1 /* 2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/bcEscapeAnalyzer.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/allocation.hpp" 30 #include "opto/c2compiler.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/compile.hpp" 35 #include "opto/escape.hpp" 36 #include "opto/phaseX.hpp" 37 #include "opto/movenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/shenandoahSupport.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_GetAndSetP: 495 case Op_GetAndSetN: { 496 add_objload_to_connection_graph(n, delayed_worklist); 497 // fallthrough 498 } 499 case Op_StoreP: 500 case Op_StoreN: 501 case Op_StoreNKlass: 502 case Op_StorePConditional: 503 case Op_CompareAndSwapP: 504 case Op_CompareAndSwapN: { 505 Node* adr = n->in(MemNode::Address); 506 const Type *adr_type = igvn->type(adr); 507 adr_type = adr_type->make_ptr(); 508 if (adr_type == NULL) { 509 break; // skip dead nodes 510 } 511 if (adr_type->isa_oopptr() || 512 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 513 (adr_type == TypeRawPtr::NOTNULL && 514 adr->in(AddPNode::Address)->is_Proj() && 515 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 516 delayed_worklist->push(n); // Process it later. 517 #ifdef ASSERT 518 assert(adr->is_AddP(), "expecting an AddP"); 519 if (adr_type == TypeRawPtr::NOTNULL) { 520 // Verify a raw address for a store captured by Initialize node. 521 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 522 assert(offs != Type::OffsetBot, "offset must be a constant"); 523 } 524 #endif 525 } else { 526 // Ignore copy the displaced header to the BoxNode (OSR compilation). 527 if (adr->is_BoxLock()) 528 break; 529 // Stored value escapes in unsafe access. 530 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 531 // Pointer stores in G1 barriers looks like unsafe access. 532 // Ignore such stores to be able scalar replace non-escaping 533 // allocations. 534 if ((UseG1GC || UseShenandoahGC) && adr->is_AddP()) { 535 Node* base = get_addp_base(adr); 536 if (base->Opcode() == Op_LoadP && 537 base->in(MemNode::Address)->is_AddP()) { 538 adr = base->in(MemNode::Address); 539 Node* tls = get_addp_base(adr); 540 if (tls->Opcode() == Op_ThreadLocal) { 541 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 542 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() + 543 PtrQueue::byte_offset_of_buf())) { 544 break; // G1 pre barrier previous oop value store. 545 } 546 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() + 547 PtrQueue::byte_offset_of_buf())) { 548 break; // G1 post barrier card address store. 549 } 550 } 551 } 552 } 553 delayed_worklist->push(n); // Process unsafe access later. 554 break; 555 } 556 #ifdef ASSERT 557 n->dump(1); 558 assert(false, "not unsafe or G1 barrier raw StoreP"); 559 #endif 560 } 561 break; 562 } 563 case Op_AryEq: 564 case Op_StrComp: 565 case Op_StrEquals: 566 case Op_StrIndexOf: 567 case Op_EncodeISOArray: { 568 add_local_var(n, PointsToNode::ArgEscape); 569 delayed_worklist->push(n); // Process it later. 570 break; 571 } 572 case Op_ThreadLocal: { 573 add_java_object(n, PointsToNode::ArgEscape); 574 break; 575 } 576 case Op_ShenandoahReadBarrier: 577 case Op_ShenandoahWriteBarrier: 578 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 579 // It doesn't escape. 580 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), delayed_worklist); 581 break; 582 default: 583 ; // Do nothing for nodes not related to EA. 584 } 585 return; 586 } 587 588 #ifdef ASSERT 589 #define ELSE_FAIL(name) \ 590 /* Should not be called for not pointer type. */ \ 591 n->dump(1); \ 592 assert(false, name); \ 593 break; 594 #else 595 #define ELSE_FAIL(name) \ 596 break; 597 #endif 598 599 // Add final simple edges to graph. 600 void ConnectionGraph::add_final_edges(Node *n) { 601 PointsToNode* n_ptn = ptnode_adr(n->_idx); 602 #ifdef ASSERT 603 if (_verify && n_ptn->is_JavaObject()) 604 return; // This method does not change graph for JavaObject. 605 #endif 606 607 if (n->is_Call()) { 608 process_call_arguments(n->as_Call()); 609 return; 610 } 611 assert(n->is_Store() || n->is_LoadStore() || 612 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 613 "node should be registered already"); 614 int opcode = n->Opcode(); 615 switch (opcode) { 616 case Op_AddP: { 617 Node* base = get_addp_base(n); 618 PointsToNode* ptn_base = ptnode_adr(base->_idx); 619 assert(ptn_base != NULL, "field's base should be registered"); 620 add_base(n_ptn->as_Field(), ptn_base); 621 break; 622 } 623 case Op_CastPP: 624 case Op_CheckCastPP: 625 case Op_EncodeP: 626 case Op_DecodeN: 627 case Op_EncodePKlass: 628 case Op_DecodeNKlass: { 629 add_local_var_and_edge(n, PointsToNode::NoEscape, 630 n->in(1), NULL); 631 break; 632 } 633 case Op_CMoveP: { 634 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 635 Node* in = n->in(i); 636 if (in == NULL) 637 continue; // ignore NULL 638 Node* uncast_in = in->uncast(); 639 if (uncast_in->is_top() || uncast_in == n) 640 continue; // ignore top or inputs which go back this node 641 PointsToNode* ptn = ptnode_adr(in->_idx); 642 assert(ptn != NULL, "node should be registered"); 643 add_edge(n_ptn, ptn); 644 } 645 break; 646 } 647 case Op_LoadP: 648 case Op_LoadN: 649 case Op_LoadPLocked: { 650 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 651 // ThreadLocal has RawPtr type. 652 const Type* t = _igvn->type(n); 653 if (t->make_ptr() != NULL) { 654 Node* adr = n->in(MemNode::Address); 655 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 656 break; 657 } 658 ELSE_FAIL("Op_LoadP"); 659 } 660 case Op_Phi: { 661 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 662 // ThreadLocal has RawPtr type. 663 const Type* t = n->as_Phi()->type(); 664 if (t->make_ptr() != NULL) { 665 for (uint i = 1; i < n->req(); i++) { 666 Node* in = n->in(i); 667 if (in == NULL) 668 continue; // ignore NULL 669 Node* uncast_in = in->uncast(); 670 if (uncast_in->is_top() || uncast_in == n) 671 continue; // ignore top or inputs which go back this node 672 PointsToNode* ptn = ptnode_adr(in->_idx); 673 assert(ptn != NULL, "node should be registered"); 674 add_edge(n_ptn, ptn); 675 } 676 break; 677 } 678 ELSE_FAIL("Op_Phi"); 679 } 680 case Op_Proj: { 681 // we are only interested in the oop result projection from a call 682 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 683 n->in(0)->as_Call()->returns_pointer()) { 684 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 685 break; 686 } 687 ELSE_FAIL("Op_Proj"); 688 } 689 case Op_Rethrow: // Exception object escapes 690 case Op_Return: { 691 if (n->req() > TypeFunc::Parms && 692 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 693 // Treat Return value as LocalVar with GlobalEscape escape state. 694 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 695 n->in(TypeFunc::Parms), NULL); 696 break; 697 } 698 ELSE_FAIL("Op_Return"); 699 } 700 case Op_StoreP: 701 case Op_StoreN: 702 case Op_StoreNKlass: 703 case Op_StorePConditional: 704 case Op_CompareAndSwapP: 705 case Op_CompareAndSwapN: 706 case Op_GetAndSetP: 707 case Op_GetAndSetN: { 708 Node* adr = n->in(MemNode::Address); 709 const Type *adr_type = _igvn->type(adr); 710 adr_type = adr_type->make_ptr(); 711 #ifdef ASSERT 712 if (adr_type == NULL) { 713 n->dump(1); 714 assert(adr_type != NULL, "dead node should not be on list"); 715 break; 716 } 717 #endif 718 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) { 719 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 720 } 721 if (adr_type->isa_oopptr() || 722 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) && 723 (adr_type == TypeRawPtr::NOTNULL && 724 adr->in(AddPNode::Address)->is_Proj() && 725 adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 726 // Point Address to Value 727 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 728 assert(adr_ptn != NULL && 729 adr_ptn->as_Field()->is_oop(), "node should be registered"); 730 Node *val = n->in(MemNode::ValueIn); 731 PointsToNode* ptn = ptnode_adr(val->_idx); 732 assert(ptn != NULL, "node should be registered"); 733 add_edge(adr_ptn, ptn); 734 break; 735 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 736 // Stored value escapes in unsafe access. 737 Node *val = n->in(MemNode::ValueIn); 738 PointsToNode* ptn = ptnode_adr(val->_idx); 739 assert(ptn != NULL, "node should be registered"); 740 set_escape_state(ptn, PointsToNode::GlobalEscape); 741 // Add edge to object for unsafe access with offset. 742 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 743 assert(adr_ptn != NULL, "node should be registered"); 744 if (adr_ptn->is_Field()) { 745 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 746 add_edge(adr_ptn, ptn); 747 } 748 break; 749 } 750 ELSE_FAIL("Op_StoreP"); 751 } 752 case Op_AryEq: 753 case Op_StrComp: 754 case Op_StrEquals: 755 case Op_StrIndexOf: 756 case Op_EncodeISOArray: { 757 // char[] arrays passed to string intrinsic do not escape but 758 // they are not scalar replaceable. Adjust escape state for them. 759 // Start from in(2) edge since in(1) is memory edge. 760 for (uint i = 2; i < n->req(); i++) { 761 Node* adr = n->in(i); 762 const Type* at = _igvn->type(adr); 763 if (!adr->is_top() && at->isa_ptr()) { 764 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 765 at->isa_ptr() != NULL, "expecting a pointer"); 766 if (adr->is_AddP()) { 767 adr = get_addp_base(adr); 768 } 769 PointsToNode* ptn = ptnode_adr(adr->_idx); 770 assert(ptn != NULL, "node should be registered"); 771 add_edge(n_ptn, ptn); 772 } 773 } 774 break; 775 } 776 case Op_ShenandoahReadBarrier: 777 case Op_ShenandoahWriteBarrier: 778 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 779 // It doesn't escape. 780 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), NULL); 781 break; 782 default: { 783 // This method should be called only for EA specific nodes which may 784 // miss some edges when they were created. 785 #ifdef ASSERT 786 n->dump(1); 787 #endif 788 guarantee(false, "unknown node"); 789 } 790 } 791 return; 792 } 793 794 void ConnectionGraph::add_call_node(CallNode* call) { 795 assert(call->returns_pointer(), "only for call which returns pointer"); 796 uint call_idx = call->_idx; 797 if (call->is_Allocate()) { 798 Node* k = call->in(AllocateNode::KlassNode); 799 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 800 assert(kt != NULL, "TypeKlassPtr required."); 801 ciKlass* cik = kt->klass(); 802 PointsToNode::EscapeState es = PointsToNode::NoEscape; 803 bool scalar_replaceable = true; 804 if (call->is_AllocateArray()) { 805 if (!cik->is_array_klass()) { // StressReflectiveCode 806 es = PointsToNode::GlobalEscape; 807 } else { 808 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 809 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 810 // Not scalar replaceable if the length is not constant or too big. 811 scalar_replaceable = false; 812 } 813 } 814 } else { // Allocate instance 815 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 816 cik->is_subclass_of(_compile->env()->Reference_klass()) || 817 !cik->is_instance_klass() || // StressReflectiveCode 818 cik->as_instance_klass()->has_finalizer()) { 819 es = PointsToNode::GlobalEscape; 820 } 821 } 822 add_java_object(call, es); 823 PointsToNode* ptn = ptnode_adr(call_idx); 824 if (!scalar_replaceable && ptn->scalar_replaceable()) { 825 ptn->set_scalar_replaceable(false); 826 } 827 } else if (call->is_CallStaticJava()) { 828 // Call nodes could be different types: 829 // 830 // 1. CallDynamicJavaNode (what happened during call is unknown): 831 // 832 // - mapped to GlobalEscape JavaObject node if oop is returned; 833 // 834 // - all oop arguments are escaping globally; 835 // 836 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 837 // 838 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 839 // 840 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 841 // - mapped to NoEscape JavaObject node if non-escaping object allocated 842 // during call is returned; 843 // - mapped to ArgEscape LocalVar node pointed to object arguments 844 // which are returned and does not escape during call; 845 // 846 // - oop arguments escaping status is defined by bytecode analysis; 847 // 848 // For a static call, we know exactly what method is being called. 849 // Use bytecode estimator to record whether the call's return value escapes. 850 ciMethod* meth = call->as_CallJava()->method(); 851 if (meth == NULL) { 852 const char* name = call->as_CallStaticJava()->_name; 853 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 854 // Returns a newly allocated unescaped object. 855 add_java_object(call, PointsToNode::NoEscape); 856 ptnode_adr(call_idx)->set_scalar_replaceable(false); 857 } else if (meth->is_boxing_method()) { 858 // Returns boxing object 859 PointsToNode::EscapeState es; 860 vmIntrinsics::ID intr = meth->intrinsic_id(); 861 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 862 // It does not escape if object is always allocated. 863 es = PointsToNode::NoEscape; 864 } else { 865 // It escapes globally if object could be loaded from cache. 866 es = PointsToNode::GlobalEscape; 867 } 868 add_java_object(call, es); 869 } else { 870 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 871 call_analyzer->copy_dependencies(_compile->dependencies()); 872 if (call_analyzer->is_return_allocated()) { 873 // Returns a newly allocated unescaped object, simply 874 // update dependency information. 875 // Mark it as NoEscape so that objects referenced by 876 // it's fields will be marked as NoEscape at least. 877 add_java_object(call, PointsToNode::NoEscape); 878 ptnode_adr(call_idx)->set_scalar_replaceable(false); 879 } else { 880 // Determine whether any arguments are returned. 881 const TypeTuple* d = call->tf()->domain(); 882 bool ret_arg = false; 883 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 884 if (d->field_at(i)->isa_ptr() != NULL && 885 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 886 ret_arg = true; 887 break; 888 } 889 } 890 if (ret_arg) { 891 add_local_var(call, PointsToNode::ArgEscape); 892 } else { 893 // Returns unknown object. 894 map_ideal_node(call, phantom_obj); 895 } 896 } 897 } 898 } else { 899 // An other type of call, assume the worst case: 900 // returned value is unknown and globally escapes. 901 assert(call->Opcode() == Op_CallDynamicJava || call->Opcode() == Op_CallLeaf, "add failed case check"); 902 map_ideal_node(call, phantom_obj); 903 } 904 } 905 906 void ConnectionGraph::process_call_arguments(CallNode *call) { 907 bool is_arraycopy = false; 908 switch (call->Opcode()) { 909 #ifdef ASSERT 910 case Op_Allocate: 911 case Op_AllocateArray: 912 case Op_Lock: 913 case Op_Unlock: 914 assert(false, "should be done already"); 915 break; 916 #endif 917 case Op_ArrayCopy: 918 case Op_CallLeafNoFP: 919 // Most array copies are ArrayCopy nodes at this point but there 920 // are still a few direct calls to the copy subroutines (See 921 // PhaseStringOpts::copy_string()) 922 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 923 call->as_CallLeaf()->is_call_to_arraycopystub(); 924 // fall through 925 case Op_CallLeaf: { 926 // Stub calls, objects do not escape but they are not scale replaceable. 927 // Adjust escape state for outgoing arguments. 928 const TypeTuple * d = call->tf()->domain(); 929 bool src_has_oops = false; 930 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 931 const Type* at = d->field_at(i); 932 Node *arg = call->in(i); 933 if (arg == NULL) { 934 continue; 935 } 936 const Type *aat = _igvn->type(arg); 937 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 938 continue; 939 if (arg->is_AddP()) { 940 // 941 // The inline_native_clone() case when the arraycopy stub is called 942 // after the allocation before Initialize and CheckCastPP nodes. 943 // Or normal arraycopy for object arrays case. 944 // 945 // Set AddP's base (Allocate) as not scalar replaceable since 946 // pointer to the base (with offset) is passed as argument. 947 // 948 arg = get_addp_base(arg); 949 } 950 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 951 assert(arg_ptn != NULL, "should be registered"); 952 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 953 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 954 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 955 aat->isa_ptr() != NULL, "expecting an Ptr"); 956 bool arg_has_oops = aat->isa_oopptr() && 957 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 958 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 959 if (i == TypeFunc::Parms) { 960 src_has_oops = arg_has_oops; 961 } 962 // 963 // src or dst could be j.l.Object when other is basic type array: 964 // 965 // arraycopy(char[],0,Object*,0,size); 966 // arraycopy(Object*,0,char[],0,size); 967 // 968 // Don't add edges in such cases. 969 // 970 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 971 arg_has_oops && (i > TypeFunc::Parms); 972 #ifdef ASSERT 973 if (!(is_arraycopy || 974 (call->as_CallLeaf()->_name != NULL && 975 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 976 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 || 977 strcmp(call->as_CallLeaf()->_name, "shenandoah_clone_barrier") == 0 || 978 strcmp(call->as_CallLeaf()->_name, "shenandoah_read_barrier") == 0 || 979 strcmp(call->as_CallLeaf()->_name, "shenandoah_cas_obj") == 0 || 980 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 981 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 982 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 983 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 984 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 985 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 986 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 987 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 988 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 989 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 990 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 991 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 992 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 993 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 994 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 995 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 996 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 997 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 998 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0) 999 ))) { 1000 call->dump(); 1001 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name)); 1002 } 1003 #endif 1004 // Always process arraycopy's destination object since 1005 // we need to add all possible edges to references in 1006 // source object. 1007 if (arg_esc >= PointsToNode::ArgEscape && 1008 !arg_is_arraycopy_dest) { 1009 continue; 1010 } 1011 PointsToNode::EscapeState es = PointsToNode::ArgEscape; 1012 if (call->is_ArrayCopy()) { 1013 ArrayCopyNode* ac = call->as_ArrayCopy(); 1014 if (ac->is_clonebasic() || 1015 ac->is_arraycopy_validated() || 1016 ac->is_copyof_validated() || 1017 ac->is_copyofrange_validated()) { 1018 es = PointsToNode::NoEscape; 1019 } 1020 } 1021 set_escape_state(arg_ptn, es); 1022 if (arg_is_arraycopy_dest) { 1023 Node* src = call->in(TypeFunc::Parms); 1024 if (src->is_AddP()) { 1025 src = get_addp_base(src); 1026 } 1027 PointsToNode* src_ptn = ptnode_adr(src->_idx); 1028 assert(src_ptn != NULL, "should be registered"); 1029 if (arg_ptn != src_ptn) { 1030 // Special arraycopy edge: 1031 // A destination object's field can't have the source object 1032 // as base since objects escape states are not related. 1033 // Only escape state of destination object's fields affects 1034 // escape state of fields in source object. 1035 add_arraycopy(call, es, src_ptn, arg_ptn); 1036 } 1037 } 1038 } 1039 } 1040 break; 1041 } 1042 case Op_CallStaticJava: { 1043 // For a static call, we know exactly what method is being called. 1044 // Use bytecode estimator to record the call's escape affects 1045 #ifdef ASSERT 1046 const char* name = call->as_CallStaticJava()->_name; 1047 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1048 #endif 1049 ciMethod* meth = call->as_CallJava()->method(); 1050 if ((meth != NULL) && meth->is_boxing_method()) { 1051 break; // Boxing methods do not modify any oops. 1052 } 1053 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1054 // fall-through if not a Java method or no analyzer information 1055 if (call_analyzer != NULL) { 1056 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1057 const TypeTuple* d = call->tf()->domain(); 1058 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1059 const Type* at = d->field_at(i); 1060 int k = i - TypeFunc::Parms; 1061 Node* arg = call->in(i); 1062 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1063 if (at->isa_ptr() != NULL && 1064 call_analyzer->is_arg_returned(k)) { 1065 // The call returns arguments. 1066 if (call_ptn != NULL) { // Is call's result used? 1067 assert(call_ptn->is_LocalVar(), "node should be registered"); 1068 assert(arg_ptn != NULL, "node should be registered"); 1069 add_edge(call_ptn, arg_ptn); 1070 } 1071 } 1072 if (at->isa_oopptr() != NULL && 1073 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1074 if (!call_analyzer->is_arg_stack(k)) { 1075 // The argument global escapes 1076 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1077 } else { 1078 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1079 if (!call_analyzer->is_arg_local(k)) { 1080 // The argument itself doesn't escape, but any fields might 1081 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1082 } 1083 } 1084 } 1085 } 1086 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1087 // The call returns arguments. 1088 assert(call_ptn->edge_count() > 0, "sanity"); 1089 if (!call_analyzer->is_return_local()) { 1090 // Returns also unknown object. 1091 add_edge(call_ptn, phantom_obj); 1092 } 1093 } 1094 break; 1095 } 1096 } 1097 default: { 1098 // Fall-through here if not a Java method or no analyzer information 1099 // or some other type of call, assume the worst case: all arguments 1100 // globally escape. 1101 const TypeTuple* d = call->tf()->domain(); 1102 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1103 const Type* at = d->field_at(i); 1104 if (at->isa_oopptr() != NULL) { 1105 Node* arg = call->in(i); 1106 if (arg->is_AddP()) { 1107 arg = get_addp_base(arg); 1108 } 1109 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1110 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1111 } 1112 } 1113 } 1114 } 1115 } 1116 1117 1118 // Finish Graph construction. 1119 bool ConnectionGraph::complete_connection_graph( 1120 GrowableArray<PointsToNode*>& ptnodes_worklist, 1121 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1122 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1123 GrowableArray<FieldNode*>& oop_fields_worklist) { 1124 // Normally only 1-3 passes needed to build Connection Graph depending 1125 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1126 // Set limit to 20 to catch situation when something did go wrong and 1127 // bailout Escape Analysis. 1128 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1129 #define CG_BUILD_ITER_LIMIT 20 1130 1131 // Propagate GlobalEscape and ArgEscape escape states and check that 1132 // we still have non-escaping objects. The method pushs on _worklist 1133 // Field nodes which reference phantom_object. 1134 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1135 return false; // Nothing to do. 1136 } 1137 // Now propagate references to all JavaObject nodes. 1138 int java_objects_length = java_objects_worklist.length(); 1139 elapsedTimer time; 1140 bool timeout = false; 1141 int new_edges = 1; 1142 int iterations = 0; 1143 do { 1144 while ((new_edges > 0) && 1145 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1146 double start_time = time.seconds(); 1147 time.start(); 1148 new_edges = 0; 1149 // Propagate references to phantom_object for nodes pushed on _worklist 1150 // by find_non_escaped_objects() and find_field_value(). 1151 new_edges += add_java_object_edges(phantom_obj, false); 1152 for (int next = 0; next < java_objects_length; ++next) { 1153 JavaObjectNode* ptn = java_objects_worklist.at(next); 1154 new_edges += add_java_object_edges(ptn, true); 1155 1156 #define SAMPLE_SIZE 4 1157 if ((next % SAMPLE_SIZE) == 0) { 1158 // Each 4 iterations calculate how much time it will take 1159 // to complete graph construction. 1160 time.stop(); 1161 // Poll for requests from shutdown mechanism to quiesce compiler 1162 // because Connection graph construction may take long time. 1163 CompileBroker::maybe_block(); 1164 double stop_time = time.seconds(); 1165 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1166 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1167 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1168 timeout = true; 1169 break; // Timeout 1170 } 1171 start_time = stop_time; 1172 time.start(); 1173 } 1174 #undef SAMPLE_SIZE 1175 1176 } 1177 if (timeout) break; 1178 if (new_edges > 0) { 1179 // Update escape states on each iteration if graph was updated. 1180 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1181 return false; // Nothing to do. 1182 } 1183 } 1184 time.stop(); 1185 if (time.seconds() >= EscapeAnalysisTimeout) { 1186 timeout = true; 1187 break; 1188 } 1189 } 1190 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1191 time.start(); 1192 // Find fields which have unknown value. 1193 int fields_length = oop_fields_worklist.length(); 1194 for (int next = 0; next < fields_length; next++) { 1195 FieldNode* field = oop_fields_worklist.at(next); 1196 if (field->edge_count() == 0) { 1197 new_edges += find_field_value(field); 1198 // This code may added new edges to phantom_object. 1199 // Need an other cycle to propagate references to phantom_object. 1200 } 1201 } 1202 time.stop(); 1203 if (time.seconds() >= EscapeAnalysisTimeout) { 1204 timeout = true; 1205 break; 1206 } 1207 } else { 1208 new_edges = 0; // Bailout 1209 } 1210 } while (new_edges > 0); 1211 1212 // Bailout if passed limits. 1213 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1214 Compile* C = _compile; 1215 if (C->log() != NULL) { 1216 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1217 C->log()->text("%s", timeout ? "time" : "iterations"); 1218 C->log()->end_elem(" limit'"); 1219 } 1220 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1221 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length())); 1222 // Possible infinite build_connection_graph loop, 1223 // bailout (no changes to ideal graph were made). 1224 return false; 1225 } 1226 #ifdef ASSERT 1227 if (Verbose && PrintEscapeAnalysis) { 1228 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1229 iterations, nodes_size(), ptnodes_worklist.length()); 1230 } 1231 #endif 1232 1233 #undef CG_BUILD_ITER_LIMIT 1234 1235 // Find fields initialized by NULL for non-escaping Allocations. 1236 int non_escaped_length = non_escaped_worklist.length(); 1237 for (int next = 0; next < non_escaped_length; next++) { 1238 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1239 PointsToNode::EscapeState es = ptn->escape_state(); 1240 assert(es <= PointsToNode::ArgEscape, "sanity"); 1241 if (es == PointsToNode::NoEscape) { 1242 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1243 // Adding references to NULL object does not change escape states 1244 // since it does not escape. Also no fields are added to NULL object. 1245 add_java_object_edges(null_obj, false); 1246 } 1247 } 1248 Node* n = ptn->ideal_node(); 1249 if (n->is_Allocate()) { 1250 // The object allocated by this Allocate node will never be 1251 // seen by an other thread. Mark it so that when it is 1252 // expanded no MemBarStoreStore is added. 1253 InitializeNode* ini = n->as_Allocate()->initialization(); 1254 if (ini != NULL) 1255 ini->set_does_not_escape(); 1256 } 1257 } 1258 return true; // Finished graph construction. 1259 } 1260 1261 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1262 // and check that we still have non-escaping java objects. 1263 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1264 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1265 GrowableArray<PointsToNode*> escape_worklist; 1266 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1267 int ptnodes_length = ptnodes_worklist.length(); 1268 for (int next = 0; next < ptnodes_length; ++next) { 1269 PointsToNode* ptn = ptnodes_worklist.at(next); 1270 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1271 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1272 escape_worklist.push(ptn); 1273 } 1274 } 1275 // Set escape states to referenced nodes (edges list). 1276 while (escape_worklist.length() > 0) { 1277 PointsToNode* ptn = escape_worklist.pop(); 1278 PointsToNode::EscapeState es = ptn->escape_state(); 1279 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1280 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1281 es >= PointsToNode::ArgEscape) { 1282 // GlobalEscape or ArgEscape state of field means it has unknown value. 1283 if (add_edge(ptn, phantom_obj)) { 1284 // New edge was added 1285 add_field_uses_to_worklist(ptn->as_Field()); 1286 } 1287 } 1288 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1289 PointsToNode* e = i.get(); 1290 if (e->is_Arraycopy()) { 1291 assert(ptn->arraycopy_dst(), "sanity"); 1292 // Propagate only fields escape state through arraycopy edge. 1293 if (e->fields_escape_state() < field_es) { 1294 set_fields_escape_state(e, field_es); 1295 escape_worklist.push(e); 1296 } 1297 } else if (es >= field_es) { 1298 // fields_escape_state is also set to 'es' if it is less than 'es'. 1299 if (e->escape_state() < es) { 1300 set_escape_state(e, es); 1301 escape_worklist.push(e); 1302 } 1303 } else { 1304 // Propagate field escape state. 1305 bool es_changed = false; 1306 if (e->fields_escape_state() < field_es) { 1307 set_fields_escape_state(e, field_es); 1308 es_changed = true; 1309 } 1310 if ((e->escape_state() < field_es) && 1311 e->is_Field() && ptn->is_JavaObject() && 1312 e->as_Field()->is_oop()) { 1313 // Change escape state of referenced fields. 1314 set_escape_state(e, field_es); 1315 es_changed = true; 1316 } else if (e->escape_state() < es) { 1317 set_escape_state(e, es); 1318 es_changed = true; 1319 } 1320 if (es_changed) { 1321 escape_worklist.push(e); 1322 } 1323 } 1324 } 1325 } 1326 // Remove escaped objects from non_escaped list. 1327 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1328 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1329 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1330 non_escaped_worklist.delete_at(next); 1331 } 1332 if (ptn->escape_state() == PointsToNode::NoEscape) { 1333 // Find fields in non-escaped allocations which have unknown value. 1334 find_init_values(ptn, phantom_obj, NULL); 1335 } 1336 } 1337 return (non_escaped_worklist.length() > 0); 1338 } 1339 1340 // Add all references to JavaObject node by walking over all uses. 1341 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1342 int new_edges = 0; 1343 if (populate_worklist) { 1344 // Populate _worklist by uses of jobj's uses. 1345 for (UseIterator i(jobj); i.has_next(); i.next()) { 1346 PointsToNode* use = i.get(); 1347 if (use->is_Arraycopy()) 1348 continue; 1349 add_uses_to_worklist(use); 1350 if (use->is_Field() && use->as_Field()->is_oop()) { 1351 // Put on worklist all field's uses (loads) and 1352 // related field nodes (same base and offset). 1353 add_field_uses_to_worklist(use->as_Field()); 1354 } 1355 } 1356 } 1357 for (int l = 0; l < _worklist.length(); l++) { 1358 PointsToNode* use = _worklist.at(l); 1359 if (PointsToNode::is_base_use(use)) { 1360 // Add reference from jobj to field and from field to jobj (field's base). 1361 use = PointsToNode::get_use_node(use)->as_Field(); 1362 if (add_base(use->as_Field(), jobj)) { 1363 new_edges++; 1364 } 1365 continue; 1366 } 1367 assert(!use->is_JavaObject(), "sanity"); 1368 if (use->is_Arraycopy()) { 1369 if (jobj == null_obj) // NULL object does not have field edges 1370 continue; 1371 // Added edge from Arraycopy node to arraycopy's source java object 1372 if (add_edge(use, jobj)) { 1373 jobj->set_arraycopy_src(); 1374 new_edges++; 1375 } 1376 // and stop here. 1377 continue; 1378 } 1379 if (!add_edge(use, jobj)) 1380 continue; // No new edge added, there was such edge already. 1381 new_edges++; 1382 if (use->is_LocalVar()) { 1383 add_uses_to_worklist(use); 1384 if (use->arraycopy_dst()) { 1385 for (EdgeIterator i(use); i.has_next(); i.next()) { 1386 PointsToNode* e = i.get(); 1387 if (e->is_Arraycopy()) { 1388 if (jobj == null_obj) // NULL object does not have field edges 1389 continue; 1390 // Add edge from arraycopy's destination java object to Arraycopy node. 1391 if (add_edge(jobj, e)) { 1392 new_edges++; 1393 jobj->set_arraycopy_dst(); 1394 } 1395 } 1396 } 1397 } 1398 } else { 1399 // Added new edge to stored in field values. 1400 // Put on worklist all field's uses (loads) and 1401 // related field nodes (same base and offset). 1402 add_field_uses_to_worklist(use->as_Field()); 1403 } 1404 } 1405 _worklist.clear(); 1406 _in_worklist.Reset(); 1407 return new_edges; 1408 } 1409 1410 // Put on worklist all related field nodes. 1411 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1412 assert(field->is_oop(), "sanity"); 1413 int offset = field->offset(); 1414 add_uses_to_worklist(field); 1415 // Loop over all bases of this field and push on worklist Field nodes 1416 // with the same offset and base (since they may reference the same field). 1417 for (BaseIterator i(field); i.has_next(); i.next()) { 1418 PointsToNode* base = i.get(); 1419 add_fields_to_worklist(field, base); 1420 // Check if the base was source object of arraycopy and go over arraycopy's 1421 // destination objects since values stored to a field of source object are 1422 // accessable by uses (loads) of fields of destination objects. 1423 if (base->arraycopy_src()) { 1424 for (UseIterator j(base); j.has_next(); j.next()) { 1425 PointsToNode* arycp = j.get(); 1426 if (arycp->is_Arraycopy()) { 1427 for (UseIterator k(arycp); k.has_next(); k.next()) { 1428 PointsToNode* abase = k.get(); 1429 if (abase->arraycopy_dst() && abase != base) { 1430 // Look for the same arraycopy reference. 1431 add_fields_to_worklist(field, abase); 1432 } 1433 } 1434 } 1435 } 1436 } 1437 } 1438 } 1439 1440 // Put on worklist all related field nodes. 1441 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1442 int offset = field->offset(); 1443 if (base->is_LocalVar()) { 1444 for (UseIterator j(base); j.has_next(); j.next()) { 1445 PointsToNode* f = j.get(); 1446 if (PointsToNode::is_base_use(f)) { // Field 1447 f = PointsToNode::get_use_node(f); 1448 if (f == field || !f->as_Field()->is_oop()) 1449 continue; 1450 int offs = f->as_Field()->offset(); 1451 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1452 add_to_worklist(f); 1453 } 1454 } 1455 } 1456 } else { 1457 assert(base->is_JavaObject(), "sanity"); 1458 if (// Skip phantom_object since it is only used to indicate that 1459 // this field's content globally escapes. 1460 (base != phantom_obj) && 1461 // NULL object node does not have fields. 1462 (base != null_obj)) { 1463 for (EdgeIterator i(base); i.has_next(); i.next()) { 1464 PointsToNode* f = i.get(); 1465 // Skip arraycopy edge since store to destination object field 1466 // does not update value in source object field. 1467 if (f->is_Arraycopy()) { 1468 assert(base->arraycopy_dst(), "sanity"); 1469 continue; 1470 } 1471 if (f == field || !f->as_Field()->is_oop()) 1472 continue; 1473 int offs = f->as_Field()->offset(); 1474 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1475 add_to_worklist(f); 1476 } 1477 } 1478 } 1479 } 1480 } 1481 1482 // Find fields which have unknown value. 1483 int ConnectionGraph::find_field_value(FieldNode* field) { 1484 // Escaped fields should have init value already. 1485 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1486 int new_edges = 0; 1487 for (BaseIterator i(field); i.has_next(); i.next()) { 1488 PointsToNode* base = i.get(); 1489 if (base->is_JavaObject()) { 1490 // Skip Allocate's fields which will be processed later. 1491 if (base->ideal_node()->is_Allocate()) 1492 return 0; 1493 assert(base == null_obj, "only NULL ptr base expected here"); 1494 } 1495 } 1496 if (add_edge(field, phantom_obj)) { 1497 // New edge was added 1498 new_edges++; 1499 add_field_uses_to_worklist(field); 1500 } 1501 return new_edges; 1502 } 1503 1504 // Find fields initializing values for allocations. 1505 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1506 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1507 int new_edges = 0; 1508 Node* alloc = pta->ideal_node(); 1509 if (init_val == phantom_obj) { 1510 // Do nothing for Allocate nodes since its fields values are 1511 // "known" unless they are initialized by arraycopy/clone. 1512 if (alloc->is_Allocate() && !pta->arraycopy_dst()) 1513 return 0; 1514 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity"); 1515 #ifdef ASSERT 1516 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) { 1517 const char* name = alloc->as_CallStaticJava()->_name; 1518 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1519 } 1520 #endif 1521 // Non-escaped allocation returned from Java or runtime call have 1522 // unknown values in fields. 1523 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1524 PointsToNode* field = i.get(); 1525 if (field->is_Field() && field->as_Field()->is_oop()) { 1526 if (add_edge(field, phantom_obj)) { 1527 // New edge was added 1528 new_edges++; 1529 add_field_uses_to_worklist(field->as_Field()); 1530 } 1531 } 1532 } 1533 return new_edges; 1534 } 1535 assert(init_val == null_obj, "sanity"); 1536 // Do nothing for Call nodes since its fields values are unknown. 1537 if (!alloc->is_Allocate()) 1538 return 0; 1539 1540 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1541 bool visited_bottom_offset = false; 1542 GrowableArray<int> offsets_worklist; 1543 1544 // Check if an oop field's initializing value is recorded and add 1545 // a corresponding NULL if field's value if it is not recorded. 1546 // Connection Graph does not record a default initialization by NULL 1547 // captured by Initialize node. 1548 // 1549 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1550 PointsToNode* field = i.get(); // Field (AddP) 1551 if (!field->is_Field() || !field->as_Field()->is_oop()) 1552 continue; // Not oop field 1553 int offset = field->as_Field()->offset(); 1554 if (offset == Type::OffsetBot) { 1555 if (!visited_bottom_offset) { 1556 // OffsetBot is used to reference array's element, 1557 // always add reference to NULL to all Field nodes since we don't 1558 // known which element is referenced. 1559 if (add_edge(field, null_obj)) { 1560 // New edge was added 1561 new_edges++; 1562 add_field_uses_to_worklist(field->as_Field()); 1563 visited_bottom_offset = true; 1564 } 1565 } 1566 } else { 1567 // Check only oop fields. 1568 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1569 if (adr_type->isa_rawptr()) { 1570 #ifdef ASSERT 1571 // Raw pointers are used for initializing stores so skip it 1572 // since it should be recorded already 1573 Node* base = get_addp_base(field->ideal_node()); 1574 assert(adr_type->isa_rawptr() && base->is_Proj() && 1575 (base->in(0) == alloc),"unexpected pointer type"); 1576 #endif 1577 continue; 1578 } 1579 if (!offsets_worklist.contains(offset)) { 1580 offsets_worklist.append(offset); 1581 Node* value = NULL; 1582 if (ini != NULL) { 1583 // StoreP::memory_type() == T_ADDRESS 1584 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1585 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1586 // Make sure initializing store has the same type as this AddP. 1587 // This AddP may reference non existing field because it is on a 1588 // dead branch of bimorphic call which is not eliminated yet. 1589 if (store != NULL && store->is_Store() && 1590 store->as_Store()->memory_type() == ft) { 1591 value = store->in(MemNode::ValueIn); 1592 #ifdef ASSERT 1593 if (VerifyConnectionGraph) { 1594 // Verify that AddP already points to all objects the value points to. 1595 PointsToNode* val = ptnode_adr(value->_idx); 1596 assert((val != NULL), "should be processed already"); 1597 PointsToNode* missed_obj = NULL; 1598 if (val->is_JavaObject()) { 1599 if (!field->points_to(val->as_JavaObject())) { 1600 missed_obj = val; 1601 } 1602 } else { 1603 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1604 tty->print_cr("----------init store has invalid value -----"); 1605 store->dump(); 1606 val->dump(); 1607 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1608 } 1609 for (EdgeIterator j(val); j.has_next(); j.next()) { 1610 PointsToNode* obj = j.get(); 1611 if (obj->is_JavaObject()) { 1612 if (!field->points_to(obj->as_JavaObject())) { 1613 missed_obj = obj; 1614 break; 1615 } 1616 } 1617 } 1618 } 1619 if (missed_obj != NULL) { 1620 tty->print_cr("----------field---------------------------------"); 1621 field->dump(); 1622 tty->print_cr("----------missed referernce to object-----------"); 1623 missed_obj->dump(); 1624 tty->print_cr("----------object referernced by init store -----"); 1625 store->dump(); 1626 val->dump(); 1627 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1628 } 1629 } 1630 #endif 1631 } else { 1632 // There could be initializing stores which follow allocation. 1633 // For example, a volatile field store is not collected 1634 // by Initialize node. 1635 // 1636 // Need to check for dependent loads to separate such stores from 1637 // stores which follow loads. For now, add initial value NULL so 1638 // that compare pointers optimization works correctly. 1639 } 1640 } 1641 if (value == NULL) { 1642 // A field's initializing value was not recorded. Add NULL. 1643 if (add_edge(field, null_obj)) { 1644 // New edge was added 1645 new_edges++; 1646 add_field_uses_to_worklist(field->as_Field()); 1647 } 1648 } 1649 } 1650 } 1651 } 1652 return new_edges; 1653 } 1654 1655 // Adjust scalar_replaceable state after Connection Graph is built. 1656 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1657 // Search for non-escaping objects which are not scalar replaceable 1658 // and mark them to propagate the state to referenced objects. 1659 1660 // 1. An object is not scalar replaceable if the field into which it is 1661 // stored has unknown offset (stored into unknown element of an array). 1662 // 1663 for (UseIterator i(jobj); i.has_next(); i.next()) { 1664 PointsToNode* use = i.get(); 1665 if (use->is_Arraycopy()) { 1666 continue; 1667 } 1668 if (use->is_Field()) { 1669 FieldNode* field = use->as_Field(); 1670 assert(field->is_oop() && field->scalar_replaceable(), "sanity"); 1671 if (field->offset() == Type::OffsetBot) { 1672 jobj->set_scalar_replaceable(false); 1673 return; 1674 } 1675 // 2. An object is not scalar replaceable if the field into which it is 1676 // stored has multiple bases one of which is null. 1677 if (field->base_count() > 1) { 1678 for (BaseIterator i(field); i.has_next(); i.next()) { 1679 PointsToNode* base = i.get(); 1680 if (base == null_obj) { 1681 jobj->set_scalar_replaceable(false); 1682 return; 1683 } 1684 } 1685 } 1686 } 1687 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1688 // 3. An object is not scalar replaceable if it is merged with other objects. 1689 for (EdgeIterator j(use); j.has_next(); j.next()) { 1690 PointsToNode* ptn = j.get(); 1691 if (ptn->is_JavaObject() && ptn != jobj) { 1692 // Mark all objects. 1693 jobj->set_scalar_replaceable(false); 1694 ptn->set_scalar_replaceable(false); 1695 } 1696 } 1697 if (!jobj->scalar_replaceable()) { 1698 return; 1699 } 1700 } 1701 1702 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1703 if (j.get()->is_Arraycopy()) { 1704 continue; 1705 } 1706 1707 // Non-escaping object node should point only to field nodes. 1708 FieldNode* field = j.get()->as_Field(); 1709 int offset = field->as_Field()->offset(); 1710 1711 // 4. An object is not scalar replaceable if it has a field with unknown 1712 // offset (array's element is accessed in loop). 1713 if (offset == Type::OffsetBot) { 1714 jobj->set_scalar_replaceable(false); 1715 return; 1716 } 1717 // 5. Currently an object is not scalar replaceable if a LoadStore node 1718 // access its field since the field value is unknown after it. 1719 // 1720 Node* n = field->ideal_node(); 1721 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1722 if (n->fast_out(i)->is_LoadStore()) { 1723 jobj->set_scalar_replaceable(false); 1724 return; 1725 } 1726 } 1727 1728 // 6. Or the address may point to more then one object. This may produce 1729 // the false positive result (set not scalar replaceable) 1730 // since the flow-insensitive escape analysis can't separate 1731 // the case when stores overwrite the field's value from the case 1732 // when stores happened on different control branches. 1733 // 1734 // Note: it will disable scalar replacement in some cases: 1735 // 1736 // Point p[] = new Point[1]; 1737 // p[0] = new Point(); // Will be not scalar replaced 1738 // 1739 // but it will save us from incorrect optimizations in next cases: 1740 // 1741 // Point p[] = new Point[1]; 1742 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1743 // 1744 if (field->base_count() > 1) { 1745 for (BaseIterator i(field); i.has_next(); i.next()) { 1746 PointsToNode* base = i.get(); 1747 // Don't take into account LocalVar nodes which 1748 // may point to only one object which should be also 1749 // this field's base by now. 1750 if (base->is_JavaObject() && base != jobj) { 1751 // Mark all bases. 1752 jobj->set_scalar_replaceable(false); 1753 base->set_scalar_replaceable(false); 1754 } 1755 } 1756 } 1757 } 1758 } 1759 1760 #ifdef ASSERT 1761 void ConnectionGraph::verify_connection_graph( 1762 GrowableArray<PointsToNode*>& ptnodes_worklist, 1763 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1764 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1765 GrowableArray<Node*>& addp_worklist) { 1766 // Verify that graph is complete - no new edges could be added. 1767 int java_objects_length = java_objects_worklist.length(); 1768 int non_escaped_length = non_escaped_worklist.length(); 1769 int new_edges = 0; 1770 for (int next = 0; next < java_objects_length; ++next) { 1771 JavaObjectNode* ptn = java_objects_worklist.at(next); 1772 new_edges += add_java_object_edges(ptn, true); 1773 } 1774 assert(new_edges == 0, "graph was not complete"); 1775 // Verify that escape state is final. 1776 int length = non_escaped_worklist.length(); 1777 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1778 assert((non_escaped_length == non_escaped_worklist.length()) && 1779 (non_escaped_length == length) && 1780 (_worklist.length() == 0), "escape state was not final"); 1781 1782 // Verify fields information. 1783 int addp_length = addp_worklist.length(); 1784 for (int next = 0; next < addp_length; ++next ) { 1785 Node* n = addp_worklist.at(next); 1786 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1787 if (field->is_oop()) { 1788 // Verify that field has all bases 1789 Node* base = get_addp_base(n); 1790 PointsToNode* ptn = ptnode_adr(base->_idx); 1791 if (ptn->is_JavaObject()) { 1792 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1793 } else { 1794 assert(ptn->is_LocalVar(), "sanity"); 1795 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1796 PointsToNode* e = i.get(); 1797 if (e->is_JavaObject()) { 1798 assert(field->has_base(e->as_JavaObject()), "sanity"); 1799 } 1800 } 1801 } 1802 // Verify that all fields have initializing values. 1803 if (field->edge_count() == 0) { 1804 tty->print_cr("----------field does not have references----------"); 1805 field->dump(); 1806 for (BaseIterator i(field); i.has_next(); i.next()) { 1807 PointsToNode* base = i.get(); 1808 tty->print_cr("----------field has next base---------------------"); 1809 base->dump(); 1810 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1811 tty->print_cr("----------base has fields-------------------------"); 1812 for (EdgeIterator j(base); j.has_next(); j.next()) { 1813 j.get()->dump(); 1814 } 1815 tty->print_cr("----------base has references---------------------"); 1816 for (UseIterator j(base); j.has_next(); j.next()) { 1817 j.get()->dump(); 1818 } 1819 } 1820 } 1821 for (UseIterator i(field); i.has_next(); i.next()) { 1822 i.get()->dump(); 1823 } 1824 assert(field->edge_count() > 0, "sanity"); 1825 } 1826 } 1827 } 1828 } 1829 #endif 1830 1831 // Optimize ideal graph. 1832 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1833 GrowableArray<Node*>& storestore_worklist) { 1834 Compile* C = _compile; 1835 PhaseIterGVN* igvn = _igvn; 1836 if (EliminateLocks) { 1837 // Mark locks before changing ideal graph. 1838 int cnt = C->macro_count(); 1839 for( int i=0; i < cnt; i++ ) { 1840 Node *n = C->macro_node(i); 1841 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1842 AbstractLockNode* alock = n->as_AbstractLock(); 1843 if (!alock->is_non_esc_obj()) { 1844 if (not_global_escape(alock->obj_node())) { 1845 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1846 // The lock could be marked eliminated by lock coarsening 1847 // code during first IGVN before EA. Replace coarsened flag 1848 // to eliminate all associated locks/unlocks. 1849 #ifdef ASSERT 1850 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1851 #endif 1852 alock->set_non_esc_obj(); 1853 } 1854 } 1855 } 1856 } 1857 } 1858 1859 if (OptimizePtrCompare) { 1860 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1861 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1862 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1863 // Optimize objects compare. 1864 while (ptr_cmp_worklist.length() != 0) { 1865 Node *n = ptr_cmp_worklist.pop(); 1866 Node *res = optimize_ptr_compare(n); 1867 if (res != NULL) { 1868 #ifndef PRODUCT 1869 if (PrintOptimizePtrCompare) { 1870 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")); 1871 if (Verbose) { 1872 n->dump(1); 1873 } 1874 } 1875 #endif 1876 igvn->replace_node(n, res); 1877 } 1878 } 1879 // cleanup 1880 if (_pcmp_neq->outcnt() == 0) 1881 igvn->hash_delete(_pcmp_neq); 1882 if (_pcmp_eq->outcnt() == 0) 1883 igvn->hash_delete(_pcmp_eq); 1884 } 1885 1886 // For MemBarStoreStore nodes added in library_call.cpp, check 1887 // escape status of associated AllocateNode and optimize out 1888 // MemBarStoreStore node if the allocated object never escapes. 1889 while (storestore_worklist.length() != 0) { 1890 Node *n = storestore_worklist.pop(); 1891 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1892 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1893 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1894 if (not_global_escape(alloc)) { 1895 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1896 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1897 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1898 igvn->register_new_node_with_optimizer(mb); 1899 igvn->replace_node(storestore, mb); 1900 } 1901 } 1902 } 1903 1904 // Optimize objects compare. 1905 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1906 assert(OptimizePtrCompare, "sanity"); 1907 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1908 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1909 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1910 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1911 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1912 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1913 1914 // Check simple cases first. 1915 if (jobj1 != NULL) { 1916 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1917 if (jobj1 == jobj2) { 1918 // Comparing the same not escaping object. 1919 return _pcmp_eq; 1920 } 1921 Node* obj = jobj1->ideal_node(); 1922 // Comparing not escaping allocation. 1923 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1924 !ptn2->points_to(jobj1)) { 1925 return _pcmp_neq; // This includes nullness check. 1926 } 1927 } 1928 } 1929 if (jobj2 != NULL) { 1930 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1931 Node* obj = jobj2->ideal_node(); 1932 // Comparing not escaping allocation. 1933 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1934 !ptn1->points_to(jobj2)) { 1935 return _pcmp_neq; // This includes nullness check. 1936 } 1937 } 1938 } 1939 if (jobj1 != NULL && jobj1 != phantom_obj && 1940 jobj2 != NULL && jobj2 != phantom_obj && 1941 jobj1->ideal_node()->is_Con() && 1942 jobj2->ideal_node()->is_Con()) { 1943 // Klass or String constants compare. Need to be careful with 1944 // compressed pointers - compare types of ConN and ConP instead of nodes. 1945 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1946 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1947 if (t1->make_ptr() == t2->make_ptr()) { 1948 return _pcmp_eq; 1949 } else { 1950 return _pcmp_neq; 1951 } 1952 } 1953 if (ptn1->meet(ptn2)) { 1954 return NULL; // Sets are not disjoint 1955 } 1956 1957 // Sets are disjoint. 1958 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1959 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1960 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1961 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1962 if (set1_has_unknown_ptr && set2_has_null_ptr || 1963 set2_has_unknown_ptr && set1_has_null_ptr) { 1964 // Check nullness of unknown object. 1965 return NULL; 1966 } 1967 1968 // Disjointness by itself is not sufficient since 1969 // alias analysis is not complete for escaped objects. 1970 // Disjoint sets are definitely unrelated only when 1971 // at least one set has only not escaping allocations. 1972 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1973 if (ptn1->non_escaping_allocation()) { 1974 return _pcmp_neq; 1975 } 1976 } 1977 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 1978 if (ptn2->non_escaping_allocation()) { 1979 return _pcmp_neq; 1980 } 1981 } 1982 return NULL; 1983 } 1984 1985 // Connection Graph constuction functions. 1986 1987 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 1988 PointsToNode* ptadr = _nodes.at(n->_idx); 1989 if (ptadr != NULL) { 1990 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 1991 return; 1992 } 1993 Compile* C = _compile; 1994 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 1995 _nodes.at_put(n->_idx, ptadr); 1996 } 1997 1998 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 1999 PointsToNode* ptadr = _nodes.at(n->_idx); 2000 if (ptadr != NULL) { 2001 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2002 return; 2003 } 2004 Compile* C = _compile; 2005 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2006 _nodes.at_put(n->_idx, ptadr); 2007 } 2008 2009 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2010 PointsToNode* ptadr = _nodes.at(n->_idx); 2011 if (ptadr != NULL) { 2012 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2013 return; 2014 } 2015 bool unsafe = false; 2016 bool is_oop = is_oop_field(n, offset, &unsafe); 2017 if (unsafe) { 2018 es = PointsToNode::GlobalEscape; 2019 } 2020 Compile* C = _compile; 2021 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2022 _nodes.at_put(n->_idx, field); 2023 } 2024 2025 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2026 PointsToNode* src, PointsToNode* dst) { 2027 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2028 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2029 PointsToNode* ptadr = _nodes.at(n->_idx); 2030 if (ptadr != NULL) { 2031 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2032 return; 2033 } 2034 Compile* C = _compile; 2035 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2036 _nodes.at_put(n->_idx, ptadr); 2037 // Add edge from arraycopy node to source object. 2038 (void)add_edge(ptadr, src); 2039 src->set_arraycopy_src(); 2040 // Add edge from destination object to arraycopy node. 2041 (void)add_edge(dst, ptadr); 2042 dst->set_arraycopy_dst(); 2043 } 2044 2045 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2046 const Type* adr_type = n->as_AddP()->bottom_type(); 2047 BasicType bt = T_INT; 2048 if (offset == Type::OffsetBot) { 2049 // Check only oop fields. 2050 if (!adr_type->isa_aryptr() || 2051 (adr_type->isa_aryptr()->klass() == NULL) || 2052 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2053 // OffsetBot is used to reference array's element. Ignore first AddP. 2054 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2055 bt = T_OBJECT; 2056 } 2057 } 2058 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2059 if (adr_type->isa_instptr()) { 2060 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2061 if (field != NULL) { 2062 bt = field->layout_type(); 2063 } else { 2064 // Check for unsafe oop field access 2065 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN)) { 2066 bt = T_OBJECT; 2067 (*unsafe) = true; 2068 } 2069 } 2070 } else if (adr_type->isa_aryptr()) { 2071 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2072 // Ignore array length load. 2073 } else if (UseShenandoahGC && offset == -8) { 2074 // Shenandoah read barrier. 2075 bt = T_ARRAY; 2076 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2077 // Ignore first AddP. 2078 } else { 2079 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2080 bt = elemtype->array_element_basic_type(); 2081 } 2082 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2083 // Allocation initialization, ThreadLocal field access, unsafe access 2084 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN)) { 2085 bt = T_OBJECT; 2086 } 2087 } 2088 } 2089 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 2090 } 2091 2092 // Returns unique pointed java object or NULL. 2093 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2094 assert(!_collecting, "should not call when contructed graph"); 2095 // If the node was created after the escape computation we can't answer. 2096 uint idx = n->_idx; 2097 if (idx >= nodes_size()) { 2098 return NULL; 2099 } 2100 PointsToNode* ptn = ptnode_adr(idx); 2101 if (ptn->is_JavaObject()) { 2102 return ptn->as_JavaObject(); 2103 } 2104 assert(ptn->is_LocalVar(), "sanity"); 2105 // Check all java objects it points to. 2106 JavaObjectNode* jobj = NULL; 2107 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2108 PointsToNode* e = i.get(); 2109 if (e->is_JavaObject()) { 2110 if (jobj == NULL) { 2111 jobj = e->as_JavaObject(); 2112 } else if (jobj != e) { 2113 return NULL; 2114 } 2115 } 2116 } 2117 return jobj; 2118 } 2119 2120 // Return true if this node points only to non-escaping allocations. 2121 bool PointsToNode::non_escaping_allocation() { 2122 if (is_JavaObject()) { 2123 Node* n = ideal_node(); 2124 if (n->is_Allocate() || n->is_CallStaticJava()) { 2125 return (escape_state() == PointsToNode::NoEscape); 2126 } else { 2127 return false; 2128 } 2129 } 2130 assert(is_LocalVar(), "sanity"); 2131 // Check all java objects it points to. 2132 for (EdgeIterator i(this); i.has_next(); i.next()) { 2133 PointsToNode* e = i.get(); 2134 if (e->is_JavaObject()) { 2135 Node* n = e->ideal_node(); 2136 if ((e->escape_state() != PointsToNode::NoEscape) || 2137 !(n->is_Allocate() || n->is_CallStaticJava())) { 2138 return false; 2139 } 2140 } 2141 } 2142 return true; 2143 } 2144 2145 // Return true if we know the node does not escape globally. 2146 bool ConnectionGraph::not_global_escape(Node *n) { 2147 assert(!_collecting, "should not call during graph construction"); 2148 // If the node was created after the escape computation we can't answer. 2149 uint idx = n->_idx; 2150 if (idx >= nodes_size()) { 2151 return false; 2152 } 2153 PointsToNode* ptn = ptnode_adr(idx); 2154 PointsToNode::EscapeState es = ptn->escape_state(); 2155 // If we have already computed a value, return it. 2156 if (es >= PointsToNode::GlobalEscape) 2157 return false; 2158 if (ptn->is_JavaObject()) { 2159 return true; // (es < PointsToNode::GlobalEscape); 2160 } 2161 assert(ptn->is_LocalVar(), "sanity"); 2162 // Check all java objects it points to. 2163 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2164 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2165 return false; 2166 } 2167 return true; 2168 } 2169 2170 2171 // Helper functions 2172 2173 // Return true if this node points to specified node or nodes it points to. 2174 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2175 if (is_JavaObject()) { 2176 return (this == ptn); 2177 } 2178 assert(is_LocalVar() || is_Field(), "sanity"); 2179 for (EdgeIterator i(this); i.has_next(); i.next()) { 2180 if (i.get() == ptn) 2181 return true; 2182 } 2183 return false; 2184 } 2185 2186 // Return true if one node points to an other. 2187 bool PointsToNode::meet(PointsToNode* ptn) { 2188 if (this == ptn) { 2189 return true; 2190 } else if (ptn->is_JavaObject()) { 2191 return this->points_to(ptn->as_JavaObject()); 2192 } else if (this->is_JavaObject()) { 2193 return ptn->points_to(this->as_JavaObject()); 2194 } 2195 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2196 int ptn_count = ptn->edge_count(); 2197 for (EdgeIterator i(this); i.has_next(); i.next()) { 2198 PointsToNode* this_e = i.get(); 2199 for (int j = 0; j < ptn_count; j++) { 2200 if (this_e == ptn->edge(j)) 2201 return true; 2202 } 2203 } 2204 return false; 2205 } 2206 2207 #ifdef ASSERT 2208 // Return true if bases point to this java object. 2209 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2210 for (BaseIterator i(this); i.has_next(); i.next()) { 2211 if (i.get() == jobj) 2212 return true; 2213 } 2214 return false; 2215 } 2216 #endif 2217 2218 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2219 const Type *adr_type = phase->type(adr); 2220 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2221 adr->in(AddPNode::Address)->is_Proj() && 2222 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2223 // We are computing a raw address for a store captured by an Initialize 2224 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2225 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2226 assert(offs != Type::OffsetBot || 2227 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2228 "offset must be a constant or it is initialization of array"); 2229 return offs; 2230 } 2231 const TypePtr *t_ptr = adr_type->isa_ptr(); 2232 assert(t_ptr != NULL, "must be a pointer type"); 2233 return t_ptr->offset(); 2234 } 2235 2236 Node* ConnectionGraph::get_addp_base(Node *addp) { 2237 assert(addp->is_AddP(), "must be AddP"); 2238 // 2239 // AddP cases for Base and Address inputs: 2240 // case #1. Direct object's field reference: 2241 // Allocate 2242 // | 2243 // Proj #5 ( oop result ) 2244 // | 2245 // CheckCastPP (cast to instance type) 2246 // | | 2247 // AddP ( base == address ) 2248 // 2249 // case #2. Indirect object's field reference: 2250 // Phi 2251 // | 2252 // CastPP (cast to instance type) 2253 // | | 2254 // AddP ( base == address ) 2255 // 2256 // case #3. Raw object's field reference for Initialize node: 2257 // Allocate 2258 // | 2259 // Proj #5 ( oop result ) 2260 // top | 2261 // \ | 2262 // AddP ( base == top ) 2263 // 2264 // case #4. Array's element reference: 2265 // {CheckCastPP | CastPP} 2266 // | | | 2267 // | AddP ( array's element offset ) 2268 // | | 2269 // AddP ( array's offset ) 2270 // 2271 // case #5. Raw object's field reference for arraycopy stub call: 2272 // The inline_native_clone() case when the arraycopy stub is called 2273 // after the allocation before Initialize and CheckCastPP nodes. 2274 // Allocate 2275 // | 2276 // Proj #5 ( oop result ) 2277 // | | 2278 // AddP ( base == address ) 2279 // 2280 // case #6. Constant Pool, ThreadLocal, CastX2P or 2281 // Raw object's field reference: 2282 // {ConP, ThreadLocal, CastX2P, raw Load} 2283 // top | 2284 // \ | 2285 // AddP ( base == top ) 2286 // 2287 // case #7. Klass's field reference. 2288 // LoadKlass 2289 // | | 2290 // AddP ( base == address ) 2291 // 2292 // case #8. narrow Klass's field reference. 2293 // LoadNKlass 2294 // | 2295 // DecodeN 2296 // | | 2297 // AddP ( base == address ) 2298 // 2299 Node *base = addp->in(AddPNode::Base); 2300 if (base->uncast()->is_top()) { // The AddP case #3 and #6. 2301 base = addp->in(AddPNode::Address); 2302 while (base->is_AddP()) { 2303 // Case #6 (unsafe access) may have several chained AddP nodes. 2304 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2305 base = base->in(AddPNode::Address); 2306 } 2307 Node* uncast_base = base->uncast(); 2308 int opcode = uncast_base->Opcode(); 2309 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2310 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2311 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2312 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity"); 2313 } 2314 return base; 2315 } 2316 2317 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2318 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2319 Node* addp2 = addp->raw_out(0); 2320 if (addp->outcnt() == 1 && addp2->is_AddP() && 2321 addp2->in(AddPNode::Base) == n && 2322 addp2->in(AddPNode::Address) == addp) { 2323 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2324 // 2325 // Find array's offset to push it on worklist first and 2326 // as result process an array's element offset first (pushed second) 2327 // to avoid CastPP for the array's offset. 2328 // Otherwise the inserted CastPP (LocalVar) will point to what 2329 // the AddP (Field) points to. Which would be wrong since 2330 // the algorithm expects the CastPP has the same point as 2331 // as AddP's base CheckCastPP (LocalVar). 2332 // 2333 // ArrayAllocation 2334 // | 2335 // CheckCastPP 2336 // | 2337 // memProj (from ArrayAllocation CheckCastPP) 2338 // | || 2339 // | || Int (element index) 2340 // | || | ConI (log(element size)) 2341 // | || | / 2342 // | || LShift 2343 // | || / 2344 // | AddP (array's element offset) 2345 // | | 2346 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2347 // | / / 2348 // AddP (array's offset) 2349 // | 2350 // Load/Store (memory operation on array's element) 2351 // 2352 return addp2; 2353 } 2354 return NULL; 2355 } 2356 2357 // 2358 // Adjust the type and inputs of an AddP which computes the 2359 // address of a field of an instance 2360 // 2361 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2362 PhaseGVN* igvn = _igvn; 2363 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2364 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2365 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2366 if (t == NULL) { 2367 // We are computing a raw address for a store captured by an Initialize 2368 // compute an appropriate address type (cases #3 and #5). 2369 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2370 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2371 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2372 assert(offs != Type::OffsetBot, "offset must be a constant"); 2373 t = base_t->add_offset(offs)->is_oopptr(); 2374 } 2375 int inst_id = base_t->instance_id(); 2376 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2377 "old type must be non-instance or match new type"); 2378 2379 // The type 't' could be subclass of 'base_t'. 2380 // As result t->offset() could be large then base_t's size and it will 2381 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2382 // constructor verifies correctness of the offset. 2383 // 2384 // It could happened on subclass's branch (from the type profiling 2385 // inlining) which was not eliminated during parsing since the exactness 2386 // of the allocation type was not propagated to the subclass type check. 2387 // 2388 // Or the type 't' could be not related to 'base_t' at all. 2389 // It could happened when CHA type is different from MDO type on a dead path 2390 // (for example, from instanceof check) which is not collapsed during parsing. 2391 // 2392 // Do nothing for such AddP node and don't process its users since 2393 // this code branch will go away. 2394 // 2395 if (!t->is_known_instance() && 2396 !base_t->klass()->is_subtype_of(t->klass())) { 2397 return false; // bail out 2398 } 2399 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2400 // Do NOT remove the next line: ensure a new alias index is allocated 2401 // for the instance type. Note: C++ will not remove it since the call 2402 // has side effect. 2403 int alias_idx = _compile->get_alias_index(tinst); 2404 igvn->set_type(addp, tinst); 2405 // record the allocation in the node map 2406 set_map(addp, get_map(base->_idx)); 2407 // Set addp's Base and Address to 'base'. 2408 Node *abase = addp->in(AddPNode::Base); 2409 Node *adr = addp->in(AddPNode::Address); 2410 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2411 adr->in(0)->_idx == (uint)inst_id) { 2412 // Skip AddP cases #3 and #5. 2413 } else { 2414 assert(!abase->is_top(), "sanity"); // AddP case #3 2415 if (abase != base) { 2416 igvn->hash_delete(addp); 2417 addp->set_req(AddPNode::Base, base); 2418 if (abase == adr) { 2419 addp->set_req(AddPNode::Address, base); 2420 } else { 2421 // AddP case #4 (adr is array's element offset AddP node) 2422 #ifdef ASSERT 2423 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2424 assert(adr->is_AddP() && atype != NULL && 2425 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2426 #endif 2427 } 2428 igvn->hash_insert(addp); 2429 } 2430 } 2431 // Put on IGVN worklist since at least addp's type was changed above. 2432 record_for_optimizer(addp); 2433 return true; 2434 } 2435 2436 // 2437 // Create a new version of orig_phi if necessary. Returns either the newly 2438 // created phi or an existing phi. Sets create_new to indicate whether a new 2439 // phi was created. Cache the last newly created phi in the node map. 2440 // 2441 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2442 Compile *C = _compile; 2443 PhaseGVN* igvn = _igvn; 2444 new_created = false; 2445 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2446 // nothing to do if orig_phi is bottom memory or matches alias_idx 2447 if (phi_alias_idx == alias_idx) { 2448 return orig_phi; 2449 } 2450 // Have we recently created a Phi for this alias index? 2451 PhiNode *result = get_map_phi(orig_phi->_idx); 2452 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2453 return result; 2454 } 2455 // Previous check may fail when the same wide memory Phi was split into Phis 2456 // for different memory slices. Search all Phis for this region. 2457 if (result != NULL) { 2458 Node* region = orig_phi->in(0); 2459 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2460 Node* phi = region->fast_out(i); 2461 if (phi->is_Phi() && 2462 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2463 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2464 return phi->as_Phi(); 2465 } 2466 } 2467 } 2468 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2469 if (C->do_escape_analysis() == true && !C->failing()) { 2470 // Retry compilation without escape analysis. 2471 // If this is the first failure, the sentinel string will "stick" 2472 // to the Compile object, and the C2Compiler will see it and retry. 2473 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2474 } 2475 return NULL; 2476 } 2477 orig_phi_worklist.append_if_missing(orig_phi); 2478 const TypePtr *atype = C->get_adr_type(alias_idx); 2479 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2480 C->copy_node_notes_to(result, orig_phi); 2481 igvn->set_type(result, result->bottom_type()); 2482 record_for_optimizer(result); 2483 set_map(orig_phi, result); 2484 new_created = true; 2485 return result; 2486 } 2487 2488 // 2489 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2490 // specified alias index. 2491 // 2492 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2493 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2494 Compile *C = _compile; 2495 PhaseGVN* igvn = _igvn; 2496 bool new_phi_created; 2497 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2498 if (!new_phi_created) { 2499 return result; 2500 } 2501 GrowableArray<PhiNode *> phi_list; 2502 GrowableArray<uint> cur_input; 2503 PhiNode *phi = orig_phi; 2504 uint idx = 1; 2505 bool finished = false; 2506 while(!finished) { 2507 while (idx < phi->req()) { 2508 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2509 if (mem != NULL && mem->is_Phi()) { 2510 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2511 if (new_phi_created) { 2512 // found an phi for which we created a new split, push current one on worklist and begin 2513 // processing new one 2514 phi_list.push(phi); 2515 cur_input.push(idx); 2516 phi = mem->as_Phi(); 2517 result = newphi; 2518 idx = 1; 2519 continue; 2520 } else { 2521 mem = newphi; 2522 } 2523 } 2524 if (C->failing()) { 2525 return NULL; 2526 } 2527 result->set_req(idx++, mem); 2528 } 2529 #ifdef ASSERT 2530 // verify that the new Phi has an input for each input of the original 2531 assert( phi->req() == result->req(), "must have same number of inputs."); 2532 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2533 #endif 2534 // Check if all new phi's inputs have specified alias index. 2535 // Otherwise use old phi. 2536 for (uint i = 1; i < phi->req(); i++) { 2537 Node* in = result->in(i); 2538 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2539 } 2540 // we have finished processing a Phi, see if there are any more to do 2541 finished = (phi_list.length() == 0 ); 2542 if (!finished) { 2543 phi = phi_list.pop(); 2544 idx = cur_input.pop(); 2545 PhiNode *prev_result = get_map_phi(phi->_idx); 2546 prev_result->set_req(idx++, result); 2547 result = prev_result; 2548 } 2549 } 2550 return result; 2551 } 2552 2553 // 2554 // The next methods are derived from methods in MemNode. 2555 // 2556 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2557 Node *mem = mmem; 2558 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2559 // means an array I have not precisely typed yet. Do not do any 2560 // alias stuff with it any time soon. 2561 if (toop->base() != Type::AnyPtr && 2562 !(toop->klass() != NULL && 2563 toop->klass()->is_java_lang_Object() && 2564 toop->offset() == Type::OffsetBot)) { 2565 mem = mmem->memory_at(alias_idx); 2566 // Update input if it is progress over what we have now 2567 } 2568 return mem; 2569 } 2570 2571 // 2572 // Move memory users to their memory slices. 2573 // 2574 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2575 Compile* C = _compile; 2576 PhaseGVN* igvn = _igvn; 2577 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2578 assert(tp != NULL, "ptr type"); 2579 int alias_idx = C->get_alias_index(tp); 2580 int general_idx = C->get_general_index(alias_idx); 2581 2582 // Move users first 2583 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2584 Node* use = n->fast_out(i); 2585 if (use->is_MergeMem()) { 2586 MergeMemNode* mmem = use->as_MergeMem(); 2587 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2588 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2589 continue; // Nothing to do 2590 } 2591 // Replace previous general reference to mem node. 2592 uint orig_uniq = C->unique(); 2593 Node* m = find_inst_mem(n, general_idx, orig_phis); 2594 assert(orig_uniq == C->unique(), "no new nodes"); 2595 mmem->set_memory_at(general_idx, m); 2596 --imax; 2597 --i; 2598 } else if (use->is_MemBar()) { 2599 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2600 if (use->req() > MemBarNode::Precedent && 2601 use->in(MemBarNode::Precedent) == n) { 2602 // Don't move related membars. 2603 record_for_optimizer(use); 2604 continue; 2605 } 2606 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2607 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 2608 alias_idx == general_idx) { 2609 continue; // Nothing to do 2610 } 2611 // Move to general memory slice. 2612 uint orig_uniq = C->unique(); 2613 Node* m = find_inst_mem(n, general_idx, orig_phis); 2614 assert(orig_uniq == C->unique(), "no new nodes"); 2615 igvn->hash_delete(use); 2616 imax -= use->replace_edge(n, m); 2617 igvn->hash_insert(use); 2618 record_for_optimizer(use); 2619 --i; 2620 #ifdef ASSERT 2621 } else if (use->is_Mem()) { 2622 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2623 // Don't move related cardmark. 2624 continue; 2625 } 2626 // Memory nodes should have new memory input. 2627 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2628 assert(tp != NULL, "ptr type"); 2629 int idx = C->get_alias_index(tp); 2630 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2631 "Following memory nodes should have new memory input or be on the same memory slice"); 2632 } else if (use->is_Phi()) { 2633 // Phi nodes should be split and moved already. 2634 tp = use->as_Phi()->adr_type()->isa_ptr(); 2635 assert(tp != NULL, "ptr type"); 2636 int idx = C->get_alias_index(tp); 2637 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2638 } else { 2639 use->dump(); 2640 assert(false, "should not be here"); 2641 #endif 2642 } 2643 } 2644 } 2645 2646 // 2647 // Search memory chain of "mem" to find a MemNode whose address 2648 // is the specified alias index. 2649 // 2650 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2651 if (orig_mem == NULL) 2652 return orig_mem; 2653 Compile* C = _compile; 2654 PhaseGVN* igvn = _igvn; 2655 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2656 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2657 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 2658 Node *prev = NULL; 2659 Node *result = orig_mem; 2660 while (prev != result) { 2661 prev = result; 2662 if (result == start_mem) 2663 break; // hit one of our sentinels 2664 assert(result->Opcode() != Op_ShenandoahWBMemProj, "unexpected memory slice"); 2665 if (result->is_Mem()) { 2666 const Type *at = igvn->type(result->in(MemNode::Address)); 2667 if (at == Type::TOP) 2668 break; // Dead 2669 assert (at->isa_ptr() != NULL, "pointer type required."); 2670 int idx = C->get_alias_index(at->is_ptr()); 2671 if (idx == alias_idx) 2672 break; // Found 2673 if (!is_instance && (at->isa_oopptr() == NULL || 2674 !at->is_oopptr()->is_known_instance())) { 2675 break; // Do not skip store to general memory slice. 2676 } 2677 result = result->in(MemNode::Memory); 2678 } 2679 if (!is_instance) 2680 continue; // don't search further for non-instance types 2681 // skip over a call which does not affect this memory slice 2682 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2683 Node *proj_in = result->in(0); 2684 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2685 break; // hit one of our sentinels 2686 } else if (proj_in->is_Call()) { 2687 // ArrayCopy node processed here as well 2688 CallNode *call = proj_in->as_Call(); 2689 if (!call->may_modify(toop, igvn)) { 2690 result = call->in(TypeFunc::Memory); 2691 } 2692 } else if (proj_in->is_Initialize()) { 2693 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2694 // Stop if this is the initialization for the object instance which 2695 // which contains this memory slice, otherwise skip over it. 2696 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2697 result = proj_in->in(TypeFunc::Memory); 2698 } 2699 } else if (proj_in->is_MemBar()) { 2700 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() && 2701 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() && 2702 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) { 2703 // clone 2704 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy(); 2705 if (ac->may_modify(toop, igvn)) { 2706 break; 2707 } 2708 } 2709 result = proj_in->in(TypeFunc::Memory); 2710 } 2711 } else if (result->is_MergeMem()) { 2712 MergeMemNode *mmem = result->as_MergeMem(); 2713 result = step_through_mergemem(mmem, alias_idx, toop); 2714 if (result == mmem->base_memory()) { 2715 // Didn't find instance memory, search through general slice recursively. 2716 result = mmem->memory_at(C->get_general_index(alias_idx)); 2717 result = find_inst_mem(result, alias_idx, orig_phis); 2718 if (C->failing()) { 2719 return NULL; 2720 } 2721 mmem->set_memory_at(alias_idx, result); 2722 } 2723 } else if (result->is_Phi() && 2724 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2725 Node *un = result->as_Phi()->unique_input(igvn); 2726 if (un != NULL) { 2727 orig_phis.append_if_missing(result->as_Phi()); 2728 result = un; 2729 } else { 2730 break; 2731 } 2732 } else if (result->is_ClearArray()) { 2733 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2734 // Can not bypass initialization of the instance 2735 // we are looking for. 2736 break; 2737 } 2738 // Otherwise skip it (the call updated 'result' value). 2739 } else if (result->Opcode() == Op_SCMemProj) { 2740 Node* mem = result->in(0); 2741 Node* adr = NULL; 2742 if (mem->is_LoadStore()) { 2743 adr = mem->in(MemNode::Address); 2744 } else { 2745 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 2746 adr = mem->in(3); // Memory edge corresponds to destination array 2747 } 2748 const Type *at = igvn->type(adr); 2749 if (at != Type::TOP) { 2750 assert (at->isa_ptr() != NULL, "pointer type required."); 2751 int idx = C->get_alias_index(at->is_ptr()); 2752 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 2753 break; 2754 } 2755 result = mem->in(MemNode::Memory); 2756 } 2757 } 2758 if (result->is_Phi()) { 2759 PhiNode *mphi = result->as_Phi(); 2760 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2761 const TypePtr *t = mphi->adr_type(); 2762 if (!is_instance) { 2763 // Push all non-instance Phis on the orig_phis worklist to update inputs 2764 // during Phase 4 if needed. 2765 orig_phis.append_if_missing(mphi); 2766 } else if (C->get_alias_index(t) != alias_idx) { 2767 // Create a new Phi with the specified alias index type. 2768 result = split_memory_phi(mphi, alias_idx, orig_phis); 2769 } 2770 } 2771 // the result is either MemNode, PhiNode, InitializeNode. 2772 return result; 2773 } 2774 2775 // 2776 // Convert the types of unescaped object to instance types where possible, 2777 // propagate the new type information through the graph, and update memory 2778 // edges and MergeMem inputs to reflect the new type. 2779 // 2780 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2781 // The processing is done in 4 phases: 2782 // 2783 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2784 // types for the CheckCastPP for allocations where possible. 2785 // Propagate the new types through users as follows: 2786 // casts and Phi: push users on alloc_worklist 2787 // AddP: cast Base and Address inputs to the instance type 2788 // push any AddP users on alloc_worklist and push any memnode 2789 // users onto memnode_worklist. 2790 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2791 // search the Memory chain for a store with the appropriate type 2792 // address type. If a Phi is found, create a new version with 2793 // the appropriate memory slices from each of the Phi inputs. 2794 // For stores, process the users as follows: 2795 // MemNode: push on memnode_worklist 2796 // MergeMem: push on mergemem_worklist 2797 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2798 // moving the first node encountered of each instance type to the 2799 // the input corresponding to its alias index. 2800 // appropriate memory slice. 2801 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2802 // 2803 // In the following example, the CheckCastPP nodes are the cast of allocation 2804 // results and the allocation of node 29 is unescaped and eligible to be an 2805 // instance type. 2806 // 2807 // We start with: 2808 // 2809 // 7 Parm #memory 2810 // 10 ConI "12" 2811 // 19 CheckCastPP "Foo" 2812 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2813 // 29 CheckCastPP "Foo" 2814 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2815 // 2816 // 40 StoreP 25 7 20 ... alias_index=4 2817 // 50 StoreP 35 40 30 ... alias_index=4 2818 // 60 StoreP 45 50 20 ... alias_index=4 2819 // 70 LoadP _ 60 30 ... alias_index=4 2820 // 80 Phi 75 50 60 Memory alias_index=4 2821 // 90 LoadP _ 80 30 ... alias_index=4 2822 // 100 LoadP _ 80 20 ... alias_index=4 2823 // 2824 // 2825 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2826 // and creating a new alias index for node 30. This gives: 2827 // 2828 // 7 Parm #memory 2829 // 10 ConI "12" 2830 // 19 CheckCastPP "Foo" 2831 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2832 // 29 CheckCastPP "Foo" iid=24 2833 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2834 // 2835 // 40 StoreP 25 7 20 ... alias_index=4 2836 // 50 StoreP 35 40 30 ... alias_index=6 2837 // 60 StoreP 45 50 20 ... alias_index=4 2838 // 70 LoadP _ 60 30 ... alias_index=6 2839 // 80 Phi 75 50 60 Memory alias_index=4 2840 // 90 LoadP _ 80 30 ... alias_index=6 2841 // 100 LoadP _ 80 20 ... alias_index=4 2842 // 2843 // In phase 2, new memory inputs are computed for the loads and stores, 2844 // And a new version of the phi is created. In phase 4, the inputs to 2845 // node 80 are updated and then the memory nodes are updated with the 2846 // values computed in phase 2. This results in: 2847 // 2848 // 7 Parm #memory 2849 // 10 ConI "12" 2850 // 19 CheckCastPP "Foo" 2851 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2852 // 29 CheckCastPP "Foo" iid=24 2853 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2854 // 2855 // 40 StoreP 25 7 20 ... alias_index=4 2856 // 50 StoreP 35 7 30 ... alias_index=6 2857 // 60 StoreP 45 40 20 ... alias_index=4 2858 // 70 LoadP _ 50 30 ... alias_index=6 2859 // 80 Phi 75 40 60 Memory alias_index=4 2860 // 120 Phi 75 50 50 Memory alias_index=6 2861 // 90 LoadP _ 120 30 ... alias_index=6 2862 // 100 LoadP _ 80 20 ... alias_index=4 2863 // 2864 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 2865 GrowableArray<Node *> memnode_worklist; 2866 GrowableArray<PhiNode *> orig_phis; 2867 PhaseIterGVN *igvn = _igvn; 2868 uint new_index_start = (uint) _compile->num_alias_types(); 2869 Arena* arena = Thread::current()->resource_area(); 2870 VectorSet visited(arena); 2871 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2872 uint unique_old = _compile->unique(); 2873 2874 // Phase 1: Process possible allocations from alloc_worklist. 2875 // Create instance types for the CheckCastPP for allocations where possible. 2876 // 2877 // (Note: don't forget to change the order of the second AddP node on 2878 // the alloc_worklist if the order of the worklist processing is changed, 2879 // see the comment in find_second_addp().) 2880 // 2881 while (alloc_worklist.length() != 0) { 2882 Node *n = alloc_worklist.pop(); 2883 uint ni = n->_idx; 2884 if (n->is_Call()) { 2885 CallNode *alloc = n->as_Call(); 2886 // copy escape information to call node 2887 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2888 PointsToNode::EscapeState es = ptn->escape_state(); 2889 // We have an allocation or call which returns a Java object, 2890 // see if it is unescaped. 2891 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2892 continue; 2893 // Find CheckCastPP for the allocate or for the return value of a call 2894 n = alloc->result_cast(); 2895 if (n == NULL) { // No uses except Initialize node 2896 if (alloc->is_Allocate()) { 2897 // Set the scalar_replaceable flag for allocation 2898 // so it could be eliminated if it has no uses. 2899 alloc->as_Allocate()->_is_scalar_replaceable = true; 2900 } 2901 if (alloc->is_CallStaticJava()) { 2902 // Set the scalar_replaceable flag for boxing method 2903 // so it could be eliminated if it has no uses. 2904 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2905 } 2906 continue; 2907 } 2908 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2909 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2910 continue; 2911 } 2912 2913 // The inline code for Object.clone() casts the allocation result to 2914 // java.lang.Object and then to the actual type of the allocated 2915 // object. Detect this case and use the second cast. 2916 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2917 // the allocation result is cast to java.lang.Object and then 2918 // to the actual Array type. 2919 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2920 && (alloc->is_AllocateArray() || 2921 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2922 Node *cast2 = NULL; 2923 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2924 Node *use = n->fast_out(i); 2925 if (use->is_CheckCastPP()) { 2926 cast2 = use; 2927 break; 2928 } 2929 } 2930 if (cast2 != NULL) { 2931 n = cast2; 2932 } else { 2933 // Non-scalar replaceable if the allocation type is unknown statically 2934 // (reflection allocation), the object can't be restored during 2935 // deoptimization without precise type. 2936 continue; 2937 } 2938 } 2939 2940 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 2941 if (t == NULL) 2942 continue; // not a TypeOopPtr 2943 if (!t->klass_is_exact()) 2944 continue; // not an unique type 2945 2946 if (alloc->is_Allocate()) { 2947 // Set the scalar_replaceable flag for allocation 2948 // so it could be eliminated. 2949 alloc->as_Allocate()->_is_scalar_replaceable = true; 2950 } 2951 if (alloc->is_CallStaticJava()) { 2952 // Set the scalar_replaceable flag for boxing method 2953 // so it could be eliminated. 2954 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2955 } 2956 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 2957 // in order for an object to be scalar-replaceable, it must be: 2958 // - a direct allocation (not a call returning an object) 2959 // - non-escaping 2960 // - eligible to be a unique type 2961 // - not determined to be ineligible by escape analysis 2962 set_map(alloc, n); 2963 set_map(n, alloc); 2964 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 2965 igvn->hash_delete(n); 2966 igvn->set_type(n, tinst); 2967 n->raise_bottom_type(tinst); 2968 igvn->hash_insert(n); 2969 record_for_optimizer(n); 2970 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 2971 2972 // First, put on the worklist all Field edges from Connection Graph 2973 // which is more accurate than putting immediate users from Ideal Graph. 2974 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 2975 PointsToNode* tgt = e.get(); 2976 if (tgt->is_Arraycopy()) { 2977 continue; 2978 } 2979 Node* use = tgt->ideal_node(); 2980 assert(tgt->is_Field() && use->is_AddP(), 2981 "only AddP nodes are Field edges in CG"); 2982 if (use->outcnt() > 0) { // Don't process dead nodes 2983 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 2984 if (addp2 != NULL) { 2985 assert(alloc->is_AllocateArray(),"array allocation was expected"); 2986 alloc_worklist.append_if_missing(addp2); 2987 } 2988 alloc_worklist.append_if_missing(use); 2989 } 2990 } 2991 2992 // An allocation may have an Initialize which has raw stores. Scan 2993 // the users of the raw allocation result and push AddP users 2994 // on alloc_worklist. 2995 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 2996 assert (raw_result != NULL, "must have an allocation result"); 2997 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 2998 Node *use = raw_result->fast_out(i); 2999 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3000 Node* addp2 = find_second_addp(use, raw_result); 3001 if (addp2 != NULL) { 3002 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3003 alloc_worklist.append_if_missing(addp2); 3004 } 3005 alloc_worklist.append_if_missing(use); 3006 } else if (use->is_MemBar()) { 3007 memnode_worklist.append_if_missing(use); 3008 } 3009 } 3010 } 3011 } else if (n->is_AddP()) { 3012 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3013 if (jobj == NULL || jobj == phantom_obj) { 3014 #ifdef ASSERT 3015 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3016 ptnode_adr(n->_idx)->dump(); 3017 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3018 #endif 3019 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3020 return; 3021 } 3022 Node *base = get_map(jobj->idx()); // CheckCastPP node 3023 if (!split_AddP(n, base)) continue; // wrong type from dead path 3024 } else if (n->is_Phi() || 3025 n->is_CheckCastPP() || 3026 n->is_EncodeP() || 3027 n->is_DecodeN() || 3028 n->is_ShenandoahBarrier() || 3029 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3030 if (visited.test_set(n->_idx)) { 3031 assert(n->is_Phi(), "loops only through Phi's"); 3032 continue; // already processed 3033 } 3034 JavaObjectNode* jobj = unique_java_object(n); 3035 if (jobj == NULL || jobj == phantom_obj) { 3036 #ifdef ASSERT 3037 ptnode_adr(n->_idx)->dump(); 3038 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3039 #endif 3040 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3041 return; 3042 } else { 3043 Node *val = get_map(jobj->idx()); // CheckCastPP node 3044 TypeNode *tn = n->as_Type(); 3045 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3046 assert(tinst != NULL && tinst->is_known_instance() && 3047 tinst->instance_id() == jobj->idx() , "instance type expected."); 3048 3049 const Type *tn_type = igvn->type(tn); 3050 const TypeOopPtr *tn_t; 3051 if (tn_type->isa_narrowoop()) { 3052 tn_t = tn_type->make_ptr()->isa_oopptr(); 3053 } else { 3054 tn_t = tn_type->isa_oopptr(); 3055 } 3056 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3057 if (tn_type->isa_narrowoop()) { 3058 tn_type = tinst->make_narrowoop(); 3059 } else { 3060 tn_type = tinst; 3061 } 3062 igvn->hash_delete(tn); 3063 igvn->set_type(tn, tn_type); 3064 tn->set_type(tn_type); 3065 igvn->hash_insert(tn); 3066 record_for_optimizer(n); 3067 } else { 3068 assert(tn_type == TypePtr::NULL_PTR || 3069 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3070 "unexpected type"); 3071 continue; // Skip dead path with different type 3072 } 3073 } 3074 } else { 3075 debug_only(n->dump();) 3076 assert(false, "EA: unexpected node"); 3077 continue; 3078 } 3079 // push allocation's users on appropriate worklist 3080 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3081 Node *use = n->fast_out(i); 3082 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3083 // Load/store to instance's field 3084 memnode_worklist.append_if_missing(use); 3085 } else if (use->is_MemBar()) { 3086 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3087 memnode_worklist.append_if_missing(use); 3088 } 3089 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3090 Node* addp2 = find_second_addp(use, n); 3091 if (addp2 != NULL) { 3092 alloc_worklist.append_if_missing(addp2); 3093 } 3094 alloc_worklist.append_if_missing(use); 3095 } else if (use->is_Phi() || 3096 use->is_CheckCastPP() || 3097 use->is_EncodeNarrowPtr() || 3098 use->is_DecodeNarrowPtr() || 3099 use->is_ShenandoahBarrier() || 3100 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3101 alloc_worklist.append_if_missing(use); 3102 #ifdef ASSERT 3103 } else if (use->is_Mem()) { 3104 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3105 } else if (use->is_MergeMem()) { 3106 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3107 } else if (use->is_SafePoint()) { 3108 // Look for MergeMem nodes for calls which reference unique allocation 3109 // (through CheckCastPP nodes) even for debug info. 3110 Node* m = use->in(TypeFunc::Memory); 3111 if (m->is_MergeMem()) { 3112 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3113 } 3114 } else if (use->Opcode() == Op_EncodeISOArray) { 3115 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3116 // EncodeISOArray overwrites destination array 3117 memnode_worklist.append_if_missing(use); 3118 } 3119 } else { 3120 uint op = use->Opcode(); 3121 if (!(op == Op_CmpP || op == Op_Conv2B || 3122 op == Op_CastP2X || op == Op_StoreCM || 3123 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 3124 op == Op_StrEquals || op == Op_StrIndexOf || 3125 op == Op_ShenandoahWBMemProj)) { 3126 n->dump(); 3127 use->dump(); 3128 assert(false, "EA: missing allocation reference path"); 3129 } 3130 #endif 3131 } 3132 } 3133 3134 } 3135 3136 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3137 // type, record it in the ArrayCopy node so we know what memory this 3138 // node uses/modified. 3139 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3140 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3141 Node* dest = ac->in(ArrayCopyNode::Dest); 3142 if (dest->is_AddP()) { 3143 dest = get_addp_base(dest); 3144 } 3145 JavaObjectNode* jobj = unique_java_object(dest); 3146 if (jobj != NULL) { 3147 Node *base = get_map(jobj->idx()); 3148 if (base != NULL) { 3149 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3150 ac->_dest_type = base_t; 3151 } 3152 } 3153 Node* src = ac->in(ArrayCopyNode::Src); 3154 if (src->is_AddP()) { 3155 src = get_addp_base(src); 3156 } 3157 jobj = unique_java_object(src); 3158 if (jobj != NULL) { 3159 Node* base = get_map(jobj->idx()); 3160 if (base != NULL) { 3161 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3162 ac->_src_type = base_t; 3163 } 3164 } 3165 } 3166 3167 // New alias types were created in split_AddP(). 3168 uint new_index_end = (uint) _compile->num_alias_types(); 3169 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3170 3171 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3172 // compute new values for Memory inputs (the Memory inputs are not 3173 // actually updated until phase 4.) 3174 if (memnode_worklist.length() == 0) 3175 return; // nothing to do 3176 while (memnode_worklist.length() != 0) { 3177 Node *n = memnode_worklist.pop(); 3178 if (visited.test_set(n->_idx)) 3179 continue; 3180 if (n->is_Phi() || n->is_ClearArray()) { 3181 // we don't need to do anything, but the users must be pushed 3182 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3183 // we don't need to do anything, but the users must be pushed 3184 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 3185 if (n == NULL) 3186 continue; 3187 } else if (n->Opcode() == Op_EncodeISOArray) { 3188 // get the memory projection 3189 n = n->find_out_with(Op_SCMemProj); 3190 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3191 } else { 3192 assert(n->is_Mem(), "memory node required."); 3193 Node *addr = n->in(MemNode::Address); 3194 const Type *addr_t = igvn->type(addr); 3195 if (addr_t == Type::TOP) 3196 continue; 3197 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3198 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3199 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3200 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3201 if (_compile->failing()) { 3202 return; 3203 } 3204 if (mem != n->in(MemNode::Memory)) { 3205 // We delay the memory edge update since we need old one in 3206 // MergeMem code below when instances memory slices are separated. 3207 set_map(n, mem); 3208 } 3209 if (n->is_Load()) { 3210 continue; // don't push users 3211 } else if (n->is_LoadStore()) { 3212 // get the memory projection 3213 n = n->find_out_with(Op_SCMemProj); 3214 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 3215 } 3216 } 3217 // push user on appropriate worklist 3218 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3219 Node *use = n->fast_out(i); 3220 if (use->is_Phi() || use->is_ClearArray()) { 3221 memnode_worklist.append_if_missing(use); 3222 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3223 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3224 continue; 3225 memnode_worklist.append_if_missing(use); 3226 } else if (use->is_MemBar()) { 3227 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3228 memnode_worklist.append_if_missing(use); 3229 } 3230 #ifdef ASSERT 3231 } else if(use->is_Mem()) { 3232 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3233 } else if (use->is_MergeMem()) { 3234 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3235 } else if (use->Opcode() == Op_EncodeISOArray) { 3236 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3237 // EncodeISOArray overwrites destination array 3238 memnode_worklist.append_if_missing(use); 3239 } 3240 } else { 3241 uint op = use->Opcode(); 3242 if (!(op == Op_StoreCM || 3243 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 3244 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 3245 op == Op_AryEq || op == Op_StrComp || 3246 op == Op_StrEquals || op == Op_StrIndexOf)) { 3247 n->dump(); 3248 use->dump(); 3249 assert(false, "EA: missing memory path"); 3250 } 3251 #endif 3252 } 3253 } 3254 } 3255 3256 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3257 // Walk each memory slice moving the first node encountered of each 3258 // instance type to the the input corresponding to its alias index. 3259 uint length = _mergemem_worklist.length(); 3260 for( uint next = 0; next < length; ++next ) { 3261 MergeMemNode* nmm = _mergemem_worklist.at(next); 3262 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3263 // Note: we don't want to use MergeMemStream here because we only want to 3264 // scan inputs which exist at the start, not ones we add during processing. 3265 // Note 2: MergeMem may already contains instance memory slices added 3266 // during find_inst_mem() call when memory nodes were processed above. 3267 igvn->hash_delete(nmm); 3268 uint nslices = MIN2(nmm->req(), new_index_start); 3269 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3270 Node* mem = nmm->in(i); 3271 Node* cur = NULL; 3272 if (mem == NULL || mem->is_top()) 3273 continue; 3274 // First, update mergemem by moving memory nodes to corresponding slices 3275 // if their type became more precise since this mergemem was created. 3276 while (mem->is_Mem()) { 3277 const Type *at = igvn->type(mem->in(MemNode::Address)); 3278 if (at != Type::TOP) { 3279 assert (at->isa_ptr() != NULL, "pointer type required."); 3280 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3281 if (idx == i) { 3282 if (cur == NULL) 3283 cur = mem; 3284 } else { 3285 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3286 nmm->set_memory_at(idx, mem); 3287 } 3288 } 3289 } 3290 mem = mem->in(MemNode::Memory); 3291 } 3292 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3293 // Find any instance of the current type if we haven't encountered 3294 // already a memory slice of the instance along the memory chain. 3295 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3296 if((uint)_compile->get_general_index(ni) == i) { 3297 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3298 if (nmm->is_empty_memory(m)) { 3299 Node* result = find_inst_mem(mem, ni, orig_phis); 3300 if (_compile->failing()) { 3301 return; 3302 } 3303 nmm->set_memory_at(ni, result); 3304 } 3305 } 3306 } 3307 } 3308 // Find the rest of instances values 3309 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3310 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3311 Node* result = step_through_mergemem(nmm, ni, tinst); 3312 if (result == nmm->base_memory()) { 3313 // Didn't find instance memory, search through general slice recursively. 3314 result = nmm->memory_at(_compile->get_general_index(ni)); 3315 result = find_inst_mem(result, ni, orig_phis); 3316 if (_compile->failing()) { 3317 return; 3318 } 3319 nmm->set_memory_at(ni, result); 3320 } 3321 } 3322 igvn->hash_insert(nmm); 3323 record_for_optimizer(nmm); 3324 } 3325 3326 // Phase 4: Update the inputs of non-instance memory Phis and 3327 // the Memory input of memnodes 3328 // First update the inputs of any non-instance Phi's from 3329 // which we split out an instance Phi. Note we don't have 3330 // to recursively process Phi's encounted on the input memory 3331 // chains as is done in split_memory_phi() since they will 3332 // also be processed here. 3333 for (int j = 0; j < orig_phis.length(); j++) { 3334 PhiNode *phi = orig_phis.at(j); 3335 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3336 igvn->hash_delete(phi); 3337 for (uint i = 1; i < phi->req(); i++) { 3338 Node *mem = phi->in(i); 3339 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3340 if (_compile->failing()) { 3341 return; 3342 } 3343 if (mem != new_mem) { 3344 phi->set_req(i, new_mem); 3345 } 3346 } 3347 igvn->hash_insert(phi); 3348 record_for_optimizer(phi); 3349 } 3350 3351 // Update the memory inputs of MemNodes with the value we computed 3352 // in Phase 2 and move stores memory users to corresponding memory slices. 3353 // Disable memory split verification code until the fix for 6984348. 3354 // Currently it produces false negative results since it does not cover all cases. 3355 #if 0 // ifdef ASSERT 3356 visited.Reset(); 3357 Node_Stack old_mems(arena, _compile->unique() >> 2); 3358 #endif 3359 for (uint i = 0; i < ideal_nodes.size(); i++) { 3360 Node* n = ideal_nodes.at(i); 3361 Node* nmem = get_map(n->_idx); 3362 assert(nmem != NULL, "sanity"); 3363 if (n->is_Mem()) { 3364 #if 0 // ifdef ASSERT 3365 Node* old_mem = n->in(MemNode::Memory); 3366 if (!visited.test_set(old_mem->_idx)) { 3367 old_mems.push(old_mem, old_mem->outcnt()); 3368 } 3369 #endif 3370 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3371 if (!n->is_Load()) { 3372 // Move memory users of a store first. 3373 move_inst_mem(n, orig_phis); 3374 } 3375 // Now update memory input 3376 igvn->hash_delete(n); 3377 n->set_req(MemNode::Memory, nmem); 3378 igvn->hash_insert(n); 3379 record_for_optimizer(n); 3380 } else { 3381 assert(n->is_Allocate() || n->is_CheckCastPP() || 3382 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3383 } 3384 } 3385 #if 0 // ifdef ASSERT 3386 // Verify that memory was split correctly 3387 while (old_mems.is_nonempty()) { 3388 Node* old_mem = old_mems.node(); 3389 uint old_cnt = old_mems.index(); 3390 old_mems.pop(); 3391 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3392 } 3393 #endif 3394 } 3395 3396 #ifndef PRODUCT 3397 static const char *node_type_names[] = { 3398 "UnknownType", 3399 "JavaObject", 3400 "LocalVar", 3401 "Field", 3402 "Arraycopy" 3403 }; 3404 3405 static const char *esc_names[] = { 3406 "UnknownEscape", 3407 "NoEscape", 3408 "ArgEscape", 3409 "GlobalEscape" 3410 }; 3411 3412 void PointsToNode::dump(bool print_state) const { 3413 NodeType nt = node_type(); 3414 tty->print("%s ", node_type_names[(int) nt]); 3415 if (print_state) { 3416 EscapeState es = escape_state(); 3417 EscapeState fields_es = fields_escape_state(); 3418 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3419 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3420 tty->print("NSR "); 3421 } 3422 if (is_Field()) { 3423 FieldNode* f = (FieldNode*)this; 3424 if (f->is_oop()) 3425 tty->print("oop "); 3426 if (f->offset() > 0) 3427 tty->print("+%d ", f->offset()); 3428 tty->print("("); 3429 for (BaseIterator i(f); i.has_next(); i.next()) { 3430 PointsToNode* b = i.get(); 3431 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3432 } 3433 tty->print(" )"); 3434 } 3435 tty->print("["); 3436 for (EdgeIterator i(this); i.has_next(); i.next()) { 3437 PointsToNode* e = i.get(); 3438 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3439 } 3440 tty->print(" ["); 3441 for (UseIterator i(this); i.has_next(); i.next()) { 3442 PointsToNode* u = i.get(); 3443 bool is_base = false; 3444 if (PointsToNode::is_base_use(u)) { 3445 is_base = true; 3446 u = PointsToNode::get_use_node(u)->as_Field(); 3447 } 3448 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3449 } 3450 tty->print(" ]] "); 3451 if (_node == NULL) 3452 tty->print_cr("<null>"); 3453 else 3454 _node->dump(); 3455 } 3456 3457 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3458 bool first = true; 3459 int ptnodes_length = ptnodes_worklist.length(); 3460 for (int i = 0; i < ptnodes_length; i++) { 3461 PointsToNode *ptn = ptnodes_worklist.at(i); 3462 if (ptn == NULL || !ptn->is_JavaObject()) 3463 continue; 3464 PointsToNode::EscapeState es = ptn->escape_state(); 3465 if ((es != PointsToNode::NoEscape) && !Verbose) { 3466 continue; 3467 } 3468 Node* n = ptn->ideal_node(); 3469 if (n->is_Allocate() || (n->is_CallStaticJava() && 3470 n->as_CallStaticJava()->is_boxing_method())) { 3471 if (first) { 3472 tty->cr(); 3473 tty->print("======== Connection graph for "); 3474 _compile->method()->print_short_name(); 3475 tty->cr(); 3476 first = false; 3477 } 3478 ptn->dump(); 3479 // Print all locals and fields which reference this allocation 3480 for (UseIterator j(ptn); j.has_next(); j.next()) { 3481 PointsToNode* use = j.get(); 3482 if (use->is_LocalVar()) { 3483 use->dump(Verbose); 3484 } else if (Verbose) { 3485 use->dump(); 3486 } 3487 } 3488 tty->cr(); 3489 } 3490 } 3491 } 3492 #endif