escape.cpp revision 1472:c18cbe5936b8
1/* 2 * Copyright (c) 2005, 2009, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "incls/_precompiled.incl" 26#include "incls/_escape.cpp.incl" 27 28void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) { 29 uint v = (targIdx << EdgeShift) + ((uint) et); 30 if (_edges == NULL) { 31 Arena *a = Compile::current()->comp_arena(); 32 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0); 33 } 34 _edges->append_if_missing(v); 35} 36 37void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) { 38 uint v = (targIdx << EdgeShift) + ((uint) et); 39 40 _edges->remove(v); 41} 42 43#ifndef PRODUCT 44static const char *node_type_names[] = { 45 "UnknownType", 46 "JavaObject", 47 "LocalVar", 48 "Field" 49}; 50 51static const char *esc_names[] = { 52 "UnknownEscape", 53 "NoEscape", 54 "ArgEscape", 55 "GlobalEscape" 56}; 57 58static const char *edge_type_suffix[] = { 59 "?", // UnknownEdge 60 "P", // PointsToEdge 61 "D", // DeferredEdge 62 "F" // FieldEdge 63}; 64 65void PointsToNode::dump(bool print_state) const { 66 NodeType nt = node_type(); 67 tty->print("%s ", node_type_names[(int) nt]); 68 if (print_state) { 69 EscapeState es = escape_state(); 70 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR"); 71 } 72 tty->print("[["); 73 for (uint i = 0; i < edge_count(); i++) { 74 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]); 75 } 76 tty->print("]] "); 77 if (_node == NULL) 78 tty->print_cr("<null>"); 79 else 80 _node->dump(); 81} 82#endif 83 84ConnectionGraph::ConnectionGraph(Compile * C) : 85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()), 86 _processed(C->comp_arena()), 87 _collecting(true), 88 _compile(C), 89 _node_map(C->comp_arena()) { 90 91 _phantom_object = C->top()->_idx, 92 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true); 93 94 // Add ConP(#NULL) and ConN(#NULL) nodes. 95 PhaseGVN* igvn = C->initial_gvn(); 96 Node* oop_null = igvn->zerocon(T_OBJECT); 97 _oop_null = oop_null->_idx; 98 assert(_oop_null < C->unique(), "should be created already"); 99 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 100 101 if (UseCompressedOops) { 102 Node* noop_null = igvn->zerocon(T_NARROWOOP); 103 _noop_null = noop_null->_idx; 104 assert(_noop_null < C->unique(), "should be created already"); 105 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true); 106 } 107} 108 109void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) { 110 PointsToNode *f = ptnode_adr(from_i); 111 PointsToNode *t = ptnode_adr(to_i); 112 113 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 114 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge"); 115 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge"); 116 f->add_edge(to_i, PointsToNode::PointsToEdge); 117} 118 119void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) { 120 PointsToNode *f = ptnode_adr(from_i); 121 PointsToNode *t = ptnode_adr(to_i); 122 123 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 124 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge"); 125 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge"); 126 // don't add a self-referential edge, this can occur during removal of 127 // deferred edges 128 if (from_i != to_i) 129 f->add_edge(to_i, PointsToNode::DeferredEdge); 130} 131 132int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) { 133 const Type *adr_type = phase->type(adr); 134 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL && 135 adr->in(AddPNode::Address)->is_Proj() && 136 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 137 // We are computing a raw address for a store captured by an Initialize 138 // compute an appropriate address type. AddP cases #3 and #5 (see below). 139 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 140 assert(offs != Type::OffsetBot || 141 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(), 142 "offset must be a constant or it is initialization of array"); 143 return offs; 144 } 145 const TypePtr *t_ptr = adr_type->isa_ptr(); 146 assert(t_ptr != NULL, "must be a pointer type"); 147 return t_ptr->offset(); 148} 149 150void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) { 151 PointsToNode *f = ptnode_adr(from_i); 152 PointsToNode *t = ptnode_adr(to_i); 153 154 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set"); 155 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge"); 156 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge"); 157 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets"); 158 t->set_offset(offset); 159 160 f->add_edge(to_i, PointsToNode::FieldEdge); 161} 162 163void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) { 164 PointsToNode *npt = ptnode_adr(ni); 165 PointsToNode::EscapeState old_es = npt->escape_state(); 166 if (es > old_es) 167 npt->set_escape_state(es); 168} 169 170void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt, 171 PointsToNode::EscapeState es, bool done) { 172 PointsToNode* ptadr = ptnode_adr(n->_idx); 173 ptadr->_node = n; 174 ptadr->set_node_type(nt); 175 176 // inline set_escape_state(idx, es); 177 PointsToNode::EscapeState old_es = ptadr->escape_state(); 178 if (es > old_es) 179 ptadr->set_escape_state(es); 180 181 if (done) 182 _processed.set(n->_idx); 183} 184 185PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n, PhaseTransform *phase) { 186 uint idx = n->_idx; 187 PointsToNode::EscapeState es; 188 189 // If we are still collecting or there were no non-escaping allocations 190 // we don't know the answer yet 191 if (_collecting) 192 return PointsToNode::UnknownEscape; 193 194 // if the node was created after the escape computation, return 195 // UnknownEscape 196 if (idx >= nodes_size()) 197 return PointsToNode::UnknownEscape; 198 199 es = ptnode_adr(idx)->escape_state(); 200 201 // if we have already computed a value, return it 202 if (es != PointsToNode::UnknownEscape && 203 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject) 204 return es; 205 206 // PointsTo() calls n->uncast() which can return a new ideal node. 207 if (n->uncast()->_idx >= nodes_size()) 208 return PointsToNode::UnknownEscape; 209 210 // compute max escape state of anything this node could point to 211 VectorSet ptset(Thread::current()->resource_area()); 212 PointsTo(ptset, n, phase); 213 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) { 214 uint pt = i.elem; 215 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state(); 216 if (pes > es) 217 es = pes; 218 } 219 // cache the computed escape state 220 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state"); 221 ptnode_adr(idx)->set_escape_state(es); 222 return es; 223} 224 225void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n, PhaseTransform *phase) { 226 VectorSet visited(Thread::current()->resource_area()); 227 GrowableArray<uint> worklist; 228 229#ifdef ASSERT 230 Node *orig_n = n; 231#endif 232 233 n = n->uncast(); 234 PointsToNode* npt = ptnode_adr(n->_idx); 235 236 // If we have a JavaObject, return just that object 237 if (npt->node_type() == PointsToNode::JavaObject) { 238 ptset.set(n->_idx); 239 return; 240 } 241#ifdef ASSERT 242 if (npt->_node == NULL) { 243 if (orig_n != n) 244 orig_n->dump(); 245 n->dump(); 246 assert(npt->_node != NULL, "unregistered node"); 247 } 248#endif 249 worklist.push(n->_idx); 250 while(worklist.length() > 0) { 251 int ni = worklist.pop(); 252 if (visited.test_set(ni)) 253 continue; 254 255 PointsToNode* pn = ptnode_adr(ni); 256 // ensure that all inputs of a Phi have been processed 257 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),""); 258 259 int edges_processed = 0; 260 uint e_cnt = pn->edge_count(); 261 for (uint e = 0; e < e_cnt; e++) { 262 uint etgt = pn->edge_target(e); 263 PointsToNode::EdgeType et = pn->edge_type(e); 264 if (et == PointsToNode::PointsToEdge) { 265 ptset.set(etgt); 266 edges_processed++; 267 } else if (et == PointsToNode::DeferredEdge) { 268 worklist.push(etgt); 269 edges_processed++; 270 } else { 271 assert(false,"neither PointsToEdge or DeferredEdge"); 272 } 273 } 274 if (edges_processed == 0) { 275 // no deferred or pointsto edges found. Assume the value was set 276 // outside this method. Add the phantom object to the pointsto set. 277 ptset.set(_phantom_object); 278 } 279 } 280} 281 282void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) { 283 // This method is most expensive during ConnectionGraph construction. 284 // Reuse vectorSet and an additional growable array for deferred edges. 285 deferred_edges->clear(); 286 visited->Clear(); 287 288 visited->set(ni); 289 PointsToNode *ptn = ptnode_adr(ni); 290 291 // Mark current edges as visited and move deferred edges to separate array. 292 for (uint i = 0; i < ptn->edge_count(); ) { 293 uint t = ptn->edge_target(i); 294#ifdef ASSERT 295 assert(!visited->test_set(t), "expecting no duplications"); 296#else 297 visited->set(t); 298#endif 299 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) { 300 ptn->remove_edge(t, PointsToNode::DeferredEdge); 301 deferred_edges->append(t); 302 } else { 303 i++; 304 } 305 } 306 for (int next = 0; next < deferred_edges->length(); ++next) { 307 uint t = deferred_edges->at(next); 308 PointsToNode *ptt = ptnode_adr(t); 309 uint e_cnt = ptt->edge_count(); 310 for (uint e = 0; e < e_cnt; e++) { 311 uint etgt = ptt->edge_target(e); 312 if (visited->test_set(etgt)) 313 continue; 314 315 PointsToNode::EdgeType et = ptt->edge_type(e); 316 if (et == PointsToNode::PointsToEdge) { 317 add_pointsto_edge(ni, etgt); 318 if(etgt == _phantom_object) { 319 // Special case - field set outside (globally escaping). 320 ptn->set_escape_state(PointsToNode::GlobalEscape); 321 } 322 } else if (et == PointsToNode::DeferredEdge) { 323 deferred_edges->append(etgt); 324 } else { 325 assert(false,"invalid connection graph"); 326 } 327 } 328 } 329} 330 331 332// Add an edge to node given by "to_i" from any field of adr_i whose offset 333// matches "offset" A deferred edge is added if to_i is a LocalVar, and 334// a pointsto edge is added if it is a JavaObject 335 336void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) { 337 PointsToNode* an = ptnode_adr(adr_i); 338 PointsToNode* to = ptnode_adr(to_i); 339 bool deferred = (to->node_type() == PointsToNode::LocalVar); 340 341 for (uint fe = 0; fe < an->edge_count(); fe++) { 342 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 343 int fi = an->edge_target(fe); 344 PointsToNode* pf = ptnode_adr(fi); 345 int po = pf->offset(); 346 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 347 if (deferred) 348 add_deferred_edge(fi, to_i); 349 else 350 add_pointsto_edge(fi, to_i); 351 } 352 } 353} 354 355// Add a deferred edge from node given by "from_i" to any field of adr_i 356// whose offset matches "offset". 357void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) { 358 PointsToNode* an = ptnode_adr(adr_i); 359 for (uint fe = 0; fe < an->edge_count(); fe++) { 360 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge"); 361 int fi = an->edge_target(fe); 362 PointsToNode* pf = ptnode_adr(fi); 363 int po = pf->offset(); 364 if (pf->edge_count() == 0) { 365 // we have not seen any stores to this field, assume it was set outside this method 366 add_pointsto_edge(fi, _phantom_object); 367 } 368 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) { 369 add_deferred_edge(from_i, fi); 370 } 371 } 372} 373 374// Helper functions 375 376static Node* get_addp_base(Node *addp) { 377 assert(addp->is_AddP(), "must be AddP"); 378 // 379 // AddP cases for Base and Address inputs: 380 // case #1. Direct object's field reference: 381 // Allocate 382 // | 383 // Proj #5 ( oop result ) 384 // | 385 // CheckCastPP (cast to instance type) 386 // | | 387 // AddP ( base == address ) 388 // 389 // case #2. Indirect object's field reference: 390 // Phi 391 // | 392 // CastPP (cast to instance type) 393 // | | 394 // AddP ( base == address ) 395 // 396 // case #3. Raw object's field reference for Initialize node: 397 // Allocate 398 // | 399 // Proj #5 ( oop result ) 400 // top | 401 // \ | 402 // AddP ( base == top ) 403 // 404 // case #4. Array's element reference: 405 // {CheckCastPP | CastPP} 406 // | | | 407 // | AddP ( array's element offset ) 408 // | | 409 // AddP ( array's offset ) 410 // 411 // case #5. Raw object's field reference for arraycopy stub call: 412 // The inline_native_clone() case when the arraycopy stub is called 413 // after the allocation before Initialize and CheckCastPP nodes. 414 // Allocate 415 // | 416 // Proj #5 ( oop result ) 417 // | | 418 // AddP ( base == address ) 419 // 420 // case #6. Constant Pool, ThreadLocal, CastX2P or 421 // Raw object's field reference: 422 // {ConP, ThreadLocal, CastX2P, raw Load} 423 // top | 424 // \ | 425 // AddP ( base == top ) 426 // 427 // case #7. Klass's field reference. 428 // LoadKlass 429 // | | 430 // AddP ( base == address ) 431 // 432 // case #8. narrow Klass's field reference. 433 // LoadNKlass 434 // | 435 // DecodeN 436 // | | 437 // AddP ( base == address ) 438 // 439 Node *base = addp->in(AddPNode::Base)->uncast(); 440 if (base->is_top()) { // The AddP case #3 and #6. 441 base = addp->in(AddPNode::Address)->uncast(); 442 while (base->is_AddP()) { 443 // Case #6 (unsafe access) may have several chained AddP nodes. 444 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only"); 445 base = base->in(AddPNode::Address)->uncast(); 446 } 447 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal || 448 base->Opcode() == Op_CastX2P || base->is_DecodeN() || 449 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) || 450 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity"); 451 } 452 return base; 453} 454 455static Node* find_second_addp(Node* addp, Node* n) { 456 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes"); 457 458 Node* addp2 = addp->raw_out(0); 459 if (addp->outcnt() == 1 && addp2->is_AddP() && 460 addp2->in(AddPNode::Base) == n && 461 addp2->in(AddPNode::Address) == addp) { 462 463 assert(addp->in(AddPNode::Base) == n, "expecting the same base"); 464 // 465 // Find array's offset to push it on worklist first and 466 // as result process an array's element offset first (pushed second) 467 // to avoid CastPP for the array's offset. 468 // Otherwise the inserted CastPP (LocalVar) will point to what 469 // the AddP (Field) points to. Which would be wrong since 470 // the algorithm expects the CastPP has the same point as 471 // as AddP's base CheckCastPP (LocalVar). 472 // 473 // ArrayAllocation 474 // | 475 // CheckCastPP 476 // | 477 // memProj (from ArrayAllocation CheckCastPP) 478 // | || 479 // | || Int (element index) 480 // | || | ConI (log(element size)) 481 // | || | / 482 // | || LShift 483 // | || / 484 // | AddP (array's element offset) 485 // | | 486 // | | ConI (array's offset: #12(32-bits) or #24(64-bits)) 487 // | / / 488 // AddP (array's offset) 489 // | 490 // Load/Store (memory operation on array's element) 491 // 492 return addp2; 493 } 494 return NULL; 495} 496 497// 498// Adjust the type and inputs of an AddP which computes the 499// address of a field of an instance 500// 501bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) { 502 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr(); 503 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr"); 504 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr(); 505 if (t == NULL) { 506 // We are computing a raw address for a store captured by an Initialize 507 // compute an appropriate address type (cases #3 and #5). 508 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer"); 509 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation"); 510 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot); 511 assert(offs != Type::OffsetBot, "offset must be a constant"); 512 t = base_t->add_offset(offs)->is_oopptr(); 513 } 514 int inst_id = base_t->instance_id(); 515 assert(!t->is_known_instance() || t->instance_id() == inst_id, 516 "old type must be non-instance or match new type"); 517 518 // The type 't' could be subclass of 'base_t'. 519 // As result t->offset() could be large then base_t's size and it will 520 // cause the failure in add_offset() with narrow oops since TypeOopPtr() 521 // constructor verifies correctness of the offset. 522 // 523 // It could happened on subclass's branch (from the type profiling 524 // inlining) which was not eliminated during parsing since the exactness 525 // of the allocation type was not propagated to the subclass type check. 526 // 527 // Or the type 't' could be not related to 'base_t' at all. 528 // It could happened when CHA type is different from MDO type on a dead path 529 // (for example, from instanceof check) which is not collapsed during parsing. 530 // 531 // Do nothing for such AddP node and don't process its users since 532 // this code branch will go away. 533 // 534 if (!t->is_known_instance() && 535 !base_t->klass()->is_subtype_of(t->klass())) { 536 return false; // bail out 537 } 538 539 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr(); 540 // Do NOT remove the next line: ensure a new alias index is allocated 541 // for the instance type. Note: C++ will not remove it since the call 542 // has side effect. 543 int alias_idx = _compile->get_alias_index(tinst); 544 igvn->set_type(addp, tinst); 545 // record the allocation in the node map 546 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered"); 547 set_map(addp->_idx, get_map(base->_idx)); 548 549 // Set addp's Base and Address to 'base'. 550 Node *abase = addp->in(AddPNode::Base); 551 Node *adr = addp->in(AddPNode::Address); 552 if (adr->is_Proj() && adr->in(0)->is_Allocate() && 553 adr->in(0)->_idx == (uint)inst_id) { 554 // Skip AddP cases #3 and #5. 555 } else { 556 assert(!abase->is_top(), "sanity"); // AddP case #3 557 if (abase != base) { 558 igvn->hash_delete(addp); 559 addp->set_req(AddPNode::Base, base); 560 if (abase == adr) { 561 addp->set_req(AddPNode::Address, base); 562 } else { 563 // AddP case #4 (adr is array's element offset AddP node) 564#ifdef ASSERT 565 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr(); 566 assert(adr->is_AddP() && atype != NULL && 567 atype->instance_id() == inst_id, "array's element offset should be processed first"); 568#endif 569 } 570 igvn->hash_insert(addp); 571 } 572 } 573 // Put on IGVN worklist since at least addp's type was changed above. 574 record_for_optimizer(addp); 575 return true; 576} 577 578// 579// Create a new version of orig_phi if necessary. Returns either the newly 580// created phi or an existing phi. Sets create_new to indicate wheter a new 581// phi was created. Cache the last newly created phi in the node map. 582// 583PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) { 584 Compile *C = _compile; 585 new_created = false; 586 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type()); 587 // nothing to do if orig_phi is bottom memory or matches alias_idx 588 if (phi_alias_idx == alias_idx) { 589 return orig_phi; 590 } 591 // Have we recently created a Phi for this alias index? 592 PhiNode *result = get_map_phi(orig_phi->_idx); 593 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) { 594 return result; 595 } 596 // Previous check may fail when the same wide memory Phi was split into Phis 597 // for different memory slices. Search all Phis for this region. 598 if (result != NULL) { 599 Node* region = orig_phi->in(0); 600 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 601 Node* phi = region->fast_out(i); 602 if (phi->is_Phi() && 603 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) { 604 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice"); 605 return phi->as_Phi(); 606 } 607 } 608 } 609 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) { 610 if (C->do_escape_analysis() == true && !C->failing()) { 611 // Retry compilation without escape analysis. 612 // If this is the first failure, the sentinel string will "stick" 613 // to the Compile object, and the C2Compiler will see it and retry. 614 C->record_failure(C2Compiler::retry_no_escape_analysis()); 615 } 616 return NULL; 617 } 618 orig_phi_worklist.append_if_missing(orig_phi); 619 const TypePtr *atype = C->get_adr_type(alias_idx); 620 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype); 621 C->copy_node_notes_to(result, orig_phi); 622 igvn->set_type(result, result->bottom_type()); 623 record_for_optimizer(result); 624 625 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;) 626 assert(pn == NULL || pn == orig_phi, "wrong node"); 627 set_map(orig_phi->_idx, result); 628 ptnode_adr(orig_phi->_idx)->_node = orig_phi; 629 630 new_created = true; 631 return result; 632} 633 634// 635// Return a new version of Memory Phi "orig_phi" with the inputs having the 636// specified alias index. 637// 638PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) { 639 640 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory"); 641 Compile *C = _compile; 642 bool new_phi_created; 643 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created); 644 if (!new_phi_created) { 645 return result; 646 } 647 648 GrowableArray<PhiNode *> phi_list; 649 GrowableArray<uint> cur_input; 650 651 PhiNode *phi = orig_phi; 652 uint idx = 1; 653 bool finished = false; 654 while(!finished) { 655 while (idx < phi->req()) { 656 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn); 657 if (mem != NULL && mem->is_Phi()) { 658 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created); 659 if (new_phi_created) { 660 // found an phi for which we created a new split, push current one on worklist and begin 661 // processing new one 662 phi_list.push(phi); 663 cur_input.push(idx); 664 phi = mem->as_Phi(); 665 result = newphi; 666 idx = 1; 667 continue; 668 } else { 669 mem = newphi; 670 } 671 } 672 if (C->failing()) { 673 return NULL; 674 } 675 result->set_req(idx++, mem); 676 } 677#ifdef ASSERT 678 // verify that the new Phi has an input for each input of the original 679 assert( phi->req() == result->req(), "must have same number of inputs."); 680 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match"); 681#endif 682 // Check if all new phi's inputs have specified alias index. 683 // Otherwise use old phi. 684 for (uint i = 1; i < phi->req(); i++) { 685 Node* in = result->in(i); 686 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond."); 687 } 688 // we have finished processing a Phi, see if there are any more to do 689 finished = (phi_list.length() == 0 ); 690 if (!finished) { 691 phi = phi_list.pop(); 692 idx = cur_input.pop(); 693 PhiNode *prev_result = get_map_phi(phi->_idx); 694 prev_result->set_req(idx++, result); 695 result = prev_result; 696 } 697 } 698 return result; 699} 700 701 702// 703// The next methods are derived from methods in MemNode. 704// 705static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) { 706 Node *mem = mmem; 707 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 708 // means an array I have not precisely typed yet. Do not do any 709 // alias stuff with it any time soon. 710 if( tinst->base() != Type::AnyPtr && 711 !(tinst->klass()->is_java_lang_Object() && 712 tinst->offset() == Type::OffsetBot) ) { 713 mem = mmem->memory_at(alias_idx); 714 // Update input if it is progress over what we have now 715 } 716 return mem; 717} 718 719// 720// Move memory users to their memory slices. 721// 722void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) { 723 Compile* C = _compile; 724 725 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr(); 726 assert(tp != NULL, "ptr type"); 727 int alias_idx = C->get_alias_index(tp); 728 int general_idx = C->get_general_index(alias_idx); 729 730 // Move users first 731 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 732 Node* use = n->fast_out(i); 733 if (use->is_MergeMem()) { 734 MergeMemNode* mmem = use->as_MergeMem(); 735 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice"); 736 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) { 737 continue; // Nothing to do 738 } 739 // Replace previous general reference to mem node. 740 uint orig_uniq = C->unique(); 741 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 742 assert(orig_uniq == C->unique(), "no new nodes"); 743 mmem->set_memory_at(general_idx, m); 744 --imax; 745 --i; 746 } else if (use->is_MemBar()) { 747 assert(!use->is_Initialize(), "initializing stores should not be moved"); 748 if (use->req() > MemBarNode::Precedent && 749 use->in(MemBarNode::Precedent) == n) { 750 // Don't move related membars. 751 record_for_optimizer(use); 752 continue; 753 } 754 tp = use->as_MemBar()->adr_type()->isa_ptr(); 755 if (tp != NULL && C->get_alias_index(tp) == alias_idx || 756 alias_idx == general_idx) { 757 continue; // Nothing to do 758 } 759 // Move to general memory slice. 760 uint orig_uniq = C->unique(); 761 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn); 762 assert(orig_uniq == C->unique(), "no new nodes"); 763 igvn->hash_delete(use); 764 imax -= use->replace_edge(n, m); 765 igvn->hash_insert(use); 766 record_for_optimizer(use); 767 --i; 768#ifdef ASSERT 769 } else if (use->is_Mem()) { 770 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) { 771 // Don't move related cardmark. 772 continue; 773 } 774 // Memory nodes should have new memory input. 775 tp = igvn->type(use->in(MemNode::Address))->isa_ptr(); 776 assert(tp != NULL, "ptr type"); 777 int idx = C->get_alias_index(tp); 778 assert(get_map(use->_idx) != NULL || idx == alias_idx, 779 "Following memory nodes should have new memory input or be on the same memory slice"); 780 } else if (use->is_Phi()) { 781 // Phi nodes should be split and moved already. 782 tp = use->as_Phi()->adr_type()->isa_ptr(); 783 assert(tp != NULL, "ptr type"); 784 int idx = C->get_alias_index(tp); 785 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice"); 786 } else { 787 use->dump(); 788 assert(false, "should not be here"); 789#endif 790 } 791 } 792} 793 794// 795// Search memory chain of "mem" to find a MemNode whose address 796// is the specified alias index. 797// 798Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) { 799 if (orig_mem == NULL) 800 return orig_mem; 801 Compile* C = phase->C; 802 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr(); 803 bool is_instance = (tinst != NULL) && tinst->is_known_instance(); 804 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 805 Node *prev = NULL; 806 Node *result = orig_mem; 807 while (prev != result) { 808 prev = result; 809 if (result == start_mem) 810 break; // hit one of our sentinels 811 if (result->is_Mem()) { 812 const Type *at = phase->type(result->in(MemNode::Address)); 813 if (at != Type::TOP) { 814 assert (at->isa_ptr() != NULL, "pointer type required."); 815 int idx = C->get_alias_index(at->is_ptr()); 816 if (idx == alias_idx) 817 break; 818 } 819 result = result->in(MemNode::Memory); 820 } 821 if (!is_instance) 822 continue; // don't search further for non-instance types 823 // skip over a call which does not affect this memory slice 824 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 825 Node *proj_in = result->in(0); 826 if (proj_in->is_Allocate() && proj_in->_idx == (uint)tinst->instance_id()) { 827 break; // hit one of our sentinels 828 } else if (proj_in->is_Call()) { 829 CallNode *call = proj_in->as_Call(); 830 if (!call->may_modify(tinst, phase)) { 831 result = call->in(TypeFunc::Memory); 832 } 833 } else if (proj_in->is_Initialize()) { 834 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 835 // Stop if this is the initialization for the object instance which 836 // which contains this memory slice, otherwise skip over it. 837 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) { 838 result = proj_in->in(TypeFunc::Memory); 839 } 840 } else if (proj_in->is_MemBar()) { 841 result = proj_in->in(TypeFunc::Memory); 842 } 843 } else if (result->is_MergeMem()) { 844 MergeMemNode *mmem = result->as_MergeMem(); 845 result = step_through_mergemem(mmem, alias_idx, tinst); 846 if (result == mmem->base_memory()) { 847 // Didn't find instance memory, search through general slice recursively. 848 result = mmem->memory_at(C->get_general_index(alias_idx)); 849 result = find_inst_mem(result, alias_idx, orig_phis, phase); 850 if (C->failing()) { 851 return NULL; 852 } 853 mmem->set_memory_at(alias_idx, result); 854 } 855 } else if (result->is_Phi() && 856 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) { 857 Node *un = result->as_Phi()->unique_input(phase); 858 if (un != NULL) { 859 orig_phis.append_if_missing(result->as_Phi()); 860 result = un; 861 } else { 862 break; 863 } 864 } else if (result->is_ClearArray()) { 865 if (!ClearArrayNode::step_through(&result, (uint)tinst->instance_id(), phase)) { 866 // Can not bypass initialization of the instance 867 // we are looking for. 868 break; 869 } 870 // Otherwise skip it (the call updated 'result' value). 871 } else if (result->Opcode() == Op_SCMemProj) { 872 assert(result->in(0)->is_LoadStore(), "sanity"); 873 const Type *at = phase->type(result->in(0)->in(MemNode::Address)); 874 if (at != Type::TOP) { 875 assert (at->isa_ptr() != NULL, "pointer type required."); 876 int idx = C->get_alias_index(at->is_ptr()); 877 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field"); 878 break; 879 } 880 result = result->in(0)->in(MemNode::Memory); 881 } 882 } 883 if (result->is_Phi()) { 884 PhiNode *mphi = result->as_Phi(); 885 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 886 const TypePtr *t = mphi->adr_type(); 887 if (C->get_alias_index(t) != alias_idx) { 888 // Create a new Phi with the specified alias index type. 889 result = split_memory_phi(mphi, alias_idx, orig_phis, phase); 890 } else if (!is_instance) { 891 // Push all non-instance Phis on the orig_phis worklist to update inputs 892 // during Phase 4 if needed. 893 orig_phis.append_if_missing(mphi); 894 } 895 } 896 // the result is either MemNode, PhiNode, InitializeNode. 897 return result; 898} 899 900// 901// Convert the types of unescaped object to instance types where possible, 902// propagate the new type information through the graph, and update memory 903// edges and MergeMem inputs to reflect the new type. 904// 905// We start with allocations (and calls which may be allocations) on alloc_worklist. 906// The processing is done in 4 phases: 907// 908// Phase 1: Process possible allocations from alloc_worklist. Create instance 909// types for the CheckCastPP for allocations where possible. 910// Propagate the the new types through users as follows: 911// casts and Phi: push users on alloc_worklist 912// AddP: cast Base and Address inputs to the instance type 913// push any AddP users on alloc_worklist and push any memnode 914// users onto memnode_worklist. 915// Phase 2: Process MemNode's from memnode_worklist. compute new address type and 916// search the Memory chain for a store with the appropriate type 917// address type. If a Phi is found, create a new version with 918// the appropriate memory slices from each of the Phi inputs. 919// For stores, process the users as follows: 920// MemNode: push on memnode_worklist 921// MergeMem: push on mergemem_worklist 922// Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice 923// moving the first node encountered of each instance type to the 924// the input corresponding to its alias index. 925// appropriate memory slice. 926// Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes. 927// 928// In the following example, the CheckCastPP nodes are the cast of allocation 929// results and the allocation of node 29 is unescaped and eligible to be an 930// instance type. 931// 932// We start with: 933// 934// 7 Parm #memory 935// 10 ConI "12" 936// 19 CheckCastPP "Foo" 937// 20 AddP _ 19 19 10 Foo+12 alias_index=4 938// 29 CheckCastPP "Foo" 939// 30 AddP _ 29 29 10 Foo+12 alias_index=4 940// 941// 40 StoreP 25 7 20 ... alias_index=4 942// 50 StoreP 35 40 30 ... alias_index=4 943// 60 StoreP 45 50 20 ... alias_index=4 944// 70 LoadP _ 60 30 ... alias_index=4 945// 80 Phi 75 50 60 Memory alias_index=4 946// 90 LoadP _ 80 30 ... alias_index=4 947// 100 LoadP _ 80 20 ... alias_index=4 948// 949// 950// Phase 1 creates an instance type for node 29 assigning it an instance id of 24 951// and creating a new alias index for node 30. This gives: 952// 953// 7 Parm #memory 954// 10 ConI "12" 955// 19 CheckCastPP "Foo" 956// 20 AddP _ 19 19 10 Foo+12 alias_index=4 957// 29 CheckCastPP "Foo" iid=24 958// 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 959// 960// 40 StoreP 25 7 20 ... alias_index=4 961// 50 StoreP 35 40 30 ... alias_index=6 962// 60 StoreP 45 50 20 ... alias_index=4 963// 70 LoadP _ 60 30 ... alias_index=6 964// 80 Phi 75 50 60 Memory alias_index=4 965// 90 LoadP _ 80 30 ... alias_index=6 966// 100 LoadP _ 80 20 ... alias_index=4 967// 968// In phase 2, new memory inputs are computed for the loads and stores, 969// And a new version of the phi is created. In phase 4, the inputs to 970// node 80 are updated and then the memory nodes are updated with the 971// values computed in phase 2. This results in: 972// 973// 7 Parm #memory 974// 10 ConI "12" 975// 19 CheckCastPP "Foo" 976// 20 AddP _ 19 19 10 Foo+12 alias_index=4 977// 29 CheckCastPP "Foo" iid=24 978// 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24 979// 980// 40 StoreP 25 7 20 ... alias_index=4 981// 50 StoreP 35 7 30 ... alias_index=6 982// 60 StoreP 45 40 20 ... alias_index=4 983// 70 LoadP _ 50 30 ... alias_index=6 984// 80 Phi 75 40 60 Memory alias_index=4 985// 120 Phi 75 50 50 Memory alias_index=6 986// 90 LoadP _ 120 30 ... alias_index=6 987// 100 LoadP _ 80 20 ... alias_index=4 988// 989void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) { 990 GrowableArray<Node *> memnode_worklist; 991 GrowableArray<PhiNode *> orig_phis; 992 993 PhaseGVN *igvn = _compile->initial_gvn(); 994 uint new_index_start = (uint) _compile->num_alias_types(); 995 Arena* arena = Thread::current()->resource_area(); 996 VectorSet visited(arena); 997 VectorSet ptset(arena); 998 999 1000 // Phase 1: Process possible allocations from alloc_worklist. 1001 // Create instance types for the CheckCastPP for allocations where possible. 1002 // 1003 // (Note: don't forget to change the order of the second AddP node on 1004 // the alloc_worklist if the order of the worklist processing is changed, 1005 // see the comment in find_second_addp().) 1006 // 1007 while (alloc_worklist.length() != 0) { 1008 Node *n = alloc_worklist.pop(); 1009 uint ni = n->_idx; 1010 const TypeOopPtr* tinst = NULL; 1011 if (n->is_Call()) { 1012 CallNode *alloc = n->as_Call(); 1013 // copy escape information to call node 1014 PointsToNode* ptn = ptnode_adr(alloc->_idx); 1015 PointsToNode::EscapeState es = escape_state(alloc, igvn); 1016 // We have an allocation or call which returns a Java object, 1017 // see if it is unescaped. 1018 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable) 1019 continue; 1020 1021 // Find CheckCastPP for the allocate or for the return value of a call 1022 n = alloc->result_cast(); 1023 if (n == NULL) { // No uses except Initialize node 1024 if (alloc->is_Allocate()) { 1025 // Set the scalar_replaceable flag for allocation 1026 // so it could be eliminated if it has no uses. 1027 alloc->as_Allocate()->_is_scalar_replaceable = true; 1028 } 1029 continue; 1030 } 1031 if (!n->is_CheckCastPP()) { // not unique CheckCastPP. 1032 assert(!alloc->is_Allocate(), "allocation should have unique type"); 1033 continue; 1034 } 1035 1036 // The inline code for Object.clone() casts the allocation result to 1037 // java.lang.Object and then to the actual type of the allocated 1038 // object. Detect this case and use the second cast. 1039 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when 1040 // the allocation result is cast to java.lang.Object and then 1041 // to the actual Array type. 1042 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL 1043 && (alloc->is_AllocateArray() || 1044 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) { 1045 Node *cast2 = NULL; 1046 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1047 Node *use = n->fast_out(i); 1048 if (use->is_CheckCastPP()) { 1049 cast2 = use; 1050 break; 1051 } 1052 } 1053 if (cast2 != NULL) { 1054 n = cast2; 1055 } else { 1056 // Non-scalar replaceable if the allocation type is unknown statically 1057 // (reflection allocation), the object can't be restored during 1058 // deoptimization without precise type. 1059 continue; 1060 } 1061 } 1062 if (alloc->is_Allocate()) { 1063 // Set the scalar_replaceable flag for allocation 1064 // so it could be eliminated. 1065 alloc->as_Allocate()->_is_scalar_replaceable = true; 1066 } 1067 set_escape_state(n->_idx, es); 1068 // in order for an object to be scalar-replaceable, it must be: 1069 // - a direct allocation (not a call returning an object) 1070 // - non-escaping 1071 // - eligible to be a unique type 1072 // - not determined to be ineligible by escape analysis 1073 assert(ptnode_adr(alloc->_idx)->_node != NULL && 1074 ptnode_adr(n->_idx)->_node != NULL, "should be registered"); 1075 set_map(alloc->_idx, n); 1076 set_map(n->_idx, alloc); 1077 const TypeOopPtr *t = igvn->type(n)->isa_oopptr(); 1078 if (t == NULL) 1079 continue; // not a TypeInstPtr 1080 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni); 1081 igvn->hash_delete(n); 1082 igvn->set_type(n, tinst); 1083 n->raise_bottom_type(tinst); 1084 igvn->hash_insert(n); 1085 record_for_optimizer(n); 1086 if (alloc->is_Allocate() && ptn->_scalar_replaceable && 1087 (t->isa_instptr() || t->isa_aryptr())) { 1088 1089 // First, put on the worklist all Field edges from Connection Graph 1090 // which is more accurate then putting immediate users from Ideal Graph. 1091 for (uint e = 0; e < ptn->edge_count(); e++) { 1092 Node *use = ptnode_adr(ptn->edge_target(e))->_node; 1093 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(), 1094 "only AddP nodes are Field edges in CG"); 1095 if (use->outcnt() > 0) { // Don't process dead nodes 1096 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base)); 1097 if (addp2 != NULL) { 1098 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1099 alloc_worklist.append_if_missing(addp2); 1100 } 1101 alloc_worklist.append_if_missing(use); 1102 } 1103 } 1104 1105 // An allocation may have an Initialize which has raw stores. Scan 1106 // the users of the raw allocation result and push AddP users 1107 // on alloc_worklist. 1108 Node *raw_result = alloc->proj_out(TypeFunc::Parms); 1109 assert (raw_result != NULL, "must have an allocation result"); 1110 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) { 1111 Node *use = raw_result->fast_out(i); 1112 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes 1113 Node* addp2 = find_second_addp(use, raw_result); 1114 if (addp2 != NULL) { 1115 assert(alloc->is_AllocateArray(),"array allocation was expected"); 1116 alloc_worklist.append_if_missing(addp2); 1117 } 1118 alloc_worklist.append_if_missing(use); 1119 } else if (use->is_MemBar()) { 1120 memnode_worklist.append_if_missing(use); 1121 } 1122 } 1123 } 1124 } else if (n->is_AddP()) { 1125 ptset.Clear(); 1126 PointsTo(ptset, get_addp_base(n), igvn); 1127 assert(ptset.Size() == 1, "AddP address is unique"); 1128 uint elem = ptset.getelem(); // Allocation node's index 1129 if (elem == _phantom_object) { 1130 assert(false, "escaped allocation"); 1131 continue; // Assume the value was set outside this method. 1132 } 1133 Node *base = get_map(elem); // CheckCastPP node 1134 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path 1135 tinst = igvn->type(base)->isa_oopptr(); 1136 } else if (n->is_Phi() || 1137 n->is_CheckCastPP() || 1138 n->is_EncodeP() || 1139 n->is_DecodeN() || 1140 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) { 1141 if (visited.test_set(n->_idx)) { 1142 assert(n->is_Phi(), "loops only through Phi's"); 1143 continue; // already processed 1144 } 1145 ptset.Clear(); 1146 PointsTo(ptset, n, igvn); 1147 if (ptset.Size() == 1) { 1148 uint elem = ptset.getelem(); // Allocation node's index 1149 if (elem == _phantom_object) { 1150 assert(false, "escaped allocation"); 1151 continue; // Assume the value was set outside this method. 1152 } 1153 Node *val = get_map(elem); // CheckCastPP node 1154 TypeNode *tn = n->as_Type(); 1155 tinst = igvn->type(val)->isa_oopptr(); 1156 assert(tinst != NULL && tinst->is_known_instance() && 1157 (uint)tinst->instance_id() == elem , "instance type expected."); 1158 1159 const Type *tn_type = igvn->type(tn); 1160 const TypeOopPtr *tn_t; 1161 if (tn_type->isa_narrowoop()) { 1162 tn_t = tn_type->make_ptr()->isa_oopptr(); 1163 } else { 1164 tn_t = tn_type->isa_oopptr(); 1165 } 1166 1167 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) { 1168 if (tn_type->isa_narrowoop()) { 1169 tn_type = tinst->make_narrowoop(); 1170 } else { 1171 tn_type = tinst; 1172 } 1173 igvn->hash_delete(tn); 1174 igvn->set_type(tn, tn_type); 1175 tn->set_type(tn_type); 1176 igvn->hash_insert(tn); 1177 record_for_optimizer(n); 1178 } else { 1179 assert(tn_type == TypePtr::NULL_PTR || 1180 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()), 1181 "unexpected type"); 1182 continue; // Skip dead path with different type 1183 } 1184 } 1185 } else { 1186 debug_only(n->dump();) 1187 assert(false, "EA: unexpected node"); 1188 continue; 1189 } 1190 // push allocation's users on appropriate worklist 1191 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1192 Node *use = n->fast_out(i); 1193 if(use->is_Mem() && use->in(MemNode::Address) == n) { 1194 // Load/store to instance's field 1195 memnode_worklist.append_if_missing(use); 1196 } else if (use->is_MemBar()) { 1197 memnode_worklist.append_if_missing(use); 1198 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes 1199 Node* addp2 = find_second_addp(use, n); 1200 if (addp2 != NULL) { 1201 alloc_worklist.append_if_missing(addp2); 1202 } 1203 alloc_worklist.append_if_missing(use); 1204 } else if (use->is_Phi() || 1205 use->is_CheckCastPP() || 1206 use->is_EncodeP() || 1207 use->is_DecodeN() || 1208 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) { 1209 alloc_worklist.append_if_missing(use); 1210#ifdef ASSERT 1211 } else if (use->is_Mem()) { 1212 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path"); 1213 } else if (use->is_MergeMem()) { 1214 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1215 } else if (use->is_SafePoint()) { 1216 // Look for MergeMem nodes for calls which reference unique allocation 1217 // (through CheckCastPP nodes) even for debug info. 1218 Node* m = use->in(TypeFunc::Memory); 1219 if (m->is_MergeMem()) { 1220 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1221 } 1222 } else { 1223 uint op = use->Opcode(); 1224 if (!(op == Op_CmpP || op == Op_Conv2B || 1225 op == Op_CastP2X || op == Op_StoreCM || 1226 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || 1227 op == Op_StrEquals || op == Op_StrIndexOf)) { 1228 n->dump(); 1229 use->dump(); 1230 assert(false, "EA: missing allocation reference path"); 1231 } 1232#endif 1233 } 1234 } 1235 1236 } 1237 // New alias types were created in split_AddP(). 1238 uint new_index_end = (uint) _compile->num_alias_types(); 1239 1240 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and 1241 // compute new values for Memory inputs (the Memory inputs are not 1242 // actually updated until phase 4.) 1243 if (memnode_worklist.length() == 0) 1244 return; // nothing to do 1245 1246 while (memnode_worklist.length() != 0) { 1247 Node *n = memnode_worklist.pop(); 1248 if (visited.test_set(n->_idx)) 1249 continue; 1250 if (n->is_Phi() || n->is_ClearArray()) { 1251 // we don't need to do anything, but the users must be pushed 1252 } else if (n->is_MemBar()) { // Initialize, MemBar nodes 1253 // we don't need to do anything, but the users must be pushed 1254 n = n->as_MemBar()->proj_out(TypeFunc::Memory); 1255 if (n == NULL) 1256 continue; 1257 } else { 1258 assert(n->is_Mem(), "memory node required."); 1259 Node *addr = n->in(MemNode::Address); 1260 const Type *addr_t = igvn->type(addr); 1261 if (addr_t == Type::TOP) 1262 continue; 1263 assert (addr_t->isa_ptr() != NULL, "pointer type required."); 1264 int alias_idx = _compile->get_alias_index(addr_t->is_ptr()); 1265 assert ((uint)alias_idx < new_index_end, "wrong alias index"); 1266 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn); 1267 if (_compile->failing()) { 1268 return; 1269 } 1270 if (mem != n->in(MemNode::Memory)) { 1271 // We delay the memory edge update since we need old one in 1272 // MergeMem code below when instances memory slices are separated. 1273 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;) 1274 assert(pn == NULL || pn == n, "wrong node"); 1275 set_map(n->_idx, mem); 1276 ptnode_adr(n->_idx)->_node = n; 1277 } 1278 if (n->is_Load()) { 1279 continue; // don't push users 1280 } else if (n->is_LoadStore()) { 1281 // get the memory projection 1282 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1283 Node *use = n->fast_out(i); 1284 if (use->Opcode() == Op_SCMemProj) { 1285 n = use; 1286 break; 1287 } 1288 } 1289 assert(n->Opcode() == Op_SCMemProj, "memory projection required"); 1290 } 1291 } 1292 // push user on appropriate worklist 1293 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1294 Node *use = n->fast_out(i); 1295 if (use->is_Phi() || use->is_ClearArray()) { 1296 memnode_worklist.append_if_missing(use); 1297 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) { 1298 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores 1299 continue; 1300 memnode_worklist.append_if_missing(use); 1301 } else if (use->is_MemBar()) { 1302 memnode_worklist.append_if_missing(use); 1303#ifdef ASSERT 1304 } else if(use->is_Mem()) { 1305 assert(use->in(MemNode::Memory) != n, "EA: missing memory path"); 1306 } else if (use->is_MergeMem()) { 1307 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist"); 1308 } else { 1309 uint op = use->Opcode(); 1310 if (!(op == Op_StoreCM || 1311 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL && 1312 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) || 1313 op == Op_AryEq || op == Op_StrComp || 1314 op == Op_StrEquals || op == Op_StrIndexOf)) { 1315 n->dump(); 1316 use->dump(); 1317 assert(false, "EA: missing memory path"); 1318 } 1319#endif 1320 } 1321 } 1322 } 1323 1324 // Phase 3: Process MergeMem nodes from mergemem_worklist. 1325 // Walk each memory slice moving the first node encountered of each 1326 // instance type to the the input corresponding to its alias index. 1327 uint length = _mergemem_worklist.length(); 1328 for( uint next = 0; next < length; ++next ) { 1329 MergeMemNode* nmm = _mergemem_worklist.at(next); 1330 assert(!visited.test_set(nmm->_idx), "should not be visited before"); 1331 // Note: we don't want to use MergeMemStream here because we only want to 1332 // scan inputs which exist at the start, not ones we add during processing. 1333 // Note 2: MergeMem may already contains instance memory slices added 1334 // during find_inst_mem() call when memory nodes were processed above. 1335 igvn->hash_delete(nmm); 1336 uint nslices = nmm->req(); 1337 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) { 1338 Node* mem = nmm->in(i); 1339 Node* cur = NULL; 1340 if (mem == NULL || mem->is_top()) 1341 continue; 1342 // First, update mergemem by moving memory nodes to corresponding slices 1343 // if their type became more precise since this mergemem was created. 1344 while (mem->is_Mem()) { 1345 const Type *at = igvn->type(mem->in(MemNode::Address)); 1346 if (at != Type::TOP) { 1347 assert (at->isa_ptr() != NULL, "pointer type required."); 1348 uint idx = (uint)_compile->get_alias_index(at->is_ptr()); 1349 if (idx == i) { 1350 if (cur == NULL) 1351 cur = mem; 1352 } else { 1353 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) { 1354 nmm->set_memory_at(idx, mem); 1355 } 1356 } 1357 } 1358 mem = mem->in(MemNode::Memory); 1359 } 1360 nmm->set_memory_at(i, (cur != NULL) ? cur : mem); 1361 // Find any instance of the current type if we haven't encountered 1362 // already a memory slice of the instance along the memory chain. 1363 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1364 if((uint)_compile->get_general_index(ni) == i) { 1365 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni); 1366 if (nmm->is_empty_memory(m)) { 1367 Node* result = find_inst_mem(mem, ni, orig_phis, igvn); 1368 if (_compile->failing()) { 1369 return; 1370 } 1371 nmm->set_memory_at(ni, result); 1372 } 1373 } 1374 } 1375 } 1376 // Find the rest of instances values 1377 for (uint ni = new_index_start; ni < new_index_end; ni++) { 1378 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr(); 1379 Node* result = step_through_mergemem(nmm, ni, tinst); 1380 if (result == nmm->base_memory()) { 1381 // Didn't find instance memory, search through general slice recursively. 1382 result = nmm->memory_at(_compile->get_general_index(ni)); 1383 result = find_inst_mem(result, ni, orig_phis, igvn); 1384 if (_compile->failing()) { 1385 return; 1386 } 1387 nmm->set_memory_at(ni, result); 1388 } 1389 } 1390 igvn->hash_insert(nmm); 1391 record_for_optimizer(nmm); 1392 } 1393 1394 // Phase 4: Update the inputs of non-instance memory Phis and 1395 // the Memory input of memnodes 1396 // First update the inputs of any non-instance Phi's from 1397 // which we split out an instance Phi. Note we don't have 1398 // to recursively process Phi's encounted on the input memory 1399 // chains as is done in split_memory_phi() since they will 1400 // also be processed here. 1401 for (int j = 0; j < orig_phis.length(); j++) { 1402 PhiNode *phi = orig_phis.at(j); 1403 int alias_idx = _compile->get_alias_index(phi->adr_type()); 1404 igvn->hash_delete(phi); 1405 for (uint i = 1; i < phi->req(); i++) { 1406 Node *mem = phi->in(i); 1407 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn); 1408 if (_compile->failing()) { 1409 return; 1410 } 1411 if (mem != new_mem) { 1412 phi->set_req(i, new_mem); 1413 } 1414 } 1415 igvn->hash_insert(phi); 1416 record_for_optimizer(phi); 1417 } 1418 1419 // Update the memory inputs of MemNodes with the value we computed 1420 // in Phase 2 and move stores memory users to corresponding memory slices. 1421#ifdef ASSERT 1422 visited.Clear(); 1423 Node_Stack old_mems(arena, _compile->unique() >> 2); 1424#endif 1425 for (uint i = 0; i < nodes_size(); i++) { 1426 Node *nmem = get_map(i); 1427 if (nmem != NULL) { 1428 Node *n = ptnode_adr(i)->_node; 1429 assert(n != NULL, "sanity"); 1430 if (n->is_Mem()) { 1431#ifdef ASSERT 1432 Node* old_mem = n->in(MemNode::Memory); 1433 if (!visited.test_set(old_mem->_idx)) { 1434 old_mems.push(old_mem, old_mem->outcnt()); 1435 } 1436#endif 1437 assert(n->in(MemNode::Memory) != nmem, "sanity"); 1438 if (!n->is_Load()) { 1439 // Move memory users of a store first. 1440 move_inst_mem(n, orig_phis, igvn); 1441 } 1442 // Now update memory input 1443 igvn->hash_delete(n); 1444 n->set_req(MemNode::Memory, nmem); 1445 igvn->hash_insert(n); 1446 record_for_optimizer(n); 1447 } else { 1448 assert(n->is_Allocate() || n->is_CheckCastPP() || 1449 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()"); 1450 } 1451 } 1452 } 1453#ifdef ASSERT 1454 // Verify that memory was split correctly 1455 while (old_mems.is_nonempty()) { 1456 Node* old_mem = old_mems.node(); 1457 uint old_cnt = old_mems.index(); 1458 old_mems.pop(); 1459 assert(old_cnt = old_mem->outcnt(), "old mem could be lost"); 1460 } 1461#endif 1462} 1463 1464bool ConnectionGraph::has_candidates(Compile *C) { 1465 // EA brings benefits only when the code has allocations and/or locks which 1466 // are represented by ideal Macro nodes. 1467 int cnt = C->macro_count(); 1468 for( int i=0; i < cnt; i++ ) { 1469 Node *n = C->macro_node(i); 1470 if ( n->is_Allocate() ) 1471 return true; 1472 if( n->is_Lock() ) { 1473 Node* obj = n->as_Lock()->obj_node()->uncast(); 1474 if( !(obj->is_Parm() || obj->is_Con()) ) 1475 return true; 1476 } 1477 } 1478 return false; 1479} 1480 1481bool ConnectionGraph::compute_escape() { 1482 Compile* C = _compile; 1483 1484 // 1. Populate Connection Graph (CG) with Ideal nodes. 1485 1486 Unique_Node_List worklist_init; 1487 worklist_init.map(C->unique(), NULL); // preallocate space 1488 1489 // Initialize worklist 1490 if (C->root() != NULL) { 1491 worklist_init.push(C->root()); 1492 } 1493 1494 GrowableArray<int> cg_worklist; 1495 PhaseGVN* igvn = C->initial_gvn(); 1496 bool has_allocations = false; 1497 1498 // Push all useful nodes onto CG list and set their type. 1499 for( uint next = 0; next < worklist_init.size(); ++next ) { 1500 Node* n = worklist_init.at(next); 1501 record_for_escape_analysis(n, igvn); 1502 // Only allocations and java static calls results are checked 1503 // for an escape status. See process_call_result() below. 1504 if (n->is_Allocate() || n->is_CallStaticJava() && 1505 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) { 1506 has_allocations = true; 1507 } 1508 if(n->is_AddP()) { 1509 // Collect address nodes which directly reference an allocation. 1510 // Use them during stage 3 below to build initial connection graph 1511 // field edges. Other field edges could be added after StoreP/LoadP 1512 // nodes are processed during stage 4 below. 1513 Node* base = get_addp_base(n); 1514 if(base->is_Proj() && base->in(0)->is_Allocate()) { 1515 cg_worklist.append(n->_idx); 1516 } 1517 } else if (n->is_MergeMem()) { 1518 // Collect all MergeMem nodes to add memory slices for 1519 // scalar replaceable objects in split_unique_types(). 1520 _mergemem_worklist.append(n->as_MergeMem()); 1521 } 1522 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1523 Node* m = n->fast_out(i); // Get user 1524 worklist_init.push(m); 1525 } 1526 } 1527 1528 if (!has_allocations) { 1529 _collecting = false; 1530 return false; // Nothing to do. 1531 } 1532 1533 // 2. First pass to create simple CG edges (doesn't require to walk CG). 1534 uint delayed_size = _delayed_worklist.size(); 1535 for( uint next = 0; next < delayed_size; ++next ) { 1536 Node* n = _delayed_worklist.at(next); 1537 build_connection_graph(n, igvn); 1538 } 1539 1540 // 3. Pass to create fields edges (Allocate -F-> AddP). 1541 uint cg_length = cg_worklist.length(); 1542 for( uint next = 0; next < cg_length; ++next ) { 1543 int ni = cg_worklist.at(next); 1544 build_connection_graph(ptnode_adr(ni)->_node, igvn); 1545 } 1546 1547 cg_worklist.clear(); 1548 cg_worklist.append(_phantom_object); 1549 1550 // 4. Build Connection Graph which need 1551 // to walk the connection graph. 1552 for (uint ni = 0; ni < nodes_size(); ni++) { 1553 PointsToNode* ptn = ptnode_adr(ni); 1554 Node *n = ptn->_node; 1555 if (n != NULL) { // Call, AddP, LoadP, StoreP 1556 build_connection_graph(n, igvn); 1557 if (ptn->node_type() != PointsToNode::UnknownType) 1558 cg_worklist.append(n->_idx); // Collect CG nodes 1559 } 1560 } 1561 1562 Arena* arena = Thread::current()->resource_area(); 1563 VectorSet ptset(arena); 1564 GrowableArray<uint> deferred_edges; 1565 VectorSet visited(arena); 1566 1567 // 5. Remove deferred edges from the graph and adjust 1568 // escape state of nonescaping objects. 1569 cg_length = cg_worklist.length(); 1570 for( uint next = 0; next < cg_length; ++next ) { 1571 int ni = cg_worklist.at(next); 1572 PointsToNode* ptn = ptnode_adr(ni); 1573 PointsToNode::NodeType nt = ptn->node_type(); 1574 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) { 1575 remove_deferred(ni, &deferred_edges, &visited); 1576 Node *n = ptn->_node; 1577 if (n->is_AddP()) { 1578 // Search for objects which are not scalar replaceable 1579 // and adjust their escape state. 1580 verify_escape_state(ni, ptset, igvn); 1581 } 1582 } 1583 } 1584 1585 // 6. Propagate escape states. 1586 GrowableArray<int> worklist; 1587 bool has_non_escaping_obj = false; 1588 1589 // push all GlobalEscape nodes on the worklist 1590 for( uint next = 0; next < cg_length; ++next ) { 1591 int nk = cg_worklist.at(next); 1592 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape) 1593 worklist.push(nk); 1594 } 1595 // mark all nodes reachable from GlobalEscape nodes 1596 while(worklist.length() > 0) { 1597 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1598 uint e_cnt = ptn->edge_count(); 1599 for (uint ei = 0; ei < e_cnt; ei++) { 1600 uint npi = ptn->edge_target(ei); 1601 PointsToNode *np = ptnode_adr(npi); 1602 if (np->escape_state() < PointsToNode::GlobalEscape) { 1603 np->set_escape_state(PointsToNode::GlobalEscape); 1604 worklist.push(npi); 1605 } 1606 } 1607 } 1608 1609 // push all ArgEscape nodes on the worklist 1610 for( uint next = 0; next < cg_length; ++next ) { 1611 int nk = cg_worklist.at(next); 1612 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape) 1613 worklist.push(nk); 1614 } 1615 // mark all nodes reachable from ArgEscape nodes 1616 while(worklist.length() > 0) { 1617 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1618 if (ptn->node_type() == PointsToNode::JavaObject) 1619 has_non_escaping_obj = true; // Non GlobalEscape 1620 uint e_cnt = ptn->edge_count(); 1621 for (uint ei = 0; ei < e_cnt; ei++) { 1622 uint npi = ptn->edge_target(ei); 1623 PointsToNode *np = ptnode_adr(npi); 1624 if (np->escape_state() < PointsToNode::ArgEscape) { 1625 np->set_escape_state(PointsToNode::ArgEscape); 1626 worklist.push(npi); 1627 } 1628 } 1629 } 1630 1631 GrowableArray<Node*> alloc_worklist; 1632 1633 // push all NoEscape nodes on the worklist 1634 for( uint next = 0; next < cg_length; ++next ) { 1635 int nk = cg_worklist.at(next); 1636 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape) 1637 worklist.push(nk); 1638 } 1639 // mark all nodes reachable from NoEscape nodes 1640 while(worklist.length() > 0) { 1641 PointsToNode* ptn = ptnode_adr(worklist.pop()); 1642 if (ptn->node_type() == PointsToNode::JavaObject) 1643 has_non_escaping_obj = true; // Non GlobalEscape 1644 Node* n = ptn->_node; 1645 if (n->is_Allocate() && ptn->_scalar_replaceable ) { 1646 // Push scalar replaceable allocations on alloc_worklist 1647 // for processing in split_unique_types(). 1648 alloc_worklist.append(n); 1649 } 1650 uint e_cnt = ptn->edge_count(); 1651 for (uint ei = 0; ei < e_cnt; ei++) { 1652 uint npi = ptn->edge_target(ei); 1653 PointsToNode *np = ptnode_adr(npi); 1654 if (np->escape_state() < PointsToNode::NoEscape) { 1655 np->set_escape_state(PointsToNode::NoEscape); 1656 worklist.push(npi); 1657 } 1658 } 1659 } 1660 1661 _collecting = false; 1662 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build"); 1663 1664 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0; 1665 if ( has_scalar_replaceable_candidates && 1666 C->AliasLevel() >= 3 && EliminateAllocations ) { 1667 1668 // Now use the escape information to create unique types for 1669 // scalar replaceable objects. 1670 split_unique_types(alloc_worklist); 1671 1672 if (C->failing()) return false; 1673 1674 // Clean up after split unique types. 1675 ResourceMark rm; 1676 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn()); 1677 1678 C->print_method("After Escape Analysis", 2); 1679 1680#ifdef ASSERT 1681 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) { 1682 tty->print("=== No allocations eliminated for "); 1683 C->method()->print_short_name(); 1684 if(!EliminateAllocations) { 1685 tty->print(" since EliminateAllocations is off ==="); 1686 } else if(!has_scalar_replaceable_candidates) { 1687 tty->print(" since there are no scalar replaceable candidates ==="); 1688 } else if(C->AliasLevel() < 3) { 1689 tty->print(" since AliasLevel < 3 ==="); 1690 } 1691 tty->cr(); 1692#endif 1693 } 1694 return has_non_escaping_obj; 1695} 1696 1697// Search for objects which are not scalar replaceable. 1698void ConnectionGraph::verify_escape_state(int nidx, VectorSet& ptset, PhaseTransform* phase) { 1699 PointsToNode* ptn = ptnode_adr(nidx); 1700 Node* n = ptn->_node; 1701 assert(n->is_AddP(), "Should be called for AddP nodes only"); 1702 // Search for objects which are not scalar replaceable. 1703 // Mark their escape state as ArgEscape to propagate the state 1704 // to referenced objects. 1705 // Note: currently there are no difference in compiler optimizations 1706 // for ArgEscape objects and NoEscape objects which are not 1707 // scalar replaceable. 1708 1709 Compile* C = _compile; 1710 1711 int offset = ptn->offset(); 1712 Node* base = get_addp_base(n); 1713 ptset.Clear(); 1714 PointsTo(ptset, base, phase); 1715 int ptset_size = ptset.Size(); 1716 1717 // Check if a oop field's initializing value is recorded and add 1718 // a corresponding NULL field's value if it is not recorded. 1719 // Connection Graph does not record a default initialization by NULL 1720 // captured by Initialize node. 1721 // 1722 // Note: it will disable scalar replacement in some cases: 1723 // 1724 // Point p[] = new Point[1]; 1725 // p[0] = new Point(); // Will be not scalar replaced 1726 // 1727 // but it will save us from incorrect optimizations in next cases: 1728 // 1729 // Point p[] = new Point[1]; 1730 // if ( x ) p[0] = new Point(); // Will be not scalar replaced 1731 // 1732 // Do a simple control flow analysis to distinguish above cases. 1733 // 1734 if (offset != Type::OffsetBot && ptset_size == 1) { 1735 uint elem = ptset.getelem(); // Allocation node's index 1736 // It does not matter if it is not Allocation node since 1737 // only non-escaping allocations are scalar replaced. 1738 if (ptnode_adr(elem)->_node->is_Allocate() && 1739 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) { 1740 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate(); 1741 InitializeNode* ini = alloc->initialization(); 1742 1743 // Check only oop fields. 1744 const Type* adr_type = n->as_AddP()->bottom_type(); 1745 BasicType basic_field_type = T_INT; 1746 if (adr_type->isa_instptr()) { 1747 ciField* field = C->alias_type(adr_type->isa_instptr())->field(); 1748 if (field != NULL) { 1749 basic_field_type = field->layout_type(); 1750 } else { 1751 // Ignore non field load (for example, klass load) 1752 } 1753 } else if (adr_type->isa_aryptr()) { 1754 const Type* elemtype = adr_type->isa_aryptr()->elem(); 1755 basic_field_type = elemtype->array_element_basic_type(); 1756 } else { 1757 // Raw pointers are used for initializing stores so skip it. 1758 assert(adr_type->isa_rawptr() && base->is_Proj() && 1759 (base->in(0) == alloc),"unexpected pointer type"); 1760 } 1761 if (basic_field_type == T_OBJECT || 1762 basic_field_type == T_NARROWOOP || 1763 basic_field_type == T_ARRAY) { 1764 Node* value = NULL; 1765 if (ini != NULL) { 1766 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT; 1767 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase); 1768 if (store != NULL && store->is_Store()) { 1769 value = store->in(MemNode::ValueIn); 1770 } else if (ptn->edge_count() > 0) { // Are there oop stores? 1771 // Check for a store which follows allocation without branches. 1772 // For example, a volatile field store is not collected 1773 // by Initialize node. TODO: it would be nice to use idom() here. 1774 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1775 store = n->fast_out(i); 1776 if (store->is_Store() && store->in(0) != NULL) { 1777 Node* ctrl = store->in(0); 1778 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL || 1779 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() || 1780 ctrl->is_IfTrue() || ctrl->is_IfFalse())) { 1781 ctrl = ctrl->in(0); 1782 } 1783 if (ctrl == ini || ctrl == alloc) { 1784 value = store->in(MemNode::ValueIn); 1785 break; 1786 } 1787 } 1788 } 1789 } 1790 } 1791 if (value == NULL || value != ptnode_adr(value->_idx)->_node) { 1792 // A field's initializing value was not recorded. Add NULL. 1793 uint null_idx = UseCompressedOops ? _noop_null : _oop_null; 1794 add_pointsto_edge(nidx, null_idx); 1795 } 1796 } 1797 } 1798 } 1799 1800 // An object is not scalar replaceable if the field which may point 1801 // to it has unknown offset (unknown element of an array of objects). 1802 // 1803 if (offset == Type::OffsetBot) { 1804 uint e_cnt = ptn->edge_count(); 1805 for (uint ei = 0; ei < e_cnt; ei++) { 1806 uint npi = ptn->edge_target(ei); 1807 set_escape_state(npi, PointsToNode::ArgEscape); 1808 ptnode_adr(npi)->_scalar_replaceable = false; 1809 } 1810 } 1811 1812 // Currently an object is not scalar replaceable if a LoadStore node 1813 // access its field since the field value is unknown after it. 1814 // 1815 bool has_LoadStore = false; 1816 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1817 Node *use = n->fast_out(i); 1818 if (use->is_LoadStore()) { 1819 has_LoadStore = true; 1820 break; 1821 } 1822 } 1823 // An object is not scalar replaceable if the address points 1824 // to unknown field (unknown element for arrays, offset is OffsetBot). 1825 // 1826 // Or the address may point to more then one object. This may produce 1827 // the false positive result (set scalar_replaceable to false) 1828 // since the flow-insensitive escape analysis can't separate 1829 // the case when stores overwrite the field's value from the case 1830 // when stores happened on different control branches. 1831 // 1832 if (ptset_size > 1 || ptset_size != 0 && 1833 (has_LoadStore || offset == Type::OffsetBot)) { 1834 for( VectorSetI j(&ptset); j.test(); ++j ) { 1835 set_escape_state(j.elem, PointsToNode::ArgEscape); 1836 ptnode_adr(j.elem)->_scalar_replaceable = false; 1837 } 1838 } 1839} 1840 1841void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) { 1842 1843 switch (call->Opcode()) { 1844#ifdef ASSERT 1845 case Op_Allocate: 1846 case Op_AllocateArray: 1847 case Op_Lock: 1848 case Op_Unlock: 1849 assert(false, "should be done already"); 1850 break; 1851#endif 1852 case Op_CallLeaf: 1853 case Op_CallLeafNoFP: 1854 { 1855 // Stub calls, objects do not escape but they are not scale replaceable. 1856 // Adjust escape state for outgoing arguments. 1857 const TypeTuple * d = call->tf()->domain(); 1858 VectorSet ptset(Thread::current()->resource_area()); 1859 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1860 const Type* at = d->field_at(i); 1861 Node *arg = call->in(i)->uncast(); 1862 const Type *aat = phase->type(arg); 1863 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() && 1864 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) { 1865 1866 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR || 1867 aat->isa_ptr() != NULL, "expecting an Ptr"); 1868#ifdef ASSERT 1869 if (!(call->Opcode() == Op_CallLeafNoFP && 1870 call->as_CallLeaf()->_name != NULL && 1871 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) || 1872 call->as_CallLeaf()->_name != NULL && 1873 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 || 1874 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 )) 1875 ) { 1876 call->dump(); 1877 assert(false, "EA: unexpected CallLeaf"); 1878 } 1879#endif 1880 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1881 if (arg->is_AddP()) { 1882 // 1883 // The inline_native_clone() case when the arraycopy stub is called 1884 // after the allocation before Initialize and CheckCastPP nodes. 1885 // 1886 // Set AddP's base (Allocate) as not scalar replaceable since 1887 // pointer to the base (with offset) is passed as argument. 1888 // 1889 arg = get_addp_base(arg); 1890 } 1891 ptset.Clear(); 1892 PointsTo(ptset, arg, phase); 1893 for( VectorSetI j(&ptset); j.test(); ++j ) { 1894 uint pt = j.elem; 1895 set_escape_state(pt, PointsToNode::ArgEscape); 1896 } 1897 } 1898 } 1899 break; 1900 } 1901 1902 case Op_CallStaticJava: 1903 // For a static call, we know exactly what method is being called. 1904 // Use bytecode estimator to record the call's escape affects 1905 { 1906 ciMethod *meth = call->as_CallJava()->method(); 1907 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL; 1908 // fall-through if not a Java method or no analyzer information 1909 if (call_analyzer != NULL) { 1910 const TypeTuple * d = call->tf()->domain(); 1911 VectorSet ptset(Thread::current()->resource_area()); 1912 bool copy_dependencies = false; 1913 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1914 const Type* at = d->field_at(i); 1915 int k = i - TypeFunc::Parms; 1916 Node *arg = call->in(i)->uncast(); 1917 1918 if (at->isa_oopptr() != NULL && 1919 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) { 1920 1921 bool global_escapes = false; 1922 bool fields_escapes = false; 1923 if (!call_analyzer->is_arg_stack(k)) { 1924 // The argument global escapes, mark everything it could point to 1925 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1926 global_escapes = true; 1927 } else { 1928 if (!call_analyzer->is_arg_local(k)) { 1929 // The argument itself doesn't escape, but any fields might 1930 fields_escapes = true; 1931 } 1932 set_escape_state(arg->_idx, PointsToNode::ArgEscape); 1933 copy_dependencies = true; 1934 } 1935 1936 ptset.Clear(); 1937 PointsTo(ptset, arg, phase); 1938 for( VectorSetI j(&ptset); j.test(); ++j ) { 1939 uint pt = j.elem; 1940 if (global_escapes) { 1941 //The argument global escapes, mark everything it could point to 1942 set_escape_state(pt, PointsToNode::GlobalEscape); 1943 } else { 1944 if (fields_escapes) { 1945 // The argument itself doesn't escape, but any fields might 1946 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot); 1947 } 1948 set_escape_state(pt, PointsToNode::ArgEscape); 1949 } 1950 } 1951 } 1952 } 1953 if (copy_dependencies) 1954 call_analyzer->copy_dependencies(_compile->dependencies()); 1955 break; 1956 } 1957 } 1958 1959 default: 1960 // Fall-through here if not a Java method or no analyzer information 1961 // or some other type of call, assume the worst case: all arguments 1962 // globally escape. 1963 { 1964 // adjust escape state for outgoing arguments 1965 const TypeTuple * d = call->tf()->domain(); 1966 VectorSet ptset(Thread::current()->resource_area()); 1967 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1968 const Type* at = d->field_at(i); 1969 if (at->isa_oopptr() != NULL) { 1970 Node *arg = call->in(i)->uncast(); 1971 set_escape_state(arg->_idx, PointsToNode::GlobalEscape); 1972 ptset.Clear(); 1973 PointsTo(ptset, arg, phase); 1974 for( VectorSetI j(&ptset); j.test(); ++j ) { 1975 uint pt = j.elem; 1976 set_escape_state(pt, PointsToNode::GlobalEscape); 1977 } 1978 } 1979 } 1980 } 1981 } 1982} 1983void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) { 1984 CallNode *call = resproj->in(0)->as_Call(); 1985 uint call_idx = call->_idx; 1986 uint resproj_idx = resproj->_idx; 1987 1988 switch (call->Opcode()) { 1989 case Op_Allocate: 1990 { 1991 Node *k = call->in(AllocateNode::KlassNode); 1992 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 1993 assert(kt != NULL, "TypeKlassPtr required."); 1994 ciKlass* cik = kt->klass(); 1995 1996 PointsToNode::EscapeState es; 1997 uint edge_to; 1998 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || 1999 !cik->is_instance_klass() || // StressReflectiveCode 2000 cik->as_instance_klass()->has_finalizer()) { 2001 es = PointsToNode::GlobalEscape; 2002 edge_to = _phantom_object; // Could not be worse 2003 } else { 2004 es = PointsToNode::NoEscape; 2005 edge_to = call_idx; 2006 } 2007 set_escape_state(call_idx, es); 2008 add_pointsto_edge(resproj_idx, edge_to); 2009 _processed.set(resproj_idx); 2010 break; 2011 } 2012 2013 case Op_AllocateArray: 2014 { 2015 2016 Node *k = call->in(AllocateNode::KlassNode); 2017 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr(); 2018 assert(kt != NULL, "TypeKlassPtr required."); 2019 ciKlass* cik = kt->klass(); 2020 2021 PointsToNode::EscapeState es; 2022 uint edge_to; 2023 if (!cik->is_array_klass()) { // StressReflectiveCode 2024 es = PointsToNode::GlobalEscape; 2025 edge_to = _phantom_object; 2026 } else { 2027 es = PointsToNode::NoEscape; 2028 edge_to = call_idx; 2029 int length = call->in(AllocateNode::ALength)->find_int_con(-1); 2030 if (length < 0 || length > EliminateAllocationArraySizeLimit) { 2031 // Not scalar replaceable if the length is not constant or too big. 2032 ptnode_adr(call_idx)->_scalar_replaceable = false; 2033 } 2034 } 2035 set_escape_state(call_idx, es); 2036 add_pointsto_edge(resproj_idx, edge_to); 2037 _processed.set(resproj_idx); 2038 break; 2039 } 2040 2041 case Op_CallStaticJava: 2042 // For a static call, we know exactly what method is being called. 2043 // Use bytecode estimator to record whether the call's return value escapes 2044 { 2045 bool done = true; 2046 const TypeTuple *r = call->tf()->range(); 2047 const Type* ret_type = NULL; 2048 2049 if (r->cnt() > TypeFunc::Parms) 2050 ret_type = r->field_at(TypeFunc::Parms); 2051 2052 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2053 // _multianewarray functions return a TypeRawPtr. 2054 if (ret_type == NULL || ret_type->isa_ptr() == NULL) { 2055 _processed.set(resproj_idx); 2056 break; // doesn't return a pointer type 2057 } 2058 ciMethod *meth = call->as_CallJava()->method(); 2059 const TypeTuple * d = call->tf()->domain(); 2060 if (meth == NULL) { 2061 // not a Java method, assume global escape 2062 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2063 add_pointsto_edge(resproj_idx, _phantom_object); 2064 } else { 2065 BCEscapeAnalyzer *call_analyzer = meth->get_bcea(); 2066 bool copy_dependencies = false; 2067 2068 if (call_analyzer->is_return_allocated()) { 2069 // Returns a newly allocated unescaped object, simply 2070 // update dependency information. 2071 // Mark it as NoEscape so that objects referenced by 2072 // it's fields will be marked as NoEscape at least. 2073 set_escape_state(call_idx, PointsToNode::NoEscape); 2074 add_pointsto_edge(resproj_idx, call_idx); 2075 copy_dependencies = true; 2076 } else if (call_analyzer->is_return_local()) { 2077 // determine whether any arguments are returned 2078 set_escape_state(call_idx, PointsToNode::NoEscape); 2079 bool ret_arg = false; 2080 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2081 const Type* at = d->field_at(i); 2082 2083 if (at->isa_oopptr() != NULL) { 2084 Node *arg = call->in(i)->uncast(); 2085 2086 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) { 2087 ret_arg = true; 2088 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2089 if (arg_esp->node_type() == PointsToNode::UnknownType) 2090 done = false; 2091 else if (arg_esp->node_type() == PointsToNode::JavaObject) 2092 add_pointsto_edge(resproj_idx, arg->_idx); 2093 else 2094 add_deferred_edge(resproj_idx, arg->_idx); 2095 arg_esp->_hidden_alias = true; 2096 } 2097 } 2098 } 2099 if (done && !ret_arg) { 2100 // Returns unknown object. 2101 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2102 add_pointsto_edge(resproj_idx, _phantom_object); 2103 } 2104 copy_dependencies = true; 2105 } else { 2106 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2107 add_pointsto_edge(resproj_idx, _phantom_object); 2108 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 2109 const Type* at = d->field_at(i); 2110 if (at->isa_oopptr() != NULL) { 2111 Node *arg = call->in(i)->uncast(); 2112 PointsToNode *arg_esp = ptnode_adr(arg->_idx); 2113 arg_esp->_hidden_alias = true; 2114 } 2115 } 2116 } 2117 if (copy_dependencies) 2118 call_analyzer->copy_dependencies(_compile->dependencies()); 2119 } 2120 if (done) 2121 _processed.set(resproj_idx); 2122 break; 2123 } 2124 2125 default: 2126 // Some other type of call, assume the worst case that the 2127 // returned value, if any, globally escapes. 2128 { 2129 const TypeTuple *r = call->tf()->range(); 2130 if (r->cnt() > TypeFunc::Parms) { 2131 const Type* ret_type = r->field_at(TypeFunc::Parms); 2132 2133 // Note: we use isa_ptr() instead of isa_oopptr() here because the 2134 // _multianewarray functions return a TypeRawPtr. 2135 if (ret_type->isa_ptr() != NULL) { 2136 set_escape_state(call_idx, PointsToNode::GlobalEscape); 2137 add_pointsto_edge(resproj_idx, _phantom_object); 2138 } 2139 } 2140 _processed.set(resproj_idx); 2141 } 2142 } 2143} 2144 2145// Populate Connection Graph with Ideal nodes and create simple 2146// connection graph edges (do not need to check the node_type of inputs 2147// or to call PointsTo() to walk the connection graph). 2148void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) { 2149 if (_processed.test(n->_idx)) 2150 return; // No need to redefine node's state. 2151 2152 if (n->is_Call()) { 2153 // Arguments to allocation and locking don't escape. 2154 if (n->is_Allocate()) { 2155 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true); 2156 record_for_optimizer(n); 2157 } else if (n->is_Lock() || n->is_Unlock()) { 2158 // Put Lock and Unlock nodes on IGVN worklist to process them during 2159 // the first IGVN optimization when escape information is still available. 2160 record_for_optimizer(n); 2161 _processed.set(n->_idx); 2162 } else { 2163 // Don't mark as processed since call's arguments have to be processed. 2164 PointsToNode::NodeType nt = PointsToNode::UnknownType; 2165 PointsToNode::EscapeState es = PointsToNode::UnknownEscape; 2166 2167 // Check if a call returns an object. 2168 const TypeTuple *r = n->as_Call()->tf()->range(); 2169 if (r->cnt() > TypeFunc::Parms && 2170 r->field_at(TypeFunc::Parms)->isa_ptr() && 2171 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) { 2172 nt = PointsToNode::JavaObject; 2173 if (!n->is_CallStaticJava()) { 2174 // Since the called mathod is statically unknown assume 2175 // the worst case that the returned value globally escapes. 2176 es = PointsToNode::GlobalEscape; 2177 } 2178 } 2179 add_node(n, nt, es, false); 2180 } 2181 return; 2182 } 2183 2184 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because 2185 // ThreadLocal has RawPrt type. 2186 switch (n->Opcode()) { 2187 case Op_AddP: 2188 { 2189 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false); 2190 break; 2191 } 2192 case Op_CastX2P: 2193 { // "Unsafe" memory access. 2194 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2195 break; 2196 } 2197 case Op_CastPP: 2198 case Op_CheckCastPP: 2199 case Op_EncodeP: 2200 case Op_DecodeN: 2201 { 2202 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2203 int ti = n->in(1)->_idx; 2204 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2205 if (nt == PointsToNode::UnknownType) { 2206 _delayed_worklist.push(n); // Process it later. 2207 break; 2208 } else if (nt == PointsToNode::JavaObject) { 2209 add_pointsto_edge(n->_idx, ti); 2210 } else { 2211 add_deferred_edge(n->_idx, ti); 2212 } 2213 _processed.set(n->_idx); 2214 break; 2215 } 2216 case Op_ConP: 2217 { 2218 // assume all pointer constants globally escape except for null 2219 PointsToNode::EscapeState es; 2220 if (phase->type(n) == TypePtr::NULL_PTR) 2221 es = PointsToNode::NoEscape; 2222 else 2223 es = PointsToNode::GlobalEscape; 2224 2225 add_node(n, PointsToNode::JavaObject, es, true); 2226 break; 2227 } 2228 case Op_ConN: 2229 { 2230 // assume all narrow oop constants globally escape except for null 2231 PointsToNode::EscapeState es; 2232 if (phase->type(n) == TypeNarrowOop::NULL_PTR) 2233 es = PointsToNode::NoEscape; 2234 else 2235 es = PointsToNode::GlobalEscape; 2236 2237 add_node(n, PointsToNode::JavaObject, es, true); 2238 break; 2239 } 2240 case Op_CreateEx: 2241 { 2242 // assume that all exception objects globally escape 2243 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2244 break; 2245 } 2246 case Op_LoadKlass: 2247 case Op_LoadNKlass: 2248 { 2249 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true); 2250 break; 2251 } 2252 case Op_LoadP: 2253 case Op_LoadN: 2254 { 2255 const Type *t = phase->type(n); 2256 if (t->make_ptr() == NULL) { 2257 _processed.set(n->_idx); 2258 return; 2259 } 2260 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2261 break; 2262 } 2263 case Op_Parm: 2264 { 2265 _processed.set(n->_idx); // No need to redefine it state. 2266 uint con = n->as_Proj()->_con; 2267 if (con < TypeFunc::Parms) 2268 return; 2269 const Type *t = n->in(0)->as_Start()->_domain->field_at(con); 2270 if (t->isa_ptr() == NULL) 2271 return; 2272 // We have to assume all input parameters globally escape 2273 // (Note: passing 'false' since _processed is already set). 2274 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false); 2275 break; 2276 } 2277 case Op_Phi: 2278 { 2279 const Type *t = n->as_Phi()->type(); 2280 if (t->make_ptr() == NULL) { 2281 // nothing to do if not an oop or narrow oop 2282 _processed.set(n->_idx); 2283 return; 2284 } 2285 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2286 uint i; 2287 for (i = 1; i < n->req() ; i++) { 2288 Node* in = n->in(i); 2289 if (in == NULL) 2290 continue; // ignore NULL 2291 in = in->uncast(); 2292 if (in->is_top() || in == n) 2293 continue; // ignore top or inputs which go back this node 2294 int ti = in->_idx; 2295 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2296 if (nt == PointsToNode::UnknownType) { 2297 break; 2298 } else if (nt == PointsToNode::JavaObject) { 2299 add_pointsto_edge(n->_idx, ti); 2300 } else { 2301 add_deferred_edge(n->_idx, ti); 2302 } 2303 } 2304 if (i >= n->req()) 2305 _processed.set(n->_idx); 2306 else 2307 _delayed_worklist.push(n); 2308 break; 2309 } 2310 case Op_Proj: 2311 { 2312 // we are only interested in the oop result projection from a call 2313 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2314 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2315 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2316 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2317 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false); 2318 int ti = n->in(0)->_idx; 2319 // The call may not be registered yet (since not all its inputs are registered) 2320 // if this is the projection from backbranch edge of Phi. 2321 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) { 2322 process_call_result(n->as_Proj(), phase); 2323 } 2324 if (!_processed.test(n->_idx)) { 2325 // The call's result may need to be processed later if the call 2326 // returns it's argument and the argument is not processed yet. 2327 _delayed_worklist.push(n); 2328 } 2329 break; 2330 } 2331 } 2332 _processed.set(n->_idx); 2333 break; 2334 } 2335 case Op_Return: 2336 { 2337 if( n->req() > TypeFunc::Parms && 2338 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2339 // Treat Return value as LocalVar with GlobalEscape escape state. 2340 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false); 2341 int ti = n->in(TypeFunc::Parms)->_idx; 2342 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2343 if (nt == PointsToNode::UnknownType) { 2344 _delayed_worklist.push(n); // Process it later. 2345 break; 2346 } else if (nt == PointsToNode::JavaObject) { 2347 add_pointsto_edge(n->_idx, ti); 2348 } else { 2349 add_deferred_edge(n->_idx, ti); 2350 } 2351 } 2352 _processed.set(n->_idx); 2353 break; 2354 } 2355 case Op_StoreP: 2356 case Op_StoreN: 2357 { 2358 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2359 adr_type = adr_type->make_ptr(); 2360 if (adr_type->isa_oopptr()) { 2361 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2362 } else { 2363 Node* adr = n->in(MemNode::Address); 2364 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL && 2365 adr->in(AddPNode::Address)->is_Proj() && 2366 adr->in(AddPNode::Address)->in(0)->is_Allocate()) { 2367 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2368 // We are computing a raw address for a store captured 2369 // by an Initialize compute an appropriate address type. 2370 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 2371 assert(offs != Type::OffsetBot, "offset must be a constant"); 2372 } else { 2373 _processed.set(n->_idx); 2374 return; 2375 } 2376 } 2377 break; 2378 } 2379 case Op_StorePConditional: 2380 case Op_CompareAndSwapP: 2381 case Op_CompareAndSwapN: 2382 { 2383 const Type *adr_type = phase->type(n->in(MemNode::Address)); 2384 adr_type = adr_type->make_ptr(); 2385 if (adr_type->isa_oopptr()) { 2386 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2387 } else { 2388 _processed.set(n->_idx); 2389 return; 2390 } 2391 break; 2392 } 2393 case Op_AryEq: 2394 case Op_StrComp: 2395 case Op_StrEquals: 2396 case Op_StrIndexOf: 2397 { 2398 // char[] arrays passed to string intrinsics are not scalar replaceable. 2399 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false); 2400 break; 2401 } 2402 case Op_ThreadLocal: 2403 { 2404 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true); 2405 break; 2406 } 2407 default: 2408 ; 2409 // nothing to do 2410 } 2411 return; 2412} 2413 2414void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) { 2415 uint n_idx = n->_idx; 2416 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered"); 2417 2418 // Don't set processed bit for AddP, LoadP, StoreP since 2419 // they may need more then one pass to process. 2420 if (_processed.test(n_idx)) 2421 return; // No need to redefine node's state. 2422 2423 if (n->is_Call()) { 2424 CallNode *call = n->as_Call(); 2425 process_call_arguments(call, phase); 2426 _processed.set(n_idx); 2427 return; 2428 } 2429 2430 switch (n->Opcode()) { 2431 case Op_AddP: 2432 { 2433 Node *base = get_addp_base(n); 2434 // Create a field edge to this node from everything base could point to. 2435 VectorSet ptset(Thread::current()->resource_area()); 2436 PointsTo(ptset, base, phase); 2437 for( VectorSetI i(&ptset); i.test(); ++i ) { 2438 uint pt = i.elem; 2439 add_field_edge(pt, n_idx, address_offset(n, phase)); 2440 } 2441 break; 2442 } 2443 case Op_CastX2P: 2444 { 2445 assert(false, "Op_CastX2P"); 2446 break; 2447 } 2448 case Op_CastPP: 2449 case Op_CheckCastPP: 2450 case Op_EncodeP: 2451 case Op_DecodeN: 2452 { 2453 int ti = n->in(1)->_idx; 2454 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered"); 2455 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2456 add_pointsto_edge(n_idx, ti); 2457 } else { 2458 add_deferred_edge(n_idx, ti); 2459 } 2460 _processed.set(n_idx); 2461 break; 2462 } 2463 case Op_ConP: 2464 { 2465 assert(false, "Op_ConP"); 2466 break; 2467 } 2468 case Op_ConN: 2469 { 2470 assert(false, "Op_ConN"); 2471 break; 2472 } 2473 case Op_CreateEx: 2474 { 2475 assert(false, "Op_CreateEx"); 2476 break; 2477 } 2478 case Op_LoadKlass: 2479 case Op_LoadNKlass: 2480 { 2481 assert(false, "Op_LoadKlass"); 2482 break; 2483 } 2484 case Op_LoadP: 2485 case Op_LoadN: 2486 { 2487 const Type *t = phase->type(n); 2488#ifdef ASSERT 2489 if (t->make_ptr() == NULL) 2490 assert(false, "Op_LoadP"); 2491#endif 2492 2493 Node* adr = n->in(MemNode::Address)->uncast(); 2494 Node* adr_base; 2495 if (adr->is_AddP()) { 2496 adr_base = get_addp_base(adr); 2497 } else { 2498 adr_base = adr; 2499 } 2500 2501 // For everything "adr_base" could point to, create a deferred edge from 2502 // this node to each field with the same offset. 2503 VectorSet ptset(Thread::current()->resource_area()); 2504 PointsTo(ptset, adr_base, phase); 2505 int offset = address_offset(adr, phase); 2506 for( VectorSetI i(&ptset); i.test(); ++i ) { 2507 uint pt = i.elem; 2508 add_deferred_edge_to_fields(n_idx, pt, offset); 2509 } 2510 break; 2511 } 2512 case Op_Parm: 2513 { 2514 assert(false, "Op_Parm"); 2515 break; 2516 } 2517 case Op_Phi: 2518 { 2519#ifdef ASSERT 2520 const Type *t = n->as_Phi()->type(); 2521 if (t->make_ptr() == NULL) 2522 assert(false, "Op_Phi"); 2523#endif 2524 for (uint i = 1; i < n->req() ; i++) { 2525 Node* in = n->in(i); 2526 if (in == NULL) 2527 continue; // ignore NULL 2528 in = in->uncast(); 2529 if (in->is_top() || in == n) 2530 continue; // ignore top or inputs which go back this node 2531 int ti = in->_idx; 2532 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type(); 2533 assert(nt != PointsToNode::UnknownType, "all nodes should be known"); 2534 if (nt == PointsToNode::JavaObject) { 2535 add_pointsto_edge(n_idx, ti); 2536 } else { 2537 add_deferred_edge(n_idx, ti); 2538 } 2539 } 2540 _processed.set(n_idx); 2541 break; 2542 } 2543 case Op_Proj: 2544 { 2545 // we are only interested in the oop result projection from a call 2546 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) { 2547 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType, 2548 "all nodes should be registered"); 2549 const TypeTuple *r = n->in(0)->as_Call()->tf()->range(); 2550 assert(r->cnt() > TypeFunc::Parms, "sanity"); 2551 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) { 2552 process_call_result(n->as_Proj(), phase); 2553 assert(_processed.test(n_idx), "all call results should be processed"); 2554 break; 2555 } 2556 } 2557 assert(false, "Op_Proj"); 2558 break; 2559 } 2560 case Op_Return: 2561 { 2562#ifdef ASSERT 2563 if( n->req() <= TypeFunc::Parms || 2564 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) { 2565 assert(false, "Op_Return"); 2566 } 2567#endif 2568 int ti = n->in(TypeFunc::Parms)->_idx; 2569 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered"); 2570 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) { 2571 add_pointsto_edge(n_idx, ti); 2572 } else { 2573 add_deferred_edge(n_idx, ti); 2574 } 2575 _processed.set(n_idx); 2576 break; 2577 } 2578 case Op_StoreP: 2579 case Op_StoreN: 2580 case Op_StorePConditional: 2581 case Op_CompareAndSwapP: 2582 case Op_CompareAndSwapN: 2583 { 2584 Node *adr = n->in(MemNode::Address); 2585 const Type *adr_type = phase->type(adr)->make_ptr(); 2586#ifdef ASSERT 2587 if (!adr_type->isa_oopptr()) 2588 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP"); 2589#endif 2590 2591 assert(adr->is_AddP(), "expecting an AddP"); 2592 Node *adr_base = get_addp_base(adr); 2593 Node *val = n->in(MemNode::ValueIn)->uncast(); 2594 // For everything "adr_base" could point to, create a deferred edge 2595 // to "val" from each field with the same offset. 2596 VectorSet ptset(Thread::current()->resource_area()); 2597 PointsTo(ptset, adr_base, phase); 2598 for( VectorSetI i(&ptset); i.test(); ++i ) { 2599 uint pt = i.elem; 2600 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase)); 2601 } 2602 break; 2603 } 2604 case Op_AryEq: 2605 case Op_StrComp: 2606 case Op_StrEquals: 2607 case Op_StrIndexOf: 2608 { 2609 // char[] arrays passed to string intrinsic do not escape but 2610 // they are not scalar replaceable. Adjust escape state for them. 2611 // Start from in(2) edge since in(1) is memory edge. 2612 for (uint i = 2; i < n->req(); i++) { 2613 Node* adr = n->in(i)->uncast(); 2614 const Type *at = phase->type(adr); 2615 if (!adr->is_top() && at->isa_ptr()) { 2616 assert(at == Type::TOP || at == TypePtr::NULL_PTR || 2617 at->isa_ptr() != NULL, "expecting an Ptr"); 2618 if (adr->is_AddP()) { 2619 adr = get_addp_base(adr); 2620 } 2621 // Mark as ArgEscape everything "adr" could point to. 2622 set_escape_state(adr->_idx, PointsToNode::ArgEscape); 2623 } 2624 } 2625 _processed.set(n_idx); 2626 break; 2627 } 2628 case Op_ThreadLocal: 2629 { 2630 assert(false, "Op_ThreadLocal"); 2631 break; 2632 } 2633 default: 2634 // This method should be called only for EA specific nodes. 2635 ShouldNotReachHere(); 2636 } 2637} 2638 2639#ifndef PRODUCT 2640void ConnectionGraph::dump() { 2641 PhaseGVN *igvn = _compile->initial_gvn(); 2642 bool first = true; 2643 2644 uint size = nodes_size(); 2645 for (uint ni = 0; ni < size; ni++) { 2646 PointsToNode *ptn = ptnode_adr(ni); 2647 PointsToNode::NodeType ptn_type = ptn->node_type(); 2648 2649 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL) 2650 continue; 2651 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn); 2652 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) { 2653 if (first) { 2654 tty->cr(); 2655 tty->print("======== Connection graph for "); 2656 _compile->method()->print_short_name(); 2657 tty->cr(); 2658 first = false; 2659 } 2660 tty->print("%6d ", ni); 2661 ptn->dump(); 2662 // Print all locals which reference this allocation 2663 for (uint li = ni; li < size; li++) { 2664 PointsToNode *ptn_loc = ptnode_adr(li); 2665 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type(); 2666 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL && 2667 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) { 2668 ptnode_adr(li)->dump(false); 2669 } 2670 } 2671 if (Verbose) { 2672 // Print all fields which reference this allocation 2673 for (uint i = 0; i < ptn->edge_count(); i++) { 2674 uint ei = ptn->edge_target(i); 2675 ptnode_adr(ei)->dump(false); 2676 } 2677 } 2678 tty->cr(); 2679 } 2680 } 2681} 2682#endif 2683