1 /* 2 * Copyright (c) 2005, 2018, 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 if (n->is_Call()) { 354 // Arguments to allocation and locking don't escape. 355 if (n->is_AbstractLock()) { 356 // Put Lock and Unlock nodes on IGVN worklist to process them during 357 // first IGVN optimization when escape information is still available. 358 record_for_optimizer(n); 359 } else if (n->is_Allocate()) { 360 add_call_node(n->as_Call()); 361 record_for_optimizer(n); 362 } else { 363 if (n->is_CallStaticJava()) { 364 const char* name = n->as_CallStaticJava()->_name; 365 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 366 return; // Skip uncommon traps 367 } 368 // Don't mark as processed since call's arguments have to be processed. 369 delayed_worklist->push(n); 370 // Check if a call returns an object. 371 if ((n->as_Call()->returns_pointer() && 372 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) || 373 (n->is_CallStaticJava() && 374 n->as_CallStaticJava()->is_boxing_method())) { 375 add_call_node(n->as_Call()); 376 } 377 } 378 return; 379 } 380 // Put this check here to process call arguments since some call nodes 381 // point to phantom_obj. 382 if (n_ptn == phantom_obj || n_ptn == null_obj) 383 return; // Skip predefined nodes. 384 385 int opcode = n->Opcode(); 386 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode); 387 if (gc_handled) { 388 return; // Ignore node if already handled by GC. 389 } 390 switch (opcode) { 391 case Op_AddP: { 392 Node* base = get_addp_base(n); 393 PointsToNode* ptn_base = ptnode_adr(base->_idx); 394 // Field nodes are created for all field types. They are used in 395 // adjust_scalar_replaceable_state() and split_unique_types(). 396 // Note, non-oop fields will have only base edges in Connection 397 // Graph because such fields are not used for oop loads and stores. 398 int offset = address_offset(n, igvn); 399 add_field(n, PointsToNode::NoEscape, offset); 400 if (ptn_base == NULL) { 401 delayed_worklist->push(n); // Process it later. 402 } else { 403 n_ptn = ptnode_adr(n_idx); 404 add_base(n_ptn->as_Field(), ptn_base); 405 } 406 break; 407 } 408 case Op_CastX2P: { 409 map_ideal_node(n, phantom_obj); 410 break; 411 } 412 case Op_CastPP: 413 case Op_CheckCastPP: 414 case Op_EncodeP: 415 case Op_DecodeN: 416 case Op_EncodePKlass: 417 case Op_DecodeNKlass: { 418 add_local_var_and_edge(n, PointsToNode::NoEscape, 419 n->in(1), delayed_worklist); 420 break; 421 } 422 case Op_CMoveP: { 423 add_local_var(n, PointsToNode::NoEscape); 424 // Do not add edges during first iteration because some could be 425 // not defined yet. 426 delayed_worklist->push(n); 427 break; 428 } 429 case Op_ConP: 430 case Op_ConN: 431 case Op_ConNKlass: { 432 // assume all oop constants globally escape except for null 433 PointsToNode::EscapeState es; 434 const Type* t = igvn->type(n); 435 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 436 es = PointsToNode::NoEscape; 437 } else { 438 es = PointsToNode::GlobalEscape; 439 } 440 add_java_object(n, es); 441 break; 442 } 443 case Op_CreateEx: { 444 // assume that all exception objects globally escape 445 map_ideal_node(n, phantom_obj); 446 break; 447 } 448 case Op_LoadKlass: 449 case Op_LoadNKlass: { 450 // Unknown class is loaded 451 map_ideal_node(n, phantom_obj); 452 break; 453 } 454 case Op_LoadP: 455 case Op_LoadN: 456 case Op_LoadPLocked: { 457 add_objload_to_connection_graph(n, delayed_worklist); 458 break; 459 } 460 case Op_Parm: { 461 map_ideal_node(n, phantom_obj); 462 break; 463 } 464 case Op_PartialSubtypeCheck: { 465 // Produces Null or notNull and is used in only in CmpP so 466 // phantom_obj could be used. 467 map_ideal_node(n, phantom_obj); // Result is unknown 468 break; 469 } 470 case Op_Phi: { 471 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 472 // ThreadLocal has RawPtr type. 473 const Type* t = n->as_Phi()->type(); 474 if (t->make_ptr() != NULL) { 475 add_local_var(n, PointsToNode::NoEscape); 476 // Do not add edges during first iteration because some could be 477 // not defined yet. 478 delayed_worklist->push(n); 479 } 480 break; 481 } 482 case Op_Proj: { 483 // we are only interested in the oop result projection from a call 484 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 485 n->in(0)->as_Call()->returns_pointer()) { 486 add_local_var_and_edge(n, PointsToNode::NoEscape, 487 n->in(0), delayed_worklist); 488 } 489 break; 490 } 491 case Op_Rethrow: // Exception object escapes 492 case Op_Return: { 493 if (n->req() > TypeFunc::Parms && 494 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 495 // Treat Return value as LocalVar with GlobalEscape escape state. 496 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 497 n->in(TypeFunc::Parms), delayed_worklist); 498 } 499 break; 500 } 501 case Op_CompareAndExchangeP: 502 case Op_CompareAndExchangeN: 503 case Op_GetAndSetP: 504 case Op_GetAndSetN: { 505 add_objload_to_connection_graph(n, delayed_worklist); 506 // fallthrough 507 } 508 case Op_StoreP: 509 case Op_StoreN: 510 case Op_StoreNKlass: 511 case Op_StorePConditional: 512 case Op_WeakCompareAndSwapP: 513 case Op_WeakCompareAndSwapN: 514 case Op_CompareAndSwapP: 515 case Op_CompareAndSwapN: { 516 add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 517 break; 518 } 519 case Op_AryEq: 520 case Op_HasNegatives: 521 case Op_StrComp: 522 case Op_StrEquals: 523 case Op_StrIndexOf: 524 case Op_StrIndexOfChar: 525 case Op_StrInflatedCopy: 526 case Op_StrCompressedCopy: 527 case Op_EncodeISOArray: { 528 add_local_var(n, PointsToNode::ArgEscape); 529 delayed_worklist->push(n); // Process it later. 530 break; 531 } 532 case Op_ThreadLocal: { 533 add_java_object(n, PointsToNode::ArgEscape); 534 break; 535 } 536 default: 537 ; // Do nothing for nodes not related to EA. 538 } 539 return; 540 } 541 542 #ifdef ASSERT 543 #define ELSE_FAIL(name) \ 544 /* Should not be called for not pointer type. */ \ 545 n->dump(1); \ 546 assert(false, name); \ 547 break; 548 #else 549 #define ELSE_FAIL(name) \ 550 break; 551 #endif 552 553 // Add final simple edges to graph. 554 void ConnectionGraph::add_final_edges(Node *n) { 555 PointsToNode* n_ptn = ptnode_adr(n->_idx); 556 #ifdef ASSERT 557 if (_verify && n_ptn->is_JavaObject()) 558 return; // This method does not change graph for JavaObject. 559 #endif 560 561 if (n->is_Call()) { 562 process_call_arguments(n->as_Call()); 563 return; 564 } 565 assert(n->is_Store() || n->is_LoadStore() || 566 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 567 "node should be registered already"); 568 int opcode = n->Opcode(); 569 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode); 570 if (gc_handled) { 571 return; // Ignore node if already handled by GC. 572 } 573 switch (opcode) { 574 case Op_AddP: { 575 Node* base = get_addp_base(n); 576 PointsToNode* ptn_base = ptnode_adr(base->_idx); 577 assert(ptn_base != NULL, "field's base should be registered"); 578 add_base(n_ptn->as_Field(), ptn_base); 579 break; 580 } 581 case Op_CastPP: 582 case Op_CheckCastPP: 583 case Op_EncodeP: 584 case Op_DecodeN: 585 case Op_EncodePKlass: 586 case Op_DecodeNKlass: { 587 add_local_var_and_edge(n, PointsToNode::NoEscape, 588 n->in(1), NULL); 589 break; 590 } 591 case Op_CMoveP: { 592 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 593 Node* in = n->in(i); 594 if (in == NULL) 595 continue; // ignore NULL 596 Node* uncast_in = in->uncast(); 597 if (uncast_in->is_top() || uncast_in == n) 598 continue; // ignore top or inputs which go back this node 599 PointsToNode* ptn = ptnode_adr(in->_idx); 600 assert(ptn != NULL, "node should be registered"); 601 add_edge(n_ptn, ptn); 602 } 603 break; 604 } 605 case Op_LoadP: 606 case Op_LoadN: 607 case Op_LoadPLocked: { 608 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 609 // ThreadLocal has RawPtr type. 610 const Type* t = _igvn->type(n); 611 if (t->make_ptr() != NULL) { 612 Node* adr = n->in(MemNode::Address); 613 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 614 break; 615 } 616 ELSE_FAIL("Op_LoadP"); 617 } 618 case Op_Phi: { 619 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 620 // ThreadLocal has RawPtr type. 621 const Type* t = n->as_Phi()->type(); 622 if (t->make_ptr() != NULL) { 623 for (uint i = 1; i < n->req(); i++) { 624 Node* in = n->in(i); 625 if (in == NULL) 626 continue; // ignore NULL 627 Node* uncast_in = in->uncast(); 628 if (uncast_in->is_top() || uncast_in == n) 629 continue; // ignore top or inputs which go back this node 630 PointsToNode* ptn = ptnode_adr(in->_idx); 631 assert(ptn != NULL, "node should be registered"); 632 add_edge(n_ptn, ptn); 633 } 634 break; 635 } 636 ELSE_FAIL("Op_Phi"); 637 } 638 case Op_Proj: { 639 // we are only interested in the oop result projection from a call 640 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 641 n->in(0)->as_Call()->returns_pointer()) { 642 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 643 break; 644 } 645 ELSE_FAIL("Op_Proj"); 646 } 647 case Op_Rethrow: // Exception object escapes 648 case Op_Return: { 649 if (n->req() > TypeFunc::Parms && 650 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 651 // Treat Return value as LocalVar with GlobalEscape escape state. 652 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 653 n->in(TypeFunc::Parms), NULL); 654 break; 655 } 656 ELSE_FAIL("Op_Return"); 657 } 658 case Op_StoreP: 659 case Op_StoreN: 660 case Op_StoreNKlass: 661 case Op_StorePConditional: 662 case Op_CompareAndExchangeP: 663 case Op_CompareAndExchangeN: 664 case Op_CompareAndSwapP: 665 case Op_CompareAndSwapN: 666 case Op_WeakCompareAndSwapP: 667 case Op_WeakCompareAndSwapN: 668 case Op_GetAndSetP: 669 case Op_GetAndSetN: { 670 Node* adr = n->in(MemNode::Address); 671 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 672 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 673 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 674 } 675 if (add_final_edges_unsafe_access(n, opcode)) { 676 break; 677 } 678 ELSE_FAIL("Op_StoreP"); 679 } 680 case Op_AryEq: 681 case Op_HasNegatives: 682 case Op_StrComp: 683 case Op_StrEquals: 684 case Op_StrIndexOf: 685 case Op_StrIndexOfChar: 686 case Op_StrInflatedCopy: 687 case Op_StrCompressedCopy: 688 case Op_EncodeISOArray: { 689 // char[]/byte[] arrays passed to string intrinsic do not escape but 690 // they are not scalar replaceable. Adjust escape state for them. 691 // Start from in(2) edge since in(1) is memory edge. 692 for (uint i = 2; i < n->req(); i++) { 693 Node* adr = n->in(i); 694 const Type* at = _igvn->type(adr); 695 if (!adr->is_top() && at->isa_ptr()) { 696 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 697 at->isa_ptr() != NULL, "expecting a pointer"); 698 if (adr->is_AddP()) { 699 adr = get_addp_base(adr); 700 } 701 PointsToNode* ptn = ptnode_adr(adr->_idx); 702 assert(ptn != NULL, "node should be registered"); 703 add_edge(n_ptn, ptn); 704 } 705 } 706 break; 707 } 708 default: { 709 // This method should be called only for EA specific nodes which may 710 // miss some edges when they were created. 711 #ifdef ASSERT 712 n->dump(1); 713 #endif 714 guarantee(false, "unknown node"); 715 } 716 } 717 return; 718 } 719 720 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) { 721 Node* adr = n->in(MemNode::Address); 722 const Type* adr_type = _igvn->type(adr); 723 adr_type = adr_type->make_ptr(); 724 if (adr_type == NULL) { 725 return; // skip dead nodes 726 } 727 if (adr_type->isa_oopptr() 728 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 729 && adr_type == TypeRawPtr::NOTNULL 730 && adr->in(AddPNode::Address)->is_Proj() 731 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 732 delayed_worklist->push(n); // Process it later. 733 #ifdef ASSERT 734 assert (adr->is_AddP(), "expecting an AddP"); 735 if (adr_type == TypeRawPtr::NOTNULL) { 736 // Verify a raw address for a store captured by Initialize node. 737 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 738 assert(offs != Type::OffsetBot, "offset must be a constant"); 739 } 740 #endif 741 } else { 742 // Ignore copy the displaced header to the BoxNode (OSR compilation). 743 if (adr->is_BoxLock()) { 744 return; 745 } 746 // Stored value escapes in unsafe access. 747 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 748 delayed_worklist->push(n); // Process unsafe access later. 749 return; 750 } 751 #ifdef ASSERT 752 n->dump(1); 753 assert(false, "not unsafe"); 754 #endif 755 } 756 } 757 758 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) { 759 Node* adr = n->in(MemNode::Address); 760 const Type *adr_type = _igvn->type(adr); 761 adr_type = adr_type->make_ptr(); 762 #ifdef ASSERT 763 if (adr_type == NULL) { 764 n->dump(1); 765 assert(adr_type != NULL, "dead node should not be on list"); 766 return true; 767 } 768 #endif 769 770 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 771 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 772 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 773 } 774 775 if (adr_type->isa_oopptr() 776 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 777 && adr_type == TypeRawPtr::NOTNULL 778 && adr->in(AddPNode::Address)->is_Proj() 779 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 780 // Point Address to Value 781 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 782 assert(adr_ptn != NULL && 783 adr_ptn->as_Field()->is_oop(), "node should be registered"); 784 Node* val = n->in(MemNode::ValueIn); 785 PointsToNode* ptn = ptnode_adr(val->_idx); 786 assert(ptn != NULL, "node should be registered"); 787 add_edge(adr_ptn, ptn); 788 return true; 789 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 790 // Stored value escapes in unsafe access. 791 Node* val = n->in(MemNode::ValueIn); 792 PointsToNode* ptn = ptnode_adr(val->_idx); 793 assert(ptn != NULL, "node should be registered"); 794 set_escape_state(ptn, PointsToNode::GlobalEscape); 795 // Add edge to object for unsafe access with offset. 796 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 797 assert(adr_ptn != NULL, "node should be registered"); 798 if (adr_ptn->is_Field()) { 799 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 800 add_edge(adr_ptn, ptn); 801 } 802 return true; 803 } 804 return false; 805 } 806 807 void ConnectionGraph::add_call_node(CallNode* call) { 808 assert(call->returns_pointer(), "only for call which returns pointer"); 809 uint call_idx = call->_idx; 810 if (call->is_Allocate()) { 811 Node* k = call->in(AllocateNode::KlassNode); 812 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 813 assert(kt != NULL, "TypeKlassPtr required."); 814 ciKlass* cik = kt->klass(); 815 PointsToNode::EscapeState es = PointsToNode::NoEscape; 816 bool scalar_replaceable = true; 817 if (call->is_AllocateArray()) { 818 if (!cik->is_array_klass()) { // StressReflectiveCode 819 es = PointsToNode::GlobalEscape; 820 } else { 821 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 822 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 823 // Not scalar replaceable if the length is not constant or too big. 824 scalar_replaceable = false; 825 } 826 } 827 } else { // Allocate instance 828 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 829 cik->is_subclass_of(_compile->env()->Reference_klass()) || 830 !cik->is_instance_klass() || // StressReflectiveCode 831 !cik->as_instance_klass()->can_be_instantiated() || 832 cik->as_instance_klass()->has_finalizer()) { 833 es = PointsToNode::GlobalEscape; 834 } 835 } 836 add_java_object(call, es); 837 PointsToNode* ptn = ptnode_adr(call_idx); 838 if (!scalar_replaceable && ptn->scalar_replaceable()) { 839 ptn->set_scalar_replaceable(false); 840 } 841 } else if (call->is_CallStaticJava()) { 842 // Call nodes could be different types: 843 // 844 // 1. CallDynamicJavaNode (what happened during call is unknown): 845 // 846 // - mapped to GlobalEscape JavaObject node if oop is returned; 847 // 848 // - all oop arguments are escaping globally; 849 // 850 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 851 // 852 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 853 // 854 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 855 // - mapped to NoEscape JavaObject node if non-escaping object allocated 856 // during call is returned; 857 // - mapped to ArgEscape LocalVar node pointed to object arguments 858 // which are returned and does not escape during call; 859 // 860 // - oop arguments escaping status is defined by bytecode analysis; 861 // 862 // For a static call, we know exactly what method is being called. 863 // Use bytecode estimator to record whether the call's return value escapes. 864 ciMethod* meth = call->as_CallJava()->method(); 865 if (meth == NULL) { 866 const char* name = call->as_CallStaticJava()->_name; 867 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 868 // Returns a newly allocated unescaped object. 869 add_java_object(call, PointsToNode::NoEscape); 870 ptnode_adr(call_idx)->set_scalar_replaceable(false); 871 } else if (meth->is_boxing_method()) { 872 // Returns boxing object 873 PointsToNode::EscapeState es; 874 vmIntrinsics::ID intr = meth->intrinsic_id(); 875 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 876 // It does not escape if object is always allocated. 877 es = PointsToNode::NoEscape; 878 } else { 879 // It escapes globally if object could be loaded from cache. 880 es = PointsToNode::GlobalEscape; 881 } 882 add_java_object(call, es); 883 } else { 884 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 885 call_analyzer->copy_dependencies(_compile->dependencies()); 886 if (call_analyzer->is_return_allocated()) { 887 // Returns a newly allocated unescaped object, simply 888 // update dependency information. 889 // Mark it as NoEscape so that objects referenced by 890 // it's fields will be marked as NoEscape at least. 891 add_java_object(call, PointsToNode::NoEscape); 892 ptnode_adr(call_idx)->set_scalar_replaceable(false); 893 } else { 894 // Determine whether any arguments are returned. 895 const TypeTuple* d = call->tf()->domain(); 896 bool ret_arg = false; 897 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 898 if (d->field_at(i)->isa_ptr() != NULL && 899 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 900 ret_arg = true; 901 break; 902 } 903 } 904 if (ret_arg) { 905 add_local_var(call, PointsToNode::ArgEscape); 906 } else { 907 // Returns unknown object. 908 map_ideal_node(call, phantom_obj); 909 } 910 } 911 } 912 } else { 913 // An other type of call, assume the worst case: 914 // returned value is unknown and globally escapes. 915 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 916 map_ideal_node(call, phantom_obj); 917 } 918 } 919 920 void ConnectionGraph::process_call_arguments(CallNode *call) { 921 bool is_arraycopy = false; 922 switch (call->Opcode()) { 923 #ifdef ASSERT 924 case Op_Allocate: 925 case Op_AllocateArray: 926 case Op_Lock: 927 case Op_Unlock: 928 assert(false, "should be done already"); 929 break; 930 #endif 931 case Op_ArrayCopy: 932 case Op_CallLeafNoFP: 933 // Most array copies are ArrayCopy nodes at this point but there 934 // are still a few direct calls to the copy subroutines (See 935 // PhaseStringOpts::copy_string()) 936 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 937 call->as_CallLeaf()->is_call_to_arraycopystub(); 938 // fall through 939 case Op_CallLeaf: { 940 // Stub calls, objects do not escape but they are not scale replaceable. 941 // Adjust escape state for outgoing arguments. 942 const TypeTuple * d = call->tf()->domain(); 943 bool src_has_oops = false; 944 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 945 const Type* at = d->field_at(i); 946 Node *arg = call->in(i); 947 if (arg == NULL) { 948 continue; 949 } 950 const Type *aat = _igvn->type(arg); 951 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 952 continue; 953 if (arg->is_AddP()) { 954 // 955 // The inline_native_clone() case when the arraycopy stub is called 956 // after the allocation before Initialize and CheckCastPP nodes. 957 // Or normal arraycopy for object arrays case. 958 // 959 // Set AddP's base (Allocate) as not scalar replaceable since 960 // pointer to the base (with offset) is passed as argument. 961 // 962 arg = get_addp_base(arg); 963 } 964 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 965 assert(arg_ptn != NULL, "should be registered"); 966 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 967 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 968 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 969 aat->isa_ptr() != NULL, "expecting an Ptr"); 970 bool arg_has_oops = aat->isa_oopptr() && 971 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 972 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 973 if (i == TypeFunc::Parms) { 974 src_has_oops = arg_has_oops; 975 } 976 // 977 // src or dst could be j.l.Object when other is basic type array: 978 // 979 // arraycopy(char[],0,Object*,0,size); 980 // arraycopy(Object*,0,char[],0,size); 981 // 982 // Don't add edges in such cases. 983 // 984 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 985 arg_has_oops && (i > TypeFunc::Parms); 986 #ifdef ASSERT 987 if (!(is_arraycopy || 988 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) || 989 (call->as_CallLeaf()->_name != NULL && 990 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 991 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 992 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 993 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 994 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 995 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 996 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 997 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 || 998 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 999 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 || 1000 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 1001 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 1002 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 1003 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 1004 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 1005 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 1006 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 1007 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 1008 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 1009 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 1010 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 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 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1735 if (n->fast_out(i)->is_LoadStore()) { 1736 jobj->set_scalar_replaceable(false); 1737 return; 1738 } 1739 } 1740 1741 // 6. Or the address may point to more then one object. This may produce 1742 // the false positive result (set not scalar replaceable) 1743 // since the flow-insensitive escape analysis can't separate 1744 // the case when stores overwrite the field's value from the case 1745 // when stores happened on different control branches. 1746 // 1747 // Note: it will disable scalar replacement in some cases: 1748 // 1749 // Point p[] = new Point[1]; 1750 // p[0] = new Point(); // Will be not scalar replaced 1751 // 1752 // but it will save us from incorrect optimizations in next cases: 1753 // 1754 // Point p[] = new Point[1]; 1755 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1756 // 1757 if (field->base_count() > 1) { 1758 for (BaseIterator i(field); i.has_next(); i.next()) { 1759 PointsToNode* base = i.get(); 1760 // Don't take into account LocalVar nodes which 1761 // may point to only one object which should be also 1762 // this field's base by now. 1763 if (base->is_JavaObject() && base != jobj) { 1764 // Mark all bases. 1765 jobj->set_scalar_replaceable(false); 1766 base->set_scalar_replaceable(false); 1767 } 1768 } 1769 } 1770 } 1771 } 1772 1773 #ifdef ASSERT 1774 void ConnectionGraph::verify_connection_graph( 1775 GrowableArray<PointsToNode*>& ptnodes_worklist, 1776 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1777 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1778 GrowableArray<Node*>& addp_worklist) { 1779 // Verify that graph is complete - no new edges could be added. 1780 int java_objects_length = java_objects_worklist.length(); 1781 int non_escaped_length = non_escaped_worklist.length(); 1782 int new_edges = 0; 1783 for (int next = 0; next < java_objects_length; ++next) { 1784 JavaObjectNode* ptn = java_objects_worklist.at(next); 1785 new_edges += add_java_object_edges(ptn, true); 1786 } 1787 assert(new_edges == 0, "graph was not complete"); 1788 // Verify that escape state is final. 1789 int length = non_escaped_worklist.length(); 1790 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1791 assert((non_escaped_length == non_escaped_worklist.length()) && 1792 (non_escaped_length == length) && 1793 (_worklist.length() == 0), "escape state was not final"); 1794 1795 // Verify fields information. 1796 int addp_length = addp_worklist.length(); 1797 for (int next = 0; next < addp_length; ++next ) { 1798 Node* n = addp_worklist.at(next); 1799 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1800 if (field->is_oop()) { 1801 // Verify that field has all bases 1802 Node* base = get_addp_base(n); 1803 PointsToNode* ptn = ptnode_adr(base->_idx); 1804 if (ptn->is_JavaObject()) { 1805 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1806 } else { 1807 assert(ptn->is_LocalVar(), "sanity"); 1808 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1809 PointsToNode* e = i.get(); 1810 if (e->is_JavaObject()) { 1811 assert(field->has_base(e->as_JavaObject()), "sanity"); 1812 } 1813 } 1814 } 1815 // Verify that all fields have initializing values. 1816 if (field->edge_count() == 0) { 1817 tty->print_cr("----------field does not have references----------"); 1818 field->dump(); 1819 for (BaseIterator i(field); i.has_next(); i.next()) { 1820 PointsToNode* base = i.get(); 1821 tty->print_cr("----------field has next base---------------------"); 1822 base->dump(); 1823 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1824 tty->print_cr("----------base has fields-------------------------"); 1825 for (EdgeIterator j(base); j.has_next(); j.next()) { 1826 j.get()->dump(); 1827 } 1828 tty->print_cr("----------base has references---------------------"); 1829 for (UseIterator j(base); j.has_next(); j.next()) { 1830 j.get()->dump(); 1831 } 1832 } 1833 } 1834 for (UseIterator i(field); i.has_next(); i.next()) { 1835 i.get()->dump(); 1836 } 1837 assert(field->edge_count() > 0, "sanity"); 1838 } 1839 } 1840 } 1841 } 1842 #endif 1843 1844 // Optimize ideal graph. 1845 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1846 GrowableArray<Node*>& storestore_worklist) { 1847 Compile* C = _compile; 1848 PhaseIterGVN* igvn = _igvn; 1849 if (EliminateLocks) { 1850 // Mark locks before changing ideal graph. 1851 int cnt = C->macro_count(); 1852 for( int i=0; i < cnt; i++ ) { 1853 Node *n = C->macro_node(i); 1854 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1855 AbstractLockNode* alock = n->as_AbstractLock(); 1856 if (!alock->is_non_esc_obj()) { 1857 if (not_global_escape(alock->obj_node())) { 1858 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1859 // The lock could be marked eliminated by lock coarsening 1860 // code during first IGVN before EA. Replace coarsened flag 1861 // to eliminate all associated locks/unlocks. 1862 #ifdef ASSERT 1863 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1864 #endif 1865 alock->set_non_esc_obj(); 1866 } 1867 } 1868 } 1869 } 1870 } 1871 1872 if (OptimizePtrCompare) { 1873 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1874 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1875 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1876 // Optimize objects compare. 1877 while (ptr_cmp_worklist.length() != 0) { 1878 Node *n = ptr_cmp_worklist.pop(); 1879 Node *res = optimize_ptr_compare(n); 1880 if (res != NULL) { 1881 #ifndef PRODUCT 1882 if (PrintOptimizePtrCompare) { 1883 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")); 1884 if (Verbose) { 1885 n->dump(1); 1886 } 1887 } 1888 #endif 1889 igvn->replace_node(n, res); 1890 } 1891 } 1892 // cleanup 1893 if (_pcmp_neq->outcnt() == 0) 1894 igvn->hash_delete(_pcmp_neq); 1895 if (_pcmp_eq->outcnt() == 0) 1896 igvn->hash_delete(_pcmp_eq); 1897 } 1898 1899 // For MemBarStoreStore nodes added in library_call.cpp, check 1900 // escape status of associated AllocateNode and optimize out 1901 // MemBarStoreStore node if the allocated object never escapes. 1902 while (storestore_worklist.length() != 0) { 1903 Node *n = storestore_worklist.pop(); 1904 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1905 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1906 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1907 if (not_global_escape(alloc)) { 1908 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1909 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1910 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 1911 igvn->register_new_node_with_optimizer(mb); 1912 igvn->replace_node(storestore, mb); 1913 } 1914 } 1915 } 1916 1917 // Optimize objects compare. 1918 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 1919 assert(OptimizePtrCompare, "sanity"); 1920 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 1921 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 1922 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 1923 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 1924 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 1925 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 1926 1927 // Check simple cases first. 1928 if (jobj1 != NULL) { 1929 if (jobj1->escape_state() == PointsToNode::NoEscape) { 1930 if (jobj1 == jobj2) { 1931 // Comparing the same not escaping object. 1932 return _pcmp_eq; 1933 } 1934 Node* obj = jobj1->ideal_node(); 1935 // Comparing not escaping allocation. 1936 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1937 !ptn2->points_to(jobj1)) { 1938 return _pcmp_neq; // This includes nullness check. 1939 } 1940 } 1941 } 1942 if (jobj2 != NULL) { 1943 if (jobj2->escape_state() == PointsToNode::NoEscape) { 1944 Node* obj = jobj2->ideal_node(); 1945 // Comparing not escaping allocation. 1946 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 1947 !ptn1->points_to(jobj2)) { 1948 return _pcmp_neq; // This includes nullness check. 1949 } 1950 } 1951 } 1952 if (jobj1 != NULL && jobj1 != phantom_obj && 1953 jobj2 != NULL && jobj2 != phantom_obj && 1954 jobj1->ideal_node()->is_Con() && 1955 jobj2->ideal_node()->is_Con()) { 1956 // Klass or String constants compare. Need to be careful with 1957 // compressed pointers - compare types of ConN and ConP instead of nodes. 1958 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 1959 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 1960 if (t1->make_ptr() == t2->make_ptr()) { 1961 return _pcmp_eq; 1962 } else { 1963 return _pcmp_neq; 1964 } 1965 } 1966 if (ptn1->meet(ptn2)) { 1967 return NULL; // Sets are not disjoint 1968 } 1969 1970 // Sets are disjoint. 1971 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 1972 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 1973 bool set1_has_null_ptr = ptn1->points_to(null_obj); 1974 bool set2_has_null_ptr = ptn2->points_to(null_obj); 1975 if ((set1_has_unknown_ptr && set2_has_null_ptr) || 1976 (set2_has_unknown_ptr && set1_has_null_ptr)) { 1977 // Check nullness of unknown object. 1978 return NULL; 1979 } 1980 1981 // Disjointness by itself is not sufficient since 1982 // alias analysis is not complete for escaped objects. 1983 // Disjoint sets are definitely unrelated only when 1984 // at least one set has only not escaping allocations. 1985 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 1986 if (ptn1->non_escaping_allocation()) { 1987 return _pcmp_neq; 1988 } 1989 } 1990 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 1991 if (ptn2->non_escaping_allocation()) { 1992 return _pcmp_neq; 1993 } 1994 } 1995 return NULL; 1996 } 1997 1998 // Connection Graph constuction functions. 1999 2000 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 2001 PointsToNode* ptadr = _nodes.at(n->_idx); 2002 if (ptadr != NULL) { 2003 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 2004 return; 2005 } 2006 Compile* C = _compile; 2007 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 2008 _nodes.at_put(n->_idx, ptadr); 2009 } 2010 2011 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 2012 PointsToNode* ptadr = _nodes.at(n->_idx); 2013 if (ptadr != NULL) { 2014 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2015 return; 2016 } 2017 Compile* C = _compile; 2018 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2019 _nodes.at_put(n->_idx, ptadr); 2020 } 2021 2022 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2023 PointsToNode* ptadr = _nodes.at(n->_idx); 2024 if (ptadr != NULL) { 2025 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2026 return; 2027 } 2028 bool unsafe = false; 2029 bool is_oop = is_oop_field(n, offset, &unsafe); 2030 if (unsafe) { 2031 es = PointsToNode::GlobalEscape; 2032 } 2033 Compile* C = _compile; 2034 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2035 _nodes.at_put(n->_idx, field); 2036 } 2037 2038 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2039 PointsToNode* src, PointsToNode* dst) { 2040 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2041 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2042 PointsToNode* ptadr = _nodes.at(n->_idx); 2043 if (ptadr != NULL) { 2044 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2045 return; 2046 } 2047 Compile* C = _compile; 2048 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2049 _nodes.at_put(n->_idx, ptadr); 2050 // Add edge from arraycopy node to source object. 2051 (void)add_edge(ptadr, src); 2052 src->set_arraycopy_src(); 2053 // Add edge from destination object to arraycopy node. 2054 (void)add_edge(dst, ptadr); 2055 dst->set_arraycopy_dst(); 2056 } 2057 2058 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2059 const Type* adr_type = n->as_AddP()->bottom_type(); 2060 BasicType bt = T_INT; 2061 if (offset == Type::OffsetBot) { 2062 // Check only oop fields. 2063 if (!adr_type->isa_aryptr() || 2064 (adr_type->isa_aryptr()->klass() == NULL) || 2065 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2066 // OffsetBot is used to reference array's element. Ignore first AddP. 2067 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2068 bt = T_OBJECT; 2069 } 2070 } 2071 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2072 if (adr_type->isa_instptr()) { 2073 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2074 if (field != NULL) { 2075 bt = field->layout_type(); 2076 } else { 2077 // Check for unsafe oop field access 2078 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2079 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2080 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2081 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2082 bt = T_OBJECT; 2083 (*unsafe) = true; 2084 } 2085 } 2086 } else if (adr_type->isa_aryptr()) { 2087 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2088 // Ignore array length load. 2089 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2090 // Ignore first AddP. 2091 } else { 2092 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2093 bt = elemtype->array_element_basic_type(); 2094 } 2095 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2096 // Allocation initialization, ThreadLocal field access, unsafe access 2097 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2098 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2099 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) || 2100 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) { 2101 bt = T_OBJECT; 2102 } 2103 } 2104 } 2105 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 2106 } 2107 2108 // Returns unique pointed java object or NULL. 2109 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2110 assert(!_collecting, "should not call when contructed graph"); 2111 // If the node was created after the escape computation we can't answer. 2112 uint idx = n->_idx; 2113 if (idx >= nodes_size()) { 2114 return NULL; 2115 } 2116 PointsToNode* ptn = ptnode_adr(idx); 2117 if (ptn->is_JavaObject()) { 2118 return ptn->as_JavaObject(); 2119 } 2120 assert(ptn->is_LocalVar(), "sanity"); 2121 // Check all java objects it points to. 2122 JavaObjectNode* jobj = NULL; 2123 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2124 PointsToNode* e = i.get(); 2125 if (e->is_JavaObject()) { 2126 if (jobj == NULL) { 2127 jobj = e->as_JavaObject(); 2128 } else if (jobj != e) { 2129 return NULL; 2130 } 2131 } 2132 } 2133 return jobj; 2134 } 2135 2136 // Return true if this node points only to non-escaping allocations. 2137 bool PointsToNode::non_escaping_allocation() { 2138 if (is_JavaObject()) { 2139 Node* n = ideal_node(); 2140 if (n->is_Allocate() || n->is_CallStaticJava()) { 2141 return (escape_state() == PointsToNode::NoEscape); 2142 } else { 2143 return false; 2144 } 2145 } 2146 assert(is_LocalVar(), "sanity"); 2147 // Check all java objects it points to. 2148 for (EdgeIterator i(this); i.has_next(); i.next()) { 2149 PointsToNode* e = i.get(); 2150 if (e->is_JavaObject()) { 2151 Node* n = e->ideal_node(); 2152 if ((e->escape_state() != PointsToNode::NoEscape) || 2153 !(n->is_Allocate() || n->is_CallStaticJava())) { 2154 return false; 2155 } 2156 } 2157 } 2158 return true; 2159 } 2160 2161 // Return true if we know the node does not escape globally. 2162 bool ConnectionGraph::not_global_escape(Node *n) { 2163 assert(!_collecting, "should not call during graph construction"); 2164 // If the node was created after the escape computation we can't answer. 2165 uint idx = n->_idx; 2166 if (idx >= nodes_size()) { 2167 return false; 2168 } 2169 PointsToNode* ptn = ptnode_adr(idx); 2170 PointsToNode::EscapeState es = ptn->escape_state(); 2171 // If we have already computed a value, return it. 2172 if (es >= PointsToNode::GlobalEscape) 2173 return false; 2174 if (ptn->is_JavaObject()) { 2175 return true; // (es < PointsToNode::GlobalEscape); 2176 } 2177 assert(ptn->is_LocalVar(), "sanity"); 2178 // Check all java objects it points to. 2179 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2180 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2181 return false; 2182 } 2183 return true; 2184 } 2185 2186 2187 // Helper functions 2188 2189 // Return true if this node points to specified node or nodes it points to. 2190 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2191 if (is_JavaObject()) { 2192 return (this == ptn); 2193 } 2194 assert(is_LocalVar() || is_Field(), "sanity"); 2195 for (EdgeIterator i(this); i.has_next(); i.next()) { 2196 if (i.get() == ptn) 2197 return true; 2198 } 2199 return false; 2200 } 2201 2202 // Return true if one node points to an other. 2203 bool PointsToNode::meet(PointsToNode* ptn) { 2204 if (this == ptn) { 2205 return true; 2206 } else if (ptn->is_JavaObject()) { 2207 return this->points_to(ptn->as_JavaObject()); 2208 } else if (this->is_JavaObject()) { 2209 return ptn->points_to(this->as_JavaObject()); 2210 } 2211 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2212 int ptn_count = ptn->edge_count(); 2213 for (EdgeIterator i(this); i.has_next(); i.next()) { 2214 PointsToNode* this_e = i.get(); 2215 for (int j = 0; j < ptn_count; j++) { 2216 if (this_e == ptn->edge(j)) 2217 return true; 2218 } 2219 } 2220 return false; 2221 } 2222 2223 #ifdef ASSERT 2224 // Return true if bases point to this java object. 2225 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2226 for (BaseIterator i(this); i.has_next(); i.next()) { 2227 if (i.get() == jobj) 2228 return true; 2229 } 2230 return false; 2231 } 2232 #endif 2233 2234 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2235 const Type *adr_type = phase->type(adr); 2236 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2237 adr->in(AddPNode::Address)->is_Proj() && 2238 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2239 // We are computing a raw address for a store captured by an Initialize 2240 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2241 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2242 assert(offs != Type::OffsetBot || 2243 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2244 "offset must be a constant or it is initialization of array"); 2245 return offs; 2246 } 2247 const TypePtr *t_ptr = adr_type->isa_ptr(); 2248 assert(t_ptr != NULL, "must be a pointer type"); 2249 return t_ptr->offset(); 2250 } 2251 2252 Node* ConnectionGraph::get_addp_base(Node *addp) { 2253 assert(addp->is_AddP(), "must be AddP"); 2254 // 2255 // AddP cases for Base and Address inputs: 2256 // case #1. Direct object's field reference: 2257 // Allocate 2258 // | 2259 // Proj #5 ( oop result ) 2260 // | 2261 // CheckCastPP (cast to instance type) 2262 // | | 2263 // AddP ( base == address ) 2264 // 2265 // case #2. Indirect object's field reference: 2266 // Phi 2267 // | 2268 // CastPP (cast to instance type) 2269 // | | 2270 // AddP ( base == address ) 2271 // 2272 // case #3. Raw object's field reference for Initialize node: 2273 // Allocate 2274 // | 2275 // Proj #5 ( oop result ) 2276 // top | 2277 // \ | 2278 // AddP ( base == top ) 2279 // 2280 // case #4. Array's element reference: 2281 // {CheckCastPP | CastPP} 2282 // | | | 2283 // | AddP ( array's element offset ) 2284 // | | 2285 // AddP ( array's offset ) 2286 // 2287 // case #5. Raw object's field reference for arraycopy stub call: 2288 // The inline_native_clone() case when the arraycopy stub is called 2289 // after the allocation before Initialize and CheckCastPP nodes. 2290 // Allocate 2291 // | 2292 // Proj #5 ( oop result ) 2293 // | | 2294 // AddP ( base == address ) 2295 // 2296 // case #6. Constant Pool, ThreadLocal, CastX2P or 2297 // Raw object's field reference: 2298 // {ConP, ThreadLocal, CastX2P, raw Load} 2299 // top | 2300 // \ | 2301 // AddP ( base == top ) 2302 // 2303 // case #7. Klass's field reference. 2304 // LoadKlass 2305 // | | 2306 // AddP ( base == address ) 2307 // 2308 // case #8. narrow Klass's field reference. 2309 // LoadNKlass 2310 // | 2311 // DecodeN 2312 // | | 2313 // AddP ( base == address ) 2314 // 2315 // case #9. Mixed unsafe access 2316 // {instance} 2317 // | 2318 // CheckCastPP (raw) 2319 // top | 2320 // \ | 2321 // AddP ( base == top ) 2322 // 2323 Node *base = addp->in(AddPNode::Base); 2324 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9. 2325 base = addp->in(AddPNode::Address); 2326 while (base->is_AddP()) { 2327 // Case #6 (unsafe access) may have several chained AddP nodes. 2328 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2329 base = base->in(AddPNode::Address); 2330 } 2331 if (base->Opcode() == Op_CheckCastPP && 2332 base->bottom_type()->isa_rawptr() && 2333 _igvn->type(base->in(1))->isa_oopptr()) { 2334 base = base->in(1); // Case #9 2335 } else { 2336 Node* uncast_base = base->uncast(); 2337 int opcode = uncast_base->Opcode(); 2338 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2339 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2340 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2341 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()) || 2342 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(uncast_base), "sanity"); 2343 } 2344 } 2345 return base; 2346 } 2347 2348 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2349 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2350 Node* addp2 = addp->raw_out(0); 2351 if (addp->outcnt() == 1 && addp2->is_AddP() && 2352 addp2->in(AddPNode::Base) == n && 2353 addp2->in(AddPNode::Address) == addp) { 2354 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2355 // 2356 // Find array's offset to push it on worklist first and 2357 // as result process an array's element offset first (pushed second) 2358 // to avoid CastPP for the array's offset. 2359 // Otherwise the inserted CastPP (LocalVar) will point to what 2360 // the AddP (Field) points to. Which would be wrong since 2361 // the algorithm expects the CastPP has the same point as 2362 // as AddP's base CheckCastPP (LocalVar). 2363 // 2364 // ArrayAllocation 2365 // | 2366 // CheckCastPP 2367 // | 2368 // memProj (from ArrayAllocation CheckCastPP) 2369 // | || 2370 // | || Int (element index) 2371 // | || | ConI (log(element size)) 2372 // | || | / 2373 // | || LShift 2374 // | || / 2375 // | AddP (array's element offset) 2376 // | | 2377 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2378 // | / / 2379 // AddP (array's offset) 2380 // | 2381 // Load/Store (memory operation on array's element) 2382 // 2383 return addp2; 2384 } 2385 return NULL; 2386 } 2387 2388 // 2389 // Adjust the type and inputs of an AddP which computes the 2390 // address of a field of an instance 2391 // 2392 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2393 PhaseGVN* igvn = _igvn; 2394 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2395 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2396 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2397 if (t == NULL) { 2398 // We are computing a raw address for a store captured by an Initialize 2399 // compute an appropriate address type (cases #3 and #5). 2400 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2401 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2402 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2403 assert(offs != Type::OffsetBot, "offset must be a constant"); 2404 t = base_t->add_offset(offs)->is_oopptr(); 2405 } 2406 int inst_id = base_t->instance_id(); 2407 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2408 "old type must be non-instance or match new type"); 2409 2410 // The type 't' could be subclass of 'base_t'. 2411 // As result t->offset() could be large then base_t's size and it will 2412 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2413 // constructor verifies correctness of the offset. 2414 // 2415 // It could happened on subclass's branch (from the type profiling 2416 // inlining) which was not eliminated during parsing since the exactness 2417 // of the allocation type was not propagated to the subclass type check. 2418 // 2419 // Or the type 't' could be not related to 'base_t' at all. 2420 // It could happened when CHA type is different from MDO type on a dead path 2421 // (for example, from instanceof check) which is not collapsed during parsing. 2422 // 2423 // Do nothing for such AddP node and don't process its users since 2424 // this code branch will go away. 2425 // 2426 if (!t->is_known_instance() && 2427 !base_t->klass()->is_subtype_of(t->klass())) { 2428 return false; // bail out 2429 } 2430 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2431 // Do NOT remove the next line: ensure a new alias index is allocated 2432 // for the instance type. Note: C++ will not remove it since the call 2433 // has side effect. 2434 int alias_idx = _compile->get_alias_index(tinst); 2435 igvn->set_type(addp, tinst); 2436 // record the allocation in the node map 2437 set_map(addp, get_map(base->_idx)); 2438 // Set addp's Base and Address to 'base'. 2439 Node *abase = addp->in(AddPNode::Base); 2440 Node *adr = addp->in(AddPNode::Address); 2441 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2442 adr->in(0)->_idx == (uint)inst_id) { 2443 // Skip AddP cases #3 and #5. 2444 } else { 2445 assert(!abase->is_top(), "sanity"); // AddP case #3 2446 if (abase != base) { 2447 igvn->hash_delete(addp); 2448 addp->set_req(AddPNode::Base, base); 2449 if (abase == adr) { 2450 addp->set_req(AddPNode::Address, base); 2451 } else { 2452 // AddP case #4 (adr is array's element offset AddP node) 2453 #ifdef ASSERT 2454 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2455 assert(adr->is_AddP() && atype != NULL && 2456 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2457 #endif 2458 } 2459 igvn->hash_insert(addp); 2460 } 2461 } 2462 // Put on IGVN worklist since at least addp's type was changed above. 2463 record_for_optimizer(addp); 2464 return true; 2465 } 2466 2467 // 2468 // Create a new version of orig_phi if necessary. Returns either the newly 2469 // created phi or an existing phi. Sets create_new to indicate whether a new 2470 // phi was created. Cache the last newly created phi in the node map. 2471 // 2472 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2473 Compile *C = _compile; 2474 PhaseGVN* igvn = _igvn; 2475 new_created = false; 2476 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2477 // nothing to do if orig_phi is bottom memory or matches alias_idx 2478 if (phi_alias_idx == alias_idx) { 2479 return orig_phi; 2480 } 2481 // Have we recently created a Phi for this alias index? 2482 PhiNode *result = get_map_phi(orig_phi->_idx); 2483 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2484 return result; 2485 } 2486 // Previous check may fail when the same wide memory Phi was split into Phis 2487 // for different memory slices. Search all Phis for this region. 2488 if (result != NULL) { 2489 Node* region = orig_phi->in(0); 2490 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2491 Node* phi = region->fast_out(i); 2492 if (phi->is_Phi() && 2493 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2494 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2495 return phi->as_Phi(); 2496 } 2497 } 2498 } 2499 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2500 if (C->do_escape_analysis() == true && !C->failing()) { 2501 // Retry compilation without escape analysis. 2502 // If this is the first failure, the sentinel string will "stick" 2503 // to the Compile object, and the C2Compiler will see it and retry. 2504 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2505 } 2506 return NULL; 2507 } 2508 orig_phi_worklist.append_if_missing(orig_phi); 2509 const TypePtr *atype = C->get_adr_type(alias_idx); 2510 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2511 C->copy_node_notes_to(result, orig_phi); 2512 igvn->set_type(result, result->bottom_type()); 2513 record_for_optimizer(result); 2514 set_map(orig_phi, result); 2515 new_created = true; 2516 return result; 2517 } 2518 2519 // 2520 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2521 // specified alias index. 2522 // 2523 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2524 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2525 Compile *C = _compile; 2526 PhaseGVN* igvn = _igvn; 2527 bool new_phi_created; 2528 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2529 if (!new_phi_created) { 2530 return result; 2531 } 2532 GrowableArray<PhiNode *> phi_list; 2533 GrowableArray<uint> cur_input; 2534 PhiNode *phi = orig_phi; 2535 uint idx = 1; 2536 bool finished = false; 2537 while(!finished) { 2538 while (idx < phi->req()) { 2539 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2540 if (mem != NULL && mem->is_Phi()) { 2541 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2542 if (new_phi_created) { 2543 // found an phi for which we created a new split, push current one on worklist and begin 2544 // processing new one 2545 phi_list.push(phi); 2546 cur_input.push(idx); 2547 phi = mem->as_Phi(); 2548 result = newphi; 2549 idx = 1; 2550 continue; 2551 } else { 2552 mem = newphi; 2553 } 2554 } 2555 if (C->failing()) { 2556 return NULL; 2557 } 2558 result->set_req(idx++, mem); 2559 } 2560 #ifdef ASSERT 2561 // verify that the new Phi has an input for each input of the original 2562 assert( phi->req() == result->req(), "must have same number of inputs."); 2563 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2564 #endif 2565 // Check if all new phi's inputs have specified alias index. 2566 // Otherwise use old phi. 2567 for (uint i = 1; i < phi->req(); i++) { 2568 Node* in = result->in(i); 2569 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2570 } 2571 // we have finished processing a Phi, see if there are any more to do 2572 finished = (phi_list.length() == 0 ); 2573 if (!finished) { 2574 phi = phi_list.pop(); 2575 idx = cur_input.pop(); 2576 PhiNode *prev_result = get_map_phi(phi->_idx); 2577 prev_result->set_req(idx++, result); 2578 result = prev_result; 2579 } 2580 } 2581 return result; 2582 } 2583 2584 // 2585 // The next methods are derived from methods in MemNode. 2586 // 2587 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2588 Node *mem = mmem; 2589 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2590 // means an array I have not precisely typed yet. Do not do any 2591 // alias stuff with it any time soon. 2592 if (toop->base() != Type::AnyPtr && 2593 !(toop->klass() != NULL && 2594 toop->klass()->is_java_lang_Object() && 2595 toop->offset() == Type::OffsetBot)) { 2596 mem = mmem->memory_at(alias_idx); 2597 // Update input if it is progress over what we have now 2598 } 2599 return mem; 2600 } 2601 2602 // 2603 // Move memory users to their memory slices. 2604 // 2605 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2606 Compile* C = _compile; 2607 PhaseGVN* igvn = _igvn; 2608 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2609 assert(tp != NULL, "ptr type"); 2610 int alias_idx = C->get_alias_index(tp); 2611 int general_idx = C->get_general_index(alias_idx); 2612 2613 // Move users first 2614 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2615 Node* use = n->fast_out(i); 2616 if (use->is_MergeMem()) { 2617 MergeMemNode* mmem = use->as_MergeMem(); 2618 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2619 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2620 continue; // Nothing to do 2621 } 2622 // Replace previous general reference to mem node. 2623 uint orig_uniq = C->unique(); 2624 Node* m = find_inst_mem(n, general_idx, orig_phis); 2625 assert(orig_uniq == C->unique(), "no new nodes"); 2626 mmem->set_memory_at(general_idx, m); 2627 --imax; 2628 --i; 2629 } else if (use->is_MemBar()) { 2630 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2631 if (use->req() > MemBarNode::Precedent && 2632 use->in(MemBarNode::Precedent) == n) { 2633 // Don't move related membars. 2634 record_for_optimizer(use); 2635 continue; 2636 } 2637 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2638 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) || 2639 alias_idx == general_idx) { 2640 continue; // Nothing to do 2641 } 2642 // Move to general memory slice. 2643 uint orig_uniq = C->unique(); 2644 Node* m = find_inst_mem(n, general_idx, orig_phis); 2645 assert(orig_uniq == C->unique(), "no new nodes"); 2646 igvn->hash_delete(use); 2647 imax -= use->replace_edge(n, m); 2648 igvn->hash_insert(use); 2649 record_for_optimizer(use); 2650 --i; 2651 #ifdef ASSERT 2652 } else if (use->is_Mem()) { 2653 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2654 // Don't move related cardmark. 2655 continue; 2656 } 2657 // Memory nodes should have new memory input. 2658 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2659 assert(tp != NULL, "ptr type"); 2660 int idx = C->get_alias_index(tp); 2661 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2662 "Following memory nodes should have new memory input or be on the same memory slice"); 2663 } else if (use->is_Phi()) { 2664 // Phi nodes should be split and moved already. 2665 tp = use->as_Phi()->adr_type()->isa_ptr(); 2666 assert(tp != NULL, "ptr type"); 2667 int idx = C->get_alias_index(tp); 2668 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2669 } else { 2670 use->dump(); 2671 assert(false, "should not be here"); 2672 #endif 2673 } 2674 } 2675 } 2676 2677 // 2678 // Search memory chain of "mem" to find a MemNode whose address 2679 // is the specified alias index. 2680 // 2681 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2682 if (orig_mem == NULL) 2683 return orig_mem; 2684 Compile* C = _compile; 2685 PhaseGVN* igvn = _igvn; 2686 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2687 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2688 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 2689 Node *prev = NULL; 2690 Node *result = orig_mem; 2691 while (prev != result) { 2692 prev = result; 2693 if (result == start_mem) 2694 break; // hit one of our sentinels 2695 if (result->is_Mem()) { 2696 const Type *at = igvn->type(result->in(MemNode::Address)); 2697 if (at == Type::TOP) 2698 break; // Dead 2699 assert (at->isa_ptr() != NULL, "pointer type required."); 2700 int idx = C->get_alias_index(at->is_ptr()); 2701 if (idx == alias_idx) 2702 break; // Found 2703 if (!is_instance && (at->isa_oopptr() == NULL || 2704 !at->is_oopptr()->is_known_instance())) { 2705 break; // Do not skip store to general memory slice. 2706 } 2707 result = result->in(MemNode::Memory); 2708 } 2709 if (!is_instance) 2710 continue; // don't search further for non-instance types 2711 // skip over a call which does not affect this memory slice 2712 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2713 Node *proj_in = result->in(0); 2714 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2715 break; // hit one of our sentinels 2716 } else if (proj_in->is_Call()) { 2717 // ArrayCopy node processed here as well 2718 CallNode *call = proj_in->as_Call(); 2719 if (!call->may_modify(toop, igvn)) { 2720 result = call->in(TypeFunc::Memory); 2721 } 2722 } else if (proj_in->is_Initialize()) { 2723 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2724 // Stop if this is the initialization for the object instance which 2725 // which contains this memory slice, otherwise skip over it. 2726 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2727 result = proj_in->in(TypeFunc::Memory); 2728 } 2729 } else if (proj_in->is_MemBar()) { 2730 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() && 2731 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() && 2732 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) { 2733 // clone 2734 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy(); 2735 if (ac->may_modify(toop, igvn)) { 2736 break; 2737 } 2738 } 2739 result = proj_in->in(TypeFunc::Memory); 2740 } 2741 } else if (result->is_MergeMem()) { 2742 MergeMemNode *mmem = result->as_MergeMem(); 2743 result = step_through_mergemem(mmem, alias_idx, toop); 2744 if (result == mmem->base_memory()) { 2745 // Didn't find instance memory, search through general slice recursively. 2746 result = mmem->memory_at(C->get_general_index(alias_idx)); 2747 result = find_inst_mem(result, alias_idx, orig_phis); 2748 if (C->failing()) { 2749 return NULL; 2750 } 2751 mmem->set_memory_at(alias_idx, result); 2752 } 2753 } else if (result->is_Phi() && 2754 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2755 Node *un = result->as_Phi()->unique_input(igvn); 2756 if (un != NULL) { 2757 orig_phis.append_if_missing(result->as_Phi()); 2758 result = un; 2759 } else { 2760 break; 2761 } 2762 } else if (result->is_ClearArray()) { 2763 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2764 // Can not bypass initialization of the instance 2765 // we are looking for. 2766 break; 2767 } 2768 // Otherwise skip it (the call updated 'result' value). 2769 } else if (result->Opcode() == Op_SCMemProj) { 2770 Node* mem = result->in(0); 2771 Node* adr = NULL; 2772 if (mem->is_LoadStore()) { 2773 adr = mem->in(MemNode::Address); 2774 } else { 2775 assert(mem->Opcode() == Op_EncodeISOArray || 2776 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 2777 adr = mem->in(3); // Memory edge corresponds to destination array 2778 } 2779 const Type *at = igvn->type(adr); 2780 if (at != Type::TOP) { 2781 assert(at->isa_ptr() != NULL, "pointer type required."); 2782 int idx = C->get_alias_index(at->is_ptr()); 2783 if (idx == alias_idx) { 2784 // Assert in debug mode 2785 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 2786 break; // In product mode return SCMemProj node 2787 } 2788 } 2789 result = mem->in(MemNode::Memory); 2790 } else if (result->Opcode() == Op_StrInflatedCopy) { 2791 Node* adr = result->in(3); // Memory edge corresponds to destination array 2792 const Type *at = igvn->type(adr); 2793 if (at != Type::TOP) { 2794 assert(at->isa_ptr() != NULL, "pointer type required."); 2795 int idx = C->get_alias_index(at->is_ptr()); 2796 if (idx == alias_idx) { 2797 // Assert in debug mode 2798 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 2799 break; // In product mode return SCMemProj node 2800 } 2801 } 2802 result = result->in(MemNode::Memory); 2803 } 2804 } 2805 if (result->is_Phi()) { 2806 PhiNode *mphi = result->as_Phi(); 2807 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2808 const TypePtr *t = mphi->adr_type(); 2809 if (!is_instance) { 2810 // Push all non-instance Phis on the orig_phis worklist to update inputs 2811 // during Phase 4 if needed. 2812 orig_phis.append_if_missing(mphi); 2813 } else if (C->get_alias_index(t) != alias_idx) { 2814 // Create a new Phi with the specified alias index type. 2815 result = split_memory_phi(mphi, alias_idx, orig_phis); 2816 } 2817 } 2818 // the result is either MemNode, PhiNode, InitializeNode. 2819 return result; 2820 } 2821 2822 // 2823 // Convert the types of unescaped object to instance types where possible, 2824 // propagate the new type information through the graph, and update memory 2825 // edges and MergeMem inputs to reflect the new type. 2826 // 2827 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2828 // The processing is done in 4 phases: 2829 // 2830 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2831 // types for the CheckCastPP for allocations where possible. 2832 // Propagate the new types through users as follows: 2833 // casts and Phi: push users on alloc_worklist 2834 // AddP: cast Base and Address inputs to the instance type 2835 // push any AddP users on alloc_worklist and push any memnode 2836 // users onto memnode_worklist. 2837 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2838 // search the Memory chain for a store with the appropriate type 2839 // address type. If a Phi is found, create a new version with 2840 // the appropriate memory slices from each of the Phi inputs. 2841 // For stores, process the users as follows: 2842 // MemNode: push on memnode_worklist 2843 // MergeMem: push on mergemem_worklist 2844 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2845 // moving the first node encountered of each instance type to the 2846 // the input corresponding to its alias index. 2847 // appropriate memory slice. 2848 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2849 // 2850 // In the following example, the CheckCastPP nodes are the cast of allocation 2851 // results and the allocation of node 29 is unescaped and eligible to be an 2852 // instance type. 2853 // 2854 // We start with: 2855 // 2856 // 7 Parm #memory 2857 // 10 ConI "12" 2858 // 19 CheckCastPP "Foo" 2859 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2860 // 29 CheckCastPP "Foo" 2861 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2862 // 2863 // 40 StoreP 25 7 20 ... alias_index=4 2864 // 50 StoreP 35 40 30 ... alias_index=4 2865 // 60 StoreP 45 50 20 ... alias_index=4 2866 // 70 LoadP _ 60 30 ... alias_index=4 2867 // 80 Phi 75 50 60 Memory alias_index=4 2868 // 90 LoadP _ 80 30 ... alias_index=4 2869 // 100 LoadP _ 80 20 ... alias_index=4 2870 // 2871 // 2872 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2873 // and creating a new alias index for node 30. This gives: 2874 // 2875 // 7 Parm #memory 2876 // 10 ConI "12" 2877 // 19 CheckCastPP "Foo" 2878 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2879 // 29 CheckCastPP "Foo" iid=24 2880 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2881 // 2882 // 40 StoreP 25 7 20 ... alias_index=4 2883 // 50 StoreP 35 40 30 ... alias_index=6 2884 // 60 StoreP 45 50 20 ... alias_index=4 2885 // 70 LoadP _ 60 30 ... alias_index=6 2886 // 80 Phi 75 50 60 Memory alias_index=4 2887 // 90 LoadP _ 80 30 ... alias_index=6 2888 // 100 LoadP _ 80 20 ... alias_index=4 2889 // 2890 // In phase 2, new memory inputs are computed for the loads and stores, 2891 // And a new version of the phi is created. In phase 4, the inputs to 2892 // node 80 are updated and then the memory nodes are updated with the 2893 // values computed in phase 2. This results in: 2894 // 2895 // 7 Parm #memory 2896 // 10 ConI "12" 2897 // 19 CheckCastPP "Foo" 2898 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2899 // 29 CheckCastPP "Foo" iid=24 2900 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2901 // 2902 // 40 StoreP 25 7 20 ... alias_index=4 2903 // 50 StoreP 35 7 30 ... alias_index=6 2904 // 60 StoreP 45 40 20 ... alias_index=4 2905 // 70 LoadP _ 50 30 ... alias_index=6 2906 // 80 Phi 75 40 60 Memory alias_index=4 2907 // 120 Phi 75 50 50 Memory alias_index=6 2908 // 90 LoadP _ 120 30 ... alias_index=6 2909 // 100 LoadP _ 80 20 ... alias_index=4 2910 // 2911 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 2912 GrowableArray<Node *> memnode_worklist; 2913 GrowableArray<PhiNode *> orig_phis; 2914 PhaseIterGVN *igvn = _igvn; 2915 uint new_index_start = (uint) _compile->num_alias_types(); 2916 Arena* arena = Thread::current()->resource_area(); 2917 VectorSet visited(arena); 2918 ideal_nodes.clear(); // Reset for use with set_map/get_map. 2919 uint unique_old = _compile->unique(); 2920 2921 // Phase 1: Process possible allocations from alloc_worklist. 2922 // Create instance types for the CheckCastPP for allocations where possible. 2923 // 2924 // (Note: don't forget to change the order of the second AddP node on 2925 // the alloc_worklist if the order of the worklist processing is changed, 2926 // see the comment in find_second_addp().) 2927 // 2928 while (alloc_worklist.length() != 0) { 2929 Node *n = alloc_worklist.pop(); 2930 uint ni = n->_idx; 2931 if (n->is_Call()) { 2932 CallNode *alloc = n->as_Call(); 2933 // copy escape information to call node 2934 PointsToNode* ptn = ptnode_adr(alloc->_idx); 2935 PointsToNode::EscapeState es = ptn->escape_state(); 2936 // We have an allocation or call which returns a Java object, 2937 // see if it is unescaped. 2938 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 2939 continue; 2940 // Find CheckCastPP for the allocate or for the return value of a call 2941 n = alloc->result_cast(); 2942 if (n == NULL) { // No uses except Initialize node 2943 if (alloc->is_Allocate()) { 2944 // Set the scalar_replaceable flag for allocation 2945 // so it could be eliminated if it has no uses. 2946 alloc->as_Allocate()->_is_scalar_replaceable = true; 2947 } 2948 if (alloc->is_CallStaticJava()) { 2949 // Set the scalar_replaceable flag for boxing method 2950 // so it could be eliminated if it has no uses. 2951 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 2952 } 2953 continue; 2954 } 2955 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 2956 assert(!alloc->is_Allocate(), "allocation should have unique type"); 2957 continue; 2958 } 2959 2960 // The inline code for Object.clone() casts the allocation result to 2961 // java.lang.Object and then to the actual type of the allocated 2962 // object. Detect this case and use the second cast. 2963 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 2964 // the allocation result is cast to java.lang.Object and then 2965 // to the actual Array type. 2966 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 2967 && (alloc->is_AllocateArray() || 2968 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 2969 Node *cast2 = NULL; 2970 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2971 Node *use = n->fast_out(i); 2972 if (use->is_CheckCastPP()) { 2973 cast2 = use; 2974 break; 2975 } 2976 } 2977 if (cast2 != NULL) { 2978 n = cast2; 2979 } else { 2980 // Non-scalar replaceable if the allocation type is unknown statically 2981 // (reflection allocation), the object can't be restored during 2982 // deoptimization without precise type. 2983 continue; 2984 } 2985 } 2986 2987 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 2988 if (t == NULL) 2989 continue; // not a TypeOopPtr 2990 if (!t->klass_is_exact()) 2991 continue; // not an unique type 2992 2993 if (alloc->is_Allocate()) { 2994 // Set the scalar_replaceable flag for allocation 2995 // so it could be eliminated. 2996 alloc->as_Allocate()->_is_scalar_replaceable = true; 2997 } 2998 if (alloc->is_CallStaticJava()) { 2999 // Set the scalar_replaceable flag for boxing method 3000 // so it could be eliminated. 3001 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3002 } 3003 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 3004 // in order for an object to be scalar-replaceable, it must be: 3005 // - a direct allocation (not a call returning an object) 3006 // - non-escaping 3007 // - eligible to be a unique type 3008 // - not determined to be ineligible by escape analysis 3009 set_map(alloc, n); 3010 set_map(n, alloc); 3011 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 3012 igvn->hash_delete(n); 3013 igvn->set_type(n, tinst); 3014 n->raise_bottom_type(tinst); 3015 igvn->hash_insert(n); 3016 record_for_optimizer(n); 3017 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 3018 3019 // First, put on the worklist all Field edges from Connection Graph 3020 // which is more accurate than putting immediate users from Ideal Graph. 3021 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 3022 PointsToNode* tgt = e.get(); 3023 if (tgt->is_Arraycopy()) { 3024 continue; 3025 } 3026 Node* use = tgt->ideal_node(); 3027 assert(tgt->is_Field() && use->is_AddP(), 3028 "only AddP nodes are Field edges in CG"); 3029 if (use->outcnt() > 0) { // Don't process dead nodes 3030 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 3031 if (addp2 != NULL) { 3032 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3033 alloc_worklist.append_if_missing(addp2); 3034 } 3035 alloc_worklist.append_if_missing(use); 3036 } 3037 } 3038 3039 // An allocation may have an Initialize which has raw stores. Scan 3040 // the users of the raw allocation result and push AddP users 3041 // on alloc_worklist. 3042 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms); 3043 assert (raw_result != NULL, "must have an allocation result"); 3044 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 3045 Node *use = raw_result->fast_out(i); 3046 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3047 Node* addp2 = find_second_addp(use, raw_result); 3048 if (addp2 != NULL) { 3049 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3050 alloc_worklist.append_if_missing(addp2); 3051 } 3052 alloc_worklist.append_if_missing(use); 3053 } else if (use->is_MemBar()) { 3054 memnode_worklist.append_if_missing(use); 3055 } 3056 } 3057 } 3058 } else if (n->is_AddP()) { 3059 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3060 if (jobj == NULL || jobj == phantom_obj) { 3061 #ifdef ASSERT 3062 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3063 ptnode_adr(n->_idx)->dump(); 3064 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3065 #endif 3066 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3067 return; 3068 } 3069 Node *base = get_map(jobj->idx()); // CheckCastPP node 3070 if (!split_AddP(n, base)) continue; // wrong type from dead path 3071 } else if (n->is_Phi() || 3072 n->is_CheckCastPP() || 3073 n->is_EncodeP() || 3074 n->is_DecodeN() || 3075 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(n) || 3076 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3077 if (visited.test_set(n->_idx)) { 3078 assert(n->is_Phi(), "loops only through Phi's"); 3079 continue; // already processed 3080 } 3081 JavaObjectNode* jobj = unique_java_object(n); 3082 if (jobj == NULL || jobj == phantom_obj) { 3083 #ifdef ASSERT 3084 ptnode_adr(n->_idx)->dump(); 3085 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3086 #endif 3087 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3088 return; 3089 } else { 3090 Node *val = get_map(jobj->idx()); // CheckCastPP node 3091 TypeNode *tn = n->as_Type(); 3092 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3093 assert(tinst != NULL && tinst->is_known_instance() && 3094 tinst->instance_id() == jobj->idx() , "instance type expected."); 3095 3096 const Type *tn_type = igvn->type(tn); 3097 const TypeOopPtr *tn_t; 3098 if (tn_type->isa_narrowoop()) { 3099 tn_t = tn_type->make_ptr()->isa_oopptr(); 3100 } else { 3101 tn_t = tn_type->isa_oopptr(); 3102 } 3103 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3104 if (tn_type->isa_narrowoop()) { 3105 tn_type = tinst->make_narrowoop(); 3106 } else { 3107 tn_type = tinst; 3108 } 3109 igvn->hash_delete(tn); 3110 igvn->set_type(tn, tn_type); 3111 tn->set_type(tn_type); 3112 igvn->hash_insert(tn); 3113 record_for_optimizer(n); 3114 } else { 3115 assert(tn_type == TypePtr::NULL_PTR || 3116 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3117 "unexpected type"); 3118 continue; // Skip dead path with different type 3119 } 3120 } 3121 } else { 3122 debug_only(n->dump();) 3123 assert(false, "EA: unexpected node"); 3124 continue; 3125 } 3126 // push allocation's users on appropriate worklist 3127 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3128 Node *use = n->fast_out(i); 3129 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3130 // Load/store to instance's field 3131 memnode_worklist.append_if_missing(use); 3132 } else if (use->is_MemBar()) { 3133 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3134 memnode_worklist.append_if_missing(use); 3135 } 3136 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3137 Node* addp2 = find_second_addp(use, n); 3138 if (addp2 != NULL) { 3139 alloc_worklist.append_if_missing(addp2); 3140 } 3141 alloc_worklist.append_if_missing(use); 3142 } else if (use->is_Phi() || 3143 use->is_CheckCastPP() || 3144 use->is_EncodeNarrowPtr() || 3145 use->is_DecodeNarrowPtr() || 3146 BarrierSet::barrier_set()->barrier_set_c2()->escape_is_barrier_node(use) || 3147 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3148 alloc_worklist.append_if_missing(use); 3149 #ifdef ASSERT 3150 } else if (use->is_Mem()) { 3151 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3152 } else if (use->is_MergeMem()) { 3153 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3154 } else if (use->is_SafePoint()) { 3155 // Look for MergeMem nodes for calls which reference unique allocation 3156 // (through CheckCastPP nodes) even for debug info. 3157 Node* m = use->in(TypeFunc::Memory); 3158 if (m->is_MergeMem()) { 3159 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3160 } 3161 } else if (use->Opcode() == Op_EncodeISOArray) { 3162 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3163 // EncodeISOArray overwrites destination array 3164 memnode_worklist.append_if_missing(use); 3165 } 3166 } else { 3167 uint op = use->Opcode(); 3168 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) && 3169 (use->in(MemNode::Memory) == n)) { 3170 // They overwrite memory edge corresponding to destination array, 3171 memnode_worklist.append_if_missing(use); 3172 } else if (!(op == Op_CmpP || op == Op_Conv2B || 3173 op == Op_CastP2X || op == Op_StoreCM || 3174 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3175 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3176 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || 3177 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) { 3178 n->dump(); 3179 use->dump(); 3180 assert(false, "EA: missing allocation reference path"); 3181 } 3182 #endif 3183 } 3184 } 3185 3186 } 3187 3188 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3189 // type, record it in the ArrayCopy node so we know what memory this 3190 // node uses/modified. 3191 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3192 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3193 Node* dest = ac->in(ArrayCopyNode::Dest); 3194 if (dest->is_AddP()) { 3195 dest = get_addp_base(dest); 3196 } 3197 JavaObjectNode* jobj = unique_java_object(dest); 3198 if (jobj != NULL) { 3199 Node *base = get_map(jobj->idx()); 3200 if (base != NULL) { 3201 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3202 ac->_dest_type = base_t; 3203 } 3204 } 3205 Node* src = ac->in(ArrayCopyNode::Src); 3206 if (src->is_AddP()) { 3207 src = get_addp_base(src); 3208 } 3209 jobj = unique_java_object(src); 3210 if (jobj != NULL) { 3211 Node* base = get_map(jobj->idx()); 3212 if (base != NULL) { 3213 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3214 ac->_src_type = base_t; 3215 } 3216 } 3217 } 3218 3219 // New alias types were created in split_AddP(). 3220 uint new_index_end = (uint) _compile->num_alias_types(); 3221 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3222 3223 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3224 // compute new values for Memory inputs (the Memory inputs are not 3225 // actually updated until phase 4.) 3226 if (memnode_worklist.length() == 0) 3227 return; // nothing to do 3228 while (memnode_worklist.length() != 0) { 3229 Node *n = memnode_worklist.pop(); 3230 if (visited.test_set(n->_idx)) 3231 continue; 3232 if (n->is_Phi() || n->is_ClearArray()) { 3233 // we don't need to do anything, but the users must be pushed 3234 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3235 // we don't need to do anything, but the users must be pushed 3236 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory); 3237 if (n == NULL) 3238 continue; 3239 } else if (n->Opcode() == Op_StrCompressedCopy || 3240 n->Opcode() == Op_EncodeISOArray) { 3241 // get the memory projection 3242 n = n->find_out_with(Op_SCMemProj); 3243 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3244 } else { 3245 assert(n->is_Mem(), "memory node required."); 3246 Node *addr = n->in(MemNode::Address); 3247 const Type *addr_t = igvn->type(addr); 3248 if (addr_t == Type::TOP) 3249 continue; 3250 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3251 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3252 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3253 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3254 if (_compile->failing()) { 3255 return; 3256 } 3257 if (mem != n->in(MemNode::Memory)) { 3258 // We delay the memory edge update since we need old one in 3259 // MergeMem code below when instances memory slices are separated. 3260 set_map(n, mem); 3261 } 3262 if (n->is_Load()) { 3263 continue; // don't push users 3264 } else if (n->is_LoadStore()) { 3265 // get the memory projection 3266 n = n->find_out_with(Op_SCMemProj); 3267 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3268 } 3269 } 3270 // push user on appropriate worklist 3271 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3272 Node *use = n->fast_out(i); 3273 if (use->is_Phi() || use->is_ClearArray()) { 3274 memnode_worklist.append_if_missing(use); 3275 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3276 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3277 continue; 3278 memnode_worklist.append_if_missing(use); 3279 } else if (use->is_MemBar()) { 3280 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3281 memnode_worklist.append_if_missing(use); 3282 } 3283 #ifdef ASSERT 3284 } else if(use->is_Mem()) { 3285 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3286 } else if (use->is_MergeMem()) { 3287 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3288 } else if (use->Opcode() == Op_EncodeISOArray) { 3289 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3290 // EncodeISOArray overwrites destination array 3291 memnode_worklist.append_if_missing(use); 3292 } 3293 } else { 3294 uint op = use->Opcode(); 3295 if ((use->in(MemNode::Memory) == n) && 3296 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) { 3297 // They overwrite memory edge corresponding to destination array, 3298 memnode_worklist.append_if_missing(use); 3299 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) || 3300 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3301 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3302 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3303 n->dump(); 3304 use->dump(); 3305 assert(false, "EA: missing memory path"); 3306 } 3307 #endif 3308 } 3309 } 3310 } 3311 3312 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3313 // Walk each memory slice moving the first node encountered of each 3314 // instance type to the the input corresponding to its alias index. 3315 uint length = _mergemem_worklist.length(); 3316 for( uint next = 0; next < length; ++next ) { 3317 MergeMemNode* nmm = _mergemem_worklist.at(next); 3318 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3319 // Note: we don't want to use MergeMemStream here because we only want to 3320 // scan inputs which exist at the start, not ones we add during processing. 3321 // Note 2: MergeMem may already contains instance memory slices added 3322 // during find_inst_mem() call when memory nodes were processed above. 3323 igvn->hash_delete(nmm); 3324 uint nslices = MIN2(nmm->req(), new_index_start); 3325 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3326 Node* mem = nmm->in(i); 3327 Node* cur = NULL; 3328 if (mem == NULL || mem->is_top()) 3329 continue; 3330 // First, update mergemem by moving memory nodes to corresponding slices 3331 // if their type became more precise since this mergemem was created. 3332 while (mem->is_Mem()) { 3333 const Type *at = igvn->type(mem->in(MemNode::Address)); 3334 if (at != Type::TOP) { 3335 assert (at->isa_ptr() != NULL, "pointer type required."); 3336 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3337 if (idx == i) { 3338 if (cur == NULL) 3339 cur = mem; 3340 } else { 3341 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3342 nmm->set_memory_at(idx, mem); 3343 } 3344 } 3345 } 3346 mem = mem->in(MemNode::Memory); 3347 } 3348 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3349 // Find any instance of the current type if we haven't encountered 3350 // already a memory slice of the instance along the memory chain. 3351 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3352 if((uint)_compile->get_general_index(ni) == i) { 3353 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3354 if (nmm->is_empty_memory(m)) { 3355 Node* result = find_inst_mem(mem, ni, orig_phis); 3356 if (_compile->failing()) { 3357 return; 3358 } 3359 nmm->set_memory_at(ni, result); 3360 } 3361 } 3362 } 3363 } 3364 // Find the rest of instances values 3365 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3366 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3367 Node* result = step_through_mergemem(nmm, ni, tinst); 3368 if (result == nmm->base_memory()) { 3369 // Didn't find instance memory, search through general slice recursively. 3370 result = nmm->memory_at(_compile->get_general_index(ni)); 3371 result = find_inst_mem(result, ni, orig_phis); 3372 if (_compile->failing()) { 3373 return; 3374 } 3375 nmm->set_memory_at(ni, result); 3376 } 3377 } 3378 igvn->hash_insert(nmm); 3379 record_for_optimizer(nmm); 3380 } 3381 3382 // Phase 4: Update the inputs of non-instance memory Phis and 3383 // the Memory input of memnodes 3384 // First update the inputs of any non-instance Phi's from 3385 // which we split out an instance Phi. Note we don't have 3386 // to recursively process Phi's encounted on the input memory 3387 // chains as is done in split_memory_phi() since they will 3388 // also be processed here. 3389 for (int j = 0; j < orig_phis.length(); j++) { 3390 PhiNode *phi = orig_phis.at(j); 3391 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3392 igvn->hash_delete(phi); 3393 for (uint i = 1; i < phi->req(); i++) { 3394 Node *mem = phi->in(i); 3395 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3396 if (_compile->failing()) { 3397 return; 3398 } 3399 if (mem != new_mem) { 3400 phi->set_req(i, new_mem); 3401 } 3402 } 3403 igvn->hash_insert(phi); 3404 record_for_optimizer(phi); 3405 } 3406 3407 // Update the memory inputs of MemNodes with the value we computed 3408 // in Phase 2 and move stores memory users to corresponding memory slices. 3409 // Disable memory split verification code until the fix for 6984348. 3410 // Currently it produces false negative results since it does not cover all cases. 3411 #if 0 // ifdef ASSERT 3412 visited.Reset(); 3413 Node_Stack old_mems(arena, _compile->unique() >> 2); 3414 #endif 3415 for (uint i = 0; i < ideal_nodes.size(); i++) { 3416 Node* n = ideal_nodes.at(i); 3417 Node* nmem = get_map(n->_idx); 3418 assert(nmem != NULL, "sanity"); 3419 if (n->is_Mem()) { 3420 #if 0 // ifdef ASSERT 3421 Node* old_mem = n->in(MemNode::Memory); 3422 if (!visited.test_set(old_mem->_idx)) { 3423 old_mems.push(old_mem, old_mem->outcnt()); 3424 } 3425 #endif 3426 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3427 if (!n->is_Load()) { 3428 // Move memory users of a store first. 3429 move_inst_mem(n, orig_phis); 3430 } 3431 // Now update memory input 3432 igvn->hash_delete(n); 3433 n->set_req(MemNode::Memory, nmem); 3434 igvn->hash_insert(n); 3435 record_for_optimizer(n); 3436 } else { 3437 assert(n->is_Allocate() || n->is_CheckCastPP() || 3438 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3439 } 3440 } 3441 #if 0 // ifdef ASSERT 3442 // Verify that memory was split correctly 3443 while (old_mems.is_nonempty()) { 3444 Node* old_mem = old_mems.node(); 3445 uint old_cnt = old_mems.index(); 3446 old_mems.pop(); 3447 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3448 } 3449 #endif 3450 } 3451 3452 #ifndef PRODUCT 3453 static const char *node_type_names[] = { 3454 "UnknownType", 3455 "JavaObject", 3456 "LocalVar", 3457 "Field", 3458 "Arraycopy" 3459 }; 3460 3461 static const char *esc_names[] = { 3462 "UnknownEscape", 3463 "NoEscape", 3464 "ArgEscape", 3465 "GlobalEscape" 3466 }; 3467 3468 void PointsToNode::dump(bool print_state) const { 3469 NodeType nt = node_type(); 3470 tty->print("%s ", node_type_names[(int) nt]); 3471 if (print_state) { 3472 EscapeState es = escape_state(); 3473 EscapeState fields_es = fields_escape_state(); 3474 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3475 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3476 tty->print("NSR "); 3477 } 3478 if (is_Field()) { 3479 FieldNode* f = (FieldNode*)this; 3480 if (f->is_oop()) 3481 tty->print("oop "); 3482 if (f->offset() > 0) 3483 tty->print("+%d ", f->offset()); 3484 tty->print("("); 3485 for (BaseIterator i(f); i.has_next(); i.next()) { 3486 PointsToNode* b = i.get(); 3487 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3488 } 3489 tty->print(" )"); 3490 } 3491 tty->print("["); 3492 for (EdgeIterator i(this); i.has_next(); i.next()) { 3493 PointsToNode* e = i.get(); 3494 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3495 } 3496 tty->print(" ["); 3497 for (UseIterator i(this); i.has_next(); i.next()) { 3498 PointsToNode* u = i.get(); 3499 bool is_base = false; 3500 if (PointsToNode::is_base_use(u)) { 3501 is_base = true; 3502 u = PointsToNode::get_use_node(u)->as_Field(); 3503 } 3504 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3505 } 3506 tty->print(" ]] "); 3507 if (_node == NULL) 3508 tty->print_cr("<null>"); 3509 else 3510 _node->dump(); 3511 } 3512 3513 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3514 bool first = true; 3515 int ptnodes_length = ptnodes_worklist.length(); 3516 for (int i = 0; i < ptnodes_length; i++) { 3517 PointsToNode *ptn = ptnodes_worklist.at(i); 3518 if (ptn == NULL || !ptn->is_JavaObject()) 3519 continue; 3520 PointsToNode::EscapeState es = ptn->escape_state(); 3521 if ((es != PointsToNode::NoEscape) && !Verbose) { 3522 continue; 3523 } 3524 Node* n = ptn->ideal_node(); 3525 if (n->is_Allocate() || (n->is_CallStaticJava() && 3526 n->as_CallStaticJava()->is_boxing_method())) { 3527 if (first) { 3528 tty->cr(); 3529 tty->print("======== Connection graph for "); 3530 _compile->method()->print_short_name(); 3531 tty->cr(); 3532 first = false; 3533 } 3534 ptn->dump(); 3535 // Print all locals and fields which reference this allocation 3536 for (UseIterator j(ptn); j.has_next(); j.next()) { 3537 PointsToNode* use = j.get(); 3538 if (use->is_LocalVar()) { 3539 use->dump(Verbose); 3540 } else if (Verbose) { 3541 use->dump(); 3542 } 3543 } 3544 tty->cr(); 3545 } 3546 } 3547 } 3548 #endif 3549 3550 void ConnectionGraph::record_for_optimizer(Node *n) { 3551 _igvn->_worklist.push(n); 3552 _igvn->add_users_to_worklist(n); 3553 }