1 /* 2 * Copyright 2005-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_escape.cpp.incl" 27 28 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) { 29 uint v = (targIdx << EdgeShift) + ((uint) et); 30 if (_edges == NULL) { 31 Arena *a = Compile::current()->comp_arena(); 32 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0); 33 } 34 _edges->append_if_missing(v); 35 } 36 37 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) { 38 uint v = (targIdx << EdgeShift) + ((uint) et); 39 40 _edges->remove(v); 41 } 42 43 #ifndef PRODUCT 44 static const char *node_type_names[] = { 45 "UnknownType", 46 "JavaObject", 47 "LocalVar", 48 "Field" 49 }; 50 51 static const char *esc_names[] = { 52 "UnknownEscape", 53 "NoEscape", 54 "ArgEscape", 55 "GlobalEscape" 56 }; 57 58 static const char *edge_type_suffix[] = { 59 "?", // UnknownEdge 60 "P", // PointsToEdge 61 "D", // DeferredEdge 62 "F" // FieldEdge 63 }; 64 65 void PointsToNode::dump(bool print_state) const { 66 NodeType nt = node_type(); 67 tty->print("%s ", node_type_names[(int) nt]); 68 if (print_state) { 69 EscapeState es = escape_state(); 70 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR"); 71 } 72 tty->print("[["); 73 for (uint i = 0; i < edge_count(); i++) { 74 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]); 75 } 76 tty->print("]] "); 77 if (_node == NULL) 78 tty->print_cr("<null>"); 79 else 80 _node->dump(); 81 } 82 #endif 83 84 ConnectionGraph::ConnectionGraph(Compile * C) : 85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()), 86 _processed(C->comp_arena()), 87 _collecting(true), 88 _compile(C), 89 _node_map(C->comp_arena()) { 90 91 _phantom_object = C->top()->_idx, 92 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true); 93 94 // Add ConP(#NULL) and ConN(#NULL) nodes. 95 PhaseGVN* igvn = C->initial_gvn(); 96 Node* oop_null = igvn->zerocon(T_OBJECT); 97 _oop_null = oop_null->_idx; 98 assert(_oop_null < C->unique(), "should be created already"); 99 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 100 101 if (UseCompressedOops) { 102 Node* noop_null = igvn->zerocon(T_NARROWOOP); 103 _noop_null = noop_null->_idx; 104 assert(_noop_null < C->unique(), "should be created already"); 105 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 106 } 107 } 108 109 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) { 110 PointsToNode *f = ptnode_adr(from_i); 111 PointsToNode *t = ptnode_adr(to_i); 112 113 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 114 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge"); 115 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge"); 116 f->add_edge(to_i, PointsToNode::PointsToEdge); 117 } 118 119 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) { 120 PointsToNode *f = ptnode_adr(from_i); 121 PointsToNode *t = ptnode_adr(to_i); 122 123 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 124 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge"); 125 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge"); 126 // don't add a self-referential edge, this can occur during removal of 127 // deferred edges 128 if (from_i != to_i) 129 f->add_edge(to_i, PointsToNode::DeferredEdge); 130 } 131 132 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 133 const Type *adr_type = phase->type(adr); 134 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 135 adr->in(AddPNode::Address)->is_Proj() && 136 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 137 // We are computing a raw address for a store captured by an Initialize 138 // compute an appropriate address type. AddP cases #3 and #5 (see below). 139 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 140 assert(offs != Type::OffsetBot || 141 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 142 "offset must be a constant or it is initialization of array"); 143 return offs; 144 } 145 const TypePtr *t_ptr = adr_type->isa_ptr(); 146 assert(t_ptr != NULL, "must be a pointer type"); 147 return t_ptr->offset(); 148 } 149 150 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) { 151 PointsToNode *f = ptnode_adr(from_i); 152 PointsToNode *t = ptnode_adr(to_i); 153 154 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 155 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge"); 156 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge"); 157 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets"); 158 t->set_offset(offset); 159 160 f->add_edge(to_i, PointsToNode::FieldEdge); 161 } 162 163 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) { 164 PointsToNode *npt = ptnode_adr(ni); 165 PointsToNode::EscapeState old_es = npt->escape_state(); 166 if (es > old_es) 167 npt->set_escape_state(es); 168 } 169 170 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt, 171 PointsToNode::EscapeState es, bool done) { 172 PointsToNode* ptadr = ptnode_adr(n->_idx); 173 ptadr->_node = n; 174 ptadr->set_node_type(nt); 175 176 // inline set_escape_state(idx, es); 177 PointsToNode::EscapeState old_es = ptadr->escape_state(); 178 if (es > old_es) 179 ptadr->set_escape_state(es); 180 181 if (done) 182 _processed.set(n->_idx); 183 } 184 185 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n, PhaseTransform *phase) { 186 uint idx = n->_idx; 187 PointsToNode::EscapeState es; 188 189 // If we are still collecting or there were no non-escaping allocations 190 // we don't know the answer yet 191 if (_collecting) 192 return PointsToNode::UnknownEscape; 193 194 // if the node was created after the escape computation, return 195 // UnknownEscape 196 if (idx >= nodes_size()) 197 return PointsToNode::UnknownEscape; 198 199 es = ptnode_adr(idx)->escape_state(); 200 201 // if we have already computed a value, return it 202 if (es != PointsToNode::UnknownEscape && 203 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject) 204 return es; 205 206 // PointsTo() calls n->uncast() which can return a new ideal node. 207 if (n->uncast()->_idx >= nodes_size()) 208 return PointsToNode::UnknownEscape; 209 210 // compute max escape state of anything this node could point to 211 VectorSet ptset(Thread::current()->resource_area()); 212 PointsTo(ptset, n, phase); 213 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) { 214 uint pt = i.elem; 215 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); 216 if (pes > es) 217 es = pes; 218 } 219 // cache the computed escape state 220 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state"); 221 ptnode_adr(idx)->set_escape_state(es); 222 return es; 223 } 224 225 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n, PhaseTransform *phase) { 226 VectorSet visited(Thread::current()->resource_area()); 227 GrowableArray<uint> worklist; 228 229 #ifdef ASSERT 230 Node *orig_n = n; 231 #endif 232 233 n = n->uncast(); 234 PointsToNode* npt = ptnode_adr(n->_idx); 235 236 // If we have a JavaObject, return just that object 237 if (npt->node_type() == PointsToNode::JavaObject) { 238 ptset.set(n->_idx); 239 return; 240 } 241 #ifdef ASSERT 242 if (npt->_node == NULL) { 243 if (orig_n != n) 244 orig_n->dump(); 245 n->dump(); 246 assert(npt->_node != NULL, "unregistered node"); 247 } 248 #endif 249 worklist.push(n->_idx); 250 while(worklist.length() > 0) { 251 int ni = worklist.pop(); 252 if (visited.test_set(ni)) 253 continue; 254 255 PointsToNode* pn = ptnode_adr(ni); 256 // ensure that all inputs of a Phi have been processed 257 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); 258 259 int edges_processed = 0; 260 uint e_cnt = pn->edge_count(); 261 for (uint e = 0; e < e_cnt; e++) { 262 uint etgt = pn->edge_target(e); 263 PointsToNode::EdgeType et = pn->edge_type(e); 264 if (et == PointsToNode::PointsToEdge) { 265 ptset.set(etgt); 266 edges_processed++; 267 } else if (et == PointsToNode::DeferredEdge) { 268 worklist.push(etgt); 269 edges_processed++; 270 } else { 271 assert(false,"neither PointsToEdge or DeferredEdge"); 272 } 273 } 274 if (edges_processed == 0) { 275 // no deferred or pointsto edges found. Assume the value was set 276 // outside this method. Add the phantom object to the pointsto set. 277 ptset.set(_phantom_object); 278 } 279 } 280 } 281 282 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { 283 // This method is most expensive during ConnectionGraph construction. 284 // Reuse vectorSet and an additional growable array for deferred edges. 285 deferred_edges->clear(); 286 visited->Clear(); 287 288 visited->set(ni); 289 PointsToNode *ptn = ptnode_adr(ni); 290 291 // Mark current edges as visited and move deferred edges to separate array. 292 for (uint i = 0; i < ptn->edge_count(); ) { 293 uint t = ptn->edge_target(i); 294 #ifdef ASSERT 295 assert(!visited->test_set(t), "expecting no duplications"); 296 #else 297 visited->set(t); 298 #endif 299 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { 300 ptn->remove_edge(t, PointsToNode::DeferredEdge); 301 deferred_edges->append(t); 302 } else { 303 i++; 304 } 305 } 306 for (int next = 0; next < deferred_edges->length(); ++next) { 307 uint t = deferred_edges->at(next); 308 PointsToNode *ptt = ptnode_adr(t); 309 uint e_cnt = ptt->edge_count(); 310 for (uint e = 0; e < e_cnt; e++) { 311 uint etgt = ptt->edge_target(e); 312 if (visited->test_set(etgt)) 313 continue; 314 315 PointsToNode::EdgeType et = ptt->edge_type(e); 316 if (et == PointsToNode::PointsToEdge) { 317 add_pointsto_edge(ni, etgt); 318 if(etgt == _phantom_object) { 319 // Special case - field set outside (globally escaping). 320 ptn->set_escape_state(PointsToNode::GlobalEscape); 321 } 322 } else if (et == PointsToNode::DeferredEdge) { 323 deferred_edges->append(etgt); 324 } else { 325 assert(false,"invalid connection graph"); 326 } 327 } 328 } 329 } 330 331 332 // Add an edge to node given by "to_i" from any field of adr_i whose offset 333 // matches "offset" A deferred edge is added if to_i is a LocalVar, and 334 // a pointsto edge is added if it is a JavaObject 335 336 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { 337 PointsToNode* an = ptnode_adr(adr_i); 338 PointsToNode* to = ptnode_adr(to_i); 339 bool deferred = (to->node_type() == PointsToNode::LocalVar); 340 341 for (uint fe = 0; fe < an->edge_count(); fe++) { 342 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 343 int fi = an->edge_target(fe); 344 PointsToNode* pf = ptnode_adr(fi); 345 int po = pf->offset(); 346 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 347 if (deferred) 348 add_deferred_edge(fi, to_i); 349 else 350 add_pointsto_edge(fi, to_i); 351 } 352 } 353 } 354 355 // Add a deferred edge from node given by "from_i" to any field of adr_i 356 // whose offset matches "offset". 357 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { 358 PointsToNode* an = ptnode_adr(adr_i); 359 for (uint fe = 0; fe < an->edge_count(); fe++) { 360 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 361 int fi = an->edge_target(fe); 362 PointsToNode* pf = ptnode_adr(fi); 363 int po = pf->offset(); 364 if (pf->edge_count() == 0) { 365 // we have not seen any stores to this field, assume it was set outside this method 366 add_pointsto_edge(fi, _phantom_object); 367 } 368 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 369 add_deferred_edge(from_i, fi); 370 } 371 } 372 } 373 374 // Helper functions 375 376 static Node* get_addp_base(Node *addp) { 377 assert(addp->is_AddP(), "must be AddP"); 378 // 379 // AddP cases for Base and Address inputs: 380 // case #1. Direct object's field reference: 381 // Allocate 382 // | 383 // Proj #5 ( oop result ) 384 // | 385 // CheckCastPP (cast to instance type) 386 // | | 387 // AddP ( base == address ) 388 // 389 // case #2. Indirect object's field reference: 390 // Phi 391 // | 392 // CastPP (cast to instance type) 393 // | | 394 // AddP ( base == address ) 395 // 396 // case #3. Raw object's field reference for Initialize node: 397 // Allocate 398 // | 399 // Proj #5 ( oop result ) 400 // top | 401 // \ | 402 // AddP ( base == top ) 403 // 404 // case #4. Array's element reference: 405 // {CheckCastPP | CastPP} 406 // | | | 407 // | AddP ( array's element offset ) 408 // | | 409 // AddP ( array's offset ) 410 // 411 // case #5. Raw object's field reference for arraycopy stub call: 412 // The inline_native_clone() case when the arraycopy stub is called 413 // after the allocation before Initialize and CheckCastPP nodes. 414 // Allocate 415 // | 416 // Proj #5 ( oop result ) 417 // | | 418 // AddP ( base == address ) 419 // 420 // case #6. Constant Pool, ThreadLocal, CastX2P or 421 // Raw object's field reference: 422 // {ConP, ThreadLocal, CastX2P, raw Load} 423 // top | 424 // \ | 425 // AddP ( base == top ) 426 // 427 // case #7. Klass's field reference. 428 // LoadKlass 429 // | | 430 // AddP ( base == address ) 431 // 432 // case #8. narrow Klass's field reference. 433 // LoadNKlass 434 // | 435 // DecodeN 436 // | | 437 // AddP ( base == address ) 438 // 439 Node *base = addp->in(AddPNode::Base)->uncast(); 440 if (base->is_top()) { // The AddP case #3 and #6. 441 base = addp->in(AddPNode::Address)->uncast(); 442 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || 443 base->Opcode() == Op_CastX2P || base->is_DecodeN() || 444 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || 445 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); 446 } 447 return base; 448 } 449 450 static Node* find_second_addp(Node* addp, Node* n) { 451 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 452 453 Node* addp2 = addp->raw_out(0); 454 if (addp->outcnt() == 1 && addp2->is_AddP() && 455 addp2->in(AddPNode::Base) == n && 456 addp2->in(AddPNode::Address) == addp) { 457 458 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 459 // 460 // Find array's offset to push it on worklist first and 461 // as result process an array's element offset first (pushed second) 462 // to avoid CastPP for the array's offset. 463 // Otherwise the inserted CastPP (LocalVar) will point to what 464 // the AddP (Field) points to. Which would be wrong since 465 // the algorithm expects the CastPP has the same point as 466 // as AddP's base CheckCastPP (LocalVar). 467 // 468 // ArrayAllocation 469 // | 470 // CheckCastPP 471 // | 472 // memProj (from ArrayAllocation CheckCastPP) 473 // | || 474 // | || Int (element index) 475 // | || | ConI (log(element size)) 476 // | || | / 477 // | || LShift 478 // | || / 479 // | AddP (array's element offset) 480 // | | 481 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 482 // | / / 483 // AddP (array's offset) 484 // | 485 // Load/Store (memory operation on array's element) 486 // 487 return addp2; 488 } 489 return NULL; 490 } 491 492 // 493 // Adjust the type and inputs of an AddP which computes the 494 // address of a field of an instance 495 // 496 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { 497 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 498 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 499 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 500 if (t == NULL) { 501 // We are computing a raw address for a store captured by an Initialize 502 // compute an appropriate address type (cases #3 and #5). 503 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 504 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 505 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 506 assert(offs != Type::OffsetBot, "offset must be a constant"); 507 t = base_t->add_offset(offs)->is_oopptr(); 508 } 509 int inst_id = base_t->instance_id(); 510 assert(!t->is_known_instance() || t->instance_id() == inst_id, 511 "old type must be non-instance or match new type"); 512 513 // The type 't' could be subclass of 'base_t'. 514 // As result t->offset() could be large then base_t's size and it will 515 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 516 // constructor verifies correctness of the offset. 517 // 518 // It could happened on subclass's branch (from the type profiling 519 // inlining) which was not eliminated during parsing since the exactness 520 // of the allocation type was not propagated to the subclass type check. 521 // 522 // Do nothing for such AddP node and don't process its users since 523 // this code branch will go away. 524 // 525 if (!t->is_known_instance() && 526 !t->klass()->equals(base_t->klass()) && 527 t->klass()->is_subtype_of(base_t->klass())) { 528 return false; // bail out 529 } 530 531 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 532 // Do NOT remove the next call: ensure an new alias index is allocated 533 // for the instance type 534 int alias_idx = _compile->get_alias_index(tinst); 535 igvn->set_type(addp, tinst); 536 // record the allocation in the node map 537 set_map(addp->_idx, get_map(base->_idx)); 538 539 // Set addp's Base and Address to 'base'. 540 Node *abase = addp->in(AddPNode::Base); 541 Node *adr = addp->in(AddPNode::Address); 542 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 543 adr->in(0)->_idx == (uint)inst_id) { 544 // Skip AddP cases #3 and #5. 545 } else { 546 assert(!abase->is_top(), "sanity"); // AddP case #3 547 if (abase != base) { 548 igvn->hash_delete(addp); 549 addp->set_req(AddPNode::Base, base); 550 if (abase == adr) { 551 addp->set_req(AddPNode::Address, base); 552 } else { 553 // AddP case #4 (adr is array's element offset AddP node) 554 #ifdef ASSERT 555 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 556 assert(adr->is_AddP() && atype != NULL && 557 atype->instance_id() == inst_id, "array's element offset should be processed first"); 558 #endif 559 } 560 igvn->hash_insert(addp); 561 } 562 } 563 // Put on IGVN worklist since at least addp's type was changed above. 564 record_for_optimizer(addp); 565 return true; 566 } 567 568 // 569 // Create a new version of orig_phi if necessary. Returns either the newly 570 // created phi or an existing phi. Sets create_new to indicate wheter a new 571 // phi was created. Cache the last newly created phi in the node map. 572 // 573 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { 574 Compile *C = _compile; 575 new_created = false; 576 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 577 // nothing to do if orig_phi is bottom memory or matches alias_idx 578 if (phi_alias_idx == alias_idx) { 579 return orig_phi; 580 } 581 // have we already created a Phi for this alias index? 582 PhiNode *result = get_map_phi(orig_phi->_idx); 583 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 584 return result; 585 } 586 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { 587 if (C->do_escape_analysis() == true && !C->failing()) { 588 // Retry compilation without escape analysis. 589 // If this is the first failure, the sentinel string will "stick" 590 // to the Compile object, and the C2Compiler will see it and retry. 591 C->record_failure(C2Compiler::retry_no_escape_analysis()); 592 } 593 return NULL; 594 } 595 orig_phi_worklist.append_if_missing(orig_phi); 596 const TypePtr *atype = C->get_adr_type(alias_idx); 597 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 598 set_map_phi(orig_phi->_idx, result); 599 igvn->set_type(result, result->bottom_type()); 600 record_for_optimizer(result); 601 new_created = true; 602 return result; 603 } 604 605 // 606 // Return a new version of Memory Phi "orig_phi" with the inputs having the 607 // specified alias index. 608 // 609 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { 610 611 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 612 Compile *C = _compile; 613 bool new_phi_created; 614 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); 615 if (!new_phi_created) { 616 return result; 617 } 618 619 GrowableArray<PhiNode *> phi_list; 620 GrowableArray<uint> cur_input; 621 622 PhiNode *phi = orig_phi; 623 uint idx = 1; 624 bool finished = false; 625 while(!finished) { 626 while (idx < phi->req()) { 627 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); 628 if (mem != NULL && mem->is_Phi()) { 629 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); 630 if (new_phi_created) { 631 // found an phi for which we created a new split, push current one on worklist and begin 632 // processing new one 633 phi_list.push(phi); 634 cur_input.push(idx); 635 phi = mem->as_Phi(); 636 result = newphi; 637 idx = 1; 638 continue; 639 } else { 640 mem = newphi; 641 } 642 } 643 if (C->failing()) { 644 return NULL; 645 } 646 result->set_req(idx++, mem); 647 } 648 #ifdef ASSERT 649 // verify that the new Phi has an input for each input of the original 650 assert( phi->req() == result->req(), "must have same number of inputs."); 651 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 652 #endif 653 // Check if all new phi's inputs have specified alias index. 654 // Otherwise use old phi. 655 for (uint i = 1; i < phi->req(); i++) { 656 Node* in = result->in(i); 657 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 658 } 659 // we have finished processing a Phi, see if there are any more to do 660 finished = (phi_list.length() == 0 ); 661 if (!finished) { 662 phi = phi_list.pop(); 663 idx = cur_input.pop(); 664 PhiNode *prev_result = get_map_phi(phi->_idx); 665 prev_result->set_req(idx++, result); 666 result = prev_result; 667 } 668 } 669 return result; 670 } 671 672 673 // 674 // The next methods are derived from methods in MemNode. 675 // 676 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) { 677 Node *mem = mmem; 678 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 679 // means an array I have not precisely typed yet. Do not do any 680 // alias stuff with it any time soon. 681 if( tinst->base() != Type::AnyPtr && 682 !(tinst->klass()->is_java_lang_Object() && 683 tinst->offset() == Type::OffsetBot) ) { 684 mem = mmem->memory_at(alias_idx); 685 // Update input if it is progress over what we have now 686 } 687 return mem; 688 } 689 690 // 691 // Search memory chain of "mem" to find a MemNode whose address 692 // is the specified alias index. 693 // 694 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { 695 if (orig_mem == NULL) 696 return orig_mem; 697 Compile* C = phase->C; 698 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr(); 699 bool is_instance = (tinst != NULL) && tinst->is_known_instance(); 700 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 701 Node *prev = NULL; 702 Node *result = orig_mem; 703 while (prev != result) { 704 prev = result; 705 if (result == start_mem) 706 break; // hit one of our sentinels 707 if (result->is_Mem()) { 708 const Type *at = phase->type(result->in(MemNode::Address)); 709 if (at != Type::TOP) { 710 assert (at->isa_ptr() != NULL, "pointer type required."); 711 int idx = C->get_alias_index(at->is_ptr()); 712 if (idx == alias_idx) 713 break; 714 } 715 result = result->in(MemNode::Memory); 716 } 717 if (!is_instance) 718 continue; // don't search further for non-instance types 719 // skip over a call which does not affect this memory slice 720 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 721 Node *proj_in = result->in(0); 722 if (proj_in->is_Allocate() && proj_in->_idx == (uint)tinst->instance_id()) { 723 break; // hit one of our sentinels 724 } else if (proj_in->is_Call()) { 725 CallNode *call = proj_in->as_Call(); 726 if (!call->may_modify(tinst, phase)) { 727 result = call->in(TypeFunc::Memory); 728 } 729 } else if (proj_in->is_Initialize()) { 730 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 731 // Stop if this is the initialization for the object instance which 732 // which contains this memory slice, otherwise skip over it. 733 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) { 734 result = proj_in->in(TypeFunc::Memory); 735 } 736 } else if (proj_in->is_MemBar()) { 737 result = proj_in->in(TypeFunc::Memory); 738 } 739 } else if (result->is_MergeMem()) { 740 MergeMemNode *mmem = result->as_MergeMem(); 741 result = step_through_mergemem(mmem, alias_idx, tinst); 742 if (result == mmem->base_memory()) { 743 // Didn't find instance memory, search through general slice recursively. 744 result = mmem->memory_at(C->get_general_index(alias_idx)); 745 result = find_inst_mem(result, alias_idx, orig_phis, phase); 746 if (C->failing()) { 747 return NULL; 748 } 749 mmem->set_memory_at(alias_idx, result); 750 } 751 } else if (result->is_Phi() && 752 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 753 Node *un = result->as_Phi()->unique_input(phase); 754 if (un != NULL) { 755 result = un; 756 } else { 757 break; 758 } 759 } else if (result->Opcode() == Op_SCMemProj) { 760 assert(result->in(0)->is_LoadStore(), "sanity"); 761 const Type *at = phase->type(result->in(0)->in(MemNode::Address)); 762 if (at != Type::TOP) { 763 assert (at->isa_ptr() != NULL, "pointer type required."); 764 int idx = C->get_alias_index(at->is_ptr()); 765 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 766 break; 767 } 768 result = result->in(0)->in(MemNode::Memory); 769 } 770 } 771 if (result->is_Phi()) { 772 PhiNode *mphi = result->as_Phi(); 773 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 774 const TypePtr *t = mphi->adr_type(); 775 if (C->get_alias_index(t) != alias_idx) { 776 // Create a new Phi with the specified alias index type. 777 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); 778 } else if (!is_instance) { 779 // Push all non-instance Phis on the orig_phis worklist to update inputs 780 // during Phase 4 if needed. 781 orig_phis.append_if_missing(mphi); 782 } 783 } 784 // the result is either MemNode, PhiNode, InitializeNode. 785 return result; 786 } 787 788 789 // 790 // Convert the types of unescaped object to instance types where possible, 791 // propagate the new type information through the graph, and update memory 792 // edges and MergeMem inputs to reflect the new type. 793 // 794 // We start with allocations (and calls which may be allocations) on alloc_worklist. 795 // The processing is done in 4 phases: 796 // 797 // Phase 1: Process possible allocations from alloc_worklist. Create instance 798 // types for the CheckCastPP for allocations where possible. 799 // Propagate the the new types through users as follows: 800 // casts and Phi: push users on alloc_worklist 801 // AddP: cast Base and Address inputs to the instance type 802 // push any AddP users on alloc_worklist and push any memnode 803 // users onto memnode_worklist. 804 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 805 // search the Memory chain for a store with the appropriate type 806 // address type. If a Phi is found, create a new version with 807 // the appropriate memory slices from each of the Phi inputs. 808 // For stores, process the users as follows: 809 // MemNode: push on memnode_worklist 810 // MergeMem: push on mergemem_worklist 811 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 812 // moving the first node encountered of each instance type to the 813 // the input corresponding to its alias index. 814 // appropriate memory slice. 815 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 816 // 817 // In the following example, the CheckCastPP nodes are the cast of allocation 818 // results and the allocation of node 29 is unescaped and eligible to be an 819 // instance type. 820 // 821 // We start with: 822 // 823 // 7 Parm #memory 824 // 10 ConI "12" 825 // 19 CheckCastPP "Foo" 826 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 827 // 29 CheckCastPP "Foo" 828 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 829 // 830 // 40 StoreP 25 7 20 ... alias_index=4 831 // 50 StoreP 35 40 30 ... alias_index=4 832 // 60 StoreP 45 50 20 ... alias_index=4 833 // 70 LoadP _ 60 30 ... alias_index=4 834 // 80 Phi 75 50 60 Memory alias_index=4 835 // 90 LoadP _ 80 30 ... alias_index=4 836 // 100 LoadP _ 80 20 ... alias_index=4 837 // 838 // 839 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 840 // and creating a new alias index for node 30. This gives: 841 // 842 // 7 Parm #memory 843 // 10 ConI "12" 844 // 19 CheckCastPP "Foo" 845 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 846 // 29 CheckCastPP "Foo" iid=24 847 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 848 // 849 // 40 StoreP 25 7 20 ... alias_index=4 850 // 50 StoreP 35 40 30 ... alias_index=6 851 // 60 StoreP 45 50 20 ... alias_index=4 852 // 70 LoadP _ 60 30 ... alias_index=6 853 // 80 Phi 75 50 60 Memory alias_index=4 854 // 90 LoadP _ 80 30 ... alias_index=6 855 // 100 LoadP _ 80 20 ... alias_index=4 856 // 857 // In phase 2, new memory inputs are computed for the loads and stores, 858 // And a new version of the phi is created. In phase 4, the inputs to 859 // node 80 are updated and then the memory nodes are updated with the 860 // values computed in phase 2. This results in: 861 // 862 // 7 Parm #memory 863 // 10 ConI "12" 864 // 19 CheckCastPP "Foo" 865 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 866 // 29 CheckCastPP "Foo" iid=24 867 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 868 // 869 // 40 StoreP 25 7 20 ... alias_index=4 870 // 50 StoreP 35 7 30 ... alias_index=6 871 // 60 StoreP 45 40 20 ... alias_index=4 872 // 70 LoadP _ 50 30 ... alias_index=6 873 // 80 Phi 75 40 60 Memory alias_index=4 874 // 120 Phi 75 50 50 Memory alias_index=6 875 // 90 LoadP _ 120 30 ... alias_index=6 876 // 100 LoadP _ 80 20 ... alias_index=4 877 // 878 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 879 GrowableArray<Node *> memnode_worklist; 880 GrowableArray<Node *> mergemem_worklist; 881 GrowableArray<PhiNode *> orig_phis; 882 PhaseGVN *igvn = _compile->initial_gvn(); 883 uint new_index_start = (uint) _compile->num_alias_types(); 884 VectorSet visited(Thread::current()->resource_area()); 885 VectorSet ptset(Thread::current()->resource_area()); 886 887 888 // Phase 1: Process possible allocations from alloc_worklist. 889 // Create instance types for the CheckCastPP for allocations where possible. 890 // 891 // (Note: don't forget to change the order of the second AddP node on 892 // the alloc_worklist if the order of the worklist processing is changed, 893 // see the comment in find_second_addp().) 894 // 895 while (alloc_worklist.length() != 0) { 896 Node *n = alloc_worklist.pop(); 897 uint ni = n->_idx; 898 const TypeOopPtr* tinst = NULL; 899 if (n->is_Call()) { 900 CallNode *alloc = n->as_Call(); 901 // copy escape information to call node 902 PointsToNode* ptn = ptnode_adr(alloc->_idx); 903 PointsToNode::EscapeState es = escape_state(alloc, igvn); 904 // We have an allocation or call which returns a Java object, 905 // see if it is unescaped. 906 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) 907 continue; 908 if (alloc->is_Allocate()) { 909 // Set the scalar_replaceable flag before the next check. 910 alloc->as_Allocate()->_is_scalar_replaceable = true; 911 } 912 // find CheckCastPP of call return value 913 n = alloc->result_cast(); 914 if (n == NULL || // No uses accept Initialize or 915 !n->is_CheckCastPP()) // not unique CheckCastPP. 916 continue; 917 // The inline code for Object.clone() casts the allocation result to 918 // java.lang.Object and then to the actual type of the allocated 919 // object. Detect this case and use the second cast. 920 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 921 // the allocation result is cast to java.lang.Object and then 922 // to the actual Array type. 923 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 924 && (alloc->is_AllocateArray() || 925 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 926 Node *cast2 = NULL; 927 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 928 Node *use = n->fast_out(i); 929 if (use->is_CheckCastPP()) { 930 cast2 = use; 931 break; 932 } 933 } 934 if (cast2 != NULL) { 935 n = cast2; 936 } else { 937 continue; 938 } 939 } 940 set_escape_state(n->_idx, es); 941 // in order for an object to be scalar-replaceable, it must be: 942 // - a direct allocation (not a call returning an object) 943 // - non-escaping 944 // - eligible to be a unique type 945 // - not determined to be ineligible by escape analysis 946 set_map(alloc->_idx, n); 947 set_map(n->_idx, alloc); 948 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 949 if (t == NULL) 950 continue; // not a TypeInstPtr 951 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 952 igvn->hash_delete(n); 953 igvn->set_type(n, tinst); 954 n->raise_bottom_type(tinst); 955 igvn->hash_insert(n); 956 record_for_optimizer(n); 957 if (alloc->is_Allocate() && ptn->_scalar_replaceable && 958 (t->isa_instptr() || t->isa_aryptr())) { 959 960 // First, put on the worklist all Field edges from Connection Graph 961 // which is more accurate then putting immediate users from Ideal Graph. 962 for (uint e = 0; e < ptn->edge_count(); e++) { 963 Node *use = ptnode_adr(ptn->edge_target(e))->_node; 964 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), 965 "only AddP nodes are Field edges in CG"); 966 if (use->outcnt() > 0) { // Don't process dead nodes 967 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 968 if (addp2 != NULL) { 969 assert(alloc->is_AllocateArray(),"array allocation was expected"); 970 alloc_worklist.append_if_missing(addp2); 971 } 972 alloc_worklist.append_if_missing(use); 973 } 974 } 975 976 // An allocation may have an Initialize which has raw stores. Scan 977 // the users of the raw allocation result and push AddP users 978 // on alloc_worklist. 979 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 980 assert (raw_result != NULL, "must have an allocation result"); 981 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 982 Node *use = raw_result->fast_out(i); 983 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 984 Node* addp2 = find_second_addp(use, raw_result); 985 if (addp2 != NULL) { 986 assert(alloc->is_AllocateArray(),"array allocation was expected"); 987 alloc_worklist.append_if_missing(addp2); 988 } 989 alloc_worklist.append_if_missing(use); 990 } else if (use->is_Initialize()) { 991 memnode_worklist.append_if_missing(use); 992 } 993 } 994 } 995 } else if (n->is_AddP()) { 996 ptset.Clear(); 997 PointsTo(ptset, get_addp_base(n), igvn); 998 assert(ptset.Size() == 1, "AddP address is unique"); 999 uint elem = ptset.getelem(); // Allocation node's index 1000 if (elem == _phantom_object) 1001 continue; // Assume the value was set outside this method. 1002 Node *base = get_map(elem); // CheckCastPP node 1003 if (!split_AddP(n, base, igvn)) continue; // wrong type 1004 tinst = igvn->type(base)->isa_oopptr(); 1005 } else if (n->is_Phi() || 1006 n->is_CheckCastPP() || 1007 n->is_EncodeP() || 1008 n->is_DecodeN() || 1009 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 1010 if (visited.test_set(n->_idx)) { 1011 assert(n->is_Phi(), "loops only through Phi's"); 1012 continue; // already processed 1013 } 1014 ptset.Clear(); 1015 PointsTo(ptset, n, igvn); 1016 if (ptset.Size() == 1) { 1017 uint elem = ptset.getelem(); // Allocation node's index 1018 if (elem == _phantom_object) 1019 continue; // Assume the value was set outside this method. 1020 Node *val = get_map(elem); // CheckCastPP node 1021 TypeNode *tn = n->as_Type(); 1022 tinst = igvn->type(val)->isa_oopptr(); 1023 assert(tinst != NULL && tinst->is_known_instance() && 1024 (uint)tinst->instance_id() == elem , "instance type expected."); 1025 1026 const Type *tn_type = igvn->type(tn); 1027 const TypeOopPtr *tn_t; 1028 if (tn_type->isa_narrowoop()) { 1029 tn_t = tn_type->make_ptr()->isa_oopptr(); 1030 } else { 1031 tn_t = tn_type->isa_oopptr(); 1032 } 1033 1034 if (tn_t != NULL && 1035 tinst->cast_to_instance_id(TypeOopPtr::InstanceBot)->higher_equal(tn_t)) { 1036 if (tn_type->isa_narrowoop()) { 1037 tn_type = tinst->make_narrowoop(); 1038 } else { 1039 tn_type = tinst; 1040 } 1041 igvn->hash_delete(tn); 1042 igvn->set_type(tn, tn_type); 1043 tn->set_type(tn_type); 1044 igvn->hash_insert(tn); 1045 record_for_optimizer(n); 1046 } else { 1047 continue; // wrong type 1048 } 1049 } 1050 } else { 1051 continue; 1052 } 1053 // push users on appropriate worklist 1054 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1055 Node *use = n->fast_out(i); 1056 if(use->is_Mem() && use->in(MemNode::Address) == n) { 1057 memnode_worklist.append_if_missing(use); 1058 } else if (use->is_Initialize()) { 1059 memnode_worklist.append_if_missing(use); 1060 } else if (use->is_MergeMem()) { 1061 mergemem_worklist.append_if_missing(use); 1062 } else if (use->is_SafePoint() && tinst != NULL) { 1063 // Look for MergeMem nodes for calls which reference unique allocation 1064 // (through CheckCastPP nodes) even for debug info. 1065 Node* m = use->in(TypeFunc::Memory); 1066 uint iid = tinst->instance_id(); 1067 while (m->is_Proj() && m->in(0)->is_SafePoint() && 1068 m->in(0) != use && !m->in(0)->_idx != iid) { 1069 m = m->in(0)->in(TypeFunc::Memory); 1070 } 1071 if (m->is_MergeMem()) { 1072 mergemem_worklist.append_if_missing(m); 1073 } 1074 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 1075 Node* addp2 = find_second_addp(use, n); 1076 if (addp2 != NULL) { 1077 alloc_worklist.append_if_missing(addp2); 1078 } 1079 alloc_worklist.append_if_missing(use); 1080 } else if (use->is_Phi() || 1081 use->is_CheckCastPP() || 1082 use->is_EncodeP() || 1083 use->is_DecodeN() || 1084 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 1085 alloc_worklist.append_if_missing(use); 1086 } 1087 } 1088 1089 } 1090 // New alias types were created in split_AddP(). 1091 uint new_index_end = (uint) _compile->num_alias_types(); 1092 1093 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 1094 // compute new values for Memory inputs (the Memory inputs are not 1095 // actually updated until phase 4.) 1096 if (memnode_worklist.length() == 0) 1097 return; // nothing to do 1098 1099 while (memnode_worklist.length() != 0) { 1100 Node *n = memnode_worklist.pop(); 1101 if (visited.test_set(n->_idx)) 1102 continue; 1103 if (n->is_Phi()) { 1104 assert(n->as_Phi()->adr_type() != TypePtr::BOTTOM, "narrow memory slice required"); 1105 // we don't need to do anything, but the users must be pushed if we haven't processed 1106 // this Phi before 1107 } else if (n->is_Initialize()) { 1108 // we don't need to do anything, but the users of the memory projection must be pushed 1109 n = n->as_Initialize()->proj_out(TypeFunc::Memory); 1110 if (n == NULL) 1111 continue; 1112 } else { 1113 assert(n->is_Mem(), "memory node required."); 1114 Node *addr = n->in(MemNode::Address); 1115 assert(addr->is_AddP(), "AddP required"); 1116 const Type *addr_t = igvn->type(addr); 1117 if (addr_t == Type::TOP) 1118 continue; 1119 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 1120 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 1121 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 1122 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); 1123 if (_compile->failing()) { 1124 return; 1125 } 1126 if (mem != n->in(MemNode::Memory)) { 1127 set_map(n->_idx, mem); 1128 ptnode_adr(n->_idx)->_node = n; 1129 } 1130 if (n->is_Load()) { 1131 continue; // don't push users 1132 } else if (n->is_LoadStore()) { 1133 // get the memory projection 1134 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1135 Node *use = n->fast_out(i); 1136 if (use->Opcode() == Op_SCMemProj) { 1137 n = use; 1138 break; 1139 } 1140 } 1141 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 1142 } 1143 } 1144 // push user on appropriate worklist 1145 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1146 Node *use = n->fast_out(i); 1147 if (use->is_Phi()) { 1148 memnode_worklist.append_if_missing(use); 1149 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { 1150 memnode_worklist.append_if_missing(use); 1151 } else if (use->is_Initialize()) { 1152 memnode_worklist.append_if_missing(use); 1153 } else if (use->is_MergeMem()) { 1154 mergemem_worklist.append_if_missing(use); 1155 } 1156 } 1157 } 1158 1159 // Phase 3: Process MergeMem nodes from mergemem_worklist. 1160 // Walk each memory moving the first node encountered of each 1161 // instance type to the the input corresponding to its alias index. 1162 while (mergemem_worklist.length() != 0) { 1163 Node *n = mergemem_worklist.pop(); 1164 assert(n->is_MergeMem(), "MergeMem node required."); 1165 if (visited.test_set(n->_idx)) 1166 continue; 1167 MergeMemNode *nmm = n->as_MergeMem(); 1168 // Note: we don't want to use MergeMemStream here because we only want to 1169 // scan inputs which exist at the start, not ones we add during processing. 1170 uint nslices = nmm->req(); 1171 igvn->hash_delete(nmm); 1172 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 1173 Node* mem = nmm->in(i); 1174 Node* cur = NULL; 1175 if (mem == NULL || mem->is_top()) 1176 continue; 1177 while (mem->is_Mem()) { 1178 const Type *at = igvn->type(mem->in(MemNode::Address)); 1179 if (at != Type::TOP) { 1180 assert (at->isa_ptr() != NULL, "pointer type required."); 1181 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 1182 if (idx == i) { 1183 if (cur == NULL) 1184 cur = mem; 1185 } else { 1186 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 1187 nmm->set_memory_at(idx, mem); 1188 } 1189 } 1190 } 1191 mem = mem->in(MemNode::Memory); 1192 } 1193 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 1194 // Find any instance of the current type if we haven't encountered 1195 // a value of the instance along the chain. 1196 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1197 if((uint)_compile->get_general_index(ni) == i) { 1198 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 1199 if (nmm->is_empty_memory(m)) { 1200 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); 1201 if (_compile->failing()) { 1202 return; 1203 } 1204 nmm->set_memory_at(ni, result); 1205 } 1206 } 1207 } 1208 } 1209 // Find the rest of instances values 1210 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1211 const TypeOopPtr *tinst = igvn->C->get_adr_type(ni)->isa_oopptr(); 1212 Node* result = step_through_mergemem(nmm, ni, tinst); 1213 if (result == nmm->base_memory()) { 1214 // Didn't find instance memory, search through general slice recursively. 1215 result = nmm->memory_at(igvn->C->get_general_index(ni)); 1216 result = find_inst_mem(result, ni, orig_phis, igvn); 1217 if (_compile->failing()) { 1218 return; 1219 } 1220 nmm->set_memory_at(ni, result); 1221 } 1222 } 1223 igvn->hash_insert(nmm); 1224 record_for_optimizer(nmm); 1225 1226 // Propagate new memory slices to following MergeMem nodes. 1227 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1228 Node *use = n->fast_out(i); 1229 if (use->is_Call()) { 1230 CallNode* in = use->as_Call(); 1231 if (in->proj_out(TypeFunc::Memory) != NULL) { 1232 Node* m = in->proj_out(TypeFunc::Memory); 1233 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 1234 Node* mm = m->fast_out(j); 1235 if (mm->is_MergeMem()) { 1236 mergemem_worklist.append_if_missing(mm); 1237 } 1238 } 1239 } 1240 if (use->is_Allocate()) { 1241 use = use->as_Allocate()->initialization(); 1242 if (use == NULL) { 1243 continue; 1244 } 1245 } 1246 } 1247 if (use->is_Initialize()) { 1248 InitializeNode* in = use->as_Initialize(); 1249 if (in->proj_out(TypeFunc::Memory) != NULL) { 1250 Node* m = in->proj_out(TypeFunc::Memory); 1251 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 1252 Node* mm = m->fast_out(j); 1253 if (mm->is_MergeMem()) { 1254 mergemem_worklist.append_if_missing(mm); 1255 } 1256 } 1257 } 1258 } 1259 } 1260 } 1261 1262 // Phase 4: Update the inputs of non-instance memory Phis and 1263 // the Memory input of memnodes 1264 // First update the inputs of any non-instance Phi's from 1265 // which we split out an instance Phi. Note we don't have 1266 // to recursively process Phi's encounted on the input memory 1267 // chains as is done in split_memory_phi() since they will 1268 // also be processed here. 1269 for (int j = 0; j < orig_phis.length(); j++) { 1270 PhiNode *phi = orig_phis.at(j); 1271 int alias_idx = _compile->get_alias_index(phi->adr_type()); 1272 igvn->hash_delete(phi); 1273 for (uint i = 1; i < phi->req(); i++) { 1274 Node *mem = phi->in(i); 1275 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); 1276 if (_compile->failing()) { 1277 return; 1278 } 1279 if (mem != new_mem) { 1280 phi->set_req(i, new_mem); 1281 } 1282 } 1283 igvn->hash_insert(phi); 1284 record_for_optimizer(phi); 1285 } 1286 1287 // Update the memory inputs of MemNodes with the value we computed 1288 // in Phase 2. 1289 for (uint i = 0; i < nodes_size(); i++) { 1290 Node *nmem = get_map(i); 1291 if (nmem != NULL) { 1292 Node *n = ptnode_adr(i)->_node; 1293 if (n != NULL && n->is_Mem()) { 1294 igvn->hash_delete(n); 1295 n->set_req(MemNode::Memory, nmem); 1296 igvn->hash_insert(n); 1297 record_for_optimizer(n); 1298 } 1299 } 1300 } 1301 } 1302 1303 bool ConnectionGraph::has_candidates(Compile *C) { 1304 // EA brings benefits only when the code has allocations and/or locks which 1305 // are represented by ideal Macro nodes. 1306 int cnt = C->macro_count(); 1307 for( int i=0; i < cnt; i++ ) { 1308 Node *n = C->macro_node(i); 1309 if ( n->is_Allocate() ) 1310 return true; 1311 if( n->is_Lock() ) { 1312 Node* obj = n->as_Lock()->obj_node()->uncast(); 1313 if( !(obj->is_Parm() || obj->is_Con()) ) 1314 return true; 1315 } 1316 } 1317 return false; 1318 } 1319 1320 bool ConnectionGraph::compute_escape() { 1321 Compile* C = _compile; 1322 1323 // 1. Populate Connection Graph (CG) with Ideal nodes. 1324 1325 Unique_Node_List worklist_init; 1326 worklist_init.map(C->unique(), NULL); // preallocate space 1327 1328 // Initialize worklist 1329 if (C->root() != NULL) { 1330 worklist_init.push(C->root()); 1331 } 1332 1333 GrowableArray<int> cg_worklist; 1334 PhaseGVN* igvn = C->initial_gvn(); 1335 bool has_allocations = false; 1336 1337 // Push all useful nodes onto CG list and set their type. 1338 for( uint next = 0; next < worklist_init.size(); ++next ) { 1339 Node* n = worklist_init.at(next); 1340 record_for_escape_analysis(n, igvn); 1341 // Only allocations and java static calls results are checked 1342 // for an escape status. See process_call_result() below. 1343 if (n->is_Allocate() || n->is_CallStaticJava() && 1344 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { 1345 has_allocations = true; 1346 } 1347 if(n->is_AddP()) 1348 cg_worklist.append(n->_idx); 1349 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1350 Node* m = n->fast_out(i); // Get user 1351 worklist_init.push(m); 1352 } 1353 } 1354 1355 if (!has_allocations) { 1356 _collecting = false; 1357 return false; // Nothing to do. 1358 } 1359 1360 // 2. First pass to create simple CG edges (doesn't require to walk CG). 1361 uint delayed_size = _delayed_worklist.size(); 1362 for( uint next = 0; next < delayed_size; ++next ) { 1363 Node* n = _delayed_worklist.at(next); 1364 build_connection_graph(n, igvn); 1365 } 1366 1367 // 3. Pass to create fields edges (Allocate -F-> AddP). 1368 uint cg_length = cg_worklist.length(); 1369 for( uint next = 0; next < cg_length; ++next ) { 1370 int ni = cg_worklist.at(next); 1371 build_connection_graph(ptnode_adr(ni)->_node, igvn); 1372 } 1373 1374 cg_worklist.clear(); 1375 cg_worklist.append(_phantom_object); 1376 1377 // 4. Build Connection Graph which need 1378 // to walk the connection graph. 1379 for (uint ni = 0; ni < nodes_size(); ni++) { 1380 PointsToNode* ptn = ptnode_adr(ni); 1381 Node *n = ptn->_node; 1382 if (n != NULL) { // Call, AddP, LoadP, StoreP 1383 build_connection_graph(n, igvn); 1384 if (ptn->node_type() != PointsToNode::UnknownType) 1385 cg_worklist.append(n->_idx); // Collect CG nodes 1386 } 1387 } 1388 1389 VectorSet ptset(Thread::current()->resource_area()); 1390 GrowableArray<uint> deferred_edges; 1391 VectorSet visited(Thread::current()->resource_area()); 1392 1393 // 5. Remove deferred edges from the graph and collect 1394 // information needed for type splitting. 1395 cg_length = cg_worklist.length(); 1396 for( uint next = 0; next < cg_length; ++next ) { 1397 int ni = cg_worklist.at(next); 1398 PointsToNode* ptn = ptnode_adr(ni); 1399 PointsToNode::NodeType nt = ptn->node_type(); 1400 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { 1401 remove_deferred(ni, &deferred_edges, &visited); 1402 Node *n = ptn->_node; 1403 if (n->is_AddP()) { 1404 // Search for objects which are not scalar replaceable. 1405 // Mark their escape state as ArgEscape to propagate the state 1406 // to referenced objects. 1407 // Note: currently there are no difference in compiler optimizations 1408 // for ArgEscape objects and NoEscape objects which are not 1409 // scalar replaceable. 1410 1411 int offset = ptn->offset(); 1412 Node *base = get_addp_base(n); 1413 ptset.Clear(); 1414 PointsTo(ptset, base, igvn); 1415 int ptset_size = ptset.Size(); 1416 1417 // Check if a field's initializing value is recorded and add 1418 // a corresponding NULL field's value if it is not recorded. 1419 // Connection Graph does not record a default initialization by NULL 1420 // captured by Initialize node. 1421 // 1422 // Note: it will disable scalar replacement in some cases: 1423 // 1424 // Point p[] = new Point[1]; 1425 // p[0] = new Point(); // Will be not scalar replaced 1426 // 1427 // but it will save us from incorrect optimizations in next cases: 1428 // 1429 // Point p[] = new Point[1]; 1430 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1431 // 1432 // Without a control flow analysis we can't distinguish above cases. 1433 // 1434 if (offset != Type::OffsetBot && ptset_size == 1) { 1435 uint elem = ptset.getelem(); // Allocation node's index 1436 // It does not matter if it is not Allocation node since 1437 // only non-escaping allocations are scalar replaced. 1438 if (ptnode_adr(elem)->_node->is_Allocate() && 1439 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) { 1440 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate(); 1441 InitializeNode* ini = alloc->initialization(); 1442 Node* value = NULL; 1443 if (ini != NULL) { 1444 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT; 1445 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), igvn); 1446 if (store != NULL && store->is_Store()) 1447 value = store->in(MemNode::ValueIn); 1448 } 1449 if (value == NULL || value != ptnode_adr(value->_idx)->_node) { 1450 // A field's initializing value was not recorded. Add NULL. 1451 uint null_idx = UseCompressedOops ? _noop_null : _oop_null; 1452 add_pointsto_edge(ni, null_idx); 1453 } 1454 } 1455 } 1456 1457 // An object is not scalar replaceable if the field which may point 1458 // to it has unknown offset (unknown element of an array of objects). 1459 // 1460 if (offset == Type::OffsetBot) { 1461 uint e_cnt = ptn->edge_count(); 1462 for (uint ei = 0; ei < e_cnt; ei++) { 1463 uint npi = ptn->edge_target(ei); 1464 set_escape_state(npi, PointsToNode::ArgEscape); 1465 ptnode_adr(npi)->_scalar_replaceable = false; 1466 } 1467 } 1468 1469 // Currently an object is not scalar replaceable if a LoadStore node 1470 // access its field since the field value is unknown after it. 1471 // 1472 bool has_LoadStore = false; 1473 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1474 Node *use = n->fast_out(i); 1475 if (use->is_LoadStore()) { 1476 has_LoadStore = true; 1477 break; 1478 } 1479 } 1480 // An object is not scalar replaceable if the address points 1481 // to unknown field (unknown element for arrays, offset is OffsetBot). 1482 // 1483 // Or the address may point to more then one object. This may produce 1484 // the false positive result (set scalar_replaceable to false) 1485 // since the flow-insensitive escape analysis can't separate 1486 // the case when stores overwrite the field's value from the case 1487 // when stores happened on different control branches. 1488 // 1489 if (ptset_size > 1 || ptset_size != 0 && 1490 (has_LoadStore || offset == Type::OffsetBot)) { 1491 for( VectorSetI j(&ptset); j.test(); ++j ) { 1492 set_escape_state(j.elem, PointsToNode::ArgEscape); 1493 ptnode_adr(j.elem)->_scalar_replaceable = false; 1494 } 1495 } 1496 } 1497 } 1498 } 1499 1500 // 6. Propagate escape states. 1501 GrowableArray<int> worklist; 1502 bool has_non_escaping_obj = false; 1503 1504 // push all GlobalEscape nodes on the worklist 1505 for( uint next = 0; next < cg_length; ++next ) { 1506 int nk = cg_worklist.at(next); 1507 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) 1508 worklist.push(nk); 1509 } 1510 // mark all nodes reachable from GlobalEscape nodes 1511 while(worklist.length() > 0) { 1512 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1513 uint e_cnt = ptn->edge_count(); 1514 for (uint ei = 0; ei < e_cnt; ei++) { 1515 uint npi = ptn->edge_target(ei); 1516 PointsToNode *np = ptnode_adr(npi); 1517 if (np->escape_state() < PointsToNode::GlobalEscape) { 1518 np->set_escape_state(PointsToNode::GlobalEscape); 1519 worklist.push(npi); 1520 } 1521 } 1522 } 1523 1524 // push all ArgEscape nodes on the worklist 1525 for( uint next = 0; next < cg_length; ++next ) { 1526 int nk = cg_worklist.at(next); 1527 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) 1528 worklist.push(nk); 1529 } 1530 // mark all nodes reachable from ArgEscape nodes 1531 while(worklist.length() > 0) { 1532 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1533 if (ptn->node_type() == PointsToNode::JavaObject) 1534 has_non_escaping_obj = true; // Non GlobalEscape 1535 uint e_cnt = ptn->edge_count(); 1536 for (uint ei = 0; ei < e_cnt; ei++) { 1537 uint npi = ptn->edge_target(ei); 1538 PointsToNode *np = ptnode_adr(npi); 1539 if (np->escape_state() < PointsToNode::ArgEscape) { 1540 np->set_escape_state(PointsToNode::ArgEscape); 1541 worklist.push(npi); 1542 } 1543 } 1544 } 1545 1546 GrowableArray<Node*> alloc_worklist; 1547 1548 // push all NoEscape nodes on the worklist 1549 for( uint next = 0; next < cg_length; ++next ) { 1550 int nk = cg_worklist.at(next); 1551 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) 1552 worklist.push(nk); 1553 } 1554 // mark all nodes reachable from NoEscape nodes 1555 while(worklist.length() > 0) { 1556 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1557 if (ptn->node_type() == PointsToNode::JavaObject) 1558 has_non_escaping_obj = true; // Non GlobalEscape 1559 Node* n = ptn->_node; 1560 if (n->is_Allocate() && ptn->_scalar_replaceable ) { 1561 // Push scalar replaceable allocations on alloc_worklist 1562 // for processing in split_unique_types(). 1563 alloc_worklist.append(n); 1564 } 1565 uint e_cnt = ptn->edge_count(); 1566 for (uint ei = 0; ei < e_cnt; ei++) { 1567 uint npi = ptn->edge_target(ei); 1568 PointsToNode *np = ptnode_adr(npi); 1569 if (np->escape_state() < PointsToNode::NoEscape) { 1570 np->set_escape_state(PointsToNode::NoEscape); 1571 worklist.push(npi); 1572 } 1573 } 1574 } 1575 1576 _collecting = false; 1577 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); 1578 1579 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; 1580 if ( has_scalar_replaceable_candidates && 1581 C->AliasLevel() >= 3 && EliminateAllocations ) { 1582 1583 // Now use the escape information to create unique types for 1584 // scalar replaceable objects. 1585 split_unique_types(alloc_worklist); 1586 1587 if (C->failing()) return false; 1588 1589 // Clean up after split unique types. 1590 ResourceMark rm; 1591 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn()); 1592 1593 C->print_method("After Escape Analysis", 2); 1594 1595 #ifdef ASSERT 1596 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 1597 tty->print("=== No allocations eliminated for "); 1598 C->method()->print_short_name(); 1599 if(!EliminateAllocations) { 1600 tty->print(" since EliminateAllocations is off ==="); 1601 } else if(!has_scalar_replaceable_candidates) { 1602 tty->print(" since there are no scalar replaceable candidates ==="); 1603 } else if(C->AliasLevel() < 3) { 1604 tty->print(" since AliasLevel < 3 ==="); 1605 } 1606 tty->cr(); 1607 #endif 1608 } 1609 return has_non_escaping_obj; 1610 } 1611 1612 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { 1613 1614 switch (call->Opcode()) { 1615 #ifdef ASSERT 1616 case Op_Allocate: 1617 case Op_AllocateArray: 1618 case Op_Lock: 1619 case Op_Unlock: 1620 assert(false, "should be done already"); 1621 break; 1622 #endif 1623 case Op_CallLeafNoFP: 1624 { 1625 // Stub calls, objects do not escape but they are not scale replaceable. 1626 // Adjust escape state for outgoing arguments. 1627 const TypeTuple * d = call->tf()->domain(); 1628 VectorSet ptset(Thread::current()->resource_area()); 1629 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1630 const Type* at = d->field_at(i); 1631 Node *arg = call->in(i)->uncast(); 1632 const Type *aat = phase->type(arg); 1633 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr()) { 1634 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 1635 aat->isa_ptr() != NULL, "expecting an Ptr"); 1636 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1637 if (arg->is_AddP()) { 1638 // 1639 // The inline_native_clone() case when the arraycopy stub is called 1640 // after the allocation before Initialize and CheckCastPP nodes. 1641 // 1642 // Set AddP's base (Allocate) as not scalar replaceable since 1643 // pointer to the base (with offset) is passed as argument. 1644 // 1645 arg = get_addp_base(arg); 1646 } 1647 ptset.Clear(); 1648 PointsTo(ptset, arg, phase); 1649 for( VectorSetI j(&ptset); j.test(); ++j ) { 1650 uint pt = j.elem; 1651 set_escape_state(pt, PointsToNode::ArgEscape); 1652 } 1653 } 1654 } 1655 break; 1656 } 1657 1658 case Op_CallStaticJava: 1659 // For a static call, we know exactly what method is being called. 1660 // Use bytecode estimator to record the call's escape affects 1661 { 1662 ciMethod *meth = call->as_CallJava()->method(); 1663 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1664 // fall-through if not a Java method or no analyzer information 1665 if (call_analyzer != NULL) { 1666 const TypeTuple * d = call->tf()->domain(); 1667 VectorSet ptset(Thread::current()->resource_area()); 1668 bool copy_dependencies = false; 1669 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1670 const Type* at = d->field_at(i); 1671 int k = i - TypeFunc::Parms; 1672 1673 if (at->isa_oopptr() != NULL) { 1674 Node *arg = call->in(i)->uncast(); 1675 1676 bool global_escapes = false; 1677 bool fields_escapes = false; 1678 if (!call_analyzer->is_arg_stack(k)) { 1679 // The argument global escapes, mark everything it could point to 1680 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1681 global_escapes = true; 1682 } else { 1683 if (!call_analyzer->is_arg_local(k)) { 1684 // The argument itself doesn't escape, but any fields might 1685 fields_escapes = true; 1686 } 1687 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1688 copy_dependencies = true; 1689 } 1690 1691 ptset.Clear(); 1692 PointsTo(ptset, arg, phase); 1693 for( VectorSetI j(&ptset); j.test(); ++j ) { 1694 uint pt = j.elem; 1695 if (global_escapes) { 1696 //The argument global escapes, mark everything it could point to 1697 set_escape_state(pt, PointsToNode::GlobalEscape); 1698 } else { 1699 if (fields_escapes) { 1700 // The argument itself doesn't escape, but any fields might 1701 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); 1702 } 1703 set_escape_state(pt, PointsToNode::ArgEscape); 1704 } 1705 } 1706 } 1707 } 1708 if (copy_dependencies) 1709 call_analyzer->copy_dependencies(_compile->dependencies()); 1710 break; 1711 } 1712 } 1713 1714 default: 1715 // Fall-through here if not a Java method or no analyzer information 1716 // or some other type of call, assume the worst case: all arguments 1717 // globally escape. 1718 { 1719 // adjust escape state for outgoing arguments 1720 const TypeTuple * d = call->tf()->domain(); 1721 VectorSet ptset(Thread::current()->resource_area()); 1722 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1723 const Type* at = d->field_at(i); 1724 if (at->isa_oopptr() != NULL) { 1725 Node *arg = call->in(i)->uncast(); 1726 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1727 ptset.Clear(); 1728 PointsTo(ptset, arg, phase); 1729 for( VectorSetI j(&ptset); j.test(); ++j ) { 1730 uint pt = j.elem; 1731 set_escape_state(pt, PointsToNode::GlobalEscape); 1732 } 1733 } 1734 } 1735 } 1736 } 1737 } 1738 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { 1739 CallNode *call = resproj->in(0)->as_Call(); 1740 uint call_idx = call->_idx; 1741 uint resproj_idx = resproj->_idx; 1742 1743 switch (call->Opcode()) { 1744 case Op_Allocate: 1745 { 1746 Node *k = call->in(AllocateNode::KlassNode); 1747 const TypeKlassPtr *kt; 1748 if (k->Opcode() == Op_LoadKlass) { 1749 kt = k->as_Load()->type()->isa_klassptr(); 1750 } else { 1751 // Also works for DecodeN(LoadNKlass). 1752 kt = k->as_Type()->type()->isa_klassptr(); 1753 } 1754 assert(kt != NULL, "TypeKlassPtr required."); 1755 ciKlass* cik = kt->klass(); 1756 ciInstanceKlass* ciik = cik->as_instance_klass(); 1757 1758 PointsToNode::EscapeState es; 1759 uint edge_to; 1760 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || ciik->has_finalizer()) { 1761 es = PointsToNode::GlobalEscape; 1762 edge_to = _phantom_object; // Could not be worse 1763 } else { 1764 es = PointsToNode::NoEscape; 1765 edge_to = call_idx; 1766 } 1767 set_escape_state(call_idx, es); 1768 add_pointsto_edge(resproj_idx, edge_to); 1769 _processed.set(resproj_idx); 1770 break; 1771 } 1772 1773 case Op_AllocateArray: 1774 { 1775 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 1776 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 1777 // Not scalar replaceable if the length is not constant or too big. 1778 ptnode_adr(call_idx)->_scalar_replaceable = false; 1779 } 1780 set_escape_state(call_idx, PointsToNode::NoEscape); 1781 add_pointsto_edge(resproj_idx, call_idx); 1782 _processed.set(resproj_idx); 1783 break; 1784 } 1785 1786 case Op_CallStaticJava: 1787 // For a static call, we know exactly what method is being called. 1788 // Use bytecode estimator to record whether the call's return value escapes 1789 { 1790 bool done = true; 1791 const TypeTuple *r = call->tf()->range(); 1792 const Type* ret_type = NULL; 1793 1794 if (r->cnt() > TypeFunc::Parms) 1795 ret_type = r->field_at(TypeFunc::Parms); 1796 1797 // Note: we use isa_ptr() instead of isa_oopptr() here because the 1798 // _multianewarray functions return a TypeRawPtr. 1799 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { 1800 _processed.set(resproj_idx); 1801 break; // doesn't return a pointer type 1802 } 1803 ciMethod *meth = call->as_CallJava()->method(); 1804 const TypeTuple * d = call->tf()->domain(); 1805 if (meth == NULL) { 1806 // not a Java method, assume global escape 1807 set_escape_state(call_idx, PointsToNode::GlobalEscape); 1808 add_pointsto_edge(resproj_idx, _phantom_object); 1809 } else { 1810 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); 1811 bool copy_dependencies = false; 1812 1813 if (call_analyzer->is_return_allocated()) { 1814 // Returns a newly allocated unescaped object, simply 1815 // update dependency information. 1816 // Mark it as NoEscape so that objects referenced by 1817 // it's fields will be marked as NoEscape at least. 1818 set_escape_state(call_idx, PointsToNode::NoEscape); 1819 add_pointsto_edge(resproj_idx, call_idx); 1820 copy_dependencies = true; 1821 } else if (call_analyzer->is_return_local()) { 1822 // determine whether any arguments are returned 1823 set_escape_state(call_idx, PointsToNode::NoEscape); 1824 bool ret_arg = false; 1825 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1826 const Type* at = d->field_at(i); 1827 1828 if (at->isa_oopptr() != NULL) { 1829 Node *arg = call->in(i)->uncast(); 1830 1831 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 1832 ret_arg = true; 1833 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 1834 if (arg_esp->node_type() == PointsToNode::UnknownType) 1835 done = false; 1836 else if (arg_esp->node_type() == PointsToNode::JavaObject) 1837 add_pointsto_edge(resproj_idx, arg->_idx); 1838 else 1839 add_deferred_edge(resproj_idx, arg->_idx); 1840 arg_esp->_hidden_alias = true; 1841 } 1842 } 1843 } 1844 if (done && !ret_arg) { 1845 // Returns unknown object. 1846 set_escape_state(call_idx, PointsToNode::GlobalEscape); 1847 add_pointsto_edge(resproj_idx, _phantom_object); 1848 } 1849 copy_dependencies = true; 1850 } else { 1851 set_escape_state(call_idx, PointsToNode::GlobalEscape); 1852 add_pointsto_edge(resproj_idx, _phantom_object); 1853 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1854 const Type* at = d->field_at(i); 1855 if (at->isa_oopptr() != NULL) { 1856 Node *arg = call->in(i)->uncast(); 1857 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 1858 arg_esp->_hidden_alias = true; 1859 } 1860 } 1861 } 1862 if (copy_dependencies) 1863 call_analyzer->copy_dependencies(_compile->dependencies()); 1864 } 1865 if (done) 1866 _processed.set(resproj_idx); 1867 break; 1868 } 1869 1870 default: 1871 // Some other type of call, assume the worst case that the 1872 // returned value, if any, globally escapes. 1873 { 1874 const TypeTuple *r = call->tf()->range(); 1875 if (r->cnt() > TypeFunc::Parms) { 1876 const Type* ret_type = r->field_at(TypeFunc::Parms); 1877 1878 // Note: we use isa_ptr() instead of isa_oopptr() here because the 1879 // _multianewarray functions return a TypeRawPtr. 1880 if (ret_type->isa_ptr() != NULL) { 1881 set_escape_state(call_idx, PointsToNode::GlobalEscape); 1882 add_pointsto_edge(resproj_idx, _phantom_object); 1883 } 1884 } 1885 _processed.set(resproj_idx); 1886 } 1887 } 1888 } 1889 1890 // Populate Connection Graph with Ideal nodes and create simple 1891 // connection graph edges (do not need to check the node_type of inputs 1892 // or to call PointsTo() to walk the connection graph). 1893 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { 1894 if (_processed.test(n->_idx)) 1895 return; // No need to redefine node's state. 1896 1897 if (n->is_Call()) { 1898 // Arguments to allocation and locking don't escape. 1899 if (n->is_Allocate()) { 1900 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); 1901 record_for_optimizer(n); 1902 } else if (n->is_Lock() || n->is_Unlock()) { 1903 // Put Lock and Unlock nodes on IGVN worklist to process them during 1904 // the first IGVN optimization when escape information is still available. 1905 record_for_optimizer(n); 1906 _processed.set(n->_idx); 1907 } else { 1908 // Have to process call's arguments first. 1909 PointsToNode::NodeType nt = PointsToNode::UnknownType; 1910 1911 // Check if a call returns an object. 1912 const TypeTuple *r = n->as_Call()->tf()->range(); 1913 if (n->is_CallStaticJava() && r->cnt() > TypeFunc::Parms && 1914 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { 1915 // Note: use isa_ptr() instead of isa_oopptr() here because 1916 // the _multianewarray functions return a TypeRawPtr. 1917 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 1918 nt = PointsToNode::JavaObject; 1919 } 1920 } 1921 add_node(n, nt, PointsToNode::UnknownEscape, false); 1922 } 1923 return; 1924 } 1925 1926 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 1927 // ThreadLocal has RawPrt type. 1928 switch (n->Opcode()) { 1929 case Op_AddP: 1930 { 1931 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); 1932 break; 1933 } 1934 case Op_CastX2P: 1935 { // "Unsafe" memory access. 1936 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 1937 break; 1938 } 1939 case Op_CastPP: 1940 case Op_CheckCastPP: 1941 case Op_EncodeP: 1942 case Op_DecodeN: 1943 { 1944 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 1945 int ti = n->in(1)->_idx; 1946 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 1947 if (nt == PointsToNode::UnknownType) { 1948 _delayed_worklist.push(n); // Process it later. 1949 break; 1950 } else if (nt == PointsToNode::JavaObject) { 1951 add_pointsto_edge(n->_idx, ti); 1952 } else { 1953 add_deferred_edge(n->_idx, ti); 1954 } 1955 _processed.set(n->_idx); 1956 break; 1957 } 1958 case Op_ConP: 1959 { 1960 // assume all pointer constants globally escape except for null 1961 PointsToNode::EscapeState es; 1962 if (phase->type(n) == TypePtr::NULL_PTR) 1963 es = PointsToNode::NoEscape; 1964 else 1965 es = PointsToNode::GlobalEscape; 1966 1967 add_node(n, PointsToNode::JavaObject, es, true); 1968 break; 1969 } 1970 case Op_ConN: 1971 { 1972 // assume all narrow oop constants globally escape except for null 1973 PointsToNode::EscapeState es; 1974 if (phase->type(n) == TypeNarrowOop::NULL_PTR) 1975 es = PointsToNode::NoEscape; 1976 else 1977 es = PointsToNode::GlobalEscape; 1978 1979 add_node(n, PointsToNode::JavaObject, es, true); 1980 break; 1981 } 1982 case Op_CreateEx: 1983 { 1984 // assume that all exception objects globally escape 1985 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 1986 break; 1987 } 1988 case Op_LoadKlass: 1989 case Op_LoadNKlass: 1990 { 1991 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 1992 break; 1993 } 1994 case Op_LoadP: 1995 case Op_LoadN: 1996 { 1997 const Type *t = phase->type(n); 1998 if (t->make_ptr() == NULL) { 1999 _processed.set(n->_idx); 2000 return; 2001 } 2002 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2003 break; 2004 } 2005 case Op_Parm: 2006 { 2007 _processed.set(n->_idx); // No need to redefine it state. 2008 uint con = n->as_Proj()->_con; 2009 if (con < TypeFunc::Parms) 2010 return; 2011 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); 2012 if (t->isa_ptr() == NULL) 2013 return; 2014 // We have to assume all input parameters globally escape 2015 // (Note: passing 'false' since _processed is already set). 2016 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); 2017 break; 2018 } 2019 case Op_Phi: 2020 { 2021 const Type *t = n->as_Phi()->type(); 2022 if (t->make_ptr() == NULL) { 2023 // nothing to do if not an oop or narrow oop 2024 _processed.set(n->_idx); 2025 return; 2026 } 2027 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2028 uint i; 2029 for (i = 1; i < n->req() ; i++) { 2030 Node* in = n->in(i); 2031 if (in == NULL) 2032 continue; // ignore NULL 2033 in = in->uncast(); 2034 if (in->is_top() || in == n) 2035 continue; // ignore top or inputs which go back this node 2036 int ti = in->_idx; 2037 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2038 if (nt == PointsToNode::UnknownType) { 2039 break; 2040 } else if (nt == PointsToNode::JavaObject) { 2041 add_pointsto_edge(n->_idx, ti); 2042 } else { 2043 add_deferred_edge(n->_idx, ti); 2044 } 2045 } 2046 if (i >= n->req()) 2047 _processed.set(n->_idx); 2048 else 2049 _delayed_worklist.push(n); 2050 break; 2051 } 2052 case Op_Proj: 2053 { 2054 // we are only interested in the result projection from a call 2055 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2056 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2057 process_call_result(n->as_Proj(), phase); 2058 if (!_processed.test(n->_idx)) { 2059 // The call's result may need to be processed later if the call 2060 // returns it's argument and the argument is not processed yet. 2061 _delayed_worklist.push(n); 2062 } 2063 } else { 2064 _processed.set(n->_idx); 2065 } 2066 break; 2067 } 2068 case Op_Return: 2069 { 2070 if( n->req() > TypeFunc::Parms && 2071 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2072 // Treat Return value as LocalVar with GlobalEscape escape state. 2073 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); 2074 int ti = n->in(TypeFunc::Parms)->_idx; 2075 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2076 if (nt == PointsToNode::UnknownType) { 2077 _delayed_worklist.push(n); // Process it later. 2078 break; 2079 } else if (nt == PointsToNode::JavaObject) { 2080 add_pointsto_edge(n->_idx, ti); 2081 } else { 2082 add_deferred_edge(n->_idx, ti); 2083 } 2084 } 2085 _processed.set(n->_idx); 2086 break; 2087 } 2088 case Op_StoreP: 2089 case Op_StoreN: 2090 { 2091 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2092 adr_type = adr_type->make_ptr(); 2093 if (adr_type->isa_oopptr()) { 2094 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2095 } else { 2096 Node* adr = n->in(MemNode::Address); 2097 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && 2098 adr->in(AddPNode::Address)->is_Proj() && 2099 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2100 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2101 // We are computing a raw address for a store captured 2102 // by an Initialize compute an appropriate address type. 2103 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2104 assert(offs != Type::OffsetBot, "offset must be a constant"); 2105 } else { 2106 _processed.set(n->_idx); 2107 return; 2108 } 2109 } 2110 break; 2111 } 2112 case Op_StorePConditional: 2113 case Op_CompareAndSwapP: 2114 case Op_CompareAndSwapN: 2115 { 2116 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2117 adr_type = adr_type->make_ptr(); 2118 if (adr_type->isa_oopptr()) { 2119 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2120 } else { 2121 _processed.set(n->_idx); 2122 return; 2123 } 2124 break; 2125 } 2126 case Op_ThreadLocal: 2127 { 2128 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); 2129 break; 2130 } 2131 default: 2132 ; 2133 // nothing to do 2134 } 2135 return; 2136 } 2137 2138 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { 2139 uint n_idx = n->_idx; 2140 2141 // Don't set processed bit for AddP, LoadP, StoreP since 2142 // they may need more then one pass to process. 2143 if (_processed.test(n_idx)) 2144 return; // No need to redefine node's state. 2145 2146 if (n->is_Call()) { 2147 CallNode *call = n->as_Call(); 2148 process_call_arguments(call, phase); 2149 _processed.set(n_idx); 2150 return; 2151 } 2152 2153 switch (n->Opcode()) { 2154 case Op_AddP: 2155 { 2156 Node *base = get_addp_base(n); 2157 // Create a field edge to this node from everything base could point to. 2158 VectorSet ptset(Thread::current()->resource_area()); 2159 PointsTo(ptset, base, phase); 2160 for( VectorSetI i(&ptset); i.test(); ++i ) { 2161 uint pt = i.elem; 2162 add_field_edge(pt, n_idx, address_offset(n, phase)); 2163 } 2164 break; 2165 } 2166 case Op_CastX2P: 2167 { 2168 assert(false, "Op_CastX2P"); 2169 break; 2170 } 2171 case Op_CastPP: 2172 case Op_CheckCastPP: 2173 case Op_EncodeP: 2174 case Op_DecodeN: 2175 { 2176 int ti = n->in(1)->_idx; 2177 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2178 add_pointsto_edge(n_idx, ti); 2179 } else { 2180 add_deferred_edge(n_idx, ti); 2181 } 2182 _processed.set(n_idx); 2183 break; 2184 } 2185 case Op_ConP: 2186 { 2187 assert(false, "Op_ConP"); 2188 break; 2189 } 2190 case Op_ConN: 2191 { 2192 assert(false, "Op_ConN"); 2193 break; 2194 } 2195 case Op_CreateEx: 2196 { 2197 assert(false, "Op_CreateEx"); 2198 break; 2199 } 2200 case Op_LoadKlass: 2201 case Op_LoadNKlass: 2202 { 2203 assert(false, "Op_LoadKlass"); 2204 break; 2205 } 2206 case Op_LoadP: 2207 case Op_LoadN: 2208 { 2209 const Type *t = phase->type(n); 2210 #ifdef ASSERT 2211 if (t->make_ptr() == NULL) 2212 assert(false, "Op_LoadP"); 2213 #endif 2214 2215 Node* adr = n->in(MemNode::Address)->uncast(); 2216 const Type *adr_type = phase->type(adr); 2217 Node* adr_base; 2218 if (adr->is_AddP()) { 2219 adr_base = get_addp_base(adr); 2220 } else { 2221 adr_base = adr; 2222 } 2223 2224 // For everything "adr_base" could point to, create a deferred edge from 2225 // this node to each field with the same offset. 2226 VectorSet ptset(Thread::current()->resource_area()); 2227 PointsTo(ptset, adr_base, phase); 2228 int offset = address_offset(adr, phase); 2229 for( VectorSetI i(&ptset); i.test(); ++i ) { 2230 uint pt = i.elem; 2231 add_deferred_edge_to_fields(n_idx, pt, offset); 2232 } 2233 break; 2234 } 2235 case Op_Parm: 2236 { 2237 assert(false, "Op_Parm"); 2238 break; 2239 } 2240 case Op_Phi: 2241 { 2242 #ifdef ASSERT 2243 const Type *t = n->as_Phi()->type(); 2244 if (t->make_ptr() == NULL) 2245 assert(false, "Op_Phi"); 2246 #endif 2247 for (uint i = 1; i < n->req() ; i++) { 2248 Node* in = n->in(i); 2249 if (in == NULL) 2250 continue; // ignore NULL 2251 in = in->uncast(); 2252 if (in->is_top() || in == n) 2253 continue; // ignore top or inputs which go back this node 2254 int ti = in->_idx; 2255 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2256 assert(nt != PointsToNode::UnknownType, "all nodes should be known"); 2257 if (nt == PointsToNode::JavaObject) { 2258 add_pointsto_edge(n_idx, ti); 2259 } else { 2260 add_deferred_edge(n_idx, ti); 2261 } 2262 } 2263 _processed.set(n_idx); 2264 break; 2265 } 2266 case Op_Proj: 2267 { 2268 // we are only interested in the result projection from a call 2269 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2270 process_call_result(n->as_Proj(), phase); 2271 assert(_processed.test(n_idx), "all call results should be processed"); 2272 } else { 2273 assert(false, "Op_Proj"); 2274 } 2275 break; 2276 } 2277 case Op_Return: 2278 { 2279 #ifdef ASSERT 2280 if( n->req() <= TypeFunc::Parms || 2281 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2282 assert(false, "Op_Return"); 2283 } 2284 #endif 2285 int ti = n->in(TypeFunc::Parms)->_idx; 2286 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2287 add_pointsto_edge(n_idx, ti); 2288 } else { 2289 add_deferred_edge(n_idx, ti); 2290 } 2291 _processed.set(n_idx); 2292 break; 2293 } 2294 case Op_StoreP: 2295 case Op_StoreN: 2296 case Op_StorePConditional: 2297 case Op_CompareAndSwapP: 2298 case Op_CompareAndSwapN: 2299 { 2300 Node *adr = n->in(MemNode::Address); 2301 const Type *adr_type = phase->type(adr)->make_ptr(); 2302 #ifdef ASSERT 2303 if (!adr_type->isa_oopptr()) 2304 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); 2305 #endif 2306 2307 assert(adr->is_AddP(), "expecting an AddP"); 2308 Node *adr_base = get_addp_base(adr); 2309 Node *val = n->in(MemNode::ValueIn)->uncast(); 2310 // For everything "adr_base" could point to, create a deferred edge 2311 // to "val" from each field with the same offset. 2312 VectorSet ptset(Thread::current()->resource_area()); 2313 PointsTo(ptset, adr_base, phase); 2314 for( VectorSetI i(&ptset); i.test(); ++i ) { 2315 uint pt = i.elem; 2316 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); 2317 } 2318 break; 2319 } 2320 case Op_ThreadLocal: 2321 { 2322 assert(false, "Op_ThreadLocal"); 2323 break; 2324 } 2325 default: 2326 ; 2327 // nothing to do 2328 } 2329 } 2330 2331 #ifndef PRODUCT 2332 void ConnectionGraph::dump() { 2333 PhaseGVN *igvn = _compile->initial_gvn(); 2334 bool first = true; 2335 2336 uint size = nodes_size(); 2337 for (uint ni = 0; ni < size; ni++) { 2338 PointsToNode *ptn = ptnode_adr(ni); 2339 PointsToNode::NodeType ptn_type = ptn->node_type(); 2340 2341 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) 2342 continue; 2343 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn); 2344 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { 2345 if (first) { 2346 tty->cr(); 2347 tty->print("======== Connection graph for "); 2348 _compile->method()->print_short_name(); 2349 tty->cr(); 2350 first = false; 2351 } 2352 tty->print("%6d ", ni); 2353 ptn->dump(); 2354 // Print all locals which reference this allocation 2355 for (uint li = ni; li < size; li++) { 2356 PointsToNode *ptn_loc = ptnode_adr(li); 2357 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); 2358 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && 2359 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { 2360 ptnode_adr(li)->dump(false); 2361 } 2362 } 2363 if (Verbose) { 2364 // Print all fields which reference this allocation 2365 for (uint i = 0; i < ptn->edge_count(); i++) { 2366 uint ei = ptn->edge_target(i); 2367 ptnode_adr(ei)->dump(false); 2368 } 2369 } 2370 tty->cr(); 2371 } 2372 } 2373 } 2374 #endif