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/c2/barrierSetC2.hpp" 29 #include "libadt/vectset.hpp" 30 #include "memory/allocation.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "opto/c2compiler.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callnode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/escape.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/rootnode.hpp" 41 #include "utilities/macros.hpp" 42 #if INCLUDE_G1GC 43 #include "gc/g1/g1ThreadLocalData.hpp" 44 #endif // INCLUDE_G1GC 45 #if INCLUDE_ZGC 46 #include "gc/z/c2/zBarrierSetC2.hpp" 47 #endif 48 #if INCLUDE_SHENANDOAHGC 49 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 50 #endif 51 52 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 53 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL), 54 _in_worklist(C->comp_arena()), 55 _next_pidx(0), 56 _collecting(true), 57 _verify(false), 58 _compile(C), 59 _igvn(igvn), 60 _node_map(C->comp_arena()) { 61 // Add unknown java object. 62 add_java_object(C->top(), PointsToNode::GlobalEscape); 63 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject(); 64 // Add ConP(#NULL) and ConN(#NULL) nodes. 65 Node* oop_null = igvn->zerocon(T_OBJECT); 66 assert(oop_null->_idx < nodes_size(), "should be created already"); 67 add_java_object(oop_null, PointsToNode::NoEscape); 68 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject(); 69 if (UseCompressedOops) { 70 Node* noop_null = igvn->zerocon(T_NARROWOOP); 71 assert(noop_null->_idx < nodes_size(), "should be created already"); 72 map_ideal_node(noop_null, null_obj); 73 } 74 _pcmp_neq = NULL; // Should be initialized 75 _pcmp_eq = NULL; 76 } 77 78 bool ConnectionGraph::has_candidates(Compile *C) { 79 // EA brings benefits only when the code has allocations and/or locks which 80 // are represented by ideal Macro nodes. 81 int cnt = C->macro_count(); 82 for (int i = 0; i < cnt; i++) { 83 Node *n = C->macro_node(i); 84 if (n->is_Allocate()) 85 return true; 86 if (n->is_Lock()) { 87 Node* obj = n->as_Lock()->obj_node()->uncast(); 88 if (!(obj->is_Parm() || obj->is_Con())) 89 return true; 90 } 91 if (n->is_CallStaticJava() && 92 n->as_CallStaticJava()->is_boxing_method()) { 93 return true; 94 } 95 } 96 return false; 97 } 98 99 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 100 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]); 101 ResourceMark rm; 102 103 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 104 // to create space for them in ConnectionGraph::_nodes[]. 105 Node* oop_null = igvn->zerocon(T_OBJECT); 106 Node* noop_null = igvn->zerocon(T_NARROWOOP); 107 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 108 // Perform escape analysis 109 if (congraph->compute_escape()) { 110 // There are non escaping objects. 111 C->set_congraph(congraph); 112 } 113 // Cleanup. 114 if (oop_null->outcnt() == 0) 115 igvn->hash_delete(oop_null); 116 if (noop_null->outcnt() == 0) 117 igvn->hash_delete(noop_null); 118 } 119 120 bool ConnectionGraph::compute_escape() { 121 Compile* C = _compile; 122 PhaseGVN* igvn = _igvn; 123 124 // Worklists used by EA. 125 Unique_Node_List delayed_worklist; 126 GrowableArray<Node*> alloc_worklist; 127 GrowableArray<Node*> ptr_cmp_worklist; 128 GrowableArray<Node*> storestore_worklist; 129 GrowableArray<ArrayCopyNode*> arraycopy_worklist; 130 GrowableArray<PointsToNode*> ptnodes_worklist; 131 GrowableArray<JavaObjectNode*> java_objects_worklist; 132 GrowableArray<JavaObjectNode*> non_escaped_worklist; 133 GrowableArray<FieldNode*> oop_fields_worklist; 134 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; ) 135 136 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]); 137 138 // 1. Populate Connection Graph (CG) with PointsTo nodes. 139 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space 140 // Initialize worklist 141 if (C->root() != NULL) { 142 ideal_nodes.push(C->root()); 143 } 144 // Processed ideal nodes are unique on ideal_nodes list 145 // but several ideal nodes are mapped to the phantom_obj. 146 // To avoid duplicated entries on the following worklists 147 // add the phantom_obj only once to them. 148 ptnodes_worklist.append(phantom_obj); 149 java_objects_worklist.append(phantom_obj); 150 for( uint next = 0; next < ideal_nodes.size(); ++next ) { 151 Node* n = ideal_nodes.at(next); 152 // Create PointsTo nodes and add them to Connection Graph. Called 153 // only once per ideal node since ideal_nodes is Unique_Node list. 154 add_node_to_connection_graph(n, &delayed_worklist); 155 PointsToNode* ptn = ptnode_adr(n->_idx); 156 if (ptn != NULL && ptn != phantom_obj) { 157 ptnodes_worklist.append(ptn); 158 if (ptn->is_JavaObject()) { 159 java_objects_worklist.append(ptn->as_JavaObject()); 160 if ((n->is_Allocate() || n->is_CallStaticJava()) && 161 (ptn->escape_state() < PointsToNode::GlobalEscape)) { 162 // Only allocations and java static calls results are interesting. 163 non_escaped_worklist.append(ptn->as_JavaObject()); 164 } 165 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) { 166 oop_fields_worklist.append(ptn->as_Field()); 167 } 168 } 169 if (n->is_MergeMem()) { 170 // Collect all MergeMem nodes to add memory slices for 171 // scalar replaceable objects in split_unique_types(). 172 _mergemem_worklist.append(n->as_MergeMem()); 173 } else if (OptimizePtrCompare && n->is_Cmp() && 174 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) { 175 // Collect compare pointers nodes. 176 ptr_cmp_worklist.append(n); 177 } else if (n->is_MemBarStoreStore()) { 178 // Collect all MemBarStoreStore nodes so that depending on the 179 // escape status of the associated Allocate node some of them 180 // may be eliminated. 181 storestore_worklist.append(n); 182 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) && 183 (n->req() > MemBarNode::Precedent)) { 184 record_for_optimizer(n); 185 #ifdef ASSERT 186 } else if (n->is_AddP()) { 187 // Collect address nodes for graph verification. 188 addp_worklist.append(n); 189 #endif 190 } else if (n->is_ArrayCopy()) { 191 // Keep a list of ArrayCopy nodes so if one of its input is non 192 // escaping, we can record a unique type 193 arraycopy_worklist.append(n->as_ArrayCopy()); 194 } 195 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 196 Node* m = n->fast_out(i); // Get user 197 ideal_nodes.push(m); 198 } 199 } 200 if (non_escaped_worklist.length() == 0) { 201 _collecting = false; 202 return false; // Nothing to do. 203 } 204 // Add final simple edges to graph. 205 while(delayed_worklist.size() > 0) { 206 Node* n = delayed_worklist.pop(); 207 add_final_edges(n); 208 } 209 int ptnodes_length = ptnodes_worklist.length(); 210 211 #ifdef ASSERT 212 if (VerifyConnectionGraph) { 213 // Verify that no new simple edges could be created and all 214 // local vars has edges. 215 _verify = true; 216 for (int next = 0; next < ptnodes_length; ++next) { 217 PointsToNode* ptn = ptnodes_worklist.at(next); 218 add_final_edges(ptn->ideal_node()); 219 if (ptn->is_LocalVar() && ptn->edge_count() == 0) { 220 ptn->dump(); 221 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity"); 222 } 223 } 224 _verify = false; 225 } 226 #endif 227 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes 228 // processing, calls to CI to resolve symbols (types, fields, methods) 229 // referenced in bytecode. During symbol resolution VM may throw 230 // an exception which CI cleans and converts to compilation failure. 231 if (C->failing()) return false; 232 233 // 2. Finish Graph construction by propagating references to all 234 // java objects through graph. 235 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist, 236 java_objects_worklist, oop_fields_worklist)) { 237 // All objects escaped or hit time or iterations limits. 238 _collecting = false; 239 return false; 240 } 241 242 // 3. Adjust scalar_replaceable state of nonescaping objects and push 243 // scalar replaceable allocations on alloc_worklist for processing 244 // in split_unique_types(). 245 int non_escaped_length = non_escaped_worklist.length(); 246 for (int next = 0; next < non_escaped_length; next++) { 247 JavaObjectNode* ptn = non_escaped_worklist.at(next); 248 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape); 249 Node* n = ptn->ideal_node(); 250 if (n->is_Allocate()) { 251 n->as_Allocate()->_is_non_escaping = noescape; 252 } 253 if (n->is_CallStaticJava()) { 254 n->as_CallStaticJava()->_is_non_escaping = noescape; 255 } 256 if (noescape && ptn->scalar_replaceable()) { 257 adjust_scalar_replaceable_state(ptn); 258 if (ptn->scalar_replaceable()) { 259 alloc_worklist.append(ptn->ideal_node()); 260 } 261 } 262 } 263 264 #ifdef ASSERT 265 if (VerifyConnectionGraph) { 266 // Verify that graph is complete - no new edges could be added or needed. 267 verify_connection_graph(ptnodes_worklist, non_escaped_worklist, 268 java_objects_worklist, addp_worklist); 269 } 270 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build"); 271 assert(null_obj->escape_state() == PointsToNode::NoEscape && 272 null_obj->edge_count() == 0 && 273 !null_obj->arraycopy_src() && 274 !null_obj->arraycopy_dst(), "sanity"); 275 #endif 276 277 _collecting = false; 278 279 } // TracePhase t3("connectionGraph") 280 281 // 4. Optimize ideal graph based on EA information. 282 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0); 283 if (has_non_escaping_obj) { 284 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist); 285 } 286 287 #ifndef PRODUCT 288 if (PrintEscapeAnalysis) { 289 dump(ptnodes_worklist); // Dump ConnectionGraph 290 } 291 #endif 292 293 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0); 294 #ifdef ASSERT 295 if (VerifyConnectionGraph) { 296 int alloc_length = alloc_worklist.length(); 297 for (int next = 0; next < alloc_length; ++next) { 298 Node* n = alloc_worklist.at(next); 299 PointsToNode* ptn = ptnode_adr(n->_idx); 300 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity"); 301 } 302 } 303 #endif 304 305 // 5. Separate memory graph for scalar replaceable allcations. 306 if (has_scalar_replaceable_candidates && 307 C->AliasLevel() >= 3 && EliminateAllocations) { 308 // Now use the escape information to create unique types for 309 // scalar replaceable objects. 310 split_unique_types(alloc_worklist, arraycopy_worklist); 311 if (C->failing()) return false; 312 C->print_method(PHASE_AFTER_EA, 2); 313 314 #ifdef ASSERT 315 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 316 tty->print("=== No allocations eliminated for "); 317 C->method()->print_short_name(); 318 if(!EliminateAllocations) { 319 tty->print(" since EliminateAllocations is off ==="); 320 } else if(!has_scalar_replaceable_candidates) { 321 tty->print(" since there are no scalar replaceable candidates ==="); 322 } else if(C->AliasLevel() < 3) { 323 tty->print(" since AliasLevel < 3 ==="); 324 } 325 tty->cr(); 326 #endif 327 } 328 return has_non_escaping_obj; 329 } 330 331 // Utility function for nodes that load an object 332 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 333 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 334 // ThreadLocal has RawPtr type. 335 const Type* t = _igvn->type(n); 336 if (t->make_ptr() != NULL) { 337 Node* adr = n->in(MemNode::Address); 338 #ifdef ASSERT 339 if (!adr->is_AddP()) { 340 assert(_igvn->type(adr)->isa_rawptr(), "sanity"); 341 } else { 342 assert((ptnode_adr(adr->_idx) == NULL || 343 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity"); 344 } 345 #endif 346 add_local_var_and_edge(n, PointsToNode::NoEscape, 347 adr, delayed_worklist); 348 } 349 } 350 351 // Populate Connection Graph with PointsTo nodes and create simple 352 // connection graph edges. 353 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) { 354 assert(!_verify, "this method should not be called for verification"); 355 PhaseGVN* igvn = _igvn; 356 uint n_idx = n->_idx; 357 PointsToNode* n_ptn = ptnode_adr(n_idx); 358 if (n_ptn != NULL) 359 return; // No need to redefine PointsTo node during first iteration. 360 361 if (n->is_Call()) { 362 // Arguments to allocation and locking don't escape. 363 if (n->is_AbstractLock()) { 364 // Put Lock and Unlock nodes on IGVN worklist to process them during 365 // first IGVN optimization when escape information is still available. 366 record_for_optimizer(n); 367 } else if (n->is_Allocate()) { 368 add_call_node(n->as_Call()); 369 record_for_optimizer(n); 370 } else { 371 if (n->is_CallStaticJava()) { 372 const char* name = n->as_CallStaticJava()->_name; 373 if (name != NULL && strcmp(name, "uncommon_trap") == 0) 374 return; // Skip uncommon traps 375 } 376 // Don't mark as processed since call's arguments have to be processed. 377 delayed_worklist->push(n); 378 // Check if a call returns an object. 379 if ((n->as_Call()->returns_pointer() && 380 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) || 381 (n->is_CallStaticJava() && 382 n->as_CallStaticJava()->is_boxing_method())) { 383 add_call_node(n->as_Call()); 384 } 385 } 386 return; 387 } 388 // Put this check here to process call arguments since some call nodes 389 // point to phantom_obj. 390 if (n_ptn == phantom_obj || n_ptn == null_obj) 391 return; // Skip predefined nodes. 392 393 int opcode = n->Opcode(); 394 switch (opcode) { 395 case Op_AddP: { 396 Node* base = get_addp_base(n); 397 PointsToNode* ptn_base = ptnode_adr(base->_idx); 398 // Field nodes are created for all field types. They are used in 399 // adjust_scalar_replaceable_state() and split_unique_types(). 400 // Note, non-oop fields will have only base edges in Connection 401 // Graph because such fields are not used for oop loads and stores. 402 int offset = address_offset(n, igvn); 403 add_field(n, PointsToNode::NoEscape, offset); 404 if (ptn_base == NULL) { 405 delayed_worklist->push(n); // Process it later. 406 } else { 407 n_ptn = ptnode_adr(n_idx); 408 add_base(n_ptn->as_Field(), ptn_base); 409 } 410 break; 411 } 412 case Op_CastX2P: { 413 map_ideal_node(n, phantom_obj); 414 break; 415 } 416 case Op_CastPP: 417 case Op_CheckCastPP: 418 case Op_EncodeP: 419 case Op_DecodeN: 420 case Op_EncodePKlass: 421 case Op_DecodeNKlass: { 422 add_local_var_and_edge(n, PointsToNode::NoEscape, 423 n->in(1), delayed_worklist); 424 break; 425 } 426 case Op_CMoveP: { 427 add_local_var(n, PointsToNode::NoEscape); 428 // Do not add edges during first iteration because some could be 429 // not defined yet. 430 delayed_worklist->push(n); 431 break; 432 } 433 case Op_ConP: 434 case Op_ConN: 435 case Op_ConNKlass: { 436 // assume all oop constants globally escape except for null 437 PointsToNode::EscapeState es; 438 const Type* t = igvn->type(n); 439 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) { 440 es = PointsToNode::NoEscape; 441 } else { 442 es = PointsToNode::GlobalEscape; 443 } 444 add_java_object(n, es); 445 break; 446 } 447 case Op_CreateEx: { 448 // assume that all exception objects globally escape 449 map_ideal_node(n, phantom_obj); 450 break; 451 } 452 case Op_LoadKlass: 453 case Op_LoadNKlass: { 454 // Unknown class is loaded 455 map_ideal_node(n, phantom_obj); 456 break; 457 } 458 case Op_LoadP: 459 #if INCLUDE_ZGC 460 case Op_LoadBarrierSlowReg: 461 case Op_LoadBarrierWeakSlowReg: 462 #endif 463 case Op_LoadN: 464 case Op_LoadPLocked: { 465 add_objload_to_connection_graph(n, delayed_worklist); 466 break; 467 } 468 case Op_Parm: { 469 map_ideal_node(n, phantom_obj); 470 break; 471 } 472 case Op_PartialSubtypeCheck: { 473 // Produces Null or notNull and is used in only in CmpP so 474 // phantom_obj could be used. 475 map_ideal_node(n, phantom_obj); // Result is unknown 476 break; 477 } 478 case Op_Phi: { 479 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 480 // ThreadLocal has RawPtr type. 481 const Type* t = n->as_Phi()->type(); 482 if (t->make_ptr() != NULL) { 483 add_local_var(n, PointsToNode::NoEscape); 484 // Do not add edges during first iteration because some could be 485 // not defined yet. 486 delayed_worklist->push(n); 487 } 488 break; 489 } 490 case Op_Proj: { 491 // we are only interested in the oop result projection from a call 492 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 493 n->in(0)->as_Call()->returns_pointer()) { 494 add_local_var_and_edge(n, PointsToNode::NoEscape, 495 n->in(0), delayed_worklist); 496 } 497 #if INCLUDE_ZGC 498 else if (UseZGC) { 499 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) { 500 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), delayed_worklist); 501 } 502 } 503 #endif 504 break; 505 } 506 case Op_Rethrow: // Exception object escapes 507 case Op_Return: { 508 if (n->req() > TypeFunc::Parms && 509 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 510 // Treat Return value as LocalVar with GlobalEscape escape state. 511 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 512 n->in(TypeFunc::Parms), delayed_worklist); 513 } 514 break; 515 } 516 case Op_CompareAndExchangeP: 517 case Op_CompareAndExchangeN: 518 #if INCLUDE_SHENANDOAHGC 519 case Op_ShenandoahCompareAndExchangeP: 520 case Op_ShenandoahCompareAndExchangeN: 521 #endif 522 case Op_GetAndSetP: 523 case Op_GetAndSetN: { 524 add_objload_to_connection_graph(n, delayed_worklist); 525 // fallthrough 526 } 527 case Op_StoreP: 528 case Op_StoreN: 529 case Op_StoreNKlass: 530 case Op_StorePConditional: 531 #if INCLUDE_SHENANDOAHGC 532 case Op_ShenandoahWeakCompareAndSwapP: 533 case Op_ShenandoahWeakCompareAndSwapN: 534 case Op_ShenandoahCompareAndSwapP: 535 case Op_ShenandoahCompareAndSwapN: 536 #endif 537 case Op_WeakCompareAndSwapP: 538 case Op_WeakCompareAndSwapN: 539 case Op_CompareAndSwapP: 540 case Op_CompareAndSwapN: { 541 Node* adr = n->in(MemNode::Address); 542 const Type *adr_type = igvn->type(adr); 543 adr_type = adr_type->make_ptr(); 544 if (adr_type == NULL) { 545 break; // skip dead nodes 546 } 547 if ( adr_type->isa_oopptr() 548 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 549 && adr_type == TypeRawPtr::NOTNULL 550 && adr->in(AddPNode::Address)->is_Proj() 551 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 552 delayed_worklist->push(n); // Process it later. 553 #ifdef ASSERT 554 assert(adr->is_AddP(), "expecting an AddP"); 555 if (adr_type == TypeRawPtr::NOTNULL) { 556 // Verify a raw address for a store captured by Initialize node. 557 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 558 assert(offs != Type::OffsetBot, "offset must be a constant"); 559 } 560 #endif 561 } else { 562 // Ignore copy the displaced header to the BoxNode (OSR compilation). 563 if (adr->is_BoxLock()) 564 break; 565 // Stored value escapes in unsafe access. 566 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 567 // Pointer stores in G1 barriers looks like unsafe access. 568 // Ignore such stores to be able scalar replace non-escaping 569 // allocations. 570 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC 571 if ((UseG1GC || UseShenandoahGC) && adr->is_AddP()) { 572 Node* base = get_addp_base(adr); 573 if (base->Opcode() == Op_LoadP && 574 base->in(MemNode::Address)->is_AddP()) { 575 adr = base->in(MemNode::Address); 576 Node* tls = get_addp_base(adr); 577 if (tls->Opcode() == Op_ThreadLocal) { 578 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 579 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC 580 const int buf_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_buffer_offset() 581 : ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 582 #elif INCLUDE_G1GC 583 const int buf_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()); 584 #else 585 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 586 #endif 587 if (offs == buf_offset) { 588 break; // G1 pre barrier previous oop value store. 589 } 590 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) { 591 break; // G1 post barrier card address store. 592 } 593 } 594 } 595 } 596 #endif 597 delayed_worklist->push(n); // Process unsafe access later. 598 break; 599 } 600 #ifdef ASSERT 601 n->dump(1); 602 assert(false, "not unsafe or G1 barrier raw StoreP"); 603 #endif 604 } 605 break; 606 } 607 case Op_AryEq: 608 case Op_HasNegatives: 609 case Op_StrComp: 610 case Op_StrEquals: 611 case Op_StrIndexOf: 612 case Op_StrIndexOfChar: 613 case Op_StrInflatedCopy: 614 case Op_StrCompressedCopy: 615 case Op_EncodeISOArray: { 616 add_local_var(n, PointsToNode::ArgEscape); 617 delayed_worklist->push(n); // Process it later. 618 break; 619 } 620 case Op_ThreadLocal: { 621 add_java_object(n, PointsToNode::ArgEscape); 622 break; 623 } 624 #if INCLUDE_SHENANDOAHGC 625 case Op_ShenandoahReadBarrier: 626 case Op_ShenandoahWriteBarrier: 627 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 628 // It doesn't escape. 629 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), delayed_worklist); 630 break; 631 case Op_ShenandoahEnqueueBarrier: 632 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist); 633 break; 634 #endif 635 default: 636 ; // Do nothing for nodes not related to EA. 637 } 638 return; 639 } 640 641 #ifdef ASSERT 642 #define ELSE_FAIL(name) \ 643 /* Should not be called for not pointer type. */ \ 644 n->dump(1); \ 645 assert(false, name); \ 646 break; 647 #else 648 #define ELSE_FAIL(name) \ 649 break; 650 #endif 651 652 // Add final simple edges to graph. 653 void ConnectionGraph::add_final_edges(Node *n) { 654 PointsToNode* n_ptn = ptnode_adr(n->_idx); 655 #ifdef ASSERT 656 if (_verify && n_ptn->is_JavaObject()) 657 return; // This method does not change graph for JavaObject. 658 #endif 659 660 if (n->is_Call()) { 661 process_call_arguments(n->as_Call()); 662 return; 663 } 664 assert(n->is_Store() || n->is_LoadStore() || 665 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL), 666 "node should be registered already"); 667 int opcode = n->Opcode(); 668 switch (opcode) { 669 case Op_AddP: { 670 Node* base = get_addp_base(n); 671 PointsToNode* ptn_base = ptnode_adr(base->_idx); 672 assert(ptn_base != NULL, "field's base should be registered"); 673 add_base(n_ptn->as_Field(), ptn_base); 674 break; 675 } 676 case Op_CastPP: 677 case Op_CheckCastPP: 678 case Op_EncodeP: 679 case Op_DecodeN: 680 case Op_EncodePKlass: 681 case Op_DecodeNKlass: { 682 add_local_var_and_edge(n, PointsToNode::NoEscape, 683 n->in(1), NULL); 684 break; 685 } 686 case Op_CMoveP: { 687 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) { 688 Node* in = n->in(i); 689 if (in == NULL) 690 continue; // ignore NULL 691 Node* uncast_in = in->uncast(); 692 if (uncast_in->is_top() || uncast_in == n) 693 continue; // ignore top or inputs which go back this node 694 PointsToNode* ptn = ptnode_adr(in->_idx); 695 assert(ptn != NULL, "node should be registered"); 696 add_edge(n_ptn, ptn); 697 } 698 break; 699 } 700 case Op_LoadP: 701 #if INCLUDE_ZGC 702 case Op_LoadBarrierSlowReg: 703 case Op_LoadBarrierWeakSlowReg: 704 #endif 705 case Op_LoadN: 706 case Op_LoadPLocked: { 707 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 708 // ThreadLocal has RawPtr type. 709 const Type* t = _igvn->type(n); 710 if (t->make_ptr() != NULL) { 711 Node* adr = n->in(MemNode::Address); 712 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 713 break; 714 } 715 ELSE_FAIL("Op_LoadP"); 716 } 717 case Op_Phi: { 718 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 719 // ThreadLocal has RawPtr type. 720 const Type* t = n->as_Phi()->type(); 721 if (t->make_ptr() != NULL) { 722 for (uint i = 1; i < n->req(); i++) { 723 Node* in = n->in(i); 724 if (in == NULL) 725 continue; // ignore NULL 726 Node* uncast_in = in->uncast(); 727 if (uncast_in->is_top() || uncast_in == n) 728 continue; // ignore top or inputs which go back this node 729 PointsToNode* ptn = ptnode_adr(in->_idx); 730 assert(ptn != NULL, "node should be registered"); 731 add_edge(n_ptn, ptn); 732 } 733 break; 734 } 735 ELSE_FAIL("Op_Phi"); 736 } 737 case Op_Proj: { 738 // we are only interested in the oop result projection from a call 739 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && 740 n->in(0)->as_Call()->returns_pointer()) { 741 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL); 742 break; 743 } 744 #if INCLUDE_ZGC 745 else if (UseZGC) { 746 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) { 747 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), NULL); 748 break; 749 } 750 } 751 #endif 752 ELSE_FAIL("Op_Proj"); 753 } 754 case Op_Rethrow: // Exception object escapes 755 case Op_Return: { 756 if (n->req() > TypeFunc::Parms && 757 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) { 758 // Treat Return value as LocalVar with GlobalEscape escape state. 759 add_local_var_and_edge(n, PointsToNode::GlobalEscape, 760 n->in(TypeFunc::Parms), NULL); 761 break; 762 } 763 ELSE_FAIL("Op_Return"); 764 } 765 case Op_StoreP: 766 case Op_StoreN: 767 case Op_StoreNKlass: 768 case Op_StorePConditional: 769 case Op_CompareAndExchangeP: 770 case Op_CompareAndExchangeN: 771 case Op_CompareAndSwapP: 772 case Op_CompareAndSwapN: 773 case Op_WeakCompareAndSwapP: 774 case Op_WeakCompareAndSwapN: 775 #if INCLUDE_SHENANDOAHGC 776 case Op_ShenandoahCompareAndExchangeP: 777 case Op_ShenandoahCompareAndExchangeN: 778 case Op_ShenandoahCompareAndSwapP: 779 case Op_ShenandoahCompareAndSwapN: 780 case Op_ShenandoahWeakCompareAndSwapP: 781 case Op_ShenandoahWeakCompareAndSwapN: 782 #endif 783 case Op_GetAndSetP: 784 case Op_GetAndSetN: { 785 Node* adr = n->in(MemNode::Address); 786 const Type *adr_type = _igvn->type(adr); 787 adr_type = adr_type->make_ptr(); 788 #ifdef ASSERT 789 if (adr_type == NULL) { 790 n->dump(1); 791 assert(adr_type != NULL, "dead node should not be on list"); 792 break; 793 } 794 #endif 795 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN || 796 #if INCLUDE_SHENANDOAHGC 797 opcode == Op_ShenandoahCompareAndExchangeN || opcode == Op_ShenandoahCompareAndExchangeP || 798 #endif 799 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) { 800 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 801 } 802 if ( adr_type->isa_oopptr() 803 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) 804 && adr_type == TypeRawPtr::NOTNULL 805 && adr->in(AddPNode::Address)->is_Proj() 806 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) { 807 // Point Address to Value 808 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 809 assert(adr_ptn != NULL && 810 adr_ptn->as_Field()->is_oop(), "node should be registered"); 811 Node *val = n->in(MemNode::ValueIn); 812 PointsToNode* ptn = ptnode_adr(val->_idx); 813 assert(ptn != NULL, "node should be registered"); 814 add_edge(adr_ptn, ptn); 815 break; 816 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) { 817 // Stored value escapes in unsafe access. 818 Node *val = n->in(MemNode::ValueIn); 819 PointsToNode* ptn = ptnode_adr(val->_idx); 820 assert(ptn != NULL, "node should be registered"); 821 set_escape_state(ptn, PointsToNode::GlobalEscape); 822 // Add edge to object for unsafe access with offset. 823 PointsToNode* adr_ptn = ptnode_adr(adr->_idx); 824 assert(adr_ptn != NULL, "node should be registered"); 825 if (adr_ptn->is_Field()) { 826 assert(adr_ptn->as_Field()->is_oop(), "should be oop field"); 827 add_edge(adr_ptn, ptn); 828 } 829 break; 830 } 831 ELSE_FAIL("Op_StoreP"); 832 } 833 case Op_AryEq: 834 case Op_HasNegatives: 835 case Op_StrComp: 836 case Op_StrEquals: 837 case Op_StrIndexOf: 838 case Op_StrIndexOfChar: 839 case Op_StrInflatedCopy: 840 case Op_StrCompressedCopy: 841 case Op_EncodeISOArray: { 842 // char[]/byte[] arrays passed to string intrinsic do not escape but 843 // they are not scalar replaceable. Adjust escape state for them. 844 // Start from in(2) edge since in(1) is memory edge. 845 for (uint i = 2; i < n->req(); i++) { 846 Node* adr = n->in(i); 847 const Type* at = _igvn->type(adr); 848 if (!adr->is_top() && at->isa_ptr()) { 849 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 850 at->isa_ptr() != NULL, "expecting a pointer"); 851 if (adr->is_AddP()) { 852 adr = get_addp_base(adr); 853 } 854 PointsToNode* ptn = ptnode_adr(adr->_idx); 855 assert(ptn != NULL, "node should be registered"); 856 add_edge(n_ptn, ptn); 857 } 858 } 859 break; 860 } 861 #if INCLUDE_SHENANDOAHGC 862 case Op_ShenandoahReadBarrier: 863 case Op_ShenandoahWriteBarrier: 864 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 865 // It doesn't escape. 866 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), NULL); 867 break; 868 case Op_ShenandoahEnqueueBarrier: 869 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL); 870 break; 871 #endif 872 default: { 873 // This method should be called only for EA specific nodes which may 874 // miss some edges when they were created. 875 #ifdef ASSERT 876 n->dump(1); 877 #endif 878 guarantee(false, "unknown node"); 879 } 880 } 881 return; 882 } 883 884 void ConnectionGraph::add_call_node(CallNode* call) { 885 assert(call->returns_pointer(), "only for call which returns pointer"); 886 uint call_idx = call->_idx; 887 if (call->is_Allocate()) { 888 Node* k = call->in(AllocateNode::KlassNode); 889 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr(); 890 assert(kt != NULL, "TypeKlassPtr required."); 891 ciKlass* cik = kt->klass(); 892 PointsToNode::EscapeState es = PointsToNode::NoEscape; 893 bool scalar_replaceable = true; 894 if (call->is_AllocateArray()) { 895 if (!cik->is_array_klass()) { // StressReflectiveCode 896 es = PointsToNode::GlobalEscape; 897 } else { 898 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 899 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 900 // Not scalar replaceable if the length is not constant or too big. 901 scalar_replaceable = false; 902 } 903 } 904 } else { // Allocate instance 905 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 906 cik->is_subclass_of(_compile->env()->Reference_klass()) || 907 !cik->is_instance_klass() || // StressReflectiveCode 908 !cik->as_instance_klass()->can_be_instantiated() || 909 cik->as_instance_klass()->has_finalizer()) { 910 es = PointsToNode::GlobalEscape; 911 } 912 } 913 add_java_object(call, es); 914 PointsToNode* ptn = ptnode_adr(call_idx); 915 if (!scalar_replaceable && ptn->scalar_replaceable()) { 916 ptn->set_scalar_replaceable(false); 917 } 918 } else if (call->is_CallStaticJava()) { 919 // Call nodes could be different types: 920 // 921 // 1. CallDynamicJavaNode (what happened during call is unknown): 922 // 923 // - mapped to GlobalEscape JavaObject node if oop is returned; 924 // 925 // - all oop arguments are escaping globally; 926 // 927 // 2. CallStaticJavaNode (execute bytecode analysis if possible): 928 // 929 // - the same as CallDynamicJavaNode if can't do bytecode analysis; 930 // 931 // - mapped to GlobalEscape JavaObject node if unknown oop is returned; 932 // - mapped to NoEscape JavaObject node if non-escaping object allocated 933 // during call is returned; 934 // - mapped to ArgEscape LocalVar node pointed to object arguments 935 // which are returned and does not escape during call; 936 // 937 // - oop arguments escaping status is defined by bytecode analysis; 938 // 939 // For a static call, we know exactly what method is being called. 940 // Use bytecode estimator to record whether the call's return value escapes. 941 ciMethod* meth = call->as_CallJava()->method(); 942 if (meth == NULL) { 943 const char* name = call->as_CallStaticJava()->_name; 944 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check"); 945 // Returns a newly allocated unescaped object. 946 add_java_object(call, PointsToNode::NoEscape); 947 ptnode_adr(call_idx)->set_scalar_replaceable(false); 948 } else if (meth->is_boxing_method()) { 949 // Returns boxing object 950 PointsToNode::EscapeState es; 951 vmIntrinsics::ID intr = meth->intrinsic_id(); 952 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) { 953 // It does not escape if object is always allocated. 954 es = PointsToNode::NoEscape; 955 } else { 956 // It escapes globally if object could be loaded from cache. 957 es = PointsToNode::GlobalEscape; 958 } 959 add_java_object(call, es); 960 } else { 961 BCEscapeAnalyzer* call_analyzer = meth->get_bcea(); 962 call_analyzer->copy_dependencies(_compile->dependencies()); 963 if (call_analyzer->is_return_allocated()) { 964 // Returns a newly allocated unescaped object, simply 965 // update dependency information. 966 // Mark it as NoEscape so that objects referenced by 967 // it's fields will be marked as NoEscape at least. 968 add_java_object(call, PointsToNode::NoEscape); 969 ptnode_adr(call_idx)->set_scalar_replaceable(false); 970 } else { 971 // Determine whether any arguments are returned. 972 const TypeTuple* d = call->tf()->domain(); 973 bool ret_arg = false; 974 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 975 if (d->field_at(i)->isa_ptr() != NULL && 976 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 977 ret_arg = true; 978 break; 979 } 980 } 981 if (ret_arg) { 982 add_local_var(call, PointsToNode::ArgEscape); 983 } else { 984 // Returns unknown object. 985 map_ideal_node(call, phantom_obj); 986 } 987 } 988 } 989 } else { 990 // An other type of call, assume the worst case: 991 // returned value is unknown and globally escapes. 992 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check"); 993 map_ideal_node(call, phantom_obj); 994 } 995 } 996 997 void ConnectionGraph::process_call_arguments(CallNode *call) { 998 bool is_arraycopy = false; 999 switch (call->Opcode()) { 1000 #ifdef ASSERT 1001 case Op_Allocate: 1002 case Op_AllocateArray: 1003 case Op_Lock: 1004 case Op_Unlock: 1005 assert(false, "should be done already"); 1006 break; 1007 #endif 1008 case Op_ArrayCopy: 1009 case Op_CallLeafNoFP: 1010 // Most array copies are ArrayCopy nodes at this point but there 1011 // are still a few direct calls to the copy subroutines (See 1012 // PhaseStringOpts::copy_string()) 1013 is_arraycopy = (call->Opcode() == Op_ArrayCopy) || 1014 call->as_CallLeaf()->is_call_to_arraycopystub(); 1015 // fall through 1016 case Op_CallLeaf: { 1017 // Stub calls, objects do not escape but they are not scale replaceable. 1018 // Adjust escape state for outgoing arguments. 1019 const TypeTuple * d = call->tf()->domain(); 1020 bool src_has_oops = false; 1021 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1022 const Type* at = d->field_at(i); 1023 Node *arg = call->in(i); 1024 if (arg == NULL) { 1025 continue; 1026 } 1027 const Type *aat = _igvn->type(arg); 1028 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) 1029 continue; 1030 if (arg->is_AddP()) { 1031 // 1032 // The inline_native_clone() case when the arraycopy stub is called 1033 // after the allocation before Initialize and CheckCastPP nodes. 1034 // Or normal arraycopy for object arrays case. 1035 // 1036 // Set AddP's base (Allocate) as not scalar replaceable since 1037 // pointer to the base (with offset) is passed as argument. 1038 // 1039 arg = get_addp_base(arg); 1040 } 1041 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1042 assert(arg_ptn != NULL, "should be registered"); 1043 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state(); 1044 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) { 1045 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 1046 aat->isa_ptr() != NULL, "expecting an Ptr"); 1047 bool arg_has_oops = aat->isa_oopptr() && 1048 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() || 1049 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass())); 1050 if (i == TypeFunc::Parms) { 1051 src_has_oops = arg_has_oops; 1052 } 1053 // 1054 // src or dst could be j.l.Object when other is basic type array: 1055 // 1056 // arraycopy(char[],0,Object*,0,size); 1057 // arraycopy(Object*,0,char[],0,size); 1058 // 1059 // Don't add edges in such cases. 1060 // 1061 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy && 1062 arg_has_oops && (i > TypeFunc::Parms); 1063 #ifdef ASSERT 1064 if (!(is_arraycopy || 1065 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) || 1066 (call->as_CallLeaf()->_name != NULL && 1067 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 || 1068 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 || 1069 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 || 1070 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 || 1071 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 || 1072 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 || 1073 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 || 1074 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 || 1075 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 || 1076 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 || 1077 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 || 1078 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 || 1079 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 || 1080 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 || 1081 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 || 1082 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 || 1083 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 || 1084 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 || 1085 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 || 1086 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 || 1087 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 || 1088 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0) 1089 ))) { 1090 call->dump(); 1091 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name); 1092 } 1093 #endif 1094 // Always process arraycopy's destination object since 1095 // we need to add all possible edges to references in 1096 // source object. 1097 if (arg_esc >= PointsToNode::ArgEscape && 1098 !arg_is_arraycopy_dest) { 1099 continue; 1100 } 1101 PointsToNode::EscapeState es = PointsToNode::ArgEscape; 1102 if (call->is_ArrayCopy()) { 1103 ArrayCopyNode* ac = call->as_ArrayCopy(); 1104 if (ac->is_clonebasic() || 1105 ac->is_arraycopy_validated() || 1106 ac->is_copyof_validated() || 1107 ac->is_copyofrange_validated()) { 1108 es = PointsToNode::NoEscape; 1109 } 1110 } 1111 set_escape_state(arg_ptn, es); 1112 if (arg_is_arraycopy_dest) { 1113 Node* src = call->in(TypeFunc::Parms); 1114 if (src->is_AddP()) { 1115 src = get_addp_base(src); 1116 } 1117 PointsToNode* src_ptn = ptnode_adr(src->_idx); 1118 assert(src_ptn != NULL, "should be registered"); 1119 if (arg_ptn != src_ptn) { 1120 // Special arraycopy edge: 1121 // A destination object's field can't have the source object 1122 // as base since objects escape states are not related. 1123 // Only escape state of destination object's fields affects 1124 // escape state of fields in source object. 1125 add_arraycopy(call, es, src_ptn, arg_ptn); 1126 } 1127 } 1128 } 1129 } 1130 break; 1131 } 1132 case Op_CallStaticJava: { 1133 // For a static call, we know exactly what method is being called. 1134 // Use bytecode estimator to record the call's escape affects 1135 #ifdef ASSERT 1136 const char* name = call->as_CallStaticJava()->_name; 1137 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only"); 1138 #endif 1139 ciMethod* meth = call->as_CallJava()->method(); 1140 if ((meth != NULL) && meth->is_boxing_method()) { 1141 break; // Boxing methods do not modify any oops. 1142 } 1143 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1144 // fall-through if not a Java method or no analyzer information 1145 if (call_analyzer != NULL) { 1146 PointsToNode* call_ptn = ptnode_adr(call->_idx); 1147 const TypeTuple* d = call->tf()->domain(); 1148 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1149 const Type* at = d->field_at(i); 1150 int k = i - TypeFunc::Parms; 1151 Node* arg = call->in(i); 1152 PointsToNode* arg_ptn = ptnode_adr(arg->_idx); 1153 if (at->isa_ptr() != NULL && 1154 call_analyzer->is_arg_returned(k)) { 1155 // The call returns arguments. 1156 if (call_ptn != NULL) { // Is call's result used? 1157 assert(call_ptn->is_LocalVar(), "node should be registered"); 1158 assert(arg_ptn != NULL, "node should be registered"); 1159 add_edge(call_ptn, arg_ptn); 1160 } 1161 } 1162 if (at->isa_oopptr() != NULL && 1163 arg_ptn->escape_state() < PointsToNode::GlobalEscape) { 1164 if (!call_analyzer->is_arg_stack(k)) { 1165 // The argument global escapes 1166 set_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1167 } else { 1168 set_escape_state(arg_ptn, PointsToNode::ArgEscape); 1169 if (!call_analyzer->is_arg_local(k)) { 1170 // The argument itself doesn't escape, but any fields might 1171 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape); 1172 } 1173 } 1174 } 1175 } 1176 if (call_ptn != NULL && call_ptn->is_LocalVar()) { 1177 // The call returns arguments. 1178 assert(call_ptn->edge_count() > 0, "sanity"); 1179 if (!call_analyzer->is_return_local()) { 1180 // Returns also unknown object. 1181 add_edge(call_ptn, phantom_obj); 1182 } 1183 } 1184 break; 1185 } 1186 } 1187 default: { 1188 // Fall-through here if not a Java method or no analyzer information 1189 // or some other type of call, assume the worst case: all arguments 1190 // globally escape. 1191 const TypeTuple* d = call->tf()->domain(); 1192 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1193 const Type* at = d->field_at(i); 1194 if (at->isa_oopptr() != NULL) { 1195 Node* arg = call->in(i); 1196 if (arg->is_AddP()) { 1197 arg = get_addp_base(arg); 1198 } 1199 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already"); 1200 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape); 1201 } 1202 } 1203 } 1204 } 1205 } 1206 1207 1208 // Finish Graph construction. 1209 bool ConnectionGraph::complete_connection_graph( 1210 GrowableArray<PointsToNode*>& ptnodes_worklist, 1211 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1212 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1213 GrowableArray<FieldNode*>& oop_fields_worklist) { 1214 // Normally only 1-3 passes needed to build Connection Graph depending 1215 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler. 1216 // Set limit to 20 to catch situation when something did go wrong and 1217 // bailout Escape Analysis. 1218 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag. 1219 #define CG_BUILD_ITER_LIMIT 20 1220 1221 // Propagate GlobalEscape and ArgEscape escape states and check that 1222 // we still have non-escaping objects. The method pushs on _worklist 1223 // Field nodes which reference phantom_object. 1224 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1225 return false; // Nothing to do. 1226 } 1227 // Now propagate references to all JavaObject nodes. 1228 int java_objects_length = java_objects_worklist.length(); 1229 elapsedTimer time; 1230 bool timeout = false; 1231 int new_edges = 1; 1232 int iterations = 0; 1233 do { 1234 while ((new_edges > 0) && 1235 (iterations++ < CG_BUILD_ITER_LIMIT)) { 1236 double start_time = time.seconds(); 1237 time.start(); 1238 new_edges = 0; 1239 // Propagate references to phantom_object for nodes pushed on _worklist 1240 // by find_non_escaped_objects() and find_field_value(). 1241 new_edges += add_java_object_edges(phantom_obj, false); 1242 for (int next = 0; next < java_objects_length; ++next) { 1243 JavaObjectNode* ptn = java_objects_worklist.at(next); 1244 new_edges += add_java_object_edges(ptn, true); 1245 1246 #define SAMPLE_SIZE 4 1247 if ((next % SAMPLE_SIZE) == 0) { 1248 // Each 4 iterations calculate how much time it will take 1249 // to complete graph construction. 1250 time.stop(); 1251 // Poll for requests from shutdown mechanism to quiesce compiler 1252 // because Connection graph construction may take long time. 1253 CompileBroker::maybe_block(); 1254 double stop_time = time.seconds(); 1255 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE; 1256 double time_until_end = time_per_iter * (double)(java_objects_length - next); 1257 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) { 1258 timeout = true; 1259 break; // Timeout 1260 } 1261 start_time = stop_time; 1262 time.start(); 1263 } 1264 #undef SAMPLE_SIZE 1265 1266 } 1267 if (timeout) break; 1268 if (new_edges > 0) { 1269 // Update escape states on each iteration if graph was updated. 1270 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) { 1271 return false; // Nothing to do. 1272 } 1273 } 1274 time.stop(); 1275 if (time.seconds() >= EscapeAnalysisTimeout) { 1276 timeout = true; 1277 break; 1278 } 1279 } 1280 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) { 1281 time.start(); 1282 // Find fields which have unknown value. 1283 int fields_length = oop_fields_worklist.length(); 1284 for (int next = 0; next < fields_length; next++) { 1285 FieldNode* field = oop_fields_worklist.at(next); 1286 if (field->edge_count() == 0) { 1287 new_edges += find_field_value(field); 1288 // This code may added new edges to phantom_object. 1289 // Need an other cycle to propagate references to phantom_object. 1290 } 1291 } 1292 time.stop(); 1293 if (time.seconds() >= EscapeAnalysisTimeout) { 1294 timeout = true; 1295 break; 1296 } 1297 } else { 1298 new_edges = 0; // Bailout 1299 } 1300 } while (new_edges > 0); 1301 1302 // Bailout if passed limits. 1303 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) { 1304 Compile* C = _compile; 1305 if (C->log() != NULL) { 1306 C->log()->begin_elem("connectionGraph_bailout reason='reached "); 1307 C->log()->text("%s", timeout ? "time" : "iterations"); 1308 C->log()->end_elem(" limit'"); 1309 } 1310 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d", 1311 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()); 1312 // Possible infinite build_connection_graph loop, 1313 // bailout (no changes to ideal graph were made). 1314 return false; 1315 } 1316 #ifdef ASSERT 1317 if (Verbose && PrintEscapeAnalysis) { 1318 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d", 1319 iterations, nodes_size(), ptnodes_worklist.length()); 1320 } 1321 #endif 1322 1323 #undef CG_BUILD_ITER_LIMIT 1324 1325 // Find fields initialized by NULL for non-escaping Allocations. 1326 int non_escaped_length = non_escaped_worklist.length(); 1327 for (int next = 0; next < non_escaped_length; next++) { 1328 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1329 PointsToNode::EscapeState es = ptn->escape_state(); 1330 assert(es <= PointsToNode::ArgEscape, "sanity"); 1331 if (es == PointsToNode::NoEscape) { 1332 if (find_init_values(ptn, null_obj, _igvn) > 0) { 1333 // Adding references to NULL object does not change escape states 1334 // since it does not escape. Also no fields are added to NULL object. 1335 add_java_object_edges(null_obj, false); 1336 } 1337 } 1338 Node* n = ptn->ideal_node(); 1339 if (n->is_Allocate()) { 1340 // The object allocated by this Allocate node will never be 1341 // seen by an other thread. Mark it so that when it is 1342 // expanded no MemBarStoreStore is added. 1343 InitializeNode* ini = n->as_Allocate()->initialization(); 1344 if (ini != NULL) 1345 ini->set_does_not_escape(); 1346 } 1347 } 1348 return true; // Finished graph construction. 1349 } 1350 1351 // Propagate GlobalEscape and ArgEscape escape states to all nodes 1352 // and check that we still have non-escaping java objects. 1353 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist, 1354 GrowableArray<JavaObjectNode*>& non_escaped_worklist) { 1355 GrowableArray<PointsToNode*> escape_worklist; 1356 // First, put all nodes with GlobalEscape and ArgEscape states on worklist. 1357 int ptnodes_length = ptnodes_worklist.length(); 1358 for (int next = 0; next < ptnodes_length; ++next) { 1359 PointsToNode* ptn = ptnodes_worklist.at(next); 1360 if (ptn->escape_state() >= PointsToNode::ArgEscape || 1361 ptn->fields_escape_state() >= PointsToNode::ArgEscape) { 1362 escape_worklist.push(ptn); 1363 } 1364 } 1365 // Set escape states to referenced nodes (edges list). 1366 while (escape_worklist.length() > 0) { 1367 PointsToNode* ptn = escape_worklist.pop(); 1368 PointsToNode::EscapeState es = ptn->escape_state(); 1369 PointsToNode::EscapeState field_es = ptn->fields_escape_state(); 1370 if (ptn->is_Field() && ptn->as_Field()->is_oop() && 1371 es >= PointsToNode::ArgEscape) { 1372 // GlobalEscape or ArgEscape state of field means it has unknown value. 1373 if (add_edge(ptn, phantom_obj)) { 1374 // New edge was added 1375 add_field_uses_to_worklist(ptn->as_Field()); 1376 } 1377 } 1378 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1379 PointsToNode* e = i.get(); 1380 if (e->is_Arraycopy()) { 1381 assert(ptn->arraycopy_dst(), "sanity"); 1382 // Propagate only fields escape state through arraycopy edge. 1383 if (e->fields_escape_state() < field_es) { 1384 set_fields_escape_state(e, field_es); 1385 escape_worklist.push(e); 1386 } 1387 } else if (es >= field_es) { 1388 // fields_escape_state is also set to 'es' if it is less than 'es'. 1389 if (e->escape_state() < es) { 1390 set_escape_state(e, es); 1391 escape_worklist.push(e); 1392 } 1393 } else { 1394 // Propagate field escape state. 1395 bool es_changed = false; 1396 if (e->fields_escape_state() < field_es) { 1397 set_fields_escape_state(e, field_es); 1398 es_changed = true; 1399 } 1400 if ((e->escape_state() < field_es) && 1401 e->is_Field() && ptn->is_JavaObject() && 1402 e->as_Field()->is_oop()) { 1403 // Change escape state of referenced fields. 1404 set_escape_state(e, field_es); 1405 es_changed = true; 1406 } else if (e->escape_state() < es) { 1407 set_escape_state(e, es); 1408 es_changed = true; 1409 } 1410 if (es_changed) { 1411 escape_worklist.push(e); 1412 } 1413 } 1414 } 1415 } 1416 // Remove escaped objects from non_escaped list. 1417 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) { 1418 JavaObjectNode* ptn = non_escaped_worklist.at(next); 1419 if (ptn->escape_state() >= PointsToNode::GlobalEscape) { 1420 non_escaped_worklist.delete_at(next); 1421 } 1422 if (ptn->escape_state() == PointsToNode::NoEscape) { 1423 // Find fields in non-escaped allocations which have unknown value. 1424 find_init_values(ptn, phantom_obj, NULL); 1425 } 1426 } 1427 return (non_escaped_worklist.length() > 0); 1428 } 1429 1430 // Add all references to JavaObject node by walking over all uses. 1431 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) { 1432 int new_edges = 0; 1433 if (populate_worklist) { 1434 // Populate _worklist by uses of jobj's uses. 1435 for (UseIterator i(jobj); i.has_next(); i.next()) { 1436 PointsToNode* use = i.get(); 1437 if (use->is_Arraycopy()) 1438 continue; 1439 add_uses_to_worklist(use); 1440 if (use->is_Field() && use->as_Field()->is_oop()) { 1441 // Put on worklist all field's uses (loads) and 1442 // related field nodes (same base and offset). 1443 add_field_uses_to_worklist(use->as_Field()); 1444 } 1445 } 1446 } 1447 for (int l = 0; l < _worklist.length(); l++) { 1448 PointsToNode* use = _worklist.at(l); 1449 if (PointsToNode::is_base_use(use)) { 1450 // Add reference from jobj to field and from field to jobj (field's base). 1451 use = PointsToNode::get_use_node(use)->as_Field(); 1452 if (add_base(use->as_Field(), jobj)) { 1453 new_edges++; 1454 } 1455 continue; 1456 } 1457 assert(!use->is_JavaObject(), "sanity"); 1458 if (use->is_Arraycopy()) { 1459 if (jobj == null_obj) // NULL object does not have field edges 1460 continue; 1461 // Added edge from Arraycopy node to arraycopy's source java object 1462 if (add_edge(use, jobj)) { 1463 jobj->set_arraycopy_src(); 1464 new_edges++; 1465 } 1466 // and stop here. 1467 continue; 1468 } 1469 if (!add_edge(use, jobj)) 1470 continue; // No new edge added, there was such edge already. 1471 new_edges++; 1472 if (use->is_LocalVar()) { 1473 add_uses_to_worklist(use); 1474 if (use->arraycopy_dst()) { 1475 for (EdgeIterator i(use); i.has_next(); i.next()) { 1476 PointsToNode* e = i.get(); 1477 if (e->is_Arraycopy()) { 1478 if (jobj == null_obj) // NULL object does not have field edges 1479 continue; 1480 // Add edge from arraycopy's destination java object to Arraycopy node. 1481 if (add_edge(jobj, e)) { 1482 new_edges++; 1483 jobj->set_arraycopy_dst(); 1484 } 1485 } 1486 } 1487 } 1488 } else { 1489 // Added new edge to stored in field values. 1490 // Put on worklist all field's uses (loads) and 1491 // related field nodes (same base and offset). 1492 add_field_uses_to_worklist(use->as_Field()); 1493 } 1494 } 1495 _worklist.clear(); 1496 _in_worklist.Reset(); 1497 return new_edges; 1498 } 1499 1500 // Put on worklist all related field nodes. 1501 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) { 1502 assert(field->is_oop(), "sanity"); 1503 int offset = field->offset(); 1504 add_uses_to_worklist(field); 1505 // Loop over all bases of this field and push on worklist Field nodes 1506 // with the same offset and base (since they may reference the same field). 1507 for (BaseIterator i(field); i.has_next(); i.next()) { 1508 PointsToNode* base = i.get(); 1509 add_fields_to_worklist(field, base); 1510 // Check if the base was source object of arraycopy and go over arraycopy's 1511 // destination objects since values stored to a field of source object are 1512 // accessable by uses (loads) of fields of destination objects. 1513 if (base->arraycopy_src()) { 1514 for (UseIterator j(base); j.has_next(); j.next()) { 1515 PointsToNode* arycp = j.get(); 1516 if (arycp->is_Arraycopy()) { 1517 for (UseIterator k(arycp); k.has_next(); k.next()) { 1518 PointsToNode* abase = k.get(); 1519 if (abase->arraycopy_dst() && abase != base) { 1520 // Look for the same arraycopy reference. 1521 add_fields_to_worklist(field, abase); 1522 } 1523 } 1524 } 1525 } 1526 } 1527 } 1528 } 1529 1530 // Put on worklist all related field nodes. 1531 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) { 1532 int offset = field->offset(); 1533 if (base->is_LocalVar()) { 1534 for (UseIterator j(base); j.has_next(); j.next()) { 1535 PointsToNode* f = j.get(); 1536 if (PointsToNode::is_base_use(f)) { // Field 1537 f = PointsToNode::get_use_node(f); 1538 if (f == field || !f->as_Field()->is_oop()) 1539 continue; 1540 int offs = f->as_Field()->offset(); 1541 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1542 add_to_worklist(f); 1543 } 1544 } 1545 } 1546 } else { 1547 assert(base->is_JavaObject(), "sanity"); 1548 if (// Skip phantom_object since it is only used to indicate that 1549 // this field's content globally escapes. 1550 (base != phantom_obj) && 1551 // NULL object node does not have fields. 1552 (base != null_obj)) { 1553 for (EdgeIterator i(base); i.has_next(); i.next()) { 1554 PointsToNode* f = i.get(); 1555 // Skip arraycopy edge since store to destination object field 1556 // does not update value in source object field. 1557 if (f->is_Arraycopy()) { 1558 assert(base->arraycopy_dst(), "sanity"); 1559 continue; 1560 } 1561 if (f == field || !f->as_Field()->is_oop()) 1562 continue; 1563 int offs = f->as_Field()->offset(); 1564 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) { 1565 add_to_worklist(f); 1566 } 1567 } 1568 } 1569 } 1570 } 1571 1572 // Find fields which have unknown value. 1573 int ConnectionGraph::find_field_value(FieldNode* field) { 1574 // Escaped fields should have init value already. 1575 assert(field->escape_state() == PointsToNode::NoEscape, "sanity"); 1576 int new_edges = 0; 1577 for (BaseIterator i(field); i.has_next(); i.next()) { 1578 PointsToNode* base = i.get(); 1579 if (base->is_JavaObject()) { 1580 // Skip Allocate's fields which will be processed later. 1581 if (base->ideal_node()->is_Allocate()) 1582 return 0; 1583 assert(base == null_obj, "only NULL ptr base expected here"); 1584 } 1585 } 1586 if (add_edge(field, phantom_obj)) { 1587 // New edge was added 1588 new_edges++; 1589 add_field_uses_to_worklist(field); 1590 } 1591 return new_edges; 1592 } 1593 1594 // Find fields initializing values for allocations. 1595 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) { 1596 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1597 int new_edges = 0; 1598 Node* alloc = pta->ideal_node(); 1599 if (init_val == phantom_obj) { 1600 // Do nothing for Allocate nodes since its fields values are 1601 // "known" unless they are initialized by arraycopy/clone. 1602 if (alloc->is_Allocate() && !pta->arraycopy_dst()) 1603 return 0; 1604 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity"); 1605 #ifdef ASSERT 1606 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) { 1607 const char* name = alloc->as_CallStaticJava()->_name; 1608 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity"); 1609 } 1610 #endif 1611 // Non-escaped allocation returned from Java or runtime call have 1612 // unknown values in fields. 1613 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1614 PointsToNode* field = i.get(); 1615 if (field->is_Field() && field->as_Field()->is_oop()) { 1616 if (add_edge(field, phantom_obj)) { 1617 // New edge was added 1618 new_edges++; 1619 add_field_uses_to_worklist(field->as_Field()); 1620 } 1621 } 1622 } 1623 return new_edges; 1624 } 1625 assert(init_val == null_obj, "sanity"); 1626 // Do nothing for Call nodes since its fields values are unknown. 1627 if (!alloc->is_Allocate()) 1628 return 0; 1629 1630 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1631 bool visited_bottom_offset = false; 1632 GrowableArray<int> offsets_worklist; 1633 1634 // Check if an oop field's initializing value is recorded and add 1635 // a corresponding NULL if field's value if it is not recorded. 1636 // Connection Graph does not record a default initialization by NULL 1637 // captured by Initialize node. 1638 // 1639 for (EdgeIterator i(pta); i.has_next(); i.next()) { 1640 PointsToNode* field = i.get(); // Field (AddP) 1641 if (!field->is_Field() || !field->as_Field()->is_oop()) 1642 continue; // Not oop field 1643 int offset = field->as_Field()->offset(); 1644 if (offset == Type::OffsetBot) { 1645 if (!visited_bottom_offset) { 1646 // OffsetBot is used to reference array's element, 1647 // always add reference to NULL to all Field nodes since we don't 1648 // known which element is referenced. 1649 if (add_edge(field, null_obj)) { 1650 // New edge was added 1651 new_edges++; 1652 add_field_uses_to_worklist(field->as_Field()); 1653 visited_bottom_offset = true; 1654 } 1655 } 1656 } else { 1657 // Check only oop fields. 1658 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type(); 1659 if (adr_type->isa_rawptr()) { 1660 #ifdef ASSERT 1661 // Raw pointers are used for initializing stores so skip it 1662 // since it should be recorded already 1663 Node* base = get_addp_base(field->ideal_node()); 1664 assert(adr_type->isa_rawptr() && base->is_Proj() && 1665 (base->in(0) == alloc),"unexpected pointer type"); 1666 #endif 1667 continue; 1668 } 1669 if (!offsets_worklist.contains(offset)) { 1670 offsets_worklist.append(offset); 1671 Node* value = NULL; 1672 if (ini != NULL) { 1673 // StoreP::memory_type() == T_ADDRESS 1674 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS; 1675 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase); 1676 // Make sure initializing store has the same type as this AddP. 1677 // This AddP may reference non existing field because it is on a 1678 // dead branch of bimorphic call which is not eliminated yet. 1679 if (store != NULL && store->is_Store() && 1680 store->as_Store()->memory_type() == ft) { 1681 value = store->in(MemNode::ValueIn); 1682 #ifdef ASSERT 1683 if (VerifyConnectionGraph) { 1684 // Verify that AddP already points to all objects the value points to. 1685 PointsToNode* val = ptnode_adr(value->_idx); 1686 assert((val != NULL), "should be processed already"); 1687 PointsToNode* missed_obj = NULL; 1688 if (val->is_JavaObject()) { 1689 if (!field->points_to(val->as_JavaObject())) { 1690 missed_obj = val; 1691 } 1692 } else { 1693 if (!val->is_LocalVar() || (val->edge_count() == 0)) { 1694 tty->print_cr("----------init store has invalid value -----"); 1695 store->dump(); 1696 val->dump(); 1697 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already"); 1698 } 1699 for (EdgeIterator j(val); j.has_next(); j.next()) { 1700 PointsToNode* obj = j.get(); 1701 if (obj->is_JavaObject()) { 1702 if (!field->points_to(obj->as_JavaObject())) { 1703 missed_obj = obj; 1704 break; 1705 } 1706 } 1707 } 1708 } 1709 if (missed_obj != NULL) { 1710 tty->print_cr("----------field---------------------------------"); 1711 field->dump(); 1712 tty->print_cr("----------missed referernce to object-----------"); 1713 missed_obj->dump(); 1714 tty->print_cr("----------object referernced by init store -----"); 1715 store->dump(); 1716 val->dump(); 1717 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference"); 1718 } 1719 } 1720 #endif 1721 } else { 1722 // There could be initializing stores which follow allocation. 1723 // For example, a volatile field store is not collected 1724 // by Initialize node. 1725 // 1726 // Need to check for dependent loads to separate such stores from 1727 // stores which follow loads. For now, add initial value NULL so 1728 // that compare pointers optimization works correctly. 1729 } 1730 } 1731 if (value == NULL) { 1732 // A field's initializing value was not recorded. Add NULL. 1733 if (add_edge(field, null_obj)) { 1734 // New edge was added 1735 new_edges++; 1736 add_field_uses_to_worklist(field->as_Field()); 1737 } 1738 } 1739 } 1740 } 1741 } 1742 return new_edges; 1743 } 1744 1745 // Adjust scalar_replaceable state after Connection Graph is built. 1746 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) { 1747 // Search for non-escaping objects which are not scalar replaceable 1748 // and mark them to propagate the state to referenced objects. 1749 1750 // 1. An object is not scalar replaceable if the field into which it is 1751 // stored has unknown offset (stored into unknown element of an array). 1752 // 1753 for (UseIterator i(jobj); i.has_next(); i.next()) { 1754 PointsToNode* use = i.get(); 1755 if (use->is_Arraycopy()) { 1756 continue; 1757 } 1758 if (use->is_Field()) { 1759 FieldNode* field = use->as_Field(); 1760 assert(field->is_oop() && field->scalar_replaceable(), "sanity"); 1761 if (field->offset() == Type::OffsetBot) { 1762 jobj->set_scalar_replaceable(false); 1763 return; 1764 } 1765 // 2. An object is not scalar replaceable if the field into which it is 1766 // stored has multiple bases one of which is null. 1767 if (field->base_count() > 1) { 1768 for (BaseIterator i(field); i.has_next(); i.next()) { 1769 PointsToNode* base = i.get(); 1770 if (base == null_obj) { 1771 jobj->set_scalar_replaceable(false); 1772 return; 1773 } 1774 } 1775 } 1776 } 1777 assert(use->is_Field() || use->is_LocalVar(), "sanity"); 1778 // 3. An object is not scalar replaceable if it is merged with other objects. 1779 for (EdgeIterator j(use); j.has_next(); j.next()) { 1780 PointsToNode* ptn = j.get(); 1781 if (ptn->is_JavaObject() && ptn != jobj) { 1782 // Mark all objects. 1783 jobj->set_scalar_replaceable(false); 1784 ptn->set_scalar_replaceable(false); 1785 } 1786 } 1787 if (!jobj->scalar_replaceable()) { 1788 return; 1789 } 1790 } 1791 1792 for (EdgeIterator j(jobj); j.has_next(); j.next()) { 1793 if (j.get()->is_Arraycopy()) { 1794 continue; 1795 } 1796 1797 // Non-escaping object node should point only to field nodes. 1798 FieldNode* field = j.get()->as_Field(); 1799 int offset = field->as_Field()->offset(); 1800 1801 // 4. An object is not scalar replaceable if it has a field with unknown 1802 // offset (array's element is accessed in loop). 1803 if (offset == Type::OffsetBot) { 1804 jobj->set_scalar_replaceable(false); 1805 return; 1806 } 1807 // 5. Currently an object is not scalar replaceable if a LoadStore node 1808 // access its field since the field value is unknown after it. 1809 // 1810 Node* n = field->ideal_node(); 1811 1812 // Test for an unsafe access that was parsed as maybe off heap 1813 // (with a CheckCastPP to raw memory). 1814 assert(n->is_AddP(), "expect an address computation"); 1815 if (n->in(AddPNode::Base)->is_top() && 1816 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) { 1817 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected"); 1818 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected"); 1819 jobj->set_scalar_replaceable(false); 1820 return; 1821 } 1822 1823 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1824 if (n->fast_out(i)->is_LoadStore()) { 1825 jobj->set_scalar_replaceable(false); 1826 return; 1827 } 1828 } 1829 1830 // 6. Or the address may point to more then one object. This may produce 1831 // the false positive result (set not scalar replaceable) 1832 // since the flow-insensitive escape analysis can't separate 1833 // the case when stores overwrite the field's value from the case 1834 // when stores happened on different control branches. 1835 // 1836 // Note: it will disable scalar replacement in some cases: 1837 // 1838 // Point p[] = new Point[1]; 1839 // p[0] = new Point(); // Will be not scalar replaced 1840 // 1841 // but it will save us from incorrect optimizations in next cases: 1842 // 1843 // Point p[] = new Point[1]; 1844 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1845 // 1846 if (field->base_count() > 1) { 1847 for (BaseIterator i(field); i.has_next(); i.next()) { 1848 PointsToNode* base = i.get(); 1849 // Don't take into account LocalVar nodes which 1850 // may point to only one object which should be also 1851 // this field's base by now. 1852 if (base->is_JavaObject() && base != jobj) { 1853 // Mark all bases. 1854 jobj->set_scalar_replaceable(false); 1855 base->set_scalar_replaceable(false); 1856 } 1857 } 1858 } 1859 } 1860 } 1861 1862 #ifdef ASSERT 1863 void ConnectionGraph::verify_connection_graph( 1864 GrowableArray<PointsToNode*>& ptnodes_worklist, 1865 GrowableArray<JavaObjectNode*>& non_escaped_worklist, 1866 GrowableArray<JavaObjectNode*>& java_objects_worklist, 1867 GrowableArray<Node*>& addp_worklist) { 1868 // Verify that graph is complete - no new edges could be added. 1869 int java_objects_length = java_objects_worklist.length(); 1870 int non_escaped_length = non_escaped_worklist.length(); 1871 int new_edges = 0; 1872 for (int next = 0; next < java_objects_length; ++next) { 1873 JavaObjectNode* ptn = java_objects_worklist.at(next); 1874 new_edges += add_java_object_edges(ptn, true); 1875 } 1876 assert(new_edges == 0, "graph was not complete"); 1877 // Verify that escape state is final. 1878 int length = non_escaped_worklist.length(); 1879 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist); 1880 assert((non_escaped_length == non_escaped_worklist.length()) && 1881 (non_escaped_length == length) && 1882 (_worklist.length() == 0), "escape state was not final"); 1883 1884 // Verify fields information. 1885 int addp_length = addp_worklist.length(); 1886 for (int next = 0; next < addp_length; ++next ) { 1887 Node* n = addp_worklist.at(next); 1888 FieldNode* field = ptnode_adr(n->_idx)->as_Field(); 1889 if (field->is_oop()) { 1890 // Verify that field has all bases 1891 Node* base = get_addp_base(n); 1892 PointsToNode* ptn = ptnode_adr(base->_idx); 1893 if (ptn->is_JavaObject()) { 1894 assert(field->has_base(ptn->as_JavaObject()), "sanity"); 1895 } else { 1896 assert(ptn->is_LocalVar(), "sanity"); 1897 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 1898 PointsToNode* e = i.get(); 1899 if (e->is_JavaObject()) { 1900 assert(field->has_base(e->as_JavaObject()), "sanity"); 1901 } 1902 } 1903 } 1904 // Verify that all fields have initializing values. 1905 if (field->edge_count() == 0) { 1906 tty->print_cr("----------field does not have references----------"); 1907 field->dump(); 1908 for (BaseIterator i(field); i.has_next(); i.next()) { 1909 PointsToNode* base = i.get(); 1910 tty->print_cr("----------field has next base---------------------"); 1911 base->dump(); 1912 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) { 1913 tty->print_cr("----------base has fields-------------------------"); 1914 for (EdgeIterator j(base); j.has_next(); j.next()) { 1915 j.get()->dump(); 1916 } 1917 tty->print_cr("----------base has references---------------------"); 1918 for (UseIterator j(base); j.has_next(); j.next()) { 1919 j.get()->dump(); 1920 } 1921 } 1922 } 1923 for (UseIterator i(field); i.has_next(); i.next()) { 1924 i.get()->dump(); 1925 } 1926 assert(field->edge_count() > 0, "sanity"); 1927 } 1928 } 1929 } 1930 } 1931 #endif 1932 1933 // Optimize ideal graph. 1934 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist, 1935 GrowableArray<Node*>& storestore_worklist) { 1936 Compile* C = _compile; 1937 PhaseIterGVN* igvn = _igvn; 1938 if (EliminateLocks) { 1939 // Mark locks before changing ideal graph. 1940 int cnt = C->macro_count(); 1941 for( int i=0; i < cnt; i++ ) { 1942 Node *n = C->macro_node(i); 1943 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1944 AbstractLockNode* alock = n->as_AbstractLock(); 1945 if (!alock->is_non_esc_obj()) { 1946 if (not_global_escape(alock->obj_node())) { 1947 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity"); 1948 // The lock could be marked eliminated by lock coarsening 1949 // code during first IGVN before EA. Replace coarsened flag 1950 // to eliminate all associated locks/unlocks. 1951 #ifdef ASSERT 1952 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3"); 1953 #endif 1954 alock->set_non_esc_obj(); 1955 } 1956 } 1957 } 1958 } 1959 } 1960 1961 if (OptimizePtrCompare) { 1962 // Add ConI(#CC_GT) and ConI(#CC_EQ). 1963 _pcmp_neq = igvn->makecon(TypeInt::CC_GT); 1964 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ); 1965 // Optimize objects compare. 1966 while (ptr_cmp_worklist.length() != 0) { 1967 Node *n = ptr_cmp_worklist.pop(); 1968 Node *res = optimize_ptr_compare(n); 1969 if (res != NULL) { 1970 #ifndef PRODUCT 1971 if (PrintOptimizePtrCompare) { 1972 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")); 1973 if (Verbose) { 1974 n->dump(1); 1975 } 1976 } 1977 #endif 1978 igvn->replace_node(n, res); 1979 } 1980 } 1981 // cleanup 1982 if (_pcmp_neq->outcnt() == 0) 1983 igvn->hash_delete(_pcmp_neq); 1984 if (_pcmp_eq->outcnt() == 0) 1985 igvn->hash_delete(_pcmp_eq); 1986 } 1987 1988 // For MemBarStoreStore nodes added in library_call.cpp, check 1989 // escape status of associated AllocateNode and optimize out 1990 // MemBarStoreStore node if the allocated object never escapes. 1991 while (storestore_worklist.length() != 0) { 1992 Node *n = storestore_worklist.pop(); 1993 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore(); 1994 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0); 1995 assert (alloc->is_Allocate(), "storestore should point to AllocateNode"); 1996 if (not_global_escape(alloc)) { 1997 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot); 1998 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory)); 1999 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control)); 2000 igvn->register_new_node_with_optimizer(mb); 2001 igvn->replace_node(storestore, mb); 2002 } 2003 } 2004 } 2005 2006 // Optimize objects compare. 2007 Node* ConnectionGraph::optimize_ptr_compare(Node* n) { 2008 assert(OptimizePtrCompare, "sanity"); 2009 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx); 2010 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx); 2011 JavaObjectNode* jobj1 = unique_java_object(n->in(1)); 2012 JavaObjectNode* jobj2 = unique_java_object(n->in(2)); 2013 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity"); 2014 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity"); 2015 2016 // Check simple cases first. 2017 if (jobj1 != NULL) { 2018 if (jobj1->escape_state() == PointsToNode::NoEscape) { 2019 if (jobj1 == jobj2) { 2020 // Comparing the same not escaping object. 2021 return _pcmp_eq; 2022 } 2023 Node* obj = jobj1->ideal_node(); 2024 // Comparing not escaping allocation. 2025 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 2026 !ptn2->points_to(jobj1)) { 2027 return _pcmp_neq; // This includes nullness check. 2028 } 2029 } 2030 } 2031 if (jobj2 != NULL) { 2032 if (jobj2->escape_state() == PointsToNode::NoEscape) { 2033 Node* obj = jobj2->ideal_node(); 2034 // Comparing not escaping allocation. 2035 if ((obj->is_Allocate() || obj->is_CallStaticJava()) && 2036 !ptn1->points_to(jobj2)) { 2037 return _pcmp_neq; // This includes nullness check. 2038 } 2039 } 2040 } 2041 if (jobj1 != NULL && jobj1 != phantom_obj && 2042 jobj2 != NULL && jobj2 != phantom_obj && 2043 jobj1->ideal_node()->is_Con() && 2044 jobj2->ideal_node()->is_Con()) { 2045 // Klass or String constants compare. Need to be careful with 2046 // compressed pointers - compare types of ConN and ConP instead of nodes. 2047 const Type* t1 = jobj1->ideal_node()->get_ptr_type(); 2048 const Type* t2 = jobj2->ideal_node()->get_ptr_type(); 2049 if (t1->make_ptr() == t2->make_ptr()) { 2050 return _pcmp_eq; 2051 } else { 2052 return _pcmp_neq; 2053 } 2054 } 2055 if (ptn1->meet(ptn2)) { 2056 return NULL; // Sets are not disjoint 2057 } 2058 2059 // Sets are disjoint. 2060 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj); 2061 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj); 2062 bool set1_has_null_ptr = ptn1->points_to(null_obj); 2063 bool set2_has_null_ptr = ptn2->points_to(null_obj); 2064 if ((set1_has_unknown_ptr && set2_has_null_ptr) || 2065 (set2_has_unknown_ptr && set1_has_null_ptr)) { 2066 // Check nullness of unknown object. 2067 return NULL; 2068 } 2069 2070 // Disjointness by itself is not sufficient since 2071 // alias analysis is not complete for escaped objects. 2072 // Disjoint sets are definitely unrelated only when 2073 // at least one set has only not escaping allocations. 2074 if (!set1_has_unknown_ptr && !set1_has_null_ptr) { 2075 if (ptn1->non_escaping_allocation()) { 2076 return _pcmp_neq; 2077 } 2078 } 2079 if (!set2_has_unknown_ptr && !set2_has_null_ptr) { 2080 if (ptn2->non_escaping_allocation()) { 2081 return _pcmp_neq; 2082 } 2083 } 2084 return NULL; 2085 } 2086 2087 // Connection Graph constuction functions. 2088 2089 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) { 2090 PointsToNode* ptadr = _nodes.at(n->_idx); 2091 if (ptadr != NULL) { 2092 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity"); 2093 return; 2094 } 2095 Compile* C = _compile; 2096 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es); 2097 _nodes.at_put(n->_idx, ptadr); 2098 } 2099 2100 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) { 2101 PointsToNode* ptadr = _nodes.at(n->_idx); 2102 if (ptadr != NULL) { 2103 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity"); 2104 return; 2105 } 2106 Compile* C = _compile; 2107 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es); 2108 _nodes.at_put(n->_idx, ptadr); 2109 } 2110 2111 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) { 2112 PointsToNode* ptadr = _nodes.at(n->_idx); 2113 if (ptadr != NULL) { 2114 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity"); 2115 return; 2116 } 2117 bool unsafe = false; 2118 bool is_oop = is_oop_field(n, offset, &unsafe); 2119 if (unsafe) { 2120 es = PointsToNode::GlobalEscape; 2121 } 2122 Compile* C = _compile; 2123 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop); 2124 _nodes.at_put(n->_idx, field); 2125 } 2126 2127 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es, 2128 PointsToNode* src, PointsToNode* dst) { 2129 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar"); 2130 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL"); 2131 PointsToNode* ptadr = _nodes.at(n->_idx); 2132 if (ptadr != NULL) { 2133 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity"); 2134 return; 2135 } 2136 Compile* C = _compile; 2137 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es); 2138 _nodes.at_put(n->_idx, ptadr); 2139 // Add edge from arraycopy node to source object. 2140 (void)add_edge(ptadr, src); 2141 src->set_arraycopy_src(); 2142 // Add edge from destination object to arraycopy node. 2143 (void)add_edge(dst, ptadr); 2144 dst->set_arraycopy_dst(); 2145 } 2146 2147 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) { 2148 const Type* adr_type = n->as_AddP()->bottom_type(); 2149 BasicType bt = T_INT; 2150 if (offset == Type::OffsetBot) { 2151 // Check only oop fields. 2152 if (!adr_type->isa_aryptr() || 2153 (adr_type->isa_aryptr()->klass() == NULL) || 2154 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) { 2155 // OffsetBot is used to reference array's element. Ignore first AddP. 2156 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) { 2157 bt = T_OBJECT; 2158 } 2159 } 2160 } else if (offset != oopDesc::klass_offset_in_bytes()) { 2161 if (adr_type->isa_instptr()) { 2162 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field(); 2163 if (field != NULL) { 2164 bt = field->layout_type(); 2165 } else { 2166 // Check for unsafe oop field access 2167 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2168 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2169 #if INCLUDE_SHENANDOAHGC 2170 n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) || 2171 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN) || 2172 #endif 2173 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) { 2174 bt = T_OBJECT; 2175 (*unsafe) = true; 2176 } 2177 } 2178 } else if (adr_type->isa_aryptr()) { 2179 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2180 // Ignore array length load. 2181 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) { 2182 // Ignore first AddP. 2183 } else { 2184 const Type* elemtype = adr_type->isa_aryptr()->elem(); 2185 bt = elemtype->array_element_basic_type(); 2186 } 2187 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) { 2188 // Allocation initialization, ThreadLocal field access, unsafe access 2189 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) || 2190 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) || 2191 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) { 2192 bt = T_OBJECT; 2193 } 2194 } 2195 } 2196 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY); 2197 } 2198 2199 // Returns unique pointed java object or NULL. 2200 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) { 2201 assert(!_collecting, "should not call when contructed graph"); 2202 // If the node was created after the escape computation we can't answer. 2203 uint idx = n->_idx; 2204 if (idx >= nodes_size()) { 2205 return NULL; 2206 } 2207 PointsToNode* ptn = ptnode_adr(idx); 2208 if (ptn->is_JavaObject()) { 2209 return ptn->as_JavaObject(); 2210 } 2211 assert(ptn->is_LocalVar(), "sanity"); 2212 // Check all java objects it points to. 2213 JavaObjectNode* jobj = NULL; 2214 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2215 PointsToNode* e = i.get(); 2216 if (e->is_JavaObject()) { 2217 if (jobj == NULL) { 2218 jobj = e->as_JavaObject(); 2219 } else if (jobj != e) { 2220 return NULL; 2221 } 2222 } 2223 } 2224 return jobj; 2225 } 2226 2227 // Return true if this node points only to non-escaping allocations. 2228 bool PointsToNode::non_escaping_allocation() { 2229 if (is_JavaObject()) { 2230 Node* n = ideal_node(); 2231 if (n->is_Allocate() || n->is_CallStaticJava()) { 2232 return (escape_state() == PointsToNode::NoEscape); 2233 } else { 2234 return false; 2235 } 2236 } 2237 assert(is_LocalVar(), "sanity"); 2238 // Check all java objects it points to. 2239 for (EdgeIterator i(this); i.has_next(); i.next()) { 2240 PointsToNode* e = i.get(); 2241 if (e->is_JavaObject()) { 2242 Node* n = e->ideal_node(); 2243 if ((e->escape_state() != PointsToNode::NoEscape) || 2244 !(n->is_Allocate() || n->is_CallStaticJava())) { 2245 return false; 2246 } 2247 } 2248 } 2249 return true; 2250 } 2251 2252 // Return true if we know the node does not escape globally. 2253 bool ConnectionGraph::not_global_escape(Node *n) { 2254 assert(!_collecting, "should not call during graph construction"); 2255 // If the node was created after the escape computation we can't answer. 2256 uint idx = n->_idx; 2257 if (idx >= nodes_size()) { 2258 return false; 2259 } 2260 PointsToNode* ptn = ptnode_adr(idx); 2261 PointsToNode::EscapeState es = ptn->escape_state(); 2262 // If we have already computed a value, return it. 2263 if (es >= PointsToNode::GlobalEscape) 2264 return false; 2265 if (ptn->is_JavaObject()) { 2266 return true; // (es < PointsToNode::GlobalEscape); 2267 } 2268 assert(ptn->is_LocalVar(), "sanity"); 2269 // Check all java objects it points to. 2270 for (EdgeIterator i(ptn); i.has_next(); i.next()) { 2271 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) 2272 return false; 2273 } 2274 return true; 2275 } 2276 2277 2278 // Helper functions 2279 2280 // Return true if this node points to specified node or nodes it points to. 2281 bool PointsToNode::points_to(JavaObjectNode* ptn) const { 2282 if (is_JavaObject()) { 2283 return (this == ptn); 2284 } 2285 assert(is_LocalVar() || is_Field(), "sanity"); 2286 for (EdgeIterator i(this); i.has_next(); i.next()) { 2287 if (i.get() == ptn) 2288 return true; 2289 } 2290 return false; 2291 } 2292 2293 // Return true if one node points to an other. 2294 bool PointsToNode::meet(PointsToNode* ptn) { 2295 if (this == ptn) { 2296 return true; 2297 } else if (ptn->is_JavaObject()) { 2298 return this->points_to(ptn->as_JavaObject()); 2299 } else if (this->is_JavaObject()) { 2300 return ptn->points_to(this->as_JavaObject()); 2301 } 2302 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity"); 2303 int ptn_count = ptn->edge_count(); 2304 for (EdgeIterator i(this); i.has_next(); i.next()) { 2305 PointsToNode* this_e = i.get(); 2306 for (int j = 0; j < ptn_count; j++) { 2307 if (this_e == ptn->edge(j)) 2308 return true; 2309 } 2310 } 2311 return false; 2312 } 2313 2314 #ifdef ASSERT 2315 // Return true if bases point to this java object. 2316 bool FieldNode::has_base(JavaObjectNode* jobj) const { 2317 for (BaseIterator i(this); i.has_next(); i.next()) { 2318 if (i.get() == jobj) 2319 return true; 2320 } 2321 return false; 2322 } 2323 #endif 2324 2325 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 2326 const Type *adr_type = phase->type(adr); 2327 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 2328 adr->in(AddPNode::Address)->is_Proj() && 2329 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2330 // We are computing a raw address for a store captured by an Initialize 2331 // compute an appropriate address type. AddP cases #3 and #5 (see below). 2332 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2333 assert(offs != Type::OffsetBot || 2334 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 2335 "offset must be a constant or it is initialization of array"); 2336 return offs; 2337 } 2338 const TypePtr *t_ptr = adr_type->isa_ptr(); 2339 assert(t_ptr != NULL, "must be a pointer type"); 2340 return t_ptr->offset(); 2341 } 2342 2343 Node* ConnectionGraph::get_addp_base(Node *addp) { 2344 assert(addp->is_AddP(), "must be AddP"); 2345 // 2346 // AddP cases for Base and Address inputs: 2347 // case #1. Direct object's field reference: 2348 // Allocate 2349 // | 2350 // Proj #5 ( oop result ) 2351 // | 2352 // CheckCastPP (cast to instance type) 2353 // | | 2354 // AddP ( base == address ) 2355 // 2356 // case #2. Indirect object's field reference: 2357 // Phi 2358 // | 2359 // CastPP (cast to instance type) 2360 // | | 2361 // AddP ( base == address ) 2362 // 2363 // case #3. Raw object's field reference for Initialize node: 2364 // Allocate 2365 // | 2366 // Proj #5 ( oop result ) 2367 // top | 2368 // \ | 2369 // AddP ( base == top ) 2370 // 2371 // case #4. Array's element reference: 2372 // {CheckCastPP | CastPP} 2373 // | | | 2374 // | AddP ( array's element offset ) 2375 // | | 2376 // AddP ( array's offset ) 2377 // 2378 // case #5. Raw object's field reference for arraycopy stub call: 2379 // The inline_native_clone() case when the arraycopy stub is called 2380 // after the allocation before Initialize and CheckCastPP nodes. 2381 // Allocate 2382 // | 2383 // Proj #5 ( oop result ) 2384 // | | 2385 // AddP ( base == address ) 2386 // 2387 // case #6. Constant Pool, ThreadLocal, CastX2P or 2388 // Raw object's field reference: 2389 // {ConP, ThreadLocal, CastX2P, raw Load} 2390 // top | 2391 // \ | 2392 // AddP ( base == top ) 2393 // 2394 // case #7. Klass's field reference. 2395 // LoadKlass 2396 // | | 2397 // AddP ( base == address ) 2398 // 2399 // case #8. narrow Klass's field reference. 2400 // LoadNKlass 2401 // | 2402 // DecodeN 2403 // | | 2404 // AddP ( base == address ) 2405 // 2406 // case #9. Mixed unsafe access 2407 // {instance} 2408 // | 2409 // CheckCastPP (raw) 2410 // top | 2411 // \ | 2412 // AddP ( base == top ) 2413 // 2414 Node *base = addp->in(AddPNode::Base); 2415 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9. 2416 base = addp->in(AddPNode::Address); 2417 while (base->is_AddP()) { 2418 // Case #6 (unsafe access) may have several chained AddP nodes. 2419 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only"); 2420 base = base->in(AddPNode::Address); 2421 } 2422 if (base->Opcode() == Op_CheckCastPP && 2423 base->bottom_type()->isa_rawptr() && 2424 _igvn->type(base->in(1))->isa_oopptr()) { 2425 base = base->in(1); // Case #9 2426 } else { 2427 Node* uncast_base = base->uncast(); 2428 int opcode = uncast_base->Opcode(); 2429 assert(opcode == Op_ConP || opcode == Op_ThreadLocal || 2430 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() || 2431 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) || 2432 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()) || 2433 uncast_base->is_ShenandoahBarrier(), "sanity"); 2434 } 2435 } 2436 return base; 2437 } 2438 2439 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) { 2440 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 2441 Node* addp2 = addp->raw_out(0); 2442 if (addp->outcnt() == 1 && addp2->is_AddP() && 2443 addp2->in(AddPNode::Base) == n && 2444 addp2->in(AddPNode::Address) == addp) { 2445 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 2446 // 2447 // Find array's offset to push it on worklist first and 2448 // as result process an array's element offset first (pushed second) 2449 // to avoid CastPP for the array's offset. 2450 // Otherwise the inserted CastPP (LocalVar) will point to what 2451 // the AddP (Field) points to. Which would be wrong since 2452 // the algorithm expects the CastPP has the same point as 2453 // as AddP's base CheckCastPP (LocalVar). 2454 // 2455 // ArrayAllocation 2456 // | 2457 // CheckCastPP 2458 // | 2459 // memProj (from ArrayAllocation CheckCastPP) 2460 // | || 2461 // | || Int (element index) 2462 // | || | ConI (log(element size)) 2463 // | || | / 2464 // | || LShift 2465 // | || / 2466 // | AddP (array's element offset) 2467 // | | 2468 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 2469 // | / / 2470 // AddP (array's offset) 2471 // | 2472 // Load/Store (memory operation on array's element) 2473 // 2474 return addp2; 2475 } 2476 return NULL; 2477 } 2478 2479 // 2480 // Adjust the type and inputs of an AddP which computes the 2481 // address of a field of an instance 2482 // 2483 bool ConnectionGraph::split_AddP(Node *addp, Node *base) { 2484 PhaseGVN* igvn = _igvn; 2485 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 2486 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 2487 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 2488 if (t == NULL) { 2489 // We are computing a raw address for a store captured by an Initialize 2490 // compute an appropriate address type (cases #3 and #5). 2491 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 2492 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 2493 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 2494 assert(offs != Type::OffsetBot, "offset must be a constant"); 2495 t = base_t->add_offset(offs)->is_oopptr(); 2496 } 2497 int inst_id = base_t->instance_id(); 2498 assert(!t->is_known_instance() || t->instance_id() == inst_id, 2499 "old type must be non-instance or match new type"); 2500 2501 // The type 't' could be subclass of 'base_t'. 2502 // As result t->offset() could be large then base_t's size and it will 2503 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 2504 // constructor verifies correctness of the offset. 2505 // 2506 // It could happened on subclass's branch (from the type profiling 2507 // inlining) which was not eliminated during parsing since the exactness 2508 // of the allocation type was not propagated to the subclass type check. 2509 // 2510 // Or the type 't' could be not related to 'base_t' at all. 2511 // It could happened when CHA type is different from MDO type on a dead path 2512 // (for example, from instanceof check) which is not collapsed during parsing. 2513 // 2514 // Do nothing for such AddP node and don't process its users since 2515 // this code branch will go away. 2516 // 2517 if (!t->is_known_instance() && 2518 !base_t->klass()->is_subtype_of(t->klass())) { 2519 return false; // bail out 2520 } 2521 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 2522 // Do NOT remove the next line: ensure a new alias index is allocated 2523 // for the instance type. Note: C++ will not remove it since the call 2524 // has side effect. 2525 int alias_idx = _compile->get_alias_index(tinst); 2526 igvn->set_type(addp, tinst); 2527 // record the allocation in the node map 2528 set_map(addp, get_map(base->_idx)); 2529 // Set addp's Base and Address to 'base'. 2530 Node *abase = addp->in(AddPNode::Base); 2531 Node *adr = addp->in(AddPNode::Address); 2532 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 2533 adr->in(0)->_idx == (uint)inst_id) { 2534 // Skip AddP cases #3 and #5. 2535 } else { 2536 assert(!abase->is_top(), "sanity"); // AddP case #3 2537 if (abase != base) { 2538 igvn->hash_delete(addp); 2539 addp->set_req(AddPNode::Base, base); 2540 if (abase == adr) { 2541 addp->set_req(AddPNode::Address, base); 2542 } else { 2543 // AddP case #4 (adr is array's element offset AddP node) 2544 #ifdef ASSERT 2545 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 2546 assert(adr->is_AddP() && atype != NULL && 2547 atype->instance_id() == inst_id, "array's element offset should be processed first"); 2548 #endif 2549 } 2550 igvn->hash_insert(addp); 2551 } 2552 } 2553 // Put on IGVN worklist since at least addp's type was changed above. 2554 record_for_optimizer(addp); 2555 return true; 2556 } 2557 2558 // 2559 // Create a new version of orig_phi if necessary. Returns either the newly 2560 // created phi or an existing phi. Sets create_new to indicate whether a new 2561 // phi was created. Cache the last newly created phi in the node map. 2562 // 2563 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) { 2564 Compile *C = _compile; 2565 PhaseGVN* igvn = _igvn; 2566 new_created = false; 2567 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 2568 // nothing to do if orig_phi is bottom memory or matches alias_idx 2569 if (phi_alias_idx == alias_idx) { 2570 return orig_phi; 2571 } 2572 // Have we recently created a Phi for this alias index? 2573 PhiNode *result = get_map_phi(orig_phi->_idx); 2574 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 2575 return result; 2576 } 2577 // Previous check may fail when the same wide memory Phi was split into Phis 2578 // for different memory slices. Search all Phis for this region. 2579 if (result != NULL) { 2580 Node* region = orig_phi->in(0); 2581 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2582 Node* phi = region->fast_out(i); 2583 if (phi->is_Phi() && 2584 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 2585 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 2586 return phi->as_Phi(); 2587 } 2588 } 2589 } 2590 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) { 2591 if (C->do_escape_analysis() == true && !C->failing()) { 2592 // Retry compilation without escape analysis. 2593 // If this is the first failure, the sentinel string will "stick" 2594 // to the Compile object, and the C2Compiler will see it and retry. 2595 C->record_failure(C2Compiler::retry_no_escape_analysis()); 2596 } 2597 return NULL; 2598 } 2599 orig_phi_worklist.append_if_missing(orig_phi); 2600 const TypePtr *atype = C->get_adr_type(alias_idx); 2601 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 2602 C->copy_node_notes_to(result, orig_phi); 2603 igvn->set_type(result, result->bottom_type()); 2604 record_for_optimizer(result); 2605 set_map(orig_phi, result); 2606 new_created = true; 2607 return result; 2608 } 2609 2610 // 2611 // Return a new version of Memory Phi "orig_phi" with the inputs having the 2612 // specified alias index. 2613 // 2614 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) { 2615 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 2616 Compile *C = _compile; 2617 PhaseGVN* igvn = _igvn; 2618 bool new_phi_created; 2619 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created); 2620 if (!new_phi_created) { 2621 return result; 2622 } 2623 GrowableArray<PhiNode *> phi_list; 2624 GrowableArray<uint> cur_input; 2625 PhiNode *phi = orig_phi; 2626 uint idx = 1; 2627 bool finished = false; 2628 while(!finished) { 2629 while (idx < phi->req()) { 2630 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist); 2631 if (mem != NULL && mem->is_Phi()) { 2632 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created); 2633 if (new_phi_created) { 2634 // found an phi for which we created a new split, push current one on worklist and begin 2635 // processing new one 2636 phi_list.push(phi); 2637 cur_input.push(idx); 2638 phi = mem->as_Phi(); 2639 result = newphi; 2640 idx = 1; 2641 continue; 2642 } else { 2643 mem = newphi; 2644 } 2645 } 2646 if (C->failing()) { 2647 return NULL; 2648 } 2649 result->set_req(idx++, mem); 2650 } 2651 #ifdef ASSERT 2652 // verify that the new Phi has an input for each input of the original 2653 assert( phi->req() == result->req(), "must have same number of inputs."); 2654 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 2655 #endif 2656 // Check if all new phi's inputs have specified alias index. 2657 // Otherwise use old phi. 2658 for (uint i = 1; i < phi->req(); i++) { 2659 Node* in = result->in(i); 2660 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 2661 } 2662 // we have finished processing a Phi, see if there are any more to do 2663 finished = (phi_list.length() == 0 ); 2664 if (!finished) { 2665 phi = phi_list.pop(); 2666 idx = cur_input.pop(); 2667 PhiNode *prev_result = get_map_phi(phi->_idx); 2668 prev_result->set_req(idx++, result); 2669 result = prev_result; 2670 } 2671 } 2672 return result; 2673 } 2674 2675 // 2676 // The next methods are derived from methods in MemNode. 2677 // 2678 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 2679 Node *mem = mmem; 2680 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 2681 // means an array I have not precisely typed yet. Do not do any 2682 // alias stuff with it any time soon. 2683 if (toop->base() != Type::AnyPtr && 2684 !(toop->klass() != NULL && 2685 toop->klass()->is_java_lang_Object() && 2686 toop->offset() == Type::OffsetBot)) { 2687 mem = mmem->memory_at(alias_idx); 2688 // Update input if it is progress over what we have now 2689 } 2690 return mem; 2691 } 2692 2693 // 2694 // Move memory users to their memory slices. 2695 // 2696 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) { 2697 Compile* C = _compile; 2698 PhaseGVN* igvn = _igvn; 2699 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 2700 assert(tp != NULL, "ptr type"); 2701 int alias_idx = C->get_alias_index(tp); 2702 int general_idx = C->get_general_index(alias_idx); 2703 2704 // Move users first 2705 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2706 Node* use = n->fast_out(i); 2707 if (use->is_MergeMem()) { 2708 MergeMemNode* mmem = use->as_MergeMem(); 2709 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 2710 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 2711 continue; // Nothing to do 2712 } 2713 // Replace previous general reference to mem node. 2714 uint orig_uniq = C->unique(); 2715 Node* m = find_inst_mem(n, general_idx, orig_phis); 2716 assert(orig_uniq == C->unique(), "no new nodes"); 2717 mmem->set_memory_at(general_idx, m); 2718 --imax; 2719 --i; 2720 } else if (use->is_MemBar()) { 2721 assert(!use->is_Initialize(), "initializing stores should not be moved"); 2722 if (use->req() > MemBarNode::Precedent && 2723 use->in(MemBarNode::Precedent) == n) { 2724 // Don't move related membars. 2725 record_for_optimizer(use); 2726 continue; 2727 } 2728 tp = use->as_MemBar()->adr_type()->isa_ptr(); 2729 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) || 2730 alias_idx == general_idx) { 2731 continue; // Nothing to do 2732 } 2733 // Move to general memory slice. 2734 uint orig_uniq = C->unique(); 2735 Node* m = find_inst_mem(n, general_idx, orig_phis); 2736 assert(orig_uniq == C->unique(), "no new nodes"); 2737 igvn->hash_delete(use); 2738 imax -= use->replace_edge(n, m); 2739 igvn->hash_insert(use); 2740 record_for_optimizer(use); 2741 --i; 2742 #ifdef ASSERT 2743 } else if (use->is_Mem()) { 2744 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 2745 // Don't move related cardmark. 2746 continue; 2747 } 2748 // Memory nodes should have new memory input. 2749 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 2750 assert(tp != NULL, "ptr type"); 2751 int idx = C->get_alias_index(tp); 2752 assert(get_map(use->_idx) != NULL || idx == alias_idx, 2753 "Following memory nodes should have new memory input or be on the same memory slice"); 2754 } else if (use->is_Phi()) { 2755 // Phi nodes should be split and moved already. 2756 tp = use->as_Phi()->adr_type()->isa_ptr(); 2757 assert(tp != NULL, "ptr type"); 2758 int idx = C->get_alias_index(tp); 2759 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 2760 } else { 2761 use->dump(); 2762 assert(false, "should not be here"); 2763 #endif 2764 } 2765 } 2766 } 2767 2768 // 2769 // Search memory chain of "mem" to find a MemNode whose address 2770 // is the specified alias index. 2771 // 2772 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) { 2773 if (orig_mem == NULL) 2774 return orig_mem; 2775 Compile* C = _compile; 2776 PhaseGVN* igvn = _igvn; 2777 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 2778 bool is_instance = (toop != NULL) && toop->is_known_instance(); 2779 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 2780 Node *prev = NULL; 2781 Node *result = orig_mem; 2782 while (prev != result) { 2783 prev = result; 2784 if (result == start_mem) 2785 break; // hit one of our sentinels 2786 if (result->is_Mem()) { 2787 const Type *at = igvn->type(result->in(MemNode::Address)); 2788 if (at == Type::TOP) 2789 break; // Dead 2790 assert (at->isa_ptr() != NULL, "pointer type required."); 2791 int idx = C->get_alias_index(at->is_ptr()); 2792 if (idx == alias_idx) 2793 break; // Found 2794 if (!is_instance && (at->isa_oopptr() == NULL || 2795 !at->is_oopptr()->is_known_instance())) { 2796 break; // Do not skip store to general memory slice. 2797 } 2798 result = result->in(MemNode::Memory); 2799 } 2800 if (!is_instance) 2801 continue; // don't search further for non-instance types 2802 // skip over a call which does not affect this memory slice 2803 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 2804 Node *proj_in = result->in(0); 2805 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 2806 break; // hit one of our sentinels 2807 } else if (proj_in->is_Call()) { 2808 // ArrayCopy node processed here as well 2809 CallNode *call = proj_in->as_Call(); 2810 if (!call->may_modify(toop, igvn)) { 2811 result = call->in(TypeFunc::Memory); 2812 } 2813 } else if (proj_in->is_Initialize()) { 2814 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 2815 // Stop if this is the initialization for the object instance which 2816 // which contains this memory slice, otherwise skip over it. 2817 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 2818 result = proj_in->in(TypeFunc::Memory); 2819 } 2820 } else if (proj_in->is_MemBar()) { 2821 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() && 2822 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() && 2823 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) { 2824 // clone 2825 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy(); 2826 if (ac->may_modify(toop, igvn)) { 2827 break; 2828 } 2829 } 2830 result = proj_in->in(TypeFunc::Memory); 2831 } 2832 } else if (result->is_MergeMem()) { 2833 MergeMemNode *mmem = result->as_MergeMem(); 2834 result = step_through_mergemem(mmem, alias_idx, toop); 2835 if (result == mmem->base_memory()) { 2836 // Didn't find instance memory, search through general slice recursively. 2837 result = mmem->memory_at(C->get_general_index(alias_idx)); 2838 result = find_inst_mem(result, alias_idx, orig_phis); 2839 if (C->failing()) { 2840 return NULL; 2841 } 2842 mmem->set_memory_at(alias_idx, result); 2843 } 2844 } else if (result->is_Phi() && 2845 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 2846 Node *un = result->as_Phi()->unique_input(igvn); 2847 if (un != NULL) { 2848 orig_phis.append_if_missing(result->as_Phi()); 2849 result = un; 2850 } else { 2851 break; 2852 } 2853 } else if (result->is_ClearArray()) { 2854 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) { 2855 // Can not bypass initialization of the instance 2856 // we are looking for. 2857 break; 2858 } 2859 // Otherwise skip it (the call updated 'result' value). 2860 } else if (result->Opcode() == Op_SCMemProj) { 2861 Node* mem = result->in(0); 2862 Node* adr = NULL; 2863 if (mem->is_LoadStore()) { 2864 adr = mem->in(MemNode::Address); 2865 } else { 2866 assert(mem->Opcode() == Op_EncodeISOArray || 2867 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 2868 adr = mem->in(3); // Memory edge corresponds to destination array 2869 } 2870 const Type *at = igvn->type(adr); 2871 if (at != Type::TOP) { 2872 assert(at->isa_ptr() != NULL, "pointer type required."); 2873 int idx = C->get_alias_index(at->is_ptr()); 2874 if (idx == alias_idx) { 2875 // Assert in debug mode 2876 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 2877 break; // In product mode return SCMemProj node 2878 } 2879 } 2880 result = mem->in(MemNode::Memory); 2881 } else if (result->Opcode() == Op_StrInflatedCopy) { 2882 Node* adr = result->in(3); // Memory edge corresponds to destination array 2883 const Type *at = igvn->type(adr); 2884 if (at != Type::TOP) { 2885 assert(at->isa_ptr() != NULL, "pointer type required."); 2886 int idx = C->get_alias_index(at->is_ptr()); 2887 if (idx == alias_idx) { 2888 // Assert in debug mode 2889 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 2890 break; // In product mode return SCMemProj node 2891 } 2892 } 2893 result = result->in(MemNode::Memory); 2894 } 2895 } 2896 if (result->is_Phi()) { 2897 PhiNode *mphi = result->as_Phi(); 2898 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 2899 const TypePtr *t = mphi->adr_type(); 2900 if (!is_instance) { 2901 // Push all non-instance Phis on the orig_phis worklist to update inputs 2902 // during Phase 4 if needed. 2903 orig_phis.append_if_missing(mphi); 2904 } else if (C->get_alias_index(t) != alias_idx) { 2905 // Create a new Phi with the specified alias index type. 2906 result = split_memory_phi(mphi, alias_idx, orig_phis); 2907 } 2908 } 2909 // the result is either MemNode, PhiNode, InitializeNode. 2910 return result; 2911 } 2912 2913 // 2914 // Convert the types of unescaped object to instance types where possible, 2915 // propagate the new type information through the graph, and update memory 2916 // edges and MergeMem inputs to reflect the new type. 2917 // 2918 // We start with allocations (and calls which may be allocations) on alloc_worklist. 2919 // The processing is done in 4 phases: 2920 // 2921 // Phase 1: Process possible allocations from alloc_worklist. Create instance 2922 // types for the CheckCastPP for allocations where possible. 2923 // Propagate the new types through users as follows: 2924 // casts and Phi: push users on alloc_worklist 2925 // AddP: cast Base and Address inputs to the instance type 2926 // push any AddP users on alloc_worklist and push any memnode 2927 // users onto memnode_worklist. 2928 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 2929 // search the Memory chain for a store with the appropriate type 2930 // address type. If a Phi is found, create a new version with 2931 // the appropriate memory slices from each of the Phi inputs. 2932 // For stores, process the users as follows: 2933 // MemNode: push on memnode_worklist 2934 // MergeMem: push on mergemem_worklist 2935 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 2936 // moving the first node encountered of each instance type to the 2937 // the input corresponding to its alias index. 2938 // appropriate memory slice. 2939 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 2940 // 2941 // In the following example, the CheckCastPP nodes are the cast of allocation 2942 // results and the allocation of node 29 is unescaped and eligible to be an 2943 // instance type. 2944 // 2945 // We start with: 2946 // 2947 // 7 Parm #memory 2948 // 10 ConI "12" 2949 // 19 CheckCastPP "Foo" 2950 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2951 // 29 CheckCastPP "Foo" 2952 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 2953 // 2954 // 40 StoreP 25 7 20 ... alias_index=4 2955 // 50 StoreP 35 40 30 ... alias_index=4 2956 // 60 StoreP 45 50 20 ... alias_index=4 2957 // 70 LoadP _ 60 30 ... alias_index=4 2958 // 80 Phi 75 50 60 Memory alias_index=4 2959 // 90 LoadP _ 80 30 ... alias_index=4 2960 // 100 LoadP _ 80 20 ... alias_index=4 2961 // 2962 // 2963 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 2964 // and creating a new alias index for node 30. This gives: 2965 // 2966 // 7 Parm #memory 2967 // 10 ConI "12" 2968 // 19 CheckCastPP "Foo" 2969 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2970 // 29 CheckCastPP "Foo" iid=24 2971 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2972 // 2973 // 40 StoreP 25 7 20 ... alias_index=4 2974 // 50 StoreP 35 40 30 ... alias_index=6 2975 // 60 StoreP 45 50 20 ... alias_index=4 2976 // 70 LoadP _ 60 30 ... alias_index=6 2977 // 80 Phi 75 50 60 Memory alias_index=4 2978 // 90 LoadP _ 80 30 ... alias_index=6 2979 // 100 LoadP _ 80 20 ... alias_index=4 2980 // 2981 // In phase 2, new memory inputs are computed for the loads and stores, 2982 // And a new version of the phi is created. In phase 4, the inputs to 2983 // node 80 are updated and then the memory nodes are updated with the 2984 // values computed in phase 2. This results in: 2985 // 2986 // 7 Parm #memory 2987 // 10 ConI "12" 2988 // 19 CheckCastPP "Foo" 2989 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 2990 // 29 CheckCastPP "Foo" iid=24 2991 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 2992 // 2993 // 40 StoreP 25 7 20 ... alias_index=4 2994 // 50 StoreP 35 7 30 ... alias_index=6 2995 // 60 StoreP 45 40 20 ... alias_index=4 2996 // 70 LoadP _ 50 30 ... alias_index=6 2997 // 80 Phi 75 40 60 Memory alias_index=4 2998 // 120 Phi 75 50 50 Memory alias_index=6 2999 // 90 LoadP _ 120 30 ... alias_index=6 3000 // 100 LoadP _ 80 20 ... alias_index=4 3001 // 3002 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) { 3003 GrowableArray<Node *> memnode_worklist; 3004 GrowableArray<PhiNode *> orig_phis; 3005 PhaseIterGVN *igvn = _igvn; 3006 uint new_index_start = (uint) _compile->num_alias_types(); 3007 Arena* arena = Thread::current()->resource_area(); 3008 VectorSet visited(arena); 3009 ideal_nodes.clear(); // Reset for use with set_map/get_map. 3010 uint unique_old = _compile->unique(); 3011 3012 // Phase 1: Process possible allocations from alloc_worklist. 3013 // Create instance types for the CheckCastPP for allocations where possible. 3014 // 3015 // (Note: don't forget to change the order of the second AddP node on 3016 // the alloc_worklist if the order of the worklist processing is changed, 3017 // see the comment in find_second_addp().) 3018 // 3019 while (alloc_worklist.length() != 0) { 3020 Node *n = alloc_worklist.pop(); 3021 uint ni = n->_idx; 3022 if (n->is_Call()) { 3023 CallNode *alloc = n->as_Call(); 3024 // copy escape information to call node 3025 PointsToNode* ptn = ptnode_adr(alloc->_idx); 3026 PointsToNode::EscapeState es = ptn->escape_state(); 3027 // We have an allocation or call which returns a Java object, 3028 // see if it is unescaped. 3029 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) 3030 continue; 3031 // Find CheckCastPP for the allocate or for the return value of a call 3032 n = alloc->result_cast(); 3033 if (n == NULL) { // No uses except Initialize node 3034 if (alloc->is_Allocate()) { 3035 // Set the scalar_replaceable flag for allocation 3036 // so it could be eliminated if it has no uses. 3037 alloc->as_Allocate()->_is_scalar_replaceable = true; 3038 } 3039 if (alloc->is_CallStaticJava()) { 3040 // Set the scalar_replaceable flag for boxing method 3041 // so it could be eliminated if it has no uses. 3042 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3043 } 3044 continue; 3045 } 3046 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 3047 assert(!alloc->is_Allocate(), "allocation should have unique type"); 3048 continue; 3049 } 3050 3051 // The inline code for Object.clone() casts the allocation result to 3052 // java.lang.Object and then to the actual type of the allocated 3053 // object. Detect this case and use the second cast. 3054 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 3055 // the allocation result is cast to java.lang.Object and then 3056 // to the actual Array type. 3057 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 3058 && (alloc->is_AllocateArray() || 3059 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 3060 Node *cast2 = NULL; 3061 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3062 Node *use = n->fast_out(i); 3063 if (use->is_CheckCastPP()) { 3064 cast2 = use; 3065 break; 3066 } 3067 } 3068 if (cast2 != NULL) { 3069 n = cast2; 3070 } else { 3071 // Non-scalar replaceable if the allocation type is unknown statically 3072 // (reflection allocation), the object can't be restored during 3073 // deoptimization without precise type. 3074 continue; 3075 } 3076 } 3077 3078 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 3079 if (t == NULL) 3080 continue; // not a TypeOopPtr 3081 if (!t->klass_is_exact()) 3082 continue; // not an unique type 3083 3084 if (alloc->is_Allocate()) { 3085 // Set the scalar_replaceable flag for allocation 3086 // so it could be eliminated. 3087 alloc->as_Allocate()->_is_scalar_replaceable = true; 3088 } 3089 if (alloc->is_CallStaticJava()) { 3090 // Set the scalar_replaceable flag for boxing method 3091 // so it could be eliminated. 3092 alloc->as_CallStaticJava()->_is_scalar_replaceable = true; 3093 } 3094 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state 3095 // in order for an object to be scalar-replaceable, it must be: 3096 // - a direct allocation (not a call returning an object) 3097 // - non-escaping 3098 // - eligible to be a unique type 3099 // - not determined to be ineligible by escape analysis 3100 set_map(alloc, n); 3101 set_map(n, alloc); 3102 const TypeOopPtr* tinst = t->cast_to_instance_id(ni); 3103 igvn->hash_delete(n); 3104 igvn->set_type(n, tinst); 3105 n->raise_bottom_type(tinst); 3106 igvn->hash_insert(n); 3107 record_for_optimizer(n); 3108 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 3109 3110 // First, put on the worklist all Field edges from Connection Graph 3111 // which is more accurate than putting immediate users from Ideal Graph. 3112 for (EdgeIterator e(ptn); e.has_next(); e.next()) { 3113 PointsToNode* tgt = e.get(); 3114 if (tgt->is_Arraycopy()) { 3115 continue; 3116 } 3117 Node* use = tgt->ideal_node(); 3118 assert(tgt->is_Field() && use->is_AddP(), 3119 "only AddP nodes are Field edges in CG"); 3120 if (use->outcnt() > 0) { // Don't process dead nodes 3121 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 3122 if (addp2 != NULL) { 3123 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3124 alloc_worklist.append_if_missing(addp2); 3125 } 3126 alloc_worklist.append_if_missing(use); 3127 } 3128 } 3129 3130 // An allocation may have an Initialize which has raw stores. Scan 3131 // the users of the raw allocation result and push AddP users 3132 // on alloc_worklist. 3133 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms); 3134 assert (raw_result != NULL, "must have an allocation result"); 3135 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 3136 Node *use = raw_result->fast_out(i); 3137 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 3138 Node* addp2 = find_second_addp(use, raw_result); 3139 if (addp2 != NULL) { 3140 assert(alloc->is_AllocateArray(),"array allocation was expected"); 3141 alloc_worklist.append_if_missing(addp2); 3142 } 3143 alloc_worklist.append_if_missing(use); 3144 } else if (use->is_MemBar()) { 3145 memnode_worklist.append_if_missing(use); 3146 } 3147 } 3148 } 3149 } else if (n->is_AddP()) { 3150 JavaObjectNode* jobj = unique_java_object(get_addp_base(n)); 3151 if (jobj == NULL || jobj == phantom_obj) { 3152 #ifdef ASSERT 3153 ptnode_adr(get_addp_base(n)->_idx)->dump(); 3154 ptnode_adr(n->_idx)->dump(); 3155 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3156 #endif 3157 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3158 return; 3159 } 3160 Node *base = get_map(jobj->idx()); // CheckCastPP node 3161 if (!split_AddP(n, base)) continue; // wrong type from dead path 3162 } else if (n->is_Phi() || 3163 n->is_CheckCastPP() || 3164 n->is_EncodeP() || 3165 n->is_DecodeN() || 3166 n->is_ShenandoahBarrier() || 3167 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 3168 if (visited.test_set(n->_idx)) { 3169 assert(n->is_Phi(), "loops only through Phi's"); 3170 continue; // already processed 3171 } 3172 JavaObjectNode* jobj = unique_java_object(n); 3173 if (jobj == NULL || jobj == phantom_obj) { 3174 #ifdef ASSERT 3175 ptnode_adr(n->_idx)->dump(); 3176 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation"); 3177 #endif 3178 _compile->record_failure(C2Compiler::retry_no_escape_analysis()); 3179 return; 3180 } else { 3181 Node *val = get_map(jobj->idx()); // CheckCastPP node 3182 TypeNode *tn = n->as_Type(); 3183 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr(); 3184 assert(tinst != NULL && tinst->is_known_instance() && 3185 tinst->instance_id() == jobj->idx() , "instance type expected."); 3186 3187 const Type *tn_type = igvn->type(tn); 3188 const TypeOopPtr *tn_t; 3189 if (tn_type->isa_narrowoop()) { 3190 tn_t = tn_type->make_ptr()->isa_oopptr(); 3191 } else { 3192 tn_t = tn_type->isa_oopptr(); 3193 } 3194 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 3195 if (tn_type->isa_narrowoop()) { 3196 tn_type = tinst->make_narrowoop(); 3197 } else { 3198 tn_type = tinst; 3199 } 3200 igvn->hash_delete(tn); 3201 igvn->set_type(tn, tn_type); 3202 tn->set_type(tn_type); 3203 igvn->hash_insert(tn); 3204 record_for_optimizer(n); 3205 } else { 3206 assert(tn_type == TypePtr::NULL_PTR || 3207 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 3208 "unexpected type"); 3209 continue; // Skip dead path with different type 3210 } 3211 } 3212 } else { 3213 debug_only(n->dump();) 3214 assert(false, "EA: unexpected node"); 3215 continue; 3216 } 3217 // push allocation's users on appropriate worklist 3218 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3219 Node *use = n->fast_out(i); 3220 if(use->is_Mem() && use->in(MemNode::Address) == n) { 3221 // Load/store to instance's field 3222 memnode_worklist.append_if_missing(use); 3223 } else if (use->is_MemBar()) { 3224 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3225 memnode_worklist.append_if_missing(use); 3226 } 3227 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 3228 Node* addp2 = find_second_addp(use, n); 3229 if (addp2 != NULL) { 3230 alloc_worklist.append_if_missing(addp2); 3231 } 3232 alloc_worklist.append_if_missing(use); 3233 } else if (use->is_Phi() || 3234 use->is_CheckCastPP() || 3235 use->is_EncodeNarrowPtr() || 3236 use->is_DecodeNarrowPtr() || 3237 use->is_ShenandoahBarrier() || 3238 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 3239 alloc_worklist.append_if_missing(use); 3240 #ifdef ASSERT 3241 } else if (use->is_Mem()) { 3242 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 3243 } else if (use->is_MergeMem()) { 3244 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3245 } else if (use->is_SafePoint()) { 3246 // Look for MergeMem nodes for calls which reference unique allocation 3247 // (through CheckCastPP nodes) even for debug info. 3248 Node* m = use->in(TypeFunc::Memory); 3249 if (m->is_MergeMem()) { 3250 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3251 } 3252 } else if (use->Opcode() == Op_EncodeISOArray) { 3253 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3254 // EncodeISOArray overwrites destination array 3255 memnode_worklist.append_if_missing(use); 3256 } 3257 } else { 3258 uint op = use->Opcode(); 3259 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) && 3260 (use->in(MemNode::Memory) == n)) { 3261 // They overwrite memory edge corresponding to destination array, 3262 memnode_worklist.append_if_missing(use); 3263 } else if (!(op == Op_CmpP || op == Op_Conv2B || 3264 op == Op_CastP2X || op == Op_StoreCM || 3265 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3266 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3267 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || 3268 op == Op_ShenandoahWBMemProj || 3269 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) { 3270 n->dump(); 3271 use->dump(); 3272 assert(false, "EA: missing allocation reference path"); 3273 } 3274 #endif 3275 } 3276 } 3277 3278 } 3279 3280 // Go over all ArrayCopy nodes and if one of the inputs has a unique 3281 // type, record it in the ArrayCopy node so we know what memory this 3282 // node uses/modified. 3283 for (int next = 0; next < arraycopy_worklist.length(); next++) { 3284 ArrayCopyNode* ac = arraycopy_worklist.at(next); 3285 Node* dest = ac->in(ArrayCopyNode::Dest); 3286 if (dest->is_AddP()) { 3287 dest = get_addp_base(dest); 3288 } 3289 JavaObjectNode* jobj = unique_java_object(dest); 3290 if (jobj != NULL) { 3291 Node *base = get_map(jobj->idx()); 3292 if (base != NULL) { 3293 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3294 ac->_dest_type = base_t; 3295 } 3296 } 3297 Node* src = ac->in(ArrayCopyNode::Src); 3298 if (src->is_AddP()) { 3299 src = get_addp_base(src); 3300 } 3301 jobj = unique_java_object(src); 3302 if (jobj != NULL) { 3303 Node* base = get_map(jobj->idx()); 3304 if (base != NULL) { 3305 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr(); 3306 ac->_src_type = base_t; 3307 } 3308 } 3309 } 3310 3311 // New alias types were created in split_AddP(). 3312 uint new_index_end = (uint) _compile->num_alias_types(); 3313 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1"); 3314 3315 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 3316 // compute new values for Memory inputs (the Memory inputs are not 3317 // actually updated until phase 4.) 3318 if (memnode_worklist.length() == 0) 3319 return; // nothing to do 3320 while (memnode_worklist.length() != 0) { 3321 Node *n = memnode_worklist.pop(); 3322 if (visited.test_set(n->_idx)) 3323 continue; 3324 if (n->is_Phi() || n->is_ClearArray()) { 3325 // we don't need to do anything, but the users must be pushed 3326 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 3327 // we don't need to do anything, but the users must be pushed 3328 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory); 3329 if (n == NULL) 3330 continue; 3331 } else if (n->Opcode() == Op_StrCompressedCopy || 3332 n->Opcode() == Op_EncodeISOArray) { 3333 // get the memory projection 3334 n = n->find_out_with(Op_SCMemProj); 3335 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3336 } else { 3337 assert(n->is_Mem(), "memory node required."); 3338 Node *addr = n->in(MemNode::Address); 3339 const Type *addr_t = igvn->type(addr); 3340 if (addr_t == Type::TOP) 3341 continue; 3342 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 3343 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 3344 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 3345 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis); 3346 if (_compile->failing()) { 3347 return; 3348 } 3349 if (mem != n->in(MemNode::Memory)) { 3350 // We delay the memory edge update since we need old one in 3351 // MergeMem code below when instances memory slices are separated. 3352 set_map(n, mem); 3353 } 3354 if (n->is_Load()) { 3355 continue; // don't push users 3356 } else if (n->is_LoadStore()) { 3357 // get the memory projection 3358 n = n->find_out_with(Op_SCMemProj); 3359 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required"); 3360 } 3361 } 3362 // push user on appropriate worklist 3363 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3364 Node *use = n->fast_out(i); 3365 if (use->is_Phi() || use->is_ClearArray()) { 3366 memnode_worklist.append_if_missing(use); 3367 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) { 3368 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 3369 continue; 3370 memnode_worklist.append_if_missing(use); 3371 } else if (use->is_MemBar()) { 3372 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge 3373 memnode_worklist.append_if_missing(use); 3374 } 3375 #ifdef ASSERT 3376 } else if(use->is_Mem()) { 3377 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 3378 } else if (use->is_MergeMem()) { 3379 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 3380 } else if (use->Opcode() == Op_EncodeISOArray) { 3381 if (use->in(MemNode::Memory) == n || use->in(3) == n) { 3382 // EncodeISOArray overwrites destination array 3383 memnode_worklist.append_if_missing(use); 3384 } 3385 } else { 3386 uint op = use->Opcode(); 3387 if ((use->in(MemNode::Memory) == n) && 3388 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) { 3389 // They overwrite memory edge corresponding to destination array, 3390 memnode_worklist.append_if_missing(use); 3391 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) || 3392 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives || 3393 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || 3394 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) { 3395 n->dump(); 3396 use->dump(); 3397 assert(false, "EA: missing memory path"); 3398 } 3399 #endif 3400 } 3401 } 3402 } 3403 3404 // Phase 3: Process MergeMem nodes from mergemem_worklist. 3405 // Walk each memory slice moving the first node encountered of each 3406 // instance type to the the input corresponding to its alias index. 3407 uint length = _mergemem_worklist.length(); 3408 for( uint next = 0; next < length; ++next ) { 3409 MergeMemNode* nmm = _mergemem_worklist.at(next); 3410 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 3411 // Note: we don't want to use MergeMemStream here because we only want to 3412 // scan inputs which exist at the start, not ones we add during processing. 3413 // Note 2: MergeMem may already contains instance memory slices added 3414 // during find_inst_mem() call when memory nodes were processed above. 3415 igvn->hash_delete(nmm); 3416 uint nslices = MIN2(nmm->req(), new_index_start); 3417 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 3418 Node* mem = nmm->in(i); 3419 Node* cur = NULL; 3420 if (mem == NULL || mem->is_top()) 3421 continue; 3422 // First, update mergemem by moving memory nodes to corresponding slices 3423 // if their type became more precise since this mergemem was created. 3424 while (mem->is_Mem()) { 3425 const Type *at = igvn->type(mem->in(MemNode::Address)); 3426 if (at != Type::TOP) { 3427 assert (at->isa_ptr() != NULL, "pointer type required."); 3428 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 3429 if (idx == i) { 3430 if (cur == NULL) 3431 cur = mem; 3432 } else { 3433 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 3434 nmm->set_memory_at(idx, mem); 3435 } 3436 } 3437 } 3438 mem = mem->in(MemNode::Memory); 3439 } 3440 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 3441 // Find any instance of the current type if we haven't encountered 3442 // already a memory slice of the instance along the memory chain. 3443 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3444 if((uint)_compile->get_general_index(ni) == i) { 3445 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 3446 if (nmm->is_empty_memory(m)) { 3447 Node* result = find_inst_mem(mem, ni, orig_phis); 3448 if (_compile->failing()) { 3449 return; 3450 } 3451 nmm->set_memory_at(ni, result); 3452 } 3453 } 3454 } 3455 } 3456 // Find the rest of instances values 3457 for (uint ni = new_index_start; ni < new_index_end; ni++) { 3458 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 3459 Node* result = step_through_mergemem(nmm, ni, tinst); 3460 if (result == nmm->base_memory()) { 3461 // Didn't find instance memory, search through general slice recursively. 3462 result = nmm->memory_at(_compile->get_general_index(ni)); 3463 result = find_inst_mem(result, ni, orig_phis); 3464 if (_compile->failing()) { 3465 return; 3466 } 3467 nmm->set_memory_at(ni, result); 3468 } 3469 } 3470 igvn->hash_insert(nmm); 3471 record_for_optimizer(nmm); 3472 } 3473 3474 // Phase 4: Update the inputs of non-instance memory Phis and 3475 // the Memory input of memnodes 3476 // First update the inputs of any non-instance Phi's from 3477 // which we split out an instance Phi. Note we don't have 3478 // to recursively process Phi's encounted on the input memory 3479 // chains as is done in split_memory_phi() since they will 3480 // also be processed here. 3481 for (int j = 0; j < orig_phis.length(); j++) { 3482 PhiNode *phi = orig_phis.at(j); 3483 int alias_idx = _compile->get_alias_index(phi->adr_type()); 3484 igvn->hash_delete(phi); 3485 for (uint i = 1; i < phi->req(); i++) { 3486 Node *mem = phi->in(i); 3487 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis); 3488 if (_compile->failing()) { 3489 return; 3490 } 3491 if (mem != new_mem) { 3492 phi->set_req(i, new_mem); 3493 } 3494 } 3495 igvn->hash_insert(phi); 3496 record_for_optimizer(phi); 3497 } 3498 3499 // Update the memory inputs of MemNodes with the value we computed 3500 // in Phase 2 and move stores memory users to corresponding memory slices. 3501 // Disable memory split verification code until the fix for 6984348. 3502 // Currently it produces false negative results since it does not cover all cases. 3503 #if 0 // ifdef ASSERT 3504 visited.Reset(); 3505 Node_Stack old_mems(arena, _compile->unique() >> 2); 3506 #endif 3507 for (uint i = 0; i < ideal_nodes.size(); i++) { 3508 Node* n = ideal_nodes.at(i); 3509 Node* nmem = get_map(n->_idx); 3510 assert(nmem != NULL, "sanity"); 3511 if (n->is_Mem()) { 3512 #if 0 // ifdef ASSERT 3513 Node* old_mem = n->in(MemNode::Memory); 3514 if (!visited.test_set(old_mem->_idx)) { 3515 old_mems.push(old_mem, old_mem->outcnt()); 3516 } 3517 #endif 3518 assert(n->in(MemNode::Memory) != nmem, "sanity"); 3519 if (!n->is_Load()) { 3520 // Move memory users of a store first. 3521 move_inst_mem(n, orig_phis); 3522 } 3523 // Now update memory input 3524 igvn->hash_delete(n); 3525 n->set_req(MemNode::Memory, nmem); 3526 igvn->hash_insert(n); 3527 record_for_optimizer(n); 3528 } else { 3529 assert(n->is_Allocate() || n->is_CheckCastPP() || 3530 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 3531 } 3532 } 3533 #if 0 // ifdef ASSERT 3534 // Verify that memory was split correctly 3535 while (old_mems.is_nonempty()) { 3536 Node* old_mem = old_mems.node(); 3537 uint old_cnt = old_mems.index(); 3538 old_mems.pop(); 3539 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 3540 } 3541 #endif 3542 } 3543 3544 #ifndef PRODUCT 3545 static const char *node_type_names[] = { 3546 "UnknownType", 3547 "JavaObject", 3548 "LocalVar", 3549 "Field", 3550 "Arraycopy" 3551 }; 3552 3553 static const char *esc_names[] = { 3554 "UnknownEscape", 3555 "NoEscape", 3556 "ArgEscape", 3557 "GlobalEscape" 3558 }; 3559 3560 void PointsToNode::dump(bool print_state) const { 3561 NodeType nt = node_type(); 3562 tty->print("%s ", node_type_names[(int) nt]); 3563 if (print_state) { 3564 EscapeState es = escape_state(); 3565 EscapeState fields_es = fields_escape_state(); 3566 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]); 3567 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) 3568 tty->print("NSR "); 3569 } 3570 if (is_Field()) { 3571 FieldNode* f = (FieldNode*)this; 3572 if (f->is_oop()) 3573 tty->print("oop "); 3574 if (f->offset() > 0) 3575 tty->print("+%d ", f->offset()); 3576 tty->print("("); 3577 for (BaseIterator i(f); i.has_next(); i.next()) { 3578 PointsToNode* b = i.get(); 3579 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : "")); 3580 } 3581 tty->print(" )"); 3582 } 3583 tty->print("["); 3584 for (EdgeIterator i(this); i.has_next(); i.next()) { 3585 PointsToNode* e = i.get(); 3586 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : ""); 3587 } 3588 tty->print(" ["); 3589 for (UseIterator i(this); i.has_next(); i.next()) { 3590 PointsToNode* u = i.get(); 3591 bool is_base = false; 3592 if (PointsToNode::is_base_use(u)) { 3593 is_base = true; 3594 u = PointsToNode::get_use_node(u)->as_Field(); 3595 } 3596 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : ""); 3597 } 3598 tty->print(" ]] "); 3599 if (_node == NULL) 3600 tty->print_cr("<null>"); 3601 else 3602 _node->dump(); 3603 } 3604 3605 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) { 3606 bool first = true; 3607 int ptnodes_length = ptnodes_worklist.length(); 3608 for (int i = 0; i < ptnodes_length; i++) { 3609 PointsToNode *ptn = ptnodes_worklist.at(i); 3610 if (ptn == NULL || !ptn->is_JavaObject()) 3611 continue; 3612 PointsToNode::EscapeState es = ptn->escape_state(); 3613 if ((es != PointsToNode::NoEscape) && !Verbose) { 3614 continue; 3615 } 3616 Node* n = ptn->ideal_node(); 3617 if (n->is_Allocate() || (n->is_CallStaticJava() && 3618 n->as_CallStaticJava()->is_boxing_method())) { 3619 if (first) { 3620 tty->cr(); 3621 tty->print("======== Connection graph for "); 3622 _compile->method()->print_short_name(); 3623 tty->cr(); 3624 first = false; 3625 } 3626 ptn->dump(); 3627 // Print all locals and fields which reference this allocation 3628 for (UseIterator j(ptn); j.has_next(); j.next()) { 3629 PointsToNode* use = j.get(); 3630 if (use->is_LocalVar()) { 3631 use->dump(Verbose); 3632 } else if (Verbose) { 3633 use->dump(); 3634 } 3635 } 3636 tty->cr(); 3637 } 3638 } 3639 } 3640 #endif