1 /* 2 * Copyright (c) 2005, 2013, 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 _collecting(true), 42 _verify(false), 43 _compile(C), 44 _igvn(igvn), 45 _node_map(C->comp_arena()) { 46 // Add unknown java object. 47 add_java_object(C->top(), PointsToNode::GlobalEscape); 48 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 49 // Add ConP(#NULL) and ConN(#NULL) nodes. 50 Node* oop_null = igvn->zerocon(T_OBJECT); 51 assert(oop_null->_idx < nodes_size(), "should be created already"); 52 add_java_object(oop_null, PointsToNode::NoEscape); 53 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 54 if (UseCompressedOops) { 55 Node* noop_null = igvn->zerocon(T_NARROWOOP); 56 assert(noop_null->_idx < nodes_size(), "should be created already"); 57 map_ideal_node(noop_null, null_obj); 58 } 59 _pcmp_neq = NULL; // Should be initialized 60 _pcmp_eq = NULL; 61 } 62 63 bool ConnectionGraph::has_candidates(Compile *C) { 64 // EA brings benefits only when the code has allocations and/or locks which 65 // are represented by ideal Macro nodes. 66 int cnt = C->macro_count(); 67 for (int i = 0; i < cnt; i++) { 68 Node *n = C->macro_node(i); 69 if (n->is_Allocate()) 70 return true; 71 if (n->is_Lock()) { 72 Node* obj = n->as_Lock()->obj_node()->uncast(); 73 if (!(obj->is_Parm() || obj->is_Con())) 74 return true; 75 } 76 if (n->is_CallStaticJava() && 77 n->as_CallStaticJava()->is_boxing_method()) { 78 return true; 79 } 80 } 81 return false; 82 } 83 84 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 85 Compile::TracePhase t2("escapeAnalysis", &Phase::_t_escapeAnalysis, true); 86 ResourceMark rm; 87 88 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 89 // to create space for them in ConnectionGraph::_nodes[]. 90 Node* oop_null = igvn->zerocon(T_OBJECT); 91 Node* noop_null = igvn->zerocon(T_NARROWOOP); 92 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 93 // Perform escape analysis 94 if (congraph->compute_escape()) { 95 // There are non escaping objects. 96 C->set_congraph(congraph); 97 } 98 // Cleanup. 99 if (oop_null->outcnt() == 0) 100 igvn->hash_delete(oop_null); 101 if (noop_null->outcnt() == 0) 102 igvn->hash_delete(noop_null); 103 } 104 105 bool ConnectionGraph::compute_escape() { 106 Compile* C = _compile; 107 PhaseGVN* igvn = _igvn; 108 109 // Worklists used by EA. 110 Unique_Node_List delayed_worklist; 111 GrowableArray<Node*> alloc_worklist; 112 GrowableArray<Node*> ptr_cmp_worklist; 113 GrowableArray<Node*> storestore_worklist; 114 GrowableArray<PointsToNode*> ptnodes_worklist; 115 GrowableArray<JavaObjectNode*> java_objects_worklist; 116 GrowableArray<JavaObjectNode*> non_escaped_worklist; 117 GrowableArray<FieldNode*> oop_fields_worklist; 118 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 119 120 { Compile::TracePhase t3("connectionGraph", &Phase::_t_connectionGraph, true); 121 122 // 1. Populate Connection Graph (CG) with PointsTo nodes. 123 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 124 // Initialize worklist 125 if (C->root() != NULL) { 126 ideal_nodes.push(C->root()); 127 } 128 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 129 Node* n = ideal_nodes.at(next); 130 // Create PointsTo nodes and add them to Connection Graph. Called 131 // only once per ideal node since ideal_nodes is Unique_Node list. 132 add_node_to_connection_graph(n, &delayed_worklist); 133 PointsToNode* ptn = ptnode_adr(n->_idx); 134 if (ptn != NULL) { 135 ptnodes_worklist.append(ptn); 136 if (ptn->is_JavaObject()) { 137 java_objects_worklist.append(ptn->as_JavaObject()); 138 if ((n->is_Allocate() || n->is_CallStaticJava()) && 139 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 140 // Only allocations and java static calls results are interesting. 141 non_escaped_worklist.append(ptn->as_JavaObject()); 142 } 143 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 144 oop_fields_worklist.append(ptn->as_Field()); 145 } 146 } 147 if (n->is_MergeMem()) { 148 // Collect all MergeMem nodes to add memory slices for 149 // scalar replaceable objects in split_unique_types(). 150 _mergemem_worklist.append(n->as_MergeMem()); 151 } else if (OptimizePtrCompare && n->is_Cmp() && 152 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) { 153 // Collect compare pointers nodes. 154 ptr_cmp_worklist.append(n); 155 } else if (n->is_MemBarStoreStore()) { 156 // Collect all MemBarStoreStore nodes so that depending on the 157 // escape status of the associated Allocate node some of them 158 // may be eliminated. 159 storestore_worklist.append(n); 160 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 161 (n->req() > MemBarNode::Precedent)) { 162 record_for_optimizer(n); 163 #ifdef ASSERT 164 } else if (n->is_AddP()) { 165 // Collect address nodes for graph verification. 166 addp_worklist.append(n); 167 #endif 168 } 169 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 170 Node* m = n->fast_out(i); // Get user 171 ideal_nodes.push(m); 172 } 173 } 174 if (non_escaped_worklist.length() == 0) { 175 _collecting = false; 176 return false; // Nothing to do. 177 } 178 // Add final simple edges to graph. 179 while(delayed_worklist.size() > 0) { 180 Node* n = delayed_worklist.pop(); 181 add_final_edges(n); 182 } 183 int ptnodes_length = ptnodes_worklist.length(); 184 185 #ifdef ASSERT 186 if (VerifyConnectionGraph) { 187 // Verify that no new simple edges could be created and all 188 // local vars has edges. 189 _verify = true; 190 for (int next = 0; next < ptnodes_length; ++next) { 191 PointsToNode* ptn = ptnodes_worklist.at(next); 192 add_final_edges(ptn->ideal_node()); 193 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 194 ptn->dump(); 195 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 196 } 197 } 198 _verify = false; 199 } 200 #endif 201 202 // 2. Finish Graph construction by propagating references to all 203 // java objects through graph. 204 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 205 java_objects_worklist, oop_fields_worklist)) { 206 // All objects escaped or hit time or iterations limits. 207 _collecting = false; 208 return false; 209 } 210 211 // 3. Adjust scalar_replaceable state of nonescaping objects and push 212 // scalar replaceable allocations on alloc_worklist for processing 213 // in split_unique_types(). 214 int non_escaped_length = non_escaped_worklist.length(); 215 for (int next = 0; next < non_escaped_length; next++) { 216 JavaObjectNode* ptn = non_escaped_worklist.at(next); 217 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 218 Node* n = ptn->ideal_node(); 219 if (n->is_Allocate()) { 220 n->as_Allocate()->_is_non_escaping = noescape; 221 } 222 if (n->is_CallStaticJava()) { 223 n->as_CallStaticJava()->_is_non_escaping = noescape; 224 } 225 if (noescape && ptn->scalar_replaceable()) { 226 adjust_scalar_replaceable_state(ptn); 227 if (ptn->scalar_replaceable()) { 228 alloc_worklist.append(ptn->ideal_node()); 229 } 230 } 231 } 232 233 #ifdef ASSERT 234 if (VerifyConnectionGraph) { 235 // Verify that graph is complete - no new edges could be added or needed. 236 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 237 java_objects_worklist, addp_worklist); 238 } 239 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 240 assert(null_obj->escape_state() == PointsToNode::NoEscape && 241 null_obj->edge_count() == 0 && 242 !null_obj->arraycopy_src() && 243 !null_obj->arraycopy_dst(), "sanity"); 244 #endif 245 246 _collecting = false; 247 248 } // TracePhase t3("connectionGraph") 249 250 // 4. Optimize ideal graph based on EA information. 251 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 252 if (has_non_escaping_obj) { 253 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 254 } 255 256 #ifndef PRODUCT 257 if (PrintEscapeAnalysis) { 258 dump(ptnodes_worklist); // Dump ConnectionGraph 259 } 260 #endif 261 262 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 263 #ifdef ASSERT 264 if (VerifyConnectionGraph) { 265 int alloc_length = alloc_worklist.length(); 266 for (int next = 0; next < alloc_length; ++next) { 267 Node* n = alloc_worklist.at(next); 268 PointsToNode* ptn = ptnode_adr(n->_idx); 269 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 270 } 271 } 272 #endif 273 274 // 5. Separate memory graph for scalar replaceable allcations. 275 if (has_scalar_replaceable_candidates && 276 C->AliasLevel() >= 3 && EliminateAllocations) { 277 // Now use the escape information to create unique types for 278 // scalar replaceable objects. 279 split_unique_types(alloc_worklist); 280 if (C->failing()) return false; 281 C->print_method(PHASE_AFTER_EA, 2); 282 283 #ifdef ASSERT 284 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 285 tty->print("=== No allocations eliminated for "); 286 C->method()->print_short_name(); 287 if(!EliminateAllocations) { 288 tty->print(" since EliminateAllocations is off ==="); 289 } else if(!has_scalar_replaceable_candidates) { 290 tty->print(" since there are no scalar replaceable candidates ==="); 291 } else if(C->AliasLevel() < 3) { 292 tty->print(" since AliasLevel < 3 ==="); 293 } 294 tty->cr(); 295 #endif 296 } 297 return has_non_escaping_obj; 298 } 299 300 // Utility function for nodes that load an object 301 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 302 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 303 // ThreadLocal has RawPtr type. 304 const Type* t = _igvn->type(n); 305 if (t->make_ptr() != NULL) { 306 Node* adr = n->in(MemNode::Address); 307 #ifdef ASSERT 308 if (!adr->is_AddP()) { 309 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 310 } else { 311 assert((ptnode_adr(adr->_idx) == NULL || 312 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 313 } 314 #endif 315 add_local_var_and_edge(n, PointsToNode::NoEscape, 316 adr, delayed_worklist); 317 } 318 } 319 320 // Populate Connection Graph with PointsTo nodes and create simple 321 // connection graph edges. 322 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 323 assert(!_verify, "this method sould not be called for verification"); 324 PhaseGVN* igvn = _igvn; 325 uint n_idx = n->_idx; 326 PointsToNode* n_ptn = ptnode_adr(n_idx); 327 if (n_ptn != NULL) 328 return; // No need to redefine PointsTo node during first iteration. 329 330 if (n->is_Call()) { 331 // Arguments to allocation and locking don't escape. 332 if (n->is_AbstractLock()) { 333 // Put Lock and Unlock nodes on IGVN worklist to process them during 334 // first IGVN optimization when escape information is still available. 335 record_for_optimizer(n); 336 } else if (n->is_Allocate()) { 337 add_call_node(n->as_Call()); 338 record_for_optimizer(n); 339 } else { 340 if (n->is_CallStaticJava()) { 341 const char* name = n->as_CallStaticJava()->_name; 342 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 343 return; // Skip uncommon traps 344 } 345 // Don't mark as processed since call's arguments have to be processed. 346 delayed_worklist->push(n); 347 // Check if a call returns an object. 348 if ((n->as_Call()->returns_pointer() && 349 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) || 350 (n->is_CallStaticJava() && 351 n->as_CallStaticJava()->is_boxing_method())) { 352 add_call_node(n->as_Call()); 353 } 354 } 355 return; 356 } 357 // Put this check here to process call arguments since some call nodes 358 // point to phantom_obj. 359 if (n_ptn == phantom_obj || n_ptn == null_obj) 360 return; // Skip predefined nodes. 361 362 int opcode = n->Opcode(); 363 switch (opcode) { 364 case Op_AddP: { 365 Node* base = get_addp_base(n); 366 PointsToNode* ptn_base = ptnode_adr(base->_idx); 367 // Field nodes are created for all field types. They are used in 368 // adjust_scalar_replaceable_state() and split_unique_types(). 369 // Note, non-oop fields will have only base edges in Connection 370 // Graph because such fields are not used for oop loads and stores. 371 int offset = address_offset(n, igvn); 372 add_field(n, PointsToNode::NoEscape, offset); 373 if (ptn_base == NULL) { 374 delayed_worklist->push(n); // Process it later. 375 } else { 376 n_ptn = ptnode_adr(n_idx); 377 add_base(n_ptn->as_Field(), ptn_base); 378 } 379 break; 380 } 381 case Op_CastX2P: { 382 map_ideal_node(n, phantom_obj); 383 break; 384 } 385 case Op_CastPP: 386 case Op_CheckCastPP: 387 case Op_EncodeP: 388 case Op_DecodeN: 389 case Op_EncodePKlass: 390 case Op_DecodeNKlass: { 391 add_local_var_and_edge(n, PointsToNode::NoEscape, 392 n->in(1), delayed_worklist); 393 break; 394 } 395 case Op_CMoveP: { 396 add_local_var(n, PointsToNode::NoEscape); 397 // Do not add edges during first iteration because some could be 398 // not defined yet. 399 delayed_worklist->push(n); 400 break; 401 } 402 case Op_ConP: 403 case Op_ConN: 404 case Op_ConNKlass: { 405 // assume all oop constants globally escape except for null 406 PointsToNode::EscapeState es; 407 const Type* t = igvn->type(n); 408 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 409 es = PointsToNode::NoEscape; 410 } else { 411 es = PointsToNode::GlobalEscape; 412 } 413 add_java_object(n, es); 414 break; 415 } 416 case Op_CreateEx: { 417 // assume that all exception objects globally escape 418 add_java_object(n, PointsToNode::GlobalEscape); 419 break; 420 } 421 case Op_LoadKlass: 422 case Op_LoadNKlass: { 423 // Unknown class is loaded 424 map_ideal_node(n, phantom_obj); 425 break; 426 } 427 case Op_LoadP: 428 case Op_LoadN: 429 case Op_LoadPLocked: { 430 add_objload_to_connection_graph(n, delayed_worklist); 431 break; 432 } 433 case Op_Parm: { 434 map_ideal_node(n, phantom_obj); 435 break; 436 } 437 case Op_PartialSubtypeCheck: { 438 // Produces Null or notNull and is used in only in CmpP so 439 // phantom_obj could be used. 440 map_ideal_node(n, phantom_obj); // Result is unknown 441 break; 442 } 443 case Op_Phi: { 444 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 445 // ThreadLocal has RawPtr type. 446 const Type* t = n->as_Phi()->type(); 447 if (t->make_ptr() != NULL) { 448 add_local_var(n, PointsToNode::NoEscape); 449 // Do not add edges during first iteration because some could be 450 // not defined yet. 451 delayed_worklist->push(n); 452 } 453 break; 454 } 455 case Op_Proj: { 456 // we are only interested in the oop result projection from a call 457 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 458 n->in(0)->as_Call()->returns_pointer()) { 459 add_local_var_and_edge(n, PointsToNode::NoEscape, 460 n->in(0), delayed_worklist); 461 } 462 break; 463 } 464 case Op_Rethrow: // Exception object escapes 465 case Op_Return: { 466 if (n->req() > TypeFunc::Parms && 467 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 468 // Treat Return value as LocalVar with GlobalEscape escape state. 469 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 470 n->in(TypeFunc::Parms), delayed_worklist); 471 } 472 break; 473 } 474 case Op_GetAndSetP: 475 case Op_GetAndSetN: { 476 add_objload_to_connection_graph(n, delayed_worklist); 477 // fallthrough 478 } 479 case Op_StoreP: 480 case Op_StoreN: 481 case Op_StoreNKlass: 482 case Op_StorePConditional: 483 case Op_CompareAndSwapP: 484 case Op_CompareAndSwapN: { 485 Node* adr = n->in(MemNode::Address); 486 const Type *adr_type = igvn->type(adr); 487 adr_type = adr_type->make_ptr(); 488 if (adr_type == NULL) { 489 break; // skip dead nodes 490 } 491 if (adr_type->isa_oopptr() || 492 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 493 (adr_type == TypeRawPtr::NOTNULL && 494 adr->in(AddPNode::Address)->is_Proj() && 495 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 496 delayed_worklist->push(n); // Process it later. 497 #ifdef ASSERT 498 assert(adr->is_AddP(), "expecting an AddP"); 499 if (adr_type == TypeRawPtr::NOTNULL) { 500 // Verify a raw address for a store captured by Initialize node. 501 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 502 assert(offs != Type::OffsetBot, "offset must be a constant"); 503 } 504 #endif 505 } else { 506 // Ignore copy the displaced header to the BoxNode (OSR compilation). 507 if (adr->is_BoxLock()) 508 break; 509 // Stored value escapes in unsafe access. 510 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) { 511 // Pointer stores in G1 barriers looks like unsafe access. 512 // Ignore such stores to be able scalar replace non-escaping 513 // allocations. 514 if (UseG1GC && adr->is_AddP()) { 515 Node* base = get_addp_base(adr); 516 if (base->Opcode() == Op_LoadP && 517 base->in(MemNode::Address)->is_AddP()) { 518 adr = base->in(MemNode::Address); 519 Node* tls = get_addp_base(adr); 520 if (tls->Opcode() == Op_ThreadLocal) { 521 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 522 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() + 523 PtrQueue::byte_offset_of_buf())) { 524 break; // G1 pre barier previous oop value store. 525 } 526 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() + 527 PtrQueue::byte_offset_of_buf())) { 528 break; // G1 post barier card address store. 529 } 530 } 531 } 532 } 533 delayed_worklist->push(n); // Process unsafe access later. 534 break; 535 } 536 #ifdef ASSERT 537 n->dump(1); 538 assert(false, "not unsafe or G1 barrier raw StoreP"); 539 #endif 540 } 541 break; 542 } 543 case Op_AryEq: 544 case Op_StrComp: 545 case Op_StrEquals: 546 case Op_StrIndexOf: 547 case Op_EncodeISOArray: { 548 add_local_var(n, PointsToNode::ArgEscape); 549 delayed_worklist->push(n); // Process it later. 550 break; 551 } 552 case Op_ThreadLocal: { 553 add_java_object(n, PointsToNode::ArgEscape); 554 break; 555 } 556 default: 557 ; // Do nothing for nodes not related to EA. 558 } 559 return; 560 } 561 562 #ifdef ASSERT 563 #define ELSE_FAIL(name) \ 564 /* Should not be called for not pointer type. */ \ 565 n->dump(1); \ 566 assert(false, name); \ 567 break; 568 #else 569 #define ELSE_FAIL(name) \ 570 break; 571 #endif 572 573 // Add final simple edges to graph. 574 void ConnectionGraph::add_final_edges(Node *n) { 575 PointsToNode* n_ptn = ptnode_adr(n->_idx); 576 #ifdef ASSERT 577 if (_verify && n_ptn->is_JavaObject()) 578 return; // This method does not change graph for JavaObject. 579 #endif 580 581 if (n->is_Call()) { 582 process_call_arguments(n->as_Call()); 583 return; 584 } 585 assert(n->is_Store() || n->is_LoadStore() || 586 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 587 "node should be registered already"); 588 int opcode = n->Opcode(); 589 switch (opcode) { 590 case Op_AddP: { 591 Node* base = get_addp_base(n); 592 PointsToNode* ptn_base = ptnode_adr(base->_idx); 593 assert(ptn_base != NULL, "field's base should be registered"); 594 add_base(n_ptn->as_Field(), ptn_base); 595 break; 596 } 597 case Op_CastPP: 598 case Op_CheckCastPP: 599 case Op_EncodeP: 600 case Op_DecodeN: 601 case Op_EncodePKlass: 602 case Op_DecodeNKlass: { 603 add_local_var_and_edge(n, PointsToNode::NoEscape, 604 n->in(1), NULL); 605 break; 606 } 607 case Op_CMoveP: { 608 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 609 Node* in = n->in(i); 610 if (in == NULL) 611 continue; // ignore NULL 612 Node* uncast_in = in->uncast(); 613 if (uncast_in->is_top() || uncast_in == n) 614 continue; // ignore top or inputs which go back this node 615 PointsToNode* ptn = ptnode_adr(in->_idx); 616 assert(ptn != NULL, "node should be registered"); 617 add_edge(n_ptn, ptn); 618 } 619 break; 620 } 621 case Op_LoadP: 622 case Op_LoadN: 623 case Op_LoadPLocked: { 624 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 625 // ThreadLocal has RawPtr type. 626 const Type* t = _igvn->type(n); 627 if (t->make_ptr() != NULL) { 628 Node* adr = n->in(MemNode::Address); 629 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 630 break; 631 } 632 ELSE_FAIL("Op_LoadP"); 633 } 634 case Op_Phi: { 635 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 636 // ThreadLocal has RawPtr type. 637 const Type* t = n->as_Phi()->type(); 638 if (t->make_ptr() != NULL) { 639 for (uint i = 1; i < n->req(); i++) { 640 Node* in = n->in(i); 641 if (in == NULL) 642 continue; // ignore NULL 643 Node* uncast_in = in->uncast(); 644 if (uncast_in->is_top() || uncast_in == n) 645 continue; // ignore top or inputs which go back this node 646 PointsToNode* ptn = ptnode_adr(in->_idx); 647 assert(ptn != NULL, "node should be registered"); 648 add_edge(n_ptn, ptn); 649 } 650 break; 651 } 652 ELSE_FAIL("Op_Phi"); 653 } 654 case Op_Proj: { 655 // we are only interested in the oop result projection from a call 656 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 657 n->in(0)->as_Call()->returns_pointer()) { 658 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 659 break; 660 } 661 ELSE_FAIL("Op_Proj"); 662 } 663 case Op_Rethrow: // Exception object escapes 664 case Op_Return: { 665 if (n->req() > TypeFunc::Parms && 666 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 667 // Treat Return value as LocalVar with GlobalEscape escape state. 668 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 669 n->in(TypeFunc::Parms), NULL); 670 break; 671 } 672 ELSE_FAIL("Op_Return"); 673 } 674 case Op_StoreP: 675 case Op_StoreN: 676 case Op_StoreNKlass: 677 case Op_StorePConditional: 678 case Op_CompareAndSwapP: 679 case Op_CompareAndSwapN: 680 case Op_GetAndSetP: 681 case Op_GetAndSetN: { 682 Node* adr = n->in(MemNode::Address); 683 const Type *adr_type = _igvn->type(adr); 684 adr_type = adr_type->make_ptr(); 685 #ifdef ASSERT 686 if (adr_type == NULL) { 687 n->dump(1); 688 assert(adr_type != NULL, "dead node should not be on list"); 689 break; 690 } 691 #endif 692 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) { 693 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 694 } 695 if (adr_type->isa_oopptr() || 696 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 697 (adr_type == TypeRawPtr::NOTNULL && 698 adr->in(AddPNode::Address)->is_Proj() && 699 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 700 // Point Address to Value 701 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 702 assert(adr_ptn != NULL && 703 adr_ptn->as_Field()->is_oop(), "node should be registered"); 704 Node *val = n->in(MemNode::ValueIn); 705 PointsToNode* ptn = ptnode_adr(val->_idx); 706 assert(ptn != NULL, "node should be registered"); 707 add_edge(adr_ptn, ptn); 708 break; 709 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) { 710 // Stored value escapes in unsafe access. 711 Node *val = n->in(MemNode::ValueIn); 712 PointsToNode* ptn = ptnode_adr(val->_idx); 713 assert(ptn != NULL, "node should be registered"); 714 ptn->set_escape_state(PointsToNode::GlobalEscape); 715 // Add edge to object for unsafe access with offset. 716 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 717 assert(adr_ptn != NULL, "node should be registered"); 718 if (adr_ptn->is_Field()) { 719 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 720 add_edge(adr_ptn, ptn); 721 } 722 break; 723 } 724 ELSE_FAIL("Op_StoreP"); 725 } 726 case Op_AryEq: 727 case Op_StrComp: 728 case Op_StrEquals: 729 case Op_StrIndexOf: 730 case Op_EncodeISOArray: { 731 // char[] arrays passed to string intrinsic do not escape but 732 // they are not scalar replaceable. Adjust escape state for them. 733 // Start from in(2) edge since in(1) is memory edge. 734 for (uint i = 2; i < n->req(); i++) { 735 Node* adr = n->in(i); 736 const Type* at = _igvn->type(adr); 737 if (!adr->is_top() && at->isa_ptr()) { 738 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 739 at->isa_ptr() != NULL, "expecting a pointer"); 740 if (adr->is_AddP()) { 741 adr = get_addp_base(adr); 742 } 743 PointsToNode* ptn = ptnode_adr(adr->_idx); 744 assert(ptn != NULL, "node should be registered"); 745 add_edge(n_ptn, ptn); 746 } 747 } 748 break; 749 } 750 default: { 751 // This method should be called only for EA specific nodes which may 752 // miss some edges when they were created. 753 #ifdef ASSERT 754 n->dump(1); 755 #endif 756 guarantee(false, "unknown node"); 757 } 758 } 759 return; 760 } 761 762 void ConnectionGraph::add_call_node(CallNode* call) { 763 assert(call->returns_pointer(), "only for call which returns pointer"); 764 uint call_idx = call->_idx; 765 if (call->is_Allocate()) { 766 Node* k = call->in(AllocateNode::KlassNode); 767 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 768 assert(kt != NULL, "TypeKlassPtr required."); 769 ciKlass* cik = kt->klass(); 770 PointsToNode::EscapeState es = PointsToNode::NoEscape; 771 bool scalar_replaceable = true; 772 if (call->is_AllocateArray()) { 773 if (!cik->is_array_klass()) { // StressReflectiveCode 774 es = PointsToNode::GlobalEscape; 775 } else { 776 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 777 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 778 // Not scalar replaceable if the length is not constant or too big. 779 scalar_replaceable = false; 780 } 781 } 782 } else { // Allocate instance 783 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 784 cik->is_subclass_of(_compile->env()->Reference_klass()) || 785 !cik->is_instance_klass() || // StressReflectiveCode 786 cik->as_instance_klass()->has_finalizer()) { 787 es = PointsToNode::GlobalEscape; 788 } 789 } 790 add_java_object(call, es); 791 PointsToNode* ptn = ptnode_adr(call_idx); 792 if (!scalar_replaceable && ptn->scalar_replaceable()) { 793 ptn->set_scalar_replaceable(false); 794 } 795 } else if (call->is_CallStaticJava()) { 796 // Call nodes could be different types: 797 // 798 // 1. CallDynamicJavaNode (what happened during call is unknown): 799 // 800 // - mapped to GlobalEscape JavaObject node if oop is returned; 801 // 802 // - all oop arguments are escaping globally; 803 // 804 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 805 // 806 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 807 // 808 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 809 // - mapped to NoEscape JavaObject node if non-escaping object allocated 810 // during call is returned; 811 // - mapped to ArgEscape LocalVar node pointed to object arguments 812 // which are returned and does not escape during call; 813 // 814 // - oop arguments escaping status is defined by bytecode analysis; 815 // 816 // For a static call, we know exactly what method is being called. 817 // Use bytecode estimator to record whether the call's return value escapes. 818 ciMethod* meth = call->as_CallJava()->method(); 819 if (meth == NULL) { 820 const char* name = call->as_CallStaticJava()->_name; 821 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 822 // Returns a newly allocated unescaped object. 823 add_java_object(call, PointsToNode::NoEscape); 824 ptnode_adr(call_idx)->set_scalar_replaceable(false); 825 } else if (meth->is_boxing_method()) { 826 // Returns boxing object 827 PointsToNode::EscapeState es; 828 vmIntrinsics::ID intr = meth->intrinsic_id(); 829 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 830 // It does not escape if object is always allocated. 831 es = PointsToNode::NoEscape; 832 } else { 833 // It escapes globally if object could be loaded from cache. 834 es = PointsToNode::GlobalEscape; 835 } 836 add_java_object(call, es); 837 } else { 838 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 839 call_analyzer->copy_dependencies(_compile->dependencies()); 840 if (call_analyzer->is_return_allocated()) { 841 // Returns a newly allocated unescaped object, simply 842 // update dependency information. 843 // Mark it as NoEscape so that objects referenced by 844 // it's fields will be marked as NoEscape at least. 845 add_java_object(call, PointsToNode::NoEscape); 846 ptnode_adr(call_idx)->set_scalar_replaceable(false); 847 } else { 848 // Determine whether any arguments are returned. 849 const TypeTuple* d = call->tf()->domain(); 850 bool ret_arg = false; 851 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 852 if (d->field_at(i)->isa_ptr() != NULL && 853 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 854 ret_arg = true; 855 break; 856 } 857 } 858 if (ret_arg) { 859 add_local_var(call, PointsToNode::ArgEscape); 860 } else { 861 // Returns unknown object. 862 map_ideal_node(call, phantom_obj); 863 } 864 } 865 } 866 } else { 867 // An other type of call, assume the worst case: 868 // returned value is unknown and globally escapes. 869 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 870 map_ideal_node(call, phantom_obj); 871 } 872 } 873 874 void ConnectionGraph::process_call_arguments(CallNode *call) { 875 bool is_arraycopy = false; 876 switch (call->Opcode()) { 877 #ifdef ASSERT 878 case Op_Allocate: 879 case Op_AllocateArray: 880 case Op_Lock: 881 case Op_Unlock: 882 assert(false, "should be done already"); 883 break; 884 #endif 885 case Op_CallLeafNoFP: 886 is_arraycopy = (call->as_CallLeaf()->_name != NULL && 887 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0); 888 // fall through 889 case Op_CallLeaf: { 890 // Stub calls, objects do not escape but they are not scale replaceable. 891 // Adjust escape state for outgoing arguments. 892 const TypeTuple * d = call->tf()->domain(); 893 bool src_has_oops = false; 894 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 895 const Type* at = d->field_at(i); 896 Node *arg = call->in(i); 897 const Type *aat = _igvn->type(arg); 898 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 899 continue; 900 if (arg->is_AddP()) { 901 // 902 // The inline_native_clone() case when the arraycopy stub is called 903 // after the allocation before Initialize and CheckCastPP nodes. 904 // Or normal arraycopy for object arrays case. 905 // 906 // Set AddP's base (Allocate) as not scalar replaceable since 907 // pointer to the base (with offset) is passed as argument. 908 // 909 arg = get_addp_base(arg); 910 } 911 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 912 assert(arg_ptn != NULL, "should be registered"); 913 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 914 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 915 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 916 aat->isa_ptr() != NULL, "expecting an Ptr"); 917 bool arg_has_oops = aat->isa_oopptr() && 918 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 919 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 920 if (i == TypeFunc::Parms) { 921 src_has_oops = arg_has_oops; 922 } 923 // 924 // src or dst could be j.l.Object when other is basic type array: 925 // 926 // arraycopy(char[],0,Object*,0,size); 927 // arraycopy(Object*,0,char[],0,size); 928 // 929 // Don't add edges in such cases. 930 // 931 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 932 arg_has_oops && (i > TypeFunc::Parms); 933 #ifdef ASSERT 934 if (!(is_arraycopy || 935 (call->as_CallLeaf()->_name != NULL && 936 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 937 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 || 938 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 939 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 940 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 941 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 942 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0) 943 ))) { 944 call->dump(); 945 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name)); 946 } 947 #endif 948 // Always process arraycopy's destination object since 949 // we need to add all possible edges to references in 950 // source object. 951 if (arg_esc >= PointsToNode::ArgEscape && 952 !arg_is_arraycopy_dest) { 953 continue; 954 } 955 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 956 if (arg_is_arraycopy_dest) { 957 Node* src = call->in(TypeFunc::Parms); 958 if (src->is_AddP()) { 959 src = get_addp_base(src); 960 } 961 PointsToNode* src_ptn = ptnode_adr(src->_idx); 962 assert(src_ptn != NULL, "should be registered"); 963 if (arg_ptn != src_ptn) { 964 // Special arraycopy edge: 965 // A destination object's field can't have the source object 966 // as base since objects escape states are not related. 967 // Only escape state of destination object's fields affects 968 // escape state of fields in source object. 969 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn); 970 } 971 } 972 } 973 } 974 break; 975 } 976 case Op_CallStaticJava: { 977 // For a static call, we know exactly what method is being called. 978 // Use bytecode estimator to record the call's escape affects 979 #ifdef ASSERT 980 const char* name = call->as_CallStaticJava()->_name; 981 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 982 #endif 983 ciMethod* meth = call->as_CallJava()->method(); 984 if ((meth != NULL) && meth->is_boxing_method()) { 985 break; // Boxing methods do not modify any oops. 986 } 987 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 988 // fall-through if not a Java method or no analyzer information 989 if (call_analyzer != NULL) { 990 PointsToNode* call_ptn = ptnode_adr(call->_idx); 991 const TypeTuple* d = call->tf()->domain(); 992 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 993 const Type* at = d->field_at(i); 994 int k = i - TypeFunc::Parms; 995 Node* arg = call->in(i); 996 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 997 if (at->isa_ptr() != NULL && 998 call_analyzer->is_arg_returned(k)) { 999 // The call returns arguments. 1000 if (call_ptn != NULL) { // Is call's result used? 1001 assert(call_ptn->is_LocalVar(), "node should be registered"); 1002 assert(arg_ptn != NULL, "node should be registered"); 1003 add_edge(call_ptn, arg_ptn); 1004 } 1005 } 1006 if (at->isa_oopptr() != NULL && 1007 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1008 if (!call_analyzer->is_arg_stack(k)) { 1009 // The argument global escapes 1010 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1011 } else { 1012 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1013 if (!call_analyzer->is_arg_local(k)) { 1014 // The argument itself doesn't escape, but any fields might 1015 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1016 } 1017 } 1018 } 1019 } 1020 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1021 // The call returns arguments. 1022 assert(call_ptn->edge_count() > 0, "sanity"); 1023 if (!call_analyzer->is_return_local()) { 1024 // Returns also unknown object. 1025 add_edge(call_ptn, phantom_obj); 1026 } 1027 } 1028 break; 1029 } 1030 } 1031 default: { 1032 // Fall-through here if not a Java method or no analyzer information 1033 // or some other type of call, assume the worst case: all arguments 1034 // globally escape. 1035 const TypeTuple* d = call->tf()->domain(); 1036 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1037 const Type* at = d->field_at(i); 1038 if (at->isa_oopptr() != NULL) { 1039 Node* arg = call->in(i); 1040 if (arg->is_AddP()) { 1041 arg = get_addp_base(arg); 1042 } 1043 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1044 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1045 } 1046 } 1047 } 1048 } 1049 } 1050 1051 1052 // Finish Graph construction. 1053 bool ConnectionGraph::complete_connection_graph( 1054 GrowableArray<PointsToNode*>& ptnodes_worklist, 1055 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1056 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1057 GrowableArray<FieldNode*>& oop_fields_worklist) { 1058 // Normally only 1-3 passes needed to build Connection Graph depending 1059 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1060 // Set limit to 20 to catch situation when something did go wrong and 1061 // bailout Escape Analysis. 1062 // Also limit build time to 30 sec (60 in debug VM). 1063 #define CG_BUILD_ITER_LIMIT 20 1064 #ifdef ASSERT 1065 #define CG_BUILD_TIME_LIMIT 60.0 1066 #else 1067 #define CG_BUILD_TIME_LIMIT 30.0 1068 #endif 1069 1070 // Propagate GlobalEscape and ArgEscape escape states and check that 1071 // we still have non-escaping objects. The method pushs on _worklist 1072 // Field nodes which reference phantom_object. 1073 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1074 return false; // Nothing to do. 1075 } 1076 // Now propagate references to all JavaObject nodes. 1077 int java_objects_length = java_objects_worklist.length(); 1078 elapsedTimer time; 1079 int new_edges = 1; 1080 int iterations = 0; 1081 do { 1082 while ((new_edges > 0) && 1083 (iterations++ < CG_BUILD_ITER_LIMIT) && 1084 (time.seconds() < CG_BUILD_TIME_LIMIT)) { 1085 time.start(); 1086 new_edges = 0; 1087 // Propagate references to phantom_object for nodes pushed on _worklist 1088 // by find_non_escaped_objects() and find_field_value(). 1089 new_edges += add_java_object_edges(phantom_obj, false); 1090 for (int next = 0; next < java_objects_length; ++next) { 1091 JavaObjectNode* ptn = java_objects_worklist.at(next); 1092 new_edges += add_java_object_edges(ptn, true); 1093 } 1094 if (new_edges > 0) { 1095 // Update escape states on each iteration if graph was updated. 1096 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1097 return false; // Nothing to do. 1098 } 1099 } 1100 time.stop(); 1101 } 1102 if ((iterations < CG_BUILD_ITER_LIMIT) && 1103 (time.seconds() < CG_BUILD_TIME_LIMIT)) { 1104 time.start(); 1105 // Find fields which have unknown value. 1106 int fields_length = oop_fields_worklist.length(); 1107 for (int next = 0; next < fields_length; next++) { 1108 FieldNode* field = oop_fields_worklist.at(next); 1109 if (field->edge_count() == 0) { 1110 new_edges += find_field_value(field); 1111 // This code may added new edges to phantom_object. 1112 // Need an other cycle to propagate references to phantom_object. 1113 } 1114 } 1115 time.stop(); 1116 } else { 1117 new_edges = 0; // Bailout 1118 } 1119 } while (new_edges > 0); 1120 1121 // Bailout if passed limits. 1122 if ((iterations >= CG_BUILD_ITER_LIMIT) || 1123 (time.seconds() >= CG_BUILD_TIME_LIMIT)) { 1124 Compile* C = _compile; 1125 if (C->log() != NULL) { 1126 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1127 C->log()->text("%s", (iterations >= CG_BUILD_ITER_LIMIT) ? "iterations" : "time"); 1128 C->log()->end_elem(" limit'"); 1129 } 1130 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1131 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length())); 1132 // Possible infinite build_connection_graph loop, 1133 // bailout (no changes to ideal graph were made). 1134 return false; 1135 } 1136 #ifdef ASSERT 1137 if (Verbose && PrintEscapeAnalysis) { 1138 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1139 iterations, nodes_size(), ptnodes_worklist.length()); 1140 } 1141 #endif 1142 1143 #undef CG_BUILD_ITER_LIMIT 1144 #undef CG_BUILD_TIME_LIMIT 1145 1146 // Find fields initialized by NULL for non-escaping Allocations. 1147 int non_escaped_length = non_escaped_worklist.length(); 1148 for (int next = 0; next < non_escaped_length; next++) { 1149 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1150 PointsToNode::EscapeState es = ptn->escape_state(); 1151 assert(es <= PointsToNode::ArgEscape, "sanity"); 1152 if (es == PointsToNode::NoEscape) { 1153 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1154 // Adding references to NULL object does not change escape states 1155 // since it does not escape. Also no fields are added to NULL object. 1156 add_java_object_edges(null_obj, false); 1157 } 1158 } 1159 Node* n = ptn->ideal_node(); 1160 if (n->is_Allocate()) { 1161 // The object allocated by this Allocate node will never be 1162 // seen by an other thread. Mark it so that when it is 1163 // expanded no MemBarStoreStore is added. 1164 InitializeNode* ini = n->as_Allocate()->initialization(); 1165 if (ini != NULL) 1166 ini->set_does_not_escape(); 1167 } 1168 } 1169 return true; // Finished graph construction. 1170 } 1171 1172 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1173 // and check that we still have non-escaping java objects. 1174 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1175 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1176 GrowableArray<PointsToNode*> escape_worklist; 1177 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1178 int ptnodes_length = ptnodes_worklist.length(); 1179 for (int next = 0; next < ptnodes_length; ++next) { 1180 PointsToNode* ptn = ptnodes_worklist.at(next); 1181 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1182 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1183 escape_worklist.push(ptn); 1184 } 1185 } 1186 // Set escape states to referenced nodes (edges list). 1187 while (escape_worklist.length() > 0) { 1188 PointsToNode* ptn = escape_worklist.pop(); 1189 PointsToNode::EscapeState es = ptn->escape_state(); 1190 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1191 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1192 es >= PointsToNode::ArgEscape) { 1193 // GlobalEscape or ArgEscape state of field means it has unknown value. 1194 if (add_edge(ptn, phantom_obj)) { 1195 // New edge was added 1196 add_field_uses_to_worklist(ptn->as_Field()); 1197 } 1198 } 1199 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1200 PointsToNode* e = i.get(); 1201 if (e->is_Arraycopy()) { 1202 assert(ptn->arraycopy_dst(), "sanity"); 1203 // Propagate only fields escape state through arraycopy edge. 1204 if (e->fields_escape_state() < field_es) { 1205 set_fields_escape_state(e, field_es); 1206 escape_worklist.push(e); 1207 } 1208 } else if (es >= field_es) { 1209 // fields_escape_state is also set to 'es' if it is less than 'es'. 1210 if (e->escape_state() < es) { 1211 set_escape_state(e, es); 1212 escape_worklist.push(e); 1213 } 1214 } else { 1215 // Propagate field escape state. 1216 bool es_changed = false; 1217 if (e->fields_escape_state() < field_es) { 1218 set_fields_escape_state(e, field_es); 1219 es_changed = true; 1220 } 1221 if ((e->escape_state() < field_es) && 1222 e->is_Field() && ptn->is_JavaObject() && 1223 e->as_Field()->is_oop()) { 1224 // Change escape state of referenced fileds. 1225 set_escape_state(e, field_es); 1226 es_changed = true;; 1227 } else if (e->escape_state() < es) { 1228 set_escape_state(e, es); 1229 es_changed = true;; 1230 } 1231 if (es_changed) { 1232 escape_worklist.push(e); 1233 } 1234 } 1235 } 1236 } 1237 // Remove escaped objects from non_escaped list. 1238 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1239 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1240 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1241 non_escaped_worklist.delete_at(next); 1242 } 1243 if (ptn->escape_state() == PointsToNode::NoEscape) { 1244 // Find fields in non-escaped allocations which have unknown value. 1245 find_init_values(ptn, phantom_obj, NULL); 1246 } 1247 } 1248 return (non_escaped_worklist.length() > 0); 1249 } 1250 1251 // Add all references to JavaObject node by walking over all uses. 1252 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1253 int new_edges = 0; 1254 if (populate_worklist) { 1255 // Populate _worklist by uses of jobj's uses. 1256 for (UseIterator i(jobj); i.has_next(); i.next()) { 1257 PointsToNode* use = i.get(); 1258 if (use->is_Arraycopy()) 1259 continue; 1260 add_uses_to_worklist(use); 1261 if (use->is_Field() && use->as_Field()->is_oop()) { 1262 // Put on worklist all field's uses (loads) and 1263 // related field nodes (same base and offset). 1264 add_field_uses_to_worklist(use->as_Field()); 1265 } 1266 } 1267 } 1268 while(_worklist.length() > 0) { 1269 PointsToNode* use = _worklist.pop(); 1270 if (PointsToNode::is_base_use(use)) { 1271 // Add reference from jobj to field and from field to jobj (field's base). 1272 use = PointsToNode::get_use_node(use)->as_Field(); 1273 if (add_base(use->as_Field(), jobj)) { 1274 new_edges++; 1275 } 1276 continue; 1277 } 1278 assert(!use->is_JavaObject(), "sanity"); 1279 if (use->is_Arraycopy()) { 1280 if (jobj == null_obj) // NULL object does not have field edges 1281 continue; 1282 // Added edge from Arraycopy node to arraycopy's source java object 1283 if (add_edge(use, jobj)) { 1284 jobj->set_arraycopy_src(); 1285 new_edges++; 1286 } 1287 // and stop here. 1288 continue; 1289 } 1290 if (!add_edge(use, jobj)) 1291 continue; // No new edge added, there was such edge already. 1292 new_edges++; 1293 if (use->is_LocalVar()) { 1294 add_uses_to_worklist(use); 1295 if (use->arraycopy_dst()) { 1296 for (EdgeIterator i(use); i.has_next(); i.next()) { 1297 PointsToNode* e = i.get(); 1298 if (e->is_Arraycopy()) { 1299 if (jobj == null_obj) // NULL object does not have field edges 1300 continue; 1301 // Add edge from arraycopy's destination java object to Arraycopy node. 1302 if (add_edge(jobj, e)) { 1303 new_edges++; 1304 jobj->set_arraycopy_dst(); 1305 } 1306 } 1307 } 1308 } 1309 } else { 1310 // Added new edge to stored in field values. 1311 // Put on worklist all field's uses (loads) and 1312 // related field nodes (same base and offset). 1313 add_field_uses_to_worklist(use->as_Field()); 1314 } 1315 } 1316 return new_edges; 1317 } 1318 1319 // Put on worklist all related field nodes. 1320 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1321 assert(field->is_oop(), "sanity"); 1322 int offset = field->offset(); 1323 add_uses_to_worklist(field); 1324 // Loop over all bases of this field and push on worklist Field nodes 1325 // with the same offset and base (since they may reference the same field). 1326 for (BaseIterator i(field); i.has_next(); i.next()) { 1327 PointsToNode* base = i.get(); 1328 add_fields_to_worklist(field, base); 1329 // Check if the base was source object of arraycopy and go over arraycopy's 1330 // destination objects since values stored to a field of source object are 1331 // accessable by uses (loads) of fields of destination objects. 1332 if (base->arraycopy_src()) { 1333 for (UseIterator j(base); j.has_next(); j.next()) { 1334 PointsToNode* arycp = j.get(); 1335 if (arycp->is_Arraycopy()) { 1336 for (UseIterator k(arycp); k.has_next(); k.next()) { 1337 PointsToNode* abase = k.get(); 1338 if (abase->arraycopy_dst() && abase != base) { 1339 // Look for the same arracopy reference. 1340 add_fields_to_worklist(field, abase); 1341 } 1342 } 1343 } 1344 } 1345 } 1346 } 1347 } 1348 1349 // Put on worklist all related field nodes. 1350 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1351 int offset = field->offset(); 1352 if (base->is_LocalVar()) { 1353 for (UseIterator j(base); j.has_next(); j.next()) { 1354 PointsToNode* f = j.get(); 1355 if (PointsToNode::is_base_use(f)) { // Field 1356 f = PointsToNode::get_use_node(f); 1357 if (f == field || !f->as_Field()->is_oop()) 1358 continue; 1359 int offs = f->as_Field()->offset(); 1360 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1361 add_to_worklist(f); 1362 } 1363 } 1364 } 1365 } else { 1366 assert(base->is_JavaObject(), "sanity"); 1367 if (// Skip phantom_object since it is only used to indicate that 1368 // this field's content globally escapes. 1369 (base != phantom_obj) && 1370 // NULL object node does not have fields. 1371 (base != null_obj)) { 1372 for (EdgeIterator i(base); i.has_next(); i.next()) { 1373 PointsToNode* f = i.get(); 1374 // Skip arraycopy edge since store to destination object field 1375 // does not update value in source object field. 1376 if (f->is_Arraycopy()) { 1377 assert(base->arraycopy_dst(), "sanity"); 1378 continue; 1379 } 1380 if (f == field || !f->as_Field()->is_oop()) 1381 continue; 1382 int offs = f->as_Field()->offset(); 1383 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1384 add_to_worklist(f); 1385 } 1386 } 1387 } 1388 } 1389 } 1390 1391 // Find fields which have unknown value. 1392 int ConnectionGraph::find_field_value(FieldNode* field) { 1393 // Escaped fields should have init value already. 1394 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1395 int new_edges = 0; 1396 for (BaseIterator i(field); i.has_next(); i.next()) { 1397 PointsToNode* base = i.get(); 1398 if (base->is_JavaObject()) { 1399 // Skip Allocate's fields which will be processed later. 1400 if (base->ideal_node()->is_Allocate()) 1401 return 0; 1402 assert(base == null_obj, "only NULL ptr base expected here"); 1403 } 1404 } 1405 if (add_edge(field, phantom_obj)) { 1406 // New edge was added 1407 new_edges++; 1408 add_field_uses_to_worklist(field); 1409 } 1410 return new_edges; 1411 } 1412 1413 // Find fields initializing values for allocations. 1414 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1415 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1416 int new_edges = 0; 1417 Node* alloc = pta->ideal_node(); 1418 if (init_val == phantom_obj) { 1419 // Do nothing for Allocate nodes since its fields values are "known". 1420 if (alloc->is_Allocate()) 1421 return 0; 1422 assert(alloc->as_CallStaticJava(), "sanity"); 1423 #ifdef ASSERT 1424 if (alloc->as_CallStaticJava()->method() == NULL) { 1425 const char* name = alloc->as_CallStaticJava()->_name; 1426 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1427 } 1428 #endif 1429 // Non-escaped allocation returned from Java or runtime call have 1430 // unknown values in fields. 1431 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1432 PointsToNode* field = i.get(); 1433 if (field->is_Field() && field->as_Field()->is_oop()) { 1434 if (add_edge(field, phantom_obj)) { 1435 // New edge was added 1436 new_edges++; 1437 add_field_uses_to_worklist(field->as_Field()); 1438 } 1439 } 1440 } 1441 return new_edges; 1442 } 1443 assert(init_val == null_obj, "sanity"); 1444 // Do nothing for Call nodes since its fields values are unknown. 1445 if (!alloc->is_Allocate()) 1446 return 0; 1447 1448 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1449 Compile* C = _compile; 1450 bool visited_bottom_offset = false; 1451 GrowableArray<int> offsets_worklist; 1452 1453 // Check if an oop field's initializing value is recorded and add 1454 // a corresponding NULL if field's value if it is not recorded. 1455 // Connection Graph does not record a default initialization by NULL 1456 // captured by Initialize node. 1457 // 1458 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1459 PointsToNode* field = i.get(); // Field (AddP) 1460 if (!field->is_Field() || !field->as_Field()->is_oop()) 1461 continue; // Not oop field 1462 int offset = field->as_Field()->offset(); 1463 if (offset == Type::OffsetBot) { 1464 if (!visited_bottom_offset) { 1465 // OffsetBot is used to reference array's element, 1466 // always add reference to NULL to all Field nodes since we don't 1467 // known which element is referenced. 1468 if (add_edge(field, null_obj)) { 1469 // New edge was added 1470 new_edges++; 1471 add_field_uses_to_worklist(field->as_Field()); 1472 visited_bottom_offset = true; 1473 } 1474 } 1475 } else { 1476 // Check only oop fields. 1477 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1478 if (adr_type->isa_rawptr()) { 1479 #ifdef ASSERT 1480 // Raw pointers are used for initializing stores so skip it 1481 // since it should be recorded already 1482 Node* base = get_addp_base(field->ideal_node()); 1483 assert(adr_type->isa_rawptr() && base->is_Proj() && 1484 (base->in(0) == alloc),"unexpected pointer type"); 1485 #endif 1486 continue; 1487 } 1488 if (!offsets_worklist.contains(offset)) { 1489 offsets_worklist.append(offset); 1490 Node* value = NULL; 1491 if (ini != NULL) { 1492 // StoreP::memory_type() == T_ADDRESS 1493 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1494 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1495 // Make sure initializing store has the same type as this AddP. 1496 // This AddP may reference non existing field because it is on a 1497 // dead branch of bimorphic call which is not eliminated yet. 1498 if (store != NULL && store->is_Store() && 1499 store->as_Store()->memory_type() == ft) { 1500 value = store->in(MemNode::ValueIn); 1501 #ifdef ASSERT 1502 if (VerifyConnectionGraph) { 1503 // Verify that AddP already points to all objects the value points to. 1504 PointsToNode* val = ptnode_adr(value->_idx); 1505 assert((val != NULL), "should be processed already"); 1506 PointsToNode* missed_obj = NULL; 1507 if (val->is_JavaObject()) { 1508 if (!field->points_to(val->as_JavaObject())) { 1509 missed_obj = val; 1510 } 1511 } else { 1512 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1513 tty->print_cr("----------init store has invalid value -----"); 1514 store->dump(); 1515 val->dump(); 1516 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1517 } 1518 for (EdgeIterator j(val); j.has_next(); j.next()) { 1519 PointsToNode* obj = j.get(); 1520 if (obj->is_JavaObject()) { 1521 if (!field->points_to(obj->as_JavaObject())) { 1522 missed_obj = obj; 1523 break; 1524 } 1525 } 1526 } 1527 } 1528 if (missed_obj != NULL) { 1529 tty->print_cr("----------field---------------------------------"); 1530 field->dump(); 1531 tty->print_cr("----------missed referernce to object-----------"); 1532 missed_obj->dump(); 1533 tty->print_cr("----------object referernced by init store -----"); 1534 store->dump(); 1535 val->dump(); 1536 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1537 } 1538 } 1539 #endif 1540 } else { 1541 // There could be initializing stores which follow allocation. 1542 // For example, a volatile field store is not collected 1543 // by Initialize node. 1544 // 1545 // Need to check for dependent loads to separate such stores from 1546 // stores which follow loads. For now, add initial value NULL so 1547 // that compare pointers optimization works correctly. 1548 } 1549 } 1550 if (value == NULL) { 1551 // A field's initializing value was not recorded. Add NULL. 1552 if (add_edge(field, null_obj)) { 1553 // New edge was added 1554 new_edges++; 1555 add_field_uses_to_worklist(field->as_Field()); 1556 } 1557 } 1558 } 1559 } 1560 } 1561 return new_edges; 1562 } 1563 1564 // Adjust scalar_replaceable state after Connection Graph is built. 1565 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1566 // Search for non-escaping objects which are not scalar replaceable 1567 // and mark them to propagate the state to referenced objects. 1568 1569 // 1. An object is not scalar replaceable if the field into which it is 1570 // stored has unknown offset (stored into unknown element of an array). 1571 // 1572 for (UseIterator i(jobj); i.has_next(); i.next()) { 1573 PointsToNode* use = i.get(); 1574 assert(!use->is_Arraycopy(), "sanity"); 1575 if (use->is_Field()) { 1576 FieldNode* field = use->as_Field(); 1577 assert(field->is_oop() && field->scalar_replaceable() && 1578 field->fields_escape_state() == PointsToNode::NoEscape, "sanity"); 1579 if (field->offset() == Type::OffsetBot) { 1580 jobj->set_scalar_replaceable(false); 1581 return; 1582 } 1583 // 2. An object is not scalar replaceable if the field into which it is 1584 // stored has multiple bases one of which is null. 1585 if (field->base_count() > 1) { 1586 for (BaseIterator i(field); i.has_next(); i.next()) { 1587 PointsToNode* base = i.get(); 1588 if (base == null_obj) { 1589 jobj->set_scalar_replaceable(false); 1590 return; 1591 } 1592 } 1593 } 1594 } 1595 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1596 // 3. An object is not scalar replaceable if it is merged with other objects. 1597 for (EdgeIterator j(use); j.has_next(); j.next()) { 1598 PointsToNode* ptn = j.get(); 1599 if (ptn->is_JavaObject() && ptn != jobj) { 1600 // Mark all objects. 1601 jobj->set_scalar_replaceable(false); 1602 ptn->set_scalar_replaceable(false); 1603 } 1604 } 1605 if (!jobj->scalar_replaceable()) { 1606 return; 1607 } 1608 } 1609 1610 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1611 // Non-escaping object node should point only to field nodes. 1612 FieldNode* field = j.get()->as_Field(); 1613 int offset = field->as_Field()->offset(); 1614 1615 // 4. An object is not scalar replaceable if it has a field with unknown 1616 // offset (array's element is accessed in loop). 1617 if (offset == Type::OffsetBot) { 1618 jobj->set_scalar_replaceable(false); 1619 return; 1620 } 1621 // 5. Currently an object is not scalar replaceable if a LoadStore node 1622 // access its field since the field value is unknown after it. 1623 // 1624 Node* n = field->ideal_node(); 1625 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1626 if (n->fast_out(i)->is_LoadStore()) { 1627 jobj->set_scalar_replaceable(false); 1628 return; 1629 } 1630 } 1631 1632 // 6. Or the address may point to more then one object. This may produce 1633 // the false positive result (set not scalar replaceable) 1634 // since the flow-insensitive escape analysis can't separate 1635 // the case when stores overwrite the field's value from the case 1636 // when stores happened on different control branches. 1637 // 1638 // Note: it will disable scalar replacement in some cases: 1639 // 1640 // Point p[] = new Point[1]; 1641 // p[0] = new Point(); // Will be not scalar replaced 1642 // 1643 // but it will save us from incorrect optimizations in next cases: 1644 // 1645 // Point p[] = new Point[1]; 1646 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1647 // 1648 if (field->base_count() > 1) { 1649 for (BaseIterator i(field); i.has_next(); i.next()) { 1650 PointsToNode* base = i.get(); 1651 // Don't take into account LocalVar nodes which 1652 // may point to only one object which should be also 1653 // this field's base by now. 1654 if (base->is_JavaObject() && base != jobj) { 1655 // Mark all bases. 1656 jobj->set_scalar_replaceable(false); 1657 base->set_scalar_replaceable(false); 1658 } 1659 } 1660 } 1661 } 1662 } 1663 1664 #ifdef ASSERT 1665 void ConnectionGraph::verify_connection_graph( 1666 GrowableArray<PointsToNode*>& ptnodes_worklist, 1667 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1668 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1669 GrowableArray<Node*>& addp_worklist) { 1670 // Verify that graph is complete - no new edges could be added. 1671 int java_objects_length = java_objects_worklist.length(); 1672 int non_escaped_length = non_escaped_worklist.length(); 1673 int new_edges = 0; 1674 for (int next = 0; next < java_objects_length; ++next) { 1675 JavaObjectNode* ptn = java_objects_worklist.at(next); 1676 new_edges += add_java_object_edges(ptn, true); 1677 } 1678 assert(new_edges == 0, "graph was not complete"); 1679 // Verify that escape state is final. 1680 int length = non_escaped_worklist.length(); 1681 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1682 assert((non_escaped_length == non_escaped_worklist.length()) && 1683 (non_escaped_length == length) && 1684 (_worklist.length() == 0), "escape state was not final"); 1685 1686 // Verify fields information. 1687 int addp_length = addp_worklist.length(); 1688 for (int next = 0; next < addp_length; ++next ) { 1689 Node* n = addp_worklist.at(next); 1690 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1691 if (field->is_oop()) { 1692 // Verify that field has all bases 1693 Node* base = get_addp_base(n); 1694 PointsToNode* ptn = ptnode_adr(base->_idx); 1695 if (ptn->is_JavaObject()) { 1696 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1697 } else { 1698 assert(ptn->is_LocalVar(), "sanity"); 1699 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1700 PointsToNode* e = i.get(); 1701 if (e->is_JavaObject()) { 1702 assert(field->has_base(e->as_JavaObject()), "sanity"); 1703 } 1704 } 1705 } 1706 // Verify that all fields have initializing values. 1707 if (field->edge_count() == 0) { 1708 tty->print_cr("----------field does not have references----------"); 1709 field->dump(); 1710 for (BaseIterator i(field); i.has_next(); i.next()) { 1711 PointsToNode* base = i.get(); 1712 tty->print_cr("----------field has next base---------------------"); 1713 base->dump(); 1714 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1715 tty->print_cr("----------base has fields-------------------------"); 1716 for (EdgeIterator j(base); j.has_next(); j.next()) { 1717 j.get()->dump(); 1718 } 1719 tty->print_cr("----------base has references---------------------"); 1720 for (UseIterator j(base); j.has_next(); j.next()) { 1721 j.get()->dump(); 1722 } 1723 } 1724 } 1725 for (UseIterator i(field); i.has_next(); i.next()) { 1726 i.get()->dump(); 1727 } 1728 assert(field->edge_count() > 0, "sanity"); 1729 } 1730 } 1731 } 1732 } 1733 #endif 1734 1735 // Optimize ideal graph. 1736 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1737 GrowableArray<Node*>& storestore_worklist) { 1738 Compile* C = _compile; 1739 PhaseIterGVN* igvn = _igvn; 1740 if (EliminateLocks) { 1741 // Mark locks before changing ideal graph. 1742 int cnt = C->macro_count(); 1743 for( int i=0; i < cnt; i++ ) { 1744 Node *n = C->macro_node(i); 1745 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1746 AbstractLockNode* alock = n->as_AbstractLock(); 1747 if (!alock->is_non_esc_obj()) { 1748 if (not_global_escape(alock->obj_node())) { 1749 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1750 // The lock could be marked eliminated by lock coarsening 1751 // code during first IGVN before EA. Replace coarsened flag 1752 // to eliminate all associated locks/unlocks. 1753 alock->set_non_esc_obj(); 1754 } 1755 } 1756 } 1757 } 1758 } 1759 1760 if (OptimizePtrCompare) { 1761 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1762 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1763 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1764 // Optimize objects compare. 1765 while (ptr_cmp_worklist.length() != 0) { 1766 Node *n = ptr_cmp_worklist.pop(); 1767 Node *res = optimize_ptr_compare(n); 1768 if (res != NULL) { 1769 #ifndef PRODUCT 1770 if (PrintOptimizePtrCompare) { 1771 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")); 1772 if (Verbose) { 1773 n->dump(1); 1774 } 1775 } 1776 #endif 1777 igvn->replace_node(n, res); 1778 } 1779 } 1780 // cleanup 1781 if (_pcmp_neq->outcnt() == 0) 1782 igvn->hash_delete(_pcmp_neq); 1783 if (_pcmp_eq->outcnt() == 0) 1784 igvn->hash_delete(_pcmp_eq); 1785 } 1786 1787 // For MemBarStoreStore nodes added in library_call.cpp, check 1788 // escape status of associated AllocateNode and optimize out 1789 // MemBarStoreStore node if the allocated object never escapes. 1790 while (storestore_worklist.length() != 0) { 1791 Node *n = storestore_worklist.pop(); 1792 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1793 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1794 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1795 if (not_global_escape(alloc)) { 1796 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1797 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1798 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1799 igvn->register_new_node_with_optimizer(mb); 1800 igvn->replace_node(storestore, mb); 1801 } 1802 } 1803 } 1804 1805 // Optimize objects compare. 1806 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1807 assert(OptimizePtrCompare, "sanity"); 1808 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1809 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1810 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1811 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1812 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1813 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1814 1815 // Check simple cases first. 1816 if (jobj1 != NULL) { 1817 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1818 if (jobj1 == jobj2) { 1819 // Comparing the same not escaping object. 1820 return _pcmp_eq; 1821 } 1822 Node* obj = jobj1->ideal_node(); 1823 // Comparing not escaping allocation. 1824 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1825 !ptn2->points_to(jobj1)) { 1826 return _pcmp_neq; // This includes nullness check. 1827 } 1828 } 1829 } 1830 if (jobj2 != NULL) { 1831 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1832 Node* obj = jobj2->ideal_node(); 1833 // Comparing not escaping allocation. 1834 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1835 !ptn1->points_to(jobj2)) { 1836 return _pcmp_neq; // This includes nullness check. 1837 } 1838 } 1839 } 1840 if (jobj1 != NULL && jobj1 != phantom_obj && 1841 jobj2 != NULL && jobj2 != phantom_obj && 1842 jobj1->ideal_node()->is_Con() && 1843 jobj2->ideal_node()->is_Con()) { 1844 // Klass or String constants compare. Need to be careful with 1845 // compressed pointers - compare types of ConN and ConP instead of nodes. 1846 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1847 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1848 if (t1->make_ptr() == t2->make_ptr()) { 1849 return _pcmp_eq; 1850 } else { 1851 return _pcmp_neq; 1852 } 1853 } 1854 if (ptn1->meet(ptn2)) { 1855 return NULL; // Sets are not disjoint 1856 } 1857 1858 // Sets are disjoint. 1859 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1860 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1861 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1862 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1863 if (set1_has_unknown_ptr && set2_has_null_ptr || 1864 set2_has_unknown_ptr && set1_has_null_ptr) { 1865 // Check nullness of unknown object. 1866 return NULL; 1867 } 1868 1869 // Disjointness by itself is not sufficient since 1870 // alias analysis is not complete for escaped objects. 1871 // Disjoint sets are definitely unrelated only when 1872 // at least one set has only not escaping allocations. 1873 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1874 if (ptn1->non_escaping_allocation()) { 1875 return _pcmp_neq; 1876 } 1877 } 1878 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 1879 if (ptn2->non_escaping_allocation()) { 1880 return _pcmp_neq; 1881 } 1882 } 1883 return NULL; 1884 } 1885 1886 // Connection Graph constuction functions. 1887 1888 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 1889 PointsToNode* ptadr = _nodes.at(n->_idx); 1890 if (ptadr != NULL) { 1891 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 1892 return; 1893 } 1894 Compile* C = _compile; 1895 ptadr = new (C->comp_arena()) LocalVarNode(C, n, es); 1896 _nodes.at_put(n->_idx, ptadr); 1897 } 1898 1899 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 1900 PointsToNode* ptadr = _nodes.at(n->_idx); 1901 if (ptadr != NULL) { 1902 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 1903 return; 1904 } 1905 Compile* C = _compile; 1906 ptadr = new (C->comp_arena()) JavaObjectNode(C, n, es); 1907 _nodes.at_put(n->_idx, ptadr); 1908 } 1909 1910 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 1911 PointsToNode* ptadr = _nodes.at(n->_idx); 1912 if (ptadr != NULL) { 1913 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 1914 return; 1915 } 1916 bool unsafe = false; 1917 bool is_oop = is_oop_field(n, offset, &unsafe); 1918 if (unsafe) { 1919 es = PointsToNode::GlobalEscape; 1920 } 1921 Compile* C = _compile; 1922 FieldNode* field = new (C->comp_arena()) FieldNode(C, n, es, offset, is_oop); 1923 _nodes.at_put(n->_idx, field); 1924 } 1925 1926 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 1927 PointsToNode* src, PointsToNode* dst) { 1928 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 1929 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 1930 PointsToNode* ptadr = _nodes.at(n->_idx); 1931 if (ptadr != NULL) { 1932 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 1933 return; 1934 } 1935 Compile* C = _compile; 1936 ptadr = new (C->comp_arena()) ArraycopyNode(C, n, es); 1937 _nodes.at_put(n->_idx, ptadr); 1938 // Add edge from arraycopy node to source object. 1939 (void)add_edge(ptadr, src); 1940 src->set_arraycopy_src(); 1941 // Add edge from destination object to arraycopy node. 1942 (void)add_edge(dst, ptadr); 1943 dst->set_arraycopy_dst(); 1944 } 1945 1946 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 1947 const Type* adr_type = n->as_AddP()->bottom_type(); 1948 BasicType bt = T_INT; 1949 if (offset == Type::OffsetBot) { 1950 // Check only oop fields. 1951 if (!adr_type->isa_aryptr() || 1952 (adr_type->isa_aryptr()->klass() == NULL) || 1953 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 1954 // OffsetBot is used to reference array's element. Ignore first AddP. 1955 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 1956 bt = T_OBJECT; 1957 } 1958 } 1959 } else if (offset != oopDesc::klass_offset_in_bytes()) { 1960 if (adr_type->isa_instptr()) { 1961 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 1962 if (field != NULL) { 1963 bt = field->layout_type(); 1964 } else { 1965 // Check for unsafe oop field access 1966 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1967 int opcode = n->fast_out(i)->Opcode(); 1968 if (opcode == Op_StoreP || opcode == Op_LoadP || 1969 opcode == Op_StoreN || opcode == Op_LoadN) { 1970 bt = T_OBJECT; 1971 (*unsafe) = true; 1972 break; 1973 } 1974 } 1975 } 1976 } else if (adr_type->isa_aryptr()) { 1977 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1978 // Ignore array length load. 1979 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 1980 // Ignore first AddP. 1981 } else { 1982 const Type* elemtype = adr_type->isa_aryptr()->elem(); 1983 bt = elemtype->array_element_basic_type(); 1984 } 1985 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 1986 // Allocation initialization, ThreadLocal field access, unsafe access 1987 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1988 int opcode = n->fast_out(i)->Opcode(); 1989 if (opcode == Op_StoreP || opcode == Op_LoadP || 1990 opcode == Op_StoreN || opcode == Op_LoadN) { 1991 bt = T_OBJECT; 1992 break; 1993 } 1994 } 1995 } 1996 } 1997 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 1998 } 1999 2000 // Returns unique pointed java object or NULL. 2001 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2002 assert(!_collecting, "should not call when contructed graph"); 2003 // If the node was created after the escape computation we can't answer. 2004 uint idx = n->_idx; 2005 if (idx >= nodes_size()) { 2006 return NULL; 2007 } 2008 PointsToNode* ptn = ptnode_adr(idx); 2009 if (ptn->is_JavaObject()) { 2010 return ptn->as_JavaObject(); 2011 } 2012 assert(ptn->is_LocalVar(), "sanity"); 2013 // Check all java objects it points to. 2014 JavaObjectNode* jobj = NULL; 2015 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2016 PointsToNode* e = i.get(); 2017 if (e->is_JavaObject()) { 2018 if (jobj == NULL) { 2019 jobj = e->as_JavaObject(); 2020 } else if (jobj != e) { 2021 return NULL; 2022 } 2023 } 2024 } 2025 return jobj; 2026 } 2027 2028 // Return true if this node points only to non-escaping allocations. 2029 bool PointsToNode::non_escaping_allocation() { 2030 if (is_JavaObject()) { 2031 Node* n = ideal_node(); 2032 if (n->is_Allocate() || n->is_CallStaticJava()) { 2033 return (escape_state() == PointsToNode::NoEscape); 2034 } else { 2035 return false; 2036 } 2037 } 2038 assert(is_LocalVar(), "sanity"); 2039 // Check all java objects it points to. 2040 for (EdgeIterator i(this); i.has_next(); i.next()) { 2041 PointsToNode* e = i.get(); 2042 if (e->is_JavaObject()) { 2043 Node* n = e->ideal_node(); 2044 if ((e->escape_state() != PointsToNode::NoEscape) || 2045 !(n->is_Allocate() || n->is_CallStaticJava())) { 2046 return false; 2047 } 2048 } 2049 } 2050 return true; 2051 } 2052 2053 // Return true if we know the node does not escape globally. 2054 bool ConnectionGraph::not_global_escape(Node *n) { 2055 assert(!_collecting, "should not call during graph construction"); 2056 // If the node was created after the escape computation we can't answer. 2057 uint idx = n->_idx; 2058 if (idx >= nodes_size()) { 2059 return false; 2060 } 2061 PointsToNode* ptn = ptnode_adr(idx); 2062 PointsToNode::EscapeState es = ptn->escape_state(); 2063 // If we have already computed a value, return it. 2064 if (es >= PointsToNode::GlobalEscape) 2065 return false; 2066 if (ptn->is_JavaObject()) { 2067 return true; // (es < PointsToNode::GlobalEscape); 2068 } 2069 assert(ptn->is_LocalVar(), "sanity"); 2070 // Check all java objects it points to. 2071 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2072 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2073 return false; 2074 } 2075 return true; 2076 } 2077 2078 2079 // Helper functions 2080 2081 // Return true if this node points to specified node or nodes it points to. 2082 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2083 if (is_JavaObject()) { 2084 return (this == ptn); 2085 } 2086 assert(is_LocalVar() || is_Field(), "sanity"); 2087 for (EdgeIterator i(this); i.has_next(); i.next()) { 2088 if (i.get() == ptn) 2089 return true; 2090 } 2091 return false; 2092 } 2093 2094 // Return true if one node points to an other. 2095 bool PointsToNode::meet(PointsToNode* ptn) { 2096 if (this == ptn) { 2097 return true; 2098 } else if (ptn->is_JavaObject()) { 2099 return this->points_to(ptn->as_JavaObject()); 2100 } else if (this->is_JavaObject()) { 2101 return ptn->points_to(this->as_JavaObject()); 2102 } 2103 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2104 int ptn_count = ptn->edge_count(); 2105 for (EdgeIterator i(this); i.has_next(); i.next()) { 2106 PointsToNode* this_e = i.get(); 2107 for (int j = 0; j < ptn_count; j++) { 2108 if (this_e == ptn->edge(j)) 2109 return true; 2110 } 2111 } 2112 return false; 2113 } 2114 2115 #ifdef ASSERT 2116 // Return true if bases point to this java object. 2117 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2118 for (BaseIterator i(this); i.has_next(); i.next()) { 2119 if (i.get() == jobj) 2120 return true; 2121 } 2122 return false; 2123 } 2124 #endif 2125 2126 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2127 const Type *adr_type = phase->type(adr); 2128 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2129 adr->in(AddPNode::Address)->is_Proj() && 2130 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2131 // We are computing a raw address for a store captured by an Initialize 2132 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2133 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2134 assert(offs != Type::OffsetBot || 2135 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2136 "offset must be a constant or it is initialization of array"); 2137 return offs; 2138 } 2139 const TypePtr *t_ptr = adr_type->isa_ptr(); 2140 assert(t_ptr != NULL, "must be a pointer type"); 2141 return t_ptr->offset(); 2142 } 2143 2144 Node* ConnectionGraph::get_addp_base(Node *addp) { 2145 assert(addp->is_AddP(), "must be AddP"); 2146 // 2147 // AddP cases for Base and Address inputs: 2148 // case #1. Direct object's field reference: 2149 // Allocate 2150 // | 2151 // Proj #5 ( oop result ) 2152 // | 2153 // CheckCastPP (cast to instance type) 2154 // | | 2155 // AddP ( base == address ) 2156 // 2157 // case #2. Indirect object's field reference: 2158 // Phi 2159 // | 2160 // CastPP (cast to instance type) 2161 // | | 2162 // AddP ( base == address ) 2163 // 2164 // case #3. Raw object's field reference for Initialize node: 2165 // Allocate 2166 // | 2167 // Proj #5 ( oop result ) 2168 // top | 2169 // \ | 2170 // AddP ( base == top ) 2171 // 2172 // case #4. Array's element reference: 2173 // {CheckCastPP | CastPP} 2174 // | | | 2175 // | AddP ( array's element offset ) 2176 // | | 2177 // AddP ( array's offset ) 2178 // 2179 // case #5. Raw object's field reference for arraycopy stub call: 2180 // The inline_native_clone() case when the arraycopy stub is called 2181 // after the allocation before Initialize and CheckCastPP nodes. 2182 // Allocate 2183 // | 2184 // Proj #5 ( oop result ) 2185 // | | 2186 // AddP ( base == address ) 2187 // 2188 // case #6. Constant Pool, ThreadLocal, CastX2P or 2189 // Raw object's field reference: 2190 // {ConP, ThreadLocal, CastX2P, raw Load} 2191 // top | 2192 // \ | 2193 // AddP ( base == top ) 2194 // 2195 // case #7. Klass's field reference. 2196 // LoadKlass 2197 // | | 2198 // AddP ( base == address ) 2199 // 2200 // case #8. narrow Klass's field reference. 2201 // LoadNKlass 2202 // | 2203 // DecodeN 2204 // | | 2205 // AddP ( base == address ) 2206 // 2207 Node *base = addp->in(AddPNode::Base); 2208 if (base->uncast()->is_top()) { // The AddP case #3 and #6. 2209 base = addp->in(AddPNode::Address); 2210 while (base->is_AddP()) { 2211 // Case #6 (unsafe access) may have several chained AddP nodes. 2212 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2213 base = base->in(AddPNode::Address); 2214 } 2215 Node* uncast_base = base->uncast(); 2216 int opcode = uncast_base->Opcode(); 2217 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2218 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2219 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2220 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity"); 2221 } 2222 return base; 2223 } 2224 2225 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2226 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2227 Node* addp2 = addp->raw_out(0); 2228 if (addp->outcnt() == 1 && addp2->is_AddP() && 2229 addp2->in(AddPNode::Base) == n && 2230 addp2->in(AddPNode::Address) == addp) { 2231 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2232 // 2233 // Find array's offset to push it on worklist first and 2234 // as result process an array's element offset first (pushed second) 2235 // to avoid CastPP for the array's offset. 2236 // Otherwise the inserted CastPP (LocalVar) will point to what 2237 // the AddP (Field) points to. Which would be wrong since 2238 // the algorithm expects the CastPP has the same point as 2239 // as AddP's base CheckCastPP (LocalVar). 2240 // 2241 // ArrayAllocation 2242 // | 2243 // CheckCastPP 2244 // | 2245 // memProj (from ArrayAllocation CheckCastPP) 2246 // | || 2247 // | || Int (element index) 2248 // | || | ConI (log(element size)) 2249 // | || | / 2250 // | || LShift 2251 // | || / 2252 // | AddP (array's element offset) 2253 // | | 2254 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2255 // | / / 2256 // AddP (array's offset) 2257 // | 2258 // Load/Store (memory operation on array's element) 2259 // 2260 return addp2; 2261 } 2262 return NULL; 2263 } 2264 2265 // 2266 // Adjust the type and inputs of an AddP which computes the 2267 // address of a field of an instance 2268 // 2269 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2270 PhaseGVN* igvn = _igvn; 2271 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2272 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2273 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2274 if (t == NULL) { 2275 // We are computing a raw address for a store captured by an Initialize 2276 // compute an appropriate address type (cases #3 and #5). 2277 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2278 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2279 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2280 assert(offs != Type::OffsetBot, "offset must be a constant"); 2281 t = base_t->add_offset(offs)->is_oopptr(); 2282 } 2283 int inst_id = base_t->instance_id(); 2284 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2285 "old type must be non-instance or match new type"); 2286 2287 // The type 't' could be subclass of 'base_t'. 2288 // As result t->offset() could be large then base_t's size and it will 2289 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2290 // constructor verifies correctness of the offset. 2291 // 2292 // It could happened on subclass's branch (from the type profiling 2293 // inlining) which was not eliminated during parsing since the exactness 2294 // of the allocation type was not propagated to the subclass type check. 2295 // 2296 // Or the type 't' could be not related to 'base_t' at all. 2297 // It could happened when CHA type is different from MDO type on a dead path 2298 // (for example, from instanceof check) which is not collapsed during parsing. 2299 // 2300 // Do nothing for such AddP node and don't process its users since 2301 // this code branch will go away. 2302 // 2303 if (!t->is_known_instance() && 2304 !base_t->klass()->is_subtype_of(t->klass())) { 2305 return false; // bail out 2306 } 2307 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2308 // Do NOT remove the next line: ensure a new alias index is allocated 2309 // for the instance type. Note: C++ will not remove it since the call 2310 // has side effect. 2311 int alias_idx = _compile->get_alias_index(tinst); 2312 igvn->set_type(addp, tinst); 2313 // record the allocation in the node map 2314 set_map(addp, get_map(base->_idx)); 2315 // Set addp's Base and Address to 'base'. 2316 Node *abase = addp->in(AddPNode::Base); 2317 Node *adr = addp->in(AddPNode::Address); 2318 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2319 adr->in(0)->_idx == (uint)inst_id) { 2320 // Skip AddP cases #3 and #5. 2321 } else { 2322 assert(!abase->is_top(), "sanity"); // AddP case #3 2323 if (abase != base) { 2324 igvn->hash_delete(addp); 2325 addp->set_req(AddPNode::Base, base); 2326 if (abase == adr) { 2327 addp->set_req(AddPNode::Address, base); 2328 } else { 2329 // AddP case #4 (adr is array's element offset AddP node) 2330 #ifdef ASSERT 2331 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2332 assert(adr->is_AddP() && atype != NULL && 2333 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2334 #endif 2335 } 2336 igvn->hash_insert(addp); 2337 } 2338 } 2339 // Put on IGVN worklist since at least addp's type was changed above. 2340 record_for_optimizer(addp); 2341 return true; 2342 } 2343 2344 // 2345 // Create a new version of orig_phi if necessary. Returns either the newly 2346 // created phi or an existing phi. Sets create_new to indicate whether a new 2347 // phi was created. Cache the last newly created phi in the node map. 2348 // 2349 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2350 Compile *C = _compile; 2351 PhaseGVN* igvn = _igvn; 2352 new_created = false; 2353 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2354 // nothing to do if orig_phi is bottom memory or matches alias_idx 2355 if (phi_alias_idx == alias_idx) { 2356 return orig_phi; 2357 } 2358 // Have we recently created a Phi for this alias index? 2359 PhiNode *result = get_map_phi(orig_phi->_idx); 2360 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2361 return result; 2362 } 2363 // Previous check may fail when the same wide memory Phi was split into Phis 2364 // for different memory slices. Search all Phis for this region. 2365 if (result != NULL) { 2366 Node* region = orig_phi->in(0); 2367 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2368 Node* phi = region->fast_out(i); 2369 if (phi->is_Phi() && 2370 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2371 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2372 return phi->as_Phi(); 2373 } 2374 } 2375 } 2376 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) { 2377 if (C->do_escape_analysis() == true && !C->failing()) { 2378 // Retry compilation without escape analysis. 2379 // If this is the first failure, the sentinel string will "stick" 2380 // to the Compile object, and the C2Compiler will see it and retry. 2381 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2382 } 2383 return NULL; 2384 } 2385 orig_phi_worklist.append_if_missing(orig_phi); 2386 const TypePtr *atype = C->get_adr_type(alias_idx); 2387 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2388 C->copy_node_notes_to(result, orig_phi); 2389 igvn->set_type(result, result->bottom_type()); 2390 record_for_optimizer(result); 2391 set_map(orig_phi, result); 2392 new_created = true; 2393 return result; 2394 } 2395 2396 // 2397 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2398 // specified alias index. 2399 // 2400 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2401 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2402 Compile *C = _compile; 2403 PhaseGVN* igvn = _igvn; 2404 bool new_phi_created; 2405 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2406 if (!new_phi_created) { 2407 return result; 2408 } 2409 GrowableArray<PhiNode *> phi_list; 2410 GrowableArray<uint> cur_input; 2411 PhiNode *phi = orig_phi; 2412 uint idx = 1; 2413 bool finished = false; 2414 while(!finished) { 2415 while (idx < phi->req()) { 2416 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2417 if (mem != NULL && mem->is_Phi()) { 2418 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2419 if (new_phi_created) { 2420 // found an phi for which we created a new split, push current one on worklist and begin 2421 // processing new one 2422 phi_list.push(phi); 2423 cur_input.push(idx); 2424 phi = mem->as_Phi(); 2425 result = newphi; 2426 idx = 1; 2427 continue; 2428 } else { 2429 mem = newphi; 2430 } 2431 } 2432 if (C->failing()) { 2433 return NULL; 2434 } 2435 result->set_req(idx++, mem); 2436 } 2437 #ifdef ASSERT 2438 // verify that the new Phi has an input for each input of the original 2439 assert( phi->req() == result->req(), "must have same number of inputs."); 2440 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2441 #endif 2442 // Check if all new phi's inputs have specified alias index. 2443 // Otherwise use old phi. 2444 for (uint i = 1; i < phi->req(); i++) { 2445 Node* in = result->in(i); 2446 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2447 } 2448 // we have finished processing a Phi, see if there are any more to do 2449 finished = (phi_list.length() == 0 ); 2450 if (!finished) { 2451 phi = phi_list.pop(); 2452 idx = cur_input.pop(); 2453 PhiNode *prev_result = get_map_phi(phi->_idx); 2454 prev_result->set_req(idx++, result); 2455 result = prev_result; 2456 } 2457 } 2458 return result; 2459 } 2460 2461 // 2462 // The next methods are derived from methods in MemNode. 2463 // 2464 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2465 Node *mem = mmem; 2466 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2467 // means an array I have not precisely typed yet. Do not do any 2468 // alias stuff with it any time soon. 2469 if (toop->base() != Type::AnyPtr && 2470 !(toop->klass() != NULL && 2471 toop->klass()->is_java_lang_Object() && 2472 toop->offset() == Type::OffsetBot)) { 2473 mem = mmem->memory_at(alias_idx); 2474 // Update input if it is progress over what we have now 2475 } 2476 return mem; 2477 } 2478 2479 // 2480 // Move memory users to their memory slices. 2481 // 2482 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2483 Compile* C = _compile; 2484 PhaseGVN* igvn = _igvn; 2485 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2486 assert(tp != NULL, "ptr type"); 2487 int alias_idx = C->get_alias_index(tp); 2488 int general_idx = C->get_general_index(alias_idx); 2489 2490 // Move users first 2491 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2492 Node* use = n->fast_out(i); 2493 if (use->is_MergeMem()) { 2494 MergeMemNode* mmem = use->as_MergeMem(); 2495 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2496 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2497 continue; // Nothing to do 2498 } 2499 // Replace previous general reference to mem node. 2500 uint orig_uniq = C->unique(); 2501 Node* m = find_inst_mem(n, general_idx, orig_phis); 2502 assert(orig_uniq == C->unique(), "no new nodes"); 2503 mmem->set_memory_at(general_idx, m); 2504 --imax; 2505 --i; 2506 } else if (use->is_MemBar()) { 2507 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2508 if (use->req() > MemBarNode::Precedent && 2509 use->in(MemBarNode::Precedent) == n) { 2510 // Don't move related membars. 2511 record_for_optimizer(use); 2512 continue; 2513 } 2514 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2515 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 2516 alias_idx == general_idx) { 2517 continue; // Nothing to do 2518 } 2519 // Move to general memory slice. 2520 uint orig_uniq = C->unique(); 2521 Node* m = find_inst_mem(n, general_idx, orig_phis); 2522 assert(orig_uniq == C->unique(), "no new nodes"); 2523 igvn->hash_delete(use); 2524 imax -= use->replace_edge(n, m); 2525 igvn->hash_insert(use); 2526 record_for_optimizer(use); 2527 --i; 2528 #ifdef ASSERT 2529 } else if (use->is_Mem()) { 2530 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2531 // Don't move related cardmark. 2532 continue; 2533 } 2534 // Memory nodes should have new memory input. 2535 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2536 assert(tp != NULL, "ptr type"); 2537 int idx = C->get_alias_index(tp); 2538 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2539 "Following memory nodes should have new memory input or be on the same memory slice"); 2540 } else if (use->is_Phi()) { 2541 // Phi nodes should be split and moved already. 2542 tp = use->as_Phi()->adr_type()->isa_ptr(); 2543 assert(tp != NULL, "ptr type"); 2544 int idx = C->get_alias_index(tp); 2545 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2546 } else { 2547 use->dump(); 2548 assert(false, "should not be here"); 2549 #endif 2550 } 2551 } 2552 } 2553 2554 // 2555 // Search memory chain of "mem" to find a MemNode whose address 2556 // is the specified alias index. 2557 // 2558 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2559 if (orig_mem == NULL) 2560 return orig_mem; 2561 Compile* C = _compile; 2562 PhaseGVN* igvn = _igvn; 2563 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2564 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2565 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 2566 Node *prev = NULL; 2567 Node *result = orig_mem; 2568 while (prev != result) { 2569 prev = result; 2570 if (result == start_mem) 2571 break; // hit one of our sentinels 2572 if (result->is_Mem()) { 2573 const Type *at = igvn->type(result->in(MemNode::Address)); 2574 if (at == Type::TOP) 2575 break; // Dead 2576 assert (at->isa_ptr() != NULL, "pointer type required."); 2577 int idx = C->get_alias_index(at->is_ptr()); 2578 if (idx == alias_idx) 2579 break; // Found 2580 if (!is_instance && (at->isa_oopptr() == NULL || 2581 !at->is_oopptr()->is_known_instance())) { 2582 break; // Do not skip store to general memory slice. 2583 } 2584 result = result->in(MemNode::Memory); 2585 } 2586 if (!is_instance) 2587 continue; // don't search further for non-instance types 2588 // skip over a call which does not affect this memory slice 2589 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2590 Node *proj_in = result->in(0); 2591 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2592 break; // hit one of our sentinels 2593 } else if (proj_in->is_Call()) { 2594 CallNode *call = proj_in->as_Call(); 2595 if (!call->may_modify(toop, igvn)) { 2596 result = call->in(TypeFunc::Memory); 2597 } 2598 } else if (proj_in->is_Initialize()) { 2599 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2600 // Stop if this is the initialization for the object instance which 2601 // which contains this memory slice, otherwise skip over it. 2602 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2603 result = proj_in->in(TypeFunc::Memory); 2604 } 2605 } else if (proj_in->is_MemBar()) { 2606 result = proj_in->in(TypeFunc::Memory); 2607 } 2608 } else if (result->is_MergeMem()) { 2609 MergeMemNode *mmem = result->as_MergeMem(); 2610 result = step_through_mergemem(mmem, alias_idx, toop); 2611 if (result == mmem->base_memory()) { 2612 // Didn't find instance memory, search through general slice recursively. 2613 result = mmem->memory_at(C->get_general_index(alias_idx)); 2614 result = find_inst_mem(result, alias_idx, orig_phis); 2615 if (C->failing()) { 2616 return NULL; 2617 } 2618 mmem->set_memory_at(alias_idx, result); 2619 } 2620 } else if (result->is_Phi() && 2621 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2622 Node *un = result->as_Phi()->unique_input(igvn); 2623 if (un != NULL) { 2624 orig_phis.append_if_missing(result->as_Phi()); 2625 result = un; 2626 } else { 2627 break; 2628 } 2629 } else if (result->is_ClearArray()) { 2630 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2631 // Can not bypass initialization of the instance 2632 // we are looking for. 2633 break; 2634 } 2635 // Otherwise skip it (the call updated 'result' value). 2636 } else if (result->Opcode() == Op_SCMemProj) { 2637 Node* mem = result->in(0); 2638 Node* adr = NULL; 2639 if (mem->is_LoadStore()) { 2640 adr = mem->in(MemNode::Address); 2641 } else { 2642 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 2643 adr = mem->in(3); // Memory edge corresponds to destination array 2644 } 2645 const Type *at = igvn->type(adr); 2646 if (at != Type::TOP) { 2647 assert (at->isa_ptr() != NULL, "pointer type required."); 2648 int idx = C->get_alias_index(at->is_ptr()); 2649 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 2650 break; 2651 } 2652 result = mem->in(MemNode::Memory); 2653 } 2654 } 2655 if (result->is_Phi()) { 2656 PhiNode *mphi = result->as_Phi(); 2657 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2658 const TypePtr *t = mphi->adr_type(); 2659 if (!is_instance) { 2660 // Push all non-instance Phis on the orig_phis worklist to update inputs 2661 // during Phase 4 if needed. 2662 orig_phis.append_if_missing(mphi); 2663 } else if (C->get_alias_index(t) != alias_idx) { 2664 // Create a new Phi with the specified alias index type. 2665 result = split_memory_phi(mphi, alias_idx, orig_phis); 2666 } 2667 } 2668 // the result is either MemNode, PhiNode, InitializeNode. 2669 return result; 2670 } 2671 2672 // 2673 // Convert the types of unescaped object to instance types where possible, 2674 // propagate the new type information through the graph, and update memory 2675 // edges and MergeMem inputs to reflect the new type. 2676 // 2677 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2678 // The processing is done in 4 phases: 2679 // 2680 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2681 // types for the CheckCastPP for allocations where possible. 2682 // Propagate the the new types through users as follows: 2683 // casts and Phi: push users on alloc_worklist 2684 // AddP: cast Base and Address inputs to the instance type 2685 // push any AddP users on alloc_worklist and push any memnode 2686 // users onto memnode_worklist. 2687 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2688 // search the Memory chain for a store with the appropriate type 2689 // address type. If a Phi is found, create a new version with 2690 // the appropriate memory slices from each of the Phi inputs. 2691 // For stores, process the users as follows: 2692 // MemNode: push on memnode_worklist 2693 // MergeMem: push on mergemem_worklist 2694 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2695 // moving the first node encountered of each instance type to the 2696 // the input corresponding to its alias index. 2697 // appropriate memory slice. 2698 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2699 // 2700 // In the following example, the CheckCastPP nodes are the cast of allocation 2701 // results and the allocation of node 29 is unescaped and eligible to be an 2702 // instance type. 2703 // 2704 // We start with: 2705 // 2706 // 7 Parm #memory 2707 // 10 ConI "12" 2708 // 19 CheckCastPP "Foo" 2709 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2710 // 29 CheckCastPP "Foo" 2711 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2712 // 2713 // 40 StoreP 25 7 20 ... alias_index=4 2714 // 50 StoreP 35 40 30 ... alias_index=4 2715 // 60 StoreP 45 50 20 ... alias_index=4 2716 // 70 LoadP _ 60 30 ... alias_index=4 2717 // 80 Phi 75 50 60 Memory alias_index=4 2718 // 90 LoadP _ 80 30 ... alias_index=4 2719 // 100 LoadP _ 80 20 ... alias_index=4 2720 // 2721 // 2722 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2723 // and creating a new alias index for node 30. This gives: 2724 // 2725 // 7 Parm #memory 2726 // 10 ConI "12" 2727 // 19 CheckCastPP "Foo" 2728 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2729 // 29 CheckCastPP "Foo" iid=24 2730 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2731 // 2732 // 40 StoreP 25 7 20 ... alias_index=4 2733 // 50 StoreP 35 40 30 ... alias_index=6 2734 // 60 StoreP 45 50 20 ... alias_index=4 2735 // 70 LoadP _ 60 30 ... alias_index=6 2736 // 80 Phi 75 50 60 Memory alias_index=4 2737 // 90 LoadP _ 80 30 ... alias_index=6 2738 // 100 LoadP _ 80 20 ... alias_index=4 2739 // 2740 // In phase 2, new memory inputs are computed for the loads and stores, 2741 // And a new version of the phi is created. In phase 4, the inputs to 2742 // node 80 are updated and then the memory nodes are updated with the 2743 // values computed in phase 2. This results in: 2744 // 2745 // 7 Parm #memory 2746 // 10 ConI "12" 2747 // 19 CheckCastPP "Foo" 2748 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2749 // 29 CheckCastPP "Foo" iid=24 2750 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2751 // 2752 // 40 StoreP 25 7 20 ... alias_index=4 2753 // 50 StoreP 35 7 30 ... alias_index=6 2754 // 60 StoreP 45 40 20 ... alias_index=4 2755 // 70 LoadP _ 50 30 ... alias_index=6 2756 // 80 Phi 75 40 60 Memory alias_index=4 2757 // 120 Phi 75 50 50 Memory alias_index=6 2758 // 90 LoadP _ 120 30 ... alias_index=6 2759 // 100 LoadP _ 80 20 ... alias_index=4 2760 // 2761 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 2762 GrowableArray<Node *> memnode_worklist; 2763 GrowableArray<PhiNode *> orig_phis; 2764 PhaseIterGVN *igvn = _igvn; 2765 uint new_index_start = (uint) _compile->num_alias_types(); 2766 Arena* arena = Thread::current()->resource_area(); 2767 VectorSet visited(arena); 2768 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2769 uint unique_old = _compile->unique(); 2770 2771 // Phase 1: Process possible allocations from alloc_worklist. 2772 // Create instance types for the CheckCastPP for allocations where possible. 2773 // 2774 // (Note: don't forget to change the order of the second AddP node on 2775 // the alloc_worklist if the order of the worklist processing is changed, 2776 // see the comment in find_second_addp().) 2777 // 2778 while (alloc_worklist.length() != 0) { 2779 Node *n = alloc_worklist.pop(); 2780 uint ni = n->_idx; 2781 if (n->is_Call()) { 2782 CallNode *alloc = n->as_Call(); 2783 // copy escape information to call node 2784 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2785 PointsToNode::EscapeState es = ptn->escape_state(); 2786 // We have an allocation or call which returns a Java object, 2787 // see if it is unescaped. 2788 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2789 continue; 2790 // Find CheckCastPP for the allocate or for the return value of a call 2791 n = alloc->result_cast(); 2792 if (n == NULL) { // No uses except Initialize node 2793 if (alloc->is_Allocate()) { 2794 // Set the scalar_replaceable flag for allocation 2795 // so it could be eliminated if it has no uses. 2796 alloc->as_Allocate()->_is_scalar_replaceable = true; 2797 } 2798 if (alloc->is_CallStaticJava()) { 2799 // Set the scalar_replaceable flag for boxing method 2800 // so it could be eliminated if it has no uses. 2801 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2802 } 2803 continue; 2804 } 2805 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2806 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2807 continue; 2808 } 2809 2810 // The inline code for Object.clone() casts the allocation result to 2811 // java.lang.Object and then to the actual type of the allocated 2812 // object. Detect this case and use the second cast. 2813 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2814 // the allocation result is cast to java.lang.Object and then 2815 // to the actual Array type. 2816 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2817 && (alloc->is_AllocateArray() || 2818 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2819 Node *cast2 = NULL; 2820 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2821 Node *use = n->fast_out(i); 2822 if (use->is_CheckCastPP()) { 2823 cast2 = use; 2824 break; 2825 } 2826 } 2827 if (cast2 != NULL) { 2828 n = cast2; 2829 } else { 2830 // Non-scalar replaceable if the allocation type is unknown statically 2831 // (reflection allocation), the object can't be restored during 2832 // deoptimization without precise type. 2833 continue; 2834 } 2835 } 2836 if (alloc->is_Allocate()) { 2837 // Set the scalar_replaceable flag for allocation 2838 // so it could be eliminated. 2839 alloc->as_Allocate()->_is_scalar_replaceable = true; 2840 } 2841 if (alloc->is_CallStaticJava()) { 2842 // Set the scalar_replaceable flag for boxing method 2843 // so it could be eliminated. 2844 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2845 } 2846 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 2847 // in order for an object to be scalar-replaceable, it must be: 2848 // - a direct allocation (not a call returning an object) 2849 // - non-escaping 2850 // - eligible to be a unique type 2851 // - not determined to be ineligible by escape analysis 2852 set_map(alloc, n); 2853 set_map(n, alloc); 2854 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 2855 if (t == NULL) 2856 continue; // not a TypeOopPtr 2857 const TypeOopPtr* tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 2858 igvn->hash_delete(n); 2859 igvn->set_type(n, tinst); 2860 n->raise_bottom_type(tinst); 2861 igvn->hash_insert(n); 2862 record_for_optimizer(n); 2863 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 2864 2865 // First, put on the worklist all Field edges from Connection Graph 2866 // which is more accurate then putting immediate users from Ideal Graph. 2867 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 2868 PointsToNode* tgt = e.get(); 2869 Node* use = tgt->ideal_node(); 2870 assert(tgt->is_Field() && use->is_AddP(), 2871 "only AddP nodes are Field edges in CG"); 2872 if (use->outcnt() > 0) { // Don't process dead nodes 2873 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 2874 if (addp2 != NULL) { 2875 assert(alloc->is_AllocateArray(),"array allocation was expected"); 2876 alloc_worklist.append_if_missing(addp2); 2877 } 2878 alloc_worklist.append_if_missing(use); 2879 } 2880 } 2881 2882 // An allocation may have an Initialize which has raw stores. Scan 2883 // the users of the raw allocation result and push AddP users 2884 // on alloc_worklist. 2885 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 2886 assert (raw_result != NULL, "must have an allocation result"); 2887 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 2888 Node *use = raw_result->fast_out(i); 2889 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 2890 Node* addp2 = find_second_addp(use, raw_result); 2891 if (addp2 != NULL) { 2892 assert(alloc->is_AllocateArray(),"array allocation was expected"); 2893 alloc_worklist.append_if_missing(addp2); 2894 } 2895 alloc_worklist.append_if_missing(use); 2896 } else if (use->is_MemBar()) { 2897 memnode_worklist.append_if_missing(use); 2898 } 2899 } 2900 } 2901 } else if (n->is_AddP()) { 2902 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 2903 if (jobj == NULL || jobj == phantom_obj) { 2904 #ifdef ASSERT 2905 ptnode_adr(get_addp_base(n)->_idx)->dump(); 2906 ptnode_adr(n->_idx)->dump(); 2907 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 2908 #endif 2909 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 2910 return; 2911 } 2912 Node *base = get_map(jobj->idx()); // CheckCastPP node 2913 if (!split_AddP(n, base)) continue; // wrong type from dead path 2914 } else if (n->is_Phi() || 2915 n->is_CheckCastPP() || 2916 n->is_EncodeP() || 2917 n->is_DecodeN() || 2918 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 2919 if (visited.test_set(n->_idx)) { 2920 assert(n->is_Phi(), "loops only through Phi's"); 2921 continue; // already processed 2922 } 2923 JavaObjectNode* jobj = unique_java_object(n); 2924 if (jobj == NULL || jobj == phantom_obj) { 2925 #ifdef ASSERT 2926 ptnode_adr(n->_idx)->dump(); 2927 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 2928 #endif 2929 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 2930 return; 2931 } else { 2932 Node *val = get_map(jobj->idx()); // CheckCastPP node 2933 TypeNode *tn = n->as_Type(); 2934 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 2935 assert(tinst != NULL && tinst->is_known_instance() && 2936 tinst->instance_id() == jobj->idx() , "instance type expected."); 2937 2938 const Type *tn_type = igvn->type(tn); 2939 const TypeOopPtr *tn_t; 2940 if (tn_type->isa_narrowoop()) { 2941 tn_t = tn_type->make_ptr()->isa_oopptr(); 2942 } else { 2943 tn_t = tn_type->isa_oopptr(); 2944 } 2945 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 2946 if (tn_type->isa_narrowoop()) { 2947 tn_type = tinst->make_narrowoop(); 2948 } else { 2949 tn_type = tinst; 2950 } 2951 igvn->hash_delete(tn); 2952 igvn->set_type(tn, tn_type); 2953 tn->set_type(tn_type); 2954 igvn->hash_insert(tn); 2955 record_for_optimizer(n); 2956 } else { 2957 assert(tn_type == TypePtr::NULL_PTR || 2958 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 2959 "unexpected type"); 2960 continue; // Skip dead path with different type 2961 } 2962 } 2963 } else { 2964 debug_only(n->dump();) 2965 assert(false, "EA: unexpected node"); 2966 continue; 2967 } 2968 // push allocation's users on appropriate worklist 2969 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2970 Node *use = n->fast_out(i); 2971 if(use->is_Mem() && use->in(MemNode::Address) == n) { 2972 // Load/store to instance's field 2973 memnode_worklist.append_if_missing(use); 2974 } else if (use->is_MemBar()) { 2975 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 2976 memnode_worklist.append_if_missing(use); 2977 } 2978 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 2979 Node* addp2 = find_second_addp(use, n); 2980 if (addp2 != NULL) { 2981 alloc_worklist.append_if_missing(addp2); 2982 } 2983 alloc_worklist.append_if_missing(use); 2984 } else if (use->is_Phi() || 2985 use->is_CheckCastPP() || 2986 use->is_EncodeNarrowPtr() || 2987 use->is_DecodeNarrowPtr() || 2988 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 2989 alloc_worklist.append_if_missing(use); 2990 #ifdef ASSERT 2991 } else if (use->is_Mem()) { 2992 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 2993 } else if (use->is_MergeMem()) { 2994 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 2995 } else if (use->is_SafePoint()) { 2996 // Look for MergeMem nodes for calls which reference unique allocation 2997 // (through CheckCastPP nodes) even for debug info. 2998 Node* m = use->in(TypeFunc::Memory); 2999 if (m->is_MergeMem()) { 3000 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3001 } 3002 } else if (use->Opcode() == Op_EncodeISOArray) { 3003 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3004 // EncodeISOArray overwrites destination array 3005 memnode_worklist.append_if_missing(use); 3006 } 3007 } else { 3008 uint op = use->Opcode(); 3009 if (!(op == Op_CmpP || op == Op_Conv2B || 3010 op == Op_CastP2X || op == Op_StoreCM || 3011 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 3012 op == Op_StrEquals || op == Op_StrIndexOf)) { 3013 n->dump(); 3014 use->dump(); 3015 assert(false, "EA: missing allocation reference path"); 3016 } 3017 #endif 3018 } 3019 } 3020 3021 } 3022 // New alias types were created in split_AddP(). 3023 uint new_index_end = (uint) _compile->num_alias_types(); 3024 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3025 3026 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3027 // compute new values for Memory inputs (the Memory inputs are not 3028 // actually updated until phase 4.) 3029 if (memnode_worklist.length() == 0) 3030 return; // nothing to do 3031 while (memnode_worklist.length() != 0) { 3032 Node *n = memnode_worklist.pop(); 3033 if (visited.test_set(n->_idx)) 3034 continue; 3035 if (n->is_Phi() || n->is_ClearArray()) { 3036 // we don't need to do anything, but the users must be pushed 3037 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3038 // we don't need to do anything, but the users must be pushed 3039 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 3040 if (n == NULL) 3041 continue; 3042 } else if (n->Opcode() == Op_EncodeISOArray) { 3043 // get the memory projection 3044 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3045 Node *use = n->fast_out(i); 3046 if (use->Opcode() == Op_SCMemProj) { 3047 n = use; 3048 break; 3049 } 3050 } 3051 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3052 } else { 3053 assert(n->is_Mem(), "memory node required."); 3054 Node *addr = n->in(MemNode::Address); 3055 const Type *addr_t = igvn->type(addr); 3056 if (addr_t == Type::TOP) 3057 continue; 3058 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3059 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3060 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3061 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3062 if (_compile->failing()) { 3063 return; 3064 } 3065 if (mem != n->in(MemNode::Memory)) { 3066 // We delay the memory edge update since we need old one in 3067 // MergeMem code below when instances memory slices are separated. 3068 set_map(n, mem); 3069 } 3070 if (n->is_Load()) { 3071 continue; // don't push users 3072 } else if (n->is_LoadStore()) { 3073 // get the memory projection 3074 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3075 Node *use = n->fast_out(i); 3076 if (use->Opcode() == Op_SCMemProj) { 3077 n = use; 3078 break; 3079 } 3080 } 3081 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3082 } 3083 } 3084 // push user on appropriate worklist 3085 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3086 Node *use = n->fast_out(i); 3087 if (use->is_Phi() || use->is_ClearArray()) { 3088 memnode_worklist.append_if_missing(use); 3089 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3090 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3091 continue; 3092 memnode_worklist.append_if_missing(use); 3093 } else if (use->is_MemBar()) { 3094 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3095 memnode_worklist.append_if_missing(use); 3096 } 3097 #ifdef ASSERT 3098 } else if(use->is_Mem()) { 3099 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3100 } else if (use->is_MergeMem()) { 3101 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3102 } else if (use->Opcode() == Op_EncodeISOArray) { 3103 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3104 // EncodeISOArray overwrites destination array 3105 memnode_worklist.append_if_missing(use); 3106 } 3107 } else { 3108 uint op = use->Opcode(); 3109 if (!(op == Op_StoreCM || 3110 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 3111 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 3112 op == Op_AryEq || op == Op_StrComp || 3113 op == Op_StrEquals || op == Op_StrIndexOf)) { 3114 n->dump(); 3115 use->dump(); 3116 assert(false, "EA: missing memory path"); 3117 } 3118 #endif 3119 } 3120 } 3121 } 3122 3123 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3124 // Walk each memory slice moving the first node encountered of each 3125 // instance type to the the input corresponding to its alias index. 3126 uint length = _mergemem_worklist.length(); 3127 for( uint next = 0; next < length; ++next ) { 3128 MergeMemNode* nmm = _mergemem_worklist.at(next); 3129 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3130 // Note: we don't want to use MergeMemStream here because we only want to 3131 // scan inputs which exist at the start, not ones we add during processing. 3132 // Note 2: MergeMem may already contains instance memory slices added 3133 // during find_inst_mem() call when memory nodes were processed above. 3134 igvn->hash_delete(nmm); 3135 uint nslices = nmm->req(); 3136 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3137 Node* mem = nmm->in(i); 3138 Node* cur = NULL; 3139 if (mem == NULL || mem->is_top()) 3140 continue; 3141 // First, update mergemem by moving memory nodes to corresponding slices 3142 // if their type became more precise since this mergemem was created. 3143 while (mem->is_Mem()) { 3144 const Type *at = igvn->type(mem->in(MemNode::Address)); 3145 if (at != Type::TOP) { 3146 assert (at->isa_ptr() != NULL, "pointer type required."); 3147 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3148 if (idx == i) { 3149 if (cur == NULL) 3150 cur = mem; 3151 } else { 3152 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3153 nmm->set_memory_at(idx, mem); 3154 } 3155 } 3156 } 3157 mem = mem->in(MemNode::Memory); 3158 } 3159 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3160 // Find any instance of the current type if we haven't encountered 3161 // already a memory slice of the instance along the memory chain. 3162 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3163 if((uint)_compile->get_general_index(ni) == i) { 3164 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3165 if (nmm->is_empty_memory(m)) { 3166 Node* result = find_inst_mem(mem, ni, orig_phis); 3167 if (_compile->failing()) { 3168 return; 3169 } 3170 nmm->set_memory_at(ni, result); 3171 } 3172 } 3173 } 3174 } 3175 // Find the rest of instances values 3176 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3177 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3178 Node* result = step_through_mergemem(nmm, ni, tinst); 3179 if (result == nmm->base_memory()) { 3180 // Didn't find instance memory, search through general slice recursively. 3181 result = nmm->memory_at(_compile->get_general_index(ni)); 3182 result = find_inst_mem(result, ni, orig_phis); 3183 if (_compile->failing()) { 3184 return; 3185 } 3186 nmm->set_memory_at(ni, result); 3187 } 3188 } 3189 igvn->hash_insert(nmm); 3190 record_for_optimizer(nmm); 3191 } 3192 3193 // Phase 4: Update the inputs of non-instance memory Phis and 3194 // the Memory input of memnodes 3195 // First update the inputs of any non-instance Phi's from 3196 // which we split out an instance Phi. Note we don't have 3197 // to recursively process Phi's encounted on the input memory 3198 // chains as is done in split_memory_phi() since they will 3199 // also be processed here. 3200 for (int j = 0; j < orig_phis.length(); j++) { 3201 PhiNode *phi = orig_phis.at(j); 3202 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3203 igvn->hash_delete(phi); 3204 for (uint i = 1; i < phi->req(); i++) { 3205 Node *mem = phi->in(i); 3206 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3207 if (_compile->failing()) { 3208 return; 3209 } 3210 if (mem != new_mem) { 3211 phi->set_req(i, new_mem); 3212 } 3213 } 3214 igvn->hash_insert(phi); 3215 record_for_optimizer(phi); 3216 } 3217 3218 // Update the memory inputs of MemNodes with the value we computed 3219 // in Phase 2 and move stores memory users to corresponding memory slices. 3220 // Disable memory split verification code until the fix for 6984348. 3221 // Currently it produces false negative results since it does not cover all cases. 3222 #if 0 // ifdef ASSERT 3223 visited.Reset(); 3224 Node_Stack old_mems(arena, _compile->unique() >> 2); 3225 #endif 3226 for (uint i = 0; i < ideal_nodes.size(); i++) { 3227 Node* n = ideal_nodes.at(i); 3228 Node* nmem = get_map(n->_idx); 3229 assert(nmem != NULL, "sanity"); 3230 if (n->is_Mem()) { 3231 #if 0 // ifdef ASSERT 3232 Node* old_mem = n->in(MemNode::Memory); 3233 if (!visited.test_set(old_mem->_idx)) { 3234 old_mems.push(old_mem, old_mem->outcnt()); 3235 } 3236 #endif 3237 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3238 if (!n->is_Load()) { 3239 // Move memory users of a store first. 3240 move_inst_mem(n, orig_phis); 3241 } 3242 // Now update memory input 3243 igvn->hash_delete(n); 3244 n->set_req(MemNode::Memory, nmem); 3245 igvn->hash_insert(n); 3246 record_for_optimizer(n); 3247 } else { 3248 assert(n->is_Allocate() || n->is_CheckCastPP() || 3249 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3250 } 3251 } 3252 #if 0 // ifdef ASSERT 3253 // Verify that memory was split correctly 3254 while (old_mems.is_nonempty()) { 3255 Node* old_mem = old_mems.node(); 3256 uint old_cnt = old_mems.index(); 3257 old_mems.pop(); 3258 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3259 } 3260 #endif 3261 } 3262 3263 #ifndef PRODUCT 3264 static const char *node_type_names[] = { 3265 "UnknownType", 3266 "JavaObject", 3267 "LocalVar", 3268 "Field", 3269 "Arraycopy" 3270 }; 3271 3272 static const char *esc_names[] = { 3273 "UnknownEscape", 3274 "NoEscape", 3275 "ArgEscape", 3276 "GlobalEscape" 3277 }; 3278 3279 void PointsToNode::dump(bool print_state) const { 3280 NodeType nt = node_type(); 3281 tty->print("%s ", node_type_names[(int) nt]); 3282 if (print_state) { 3283 EscapeState es = escape_state(); 3284 EscapeState fields_es = fields_escape_state(); 3285 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3286 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3287 tty->print("NSR "); 3288 } 3289 if (is_Field()) { 3290 FieldNode* f = (FieldNode*)this; 3291 if (f->is_oop()) 3292 tty->print("oop "); 3293 if (f->offset() > 0) 3294 tty->print("+%d ", f->offset()); 3295 tty->print("("); 3296 for (BaseIterator i(f); i.has_next(); i.next()) { 3297 PointsToNode* b = i.get(); 3298 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3299 } 3300 tty->print(" )"); 3301 } 3302 tty->print("["); 3303 for (EdgeIterator i(this); i.has_next(); i.next()) { 3304 PointsToNode* e = i.get(); 3305 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3306 } 3307 tty->print(" ["); 3308 for (UseIterator i(this); i.has_next(); i.next()) { 3309 PointsToNode* u = i.get(); 3310 bool is_base = false; 3311 if (PointsToNode::is_base_use(u)) { 3312 is_base = true; 3313 u = PointsToNode::get_use_node(u)->as_Field(); 3314 } 3315 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3316 } 3317 tty->print(" ]] "); 3318 if (_node == NULL) 3319 tty->print_cr("<null>"); 3320 else 3321 _node->dump(); 3322 } 3323 3324 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3325 bool first = true; 3326 int ptnodes_length = ptnodes_worklist.length(); 3327 for (int i = 0; i < ptnodes_length; i++) { 3328 PointsToNode *ptn = ptnodes_worklist.at(i); 3329 if (ptn == NULL || !ptn->is_JavaObject()) 3330 continue; 3331 PointsToNode::EscapeState es = ptn->escape_state(); 3332 if ((es != PointsToNode::NoEscape) && !Verbose) { 3333 continue; 3334 } 3335 Node* n = ptn->ideal_node(); 3336 if (n->is_Allocate() || (n->is_CallStaticJava() && 3337 n->as_CallStaticJava()->is_boxing_method())) { 3338 if (first) { 3339 tty->cr(); 3340 tty->print("======== Connection graph for "); 3341 _compile->method()->print_short_name(); 3342 tty->cr(); 3343 first = false; 3344 } 3345 ptn->dump(); 3346 // Print all locals and fields which reference this allocation 3347 for (UseIterator j(ptn); j.has_next(); j.next()) { 3348 PointsToNode* use = j.get(); 3349 if (use->is_LocalVar()) { 3350 use->dump(Verbose); 3351 } else if (Verbose) { 3352 use->dump(); 3353 } 3354 } 3355 tty->cr(); 3356 } 3357 } 3358 } 3359 #endif