1 /* 2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/bcEscapeAnalyzer.hpp" 27 #include "libadt/vectset.hpp" 28 #include "memory/allocation.hpp" 29 #include "opto/c2compiler.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/escape.hpp" 34 #include "opto/phaseX.hpp" 35 #include "opto/rootnode.hpp" 36 37 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) { 38 uint v = (targIdx << EdgeShift) + ((uint) et); 39 if (_edges == NULL) { 40 Arena *a = Compile::current()->comp_arena(); 41 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0); 42 } 43 _edges->append_if_missing(v); 44 } 45 46 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) { 47 uint v = (targIdx << EdgeShift) + ((uint) et); 48 49 _edges->remove(v); 50 } 51 52 #ifndef PRODUCT 53 static const char *node_type_names[] = { 54 "UnknownType", 55 "JavaObject", 56 "LocalVar", 57 "Field" 58 }; 59 60 static const char *esc_names[] = { 61 "UnknownEscape", 62 "NoEscape", 63 "ControlEscape", 64 "ArgEscape", 65 "GlobalEscape" 66 }; 67 68 static const char *edge_type_suffix[] = { 69 "?", // UnknownEdge 70 "P", // PointsToEdge 71 "D", // DeferredEdge 72 "F" // FieldEdge 73 }; 74 75 void PointsToNode::dump(bool print_state) const { 76 NodeType nt = node_type(); 77 tty->print("%s ", node_type_names[(int) nt]); 78 if (print_state) { 79 EscapeState es = escape_state(); 80 tty->print("%s ", esc_names[(int) es]); 81 } 82 tty->print("[["); 83 for (uint i = 0; i < edge_count(); i++) { 84 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]); 85 } 86 tty->print("]] "); 87 if (_node == NULL) 88 tty->print_cr("<null>"); 89 else 90 _node->dump(); 91 } 92 #endif 93 94 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) : 95 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()), 96 _processed(C->comp_arena()), 97 pt_ptset(C->comp_arena()), 98 pt_visited(C->comp_arena()), 99 pt_worklist(C->comp_arena(), 4, 0, 0), 100 _collecting(true), 101 _progress(false), 102 _compile(C), 103 _igvn(igvn), 104 _node_map(C->comp_arena()) { 105 106 _phantom_object = C->top()->_idx, 107 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true); 108 109 // Add ConP(#NULL) and ConN(#NULL) nodes. 110 Node* oop_null = igvn->zerocon(T_OBJECT); 111 _oop_null = oop_null->_idx; 112 assert(_oop_null < nodes_size(), "should be created already"); 113 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 114 115 if (UseCompressedOops) { 116 Node* noop_null = igvn->zerocon(T_NARROWOOP); 117 _noop_null = noop_null->_idx; 118 assert(_noop_null < nodes_size(), "should be created already"); 119 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 120 } else { 121 _noop_null = _oop_null; // Should be initialized 122 } 123 } 124 125 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) { 126 PointsToNode *f = ptnode_adr(from_i); 127 PointsToNode *t = ptnode_adr(to_i); 128 129 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 130 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge"); 131 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge"); 132 add_edge(f, to_i, PointsToNode::PointsToEdge); 133 } 134 135 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) { 136 PointsToNode *f = ptnode_adr(from_i); 137 PointsToNode *t = ptnode_adr(to_i); 138 139 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 140 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge"); 141 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge"); 142 // don't add a self-referential edge, this can occur during removal of 143 // deferred edges 144 if (from_i != to_i) 145 add_edge(f, to_i, PointsToNode::DeferredEdge); 146 } 147 148 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 149 const Type *adr_type = phase->type(adr); 150 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 151 adr->in(AddPNode::Address)->is_Proj() && 152 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 153 // We are computing a raw address for a store captured by an Initialize 154 // compute an appropriate address type. AddP cases #3 and #5 (see below). 155 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 156 assert(offs != Type::OffsetBot || 157 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 158 "offset must be a constant or it is initialization of array"); 159 return offs; 160 } 161 const TypePtr *t_ptr = adr_type->isa_ptr(); 162 assert(t_ptr != NULL, "must be a pointer type"); 163 return t_ptr->offset(); 164 } 165 166 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) { 167 PointsToNode *f = ptnode_adr(from_i); 168 PointsToNode *t = ptnode_adr(to_i); 169 170 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 171 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge"); 172 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge"); 173 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets"); 174 t->set_offset(offset); 175 176 add_edge(f, to_i, PointsToNode::FieldEdge); 177 } 178 179 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) { 180 // Don't change non-escaping state of NULL pointer. 181 if (ni == _noop_null || ni == _oop_null) 182 return; 183 PointsToNode *npt = ptnode_adr(ni); 184 PointsToNode::EscapeState old_es = npt->escape_state(); 185 if (es > old_es) 186 npt->set_escape_state(es); 187 } 188 189 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt, 190 PointsToNode::EscapeState es, bool done) { 191 PointsToNode* ptadr = ptnode_adr(n->_idx); 192 ptadr->_node = n; 193 ptadr->set_node_type(nt); 194 195 // inline set_escape_state(idx, es); 196 PointsToNode::EscapeState old_es = ptadr->escape_state(); 197 if (es > old_es) 198 ptadr->set_escape_state(es); 199 200 if (done) 201 _processed.set(n->_idx); 202 } 203 204 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) { 205 uint idx = n->_idx; 206 PointsToNode::EscapeState es; 207 208 // If we are still collecting or there were no non-escaping allocations 209 // we don't know the answer yet 210 if (_collecting) 211 return PointsToNode::UnknownEscape; 212 213 // if the node was created after the escape computation, return 214 // UnknownEscape 215 if (idx >= nodes_size()) 216 return PointsToNode::UnknownEscape; 217 218 es = ptnode_adr(idx)->escape_state(); 219 220 // if we have already computed a value, return it 221 if (es != PointsToNode::UnknownEscape && 222 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject) 223 return es; 224 225 // PointsTo() calls n->uncast() which can return a new ideal node. 226 if (n->uncast()->_idx >= nodes_size()) 227 return PointsToNode::UnknownEscape; 228 229 PointsToNode::EscapeState orig_es = es; 230 231 // compute max escape state of anything this node could point to 232 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) { 233 uint pt = i.elem; 234 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); 235 if (pes > es) 236 es = pes; 237 } 238 if (orig_es != es) { 239 // cache the computed escape state 240 assert(es > orig_es, "should have computed an escape state"); 241 set_escape_state(idx, es); 242 } // orig_es could be PointsToNode::UnknownEscape 243 return es; 244 } 245 246 VectorSet* ConnectionGraph::PointsTo(Node * n) { 247 pt_ptset.Reset(); 248 pt_visited.Reset(); 249 pt_worklist.clear(); 250 251 #ifdef ASSERT 252 Node *orig_n = n; 253 #endif 254 255 n = n->uncast(); 256 PointsToNode* npt = ptnode_adr(n->_idx); 257 258 // If we have a JavaObject, return just that object 259 if (npt->node_type() == PointsToNode::JavaObject) { 260 pt_ptset.set(n->_idx); 261 return &pt_ptset; 262 } 263 #ifdef ASSERT 264 if (npt->_node == NULL) { 265 if (orig_n != n) 266 orig_n->dump(); 267 n->dump(); 268 assert(npt->_node != NULL, "unregistered node"); 269 } 270 #endif 271 pt_worklist.push(n->_idx); 272 while(pt_worklist.length() > 0) { 273 int ni = pt_worklist.pop(); 274 if (pt_visited.test_set(ni)) 275 continue; 276 277 PointsToNode* pn = ptnode_adr(ni); 278 // ensure that all inputs of a Phi have been processed 279 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); 280 281 int edges_processed = 0; 282 uint e_cnt = pn->edge_count(); 283 for (uint e = 0; e < e_cnt; e++) { 284 uint etgt = pn->edge_target(e); 285 PointsToNode::EdgeType et = pn->edge_type(e); 286 if (et == PointsToNode::PointsToEdge) { 287 pt_ptset.set(etgt); 288 edges_processed++; 289 } else if (et == PointsToNode::DeferredEdge) { 290 pt_worklist.push(etgt); 291 edges_processed++; 292 } else { 293 assert(false,"neither PointsToEdge or DeferredEdge"); 294 } 295 } 296 if (edges_processed == 0) { 297 // no deferred or pointsto edges found. Assume the value was set 298 // outside this method. Add the phantom object to the pointsto set. 299 pt_ptset.set(_phantom_object); 300 } 301 } 302 return &pt_ptset; 303 } 304 305 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { 306 // This method is most expensive during ConnectionGraph construction. 307 // Reuse vectorSet and an additional growable array for deferred edges. 308 deferred_edges->clear(); 309 visited->Reset(); 310 311 visited->set(ni); 312 PointsToNode *ptn = ptnode_adr(ni); 313 314 // Mark current edges as visited and move deferred edges to separate array. 315 for (uint i = 0; i < ptn->edge_count(); ) { 316 uint t = ptn->edge_target(i); 317 #ifdef ASSERT 318 assert(!visited->test_set(t), "expecting no duplications"); 319 #else 320 visited->set(t); 321 #endif 322 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { 323 ptn->remove_edge(t, PointsToNode::DeferredEdge); 324 deferred_edges->append(t); 325 } else { 326 i++; 327 } 328 } 329 for (int next = 0; next < deferred_edges->length(); ++next) { 330 uint t = deferred_edges->at(next); 331 PointsToNode *ptt = ptnode_adr(t); 332 uint e_cnt = ptt->edge_count(); 333 for (uint e = 0; e < e_cnt; e++) { 334 uint etgt = ptt->edge_target(e); 335 if (visited->test_set(etgt)) 336 continue; 337 338 PointsToNode::EdgeType et = ptt->edge_type(e); 339 if (et == PointsToNode::PointsToEdge) { 340 add_pointsto_edge(ni, etgt); 341 if(etgt == _phantom_object) { 342 // Special case - field set outside (globally escaping). 343 set_escape_state(ni, PointsToNode::GlobalEscape); 344 } 345 } else if (et == PointsToNode::DeferredEdge) { 346 deferred_edges->append(etgt); 347 } else { 348 assert(false,"invalid connection graph"); 349 } 350 } 351 } 352 } 353 354 355 // Add an edge to node given by "to_i" from any field of adr_i whose offset 356 // matches "offset" A deferred edge is added if to_i is a LocalVar, and 357 // a pointsto edge is added if it is a JavaObject 358 359 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { 360 PointsToNode* an = ptnode_adr(adr_i); 361 PointsToNode* to = ptnode_adr(to_i); 362 bool deferred = (to->node_type() == PointsToNode::LocalVar); 363 364 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 (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 370 if (deferred) 371 add_deferred_edge(fi, to_i); 372 else 373 add_pointsto_edge(fi, to_i); 374 } 375 } 376 } 377 378 // Add a deferred edge from node given by "from_i" to any field of adr_i 379 // whose offset matches "offset". 380 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { 381 PointsToNode* an = ptnode_adr(adr_i); 382 bool is_alloc = an->_node->is_Allocate(); 383 for (uint fe = 0; fe < an->edge_count(); fe++) { 384 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 385 int fi = an->edge_target(fe); 386 PointsToNode* pf = ptnode_adr(fi); 387 int offset = pf->offset(); 388 if (!is_alloc) { 389 // Assume the field was set outside this method if it is not Allocation 390 add_pointsto_edge(fi, _phantom_object); 391 } 392 if (offset == offs || offset == Type::OffsetBot || offs == Type::OffsetBot) { 393 add_deferred_edge(from_i, fi); 394 } 395 } 396 } 397 398 // Helper functions 399 400 static Node* get_addp_base(Node *addp) { 401 assert(addp->is_AddP(), "must be AddP"); 402 // 403 // AddP cases for Base and Address inputs: 404 // case #1. Direct object's field reference: 405 // Allocate 406 // | 407 // Proj #5 ( oop result ) 408 // | 409 // CheckCastPP (cast to instance type) 410 // | | 411 // AddP ( base == address ) 412 // 413 // case #2. Indirect object's field reference: 414 // Phi 415 // | 416 // CastPP (cast to instance type) 417 // | | 418 // AddP ( base == address ) 419 // 420 // case #3. Raw object's field reference for Initialize node: 421 // Allocate 422 // | 423 // Proj #5 ( oop result ) 424 // top | 425 // \ | 426 // AddP ( base == top ) 427 // 428 // case #4. Array's element reference: 429 // {CheckCastPP | CastPP} 430 // | | | 431 // | AddP ( array's element offset ) 432 // | | 433 // AddP ( array's offset ) 434 // 435 // case #5. Raw object's field reference for arraycopy stub call: 436 // The inline_native_clone() case when the arraycopy stub is called 437 // after the allocation before Initialize and CheckCastPP nodes. 438 // Allocate 439 // | 440 // Proj #5 ( oop result ) 441 // | | 442 // AddP ( base == address ) 443 // 444 // case #6. Constant Pool, ThreadLocal, CastX2P or 445 // Raw object's field reference: 446 // {ConP, ThreadLocal, CastX2P, raw Load} 447 // top | 448 // \ | 449 // AddP ( base == top ) 450 // 451 // case #7. Klass's field reference. 452 // LoadKlass 453 // | | 454 // AddP ( base == address ) 455 // 456 // case #8. narrow Klass's field reference. 457 // LoadNKlass 458 // | 459 // DecodeN 460 // | | 461 // AddP ( base == address ) 462 // 463 Node *base = addp->in(AddPNode::Base)->uncast(); 464 if (base->is_top()) { // The AddP case #3 and #6. 465 base = addp->in(AddPNode::Address)->uncast(); 466 while (base->is_AddP()) { 467 // Case #6 (unsafe access) may have several chained AddP nodes. 468 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only"); 469 base = base->in(AddPNode::Address)->uncast(); 470 } 471 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || 472 base->Opcode() == Op_CastX2P || base->is_DecodeN() || 473 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || 474 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); 475 } 476 return base; 477 } 478 479 static Node* find_second_addp(Node* addp, Node* n) { 480 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 481 482 Node* addp2 = addp->raw_out(0); 483 if (addp->outcnt() == 1 && addp2->is_AddP() && 484 addp2->in(AddPNode::Base) == n && 485 addp2->in(AddPNode::Address) == addp) { 486 487 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 488 // 489 // Find array's offset to push it on worklist first and 490 // as result process an array's element offset first (pushed second) 491 // to avoid CastPP for the array's offset. 492 // Otherwise the inserted CastPP (LocalVar) will point to what 493 // the AddP (Field) points to. Which would be wrong since 494 // the algorithm expects the CastPP has the same point as 495 // as AddP's base CheckCastPP (LocalVar). 496 // 497 // ArrayAllocation 498 // | 499 // CheckCastPP 500 // | 501 // memProj (from ArrayAllocation CheckCastPP) 502 // | || 503 // | || Int (element index) 504 // | || | ConI (log(element size)) 505 // | || | / 506 // | || LShift 507 // | || / 508 // | AddP (array's element offset) 509 // | | 510 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 511 // | / / 512 // AddP (array's offset) 513 // | 514 // Load/Store (memory operation on array's element) 515 // 516 return addp2; 517 } 518 return NULL; 519 } 520 521 // 522 // Adjust the type and inputs of an AddP which computes the 523 // address of a field of an instance 524 // 525 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { 526 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 527 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 528 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 529 if (t == NULL) { 530 // We are computing a raw address for a store captured by an Initialize 531 // compute an appropriate address type (cases #3 and #5). 532 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 533 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 534 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 535 assert(offs != Type::OffsetBot, "offset must be a constant"); 536 t = base_t->add_offset(offs)->is_oopptr(); 537 } 538 int inst_id = base_t->instance_id(); 539 assert(!t->is_known_instance() || t->instance_id() == inst_id, 540 "old type must be non-instance or match new type"); 541 542 // The type 't' could be subclass of 'base_t'. 543 // As result t->offset() could be large then base_t's size and it will 544 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 545 // constructor verifies correctness of the offset. 546 // 547 // It could happened on subclass's branch (from the type profiling 548 // inlining) which was not eliminated during parsing since the exactness 549 // of the allocation type was not propagated to the subclass type check. 550 // 551 // Or the type 't' could be not related to 'base_t' at all. 552 // It could happened when CHA type is different from MDO type on a dead path 553 // (for example, from instanceof check) which is not collapsed during parsing. 554 // 555 // Do nothing for such AddP node and don't process its users since 556 // this code branch will go away. 557 // 558 if (!t->is_known_instance() && 559 !base_t->klass()->is_subtype_of(t->klass())) { 560 return false; // bail out 561 } 562 563 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 564 // Do NOT remove the next line: ensure a new alias index is allocated 565 // for the instance type. Note: C++ will not remove it since the call 566 // has side effect. 567 int alias_idx = _compile->get_alias_index(tinst); 568 igvn->set_type(addp, tinst); 569 // record the allocation in the node map 570 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered"); 571 set_map(addp->_idx, get_map(base->_idx)); 572 573 // Set addp's Base and Address to 'base'. 574 Node *abase = addp->in(AddPNode::Base); 575 Node *adr = addp->in(AddPNode::Address); 576 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 577 adr->in(0)->_idx == (uint)inst_id) { 578 // Skip AddP cases #3 and #5. 579 } else { 580 assert(!abase->is_top(), "sanity"); // AddP case #3 581 if (abase != base) { 582 igvn->hash_delete(addp); 583 addp->set_req(AddPNode::Base, base); 584 if (abase == adr) { 585 addp->set_req(AddPNode::Address, base); 586 } else { 587 // AddP case #4 (adr is array's element offset AddP node) 588 #ifdef ASSERT 589 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 590 assert(adr->is_AddP() && atype != NULL && 591 atype->instance_id() == inst_id, "array's element offset should be processed first"); 592 #endif 593 } 594 igvn->hash_insert(addp); 595 } 596 } 597 // Put on IGVN worklist since at least addp's type was changed above. 598 record_for_optimizer(addp); 599 return true; 600 } 601 602 // 603 // Create a new version of orig_phi if necessary. Returns either the newly 604 // created phi or an existing phi. Sets create_new to indicate whether a new 605 // phi was created. Cache the last newly created phi in the node map. 606 // 607 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { 608 Compile *C = _compile; 609 new_created = false; 610 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 611 // nothing to do if orig_phi is bottom memory or matches alias_idx 612 if (phi_alias_idx == alias_idx) { 613 return orig_phi; 614 } 615 // Have we recently created a Phi for this alias index? 616 PhiNode *result = get_map_phi(orig_phi->_idx); 617 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 618 return result; 619 } 620 // Previous check may fail when the same wide memory Phi was split into Phis 621 // for different memory slices. Search all Phis for this region. 622 if (result != NULL) { 623 Node* region = orig_phi->in(0); 624 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 625 Node* phi = region->fast_out(i); 626 if (phi->is_Phi() && 627 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 628 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 629 return phi->as_Phi(); 630 } 631 } 632 } 633 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { 634 if (C->do_escape_analysis() == true && !C->failing()) { 635 // Retry compilation without escape analysis. 636 // If this is the first failure, the sentinel string will "stick" 637 // to the Compile object, and the C2Compiler will see it and retry. 638 C->record_failure(C2Compiler::retry_no_escape_analysis()); 639 } 640 return NULL; 641 } 642 orig_phi_worklist.append_if_missing(orig_phi); 643 const TypePtr *atype = C->get_adr_type(alias_idx); 644 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 645 C->copy_node_notes_to(result, orig_phi); 646 igvn->set_type(result, result->bottom_type()); 647 record_for_optimizer(result); 648 649 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;) 650 assert(pn == NULL || pn == orig_phi, "wrong node"); 651 set_map(orig_phi->_idx, result); 652 ptnode_adr(orig_phi->_idx)->_node = orig_phi; 653 654 new_created = true; 655 return result; 656 } 657 658 // 659 // Return a new version of Memory Phi "orig_phi" with the inputs having the 660 // specified alias index. 661 // 662 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { 663 664 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 665 Compile *C = _compile; 666 bool new_phi_created; 667 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); 668 if (!new_phi_created) { 669 return result; 670 } 671 672 GrowableArray<PhiNode *> phi_list; 673 GrowableArray<uint> cur_input; 674 675 PhiNode *phi = orig_phi; 676 uint idx = 1; 677 bool finished = false; 678 while(!finished) { 679 while (idx < phi->req()) { 680 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); 681 if (mem != NULL && mem->is_Phi()) { 682 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); 683 if (new_phi_created) { 684 // found an phi for which we created a new split, push current one on worklist and begin 685 // processing new one 686 phi_list.push(phi); 687 cur_input.push(idx); 688 phi = mem->as_Phi(); 689 result = newphi; 690 idx = 1; 691 continue; 692 } else { 693 mem = newphi; 694 } 695 } 696 if (C->failing()) { 697 return NULL; 698 } 699 result->set_req(idx++, mem); 700 } 701 #ifdef ASSERT 702 // verify that the new Phi has an input for each input of the original 703 assert( phi->req() == result->req(), "must have same number of inputs."); 704 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 705 #endif 706 // Check if all new phi's inputs have specified alias index. 707 // Otherwise use old phi. 708 for (uint i = 1; i < phi->req(); i++) { 709 Node* in = result->in(i); 710 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 711 } 712 // we have finished processing a Phi, see if there are any more to do 713 finished = (phi_list.length() == 0 ); 714 if (!finished) { 715 phi = phi_list.pop(); 716 idx = cur_input.pop(); 717 PhiNode *prev_result = get_map_phi(phi->_idx); 718 prev_result->set_req(idx++, result); 719 result = prev_result; 720 } 721 } 722 return result; 723 } 724 725 726 // 727 // The next methods are derived from methods in MemNode. 728 // 729 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) { 730 Node *mem = mmem; 731 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 732 // means an array I have not precisely typed yet. Do not do any 733 // alias stuff with it any time soon. 734 if( toop->base() != Type::AnyPtr && 735 !(toop->klass() != NULL && 736 toop->klass()->is_java_lang_Object() && 737 toop->offset() == Type::OffsetBot) ) { 738 mem = mmem->memory_at(alias_idx); 739 // Update input if it is progress over what we have now 740 } 741 return mem; 742 } 743 744 // 745 // Move memory users to their memory slices. 746 // 747 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) { 748 Compile* C = _compile; 749 750 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 751 assert(tp != NULL, "ptr type"); 752 int alias_idx = C->get_alias_index(tp); 753 int general_idx = C->get_general_index(alias_idx); 754 755 // Move users first 756 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 757 Node* use = n->fast_out(i); 758 if (use->is_MergeMem()) { 759 MergeMemNode* mmem = use->as_MergeMem(); 760 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 761 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 762 continue; // Nothing to do 763 } 764 // Replace previous general reference to mem node. 765 uint orig_uniq = C->unique(); 766 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 767 assert(orig_uniq == C->unique(), "no new nodes"); 768 mmem->set_memory_at(general_idx, m); 769 --imax; 770 --i; 771 } else if (use->is_MemBar()) { 772 assert(!use->is_Initialize(), "initializing stores should not be moved"); 773 if (use->req() > MemBarNode::Precedent && 774 use->in(MemBarNode::Precedent) == n) { 775 // Don't move related membars. 776 record_for_optimizer(use); 777 continue; 778 } 779 tp = use->as_MemBar()->adr_type()->isa_ptr(); 780 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 781 alias_idx == general_idx) { 782 continue; // Nothing to do 783 } 784 // Move to general memory slice. 785 uint orig_uniq = C->unique(); 786 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 787 assert(orig_uniq == C->unique(), "no new nodes"); 788 igvn->hash_delete(use); 789 imax -= use->replace_edge(n, m); 790 igvn->hash_insert(use); 791 record_for_optimizer(use); 792 --i; 793 #ifdef ASSERT 794 } else if (use->is_Mem()) { 795 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 796 // Don't move related cardmark. 797 continue; 798 } 799 // Memory nodes should have new memory input. 800 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 801 assert(tp != NULL, "ptr type"); 802 int idx = C->get_alias_index(tp); 803 assert(get_map(use->_idx) != NULL || idx == alias_idx, 804 "Following memory nodes should have new memory input or be on the same memory slice"); 805 } else if (use->is_Phi()) { 806 // Phi nodes should be split and moved already. 807 tp = use->as_Phi()->adr_type()->isa_ptr(); 808 assert(tp != NULL, "ptr type"); 809 int idx = C->get_alias_index(tp); 810 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 811 } else { 812 use->dump(); 813 assert(false, "should not be here"); 814 #endif 815 } 816 } 817 } 818 819 // 820 // Search memory chain of "mem" to find a MemNode whose address 821 // is the specified alias index. 822 // 823 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { 824 if (orig_mem == NULL) 825 return orig_mem; 826 Compile* C = phase->C; 827 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr(); 828 bool is_instance = (toop != NULL) && toop->is_known_instance(); 829 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 830 Node *prev = NULL; 831 Node *result = orig_mem; 832 while (prev != result) { 833 prev = result; 834 if (result == start_mem) 835 break; // hit one of our sentinels 836 if (result->is_Mem()) { 837 const Type *at = phase->type(result->in(MemNode::Address)); 838 if (at == Type::TOP) 839 break; // Dead 840 assert (at->isa_ptr() != NULL, "pointer type required."); 841 int idx = C->get_alias_index(at->is_ptr()); 842 if (idx == alias_idx) 843 break; // Found 844 if (!is_instance && (at->isa_oopptr() == NULL || 845 !at->is_oopptr()->is_known_instance())) { 846 break; // Do not skip store to general memory slice. 847 } 848 result = result->in(MemNode::Memory); 849 } 850 if (!is_instance) 851 continue; // don't search further for non-instance types 852 // skip over a call which does not affect this memory slice 853 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 854 Node *proj_in = result->in(0); 855 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) { 856 break; // hit one of our sentinels 857 } else if (proj_in->is_Call()) { 858 CallNode *call = proj_in->as_Call(); 859 if (!call->may_modify(toop, phase)) { 860 result = call->in(TypeFunc::Memory); 861 } 862 } else if (proj_in->is_Initialize()) { 863 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 864 // Stop if this is the initialization for the object instance which 865 // which contains this memory slice, otherwise skip over it. 866 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) { 867 result = proj_in->in(TypeFunc::Memory); 868 } 869 } else if (proj_in->is_MemBar()) { 870 result = proj_in->in(TypeFunc::Memory); 871 } 872 } else if (result->is_MergeMem()) { 873 MergeMemNode *mmem = result->as_MergeMem(); 874 result = step_through_mergemem(mmem, alias_idx, toop); 875 if (result == mmem->base_memory()) { 876 // Didn't find instance memory, search through general slice recursively. 877 result = mmem->memory_at(C->get_general_index(alias_idx)); 878 result = find_inst_mem(result, alias_idx, orig_phis, phase); 879 if (C->failing()) { 880 return NULL; 881 } 882 mmem->set_memory_at(alias_idx, result); 883 } 884 } else if (result->is_Phi() && 885 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 886 Node *un = result->as_Phi()->unique_input(phase); 887 if (un != NULL) { 888 orig_phis.append_if_missing(result->as_Phi()); 889 result = un; 890 } else { 891 break; 892 } 893 } else if (result->is_ClearArray()) { 894 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) { 895 // Can not bypass initialization of the instance 896 // we are looking for. 897 break; 898 } 899 // Otherwise skip it (the call updated 'result' value). 900 } else if (result->Opcode() == Op_SCMemProj) { 901 assert(result->in(0)->is_LoadStore(), "sanity"); 902 const Type *at = phase->type(result->in(0)->in(MemNode::Address)); 903 if (at != Type::TOP) { 904 assert (at->isa_ptr() != NULL, "pointer type required."); 905 int idx = C->get_alias_index(at->is_ptr()); 906 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 907 break; 908 } 909 result = result->in(0)->in(MemNode::Memory); 910 } 911 } 912 if (result->is_Phi()) { 913 PhiNode *mphi = result->as_Phi(); 914 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 915 const TypePtr *t = mphi->adr_type(); 916 if (!is_instance) { 917 // Push all non-instance Phis on the orig_phis worklist to update inputs 918 // during Phase 4 if needed. 919 orig_phis.append_if_missing(mphi); 920 } else if (C->get_alias_index(t) != alias_idx) { 921 // Create a new Phi with the specified alias index type. 922 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); 923 } 924 } 925 // the result is either MemNode, PhiNode, InitializeNode. 926 return result; 927 } 928 929 // 930 // Convert the types of unescaped object to instance types where possible, 931 // propagate the new type information through the graph, and update memory 932 // edges and MergeMem inputs to reflect the new type. 933 // 934 // We start with allocations (and calls which may be allocations) on alloc_worklist. 935 // The processing is done in 4 phases: 936 // 937 // Phase 1: Process possible allocations from alloc_worklist. Create instance 938 // types for the CheckCastPP for allocations where possible. 939 // Propagate the the new types through users as follows: 940 // casts and Phi: push users on alloc_worklist 941 // AddP: cast Base and Address inputs to the instance type 942 // push any AddP users on alloc_worklist and push any memnode 943 // users onto memnode_worklist. 944 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 945 // search the Memory chain for a store with the appropriate type 946 // address type. If a Phi is found, create a new version with 947 // the appropriate memory slices from each of the Phi inputs. 948 // For stores, process the users as follows: 949 // MemNode: push on memnode_worklist 950 // MergeMem: push on mergemem_worklist 951 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 952 // moving the first node encountered of each instance type to the 953 // the input corresponding to its alias index. 954 // appropriate memory slice. 955 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 956 // 957 // In the following example, the CheckCastPP nodes are the cast of allocation 958 // results and the allocation of node 29 is unescaped and eligible to be an 959 // instance type. 960 // 961 // We start with: 962 // 963 // 7 Parm #memory 964 // 10 ConI "12" 965 // 19 CheckCastPP "Foo" 966 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 967 // 29 CheckCastPP "Foo" 968 // 30 AddP _ 29 29 10 Foo+12 alias_index=4 969 // 970 // 40 StoreP 25 7 20 ... alias_index=4 971 // 50 StoreP 35 40 30 ... alias_index=4 972 // 60 StoreP 45 50 20 ... alias_index=4 973 // 70 LoadP _ 60 30 ... alias_index=4 974 // 80 Phi 75 50 60 Memory alias_index=4 975 // 90 LoadP _ 80 30 ... alias_index=4 976 // 100 LoadP _ 80 20 ... alias_index=4 977 // 978 // 979 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24 980 // and creating a new alias index for node 30. This gives: 981 // 982 // 7 Parm #memory 983 // 10 ConI "12" 984 // 19 CheckCastPP "Foo" 985 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 986 // 29 CheckCastPP "Foo" iid=24 987 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 988 // 989 // 40 StoreP 25 7 20 ... alias_index=4 990 // 50 StoreP 35 40 30 ... alias_index=6 991 // 60 StoreP 45 50 20 ... alias_index=4 992 // 70 LoadP _ 60 30 ... alias_index=6 993 // 80 Phi 75 50 60 Memory alias_index=4 994 // 90 LoadP _ 80 30 ... alias_index=6 995 // 100 LoadP _ 80 20 ... alias_index=4 996 // 997 // In phase 2, new memory inputs are computed for the loads and stores, 998 // And a new version of the phi is created. In phase 4, the inputs to 999 // node 80 are updated and then the memory nodes are updated with the 1000 // values computed in phase 2. This results in: 1001 // 1002 // 7 Parm #memory 1003 // 10 ConI "12" 1004 // 19 CheckCastPP "Foo" 1005 // 20 AddP _ 19 19 10 Foo+12 alias_index=4 1006 // 29 CheckCastPP "Foo" iid=24 1007 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 1008 // 1009 // 40 StoreP 25 7 20 ... alias_index=4 1010 // 50 StoreP 35 7 30 ... alias_index=6 1011 // 60 StoreP 45 40 20 ... alias_index=4 1012 // 70 LoadP _ 50 30 ... alias_index=6 1013 // 80 Phi 75 40 60 Memory alias_index=4 1014 // 120 Phi 75 50 50 Memory alias_index=6 1015 // 90 LoadP _ 120 30 ... alias_index=6 1016 // 100 LoadP _ 80 20 ... alias_index=4 1017 // 1018 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 1019 GrowableArray<Node *> memnode_worklist; 1020 GrowableArray<PhiNode *> orig_phis; 1021 1022 PhaseIterGVN *igvn = _igvn; 1023 uint new_index_start = (uint) _compile->num_alias_types(); 1024 Arena* arena = Thread::current()->resource_area(); 1025 VectorSet visited(arena); 1026 1027 1028 // Phase 1: Process possible allocations from alloc_worklist. 1029 // Create instance types for the CheckCastPP for allocations where possible. 1030 // 1031 // (Note: don't forget to change the order of the second AddP node on 1032 // the alloc_worklist if the order of the worklist processing is changed, 1033 // see the comment in find_second_addp().) 1034 // 1035 while (alloc_worklist.length() != 0) { 1036 Node *n = alloc_worklist.pop(); 1037 uint ni = n->_idx; 1038 const TypeOopPtr* tinst = NULL; 1039 if (n->is_Call()) { 1040 CallNode *alloc = n->as_Call(); 1041 // copy escape information to call node 1042 PointsToNode* ptn = ptnode_adr(alloc->_idx); 1043 PointsToNode::EscapeState es = escape_state(alloc); 1044 // We have an allocation or call which returns a Java object, 1045 // see if it is unescaped. 1046 if (es != PointsToNode::NoEscape) 1047 continue; 1048 1049 // Find CheckCastPP for the allocate or for the return value of a call 1050 n = alloc->result_cast(); 1051 if (n == NULL) { // No uses except Initialize node 1052 if (alloc->is_Allocate()) { 1053 // Set the scalar_replaceable flag for allocation 1054 // so it could be eliminated if it has no uses. 1055 alloc->as_Allocate()->_is_scalar_replaceable = true; 1056 } 1057 continue; 1058 } 1059 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 1060 assert(!alloc->is_Allocate(), "allocation should have unique type"); 1061 continue; 1062 } 1063 1064 // The inline code for Object.clone() casts the allocation result to 1065 // java.lang.Object and then to the actual type of the allocated 1066 // object. Detect this case and use the second cast. 1067 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 1068 // the allocation result is cast to java.lang.Object and then 1069 // to the actual Array type. 1070 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 1071 && (alloc->is_AllocateArray() || 1072 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 1073 Node *cast2 = NULL; 1074 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1075 Node *use = n->fast_out(i); 1076 if (use->is_CheckCastPP()) { 1077 cast2 = use; 1078 break; 1079 } 1080 } 1081 if (cast2 != NULL) { 1082 n = cast2; 1083 } else { 1084 // Non-scalar replaceable if the allocation type is unknown statically 1085 // (reflection allocation), the object can't be restored during 1086 // deoptimization without precise type. 1087 continue; 1088 } 1089 } 1090 if (alloc->is_Allocate()) { 1091 // Set the scalar_replaceable flag for allocation 1092 // so it could be eliminated. 1093 alloc->as_Allocate()->_is_scalar_replaceable = true; 1094 } 1095 set_escape_state(n->_idx, es); 1096 // in order for an object to be scalar-replaceable, it must be: 1097 // - a direct allocation (not a call returning an object) 1098 // - non-escaping 1099 // - eligible to be a unique type 1100 // - not determined to be ineligible by escape analysis 1101 assert(ptnode_adr(alloc->_idx)->_node != NULL && 1102 ptnode_adr(n->_idx)->_node != NULL, "should be registered"); 1103 set_map(alloc->_idx, n); 1104 set_map(n->_idx, alloc); 1105 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 1106 if (t == NULL) 1107 continue; // not a TypeOopPtr 1108 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 1109 igvn->hash_delete(n); 1110 igvn->set_type(n, tinst); 1111 n->raise_bottom_type(tinst); 1112 igvn->hash_insert(n); 1113 record_for_optimizer(n); 1114 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) { 1115 1116 // First, put on the worklist all Field edges from Connection Graph 1117 // which is more accurate then putting immediate users from Ideal Graph. 1118 for (uint e = 0; e < ptn->edge_count(); e++) { 1119 Node *use = ptnode_adr(ptn->edge_target(e))->_node; 1120 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), 1121 "only AddP nodes are Field edges in CG"); 1122 if (use->outcnt() > 0) { // Don't process dead nodes 1123 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 1124 if (addp2 != NULL) { 1125 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1126 alloc_worklist.append_if_missing(addp2); 1127 } 1128 alloc_worklist.append_if_missing(use); 1129 } 1130 } 1131 1132 // An allocation may have an Initialize which has raw stores. Scan 1133 // the users of the raw allocation result and push AddP users 1134 // on alloc_worklist. 1135 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 1136 assert (raw_result != NULL, "must have an allocation result"); 1137 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 1138 Node *use = raw_result->fast_out(i); 1139 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 1140 Node* addp2 = find_second_addp(use, raw_result); 1141 if (addp2 != NULL) { 1142 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1143 alloc_worklist.append_if_missing(addp2); 1144 } 1145 alloc_worklist.append_if_missing(use); 1146 } else if (use->is_MemBar()) { 1147 memnode_worklist.append_if_missing(use); 1148 } 1149 } 1150 } 1151 } else if (n->is_AddP()) { 1152 VectorSet* ptset = PointsTo(get_addp_base(n)); 1153 assert(ptset->Size() == 1, "AddP address is unique"); 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 *base = get_map(elem); // CheckCastPP node 1160 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path 1161 tinst = igvn->type(base)->isa_oopptr(); 1162 } else if (n->is_Phi() || 1163 n->is_CheckCastPP() || 1164 n->is_EncodeP() || 1165 n->is_DecodeN() || 1166 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 1167 if (visited.test_set(n->_idx)) { 1168 assert(n->is_Phi(), "loops only through Phi's"); 1169 continue; // already processed 1170 } 1171 VectorSet* ptset = PointsTo(n); 1172 if (ptset->Size() == 1) { 1173 uint elem = ptset->getelem(); // Allocation node's index 1174 if (elem == _phantom_object) { 1175 assert(false, "escaped allocation"); 1176 continue; // Assume the value was set outside this method. 1177 } 1178 Node *val = get_map(elem); // CheckCastPP node 1179 TypeNode *tn = n->as_Type(); 1180 tinst = igvn->type(val)->isa_oopptr(); 1181 assert(tinst != NULL && tinst->is_known_instance() && 1182 (uint)tinst->instance_id() == elem , "instance type expected."); 1183 1184 const Type *tn_type = igvn->type(tn); 1185 const TypeOopPtr *tn_t; 1186 if (tn_type->isa_narrowoop()) { 1187 tn_t = tn_type->make_ptr()->isa_oopptr(); 1188 } else { 1189 tn_t = tn_type->isa_oopptr(); 1190 } 1191 1192 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 1193 if (tn_type->isa_narrowoop()) { 1194 tn_type = tinst->make_narrowoop(); 1195 } else { 1196 tn_type = tinst; 1197 } 1198 igvn->hash_delete(tn); 1199 igvn->set_type(tn, tn_type); 1200 tn->set_type(tn_type); 1201 igvn->hash_insert(tn); 1202 record_for_optimizer(n); 1203 } else { 1204 assert(tn_type == TypePtr::NULL_PTR || 1205 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 1206 "unexpected type"); 1207 continue; // Skip dead path with different type 1208 } 1209 } 1210 } else { 1211 debug_only(n->dump();) 1212 assert(false, "EA: unexpected node"); 1213 continue; 1214 } 1215 // push allocation's users on appropriate worklist 1216 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1217 Node *use = n->fast_out(i); 1218 if(use->is_Mem() && use->in(MemNode::Address) == n) { 1219 // Load/store to instance's field 1220 memnode_worklist.append_if_missing(use); 1221 } else if (use->is_MemBar()) { 1222 memnode_worklist.append_if_missing(use); 1223 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 1224 Node* addp2 = find_second_addp(use, n); 1225 if (addp2 != NULL) { 1226 alloc_worklist.append_if_missing(addp2); 1227 } 1228 alloc_worklist.append_if_missing(use); 1229 } else if (use->is_Phi() || 1230 use->is_CheckCastPP() || 1231 use->is_EncodeP() || 1232 use->is_DecodeN() || 1233 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 1234 alloc_worklist.append_if_missing(use); 1235 #ifdef ASSERT 1236 } else if (use->is_Mem()) { 1237 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 1238 } else if (use->is_MergeMem()) { 1239 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1240 } else if (use->is_SafePoint()) { 1241 // Look for MergeMem nodes for calls which reference unique allocation 1242 // (through CheckCastPP nodes) even for debug info. 1243 Node* m = use->in(TypeFunc::Memory); 1244 if (m->is_MergeMem()) { 1245 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1246 } 1247 } else { 1248 uint op = use->Opcode(); 1249 if (!(op == Op_CmpP || op == Op_Conv2B || 1250 op == Op_CastP2X || op == Op_StoreCM || 1251 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 1252 op == Op_StrEquals || op == Op_StrIndexOf)) { 1253 n->dump(); 1254 use->dump(); 1255 assert(false, "EA: missing allocation reference path"); 1256 } 1257 #endif 1258 } 1259 } 1260 1261 } 1262 // New alias types were created in split_AddP(). 1263 uint new_index_end = (uint) _compile->num_alias_types(); 1264 1265 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 1266 // compute new values for Memory inputs (the Memory inputs are not 1267 // actually updated until phase 4.) 1268 if (memnode_worklist.length() == 0) 1269 return; // nothing to do 1270 1271 while (memnode_worklist.length() != 0) { 1272 Node *n = memnode_worklist.pop(); 1273 if (visited.test_set(n->_idx)) 1274 continue; 1275 if (n->is_Phi() || n->is_ClearArray()) { 1276 // we don't need to do anything, but the users must be pushed 1277 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 1278 // we don't need to do anything, but the users must be pushed 1279 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 1280 if (n == NULL) 1281 continue; 1282 } else { 1283 assert(n->is_Mem(), "memory node required."); 1284 Node *addr = n->in(MemNode::Address); 1285 const Type *addr_t = igvn->type(addr); 1286 if (addr_t == Type::TOP) 1287 continue; 1288 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 1289 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 1290 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 1291 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); 1292 if (_compile->failing()) { 1293 return; 1294 } 1295 if (mem != n->in(MemNode::Memory)) { 1296 // We delay the memory edge update since we need old one in 1297 // MergeMem code below when instances memory slices are separated. 1298 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;) 1299 assert(pn == NULL || pn == n, "wrong node"); 1300 set_map(n->_idx, mem); 1301 ptnode_adr(n->_idx)->_node = n; 1302 } 1303 if (n->is_Load()) { 1304 continue; // don't push users 1305 } else if (n->is_LoadStore()) { 1306 // get the memory projection 1307 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1308 Node *use = n->fast_out(i); 1309 if (use->Opcode() == Op_SCMemProj) { 1310 n = use; 1311 break; 1312 } 1313 } 1314 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 1315 } 1316 } 1317 // push user on appropriate worklist 1318 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1319 Node *use = n->fast_out(i); 1320 if (use->is_Phi() || use->is_ClearArray()) { 1321 memnode_worklist.append_if_missing(use); 1322 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { 1323 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 1324 continue; 1325 memnode_worklist.append_if_missing(use); 1326 } else if (use->is_MemBar()) { 1327 memnode_worklist.append_if_missing(use); 1328 #ifdef ASSERT 1329 } else if(use->is_Mem()) { 1330 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 1331 } else if (use->is_MergeMem()) { 1332 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1333 } else { 1334 uint op = use->Opcode(); 1335 if (!(op == Op_StoreCM || 1336 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 1337 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 1338 op == Op_AryEq || op == Op_StrComp || 1339 op == Op_StrEquals || op == Op_StrIndexOf)) { 1340 n->dump(); 1341 use->dump(); 1342 assert(false, "EA: missing memory path"); 1343 } 1344 #endif 1345 } 1346 } 1347 } 1348 1349 // Phase 3: Process MergeMem nodes from mergemem_worklist. 1350 // Walk each memory slice moving the first node encountered of each 1351 // instance type to the the input corresponding to its alias index. 1352 uint length = _mergemem_worklist.length(); 1353 for( uint next = 0; next < length; ++next ) { 1354 MergeMemNode* nmm = _mergemem_worklist.at(next); 1355 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 1356 // Note: we don't want to use MergeMemStream here because we only want to 1357 // scan inputs which exist at the start, not ones we add during processing. 1358 // Note 2: MergeMem may already contains instance memory slices added 1359 // during find_inst_mem() call when memory nodes were processed above. 1360 igvn->hash_delete(nmm); 1361 uint nslices = nmm->req(); 1362 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 1363 Node* mem = nmm->in(i); 1364 Node* cur = NULL; 1365 if (mem == NULL || mem->is_top()) 1366 continue; 1367 // First, update mergemem by moving memory nodes to corresponding slices 1368 // if their type became more precise since this mergemem was created. 1369 while (mem->is_Mem()) { 1370 const Type *at = igvn->type(mem->in(MemNode::Address)); 1371 if (at != Type::TOP) { 1372 assert (at->isa_ptr() != NULL, "pointer type required."); 1373 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 1374 if (idx == i) { 1375 if (cur == NULL) 1376 cur = mem; 1377 } else { 1378 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 1379 nmm->set_memory_at(idx, mem); 1380 } 1381 } 1382 } 1383 mem = mem->in(MemNode::Memory); 1384 } 1385 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 1386 // Find any instance of the current type if we haven't encountered 1387 // already a memory slice of the instance along the memory chain. 1388 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1389 if((uint)_compile->get_general_index(ni) == i) { 1390 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 1391 if (nmm->is_empty_memory(m)) { 1392 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); 1393 if (_compile->failing()) { 1394 return; 1395 } 1396 nmm->set_memory_at(ni, result); 1397 } 1398 } 1399 } 1400 } 1401 // Find the rest of instances values 1402 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1403 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 1404 Node* result = step_through_mergemem(nmm, ni, tinst); 1405 if (result == nmm->base_memory()) { 1406 // Didn't find instance memory, search through general slice recursively. 1407 result = nmm->memory_at(_compile->get_general_index(ni)); 1408 result = find_inst_mem(result, ni, orig_phis, igvn); 1409 if (_compile->failing()) { 1410 return; 1411 } 1412 nmm->set_memory_at(ni, result); 1413 } 1414 } 1415 igvn->hash_insert(nmm); 1416 record_for_optimizer(nmm); 1417 } 1418 1419 // Phase 4: Update the inputs of non-instance memory Phis and 1420 // the Memory input of memnodes 1421 // First update the inputs of any non-instance Phi's from 1422 // which we split out an instance Phi. Note we don't have 1423 // to recursively process Phi's encounted on the input memory 1424 // chains as is done in split_memory_phi() since they will 1425 // also be processed here. 1426 for (int j = 0; j < orig_phis.length(); j++) { 1427 PhiNode *phi = orig_phis.at(j); 1428 int alias_idx = _compile->get_alias_index(phi->adr_type()); 1429 igvn->hash_delete(phi); 1430 for (uint i = 1; i < phi->req(); i++) { 1431 Node *mem = phi->in(i); 1432 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); 1433 if (_compile->failing()) { 1434 return; 1435 } 1436 if (mem != new_mem) { 1437 phi->set_req(i, new_mem); 1438 } 1439 } 1440 igvn->hash_insert(phi); 1441 record_for_optimizer(phi); 1442 } 1443 1444 // Update the memory inputs of MemNodes with the value we computed 1445 // in Phase 2 and move stores memory users to corresponding memory slices. 1446 1447 // Disable memory split verification code until the fix for 6984348. 1448 // Currently it produces false negative results since it does not cover all cases. 1449 #if 0 // ifdef ASSERT 1450 visited.Reset(); 1451 Node_Stack old_mems(arena, _compile->unique() >> 2); 1452 #endif 1453 for (uint i = 0; i < nodes_size(); i++) { 1454 Node *nmem = get_map(i); 1455 if (nmem != NULL) { 1456 Node *n = ptnode_adr(i)->_node; 1457 assert(n != NULL, "sanity"); 1458 if (n->is_Mem()) { 1459 #if 0 // ifdef ASSERT 1460 Node* old_mem = n->in(MemNode::Memory); 1461 if (!visited.test_set(old_mem->_idx)) { 1462 old_mems.push(old_mem, old_mem->outcnt()); 1463 } 1464 #endif 1465 assert(n->in(MemNode::Memory) != nmem, "sanity"); 1466 if (!n->is_Load()) { 1467 // Move memory users of a store first. 1468 move_inst_mem(n, orig_phis, igvn); 1469 } 1470 // Now update memory input 1471 igvn->hash_delete(n); 1472 n->set_req(MemNode::Memory, nmem); 1473 igvn->hash_insert(n); 1474 record_for_optimizer(n); 1475 } else { 1476 assert(n->is_Allocate() || n->is_CheckCastPP() || 1477 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 1478 } 1479 } 1480 } 1481 #if 0 // ifdef ASSERT 1482 // Verify that memory was split correctly 1483 while (old_mems.is_nonempty()) { 1484 Node* old_mem = old_mems.node(); 1485 uint old_cnt = old_mems.index(); 1486 old_mems.pop(); 1487 assert(old_cnt == old_mem->outcnt(), "old mem could be lost"); 1488 } 1489 #endif 1490 } 1491 1492 bool ConnectionGraph::has_candidates(Compile *C) { 1493 // EA brings benefits only when the code has allocations and/or locks which 1494 // are represented by ideal Macro nodes. 1495 int cnt = C->macro_count(); 1496 for( int i=0; i < cnt; i++ ) { 1497 Node *n = C->macro_node(i); 1498 if ( n->is_Allocate() ) 1499 return true; 1500 if( n->is_Lock() ) { 1501 Node* obj = n->as_Lock()->obj_node()->uncast(); 1502 if( !(obj->is_Parm() || obj->is_Con()) ) 1503 return true; 1504 } 1505 } 1506 return false; 1507 } 1508 1509 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) { 1510 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction 1511 // to create space for them in ConnectionGraph::_nodes[]. 1512 Node* oop_null = igvn->zerocon(T_OBJECT); 1513 Node* noop_null = igvn->zerocon(T_NARROWOOP); 1514 1515 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn); 1516 // Perform escape analysis 1517 if (congraph->compute_escape()) { 1518 // There are non escaping objects. 1519 C->set_congraph(congraph); 1520 } 1521 1522 // Cleanup. 1523 if (oop_null->outcnt() == 0) 1524 igvn->hash_delete(oop_null); 1525 if (noop_null->outcnt() == 0) 1526 igvn->hash_delete(noop_null); 1527 } 1528 1529 bool ConnectionGraph::compute_escape() { 1530 Compile* C = _compile; 1531 1532 // 1. Populate Connection Graph (CG) with Ideal nodes. 1533 1534 Unique_Node_List worklist_init; 1535 worklist_init.map(C->unique(), NULL); // preallocate space 1536 1537 // Initialize worklist 1538 if (C->root() != NULL) { 1539 worklist_init.push(C->root()); 1540 } 1541 1542 GrowableArray<Node*> alloc_worklist; 1543 GrowableArray<int> cg_worklist; 1544 PhaseGVN* igvn = _igvn; 1545 bool has_allocations = false; 1546 1547 // Push all useful nodes onto CG list and set their type. 1548 for( uint next = 0; next < worklist_init.size(); ++next ) { 1549 Node* n = worklist_init.at(next); 1550 record_for_escape_analysis(n, igvn); 1551 // Only allocations and java static calls results are checked 1552 // for an escape status. See process_call_result() below. 1553 if (n->is_Allocate() || n->is_CallStaticJava() && 1554 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { 1555 has_allocations = true; 1556 if (n->is_Allocate()) 1557 alloc_worklist.append(n); 1558 } 1559 if(n->is_AddP()) { 1560 // Collect address nodes. Use them during stage 3 below 1561 // to build initial connection graph field edges. 1562 cg_worklist.append(n->_idx); 1563 } else if (n->is_MergeMem()) { 1564 // Collect all MergeMem nodes to add memory slices for 1565 // scalar replaceable objects in split_unique_types(). 1566 _mergemem_worklist.append(n->as_MergeMem()); 1567 } 1568 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1569 Node* m = n->fast_out(i); // Get user 1570 worklist_init.push(m); 1571 } 1572 } 1573 1574 if (!has_allocations) { 1575 _collecting = false; 1576 return false; // Nothing to do. 1577 } 1578 1579 // 2. First pass to create simple CG edges (doesn't require to walk CG). 1580 uint delayed_size = _delayed_worklist.size(); 1581 for( uint next = 0; next < delayed_size; ++next ) { 1582 Node* n = _delayed_worklist.at(next); 1583 build_connection_graph(n, igvn); 1584 } 1585 1586 // 3. Pass to create initial fields edges (JavaObject -F-> AddP) 1587 // to reduce number of iterations during stage 4 below. 1588 uint cg_length = cg_worklist.length(); 1589 for( uint next = 0; next < cg_length; ++next ) { 1590 int ni = cg_worklist.at(next); 1591 Node* n = ptnode_adr(ni)->_node; 1592 Node* base = get_addp_base(n); 1593 if (base->is_Proj()) 1594 base = base->in(0); 1595 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type(); 1596 if (nt == PointsToNode::JavaObject) { 1597 build_connection_graph(n, igvn); 1598 } 1599 } 1600 1601 cg_worklist.clear(); 1602 cg_worklist.append(_phantom_object); 1603 GrowableArray<uint> worklist; 1604 1605 // 4. Build Connection Graph which need 1606 // to walk the connection graph. 1607 _progress = false; 1608 for (uint ni = 0; ni < nodes_size(); ni++) { 1609 PointsToNode* ptn = ptnode_adr(ni); 1610 Node *n = ptn->_node; 1611 if (n != NULL) { // Call, AddP, LoadP, StoreP 1612 build_connection_graph(n, igvn); 1613 if (ptn->node_type() != PointsToNode::UnknownType) 1614 cg_worklist.append(n->_idx); // Collect CG nodes 1615 if (!_processed.test(n->_idx)) 1616 worklist.append(n->_idx); // Collect C/A/L/S nodes 1617 } 1618 } 1619 1620 // After IGVN user nodes may have smaller _idx than 1621 // their inputs so they will be processed first in 1622 // previous loop. Because of that not all Graph 1623 // edges will be created. Walk over interesting 1624 // nodes again until no new edges are created. 1625 // 1626 // Normally only 1-3 passes needed to build 1627 // Connection Graph depending on graph complexity. 1628 // Observed 8 passes in jvm2008 compiler.compiler. 1629 // Set limit to 20 to catch situation when something 1630 // did go wrong and recompile the method without EA. 1631 1632 #define CG_BUILD_ITER_LIMIT 20 1633 1634 uint length = worklist.length(); 1635 int iterations = 0; 1636 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) { 1637 _progress = false; 1638 for( uint next = 0; next < length; ++next ) { 1639 int ni = worklist.at(next); 1640 PointsToNode* ptn = ptnode_adr(ni); 1641 Node* n = ptn->_node; 1642 assert(n != NULL, "should be known node"); 1643 build_connection_graph(n, igvn); 1644 } 1645 } 1646 if (iterations >= CG_BUILD_ITER_LIMIT) { 1647 assert(iterations < CG_BUILD_ITER_LIMIT, 1648 err_msg("infinite EA connection graph build with %d nodes and worklist size %d", 1649 nodes_size(), length)); 1650 // Possible infinite build_connection_graph loop, 1651 // retry compilation without escape analysis. 1652 C->record_failure(C2Compiler::retry_no_escape_analysis()); 1653 _collecting = false; 1654 return false; 1655 } 1656 #undef CG_BUILD_ITER_LIMIT 1657 1658 Arena* arena = Thread::current()->resource_area(); 1659 VectorSet visited(arena); 1660 1661 // 5. Find fields initializing values for not escaped allocations 1662 length = alloc_worklist.length(); 1663 for (uint next = 0; next < length; ++next) { 1664 Node* n = alloc_worklist.at(next); 1665 if (ptnode_adr(n->_idx)->escape_state() == PointsToNode::NoEscape) { 1666 find_init_values(n, &visited, igvn); 1667 } 1668 } 1669 1670 worklist.clear(); 1671 worklist_init.clear(); 1672 alloc_worklist.clear(); 1673 1674 // 6. Remove deferred edges from the graph. 1675 cg_length = cg_worklist.length(); 1676 for (uint next = 0; next < cg_length; ++next) { 1677 int ni = cg_worklist.at(next); 1678 PointsToNode* ptn = ptnode_adr(ni); 1679 PointsToNode::NodeType nt = ptn->node_type(); 1680 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { 1681 remove_deferred(ni, &worklist, &visited); 1682 Node *n = ptn->_node; 1683 if (n->is_AddP()) { 1684 alloc_worklist.append(n); 1685 } 1686 } 1687 } 1688 1689 // 7. Adjust escape state of nonescaping objects. 1690 length = alloc_worklist.length(); 1691 for (uint next = 0; next < length; ++next) { 1692 Node* n = alloc_worklist.at(next); 1693 adjust_escape_state(n); 1694 } 1695 1696 // 8. Propagate escape states. 1697 worklist.clear(); 1698 bool has_non_escaping_obj = false; 1699 1700 // mark all nodes reachable from GlobalEscape nodes 1701 (void)propagate_escape_state(&cg_worklist, &worklist, PointsToNode::GlobalEscape); 1702 1703 // mark all nodes reachable from ArgEscape nodes 1704 has_non_escaping_obj = propagate_escape_state(&cg_worklist, &worklist, PointsToNode::ArgEscape); 1705 1706 // mark all nodes reachable from ControlEscape nodes 1707 has_non_escaping_obj = propagate_escape_state(&cg_worklist, &worklist, PointsToNode::ControlEscape); 1708 1709 alloc_worklist.clear(); 1710 1711 // push all NoEscape nodes on the worklist 1712 for( uint next = 0; next < cg_length; ++next ) { 1713 int nk = cg_worklist.at(next); 1714 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) 1715 worklist.push(nk); 1716 } 1717 // mark all nodes reachable from NoEscape nodes 1718 while(worklist.length() > 0) { 1719 uint nk = worklist.pop(); 1720 PointsToNode* ptn = ptnode_adr(nk); 1721 if (ptn->node_type() == PointsToNode::JavaObject && 1722 !(nk == _noop_null || nk == _oop_null)) 1723 has_non_escaping_obj = true; // Non Escape 1724 Node* n = ptn->_node; 1725 if (n->is_Allocate()) { 1726 // Push scalar replaceable allocations on alloc_worklist 1727 // for processing in split_unique_types(). 1728 alloc_worklist.append(n); 1729 } 1730 uint e_cnt = ptn->edge_count(); 1731 for (uint ei = 0; ei < e_cnt; ei++) { 1732 uint npi = ptn->edge_target(ei); 1733 PointsToNode *np = ptnode_adr(npi); 1734 if (np->escape_state() < PointsToNode::NoEscape) { 1735 set_escape_state(npi, PointsToNode::NoEscape); 1736 worklist.push(npi); 1737 } 1738 } 1739 } 1740 1741 _collecting = false; 1742 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); 1743 1744 assert(ptnode_adr(_oop_null)->escape_state() == PointsToNode::NoEscape, "sanity"); 1745 if (UseCompressedOops) { 1746 assert(ptnode_adr(_noop_null)->escape_state() == PointsToNode::NoEscape, "sanity"); 1747 } 1748 1749 if (EliminateLocks && has_non_escaping_obj) { 1750 // Mark locks before changing ideal graph. 1751 int cnt = C->macro_count(); 1752 for( int i=0; i < cnt; i++ ) { 1753 Node *n = C->macro_node(i); 1754 if (n->is_AbstractLock()) { // Lock and Unlock nodes 1755 AbstractLockNode* alock = n->as_AbstractLock(); 1756 if (!alock->is_eliminated()) { 1757 PointsToNode::EscapeState es = escape_state(alock->obj_node()); 1758 assert(es != PointsToNode::UnknownEscape, "should know"); 1759 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) { 1760 // Mark it eliminated 1761 alock->set_eliminated(); 1762 } 1763 } 1764 } 1765 } 1766 } 1767 1768 #ifndef PRODUCT 1769 if (PrintEscapeAnalysis) { 1770 dump(); // Dump ConnectionGraph 1771 } 1772 #endif 1773 1774 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; 1775 if ( has_scalar_replaceable_candidates && 1776 C->AliasLevel() >= 3 && EliminateAllocations ) { 1777 1778 // Now use the escape information to create unique types for 1779 // scalar replaceable objects. 1780 split_unique_types(alloc_worklist); 1781 1782 if (C->failing()) return false; 1783 1784 C->print_method("After Escape Analysis", 2); 1785 1786 #ifdef ASSERT 1787 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 1788 tty->print("=== No allocations eliminated for "); 1789 C->method()->print_short_name(); 1790 if(!EliminateAllocations) { 1791 tty->print(" since EliminateAllocations is off ==="); 1792 } else if(!has_scalar_replaceable_candidates) { 1793 tty->print(" since there are no scalar replaceable candidates ==="); 1794 } else if(C->AliasLevel() < 3) { 1795 tty->print(" since AliasLevel < 3 ==="); 1796 } 1797 tty->cr(); 1798 #endif 1799 } 1800 return has_non_escaping_obj; 1801 } 1802 1803 // Find fields initializing values for allocations. 1804 void ConnectionGraph::find_init_values(Node* alloc, VectorSet* visited, PhaseTransform* phase) { 1805 assert(alloc->is_Allocate(), "Should be called for Allocate nodes only"); 1806 PointsToNode* pta = ptnode_adr(alloc->_idx); 1807 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only"); 1808 InitializeNode* ini = alloc->as_Allocate()->initialization(); 1809 1810 Compile* C = _compile; 1811 visited->Reset(); 1812 // Check if a oop field's initializing value is recorded and add 1813 // a corresponding NULL field's value if it is not recorded. 1814 // Connection Graph does not record a default initialization by NULL 1815 // captured by Initialize node. 1816 // 1817 uint ae_cnt = pta->edge_count(); 1818 for (uint ei = 0; ei < ae_cnt; ei++) { 1819 uint nidx = pta->edge_target(ei); // Field (AddP) 1820 PointsToNode* ptn = ptnode_adr(nidx); 1821 assert(ptn->_node->is_AddP(), "Should be AddP nodes only"); 1822 int offset = ptn->offset(); 1823 if (offset != Type::OffsetBot && 1824 offset != oopDesc::klass_offset_in_bytes() && 1825 !visited->test_set(offset)) { 1826 1827 // Check only oop fields. 1828 const Type* adr_type = ptn->_node->as_AddP()->bottom_type(); 1829 BasicType basic_field_type = T_INT; 1830 if (adr_type->isa_instptr()) { 1831 ciField* field = C->alias_type(adr_type->isa_instptr())->field(); 1832 if (field != NULL) { 1833 basic_field_type = field->layout_type(); 1834 } else { 1835 // Ignore non field load (for example, klass load) 1836 } 1837 } else if (adr_type->isa_aryptr()) { 1838 if (offset != arrayOopDesc::length_offset_in_bytes()) { 1839 const Type* elemtype = adr_type->isa_aryptr()->elem(); 1840 basic_field_type = elemtype->array_element_basic_type(); 1841 } else { 1842 // Ignore array length load 1843 } 1844 #ifdef ASSERT 1845 } else { 1846 // Raw pointers are used for initializing stores so skip it 1847 // since it should be recorded already 1848 Node* base = get_addp_base(ptn->_node); 1849 assert(adr_type->isa_rawptr() && base->is_Proj() && 1850 (base->in(0) == alloc),"unexpected pointer type"); 1851 #endif 1852 } 1853 if (basic_field_type == T_OBJECT || 1854 basic_field_type == T_NARROWOOP || 1855 basic_field_type == T_ARRAY) { 1856 Node* value = NULL; 1857 if (ini != NULL) { 1858 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT; 1859 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase); 1860 if (store != NULL && store->is_Store()) { 1861 value = store->in(MemNode::ValueIn); 1862 } else if (ptn->edge_count() > 0) { // Are there oop stores? 1863 // Check for a store which follows allocation without branches. 1864 // For example, a volatile field store is not collected 1865 // by Initialize node. TODO: it would be nice to use idom() here. 1866 // 1867 // Search all references to the same field which use different 1868 // AddP nodes, for example, in the next case: 1869 // 1870 // Point p[] = new Point[1]; 1871 // if ( x ) { p[0] = new Point(); p[0].x = x; } 1872 // if ( p[0] != null ) { y = p[0].x; } // has CastPP 1873 // 1874 for (uint next = ei; (next < ae_cnt) && (value == NULL); next++) { 1875 uint fpi = pta->edge_target(next); // Field (AddP) 1876 PointsToNode *ptf = ptnode_adr(fpi); 1877 if (ptf->offset() == offset) { 1878 Node* nf = ptf->_node; 1879 for (DUIterator_Fast imax, i = nf->fast_outs(imax); i < imax; i++) { 1880 store = nf->fast_out(i); 1881 if (store->is_Store() && store->in(0) != NULL) { 1882 Node* ctrl = store->in(0); 1883 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL || 1884 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() || 1885 ctrl->is_IfTrue() || ctrl->is_IfFalse())) { 1886 ctrl = ctrl->in(0); 1887 } 1888 if (ctrl == ini || ctrl == alloc) { 1889 value = store->in(MemNode::ValueIn); 1890 break; 1891 } 1892 } 1893 } 1894 } 1895 } 1896 } 1897 } 1898 if (value == NULL || value != ptnode_adr(value->_idx)->_node) { 1899 // A field's initializing value was not recorded. Add NULL. 1900 uint null_idx = UseCompressedOops ? _noop_null : _oop_null; 1901 add_edge_from_fields(alloc->_idx, null_idx, offset); 1902 } 1903 } 1904 } 1905 } 1906 } 1907 1908 // Adjust escape state after Connection Graph is built. 1909 void ConnectionGraph::adjust_escape_state(Node* n) { 1910 PointsToNode* ptn = ptnode_adr(n->_idx); 1911 assert(n->is_AddP(), "Should be called for AddP nodes only"); 1912 // Search for objects which are not scalar replaceable. 1913 // Mark their escape state as ControlEscape to propagate the state 1914 // to referenced objects. 1915 1916 int offset = ptn->offset(); 1917 Node* base = get_addp_base(n); 1918 VectorSet* ptset = PointsTo(base); 1919 int ptset_size = ptset->Size(); 1920 1921 // An object is not scalar replaceable if the field which may point 1922 // to it has unknown offset (unknown element of an array of objects). 1923 // 1924 1925 if (offset == Type::OffsetBot) { 1926 uint e_cnt = ptn->edge_count(); 1927 for (uint ei = 0; ei < e_cnt; ei++) { 1928 uint npi = ptn->edge_target(ei); 1929 set_escape_state(npi, PointsToNode::ControlEscape); 1930 } 1931 } 1932 1933 // Currently an object is not scalar replaceable if a LoadStore node 1934 // access its field since the field value is unknown after it. 1935 // 1936 bool has_LoadStore = false; 1937 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1938 Node *use = n->fast_out(i); 1939 if (use->is_LoadStore()) { 1940 has_LoadStore = true; 1941 break; 1942 } 1943 } 1944 // An object is not scalar replaceable if the address points 1945 // to unknown field (unknown element for arrays, offset is OffsetBot). 1946 // 1947 // Or the address may point to more then one object. This may produce 1948 // the false positive result (set escape to ControlEscape) 1949 // since the flow-insensitive escape analysis can't separate 1950 // the case when stores overwrite the field's value from the case 1951 // when stores happened on different control branches. 1952 // 1953 // Note: it will disable scalar replacement in some cases: 1954 // 1955 // Point p[] = new Point[1]; 1956 // p[0] = new Point(); // Will be not scalar replaced 1957 // 1958 // but it will save us from incorrect optimizations in next cases: 1959 // 1960 // Point p[] = new Point[1]; 1961 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1962 // 1963 if (ptset_size > 1 || ptset_size != 0 && 1964 (has_LoadStore || offset == Type::OffsetBot)) { 1965 for( VectorSetI j(ptset); j.test(); ++j ) { 1966 set_escape_state(j.elem, PointsToNode::ControlEscape); 1967 } 1968 } 1969 } 1970 1971 // Propagate escape states to referenced nodes. 1972 bool ConnectionGraph::propagate_escape_state(GrowableArray<int>* cg_worklist, 1973 GrowableArray<uint>* worklist, 1974 PointsToNode::EscapeState esc_state) { 1975 bool has_java_obj = false; 1976 1977 // push all nodes with the same escape state on the worklist 1978 uint cg_length = cg_worklist->length(); 1979 for (uint next = 0; next < cg_length; ++next) { 1980 int nk = cg_worklist->at(next); 1981 if (ptnode_adr(nk)->escape_state() == esc_state) 1982 worklist->push(nk); 1983 } 1984 // mark all reachable nodes 1985 while (worklist->length() > 0) { 1986 PointsToNode* ptn = ptnode_adr(worklist->pop()); 1987 if (ptn->node_type() == PointsToNode::JavaObject) { 1988 has_java_obj = true; 1989 } 1990 uint e_cnt = ptn->edge_count(); 1991 for (uint ei = 0; ei < e_cnt; ei++) { 1992 uint npi = ptn->edge_target(ei); 1993 PointsToNode *np = ptnode_adr(npi); 1994 if (np->escape_state() < esc_state) { 1995 set_escape_state(npi, esc_state); 1996 worklist->push(npi); 1997 } 1998 } 1999 } 2000 // Has not escaping java objects 2001 return has_java_obj && (esc_state < PointsToNode::GlobalEscape); 2002 } 2003 2004 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { 2005 2006 switch (call->Opcode()) { 2007 #ifdef ASSERT 2008 case Op_Allocate: 2009 case Op_AllocateArray: 2010 case Op_Lock: 2011 case Op_Unlock: 2012 assert(false, "should be done already"); 2013 break; 2014 #endif 2015 case Op_CallLeaf: 2016 case Op_CallLeafNoFP: 2017 { 2018 // Stub calls, objects do not escape but they are not scale replaceable. 2019 // Adjust escape state for outgoing arguments. 2020 const TypeTuple * d = call->tf()->domain(); 2021 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2022 const Type* at = d->field_at(i); 2023 Node *arg = call->in(i)->uncast(); 2024 const Type *aat = phase->type(arg); 2025 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() && 2026 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) { 2027 2028 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 2029 aat->isa_ptr() != NULL, "expecting an Ptr"); 2030 #ifdef ASSERT 2031 if (!(call->Opcode() == Op_CallLeafNoFP && 2032 call->as_CallLeaf()->_name != NULL && 2033 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) || 2034 call->as_CallLeaf()->_name != NULL && 2035 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 2036 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 )) 2037 ) { 2038 call->dump(); 2039 assert(false, "EA: unexpected CallLeaf"); 2040 } 2041 #endif 2042 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 2043 if (arg->is_AddP()) { 2044 // 2045 // The inline_native_clone() case when the arraycopy stub is called 2046 // after the allocation before Initialize and CheckCastPP nodes. 2047 // 2048 // Set AddP's base (Allocate) as not scalar replaceable since 2049 // pointer to the base (with offset) is passed as argument. 2050 // 2051 arg = get_addp_base(arg); 2052 } 2053 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) { 2054 uint pt = j.elem; 2055 set_escape_state(pt, PointsToNode::ArgEscape); 2056 } 2057 } 2058 } 2059 break; 2060 } 2061 2062 case Op_CallStaticJava: 2063 // For a static call, we know exactly what method is being called. 2064 // Use bytecode estimator to record the call's escape affects 2065 { 2066 ciMethod *meth = call->as_CallJava()->method(); 2067 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 2068 // fall-through if not a Java method or no analyzer information 2069 if (call_analyzer != NULL) { 2070 const TypeTuple * d = call->tf()->domain(); 2071 bool copy_dependencies = false; 2072 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2073 const Type* at = d->field_at(i); 2074 int k = i - TypeFunc::Parms; 2075 Node *arg = call->in(i)->uncast(); 2076 2077 if (at->isa_oopptr() != NULL && 2078 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) { 2079 2080 bool global_escapes = false; 2081 bool fields_escapes = false; 2082 if (!call_analyzer->is_arg_stack(k)) { 2083 // The argument global escapes, mark everything it could point to 2084 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 2085 global_escapes = true; 2086 } else { 2087 if (!call_analyzer->is_arg_local(k)) { 2088 // The argument itself doesn't escape, but any fields might 2089 fields_escapes = true; 2090 } 2091 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 2092 copy_dependencies = true; 2093 } 2094 2095 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) { 2096 uint pt = j.elem; 2097 if (global_escapes) { 2098 //The argument global escapes, mark everything it could point to 2099 set_escape_state(pt, PointsToNode::GlobalEscape); 2100 } else { 2101 if (fields_escapes) { 2102 // The argument itself doesn't escape, but any fields might 2103 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); 2104 } 2105 set_escape_state(pt, PointsToNode::ArgEscape); 2106 } 2107 } 2108 } 2109 } 2110 if (copy_dependencies) 2111 call_analyzer->copy_dependencies(_compile->dependencies()); 2112 break; 2113 } 2114 } 2115 2116 default: 2117 // Fall-through here if not a Java method or no analyzer information 2118 // or some other type of call, assume the worst case: all arguments 2119 // globally escape. 2120 { 2121 // adjust escape state for outgoing arguments 2122 const TypeTuple * d = call->tf()->domain(); 2123 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2124 const Type* at = d->field_at(i); 2125 if (at->isa_oopptr() != NULL) { 2126 Node *arg = call->in(i)->uncast(); 2127 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 2128 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) { 2129 uint pt = j.elem; 2130 set_escape_state(pt, PointsToNode::GlobalEscape); 2131 } 2132 } 2133 } 2134 } 2135 } 2136 } 2137 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { 2138 CallNode *call = resproj->in(0)->as_Call(); 2139 uint call_idx = call->_idx; 2140 uint resproj_idx = resproj->_idx; 2141 2142 switch (call->Opcode()) { 2143 case Op_Allocate: 2144 { 2145 Node *k = call->in(AllocateNode::KlassNode); 2146 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2147 assert(kt != NULL, "TypeKlassPtr required."); 2148 ciKlass* cik = kt->klass(); 2149 2150 PointsToNode::EscapeState es; 2151 uint edge_to; 2152 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 2153 !cik->is_instance_klass() || // StressReflectiveCode 2154 cik->as_instance_klass()->has_finalizer()) { 2155 es = PointsToNode::GlobalEscape; 2156 edge_to = _phantom_object; // Could not be worse 2157 } else { 2158 es = PointsToNode::NoEscape; 2159 edge_to = call_idx; 2160 } 2161 set_escape_state(call_idx, es); 2162 add_pointsto_edge(resproj_idx, edge_to); 2163 _processed.set(resproj_idx); 2164 break; 2165 } 2166 2167 case Op_AllocateArray: 2168 { 2169 2170 Node *k = call->in(AllocateNode::KlassNode); 2171 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2172 assert(kt != NULL, "TypeKlassPtr required."); 2173 ciKlass* cik = kt->klass(); 2174 2175 PointsToNode::EscapeState es; 2176 uint edge_to; 2177 if (!cik->is_array_klass()) { // StressReflectiveCode 2178 es = PointsToNode::GlobalEscape; 2179 edge_to = _phantom_object; 2180 } else { 2181 es = PointsToNode::NoEscape; 2182 edge_to = call_idx; 2183 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 2184 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 2185 // Not scalar replaceable if the length is not constant or too big. 2186 es = PointsToNode::ControlEscape; 2187 } 2188 } 2189 set_escape_state(call_idx, es); 2190 add_pointsto_edge(resproj_idx, edge_to); 2191 _processed.set(resproj_idx); 2192 break; 2193 } 2194 2195 case Op_CallStaticJava: 2196 // For a static call, we know exactly what method is being called. 2197 // Use bytecode estimator to record whether the call's return value escapes 2198 { 2199 bool done = true; 2200 const TypeTuple *r = call->tf()->range(); 2201 const Type* ret_type = NULL; 2202 2203 if (r->cnt() > TypeFunc::Parms) 2204 ret_type = r->field_at(TypeFunc::Parms); 2205 2206 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2207 // _multianewarray functions return a TypeRawPtr. 2208 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { 2209 _processed.set(resproj_idx); 2210 break; // doesn't return a pointer type 2211 } 2212 ciMethod *meth = call->as_CallJava()->method(); 2213 const TypeTuple * d = call->tf()->domain(); 2214 if (meth == NULL) { 2215 // not a Java method, assume global escape 2216 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2217 add_pointsto_edge(resproj_idx, _phantom_object); 2218 } else { 2219 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); 2220 bool copy_dependencies = false; 2221 2222 if (call_analyzer->is_return_allocated()) { 2223 // Returns a newly allocated unescaped object, simply 2224 // update dependency information. 2225 // Mark it as ControlEscape so that objects referenced by 2226 // it's fields will be marked as ControlEscape at least. 2227 set_escape_state(call_idx, PointsToNode::ControlEscape); 2228 add_pointsto_edge(resproj_idx, call_idx); 2229 copy_dependencies = true; 2230 } else if (call_analyzer->is_return_local()) { 2231 // determine whether any arguments are returned 2232 set_escape_state(call_idx, PointsToNode::ArgEscape); 2233 bool ret_arg = false; 2234 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2235 const Type* at = d->field_at(i); 2236 2237 if (at->isa_oopptr() != NULL) { 2238 Node *arg = call->in(i)->uncast(); 2239 2240 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 2241 ret_arg = true; 2242 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2243 if (arg_esp->node_type() == PointsToNode::UnknownType) 2244 done = false; 2245 else if (arg_esp->node_type() == PointsToNode::JavaObject) 2246 add_pointsto_edge(resproj_idx, arg->_idx); 2247 else 2248 add_deferred_edge(resproj_idx, arg->_idx); 2249 } 2250 } 2251 } 2252 if (done && !ret_arg) { 2253 // Returns unknown object. 2254 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2255 add_pointsto_edge(resproj_idx, _phantom_object); 2256 } 2257 if (done) { 2258 copy_dependencies = true; 2259 } 2260 } else { 2261 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2262 add_pointsto_edge(resproj_idx, _phantom_object); 2263 } 2264 if (copy_dependencies) 2265 call_analyzer->copy_dependencies(_compile->dependencies()); 2266 } 2267 if (done) 2268 _processed.set(resproj_idx); 2269 break; 2270 } 2271 2272 default: 2273 // Some other type of call, assume the worst case that the 2274 // returned value, if any, globally escapes. 2275 { 2276 const TypeTuple *r = call->tf()->range(); 2277 if (r->cnt() > TypeFunc::Parms) { 2278 const Type* ret_type = r->field_at(TypeFunc::Parms); 2279 2280 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2281 // _multianewarray functions return a TypeRawPtr. 2282 if (ret_type->isa_ptr() != NULL) { 2283 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2284 add_pointsto_edge(resproj_idx, _phantom_object); 2285 } 2286 } 2287 _processed.set(resproj_idx); 2288 } 2289 } 2290 } 2291 2292 // Populate Connection Graph with Ideal nodes and create simple 2293 // connection graph edges (do not need to check the node_type of inputs 2294 // or to call PointsTo() to walk the connection graph). 2295 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { 2296 if (_processed.test(n->_idx)) 2297 return; // No need to redefine node's state. 2298 2299 if (n->is_Call()) { 2300 // Arguments to allocation and locking don't escape. 2301 if (n->is_Allocate()) { 2302 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); 2303 record_for_optimizer(n); 2304 } else if (n->is_Lock() || n->is_Unlock()) { 2305 // Put Lock and Unlock nodes on IGVN worklist to process them during 2306 // the first IGVN optimization when escape information is still available. 2307 record_for_optimizer(n); 2308 _processed.set(n->_idx); 2309 } else { 2310 // Don't mark as processed since call's arguments have to be processed. 2311 PointsToNode::NodeType nt = PointsToNode::UnknownType; 2312 PointsToNode::EscapeState es = PointsToNode::UnknownEscape; 2313 2314 // Check if a call returns an object. 2315 const TypeTuple *r = n->as_Call()->tf()->range(); 2316 if (r->cnt() > TypeFunc::Parms && 2317 r->field_at(TypeFunc::Parms)->isa_ptr() && 2318 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { 2319 nt = PointsToNode::JavaObject; 2320 if (!n->is_CallStaticJava()) { 2321 // Since the called mathod is statically unknown assume 2322 // the worst case that the returned value globally escapes. 2323 es = PointsToNode::GlobalEscape; 2324 } 2325 } 2326 add_node(n, nt, es, false); 2327 } 2328 return; 2329 } 2330 2331 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 2332 // ThreadLocal has RawPrt type. 2333 switch (n->Opcode()) { 2334 case Op_AddP: 2335 { 2336 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); 2337 break; 2338 } 2339 case Op_CastX2P: 2340 { // "Unsafe" memory access. 2341 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2342 break; 2343 } 2344 case Op_CastPP: 2345 case Op_CheckCastPP: 2346 case Op_EncodeP: 2347 case Op_DecodeN: 2348 { 2349 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2350 int ti = n->in(1)->_idx; 2351 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2352 if (nt == PointsToNode::UnknownType) { 2353 _delayed_worklist.push(n); // Process it later. 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 _processed.set(n->_idx); 2361 break; 2362 } 2363 case Op_ConP: 2364 { 2365 // assume all pointer constants globally escape except for null 2366 PointsToNode::EscapeState es; 2367 if (phase->type(n) == TypePtr::NULL_PTR) 2368 es = PointsToNode::NoEscape; 2369 else 2370 es = PointsToNode::GlobalEscape; 2371 2372 add_node(n, PointsToNode::JavaObject, es, true); 2373 break; 2374 } 2375 case Op_ConN: 2376 { 2377 // assume all narrow oop constants globally escape except for null 2378 PointsToNode::EscapeState es; 2379 if (phase->type(n) == TypeNarrowOop::NULL_PTR) 2380 es = PointsToNode::NoEscape; 2381 else 2382 es = PointsToNode::GlobalEscape; 2383 2384 add_node(n, PointsToNode::JavaObject, es, true); 2385 break; 2386 } 2387 case Op_CreateEx: 2388 { 2389 // assume that all exception objects globally escape 2390 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2391 break; 2392 } 2393 case Op_LoadKlass: 2394 case Op_LoadNKlass: 2395 { 2396 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2397 break; 2398 } 2399 case Op_LoadP: 2400 case Op_LoadN: 2401 { 2402 const Type *t = phase->type(n); 2403 if (t->make_ptr() == NULL) { 2404 _processed.set(n->_idx); 2405 return; 2406 } 2407 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2408 break; 2409 } 2410 case Op_Parm: 2411 { 2412 _processed.set(n->_idx); // No need to redefine it state. 2413 uint con = n->as_Proj()->_con; 2414 if (con < TypeFunc::Parms) 2415 return; 2416 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); 2417 if (t->isa_ptr() == NULL) 2418 return; 2419 // We have to assume all input parameters globally escape 2420 // (Note: passing 'false' since _processed is already set). 2421 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); 2422 break; 2423 } 2424 case Op_Phi: 2425 { 2426 const Type *t = n->as_Phi()->type(); 2427 if (t->make_ptr() == NULL) { 2428 // nothing to do if not an oop or narrow oop 2429 _processed.set(n->_idx); 2430 return; 2431 } 2432 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2433 uint i; 2434 for (i = 1; i < n->req() ; i++) { 2435 Node* in = n->in(i); 2436 if (in == NULL) 2437 continue; // ignore NULL 2438 in = in->uncast(); 2439 if (in->is_top() || in == n) 2440 continue; // ignore top or inputs which go back this node 2441 int ti = in->_idx; 2442 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2443 if (nt == PointsToNode::UnknownType) { 2444 break; 2445 } else if (nt == PointsToNode::JavaObject) { 2446 add_pointsto_edge(n->_idx, ti); 2447 } else { 2448 add_deferred_edge(n->_idx, ti); 2449 } 2450 } 2451 if (i >= n->req()) 2452 _processed.set(n->_idx); 2453 else 2454 _delayed_worklist.push(n); 2455 break; 2456 } 2457 case Op_Proj: 2458 { 2459 // we are only interested in the oop result projection from a call 2460 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2461 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2462 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2463 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2464 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2465 int ti = n->in(0)->_idx; 2466 // The call may not be registered yet (since not all its inputs are registered) 2467 // if this is the projection from backbranch edge of Phi. 2468 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) { 2469 process_call_result(n->as_Proj(), phase); 2470 } 2471 if (!_processed.test(n->_idx)) { 2472 // The call's result may need to be processed later if the call 2473 // returns it's argument and the argument is not processed yet. 2474 _delayed_worklist.push(n); 2475 } 2476 break; 2477 } 2478 } 2479 _processed.set(n->_idx); 2480 break; 2481 } 2482 case Op_Return: 2483 { 2484 if( n->req() > TypeFunc::Parms && 2485 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2486 // Treat Return value as LocalVar with GlobalEscape escape state. 2487 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); 2488 int ti = n->in(TypeFunc::Parms)->_idx; 2489 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2490 if (nt == PointsToNode::UnknownType) { 2491 _delayed_worklist.push(n); // Process it later. 2492 break; 2493 } else if (nt == PointsToNode::JavaObject) { 2494 add_pointsto_edge(n->_idx, ti); 2495 } else { 2496 add_deferred_edge(n->_idx, ti); 2497 } 2498 } 2499 _processed.set(n->_idx); 2500 break; 2501 } 2502 case Op_StoreP: 2503 case Op_StoreN: 2504 { 2505 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2506 adr_type = adr_type->make_ptr(); 2507 if (adr_type->isa_oopptr()) { 2508 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2509 } else { 2510 Node* adr = n->in(MemNode::Address); 2511 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && 2512 adr->in(AddPNode::Address)->is_Proj() && 2513 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2514 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2515 // We are computing a raw address for a store captured 2516 // by an Initialize compute an appropriate address type. 2517 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2518 assert(offs != Type::OffsetBot, "offset must be a constant"); 2519 } else { 2520 _processed.set(n->_idx); 2521 return; 2522 } 2523 } 2524 break; 2525 } 2526 case Op_StorePConditional: 2527 case Op_CompareAndSwapP: 2528 case Op_CompareAndSwapN: 2529 { 2530 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2531 adr_type = adr_type->make_ptr(); 2532 if (adr_type->isa_oopptr()) { 2533 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2534 } else { 2535 _processed.set(n->_idx); 2536 return; 2537 } 2538 break; 2539 } 2540 case Op_AryEq: 2541 case Op_StrComp: 2542 case Op_StrEquals: 2543 case Op_StrIndexOf: 2544 { 2545 // char[] arrays passed to string intrinsics are not scalar replaceable. 2546 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2547 break; 2548 } 2549 case Op_ThreadLocal: 2550 { 2551 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); 2552 break; 2553 } 2554 default: 2555 ; 2556 // nothing to do 2557 } 2558 return; 2559 } 2560 2561 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { 2562 uint n_idx = n->_idx; 2563 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered"); 2564 2565 // Don't set processed bit for AddP, LoadP, StoreP since 2566 // they may need more then one pass to process. 2567 // Also don't mark as processed Call nodes since their 2568 // arguments may need more then one pass to process. 2569 if (_processed.test(n_idx)) 2570 return; // No need to redefine node's state. 2571 2572 if (n->is_Call()) { 2573 CallNode *call = n->as_Call(); 2574 process_call_arguments(call, phase); 2575 return; 2576 } 2577 2578 switch (n->Opcode()) { 2579 case Op_AddP: 2580 { 2581 Node *base = get_addp_base(n); 2582 // Create a field edge to this node from everything base could point to. 2583 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) { 2584 uint pt = i.elem; 2585 add_field_edge(pt, n_idx, address_offset(n, phase)); 2586 } 2587 break; 2588 } 2589 case Op_CastX2P: 2590 { 2591 assert(false, "Op_CastX2P"); 2592 break; 2593 } 2594 case Op_CastPP: 2595 case Op_CheckCastPP: 2596 case Op_EncodeP: 2597 case Op_DecodeN: 2598 { 2599 int ti = n->in(1)->_idx; 2600 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered"); 2601 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2602 add_pointsto_edge(n_idx, ti); 2603 } else { 2604 add_deferred_edge(n_idx, ti); 2605 } 2606 _processed.set(n_idx); 2607 break; 2608 } 2609 case Op_ConP: 2610 { 2611 assert(false, "Op_ConP"); 2612 break; 2613 } 2614 case Op_ConN: 2615 { 2616 assert(false, "Op_ConN"); 2617 break; 2618 } 2619 case Op_CreateEx: 2620 { 2621 assert(false, "Op_CreateEx"); 2622 break; 2623 } 2624 case Op_LoadKlass: 2625 case Op_LoadNKlass: 2626 { 2627 assert(false, "Op_LoadKlass"); 2628 break; 2629 } 2630 case Op_LoadP: 2631 case Op_LoadN: 2632 { 2633 const Type *t = phase->type(n); 2634 #ifdef ASSERT 2635 if (t->make_ptr() == NULL) 2636 assert(false, "Op_LoadP"); 2637 #endif 2638 2639 Node* adr = n->in(MemNode::Address)->uncast(); 2640 Node* adr_base; 2641 if (adr->is_AddP()) { 2642 adr_base = get_addp_base(adr); 2643 } else { 2644 adr_base = adr; 2645 } 2646 2647 // For everything "adr_base" could point to, create a deferred edge from 2648 // this node to each field with the same offset. 2649 int offset = address_offset(adr, phase); 2650 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) { 2651 uint pt = i.elem; 2652 add_deferred_edge_to_fields(n_idx, pt, offset); 2653 } 2654 break; 2655 } 2656 case Op_Parm: 2657 { 2658 assert(false, "Op_Parm"); 2659 break; 2660 } 2661 case Op_Phi: 2662 { 2663 #ifdef ASSERT 2664 const Type *t = n->as_Phi()->type(); 2665 if (t->make_ptr() == NULL) 2666 assert(false, "Op_Phi"); 2667 #endif 2668 for (uint i = 1; i < n->req() ; i++) { 2669 Node* in = n->in(i); 2670 if (in == NULL) 2671 continue; // ignore NULL 2672 in = in->uncast(); 2673 if (in->is_top() || in == n) 2674 continue; // ignore top or inputs which go back this node 2675 int ti = in->_idx; 2676 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2677 assert(nt != PointsToNode::UnknownType, "all nodes should be known"); 2678 if (nt == PointsToNode::JavaObject) { 2679 add_pointsto_edge(n_idx, ti); 2680 } else { 2681 add_deferred_edge(n_idx, ti); 2682 } 2683 } 2684 _processed.set(n_idx); 2685 break; 2686 } 2687 case Op_Proj: 2688 { 2689 // we are only interested in the oop result projection from a call 2690 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2691 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType, 2692 "all nodes should be registered"); 2693 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2694 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2695 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2696 process_call_result(n->as_Proj(), phase); 2697 assert(_processed.test(n_idx), "all call results should be processed"); 2698 break; 2699 } 2700 } 2701 assert(false, "Op_Proj"); 2702 break; 2703 } 2704 case Op_Return: 2705 { 2706 #ifdef ASSERT 2707 if( n->req() <= TypeFunc::Parms || 2708 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2709 assert(false, "Op_Return"); 2710 } 2711 #endif 2712 int ti = n->in(TypeFunc::Parms)->_idx; 2713 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered"); 2714 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2715 add_pointsto_edge(n_idx, ti); 2716 } else { 2717 add_deferred_edge(n_idx, ti); 2718 } 2719 _processed.set(n_idx); 2720 break; 2721 } 2722 case Op_StoreP: 2723 case Op_StoreN: 2724 case Op_StorePConditional: 2725 case Op_CompareAndSwapP: 2726 case Op_CompareAndSwapN: 2727 { 2728 Node *adr = n->in(MemNode::Address); 2729 const Type *adr_type = phase->type(adr)->make_ptr(); 2730 #ifdef ASSERT 2731 if (!adr_type->isa_oopptr()) 2732 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); 2733 #endif 2734 2735 assert(adr->is_AddP(), "expecting an AddP"); 2736 Node *adr_base = get_addp_base(adr); 2737 Node *val = n->in(MemNode::ValueIn)->uncast(); 2738 // For everything "adr_base" could point to, create a deferred edge 2739 // to "val" from each field with the same offset. 2740 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) { 2741 uint pt = i.elem; 2742 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); 2743 } 2744 break; 2745 } 2746 case Op_AryEq: 2747 case Op_StrComp: 2748 case Op_StrEquals: 2749 case Op_StrIndexOf: 2750 { 2751 // char[] arrays passed to string intrinsic do not escape but 2752 // they are not scalar replaceable. Adjust escape state for them. 2753 // Start from in(2) edge since in(1) is memory edge. 2754 for (uint i = 2; i < n->req(); i++) { 2755 Node* adr = n->in(i)->uncast(); 2756 const Type *at = phase->type(adr); 2757 if (!adr->is_top() && at->isa_ptr()) { 2758 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 2759 at->isa_ptr() != NULL, "expecting an Ptr"); 2760 if (adr->is_AddP()) { 2761 adr = get_addp_base(adr); 2762 } 2763 // Mark as ArgEscape everything "adr" could point to. 2764 set_escape_state(adr->_idx, PointsToNode::ArgEscape); 2765 } 2766 } 2767 _processed.set(n_idx); 2768 break; 2769 } 2770 case Op_ThreadLocal: 2771 { 2772 assert(false, "Op_ThreadLocal"); 2773 break; 2774 } 2775 default: 2776 // This method should be called only for EA specific nodes. 2777 ShouldNotReachHere(); 2778 } 2779 } 2780 2781 #ifndef PRODUCT 2782 void ConnectionGraph::dump() { 2783 bool first = true; 2784 2785 uint size = nodes_size(); 2786 for (uint ni = 0; ni < size; ni++) { 2787 PointsToNode *ptn = ptnode_adr(ni); 2788 PointsToNode::NodeType ptn_type = ptn->node_type(); 2789 2790 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) 2791 continue; 2792 PointsToNode::EscapeState es = escape_state(ptn->_node); 2793 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { 2794 if (first) { 2795 tty->cr(); 2796 tty->print("======== Connection graph for "); 2797 _compile->method()->print_short_name(); 2798 tty->cr(); 2799 first = false; 2800 } 2801 tty->print("%6d ", ni); 2802 ptn->dump(); 2803 // Print all locals which reference this allocation 2804 for (uint li = ni; li < size; li++) { 2805 PointsToNode *ptn_loc = ptnode_adr(li); 2806 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); 2807 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && 2808 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { 2809 ptnode_adr(li)->dump(false); 2810 } 2811 } 2812 if (Verbose) { 2813 // Print all fields which reference this allocation 2814 for (uint i = 0; i < ptn->edge_count(); i++) { 2815 uint ei = ptn->edge_target(i); 2816 ptnode_adr(ei)->dump(false); 2817 } 2818 } 2819 tty->cr(); 2820 } 2821 } 2822 } 2823 #endif