 2612 if (at == Type::TOP) 2613 break; // Dead 2614 assert (at->isa_ptr() != NULL, "pointer type required."); 2615 int idx = C->get_alias_index(at->is_ptr()); 2616 if (idx == alias_idx) 2617 break; // Found 2618 if (!is_instance && (at->isa_oopptr() == NULL || 2619 !at->is_oopptr()->is_known_instance())) { 2620 break; // Do not skip store to general memory slice. 2621 } 2622 result = result->in(MemNode::Memory); 2623 } 2624 if (!is_instance) 2625 continue; // don't search further for non-instance types 2626 // skip over a call which does not affect this memory slice 2627 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2628 Node *proj_in = result->in(0); 2629 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2630 break; // hit one of our sentinels 2631 } else if (proj_in->is_Call()) { 2632 CallNode *call = proj_in->as_Call(); 2633 if (!call->may_modify(toop, igvn)) { 2634 result = call->in(TypeFunc::Memory); 2635 } 2636 } else if (proj_in->is_Initialize()) { 2637 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2638 // Stop if this is the initialization for the object instance which 2639 // which contains this memory slice, otherwise skip over it. 2640 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2641 result = proj_in->in(TypeFunc::Memory); 2642 } 2643 } else if (proj_in->is_MemBar()) { 2644 result = proj_in->in(TypeFunc::Memory); 2645 } 2646 } else if (result->is_MergeMem()) { 2647 MergeMemNode *mmem = result->as_MergeMem(); 2648 result = step_through_mergemem(mmem, alias_idx, toop); 2649 if (result == mmem->base_memory()) { 2650 // Didn't find instance memory, search through general slice recursively. 2651 result = mmem->memory_at(C->get_general_index(alias_idx)); 2652 result = find_inst_mem(result, alias_idx, orig_phis); 2653 if (C->failing()) { 2654 return NULL; 2655 } 2656 mmem->set_memory_at(alias_idx, result); 2657 } 2658 } else if (result->is_Phi() && 2659 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2660 Node *un = result->as_Phi()->unique_input(igvn); 2661 if (un != NULL) { 2662 orig_phis.append_if_missing(result->as_Phi()); 2663 result = un; 2664 } else { 2665 break; 2666 } 2667 } else if (result->is_ClearArray()) { 2668 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2669 // Can not bypass initialization of the instance 2670 // we are looking for. 2671 break; 2672 } 2673 // Otherwise skip it (the call updated 'result' value). 2674 } else if (result->Opcode() == Op_SCMemProj) { 2675 Node* mem = result->in(0); 2676 Node* adr = NULL; 2677 if (mem->is_LoadStore()) { 2678 adr = mem->in(MemNode::Address); 2679 } else { 2680 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 2681 adr = mem->in(3); // Memory edge corresponds to destination array 2682 } 2683 const Type *at = igvn->type(adr); 2684 if (at != Type::TOP) { 2685 assert (at->isa_ptr() != NULL, "pointer type required."); 2686 int idx = C->get_alias_index(at->is_ptr()); 2687 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 2688 break; 2689 } 2690 result = mem->in(MemNode::Memory); 2691 } 2692 } 2693 if (result->is_Phi()) { 2694 PhiNode *mphi = result->as_Phi(); 2695 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2696 const TypePtr *t = mphi->adr_type(); 2697 if (!is_instance) { 2698 // Push all non-instance Phis on the orig_phis worklist to update inputs 2699 // during Phase 4 if needed. 2700 orig_phis.append_if_missing(mphi); 2701 } else if (C->get_alias_index(t) != alias_idx) { 2702 // Create a new Phi with the specified alias index type. 2703 result = split_memory_phi(mphi, alias_idx, orig_phis); 2704 } 2705 } 2706 // the result is either MemNode, PhiNode, InitializeNode. 2707 return result; 2708 } 2709 2710 // 2711 // Convert the types of unescaped object to instance types where possible, 2712 // propagate the new type information through the graph, and update memory 2713 // edges and MergeMem inputs to reflect the new type. 2714 // 2715 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2716 // The processing is done in 4 phases: 2717 // 2718 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2719 // types for the CheckCastPP for allocations where possible. 2720 // Propagate the the new types through users as follows: 2721 // casts and Phi: push users on alloc_worklist 2722 // AddP: cast Base and Address inputs to the instance type 2723 // push any AddP users on alloc_worklist and push any memnode 2724 // users onto memnode_worklist. 2725 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2726 // search the Memory chain for a store with the appropriate type 2727 // address type. If a Phi is found, create a new version with 2728 // the appropriate memory slices from each of the Phi inputs. 2729 // For stores, process the users as follows: 2730 // MemNode: push on memnode_worklist 2731 // MergeMem: push on mergemem_worklist 2732 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2733 // moving the first node encountered of each instance type to the 2734 // the input corresponding to its alias index. 2735 // appropriate memory slice. 2736 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2737 // 2738 // In the following example, the CheckCastPP nodes are the cast of allocation 2739 // results and the allocation of node 29 is unescaped and eligible to be an 2740 // instance type. 2741 // 2742 // We start with: 2743 // 2744 // 7 Parm #memory 2745 // 10 ConI "12" 2746 // 19 CheckCastPP "Foo" 2747 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2748 // 29 CheckCastPP "Foo" 2749 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2750 // 2751 // 40 StoreP 25 7 20 ... alias_index=4 2752 // 50 StoreP 35 40 30 ... alias_index=4 2753 // 60 StoreP 45 50 20 ... alias_index=4 2754 // 70 LoadP _ 60 30 ... alias_index=4 2755 // 80 Phi 75 50 60 Memory alias_index=4 2756 // 90 LoadP _ 80 30 ... alias_index=4 2757 // 100 LoadP _ 80 20 ... alias_index=4 2758 // 2759 // 2760 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2761 // and creating a new alias index for node 30. This gives: 2762 // 2763 // 7 Parm #memory 2764 // 10 ConI "12" 2765 // 19 CheckCastPP "Foo" 2766 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2767 // 29 CheckCastPP "Foo" iid=24 2768 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2769 // 2770 // 40 StoreP 25 7 20 ... alias_index=4 2771 // 50 StoreP 35 40 30 ... alias_index=6 2772 // 60 StoreP 45 50 20 ... alias_index=4 2773 // 70 LoadP _ 60 30 ... alias_index=6 2774 // 80 Phi 75 50 60 Memory alias_index=4 2775 // 90 LoadP _ 80 30 ... alias_index=6 2776 // 100 LoadP _ 80 20 ... alias_index=4 2777 // 2778 // In phase 2, new memory inputs are computed for the loads and stores, 2779 // And a new version of the phi is created. In phase 4, the inputs to 2780 // node 80 are updated and then the memory nodes are updated with the 2781 // values computed in phase 2. This results in: 2782 // 2783 // 7 Parm #memory 2784 // 10 ConI "12" 2785 // 19 CheckCastPP "Foo" 2786 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2787 // 29 CheckCastPP "Foo" iid=24 2788 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2789 // 2790 // 40 StoreP 25 7 20 ... alias_index=4 2791 // 50 StoreP 35 7 30 ... alias_index=6 2792 // 60 StoreP 45 40 20 ... alias_index=4 2793 // 70 LoadP _ 50 30 ... alias_index=6 2794 // 80 Phi 75 40 60 Memory alias_index=4 2795 // 120 Phi 75 50 50 Memory alias_index=6 2796 // 90 LoadP _ 120 30 ... alias_index=6 2797 // 100 LoadP _ 80 20 ... alias_index=4 2798 // 2799 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 2800 GrowableArray<Node *> memnode_worklist; 2801 GrowableArray<PhiNode *> orig_phis; 2802 PhaseIterGVN *igvn = _igvn; 2803 uint new_index_start = (uint) _compile->num_alias_types(); 2804 Arena* arena = Thread::current()->resource_area(); 2805 VectorSet visited(arena); 2806 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2807 uint unique_old = _compile->unique(); 2808 2809 // Phase 1: Process possible allocations from alloc_worklist. 2810 // Create instance types for the CheckCastPP for allocations where possible. 2811 // 2812 // (Note: don't forget to change the order of the second AddP node on 2813 // the alloc_worklist if the order of the worklist processing is changed, 2814 // see the comment in find_second_addp().) 2815 // 2816 while (alloc_worklist.length() != 0) { 2817 Node *n = alloc_worklist.pop(); 2818 uint ni = n->_idx; 2819 if (n->is_Call()) { 2820 CallNode *alloc = n->as_Call(); 2821 // copy escape information to call node 2822 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2823 PointsToNode::EscapeState es = ptn->escape_state(); 2824 // We have an allocation or call which returns a Java object, 2825 // see if it is unescaped. 2826 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2827 continue; 2828 // Find CheckCastPP for the allocate or for the return value of a call 2829 n = alloc->result_cast(); 2830 if (n == NULL) { // No uses except Initialize node 2831 if (alloc->is_Allocate()) { 2832 // Set the scalar_replaceable flag for allocation 2833 // so it could be eliminated if it has no uses. 2834 alloc->as_Allocate()->_is_scalar_replaceable = true; 2835 } 2836 if (alloc->is_CallStaticJava()) { 2837 // Set the scalar_replaceable flag for boxing method 2838 // so it could be eliminated if it has no uses. 2839 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2840 } 2841 continue; 2842 } 2843 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2844 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2845 continue; 2846 } 2847 2848 // The inline code for Object.clone() casts the allocation result to 2849 // java.lang.Object and then to the actual type of the allocated 2850 // object. Detect this case and use the second cast. 2851 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2852 // the allocation result is cast to java.lang.Object and then 2853 // to the actual Array type. 2854 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2855 && (alloc->is_AllocateArray() || 2856 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2857 Node *cast2 = NULL; 2858 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2859 Node *use = n->fast_out(i); 2860 if (use->is_CheckCastPP()) { 2861 cast2 = use; 2862 break; 2863 } 2864 } 2865 if (cast2 != NULL) { 2866 n = cast2; 2867 } else { 2868 // Non-scalar replaceable if the allocation type is unknown statically 2869 // (reflection allocation), the object can't be restored during 2870 // deoptimization without precise type. 2871 continue; 2872 } 2873 } 2874 2875 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 2876 if (t == NULL) 2877 continue; // not a TypeOopPtr 2878 if (!t->klass_is_exact()) 2879 continue; // not an unique type 2880 2881 if (alloc->is_Allocate()) { 2882 // Set the scalar_replaceable flag for allocation 2883 // so it could be eliminated. 2884 alloc->as_Allocate()->_is_scalar_replaceable = true; 2885 } 2886 if (alloc->is_CallStaticJava()) { 2887 // Set the scalar_replaceable flag for boxing method 2888 // so it could be eliminated. 2889 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2890 } 2891 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 2892 // in order for an object to be scalar-replaceable, it must be: 2893 // - a direct allocation (not a call returning an object) 2894 // - non-escaping 2895 // - eligible to be a unique type 2896 // - not determined to be ineligible by escape analysis 2897 set_map(alloc, n); 2898 set_map(n, alloc); 2899 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 2900 igvn->hash_delete(n); 2901 igvn->set_type(n, tinst); 2902 n->raise_bottom_type(tinst); 2903 igvn->hash_insert(n); 2904 record_for_optimizer(n); 2905 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 2906 2907 // First, put on the worklist all Field edges from Connection Graph 2908 // which is more accurate then putting immediate users from Ideal Graph. 2909 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 2910 PointsToNode* tgt = e.get(); 2911 Node* use = tgt->ideal_node(); 2912 assert(tgt->is_Field() && use->is_AddP(), 2913 "only AddP nodes are Field edges in CG"); 2914 if (use->outcnt() > 0) { // Don't process dead nodes 2915 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 2916 if (addp2 != NULL) { 2917 assert(alloc->is_AllocateArray(),"array allocation was expected"); 2918 alloc_worklist.append_if_missing(addp2); 2919 } 2920 alloc_worklist.append_if_missing(use); 2921 } 2922 } 2923 2924 // An allocation may have an Initialize which has raw stores. Scan 2925 // the users of the raw allocation result and push AddP users 2926 // on alloc_worklist. 2927 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 2928 assert (raw_result != NULL, "must have an allocation result"); 2929 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 2930 Node *use = raw_result->fast_out(i); 2931 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 2932 Node* addp2 = find_second_addp(use, raw_result); 2933 if (addp2 != NULL) { 2934 assert(alloc->is_AllocateArray(),"array allocation was expected"); 2935 alloc_worklist.append_if_missing(addp2); 2936 } 2937 alloc_worklist.append_if_missing(use); 2938 } else if (use->is_MemBar()) { 2939 memnode_worklist.append_if_missing(use); 2940 } 2941 } 2942 } 2943 } else if (n->is_AddP()) { 2944 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 2945 if (jobj == NULL || jobj == phantom_obj) { 2946 #ifdef ASSERT 2947 ptnode_adr(get_addp_base(n)->_idx)->dump(); 2948 ptnode_adr(n->_idx)->dump(); 2949 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 2950 #endif 2951 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 2952 return; 2953 } 2954 Node *base = get_map(jobj->idx()); // CheckCastPP node 2955 if (!split_AddP(n, base)) continue; // wrong type from dead path 2956 } else if (n->is_Phi() || 2957 n->is_CheckCastPP() || 2958 n->is_EncodeP() || 2959 n->is_DecodeN() || 2960 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 2961 if (visited.test_set(n->_idx)) { 2962 assert(n->is_Phi(), "loops only through Phi's"); 2963 continue; // already processed 2964 } 2965 JavaObjectNode* jobj = unique_java_object(n); 2966 if (jobj == NULL || jobj == phantom_obj) { 2967 #ifdef ASSERT 2968 ptnode_adr(n->_idx)->dump(); 2969 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 2970 #endif 2971 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 2972 return; 2973 } else { 2974 Node *val = get_map(jobj->idx()); // CheckCastPP node 2975 TypeNode *tn = n->as_Type(); 2976 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 2977 assert(tinst != NULL && tinst->is_known_instance() && 2978 tinst->instance_id() == jobj->idx() , "instance type expected."); 2979 2980 const Type *tn_type = igvn->type(tn); 2981 const TypeOopPtr *tn_t; 2982 if (tn_type->isa_narrowoop()) { 2983 tn_t = tn_type->make_ptr()->isa_oopptr(); 2984 } else { 2985 tn_t = tn_type->isa_oopptr(); 2986 } 2987 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 2988 if (tn_type->isa_narrowoop()) { 2989 tn_type = tinst->make_narrowoop(); 2990 } else { 2991 tn_type = tinst; 2992 } 2993 igvn->hash_delete(tn); 2994 igvn->set_type(tn, tn_type); 2995 tn->set_type(tn_type); 2996 igvn->hash_insert(tn); 2997 record_for_optimizer(n); 2998 } else { 2999 assert(tn_type == TypePtr::NULL_PTR || 3000 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3001 "unexpected type"); 3002 continue; // Skip dead path with different type 3003 } 3004 } 3005 } else { 3006 debug_only(n->dump();) 3007 assert(false, "EA: unexpected node"); 3008 continue; 3009 } 3010 // push allocation's users on appropriate worklist 3011 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3012 Node *use = n->fast_out(i); 3013 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3014 // Load/store to instance's field 3015 memnode_worklist.append_if_missing(use); 3016 } else if (use->is_MemBar()) { 3017 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3018 memnode_worklist.append_if_missing(use); 3019 } 3020 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3021 Node* addp2 = find_second_addp(use, n); 3022 if (addp2 != NULL) { 3023 alloc_worklist.append_if_missing(addp2); 3024 } 3025 alloc_worklist.append_if_missing(use); 3026 } else if (use->is_Phi() || 3027 use->is_CheckCastPP() || 3028 use->is_EncodeNarrowPtr() || 3029 use->is_DecodeNarrowPtr() || 3030 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3031 alloc_worklist.append_if_missing(use); 3032 #ifdef ASSERT 3033 } else if (use->is_Mem()) { 3034 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3035 } else if (use->is_MergeMem()) { 3036 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3037 } else if (use->is_SafePoint()) { 3038 // Look for MergeMem nodes for calls which reference unique allocation 3039 // (through CheckCastPP nodes) even for debug info. 3040 Node* m = use->in(TypeFunc::Memory); 3041 if (m->is_MergeMem()) { 3042 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3043 } 3044 } else if (use->Opcode() == Op_EncodeISOArray) { 3045 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3046 // EncodeISOArray overwrites destination array 3047 memnode_worklist.append_if_missing(use); 3048 } 3049 } else { 3050 uint op = use->Opcode(); 3051 if (!(op == Op_CmpP || op == Op_Conv2B || 3052 op == Op_CastP2X || op == Op_StoreCM || 3053 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 3054 op == Op_StrEquals || op == Op_StrIndexOf)) { 3055 n->dump(); 3056 use->dump(); 3057 assert(false, "EA: missing allocation reference path"); 3058 } 3059 #endif 3060 } 3061 } 3062 3063 } 3064 // New alias types were created in split_AddP(). 3065 uint new_index_end = (uint) _compile->num_alias_types(); 3066 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3067 3068 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3069 // compute new values for Memory inputs (the Memory inputs are not 3070 // actually updated until phase 4.) 3071 if (memnode_worklist.length() == 0) 3072 return; // nothing to do 3073 while (memnode_worklist.length() != 0) { 3074 Node *n = memnode_worklist.pop(); 3075 if (visited.test_set(n->_idx)) 3076 continue; 3077 if (n->is_Phi() || n->is_ClearArray()) { 3078 // we don't need to do anything, but the users must be pushed 3079 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3080 // we don't need to do anything, but the users must be pushed 3081 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 3082 if (n == NULL) 3083 continue; 3084 } else if (n->Opcode() == Op_EncodeISOArray) { 3085 // get the memory projection 3086 n = n->find_out_with(Op_SCMemProj); 3087 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3088 } else { 3089 assert(n->is_Mem(), "memory node required."); 3090 Node *addr = n->in(MemNode::Address); 3091 const Type *addr_t = igvn->type(addr); 3092 if (addr_t == Type::TOP) 3093 continue; 3094 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3095 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3096 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3097 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3098 if (_compile->failing()) { 3099 return; 3100 } 3101 if (mem != n->in(MemNode::Memory)) { 3102 // We delay the memory edge update since we need old one in 3103 // MergeMem code below when instances memory slices are separated. 3104 set_map(n, mem); 3105 } 3106 if (n->is_Load()) { 3107 continue; // don't push users 3108 } else if (n->is_LoadStore()) { 3109 // get the memory projection 3110 n = n->find_out_with(Op_SCMemProj); 3111 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3112 } 3113 } 3114 // push user on appropriate worklist 3115 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3116 Node *use = n->fast_out(i); 3117 if (use->is_Phi() || use->is_ClearArray()) { 3118 memnode_worklist.append_if_missing(use); 3119 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3120 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3121 continue; 3122 memnode_worklist.append_if_missing(use); 3123 } else if (use->is_MemBar()) { 3124 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3125 memnode_worklist.append_if_missing(use); 3126 } 3127 #ifdef ASSERT 3128 } else if(use->is_Mem()) { 3129 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3130 } else if (use->is_MergeMem()) { 3131 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3132 } else if (use->Opcode() == Op_EncodeISOArray) { 3133 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3134 // EncodeISOArray overwrites destination array 3135 memnode_worklist.append_if_missing(use); 3136 } 3137 } else { 3138 uint op = use->Opcode(); 3139 if (!(op == Op_StoreCM || 3140 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 3141 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 3142 op == Op_AryEq || op == Op_StrComp || 3143 op == Op_StrEquals || op == Op_StrIndexOf)) { 3144 n->dump(); 3145 use->dump(); 3146 assert(false, "EA: missing memory path"); 3147 } 3148 #endif 3149 } 3150 } 3151 } 3152 3153 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3154 // Walk each memory slice moving the first node encountered of each 3155 // instance type to the the input corresponding to its alias index. 3156 uint length = _mergemem_worklist.length(); 3157 for( uint next = 0; next < length; ++next ) { 3158 MergeMemNode* nmm = _mergemem_worklist.at(next); 3159 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3160 // Note: we don't want to use MergeMemStream here because we only want to 3161 // scan inputs which exist at the start, not ones we add during processing. 3162 // Note 2: MergeMem may already contains instance memory slices added 3163 // during find_inst_mem() call when memory nodes were processed above. 3164 igvn->hash_delete(nmm); 3165 uint nslices = nmm->req(); 3166 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3167 Node* mem = nmm->in(i); 3168 Node* cur = NULL; 3169 if (mem == NULL || mem->is_top()) 3170 continue; 3171 // First, update mergemem by moving memory nodes to corresponding slices 3172 // if their type became more precise since this mergemem was created. 3173 while (mem->is_Mem()) { 3174 const Type *at = igvn->type(mem->in(MemNode::Address)); 3175 if (at != Type::TOP) { 3176 assert (at->isa_ptr() != NULL, "pointer type required."); 3177 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3178 if (idx == i) { 3179 if (cur == NULL) 3180 cur = mem; 3181 } else { 3182 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3183 nmm->set_memory_at(idx, mem); 3184 } 3185 } 3186 } 3187 mem = mem->in(MemNode::Memory); 3188 } 3189 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3190 // Find any instance of the current type if we haven't encountered 3191 // already a memory slice of the instance along the memory chain. 3192 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3193 if((uint)_compile->get_general_index(ni) == i) { 3194 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3195 if (nmm->is_empty_memory(m)) { 3196 Node* result = find_inst_mem(mem, ni, orig_phis); 3197 if (_compile->failing()) { 3198 return; 3199 } 3200 nmm->set_memory_at(ni, result); 3201 } 3202 } 3203 } 3204 } 3205 // Find the rest of instances values 3206 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3207 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3208 Node* result = step_through_mergemem(nmm, ni, tinst); 3209 if (result == nmm->base_memory()) { 3210 // Didn't find instance memory, search through general slice recursively. 3211 result = nmm->memory_at(_compile->get_general_index(ni)); 3212 result = find_inst_mem(result, ni, orig_phis); 3213 if (_compile->failing()) { 3214 return; 3215 } 3216 nmm->set_memory_at(ni, result); 3217 } 3218 } 3219 igvn->hash_insert(nmm); 3220 record_for_optimizer(nmm); 3221 } 3222 3223 // Phase 4: Update the inputs of non-instance memory Phis and 3224 // the Memory input of memnodes 3225 // First update the inputs of any non-instance Phi's from 3226 // which we split out an instance Phi. Note we don't have 3227 // to recursively process Phi's encounted on the input memory 3228 // chains as is done in split_memory_phi() since they will 3229 // also be processed here. 3230 for (int j = 0; j < orig_phis.length(); j++) { 3231 PhiNode *phi = orig_phis.at(j); 3232 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3233 igvn->hash_delete(phi); 3234 for (uint i = 1; i < phi->req(); i++) { 3235 Node *mem = phi->in(i); 3236 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3237 if (_compile->failing()) { 3238 return; 3239 } 3240 if (mem != new_mem) { 3241 phi->set_req(i, new_mem); 3242 } 3243 } 3244 igvn->hash_insert(phi); 3245 record_for_optimizer(phi); 3246 } 3247 3248 // Update the memory inputs of MemNodes with the value we computed 3249 // in Phase 2 and move stores memory users to corresponding memory slices. 3250 // Disable memory split verification code until the fix for 6984348. 3251 // Currently it produces false negative results since it does not cover all cases. 3252 #if 0 // ifdef ASSERT 3253 visited.Reset(); 3254 Node_Stack old_mems(arena, _compile->unique() >> 2); 3255 #endif 3256 for (uint i = 0; i < ideal_nodes.size(); i++) { 3257 Node* n = ideal_nodes.at(i); 3258 Node* nmem = get_map(n->_idx); 3259 assert(nmem != NULL, "sanity"); 3260 if (n->is_Mem()) { 3261 #if 0 // ifdef ASSERT 3262 Node* old_mem = n->in(MemNode::Memory); 3263 if (!visited.test_set(old_mem->_idx)) { 3264 old_mems.push(old_mem, old_mem->outcnt()); 3265 } 3266 #endif 3267 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3268 if (!n->is_Load()) { 3269 // Move memory users of a store first. 3270 move_inst_mem(n, orig_phis); 3271 } 3272 // Now update memory input 3273 igvn->hash_delete(n); 3274 n->set_req(MemNode::Memory, nmem); 3275 igvn->hash_insert(n); 3276 record_for_optimizer(n); 3277 } else { 3278 assert(n->is_Allocate() || n->is_CheckCastPP() || 3279 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3280 } 3281 } 3282 #if 0 // ifdef ASSERT 3283 // Verify that memory was split correctly 3284 while (old_mems.is_nonempty()) { 3285 Node* old_mem = old_mems.node(); 3286 uint old_cnt = old_mems.index(); 3287 old_mems.pop(); 3288 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3289 } 3290 #endif 3291 } 3292 3293 #ifndef PRODUCT 3294 static const char *node_type_names[] = { 3295 "UnknownType", 3296 "JavaObject", 3297 "LocalVar", 3298 "Field", 3299 "Arraycopy" 3300 }; 3301 3302 static const char *esc_names[] = { 3303 "UnknownEscape", 3304 "NoEscape", 3305 "ArgEscape", 3306 "GlobalEscape" 3307 }; 3308 3309 void PointsToNode::dump(bool print_state) const { 3310 NodeType nt = node_type(); 3311 tty->print("%s ", node_type_names[(int) nt]); 3312 if (print_state) { 3313 EscapeState es = escape_state(); 3314 EscapeState fields_es = fields_escape_state(); 3315 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3316 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3317 tty->print("NSR "); 3318 } 3319 if (is_Field()) { 3320 FieldNode* f = (FieldNode*)this; 3321 if (f->is_oop()) 3322 tty->print("oop "); 3323 if (f->offset() > 0) 3324 tty->print("+%d ", f->offset()); 3325 tty->print("("); 3326 for (BaseIterator i(f); i.has_next(); i.next()) { 3327 PointsToNode* b = i.get(); 3328 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3329 } 3330 tty->print(" )"); 3331 } 3332 tty->print("["); 3333 for (EdgeIterator i(this); i.has_next(); i.next()) { 3334 PointsToNode* e = i.get(); 3335 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3336 } 3337 tty->print(" ["); 3338 for (UseIterator i(this); i.has_next(); i.next()) { 3339 PointsToNode* u = i.get(); 3340 bool is_base = false; 3341 if (PointsToNode::is_base_use(u)) { 3342 is_base = true; 3343 u = PointsToNode::get_use_node(u)->as_Field(); 3344 } 3345 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3346 } 3347 tty->print(" ]] "); 3348 if (_node == NULL) 3349 tty->print_cr("<null>"); 3350 else 3351 _node->dump(); 3352 } 3353 3354 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3355 bool first = true; 3356 int ptnodes_length = ptnodes_worklist.length(); 3357 for (int i = 0; i < ptnodes_length; i++) { 3358 PointsToNode *ptn = ptnodes_worklist.at(i); 3359 if (ptn == NULL || !ptn->is_JavaObject()) 3360 continue; 3361 PointsToNode::EscapeState es = ptn->escape_state(); 3362 if ((es != PointsToNode::NoEscape) && !Verbose) { 3363 continue; 3364 } 3365 Node* n = ptn->ideal_node(); 3366 if (n->is_Allocate() || (n->is_CallStaticJava() && 3367 n->as_CallStaticJava()->is_boxing_method())) { 3368 if (first) { 3369 tty->cr(); 3370 tty->print("======== Connection graph for "); 3371 _compile->method()->print_short_name(); 3372 tty->cr(); 3373 first = false; 3374 } 3375 ptn->dump(); 3376 // Print all locals and fields which reference this allocation 3377 for (UseIterator j(ptn); j.has_next(); j.next()) { 3378 PointsToNode* use = j.get(); 3379 if (use->is_LocalVar()) { 3380 use->dump(Verbose); 3381 } else if (Verbose) { 3382 use->dump(); 3383 } 3384 } 3385 tty->cr(); 3386 } 3387 } 3388 } 3389 #endif