1 /* 2 * Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/bcEscapeAnalyzer.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "libadt/vectset.hpp" 31 #include "memory/allocation.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/c2compiler.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/callnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/compile.hpp" 38 #include "opto/escape.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/movenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "utilities/macros.hpp" 43 44 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 45 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL), 46 _in_worklist(C->comp_arena()), 47 _next_pidx(0), 48 _collecting(true), 49 _verify(false), 50 _compile(C), 51 _igvn(igvn), 52 _node_map(C->comp_arena()) { 53 // Add unknown java object. 54 add_java_object(C->top(), PointsToNode::GlobalEscape); 55 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 56 // Add ConP(#NULL) and ConN(#NULL) nodes. 57 Node* oop_null = igvn->zerocon(T_OBJECT); 58 assert(oop_null->_idx < nodes_size(), "should be created already"); 59 add_java_object(oop_null, PointsToNode::NoEscape); 60 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 61 if (UseCompressedOops) { 62 Node* noop_null = igvn->zerocon(T_NARROWOOP); 63 assert(noop_null->_idx < nodes_size(), "should be created already"); 64 map_ideal_node(noop_null, null_obj); 65 } 66 _pcmp_neq = NULL; // Should be initialized 67 _pcmp_eq = NULL; 68 } 69 70 bool ConnectionGraph::has_candidates(Compile *C) { 71 // EA brings benefits only when the code has allocations and/or locks which 72 // are represented by ideal Macro nodes. 73 int cnt = C->macro_count(); 74 for (int i = 0; i < cnt; i++) { 75 Node *n = C->macro_node(i); 76 if (n->is_Allocate()) 77 return true; 78 if (n->is_Lock()) { 79 Node* obj = n->as_Lock()->obj_node()->uncast(); 80 if (!(obj->is_Parm() || obj->is_Con())) 81 return true; 82 } 83 if (n->is_CallStaticJava() && 84 n->as_CallStaticJava()->is_boxing_method()) { 85 return true; 86 } 87 } 88 return false; 89 } 90 91 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 92 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]); 93 ResourceMark rm; 94 95 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 96 // to create space for them in ConnectionGraph::_nodes[]. 97 Node* oop_null = igvn->zerocon(T_OBJECT); 98 Node* noop_null = igvn->zerocon(T_NARROWOOP); 99 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 100 // Perform escape analysis 101 if (congraph->compute_escape()) { 102 // There are non escaping objects. 103 C->set_congraph(congraph); 104 } 105 // Cleanup. 106 if (oop_null->outcnt() == 0) 107 igvn->hash_delete(oop_null); 108 if (noop_null->outcnt() == 0) 109 igvn->hash_delete(noop_null); 110 } 111 112 bool ConnectionGraph::compute_escape() { 113 Compile* C = _compile; 114 PhaseGVN* igvn = _igvn; 115 116 // Worklists used by EA. 117 Unique_Node_List delayed_worklist; 118 GrowableArray<Node*> alloc_worklist; 119 GrowableArray<Node*> ptr_cmp_worklist; 120 GrowableArray<Node*> storestore_worklist; 121 GrowableArray<ArrayCopyNode*> arraycopy_worklist; 122 GrowableArray<PointsToNode*> ptnodes_worklist; 123 GrowableArray<JavaObjectNode*> java_objects_worklist; 124 GrowableArray<JavaObjectNode*> non_escaped_worklist; 125 GrowableArray<FieldNode*> oop_fields_worklist; 126 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 127 128 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]); 129 130 // 1. Populate Connection Graph (CG) with PointsTo nodes. 131 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 132 // Initialize worklist 133 if (C->root() != NULL) { 134 ideal_nodes.push(C->root()); 135 } 136 // Processed ideal nodes are unique on ideal_nodes list 137 // but several ideal nodes are mapped to the phantom_obj. 138 // To avoid duplicated entries on the following worklists 139 // add the phantom_obj only once to them. 140 ptnodes_worklist.append(phantom_obj); 141 java_objects_worklist.append(phantom_obj); 142 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 143 Node* n = ideal_nodes.at(next); 144 // Create PointsTo nodes and add them to Connection Graph. Called 145 // only once per ideal node since ideal_nodes is Unique_Node list. 146 add_node_to_connection_graph(n, &delayed_worklist); 147 PointsToNode* ptn = ptnode_adr(n->_idx); 148 if (ptn != NULL && ptn != phantom_obj) { 149 ptnodes_worklist.append(ptn); 150 if (ptn->is_JavaObject()) { 151 java_objects_worklist.append(ptn->as_JavaObject()); 152 if ((n->is_Allocate() || n->is_CallStaticJava()) && 153 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 154 // Only allocations and java static calls results are interesting. 155 non_escaped_worklist.append(ptn->as_JavaObject()); 156 } 157 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 158 oop_fields_worklist.append(ptn->as_Field()); 159 } 160 } 161 if (n->is_MergeMem()) { 162 // Collect all MergeMem nodes to add memory slices for 163 // scalar replaceable objects in split_unique_types(). 164 _mergemem_worklist.append(n->as_MergeMem()); 165 } else if (OptimizePtrCompare && n->is_Cmp() && 166 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) { 167 // Collect compare pointers nodes. 168 ptr_cmp_worklist.append(n); 169 } else if (n->is_MemBarStoreStore()) { 170 // Collect all MemBarStoreStore nodes so that depending on the 171 // escape status of the associated Allocate node some of them 172 // may be eliminated. 173 storestore_worklist.append(n); 174 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 175 (n->req() > MemBarNode::Precedent)) { 176 record_for_optimizer(n); 177 #ifdef ASSERT 178 } else if (n->is_AddP()) { 179 // Collect address nodes for graph verification. 180 addp_worklist.append(n); 181 #endif 182 } else if (n->is_ArrayCopy()) { 183 // Keep a list of ArrayCopy nodes so if one of its input is non 184 // escaping, we can record a unique type 185 arraycopy_worklist.append(n->as_ArrayCopy()); 186 } 187 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 188 Node* m = n->fast_out(i); // Get user 189 ideal_nodes.push(m); 190 } 191 } 192 if (non_escaped_worklist.length() == 0) { 193 _collecting = false; 194 return false; // Nothing to do. 195 } 196 // Add final simple edges to graph. 197 while(delayed_worklist.size() > 0) { 198 Node* n = delayed_worklist.pop(); 199 add_final_edges(n); 200 } 201 int ptnodes_length = ptnodes_worklist.length(); 202 203 #ifdef ASSERT 204 if (VerifyConnectionGraph) { 205 // Verify that no new simple edges could be created and all 206 // local vars has edges. 207 _verify = true; 208 for (int next = 0; next < ptnodes_length; ++next) { 209 PointsToNode* ptn = ptnodes_worklist.at(next); 210 add_final_edges(ptn->ideal_node()); 211 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 212 ptn->dump(); 213 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 214 } 215 } 216 _verify = false; 217 } 218 #endif 219 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes 220 // processing, calls to CI to resolve symbols (types, fields, methods) 221 // referenced in bytecode. During symbol resolution VM may throw 222 // an exception which CI cleans and converts to compilation failure. 223 if (C->failing()) return false; 224 225 // 2. Finish Graph construction by propagating references to all 226 // java objects through graph. 227 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 228 java_objects_worklist, oop_fields_worklist)) { 229 // All objects escaped or hit time or iterations limits. 230 _collecting = false; 231 return false; 232 } 233 234 // 3. Adjust scalar_replaceable state of nonescaping objects and push 235 // scalar replaceable allocations on alloc_worklist for processing 236 // in split_unique_types(). 237 int non_escaped_length = non_escaped_worklist.length(); 238 for (int next = 0; next < non_escaped_length; next++) { 239 JavaObjectNode* ptn = non_escaped_worklist.at(next); 240 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 241 Node* n = ptn->ideal_node(); 242 if (n->is_Allocate()) { 243 n->as_Allocate()->_is_non_escaping = noescape; 244 } 245 if (n->is_CallStaticJava()) { 246 n->as_CallStaticJava()->_is_non_escaping = noescape; 247 } 248 if (noescape && ptn->scalar_replaceable()) { 249 adjust_scalar_replaceable_state(ptn); 250 if (ptn->scalar_replaceable()) { 251 alloc_worklist.append(ptn->ideal_node()); 252 } 253 } 254 } 255 256 #ifdef ASSERT 257 if (VerifyConnectionGraph) { 258 // Verify that graph is complete - no new edges could be added or needed. 259 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 260 java_objects_worklist, addp_worklist); 261 } 262 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 263 assert(null_obj->escape_state() == PointsToNode::NoEscape && 264 null_obj->edge_count() == 0 && 265 !null_obj->arraycopy_src() && 266 !null_obj->arraycopy_dst(), "sanity"); 267 #endif 268 269 _collecting = false; 270 271 } // TracePhase t3("connectionGraph") 272 273 // 4. Optimize ideal graph based on EA information. 274 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 275 if (has_non_escaping_obj) { 276 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 277 } 278 279 #ifndef PRODUCT 280 if (PrintEscapeAnalysis) { 281 dump(ptnodes_worklist); // Dump ConnectionGraph 282 } 283 #endif 284 285 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 286 #ifdef ASSERT 287 if (VerifyConnectionGraph) { 288 int alloc_length = alloc_worklist.length(); 289 for (int next = 0; next < alloc_length; ++next) { 290 Node* n = alloc_worklist.at(next); 291 PointsToNode* ptn = ptnode_adr(n->_idx); 292 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 293 } 294 } 295 #endif 296 297 // 5. Separate memory graph for scalar replaceable allcations. 298 if (has_scalar_replaceable_candidates && 299 C->AliasLevel() >= 3 && EliminateAllocations) { 300 // Now use the escape information to create unique types for 301 // scalar replaceable objects. 302 split_unique_types(alloc_worklist, arraycopy_worklist); 303 if (C->failing()) return false; 304 C->print_method(PHASE_AFTER_EA, 2); 305 306 #ifdef ASSERT 307 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 308 tty->print("=== No allocations eliminated for "); 309 C->method()->print_short_name(); 310 if(!EliminateAllocations) { 311 tty->print(" since EliminateAllocations is off ==="); 312 } else if(!has_scalar_replaceable_candidates) { 313 tty->print(" since there are no scalar replaceable candidates ==="); 314 } else if(C->AliasLevel() < 3) { 315 tty->print(" since AliasLevel < 3 ==="); 316 } 317 tty->cr(); 318 #endif 319 } 320 return has_non_escaping_obj; 321 } 322 323 // Utility function for nodes that load an object 324 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 325 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 326 // ThreadLocal has RawPtr type. 327 const Type* t = _igvn->type(n); 328 if (t->make_ptr() != NULL) { 329 Node* adr = n->in(MemNode::Address); 330 #ifdef ASSERT 331 if (!adr->is_AddP()) { 332 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 333 } else { 334 assert((ptnode_adr(adr->_idx) == NULL || 335 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 336 } 337 #endif 338 add_local_var_and_edge(n, PointsToNode::NoEscape, 339 adr, delayed_worklist); 340 } 341 } 342 343 // Populate Connection Graph with PointsTo nodes and create simple 344 // connection graph edges. 345 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 346 assert(!_verify, "this method should not be called for verification"); 347 PhaseGVN* igvn = _igvn; 348 uint n_idx = n->_idx; 349 PointsToNode* n_ptn = ptnode_adr(n_idx); 350 if (n_ptn != NULL) 351 return; // No need to redefine PointsTo node during first iteration. 352 353 int opcode = n->Opcode(); 354 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode); 355 if (gc_handled) { 356 return; // Ignore node if already handled by GC. 357 } 358 359 if (n->is_Call()) { 360 // Arguments to allocation and locking don't escape. 361 if (n->is_AbstractLock()) { 362 // Put Lock and Unlock nodes on IGVN worklist to process them during 363 // first IGVN optimization when escape information is still available. 364 record_for_optimizer(n); 365 } else if (n->is_Allocate()) { 366 add_call_node(n->as_Call()); 367 record_for_optimizer(n); 368 } else { 369 if (n->is_CallStaticJava()) { 370 const char* name = n->as_CallStaticJava()->_name; 371 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 372 return; // Skip uncommon traps 373 } 374 // Don't mark as processed since call's arguments have to be processed. 375 delayed_worklist->push(n); 376 // Check if a call returns an object. 377 if ((n->as_Call()->returns_pointer() && 378 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) || 379 (n->is_CallStaticJava() && 380 n->as_CallStaticJava()->is_boxing_method())) { 381 add_call_node(n->as_Call()); 382 } 383 } 384 return; 385 } 386 // Put this check here to process call arguments since some call nodes 387 // point to phantom_obj. 388 if (n_ptn == phantom_obj || n_ptn == null_obj) 389 return; // Skip predefined nodes. 390 391 switch (opcode) { 392 case Op_AddP: { 393 Node* base = get_addp_base(n); 394 PointsToNode* ptn_base = ptnode_adr(base->_idx); 395 // Field nodes are created for all field types. They are used in 396 // adjust_scalar_replaceable_state() and split_unique_types(). 397 // Note, non-oop fields will have only base edges in Connection 398 // Graph because such fields are not used for oop loads and stores. 399 int offset = address_offset(n, igvn); 400 add_field(n, PointsToNode::NoEscape, offset); 401 if (ptn_base == NULL) { 402 delayed_worklist->push(n); // Process it later. 403 } else { 404 n_ptn = ptnode_adr(n_idx); 405 add_base(n_ptn->as_Field(), ptn_base); 406 } 407 break; 408 } 409 case Op_CastX2P: { 410 map_ideal_node(n, phantom_obj); 411 break; 412 } 413 case Op_CastPP: 414 case Op_CheckCastPP: 415 case Op_EncodeP: 416 case Op_DecodeN: 417 case Op_EncodePKlass: 418 case Op_DecodeNKlass: { 419 add_local_var_and_edge(n, PointsToNode::NoEscape, 420 n->in(1), delayed_worklist); 421 break; 422 } 423 case Op_CMoveP: { 424 add_local_var(n, PointsToNode::NoEscape); 425 // Do not add edges during first iteration because some could be 426 // not defined yet. 427 delayed_worklist->push(n); 428 break; 429 } 430 case Op_ConP: 431 case Op_ConN: 432 case Op_ConNKlass: { 433 // assume all oop constants globally escape except for null 434 PointsToNode::EscapeState es; 435 const Type* t = igvn->type(n); 436 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 437 es = PointsToNode::NoEscape; 438 } else { 439 es = PointsToNode::GlobalEscape; 440 } 441 add_java_object(n, es); 442 break; 443 } 444 case Op_CreateEx: { 445 // assume that all exception objects globally escape 446 map_ideal_node(n, phantom_obj); 447 break; 448 } 449 case Op_LoadKlass: 450 case Op_LoadNKlass: { 451 // Unknown class is loaded 452 map_ideal_node(n, phantom_obj); 453 break; 454 } 455 case Op_LoadP: 456 case Op_LoadN: 457 case Op_LoadPLocked: { 458 add_objload_to_connection_graph(n, delayed_worklist); 459 break; 460 } 461 case Op_Parm: { 462 map_ideal_node(n, phantom_obj); 463 break; 464 } 465 case Op_PartialSubtypeCheck: { 466 // Produces Null or notNull and is used in only in CmpP so 467 // phantom_obj could be used. 468 map_ideal_node(n, phantom_obj); // Result is unknown 469 break; 470 } 471 case Op_Phi: { 472 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 473 // ThreadLocal has RawPtr type. 474 const Type* t = n->as_Phi()->type(); 475 if (t->make_ptr() != NULL) { 476 add_local_var(n, PointsToNode::NoEscape); 477 // Do not add edges during first iteration because some could be 478 // not defined yet. 479 delayed_worklist->push(n); 480 } 481 break; 482 } 483 case Op_Proj: { 484 // we are only interested in the oop result projection from a call 485 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 486 n->in(0)->as_Call()->returns_pointer()) { 487 add_local_var_and_edge(n, PointsToNode::NoEscape, 488 n->in(0), delayed_worklist); 489 } 490 break; 491 } 492 case Op_Rethrow: // Exception object escapes 493 case Op_Return: { 494 if (n->req() > TypeFunc::Parms && 495 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 496 // Treat Return value as LocalVar with GlobalEscape escape state. 497 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 498 n->in(TypeFunc::Parms), delayed_worklist); 499 } 500 break; 501 } 502 case Op_CompareAndExchangeP: 503 case Op_CompareAndExchangeN: 504 case Op_GetAndSetP: 505 case Op_GetAndSetN: { 506 add_objload_to_connection_graph(n, delayed_worklist); 507 // fallthrough 508 } 509 case Op_StoreP: 510 case Op_StoreN: 511 case Op_StoreNKlass: 512 case Op_StorePConditional: 513 case Op_WeakCompareAndSwapP: 514 case Op_WeakCompareAndSwapN: 515 case Op_CompareAndSwapP: 516 case Op_CompareAndSwapN: { 517 add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 518 break; 519 } 520 case Op_AryEq: 521 case Op_HasNegatives: 522 case Op_StrComp: 523 case Op_StrEquals: 524 case Op_StrIndexOf: 525 case Op_StrIndexOfChar: 526 case Op_StrInflatedCopy: 527 case Op_StrCompressedCopy: 528 case Op_EncodeISOArray: { 529 add_local_var(n, PointsToNode::ArgEscape); 530 delayed_worklist->push(n); // Process it later. 531 break; 532 } 533 case Op_ThreadLocal: { 534 add_java_object(n, PointsToNode::ArgEscape); 535 break; 536 } 537 default: 538 ; // Do nothing for nodes not related to EA. 539 } 540 return; 541 } 542 543 #ifdef ASSERT 544 #define ELSE_FAIL(name) \ 545 /* Should not be called for not pointer type. */ \ 546 n->dump(1); \ 547 assert(false, name); \ 548 break; 549 #else 550 #define ELSE_FAIL(name) \ 551 break; 552 #endif 553 554 // Add final simple edges to graph. 555 void ConnectionGraph::add_final_edges(Node *n) { 556 PointsToNode* n_ptn = ptnode_adr(n->_idx); 557 #ifdef ASSERT 558 if (_verify && n_ptn->is_JavaObject()) 559 return; // This method does not change graph for JavaObject. 560 #endif 561 562 if (n->is_Call()) { 563 process_call_arguments(n->as_Call()); 564 return; 565 } 566 assert(n->is_Store() || n->is_LoadStore() || 567 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 568 "node should be registered already"); 569 int opcode = n->Opcode(); 570 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode); 571 if (gc_handled) { 572 return; // Ignore node if already handled by GC. 573 } 574 switch (opcode) { 575 case Op_AddP: { 576 Node* base = get_addp_base(n); 577 PointsToNode* ptn_base = ptnode_adr(base->_idx); 578 assert(ptn_base != NULL, "field's base should be registered"); 579 add_base(n_ptn->as_Field(), ptn_base); 580 break; 581 } 582 case Op_CastPP: 583 case Op_CheckCastPP: 584 case Op_EncodeP: 585 case Op_DecodeN: 586 case Op_EncodePKlass: 587 case Op_DecodeNKlass: { 588 add_local_var_and_edge(n, PointsToNode::NoEscape, 589 n->in(1), NULL); 590 break; 591 } 592 case Op_CMoveP: { 593 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 594 Node* in = n->in(i); 595 if (in == NULL) 596 continue; // ignore NULL 597 Node* uncast_in = in->uncast(); 598 if (uncast_in->is_top() || uncast_in == n) 599 continue; // ignore top or inputs which go back this node 600 PointsToNode* ptn = ptnode_adr(in->_idx); 601 assert(ptn != NULL, "node should be registered"); 602 add_edge(n_ptn, ptn); 603 } 604 break; 605 } 606 case Op_LoadP: 607 case Op_LoadN: 608 case Op_LoadPLocked: { 609 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 610 // ThreadLocal has RawPtr type. 611 const Type* t = _igvn->type(n); 612 if (t->make_ptr() != NULL) { 613 Node* adr = n->in(MemNode::Address); 614 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 615 break; 616 } 617 ELSE_FAIL("Op_LoadP"); 618 } 619 case Op_Phi: { 620 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 621 // ThreadLocal has RawPtr type. 622 const Type* t = n->as_Phi()->type(); 623 if (t->make_ptr() != NULL) { 624 for (uint i = 1; i < n->req(); i++) { 625 Node* in = n->in(i); 626 if (in == NULL) 627 continue; // ignore NULL 628 Node* uncast_in = in->uncast(); 629 if (uncast_in->is_top() || uncast_in == n) 630 continue; // ignore top or inputs which go back this node 631 PointsToNode* ptn = ptnode_adr(in->_idx); 632 assert(ptn != NULL, "node should be registered"); 633 add_edge(n_ptn, ptn); 634 } 635 break; 636 } 637 ELSE_FAIL("Op_Phi"); 638 } 639 case Op_Proj: { 640 // we are only interested in the oop result projection from a call 641 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 642 n->in(0)->as_Call()->returns_pointer()) { 643 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 644 break; 645 } 646 ELSE_FAIL("Op_Proj"); 647 } 648 case Op_Rethrow: // Exception object escapes 649 case Op_Return: { 650 if (n->req() > TypeFunc::Parms && 651 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 652 // Treat Return value as LocalVar with GlobalEscape escape state. 653 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 654 n->in(TypeFunc::Parms), NULL); 655 break; 656 } 657 ELSE_FAIL("Op_Return"); 658 } 659 case Op_StoreP: 660 case Op_StoreN: 661 case Op_StoreNKlass: 662 case Op_StorePConditional: 663 case Op_CompareAndExchangeP: 664 case Op_CompareAndExchangeN: 665 case Op_CompareAndSwapP: 666 case Op_CompareAndSwapN: 667 case Op_WeakCompareAndSwapP: 668 case Op_WeakCompareAndSwapN: 669 case Op_GetAndSetP: 670 case Op_GetAndSetN: { 671 if (add_final_edges_unsafe_access(n, opcode)) { 672 break; 673 } 674 ELSE_FAIL("Op_StoreP"); 675 } 676 case Op_AryEq: 677 case Op_HasNegatives: 678 case Op_StrComp: 679 case Op_StrEquals: 680 case Op_StrIndexOf: 681 case Op_StrIndexOfChar: 682 case Op_StrInflatedCopy: 683 case Op_StrCompressedCopy: 684 case Op_EncodeISOArray: { 685 // char[]/byte[] arrays passed to string intrinsic do not escape but 686 // they are not scalar replaceable. Adjust escape state for them. 687 // Start from in(2) edge since in(1) is memory edge. 688 for (uint i = 2; i < n->req(); i++) { 689 Node* adr = n->in(i); 690 const Type* at = _igvn->type(adr); 691 if (!adr->is_top() && at->isa_ptr()) { 692 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 693 at->isa_ptr() != NULL, "expecting a pointer"); 694 if (adr->is_AddP()) { 695 adr = get_addp_base(adr); 696 } 697 PointsToNode* ptn = ptnode_adr(adr->_idx); 698 assert(ptn != NULL, "node should be registered"); 699 add_edge(n_ptn, ptn); 700 } 701 } 702 break; 703 } 704 default: { 705 // This method should be called only for EA specific nodes which may 706 // miss some edges when they were created. 707 #ifdef ASSERT 708 n->dump(1); 709 #endif 710 guarantee(false, "unknown node"); 711 } 712 } 713 return; 714 } 715 716 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) { 717 Node* adr = n->in(MemNode::Address); 718 const Type* adr_type = _igvn->type(adr); 719 adr_type = adr_type->make_ptr(); 720 if (adr_type == NULL) { 721 return; // skip dead nodes 722 } 723 if (adr_type->isa_oopptr() 724 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 725 && adr_type == TypeRawPtr::NOTNULL 726 && adr->in(AddPNode::Address)->is_Proj() 727 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 728 delayed_worklist->push(n); // Process it later. 729 #ifdef ASSERT 730 assert (adr->is_AddP(), "expecting an AddP"); 731 if (adr_type == TypeRawPtr::NOTNULL) { 732 // Verify a raw address for a store captured by Initialize node. 733 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 734 assert(offs != Type::OffsetBot, "offset must be a constant"); 735 } 736 #endif 737 } else { 738 // Ignore copy the displaced header to the BoxNode (OSR compilation). 739 if (adr->is_BoxLock()) { 740 return; 741 } 742 // Stored value escapes in unsafe access. 743 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 744 delayed_worklist->push(n); // Process unsafe access later. 745 return; 746 } 747 #ifdef ASSERT 748 n->dump(1); 749 assert(false, "not unsafe"); 750 #endif 751 } 752 } 753 754 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) { 755 Node* adr = n->in(MemNode::Address); 756 const Type *adr_type = _igvn->type(adr); 757 adr_type = adr_type->make_ptr(); 758 #ifdef ASSERT 759 if (adr_type == NULL) { 760 n->dump(1); 761 assert(adr_type != NULL, "dead node should not be on list"); 762 return true; 763 } 764 #endif 765 766 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 767 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 768 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 769 } 770 771 if (adr_type->isa_oopptr() 772 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 773 && adr_type == TypeRawPtr::NOTNULL 774 && adr->in(AddPNode::Address)->is_Proj() 775 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 776 // Point Address to Value 777 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 778 assert(adr_ptn != NULL && 779 adr_ptn->as_Field()->is_oop(), "node should be registered"); 780 Node* val = n->in(MemNode::ValueIn); 781 PointsToNode* ptn = ptnode_adr(val->_idx); 782 assert(ptn != NULL, "node should be registered"); 783 add_edge(adr_ptn, ptn); 784 return true; 785 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 786 // Stored value escapes in unsafe access. 787 Node* val = n->in(MemNode::ValueIn); 788 PointsToNode* ptn = ptnode_adr(val->_idx); 789 assert(ptn != NULL, "node should be registered"); 790 set_escape_state(ptn, PointsToNode::GlobalEscape); 791 // Add edge to object for unsafe access with offset. 792 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 793 assert(adr_ptn != NULL, "node should be registered"); 794 if (adr_ptn->is_Field()) { 795 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 796 add_edge(adr_ptn, ptn); 797 } 798 return true; 799 } 800 return false; 801 } 802 803 void ConnectionGraph::add_call_node(CallNode* call) { 804 assert(call->returns_pointer(), "only for call which returns pointer"); 805 uint call_idx = call->_idx; 806 if (call->is_Allocate()) { 807 Node* k = call->in(AllocateNode::KlassNode); 808 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 809 assert(kt != NULL, "TypeKlassPtr required."); 810 ciKlass* cik = kt->klass(); 811 PointsToNode::EscapeState es = PointsToNode::NoEscape; 812 bool scalar_replaceable = true; 813 if (call->is_AllocateArray()) { 814 if (!cik->is_array_klass()) { // StressReflectiveCode 815 es = PointsToNode::GlobalEscape; 816 } else { 817 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 818 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 819 // Not scalar replaceable if the length is not constant or too big. 820 scalar_replaceable = false; 821 } 822 } 823 } else { // Allocate instance 824 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 825 cik->is_subclass_of(_compile->env()->Reference_klass()) || 826 !cik->is_instance_klass() || // StressReflectiveCode 827 !cik->as_instance_klass()->can_be_instantiated() || 828 cik->as_instance_klass()->has_finalizer()) { 829 es = PointsToNode::GlobalEscape; 830 } 831 } 832 add_java_object(call, es); 833 PointsToNode* ptn = ptnode_adr(call_idx); 834 if (!scalar_replaceable && ptn->scalar_replaceable()) { 835 ptn->set_scalar_replaceable(false); 836 } 837 } else if (call->is_CallStaticJava()) { 838 // Call nodes could be different types: 839 // 840 // 1. CallDynamicJavaNode (what happened during call is unknown): 841 // 842 // - mapped to GlobalEscape JavaObject node if oop is returned; 843 // 844 // - all oop arguments are escaping globally; 845 // 846 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 847 // 848 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 849 // 850 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 851 // - mapped to NoEscape JavaObject node if non-escaping object allocated 852 // during call is returned; 853 // - mapped to ArgEscape LocalVar node pointed to object arguments 854 // which are returned and does not escape during call; 855 // 856 // - oop arguments escaping status is defined by bytecode analysis; 857 // 858 // For a static call, we know exactly what method is being called. 859 // Use bytecode estimator to record whether the call's return value escapes. 860 ciMethod* meth = call->as_CallJava()->method(); 861 if (meth == NULL) { 862 const char* name = call->as_CallStaticJava()->_name; 863 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 864 // Returns a newly allocated unescaped object. 865 add_java_object(call, PointsToNode::NoEscape); 866 ptnode_adr(call_idx)->set_scalar_replaceable(false); 867 } else if (meth->is_boxing_method()) { 868 // Returns boxing object 869 PointsToNode::EscapeState es; 870 vmIntrinsics::ID intr = meth->intrinsic_id(); 871 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 872 // It does not escape if object is always allocated. 873 es = PointsToNode::NoEscape; 874 } else { 875 // It escapes globally if object could be loaded from cache. 876 es = PointsToNode::GlobalEscape; 877 } 878 add_java_object(call, es); 879 } else { 880 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 881 call_analyzer->copy_dependencies(_compile->dependencies()); 882 if (call_analyzer->is_return_allocated()) { 883 // Returns a newly allocated unescaped object, simply 884 // update dependency information. 885 // Mark it as NoEscape so that objects referenced by 886 // it's fields will be marked as NoEscape at least. 887 add_java_object(call, PointsToNode::NoEscape); 888 ptnode_adr(call_idx)->set_scalar_replaceable(false); 889 } else { 890 // Determine whether any arguments are returned. 891 const TypeTuple* d = call->tf()->domain(); 892 bool ret_arg = false; 893 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 894 if (d->field_at(i)->isa_ptr() != NULL && 895 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 896 ret_arg = true; 897 break; 898 } 899 } 900 if (ret_arg) { 901 add_local_var(call, PointsToNode::ArgEscape); 902 } else { 903 // Returns unknown object. 904 map_ideal_node(call, phantom_obj); 905 } 906 } 907 } 908 } else { 909 // An other type of call, assume the worst case: 910 // returned value is unknown and globally escapes. 911 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 912 map_ideal_node(call, phantom_obj); 913 } 914 } 915 916 void ConnectionGraph::process_call_arguments(CallNode *call) { 917 bool is_arraycopy = false; 918 switch (call->Opcode()) { 919 #ifdef ASSERT 920 case Op_Allocate: 921 case Op_AllocateArray: 922 case Op_Lock: 923 case Op_Unlock: 924 assert(false, "should be done already"); 925 break; 926 #endif 927 case Op_ArrayCopy: 928 case Op_CallLeafNoFP: 929 // Most array copies are ArrayCopy nodes at this point but there 930 // are still a few direct calls to the copy subroutines (See 931 // PhaseStringOpts::copy_string()) 932 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 933 call->as_CallLeaf()->is_call_to_arraycopystub(); 934 // fall through 935 case Op_CallLeaf: { 936 // Stub calls, objects do not escape but they are not scale replaceable. 937 // Adjust escape state for outgoing arguments. 938 const TypeTuple * d = call->tf()->domain(); 939 bool src_has_oops = false; 940 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 941 const Type* at = d->field_at(i); 942 Node *arg = call->in(i); 943 if (arg == NULL) { 944 continue; 945 } 946 const Type *aat = _igvn->type(arg); 947 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 948 continue; 949 if (arg->is_AddP()) { 950 // 951 // The inline_native_clone() case when the arraycopy stub is called 952 // after the allocation before Initialize and CheckCastPP nodes. 953 // Or normal arraycopy for object arrays case. 954 // 955 // Set AddP's base (Allocate) as not scalar replaceable since 956 // pointer to the base (with offset) is passed as argument. 957 // 958 arg = get_addp_base(arg); 959 } 960 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 961 assert(arg_ptn != NULL, "should be registered"); 962 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 963 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 964 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 965 aat->isa_ptr() != NULL, "expecting an Ptr"); 966 bool arg_has_oops = aat->isa_oopptr() && 967 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 968 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 969 if (i == TypeFunc::Parms) { 970 src_has_oops = arg_has_oops; 971 } 972 // 973 // src or dst could be j.l.Object when other is basic type array: 974 // 975 // arraycopy(char[],0,Object*,0,size); 976 // arraycopy(Object*,0,char[],0,size); 977 // 978 // Don't add edges in such cases. 979 // 980 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 981 arg_has_oops && (i > TypeFunc::Parms); 982 #ifdef ASSERT 983 if (!(is_arraycopy || 984 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) || 985 (call->as_CallLeaf()->_name != NULL && 986 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 987 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 988 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 989 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 990 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 991 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 992 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 993 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 || 994 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 || 995 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 || 996 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 997 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 || 998 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 999 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 1000 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 1001 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 1002 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 1003 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 1004 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 1005 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 1006 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 1007 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 1008 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 || 1009 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 || 1010 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 || 1011 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0) 1012 ))) { 1013 call->dump(); 1014 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name); 1015 } 1016 #endif 1017 // Always process arraycopy's destination object since 1018 // we need to add all possible edges to references in 1019 // source object. 1020 if (arg_esc >= PointsToNode::ArgEscape && 1021 !arg_is_arraycopy_dest) { 1022 continue; 1023 } 1024 PointsToNode::EscapeState es = PointsToNode::ArgEscape; 1025 if (call->is_ArrayCopy()) { 1026 ArrayCopyNode* ac = call->as_ArrayCopy(); 1027 if (ac->is_clonebasic() || 1028 ac->is_arraycopy_validated() || 1029 ac->is_copyof_validated() || 1030 ac->is_copyofrange_validated()) { 1031 es = PointsToNode::NoEscape; 1032 } 1033 } 1034 set_escape_state(arg_ptn, es); 1035 if (arg_is_arraycopy_dest) { 1036 Node* src = call->in(TypeFunc::Parms); 1037 if (src->is_AddP()) { 1038 src = get_addp_base(src); 1039 } 1040 PointsToNode* src_ptn = ptnode_adr(src->_idx); 1041 assert(src_ptn != NULL, "should be registered"); 1042 if (arg_ptn != src_ptn) { 1043 // Special arraycopy edge: 1044 // A destination object's field can't have the source object 1045 // as base since objects escape states are not related. 1046 // Only escape state of destination object's fields affects 1047 // escape state of fields in source object. 1048 add_arraycopy(call, es, src_ptn, arg_ptn); 1049 } 1050 } 1051 } 1052 } 1053 break; 1054 } 1055 case Op_CallStaticJava: { 1056 // For a static call, we know exactly what method is being called. 1057 // Use bytecode estimator to record the call's escape affects 1058 #ifdef ASSERT 1059 const char* name = call->as_CallStaticJava()->_name; 1060 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1061 #endif 1062 ciMethod* meth = call->as_CallJava()->method(); 1063 if ((meth != NULL) && meth->is_boxing_method()) { 1064 break; // Boxing methods do not modify any oops. 1065 } 1066 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1067 // fall-through if not a Java method or no analyzer information 1068 if (call_analyzer != NULL) { 1069 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1070 const TypeTuple* d = call->tf()->domain(); 1071 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1072 const Type* at = d->field_at(i); 1073 int k = i - TypeFunc::Parms; 1074 Node* arg = call->in(i); 1075 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1076 if (at->isa_ptr() != NULL && 1077 call_analyzer->is_arg_returned(k)) { 1078 // The call returns arguments. 1079 if (call_ptn != NULL) { // Is call's result used? 1080 assert(call_ptn->is_LocalVar(), "node should be registered"); 1081 assert(arg_ptn != NULL, "node should be registered"); 1082 add_edge(call_ptn, arg_ptn); 1083 } 1084 } 1085 if (at->isa_oopptr() != NULL && 1086 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1087 if (!call_analyzer->is_arg_stack(k)) { 1088 // The argument global escapes 1089 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1090 } else { 1091 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1092 if (!call_analyzer->is_arg_local(k)) { 1093 // The argument itself doesn't escape, but any fields might 1094 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1095 } 1096 } 1097 } 1098 } 1099 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1100 // The call returns arguments. 1101 assert(call_ptn->edge_count() > 0, "sanity"); 1102 if (!call_analyzer->is_return_local()) { 1103 // Returns also unknown object. 1104 add_edge(call_ptn, phantom_obj); 1105 } 1106 } 1107 break; 1108 } 1109 } 1110 default: { 1111 // Fall-through here if not a Java method or no analyzer information 1112 // or some other type of call, assume the worst case: all arguments 1113 // globally escape. 1114 const TypeTuple* d = call->tf()->domain(); 1115 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1116 const Type* at = d->field_at(i); 1117 if (at->isa_oopptr() != NULL) { 1118 Node* arg = call->in(i); 1119 if (arg->is_AddP()) { 1120 arg = get_addp_base(arg); 1121 } 1122 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1123 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1124 } 1125 } 1126 } 1127 } 1128 } 1129 1130 1131 // Finish Graph construction. 1132 bool ConnectionGraph::complete_connection_graph( 1133 GrowableArray<PointsToNode*>& ptnodes_worklist, 1134 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1135 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1136 GrowableArray<FieldNode*>& oop_fields_worklist) { 1137 // Normally only 1-3 passes needed to build Connection Graph depending 1138 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1139 // Set limit to 20 to catch situation when something did go wrong and 1140 // bailout Escape Analysis. 1141 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1142 #define CG_BUILD_ITER_LIMIT 20 1143 1144 // Propagate GlobalEscape and ArgEscape escape states and check that 1145 // we still have non-escaping objects. The method pushs on _worklist 1146 // Field nodes which reference phantom_object. 1147 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1148 return false; // Nothing to do. 1149 } 1150 // Now propagate references to all JavaObject nodes. 1151 int java_objects_length = java_objects_worklist.length(); 1152 elapsedTimer time; 1153 bool timeout = false; 1154 int new_edges = 1; 1155 int iterations = 0; 1156 do { 1157 while ((new_edges > 0) && 1158 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1159 double start_time = time.seconds(); 1160 time.start(); 1161 new_edges = 0; 1162 // Propagate references to phantom_object for nodes pushed on _worklist 1163 // by find_non_escaped_objects() and find_field_value(). 1164 new_edges += add_java_object_edges(phantom_obj, false); 1165 for (int next = 0; next < java_objects_length; ++next) { 1166 JavaObjectNode* ptn = java_objects_worklist.at(next); 1167 new_edges += add_java_object_edges(ptn, true); 1168 1169 #define SAMPLE_SIZE 4 1170 if ((next % SAMPLE_SIZE) == 0) { 1171 // Each 4 iterations calculate how much time it will take 1172 // to complete graph construction. 1173 time.stop(); 1174 // Poll for requests from shutdown mechanism to quiesce compiler 1175 // because Connection graph construction may take long time. 1176 CompileBroker::maybe_block(); 1177 double stop_time = time.seconds(); 1178 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1179 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1180 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1181 timeout = true; 1182 break; // Timeout 1183 } 1184 start_time = stop_time; 1185 time.start(); 1186 } 1187 #undef SAMPLE_SIZE 1188 1189 } 1190 if (timeout) break; 1191 if (new_edges > 0) { 1192 // Update escape states on each iteration if graph was updated. 1193 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1194 return false; // Nothing to do. 1195 } 1196 } 1197 time.stop(); 1198 if (time.seconds() >= EscapeAnalysisTimeout) { 1199 timeout = true; 1200 break; 1201 } 1202 } 1203 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1204 time.start(); 1205 // Find fields which have unknown value. 1206 int fields_length = oop_fields_worklist.length(); 1207 for (int next = 0; next < fields_length; next++) { 1208 FieldNode* field = oop_fields_worklist.at(next); 1209 if (field->edge_count() == 0) { 1210 new_edges += find_field_value(field); 1211 // This code may added new edges to phantom_object. 1212 // Need an other cycle to propagate references to phantom_object. 1213 } 1214 } 1215 time.stop(); 1216 if (time.seconds() >= EscapeAnalysisTimeout) { 1217 timeout = true; 1218 break; 1219 } 1220 } else { 1221 new_edges = 0; // Bailout 1222 } 1223 } while (new_edges > 0); 1224 1225 // Bailout if passed limits. 1226 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1227 Compile* C = _compile; 1228 if (C->log() != NULL) { 1229 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1230 C->log()->text("%s", timeout ? "time" : "iterations"); 1231 C->log()->end_elem(" limit'"); 1232 } 1233 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1234 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()); 1235 // Possible infinite build_connection_graph loop, 1236 // bailout (no changes to ideal graph were made). 1237 return false; 1238 } 1239 #ifdef ASSERT 1240 if (Verbose && PrintEscapeAnalysis) { 1241 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1242 iterations, nodes_size(), ptnodes_worklist.length()); 1243 } 1244 #endif 1245 1246 #undef CG_BUILD_ITER_LIMIT 1247 1248 // Find fields initialized by NULL for non-escaping Allocations. 1249 int non_escaped_length = non_escaped_worklist.length(); 1250 for (int next = 0; next < non_escaped_length; next++) { 1251 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1252 PointsToNode::EscapeState es = ptn->escape_state(); 1253 assert(es <= PointsToNode::ArgEscape, "sanity"); 1254 if (es == PointsToNode::NoEscape) { 1255 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1256 // Adding references to NULL object does not change escape states 1257 // since it does not escape. Also no fields are added to NULL object. 1258 add_java_object_edges(null_obj, false); 1259 } 1260 } 1261 Node* n = ptn->ideal_node(); 1262 if (n->is_Allocate()) { 1263 // The object allocated by this Allocate node will never be 1264 // seen by an other thread. Mark it so that when it is 1265 // expanded no MemBarStoreStore is added. 1266 InitializeNode* ini = n->as_Allocate()->initialization(); 1267 if (ini != NULL) 1268 ini->set_does_not_escape(); 1269 } 1270 } 1271 return true; // Finished graph construction. 1272 } 1273 1274 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1275 // and check that we still have non-escaping java objects. 1276 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1277 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1278 GrowableArray<PointsToNode*> escape_worklist; 1279 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1280 int ptnodes_length = ptnodes_worklist.length(); 1281 for (int next = 0; next < ptnodes_length; ++next) { 1282 PointsToNode* ptn = ptnodes_worklist.at(next); 1283 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1284 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1285 escape_worklist.push(ptn); 1286 } 1287 } 1288 // Set escape states to referenced nodes (edges list). 1289 while (escape_worklist.length() > 0) { 1290 PointsToNode* ptn = escape_worklist.pop(); 1291 PointsToNode::EscapeState es = ptn->escape_state(); 1292 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1293 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1294 es >= PointsToNode::ArgEscape) { 1295 // GlobalEscape or ArgEscape state of field means it has unknown value. 1296 if (add_edge(ptn, phantom_obj)) { 1297 // New edge was added 1298 add_field_uses_to_worklist(ptn->as_Field()); 1299 } 1300 } 1301 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1302 PointsToNode* e = i.get(); 1303 if (e->is_Arraycopy()) { 1304 assert(ptn->arraycopy_dst(), "sanity"); 1305 // Propagate only fields escape state through arraycopy edge. 1306 if (e->fields_escape_state() < field_es) { 1307 set_fields_escape_state(e, field_es); 1308 escape_worklist.push(e); 1309 } 1310 } else if (es >= field_es) { 1311 // fields_escape_state is also set to 'es' if it is less than 'es'. 1312 if (e->escape_state() < es) { 1313 set_escape_state(e, es); 1314 escape_worklist.push(e); 1315 } 1316 } else { 1317 // Propagate field escape state. 1318 bool es_changed = false; 1319 if (e->fields_escape_state() < field_es) { 1320 set_fields_escape_state(e, field_es); 1321 es_changed = true; 1322 } 1323 if ((e->escape_state() < field_es) && 1324 e->is_Field() && ptn->is_JavaObject() && 1325 e->as_Field()->is_oop()) { 1326 // Change escape state of referenced fields. 1327 set_escape_state(e, field_es); 1328 es_changed = true; 1329 } else if (e->escape_state() < es) { 1330 set_escape_state(e, es); 1331 es_changed = true; 1332 } 1333 if (es_changed) { 1334 escape_worklist.push(e); 1335 } 1336 } 1337 } 1338 } 1339 // Remove escaped objects from non_escaped list. 1340 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1341 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1342 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1343 non_escaped_worklist.delete_at(next); 1344 } 1345 if (ptn->escape_state() == PointsToNode::NoEscape) { 1346 // Find fields in non-escaped allocations which have unknown value. 1347 find_init_values(ptn, phantom_obj, NULL); 1348 } 1349 } 1350 return (non_escaped_worklist.length() > 0); 1351 } 1352 1353 // Add all references to JavaObject node by walking over all uses. 1354 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1355 int new_edges = 0; 1356 if (populate_worklist) { 1357 // Populate _worklist by uses of jobj's uses. 1358 for (UseIterator i(jobj); i.has_next(); i.next()) { 1359 PointsToNode* use = i.get(); 1360 if (use->is_Arraycopy()) 1361 continue; 1362 add_uses_to_worklist(use); 1363 if (use->is_Field() && use->as_Field()->is_oop()) { 1364 // Put on worklist all field's uses (loads) and 1365 // related field nodes (same base and offset). 1366 add_field_uses_to_worklist(use->as_Field()); 1367 } 1368 } 1369 } 1370 for (int l = 0; l < _worklist.length(); l++) { 1371 PointsToNode* use = _worklist.at(l); 1372 if (PointsToNode::is_base_use(use)) { 1373 // Add reference from jobj to field and from field to jobj (field's base). 1374 use = PointsToNode::get_use_node(use)->as_Field(); 1375 if (add_base(use->as_Field(), jobj)) { 1376 new_edges++; 1377 } 1378 continue; 1379 } 1380 assert(!use->is_JavaObject(), "sanity"); 1381 if (use->is_Arraycopy()) { 1382 if (jobj == null_obj) // NULL object does not have field edges 1383 continue; 1384 // Added edge from Arraycopy node to arraycopy's source java object 1385 if (add_edge(use, jobj)) { 1386 jobj->set_arraycopy_src(); 1387 new_edges++; 1388 } 1389 // and stop here. 1390 continue; 1391 } 1392 if (!add_edge(use, jobj)) 1393 continue; // No new edge added, there was such edge already. 1394 new_edges++; 1395 if (use->is_LocalVar()) { 1396 add_uses_to_worklist(use); 1397 if (use->arraycopy_dst()) { 1398 for (EdgeIterator i(use); i.has_next(); i.next()) { 1399 PointsToNode* e = i.get(); 1400 if (e->is_Arraycopy()) { 1401 if (jobj == null_obj) // NULL object does not have field edges 1402 continue; 1403 // Add edge from arraycopy's destination java object to Arraycopy node. 1404 if (add_edge(jobj, e)) { 1405 new_edges++; 1406 jobj->set_arraycopy_dst(); 1407 } 1408 } 1409 } 1410 } 1411 } else { 1412 // Added new edge to stored in field values. 1413 // Put on worklist all field's uses (loads) and 1414 // related field nodes (same base and offset). 1415 add_field_uses_to_worklist(use->as_Field()); 1416 } 1417 } 1418 _worklist.clear(); 1419 _in_worklist.reset(); 1420 return new_edges; 1421 } 1422 1423 // Put on worklist all related field nodes. 1424 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1425 assert(field->is_oop(), "sanity"); 1426 int offset = field->offset(); 1427 add_uses_to_worklist(field); 1428 // Loop over all bases of this field and push on worklist Field nodes 1429 // with the same offset and base (since they may reference the same field). 1430 for (BaseIterator i(field); i.has_next(); i.next()) { 1431 PointsToNode* base = i.get(); 1432 add_fields_to_worklist(field, base); 1433 // Check if the base was source object of arraycopy and go over arraycopy's 1434 // destination objects since values stored to a field of source object are 1435 // accessable by uses (loads) of fields of destination objects. 1436 if (base->arraycopy_src()) { 1437 for (UseIterator j(base); j.has_next(); j.next()) { 1438 PointsToNode* arycp = j.get(); 1439 if (arycp->is_Arraycopy()) { 1440 for (UseIterator k(arycp); k.has_next(); k.next()) { 1441 PointsToNode* abase = k.get(); 1442 if (abase->arraycopy_dst() && abase != base) { 1443 // Look for the same arraycopy reference. 1444 add_fields_to_worklist(field, abase); 1445 } 1446 } 1447 } 1448 } 1449 } 1450 } 1451 } 1452 1453 // Put on worklist all related field nodes. 1454 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1455 int offset = field->offset(); 1456 if (base->is_LocalVar()) { 1457 for (UseIterator j(base); j.has_next(); j.next()) { 1458 PointsToNode* f = j.get(); 1459 if (PointsToNode::is_base_use(f)) { // Field 1460 f = PointsToNode::get_use_node(f); 1461 if (f == field || !f->as_Field()->is_oop()) 1462 continue; 1463 int offs = f->as_Field()->offset(); 1464 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1465 add_to_worklist(f); 1466 } 1467 } 1468 } 1469 } else { 1470 assert(base->is_JavaObject(), "sanity"); 1471 if (// Skip phantom_object since it is only used to indicate that 1472 // this field's content globally escapes. 1473 (base != phantom_obj) && 1474 // NULL object node does not have fields. 1475 (base != null_obj)) { 1476 for (EdgeIterator i(base); i.has_next(); i.next()) { 1477 PointsToNode* f = i.get(); 1478 // Skip arraycopy edge since store to destination object field 1479 // does not update value in source object field. 1480 if (f->is_Arraycopy()) { 1481 assert(base->arraycopy_dst(), "sanity"); 1482 continue; 1483 } 1484 if (f == field || !f->as_Field()->is_oop()) 1485 continue; 1486 int offs = f->as_Field()->offset(); 1487 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1488 add_to_worklist(f); 1489 } 1490 } 1491 } 1492 } 1493 } 1494 1495 // Find fields which have unknown value. 1496 int ConnectionGraph::find_field_value(FieldNode* field) { 1497 // Escaped fields should have init value already. 1498 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1499 int new_edges = 0; 1500 for (BaseIterator i(field); i.has_next(); i.next()) { 1501 PointsToNode* base = i.get(); 1502 if (base->is_JavaObject()) { 1503 // Skip Allocate's fields which will be processed later. 1504 if (base->ideal_node()->is_Allocate()) 1505 return 0; 1506 assert(base == null_obj, "only NULL ptr base expected here"); 1507 } 1508 } 1509 if (add_edge(field, phantom_obj)) { 1510 // New edge was added 1511 new_edges++; 1512 add_field_uses_to_worklist(field); 1513 } 1514 return new_edges; 1515 } 1516 1517 // Find fields initializing values for allocations. 1518 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1519 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1520 int new_edges = 0; 1521 Node* alloc = pta->ideal_node(); 1522 if (init_val == phantom_obj) { 1523 // Do nothing for Allocate nodes since its fields values are 1524 // "known" unless they are initialized by arraycopy/clone. 1525 if (alloc->is_Allocate() && !pta->arraycopy_dst()) 1526 return 0; 1527 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity"); 1528 #ifdef ASSERT 1529 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) { 1530 const char* name = alloc->as_CallStaticJava()->_name; 1531 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1532 } 1533 #endif 1534 // Non-escaped allocation returned from Java or runtime call have 1535 // unknown values in fields. 1536 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1537 PointsToNode* field = i.get(); 1538 if (field->is_Field() && field->as_Field()->is_oop()) { 1539 if (add_edge(field, phantom_obj)) { 1540 // New edge was added 1541 new_edges++; 1542 add_field_uses_to_worklist(field->as_Field()); 1543 } 1544 } 1545 } 1546 return new_edges; 1547 } 1548 assert(init_val == null_obj, "sanity"); 1549 // Do nothing for Call nodes since its fields values are unknown. 1550 if (!alloc->is_Allocate()) 1551 return 0; 1552 1553 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1554 bool visited_bottom_offset = false; 1555 GrowableArray<int> offsets_worklist; 1556 1557 // Check if an oop field's initializing value is recorded and add 1558 // a corresponding NULL if field's value if it is not recorded. 1559 // Connection Graph does not record a default initialization by NULL 1560 // captured by Initialize node. 1561 // 1562 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1563 PointsToNode* field = i.get(); // Field (AddP) 1564 if (!field->is_Field() || !field->as_Field()->is_oop()) 1565 continue; // Not oop field 1566 int offset = field->as_Field()->offset(); 1567 if (offset == Type::OffsetBot) { 1568 if (!visited_bottom_offset) { 1569 // OffsetBot is used to reference array's element, 1570 // always add reference to NULL to all Field nodes since we don't 1571 // known which element is referenced. 1572 if (add_edge(field, null_obj)) { 1573 // New edge was added 1574 new_edges++; 1575 add_field_uses_to_worklist(field->as_Field()); 1576 visited_bottom_offset = true; 1577 } 1578 } 1579 } else { 1580 // Check only oop fields. 1581 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1582 if (adr_type->isa_rawptr()) { 1583 #ifdef ASSERT 1584 // Raw pointers are used for initializing stores so skip it 1585 // since it should be recorded already 1586 Node* base = get_addp_base(field->ideal_node()); 1587 assert(adr_type->isa_rawptr() && base->is_Proj() && 1588 (base->in(0) == alloc),"unexpected pointer type"); 1589 #endif 1590 continue; 1591 } 1592 if (!offsets_worklist.contains(offset)) { 1593 offsets_worklist.append(offset); 1594 Node* value = NULL; 1595 if (ini != NULL) { 1596 // StoreP::memory_type() == T_ADDRESS 1597 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1598 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1599 // Make sure initializing store has the same type as this AddP. 1600 // This AddP may reference non existing field because it is on a 1601 // dead branch of bimorphic call which is not eliminated yet. 1602 if (store != NULL && store->is_Store() && 1603 store->as_Store()->memory_type() == ft) { 1604 value = store->in(MemNode::ValueIn); 1605 #ifdef ASSERT 1606 if (VerifyConnectionGraph) { 1607 // Verify that AddP already points to all objects the value points to. 1608 PointsToNode* val = ptnode_adr(value->_idx); 1609 assert((val != NULL), "should be processed already"); 1610 PointsToNode* missed_obj = NULL; 1611 if (val->is_JavaObject()) { 1612 if (!field->points_to(val->as_JavaObject())) { 1613 missed_obj = val; 1614 } 1615 } else { 1616 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1617 tty->print_cr("----------init store has invalid value -----"); 1618 store->dump(); 1619 val->dump(); 1620 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1621 } 1622 for (EdgeIterator j(val); j.has_next(); j.next()) { 1623 PointsToNode* obj = j.get(); 1624 if (obj->is_JavaObject()) { 1625 if (!field->points_to(obj->as_JavaObject())) { 1626 missed_obj = obj; 1627 break; 1628 } 1629 } 1630 } 1631 } 1632 if (missed_obj != NULL) { 1633 tty->print_cr("----------field---------------------------------"); 1634 field->dump(); 1635 tty->print_cr("----------missed referernce to object-----------"); 1636 missed_obj->dump(); 1637 tty->print_cr("----------object referernced by init store -----"); 1638 store->dump(); 1639 val->dump(); 1640 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1641 } 1642 } 1643 #endif 1644 } else { 1645 // There could be initializing stores which follow allocation. 1646 // For example, a volatile field store is not collected 1647 // by Initialize node. 1648 // 1649 // Need to check for dependent loads to separate such stores from 1650 // stores which follow loads. For now, add initial value NULL so 1651 // that compare pointers optimization works correctly. 1652 } 1653 } 1654 if (value == NULL) { 1655 // A field's initializing value was not recorded. Add NULL. 1656 if (add_edge(field, null_obj)) { 1657 // New edge was added 1658 new_edges++; 1659 add_field_uses_to_worklist(field->as_Field()); 1660 } 1661 } 1662 } 1663 } 1664 } 1665 return new_edges; 1666 } 1667 1668 // Adjust scalar_replaceable state after Connection Graph is built. 1669 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1670 // Search for non-escaping objects which are not scalar replaceable 1671 // and mark them to propagate the state to referenced objects. 1672 1673 // 1. An object is not scalar replaceable if the field into which it is 1674 // stored has unknown offset (stored into unknown element of an array). 1675 // 1676 for (UseIterator i(jobj); i.has_next(); i.next()) { 1677 PointsToNode* use = i.get(); 1678 if (use->is_Arraycopy()) { 1679 continue; 1680 } 1681 if (use->is_Field()) { 1682 FieldNode* field = use->as_Field(); 1683 assert(field->is_oop() && field->scalar_replaceable(), "sanity"); 1684 if (field->offset() == Type::OffsetBot) { 1685 jobj->set_scalar_replaceable(false); 1686 return; 1687 } 1688 // 2. An object is not scalar replaceable if the field into which it is 1689 // stored has multiple bases one of which is null. 1690 if (field->base_count() > 1) { 1691 for (BaseIterator i(field); i.has_next(); i.next()) { 1692 PointsToNode* base = i.get(); 1693 if (base == null_obj) { 1694 jobj->set_scalar_replaceable(false); 1695 return; 1696 } 1697 } 1698 } 1699 } 1700 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1701 // 3. An object is not scalar replaceable if it is merged with other objects. 1702 for (EdgeIterator j(use); j.has_next(); j.next()) { 1703 PointsToNode* ptn = j.get(); 1704 if (ptn->is_JavaObject() && ptn != jobj) { 1705 // Mark all objects. 1706 jobj->set_scalar_replaceable(false); 1707 ptn->set_scalar_replaceable(false); 1708 } 1709 } 1710 if (!jobj->scalar_replaceable()) { 1711 return; 1712 } 1713 } 1714 1715 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1716 if (j.get()->is_Arraycopy()) { 1717 continue; 1718 } 1719 1720 // Non-escaping object node should point only to field nodes. 1721 FieldNode* field = j.get()->as_Field(); 1722 int offset = field->as_Field()->offset(); 1723 1724 // 4. An object is not scalar replaceable if it has a field with unknown 1725 // offset (array's element is accessed in loop). 1726 if (offset == Type::OffsetBot) { 1727 jobj->set_scalar_replaceable(false); 1728 return; 1729 } 1730 // 5. Currently an object is not scalar replaceable if a LoadStore node 1731 // access its field since the field value is unknown after it. 1732 // 1733 Node* n = field->ideal_node(); 1734 1735 // Test for an unsafe access that was parsed as maybe off heap 1736 // (with a CheckCastPP to raw memory). 1737 assert(n->is_AddP(), "expect an address computation"); 1738 if (n->in(AddPNode::Base)->is_top() && 1739 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) { 1740 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected"); 1741 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected"); 1742 jobj->set_scalar_replaceable(false); 1743 return; 1744 } 1745 1746 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1747 Node* u = n->fast_out(i); 1748 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) { 1749 jobj->set_scalar_replaceable(false); 1750 return; 1751 } 1752 } 1753 1754 // 6. Or the address may point to more then one object. This may produce 1755 // the false positive result (set not scalar replaceable) 1756 // since the flow-insensitive escape analysis can't separate 1757 // the case when stores overwrite the field's value from the case 1758 // when stores happened on different control branches. 1759 // 1760 // Note: it will disable scalar replacement in some cases: 1761 // 1762 // Point p[] = new Point[1]; 1763 // p[0] = new Point(); // Will be not scalar replaced 1764 // 1765 // but it will save us from incorrect optimizations in next cases: 1766 // 1767 // Point p[] = new Point[1]; 1768 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1769 // 1770 if (field->base_count() > 1) { 1771 for (BaseIterator i(field); i.has_next(); i.next()) { 1772 PointsToNode* base = i.get(); 1773 // Don't take into account LocalVar nodes which 1774 // may point to only one object which should be also 1775 // this field's base by now. 1776 if (base->is_JavaObject() && base != jobj) { 1777 // Mark all bases. 1778 jobj->set_scalar_replaceable(false); 1779 base->set_scalar_replaceable(false); 1780 } 1781 } 1782 } 1783 } 1784 } 1785 1786 #ifdef ASSERT 1787 void ConnectionGraph::verify_connection_graph( 1788 GrowableArray<PointsToNode*>& ptnodes_worklist, 1789 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1790 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1791 GrowableArray<Node*>& addp_worklist) { 1792 // Verify that graph is complete - no new edges could be added. 1793 int java_objects_length = java_objects_worklist.length(); 1794 int non_escaped_length = non_escaped_worklist.length(); 1795 int new_edges = 0; 1796 for (int next = 0; next < java_objects_length; ++next) { 1797 JavaObjectNode* ptn = java_objects_worklist.at(next); 1798 new_edges += add_java_object_edges(ptn, true); 1799 } 1800 assert(new_edges == 0, "graph was not complete"); 1801 // Verify that escape state is final. 1802 int length = non_escaped_worklist.length(); 1803 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1804 assert((non_escaped_length == non_escaped_worklist.length()) && 1805 (non_escaped_length == length) && 1806 (_worklist.length() == 0), "escape state was not final"); 1807 1808 // Verify fields information. 1809 int addp_length = addp_worklist.length(); 1810 for (int next = 0; next < addp_length; ++next ) { 1811 Node* n = addp_worklist.at(next); 1812 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1813 if (field->is_oop()) { 1814 // Verify that field has all bases 1815 Node* base = get_addp_base(n); 1816 PointsToNode* ptn = ptnode_adr(base->_idx); 1817 if (ptn->is_JavaObject()) { 1818 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1819 } else { 1820 assert(ptn->is_LocalVar(), "sanity"); 1821 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1822 PointsToNode* e = i.get(); 1823 if (e->is_JavaObject()) { 1824 assert(field->has_base(e->as_JavaObject()), "sanity"); 1825 } 1826 } 1827 } 1828 // Verify that all fields have initializing values. 1829 if (field->edge_count() == 0) { 1830 tty->print_cr("----------field does not have references----------"); 1831 field->dump(); 1832 for (BaseIterator i(field); i.has_next(); i.next()) { 1833 PointsToNode* base = i.get(); 1834 tty->print_cr("----------field has next base---------------------"); 1835 base->dump(); 1836 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1837 tty->print_cr("----------base has fields-------------------------"); 1838 for (EdgeIterator j(base); j.has_next(); j.next()) { 1839 j.get()->dump(); 1840 } 1841 tty->print_cr("----------base has references---------------------"); 1842 for (UseIterator j(base); j.has_next(); j.next()) { 1843 j.get()->dump(); 1844 } 1845 } 1846 } 1847 for (UseIterator i(field); i.has_next(); i.next()) { 1848 i.get()->dump(); 1849 } 1850 assert(field->edge_count() > 0, "sanity"); 1851 } 1852 } 1853 } 1854 } 1855 #endif 1856 1857 // Optimize ideal graph. 1858 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1859 GrowableArray<Node*>& storestore_worklist) { 1860 Compile* C = _compile; 1861 PhaseIterGVN* igvn = _igvn; 1862 if (EliminateLocks) { 1863 // Mark locks before changing ideal graph. 1864 int cnt = C->macro_count(); 1865 for( int i=0; i < cnt; i++ ) { 1866 Node *n = C->macro_node(i); 1867 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1868 AbstractLockNode* alock = n->as_AbstractLock(); 1869 if (!alock->is_non_esc_obj()) { 1870 if (not_global_escape(alock->obj_node())) { 1871 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1872 // The lock could be marked eliminated by lock coarsening 1873 // code during first IGVN before EA. Replace coarsened flag 1874 // to eliminate all associated locks/unlocks. 1875 #ifdef ASSERT 1876 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1877 #endif 1878 alock->set_non_esc_obj(); 1879 } 1880 } 1881 } 1882 } 1883 } 1884 1885 if (OptimizePtrCompare) { 1886 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1887 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1888 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1889 // Optimize objects compare. 1890 while (ptr_cmp_worklist.length() != 0) { 1891 Node *n = ptr_cmp_worklist.pop(); 1892 Node *res = optimize_ptr_compare(n); 1893 if (res != NULL) { 1894 #ifndef PRODUCT 1895 if (PrintOptimizePtrCompare) { 1896 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")); 1897 if (Verbose) { 1898 n->dump(1); 1899 } 1900 } 1901 #endif 1902 igvn->replace_node(n, res); 1903 } 1904 } 1905 // cleanup 1906 if (_pcmp_neq->outcnt() == 0) 1907 igvn->hash_delete(_pcmp_neq); 1908 if (_pcmp_eq->outcnt() == 0) 1909 igvn->hash_delete(_pcmp_eq); 1910 } 1911 1912 // For MemBarStoreStore nodes added in library_call.cpp, check 1913 // escape status of associated AllocateNode and optimize out 1914 // MemBarStoreStore node if the allocated object never escapes. 1915 while (storestore_worklist.length() != 0) { 1916 Node *n = storestore_worklist.pop(); 1917 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1918 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1919 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1920 if (not_global_escape(alloc)) { 1921 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1922 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1923 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1924 igvn->register_new_node_with_optimizer(mb); 1925 igvn->replace_node(storestore, mb); 1926 } 1927 } 1928 } 1929 1930 // Optimize objects compare. 1931 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1932 assert(OptimizePtrCompare, "sanity"); 1933 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1934 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1935 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1936 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1937 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1938 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1939 1940 // Check simple cases first. 1941 if (jobj1 != NULL) { 1942 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1943 if (jobj1 == jobj2) { 1944 // Comparing the same not escaping object. 1945 return _pcmp_eq; 1946 } 1947 Node* obj = jobj1->ideal_node(); 1948 // Comparing not escaping allocation. 1949 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1950 !ptn2->points_to(jobj1)) { 1951 return _pcmp_neq; // This includes nullness check. 1952 } 1953 } 1954 } 1955 if (jobj2 != NULL) { 1956 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1957 Node* obj = jobj2->ideal_node(); 1958 // Comparing not escaping allocation. 1959 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1960 !ptn1->points_to(jobj2)) { 1961 return _pcmp_neq; // This includes nullness check. 1962 } 1963 } 1964 } 1965 if (jobj1 != NULL && jobj1 != phantom_obj && 1966 jobj2 != NULL && jobj2 != phantom_obj && 1967 jobj1->ideal_node()->is_Con() && 1968 jobj2->ideal_node()->is_Con()) { 1969 // Klass or String constants compare. Need to be careful with 1970 // compressed pointers - compare types of ConN and ConP instead of nodes. 1971 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1972 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1973 if (t1->make_ptr() == t2->make_ptr()) { 1974 return _pcmp_eq; 1975 } else { 1976 return _pcmp_neq; 1977 } 1978 } 1979 if (ptn1->meet(ptn2)) { 1980 return NULL; // Sets are not disjoint 1981 } 1982 1983 // Sets are disjoint. 1984 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1985 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1986 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1987 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1988 if ((set1_has_unknown_ptr && set2_has_null_ptr) || 1989 (set2_has_unknown_ptr && set1_has_null_ptr)) { 1990 // Check nullness of unknown object. 1991 return NULL; 1992 } 1993 1994 // Disjointness by itself is not sufficient since 1995 // alias analysis is not complete for escaped objects. 1996 // Disjoint sets are definitely unrelated only when 1997 // at least one set has only not escaping allocations. 1998 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1999 if (ptn1->non_escaping_allocation()) { 2000 return _pcmp_neq; 2001 } 2002 } 2003 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 2004 if (ptn2->non_escaping_allocation()) { 2005 return _pcmp_neq; 2006 } 2007 } 2008 return NULL; 2009 } 2010 2011 // Connection Graph constuction functions. 2012 2013 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 2014 PointsToNode* ptadr = _nodes.at(n->_idx); 2015 if (ptadr != NULL) { 2016 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 2017 return; 2018 } 2019 Compile* C = _compile; 2020 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 2021 _nodes.at_put(n->_idx, ptadr); 2022 } 2023 2024 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 2025 PointsToNode* ptadr = _nodes.at(n->_idx); 2026 if (ptadr != NULL) { 2027 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2028 return; 2029 } 2030 Compile* C = _compile; 2031 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2032 _nodes.at_put(n->_idx, ptadr); 2033 } 2034 2035 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2036 PointsToNode* ptadr = _nodes.at(n->_idx); 2037 if (ptadr != NULL) { 2038 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2039 return; 2040 } 2041 bool unsafe = false; 2042 bool is_oop = is_oop_field(n, offset, &unsafe); 2043 if (unsafe) { 2044 es = PointsToNode::GlobalEscape; 2045 } 2046 Compile* C = _compile; 2047 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2048 _nodes.at_put(n->_idx, field); 2049 } 2050 2051 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2052 PointsToNode* src, PointsToNode* dst) { 2053 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2054 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2055 PointsToNode* ptadr = _nodes.at(n->_idx); 2056 if (ptadr != NULL) { 2057 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2058 return; 2059 } 2060 Compile* C = _compile; 2061 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2062 _nodes.at_put(n->_idx, ptadr); 2063 // Add edge from arraycopy node to source object. 2064 (void)add_edge(ptadr, src); 2065 src->set_arraycopy_src(); 2066 // Add edge from destination object to arraycopy node. 2067 (void)add_edge(dst, ptadr); 2068 dst->set_arraycopy_dst(); 2069 } 2070 2071 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2072 const Type* adr_type = n->as_AddP()->bottom_type(); 2073 BasicType bt = T_INT; 2074 if (offset == Type::OffsetBot) { 2075 // Check only oop fields. 2076 if (!adr_type->isa_aryptr() || 2077 (adr_type->isa_aryptr()->klass() == NULL) || 2078 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2079 // OffsetBot is used to reference array's element. Ignore first AddP. 2080 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2081 bt = T_OBJECT; 2082 } 2083 } 2084 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2085 if (adr_type->isa_instptr()) { 2086 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2087 if (field != NULL) { 2088 bt = field->layout_type(); 2089 } else { 2090 // Check for unsafe oop field access 2091 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2092 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2093 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2094 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2095 bt = T_OBJECT; 2096 (*unsafe) = true; 2097 } 2098 } 2099 } else if (adr_type->isa_aryptr()) { 2100 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2101 // Ignore array length load. 2102 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2103 // Ignore first AddP. 2104 } else { 2105 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2106 bt = elemtype->array_element_basic_type(); 2107 } 2108 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2109 // Allocation initialization, ThreadLocal field access, unsafe access 2110 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2111 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2112 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2113 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2114 bt = T_OBJECT; 2115 } 2116 } 2117 } 2118 // Note: T_NARROWOOP is not classed as a real reference type 2119 return (is_reference_type(bt) || bt == T_NARROWOOP); 2120 } 2121 2122 // Returns unique pointed java object or NULL. 2123 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2124 assert(!_collecting, "should not call when contructed graph"); 2125 // If the node was created after the escape computation we can't answer. 2126 uint idx = n->_idx; 2127 if (idx >= nodes_size()) { 2128 return NULL; 2129 } 2130 PointsToNode* ptn = ptnode_adr(idx); 2131 if (ptn == NULL) { 2132 return NULL; 2133 } 2134 if (ptn->is_JavaObject()) { 2135 return ptn->as_JavaObject(); 2136 } 2137 assert(ptn->is_LocalVar(), "sanity"); 2138 // Check all java objects it points to. 2139 JavaObjectNode* jobj = NULL; 2140 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2141 PointsToNode* e = i.get(); 2142 if (e->is_JavaObject()) { 2143 if (jobj == NULL) { 2144 jobj = e->as_JavaObject(); 2145 } else if (jobj != e) { 2146 return NULL; 2147 } 2148 } 2149 } 2150 return jobj; 2151 } 2152 2153 // Return true if this node points only to non-escaping allocations. 2154 bool PointsToNode::non_escaping_allocation() { 2155 if (is_JavaObject()) { 2156 Node* n = ideal_node(); 2157 if (n->is_Allocate() || n->is_CallStaticJava()) { 2158 return (escape_state() == PointsToNode::NoEscape); 2159 } else { 2160 return false; 2161 } 2162 } 2163 assert(is_LocalVar(), "sanity"); 2164 // Check all java objects it points to. 2165 for (EdgeIterator i(this); i.has_next(); i.next()) { 2166 PointsToNode* e = i.get(); 2167 if (e->is_JavaObject()) { 2168 Node* n = e->ideal_node(); 2169 if ((e->escape_state() != PointsToNode::NoEscape) || 2170 !(n->is_Allocate() || n->is_CallStaticJava())) { 2171 return false; 2172 } 2173 } 2174 } 2175 return true; 2176 } 2177 2178 // Return true if we know the node does not escape globally. 2179 bool ConnectionGraph::not_global_escape(Node *n) { 2180 assert(!_collecting, "should not call during graph construction"); 2181 // If the node was created after the escape computation we can't answer. 2182 uint idx = n->_idx; 2183 if (idx >= nodes_size()) { 2184 return false; 2185 } 2186 PointsToNode* ptn = ptnode_adr(idx); 2187 if (ptn == NULL) { 2188 return false; // not in congraph (e.g. ConI) 2189 } 2190 PointsToNode::EscapeState es = ptn->escape_state(); 2191 // If we have already computed a value, return it. 2192 if (es >= PointsToNode::GlobalEscape) 2193 return false; 2194 if (ptn->is_JavaObject()) { 2195 return true; // (es < PointsToNode::GlobalEscape); 2196 } 2197 assert(ptn->is_LocalVar(), "sanity"); 2198 // Check all java objects it points to. 2199 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2200 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2201 return false; 2202 } 2203 return true; 2204 } 2205 2206 2207 // Helper functions 2208 2209 // Return true if this node points to specified node or nodes it points to. 2210 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2211 if (is_JavaObject()) { 2212 return (this == ptn); 2213 } 2214 assert(is_LocalVar() || is_Field(), "sanity"); 2215 for (EdgeIterator i(this); i.has_next(); i.next()) { 2216 if (i.get() == ptn) 2217 return true; 2218 } 2219 return false; 2220 } 2221 2222 // Return true if one node points to an other. 2223 bool PointsToNode::meet(PointsToNode* ptn) { 2224 if (this == ptn) { 2225 return true; 2226 } else if (ptn->is_JavaObject()) { 2227 return this->points_to(ptn->as_JavaObject()); 2228 } else if (this->is_JavaObject()) { 2229 return ptn->points_to(this->as_JavaObject()); 2230 } 2231 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2232 int ptn_count = ptn->edge_count(); 2233 for (EdgeIterator i(this); i.has_next(); i.next()) { 2234 PointsToNode* this_e = i.get(); 2235 for (int j = 0; j < ptn_count; j++) { 2236 if (this_e == ptn->edge(j)) 2237 return true; 2238 } 2239 } 2240 return false; 2241 } 2242 2243 #ifdef ASSERT 2244 // Return true if bases point to this java object. 2245 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2246 for (BaseIterator i(this); i.has_next(); i.next()) { 2247 if (i.get() == jobj) 2248 return true; 2249 } 2250 return false; 2251 } 2252 #endif 2253 2254 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2255 const Type *adr_type = phase->type(adr); 2256 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2257 adr->in(AddPNode::Address)->is_Proj() && 2258 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2259 // We are computing a raw address for a store captured by an Initialize 2260 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2261 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2262 assert(offs != Type::OffsetBot || 2263 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2264 "offset must be a constant or it is initialization of array"); 2265 return offs; 2266 } 2267 const TypePtr *t_ptr = adr_type->isa_ptr(); 2268 assert(t_ptr != NULL, "must be a pointer type"); 2269 return t_ptr->offset(); 2270 } 2271 2272 Node* ConnectionGraph::get_addp_base(Node *addp) { 2273 assert(addp->is_AddP(), "must be AddP"); 2274 // 2275 // AddP cases for Base and Address inputs: 2276 // case #1. Direct object's field reference: 2277 // Allocate 2278 // | 2279 // Proj #5 ( oop result ) 2280 // | 2281 // CheckCastPP (cast to instance type) 2282 // | | 2283 // AddP ( base == address ) 2284 // 2285 // case #2. Indirect object's field reference: 2286 // Phi 2287 // | 2288 // CastPP (cast to instance type) 2289 // | | 2290 // AddP ( base == address ) 2291 // 2292 // case #3. Raw object's field reference for Initialize node: 2293 // Allocate 2294 // | 2295 // Proj #5 ( oop result ) 2296 // top | 2297 // \ | 2298 // AddP ( base == top ) 2299 // 2300 // case #4. Array's element reference: 2301 // {CheckCastPP | CastPP} 2302 // | | | 2303 // | AddP ( array's element offset ) 2304 // | | 2305 // AddP ( array's offset ) 2306 // 2307 // case #5. Raw object's field reference for arraycopy stub call: 2308 // The inline_native_clone() case when the arraycopy stub is called 2309 // after the allocation before Initialize and CheckCastPP nodes. 2310 // Allocate 2311 // | 2312 // Proj #5 ( oop result ) 2313 // | | 2314 // AddP ( base == address ) 2315 // 2316 // case #6. Constant Pool, ThreadLocal, CastX2P or 2317 // Raw object's field reference: 2318 // {ConP, ThreadLocal, CastX2P, raw Load} 2319 // top | 2320 // \ | 2321 // AddP ( base == top ) 2322 // 2323 // case #7. Klass's field reference. 2324 // LoadKlass 2325 // | | 2326 // AddP ( base == address ) 2327 // 2328 // case #8. narrow Klass's field reference. 2329 // LoadNKlass 2330 // | 2331 // DecodeN 2332 // | | 2333 // AddP ( base == address ) 2334 // 2335 // case #9. Mixed unsafe access 2336 // {instance} 2337 // | 2338 // CheckCastPP (raw) 2339 // top | 2340 // \ | 2341 // AddP ( base == top ) 2342 // 2343 Node *base = addp->in(AddPNode::Base); 2344 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9. 2345 base = addp->in(AddPNode::Address); 2346 while (base->is_AddP()) { 2347 // Case #6 (unsafe access) may have several chained AddP nodes. 2348 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2349 base = base->in(AddPNode::Address); 2350 } 2351 if (base->Opcode() == Op_CheckCastPP && 2352 base->bottom_type()->isa_rawptr() && 2353 _igvn->type(base->in(1))->isa_oopptr()) { 2354 base = base->in(1); // Case #9 2355 } else { 2356 Node* uncast_base = base->uncast(); 2357 int opcode = uncast_base->Opcode(); 2358 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2359 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2360 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2361 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity"); 2362 } 2363 } 2364 return base; 2365 } 2366 2367 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2368 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2369 Node* addp2 = addp->raw_out(0); 2370 if (addp->outcnt() == 1 && addp2->is_AddP() && 2371 addp2->in(AddPNode::Base) == n && 2372 addp2->in(AddPNode::Address) == addp) { 2373 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2374 // 2375 // Find array's offset to push it on worklist first and 2376 // as result process an array's element offset first (pushed second) 2377 // to avoid CastPP for the array's offset. 2378 // Otherwise the inserted CastPP (LocalVar) will point to what 2379 // the AddP (Field) points to. Which would be wrong since 2380 // the algorithm expects the CastPP has the same point as 2381 // as AddP's base CheckCastPP (LocalVar). 2382 // 2383 // ArrayAllocation 2384 // | 2385 // CheckCastPP 2386 // | 2387 // memProj (from ArrayAllocation CheckCastPP) 2388 // | || 2389 // | || Int (element index) 2390 // | || | ConI (log(element size)) 2391 // | || | / 2392 // | || LShift 2393 // | || / 2394 // | AddP (array's element offset) 2395 // | | 2396 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2397 // | / / 2398 // AddP (array's offset) 2399 // | 2400 // Load/Store (memory operation on array's element) 2401 // 2402 return addp2; 2403 } 2404 return NULL; 2405 } 2406 2407 // 2408 // Adjust the type and inputs of an AddP which computes the 2409 // address of a field of an instance 2410 // 2411 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2412 PhaseGVN* igvn = _igvn; 2413 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2414 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2415 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2416 if (t == NULL) { 2417 // We are computing a raw address for a store captured by an Initialize 2418 // compute an appropriate address type (cases #3 and #5). 2419 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2420 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2421 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2422 assert(offs != Type::OffsetBot, "offset must be a constant"); 2423 t = base_t->add_offset(offs)->is_oopptr(); 2424 } 2425 int inst_id = base_t->instance_id(); 2426 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2427 "old type must be non-instance or match new type"); 2428 2429 // The type 't' could be subclass of 'base_t'. 2430 // As result t->offset() could be large then base_t's size and it will 2431 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2432 // constructor verifies correctness of the offset. 2433 // 2434 // It could happened on subclass's branch (from the type profiling 2435 // inlining) which was not eliminated during parsing since the exactness 2436 // of the allocation type was not propagated to the subclass type check. 2437 // 2438 // Or the type 't' could be not related to 'base_t' at all. 2439 // It could happened when CHA type is different from MDO type on a dead path 2440 // (for example, from instanceof check) which is not collapsed during parsing. 2441 // 2442 // Do nothing for such AddP node and don't process its users since 2443 // this code branch will go away. 2444 // 2445 if (!t->is_known_instance() && 2446 !base_t->klass()->is_subtype_of(t->klass())) { 2447 return false; // bail out 2448 } 2449 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2450 // Do NOT remove the next line: ensure a new alias index is allocated 2451 // for the instance type. Note: C++ will not remove it since the call 2452 // has side effect. 2453 int alias_idx = _compile->get_alias_index(tinst); 2454 igvn->set_type(addp, tinst); 2455 // record the allocation in the node map 2456 set_map(addp, get_map(base->_idx)); 2457 // Set addp's Base and Address to 'base'. 2458 Node *abase = addp->in(AddPNode::Base); 2459 Node *adr = addp->in(AddPNode::Address); 2460 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2461 adr->in(0)->_idx == (uint)inst_id) { 2462 // Skip AddP cases #3 and #5. 2463 } else { 2464 assert(!abase->is_top(), "sanity"); // AddP case #3 2465 if (abase != base) { 2466 igvn->hash_delete(addp); 2467 addp->set_req(AddPNode::Base, base); 2468 if (abase == adr) { 2469 addp->set_req(AddPNode::Address, base); 2470 } else { 2471 // AddP case #4 (adr is array's element offset AddP node) 2472 #ifdef ASSERT 2473 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2474 assert(adr->is_AddP() && atype != NULL && 2475 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2476 #endif 2477 } 2478 igvn->hash_insert(addp); 2479 } 2480 } 2481 // Put on IGVN worklist since at least addp's type was changed above. 2482 record_for_optimizer(addp); 2483 return true; 2484 } 2485 2486 // 2487 // Create a new version of orig_phi if necessary. Returns either the newly 2488 // created phi or an existing phi. Sets create_new to indicate whether a new 2489 // phi was created. Cache the last newly created phi in the node map. 2490 // 2491 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2492 Compile *C = _compile; 2493 PhaseGVN* igvn = _igvn; 2494 new_created = false; 2495 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2496 // nothing to do if orig_phi is bottom memory or matches alias_idx 2497 if (phi_alias_idx == alias_idx) { 2498 return orig_phi; 2499 } 2500 // Have we recently created a Phi for this alias index? 2501 PhiNode *result = get_map_phi(orig_phi->_idx); 2502 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2503 return result; 2504 } 2505 // Previous check may fail when the same wide memory Phi was split into Phis 2506 // for different memory slices. Search all Phis for this region. 2507 if (result != NULL) { 2508 Node* region = orig_phi->in(0); 2509 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2510 Node* phi = region->fast_out(i); 2511 if (phi->is_Phi() && 2512 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2513 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2514 return phi->as_Phi(); 2515 } 2516 } 2517 } 2518 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2519 if (C->do_escape_analysis() == true && !C->failing()) { 2520 // Retry compilation without escape analysis. 2521 // If this is the first failure, the sentinel string will "stick" 2522 // to the Compile object, and the C2Compiler will see it and retry. 2523 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2524 } 2525 return NULL; 2526 } 2527 orig_phi_worklist.append_if_missing(orig_phi); 2528 const TypePtr *atype = C->get_adr_type(alias_idx); 2529 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2530 C->copy_node_notes_to(result, orig_phi); 2531 igvn->set_type(result, result->bottom_type()); 2532 record_for_optimizer(result); 2533 set_map(orig_phi, result); 2534 new_created = true; 2535 return result; 2536 } 2537 2538 // 2539 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2540 // specified alias index. 2541 // 2542 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2543 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2544 Compile *C = _compile; 2545 PhaseGVN* igvn = _igvn; 2546 bool new_phi_created; 2547 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2548 if (!new_phi_created) { 2549 return result; 2550 } 2551 GrowableArray<PhiNode *> phi_list; 2552 GrowableArray<uint> cur_input; 2553 PhiNode *phi = orig_phi; 2554 uint idx = 1; 2555 bool finished = false; 2556 while(!finished) { 2557 while (idx < phi->req()) { 2558 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2559 if (mem != NULL && mem->is_Phi()) { 2560 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2561 if (new_phi_created) { 2562 // found an phi for which we created a new split, push current one on worklist and begin 2563 // processing new one 2564 phi_list.push(phi); 2565 cur_input.push(idx); 2566 phi = mem->as_Phi(); 2567 result = newphi; 2568 idx = 1; 2569 continue; 2570 } else { 2571 mem = newphi; 2572 } 2573 } 2574 if (C->failing()) { 2575 return NULL; 2576 } 2577 result->set_req(idx++, mem); 2578 } 2579 #ifdef ASSERT 2580 // verify that the new Phi has an input for each input of the original 2581 assert( phi->req() == result->req(), "must have same number of inputs."); 2582 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2583 #endif 2584 // Check if all new phi's inputs have specified alias index. 2585 // Otherwise use old phi. 2586 for (uint i = 1; i < phi->req(); i++) { 2587 Node* in = result->in(i); 2588 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2589 } 2590 // we have finished processing a Phi, see if there are any more to do 2591 finished = (phi_list.length() == 0 ); 2592 if (!finished) { 2593 phi = phi_list.pop(); 2594 idx = cur_input.pop(); 2595 PhiNode *prev_result = get_map_phi(phi->_idx); 2596 prev_result->set_req(idx++, result); 2597 result = prev_result; 2598 } 2599 } 2600 return result; 2601 } 2602 2603 // 2604 // The next methods are derived from methods in MemNode. 2605 // 2606 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2607 Node *mem = mmem; 2608 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2609 // means an array I have not precisely typed yet. Do not do any 2610 // alias stuff with it any time soon. 2611 if (toop->base() != Type::AnyPtr && 2612 !(toop->klass() != NULL && 2613 toop->klass()->is_java_lang_Object() && 2614 toop->offset() == Type::OffsetBot)) { 2615 mem = mmem->memory_at(alias_idx); 2616 // Update input if it is progress over what we have now 2617 } 2618 return mem; 2619 } 2620 2621 // 2622 // Move memory users to their memory slices. 2623 // 2624 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2625 Compile* C = _compile; 2626 PhaseGVN* igvn = _igvn; 2627 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2628 assert(tp != NULL, "ptr type"); 2629 int alias_idx = C->get_alias_index(tp); 2630 int general_idx = C->get_general_index(alias_idx); 2631 2632 // Move users first 2633 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2634 Node* use = n->fast_out(i); 2635 if (use->is_MergeMem()) { 2636 MergeMemNode* mmem = use->as_MergeMem(); 2637 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2638 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2639 continue; // Nothing to do 2640 } 2641 // Replace previous general reference to mem node. 2642 uint orig_uniq = C->unique(); 2643 Node* m = find_inst_mem(n, general_idx, orig_phis); 2644 assert(orig_uniq == C->unique(), "no new nodes"); 2645 mmem->set_memory_at(general_idx, m); 2646 --imax; 2647 --i; 2648 } else if (use->is_MemBar()) { 2649 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2650 if (use->req() > MemBarNode::Precedent && 2651 use->in(MemBarNode::Precedent) == n) { 2652 // Don't move related membars. 2653 record_for_optimizer(use); 2654 continue; 2655 } 2656 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2657 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) || 2658 alias_idx == general_idx) { 2659 continue; // Nothing to do 2660 } 2661 // Move to general memory slice. 2662 uint orig_uniq = C->unique(); 2663 Node* m = find_inst_mem(n, general_idx, orig_phis); 2664 assert(orig_uniq == C->unique(), "no new nodes"); 2665 igvn->hash_delete(use); 2666 imax -= use->replace_edge(n, m); 2667 igvn->hash_insert(use); 2668 record_for_optimizer(use); 2669 --i; 2670 #ifdef ASSERT 2671 } else if (use->is_Mem()) { 2672 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2673 // Don't move related cardmark. 2674 continue; 2675 } 2676 // Memory nodes should have new memory input. 2677 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2678 assert(tp != NULL, "ptr type"); 2679 int idx = C->get_alias_index(tp); 2680 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2681 "Following memory nodes should have new memory input or be on the same memory slice"); 2682 } else if (use->is_Phi()) { 2683 // Phi nodes should be split and moved already. 2684 tp = use->as_Phi()->adr_type()->isa_ptr(); 2685 assert(tp != NULL, "ptr type"); 2686 int idx = C->get_alias_index(tp); 2687 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2688 } else { 2689 use->dump(); 2690 assert(false, "should not be here"); 2691 #endif 2692 } 2693 } 2694 } 2695 2696 // 2697 // Search memory chain of "mem" to find a MemNode whose address 2698 // is the specified alias index. 2699 // 2700 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2701 if (orig_mem == NULL) 2702 return orig_mem; 2703 Compile* C = _compile; 2704 PhaseGVN* igvn = _igvn; 2705 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2706 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2707 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 2708 Node *prev = NULL; 2709 Node *result = orig_mem; 2710 while (prev != result) { 2711 prev = result; 2712 if (result == start_mem) 2713 break; // hit one of our sentinels 2714 if (result->is_Mem()) { 2715 const Type *at = igvn->type(result->in(MemNode::Address)); 2716 if (at == Type::TOP) 2717 break; // Dead 2718 assert (at->isa_ptr() != NULL, "pointer type required."); 2719 int idx = C->get_alias_index(at->is_ptr()); 2720 if (idx == alias_idx) 2721 break; // Found 2722 if (!is_instance && (at->isa_oopptr() == NULL || 2723 !at->is_oopptr()->is_known_instance())) { 2724 break; // Do not skip store to general memory slice. 2725 } 2726 result = result->in(MemNode::Memory); 2727 } 2728 if (!is_instance) 2729 continue; // don't search further for non-instance types 2730 // skip over a call which does not affect this memory slice 2731 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2732 Node *proj_in = result->in(0); 2733 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2734 break; // hit one of our sentinels 2735 } else if (proj_in->is_Call()) { 2736 // ArrayCopy node processed here as well 2737 CallNode *call = proj_in->as_Call(); 2738 if (!call->may_modify(toop, igvn)) { 2739 result = call->in(TypeFunc::Memory); 2740 } 2741 } else if (proj_in->is_Initialize()) { 2742 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2743 // Stop if this is the initialization for the object instance which 2744 // which contains this memory slice, otherwise skip over it. 2745 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2746 result = proj_in->in(TypeFunc::Memory); 2747 } 2748 } else if (proj_in->is_MemBar()) { 2749 // Check if there is an array copy for a clone 2750 // Step over GC barrier when ReduceInitialCardMarks is disabled 2751 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2752 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0)); 2753 2754 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) { 2755 // Stop if it is a clone 2756 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy(); 2757 if (ac->may_modify(toop, igvn)) { 2758 break; 2759 } 2760 } 2761 result = proj_in->in(TypeFunc::Memory); 2762 } 2763 } else if (result->is_MergeMem()) { 2764 MergeMemNode *mmem = result->as_MergeMem(); 2765 result = step_through_mergemem(mmem, alias_idx, toop); 2766 if (result == mmem->base_memory()) { 2767 // Didn't find instance memory, search through general slice recursively. 2768 result = mmem->memory_at(C->get_general_index(alias_idx)); 2769 result = find_inst_mem(result, alias_idx, orig_phis); 2770 if (C->failing()) { 2771 return NULL; 2772 } 2773 mmem->set_memory_at(alias_idx, result); 2774 } 2775 } else if (result->is_Phi() && 2776 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2777 Node *un = result->as_Phi()->unique_input(igvn); 2778 if (un != NULL) { 2779 orig_phis.append_if_missing(result->as_Phi()); 2780 result = un; 2781 } else { 2782 break; 2783 } 2784 } else if (result->is_ClearArray()) { 2785 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2786 // Can not bypass initialization of the instance 2787 // we are looking for. 2788 break; 2789 } 2790 // Otherwise skip it (the call updated 'result' value). 2791 } else if (result->Opcode() == Op_SCMemProj) { 2792 Node* mem = result->in(0); 2793 Node* adr = NULL; 2794 if (mem->is_LoadStore()) { 2795 adr = mem->in(MemNode::Address); 2796 } else { 2797 assert(mem->Opcode() == Op_EncodeISOArray || 2798 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 2799 adr = mem->in(3); // Memory edge corresponds to destination array 2800 } 2801 const Type *at = igvn->type(adr); 2802 if (at != Type::TOP) { 2803 assert(at->isa_ptr() != NULL, "pointer type required."); 2804 int idx = C->get_alias_index(at->is_ptr()); 2805 if (idx == alias_idx) { 2806 // Assert in debug mode 2807 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 2808 break; // In product mode return SCMemProj node 2809 } 2810 } 2811 result = mem->in(MemNode::Memory); 2812 } else if (result->Opcode() == Op_StrInflatedCopy) { 2813 Node* adr = result->in(3); // Memory edge corresponds to destination array 2814 const Type *at = igvn->type(adr); 2815 if (at != Type::TOP) { 2816 assert(at->isa_ptr() != NULL, "pointer type required."); 2817 int idx = C->get_alias_index(at->is_ptr()); 2818 if (idx == alias_idx) { 2819 // Assert in debug mode 2820 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 2821 break; // In product mode return SCMemProj node 2822 } 2823 } 2824 result = result->in(MemNode::Memory); 2825 } 2826 } 2827 if (result->is_Phi()) { 2828 PhiNode *mphi = result->as_Phi(); 2829 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2830 const TypePtr *t = mphi->adr_type(); 2831 if (!is_instance) { 2832 // Push all non-instance Phis on the orig_phis worklist to update inputs 2833 // during Phase 4 if needed. 2834 orig_phis.append_if_missing(mphi); 2835 } else if (C->get_alias_index(t) != alias_idx) { 2836 // Create a new Phi with the specified alias index type. 2837 result = split_memory_phi(mphi, alias_idx, orig_phis); 2838 } 2839 } 2840 // the result is either MemNode, PhiNode, InitializeNode. 2841 return result; 2842 } 2843 2844 // 2845 // Convert the types of unescaped object to instance types where possible, 2846 // propagate the new type information through the graph, and update memory 2847 // edges and MergeMem inputs to reflect the new type. 2848 // 2849 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2850 // The processing is done in 4 phases: 2851 // 2852 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2853 // types for the CheckCastPP for allocations where possible. 2854 // Propagate the new types through users as follows: 2855 // casts and Phi: push users on alloc_worklist 2856 // AddP: cast Base and Address inputs to the instance type 2857 // push any AddP users on alloc_worklist and push any memnode 2858 // users onto memnode_worklist. 2859 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2860 // search the Memory chain for a store with the appropriate type 2861 // address type. If a Phi is found, create a new version with 2862 // the appropriate memory slices from each of the Phi inputs. 2863 // For stores, process the users as follows: 2864 // MemNode: push on memnode_worklist 2865 // MergeMem: push on mergemem_worklist 2866 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2867 // moving the first node encountered of each instance type to the 2868 // the input corresponding to its alias index. 2869 // appropriate memory slice. 2870 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2871 // 2872 // In the following example, the CheckCastPP nodes are the cast of allocation 2873 // results and the allocation of node 29 is unescaped and eligible to be an 2874 // instance type. 2875 // 2876 // We start with: 2877 // 2878 // 7 Parm #memory 2879 // 10 ConI "12" 2880 // 19 CheckCastPP "Foo" 2881 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2882 // 29 CheckCastPP "Foo" 2883 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2884 // 2885 // 40 StoreP 25 7 20 ... alias_index=4 2886 // 50 StoreP 35 40 30 ... alias_index=4 2887 // 60 StoreP 45 50 20 ... alias_index=4 2888 // 70 LoadP _ 60 30 ... alias_index=4 2889 // 80 Phi 75 50 60 Memory alias_index=4 2890 // 90 LoadP _ 80 30 ... alias_index=4 2891 // 100 LoadP _ 80 20 ... alias_index=4 2892 // 2893 // 2894 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2895 // and creating a new alias index for node 30. This gives: 2896 // 2897 // 7 Parm #memory 2898 // 10 ConI "12" 2899 // 19 CheckCastPP "Foo" 2900 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2901 // 29 CheckCastPP "Foo" iid=24 2902 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2903 // 2904 // 40 StoreP 25 7 20 ... alias_index=4 2905 // 50 StoreP 35 40 30 ... alias_index=6 2906 // 60 StoreP 45 50 20 ... alias_index=4 2907 // 70 LoadP _ 60 30 ... alias_index=6 2908 // 80 Phi 75 50 60 Memory alias_index=4 2909 // 90 LoadP _ 80 30 ... alias_index=6 2910 // 100 LoadP _ 80 20 ... alias_index=4 2911 // 2912 // In phase 2, new memory inputs are computed for the loads and stores, 2913 // And a new version of the phi is created. In phase 4, the inputs to 2914 // node 80 are updated and then the memory nodes are updated with the 2915 // values computed in phase 2. This results in: 2916 // 2917 // 7 Parm #memory 2918 // 10 ConI "12" 2919 // 19 CheckCastPP "Foo" 2920 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2921 // 29 CheckCastPP "Foo" iid=24 2922 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2923 // 2924 // 40 StoreP 25 7 20 ... alias_index=4 2925 // 50 StoreP 35 7 30 ... alias_index=6 2926 // 60 StoreP 45 40 20 ... alias_index=4 2927 // 70 LoadP _ 50 30 ... alias_index=6 2928 // 80 Phi 75 40 60 Memory alias_index=4 2929 // 120 Phi 75 50 50 Memory alias_index=6 2930 // 90 LoadP _ 120 30 ... alias_index=6 2931 // 100 LoadP _ 80 20 ... alias_index=4 2932 // 2933 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 2934 GrowableArray<Node *> memnode_worklist; 2935 GrowableArray<PhiNode *> orig_phis; 2936 PhaseIterGVN *igvn = _igvn; 2937 uint new_index_start = (uint) _compile->num_alias_types(); 2938 VectorSet visited; 2939 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2940 uint unique_old = _compile->unique(); 2941 2942 // Phase 1: Process possible allocations from alloc_worklist. 2943 // Create instance types for the CheckCastPP for allocations where possible. 2944 // 2945 // (Note: don't forget to change the order of the second AddP node on 2946 // the alloc_worklist if the order of the worklist processing is changed, 2947 // see the comment in find_second_addp().) 2948 // 2949 while (alloc_worklist.length() != 0) { 2950 Node *n = alloc_worklist.pop(); 2951 uint ni = n->_idx; 2952 if (n->is_Call()) { 2953 CallNode *alloc = n->as_Call(); 2954 // copy escape information to call node 2955 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2956 PointsToNode::EscapeState es = ptn->escape_state(); 2957 // We have an allocation or call which returns a Java object, 2958 // see if it is unescaped. 2959 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2960 continue; 2961 // Find CheckCastPP for the allocate or for the return value of a call 2962 n = alloc->result_cast(); 2963 if (n == NULL) { // No uses except Initialize node 2964 if (alloc->is_Allocate()) { 2965 // Set the scalar_replaceable flag for allocation 2966 // so it could be eliminated if it has no uses. 2967 alloc->as_Allocate()->_is_scalar_replaceable = true; 2968 } 2969 if (alloc->is_CallStaticJava()) { 2970 // Set the scalar_replaceable flag for boxing method 2971 // so it could be eliminated if it has no uses. 2972 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2973 } 2974 continue; 2975 } 2976 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2977 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2978 continue; 2979 } 2980 2981 // The inline code for Object.clone() casts the allocation result to 2982 // java.lang.Object and then to the actual type of the allocated 2983 // object. Detect this case and use the second cast. 2984 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2985 // the allocation result is cast to java.lang.Object and then 2986 // to the actual Array type. 2987 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2988 && (alloc->is_AllocateArray() || 2989 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2990 Node *cast2 = NULL; 2991 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2992 Node *use = n->fast_out(i); 2993 if (use->is_CheckCastPP()) { 2994 cast2 = use; 2995 break; 2996 } 2997 } 2998 if (cast2 != NULL) { 2999 n = cast2; 3000 } else { 3001 // Non-scalar replaceable if the allocation type is unknown statically 3002 // (reflection allocation), the object can't be restored during 3003 // deoptimization without precise type. 3004 continue; 3005 } 3006 } 3007 3008 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 3009 if (t == NULL) 3010 continue; // not a TypeOopPtr 3011 if (!t->klass_is_exact()) 3012 continue; // not an unique type 3013 3014 if (alloc->is_Allocate()) { 3015 // Set the scalar_replaceable flag for allocation 3016 // so it could be eliminated. 3017 alloc->as_Allocate()->_is_scalar_replaceable = true; 3018 } 3019 if (alloc->is_CallStaticJava()) { 3020 // Set the scalar_replaceable flag for boxing method 3021 // so it could be eliminated. 3022 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3023 } 3024 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 3025 // in order for an object to be scalar-replaceable, it must be: 3026 // - a direct allocation (not a call returning an object) 3027 // - non-escaping 3028 // - eligible to be a unique type 3029 // - not determined to be ineligible by escape analysis 3030 set_map(alloc, n); 3031 set_map(n, alloc); 3032 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 3033 igvn->hash_delete(n); 3034 igvn->set_type(n, tinst); 3035 n->raise_bottom_type(tinst); 3036 igvn->hash_insert(n); 3037 record_for_optimizer(n); 3038 // Allocate an alias index for the header fields. Accesses to 3039 // the header emitted during macro expansion wouldn't have 3040 // correct memory state otherwise. 3041 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes())); 3042 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes())); 3043 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 3044 3045 // First, put on the worklist all Field edges from Connection Graph 3046 // which is more accurate than putting immediate users from Ideal Graph. 3047 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 3048 PointsToNode* tgt = e.get(); 3049 if (tgt->is_Arraycopy()) { 3050 continue; 3051 } 3052 Node* use = tgt->ideal_node(); 3053 assert(tgt->is_Field() && use->is_AddP(), 3054 "only AddP nodes are Field edges in CG"); 3055 if (use->outcnt() > 0) { // Don't process dead nodes 3056 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 3057 if (addp2 != NULL) { 3058 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3059 alloc_worklist.append_if_missing(addp2); 3060 } 3061 alloc_worklist.append_if_missing(use); 3062 } 3063 } 3064 3065 // An allocation may have an Initialize which has raw stores. Scan 3066 // the users of the raw allocation result and push AddP users 3067 // on alloc_worklist. 3068 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms); 3069 assert (raw_result != NULL, "must have an allocation result"); 3070 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 3071 Node *use = raw_result->fast_out(i); 3072 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3073 Node* addp2 = find_second_addp(use, raw_result); 3074 if (addp2 != NULL) { 3075 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3076 alloc_worklist.append_if_missing(addp2); 3077 } 3078 alloc_worklist.append_if_missing(use); 3079 } else if (use->is_MemBar()) { 3080 memnode_worklist.append_if_missing(use); 3081 } 3082 } 3083 } 3084 } else if (n->is_AddP()) { 3085 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3086 if (jobj == NULL || jobj == phantom_obj) { 3087 #ifdef ASSERT 3088 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3089 ptnode_adr(n->_idx)->dump(); 3090 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3091 #endif 3092 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3093 return; 3094 } 3095 Node *base = get_map(jobj->idx()); // CheckCastPP node 3096 if (!split_AddP(n, base)) continue; // wrong type from dead path 3097 } else if (n->is_Phi() || 3098 n->is_CheckCastPP() || 3099 n->is_EncodeP() || 3100 n->is_DecodeN() || 3101 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3102 if (visited.test_set(n->_idx)) { 3103 assert(n->is_Phi(), "loops only through Phi's"); 3104 continue; // already processed 3105 } 3106 JavaObjectNode* jobj = unique_java_object(n); 3107 if (jobj == NULL || jobj == phantom_obj) { 3108 #ifdef ASSERT 3109 ptnode_adr(n->_idx)->dump(); 3110 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3111 #endif 3112 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3113 return; 3114 } else { 3115 Node *val = get_map(jobj->idx()); // CheckCastPP node 3116 TypeNode *tn = n->as_Type(); 3117 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3118 assert(tinst != NULL && tinst->is_known_instance() && 3119 tinst->instance_id() == jobj->idx() , "instance type expected."); 3120 3121 const Type *tn_type = igvn->type(tn); 3122 const TypeOopPtr *tn_t; 3123 if (tn_type->isa_narrowoop()) { 3124 tn_t = tn_type->make_ptr()->isa_oopptr(); 3125 } else { 3126 tn_t = tn_type->isa_oopptr(); 3127 } 3128 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3129 if (tn_type->isa_narrowoop()) { 3130 tn_type = tinst->make_narrowoop(); 3131 } else { 3132 tn_type = tinst; 3133 } 3134 igvn->hash_delete(tn); 3135 igvn->set_type(tn, tn_type); 3136 tn->set_type(tn_type); 3137 igvn->hash_insert(tn); 3138 record_for_optimizer(n); 3139 } else { 3140 assert(tn_type == TypePtr::NULL_PTR || 3141 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3142 "unexpected type"); 3143 continue; // Skip dead path with different type 3144 } 3145 } 3146 } else { 3147 debug_only(n->dump();) 3148 assert(false, "EA: unexpected node"); 3149 continue; 3150 } 3151 // push allocation's users on appropriate worklist 3152 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3153 Node *use = n->fast_out(i); 3154 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3155 // Load/store to instance's field 3156 memnode_worklist.append_if_missing(use); 3157 } else if (use->is_MemBar()) { 3158 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3159 memnode_worklist.append_if_missing(use); 3160 } 3161 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3162 Node* addp2 = find_second_addp(use, n); 3163 if (addp2 != NULL) { 3164 alloc_worklist.append_if_missing(addp2); 3165 } 3166 alloc_worklist.append_if_missing(use); 3167 } else if (use->is_Phi() || 3168 use->is_CheckCastPP() || 3169 use->is_EncodeNarrowPtr() || 3170 use->is_DecodeNarrowPtr() || 3171 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3172 alloc_worklist.append_if_missing(use); 3173 #ifdef ASSERT 3174 } else if (use->is_Mem()) { 3175 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3176 } else if (use->is_MergeMem()) { 3177 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3178 } else if (use->is_SafePoint()) { 3179 // Look for MergeMem nodes for calls which reference unique allocation 3180 // (through CheckCastPP nodes) even for debug info. 3181 Node* m = use->in(TypeFunc::Memory); 3182 if (m->is_MergeMem()) { 3183 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3184 } 3185 } else if (use->Opcode() == Op_EncodeISOArray) { 3186 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3187 // EncodeISOArray overwrites destination array 3188 memnode_worklist.append_if_missing(use); 3189 } 3190 } else { 3191 uint op = use->Opcode(); 3192 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) && 3193 (use->in(MemNode::Memory) == n)) { 3194 // They overwrite memory edge corresponding to destination array, 3195 memnode_worklist.append_if_missing(use); 3196 } else if (!(op == Op_CmpP || op == Op_Conv2B || 3197 op == Op_CastP2X || op == Op_StoreCM || 3198 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3199 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3200 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || 3201 op == Op_SubTypeCheck || 3202 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) { 3203 n->dump(); 3204 use->dump(); 3205 assert(false, "EA: missing allocation reference path"); 3206 } 3207 #endif 3208 } 3209 } 3210 3211 } 3212 3213 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3214 // type, record it in the ArrayCopy node so we know what memory this 3215 // node uses/modified. 3216 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3217 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3218 Node* dest = ac->in(ArrayCopyNode::Dest); 3219 if (dest->is_AddP()) { 3220 dest = get_addp_base(dest); 3221 } 3222 JavaObjectNode* jobj = unique_java_object(dest); 3223 if (jobj != NULL) { 3224 Node *base = get_map(jobj->idx()); 3225 if (base != NULL) { 3226 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3227 ac->_dest_type = base_t; 3228 } 3229 } 3230 Node* src = ac->in(ArrayCopyNode::Src); 3231 if (src->is_AddP()) { 3232 src = get_addp_base(src); 3233 } 3234 jobj = unique_java_object(src); 3235 if (jobj != NULL) { 3236 Node* base = get_map(jobj->idx()); 3237 if (base != NULL) { 3238 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3239 ac->_src_type = base_t; 3240 } 3241 } 3242 } 3243 3244 // New alias types were created in split_AddP(). 3245 uint new_index_end = (uint) _compile->num_alias_types(); 3246 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3247 3248 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3249 // compute new values for Memory inputs (the Memory inputs are not 3250 // actually updated until phase 4.) 3251 if (memnode_worklist.length() == 0) 3252 return; // nothing to do 3253 while (memnode_worklist.length() != 0) { 3254 Node *n = memnode_worklist.pop(); 3255 if (visited.test_set(n->_idx)) 3256 continue; 3257 if (n->is_Phi() || n->is_ClearArray()) { 3258 // we don't need to do anything, but the users must be pushed 3259 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3260 // we don't need to do anything, but the users must be pushed 3261 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory); 3262 if (n == NULL) 3263 continue; 3264 } else if (n->Opcode() == Op_StrCompressedCopy || 3265 n->Opcode() == Op_EncodeISOArray) { 3266 // get the memory projection 3267 n = n->find_out_with(Op_SCMemProj); 3268 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3269 } else { 3270 assert(n->is_Mem(), "memory node required."); 3271 Node *addr = n->in(MemNode::Address); 3272 const Type *addr_t = igvn->type(addr); 3273 if (addr_t == Type::TOP) 3274 continue; 3275 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3276 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3277 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3278 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3279 if (_compile->failing()) { 3280 return; 3281 } 3282 if (mem != n->in(MemNode::Memory)) { 3283 // We delay the memory edge update since we need old one in 3284 // MergeMem code below when instances memory slices are separated. 3285 set_map(n, mem); 3286 } 3287 if (n->is_Load()) { 3288 continue; // don't push users 3289 } else if (n->is_LoadStore()) { 3290 // get the memory projection 3291 n = n->find_out_with(Op_SCMemProj); 3292 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3293 } 3294 } 3295 // push user on appropriate worklist 3296 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3297 Node *use = n->fast_out(i); 3298 if (use->is_Phi() || use->is_ClearArray()) { 3299 memnode_worklist.append_if_missing(use); 3300 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3301 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3302 continue; 3303 memnode_worklist.append_if_missing(use); 3304 } else if (use->is_MemBar()) { 3305 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3306 memnode_worklist.append_if_missing(use); 3307 } 3308 #ifdef ASSERT 3309 } else if(use->is_Mem()) { 3310 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3311 } else if (use->is_MergeMem()) { 3312 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3313 } else if (use->Opcode() == Op_EncodeISOArray) { 3314 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3315 // EncodeISOArray overwrites destination array 3316 memnode_worklist.append_if_missing(use); 3317 } 3318 } else { 3319 uint op = use->Opcode(); 3320 if ((use->in(MemNode::Memory) == n) && 3321 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) { 3322 // They overwrite memory edge corresponding to destination array, 3323 memnode_worklist.append_if_missing(use); 3324 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) || 3325 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3326 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3327 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3328 n->dump(); 3329 use->dump(); 3330 assert(false, "EA: missing memory path"); 3331 } 3332 #endif 3333 } 3334 } 3335 } 3336 3337 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3338 // Walk each memory slice moving the first node encountered of each 3339 // instance type to the the input corresponding to its alias index. 3340 uint length = _mergemem_worklist.length(); 3341 for( uint next = 0; next < length; ++next ) { 3342 MergeMemNode* nmm = _mergemem_worklist.at(next); 3343 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3344 // Note: we don't want to use MergeMemStream here because we only want to 3345 // scan inputs which exist at the start, not ones we add during processing. 3346 // Note 2: MergeMem may already contains instance memory slices added 3347 // during find_inst_mem() call when memory nodes were processed above. 3348 igvn->hash_delete(nmm); 3349 uint nslices = MIN2(nmm->req(), new_index_start); 3350 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3351 Node* mem = nmm->in(i); 3352 Node* cur = NULL; 3353 if (mem == NULL || mem->is_top()) 3354 continue; 3355 // First, update mergemem by moving memory nodes to corresponding slices 3356 // if their type became more precise since this mergemem was created. 3357 while (mem->is_Mem()) { 3358 const Type *at = igvn->type(mem->in(MemNode::Address)); 3359 if (at != Type::TOP) { 3360 assert (at->isa_ptr() != NULL, "pointer type required."); 3361 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3362 if (idx == i) { 3363 if (cur == NULL) 3364 cur = mem; 3365 } else { 3366 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3367 nmm->set_memory_at(idx, mem); 3368 } 3369 } 3370 } 3371 mem = mem->in(MemNode::Memory); 3372 } 3373 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3374 // Find any instance of the current type if we haven't encountered 3375 // already a memory slice of the instance along the memory chain. 3376 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3377 if((uint)_compile->get_general_index(ni) == i) { 3378 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3379 if (nmm->is_empty_memory(m)) { 3380 Node* result = find_inst_mem(mem, ni, orig_phis); 3381 if (_compile->failing()) { 3382 return; 3383 } 3384 nmm->set_memory_at(ni, result); 3385 } 3386 } 3387 } 3388 } 3389 // Find the rest of instances values 3390 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3391 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3392 Node* result = step_through_mergemem(nmm, ni, tinst); 3393 if (result == nmm->base_memory()) { 3394 // Didn't find instance memory, search through general slice recursively. 3395 result = nmm->memory_at(_compile->get_general_index(ni)); 3396 result = find_inst_mem(result, ni, orig_phis); 3397 if (_compile->failing()) { 3398 return; 3399 } 3400 nmm->set_memory_at(ni, result); 3401 } 3402 } 3403 igvn->hash_insert(nmm); 3404 record_for_optimizer(nmm); 3405 } 3406 3407 // Phase 4: Update the inputs of non-instance memory Phis and 3408 // the Memory input of memnodes 3409 // First update the inputs of any non-instance Phi's from 3410 // which we split out an instance Phi. Note we don't have 3411 // to recursively process Phi's encounted on the input memory 3412 // chains as is done in split_memory_phi() since they will 3413 // also be processed here. 3414 for (int j = 0; j < orig_phis.length(); j++) { 3415 PhiNode *phi = orig_phis.at(j); 3416 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3417 igvn->hash_delete(phi); 3418 for (uint i = 1; i < phi->req(); i++) { 3419 Node *mem = phi->in(i); 3420 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3421 if (_compile->failing()) { 3422 return; 3423 } 3424 if (mem != new_mem) { 3425 phi->set_req(i, new_mem); 3426 } 3427 } 3428 igvn->hash_insert(phi); 3429 record_for_optimizer(phi); 3430 } 3431 3432 // Update the memory inputs of MemNodes with the value we computed 3433 // in Phase 2 and move stores memory users to corresponding memory slices. 3434 // Disable memory split verification code until the fix for 6984348. 3435 // Currently it produces false negative results since it does not cover all cases. 3436 #if 0 // ifdef ASSERT 3437 visited.Reset(); 3438 Node_Stack old_mems(arena, _compile->unique() >> 2); 3439 #endif 3440 for (uint i = 0; i < ideal_nodes.size(); i++) { 3441 Node* n = ideal_nodes.at(i); 3442 Node* nmem = get_map(n->_idx); 3443 assert(nmem != NULL, "sanity"); 3444 if (n->is_Mem()) { 3445 #if 0 // ifdef ASSERT 3446 Node* old_mem = n->in(MemNode::Memory); 3447 if (!visited.test_set(old_mem->_idx)) { 3448 old_mems.push(old_mem, old_mem->outcnt()); 3449 } 3450 #endif 3451 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3452 if (!n->is_Load()) { 3453 // Move memory users of a store first. 3454 move_inst_mem(n, orig_phis); 3455 } 3456 // Now update memory input 3457 igvn->hash_delete(n); 3458 n->set_req(MemNode::Memory, nmem); 3459 igvn->hash_insert(n); 3460 record_for_optimizer(n); 3461 } else { 3462 assert(n->is_Allocate() || n->is_CheckCastPP() || 3463 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3464 } 3465 } 3466 #if 0 // ifdef ASSERT 3467 // Verify that memory was split correctly 3468 while (old_mems.is_nonempty()) { 3469 Node* old_mem = old_mems.node(); 3470 uint old_cnt = old_mems.index(); 3471 old_mems.pop(); 3472 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3473 } 3474 #endif 3475 } 3476 3477 #ifndef PRODUCT 3478 static const char *node_type_names[] = { 3479 "UnknownType", 3480 "JavaObject", 3481 "LocalVar", 3482 "Field", 3483 "Arraycopy" 3484 }; 3485 3486 static const char *esc_names[] = { 3487 "UnknownEscape", 3488 "NoEscape", 3489 "ArgEscape", 3490 "GlobalEscape" 3491 }; 3492 3493 void PointsToNode::dump(bool print_state) const { 3494 NodeType nt = node_type(); 3495 tty->print("%s ", node_type_names[(int) nt]); 3496 if (print_state) { 3497 EscapeState es = escape_state(); 3498 EscapeState fields_es = fields_escape_state(); 3499 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3500 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3501 tty->print("NSR "); 3502 } 3503 if (is_Field()) { 3504 FieldNode* f = (FieldNode*)this; 3505 if (f->is_oop()) 3506 tty->print("oop "); 3507 if (f->offset() > 0) 3508 tty->print("+%d ", f->offset()); 3509 tty->print("("); 3510 for (BaseIterator i(f); i.has_next(); i.next()) { 3511 PointsToNode* b = i.get(); 3512 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3513 } 3514 tty->print(" )"); 3515 } 3516 tty->print("["); 3517 for (EdgeIterator i(this); i.has_next(); i.next()) { 3518 PointsToNode* e = i.get(); 3519 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3520 } 3521 tty->print(" ["); 3522 for (UseIterator i(this); i.has_next(); i.next()) { 3523 PointsToNode* u = i.get(); 3524 bool is_base = false; 3525 if (PointsToNode::is_base_use(u)) { 3526 is_base = true; 3527 u = PointsToNode::get_use_node(u)->as_Field(); 3528 } 3529 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3530 } 3531 tty->print(" ]] "); 3532 if (_node == NULL) 3533 tty->print_cr("<null>"); 3534 else 3535 _node->dump(); 3536 } 3537 3538 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3539 bool first = true; 3540 int ptnodes_length = ptnodes_worklist.length(); 3541 for (int i = 0; i < ptnodes_length; i++) { 3542 PointsToNode *ptn = ptnodes_worklist.at(i); 3543 if (ptn == NULL || !ptn->is_JavaObject()) 3544 continue; 3545 PointsToNode::EscapeState es = ptn->escape_state(); 3546 if ((es != PointsToNode::NoEscape) && !Verbose) { 3547 continue; 3548 } 3549 Node* n = ptn->ideal_node(); 3550 if (n->is_Allocate() || (n->is_CallStaticJava() && 3551 n->as_CallStaticJava()->is_boxing_method())) { 3552 if (first) { 3553 tty->cr(); 3554 tty->print("======== Connection graph for "); 3555 _compile->method()->print_short_name(); 3556 tty->cr(); 3557 first = false; 3558 } 3559 ptn->dump(); 3560 // Print all locals and fields which reference this allocation 3561 for (UseIterator j(ptn); j.has_next(); j.next()) { 3562 PointsToNode* use = j.get(); 3563 if (use->is_LocalVar()) { 3564 use->dump(Verbose); 3565 } else if (Verbose) { 3566 use->dump(); 3567 } 3568 } 3569 tty->cr(); 3570 } 3571 } 3572 } 3573 #endif 3574 3575 void ConnectionGraph::record_for_optimizer(Node *n) { 3576 _igvn->_worklist.push(n); 3577 _igvn->add_users_to_worklist(n); 3578 }