1 /* 2 * Copyright (c) 2005, 2009, 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 "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, PhaseIterGVN *igvn) : 85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()), 86 _processed(C->comp_arena()), 87 _collecting(true), 88 _compile(C), 89 _igvn(igvn), 90 _node_map(C->comp_arena()) { 91 92 _phantom_object = C->top()->_idx, 93 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true); 94 95 // Add ConP(#NULL) and ConN(#NULL) nodes. 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) { 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 PointsToNode::EscapeState orig_es = es; 211 212 // compute max escape state of anything this node could point to 213 VectorSet ptset(Thread::current()->resource_area()); 214 PointsTo(ptset, n); 215 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) { 216 uint pt = i.elem; 217 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); 218 if (pes > es) 219 es = pes; 220 } 221 if (orig_es != es) { 222 // cache the computed escape state 223 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state"); 224 ptnode_adr(idx)->set_escape_state(es); 225 } // orig_es could be PointsToNode::UnknownEscape 226 return es; 227 } 228 229 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n) { 230 VectorSet visited(Thread::current()->resource_area()); 231 GrowableArray<uint> worklist; 232 233 #ifdef ASSERT 234 Node *orig_n = n; 235 #endif 236 237 n = n->uncast(); 238 PointsToNode* npt = ptnode_adr(n->_idx); 239 240 // If we have a JavaObject, return just that object 241 if (npt->node_type() == PointsToNode::JavaObject) { 242 ptset.set(n->_idx); 243 return; 244 } 245 #ifdef ASSERT 246 if (npt->_node == NULL) { 247 if (orig_n != n) 248 orig_n->dump(); 249 n->dump(); 250 assert(npt->_node != NULL, "unregistered node"); 251 } 252 #endif 253 worklist.push(n->_idx); 254 while(worklist.length() > 0) { 255 int ni = worklist.pop(); 256 if (visited.test_set(ni)) 257 continue; 258 259 PointsToNode* pn = ptnode_adr(ni); 260 // ensure that all inputs of a Phi have been processed 261 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); 262 263 int edges_processed = 0; 264 uint e_cnt = pn->edge_count(); 265 for (uint e = 0; e < e_cnt; e++) { 266 uint etgt = pn->edge_target(e); 267 PointsToNode::EdgeType et = pn->edge_type(e); 268 if (et == PointsToNode::PointsToEdge) { 269 ptset.set(etgt); 270 edges_processed++; 271 } else if (et == PointsToNode::DeferredEdge) { 272 worklist.push(etgt); 273 edges_processed++; 274 } else { 275 assert(false,"neither PointsToEdge or DeferredEdge"); 276 } 277 } 278 if (edges_processed == 0) { 279 // no deferred or pointsto edges found. Assume the value was set 280 // outside this method. Add the phantom object to the pointsto set. 281 ptset.set(_phantom_object); 282 } 283 } 284 } 285 286 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { 287 // This method is most expensive during ConnectionGraph construction. 288 // Reuse vectorSet and an additional growable array for deferred edges. 289 deferred_edges->clear(); 290 visited->Clear(); 291 292 visited->set(ni); 293 PointsToNode *ptn = ptnode_adr(ni); 294 295 // Mark current edges as visited and move deferred edges to separate array. 296 for (uint i = 0; i < ptn->edge_count(); ) { 297 uint t = ptn->edge_target(i); 298 #ifdef ASSERT 299 assert(!visited->test_set(t), "expecting no duplications"); 300 #else 301 visited->set(t); 302 #endif 303 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { 304 ptn->remove_edge(t, PointsToNode::DeferredEdge); 305 deferred_edges->append(t); 306 } else { 307 i++; 308 } 309 } 310 for (int next = 0; next < deferred_edges->length(); ++next) { 311 uint t = deferred_edges->at(next); 312 PointsToNode *ptt = ptnode_adr(t); 313 uint e_cnt = ptt->edge_count(); 314 for (uint e = 0; e < e_cnt; e++) { 315 uint etgt = ptt->edge_target(e); 316 if (visited->test_set(etgt)) 317 continue; 318 319 PointsToNode::EdgeType et = ptt->edge_type(e); 320 if (et == PointsToNode::PointsToEdge) { 321 add_pointsto_edge(ni, etgt); 322 if(etgt == _phantom_object) { 323 // Special case - field set outside (globally escaping). 324 ptn->set_escape_state(PointsToNode::GlobalEscape); 325 } 326 } else if (et == PointsToNode::DeferredEdge) { 327 deferred_edges->append(etgt); 328 } else { 329 assert(false,"invalid connection graph"); 330 } 331 } 332 } 333 } 334 335 336 // Add an edge to node given by "to_i" from any field of adr_i whose offset 337 // matches "offset" A deferred edge is added if to_i is a LocalVar, and 338 // a pointsto edge is added if it is a JavaObject 339 340 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { 341 PointsToNode* an = ptnode_adr(adr_i); 342 PointsToNode* to = ptnode_adr(to_i); 343 bool deferred = (to->node_type() == PointsToNode::LocalVar); 344 345 for (uint fe = 0; fe < an->edge_count(); fe++) { 346 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 347 int fi = an->edge_target(fe); 348 PointsToNode* pf = ptnode_adr(fi); 349 int po = pf->offset(); 350 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 351 if (deferred) 352 add_deferred_edge(fi, to_i); 353 else 354 add_pointsto_edge(fi, to_i); 355 } 356 } 357 } 358 359 // Add a deferred edge from node given by "from_i" to any field of adr_i 360 // whose offset matches "offset". 361 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { 362 PointsToNode* an = ptnode_adr(adr_i); 363 for (uint fe = 0; fe < an->edge_count(); fe++) { 364 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 365 int fi = an->edge_target(fe); 366 PointsToNode* pf = ptnode_adr(fi); 367 int po = pf->offset(); 368 if (pf->edge_count() == 0) { 369 // we have not seen any stores to this field, assume it was set outside this method 370 add_pointsto_edge(fi, _phantom_object); 371 } 372 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 373 add_deferred_edge(from_i, fi); 374 } 375 } 376 } 377 378 // Helper functions 379 380 static Node* get_addp_base(Node *addp) { 381 assert(addp->is_AddP(), "must be AddP"); 382 // 383 // AddP cases for Base and Address inputs: 384 // case #1. Direct object's field reference: 385 // Allocate 386 // | 387 // Proj #5 ( oop result ) 388 // | 389 // CheckCastPP (cast to instance type) 390 // | | 391 // AddP ( base == address ) 392 // 393 // case #2. Indirect object's field reference: 394 // Phi 395 // | 396 // CastPP (cast to instance type) 397 // | | 398 // AddP ( base == address ) 399 // 400 // case #3. Raw object's field reference for Initialize node: 401 // Allocate 402 // | 403 // Proj #5 ( oop result ) 404 // top | 405 // \ | 406 // AddP ( base == top ) 407 // 408 // case #4. Array's element reference: 409 // {CheckCastPP | CastPP} 410 // | | | 411 // | AddP ( array's element offset ) 412 // | | 413 // AddP ( array's offset ) 414 // 415 // case #5. Raw object's field reference for arraycopy stub call: 416 // The inline_native_clone() case when the arraycopy stub is called 417 // after the allocation before Initialize and CheckCastPP nodes. 418 // Allocate 419 // | 420 // Proj #5 ( oop result ) 421 // | | 422 // AddP ( base == address ) 423 // 424 // case #6. Constant Pool, ThreadLocal, CastX2P or 425 // Raw object's field reference: 426 // {ConP, ThreadLocal, CastX2P, raw Load} 427 // top | 428 // \ | 429 // AddP ( base == top ) 430 // 431 // case #7. Klass's field reference. 432 // LoadKlass 433 // | | 434 // AddP ( base == address ) 435 // 436 // case #8. narrow Klass's field reference. 437 // LoadNKlass 438 // | 439 // DecodeN 440 // | | 441 // AddP ( base == address ) 442 // 443 Node *base = addp->in(AddPNode::Base)->uncast(); 444 if (base->is_top()) { // The AddP case #3 and #6. 445 base = addp->in(AddPNode::Address)->uncast(); 446 while (base->is_AddP()) { 447 // Case #6 (unsafe access) may have several chained AddP nodes. 448 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only"); 449 base = base->in(AddPNode::Address)->uncast(); 450 } 451 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || 452 base->Opcode() == Op_CastX2P || base->is_DecodeN() || 453 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || 454 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); 455 } 456 return base; 457 } 458 459 static Node* find_second_addp(Node* addp, Node* n) { 460 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 461 462 Node* addp2 = addp->raw_out(0); 463 if (addp->outcnt() == 1 && addp2->is_AddP() && 464 addp2->in(AddPNode::Base) == n && 465 addp2->in(AddPNode::Address) == addp) { 466 467 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 468 // 469 // Find array's offset to push it on worklist first and 470 // as result process an array's element offset first (pushed second) 471 // to avoid CastPP for the array's offset. 472 // Otherwise the inserted CastPP (LocalVar) will point to what 473 // the AddP (Field) points to. Which would be wrong since 474 // the algorithm expects the CastPP has the same point as 475 // as AddP's base CheckCastPP (LocalVar). 476 // 477 // ArrayAllocation 478 // | 479 // CheckCastPP 480 // | 481 // memProj (from ArrayAllocation CheckCastPP) 482 // | || 483 // | || Int (element index) 484 // | || | ConI (log(element size)) 485 // | || | / 486 // | || LShift 487 // | || / 488 // | AddP (array's element offset) 489 // | | 490 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 491 // | / / 492 // AddP (array's offset) 493 // | 494 // Load/Store (memory operation on array's element) 495 // 496 return addp2; 497 } 498 return NULL; 499 } 500 501 // 502 // Adjust the type and inputs of an AddP which computes the 503 // address of a field of an instance 504 // 505 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { 506 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 507 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 508 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 509 if (t == NULL) { 510 // We are computing a raw address for a store captured by an Initialize 511 // compute an appropriate address type (cases #3 and #5). 512 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 513 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 514 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 515 assert(offs != Type::OffsetBot, "offset must be a constant"); 516 t = base_t->add_offset(offs)->is_oopptr(); 517 } 518 int inst_id = base_t->instance_id(); 519 assert(!t->is_known_instance() || t->instance_id() == inst_id, 520 "old type must be non-instance or match new type"); 521 522 // The type 't' could be subclass of 'base_t'. 523 // As result t->offset() could be large then base_t's size and it will 524 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 525 // constructor verifies correctness of the offset. 526 // 527 // It could happened on subclass's branch (from the type profiling 528 // inlining) which was not eliminated during parsing since the exactness 529 // of the allocation type was not propagated to the subclass type check. 530 // 531 // Or the type 't' could be not related to 'base_t' at all. 532 // It could happened when CHA type is different from MDO type on a dead path 533 // (for example, from instanceof check) which is not collapsed during parsing. 534 // 535 // Do nothing for such AddP node and don't process its users since 536 // this code branch will go away. 537 // 538 if (!t->is_known_instance() && 539 !base_t->klass()->is_subtype_of(t->klass())) { 540 return false; // bail out 541 } 542 543 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 544 // Do NOT remove the next line: ensure a new alias index is allocated 545 // for the instance type. Note: C++ will not remove it since the call 546 // has side effect. 547 int alias_idx = _compile->get_alias_index(tinst); 548 igvn->set_type(addp, tinst); 549 // record the allocation in the node map 550 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered"); 551 set_map(addp->_idx, get_map(base->_idx)); 552 553 // Set addp's Base and Address to 'base'. 554 Node *abase = addp->in(AddPNode::Base); 555 Node *adr = addp->in(AddPNode::Address); 556 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 557 adr->in(0)->_idx == (uint)inst_id) { 558 // Skip AddP cases #3 and #5. 559 } else { 560 assert(!abase->is_top(), "sanity"); // AddP case #3 561 if (abase != base) { 562 igvn->hash_delete(addp); 563 addp->set_req(AddPNode::Base, base); 564 if (abase == adr) { 565 addp->set_req(AddPNode::Address, base); 566 } else { 567 // AddP case #4 (adr is array's element offset AddP node) 568 #ifdef ASSERT 569 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 570 assert(adr->is_AddP() && atype != NULL && 571 atype->instance_id() == inst_id, "array's element offset should be processed first"); 572 #endif 573 } 574 igvn->hash_insert(addp); 575 } 576 } 577 // Put on IGVN worklist since at least addp's type was changed above. 578 record_for_optimizer(addp); 579 return true; 580 } 581 582 // 583 // Create a new version of orig_phi if necessary. Returns either the newly 584 // created phi or an existing phi. Sets create_new to indicate wheter a new 585 // phi was created. Cache the last newly created phi in the node map. 586 // 587 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { 588 Compile *C = _compile; 589 new_created = false; 590 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 591 // nothing to do if orig_phi is bottom memory or matches alias_idx 592 if (phi_alias_idx == alias_idx) { 593 return orig_phi; 594 } 595 // Have we recently created a Phi for this alias index? 596 PhiNode *result = get_map_phi(orig_phi->_idx); 597 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 598 return result; 599 } 600 // Previous check may fail when the same wide memory Phi was split into Phis 601 // for different memory slices. Search all Phis for this region. 602 if (result != NULL) { 603 Node* region = orig_phi->in(0); 604 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 605 Node* phi = region->fast_out(i); 606 if (phi->is_Phi() && 607 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 608 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 609 return phi->as_Phi(); 610 } 611 } 612 } 613 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { 614 if (C->do_escape_analysis() == true && !C->failing()) { 615 // Retry compilation without escape analysis. 616 // If this is the first failure, the sentinel string will "stick" 617 // to the Compile object, and the C2Compiler will see it and retry. 618 C->record_failure(C2Compiler::retry_no_escape_analysis()); 619 } 620 return NULL; 621 } 622 orig_phi_worklist.append_if_missing(orig_phi); 623 const TypePtr *atype = C->get_adr_type(alias_idx); 624 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 625 C->copy_node_notes_to(result, orig_phi); 626 igvn->set_type(result, result->bottom_type()); 627 record_for_optimizer(result); 628 629 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;) 630 assert(pn == NULL || pn == orig_phi, "wrong node"); 631 set_map(orig_phi->_idx, result); 632 ptnode_adr(orig_phi->_idx)->_node = orig_phi; 633 634 new_created = true; 635 return result; 636 } 637 638 // 639 // Return a new version of Memory Phi "orig_phi" with the inputs having the 640 // specified alias index. 641 // 642 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { 643 644 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 645 Compile *C = _compile; 646 bool new_phi_created; 647 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); 648 if (!new_phi_created) { 649 return result; 650 } 651 652 GrowableArray<PhiNode *> phi_list; 653 GrowableArray<uint> cur_input; 654 655 PhiNode *phi = orig_phi; 656 uint idx = 1; 657 bool finished = false; 658 while(!finished) { 659 while (idx < phi->req()) { 660 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); 661 if (mem != NULL && mem->is_Phi()) { 662 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); 663 if (new_phi_created) { 664 // found an phi for which we created a new split, push current one on worklist and begin 665 // processing new one 666 phi_list.push(phi); 667 cur_input.push(idx); 668 phi = mem->as_Phi(); 669 result = newphi; 670 idx = 1; 671 continue; 672 } else { 673 mem = newphi; 674 } 675 } 676 if (C->failing()) { 677 return NULL; 678 } 679 result->set_req(idx++, mem); 680 } 681 #ifdef ASSERT 682 // verify that the new Phi has an input for each input of the original 683 assert( phi->req() == result->req(), "must have same number of inputs."); 684 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 685 #endif 686 // Check if all new phi's inputs have specified alias index. 687 // Otherwise use old phi. 688 for (uint i = 1; i < phi->req(); i++) { 689 Node* in = result->in(i); 690 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 691 } 692 // we have finished processing a Phi, see if there are any more to do 693 finished = (phi_list.length() == 0 ); 694 if (!finished) { 695 phi = phi_list.pop(); 696 idx = cur_input.pop(); 697 PhiNode *prev_result = get_map_phi(phi->_idx); 698 prev_result->set_req(idx++, result); 699 result = prev_result; 700 } 701 } 702 return result; 703 } 704 705 706 // 707 // The next methods are derived from methods in MemNode. 708 // 709 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 710 Node *mem = mmem; 711 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 712 // means an array I have not precisely typed yet. Do not do any 713 // alias stuff with it any time soon. 714 if( toop->base() != Type::AnyPtr && 715 !(toop->klass() != NULL && 716 toop->klass()->is_java_lang_Object() && 717 toop->offset() == Type::OffsetBot) ) { 718 mem = mmem->memory_at(alias_idx); 719 // Update input if it is progress over what we have now 720 } 721 return mem; 722 } 723 724 // 725 // Move memory users to their memory slices. 726 // 727 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) { 728 Compile* C = _compile; 729 730 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 731 assert(tp != NULL, "ptr type"); 732 int alias_idx = C->get_alias_index(tp); 733 int general_idx = C->get_general_index(alias_idx); 734 735 // Move users first 736 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 737 Node* use = n->fast_out(i); 738 if (use->is_MergeMem()) { 739 MergeMemNode* mmem = use->as_MergeMem(); 740 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 741 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 742 continue; // Nothing to do 743 } 744 // Replace previous general reference to mem node. 745 uint orig_uniq = C->unique(); 746 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 747 assert(orig_uniq == C->unique(), "no new nodes"); 748 mmem->set_memory_at(general_idx, m); 749 --imax; 750 --i; 751 } else if (use->is_MemBar()) { 752 assert(!use->is_Initialize(), "initializing stores should not be moved"); 753 if (use->req() > MemBarNode::Precedent && 754 use->in(MemBarNode::Precedent) == n) { 755 // Don't move related membars. 756 record_for_optimizer(use); 757 continue; 758 } 759 tp = use->as_MemBar()->adr_type()->isa_ptr(); 760 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 761 alias_idx == general_idx) { 762 continue; // Nothing to do 763 } 764 // Move to general memory slice. 765 uint orig_uniq = C->unique(); 766 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 767 assert(orig_uniq == C->unique(), "no new nodes"); 768 igvn->hash_delete(use); 769 imax -= use->replace_edge(n, m); 770 igvn->hash_insert(use); 771 record_for_optimizer(use); 772 --i; 773 #ifdef ASSERT 774 } else if (use->is_Mem()) { 775 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 776 // Don't move related cardmark. 777 continue; 778 } 779 // Memory nodes should have new memory input. 780 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 781 assert(tp != NULL, "ptr type"); 782 int idx = C->get_alias_index(tp); 783 assert(get_map(use->_idx) != NULL || idx == alias_idx, 784 "Following memory nodes should have new memory input or be on the same memory slice"); 785 } else if (use->is_Phi()) { 786 // Phi nodes should be split and moved already. 787 tp = use->as_Phi()->adr_type()->isa_ptr(); 788 assert(tp != NULL, "ptr type"); 789 int idx = C->get_alias_index(tp); 790 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 791 } else { 792 use->dump(); 793 assert(false, "should not be here"); 794 #endif 795 } 796 } 797 } 798 799 // 800 // Search memory chain of "mem" to find a MemNode whose address 801 // is the specified alias index. 802 // 803 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { 804 if (orig_mem == NULL) 805 return orig_mem; 806 Compile* C = phase->C; 807 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 808 bool is_instance = (toop != NULL) && toop->is_known_instance(); 809 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 810 Node *prev = NULL; 811 Node *result = orig_mem; 812 while (prev != result) { 813 prev = result; 814 if (result == start_mem) 815 break; // hit one of our sentinels 816 if (result->is_Mem()) { 817 const Type *at = phase->type(result->in(MemNode::Address)); 818 if (at != Type::TOP) { 819 assert (at->isa_ptr() != NULL, "pointer type required."); 820 int idx = C->get_alias_index(at->is_ptr()); 821 if (idx == alias_idx) 822 break; 823 } 824 result = result->in(MemNode::Memory); 825 } 826 if (!is_instance) 827 continue; // don't search further for non-instance types 828 // skip over a call which does not affect this memory slice 829 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 830 Node *proj_in = result->in(0); 831 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 832 break; // hit one of our sentinels 833 } else if (proj_in->is_Call()) { 834 CallNode *call = proj_in->as_Call(); 835 if (!call->may_modify(toop, phase)) { 836 result = call->in(TypeFunc::Memory); 837 } 838 } else if (proj_in->is_Initialize()) { 839 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 840 // Stop if this is the initialization for the object instance which 841 // which contains this memory slice, otherwise skip over it. 842 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 843 result = proj_in->in(TypeFunc::Memory); 844 } 845 } else if (proj_in->is_MemBar()) { 846 result = proj_in->in(TypeFunc::Memory); 847 } 848 } else if (result->is_MergeMem()) { 849 MergeMemNode *mmem = result->as_MergeMem(); 850 result = step_through_mergemem(mmem, alias_idx, toop); 851 if (result == mmem->base_memory()) { 852 // Didn't find instance memory, search through general slice recursively. 853 result = mmem->memory_at(C->get_general_index(alias_idx)); 854 result = find_inst_mem(result, alias_idx, orig_phis, phase); 855 if (C->failing()) { 856 return NULL; 857 } 858 mmem->set_memory_at(alias_idx, result); 859 } 860 } else if (result->is_Phi() && 861 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 862 Node *un = result->as_Phi()->unique_input(phase); 863 if (un != NULL) { 864 orig_phis.append_if_missing(result->as_Phi()); 865 result = un; 866 } else { 867 break; 868 } 869 } else if (result->is_ClearArray()) { 870 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) { 871 // Can not bypass initialization of the instance 872 // we are looking for. 873 break; 874 } 875 // Otherwise skip it (the call updated 'result' value). 876 } else if (result->Opcode() == Op_SCMemProj) { 877 assert(result->in(0)->is_LoadStore(), "sanity"); 878 const Type *at = phase->type(result->in(0)->in(MemNode::Address)); 879 if (at != Type::TOP) { 880 assert (at->isa_ptr() != NULL, "pointer type required."); 881 int idx = C->get_alias_index(at->is_ptr()); 882 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 883 break; 884 } 885 result = result->in(0)->in(MemNode::Memory); 886 } 887 } 888 if (result->is_Phi()) { 889 PhiNode *mphi = result->as_Phi(); 890 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 891 const TypePtr *t = mphi->adr_type(); 892 if (C->get_alias_index(t) != alias_idx) { 893 // Create a new Phi with the specified alias index type. 894 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); 895 } else if (!is_instance) { 896 // Push all non-instance Phis on the orig_phis worklist to update inputs 897 // during Phase 4 if needed. 898 orig_phis.append_if_missing(mphi); 899 } 900 } 901 // the result is either MemNode, PhiNode, InitializeNode. 902 return result; 903 } 904 905 // 906 // Convert the types of unescaped object to instance types where possible, 907 // propagate the new type information through the graph, and update memory 908 // edges and MergeMem inputs to reflect the new type. 909 // 910 // We start with allocations (and calls which may be allocations) on alloc_worklist. 911 // The processing is done in 4 phases: 912 // 913 // Phase 1: Process possible allocations from alloc_worklist. Create instance 914 // types for the CheckCastPP for allocations where possible. 915 // Propagate the the new types through users as follows: 916 // casts and Phi: push users on alloc_worklist 917 // AddP: cast Base and Address inputs to the instance type 918 // push any AddP users on alloc_worklist and push any memnode 919 // users onto memnode_worklist. 920 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 921 // search the Memory chain for a store with the appropriate type 922 // address type. If a Phi is found, create a new version with 923 // the appropriate memory slices from each of the Phi inputs. 924 // For stores, process the users as follows: 925 // MemNode: push on memnode_worklist 926 // MergeMem: push on mergemem_worklist 927 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 928 // moving the first node encountered of each instance type to the 929 // the input corresponding to its alias index. 930 // appropriate memory slice. 931 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 932 // 933 // In the following example, the CheckCastPP nodes are the cast of allocation 934 // results and the allocation of node 29 is unescaped and eligible to be an 935 // instance type. 936 // 937 // We start with: 938 // 939 // 7 Parm #memory 940 // 10 ConI "12" 941 // 19 CheckCastPP "Foo" 942 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 943 // 29 CheckCastPP "Foo" 944 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 945 // 946 // 40 StoreP 25 7 20 ... alias_index=4 947 // 50 StoreP 35 40 30 ... alias_index=4 948 // 60 StoreP 45 50 20 ... alias_index=4 949 // 70 LoadP _ 60 30 ... alias_index=4 950 // 80 Phi 75 50 60 Memory alias_index=4 951 // 90 LoadP _ 80 30 ... alias_index=4 952 // 100 LoadP _ 80 20 ... alias_index=4 953 // 954 // 955 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 956 // and creating a new alias index for node 30. This gives: 957 // 958 // 7 Parm #memory 959 // 10 ConI "12" 960 // 19 CheckCastPP "Foo" 961 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 962 // 29 CheckCastPP "Foo" iid=24 963 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 964 // 965 // 40 StoreP 25 7 20 ... alias_index=4 966 // 50 StoreP 35 40 30 ... alias_index=6 967 // 60 StoreP 45 50 20 ... alias_index=4 968 // 70 LoadP _ 60 30 ... alias_index=6 969 // 80 Phi 75 50 60 Memory alias_index=4 970 // 90 LoadP _ 80 30 ... alias_index=6 971 // 100 LoadP _ 80 20 ... alias_index=4 972 // 973 // In phase 2, new memory inputs are computed for the loads and stores, 974 // And a new version of the phi is created. In phase 4, the inputs to 975 // node 80 are updated and then the memory nodes are updated with the 976 // values computed in phase 2. This results in: 977 // 978 // 7 Parm #memory 979 // 10 ConI "12" 980 // 19 CheckCastPP "Foo" 981 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 982 // 29 CheckCastPP "Foo" iid=24 983 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 984 // 985 // 40 StoreP 25 7 20 ... alias_index=4 986 // 50 StoreP 35 7 30 ... alias_index=6 987 // 60 StoreP 45 40 20 ... alias_index=4 988 // 70 LoadP _ 50 30 ... alias_index=6 989 // 80 Phi 75 40 60 Memory alias_index=4 990 // 120 Phi 75 50 50 Memory alias_index=6 991 // 90 LoadP _ 120 30 ... alias_index=6 992 // 100 LoadP _ 80 20 ... alias_index=4 993 // 994 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 995 GrowableArray<Node *> memnode_worklist; 996 GrowableArray<PhiNode *> orig_phis; 997 998 PhaseGVN *igvn = _igvn; 999 uint new_index_start = (uint) _compile->num_alias_types(); 1000 Arena* arena = Thread::current()->resource_area(); 1001 VectorSet visited(arena); 1002 VectorSet ptset(arena); 1003 1004 1005 // Phase 1: Process possible allocations from alloc_worklist. 1006 // Create instance types for the CheckCastPP for allocations where possible. 1007 // 1008 // (Note: don't forget to change the order of the second AddP node on 1009 // the alloc_worklist if the order of the worklist processing is changed, 1010 // see the comment in find_second_addp().) 1011 // 1012 while (alloc_worklist.length() != 0) { 1013 Node *n = alloc_worklist.pop(); 1014 uint ni = n->_idx; 1015 const TypeOopPtr* tinst = NULL; 1016 if (n->is_Call()) { 1017 CallNode *alloc = n->as_Call(); 1018 // copy escape information to call node 1019 PointsToNode* ptn = ptnode_adr(alloc->_idx); 1020 PointsToNode::EscapeState es = escape_state(alloc); 1021 // We have an allocation or call which returns a Java object, 1022 // see if it is unescaped. 1023 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) 1024 continue; 1025 1026 // Find CheckCastPP for the allocate or for the return value of a call 1027 n = alloc->result_cast(); 1028 if (n == NULL) { // No uses except Initialize node 1029 if (alloc->is_Allocate()) { 1030 // Set the scalar_replaceable flag for allocation 1031 // so it could be eliminated if it has no uses. 1032 alloc->as_Allocate()->_is_scalar_replaceable = true; 1033 } 1034 continue; 1035 } 1036 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 1037 assert(!alloc->is_Allocate(), "allocation should have unique type"); 1038 continue; 1039 } 1040 1041 // The inline code for Object.clone() casts the allocation result to 1042 // java.lang.Object and then to the actual type of the allocated 1043 // object. Detect this case and use the second cast. 1044 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 1045 // the allocation result is cast to java.lang.Object and then 1046 // to the actual Array type. 1047 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 1048 && (alloc->is_AllocateArray() || 1049 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 1050 Node *cast2 = NULL; 1051 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1052 Node *use = n->fast_out(i); 1053 if (use->is_CheckCastPP()) { 1054 cast2 = use; 1055 break; 1056 } 1057 } 1058 if (cast2 != NULL) { 1059 n = cast2; 1060 } else { 1061 // Non-scalar replaceable if the allocation type is unknown statically 1062 // (reflection allocation), the object can't be restored during 1063 // deoptimization without precise type. 1064 continue; 1065 } 1066 } 1067 if (alloc->is_Allocate()) { 1068 // Set the scalar_replaceable flag for allocation 1069 // so it could be eliminated. 1070 alloc->as_Allocate()->_is_scalar_replaceable = true; 1071 } 1072 set_escape_state(n->_idx, es); 1073 // in order for an object to be scalar-replaceable, it must be: 1074 // - a direct allocation (not a call returning an object) 1075 // - non-escaping 1076 // - eligible to be a unique type 1077 // - not determined to be ineligible by escape analysis 1078 assert(ptnode_adr(alloc->_idx)->_node != NULL && 1079 ptnode_adr(n->_idx)->_node != NULL, "should be registered"); 1080 set_map(alloc->_idx, n); 1081 set_map(n->_idx, alloc); 1082 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 1083 if (t == NULL) 1084 continue; // not a TypeInstPtr 1085 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 1086 igvn->hash_delete(n); 1087 igvn->set_type(n, tinst); 1088 n->raise_bottom_type(tinst); 1089 igvn->hash_insert(n); 1090 record_for_optimizer(n); 1091 if (alloc->is_Allocate() && ptn->_scalar_replaceable && 1092 (t->isa_instptr() || t->isa_aryptr())) { 1093 1094 // First, put on the worklist all Field edges from Connection Graph 1095 // which is more accurate then putting immediate users from Ideal Graph. 1096 for (uint e = 0; e < ptn->edge_count(); e++) { 1097 Node *use = ptnode_adr(ptn->edge_target(e))->_node; 1098 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), 1099 "only AddP nodes are Field edges in CG"); 1100 if (use->outcnt() > 0) { // Don't process dead nodes 1101 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 1102 if (addp2 != NULL) { 1103 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1104 alloc_worklist.append_if_missing(addp2); 1105 } 1106 alloc_worklist.append_if_missing(use); 1107 } 1108 } 1109 1110 // An allocation may have an Initialize which has raw stores. Scan 1111 // the users of the raw allocation result and push AddP users 1112 // on alloc_worklist. 1113 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 1114 assert (raw_result != NULL, "must have an allocation result"); 1115 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 1116 Node *use = raw_result->fast_out(i); 1117 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 1118 Node* addp2 = find_second_addp(use, raw_result); 1119 if (addp2 != NULL) { 1120 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1121 alloc_worklist.append_if_missing(addp2); 1122 } 1123 alloc_worklist.append_if_missing(use); 1124 } else if (use->is_MemBar()) { 1125 memnode_worklist.append_if_missing(use); 1126 } 1127 } 1128 } 1129 } else if (n->is_AddP()) { 1130 ptset.Clear(); 1131 PointsTo(ptset, get_addp_base(n)); 1132 assert(ptset.Size() == 1, "AddP address is unique"); 1133 uint elem = ptset.getelem(); // Allocation node's index 1134 if (elem == _phantom_object) { 1135 assert(false, "escaped allocation"); 1136 continue; // Assume the value was set outside this method. 1137 } 1138 Node *base = get_map(elem); // CheckCastPP node 1139 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path 1140 tinst = igvn->type(base)->isa_oopptr(); 1141 } else if (n->is_Phi() || 1142 n->is_CheckCastPP() || 1143 n->is_EncodeP() || 1144 n->is_DecodeN() || 1145 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 1146 if (visited.test_set(n->_idx)) { 1147 assert(n->is_Phi(), "loops only through Phi's"); 1148 continue; // already processed 1149 } 1150 ptset.Clear(); 1151 PointsTo(ptset, n); 1152 if (ptset.Size() == 1) { 1153 uint elem = ptset.getelem(); // Allocation node's index 1154 if (elem == _phantom_object) { 1155 assert(false, "escaped allocation"); 1156 continue; // Assume the value was set outside this method. 1157 } 1158 Node *val = get_map(elem); // CheckCastPP node 1159 TypeNode *tn = n->as_Type(); 1160 tinst = igvn->type(val)->isa_oopptr(); 1161 assert(tinst != NULL && tinst->is_known_instance() && 1162 (uint)tinst->instance_id() == elem , "instance type expected."); 1163 1164 const Type *tn_type = igvn->type(tn); 1165 const TypeOopPtr *tn_t; 1166 if (tn_type->isa_narrowoop()) { 1167 tn_t = tn_type->make_ptr()->isa_oopptr(); 1168 } else { 1169 tn_t = tn_type->isa_oopptr(); 1170 } 1171 1172 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 1173 if (tn_type->isa_narrowoop()) { 1174 tn_type = tinst->make_narrowoop(); 1175 } else { 1176 tn_type = tinst; 1177 } 1178 igvn->hash_delete(tn); 1179 igvn->set_type(tn, tn_type); 1180 tn->set_type(tn_type); 1181 igvn->hash_insert(tn); 1182 record_for_optimizer(n); 1183 } else { 1184 assert(tn_type == TypePtr::NULL_PTR || 1185 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 1186 "unexpected type"); 1187 continue; // Skip dead path with different type 1188 } 1189 } 1190 } else { 1191 debug_only(n->dump();) 1192 assert(false, "EA: unexpected node"); 1193 continue; 1194 } 1195 // push allocation's users on appropriate worklist 1196 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1197 Node *use = n->fast_out(i); 1198 if(use->is_Mem() && use->in(MemNode::Address) == n) { 1199 // Load/store to instance's field 1200 memnode_worklist.append_if_missing(use); 1201 } else if (use->is_MemBar()) { 1202 memnode_worklist.append_if_missing(use); 1203 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 1204 Node* addp2 = find_second_addp(use, n); 1205 if (addp2 != NULL) { 1206 alloc_worklist.append_if_missing(addp2); 1207 } 1208 alloc_worklist.append_if_missing(use); 1209 } else if (use->is_Phi() || 1210 use->is_CheckCastPP() || 1211 use->is_EncodeP() || 1212 use->is_DecodeN() || 1213 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 1214 alloc_worklist.append_if_missing(use); 1215 #ifdef ASSERT 1216 } else if (use->is_Mem()) { 1217 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 1218 } else if (use->is_MergeMem()) { 1219 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1220 } else if (use->is_SafePoint()) { 1221 // Look for MergeMem nodes for calls which reference unique allocation 1222 // (through CheckCastPP nodes) even for debug info. 1223 Node* m = use->in(TypeFunc::Memory); 1224 if (m->is_MergeMem()) { 1225 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1226 } 1227 } else { 1228 uint op = use->Opcode(); 1229 if (!(op == Op_CmpP || op == Op_Conv2B || 1230 op == Op_CastP2X || op == Op_StoreCM || 1231 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 1232 op == Op_StrEquals || op == Op_StrIndexOf)) { 1233 n->dump(); 1234 use->dump(); 1235 assert(false, "EA: missing allocation reference path"); 1236 } 1237 #endif 1238 } 1239 } 1240 1241 } 1242 // New alias types were created in split_AddP(). 1243 uint new_index_end = (uint) _compile->num_alias_types(); 1244 1245 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 1246 // compute new values for Memory inputs (the Memory inputs are not 1247 // actually updated until phase 4.) 1248 if (memnode_worklist.length() == 0) 1249 return; // nothing to do 1250 1251 while (memnode_worklist.length() != 0) { 1252 Node *n = memnode_worklist.pop(); 1253 if (visited.test_set(n->_idx)) 1254 continue; 1255 if (n->is_Phi() || n->is_ClearArray()) { 1256 // we don't need to do anything, but the users must be pushed 1257 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 1258 // we don't need to do anything, but the users must be pushed 1259 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 1260 if (n == NULL) 1261 continue; 1262 } else { 1263 assert(n->is_Mem(), "memory node required."); 1264 Node *addr = n->in(MemNode::Address); 1265 const Type *addr_t = igvn->type(addr); 1266 if (addr_t == Type::TOP) 1267 continue; 1268 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 1269 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 1270 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 1271 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); 1272 if (_compile->failing()) { 1273 return; 1274 } 1275 if (mem != n->in(MemNode::Memory)) { 1276 // We delay the memory edge update since we need old one in 1277 // MergeMem code below when instances memory slices are separated. 1278 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;) 1279 assert(pn == NULL || pn == n, "wrong node"); 1280 set_map(n->_idx, mem); 1281 ptnode_adr(n->_idx)->_node = n; 1282 } 1283 if (n->is_Load()) { 1284 continue; // don't push users 1285 } else if (n->is_LoadStore()) { 1286 // get the memory projection 1287 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1288 Node *use = n->fast_out(i); 1289 if (use->Opcode() == Op_SCMemProj) { 1290 n = use; 1291 break; 1292 } 1293 } 1294 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 1295 } 1296 } 1297 // push user on appropriate worklist 1298 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1299 Node *use = n->fast_out(i); 1300 if (use->is_Phi() || use->is_ClearArray()) { 1301 memnode_worklist.append_if_missing(use); 1302 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { 1303 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 1304 continue; 1305 memnode_worklist.append_if_missing(use); 1306 } else if (use->is_MemBar()) { 1307 memnode_worklist.append_if_missing(use); 1308 #ifdef ASSERT 1309 } else if(use->is_Mem()) { 1310 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 1311 } else if (use->is_MergeMem()) { 1312 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1313 } else { 1314 uint op = use->Opcode(); 1315 if (!(op == Op_StoreCM || 1316 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 1317 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 1318 op == Op_AryEq || op == Op_StrComp || 1319 op == Op_StrEquals || op == Op_StrIndexOf)) { 1320 n->dump(); 1321 use->dump(); 1322 assert(false, "EA: missing memory path"); 1323 } 1324 #endif 1325 } 1326 } 1327 } 1328 1329 // Phase 3: Process MergeMem nodes from mergemem_worklist. 1330 // Walk each memory slice moving the first node encountered of each 1331 // instance type to the the input corresponding to its alias index. 1332 uint length = _mergemem_worklist.length(); 1333 for( uint next = 0; next < length; ++next ) { 1334 MergeMemNode* nmm = _mergemem_worklist.at(next); 1335 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 1336 // Note: we don't want to use MergeMemStream here because we only want to 1337 // scan inputs which exist at the start, not ones we add during processing. 1338 // Note 2: MergeMem may already contains instance memory slices added 1339 // during find_inst_mem() call when memory nodes were processed above. 1340 igvn->hash_delete(nmm); 1341 uint nslices = nmm->req(); 1342 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 1343 Node* mem = nmm->in(i); 1344 Node* cur = NULL; 1345 if (mem == NULL || mem->is_top()) 1346 continue; 1347 // First, update mergemem by moving memory nodes to corresponding slices 1348 // if their type became more precise since this mergemem was created. 1349 while (mem->is_Mem()) { 1350 const Type *at = igvn->type(mem->in(MemNode::Address)); 1351 if (at != Type::TOP) { 1352 assert (at->isa_ptr() != NULL, "pointer type required."); 1353 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 1354 if (idx == i) { 1355 if (cur == NULL) 1356 cur = mem; 1357 } else { 1358 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 1359 nmm->set_memory_at(idx, mem); 1360 } 1361 } 1362 } 1363 mem = mem->in(MemNode::Memory); 1364 } 1365 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 1366 // Find any instance of the current type if we haven't encountered 1367 // already a memory slice of the instance along the memory chain. 1368 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1369 if((uint)_compile->get_general_index(ni) == i) { 1370 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 1371 if (nmm->is_empty_memory(m)) { 1372 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); 1373 if (_compile->failing()) { 1374 return; 1375 } 1376 nmm->set_memory_at(ni, result); 1377 } 1378 } 1379 } 1380 } 1381 // Find the rest of instances values 1382 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1383 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 1384 Node* result = step_through_mergemem(nmm, ni, tinst); 1385 if (result == nmm->base_memory()) { 1386 // Didn't find instance memory, search through general slice recursively. 1387 result = nmm->memory_at(_compile->get_general_index(ni)); 1388 result = find_inst_mem(result, ni, orig_phis, igvn); 1389 if (_compile->failing()) { 1390 return; 1391 } 1392 nmm->set_memory_at(ni, result); 1393 } 1394 } 1395 igvn->hash_insert(nmm); 1396 record_for_optimizer(nmm); 1397 } 1398 1399 // Phase 4: Update the inputs of non-instance memory Phis and 1400 // the Memory input of memnodes 1401 // First update the inputs of any non-instance Phi's from 1402 // which we split out an instance Phi. Note we don't have 1403 // to recursively process Phi's encounted on the input memory 1404 // chains as is done in split_memory_phi() since they will 1405 // also be processed here. 1406 for (int j = 0; j < orig_phis.length(); j++) { 1407 PhiNode *phi = orig_phis.at(j); 1408 int alias_idx = _compile->get_alias_index(phi->adr_type()); 1409 igvn->hash_delete(phi); 1410 for (uint i = 1; i < phi->req(); i++) { 1411 Node *mem = phi->in(i); 1412 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); 1413 if (_compile->failing()) { 1414 return; 1415 } 1416 if (mem != new_mem) { 1417 phi->set_req(i, new_mem); 1418 } 1419 } 1420 igvn->hash_insert(phi); 1421 record_for_optimizer(phi); 1422 } 1423 1424 // Update the memory inputs of MemNodes with the value we computed 1425 // in Phase 2 and move stores memory users to corresponding memory slices. 1426 #ifdef ASSERT 1427 visited.Clear(); 1428 Node_Stack old_mems(arena, _compile->unique() >> 2); 1429 #endif 1430 for (uint i = 0; i < nodes_size(); i++) { 1431 Node *nmem = get_map(i); 1432 if (nmem != NULL) { 1433 Node *n = ptnode_adr(i)->_node; 1434 assert(n != NULL, "sanity"); 1435 if (n->is_Mem()) { 1436 #ifdef ASSERT 1437 Node* old_mem = n->in(MemNode::Memory); 1438 if (!visited.test_set(old_mem->_idx)) { 1439 old_mems.push(old_mem, old_mem->outcnt()); 1440 } 1441 #endif 1442 assert(n->in(MemNode::Memory) != nmem, "sanity"); 1443 if (!n->is_Load()) { 1444 // Move memory users of a store first. 1445 move_inst_mem(n, orig_phis, igvn); 1446 } 1447 // Now update memory input 1448 igvn->hash_delete(n); 1449 n->set_req(MemNode::Memory, nmem); 1450 igvn->hash_insert(n); 1451 record_for_optimizer(n); 1452 } else { 1453 assert(n->is_Allocate() || n->is_CheckCastPP() || 1454 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 1455 } 1456 } 1457 } 1458 #ifdef ASSERT 1459 // Verify that memory was split correctly 1460 while (old_mems.is_nonempty()) { 1461 Node* old_mem = old_mems.node(); 1462 uint old_cnt = old_mems.index(); 1463 old_mems.pop(); 1464 assert(old_cnt = old_mem->outcnt(), "old mem could be lost"); 1465 } 1466 #endif 1467 } 1468 1469 bool ConnectionGraph::has_candidates(Compile *C) { 1470 // EA brings benefits only when the code has allocations and/or locks which 1471 // are represented by ideal Macro nodes. 1472 int cnt = C->macro_count(); 1473 for( int i=0; i < cnt; i++ ) { 1474 Node *n = C->macro_node(i); 1475 if ( n->is_Allocate() ) 1476 return true; 1477 if( n->is_Lock() ) { 1478 Node* obj = n->as_Lock()->obj_node()->uncast(); 1479 if( !(obj->is_Parm() || obj->is_Con()) ) 1480 return true; 1481 } 1482 } 1483 return false; 1484 } 1485 1486 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 1487 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 1488 // to create space for them in ConnectionGraph::_nodes[]. 1489 Node* oop_null = igvn->zerocon(T_OBJECT); 1490 Node* noop_null = igvn->zerocon(T_NARROWOOP); 1491 1492 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 1493 // Perform escape analysis 1494 if (congraph->compute_escape()) { 1495 // There are non escaping objects. 1496 C->set_congraph(congraph); 1497 } 1498 1499 // Cleanup. 1500 if (oop_null->outcnt() == 0) 1501 igvn->hash_delete(oop_null); 1502 if (noop_null->outcnt() == 0) 1503 igvn->hash_delete(noop_null); 1504 } 1505 1506 bool ConnectionGraph::compute_escape() { 1507 Compile* C = _compile; 1508 1509 // 1. Populate Connection Graph (CG) with Ideal nodes. 1510 1511 Unique_Node_List worklist_init; 1512 worklist_init.map(C->unique(), NULL); // preallocate space 1513 1514 // Initialize worklist 1515 if (C->root() != NULL) { 1516 worklist_init.push(C->root()); 1517 } 1518 1519 GrowableArray<int> cg_worklist; 1520 PhaseGVN* igvn = _igvn; 1521 bool has_allocations = false; 1522 1523 // Push all useful nodes onto CG list and set their type. 1524 for( uint next = 0; next < worklist_init.size(); ++next ) { 1525 Node* n = worklist_init.at(next); 1526 record_for_escape_analysis(n, igvn); 1527 // Only allocations and java static calls results are checked 1528 // for an escape status. See process_call_result() below. 1529 if (n->is_Allocate() || n->is_CallStaticJava() && 1530 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { 1531 has_allocations = true; 1532 } 1533 if(n->is_AddP()) { 1534 // Collect address nodes which directly reference an allocation. 1535 // Use them during stage 3 below to build initial connection graph 1536 // field edges. Other field edges could be added after StoreP/LoadP 1537 // nodes are processed during stage 4 below. 1538 Node* base = get_addp_base(n); 1539 if(base->is_Proj() && base->in(0)->is_Allocate()) { 1540 cg_worklist.append(n->_idx); 1541 } 1542 } else if (n->is_MergeMem()) { 1543 // Collect all MergeMem nodes to add memory slices for 1544 // scalar replaceable objects in split_unique_types(). 1545 _mergemem_worklist.append(n->as_MergeMem()); 1546 } 1547 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1548 Node* m = n->fast_out(i); // Get user 1549 worklist_init.push(m); 1550 } 1551 } 1552 1553 if (!has_allocations) { 1554 _collecting = false; 1555 return false; // Nothing to do. 1556 } 1557 1558 // 2. First pass to create simple CG edges (doesn't require to walk CG). 1559 uint delayed_size = _delayed_worklist.size(); 1560 for( uint next = 0; next < delayed_size; ++next ) { 1561 Node* n = _delayed_worklist.at(next); 1562 build_connection_graph(n, igvn); 1563 } 1564 1565 // 3. Pass to create fields edges (Allocate -F-> AddP). 1566 uint cg_length = cg_worklist.length(); 1567 for( uint next = 0; next < cg_length; ++next ) { 1568 int ni = cg_worklist.at(next); 1569 build_connection_graph(ptnode_adr(ni)->_node, igvn); 1570 } 1571 1572 cg_worklist.clear(); 1573 cg_worklist.append(_phantom_object); 1574 1575 // 4. Build Connection Graph which need 1576 // to walk the connection graph. 1577 for (uint ni = 0; ni < nodes_size(); ni++) { 1578 PointsToNode* ptn = ptnode_adr(ni); 1579 Node *n = ptn->_node; 1580 if (n != NULL) { // Call, AddP, LoadP, StoreP 1581 build_connection_graph(n, igvn); 1582 if (ptn->node_type() != PointsToNode::UnknownType) 1583 cg_worklist.append(n->_idx); // Collect CG nodes 1584 } 1585 } 1586 1587 Arena* arena = Thread::current()->resource_area(); 1588 VectorSet ptset(arena); 1589 GrowableArray<uint> deferred_edges; 1590 VectorSet visited(arena); 1591 1592 // 5. Remove deferred edges from the graph and adjust 1593 // escape state of nonescaping objects. 1594 cg_length = cg_worklist.length(); 1595 for( uint next = 0; next < cg_length; ++next ) { 1596 int ni = cg_worklist.at(next); 1597 PointsToNode* ptn = ptnode_adr(ni); 1598 PointsToNode::NodeType nt = ptn->node_type(); 1599 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { 1600 remove_deferred(ni, &deferred_edges, &visited); 1601 Node *n = ptn->_node; 1602 if (n->is_AddP()) { 1603 // Search for objects which are not scalar replaceable 1604 // and adjust their escape state. 1605 verify_escape_state(ni, ptset, igvn); 1606 } 1607 } 1608 } 1609 1610 // 6. Propagate escape states. 1611 GrowableArray<int> worklist; 1612 bool has_non_escaping_obj = false; 1613 1614 // push all GlobalEscape nodes on the worklist 1615 for( uint next = 0; next < cg_length; ++next ) { 1616 int nk = cg_worklist.at(next); 1617 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) 1618 worklist.push(nk); 1619 } 1620 // mark all nodes reachable from GlobalEscape nodes 1621 while(worklist.length() > 0) { 1622 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1623 uint e_cnt = ptn->edge_count(); 1624 for (uint ei = 0; ei < e_cnt; ei++) { 1625 uint npi = ptn->edge_target(ei); 1626 PointsToNode *np = ptnode_adr(npi); 1627 if (np->escape_state() < PointsToNode::GlobalEscape) { 1628 np->set_escape_state(PointsToNode::GlobalEscape); 1629 worklist.push(npi); 1630 } 1631 } 1632 } 1633 1634 // push all ArgEscape nodes on the worklist 1635 for( uint next = 0; next < cg_length; ++next ) { 1636 int nk = cg_worklist.at(next); 1637 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) 1638 worklist.push(nk); 1639 } 1640 // mark all nodes reachable from ArgEscape nodes 1641 while(worklist.length() > 0) { 1642 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1643 if (ptn->node_type() == PointsToNode::JavaObject) 1644 has_non_escaping_obj = true; // Non GlobalEscape 1645 uint e_cnt = ptn->edge_count(); 1646 for (uint ei = 0; ei < e_cnt; ei++) { 1647 uint npi = ptn->edge_target(ei); 1648 PointsToNode *np = ptnode_adr(npi); 1649 if (np->escape_state() < PointsToNode::ArgEscape) { 1650 np->set_escape_state(PointsToNode::ArgEscape); 1651 worklist.push(npi); 1652 } 1653 } 1654 } 1655 1656 GrowableArray<Node*> alloc_worklist; 1657 1658 // push all NoEscape nodes on the worklist 1659 for( uint next = 0; next < cg_length; ++next ) { 1660 int nk = cg_worklist.at(next); 1661 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) 1662 worklist.push(nk); 1663 } 1664 // mark all nodes reachable from NoEscape nodes 1665 while(worklist.length() > 0) { 1666 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1667 if (ptn->node_type() == PointsToNode::JavaObject) 1668 has_non_escaping_obj = true; // Non GlobalEscape 1669 Node* n = ptn->_node; 1670 if (n->is_Allocate() && ptn->_scalar_replaceable ) { 1671 // Push scalar replaceable allocations on alloc_worklist 1672 // for processing in split_unique_types(). 1673 alloc_worklist.append(n); 1674 } 1675 uint e_cnt = ptn->edge_count(); 1676 for (uint ei = 0; ei < e_cnt; ei++) { 1677 uint npi = ptn->edge_target(ei); 1678 PointsToNode *np = ptnode_adr(npi); 1679 if (np->escape_state() < PointsToNode::NoEscape) { 1680 np->set_escape_state(PointsToNode::NoEscape); 1681 worklist.push(npi); 1682 } 1683 } 1684 } 1685 1686 _collecting = false; 1687 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); 1688 1689 #ifndef PRODUCT 1690 if (PrintEscapeAnalysis) { 1691 dump(); // Dump ConnectionGraph 1692 } 1693 #endif 1694 1695 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; 1696 if ( has_scalar_replaceable_candidates && 1697 C->AliasLevel() >= 3 && EliminateAllocations ) { 1698 1699 // Now use the escape information to create unique types for 1700 // scalar replaceable objects. 1701 split_unique_types(alloc_worklist); 1702 1703 if (C->failing()) return false; 1704 1705 C->print_method("After Escape Analysis", 2); 1706 1707 #ifdef ASSERT 1708 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 1709 tty->print("=== No allocations eliminated for "); 1710 C->method()->print_short_name(); 1711 if(!EliminateAllocations) { 1712 tty->print(" since EliminateAllocations is off ==="); 1713 } else if(!has_scalar_replaceable_candidates) { 1714 tty->print(" since there are no scalar replaceable candidates ==="); 1715 } else if(C->AliasLevel() < 3) { 1716 tty->print(" since AliasLevel < 3 ==="); 1717 } 1718 tty->cr(); 1719 #endif 1720 } 1721 return has_non_escaping_obj; 1722 } 1723 1724 // Search for objects which are not scalar replaceable. 1725 void ConnectionGraph::verify_escape_state(int nidx, VectorSet& ptset, PhaseTransform* phase) { 1726 PointsToNode* ptn = ptnode_adr(nidx); 1727 Node* n = ptn->_node; 1728 assert(n->is_AddP(), "Should be called for AddP nodes only"); 1729 // Search for objects which are not scalar replaceable. 1730 // Mark their escape state as ArgEscape to propagate the state 1731 // to referenced objects. 1732 // Note: currently there are no difference in compiler optimizations 1733 // for ArgEscape objects and NoEscape objects which are not 1734 // scalar replaceable. 1735 1736 Compile* C = _compile; 1737 1738 int offset = ptn->offset(); 1739 Node* base = get_addp_base(n); 1740 ptset.Clear(); 1741 PointsTo(ptset, base); 1742 int ptset_size = ptset.Size(); 1743 1744 // Check if a oop field's initializing value is recorded and add 1745 // a corresponding NULL field's value if it is not recorded. 1746 // Connection Graph does not record a default initialization by NULL 1747 // captured by Initialize node. 1748 // 1749 // Note: it will disable scalar replacement in some cases: 1750 // 1751 // Point p[] = new Point[1]; 1752 // p[0] = new Point(); // Will be not scalar replaced 1753 // 1754 // but it will save us from incorrect optimizations in next cases: 1755 // 1756 // Point p[] = new Point[1]; 1757 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1758 // 1759 // Do a simple control flow analysis to distinguish above cases. 1760 // 1761 if (offset != Type::OffsetBot && ptset_size == 1) { 1762 uint elem = ptset.getelem(); // Allocation node's index 1763 // It does not matter if it is not Allocation node since 1764 // only non-escaping allocations are scalar replaced. 1765 if (ptnode_adr(elem)->_node->is_Allocate() && 1766 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) { 1767 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate(); 1768 InitializeNode* ini = alloc->initialization(); 1769 1770 // Check only oop fields. 1771 const Type* adr_type = n->as_AddP()->bottom_type(); 1772 BasicType basic_field_type = T_INT; 1773 if (adr_type->isa_instptr()) { 1774 ciField* field = C->alias_type(adr_type->isa_instptr())->field(); 1775 if (field != NULL) { 1776 basic_field_type = field->layout_type(); 1777 } else { 1778 // Ignore non field load (for example, klass load) 1779 } 1780 } else if (adr_type->isa_aryptr()) { 1781 const Type* elemtype = adr_type->isa_aryptr()->elem(); 1782 basic_field_type = elemtype->array_element_basic_type(); 1783 } else { 1784 // Raw pointers are used for initializing stores so skip it. 1785 assert(adr_type->isa_rawptr() && base->is_Proj() && 1786 (base->in(0) == alloc),"unexpected pointer type"); 1787 } 1788 if (basic_field_type == T_OBJECT || 1789 basic_field_type == T_NARROWOOP || 1790 basic_field_type == T_ARRAY) { 1791 Node* value = NULL; 1792 if (ini != NULL) { 1793 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT; 1794 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase); 1795 if (store != NULL && store->is_Store()) { 1796 value = store->in(MemNode::ValueIn); 1797 } else if (ptn->edge_count() > 0) { // Are there oop stores? 1798 // Check for a store which follows allocation without branches. 1799 // For example, a volatile field store is not collected 1800 // by Initialize node. TODO: it would be nice to use idom() here. 1801 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1802 store = n->fast_out(i); 1803 if (store->is_Store() && store->in(0) != NULL) { 1804 Node* ctrl = store->in(0); 1805 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL || 1806 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() || 1807 ctrl->is_IfTrue() || ctrl->is_IfFalse())) { 1808 ctrl = ctrl->in(0); 1809 } 1810 if (ctrl == ini || ctrl == alloc) { 1811 value = store->in(MemNode::ValueIn); 1812 break; 1813 } 1814 } 1815 } 1816 } 1817 } 1818 if (value == NULL || value != ptnode_adr(value->_idx)->_node) { 1819 // A field's initializing value was not recorded. Add NULL. 1820 uint null_idx = UseCompressedOops ? _noop_null : _oop_null; 1821 add_pointsto_edge(nidx, null_idx); 1822 } 1823 } 1824 } 1825 } 1826 1827 // An object is not scalar replaceable if the field which may point 1828 // to it has unknown offset (unknown element of an array of objects). 1829 // 1830 if (offset == Type::OffsetBot) { 1831 uint e_cnt = ptn->edge_count(); 1832 for (uint ei = 0; ei < e_cnt; ei++) { 1833 uint npi = ptn->edge_target(ei); 1834 set_escape_state(npi, PointsToNode::ArgEscape); 1835 ptnode_adr(npi)->_scalar_replaceable = false; 1836 } 1837 } 1838 1839 // Currently an object is not scalar replaceable if a LoadStore node 1840 // access its field since the field value is unknown after it. 1841 // 1842 bool has_LoadStore = false; 1843 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1844 Node *use = n->fast_out(i); 1845 if (use->is_LoadStore()) { 1846 has_LoadStore = true; 1847 break; 1848 } 1849 } 1850 // An object is not scalar replaceable if the address points 1851 // to unknown field (unknown element for arrays, offset is OffsetBot). 1852 // 1853 // Or the address may point to more then one object. This may produce 1854 // the false positive result (set scalar_replaceable to false) 1855 // since the flow-insensitive escape analysis can't separate 1856 // the case when stores overwrite the field's value from the case 1857 // when stores happened on different control branches. 1858 // 1859 if (ptset_size > 1 || ptset_size != 0 && 1860 (has_LoadStore || offset == Type::OffsetBot)) { 1861 for( VectorSetI j(&ptset); j.test(); ++j ) { 1862 set_escape_state(j.elem, PointsToNode::ArgEscape); 1863 ptnode_adr(j.elem)->_scalar_replaceable = false; 1864 } 1865 } 1866 } 1867 1868 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { 1869 1870 switch (call->Opcode()) { 1871 #ifdef ASSERT 1872 case Op_Allocate: 1873 case Op_AllocateArray: 1874 case Op_Lock: 1875 case Op_Unlock: 1876 assert(false, "should be done already"); 1877 break; 1878 #endif 1879 case Op_CallLeaf: 1880 case Op_CallLeafNoFP: 1881 { 1882 // Stub calls, objects do not escape but they are not scale replaceable. 1883 // Adjust escape state for outgoing arguments. 1884 const TypeTuple * d = call->tf()->domain(); 1885 VectorSet ptset(Thread::current()->resource_area()); 1886 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1887 const Type* at = d->field_at(i); 1888 Node *arg = call->in(i)->uncast(); 1889 const Type *aat = phase->type(arg); 1890 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() && 1891 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) { 1892 1893 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 1894 aat->isa_ptr() != NULL, "expecting an Ptr"); 1895 #ifdef ASSERT 1896 if (!(call->Opcode() == Op_CallLeafNoFP && 1897 call->as_CallLeaf()->_name != NULL && 1898 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) || 1899 call->as_CallLeaf()->_name != NULL && 1900 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 1901 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 )) 1902 ) { 1903 call->dump(); 1904 assert(false, "EA: unexpected CallLeaf"); 1905 } 1906 #endif 1907 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1908 if (arg->is_AddP()) { 1909 // 1910 // The inline_native_clone() case when the arraycopy stub is called 1911 // after the allocation before Initialize and CheckCastPP nodes. 1912 // 1913 // Set AddP's base (Allocate) as not scalar replaceable since 1914 // pointer to the base (with offset) is passed as argument. 1915 // 1916 arg = get_addp_base(arg); 1917 } 1918 ptset.Clear(); 1919 PointsTo(ptset, arg); 1920 for( VectorSetI j(&ptset); j.test(); ++j ) { 1921 uint pt = j.elem; 1922 set_escape_state(pt, PointsToNode::ArgEscape); 1923 } 1924 } 1925 } 1926 break; 1927 } 1928 1929 case Op_CallStaticJava: 1930 // For a static call, we know exactly what method is being called. 1931 // Use bytecode estimator to record the call's escape affects 1932 { 1933 ciMethod *meth = call->as_CallJava()->method(); 1934 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1935 // fall-through if not a Java method or no analyzer information 1936 if (call_analyzer != NULL) { 1937 const TypeTuple * d = call->tf()->domain(); 1938 VectorSet ptset(Thread::current()->resource_area()); 1939 bool copy_dependencies = false; 1940 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1941 const Type* at = d->field_at(i); 1942 int k = i - TypeFunc::Parms; 1943 Node *arg = call->in(i)->uncast(); 1944 1945 if (at->isa_oopptr() != NULL && 1946 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) { 1947 1948 bool global_escapes = false; 1949 bool fields_escapes = false; 1950 if (!call_analyzer->is_arg_stack(k)) { 1951 // The argument global escapes, mark everything it could point to 1952 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1953 global_escapes = true; 1954 } else { 1955 if (!call_analyzer->is_arg_local(k)) { 1956 // The argument itself doesn't escape, but any fields might 1957 fields_escapes = true; 1958 } 1959 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1960 copy_dependencies = true; 1961 } 1962 1963 ptset.Clear(); 1964 PointsTo(ptset, arg); 1965 for( VectorSetI j(&ptset); j.test(); ++j ) { 1966 uint pt = j.elem; 1967 if (global_escapes) { 1968 //The argument global escapes, mark everything it could point to 1969 set_escape_state(pt, PointsToNode::GlobalEscape); 1970 } else { 1971 if (fields_escapes) { 1972 // The argument itself doesn't escape, but any fields might 1973 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); 1974 } 1975 set_escape_state(pt, PointsToNode::ArgEscape); 1976 } 1977 } 1978 } 1979 } 1980 if (copy_dependencies) 1981 call_analyzer->copy_dependencies(_compile->dependencies()); 1982 break; 1983 } 1984 } 1985 1986 default: 1987 // Fall-through here if not a Java method or no analyzer information 1988 // or some other type of call, assume the worst case: all arguments 1989 // globally escape. 1990 { 1991 // adjust escape state for outgoing arguments 1992 const TypeTuple * d = call->tf()->domain(); 1993 VectorSet ptset(Thread::current()->resource_area()); 1994 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1995 const Type* at = d->field_at(i); 1996 if (at->isa_oopptr() != NULL) { 1997 Node *arg = call->in(i)->uncast(); 1998 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1999 ptset.Clear(); 2000 PointsTo(ptset, arg); 2001 for( VectorSetI j(&ptset); j.test(); ++j ) { 2002 uint pt = j.elem; 2003 set_escape_state(pt, PointsToNode::GlobalEscape); 2004 } 2005 } 2006 } 2007 } 2008 } 2009 } 2010 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { 2011 CallNode *call = resproj->in(0)->as_Call(); 2012 uint call_idx = call->_idx; 2013 uint resproj_idx = resproj->_idx; 2014 2015 switch (call->Opcode()) { 2016 case Op_Allocate: 2017 { 2018 Node *k = call->in(AllocateNode::KlassNode); 2019 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2020 assert(kt != NULL, "TypeKlassPtr required."); 2021 ciKlass* cik = kt->klass(); 2022 2023 PointsToNode::EscapeState es; 2024 uint edge_to; 2025 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 2026 !cik->is_instance_klass() || // StressReflectiveCode 2027 cik->as_instance_klass()->has_finalizer()) { 2028 es = PointsToNode::GlobalEscape; 2029 edge_to = _phantom_object; // Could not be worse 2030 } else { 2031 es = PointsToNode::NoEscape; 2032 edge_to = call_idx; 2033 } 2034 set_escape_state(call_idx, es); 2035 add_pointsto_edge(resproj_idx, edge_to); 2036 _processed.set(resproj_idx); 2037 break; 2038 } 2039 2040 case Op_AllocateArray: 2041 { 2042 2043 Node *k = call->in(AllocateNode::KlassNode); 2044 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2045 assert(kt != NULL, "TypeKlassPtr required."); 2046 ciKlass* cik = kt->klass(); 2047 2048 PointsToNode::EscapeState es; 2049 uint edge_to; 2050 if (!cik->is_array_klass()) { // StressReflectiveCode 2051 es = PointsToNode::GlobalEscape; 2052 edge_to = _phantom_object; 2053 } else { 2054 es = PointsToNode::NoEscape; 2055 edge_to = call_idx; 2056 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 2057 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 2058 // Not scalar replaceable if the length is not constant or too big. 2059 ptnode_adr(call_idx)->_scalar_replaceable = false; 2060 } 2061 } 2062 set_escape_state(call_idx, es); 2063 add_pointsto_edge(resproj_idx, edge_to); 2064 _processed.set(resproj_idx); 2065 break; 2066 } 2067 2068 case Op_CallStaticJava: 2069 // For a static call, we know exactly what method is being called. 2070 // Use bytecode estimator to record whether the call's return value escapes 2071 { 2072 bool done = true; 2073 const TypeTuple *r = call->tf()->range(); 2074 const Type* ret_type = NULL; 2075 2076 if (r->cnt() > TypeFunc::Parms) 2077 ret_type = r->field_at(TypeFunc::Parms); 2078 2079 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2080 // _multianewarray functions return a TypeRawPtr. 2081 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { 2082 _processed.set(resproj_idx); 2083 break; // doesn't return a pointer type 2084 } 2085 ciMethod *meth = call->as_CallJava()->method(); 2086 const TypeTuple * d = call->tf()->domain(); 2087 if (meth == NULL) { 2088 // not a Java method, assume global escape 2089 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2090 add_pointsto_edge(resproj_idx, _phantom_object); 2091 } else { 2092 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); 2093 bool copy_dependencies = false; 2094 2095 if (call_analyzer->is_return_allocated()) { 2096 // Returns a newly allocated unescaped object, simply 2097 // update dependency information. 2098 // Mark it as NoEscape so that objects referenced by 2099 // it's fields will be marked as NoEscape at least. 2100 set_escape_state(call_idx, PointsToNode::NoEscape); 2101 add_pointsto_edge(resproj_idx, call_idx); 2102 copy_dependencies = true; 2103 } else if (call_analyzer->is_return_local()) { 2104 // determine whether any arguments are returned 2105 set_escape_state(call_idx, PointsToNode::NoEscape); 2106 bool ret_arg = false; 2107 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2108 const Type* at = d->field_at(i); 2109 2110 if (at->isa_oopptr() != NULL) { 2111 Node *arg = call->in(i)->uncast(); 2112 2113 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 2114 ret_arg = true; 2115 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2116 if (arg_esp->node_type() == PointsToNode::UnknownType) 2117 done = false; 2118 else if (arg_esp->node_type() == PointsToNode::JavaObject) 2119 add_pointsto_edge(resproj_idx, arg->_idx); 2120 else 2121 add_deferred_edge(resproj_idx, arg->_idx); 2122 arg_esp->_hidden_alias = true; 2123 } 2124 } 2125 } 2126 if (done && !ret_arg) { 2127 // Returns unknown object. 2128 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2129 add_pointsto_edge(resproj_idx, _phantom_object); 2130 } 2131 copy_dependencies = true; 2132 } else { 2133 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2134 add_pointsto_edge(resproj_idx, _phantom_object); 2135 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2136 const Type* at = d->field_at(i); 2137 if (at->isa_oopptr() != NULL) { 2138 Node *arg = call->in(i)->uncast(); 2139 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2140 arg_esp->_hidden_alias = true; 2141 } 2142 } 2143 } 2144 if (copy_dependencies) 2145 call_analyzer->copy_dependencies(_compile->dependencies()); 2146 } 2147 if (done) 2148 _processed.set(resproj_idx); 2149 break; 2150 } 2151 2152 default: 2153 // Some other type of call, assume the worst case that the 2154 // returned value, if any, globally escapes. 2155 { 2156 const TypeTuple *r = call->tf()->range(); 2157 if (r->cnt() > TypeFunc::Parms) { 2158 const Type* ret_type = r->field_at(TypeFunc::Parms); 2159 2160 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2161 // _multianewarray functions return a TypeRawPtr. 2162 if (ret_type->isa_ptr() != NULL) { 2163 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2164 add_pointsto_edge(resproj_idx, _phantom_object); 2165 } 2166 } 2167 _processed.set(resproj_idx); 2168 } 2169 } 2170 } 2171 2172 // Populate Connection Graph with Ideal nodes and create simple 2173 // connection graph edges (do not need to check the node_type of inputs 2174 // or to call PointsTo() to walk the connection graph). 2175 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { 2176 if (_processed.test(n->_idx)) 2177 return; // No need to redefine node's state. 2178 2179 if (n->is_Call()) { 2180 // Arguments to allocation and locking don't escape. 2181 if (n->is_Allocate()) { 2182 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); 2183 record_for_optimizer(n); 2184 } else if (n->is_Lock() || n->is_Unlock()) { 2185 // Put Lock and Unlock nodes on IGVN worklist to process them during 2186 // the first IGVN optimization when escape information is still available. 2187 record_for_optimizer(n); 2188 _processed.set(n->_idx); 2189 } else { 2190 // Don't mark as processed since call's arguments have to be processed. 2191 PointsToNode::NodeType nt = PointsToNode::UnknownType; 2192 PointsToNode::EscapeState es = PointsToNode::UnknownEscape; 2193 2194 // Check if a call returns an object. 2195 const TypeTuple *r = n->as_Call()->tf()->range(); 2196 if (r->cnt() > TypeFunc::Parms && 2197 r->field_at(TypeFunc::Parms)->isa_ptr() && 2198 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { 2199 nt = PointsToNode::JavaObject; 2200 if (!n->is_CallStaticJava()) { 2201 // Since the called mathod is statically unknown assume 2202 // the worst case that the returned value globally escapes. 2203 es = PointsToNode::GlobalEscape; 2204 } 2205 } 2206 add_node(n, nt, es, false); 2207 } 2208 return; 2209 } 2210 2211 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 2212 // ThreadLocal has RawPrt type. 2213 switch (n->Opcode()) { 2214 case Op_AddP: 2215 { 2216 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); 2217 break; 2218 } 2219 case Op_CastX2P: 2220 { // "Unsafe" memory access. 2221 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2222 break; 2223 } 2224 case Op_CastPP: 2225 case Op_CheckCastPP: 2226 case Op_EncodeP: 2227 case Op_DecodeN: 2228 { 2229 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2230 int ti = n->in(1)->_idx; 2231 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2232 if (nt == PointsToNode::UnknownType) { 2233 _delayed_worklist.push(n); // Process it later. 2234 break; 2235 } else if (nt == PointsToNode::JavaObject) { 2236 add_pointsto_edge(n->_idx, ti); 2237 } else { 2238 add_deferred_edge(n->_idx, ti); 2239 } 2240 _processed.set(n->_idx); 2241 break; 2242 } 2243 case Op_ConP: 2244 { 2245 // assume all pointer constants globally escape except for null 2246 PointsToNode::EscapeState es; 2247 if (phase->type(n) == TypePtr::NULL_PTR) 2248 es = PointsToNode::NoEscape; 2249 else 2250 es = PointsToNode::GlobalEscape; 2251 2252 add_node(n, PointsToNode::JavaObject, es, true); 2253 break; 2254 } 2255 case Op_ConN: 2256 { 2257 // assume all narrow oop constants globally escape except for null 2258 PointsToNode::EscapeState es; 2259 if (phase->type(n) == TypeNarrowOop::NULL_PTR) 2260 es = PointsToNode::NoEscape; 2261 else 2262 es = PointsToNode::GlobalEscape; 2263 2264 add_node(n, PointsToNode::JavaObject, es, true); 2265 break; 2266 } 2267 case Op_CreateEx: 2268 { 2269 // assume that all exception objects globally escape 2270 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2271 break; 2272 } 2273 case Op_LoadKlass: 2274 case Op_LoadNKlass: 2275 { 2276 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2277 break; 2278 } 2279 case Op_LoadP: 2280 case Op_LoadN: 2281 { 2282 const Type *t = phase->type(n); 2283 if (t->make_ptr() == NULL) { 2284 _processed.set(n->_idx); 2285 return; 2286 } 2287 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2288 break; 2289 } 2290 case Op_Parm: 2291 { 2292 _processed.set(n->_idx); // No need to redefine it state. 2293 uint con = n->as_Proj()->_con; 2294 if (con < TypeFunc::Parms) 2295 return; 2296 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); 2297 if (t->isa_ptr() == NULL) 2298 return; 2299 // We have to assume all input parameters globally escape 2300 // (Note: passing 'false' since _processed is already set). 2301 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); 2302 break; 2303 } 2304 case Op_Phi: 2305 { 2306 const Type *t = n->as_Phi()->type(); 2307 if (t->make_ptr() == NULL) { 2308 // nothing to do if not an oop or narrow oop 2309 _processed.set(n->_idx); 2310 return; 2311 } 2312 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2313 uint i; 2314 for (i = 1; i < n->req() ; i++) { 2315 Node* in = n->in(i); 2316 if (in == NULL) 2317 continue; // ignore NULL 2318 in = in->uncast(); 2319 if (in->is_top() || in == n) 2320 continue; // ignore top or inputs which go back this node 2321 int ti = in->_idx; 2322 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2323 if (nt == PointsToNode::UnknownType) { 2324 break; 2325 } else if (nt == PointsToNode::JavaObject) { 2326 add_pointsto_edge(n->_idx, ti); 2327 } else { 2328 add_deferred_edge(n->_idx, ti); 2329 } 2330 } 2331 if (i >= n->req()) 2332 _processed.set(n->_idx); 2333 else 2334 _delayed_worklist.push(n); 2335 break; 2336 } 2337 case Op_Proj: 2338 { 2339 // we are only interested in the oop result projection from a call 2340 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2341 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2342 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2343 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2344 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2345 int ti = n->in(0)->_idx; 2346 // The call may not be registered yet (since not all its inputs are registered) 2347 // if this is the projection from backbranch edge of Phi. 2348 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) { 2349 process_call_result(n->as_Proj(), phase); 2350 } 2351 if (!_processed.test(n->_idx)) { 2352 // The call's result may need to be processed later if the call 2353 // returns it's argument and the argument is not processed yet. 2354 _delayed_worklist.push(n); 2355 } 2356 break; 2357 } 2358 } 2359 _processed.set(n->_idx); 2360 break; 2361 } 2362 case Op_Return: 2363 { 2364 if( n->req() > TypeFunc::Parms && 2365 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2366 // Treat Return value as LocalVar with GlobalEscape escape state. 2367 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); 2368 int ti = n->in(TypeFunc::Parms)->_idx; 2369 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2370 if (nt == PointsToNode::UnknownType) { 2371 _delayed_worklist.push(n); // Process it later. 2372 break; 2373 } else if (nt == PointsToNode::JavaObject) { 2374 add_pointsto_edge(n->_idx, ti); 2375 } else { 2376 add_deferred_edge(n->_idx, ti); 2377 } 2378 } 2379 _processed.set(n->_idx); 2380 break; 2381 } 2382 case Op_StoreP: 2383 case Op_StoreN: 2384 { 2385 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2386 adr_type = adr_type->make_ptr(); 2387 if (adr_type->isa_oopptr()) { 2388 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2389 } else { 2390 Node* adr = n->in(MemNode::Address); 2391 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && 2392 adr->in(AddPNode::Address)->is_Proj() && 2393 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2394 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2395 // We are computing a raw address for a store captured 2396 // by an Initialize compute an appropriate address type. 2397 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2398 assert(offs != Type::OffsetBot, "offset must be a constant"); 2399 } else { 2400 _processed.set(n->_idx); 2401 return; 2402 } 2403 } 2404 break; 2405 } 2406 case Op_StorePConditional: 2407 case Op_CompareAndSwapP: 2408 case Op_CompareAndSwapN: 2409 { 2410 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2411 adr_type = adr_type->make_ptr(); 2412 if (adr_type->isa_oopptr()) { 2413 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2414 } else { 2415 _processed.set(n->_idx); 2416 return; 2417 } 2418 break; 2419 } 2420 case Op_AryEq: 2421 case Op_StrComp: 2422 case Op_StrEquals: 2423 case Op_StrIndexOf: 2424 { 2425 // char[] arrays passed to string intrinsics are not scalar replaceable. 2426 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2427 break; 2428 } 2429 case Op_ThreadLocal: 2430 { 2431 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); 2432 break; 2433 } 2434 default: 2435 ; 2436 // nothing to do 2437 } 2438 return; 2439 } 2440 2441 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { 2442 uint n_idx = n->_idx; 2443 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered"); 2444 2445 // Don't set processed bit for AddP, LoadP, StoreP since 2446 // they may need more then one pass to process. 2447 if (_processed.test(n_idx)) 2448 return; // No need to redefine node's state. 2449 2450 if (n->is_Call()) { 2451 CallNode *call = n->as_Call(); 2452 process_call_arguments(call, phase); 2453 _processed.set(n_idx); 2454 return; 2455 } 2456 2457 switch (n->Opcode()) { 2458 case Op_AddP: 2459 { 2460 Node *base = get_addp_base(n); 2461 // Create a field edge to this node from everything base could point to. 2462 VectorSet ptset(Thread::current()->resource_area()); 2463 PointsTo(ptset, base); 2464 for( VectorSetI i(&ptset); i.test(); ++i ) { 2465 uint pt = i.elem; 2466 add_field_edge(pt, n_idx, address_offset(n, phase)); 2467 } 2468 break; 2469 } 2470 case Op_CastX2P: 2471 { 2472 assert(false, "Op_CastX2P"); 2473 break; 2474 } 2475 case Op_CastPP: 2476 case Op_CheckCastPP: 2477 case Op_EncodeP: 2478 case Op_DecodeN: 2479 { 2480 int ti = n->in(1)->_idx; 2481 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered"); 2482 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2483 add_pointsto_edge(n_idx, ti); 2484 } else { 2485 add_deferred_edge(n_idx, ti); 2486 } 2487 _processed.set(n_idx); 2488 break; 2489 } 2490 case Op_ConP: 2491 { 2492 assert(false, "Op_ConP"); 2493 break; 2494 } 2495 case Op_ConN: 2496 { 2497 assert(false, "Op_ConN"); 2498 break; 2499 } 2500 case Op_CreateEx: 2501 { 2502 assert(false, "Op_CreateEx"); 2503 break; 2504 } 2505 case Op_LoadKlass: 2506 case Op_LoadNKlass: 2507 { 2508 assert(false, "Op_LoadKlass"); 2509 break; 2510 } 2511 case Op_LoadP: 2512 case Op_LoadN: 2513 { 2514 const Type *t = phase->type(n); 2515 #ifdef ASSERT 2516 if (t->make_ptr() == NULL) 2517 assert(false, "Op_LoadP"); 2518 #endif 2519 2520 Node* adr = n->in(MemNode::Address)->uncast(); 2521 Node* adr_base; 2522 if (adr->is_AddP()) { 2523 adr_base = get_addp_base(adr); 2524 } else { 2525 adr_base = adr; 2526 } 2527 2528 // For everything "adr_base" could point to, create a deferred edge from 2529 // this node to each field with the same offset. 2530 VectorSet ptset(Thread::current()->resource_area()); 2531 PointsTo(ptset, adr_base); 2532 int offset = address_offset(adr, phase); 2533 for( VectorSetI i(&ptset); i.test(); ++i ) { 2534 uint pt = i.elem; 2535 add_deferred_edge_to_fields(n_idx, pt, offset); 2536 } 2537 break; 2538 } 2539 case Op_Parm: 2540 { 2541 assert(false, "Op_Parm"); 2542 break; 2543 } 2544 case Op_Phi: 2545 { 2546 #ifdef ASSERT 2547 const Type *t = n->as_Phi()->type(); 2548 if (t->make_ptr() == NULL) 2549 assert(false, "Op_Phi"); 2550 #endif 2551 for (uint i = 1; i < n->req() ; i++) { 2552 Node* in = n->in(i); 2553 if (in == NULL) 2554 continue; // ignore NULL 2555 in = in->uncast(); 2556 if (in->is_top() || in == n) 2557 continue; // ignore top or inputs which go back this node 2558 int ti = in->_idx; 2559 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2560 assert(nt != PointsToNode::UnknownType, "all nodes should be known"); 2561 if (nt == PointsToNode::JavaObject) { 2562 add_pointsto_edge(n_idx, ti); 2563 } else { 2564 add_deferred_edge(n_idx, ti); 2565 } 2566 } 2567 _processed.set(n_idx); 2568 break; 2569 } 2570 case Op_Proj: 2571 { 2572 // we are only interested in the oop result projection from a call 2573 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2574 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType, 2575 "all nodes should be registered"); 2576 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2577 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2578 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2579 process_call_result(n->as_Proj(), phase); 2580 assert(_processed.test(n_idx), "all call results should be processed"); 2581 break; 2582 } 2583 } 2584 assert(false, "Op_Proj"); 2585 break; 2586 } 2587 case Op_Return: 2588 { 2589 #ifdef ASSERT 2590 if( n->req() <= TypeFunc::Parms || 2591 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2592 assert(false, "Op_Return"); 2593 } 2594 #endif 2595 int ti = n->in(TypeFunc::Parms)->_idx; 2596 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered"); 2597 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2598 add_pointsto_edge(n_idx, ti); 2599 } else { 2600 add_deferred_edge(n_idx, ti); 2601 } 2602 _processed.set(n_idx); 2603 break; 2604 } 2605 case Op_StoreP: 2606 case Op_StoreN: 2607 case Op_StorePConditional: 2608 case Op_CompareAndSwapP: 2609 case Op_CompareAndSwapN: 2610 { 2611 Node *adr = n->in(MemNode::Address); 2612 const Type *adr_type = phase->type(adr)->make_ptr(); 2613 #ifdef ASSERT 2614 if (!adr_type->isa_oopptr()) 2615 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); 2616 #endif 2617 2618 assert(adr->is_AddP(), "expecting an AddP"); 2619 Node *adr_base = get_addp_base(adr); 2620 Node *val = n->in(MemNode::ValueIn)->uncast(); 2621 // For everything "adr_base" could point to, create a deferred edge 2622 // to "val" from each field with the same offset. 2623 VectorSet ptset(Thread::current()->resource_area()); 2624 PointsTo(ptset, adr_base); 2625 for( VectorSetI i(&ptset); i.test(); ++i ) { 2626 uint pt = i.elem; 2627 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); 2628 } 2629 break; 2630 } 2631 case Op_AryEq: 2632 case Op_StrComp: 2633 case Op_StrEquals: 2634 case Op_StrIndexOf: 2635 { 2636 // char[] arrays passed to string intrinsic do not escape but 2637 // they are not scalar replaceable. Adjust escape state for them. 2638 // Start from in(2) edge since in(1) is memory edge. 2639 for (uint i = 2; i < n->req(); i++) { 2640 Node* adr = n->in(i)->uncast(); 2641 const Type *at = phase->type(adr); 2642 if (!adr->is_top() && at->isa_ptr()) { 2643 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 2644 at->isa_ptr() != NULL, "expecting an Ptr"); 2645 if (adr->is_AddP()) { 2646 adr = get_addp_base(adr); 2647 } 2648 // Mark as ArgEscape everything "adr" could point to. 2649 set_escape_state(adr->_idx, PointsToNode::ArgEscape); 2650 } 2651 } 2652 _processed.set(n_idx); 2653 break; 2654 } 2655 case Op_ThreadLocal: 2656 { 2657 assert(false, "Op_ThreadLocal"); 2658 break; 2659 } 2660 default: 2661 // This method should be called only for EA specific nodes. 2662 ShouldNotReachHere(); 2663 } 2664 } 2665 2666 #ifndef PRODUCT 2667 void ConnectionGraph::dump() { 2668 bool first = true; 2669 2670 uint size = nodes_size(); 2671 for (uint ni = 0; ni < size; ni++) { 2672 PointsToNode *ptn = ptnode_adr(ni); 2673 PointsToNode::NodeType ptn_type = ptn->node_type(); 2674 2675 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) 2676 continue; 2677 PointsToNode::EscapeState es = escape_state(ptn->_node); 2678 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { 2679 if (first) { 2680 tty->cr(); 2681 tty->print("======== Connection graph for "); 2682 _compile->method()->print_short_name(); 2683 tty->cr(); 2684 first = false; 2685 } 2686 tty->print("%6d ", ni); 2687 ptn->dump(); 2688 // Print all locals which reference this allocation 2689 for (uint li = ni; li < size; li++) { 2690 PointsToNode *ptn_loc = ptnode_adr(li); 2691 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); 2692 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && 2693 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { 2694 ptnode_adr(li)->dump(false); 2695 } 2696 } 2697 if (Verbose) { 2698 // Print all fields which reference this allocation 2699 for (uint i = 0; i < ptn->edge_count(); i++) { 2700 uint ei = ptn->edge_target(i); 2701 ptnode_adr(ei)->dump(false); 2702 } 2703 } 2704 tty->cr(); 2705 } 2706 } 2707 } 2708 #endif