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 _progress(false), 89 _compile(C), 90 _igvn(igvn), 91 _node_map(C->comp_arena()) { 92 93 _phantom_object = C->top()->_idx, 94 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true); 95 96 // Add ConP(#NULL) and ConN(#NULL) nodes. 97 Node* oop_null = igvn->zerocon(T_OBJECT); 98 _oop_null = oop_null->_idx; 99 assert(_oop_null < C->unique(), "should be created already"); 100 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 101 102 if (UseCompressedOops) { 103 Node* noop_null = igvn->zerocon(T_NARROWOOP); 104 _noop_null = noop_null->_idx; 105 assert(_noop_null < C->unique(), "should be created already"); 106 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 107 } 108 } 109 110 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) { 111 PointsToNode *f = ptnode_adr(from_i); 112 PointsToNode *t = ptnode_adr(to_i); 113 114 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 115 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge"); 116 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge"); 117 add_edge(f, to_i, PointsToNode::PointsToEdge); 118 } 119 120 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) { 121 PointsToNode *f = ptnode_adr(from_i); 122 PointsToNode *t = ptnode_adr(to_i); 123 124 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 125 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge"); 126 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge"); 127 // don't add a self-referential edge, this can occur during removal of 128 // deferred edges 129 if (from_i != to_i) 130 add_edge(f, to_i, PointsToNode::DeferredEdge); 131 } 132 133 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 134 const Type *adr_type = phase->type(adr); 135 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 136 adr->in(AddPNode::Address)->is_Proj() && 137 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 138 // We are computing a raw address for a store captured by an Initialize 139 // compute an appropriate address type. AddP cases #3 and #5 (see below). 140 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 141 assert(offs != Type::OffsetBot || 142 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 143 "offset must be a constant or it is initialization of array"); 144 return offs; 145 } 146 const TypePtr *t_ptr = adr_type->isa_ptr(); 147 assert(t_ptr != NULL, "must be a pointer type"); 148 return t_ptr->offset(); 149 } 150 151 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) { 152 PointsToNode *f = ptnode_adr(from_i); 153 PointsToNode *t = ptnode_adr(to_i); 154 155 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 156 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge"); 157 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge"); 158 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets"); 159 t->set_offset(offset); 160 161 add_edge(f, to_i, PointsToNode::FieldEdge); 162 } 163 164 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) { 165 PointsToNode *npt = ptnode_adr(ni); 166 PointsToNode::EscapeState old_es = npt->escape_state(); 167 if (es > old_es) 168 npt->set_escape_state(es); 169 } 170 171 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt, 172 PointsToNode::EscapeState es, bool done) { 173 PointsToNode* ptadr = ptnode_adr(n->_idx); 174 ptadr->_node = n; 175 ptadr->set_node_type(nt); 176 177 // inline set_escape_state(idx, es); 178 PointsToNode::EscapeState old_es = ptadr->escape_state(); 179 if (es > old_es) 180 ptadr->set_escape_state(es); 181 182 if (done) 183 _processed.set(n->_idx); 184 } 185 186 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) { 187 uint idx = n->_idx; 188 PointsToNode::EscapeState es; 189 190 // If we are still collecting or there were no non-escaping allocations 191 // we don't know the answer yet 192 if (_collecting) 193 return PointsToNode::UnknownEscape; 194 195 // if the node was created after the escape computation, return 196 // UnknownEscape 197 if (idx >= nodes_size()) 198 return PointsToNode::UnknownEscape; 199 200 es = ptnode_adr(idx)->escape_state(); 201 202 // if we have already computed a value, return it 203 if (es != PointsToNode::UnknownEscape && 204 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject) 205 return es; 206 207 // PointsTo() calls n->uncast() which can return a new ideal node. 208 if (n->uncast()->_idx >= nodes_size()) 209 return PointsToNode::UnknownEscape; 210 211 PointsToNode::EscapeState orig_es = es; 212 213 // compute max escape state of anything this node could point to 214 VectorSet ptset(Thread::current()->resource_area()); 215 PointsTo(ptset, n); 216 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) { 217 uint pt = i.elem; 218 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); 219 if (pes > es) 220 es = pes; 221 } 222 if (orig_es != es) { 223 // cache the computed escape state 224 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state"); 225 ptnode_adr(idx)->set_escape_state(es); 226 } // orig_es could be PointsToNode::UnknownEscape 227 return es; 228 } 229 230 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n) { 231 VectorSet visited(Thread::current()->resource_area()); 232 GrowableArray<uint> worklist; 233 234 #ifdef ASSERT 235 Node *orig_n = n; 236 #endif 237 238 n = n->uncast(); 239 PointsToNode* npt = ptnode_adr(n->_idx); 240 241 // If we have a JavaObject, return just that object 242 if (npt->node_type() == PointsToNode::JavaObject) { 243 ptset.set(n->_idx); 244 return; 245 } 246 #ifdef ASSERT 247 if (npt->_node == NULL) { 248 if (orig_n != n) 249 orig_n->dump(); 250 n->dump(); 251 assert(npt->_node != NULL, "unregistered node"); 252 } 253 #endif 254 worklist.push(n->_idx); 255 while(worklist.length() > 0) { 256 int ni = worklist.pop(); 257 if (visited.test_set(ni)) 258 continue; 259 260 PointsToNode* pn = ptnode_adr(ni); 261 // ensure that all inputs of a Phi have been processed 262 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); 263 264 int edges_processed = 0; 265 uint e_cnt = pn->edge_count(); 266 for (uint e = 0; e < e_cnt; e++) { 267 uint etgt = pn->edge_target(e); 268 PointsToNode::EdgeType et = pn->edge_type(e); 269 if (et == PointsToNode::PointsToEdge) { 270 ptset.set(etgt); 271 edges_processed++; 272 } else if (et == PointsToNode::DeferredEdge) { 273 worklist.push(etgt); 274 edges_processed++; 275 } else { 276 assert(false,"neither PointsToEdge or DeferredEdge"); 277 } 278 } 279 if (edges_processed == 0) { 280 // no deferred or pointsto edges found. Assume the value was set 281 // outside this method. Add the phantom object to the pointsto set. 282 ptset.set(_phantom_object); 283 } 284 } 285 } 286 287 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { 288 // This method is most expensive during ConnectionGraph construction. 289 // Reuse vectorSet and an additional growable array for deferred edges. 290 deferred_edges->clear(); 291 visited->Clear(); 292 293 visited->set(ni); 294 PointsToNode *ptn = ptnode_adr(ni); 295 296 // Mark current edges as visited and move deferred edges to separate array. 297 for (uint i = 0; i < ptn->edge_count(); ) { 298 uint t = ptn->edge_target(i); 299 #ifdef ASSERT 300 assert(!visited->test_set(t), "expecting no duplications"); 301 #else 302 visited->set(t); 303 #endif 304 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { 305 ptn->remove_edge(t, PointsToNode::DeferredEdge); 306 deferred_edges->append(t); 307 } else { 308 i++; 309 } 310 } 311 for (int next = 0; next < deferred_edges->length(); ++next) { 312 uint t = deferred_edges->at(next); 313 PointsToNode *ptt = ptnode_adr(t); 314 uint e_cnt = ptt->edge_count(); 315 for (uint e = 0; e < e_cnt; e++) { 316 uint etgt = ptt->edge_target(e); 317 if (visited->test_set(etgt)) 318 continue; 319 320 PointsToNode::EdgeType et = ptt->edge_type(e); 321 if (et == PointsToNode::PointsToEdge) { 322 add_pointsto_edge(ni, etgt); 323 if(etgt == _phantom_object) { 324 // Special case - field set outside (globally escaping). 325 ptn->set_escape_state(PointsToNode::GlobalEscape); 326 } 327 } else if (et == PointsToNode::DeferredEdge) { 328 deferred_edges->append(etgt); 329 } else { 330 assert(false,"invalid connection graph"); 331 } 332 } 333 } 334 } 335 336 337 // Add an edge to node given by "to_i" from any field of adr_i whose offset 338 // matches "offset" A deferred edge is added if to_i is a LocalVar, and 339 // a pointsto edge is added if it is a JavaObject 340 341 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { 342 PointsToNode* an = ptnode_adr(adr_i); 343 PointsToNode* to = ptnode_adr(to_i); 344 bool deferred = (to->node_type() == PointsToNode::LocalVar); 345 346 for (uint fe = 0; fe < an->edge_count(); fe++) { 347 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 348 int fi = an->edge_target(fe); 349 PointsToNode* pf = ptnode_adr(fi); 350 int po = pf->offset(); 351 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 352 if (deferred) 353 add_deferred_edge(fi, to_i); 354 else 355 add_pointsto_edge(fi, to_i); 356 } 357 } 358 } 359 360 // Add a deferred edge from node given by "from_i" to any field of adr_i 361 // whose offset matches "offset". 362 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { 363 PointsToNode* an = ptnode_adr(adr_i); 364 for (uint fe = 0; fe < an->edge_count(); fe++) { 365 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 366 int fi = an->edge_target(fe); 367 PointsToNode* pf = ptnode_adr(fi); 368 int po = pf->offset(); 369 if (pf->edge_count() == 0) { 370 // we have not seen any stores to this field, assume it was set outside this method 371 add_pointsto_edge(fi, _phantom_object); 372 } 373 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 374 add_deferred_edge(from_i, fi); 375 } 376 } 377 } 378 379 // Helper functions 380 381 static Node* get_addp_base(Node *addp) { 382 assert(addp->is_AddP(), "must be AddP"); 383 // 384 // AddP cases for Base and Address inputs: 385 // case #1. Direct object's field reference: 386 // Allocate 387 // | 388 // Proj #5 ( oop result ) 389 // | 390 // CheckCastPP (cast to instance type) 391 // | | 392 // AddP ( base == address ) 393 // 394 // case #2. Indirect object's field reference: 395 // Phi 396 // | 397 // CastPP (cast to instance type) 398 // | | 399 // AddP ( base == address ) 400 // 401 // case #3. Raw object's field reference for Initialize node: 402 // Allocate 403 // | 404 // Proj #5 ( oop result ) 405 // top | 406 // \ | 407 // AddP ( base == top ) 408 // 409 // case #4. Array's element reference: 410 // {CheckCastPP | CastPP} 411 // | | | 412 // | AddP ( array's element offset ) 413 // | | 414 // AddP ( array's offset ) 415 // 416 // case #5. Raw object's field reference for arraycopy stub call: 417 // The inline_native_clone() case when the arraycopy stub is called 418 // after the allocation before Initialize and CheckCastPP nodes. 419 // Allocate 420 // | 421 // Proj #5 ( oop result ) 422 // | | 423 // AddP ( base == address ) 424 // 425 // case #6. Constant Pool, ThreadLocal, CastX2P or 426 // Raw object's field reference: 427 // {ConP, ThreadLocal, CastX2P, raw Load} 428 // top | 429 // \ | 430 // AddP ( base == top ) 431 // 432 // case #7. Klass's field reference. 433 // LoadKlass 434 // | | 435 // AddP ( base == address ) 436 // 437 // case #8. narrow Klass's field reference. 438 // LoadNKlass 439 // | 440 // DecodeN 441 // | | 442 // AddP ( base == address ) 443 // 444 Node *base = addp->in(AddPNode::Base)->uncast(); 445 if (base->is_top()) { // The AddP case #3 and #6. 446 base = addp->in(AddPNode::Address)->uncast(); 447 while (base->is_AddP()) { 448 // Case #6 (unsafe access) may have several chained AddP nodes. 449 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only"); 450 base = base->in(AddPNode::Address)->uncast(); 451 } 452 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || 453 base->Opcode() == Op_CastX2P || base->is_DecodeN() || 454 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || 455 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); 456 } 457 return base; 458 } 459 460 static Node* find_second_addp(Node* addp, Node* n) { 461 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 462 463 Node* addp2 = addp->raw_out(0); 464 if (addp->outcnt() == 1 && addp2->is_AddP() && 465 addp2->in(AddPNode::Base) == n && 466 addp2->in(AddPNode::Address) == addp) { 467 468 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 469 // 470 // Find array's offset to push it on worklist first and 471 // as result process an array's element offset first (pushed second) 472 // to avoid CastPP for the array's offset. 473 // Otherwise the inserted CastPP (LocalVar) will point to what 474 // the AddP (Field) points to. Which would be wrong since 475 // the algorithm expects the CastPP has the same point as 476 // as AddP's base CheckCastPP (LocalVar). 477 // 478 // ArrayAllocation 479 // | 480 // CheckCastPP 481 // | 482 // memProj (from ArrayAllocation CheckCastPP) 483 // | || 484 // | || Int (element index) 485 // | || | ConI (log(element size)) 486 // | || | / 487 // | || LShift 488 // | || / 489 // | AddP (array's element offset) 490 // | | 491 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 492 // | / / 493 // AddP (array's offset) 494 // | 495 // Load/Store (memory operation on array's element) 496 // 497 return addp2; 498 } 499 return NULL; 500 } 501 502 // 503 // Adjust the type and inputs of an AddP which computes the 504 // address of a field of an instance 505 // 506 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { 507 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 508 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 509 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 510 if (t == NULL) { 511 // We are computing a raw address for a store captured by an Initialize 512 // compute an appropriate address type (cases #3 and #5). 513 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 514 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 515 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 516 assert(offs != Type::OffsetBot, "offset must be a constant"); 517 t = base_t->add_offset(offs)->is_oopptr(); 518 } 519 int inst_id = base_t->instance_id(); 520 assert(!t->is_known_instance() || t->instance_id() == inst_id, 521 "old type must be non-instance or match new type"); 522 523 // The type 't' could be subclass of 'base_t'. 524 // As result t->offset() could be large then base_t's size and it will 525 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 526 // constructor verifies correctness of the offset. 527 // 528 // It could happened on subclass's branch (from the type profiling 529 // inlining) which was not eliminated during parsing since the exactness 530 // of the allocation type was not propagated to the subclass type check. 531 // 532 // Or the type 't' could be not related to 'base_t' at all. 533 // It could happened when CHA type is different from MDO type on a dead path 534 // (for example, from instanceof check) which is not collapsed during parsing. 535 // 536 // Do nothing for such AddP node and don't process its users since 537 // this code branch will go away. 538 // 539 if (!t->is_known_instance() && 540 !base_t->klass()->is_subtype_of(t->klass())) { 541 return false; // bail out 542 } 543 544 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 545 // Do NOT remove the next line: ensure a new alias index is allocated 546 // for the instance type. Note: C++ will not remove it since the call 547 // has side effect. 548 int alias_idx = _compile->get_alias_index(tinst); 549 igvn->set_type(addp, tinst); 550 // record the allocation in the node map 551 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered"); 552 set_map(addp->_idx, get_map(base->_idx)); 553 554 // Set addp's Base and Address to 'base'. 555 Node *abase = addp->in(AddPNode::Base); 556 Node *adr = addp->in(AddPNode::Address); 557 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 558 adr->in(0)->_idx == (uint)inst_id) { 559 // Skip AddP cases #3 and #5. 560 } else { 561 assert(!abase->is_top(), "sanity"); // AddP case #3 562 if (abase != base) { 563 igvn->hash_delete(addp); 564 addp->set_req(AddPNode::Base, base); 565 if (abase == adr) { 566 addp->set_req(AddPNode::Address, base); 567 } else { 568 // AddP case #4 (adr is array's element offset AddP node) 569 #ifdef ASSERT 570 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 571 assert(adr->is_AddP() && atype != NULL && 572 atype->instance_id() == inst_id, "array's element offset should be processed first"); 573 #endif 574 } 575 igvn->hash_insert(addp); 576 } 577 } 578 // Put on IGVN worklist since at least addp's type was changed above. 579 record_for_optimizer(addp); 580 return true; 581 } 582 583 // 584 // Create a new version of orig_phi if necessary. Returns either the newly 585 // created phi or an existing phi. Sets create_new to indicate wheter a new 586 // phi was created. Cache the last newly created phi in the node map. 587 // 588 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { 589 Compile *C = _compile; 590 new_created = false; 591 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 592 // nothing to do if orig_phi is bottom memory or matches alias_idx 593 if (phi_alias_idx == alias_idx) { 594 return orig_phi; 595 } 596 // Have we recently created a Phi for this alias index? 597 PhiNode *result = get_map_phi(orig_phi->_idx); 598 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 599 return result; 600 } 601 // Previous check may fail when the same wide memory Phi was split into Phis 602 // for different memory slices. Search all Phis for this region. 603 if (result != NULL) { 604 Node* region = orig_phi->in(0); 605 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 606 Node* phi = region->fast_out(i); 607 if (phi->is_Phi() && 608 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 609 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 610 return phi->as_Phi(); 611 } 612 } 613 } 614 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { 615 if (C->do_escape_analysis() == true && !C->failing()) { 616 // Retry compilation without escape analysis. 617 // If this is the first failure, the sentinel string will "stick" 618 // to the Compile object, and the C2Compiler will see it and retry. 619 C->record_failure(C2Compiler::retry_no_escape_analysis()); 620 } 621 return NULL; 622 } 623 orig_phi_worklist.append_if_missing(orig_phi); 624 const TypePtr *atype = C->get_adr_type(alias_idx); 625 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 626 C->copy_node_notes_to(result, orig_phi); 627 igvn->set_type(result, result->bottom_type()); 628 record_for_optimizer(result); 629 630 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;) 631 assert(pn == NULL || pn == orig_phi, "wrong node"); 632 set_map(orig_phi->_idx, result); 633 ptnode_adr(orig_phi->_idx)->_node = orig_phi; 634 635 new_created = true; 636 return result; 637 } 638 639 // 640 // Return a new version of Memory Phi "orig_phi" with the inputs having the 641 // specified alias index. 642 // 643 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { 644 645 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 646 Compile *C = _compile; 647 bool new_phi_created; 648 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); 649 if (!new_phi_created) { 650 return result; 651 } 652 653 GrowableArray<PhiNode *> phi_list; 654 GrowableArray<uint> cur_input; 655 656 PhiNode *phi = orig_phi; 657 uint idx = 1; 658 bool finished = false; 659 while(!finished) { 660 while (idx < phi->req()) { 661 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); 662 if (mem != NULL && mem->is_Phi()) { 663 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); 664 if (new_phi_created) { 665 // found an phi for which we created a new split, push current one on worklist and begin 666 // processing new one 667 phi_list.push(phi); 668 cur_input.push(idx); 669 phi = mem->as_Phi(); 670 result = newphi; 671 idx = 1; 672 continue; 673 } else { 674 mem = newphi; 675 } 676 } 677 if (C->failing()) { 678 return NULL; 679 } 680 result->set_req(idx++, mem); 681 } 682 #ifdef ASSERT 683 // verify that the new Phi has an input for each input of the original 684 assert( phi->req() == result->req(), "must have same number of inputs."); 685 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 686 #endif 687 // Check if all new phi's inputs have specified alias index. 688 // Otherwise use old phi. 689 for (uint i = 1; i < phi->req(); i++) { 690 Node* in = result->in(i); 691 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 692 } 693 // we have finished processing a Phi, see if there are any more to do 694 finished = (phi_list.length() == 0 ); 695 if (!finished) { 696 phi = phi_list.pop(); 697 idx = cur_input.pop(); 698 PhiNode *prev_result = get_map_phi(phi->_idx); 699 prev_result->set_req(idx++, result); 700 result = prev_result; 701 } 702 } 703 return result; 704 } 705 706 707 // 708 // The next methods are derived from methods in MemNode. 709 // 710 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 711 Node *mem = mmem; 712 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 713 // means an array I have not precisely typed yet. Do not do any 714 // alias stuff with it any time soon. 715 if( toop->base() != Type::AnyPtr && 716 !(toop->klass() != NULL && 717 toop->klass()->is_java_lang_Object() && 718 toop->offset() == Type::OffsetBot) ) { 719 mem = mmem->memory_at(alias_idx); 720 // Update input if it is progress over what we have now 721 } 722 return mem; 723 } 724 725 // 726 // Move memory users to their memory slices. 727 // 728 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) { 729 Compile* C = _compile; 730 731 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 732 assert(tp != NULL, "ptr type"); 733 int alias_idx = C->get_alias_index(tp); 734 int general_idx = C->get_general_index(alias_idx); 735 736 // Move users first 737 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 738 Node* use = n->fast_out(i); 739 if (use->is_MergeMem()) { 740 MergeMemNode* mmem = use->as_MergeMem(); 741 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 742 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 743 continue; // Nothing to do 744 } 745 // Replace previous general reference to mem node. 746 uint orig_uniq = C->unique(); 747 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 748 assert(orig_uniq == C->unique(), "no new nodes"); 749 mmem->set_memory_at(general_idx, m); 750 --imax; 751 --i; 752 } else if (use->is_MemBar()) { 753 assert(!use->is_Initialize(), "initializing stores should not be moved"); 754 if (use->req() > MemBarNode::Precedent && 755 use->in(MemBarNode::Precedent) == n) { 756 // Don't move related membars. 757 record_for_optimizer(use); 758 continue; 759 } 760 tp = use->as_MemBar()->adr_type()->isa_ptr(); 761 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 762 alias_idx == general_idx) { 763 continue; // Nothing to do 764 } 765 // Move to general memory slice. 766 uint orig_uniq = C->unique(); 767 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 768 assert(orig_uniq == C->unique(), "no new nodes"); 769 igvn->hash_delete(use); 770 imax -= use->replace_edge(n, m); 771 igvn->hash_insert(use); 772 record_for_optimizer(use); 773 --i; 774 #ifdef ASSERT 775 } else if (use->is_Mem()) { 776 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 777 // Don't move related cardmark. 778 continue; 779 } 780 // Memory nodes should have new memory input. 781 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 782 assert(tp != NULL, "ptr type"); 783 int idx = C->get_alias_index(tp); 784 assert(get_map(use->_idx) != NULL || idx == alias_idx, 785 "Following memory nodes should have new memory input or be on the same memory slice"); 786 } else if (use->is_Phi()) { 787 // Phi nodes should be split and moved already. 788 tp = use->as_Phi()->adr_type()->isa_ptr(); 789 assert(tp != NULL, "ptr type"); 790 int idx = C->get_alias_index(tp); 791 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 792 } else { 793 use->dump(); 794 assert(false, "should not be here"); 795 #endif 796 } 797 } 798 } 799 800 // 801 // Search memory chain of "mem" to find a MemNode whose address 802 // is the specified alias index. 803 // 804 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { 805 if (orig_mem == NULL) 806 return orig_mem; 807 Compile* C = phase->C; 808 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 809 bool is_instance = (toop != NULL) && toop->is_known_instance(); 810 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 811 Node *prev = NULL; 812 Node *result = orig_mem; 813 while (prev != result) { 814 prev = result; 815 if (result == start_mem) 816 break; // hit one of our sentinels 817 if (result->is_Mem()) { 818 const Type *at = phase->type(result->in(MemNode::Address)); 819 if (at != Type::TOP) { 820 assert (at->isa_ptr() != NULL, "pointer type required."); 821 int idx = C->get_alias_index(at->is_ptr()); 822 if (idx == alias_idx) 823 break; 824 } 825 result = result->in(MemNode::Memory); 826 } 827 if (!is_instance) 828 continue; // don't search further for non-instance types 829 // skip over a call which does not affect this memory slice 830 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 831 Node *proj_in = result->in(0); 832 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 833 break; // hit one of our sentinels 834 } else if (proj_in->is_Call()) { 835 CallNode *call = proj_in->as_Call(); 836 if (!call->may_modify(toop, phase)) { 837 result = call->in(TypeFunc::Memory); 838 } 839 } else if (proj_in->is_Initialize()) { 840 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 841 // Stop if this is the initialization for the object instance which 842 // which contains this memory slice, otherwise skip over it. 843 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 844 result = proj_in->in(TypeFunc::Memory); 845 } 846 } else if (proj_in->is_MemBar()) { 847 result = proj_in->in(TypeFunc::Memory); 848 } 849 } else if (result->is_MergeMem()) { 850 MergeMemNode *mmem = result->as_MergeMem(); 851 result = step_through_mergemem(mmem, alias_idx, toop); 852 if (result == mmem->base_memory()) { 853 // Didn't find instance memory, search through general slice recursively. 854 result = mmem->memory_at(C->get_general_index(alias_idx)); 855 result = find_inst_mem(result, alias_idx, orig_phis, phase); 856 if (C->failing()) { 857 return NULL; 858 } 859 mmem->set_memory_at(alias_idx, result); 860 } 861 } else if (result->is_Phi() && 862 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 863 Node *un = result->as_Phi()->unique_input(phase); 864 if (un != NULL) { 865 orig_phis.append_if_missing(result->as_Phi()); 866 result = un; 867 } else { 868 break; 869 } 870 } else if (result->is_ClearArray()) { 871 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) { 872 // Can not bypass initialization of the instance 873 // we are looking for. 874 break; 875 } 876 // Otherwise skip it (the call updated 'result' value). 877 } else if (result->Opcode() == Op_SCMemProj) { 878 assert(result->in(0)->is_LoadStore(), "sanity"); 879 const Type *at = phase->type(result->in(0)->in(MemNode::Address)); 880 if (at != Type::TOP) { 881 assert (at->isa_ptr() != NULL, "pointer type required."); 882 int idx = C->get_alias_index(at->is_ptr()); 883 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 884 break; 885 } 886 result = result->in(0)->in(MemNode::Memory); 887 } 888 } 889 if (result->is_Phi()) { 890 PhiNode *mphi = result->as_Phi(); 891 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 892 const TypePtr *t = mphi->adr_type(); 893 if (C->get_alias_index(t) != alias_idx) { 894 // Create a new Phi with the specified alias index type. 895 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); 896 } else if (!is_instance) { 897 // Push all non-instance Phis on the orig_phis worklist to update inputs 898 // during Phase 4 if needed. 899 orig_phis.append_if_missing(mphi); 900 } 901 } 902 // the result is either MemNode, PhiNode, InitializeNode. 903 return result; 904 } 905 906 // 907 // Convert the types of unescaped object to instance types where possible, 908 // propagate the new type information through the graph, and update memory 909 // edges and MergeMem inputs to reflect the new type. 910 // 911 // We start with allocations (and calls which may be allocations) on alloc_worklist. 912 // The processing is done in 4 phases: 913 // 914 // Phase 1: Process possible allocations from alloc_worklist. Create instance 915 // types for the CheckCastPP for allocations where possible. 916 // Propagate the the new types through users as follows: 917 // casts and Phi: push users on alloc_worklist 918 // AddP: cast Base and Address inputs to the instance type 919 // push any AddP users on alloc_worklist and push any memnode 920 // users onto memnode_worklist. 921 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 922 // search the Memory chain for a store with the appropriate type 923 // address type. If a Phi is found, create a new version with 924 // the appropriate memory slices from each of the Phi inputs. 925 // For stores, process the users as follows: 926 // MemNode: push on memnode_worklist 927 // MergeMem: push on mergemem_worklist 928 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 929 // moving the first node encountered of each instance type to the 930 // the input corresponding to its alias index. 931 // appropriate memory slice. 932 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 933 // 934 // In the following example, the CheckCastPP nodes are the cast of allocation 935 // results and the allocation of node 29 is unescaped and eligible to be an 936 // instance type. 937 // 938 // We start with: 939 // 940 // 7 Parm #memory 941 // 10 ConI "12" 942 // 19 CheckCastPP "Foo" 943 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 944 // 29 CheckCastPP "Foo" 945 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 946 // 947 // 40 StoreP 25 7 20 ... alias_index=4 948 // 50 StoreP 35 40 30 ... alias_index=4 949 // 60 StoreP 45 50 20 ... alias_index=4 950 // 70 LoadP _ 60 30 ... alias_index=4 951 // 80 Phi 75 50 60 Memory alias_index=4 952 // 90 LoadP _ 80 30 ... alias_index=4 953 // 100 LoadP _ 80 20 ... alias_index=4 954 // 955 // 956 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 957 // and creating a new alias index for node 30. This gives: 958 // 959 // 7 Parm #memory 960 // 10 ConI "12" 961 // 19 CheckCastPP "Foo" 962 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 963 // 29 CheckCastPP "Foo" iid=24 964 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 965 // 966 // 40 StoreP 25 7 20 ... alias_index=4 967 // 50 StoreP 35 40 30 ... alias_index=6 968 // 60 StoreP 45 50 20 ... alias_index=4 969 // 70 LoadP _ 60 30 ... alias_index=6 970 // 80 Phi 75 50 60 Memory alias_index=4 971 // 90 LoadP _ 80 30 ... alias_index=6 972 // 100 LoadP _ 80 20 ... alias_index=4 973 // 974 // In phase 2, new memory inputs are computed for the loads and stores, 975 // And a new version of the phi is created. In phase 4, the inputs to 976 // node 80 are updated and then the memory nodes are updated with the 977 // values computed in phase 2. This results in: 978 // 979 // 7 Parm #memory 980 // 10 ConI "12" 981 // 19 CheckCastPP "Foo" 982 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 983 // 29 CheckCastPP "Foo" iid=24 984 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 985 // 986 // 40 StoreP 25 7 20 ... alias_index=4 987 // 50 StoreP 35 7 30 ... alias_index=6 988 // 60 StoreP 45 40 20 ... alias_index=4 989 // 70 LoadP _ 50 30 ... alias_index=6 990 // 80 Phi 75 40 60 Memory alias_index=4 991 // 120 Phi 75 50 50 Memory alias_index=6 992 // 90 LoadP _ 120 30 ... alias_index=6 993 // 100 LoadP _ 80 20 ... alias_index=4 994 // 995 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 996 GrowableArray<Node *> memnode_worklist; 997 GrowableArray<PhiNode *> orig_phis; 998 999 PhaseIterGVN *igvn = _igvn; 1000 uint new_index_start = (uint) _compile->num_alias_types(); 1001 Arena* arena = Thread::current()->resource_area(); 1002 VectorSet visited(arena); 1003 VectorSet ptset(arena); 1004 1005 1006 // Phase 1: Process possible allocations from alloc_worklist. 1007 // Create instance types for the CheckCastPP for allocations where possible. 1008 // 1009 // (Note: don't forget to change the order of the second AddP node on 1010 // the alloc_worklist if the order of the worklist processing is changed, 1011 // see the comment in find_second_addp().) 1012 // 1013 while (alloc_worklist.length() != 0) { 1014 Node *n = alloc_worklist.pop(); 1015 uint ni = n->_idx; 1016 const TypeOopPtr* tinst = NULL; 1017 if (n->is_Call()) { 1018 CallNode *alloc = n->as_Call(); 1019 // copy escape information to call node 1020 PointsToNode* ptn = ptnode_adr(alloc->_idx); 1021 PointsToNode::EscapeState es = escape_state(alloc); 1022 // We have an allocation or call which returns a Java object, 1023 // see if it is unescaped. 1024 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) 1025 continue; 1026 1027 // Find CheckCastPP for the allocate or for the return value of a call 1028 n = alloc->result_cast(); 1029 if (n == NULL) { // No uses except Initialize node 1030 if (alloc->is_Allocate()) { 1031 // Set the scalar_replaceable flag for allocation 1032 // so it could be eliminated if it has no uses. 1033 alloc->as_Allocate()->_is_scalar_replaceable = true; 1034 } 1035 continue; 1036 } 1037 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 1038 assert(!alloc->is_Allocate(), "allocation should have unique type"); 1039 continue; 1040 } 1041 1042 // The inline code for Object.clone() casts the allocation result to 1043 // java.lang.Object and then to the actual type of the allocated 1044 // object. Detect this case and use the second cast. 1045 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 1046 // the allocation result is cast to java.lang.Object and then 1047 // to the actual Array type. 1048 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 1049 && (alloc->is_AllocateArray() || 1050 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 1051 Node *cast2 = NULL; 1052 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1053 Node *use = n->fast_out(i); 1054 if (use->is_CheckCastPP()) { 1055 cast2 = use; 1056 break; 1057 } 1058 } 1059 if (cast2 != NULL) { 1060 n = cast2; 1061 } else { 1062 // Non-scalar replaceable if the allocation type is unknown statically 1063 // (reflection allocation), the object can't be restored during 1064 // deoptimization without precise type. 1065 continue; 1066 } 1067 } 1068 if (alloc->is_Allocate()) { 1069 // Set the scalar_replaceable flag for allocation 1070 // so it could be eliminated. 1071 alloc->as_Allocate()->_is_scalar_replaceable = true; 1072 } 1073 set_escape_state(n->_idx, es); 1074 // in order for an object to be scalar-replaceable, it must be: 1075 // - a direct allocation (not a call returning an object) 1076 // - non-escaping 1077 // - eligible to be a unique type 1078 // - not determined to be ineligible by escape analysis 1079 assert(ptnode_adr(alloc->_idx)->_node != NULL && 1080 ptnode_adr(n->_idx)->_node != NULL, "should be registered"); 1081 set_map(alloc->_idx, n); 1082 set_map(n->_idx, alloc); 1083 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 1084 if (t == NULL) 1085 continue; // not a TypeInstPtr 1086 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 1087 igvn->hash_delete(n); 1088 igvn->set_type(n, tinst); 1089 n->raise_bottom_type(tinst); 1090 igvn->hash_insert(n); 1091 record_for_optimizer(n); 1092 if (alloc->is_Allocate() && ptn->_scalar_replaceable && 1093 (t->isa_instptr() || t->isa_aryptr())) { 1094 1095 // First, put on the worklist all Field edges from Connection Graph 1096 // which is more accurate then putting immediate users from Ideal Graph. 1097 for (uint e = 0; e < ptn->edge_count(); e++) { 1098 Node *use = ptnode_adr(ptn->edge_target(e))->_node; 1099 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), 1100 "only AddP nodes are Field edges in CG"); 1101 if (use->outcnt() > 0) { // Don't process dead nodes 1102 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 1103 if (addp2 != NULL) { 1104 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1105 alloc_worklist.append_if_missing(addp2); 1106 } 1107 alloc_worklist.append_if_missing(use); 1108 } 1109 } 1110 1111 // An allocation may have an Initialize which has raw stores. Scan 1112 // the users of the raw allocation result and push AddP users 1113 // on alloc_worklist. 1114 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 1115 assert (raw_result != NULL, "must have an allocation result"); 1116 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 1117 Node *use = raw_result->fast_out(i); 1118 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 1119 Node* addp2 = find_second_addp(use, raw_result); 1120 if (addp2 != NULL) { 1121 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1122 alloc_worklist.append_if_missing(addp2); 1123 } 1124 alloc_worklist.append_if_missing(use); 1125 } else if (use->is_MemBar()) { 1126 memnode_worklist.append_if_missing(use); 1127 } 1128 } 1129 } 1130 } else if (n->is_AddP()) { 1131 ptset.Clear(); 1132 PointsTo(ptset, get_addp_base(n)); 1133 assert(ptset.Size() == 1, "AddP address is unique"); 1134 uint elem = ptset.getelem(); // Allocation node's index 1135 if (elem == _phantom_object) { 1136 assert(false, "escaped allocation"); 1137 continue; // Assume the value was set outside this method. 1138 } 1139 Node *base = get_map(elem); // CheckCastPP node 1140 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path 1141 tinst = igvn->type(base)->isa_oopptr(); 1142 } else if (n->is_Phi() || 1143 n->is_CheckCastPP() || 1144 n->is_EncodeP() || 1145 n->is_DecodeN() || 1146 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 1147 if (visited.test_set(n->_idx)) { 1148 assert(n->is_Phi(), "loops only through Phi's"); 1149 continue; // already processed 1150 } 1151 ptset.Clear(); 1152 PointsTo(ptset, n); 1153 if (ptset.Size() == 1) { 1154 uint elem = ptset.getelem(); // Allocation node's index 1155 if (elem == _phantom_object) { 1156 assert(false, "escaped allocation"); 1157 continue; // Assume the value was set outside this method. 1158 } 1159 Node *val = get_map(elem); // CheckCastPP node 1160 TypeNode *tn = n->as_Type(); 1161 tinst = igvn->type(val)->isa_oopptr(); 1162 assert(tinst != NULL && tinst->is_known_instance() && 1163 (uint)tinst->instance_id() == elem , "instance type expected."); 1164 1165 const Type *tn_type = igvn->type(tn); 1166 const TypeOopPtr *tn_t; 1167 if (tn_type->isa_narrowoop()) { 1168 tn_t = tn_type->make_ptr()->isa_oopptr(); 1169 } else { 1170 tn_t = tn_type->isa_oopptr(); 1171 } 1172 1173 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 1174 if (tn_type->isa_narrowoop()) { 1175 tn_type = tinst->make_narrowoop(); 1176 } else { 1177 tn_type = tinst; 1178 } 1179 igvn->hash_delete(tn); 1180 igvn->set_type(tn, tn_type); 1181 tn->set_type(tn_type); 1182 igvn->hash_insert(tn); 1183 record_for_optimizer(n); 1184 } else { 1185 assert(tn_type == TypePtr::NULL_PTR || 1186 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 1187 "unexpected type"); 1188 continue; // Skip dead path with different type 1189 } 1190 } 1191 } else { 1192 debug_only(n->dump();) 1193 assert(false, "EA: unexpected node"); 1194 continue; 1195 } 1196 // push allocation's users on appropriate worklist 1197 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1198 Node *use = n->fast_out(i); 1199 if(use->is_Mem() && use->in(MemNode::Address) == n) { 1200 // Load/store to instance's field 1201 memnode_worklist.append_if_missing(use); 1202 } else if (use->is_MemBar()) { 1203 memnode_worklist.append_if_missing(use); 1204 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 1205 Node* addp2 = find_second_addp(use, n); 1206 if (addp2 != NULL) { 1207 alloc_worklist.append_if_missing(addp2); 1208 } 1209 alloc_worklist.append_if_missing(use); 1210 } else if (use->is_Phi() || 1211 use->is_CheckCastPP() || 1212 use->is_EncodeP() || 1213 use->is_DecodeN() || 1214 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 1215 alloc_worklist.append_if_missing(use); 1216 #ifdef ASSERT 1217 } else if (use->is_Mem()) { 1218 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 1219 } else if (use->is_MergeMem()) { 1220 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1221 } else if (use->is_SafePoint()) { 1222 // Look for MergeMem nodes for calls which reference unique allocation 1223 // (through CheckCastPP nodes) even for debug info. 1224 Node* m = use->in(TypeFunc::Memory); 1225 if (m->is_MergeMem()) { 1226 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1227 } 1228 } else { 1229 uint op = use->Opcode(); 1230 if (!(op == Op_CmpP || op == Op_Conv2B || 1231 op == Op_CastP2X || op == Op_StoreCM || 1232 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 1233 op == Op_StrEquals || op == Op_StrIndexOf)) { 1234 n->dump(); 1235 use->dump(); 1236 assert(false, "EA: missing allocation reference path"); 1237 } 1238 #endif 1239 } 1240 } 1241 1242 } 1243 // New alias types were created in split_AddP(). 1244 uint new_index_end = (uint) _compile->num_alias_types(); 1245 1246 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 1247 // compute new values for Memory inputs (the Memory inputs are not 1248 // actually updated until phase 4.) 1249 if (memnode_worklist.length() == 0) 1250 return; // nothing to do 1251 1252 while (memnode_worklist.length() != 0) { 1253 Node *n = memnode_worklist.pop(); 1254 if (visited.test_set(n->_idx)) 1255 continue; 1256 if (n->is_Phi() || n->is_ClearArray()) { 1257 // we don't need to do anything, but the users must be pushed 1258 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 1259 // we don't need to do anything, but the users must be pushed 1260 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 1261 if (n == NULL) 1262 continue; 1263 } else { 1264 assert(n->is_Mem(), "memory node required."); 1265 Node *addr = n->in(MemNode::Address); 1266 const Type *addr_t = igvn->type(addr); 1267 if (addr_t == Type::TOP) 1268 continue; 1269 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 1270 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 1271 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 1272 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); 1273 if (_compile->failing()) { 1274 return; 1275 } 1276 if (mem != n->in(MemNode::Memory)) { 1277 // We delay the memory edge update since we need old one in 1278 // MergeMem code below when instances memory slices are separated. 1279 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;) 1280 assert(pn == NULL || pn == n, "wrong node"); 1281 set_map(n->_idx, mem); 1282 ptnode_adr(n->_idx)->_node = n; 1283 } 1284 if (n->is_Load()) { 1285 continue; // don't push users 1286 } else if (n->is_LoadStore()) { 1287 // get the memory projection 1288 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1289 Node *use = n->fast_out(i); 1290 if (use->Opcode() == Op_SCMemProj) { 1291 n = use; 1292 break; 1293 } 1294 } 1295 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 1296 } 1297 } 1298 // push user on appropriate worklist 1299 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1300 Node *use = n->fast_out(i); 1301 if (use->is_Phi() || use->is_ClearArray()) { 1302 memnode_worklist.append_if_missing(use); 1303 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { 1304 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 1305 continue; 1306 memnode_worklist.append_if_missing(use); 1307 } else if (use->is_MemBar()) { 1308 memnode_worklist.append_if_missing(use); 1309 #ifdef ASSERT 1310 } else if(use->is_Mem()) { 1311 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 1312 } else if (use->is_MergeMem()) { 1313 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1314 } else { 1315 uint op = use->Opcode(); 1316 if (!(op == Op_StoreCM || 1317 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 1318 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 1319 op == Op_AryEq || op == Op_StrComp || 1320 op == Op_StrEquals || op == Op_StrIndexOf)) { 1321 n->dump(); 1322 use->dump(); 1323 assert(false, "EA: missing memory path"); 1324 } 1325 #endif 1326 } 1327 } 1328 } 1329 1330 // Phase 3: Process MergeMem nodes from mergemem_worklist. 1331 // Walk each memory slice moving the first node encountered of each 1332 // instance type to the the input corresponding to its alias index. 1333 uint length = _mergemem_worklist.length(); 1334 for( uint next = 0; next < length; ++next ) { 1335 MergeMemNode* nmm = _mergemem_worklist.at(next); 1336 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 1337 // Note: we don't want to use MergeMemStream here because we only want to 1338 // scan inputs which exist at the start, not ones we add during processing. 1339 // Note 2: MergeMem may already contains instance memory slices added 1340 // during find_inst_mem() call when memory nodes were processed above. 1341 igvn->hash_delete(nmm); 1342 uint nslices = nmm->req(); 1343 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 1344 Node* mem = nmm->in(i); 1345 Node* cur = NULL; 1346 if (mem == NULL || mem->is_top()) 1347 continue; 1348 // First, update mergemem by moving memory nodes to corresponding slices 1349 // if their type became more precise since this mergemem was created. 1350 while (mem->is_Mem()) { 1351 const Type *at = igvn->type(mem->in(MemNode::Address)); 1352 if (at != Type::TOP) { 1353 assert (at->isa_ptr() != NULL, "pointer type required."); 1354 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 1355 if (idx == i) { 1356 if (cur == NULL) 1357 cur = mem; 1358 } else { 1359 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 1360 nmm->set_memory_at(idx, mem); 1361 } 1362 } 1363 } 1364 mem = mem->in(MemNode::Memory); 1365 } 1366 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 1367 // Find any instance of the current type if we haven't encountered 1368 // already a memory slice of the instance along the memory chain. 1369 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1370 if((uint)_compile->get_general_index(ni) == i) { 1371 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 1372 if (nmm->is_empty_memory(m)) { 1373 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); 1374 if (_compile->failing()) { 1375 return; 1376 } 1377 nmm->set_memory_at(ni, result); 1378 } 1379 } 1380 } 1381 } 1382 // Find the rest of instances values 1383 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1384 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 1385 Node* result = step_through_mergemem(nmm, ni, tinst); 1386 if (result == nmm->base_memory()) { 1387 // Didn't find instance memory, search through general slice recursively. 1388 result = nmm->memory_at(_compile->get_general_index(ni)); 1389 result = find_inst_mem(result, ni, orig_phis, igvn); 1390 if (_compile->failing()) { 1391 return; 1392 } 1393 nmm->set_memory_at(ni, result); 1394 } 1395 } 1396 igvn->hash_insert(nmm); 1397 record_for_optimizer(nmm); 1398 } 1399 1400 // Phase 4: Update the inputs of non-instance memory Phis and 1401 // the Memory input of memnodes 1402 // First update the inputs of any non-instance Phi's from 1403 // which we split out an instance Phi. Note we don't have 1404 // to recursively process Phi's encounted on the input memory 1405 // chains as is done in split_memory_phi() since they will 1406 // also be processed here. 1407 for (int j = 0; j < orig_phis.length(); j++) { 1408 PhiNode *phi = orig_phis.at(j); 1409 int alias_idx = _compile->get_alias_index(phi->adr_type()); 1410 igvn->hash_delete(phi); 1411 for (uint i = 1; i < phi->req(); i++) { 1412 Node *mem = phi->in(i); 1413 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); 1414 if (_compile->failing()) { 1415 return; 1416 } 1417 if (mem != new_mem) { 1418 phi->set_req(i, new_mem); 1419 } 1420 } 1421 igvn->hash_insert(phi); 1422 record_for_optimizer(phi); 1423 } 1424 1425 // Update the memory inputs of MemNodes with the value we computed 1426 // in Phase 2 and move stores memory users to corresponding memory slices. 1427 #ifdef ASSERT 1428 visited.Clear(); 1429 Node_Stack old_mems(arena, _compile->unique() >> 2); 1430 #endif 1431 for (uint i = 0; i < nodes_size(); i++) { 1432 Node *nmem = get_map(i); 1433 if (nmem != NULL) { 1434 Node *n = ptnode_adr(i)->_node; 1435 assert(n != NULL, "sanity"); 1436 if (n->is_Mem()) { 1437 #ifdef ASSERT 1438 Node* old_mem = n->in(MemNode::Memory); 1439 if (!visited.test_set(old_mem->_idx)) { 1440 old_mems.push(old_mem, old_mem->outcnt()); 1441 } 1442 #endif 1443 assert(n->in(MemNode::Memory) != nmem, "sanity"); 1444 if (!n->is_Load()) { 1445 // Move memory users of a store first. 1446 move_inst_mem(n, orig_phis, igvn); 1447 } 1448 // Now update memory input 1449 igvn->hash_delete(n); 1450 n->set_req(MemNode::Memory, nmem); 1451 igvn->hash_insert(n); 1452 record_for_optimizer(n); 1453 } else { 1454 assert(n->is_Allocate() || n->is_CheckCastPP() || 1455 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 1456 } 1457 } 1458 } 1459 #ifdef ASSERT 1460 // Verify that memory was split correctly 1461 while (old_mems.is_nonempty()) { 1462 Node* old_mem = old_mems.node(); 1463 uint old_cnt = old_mems.index(); 1464 old_mems.pop(); 1465 assert(old_cnt = old_mem->outcnt(), "old mem could be lost"); 1466 } 1467 #endif 1468 } 1469 1470 bool ConnectionGraph::has_candidates(Compile *C) { 1471 // EA brings benefits only when the code has allocations and/or locks which 1472 // are represented by ideal Macro nodes. 1473 int cnt = C->macro_count(); 1474 for( int i=0; i < cnt; i++ ) { 1475 Node *n = C->macro_node(i); 1476 if ( n->is_Allocate() ) 1477 return true; 1478 if( n->is_Lock() ) { 1479 Node* obj = n->as_Lock()->obj_node()->uncast(); 1480 if( !(obj->is_Parm() || obj->is_Con()) ) 1481 return true; 1482 } 1483 } 1484 return false; 1485 } 1486 1487 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 1488 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 1489 // to create space for them in ConnectionGraph::_nodes[]. 1490 Node* oop_null = igvn->zerocon(T_OBJECT); 1491 Node* noop_null = igvn->zerocon(T_NARROWOOP); 1492 1493 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 1494 // Perform escape analysis 1495 if (congraph->compute_escape()) { 1496 // There are non escaping objects. 1497 C->set_congraph(congraph); 1498 } 1499 1500 // Cleanup. 1501 if (oop_null->outcnt() == 0) 1502 igvn->hash_delete(oop_null); 1503 if (noop_null->outcnt() == 0) 1504 igvn->hash_delete(noop_null); 1505 } 1506 1507 bool ConnectionGraph::compute_escape() { 1508 Compile* C = _compile; 1509 1510 // 1. Populate Connection Graph (CG) with Ideal nodes. 1511 1512 Unique_Node_List worklist_init; 1513 worklist_init.map(C->unique(), NULL); // preallocate space 1514 1515 // Initialize worklist 1516 if (C->root() != NULL) { 1517 worklist_init.push(C->root()); 1518 } 1519 1520 GrowableArray<int> cg_worklist; 1521 PhaseGVN* igvn = _igvn; 1522 bool has_allocations = false; 1523 1524 // Push all useful nodes onto CG list and set their type. 1525 for( uint next = 0; next < worklist_init.size(); ++next ) { 1526 Node* n = worklist_init.at(next); 1527 record_for_escape_analysis(n, igvn); 1528 // Only allocations and java static calls results are checked 1529 // for an escape status. See process_call_result() below. 1530 if (n->is_Allocate() || n->is_CallStaticJava() && 1531 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { 1532 has_allocations = true; 1533 } 1534 if(n->is_AddP()) { 1535 // Collect address nodes. Use them during stage 3 below 1536 // to build initial connection graph field edges. 1537 cg_worklist.append(n->_idx); 1538 } else if (n->is_MergeMem()) { 1539 // Collect all MergeMem nodes to add memory slices for 1540 // scalar replaceable objects in split_unique_types(). 1541 _mergemem_worklist.append(n->as_MergeMem()); 1542 } 1543 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1544 Node* m = n->fast_out(i); // Get user 1545 worklist_init.push(m); 1546 } 1547 } 1548 1549 if (!has_allocations) { 1550 _collecting = false; 1551 return false; // Nothing to do. 1552 } 1553 1554 // 2. First pass to create simple CG edges (doesn't require to walk CG). 1555 uint delayed_size = _delayed_worklist.size(); 1556 for( uint next = 0; next < delayed_size; ++next ) { 1557 Node* n = _delayed_worklist.at(next); 1558 build_connection_graph(n, igvn); 1559 } 1560 1561 // 3. Pass to create initial fields edges (JavaObject -F-> AddP) 1562 // to reduce number of iterations during stage 4 below. 1563 uint cg_length = cg_worklist.length(); 1564 for( uint next = 0; next < cg_length; ++next ) { 1565 int ni = cg_worklist.at(next); 1566 Node* n = ptnode_adr(ni)->_node; 1567 Node* base = get_addp_base(n); 1568 if (base->is_Proj()) 1569 base = base->in(0); 1570 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type(); 1571 if (nt == PointsToNode::JavaObject) { 1572 build_connection_graph(n, igvn); 1573 } 1574 } 1575 1576 cg_worklist.clear(); 1577 cg_worklist.append(_phantom_object); 1578 1579 // 4. Build Connection Graph which need 1580 // to walk the connection graph. 1581 _progress = false; 1582 for (uint ni = 0; ni < nodes_size(); ni++) { 1583 PointsToNode* ptn = ptnode_adr(ni); 1584 Node *n = ptn->_node; 1585 if (n != NULL) { // Call, AddP, LoadP, StoreP 1586 build_connection_graph(n, igvn); 1587 if (ptn->node_type() != PointsToNode::UnknownType) 1588 cg_worklist.append(n->_idx); // Collect CG nodes 1589 } 1590 } 1591 // After IGVN user nodes may have smaller _idx than 1592 // their inputs so they will be processed first in 1593 // previous loop. Because of that not all Graph 1594 // edges will be created. Walk over interesting 1595 // nodes again until no new edges are created. 1596 // Set limit on iterations. 1597 cg_length = cg_worklist.length(); 1598 int iterations = 0; 1599 while(_progress && (iterations++ < 10)) { 1600 _progress = false; 1601 for( uint next = 0; next < cg_length; ++next ) { 1602 int ni = cg_worklist.at(next); 1603 PointsToNode* ptn = ptnode_adr(ni); 1604 DEBUG_ONLY(PointsToNode::NodeType nt = ptn->node_type()); 1605 Node* n = ptn->_node; 1606 assert(n != NULL && nt != PointsToNode::UnknownType, "should be known node"); 1607 build_connection_graph(n, igvn); 1608 } 1609 } 1610 if (iterations >= 10) { 1611 // Possible infinit build_connection_graph loop. 1612 // Retry compilation without escape analysis. 1613 C->record_failure(C2Compiler::retry_no_escape_analysis()); 1614 _collecting = false; 1615 return false; 1616 } 1617 1618 Arena* arena = Thread::current()->resource_area(); 1619 VectorSet ptset(arena); 1620 GrowableArray<uint> deferred_edges; 1621 VectorSet visited(arena); 1622 1623 // 5. Remove deferred edges from the graph and adjust 1624 // escape state of nonescaping objects. 1625 for( uint next = 0; next < cg_length; ++next ) { 1626 int ni = cg_worklist.at(next); 1627 PointsToNode* ptn = ptnode_adr(ni); 1628 PointsToNode::NodeType nt = ptn->node_type(); 1629 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { 1630 remove_deferred(ni, &deferred_edges, &visited); 1631 Node *n = ptn->_node; 1632 if (n->is_AddP()) { 1633 // Search for objects which are not scalar replaceable 1634 // and adjust their escape state. 1635 verify_escape_state(ni, ptset, igvn); 1636 } 1637 } 1638 } 1639 1640 // 6. Propagate escape states. 1641 GrowableArray<int> worklist; 1642 bool has_non_escaping_obj = false; 1643 1644 // push all GlobalEscape nodes on the worklist 1645 for( uint next = 0; next < cg_length; ++next ) { 1646 int nk = cg_worklist.at(next); 1647 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) 1648 worklist.push(nk); 1649 } 1650 // mark all nodes reachable from GlobalEscape nodes 1651 while(worklist.length() > 0) { 1652 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1653 uint e_cnt = ptn->edge_count(); 1654 for (uint ei = 0; ei < e_cnt; ei++) { 1655 uint npi = ptn->edge_target(ei); 1656 PointsToNode *np = ptnode_adr(npi); 1657 if (np->escape_state() < PointsToNode::GlobalEscape) { 1658 np->set_escape_state(PointsToNode::GlobalEscape); 1659 worklist.push(npi); 1660 } 1661 } 1662 } 1663 1664 // push all ArgEscape nodes on the worklist 1665 for( uint next = 0; next < cg_length; ++next ) { 1666 int nk = cg_worklist.at(next); 1667 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) 1668 worklist.push(nk); 1669 } 1670 // mark all nodes reachable from ArgEscape nodes 1671 while(worklist.length() > 0) { 1672 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1673 if (ptn->node_type() == PointsToNode::JavaObject) 1674 has_non_escaping_obj = true; // Non GlobalEscape 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::ArgEscape) { 1680 np->set_escape_state(PointsToNode::ArgEscape); 1681 worklist.push(npi); 1682 } 1683 } 1684 } 1685 1686 GrowableArray<Node*> alloc_worklist; 1687 1688 // push all NoEscape nodes on the worklist 1689 for( uint next = 0; next < cg_length; ++next ) { 1690 int nk = cg_worklist.at(next); 1691 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) 1692 worklist.push(nk); 1693 } 1694 // mark all nodes reachable from NoEscape nodes 1695 while(worklist.length() > 0) { 1696 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1697 if (ptn->node_type() == PointsToNode::JavaObject) 1698 has_non_escaping_obj = true; // Non GlobalEscape 1699 Node* n = ptn->_node; 1700 if (n->is_Allocate() && ptn->_scalar_replaceable ) { 1701 // Push scalar replaceable allocations on alloc_worklist 1702 // for processing in split_unique_types(). 1703 alloc_worklist.append(n); 1704 } 1705 uint e_cnt = ptn->edge_count(); 1706 for (uint ei = 0; ei < e_cnt; ei++) { 1707 uint npi = ptn->edge_target(ei); 1708 PointsToNode *np = ptnode_adr(npi); 1709 if (np->escape_state() < PointsToNode::NoEscape) { 1710 np->set_escape_state(PointsToNode::NoEscape); 1711 worklist.push(npi); 1712 } 1713 } 1714 } 1715 1716 _collecting = false; 1717 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); 1718 1719 #ifndef PRODUCT 1720 if (PrintEscapeAnalysis) { 1721 dump(); // Dump ConnectionGraph 1722 } 1723 #endif 1724 1725 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; 1726 if ( has_scalar_replaceable_candidates && 1727 C->AliasLevel() >= 3 && EliminateAllocations ) { 1728 1729 // Now use the escape information to create unique types for 1730 // scalar replaceable objects. 1731 split_unique_types(alloc_worklist); 1732 1733 if (C->failing()) return false; 1734 1735 C->print_method("After Escape Analysis", 2); 1736 1737 #ifdef ASSERT 1738 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 1739 tty->print("=== No allocations eliminated for "); 1740 C->method()->print_short_name(); 1741 if(!EliminateAllocations) { 1742 tty->print(" since EliminateAllocations is off ==="); 1743 } else if(!has_scalar_replaceable_candidates) { 1744 tty->print(" since there are no scalar replaceable candidates ==="); 1745 } else if(C->AliasLevel() < 3) { 1746 tty->print(" since AliasLevel < 3 ==="); 1747 } 1748 tty->cr(); 1749 #endif 1750 } 1751 return has_non_escaping_obj; 1752 } 1753 1754 // Search for objects which are not scalar replaceable. 1755 void ConnectionGraph::verify_escape_state(int nidx, VectorSet& ptset, PhaseTransform* phase) { 1756 PointsToNode* ptn = ptnode_adr(nidx); 1757 Node* n = ptn->_node; 1758 assert(n->is_AddP(), "Should be called for AddP nodes only"); 1759 // Search for objects which are not scalar replaceable. 1760 // Mark their escape state as ArgEscape to propagate the state 1761 // to referenced objects. 1762 // Note: currently there are no difference in compiler optimizations 1763 // for ArgEscape objects and NoEscape objects which are not 1764 // scalar replaceable. 1765 1766 Compile* C = _compile; 1767 1768 int offset = ptn->offset(); 1769 Node* base = get_addp_base(n); 1770 ptset.Clear(); 1771 PointsTo(ptset, base); 1772 int ptset_size = ptset.Size(); 1773 1774 // Check if a oop field's initializing value is recorded and add 1775 // a corresponding NULL field's value if it is not recorded. 1776 // Connection Graph does not record a default initialization by NULL 1777 // captured by Initialize node. 1778 // 1779 // Note: it will disable scalar replacement in some cases: 1780 // 1781 // Point p[] = new Point[1]; 1782 // p[0] = new Point(); // Will be not scalar replaced 1783 // 1784 // but it will save us from incorrect optimizations in next cases: 1785 // 1786 // Point p[] = new Point[1]; 1787 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1788 // 1789 // Do a simple control flow analysis to distinguish above cases. 1790 // 1791 if (offset != Type::OffsetBot && ptset_size == 1) { 1792 uint elem = ptset.getelem(); // Allocation node's index 1793 // It does not matter if it is not Allocation node since 1794 // only non-escaping allocations are scalar replaced. 1795 if (ptnode_adr(elem)->_node->is_Allocate() && 1796 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) { 1797 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate(); 1798 InitializeNode* ini = alloc->initialization(); 1799 1800 // Check only oop fields. 1801 const Type* adr_type = n->as_AddP()->bottom_type(); 1802 BasicType basic_field_type = T_INT; 1803 if (adr_type->isa_instptr()) { 1804 ciField* field = C->alias_type(adr_type->isa_instptr())->field(); 1805 if (field != NULL) { 1806 basic_field_type = field->layout_type(); 1807 } else { 1808 // Ignore non field load (for example, klass load) 1809 } 1810 } else if (adr_type->isa_aryptr()) { 1811 const Type* elemtype = adr_type->isa_aryptr()->elem(); 1812 basic_field_type = elemtype->array_element_basic_type(); 1813 } else { 1814 // Raw pointers are used for initializing stores so skip it. 1815 assert(adr_type->isa_rawptr() && base->is_Proj() && 1816 (base->in(0) == alloc),"unexpected pointer type"); 1817 } 1818 if (basic_field_type == T_OBJECT || 1819 basic_field_type == T_NARROWOOP || 1820 basic_field_type == T_ARRAY) { 1821 Node* value = NULL; 1822 if (ini != NULL) { 1823 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT; 1824 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase); 1825 if (store != NULL && store->is_Store()) { 1826 value = store->in(MemNode::ValueIn); 1827 } else if (ptn->edge_count() > 0) { // Are there oop stores? 1828 // Check for a store which follows allocation without branches. 1829 // For example, a volatile field store is not collected 1830 // by Initialize node. TODO: it would be nice to use idom() here. 1831 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1832 store = n->fast_out(i); 1833 if (store->is_Store() && store->in(0) != NULL) { 1834 Node* ctrl = store->in(0); 1835 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL || 1836 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() || 1837 ctrl->is_IfTrue() || ctrl->is_IfFalse())) { 1838 ctrl = ctrl->in(0); 1839 } 1840 if (ctrl == ini || ctrl == alloc) { 1841 value = store->in(MemNode::ValueIn); 1842 break; 1843 } 1844 } 1845 } 1846 } 1847 } 1848 if (value == NULL || value != ptnode_adr(value->_idx)->_node) { 1849 // A field's initializing value was not recorded. Add NULL. 1850 uint null_idx = UseCompressedOops ? _noop_null : _oop_null; 1851 add_pointsto_edge(nidx, null_idx); 1852 } 1853 } 1854 } 1855 } 1856 1857 // An object is not scalar replaceable if the field which may point 1858 // to it has unknown offset (unknown element of an array of objects). 1859 // 1860 if (offset == Type::OffsetBot) { 1861 uint e_cnt = ptn->edge_count(); 1862 for (uint ei = 0; ei < e_cnt; ei++) { 1863 uint npi = ptn->edge_target(ei); 1864 set_escape_state(npi, PointsToNode::ArgEscape); 1865 ptnode_adr(npi)->_scalar_replaceable = false; 1866 } 1867 } 1868 1869 // Currently an object is not scalar replaceable if a LoadStore node 1870 // access its field since the field value is unknown after it. 1871 // 1872 bool has_LoadStore = false; 1873 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1874 Node *use = n->fast_out(i); 1875 if (use->is_LoadStore()) { 1876 has_LoadStore = true; 1877 break; 1878 } 1879 } 1880 // An object is not scalar replaceable if the address points 1881 // to unknown field (unknown element for arrays, offset is OffsetBot). 1882 // 1883 // Or the address may point to more then one object. This may produce 1884 // the false positive result (set scalar_replaceable to false) 1885 // since the flow-insensitive escape analysis can't separate 1886 // the case when stores overwrite the field's value from the case 1887 // when stores happened on different control branches. 1888 // 1889 if (ptset_size > 1 || ptset_size != 0 && 1890 (has_LoadStore || offset == Type::OffsetBot)) { 1891 for( VectorSetI j(&ptset); j.test(); ++j ) { 1892 set_escape_state(j.elem, PointsToNode::ArgEscape); 1893 ptnode_adr(j.elem)->_scalar_replaceable = false; 1894 } 1895 } 1896 } 1897 1898 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { 1899 1900 switch (call->Opcode()) { 1901 #ifdef ASSERT 1902 case Op_Allocate: 1903 case Op_AllocateArray: 1904 case Op_Lock: 1905 case Op_Unlock: 1906 assert(false, "should be done already"); 1907 break; 1908 #endif 1909 case Op_CallLeaf: 1910 case Op_CallLeafNoFP: 1911 { 1912 // Stub calls, objects do not escape but they are not scale replaceable. 1913 // Adjust escape state for outgoing arguments. 1914 const TypeTuple * d = call->tf()->domain(); 1915 VectorSet ptset(Thread::current()->resource_area()); 1916 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1917 const Type* at = d->field_at(i); 1918 Node *arg = call->in(i)->uncast(); 1919 const Type *aat = phase->type(arg); 1920 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() && 1921 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) { 1922 1923 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 1924 aat->isa_ptr() != NULL, "expecting an Ptr"); 1925 #ifdef ASSERT 1926 if (!(call->Opcode() == Op_CallLeafNoFP && 1927 call->as_CallLeaf()->_name != NULL && 1928 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) || 1929 call->as_CallLeaf()->_name != NULL && 1930 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 1931 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 )) 1932 ) { 1933 call->dump(); 1934 assert(false, "EA: unexpected CallLeaf"); 1935 } 1936 #endif 1937 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1938 if (arg->is_AddP()) { 1939 // 1940 // The inline_native_clone() case when the arraycopy stub is called 1941 // after the allocation before Initialize and CheckCastPP nodes. 1942 // 1943 // Set AddP's base (Allocate) as not scalar replaceable since 1944 // pointer to the base (with offset) is passed as argument. 1945 // 1946 arg = get_addp_base(arg); 1947 } 1948 ptset.Clear(); 1949 PointsTo(ptset, arg); 1950 for( VectorSetI j(&ptset); j.test(); ++j ) { 1951 uint pt = j.elem; 1952 set_escape_state(pt, PointsToNode::ArgEscape); 1953 } 1954 } 1955 } 1956 break; 1957 } 1958 1959 case Op_CallStaticJava: 1960 // For a static call, we know exactly what method is being called. 1961 // Use bytecode estimator to record the call's escape affects 1962 { 1963 ciMethod *meth = call->as_CallJava()->method(); 1964 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1965 // fall-through if not a Java method or no analyzer information 1966 if (call_analyzer != NULL) { 1967 const TypeTuple * d = call->tf()->domain(); 1968 VectorSet ptset(Thread::current()->resource_area()); 1969 bool copy_dependencies = false; 1970 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1971 const Type* at = d->field_at(i); 1972 int k = i - TypeFunc::Parms; 1973 Node *arg = call->in(i)->uncast(); 1974 1975 if (at->isa_oopptr() != NULL && 1976 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) { 1977 1978 bool global_escapes = false; 1979 bool fields_escapes = false; 1980 if (!call_analyzer->is_arg_stack(k)) { 1981 // The argument global escapes, mark everything it could point to 1982 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1983 global_escapes = true; 1984 } else { 1985 if (!call_analyzer->is_arg_local(k)) { 1986 // The argument itself doesn't escape, but any fields might 1987 fields_escapes = true; 1988 } 1989 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1990 copy_dependencies = true; 1991 } 1992 1993 ptset.Clear(); 1994 PointsTo(ptset, arg); 1995 for( VectorSetI j(&ptset); j.test(); ++j ) { 1996 uint pt = j.elem; 1997 if (global_escapes) { 1998 //The argument global escapes, mark everything it could point to 1999 set_escape_state(pt, PointsToNode::GlobalEscape); 2000 } else { 2001 if (fields_escapes) { 2002 // The argument itself doesn't escape, but any fields might 2003 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); 2004 } 2005 set_escape_state(pt, PointsToNode::ArgEscape); 2006 } 2007 } 2008 } 2009 } 2010 if (copy_dependencies) 2011 call_analyzer->copy_dependencies(_compile->dependencies()); 2012 break; 2013 } 2014 } 2015 2016 default: 2017 // Fall-through here if not a Java method or no analyzer information 2018 // or some other type of call, assume the worst case: all arguments 2019 // globally escape. 2020 { 2021 // adjust escape state for outgoing arguments 2022 const TypeTuple * d = call->tf()->domain(); 2023 VectorSet ptset(Thread::current()->resource_area()); 2024 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2025 const Type* at = d->field_at(i); 2026 if (at->isa_oopptr() != NULL) { 2027 Node *arg = call->in(i)->uncast(); 2028 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 2029 ptset.Clear(); 2030 PointsTo(ptset, arg); 2031 for( VectorSetI j(&ptset); j.test(); ++j ) { 2032 uint pt = j.elem; 2033 set_escape_state(pt, PointsToNode::GlobalEscape); 2034 } 2035 } 2036 } 2037 } 2038 } 2039 } 2040 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { 2041 CallNode *call = resproj->in(0)->as_Call(); 2042 uint call_idx = call->_idx; 2043 uint resproj_idx = resproj->_idx; 2044 2045 switch (call->Opcode()) { 2046 case Op_Allocate: 2047 { 2048 Node *k = call->in(AllocateNode::KlassNode); 2049 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2050 assert(kt != NULL, "TypeKlassPtr required."); 2051 ciKlass* cik = kt->klass(); 2052 2053 PointsToNode::EscapeState es; 2054 uint edge_to; 2055 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 2056 !cik->is_instance_klass() || // StressReflectiveCode 2057 cik->as_instance_klass()->has_finalizer()) { 2058 es = PointsToNode::GlobalEscape; 2059 edge_to = _phantom_object; // Could not be worse 2060 } else { 2061 es = PointsToNode::NoEscape; 2062 edge_to = call_idx; 2063 } 2064 set_escape_state(call_idx, es); 2065 add_pointsto_edge(resproj_idx, edge_to); 2066 _processed.set(resproj_idx); 2067 break; 2068 } 2069 2070 case Op_AllocateArray: 2071 { 2072 2073 Node *k = call->in(AllocateNode::KlassNode); 2074 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2075 assert(kt != NULL, "TypeKlassPtr required."); 2076 ciKlass* cik = kt->klass(); 2077 2078 PointsToNode::EscapeState es; 2079 uint edge_to; 2080 if (!cik->is_array_klass()) { // StressReflectiveCode 2081 es = PointsToNode::GlobalEscape; 2082 edge_to = _phantom_object; 2083 } else { 2084 es = PointsToNode::NoEscape; 2085 edge_to = call_idx; 2086 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 2087 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 2088 // Not scalar replaceable if the length is not constant or too big. 2089 ptnode_adr(call_idx)->_scalar_replaceable = false; 2090 } 2091 } 2092 set_escape_state(call_idx, es); 2093 add_pointsto_edge(resproj_idx, edge_to); 2094 _processed.set(resproj_idx); 2095 break; 2096 } 2097 2098 case Op_CallStaticJava: 2099 // For a static call, we know exactly what method is being called. 2100 // Use bytecode estimator to record whether the call's return value escapes 2101 { 2102 bool done = true; 2103 const TypeTuple *r = call->tf()->range(); 2104 const Type* ret_type = NULL; 2105 2106 if (r->cnt() > TypeFunc::Parms) 2107 ret_type = r->field_at(TypeFunc::Parms); 2108 2109 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2110 // _multianewarray functions return a TypeRawPtr. 2111 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { 2112 _processed.set(resproj_idx); 2113 break; // doesn't return a pointer type 2114 } 2115 ciMethod *meth = call->as_CallJava()->method(); 2116 const TypeTuple * d = call->tf()->domain(); 2117 if (meth == NULL) { 2118 // not a Java method, assume global escape 2119 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2120 add_pointsto_edge(resproj_idx, _phantom_object); 2121 } else { 2122 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); 2123 bool copy_dependencies = false; 2124 2125 if (call_analyzer->is_return_allocated()) { 2126 // Returns a newly allocated unescaped object, simply 2127 // update dependency information. 2128 // Mark it as NoEscape so that objects referenced by 2129 // it's fields will be marked as NoEscape at least. 2130 set_escape_state(call_idx, PointsToNode::NoEscape); 2131 add_pointsto_edge(resproj_idx, call_idx); 2132 copy_dependencies = true; 2133 } else if (call_analyzer->is_return_local()) { 2134 // determine whether any arguments are returned 2135 set_escape_state(call_idx, PointsToNode::NoEscape); 2136 bool ret_arg = false; 2137 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2138 const Type* at = d->field_at(i); 2139 2140 if (at->isa_oopptr() != NULL) { 2141 Node *arg = call->in(i)->uncast(); 2142 2143 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 2144 ret_arg = true; 2145 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2146 if (arg_esp->node_type() == PointsToNode::UnknownType) 2147 done = false; 2148 else if (arg_esp->node_type() == PointsToNode::JavaObject) 2149 add_pointsto_edge(resproj_idx, arg->_idx); 2150 else 2151 add_deferred_edge(resproj_idx, arg->_idx); 2152 arg_esp->_hidden_alias = true; 2153 } 2154 } 2155 } 2156 if (done && !ret_arg) { 2157 // Returns unknown object. 2158 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2159 add_pointsto_edge(resproj_idx, _phantom_object); 2160 } 2161 copy_dependencies = true; 2162 } else { 2163 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2164 add_pointsto_edge(resproj_idx, _phantom_object); 2165 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2166 const Type* at = d->field_at(i); 2167 if (at->isa_oopptr() != NULL) { 2168 Node *arg = call->in(i)->uncast(); 2169 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2170 arg_esp->_hidden_alias = true; 2171 } 2172 } 2173 } 2174 if (copy_dependencies) 2175 call_analyzer->copy_dependencies(_compile->dependencies()); 2176 } 2177 if (done) 2178 _processed.set(resproj_idx); 2179 break; 2180 } 2181 2182 default: 2183 // Some other type of call, assume the worst case that the 2184 // returned value, if any, globally escapes. 2185 { 2186 const TypeTuple *r = call->tf()->range(); 2187 if (r->cnt() > TypeFunc::Parms) { 2188 const Type* ret_type = r->field_at(TypeFunc::Parms); 2189 2190 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2191 // _multianewarray functions return a TypeRawPtr. 2192 if (ret_type->isa_ptr() != NULL) { 2193 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2194 add_pointsto_edge(resproj_idx, _phantom_object); 2195 } 2196 } 2197 _processed.set(resproj_idx); 2198 } 2199 } 2200 } 2201 2202 // Populate Connection Graph with Ideal nodes and create simple 2203 // connection graph edges (do not need to check the node_type of inputs 2204 // or to call PointsTo() to walk the connection graph). 2205 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { 2206 if (_processed.test(n->_idx)) 2207 return; // No need to redefine node's state. 2208 2209 if (n->is_Call()) { 2210 // Arguments to allocation and locking don't escape. 2211 if (n->is_Allocate()) { 2212 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); 2213 record_for_optimizer(n); 2214 } else if (n->is_Lock() || n->is_Unlock()) { 2215 // Put Lock and Unlock nodes on IGVN worklist to process them during 2216 // the first IGVN optimization when escape information is still available. 2217 record_for_optimizer(n); 2218 _processed.set(n->_idx); 2219 } else { 2220 // Don't mark as processed since call's arguments have to be processed. 2221 PointsToNode::NodeType nt = PointsToNode::UnknownType; 2222 PointsToNode::EscapeState es = PointsToNode::UnknownEscape; 2223 2224 // Check if a call returns an object. 2225 const TypeTuple *r = n->as_Call()->tf()->range(); 2226 if (r->cnt() > TypeFunc::Parms && 2227 r->field_at(TypeFunc::Parms)->isa_ptr() && 2228 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { 2229 nt = PointsToNode::JavaObject; 2230 if (!n->is_CallStaticJava()) { 2231 // Since the called mathod is statically unknown assume 2232 // the worst case that the returned value globally escapes. 2233 es = PointsToNode::GlobalEscape; 2234 } 2235 } 2236 add_node(n, nt, es, false); 2237 } 2238 return; 2239 } 2240 2241 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 2242 // ThreadLocal has RawPrt type. 2243 switch (n->Opcode()) { 2244 case Op_AddP: 2245 { 2246 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); 2247 break; 2248 } 2249 case Op_CastX2P: 2250 { // "Unsafe" memory access. 2251 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2252 break; 2253 } 2254 case Op_CastPP: 2255 case Op_CheckCastPP: 2256 case Op_EncodeP: 2257 case Op_DecodeN: 2258 { 2259 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2260 int ti = n->in(1)->_idx; 2261 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2262 if (nt == PointsToNode::UnknownType) { 2263 _delayed_worklist.push(n); // Process it later. 2264 break; 2265 } else if (nt == PointsToNode::JavaObject) { 2266 add_pointsto_edge(n->_idx, ti); 2267 } else { 2268 add_deferred_edge(n->_idx, ti); 2269 } 2270 _processed.set(n->_idx); 2271 break; 2272 } 2273 case Op_ConP: 2274 { 2275 // assume all pointer constants globally escape except for null 2276 PointsToNode::EscapeState es; 2277 if (phase->type(n) == TypePtr::NULL_PTR) 2278 es = PointsToNode::NoEscape; 2279 else 2280 es = PointsToNode::GlobalEscape; 2281 2282 add_node(n, PointsToNode::JavaObject, es, true); 2283 break; 2284 } 2285 case Op_ConN: 2286 { 2287 // assume all narrow oop constants globally escape except for null 2288 PointsToNode::EscapeState es; 2289 if (phase->type(n) == TypeNarrowOop::NULL_PTR) 2290 es = PointsToNode::NoEscape; 2291 else 2292 es = PointsToNode::GlobalEscape; 2293 2294 add_node(n, PointsToNode::JavaObject, es, true); 2295 break; 2296 } 2297 case Op_CreateEx: 2298 { 2299 // assume that all exception objects globally escape 2300 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2301 break; 2302 } 2303 case Op_LoadKlass: 2304 case Op_LoadNKlass: 2305 { 2306 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2307 break; 2308 } 2309 case Op_LoadP: 2310 case Op_LoadN: 2311 { 2312 const Type *t = phase->type(n); 2313 if (t->make_ptr() == NULL) { 2314 _processed.set(n->_idx); 2315 return; 2316 } 2317 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2318 break; 2319 } 2320 case Op_Parm: 2321 { 2322 _processed.set(n->_idx); // No need to redefine it state. 2323 uint con = n->as_Proj()->_con; 2324 if (con < TypeFunc::Parms) 2325 return; 2326 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); 2327 if (t->isa_ptr() == NULL) 2328 return; 2329 // We have to assume all input parameters globally escape 2330 // (Note: passing 'false' since _processed is already set). 2331 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); 2332 break; 2333 } 2334 case Op_Phi: 2335 { 2336 const Type *t = n->as_Phi()->type(); 2337 if (t->make_ptr() == NULL) { 2338 // nothing to do if not an oop or narrow oop 2339 _processed.set(n->_idx); 2340 return; 2341 } 2342 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2343 uint i; 2344 for (i = 1; i < n->req() ; i++) { 2345 Node* in = n->in(i); 2346 if (in == NULL) 2347 continue; // ignore NULL 2348 in = in->uncast(); 2349 if (in->is_top() || in == n) 2350 continue; // ignore top or inputs which go back this node 2351 int ti = in->_idx; 2352 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2353 if (nt == PointsToNode::UnknownType) { 2354 break; 2355 } else if (nt == PointsToNode::JavaObject) { 2356 add_pointsto_edge(n->_idx, ti); 2357 } else { 2358 add_deferred_edge(n->_idx, ti); 2359 } 2360 } 2361 if (i >= n->req()) 2362 _processed.set(n->_idx); 2363 else 2364 _delayed_worklist.push(n); 2365 break; 2366 } 2367 case Op_Proj: 2368 { 2369 // we are only interested in the oop result projection from a call 2370 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2371 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2372 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2373 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2374 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2375 int ti = n->in(0)->_idx; 2376 // The call may not be registered yet (since not all its inputs are registered) 2377 // if this is the projection from backbranch edge of Phi. 2378 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) { 2379 process_call_result(n->as_Proj(), phase); 2380 } 2381 if (!_processed.test(n->_idx)) { 2382 // The call's result may need to be processed later if the call 2383 // returns it's argument and the argument is not processed yet. 2384 _delayed_worklist.push(n); 2385 } 2386 break; 2387 } 2388 } 2389 _processed.set(n->_idx); 2390 break; 2391 } 2392 case Op_Return: 2393 { 2394 if( n->req() > TypeFunc::Parms && 2395 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2396 // Treat Return value as LocalVar with GlobalEscape escape state. 2397 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); 2398 int ti = n->in(TypeFunc::Parms)->_idx; 2399 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2400 if (nt == PointsToNode::UnknownType) { 2401 _delayed_worklist.push(n); // Process it later. 2402 break; 2403 } else if (nt == PointsToNode::JavaObject) { 2404 add_pointsto_edge(n->_idx, ti); 2405 } else { 2406 add_deferred_edge(n->_idx, ti); 2407 } 2408 } 2409 _processed.set(n->_idx); 2410 break; 2411 } 2412 case Op_StoreP: 2413 case Op_StoreN: 2414 { 2415 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2416 adr_type = adr_type->make_ptr(); 2417 if (adr_type->isa_oopptr()) { 2418 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2419 } else { 2420 Node* adr = n->in(MemNode::Address); 2421 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && 2422 adr->in(AddPNode::Address)->is_Proj() && 2423 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2424 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2425 // We are computing a raw address for a store captured 2426 // by an Initialize compute an appropriate address type. 2427 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2428 assert(offs != Type::OffsetBot, "offset must be a constant"); 2429 } else { 2430 _processed.set(n->_idx); 2431 return; 2432 } 2433 } 2434 break; 2435 } 2436 case Op_StorePConditional: 2437 case Op_CompareAndSwapP: 2438 case Op_CompareAndSwapN: 2439 { 2440 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2441 adr_type = adr_type->make_ptr(); 2442 if (adr_type->isa_oopptr()) { 2443 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2444 } else { 2445 _processed.set(n->_idx); 2446 return; 2447 } 2448 break; 2449 } 2450 case Op_AryEq: 2451 case Op_StrComp: 2452 case Op_StrEquals: 2453 case Op_StrIndexOf: 2454 { 2455 // char[] arrays passed to string intrinsics are not scalar replaceable. 2456 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2457 break; 2458 } 2459 case Op_ThreadLocal: 2460 { 2461 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); 2462 break; 2463 } 2464 default: 2465 ; 2466 // nothing to do 2467 } 2468 return; 2469 } 2470 2471 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { 2472 uint n_idx = n->_idx; 2473 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered"); 2474 2475 // Don't set processed bit for AddP, LoadP, StoreP since 2476 // they may need more then one pass to process. 2477 if (_processed.test(n_idx)) 2478 return; // No need to redefine node's state. 2479 2480 if (n->is_Call()) { 2481 CallNode *call = n->as_Call(); 2482 process_call_arguments(call, phase); 2483 _processed.set(n_idx); 2484 return; 2485 } 2486 2487 switch (n->Opcode()) { 2488 case Op_AddP: 2489 { 2490 Node *base = get_addp_base(n); 2491 // Create a field edge to this node from everything base could point to. 2492 VectorSet ptset(Thread::current()->resource_area()); 2493 PointsTo(ptset, base); 2494 for( VectorSetI i(&ptset); i.test(); ++i ) { 2495 uint pt = i.elem; 2496 add_field_edge(pt, n_idx, address_offset(n, phase)); 2497 } 2498 break; 2499 } 2500 case Op_CastX2P: 2501 { 2502 assert(false, "Op_CastX2P"); 2503 break; 2504 } 2505 case Op_CastPP: 2506 case Op_CheckCastPP: 2507 case Op_EncodeP: 2508 case Op_DecodeN: 2509 { 2510 int ti = n->in(1)->_idx; 2511 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered"); 2512 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2513 add_pointsto_edge(n_idx, ti); 2514 } else { 2515 add_deferred_edge(n_idx, ti); 2516 } 2517 _processed.set(n_idx); 2518 break; 2519 } 2520 case Op_ConP: 2521 { 2522 assert(false, "Op_ConP"); 2523 break; 2524 } 2525 case Op_ConN: 2526 { 2527 assert(false, "Op_ConN"); 2528 break; 2529 } 2530 case Op_CreateEx: 2531 { 2532 assert(false, "Op_CreateEx"); 2533 break; 2534 } 2535 case Op_LoadKlass: 2536 case Op_LoadNKlass: 2537 { 2538 assert(false, "Op_LoadKlass"); 2539 break; 2540 } 2541 case Op_LoadP: 2542 case Op_LoadN: 2543 { 2544 const Type *t = phase->type(n); 2545 #ifdef ASSERT 2546 if (t->make_ptr() == NULL) 2547 assert(false, "Op_LoadP"); 2548 #endif 2549 2550 Node* adr = n->in(MemNode::Address)->uncast(); 2551 Node* adr_base; 2552 if (adr->is_AddP()) { 2553 adr_base = get_addp_base(adr); 2554 } else { 2555 adr_base = adr; 2556 } 2557 2558 // For everything "adr_base" could point to, create a deferred edge from 2559 // this node to each field with the same offset. 2560 VectorSet ptset(Thread::current()->resource_area()); 2561 PointsTo(ptset, adr_base); 2562 int offset = address_offset(adr, phase); 2563 for( VectorSetI i(&ptset); i.test(); ++i ) { 2564 uint pt = i.elem; 2565 add_deferred_edge_to_fields(n_idx, pt, offset); 2566 } 2567 break; 2568 } 2569 case Op_Parm: 2570 { 2571 assert(false, "Op_Parm"); 2572 break; 2573 } 2574 case Op_Phi: 2575 { 2576 #ifdef ASSERT 2577 const Type *t = n->as_Phi()->type(); 2578 if (t->make_ptr() == NULL) 2579 assert(false, "Op_Phi"); 2580 #endif 2581 for (uint i = 1; i < n->req() ; i++) { 2582 Node* in = n->in(i); 2583 if (in == NULL) 2584 continue; // ignore NULL 2585 in = in->uncast(); 2586 if (in->is_top() || in == n) 2587 continue; // ignore top or inputs which go back this node 2588 int ti = in->_idx; 2589 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2590 assert(nt != PointsToNode::UnknownType, "all nodes should be known"); 2591 if (nt == PointsToNode::JavaObject) { 2592 add_pointsto_edge(n_idx, ti); 2593 } else { 2594 add_deferred_edge(n_idx, ti); 2595 } 2596 } 2597 _processed.set(n_idx); 2598 break; 2599 } 2600 case Op_Proj: 2601 { 2602 // we are only interested in the oop result projection from a call 2603 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2604 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType, 2605 "all nodes should be registered"); 2606 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2607 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2608 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2609 process_call_result(n->as_Proj(), phase); 2610 assert(_processed.test(n_idx), "all call results should be processed"); 2611 break; 2612 } 2613 } 2614 assert(false, "Op_Proj"); 2615 break; 2616 } 2617 case Op_Return: 2618 { 2619 #ifdef ASSERT 2620 if( n->req() <= TypeFunc::Parms || 2621 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2622 assert(false, "Op_Return"); 2623 } 2624 #endif 2625 int ti = n->in(TypeFunc::Parms)->_idx; 2626 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered"); 2627 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2628 add_pointsto_edge(n_idx, ti); 2629 } else { 2630 add_deferred_edge(n_idx, ti); 2631 } 2632 _processed.set(n_idx); 2633 break; 2634 } 2635 case Op_StoreP: 2636 case Op_StoreN: 2637 case Op_StorePConditional: 2638 case Op_CompareAndSwapP: 2639 case Op_CompareAndSwapN: 2640 { 2641 Node *adr = n->in(MemNode::Address); 2642 const Type *adr_type = phase->type(adr)->make_ptr(); 2643 #ifdef ASSERT 2644 if (!adr_type->isa_oopptr()) 2645 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); 2646 #endif 2647 2648 assert(adr->is_AddP(), "expecting an AddP"); 2649 Node *adr_base = get_addp_base(adr); 2650 Node *val = n->in(MemNode::ValueIn)->uncast(); 2651 // For everything "adr_base" could point to, create a deferred edge 2652 // to "val" from each field with the same offset. 2653 VectorSet ptset(Thread::current()->resource_area()); 2654 PointsTo(ptset, adr_base); 2655 for( VectorSetI i(&ptset); i.test(); ++i ) { 2656 uint pt = i.elem; 2657 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); 2658 } 2659 break; 2660 } 2661 case Op_AryEq: 2662 case Op_StrComp: 2663 case Op_StrEquals: 2664 case Op_StrIndexOf: 2665 { 2666 // char[] arrays passed to string intrinsic do not escape but 2667 // they are not scalar replaceable. Adjust escape state for them. 2668 // Start from in(2) edge since in(1) is memory edge. 2669 for (uint i = 2; i < n->req(); i++) { 2670 Node* adr = n->in(i)->uncast(); 2671 const Type *at = phase->type(adr); 2672 if (!adr->is_top() && at->isa_ptr()) { 2673 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 2674 at->isa_ptr() != NULL, "expecting an Ptr"); 2675 if (adr->is_AddP()) { 2676 adr = get_addp_base(adr); 2677 } 2678 // Mark as ArgEscape everything "adr" could point to. 2679 set_escape_state(adr->_idx, PointsToNode::ArgEscape); 2680 } 2681 } 2682 _processed.set(n_idx); 2683 break; 2684 } 2685 case Op_ThreadLocal: 2686 { 2687 assert(false, "Op_ThreadLocal"); 2688 break; 2689 } 2690 default: 2691 // This method should be called only for EA specific nodes. 2692 ShouldNotReachHere(); 2693 } 2694 } 2695 2696 #ifndef PRODUCT 2697 void ConnectionGraph::dump() { 2698 bool first = true; 2699 2700 uint size = nodes_size(); 2701 for (uint ni = 0; ni < size; ni++) { 2702 PointsToNode *ptn = ptnode_adr(ni); 2703 PointsToNode::NodeType ptn_type = ptn->node_type(); 2704 2705 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) 2706 continue; 2707 PointsToNode::EscapeState es = escape_state(ptn->_node); 2708 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { 2709 if (first) { 2710 tty->cr(); 2711 tty->print("======== Connection graph for "); 2712 _compile->method()->print_short_name(); 2713 tty->cr(); 2714 first = false; 2715 } 2716 tty->print("%6d ", ni); 2717 ptn->dump(); 2718 // Print all locals which reference this allocation 2719 for (uint li = ni; li < size; li++) { 2720 PointsToNode *ptn_loc = ptnode_adr(li); 2721 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); 2722 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && 2723 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { 2724 ptnode_adr(li)->dump(false); 2725 } 2726 } 2727 if (Verbose) { 2728 // Print all fields which reference this allocation 2729 for (uint i = 0; i < ptn->edge_count(); i++) { 2730 uint ei = ptn->edge_target(i); 2731 ptnode_adr(ei)->dump(false); 2732 } 2733 } 2734 tty->cr(); 2735 } 2736 } 2737 } 2738 #endif