1 /* 2 * Copyright (c) 2005, 2016, 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 "memory/resourceArea.hpp" 31 #include "opto/c2compiler.hpp" 32 #include "opto/arraycopynode.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/compile.hpp" 36 #include "opto/escape.hpp" 37 #include "opto/phaseX.hpp" 38 #include "opto/movenode.hpp" 39 #include "opto/rootnode.hpp" 40 41 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 42 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL), 43 _in_worklist(C->comp_arena()), 44 _next_pidx(0), 45 _collecting(true), 46 _verify(false), 47 _compile(C), 48 _igvn(igvn), 49 _node_map(C->comp_arena()) { 50 // Add unknown java object. 51 add_java_object(C->top(), PointsToNode::GlobalEscape); 52 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 53 // Add ConP(#NULL) and ConN(#NULL) nodes. 54 Node* oop_null = igvn->zerocon(T_OBJECT); 55 assert(oop_null->_idx < nodes_size(), "should be created already"); 56 add_java_object(oop_null, PointsToNode::NoEscape); 57 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 58 if (UseCompressedOops) { 59 Node* noop_null = igvn->zerocon(T_NARROWOOP); 60 assert(noop_null->_idx < nodes_size(), "should be created already"); 61 map_ideal_node(noop_null, null_obj); 62 } 63 _pcmp_neq = NULL; // Should be initialized 64 _pcmp_eq = NULL; 65 } 66 67 bool ConnectionGraph::has_candidates(Compile *C) { 68 // EA brings benefits only when the code has allocations and/or locks which 69 // are represented by ideal Macro nodes. 70 int cnt = C->macro_count(); 71 for (int i = 0; i < cnt; i++) { 72 Node *n = C->macro_node(i); 73 if (n->is_Allocate()) 74 return true; 75 if (n->is_Lock()) { 76 Node* obj = n->as_Lock()->obj_node()->uncast(); 77 if (!(obj->is_Parm() || obj->is_Con())) 78 return true; 79 } 80 if (n->is_CallStaticJava() && 81 n->as_CallStaticJava()->is_boxing_method()) { 82 return true; 83 } 84 } 85 return false; 86 } 87 88 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 89 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]); 90 ResourceMark rm; 91 92 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 93 // to create space for them in ConnectionGraph::_nodes[]. 94 Node* oop_null = igvn->zerocon(T_OBJECT); 95 Node* noop_null = igvn->zerocon(T_NARROWOOP); 96 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 97 // Perform escape analysis 98 if (congraph->compute_escape()) { 99 // There are non escaping objects. 100 C->set_congraph(congraph); 101 } 102 // Cleanup. 103 if (oop_null->outcnt() == 0) 104 igvn->hash_delete(oop_null); 105 if (noop_null->outcnt() == 0) 106 igvn->hash_delete(noop_null); 107 } 108 109 bool ConnectionGraph::compute_escape() { 110 Compile* C = _compile; 111 PhaseGVN* igvn = _igvn; 112 113 // Worklists used by EA. 114 Unique_Node_List delayed_worklist; 115 GrowableArray<Node*> alloc_worklist; 116 GrowableArray<Node*> ptr_cmp_worklist; 117 GrowableArray<Node*> storestore_worklist; 118 GrowableArray<ArrayCopyNode*> arraycopy_worklist; 119 GrowableArray<PointsToNode*> ptnodes_worklist; 120 GrowableArray<JavaObjectNode*> java_objects_worklist; 121 GrowableArray<JavaObjectNode*> non_escaped_worklist; 122 GrowableArray<FieldNode*> oop_fields_worklist; 123 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 124 125 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]); 126 127 // 1. Populate Connection Graph (CG) with PointsTo nodes. 128 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 129 // Initialize worklist 130 if (C->root() != NULL) { 131 ideal_nodes.push(C->root()); 132 } 133 // Processed ideal nodes are unique on ideal_nodes list 134 // but several ideal nodes are mapped to the phantom_obj. 135 // To avoid duplicated entries on the following worklists 136 // add the phantom_obj only once to them. 137 ptnodes_worklist.append(phantom_obj); 138 java_objects_worklist.append(phantom_obj); 139 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 140 Node* n = ideal_nodes.at(next); 141 // Create PointsTo nodes and add them to Connection Graph. Called 142 // only once per ideal node since ideal_nodes is Unique_Node list. 143 add_node_to_connection_graph(n, &delayed_worklist); 144 PointsToNode* ptn = ptnode_adr(n->_idx); 145 if (ptn != NULL && ptn != phantom_obj) { 146 ptnodes_worklist.append(ptn); 147 if (ptn->is_JavaObject()) { 148 java_objects_worklist.append(ptn->as_JavaObject()); 149 if ((n->is_Allocate() || n->is_CallStaticJava()) && 150 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 151 // Only allocations and java static calls results are interesting. 152 non_escaped_worklist.append(ptn->as_JavaObject()); 153 } 154 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 155 oop_fields_worklist.append(ptn->as_Field()); 156 } 157 } 158 if (n->is_MergeMem()) { 159 // Collect all MergeMem nodes to add memory slices for 160 // scalar replaceable objects in split_unique_types(). 161 _mergemem_worklist.append(n->as_MergeMem()); 162 } else if (OptimizePtrCompare && n->is_Cmp() && 163 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) { 164 // Collect compare pointers nodes. 165 ptr_cmp_worklist.append(n); 166 } else if (n->is_MemBarStoreStore()) { 167 // Collect all MemBarStoreStore nodes so that depending on the 168 // escape status of the associated Allocate node some of them 169 // may be eliminated. 170 storestore_worklist.append(n); 171 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 172 (n->req() > MemBarNode::Precedent)) { 173 record_for_optimizer(n); 174 #ifdef ASSERT 175 } else if (n->is_AddP()) { 176 // Collect address nodes for graph verification. 177 addp_worklist.append(n); 178 #endif 179 } else if (n->is_ArrayCopy()) { 180 // Keep a list of ArrayCopy nodes so if one of its input is non 181 // escaping, we can record a unique type 182 arraycopy_worklist.append(n->as_ArrayCopy()); 183 } 184 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 185 Node* m = n->fast_out(i); // Get user 186 ideal_nodes.push(m); 187 } 188 } 189 if (non_escaped_worklist.length() == 0) { 190 _collecting = false; 191 return false; // Nothing to do. 192 } 193 // Add final simple edges to graph. 194 while(delayed_worklist.size() > 0) { 195 Node* n = delayed_worklist.pop(); 196 add_final_edges(n); 197 } 198 int ptnodes_length = ptnodes_worklist.length(); 199 200 #ifdef ASSERT 201 if (VerifyConnectionGraph) { 202 // Verify that no new simple edges could be created and all 203 // local vars has edges. 204 _verify = true; 205 for (int next = 0; next < ptnodes_length; ++next) { 206 PointsToNode* ptn = ptnodes_worklist.at(next); 207 add_final_edges(ptn->ideal_node()); 208 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 209 ptn->dump(); 210 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 211 } 212 } 213 _verify = false; 214 } 215 #endif 216 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes 217 // processing, calls to CI to resolve symbols (types, fields, methods) 218 // referenced in bytecode. During symbol resolution VM may throw 219 // an exception which CI cleans and converts to compilation failure. 220 if (C->failing()) return false; 221 222 // 2. Finish Graph construction by propagating references to all 223 // java objects through graph. 224 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 225 java_objects_worklist, oop_fields_worklist)) { 226 // All objects escaped or hit time or iterations limits. 227 _collecting = false; 228 return false; 229 } 230 231 // 3. Adjust scalar_replaceable state of nonescaping objects and push 232 // scalar replaceable allocations on alloc_worklist for processing 233 // in split_unique_types(). 234 int non_escaped_length = non_escaped_worklist.length(); 235 for (int next = 0; next < non_escaped_length; next++) { 236 JavaObjectNode* ptn = non_escaped_worklist.at(next); 237 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 238 Node* n = ptn->ideal_node(); 239 if (n->is_Allocate()) { 240 n->as_Allocate()->_is_non_escaping = noescape; 241 } 242 if (n->is_CallStaticJava()) { 243 n->as_CallStaticJava()->_is_non_escaping = noescape; 244 } 245 if (noescape && ptn->scalar_replaceable()) { 246 adjust_scalar_replaceable_state(ptn); 247 if (ptn->scalar_replaceable()) { 248 alloc_worklist.append(ptn->ideal_node()); 249 } 250 } 251 } 252 253 #ifdef ASSERT 254 if (VerifyConnectionGraph) { 255 // Verify that graph is complete - no new edges could be added or needed. 256 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 257 java_objects_worklist, addp_worklist); 258 } 259 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 260 assert(null_obj->escape_state() == PointsToNode::NoEscape && 261 null_obj->edge_count() == 0 && 262 !null_obj->arraycopy_src() && 263 !null_obj->arraycopy_dst(), "sanity"); 264 #endif 265 266 _collecting = false; 267 268 } // TracePhase t3("connectionGraph") 269 270 // 4. Optimize ideal graph based on EA information. 271 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 272 if (has_non_escaping_obj) { 273 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 274 } 275 276 #ifndef PRODUCT 277 if (PrintEscapeAnalysis) { 278 dump(ptnodes_worklist); // Dump ConnectionGraph 279 } 280 #endif 281 282 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 283 #ifdef ASSERT 284 if (VerifyConnectionGraph) { 285 int alloc_length = alloc_worklist.length(); 286 for (int next = 0; next < alloc_length; ++next) { 287 Node* n = alloc_worklist.at(next); 288 PointsToNode* ptn = ptnode_adr(n->_idx); 289 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 290 } 291 } 292 #endif 293 294 // 5. Separate memory graph for scalar replaceable allcations. 295 if (has_scalar_replaceable_candidates && 296 C->AliasLevel() >= 3 && EliminateAllocations) { 297 // Now use the escape information to create unique types for 298 // scalar replaceable objects. 299 split_unique_types(alloc_worklist, arraycopy_worklist); 300 if (C->failing()) return false; 301 C->print_method(PHASE_AFTER_EA, 2); 302 303 #ifdef ASSERT 304 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 305 tty->print("=== No allocations eliminated for "); 306 C->method()->print_short_name(); 307 if(!EliminateAllocations) { 308 tty->print(" since EliminateAllocations is off ==="); 309 } else if(!has_scalar_replaceable_candidates) { 310 tty->print(" since there are no scalar replaceable candidates ==="); 311 } else if(C->AliasLevel() < 3) { 312 tty->print(" since AliasLevel < 3 ==="); 313 } 314 tty->cr(); 315 #endif 316 } 317 return has_non_escaping_obj; 318 } 319 320 // Utility function for nodes that load an object 321 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 322 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 323 // ThreadLocal has RawPtr type. 324 const Type* t = _igvn->type(n); 325 if (t->make_ptr() != NULL) { 326 Node* adr = n->in(MemNode::Address); 327 #ifdef ASSERT 328 if (!adr->is_AddP()) { 329 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 330 } else { 331 assert((ptnode_adr(adr->_idx) == NULL || 332 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 333 } 334 #endif 335 add_local_var_and_edge(n, PointsToNode::NoEscape, 336 adr, delayed_worklist); 337 } 338 } 339 340 // Populate Connection Graph with PointsTo nodes and create simple 341 // connection graph edges. 342 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 343 assert(!_verify, "this method should not be called for verification"); 344 PhaseGVN* igvn = _igvn; 345 uint n_idx = n->_idx; 346 PointsToNode* n_ptn = ptnode_adr(n_idx); 347 if (n_ptn != NULL) 348 return; // No need to redefine PointsTo node during first iteration. 349 350 if (n->is_Call()) { 351 // Arguments to allocation and locking don't escape. 352 if (n->is_AbstractLock()) { 353 // Put Lock and Unlock nodes on IGVN worklist to process them during 354 // first IGVN optimization when escape information is still available. 355 record_for_optimizer(n); 356 } else if (n->is_Allocate()) { 357 add_call_node(n->as_Call()); 358 record_for_optimizer(n); 359 } else { 360 if (n->is_CallStaticJava()) { 361 const char* name = n->as_CallStaticJava()->_name; 362 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 363 return; // Skip uncommon traps 364 } 365 // Don't mark as processed since call's arguments have to be processed. 366 delayed_worklist->push(n); 367 // Check if a call returns an object. 368 if ((n->as_Call()->returns_pointer() && 369 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) || 370 (n->is_CallStaticJava() && 371 n->as_CallStaticJava()->is_boxing_method())) { 372 add_call_node(n->as_Call()); 373 } 374 } 375 return; 376 } 377 // Put this check here to process call arguments since some call nodes 378 // point to phantom_obj. 379 if (n_ptn == phantom_obj || n_ptn == null_obj) 380 return; // Skip predefined nodes. 381 382 int opcode = n->Opcode(); 383 switch (opcode) { 384 case Op_AddP: { 385 Node* base = get_addp_base(n); 386 PointsToNode* ptn_base = ptnode_adr(base->_idx); 387 // Field nodes are created for all field types. They are used in 388 // adjust_scalar_replaceable_state() and split_unique_types(). 389 // Note, non-oop fields will have only base edges in Connection 390 // Graph because such fields are not used for oop loads and stores. 391 int offset = address_offset(n, igvn); 392 add_field(n, PointsToNode::NoEscape, offset); 393 if (ptn_base == NULL) { 394 delayed_worklist->push(n); // Process it later. 395 } else { 396 n_ptn = ptnode_adr(n_idx); 397 add_base(n_ptn->as_Field(), ptn_base); 398 } 399 break; 400 } 401 case Op_CastX2P: { 402 map_ideal_node(n, phantom_obj); 403 break; 404 } 405 case Op_CastPP: 406 case Op_CheckCastPP: 407 case Op_EncodeP: 408 case Op_DecodeN: 409 case Op_EncodePKlass: 410 case Op_DecodeNKlass: { 411 add_local_var_and_edge(n, PointsToNode::NoEscape, 412 n->in(1), delayed_worklist); 413 break; 414 } 415 case Op_CMoveP: { 416 add_local_var(n, PointsToNode::NoEscape); 417 // Do not add edges during first iteration because some could be 418 // not defined yet. 419 delayed_worklist->push(n); 420 break; 421 } 422 case Op_ConP: 423 case Op_ConN: 424 case Op_ConNKlass: { 425 // assume all oop constants globally escape except for null 426 PointsToNode::EscapeState es; 427 const Type* t = igvn->type(n); 428 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 429 es = PointsToNode::NoEscape; 430 } else { 431 es = PointsToNode::GlobalEscape; 432 } 433 add_java_object(n, es); 434 break; 435 } 436 case Op_CreateEx: { 437 // assume that all exception objects globally escape 438 map_ideal_node(n, phantom_obj); 439 break; 440 } 441 case Op_LoadKlass: 442 case Op_LoadNKlass: { 443 // Unknown class is loaded 444 map_ideal_node(n, phantom_obj); 445 break; 446 } 447 case Op_LoadP: 448 case Op_LoadN: 449 case Op_LoadPLocked: { 450 add_objload_to_connection_graph(n, delayed_worklist); 451 break; 452 } 453 case Op_Parm: { 454 map_ideal_node(n, phantom_obj); 455 break; 456 } 457 case Op_PartialSubtypeCheck: { 458 // Produces Null or notNull and is used in only in CmpP so 459 // phantom_obj could be used. 460 map_ideal_node(n, phantom_obj); // Result is unknown 461 break; 462 } 463 case Op_Phi: { 464 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 465 // ThreadLocal has RawPtr type. 466 const Type* t = n->as_Phi()->type(); 467 if (t->make_ptr() != NULL) { 468 add_local_var(n, PointsToNode::NoEscape); 469 // Do not add edges during first iteration because some could be 470 // not defined yet. 471 delayed_worklist->push(n); 472 } 473 break; 474 } 475 case Op_Proj: { 476 // we are only interested in the oop result projection from a call 477 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 478 n->in(0)->as_Call()->returns_pointer()) { 479 add_local_var_and_edge(n, PointsToNode::NoEscape, 480 n->in(0), delayed_worklist); 481 } 482 break; 483 } 484 case Op_Rethrow: // Exception object escapes 485 case Op_Return: { 486 if (n->req() > TypeFunc::Parms && 487 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 488 // Treat Return value as LocalVar with GlobalEscape escape state. 489 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 490 n->in(TypeFunc::Parms), delayed_worklist); 491 } 492 break; 493 } 494 case Op_CompareAndExchangeP: 495 case Op_CompareAndExchangeN: 496 case Op_GetAndSetP: 497 case Op_GetAndSetN: { 498 add_objload_to_connection_graph(n, delayed_worklist); 499 // fallthrough 500 } 501 case Op_StoreP: 502 case Op_StoreN: 503 case Op_StoreNKlass: 504 case Op_StorePConditional: 505 case Op_WeakCompareAndSwapP: 506 case Op_WeakCompareAndSwapN: 507 case Op_CompareAndSwapP: 508 case Op_CompareAndSwapN: { 509 Node* adr = n->in(MemNode::Address); 510 const Type *adr_type = igvn->type(adr); 511 adr_type = adr_type->make_ptr(); 512 if (adr_type == NULL) { 513 break; // skip dead nodes 514 } 515 if (adr_type->isa_oopptr() || 516 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 517 (adr_type == TypeRawPtr::NOTNULL && 518 adr->in(AddPNode::Address)->is_Proj() && 519 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 520 delayed_worklist->push(n); // Process it later. 521 #ifdef ASSERT 522 assert(adr->is_AddP(), "expecting an AddP"); 523 if (adr_type == TypeRawPtr::NOTNULL) { 524 // Verify a raw address for a store captured by Initialize node. 525 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 526 assert(offs != Type::OffsetBot, "offset must be a constant"); 527 } 528 #endif 529 } else { 530 // Ignore copy the displaced header to the BoxNode (OSR compilation). 531 if (adr->is_BoxLock()) 532 break; 533 // Stored value escapes in unsafe access. 534 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 535 // Pointer stores in G1 barriers looks like unsafe access. 536 // Ignore such stores to be able scalar replace non-escaping 537 // allocations. 538 if (UseG1GC && adr->is_AddP()) { 539 Node* base = get_addp_base(adr); 540 if (base->Opcode() == Op_LoadP && 541 base->in(MemNode::Address)->is_AddP()) { 542 adr = base->in(MemNode::Address); 543 Node* tls = get_addp_base(adr); 544 if (tls->Opcode() == Op_ThreadLocal) { 545 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 546 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() + 547 SATBMarkQueue::byte_offset_of_buf())) { 548 break; // G1 pre barrier previous oop value store. 549 } 550 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() + 551 DirtyCardQueue::byte_offset_of_buf())) { 552 break; // G1 post barrier card address store. 553 } 554 } 555 } 556 } 557 delayed_worklist->push(n); // Process unsafe access later. 558 break; 559 } 560 #ifdef ASSERT 561 n->dump(1); 562 assert(false, "not unsafe or G1 barrier raw StoreP"); 563 #endif 564 } 565 break; 566 } 567 case Op_AryEq: 568 case Op_HasNegatives: 569 case Op_StrComp: 570 case Op_StrEquals: 571 case Op_StrIndexOf: 572 case Op_StrIndexOfChar: 573 case Op_StrInflatedCopy: 574 case Op_StrCompressedCopy: 575 case Op_EncodeISOArray: { 576 add_local_var(n, PointsToNode::ArgEscape); 577 delayed_worklist->push(n); // Process it later. 578 break; 579 } 580 case Op_ThreadLocal: { 581 add_java_object(n, PointsToNode::ArgEscape); 582 break; 583 } 584 default: 585 ; // Do nothing for nodes not related to EA. 586 } 587 return; 588 } 589 590 #ifdef ASSERT 591 #define ELSE_FAIL(name) \ 592 /* Should not be called for not pointer type. */ \ 593 n->dump(1); \ 594 assert(false, name); \ 595 break; 596 #else 597 #define ELSE_FAIL(name) \ 598 break; 599 #endif 600 601 // Add final simple edges to graph. 602 void ConnectionGraph::add_final_edges(Node *n) { 603 PointsToNode* n_ptn = ptnode_adr(n->_idx); 604 #ifdef ASSERT 605 if (_verify && n_ptn->is_JavaObject()) 606 return; // This method does not change graph for JavaObject. 607 #endif 608 609 if (n->is_Call()) { 610 process_call_arguments(n->as_Call()); 611 return; 612 } 613 assert(n->is_Store() || n->is_LoadStore() || 614 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 615 "node should be registered already"); 616 int opcode = n->Opcode(); 617 switch (opcode) { 618 case Op_AddP: { 619 Node* base = get_addp_base(n); 620 PointsToNode* ptn_base = ptnode_adr(base->_idx); 621 assert(ptn_base != NULL, "field's base should be registered"); 622 add_base(n_ptn->as_Field(), ptn_base); 623 break; 624 } 625 case Op_CastPP: 626 case Op_CheckCastPP: 627 case Op_EncodeP: 628 case Op_DecodeN: 629 case Op_EncodePKlass: 630 case Op_DecodeNKlass: { 631 add_local_var_and_edge(n, PointsToNode::NoEscape, 632 n->in(1), NULL); 633 break; 634 } 635 case Op_CMoveP: { 636 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 637 Node* in = n->in(i); 638 if (in == NULL) 639 continue; // ignore NULL 640 Node* uncast_in = in->uncast(); 641 if (uncast_in->is_top() || uncast_in == n) 642 continue; // ignore top or inputs which go back this node 643 PointsToNode* ptn = ptnode_adr(in->_idx); 644 assert(ptn != NULL, "node should be registered"); 645 add_edge(n_ptn, ptn); 646 } 647 break; 648 } 649 case Op_LoadP: 650 case Op_LoadN: 651 case Op_LoadPLocked: { 652 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 653 // ThreadLocal has RawPtr type. 654 const Type* t = _igvn->type(n); 655 if (t->make_ptr() != NULL) { 656 Node* adr = n->in(MemNode::Address); 657 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 658 break; 659 } 660 ELSE_FAIL("Op_LoadP"); 661 } 662 case Op_Phi: { 663 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 664 // ThreadLocal has RawPtr type. 665 const Type* t = n->as_Phi()->type(); 666 if (t->make_ptr() != NULL) { 667 for (uint i = 1; i < n->req(); i++) { 668 Node* in = n->in(i); 669 if (in == NULL) 670 continue; // ignore NULL 671 Node* uncast_in = in->uncast(); 672 if (uncast_in->is_top() || uncast_in == n) 673 continue; // ignore top or inputs which go back this node 674 PointsToNode* ptn = ptnode_adr(in->_idx); 675 assert(ptn != NULL, "node should be registered"); 676 add_edge(n_ptn, ptn); 677 } 678 break; 679 } 680 ELSE_FAIL("Op_Phi"); 681 } 682 case Op_Proj: { 683 // we are only interested in the oop result projection from a call 684 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 685 n->in(0)->as_Call()->returns_pointer()) { 686 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 687 break; 688 } 689 ELSE_FAIL("Op_Proj"); 690 } 691 case Op_Rethrow: // Exception object escapes 692 case Op_Return: { 693 if (n->req() > TypeFunc::Parms && 694 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 695 // Treat Return value as LocalVar with GlobalEscape escape state. 696 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 697 n->in(TypeFunc::Parms), NULL); 698 break; 699 } 700 ELSE_FAIL("Op_Return"); 701 } 702 case Op_StoreP: 703 case Op_StoreN: 704 case Op_StoreNKlass: 705 case Op_StorePConditional: 706 case Op_CompareAndExchangeP: 707 case Op_CompareAndExchangeN: 708 case Op_CompareAndSwapP: 709 case Op_CompareAndSwapN: 710 case Op_WeakCompareAndSwapP: 711 case Op_WeakCompareAndSwapN: 712 case Op_GetAndSetP: 713 case Op_GetAndSetN: { 714 Node* adr = n->in(MemNode::Address); 715 const Type *adr_type = _igvn->type(adr); 716 adr_type = adr_type->make_ptr(); 717 #ifdef ASSERT 718 if (adr_type == NULL) { 719 n->dump(1); 720 assert(adr_type != NULL, "dead node should not be on list"); 721 break; 722 } 723 #endif 724 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 725 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 726 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 727 } 728 if (adr_type->isa_oopptr() || 729 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 730 (adr_type == TypeRawPtr::NOTNULL && 731 adr->in(AddPNode::Address)->is_Proj() && 732 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 733 // Point Address to Value 734 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 735 assert(adr_ptn != NULL && 736 adr_ptn->as_Field()->is_oop(), "node should be registered"); 737 Node *val = n->in(MemNode::ValueIn); 738 PointsToNode* ptn = ptnode_adr(val->_idx); 739 assert(ptn != NULL, "node should be registered"); 740 add_edge(adr_ptn, ptn); 741 break; 742 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 743 // Stored value escapes in unsafe access. 744 Node *val = n->in(MemNode::ValueIn); 745 PointsToNode* ptn = ptnode_adr(val->_idx); 746 assert(ptn != NULL, "node should be registered"); 747 set_escape_state(ptn, PointsToNode::GlobalEscape); 748 // Add edge to object for unsafe access with offset. 749 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 750 assert(adr_ptn != NULL, "node should be registered"); 751 if (adr_ptn->is_Field()) { 752 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 753 add_edge(adr_ptn, ptn); 754 } 755 break; 756 } 757 ELSE_FAIL("Op_StoreP"); 758 } 759 case Op_AryEq: 760 case Op_HasNegatives: 761 case Op_StrComp: 762 case Op_StrEquals: 763 case Op_StrIndexOf: 764 case Op_StrIndexOfChar: 765 case Op_StrInflatedCopy: 766 case Op_StrCompressedCopy: 767 case Op_EncodeISOArray: { 768 // char[]/byte[] arrays passed to string intrinsic do not escape but 769 // they are not scalar replaceable. Adjust escape state for them. 770 // Start from in(2) edge since in(1) is memory edge. 771 for (uint i = 2; i < n->req(); i++) { 772 Node* adr = n->in(i); 773 const Type* at = _igvn->type(adr); 774 if (!adr->is_top() && at->isa_ptr()) { 775 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 776 at->isa_ptr() != NULL, "expecting a pointer"); 777 if (adr->is_AddP()) { 778 adr = get_addp_base(adr); 779 } 780 PointsToNode* ptn = ptnode_adr(adr->_idx); 781 assert(ptn != NULL, "node should be registered"); 782 add_edge(n_ptn, ptn); 783 } 784 } 785 break; 786 } 787 default: { 788 // This method should be called only for EA specific nodes which may 789 // miss some edges when they were created. 790 #ifdef ASSERT 791 n->dump(1); 792 #endif 793 guarantee(false, "unknown node"); 794 } 795 } 796 return; 797 } 798 799 void ConnectionGraph::add_call_node(CallNode* call) { 800 assert(call->returns_pointer(), "only for call which returns pointer"); 801 uint call_idx = call->_idx; 802 if (call->is_Allocate()) { 803 Node* k = call->in(AllocateNode::KlassNode); 804 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 805 assert(kt != NULL, "TypeKlassPtr required."); 806 ciKlass* cik = kt->klass(); 807 PointsToNode::EscapeState es = PointsToNode::NoEscape; 808 bool scalar_replaceable = true; 809 if (call->is_AllocateArray()) { 810 if (!cik->is_array_klass()) { // StressReflectiveCode 811 es = PointsToNode::GlobalEscape; 812 } else { 813 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 814 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 815 // Not scalar replaceable if the length is not constant or too big. 816 scalar_replaceable = false; 817 } 818 } 819 } else { // Allocate instance 820 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 821 cik->is_subclass_of(_compile->env()->Reference_klass()) || 822 !cik->is_instance_klass() || // StressReflectiveCode 823 !cik->as_instance_klass()->can_be_instantiated() || 824 cik->as_instance_klass()->has_finalizer()) { 825 es = PointsToNode::GlobalEscape; 826 } 827 } 828 add_java_object(call, es); 829 PointsToNode* ptn = ptnode_adr(call_idx); 830 if (!scalar_replaceable && ptn->scalar_replaceable()) { 831 ptn->set_scalar_replaceable(false); 832 } 833 } else if (call->is_CallStaticJava()) { 834 // Call nodes could be different types: 835 // 836 // 1. CallDynamicJavaNode (what happened during call is unknown): 837 // 838 // - mapped to GlobalEscape JavaObject node if oop is returned; 839 // 840 // - all oop arguments are escaping globally; 841 // 842 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 843 // 844 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 845 // 846 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 847 // - mapped to NoEscape JavaObject node if non-escaping object allocated 848 // during call is returned; 849 // - mapped to ArgEscape LocalVar node pointed to object arguments 850 // which are returned and does not escape during call; 851 // 852 // - oop arguments escaping status is defined by bytecode analysis; 853 // 854 // For a static call, we know exactly what method is being called. 855 // Use bytecode estimator to record whether the call's return value escapes. 856 ciMethod* meth = call->as_CallJava()->method(); 857 if (meth == NULL) { 858 const char* name = call->as_CallStaticJava()->_name; 859 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 860 // Returns a newly allocated unescaped object. 861 add_java_object(call, PointsToNode::NoEscape); 862 ptnode_adr(call_idx)->set_scalar_replaceable(false); 863 } else if (meth->is_boxing_method()) { 864 // Returns boxing object 865 PointsToNode::EscapeState es; 866 vmIntrinsics::ID intr = meth->intrinsic_id(); 867 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 868 // It does not escape if object is always allocated. 869 es = PointsToNode::NoEscape; 870 } else { 871 // It escapes globally if object could be loaded from cache. 872 es = PointsToNode::GlobalEscape; 873 } 874 add_java_object(call, es); 875 } else { 876 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 877 call_analyzer->copy_dependencies(_compile->dependencies()); 878 if (call_analyzer->is_return_allocated()) { 879 // Returns a newly allocated unescaped object, simply 880 // update dependency information. 881 // Mark it as NoEscape so that objects referenced by 882 // it's fields will be marked as NoEscape at least. 883 add_java_object(call, PointsToNode::NoEscape); 884 ptnode_adr(call_idx)->set_scalar_replaceable(false); 885 } else { 886 // Determine whether any arguments are returned. 887 const TypeTuple* d = call->tf()->domain(); 888 bool ret_arg = false; 889 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 890 if (d->field_at(i)->isa_ptr() != NULL && 891 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 892 ret_arg = true; 893 break; 894 } 895 } 896 if (ret_arg) { 897 add_local_var(call, PointsToNode::ArgEscape); 898 } else { 899 // Returns unknown object. 900 map_ideal_node(call, phantom_obj); 901 } 902 } 903 } 904 } else { 905 // An other type of call, assume the worst case: 906 // returned value is unknown and globally escapes. 907 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 908 map_ideal_node(call, phantom_obj); 909 } 910 } 911 912 void ConnectionGraph::process_call_arguments(CallNode *call) { 913 bool is_arraycopy = false; 914 switch (call->Opcode()) { 915 #ifdef ASSERT 916 case Op_Allocate: 917 case Op_AllocateArray: 918 case Op_Lock: 919 case Op_Unlock: 920 assert(false, "should be done already"); 921 break; 922 #endif 923 case Op_ArrayCopy: 924 case Op_CallLeafNoFP: 925 // Most array copies are ArrayCopy nodes at this point but there 926 // are still a few direct calls to the copy subroutines (See 927 // PhaseStringOpts::copy_string()) 928 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 929 call->as_CallLeaf()->is_call_to_arraycopystub(); 930 // fall through 931 case Op_CallLeaf: { 932 // Stub calls, objects do not escape but they are not scale replaceable. 933 // Adjust escape state for outgoing arguments. 934 const TypeTuple * d = call->tf()->domain(); 935 bool src_has_oops = false; 936 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 937 const Type* at = d->field_at(i); 938 Node *arg = call->in(i); 939 if (arg == NULL) { 940 continue; 941 } 942 const Type *aat = _igvn->type(arg); 943 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 944 continue; 945 if (arg->is_AddP()) { 946 // 947 // The inline_native_clone() case when the arraycopy stub is called 948 // after the allocation before Initialize and CheckCastPP nodes. 949 // Or normal arraycopy for object arrays case. 950 // 951 // Set AddP's base (Allocate) as not scalar replaceable since 952 // pointer to the base (with offset) is passed as argument. 953 // 954 arg = get_addp_base(arg); 955 } 956 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 957 assert(arg_ptn != NULL, "should be registered"); 958 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 959 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 960 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 961 aat->isa_ptr() != NULL, "expecting an Ptr"); 962 bool arg_has_oops = aat->isa_oopptr() && 963 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 964 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 965 if (i == TypeFunc::Parms) { 966 src_has_oops = arg_has_oops; 967 } 968 // 969 // src or dst could be j.l.Object when other is basic type array: 970 // 971 // arraycopy(char[],0,Object*,0,size); 972 // arraycopy(Object*,0,char[],0,size); 973 // 974 // Don't add edges in such cases. 975 // 976 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 977 arg_has_oops && (i > TypeFunc::Parms); 978 #ifdef ASSERT 979 if (!(is_arraycopy || 980 (call->as_CallLeaf()->_name != NULL && 981 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 982 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 || 983 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 984 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 985 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 986 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 987 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 988 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 989 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 990 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 || 991 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 992 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 993 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 994 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 995 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 996 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 997 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 998 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 999 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 1000 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 1001 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 1002 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 || 1003 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0) 1004 ))) { 1005 call->dump(); 1006 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name); 1007 } 1008 #endif 1009 // Always process arraycopy's destination object since 1010 // we need to add all possible edges to references in 1011 // source object. 1012 if (arg_esc >= PointsToNode::ArgEscape && 1013 !arg_is_arraycopy_dest) { 1014 continue; 1015 } 1016 PointsToNode::EscapeState es = PointsToNode::ArgEscape; 1017 if (call->is_ArrayCopy()) { 1018 ArrayCopyNode* ac = call->as_ArrayCopy(); 1019 if (ac->is_clonebasic() || 1020 ac->is_arraycopy_validated() || 1021 ac->is_copyof_validated() || 1022 ac->is_copyofrange_validated()) { 1023 es = PointsToNode::NoEscape; 1024 } 1025 } 1026 set_escape_state(arg_ptn, es); 1027 if (arg_is_arraycopy_dest) { 1028 Node* src = call->in(TypeFunc::Parms); 1029 if (src->is_AddP()) { 1030 src = get_addp_base(src); 1031 } 1032 PointsToNode* src_ptn = ptnode_adr(src->_idx); 1033 assert(src_ptn != NULL, "should be registered"); 1034 if (arg_ptn != src_ptn) { 1035 // Special arraycopy edge: 1036 // A destination object's field can't have the source object 1037 // as base since objects escape states are not related. 1038 // Only escape state of destination object's fields affects 1039 // escape state of fields in source object. 1040 add_arraycopy(call, es, src_ptn, arg_ptn); 1041 } 1042 } 1043 } 1044 } 1045 break; 1046 } 1047 case Op_CallStaticJava: { 1048 // For a static call, we know exactly what method is being called. 1049 // Use bytecode estimator to record the call's escape affects 1050 #ifdef ASSERT 1051 const char* name = call->as_CallStaticJava()->_name; 1052 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1053 #endif 1054 ciMethod* meth = call->as_CallJava()->method(); 1055 if ((meth != NULL) && meth->is_boxing_method()) { 1056 break; // Boxing methods do not modify any oops. 1057 } 1058 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1059 // fall-through if not a Java method or no analyzer information 1060 if (call_analyzer != NULL) { 1061 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1062 const TypeTuple* d = call->tf()->domain(); 1063 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1064 const Type* at = d->field_at(i); 1065 int k = i - TypeFunc::Parms; 1066 Node* arg = call->in(i); 1067 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1068 if (at->isa_ptr() != NULL && 1069 call_analyzer->is_arg_returned(k)) { 1070 // The call returns arguments. 1071 if (call_ptn != NULL) { // Is call's result used? 1072 assert(call_ptn->is_LocalVar(), "node should be registered"); 1073 assert(arg_ptn != NULL, "node should be registered"); 1074 add_edge(call_ptn, arg_ptn); 1075 } 1076 } 1077 if (at->isa_oopptr() != NULL && 1078 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1079 if (!call_analyzer->is_arg_stack(k)) { 1080 // The argument global escapes 1081 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1082 } else { 1083 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1084 if (!call_analyzer->is_arg_local(k)) { 1085 // The argument itself doesn't escape, but any fields might 1086 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1087 } 1088 } 1089 } 1090 } 1091 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1092 // The call returns arguments. 1093 assert(call_ptn->edge_count() > 0, "sanity"); 1094 if (!call_analyzer->is_return_local()) { 1095 // Returns also unknown object. 1096 add_edge(call_ptn, phantom_obj); 1097 } 1098 } 1099 break; 1100 } 1101 } 1102 default: { 1103 // Fall-through here if not a Java method or no analyzer information 1104 // or some other type of call, assume the worst case: all arguments 1105 // globally escape. 1106 const TypeTuple* d = call->tf()->domain(); 1107 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1108 const Type* at = d->field_at(i); 1109 if (at->isa_oopptr() != NULL) { 1110 Node* arg = call->in(i); 1111 if (arg->is_AddP()) { 1112 arg = get_addp_base(arg); 1113 } 1114 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1115 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1116 } 1117 } 1118 } 1119 } 1120 } 1121 1122 1123 // Finish Graph construction. 1124 bool ConnectionGraph::complete_connection_graph( 1125 GrowableArray<PointsToNode*>& ptnodes_worklist, 1126 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1127 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1128 GrowableArray<FieldNode*>& oop_fields_worklist) { 1129 // Normally only 1-3 passes needed to build Connection Graph depending 1130 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1131 // Set limit to 20 to catch situation when something did go wrong and 1132 // bailout Escape Analysis. 1133 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1134 #define CG_BUILD_ITER_LIMIT 20 1135 1136 // Propagate GlobalEscape and ArgEscape escape states and check that 1137 // we still have non-escaping objects. The method pushs on _worklist 1138 // Field nodes which reference phantom_object. 1139 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1140 return false; // Nothing to do. 1141 } 1142 // Now propagate references to all JavaObject nodes. 1143 int java_objects_length = java_objects_worklist.length(); 1144 elapsedTimer time; 1145 bool timeout = false; 1146 int new_edges = 1; 1147 int iterations = 0; 1148 do { 1149 while ((new_edges > 0) && 1150 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1151 double start_time = time.seconds(); 1152 time.start(); 1153 new_edges = 0; 1154 // Propagate references to phantom_object for nodes pushed on _worklist 1155 // by find_non_escaped_objects() and find_field_value(). 1156 new_edges += add_java_object_edges(phantom_obj, false); 1157 for (int next = 0; next < java_objects_length; ++next) { 1158 JavaObjectNode* ptn = java_objects_worklist.at(next); 1159 new_edges += add_java_object_edges(ptn, true); 1160 1161 #define SAMPLE_SIZE 4 1162 if ((next % SAMPLE_SIZE) == 0) { 1163 // Each 4 iterations calculate how much time it will take 1164 // to complete graph construction. 1165 time.stop(); 1166 // Poll for requests from shutdown mechanism to quiesce compiler 1167 // because Connection graph construction may take long time. 1168 CompileBroker::maybe_block(); 1169 double stop_time = time.seconds(); 1170 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1171 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1172 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1173 timeout = true; 1174 break; // Timeout 1175 } 1176 start_time = stop_time; 1177 time.start(); 1178 } 1179 #undef SAMPLE_SIZE 1180 1181 } 1182 if (timeout) break; 1183 if (new_edges > 0) { 1184 // Update escape states on each iteration if graph was updated. 1185 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1186 return false; // Nothing to do. 1187 } 1188 } 1189 time.stop(); 1190 if (time.seconds() >= EscapeAnalysisTimeout) { 1191 timeout = true; 1192 break; 1193 } 1194 } 1195 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1196 time.start(); 1197 // Find fields which have unknown value. 1198 int fields_length = oop_fields_worklist.length(); 1199 for (int next = 0; next < fields_length; next++) { 1200 FieldNode* field = oop_fields_worklist.at(next); 1201 if (field->edge_count() == 0) { 1202 new_edges += find_field_value(field); 1203 // This code may added new edges to phantom_object. 1204 // Need an other cycle to propagate references to phantom_object. 1205 } 1206 } 1207 time.stop(); 1208 if (time.seconds() >= EscapeAnalysisTimeout) { 1209 timeout = true; 1210 break; 1211 } 1212 } else { 1213 new_edges = 0; // Bailout 1214 } 1215 } while (new_edges > 0); 1216 1217 // Bailout if passed limits. 1218 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1219 Compile* C = _compile; 1220 if (C->log() != NULL) { 1221 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1222 C->log()->text("%s", timeout ? "time" : "iterations"); 1223 C->log()->end_elem(" limit'"); 1224 } 1225 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1226 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()); 1227 // Possible infinite build_connection_graph loop, 1228 // bailout (no changes to ideal graph were made). 1229 return false; 1230 } 1231 #ifdef ASSERT 1232 if (Verbose && PrintEscapeAnalysis) { 1233 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1234 iterations, nodes_size(), ptnodes_worklist.length()); 1235 } 1236 #endif 1237 1238 #undef CG_BUILD_ITER_LIMIT 1239 1240 // Find fields initialized by NULL for non-escaping Allocations. 1241 int non_escaped_length = non_escaped_worklist.length(); 1242 for (int next = 0; next < non_escaped_length; next++) { 1243 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1244 PointsToNode::EscapeState es = ptn->escape_state(); 1245 assert(es <= PointsToNode::ArgEscape, "sanity"); 1246 if (es == PointsToNode::NoEscape) { 1247 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1248 // Adding references to NULL object does not change escape states 1249 // since it does not escape. Also no fields are added to NULL object. 1250 add_java_object_edges(null_obj, false); 1251 } 1252 } 1253 Node* n = ptn->ideal_node(); 1254 if (n->is_Allocate()) { 1255 // The object allocated by this Allocate node will never be 1256 // seen by an other thread. Mark it so that when it is 1257 // expanded no MemBarStoreStore is added. 1258 InitializeNode* ini = n->as_Allocate()->initialization(); 1259 if (ini != NULL) 1260 ini->set_does_not_escape(); 1261 } 1262 } 1263 return true; // Finished graph construction. 1264 } 1265 1266 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1267 // and check that we still have non-escaping java objects. 1268 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1269 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1270 GrowableArray<PointsToNode*> escape_worklist; 1271 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1272 int ptnodes_length = ptnodes_worklist.length(); 1273 for (int next = 0; next < ptnodes_length; ++next) { 1274 PointsToNode* ptn = ptnodes_worklist.at(next); 1275 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1276 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1277 escape_worklist.push(ptn); 1278 } 1279 } 1280 // Set escape states to referenced nodes (edges list). 1281 while (escape_worklist.length() > 0) { 1282 PointsToNode* ptn = escape_worklist.pop(); 1283 PointsToNode::EscapeState es = ptn->escape_state(); 1284 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1285 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1286 es >= PointsToNode::ArgEscape) { 1287 // GlobalEscape or ArgEscape state of field means it has unknown value. 1288 if (add_edge(ptn, phantom_obj)) { 1289 // New edge was added 1290 add_field_uses_to_worklist(ptn->as_Field()); 1291 } 1292 } 1293 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1294 PointsToNode* e = i.get(); 1295 if (e->is_Arraycopy()) { 1296 assert(ptn->arraycopy_dst(), "sanity"); 1297 // Propagate only fields escape state through arraycopy edge. 1298 if (e->fields_escape_state() < field_es) { 1299 set_fields_escape_state(e, field_es); 1300 escape_worklist.push(e); 1301 } 1302 } else if (es >= field_es) { 1303 // fields_escape_state is also set to 'es' if it is less than 'es'. 1304 if (e->escape_state() < es) { 1305 set_escape_state(e, es); 1306 escape_worklist.push(e); 1307 } 1308 } else { 1309 // Propagate field escape state. 1310 bool es_changed = false; 1311 if (e->fields_escape_state() < field_es) { 1312 set_fields_escape_state(e, field_es); 1313 es_changed = true; 1314 } 1315 if ((e->escape_state() < field_es) && 1316 e->is_Field() && ptn->is_JavaObject() && 1317 e->as_Field()->is_oop()) { 1318 // Change escape state of referenced fields. 1319 set_escape_state(e, field_es); 1320 es_changed = true; 1321 } else if (e->escape_state() < es) { 1322 set_escape_state(e, es); 1323 es_changed = true; 1324 } 1325 if (es_changed) { 1326 escape_worklist.push(e); 1327 } 1328 } 1329 } 1330 } 1331 // Remove escaped objects from non_escaped list. 1332 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1333 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1334 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1335 non_escaped_worklist.delete_at(next); 1336 } 1337 if (ptn->escape_state() == PointsToNode::NoEscape) { 1338 // Find fields in non-escaped allocations which have unknown value. 1339 find_init_values(ptn, phantom_obj, NULL); 1340 } 1341 } 1342 return (non_escaped_worklist.length() > 0); 1343 } 1344 1345 // Add all references to JavaObject node by walking over all uses. 1346 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1347 int new_edges = 0; 1348 if (populate_worklist) { 1349 // Populate _worklist by uses of jobj's uses. 1350 for (UseIterator i(jobj); i.has_next(); i.next()) { 1351 PointsToNode* use = i.get(); 1352 if (use->is_Arraycopy()) 1353 continue; 1354 add_uses_to_worklist(use); 1355 if (use->is_Field() && use->as_Field()->is_oop()) { 1356 // Put on worklist all field's uses (loads) and 1357 // related field nodes (same base and offset). 1358 add_field_uses_to_worklist(use->as_Field()); 1359 } 1360 } 1361 } 1362 for (int l = 0; l < _worklist.length(); l++) { 1363 PointsToNode* use = _worklist.at(l); 1364 if (PointsToNode::is_base_use(use)) { 1365 // Add reference from jobj to field and from field to jobj (field's base). 1366 use = PointsToNode::get_use_node(use)->as_Field(); 1367 if (add_base(use->as_Field(), jobj)) { 1368 new_edges++; 1369 } 1370 continue; 1371 } 1372 assert(!use->is_JavaObject(), "sanity"); 1373 if (use->is_Arraycopy()) { 1374 if (jobj == null_obj) // NULL object does not have field edges 1375 continue; 1376 // Added edge from Arraycopy node to arraycopy's source java object 1377 if (add_edge(use, jobj)) { 1378 jobj->set_arraycopy_src(); 1379 new_edges++; 1380 } 1381 // and stop here. 1382 continue; 1383 } 1384 if (!add_edge(use, jobj)) 1385 continue; // No new edge added, there was such edge already. 1386 new_edges++; 1387 if (use->is_LocalVar()) { 1388 add_uses_to_worklist(use); 1389 if (use->arraycopy_dst()) { 1390 for (EdgeIterator i(use); i.has_next(); i.next()) { 1391 PointsToNode* e = i.get(); 1392 if (e->is_Arraycopy()) { 1393 if (jobj == null_obj) // NULL object does not have field edges 1394 continue; 1395 // Add edge from arraycopy's destination java object to Arraycopy node. 1396 if (add_edge(jobj, e)) { 1397 new_edges++; 1398 jobj->set_arraycopy_dst(); 1399 } 1400 } 1401 } 1402 } 1403 } else { 1404 // Added new edge to stored in field values. 1405 // Put on worklist all field's uses (loads) and 1406 // related field nodes (same base and offset). 1407 add_field_uses_to_worklist(use->as_Field()); 1408 } 1409 } 1410 _worklist.clear(); 1411 _in_worklist.Reset(); 1412 return new_edges; 1413 } 1414 1415 // Put on worklist all related field nodes. 1416 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1417 assert(field->is_oop(), "sanity"); 1418 int offset = field->offset(); 1419 add_uses_to_worklist(field); 1420 // Loop over all bases of this field and push on worklist Field nodes 1421 // with the same offset and base (since they may reference the same field). 1422 for (BaseIterator i(field); i.has_next(); i.next()) { 1423 PointsToNode* base = i.get(); 1424 add_fields_to_worklist(field, base); 1425 // Check if the base was source object of arraycopy and go over arraycopy's 1426 // destination objects since values stored to a field of source object are 1427 // accessable by uses (loads) of fields of destination objects. 1428 if (base->arraycopy_src()) { 1429 for (UseIterator j(base); j.has_next(); j.next()) { 1430 PointsToNode* arycp = j.get(); 1431 if (arycp->is_Arraycopy()) { 1432 for (UseIterator k(arycp); k.has_next(); k.next()) { 1433 PointsToNode* abase = k.get(); 1434 if (abase->arraycopy_dst() && abase != base) { 1435 // Look for the same arraycopy reference. 1436 add_fields_to_worklist(field, abase); 1437 } 1438 } 1439 } 1440 } 1441 } 1442 } 1443 } 1444 1445 // Put on worklist all related field nodes. 1446 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1447 int offset = field->offset(); 1448 if (base->is_LocalVar()) { 1449 for (UseIterator j(base); j.has_next(); j.next()) { 1450 PointsToNode* f = j.get(); 1451 if (PointsToNode::is_base_use(f)) { // Field 1452 f = PointsToNode::get_use_node(f); 1453 if (f == field || !f->as_Field()->is_oop()) 1454 continue; 1455 int offs = f->as_Field()->offset(); 1456 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1457 add_to_worklist(f); 1458 } 1459 } 1460 } 1461 } else { 1462 assert(base->is_JavaObject(), "sanity"); 1463 if (// Skip phantom_object since it is only used to indicate that 1464 // this field's content globally escapes. 1465 (base != phantom_obj) && 1466 // NULL object node does not have fields. 1467 (base != null_obj)) { 1468 for (EdgeIterator i(base); i.has_next(); i.next()) { 1469 PointsToNode* f = i.get(); 1470 // Skip arraycopy edge since store to destination object field 1471 // does not update value in source object field. 1472 if (f->is_Arraycopy()) { 1473 assert(base->arraycopy_dst(), "sanity"); 1474 continue; 1475 } 1476 if (f == field || !f->as_Field()->is_oop()) 1477 continue; 1478 int offs = f->as_Field()->offset(); 1479 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1480 add_to_worklist(f); 1481 } 1482 } 1483 } 1484 } 1485 } 1486 1487 // Find fields which have unknown value. 1488 int ConnectionGraph::find_field_value(FieldNode* field) { 1489 // Escaped fields should have init value already. 1490 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1491 int new_edges = 0; 1492 for (BaseIterator i(field); i.has_next(); i.next()) { 1493 PointsToNode* base = i.get(); 1494 if (base->is_JavaObject()) { 1495 // Skip Allocate's fields which will be processed later. 1496 if (base->ideal_node()->is_Allocate()) 1497 return 0; 1498 assert(base == null_obj, "only NULL ptr base expected here"); 1499 } 1500 } 1501 if (add_edge(field, phantom_obj)) { 1502 // New edge was added 1503 new_edges++; 1504 add_field_uses_to_worklist(field); 1505 } 1506 return new_edges; 1507 } 1508 1509 // Find fields initializing values for allocations. 1510 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1511 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1512 int new_edges = 0; 1513 Node* alloc = pta->ideal_node(); 1514 if (init_val == phantom_obj) { 1515 // Do nothing for Allocate nodes since its fields values are 1516 // "known" unless they are initialized by arraycopy/clone. 1517 if (alloc->is_Allocate() && !pta->arraycopy_dst()) 1518 return 0; 1519 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity"); 1520 #ifdef ASSERT 1521 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) { 1522 const char* name = alloc->as_CallStaticJava()->_name; 1523 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1524 } 1525 #endif 1526 // Non-escaped allocation returned from Java or runtime call have 1527 // unknown values in fields. 1528 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1529 PointsToNode* field = i.get(); 1530 if (field->is_Field() && field->as_Field()->is_oop()) { 1531 if (add_edge(field, phantom_obj)) { 1532 // New edge was added 1533 new_edges++; 1534 add_field_uses_to_worklist(field->as_Field()); 1535 } 1536 } 1537 } 1538 return new_edges; 1539 } 1540 assert(init_val == null_obj, "sanity"); 1541 // Do nothing for Call nodes since its fields values are unknown. 1542 if (!alloc->is_Allocate()) 1543 return 0; 1544 1545 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1546 bool visited_bottom_offset = false; 1547 GrowableArray<int> offsets_worklist; 1548 1549 // Check if an oop field's initializing value is recorded and add 1550 // a corresponding NULL if field's value if it is not recorded. 1551 // Connection Graph does not record a default initialization by NULL 1552 // captured by Initialize node. 1553 // 1554 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1555 PointsToNode* field = i.get(); // Field (AddP) 1556 if (!field->is_Field() || !field->as_Field()->is_oop()) 1557 continue; // Not oop field 1558 int offset = field->as_Field()->offset(); 1559 if (offset == Type::OffsetBot) { 1560 if (!visited_bottom_offset) { 1561 // OffsetBot is used to reference array's element, 1562 // always add reference to NULL to all Field nodes since we don't 1563 // known which element is referenced. 1564 if (add_edge(field, null_obj)) { 1565 // New edge was added 1566 new_edges++; 1567 add_field_uses_to_worklist(field->as_Field()); 1568 visited_bottom_offset = true; 1569 } 1570 } 1571 } else { 1572 // Check only oop fields. 1573 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1574 if (adr_type->isa_rawptr()) { 1575 #ifdef ASSERT 1576 // Raw pointers are used for initializing stores so skip it 1577 // since it should be recorded already 1578 Node* base = get_addp_base(field->ideal_node()); 1579 assert(adr_type->isa_rawptr() && base->is_Proj() && 1580 (base->in(0) == alloc),"unexpected pointer type"); 1581 #endif 1582 continue; 1583 } 1584 if (!offsets_worklist.contains(offset)) { 1585 offsets_worklist.append(offset); 1586 Node* value = NULL; 1587 if (ini != NULL) { 1588 // StoreP::memory_type() == T_ADDRESS 1589 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1590 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1591 // Make sure initializing store has the same type as this AddP. 1592 // This AddP may reference non existing field because it is on a 1593 // dead branch of bimorphic call which is not eliminated yet. 1594 if (store != NULL && store->is_Store() && 1595 store->as_Store()->memory_type() == ft) { 1596 value = store->in(MemNode::ValueIn); 1597 #ifdef ASSERT 1598 if (VerifyConnectionGraph) { 1599 // Verify that AddP already points to all objects the value points to. 1600 PointsToNode* val = ptnode_adr(value->_idx); 1601 assert((val != NULL), "should be processed already"); 1602 PointsToNode* missed_obj = NULL; 1603 if (val->is_JavaObject()) { 1604 if (!field->points_to(val->as_JavaObject())) { 1605 missed_obj = val; 1606 } 1607 } else { 1608 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1609 tty->print_cr("----------init store has invalid value -----"); 1610 store->dump(); 1611 val->dump(); 1612 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1613 } 1614 for (EdgeIterator j(val); j.has_next(); j.next()) { 1615 PointsToNode* obj = j.get(); 1616 if (obj->is_JavaObject()) { 1617 if (!field->points_to(obj->as_JavaObject())) { 1618 missed_obj = obj; 1619 break; 1620 } 1621 } 1622 } 1623 } 1624 if (missed_obj != NULL) { 1625 tty->print_cr("----------field---------------------------------"); 1626 field->dump(); 1627 tty->print_cr("----------missed referernce to object-----------"); 1628 missed_obj->dump(); 1629 tty->print_cr("----------object referernced by init store -----"); 1630 store->dump(); 1631 val->dump(); 1632 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1633 } 1634 } 1635 #endif 1636 } else { 1637 // There could be initializing stores which follow allocation. 1638 // For example, a volatile field store is not collected 1639 // by Initialize node. 1640 // 1641 // Need to check for dependent loads to separate such stores from 1642 // stores which follow loads. For now, add initial value NULL so 1643 // that compare pointers optimization works correctly. 1644 } 1645 } 1646 if (value == NULL) { 1647 // A field's initializing value was not recorded. Add NULL. 1648 if (add_edge(field, null_obj)) { 1649 // New edge was added 1650 new_edges++; 1651 add_field_uses_to_worklist(field->as_Field()); 1652 } 1653 } 1654 } 1655 } 1656 } 1657 return new_edges; 1658 } 1659 1660 // Adjust scalar_replaceable state after Connection Graph is built. 1661 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1662 // Search for non-escaping objects which are not scalar replaceable 1663 // and mark them to propagate the state to referenced objects. 1664 1665 // 1. An object is not scalar replaceable if the field into which it is 1666 // stored has unknown offset (stored into unknown element of an array). 1667 // 1668 for (UseIterator i(jobj); i.has_next(); i.next()) { 1669 PointsToNode* use = i.get(); 1670 if (use->is_Arraycopy()) { 1671 continue; 1672 } 1673 if (use->is_Field()) { 1674 FieldNode* field = use->as_Field(); 1675 assert(field->is_oop() && field->scalar_replaceable(), "sanity"); 1676 if (field->offset() == Type::OffsetBot) { 1677 jobj->set_scalar_replaceable(false); 1678 return; 1679 } 1680 // 2. An object is not scalar replaceable if the field into which it is 1681 // stored has multiple bases one of which is null. 1682 if (field->base_count() > 1) { 1683 for (BaseIterator i(field); i.has_next(); i.next()) { 1684 PointsToNode* base = i.get(); 1685 if (base == null_obj) { 1686 jobj->set_scalar_replaceable(false); 1687 return; 1688 } 1689 } 1690 } 1691 } 1692 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1693 // 3. An object is not scalar replaceable if it is merged with other objects. 1694 for (EdgeIterator j(use); j.has_next(); j.next()) { 1695 PointsToNode* ptn = j.get(); 1696 if (ptn->is_JavaObject() && ptn != jobj) { 1697 // Mark all objects. 1698 jobj->set_scalar_replaceable(false); 1699 ptn->set_scalar_replaceable(false); 1700 } 1701 } 1702 if (!jobj->scalar_replaceable()) { 1703 return; 1704 } 1705 } 1706 1707 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1708 if (j.get()->is_Arraycopy()) { 1709 continue; 1710 } 1711 1712 // Non-escaping object node should point only to field nodes. 1713 FieldNode* field = j.get()->as_Field(); 1714 int offset = field->as_Field()->offset(); 1715 1716 // 4. An object is not scalar replaceable if it has a field with unknown 1717 // offset (array's element is accessed in loop). 1718 if (offset == Type::OffsetBot) { 1719 jobj->set_scalar_replaceable(false); 1720 return; 1721 } 1722 // 5. Currently an object is not scalar replaceable if a LoadStore node 1723 // access its field since the field value is unknown after it. 1724 // 1725 Node* n = field->ideal_node(); 1726 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1727 if (n->fast_out(i)->is_LoadStore()) { 1728 jobj->set_scalar_replaceable(false); 1729 return; 1730 } 1731 } 1732 1733 // 6. Or the address may point to more then one object. This may produce 1734 // the false positive result (set not scalar replaceable) 1735 // since the flow-insensitive escape analysis can't separate 1736 // the case when stores overwrite the field's value from the case 1737 // when stores happened on different control branches. 1738 // 1739 // Note: it will disable scalar replacement in some cases: 1740 // 1741 // Point p[] = new Point[1]; 1742 // p[0] = new Point(); // Will be not scalar replaced 1743 // 1744 // but it will save us from incorrect optimizations in next cases: 1745 // 1746 // Point p[] = new Point[1]; 1747 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1748 // 1749 if (field->base_count() > 1) { 1750 for (BaseIterator i(field); i.has_next(); i.next()) { 1751 PointsToNode* base = i.get(); 1752 // Don't take into account LocalVar nodes which 1753 // may point to only one object which should be also 1754 // this field's base by now. 1755 if (base->is_JavaObject() && base != jobj) { 1756 // Mark all bases. 1757 jobj->set_scalar_replaceable(false); 1758 base->set_scalar_replaceable(false); 1759 } 1760 } 1761 } 1762 } 1763 } 1764 1765 #ifdef ASSERT 1766 void ConnectionGraph::verify_connection_graph( 1767 GrowableArray<PointsToNode*>& ptnodes_worklist, 1768 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1769 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1770 GrowableArray<Node*>& addp_worklist) { 1771 // Verify that graph is complete - no new edges could be added. 1772 int java_objects_length = java_objects_worklist.length(); 1773 int non_escaped_length = non_escaped_worklist.length(); 1774 int new_edges = 0; 1775 for (int next = 0; next < java_objects_length; ++next) { 1776 JavaObjectNode* ptn = java_objects_worklist.at(next); 1777 new_edges += add_java_object_edges(ptn, true); 1778 } 1779 assert(new_edges == 0, "graph was not complete"); 1780 // Verify that escape state is final. 1781 int length = non_escaped_worklist.length(); 1782 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1783 assert((non_escaped_length == non_escaped_worklist.length()) && 1784 (non_escaped_length == length) && 1785 (_worklist.length() == 0), "escape state was not final"); 1786 1787 // Verify fields information. 1788 int addp_length = addp_worklist.length(); 1789 for (int next = 0; next < addp_length; ++next ) { 1790 Node* n = addp_worklist.at(next); 1791 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1792 if (field->is_oop()) { 1793 // Verify that field has all bases 1794 Node* base = get_addp_base(n); 1795 PointsToNode* ptn = ptnode_adr(base->_idx); 1796 if (ptn->is_JavaObject()) { 1797 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1798 } else { 1799 assert(ptn->is_LocalVar(), "sanity"); 1800 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1801 PointsToNode* e = i.get(); 1802 if (e->is_JavaObject()) { 1803 assert(field->has_base(e->as_JavaObject()), "sanity"); 1804 } 1805 } 1806 } 1807 // Verify that all fields have initializing values. 1808 if (field->edge_count() == 0) { 1809 tty->print_cr("----------field does not have references----------"); 1810 field->dump(); 1811 for (BaseIterator i(field); i.has_next(); i.next()) { 1812 PointsToNode* base = i.get(); 1813 tty->print_cr("----------field has next base---------------------"); 1814 base->dump(); 1815 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1816 tty->print_cr("----------base has fields-------------------------"); 1817 for (EdgeIterator j(base); j.has_next(); j.next()) { 1818 j.get()->dump(); 1819 } 1820 tty->print_cr("----------base has references---------------------"); 1821 for (UseIterator j(base); j.has_next(); j.next()) { 1822 j.get()->dump(); 1823 } 1824 } 1825 } 1826 for (UseIterator i(field); i.has_next(); i.next()) { 1827 i.get()->dump(); 1828 } 1829 assert(field->edge_count() > 0, "sanity"); 1830 } 1831 } 1832 } 1833 } 1834 #endif 1835 1836 // Optimize ideal graph. 1837 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1838 GrowableArray<Node*>& storestore_worklist) { 1839 Compile* C = _compile; 1840 PhaseIterGVN* igvn = _igvn; 1841 if (EliminateLocks) { 1842 // Mark locks before changing ideal graph. 1843 int cnt = C->macro_count(); 1844 for( int i=0; i < cnt; i++ ) { 1845 Node *n = C->macro_node(i); 1846 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1847 AbstractLockNode* alock = n->as_AbstractLock(); 1848 if (!alock->is_non_esc_obj()) { 1849 if (not_global_escape(alock->obj_node())) { 1850 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1851 // The lock could be marked eliminated by lock coarsening 1852 // code during first IGVN before EA. Replace coarsened flag 1853 // to eliminate all associated locks/unlocks. 1854 #ifdef ASSERT 1855 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1856 #endif 1857 alock->set_non_esc_obj(); 1858 } 1859 } 1860 } 1861 } 1862 } 1863 1864 if (OptimizePtrCompare) { 1865 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1866 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1867 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1868 // Optimize objects compare. 1869 while (ptr_cmp_worklist.length() != 0) { 1870 Node *n = ptr_cmp_worklist.pop(); 1871 Node *res = optimize_ptr_compare(n); 1872 if (res != NULL) { 1873 #ifndef PRODUCT 1874 if (PrintOptimizePtrCompare) { 1875 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")); 1876 if (Verbose) { 1877 n->dump(1); 1878 } 1879 } 1880 #endif 1881 igvn->replace_node(n, res); 1882 } 1883 } 1884 // cleanup 1885 if (_pcmp_neq->outcnt() == 0) 1886 igvn->hash_delete(_pcmp_neq); 1887 if (_pcmp_eq->outcnt() == 0) 1888 igvn->hash_delete(_pcmp_eq); 1889 } 1890 1891 // For MemBarStoreStore nodes added in library_call.cpp, check 1892 // escape status of associated AllocateNode and optimize out 1893 // MemBarStoreStore node if the allocated object never escapes. 1894 while (storestore_worklist.length() != 0) { 1895 Node *n = storestore_worklist.pop(); 1896 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1897 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1898 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1899 if (not_global_escape(alloc)) { 1900 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1901 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1902 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1903 igvn->register_new_node_with_optimizer(mb); 1904 igvn->replace_node(storestore, mb); 1905 } 1906 } 1907 } 1908 1909 // Optimize objects compare. 1910 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1911 assert(OptimizePtrCompare, "sanity"); 1912 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1913 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1914 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1915 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1916 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1917 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1918 1919 // Check simple cases first. 1920 if (jobj1 != NULL) { 1921 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1922 if (jobj1 == jobj2) { 1923 // Comparing the same not escaping object. 1924 return _pcmp_eq; 1925 } 1926 Node* obj = jobj1->ideal_node(); 1927 // Comparing not escaping allocation. 1928 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1929 !ptn2->points_to(jobj1)) { 1930 return _pcmp_neq; // This includes nullness check. 1931 } 1932 } 1933 } 1934 if (jobj2 != NULL) { 1935 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1936 Node* obj = jobj2->ideal_node(); 1937 // Comparing not escaping allocation. 1938 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1939 !ptn1->points_to(jobj2)) { 1940 return _pcmp_neq; // This includes nullness check. 1941 } 1942 } 1943 } 1944 if (jobj1 != NULL && jobj1 != phantom_obj && 1945 jobj2 != NULL && jobj2 != phantom_obj && 1946 jobj1->ideal_node()->is_Con() && 1947 jobj2->ideal_node()->is_Con()) { 1948 // Klass or String constants compare. Need to be careful with 1949 // compressed pointers - compare types of ConN and ConP instead of nodes. 1950 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1951 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1952 if (t1->make_ptr() == t2->make_ptr()) { 1953 return _pcmp_eq; 1954 } else { 1955 return _pcmp_neq; 1956 } 1957 } 1958 if (ptn1->meet(ptn2)) { 1959 return NULL; // Sets are not disjoint 1960 } 1961 1962 // Sets are disjoint. 1963 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1964 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1965 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1966 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1967 if (set1_has_unknown_ptr && set2_has_null_ptr || 1968 set2_has_unknown_ptr && set1_has_null_ptr) { 1969 // Check nullness of unknown object. 1970 return NULL; 1971 } 1972 1973 // Disjointness by itself is not sufficient since 1974 // alias analysis is not complete for escaped objects. 1975 // Disjoint sets are definitely unrelated only when 1976 // at least one set has only not escaping allocations. 1977 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1978 if (ptn1->non_escaping_allocation()) { 1979 return _pcmp_neq; 1980 } 1981 } 1982 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 1983 if (ptn2->non_escaping_allocation()) { 1984 return _pcmp_neq; 1985 } 1986 } 1987 return NULL; 1988 } 1989 1990 // Connection Graph constuction functions. 1991 1992 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 1993 PointsToNode* ptadr = _nodes.at(n->_idx); 1994 if (ptadr != NULL) { 1995 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 1996 return; 1997 } 1998 Compile* C = _compile; 1999 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 2000 _nodes.at_put(n->_idx, ptadr); 2001 } 2002 2003 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 2004 PointsToNode* ptadr = _nodes.at(n->_idx); 2005 if (ptadr != NULL) { 2006 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2007 return; 2008 } 2009 Compile* C = _compile; 2010 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2011 _nodes.at_put(n->_idx, ptadr); 2012 } 2013 2014 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2015 PointsToNode* ptadr = _nodes.at(n->_idx); 2016 if (ptadr != NULL) { 2017 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2018 return; 2019 } 2020 bool unsafe = false; 2021 bool is_oop = is_oop_field(n, offset, &unsafe); 2022 if (unsafe) { 2023 es = PointsToNode::GlobalEscape; 2024 } 2025 Compile* C = _compile; 2026 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2027 _nodes.at_put(n->_idx, field); 2028 } 2029 2030 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2031 PointsToNode* src, PointsToNode* dst) { 2032 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2033 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2034 PointsToNode* ptadr = _nodes.at(n->_idx); 2035 if (ptadr != NULL) { 2036 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2037 return; 2038 } 2039 Compile* C = _compile; 2040 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2041 _nodes.at_put(n->_idx, ptadr); 2042 // Add edge from arraycopy node to source object. 2043 (void)add_edge(ptadr, src); 2044 src->set_arraycopy_src(); 2045 // Add edge from destination object to arraycopy node. 2046 (void)add_edge(dst, ptadr); 2047 dst->set_arraycopy_dst(); 2048 } 2049 2050 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2051 const Type* adr_type = n->as_AddP()->bottom_type(); 2052 BasicType bt = T_INT; 2053 if (offset == Type::OffsetBot) { 2054 // Check only oop fields. 2055 if (!adr_type->isa_aryptr() || 2056 (adr_type->isa_aryptr()->klass() == NULL) || 2057 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2058 // OffsetBot is used to reference array's element. Ignore first AddP. 2059 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2060 bt = T_OBJECT; 2061 } 2062 } 2063 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2064 if (adr_type->isa_instptr()) { 2065 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2066 if (field != NULL) { 2067 bt = field->layout_type(); 2068 } else { 2069 // Check for unsafe oop field access 2070 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2071 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2072 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) { 2073 bt = T_OBJECT; 2074 (*unsafe) = true; 2075 } 2076 } 2077 } else if (adr_type->isa_aryptr()) { 2078 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2079 // Ignore array length load. 2080 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2081 // Ignore first AddP. 2082 } else { 2083 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2084 bt = elemtype->array_element_basic_type(); 2085 } 2086 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2087 // Allocation initialization, ThreadLocal field access, unsafe access 2088 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2089 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2090 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) { 2091 bt = T_OBJECT; 2092 } 2093 } 2094 } 2095 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 2096 } 2097 2098 // Returns unique pointed java object or NULL. 2099 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2100 assert(!_collecting, "should not call when contructed graph"); 2101 // If the node was created after the escape computation we can't answer. 2102 uint idx = n->_idx; 2103 if (idx >= nodes_size()) { 2104 return NULL; 2105 } 2106 PointsToNode* ptn = ptnode_adr(idx); 2107 if (ptn->is_JavaObject()) { 2108 return ptn->as_JavaObject(); 2109 } 2110 assert(ptn->is_LocalVar(), "sanity"); 2111 // Check all java objects it points to. 2112 JavaObjectNode* jobj = NULL; 2113 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2114 PointsToNode* e = i.get(); 2115 if (e->is_JavaObject()) { 2116 if (jobj == NULL) { 2117 jobj = e->as_JavaObject(); 2118 } else if (jobj != e) { 2119 return NULL; 2120 } 2121 } 2122 } 2123 return jobj; 2124 } 2125 2126 // Return true if this node points only to non-escaping allocations. 2127 bool PointsToNode::non_escaping_allocation() { 2128 if (is_JavaObject()) { 2129 Node* n = ideal_node(); 2130 if (n->is_Allocate() || n->is_CallStaticJava()) { 2131 return (escape_state() == PointsToNode::NoEscape); 2132 } else { 2133 return false; 2134 } 2135 } 2136 assert(is_LocalVar(), "sanity"); 2137 // Check all java objects it points to. 2138 for (EdgeIterator i(this); i.has_next(); i.next()) { 2139 PointsToNode* e = i.get(); 2140 if (e->is_JavaObject()) { 2141 Node* n = e->ideal_node(); 2142 if ((e->escape_state() != PointsToNode::NoEscape) || 2143 !(n->is_Allocate() || n->is_CallStaticJava())) { 2144 return false; 2145 } 2146 } 2147 } 2148 return true; 2149 } 2150 2151 // Return true if we know the node does not escape globally. 2152 bool ConnectionGraph::not_global_escape(Node *n) { 2153 assert(!_collecting, "should not call during graph construction"); 2154 // If the node was created after the escape computation we can't answer. 2155 uint idx = n->_idx; 2156 if (idx >= nodes_size()) { 2157 return false; 2158 } 2159 PointsToNode* ptn = ptnode_adr(idx); 2160 PointsToNode::EscapeState es = ptn->escape_state(); 2161 // If we have already computed a value, return it. 2162 if (es >= PointsToNode::GlobalEscape) 2163 return false; 2164 if (ptn->is_JavaObject()) { 2165 return true; // (es < PointsToNode::GlobalEscape); 2166 } 2167 assert(ptn->is_LocalVar(), "sanity"); 2168 // Check all java objects it points to. 2169 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2170 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2171 return false; 2172 } 2173 return true; 2174 } 2175 2176 2177 // Helper functions 2178 2179 // Return true if this node points to specified node or nodes it points to. 2180 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2181 if (is_JavaObject()) { 2182 return (this == ptn); 2183 } 2184 assert(is_LocalVar() || is_Field(), "sanity"); 2185 for (EdgeIterator i(this); i.has_next(); i.next()) { 2186 if (i.get() == ptn) 2187 return true; 2188 } 2189 return false; 2190 } 2191 2192 // Return true if one node points to an other. 2193 bool PointsToNode::meet(PointsToNode* ptn) { 2194 if (this == ptn) { 2195 return true; 2196 } else if (ptn->is_JavaObject()) { 2197 return this->points_to(ptn->as_JavaObject()); 2198 } else if (this->is_JavaObject()) { 2199 return ptn->points_to(this->as_JavaObject()); 2200 } 2201 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2202 int ptn_count = ptn->edge_count(); 2203 for (EdgeIterator i(this); i.has_next(); i.next()) { 2204 PointsToNode* this_e = i.get(); 2205 for (int j = 0; j < ptn_count; j++) { 2206 if (this_e == ptn->edge(j)) 2207 return true; 2208 } 2209 } 2210 return false; 2211 } 2212 2213 #ifdef ASSERT 2214 // Return true if bases point to this java object. 2215 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2216 for (BaseIterator i(this); i.has_next(); i.next()) { 2217 if (i.get() == jobj) 2218 return true; 2219 } 2220 return false; 2221 } 2222 #endif 2223 2224 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2225 const Type *adr_type = phase->type(adr); 2226 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2227 adr->in(AddPNode::Address)->is_Proj() && 2228 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2229 // We are computing a raw address for a store captured by an Initialize 2230 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2231 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2232 assert(offs != Type::OffsetBot || 2233 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2234 "offset must be a constant or it is initialization of array"); 2235 return offs; 2236 } 2237 const TypePtr *t_ptr = adr_type->isa_ptr(); 2238 assert(t_ptr != NULL, "must be a pointer type"); 2239 return t_ptr->offset(); 2240 } 2241 2242 Node* ConnectionGraph::get_addp_base(Node *addp) { 2243 assert(addp->is_AddP(), "must be AddP"); 2244 // 2245 // AddP cases for Base and Address inputs: 2246 // case #1. Direct object's field reference: 2247 // Allocate 2248 // | 2249 // Proj #5 ( oop result ) 2250 // | 2251 // CheckCastPP (cast to instance type) 2252 // | | 2253 // AddP ( base == address ) 2254 // 2255 // case #2. Indirect object's field reference: 2256 // Phi 2257 // | 2258 // CastPP (cast to instance type) 2259 // | | 2260 // AddP ( base == address ) 2261 // 2262 // case #3. Raw object's field reference for Initialize node: 2263 // Allocate 2264 // | 2265 // Proj #5 ( oop result ) 2266 // top | 2267 // \ | 2268 // AddP ( base == top ) 2269 // 2270 // case #4. Array's element reference: 2271 // {CheckCastPP | CastPP} 2272 // | | | 2273 // | AddP ( array's element offset ) 2274 // | | 2275 // AddP ( array's offset ) 2276 // 2277 // case #5. Raw object's field reference for arraycopy stub call: 2278 // The inline_native_clone() case when the arraycopy stub is called 2279 // after the allocation before Initialize and CheckCastPP nodes. 2280 // Allocate 2281 // | 2282 // Proj #5 ( oop result ) 2283 // | | 2284 // AddP ( base == address ) 2285 // 2286 // case #6. Constant Pool, ThreadLocal, CastX2P or 2287 // Raw object's field reference: 2288 // {ConP, ThreadLocal, CastX2P, raw Load} 2289 // top | 2290 // \ | 2291 // AddP ( base == top ) 2292 // 2293 // case #7. Klass's field reference. 2294 // LoadKlass 2295 // | | 2296 // AddP ( base == address ) 2297 // 2298 // case #8. narrow Klass's field reference. 2299 // LoadNKlass 2300 // | 2301 // DecodeN 2302 // | | 2303 // AddP ( base == address ) 2304 // 2305 Node *base = addp->in(AddPNode::Base); 2306 if (base->uncast()->is_top()) { // The AddP case #3 and #6. 2307 base = addp->in(AddPNode::Address); 2308 while (base->is_AddP()) { 2309 // Case #6 (unsafe access) may have several chained AddP nodes. 2310 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2311 base = base->in(AddPNode::Address); 2312 } 2313 Node* uncast_base = base->uncast(); 2314 int opcode = uncast_base->Opcode(); 2315 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2316 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2317 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2318 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity"); 2319 } 2320 return base; 2321 } 2322 2323 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2324 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2325 Node* addp2 = addp->raw_out(0); 2326 if (addp->outcnt() == 1 && addp2->is_AddP() && 2327 addp2->in(AddPNode::Base) == n && 2328 addp2->in(AddPNode::Address) == addp) { 2329 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2330 // 2331 // Find array's offset to push it on worklist first and 2332 // as result process an array's element offset first (pushed second) 2333 // to avoid CastPP for the array's offset. 2334 // Otherwise the inserted CastPP (LocalVar) will point to what 2335 // the AddP (Field) points to. Which would be wrong since 2336 // the algorithm expects the CastPP has the same point as 2337 // as AddP's base CheckCastPP (LocalVar). 2338 // 2339 // ArrayAllocation 2340 // | 2341 // CheckCastPP 2342 // | 2343 // memProj (from ArrayAllocation CheckCastPP) 2344 // | || 2345 // | || Int (element index) 2346 // | || | ConI (log(element size)) 2347 // | || | / 2348 // | || LShift 2349 // | || / 2350 // | AddP (array's element offset) 2351 // | | 2352 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2353 // | / / 2354 // AddP (array's offset) 2355 // | 2356 // Load/Store (memory operation on array's element) 2357 // 2358 return addp2; 2359 } 2360 return NULL; 2361 } 2362 2363 // 2364 // Adjust the type and inputs of an AddP which computes the 2365 // address of a field of an instance 2366 // 2367 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2368 PhaseGVN* igvn = _igvn; 2369 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2370 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2371 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2372 if (t == NULL) { 2373 // We are computing a raw address for a store captured by an Initialize 2374 // compute an appropriate address type (cases #3 and #5). 2375 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2376 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2377 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2378 assert(offs != Type::OffsetBot, "offset must be a constant"); 2379 t = base_t->add_offset(offs)->is_oopptr(); 2380 } 2381 int inst_id = base_t->instance_id(); 2382 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2383 "old type must be non-instance or match new type"); 2384 2385 // The type 't' could be subclass of 'base_t'. 2386 // As result t->offset() could be large then base_t's size and it will 2387 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2388 // constructor verifies correctness of the offset. 2389 // 2390 // It could happened on subclass's branch (from the type profiling 2391 // inlining) which was not eliminated during parsing since the exactness 2392 // of the allocation type was not propagated to the subclass type check. 2393 // 2394 // Or the type 't' could be not related to 'base_t' at all. 2395 // It could happened when CHA type is different from MDO type on a dead path 2396 // (for example, from instanceof check) which is not collapsed during parsing. 2397 // 2398 // Do nothing for such AddP node and don't process its users since 2399 // this code branch will go away. 2400 // 2401 if (!t->is_known_instance() && 2402 !base_t->klass()->is_subtype_of(t->klass())) { 2403 return false; // bail out 2404 } 2405 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2406 // Do NOT remove the next line: ensure a new alias index is allocated 2407 // for the instance type. Note: C++ will not remove it since the call 2408 // has side effect. 2409 int alias_idx = _compile->get_alias_index(tinst); 2410 igvn->set_type(addp, tinst); 2411 // record the allocation in the node map 2412 set_map(addp, get_map(base->_idx)); 2413 // Set addp's Base and Address to 'base'. 2414 Node *abase = addp->in(AddPNode::Base); 2415 Node *adr = addp->in(AddPNode::Address); 2416 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2417 adr->in(0)->_idx == (uint)inst_id) { 2418 // Skip AddP cases #3 and #5. 2419 } else { 2420 assert(!abase->is_top(), "sanity"); // AddP case #3 2421 if (abase != base) { 2422 igvn->hash_delete(addp); 2423 addp->set_req(AddPNode::Base, base); 2424 if (abase == adr) { 2425 addp->set_req(AddPNode::Address, base); 2426 } else { 2427 // AddP case #4 (adr is array's element offset AddP node) 2428 #ifdef ASSERT 2429 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2430 assert(adr->is_AddP() && atype != NULL && 2431 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2432 #endif 2433 } 2434 igvn->hash_insert(addp); 2435 } 2436 } 2437 // Put on IGVN worklist since at least addp's type was changed above. 2438 record_for_optimizer(addp); 2439 return true; 2440 } 2441 2442 // 2443 // Create a new version of orig_phi if necessary. Returns either the newly 2444 // created phi or an existing phi. Sets create_new to indicate whether a new 2445 // phi was created. Cache the last newly created phi in the node map. 2446 // 2447 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2448 Compile *C = _compile; 2449 PhaseGVN* igvn = _igvn; 2450 new_created = false; 2451 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2452 // nothing to do if orig_phi is bottom memory or matches alias_idx 2453 if (phi_alias_idx == alias_idx) { 2454 return orig_phi; 2455 } 2456 // Have we recently created a Phi for this alias index? 2457 PhiNode *result = get_map_phi(orig_phi->_idx); 2458 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2459 return result; 2460 } 2461 // Previous check may fail when the same wide memory Phi was split into Phis 2462 // for different memory slices. Search all Phis for this region. 2463 if (result != NULL) { 2464 Node* region = orig_phi->in(0); 2465 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2466 Node* phi = region->fast_out(i); 2467 if (phi->is_Phi() && 2468 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2469 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2470 return phi->as_Phi(); 2471 } 2472 } 2473 } 2474 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2475 if (C->do_escape_analysis() == true && !C->failing()) { 2476 // Retry compilation without escape analysis. 2477 // If this is the first failure, the sentinel string will "stick" 2478 // to the Compile object, and the C2Compiler will see it and retry. 2479 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2480 } 2481 return NULL; 2482 } 2483 orig_phi_worklist.append_if_missing(orig_phi); 2484 const TypePtr *atype = C->get_adr_type(alias_idx); 2485 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2486 C->copy_node_notes_to(result, orig_phi); 2487 igvn->set_type(result, result->bottom_type()); 2488 record_for_optimizer(result); 2489 set_map(orig_phi, result); 2490 new_created = true; 2491 return result; 2492 } 2493 2494 // 2495 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2496 // specified alias index. 2497 // 2498 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2499 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2500 Compile *C = _compile; 2501 PhaseGVN* igvn = _igvn; 2502 bool new_phi_created; 2503 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2504 if (!new_phi_created) { 2505 return result; 2506 } 2507 GrowableArray<PhiNode *> phi_list; 2508 GrowableArray<uint> cur_input; 2509 PhiNode *phi = orig_phi; 2510 uint idx = 1; 2511 bool finished = false; 2512 while(!finished) { 2513 while (idx < phi->req()) { 2514 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2515 if (mem != NULL && mem->is_Phi()) { 2516 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2517 if (new_phi_created) { 2518 // found an phi for which we created a new split, push current one on worklist and begin 2519 // processing new one 2520 phi_list.push(phi); 2521 cur_input.push(idx); 2522 phi = mem->as_Phi(); 2523 result = newphi; 2524 idx = 1; 2525 continue; 2526 } else { 2527 mem = newphi; 2528 } 2529 } 2530 if (C->failing()) { 2531 return NULL; 2532 } 2533 result->set_req(idx++, mem); 2534 } 2535 #ifdef ASSERT 2536 // verify that the new Phi has an input for each input of the original 2537 assert( phi->req() == result->req(), "must have same number of inputs."); 2538 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2539 #endif 2540 // Check if all new phi's inputs have specified alias index. 2541 // Otherwise use old phi. 2542 for (uint i = 1; i < phi->req(); i++) { 2543 Node* in = result->in(i); 2544 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2545 } 2546 // we have finished processing a Phi, see if there are any more to do 2547 finished = (phi_list.length() == 0 ); 2548 if (!finished) { 2549 phi = phi_list.pop(); 2550 idx = cur_input.pop(); 2551 PhiNode *prev_result = get_map_phi(phi->_idx); 2552 prev_result->set_req(idx++, result); 2553 result = prev_result; 2554 } 2555 } 2556 return result; 2557 } 2558 2559 // 2560 // The next methods are derived from methods in MemNode. 2561 // 2562 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2563 Node *mem = mmem; 2564 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2565 // means an array I have not precisely typed yet. Do not do any 2566 // alias stuff with it any time soon. 2567 if (toop->base() != Type::AnyPtr && 2568 !(toop->klass() != NULL && 2569 toop->klass()->is_java_lang_Object() && 2570 toop->offset() == Type::OffsetBot)) { 2571 mem = mmem->memory_at(alias_idx); 2572 // Update input if it is progress over what we have now 2573 } 2574 return mem; 2575 } 2576 2577 // 2578 // Move memory users to their memory slices. 2579 // 2580 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2581 Compile* C = _compile; 2582 PhaseGVN* igvn = _igvn; 2583 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2584 assert(tp != NULL, "ptr type"); 2585 int alias_idx = C->get_alias_index(tp); 2586 int general_idx = C->get_general_index(alias_idx); 2587 2588 // Move users first 2589 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2590 Node* use = n->fast_out(i); 2591 if (use->is_MergeMem()) { 2592 MergeMemNode* mmem = use->as_MergeMem(); 2593 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2594 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2595 continue; // Nothing to do 2596 } 2597 // Replace previous general reference to mem node. 2598 uint orig_uniq = C->unique(); 2599 Node* m = find_inst_mem(n, general_idx, orig_phis); 2600 assert(orig_uniq == C->unique(), "no new nodes"); 2601 mmem->set_memory_at(general_idx, m); 2602 --imax; 2603 --i; 2604 } else if (use->is_MemBar()) { 2605 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2606 if (use->req() > MemBarNode::Precedent && 2607 use->in(MemBarNode::Precedent) == n) { 2608 // Don't move related membars. 2609 record_for_optimizer(use); 2610 continue; 2611 } 2612 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2613 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 2614 alias_idx == general_idx) { 2615 continue; // Nothing to do 2616 } 2617 // Move to general memory slice. 2618 uint orig_uniq = C->unique(); 2619 Node* m = find_inst_mem(n, general_idx, orig_phis); 2620 assert(orig_uniq == C->unique(), "no new nodes"); 2621 igvn->hash_delete(use); 2622 imax -= use->replace_edge(n, m); 2623 igvn->hash_insert(use); 2624 record_for_optimizer(use); 2625 --i; 2626 #ifdef ASSERT 2627 } else if (use->is_Mem()) { 2628 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2629 // Don't move related cardmark. 2630 continue; 2631 } 2632 // Memory nodes should have new memory input. 2633 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2634 assert(tp != NULL, "ptr type"); 2635 int idx = C->get_alias_index(tp); 2636 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2637 "Following memory nodes should have new memory input or be on the same memory slice"); 2638 } else if (use->is_Phi()) { 2639 // Phi nodes should be split and moved already. 2640 tp = use->as_Phi()->adr_type()->isa_ptr(); 2641 assert(tp != NULL, "ptr type"); 2642 int idx = C->get_alias_index(tp); 2643 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2644 } else { 2645 use->dump(); 2646 assert(false, "should not be here"); 2647 #endif 2648 } 2649 } 2650 } 2651 2652 // 2653 // Search memory chain of "mem" to find a MemNode whose address 2654 // is the specified alias index. 2655 // 2656 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2657 if (orig_mem == NULL) 2658 return orig_mem; 2659 Compile* C = _compile; 2660 PhaseGVN* igvn = _igvn; 2661 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2662 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2663 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 2664 Node *prev = NULL; 2665 Node *result = orig_mem; 2666 while (prev != result) { 2667 prev = result; 2668 if (result == start_mem) 2669 break; // hit one of our sentinels 2670 if (result->is_Mem()) { 2671 const Type *at = igvn->type(result->in(MemNode::Address)); 2672 if (at == Type::TOP) 2673 break; // Dead 2674 assert (at->isa_ptr() != NULL, "pointer type required."); 2675 int idx = C->get_alias_index(at->is_ptr()); 2676 if (idx == alias_idx) 2677 break; // Found 2678 if (!is_instance && (at->isa_oopptr() == NULL || 2679 !at->is_oopptr()->is_known_instance())) { 2680 break; // Do not skip store to general memory slice. 2681 } 2682 result = result->in(MemNode::Memory); 2683 } 2684 if (!is_instance) 2685 continue; // don't search further for non-instance types 2686 // skip over a call which does not affect this memory slice 2687 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2688 Node *proj_in = result->in(0); 2689 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2690 break; // hit one of our sentinels 2691 } else if (proj_in->is_Call()) { 2692 // ArrayCopy node processed here as well 2693 CallNode *call = proj_in->as_Call(); 2694 if (!call->may_modify(toop, igvn)) { 2695 result = call->in(TypeFunc::Memory); 2696 } 2697 } else if (proj_in->is_Initialize()) { 2698 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2699 // Stop if this is the initialization for the object instance which 2700 // which contains this memory slice, otherwise skip over it. 2701 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2702 result = proj_in->in(TypeFunc::Memory); 2703 } 2704 } else if (proj_in->is_MemBar()) { 2705 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() && 2706 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() && 2707 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) { 2708 // clone 2709 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy(); 2710 if (ac->may_modify(toop, igvn)) { 2711 break; 2712 } 2713 } 2714 result = proj_in->in(TypeFunc::Memory); 2715 } 2716 } else if (result->is_MergeMem()) { 2717 MergeMemNode *mmem = result->as_MergeMem(); 2718 result = step_through_mergemem(mmem, alias_idx, toop); 2719 if (result == mmem->base_memory()) { 2720 // Didn't find instance memory, search through general slice recursively. 2721 result = mmem->memory_at(C->get_general_index(alias_idx)); 2722 result = find_inst_mem(result, alias_idx, orig_phis); 2723 if (C->failing()) { 2724 return NULL; 2725 } 2726 mmem->set_memory_at(alias_idx, result); 2727 } 2728 } else if (result->is_Phi() && 2729 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2730 Node *un = result->as_Phi()->unique_input(igvn); 2731 if (un != NULL) { 2732 orig_phis.append_if_missing(result->as_Phi()); 2733 result = un; 2734 } else { 2735 break; 2736 } 2737 } else if (result->is_ClearArray()) { 2738 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2739 // Can not bypass initialization of the instance 2740 // we are looking for. 2741 break; 2742 } 2743 // Otherwise skip it (the call updated 'result' value). 2744 } else if (result->Opcode() == Op_SCMemProj) { 2745 Node* mem = result->in(0); 2746 Node* adr = NULL; 2747 if (mem->is_LoadStore()) { 2748 adr = mem->in(MemNode::Address); 2749 } else { 2750 assert(mem->Opcode() == Op_EncodeISOArray || 2751 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 2752 adr = mem->in(3); // Memory edge corresponds to destination array 2753 } 2754 const Type *at = igvn->type(adr); 2755 if (at != Type::TOP) { 2756 assert(at->isa_ptr() != NULL, "pointer type required."); 2757 int idx = C->get_alias_index(at->is_ptr()); 2758 if (idx == alias_idx) { 2759 // Assert in debug mode 2760 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 2761 break; // In product mode return SCMemProj node 2762 } 2763 } 2764 result = mem->in(MemNode::Memory); 2765 } else if (result->Opcode() == Op_StrInflatedCopy) { 2766 Node* adr = result->in(3); // Memory edge corresponds to destination array 2767 const Type *at = igvn->type(adr); 2768 if (at != Type::TOP) { 2769 assert(at->isa_ptr() != NULL, "pointer type required."); 2770 int idx = C->get_alias_index(at->is_ptr()); 2771 if (idx == alias_idx) { 2772 // Assert in debug mode 2773 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 2774 break; // In product mode return SCMemProj node 2775 } 2776 } 2777 result = result->in(MemNode::Memory); 2778 } 2779 } 2780 if (result->is_Phi()) { 2781 PhiNode *mphi = result->as_Phi(); 2782 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2783 const TypePtr *t = mphi->adr_type(); 2784 if (!is_instance) { 2785 // Push all non-instance Phis on the orig_phis worklist to update inputs 2786 // during Phase 4 if needed. 2787 orig_phis.append_if_missing(mphi); 2788 } else if (C->get_alias_index(t) != alias_idx) { 2789 // Create a new Phi with the specified alias index type. 2790 result = split_memory_phi(mphi, alias_idx, orig_phis); 2791 } 2792 } 2793 // the result is either MemNode, PhiNode, InitializeNode. 2794 return result; 2795 } 2796 2797 // 2798 // Convert the types of unescaped object to instance types where possible, 2799 // propagate the new type information through the graph, and update memory 2800 // edges and MergeMem inputs to reflect the new type. 2801 // 2802 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2803 // The processing is done in 4 phases: 2804 // 2805 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2806 // types for the CheckCastPP for allocations where possible. 2807 // Propagate the new types through users as follows: 2808 // casts and Phi: push users on alloc_worklist 2809 // AddP: cast Base and Address inputs to the instance type 2810 // push any AddP users on alloc_worklist and push any memnode 2811 // users onto memnode_worklist. 2812 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2813 // search the Memory chain for a store with the appropriate type 2814 // address type. If a Phi is found, create a new version with 2815 // the appropriate memory slices from each of the Phi inputs. 2816 // For stores, process the users as follows: 2817 // MemNode: push on memnode_worklist 2818 // MergeMem: push on mergemem_worklist 2819 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2820 // moving the first node encountered of each instance type to the 2821 // the input corresponding to its alias index. 2822 // appropriate memory slice. 2823 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2824 // 2825 // In the following example, the CheckCastPP nodes are the cast of allocation 2826 // results and the allocation of node 29 is unescaped and eligible to be an 2827 // instance type. 2828 // 2829 // We start with: 2830 // 2831 // 7 Parm #memory 2832 // 10 ConI "12" 2833 // 19 CheckCastPP "Foo" 2834 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2835 // 29 CheckCastPP "Foo" 2836 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2837 // 2838 // 40 StoreP 25 7 20 ... alias_index=4 2839 // 50 StoreP 35 40 30 ... alias_index=4 2840 // 60 StoreP 45 50 20 ... alias_index=4 2841 // 70 LoadP _ 60 30 ... alias_index=4 2842 // 80 Phi 75 50 60 Memory alias_index=4 2843 // 90 LoadP _ 80 30 ... alias_index=4 2844 // 100 LoadP _ 80 20 ... alias_index=4 2845 // 2846 // 2847 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2848 // and creating a new alias index for node 30. This gives: 2849 // 2850 // 7 Parm #memory 2851 // 10 ConI "12" 2852 // 19 CheckCastPP "Foo" 2853 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2854 // 29 CheckCastPP "Foo" iid=24 2855 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2856 // 2857 // 40 StoreP 25 7 20 ... alias_index=4 2858 // 50 StoreP 35 40 30 ... alias_index=6 2859 // 60 StoreP 45 50 20 ... alias_index=4 2860 // 70 LoadP _ 60 30 ... alias_index=6 2861 // 80 Phi 75 50 60 Memory alias_index=4 2862 // 90 LoadP _ 80 30 ... alias_index=6 2863 // 100 LoadP _ 80 20 ... alias_index=4 2864 // 2865 // In phase 2, new memory inputs are computed for the loads and stores, 2866 // And a new version of the phi is created. In phase 4, the inputs to 2867 // node 80 are updated and then the memory nodes are updated with the 2868 // values computed in phase 2. This results in: 2869 // 2870 // 7 Parm #memory 2871 // 10 ConI "12" 2872 // 19 CheckCastPP "Foo" 2873 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2874 // 29 CheckCastPP "Foo" iid=24 2875 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2876 // 2877 // 40 StoreP 25 7 20 ... alias_index=4 2878 // 50 StoreP 35 7 30 ... alias_index=6 2879 // 60 StoreP 45 40 20 ... alias_index=4 2880 // 70 LoadP _ 50 30 ... alias_index=6 2881 // 80 Phi 75 40 60 Memory alias_index=4 2882 // 120 Phi 75 50 50 Memory alias_index=6 2883 // 90 LoadP _ 120 30 ... alias_index=6 2884 // 100 LoadP _ 80 20 ... alias_index=4 2885 // 2886 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 2887 GrowableArray<Node *> memnode_worklist; 2888 GrowableArray<PhiNode *> orig_phis; 2889 PhaseIterGVN *igvn = _igvn; 2890 uint new_index_start = (uint) _compile->num_alias_types(); 2891 Arena* arena = Thread::current()->resource_area(); 2892 VectorSet visited(arena); 2893 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2894 uint unique_old = _compile->unique(); 2895 2896 // Phase 1: Process possible allocations from alloc_worklist. 2897 // Create instance types for the CheckCastPP for allocations where possible. 2898 // 2899 // (Note: don't forget to change the order of the second AddP node on 2900 // the alloc_worklist if the order of the worklist processing is changed, 2901 // see the comment in find_second_addp().) 2902 // 2903 while (alloc_worklist.length() != 0) { 2904 Node *n = alloc_worklist.pop(); 2905 uint ni = n->_idx; 2906 if (n->is_Call()) { 2907 CallNode *alloc = n->as_Call(); 2908 // copy escape information to call node 2909 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2910 PointsToNode::EscapeState es = ptn->escape_state(); 2911 // We have an allocation or call which returns a Java object, 2912 // see if it is unescaped. 2913 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2914 continue; 2915 // Find CheckCastPP for the allocate or for the return value of a call 2916 n = alloc->result_cast(); 2917 if (n == NULL) { // No uses except Initialize node 2918 if (alloc->is_Allocate()) { 2919 // Set the scalar_replaceable flag for allocation 2920 // so it could be eliminated if it has no uses. 2921 alloc->as_Allocate()->_is_scalar_replaceable = true; 2922 } 2923 if (alloc->is_CallStaticJava()) { 2924 // Set the scalar_replaceable flag for boxing method 2925 // so it could be eliminated if it has no uses. 2926 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2927 } 2928 continue; 2929 } 2930 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2931 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2932 continue; 2933 } 2934 2935 // The inline code for Object.clone() casts the allocation result to 2936 // java.lang.Object and then to the actual type of the allocated 2937 // object. Detect this case and use the second cast. 2938 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2939 // the allocation result is cast to java.lang.Object and then 2940 // to the actual Array type. 2941 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2942 && (alloc->is_AllocateArray() || 2943 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2944 Node *cast2 = NULL; 2945 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2946 Node *use = n->fast_out(i); 2947 if (use->is_CheckCastPP()) { 2948 cast2 = use; 2949 break; 2950 } 2951 } 2952 if (cast2 != NULL) { 2953 n = cast2; 2954 } else { 2955 // Non-scalar replaceable if the allocation type is unknown statically 2956 // (reflection allocation), the object can't be restored during 2957 // deoptimization without precise type. 2958 continue; 2959 } 2960 } 2961 2962 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 2963 if (t == NULL) 2964 continue; // not a TypeOopPtr 2965 if (!t->klass_is_exact()) 2966 continue; // not an unique type 2967 2968 if (alloc->is_Allocate()) { 2969 // Set the scalar_replaceable flag for allocation 2970 // so it could be eliminated. 2971 alloc->as_Allocate()->_is_scalar_replaceable = true; 2972 } 2973 if (alloc->is_CallStaticJava()) { 2974 // Set the scalar_replaceable flag for boxing method 2975 // so it could be eliminated. 2976 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2977 } 2978 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 2979 // in order for an object to be scalar-replaceable, it must be: 2980 // - a direct allocation (not a call returning an object) 2981 // - non-escaping 2982 // - eligible to be a unique type 2983 // - not determined to be ineligible by escape analysis 2984 set_map(alloc, n); 2985 set_map(n, alloc); 2986 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 2987 igvn->hash_delete(n); 2988 igvn->set_type(n, tinst); 2989 n->raise_bottom_type(tinst); 2990 igvn->hash_insert(n); 2991 record_for_optimizer(n); 2992 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 2993 2994 // First, put on the worklist all Field edges from Connection Graph 2995 // which is more accurate than putting immediate users from Ideal Graph. 2996 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 2997 PointsToNode* tgt = e.get(); 2998 if (tgt->is_Arraycopy()) { 2999 continue; 3000 } 3001 Node* use = tgt->ideal_node(); 3002 assert(tgt->is_Field() && use->is_AddP(), 3003 "only AddP nodes are Field edges in CG"); 3004 if (use->outcnt() > 0) { // Don't process dead nodes 3005 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 3006 if (addp2 != NULL) { 3007 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3008 alloc_worklist.append_if_missing(addp2); 3009 } 3010 alloc_worklist.append_if_missing(use); 3011 } 3012 } 3013 3014 // An allocation may have an Initialize which has raw stores. Scan 3015 // the users of the raw allocation result and push AddP users 3016 // on alloc_worklist. 3017 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 3018 assert (raw_result != NULL, "must have an allocation result"); 3019 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 3020 Node *use = raw_result->fast_out(i); 3021 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3022 Node* addp2 = find_second_addp(use, raw_result); 3023 if (addp2 != NULL) { 3024 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3025 alloc_worklist.append_if_missing(addp2); 3026 } 3027 alloc_worklist.append_if_missing(use); 3028 } else if (use->is_MemBar()) { 3029 memnode_worklist.append_if_missing(use); 3030 } 3031 } 3032 } 3033 } else if (n->is_AddP()) { 3034 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3035 if (jobj == NULL || jobj == phantom_obj) { 3036 #ifdef ASSERT 3037 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3038 ptnode_adr(n->_idx)->dump(); 3039 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3040 #endif 3041 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3042 return; 3043 } 3044 Node *base = get_map(jobj->idx()); // CheckCastPP node 3045 if (!split_AddP(n, base)) continue; // wrong type from dead path 3046 } else if (n->is_Phi() || 3047 n->is_CheckCastPP() || 3048 n->is_EncodeP() || 3049 n->is_DecodeN() || 3050 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3051 if (visited.test_set(n->_idx)) { 3052 assert(n->is_Phi(), "loops only through Phi's"); 3053 continue; // already processed 3054 } 3055 JavaObjectNode* jobj = unique_java_object(n); 3056 if (jobj == NULL || jobj == phantom_obj) { 3057 #ifdef ASSERT 3058 ptnode_adr(n->_idx)->dump(); 3059 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3060 #endif 3061 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3062 return; 3063 } else { 3064 Node *val = get_map(jobj->idx()); // CheckCastPP node 3065 TypeNode *tn = n->as_Type(); 3066 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3067 assert(tinst != NULL && tinst->is_known_instance() && 3068 tinst->instance_id() == jobj->idx() , "instance type expected."); 3069 3070 const Type *tn_type = igvn->type(tn); 3071 const TypeOopPtr *tn_t; 3072 if (tn_type->isa_narrowoop()) { 3073 tn_t = tn_type->make_ptr()->isa_oopptr(); 3074 } else { 3075 tn_t = tn_type->isa_oopptr(); 3076 } 3077 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3078 if (tn_type->isa_narrowoop()) { 3079 tn_type = tinst->make_narrowoop(); 3080 } else { 3081 tn_type = tinst; 3082 } 3083 igvn->hash_delete(tn); 3084 igvn->set_type(tn, tn_type); 3085 tn->set_type(tn_type); 3086 igvn->hash_insert(tn); 3087 record_for_optimizer(n); 3088 } else { 3089 assert(tn_type == TypePtr::NULL_PTR || 3090 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3091 "unexpected type"); 3092 continue; // Skip dead path with different type 3093 } 3094 } 3095 } else { 3096 debug_only(n->dump();) 3097 assert(false, "EA: unexpected node"); 3098 continue; 3099 } 3100 // push allocation's users on appropriate worklist 3101 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3102 Node *use = n->fast_out(i); 3103 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3104 // Load/store to instance's field 3105 memnode_worklist.append_if_missing(use); 3106 } else if (use->is_MemBar()) { 3107 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3108 memnode_worklist.append_if_missing(use); 3109 } 3110 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3111 Node* addp2 = find_second_addp(use, n); 3112 if (addp2 != NULL) { 3113 alloc_worklist.append_if_missing(addp2); 3114 } 3115 alloc_worklist.append_if_missing(use); 3116 } else if (use->is_Phi() || 3117 use->is_CheckCastPP() || 3118 use->is_EncodeNarrowPtr() || 3119 use->is_DecodeNarrowPtr() || 3120 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3121 alloc_worklist.append_if_missing(use); 3122 #ifdef ASSERT 3123 } else if (use->is_Mem()) { 3124 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3125 } else if (use->is_MergeMem()) { 3126 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3127 } else if (use->is_SafePoint()) { 3128 // Look for MergeMem nodes for calls which reference unique allocation 3129 // (through CheckCastPP nodes) even for debug info. 3130 Node* m = use->in(TypeFunc::Memory); 3131 if (m->is_MergeMem()) { 3132 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3133 } 3134 } else if (use->Opcode() == Op_EncodeISOArray) { 3135 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3136 // EncodeISOArray overwrites destination array 3137 memnode_worklist.append_if_missing(use); 3138 } 3139 } else { 3140 uint op = use->Opcode(); 3141 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) && 3142 (use->in(MemNode::Memory) == n)) { 3143 // They overwrite memory edge corresponding to destination array, 3144 memnode_worklist.append_if_missing(use); 3145 } else if (!(op == Op_CmpP || op == Op_Conv2B || 3146 op == Op_CastP2X || op == Op_StoreCM || 3147 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3148 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3149 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3150 n->dump(); 3151 use->dump(); 3152 assert(false, "EA: missing allocation reference path"); 3153 } 3154 #endif 3155 } 3156 } 3157 3158 } 3159 3160 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3161 // type, record it in the ArrayCopy node so we know what memory this 3162 // node uses/modified. 3163 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3164 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3165 Node* dest = ac->in(ArrayCopyNode::Dest); 3166 if (dest->is_AddP()) { 3167 dest = get_addp_base(dest); 3168 } 3169 JavaObjectNode* jobj = unique_java_object(dest); 3170 if (jobj != NULL) { 3171 Node *base = get_map(jobj->idx()); 3172 if (base != NULL) { 3173 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3174 ac->_dest_type = base_t; 3175 } 3176 } 3177 Node* src = ac->in(ArrayCopyNode::Src); 3178 if (src->is_AddP()) { 3179 src = get_addp_base(src); 3180 } 3181 jobj = unique_java_object(src); 3182 if (jobj != NULL) { 3183 Node* base = get_map(jobj->idx()); 3184 if (base != NULL) { 3185 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3186 ac->_src_type = base_t; 3187 } 3188 } 3189 } 3190 3191 // New alias types were created in split_AddP(). 3192 uint new_index_end = (uint) _compile->num_alias_types(); 3193 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3194 3195 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3196 // compute new values for Memory inputs (the Memory inputs are not 3197 // actually updated until phase 4.) 3198 if (memnode_worklist.length() == 0) 3199 return; // nothing to do 3200 while (memnode_worklist.length() != 0) { 3201 Node *n = memnode_worklist.pop(); 3202 if (visited.test_set(n->_idx)) 3203 continue; 3204 if (n->is_Phi() || n->is_ClearArray()) { 3205 // we don't need to do anything, but the users must be pushed 3206 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3207 // we don't need to do anything, but the users must be pushed 3208 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 3209 if (n == NULL) 3210 continue; 3211 } else if (n->Opcode() == Op_StrCompressedCopy || 3212 n->Opcode() == Op_EncodeISOArray) { 3213 // get the memory projection 3214 n = n->find_out_with(Op_SCMemProj); 3215 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3216 } else { 3217 assert(n->is_Mem(), "memory node required."); 3218 Node *addr = n->in(MemNode::Address); 3219 const Type *addr_t = igvn->type(addr); 3220 if (addr_t == Type::TOP) 3221 continue; 3222 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3223 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3224 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3225 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3226 if (_compile->failing()) { 3227 return; 3228 } 3229 if (mem != n->in(MemNode::Memory)) { 3230 // We delay the memory edge update since we need old one in 3231 // MergeMem code below when instances memory slices are separated. 3232 set_map(n, mem); 3233 } 3234 if (n->is_Load()) { 3235 continue; // don't push users 3236 } else if (n->is_LoadStore()) { 3237 // get the memory projection 3238 n = n->find_out_with(Op_SCMemProj); 3239 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3240 } 3241 } 3242 // push user on appropriate worklist 3243 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3244 Node *use = n->fast_out(i); 3245 if (use->is_Phi() || use->is_ClearArray()) { 3246 memnode_worklist.append_if_missing(use); 3247 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3248 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3249 continue; 3250 memnode_worklist.append_if_missing(use); 3251 } else if (use->is_MemBar()) { 3252 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3253 memnode_worklist.append_if_missing(use); 3254 } 3255 #ifdef ASSERT 3256 } else if(use->is_Mem()) { 3257 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3258 } else if (use->is_MergeMem()) { 3259 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3260 } else if (use->Opcode() == Op_EncodeISOArray) { 3261 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3262 // EncodeISOArray overwrites destination array 3263 memnode_worklist.append_if_missing(use); 3264 } 3265 } else { 3266 uint op = use->Opcode(); 3267 if ((use->in(MemNode::Memory) == n) && 3268 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) { 3269 // They overwrite memory edge corresponding to destination array, 3270 memnode_worklist.append_if_missing(use); 3271 } else if (!(op == Op_StoreCM || 3272 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 3273 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 3274 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3275 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3276 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3277 n->dump(); 3278 use->dump(); 3279 assert(false, "EA: missing memory path"); 3280 } 3281 #endif 3282 } 3283 } 3284 } 3285 3286 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3287 // Walk each memory slice moving the first node encountered of each 3288 // instance type to the the input corresponding to its alias index. 3289 uint length = _mergemem_worklist.length(); 3290 for( uint next = 0; next < length; ++next ) { 3291 MergeMemNode* nmm = _mergemem_worklist.at(next); 3292 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3293 // Note: we don't want to use MergeMemStream here because we only want to 3294 // scan inputs which exist at the start, not ones we add during processing. 3295 // Note 2: MergeMem may already contains instance memory slices added 3296 // during find_inst_mem() call when memory nodes were processed above. 3297 igvn->hash_delete(nmm); 3298 uint nslices = MIN2(nmm->req(), new_index_start); 3299 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3300 Node* mem = nmm->in(i); 3301 Node* cur = NULL; 3302 if (mem == NULL || mem->is_top()) 3303 continue; 3304 // First, update mergemem by moving memory nodes to corresponding slices 3305 // if their type became more precise since this mergemem was created. 3306 while (mem->is_Mem()) { 3307 const Type *at = igvn->type(mem->in(MemNode::Address)); 3308 if (at != Type::TOP) { 3309 assert (at->isa_ptr() != NULL, "pointer type required."); 3310 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3311 if (idx == i) { 3312 if (cur == NULL) 3313 cur = mem; 3314 } else { 3315 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3316 nmm->set_memory_at(idx, mem); 3317 } 3318 } 3319 } 3320 mem = mem->in(MemNode::Memory); 3321 } 3322 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3323 // Find any instance of the current type if we haven't encountered 3324 // already a memory slice of the instance along the memory chain. 3325 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3326 if((uint)_compile->get_general_index(ni) == i) { 3327 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3328 if (nmm->is_empty_memory(m)) { 3329 Node* result = find_inst_mem(mem, ni, orig_phis); 3330 if (_compile->failing()) { 3331 return; 3332 } 3333 nmm->set_memory_at(ni, result); 3334 } 3335 } 3336 } 3337 } 3338 // Find the rest of instances values 3339 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3340 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3341 Node* result = step_through_mergemem(nmm, ni, tinst); 3342 if (result == nmm->base_memory()) { 3343 // Didn't find instance memory, search through general slice recursively. 3344 result = nmm->memory_at(_compile->get_general_index(ni)); 3345 result = find_inst_mem(result, ni, orig_phis); 3346 if (_compile->failing()) { 3347 return; 3348 } 3349 nmm->set_memory_at(ni, result); 3350 } 3351 } 3352 igvn->hash_insert(nmm); 3353 record_for_optimizer(nmm); 3354 } 3355 3356 // Phase 4: Update the inputs of non-instance memory Phis and 3357 // the Memory input of memnodes 3358 // First update the inputs of any non-instance Phi's from 3359 // which we split out an instance Phi. Note we don't have 3360 // to recursively process Phi's encounted on the input memory 3361 // chains as is done in split_memory_phi() since they will 3362 // also be processed here. 3363 for (int j = 0; j < orig_phis.length(); j++) { 3364 PhiNode *phi = orig_phis.at(j); 3365 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3366 igvn->hash_delete(phi); 3367 for (uint i = 1; i < phi->req(); i++) { 3368 Node *mem = phi->in(i); 3369 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3370 if (_compile->failing()) { 3371 return; 3372 } 3373 if (mem != new_mem) { 3374 phi->set_req(i, new_mem); 3375 } 3376 } 3377 igvn->hash_insert(phi); 3378 record_for_optimizer(phi); 3379 } 3380 3381 // Update the memory inputs of MemNodes with the value we computed 3382 // in Phase 2 and move stores memory users to corresponding memory slices. 3383 // Disable memory split verification code until the fix for 6984348. 3384 // Currently it produces false negative results since it does not cover all cases. 3385 #if 0 // ifdef ASSERT 3386 visited.Reset(); 3387 Node_Stack old_mems(arena, _compile->unique() >> 2); 3388 #endif 3389 for (uint i = 0; i < ideal_nodes.size(); i++) { 3390 Node* n = ideal_nodes.at(i); 3391 Node* nmem = get_map(n->_idx); 3392 assert(nmem != NULL, "sanity"); 3393 if (n->is_Mem()) { 3394 #if 0 // ifdef ASSERT 3395 Node* old_mem = n->in(MemNode::Memory); 3396 if (!visited.test_set(old_mem->_idx)) { 3397 old_mems.push(old_mem, old_mem->outcnt()); 3398 } 3399 #endif 3400 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3401 if (!n->is_Load()) { 3402 // Move memory users of a store first. 3403 move_inst_mem(n, orig_phis); 3404 } 3405 // Now update memory input 3406 igvn->hash_delete(n); 3407 n->set_req(MemNode::Memory, nmem); 3408 igvn->hash_insert(n); 3409 record_for_optimizer(n); 3410 } else { 3411 assert(n->is_Allocate() || n->is_CheckCastPP() || 3412 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3413 } 3414 } 3415 #if 0 // ifdef ASSERT 3416 // Verify that memory was split correctly 3417 while (old_mems.is_nonempty()) { 3418 Node* old_mem = old_mems.node(); 3419 uint old_cnt = old_mems.index(); 3420 old_mems.pop(); 3421 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3422 } 3423 #endif 3424 } 3425 3426 #ifndef PRODUCT 3427 static const char *node_type_names[] = { 3428 "UnknownType", 3429 "JavaObject", 3430 "LocalVar", 3431 "Field", 3432 "Arraycopy" 3433 }; 3434 3435 static const char *esc_names[] = { 3436 "UnknownEscape", 3437 "NoEscape", 3438 "ArgEscape", 3439 "GlobalEscape" 3440 }; 3441 3442 void PointsToNode::dump(bool print_state) const { 3443 NodeType nt = node_type(); 3444 tty->print("%s ", node_type_names[(int) nt]); 3445 if (print_state) { 3446 EscapeState es = escape_state(); 3447 EscapeState fields_es = fields_escape_state(); 3448 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3449 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3450 tty->print("NSR "); 3451 } 3452 if (is_Field()) { 3453 FieldNode* f = (FieldNode*)this; 3454 if (f->is_oop()) 3455 tty->print("oop "); 3456 if (f->offset() > 0) 3457 tty->print("+%d ", f->offset()); 3458 tty->print("("); 3459 for (BaseIterator i(f); i.has_next(); i.next()) { 3460 PointsToNode* b = i.get(); 3461 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3462 } 3463 tty->print(" )"); 3464 } 3465 tty->print("["); 3466 for (EdgeIterator i(this); i.has_next(); i.next()) { 3467 PointsToNode* e = i.get(); 3468 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3469 } 3470 tty->print(" ["); 3471 for (UseIterator i(this); i.has_next(); i.next()) { 3472 PointsToNode* u = i.get(); 3473 bool is_base = false; 3474 if (PointsToNode::is_base_use(u)) { 3475 is_base = true; 3476 u = PointsToNode::get_use_node(u)->as_Field(); 3477 } 3478 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3479 } 3480 tty->print(" ]] "); 3481 if (_node == NULL) 3482 tty->print_cr("<null>"); 3483 else 3484 _node->dump(); 3485 } 3486 3487 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3488 bool first = true; 3489 int ptnodes_length = ptnodes_worklist.length(); 3490 for (int i = 0; i < ptnodes_length; i++) { 3491 PointsToNode *ptn = ptnodes_worklist.at(i); 3492 if (ptn == NULL || !ptn->is_JavaObject()) 3493 continue; 3494 PointsToNode::EscapeState es = ptn->escape_state(); 3495 if ((es != PointsToNode::NoEscape) && !Verbose) { 3496 continue; 3497 } 3498 Node* n = ptn->ideal_node(); 3499 if (n->is_Allocate() || (n->is_CallStaticJava() && 3500 n->as_CallStaticJava()->is_boxing_method())) { 3501 if (first) { 3502 tty->cr(); 3503 tty->print("======== Connection graph for "); 3504 _compile->method()->print_short_name(); 3505 tty->cr(); 3506 first = false; 3507 } 3508 ptn->dump(); 3509 // Print all locals and fields which reference this allocation 3510 for (UseIterator j(ptn); j.has_next(); j.next()) { 3511 PointsToNode* use = j.get(); 3512 if (use->is_LocalVar()) { 3513 use->dump(Verbose); 3514 } else if (Verbose) { 3515 use->dump(); 3516 } 3517 } 3518 tty->cr(); 3519 } 3520 } 3521 } 3522 #endif