1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/block.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/regalloc.hpp" 40 #include "opto/rootnode.hpp" 41 #include "utilities/macros.hpp" 42 43 //============================================================================= 44 #define NODE_HASH_MINIMUM_SIZE 255 45 //------------------------------NodeHash--------------------------------------- 46 NodeHash::NodeHash(uint est_max_size) : 47 _a(Thread::current()->resource_area()), 48 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ), 49 _inserts(0), _insert_limit( insert_limit() ), 50 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) // (Node**)_a->Amalloc(_max * sizeof(Node*)) ), 51 #ifndef PRODUCT 52 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0), 53 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0), 54 _total_inserts(0), _total_insert_probes(0) 55 #endif 56 { 57 // _sentinel must be in the current node space 58 _sentinel = new ProjNode(NULL, TypeFunc::Control); 59 memset(_table,0,sizeof(Node*)*_max); 60 } 61 62 //------------------------------NodeHash--------------------------------------- 63 NodeHash::NodeHash(Arena *arena, uint est_max_size) : 64 _a(arena), 65 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ), 66 _inserts(0), _insert_limit( insert_limit() ), 67 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) 68 #ifndef PRODUCT 69 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0), 70 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0), 71 _total_inserts(0), _total_insert_probes(0) 72 #endif 73 { 74 // _sentinel must be in the current node space 75 _sentinel = new ProjNode(NULL, TypeFunc::Control); 76 memset(_table,0,sizeof(Node*)*_max); 77 } 78 79 //------------------------------NodeHash--------------------------------------- 80 NodeHash::NodeHash(NodeHash *nh) { 81 debug_only(_table = (Node**)badAddress); // interact correctly w/ operator= 82 // just copy in all the fields 83 *this = *nh; 84 // nh->_sentinel must be in the current node space 85 } 86 87 void NodeHash::replace_with(NodeHash *nh) { 88 debug_only(_table = (Node**)badAddress); // interact correctly w/ operator= 89 // just copy in all the fields 90 *this = *nh; 91 // nh->_sentinel must be in the current node space 92 } 93 94 //------------------------------hash_find-------------------------------------- 95 // Find in hash table 96 Node *NodeHash::hash_find( const Node *n ) { 97 // ((Node*)n)->set_hash( n->hash() ); 98 uint hash = n->hash(); 99 if (hash == Node::NO_HASH) { 100 NOT_PRODUCT( _lookup_misses++ ); 101 return NULL; 102 } 103 uint key = hash & (_max-1); 104 uint stride = key | 0x01; 105 NOT_PRODUCT( _look_probes++ ); 106 Node *k = _table[key]; // Get hashed value 107 if( !k ) { // ?Miss? 108 NOT_PRODUCT( _lookup_misses++ ); 109 return NULL; // Miss! 110 } 111 112 int op = n->Opcode(); 113 uint req = n->req(); 114 while( 1 ) { // While probing hash table 115 if( k->req() == req && // Same count of inputs 116 k->Opcode() == op ) { // Same Opcode 117 for( uint i=0; i<req; i++ ) 118 if( n->in(i)!=k->in(i)) // Different inputs? 119 goto collision; // "goto" is a speed hack... 120 if( n->cmp(*k) ) { // Check for any special bits 121 NOT_PRODUCT( _lookup_hits++ ); 122 return k; // Hit! 123 } 124 } 125 collision: 126 NOT_PRODUCT( _look_probes++ ); 127 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime 128 k = _table[key]; // Get hashed value 129 if( !k ) { // ?Miss? 130 NOT_PRODUCT( _lookup_misses++ ); 131 return NULL; // Miss! 132 } 133 } 134 ShouldNotReachHere(); 135 return NULL; 136 } 137 138 //------------------------------hash_find_insert------------------------------- 139 // Find in hash table, insert if not already present 140 // Used to preserve unique entries in hash table 141 Node *NodeHash::hash_find_insert( Node *n ) { 142 // n->set_hash( ); 143 uint hash = n->hash(); 144 if (hash == Node::NO_HASH) { 145 NOT_PRODUCT( _lookup_misses++ ); 146 return NULL; 147 } 148 uint key = hash & (_max-1); 149 uint stride = key | 0x01; // stride must be relatively prime to table siz 150 uint first_sentinel = 0; // replace a sentinel if seen. 151 NOT_PRODUCT( _look_probes++ ); 152 Node *k = _table[key]; // Get hashed value 153 if( !k ) { // ?Miss? 154 NOT_PRODUCT( _lookup_misses++ ); 155 _table[key] = n; // Insert into table! 156 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 157 check_grow(); // Grow table if insert hit limit 158 return NULL; // Miss! 159 } 160 else if( k == _sentinel ) { 161 first_sentinel = key; // Can insert here 162 } 163 164 int op = n->Opcode(); 165 uint req = n->req(); 166 while( 1 ) { // While probing hash table 167 if( k->req() == req && // Same count of inputs 168 k->Opcode() == op ) { // Same Opcode 169 for( uint i=0; i<req; i++ ) 170 if( n->in(i)!=k->in(i)) // Different inputs? 171 goto collision; // "goto" is a speed hack... 172 if( n->cmp(*k) ) { // Check for any special bits 173 NOT_PRODUCT( _lookup_hits++ ); 174 return k; // Hit! 175 } 176 } 177 collision: 178 NOT_PRODUCT( _look_probes++ ); 179 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 180 k = _table[key]; // Get hashed value 181 if( !k ) { // ?Miss? 182 NOT_PRODUCT( _lookup_misses++ ); 183 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel? 184 _table[key] = n; // Insert into table! 185 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 186 check_grow(); // Grow table if insert hit limit 187 return NULL; // Miss! 188 } 189 else if( first_sentinel == 0 && k == _sentinel ) { 190 first_sentinel = key; // Can insert here 191 } 192 193 } 194 ShouldNotReachHere(); 195 return NULL; 196 } 197 198 //------------------------------hash_insert------------------------------------ 199 // Insert into hash table 200 void NodeHash::hash_insert( Node *n ) { 201 // // "conflict" comments -- print nodes that conflict 202 // bool conflict = false; 203 // n->set_hash(); 204 uint hash = n->hash(); 205 if (hash == Node::NO_HASH) { 206 return; 207 } 208 check_grow(); 209 uint key = hash & (_max-1); 210 uint stride = key | 0x01; 211 212 while( 1 ) { // While probing hash table 213 NOT_PRODUCT( _insert_probes++ ); 214 Node *k = _table[key]; // Get hashed value 215 if( !k || (k == _sentinel) ) break; // Found a slot 216 assert( k != n, "already inserted" ); 217 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; } 218 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 219 } 220 _table[key] = n; // Insert into table! 221 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 222 // if( conflict ) { n->dump(); } 223 } 224 225 //------------------------------hash_delete------------------------------------ 226 // Replace in hash table with sentinel 227 bool NodeHash::hash_delete( const Node *n ) { 228 Node *k; 229 uint hash = n->hash(); 230 if (hash == Node::NO_HASH) { 231 NOT_PRODUCT( _delete_misses++ ); 232 return false; 233 } 234 uint key = hash & (_max-1); 235 uint stride = key | 0x01; 236 debug_only( uint counter = 0; ); 237 for( ; /* (k != NULL) && (k != _sentinel) */; ) { 238 debug_only( counter++ ); 239 NOT_PRODUCT( _delete_probes++ ); 240 k = _table[key]; // Get hashed value 241 if( !k ) { // Miss? 242 NOT_PRODUCT( _delete_misses++ ); 243 #ifdef ASSERT 244 if( VerifyOpto ) { 245 for( uint i=0; i < _max; i++ ) 246 assert( _table[i] != n, "changed edges with rehashing" ); 247 } 248 #endif 249 return false; // Miss! Not in chain 250 } 251 else if( n == k ) { 252 NOT_PRODUCT( _delete_hits++ ); 253 _table[key] = _sentinel; // Hit! Label as deleted entry 254 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table. 255 return true; 256 } 257 else { 258 // collision: move through table with prime offset 259 key = (key + stride/*7*/) & (_max-1); 260 assert( counter <= _insert_limit, "Cycle in hash-table"); 261 } 262 } 263 ShouldNotReachHere(); 264 return false; 265 } 266 267 //------------------------------round_up--------------------------------------- 268 // Round up to nearest power of 2 269 uint NodeHash::round_up( uint x ) { 270 x += (x>>2); // Add 25% slop 271 if( x <16 ) return 16; // Small stuff 272 uint i=16; 273 while( i < x ) i <<= 1; // Double to fit 274 return i; // Return hash table size 275 } 276 277 //------------------------------grow------------------------------------------- 278 // Grow _table to next power of 2 and insert old entries 279 void NodeHash::grow() { 280 // Record old state 281 uint old_max = _max; 282 Node **old_table = _table; 283 // Construct new table with twice the space 284 #ifndef PRODUCT 285 _grows++; 286 _total_inserts += _inserts; 287 _total_insert_probes += _insert_probes; 288 _insert_probes = 0; 289 #endif 290 _inserts = 0; 291 _max = _max << 1; 292 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) ); 293 memset(_table,0,sizeof(Node*)*_max); 294 _insert_limit = insert_limit(); 295 // Insert old entries into the new table 296 for( uint i = 0; i < old_max; i++ ) { 297 Node *m = *old_table++; 298 if( !m || m == _sentinel ) continue; 299 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table. 300 hash_insert(m); 301 } 302 } 303 304 //------------------------------clear------------------------------------------ 305 // Clear all entries in _table to NULL but keep storage 306 void NodeHash::clear() { 307 #ifdef ASSERT 308 // Unlock all nodes upon removal from table. 309 for (uint i = 0; i < _max; i++) { 310 Node* n = _table[i]; 311 if (!n || n == _sentinel) continue; 312 n->exit_hash_lock(); 313 } 314 #endif 315 316 memset( _table, 0, _max * sizeof(Node*) ); 317 } 318 319 //-----------------------remove_useless_nodes---------------------------------- 320 // Remove useless nodes from value table, 321 // implementation does not depend on hash function 322 void NodeHash::remove_useless_nodes(VectorSet &useful) { 323 324 // Dead nodes in the hash table inherited from GVN should not replace 325 // existing nodes, remove dead nodes. 326 uint max = size(); 327 Node *sentinel_node = sentinel(); 328 for( uint i = 0; i < max; ++i ) { 329 Node *n = at(i); 330 if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) { 331 debug_only(n->exit_hash_lock()); // Unlock the node when removed 332 _table[i] = sentinel_node; // Replace with placeholder 333 } 334 } 335 } 336 337 338 void NodeHash::check_no_speculative_types() { 339 #ifdef ASSERT 340 uint max = size(); 341 Node *sentinel_node = sentinel(); 342 for (uint i = 0; i < max; ++i) { 343 Node *n = at(i); 344 if(n != NULL && n != sentinel_node && n->is_Type() && n->outcnt() > 0) { 345 TypeNode* tn = n->as_Type(); 346 const Type* t = tn->type(); 347 const Type* t_no_spec = t->remove_speculative(); 348 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup"); 349 } 350 } 351 #endif 352 } 353 354 #ifndef PRODUCT 355 //------------------------------dump------------------------------------------- 356 // Dump statistics for the hash table 357 void NodeHash::dump() { 358 _total_inserts += _inserts; 359 _total_insert_probes += _insert_probes; 360 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) { 361 if (WizardMode) { 362 for (uint i=0; i<_max; i++) { 363 if (_table[i]) 364 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx); 365 } 366 } 367 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max); 368 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0); 369 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses)); 370 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts); 371 // sentinels increase lookup cost, but not insert cost 372 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function"); 373 assert( _inserts+(_inserts>>3) < _max, "table too full" ); 374 assert( _inserts*3+100 >= _insert_probes, "bad hash function" ); 375 } 376 } 377 378 Node *NodeHash::find_index(uint idx) { // For debugging 379 // Find an entry by its index value 380 for( uint i = 0; i < _max; i++ ) { 381 Node *m = _table[i]; 382 if( !m || m == _sentinel ) continue; 383 if( m->_idx == (uint)idx ) return m; 384 } 385 return NULL; 386 } 387 #endif 388 389 #ifdef ASSERT 390 NodeHash::~NodeHash() { 391 // Unlock all nodes upon destruction of table. 392 if (_table != (Node**)badAddress) clear(); 393 } 394 395 void NodeHash::operator=(const NodeHash& nh) { 396 // Unlock all nodes upon replacement of table. 397 if (&nh == this) return; 398 if (_table != (Node**)badAddress) clear(); 399 memcpy((void*)this, (void*)&nh, sizeof(*this)); 400 // Do not increment hash_lock counts again. 401 // Instead, be sure we never again use the source table. 402 ((NodeHash*)&nh)->_table = (Node**)badAddress; 403 } 404 405 406 #endif 407 408 409 //============================================================================= 410 //------------------------------PhaseRemoveUseless----------------------------- 411 // 1) Use a breadthfirst walk to collect useful nodes reachable from root. 412 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN *gvn, Unique_Node_List *worklist, PhaseNumber phase_num) : Phase(phase_num), 413 _useful(Thread::current()->resource_area()) { 414 415 // Implementation requires 'UseLoopSafepoints == true' and an edge from root 416 // to each SafePointNode at a backward branch. Inserted in add_safepoint(). 417 if( !UseLoopSafepoints || !OptoRemoveUseless ) return; 418 419 // Identify nodes that are reachable from below, useful. 420 C->identify_useful_nodes(_useful); 421 // Update dead node list 422 C->update_dead_node_list(_useful); 423 424 // Remove all useless nodes from PhaseValues' recorded types 425 // Must be done before disconnecting nodes to preserve hash-table-invariant 426 gvn->remove_useless_nodes(_useful.member_set()); 427 428 // Remove all useless nodes from future worklist 429 worklist->remove_useless_nodes(_useful.member_set()); 430 431 // Disconnect 'useless' nodes that are adjacent to useful nodes 432 C->remove_useless_nodes(_useful); 433 } 434 435 //============================================================================= 436 //------------------------------PhaseRenumberLive------------------------------ 437 // First, remove useless nodes (equivalent to identifying live nodes). 438 // Then, renumber live nodes. 439 // 440 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure. 441 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the 442 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique 443 // value in the range [0, x). 444 // 445 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is 446 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes). 447 // 448 // The PhaseRenumberLive phase updates two data structures with the new node IDs. 449 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be 450 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'. 451 // (2) Type information (the field PhaseGVN::_types) maps type information to each 452 // node ID. The mapping is updated to use the new node IDs as well. Updated type 453 // information is returned in PhaseGVN::_types. 454 // 455 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist. 456 // 457 // Other data structures used by the compiler are not updated. The hash table for value 458 // numbering (the field PhaseGVN::_table) is not updated because computing the hash 459 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either 460 // because it is empty wherever PhaseRenumberLive is used. 461 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn, 462 Unique_Node_List* worklist, Unique_Node_List* new_worklist, 463 PhaseNumber phase_num) : 464 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live), 465 _new_type_array(C->comp_arena()), 466 _old2new_map(C->unique(), C->unique(), -1), 467 _delayed(Thread::current()->resource_area()), 468 _is_pass_finished(false), 469 _live_node_count(C->live_nodes()) 470 { 471 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place"); 472 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes"); 473 assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point"); 474 assert(_delayed.size() == 0, "should be empty"); 475 476 uint worklist_size = worklist->size(); 477 478 // Iterate over the set of live nodes. 479 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) { 480 Node* n = _useful.at(current_idx); 481 482 bool in_worklist = false; 483 if (worklist->member(n)) { 484 in_worklist = true; 485 } 486 487 const Type* type = gvn->type_or_null(n); 488 _new_type_array.map(current_idx, type); 489 490 assert(_old2new_map.at(n->_idx) == -1, "already seen"); 491 _old2new_map.at_put(n->_idx, current_idx); 492 493 n->set_idx(current_idx); // Update node ID. 494 495 if (in_worklist) { 496 new_worklist->push(n); 497 } 498 499 if (update_embedded_ids(n) < 0) { 500 _delayed.push(n); // has embedded IDs; handle later 501 } 502 } 503 504 assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist"); 505 assert(_live_node_count == _useful.size(), "all live nodes must be processed"); 506 507 _is_pass_finished = true; // pass finished; safe to process delayed updates 508 509 while (_delayed.size() > 0) { 510 Node* n = _delayed.pop(); 511 int no_of_updates = update_embedded_ids(n); 512 assert(no_of_updates > 0, "should be updated"); 513 } 514 515 // Replace the compiler's type information with the updated type information. 516 gvn->replace_types(_new_type_array); 517 518 // Update the unique node count of the compilation to the number of currently live nodes. 519 C->set_unique(_live_node_count); 520 521 // Set the dead node count to 0 and reset dead node list. 522 C->reset_dead_node_list(); 523 } 524 525 int PhaseRenumberLive::new_index(int old_idx) { 526 assert(_is_pass_finished, "not finished"); 527 if (_old2new_map.at(old_idx) == -1) { // absent 528 // Allocate a placeholder to preserve uniqueness 529 _old2new_map.at_put(old_idx, _live_node_count); 530 _live_node_count++; 531 } 532 return _old2new_map.at(old_idx); 533 } 534 535 int PhaseRenumberLive::update_embedded_ids(Node* n) { 536 int no_of_updates = 0; 537 if (n->is_Phi()) { 538 PhiNode* phi = n->as_Phi(); 539 if (phi->_inst_id != -1) { 540 if (!_is_pass_finished) { 541 return -1; // delay 542 } 543 int new_idx = new_index(phi->_inst_id); 544 assert(new_idx != -1, ""); 545 phi->_inst_id = new_idx; 546 no_of_updates++; 547 } 548 if (phi->_inst_mem_id != -1) { 549 if (!_is_pass_finished) { 550 return -1; // delay 551 } 552 int new_idx = new_index(phi->_inst_mem_id); 553 assert(new_idx != -1, ""); 554 phi->_inst_mem_id = new_idx; 555 no_of_updates++; 556 } 557 } 558 559 const Type* type = _new_type_array.fast_lookup(n->_idx); 560 if (type != NULL && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) { 561 if (!_is_pass_finished) { 562 return -1; // delay 563 } 564 int old_idx = type->is_oopptr()->instance_id(); 565 int new_idx = new_index(old_idx); 566 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx); 567 _new_type_array.map(n->_idx, new_type); 568 no_of_updates++; 569 } 570 571 return no_of_updates; 572 } 573 574 //============================================================================= 575 //------------------------------PhaseTransform--------------------------------- 576 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 577 _arena(Thread::current()->resource_area()), 578 _nodes(_arena), 579 _types(_arena) 580 { 581 init_con_caches(); 582 #ifndef PRODUCT 583 clear_progress(); 584 clear_transforms(); 585 set_allow_progress(true); 586 #endif 587 // Force allocation for currently existing nodes 588 _types.map(C->unique(), NULL); 589 } 590 591 //------------------------------PhaseTransform--------------------------------- 592 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 593 _arena(arena), 594 _nodes(arena), 595 _types(arena) 596 { 597 init_con_caches(); 598 #ifndef PRODUCT 599 clear_progress(); 600 clear_transforms(); 601 set_allow_progress(true); 602 #endif 603 // Force allocation for currently existing nodes 604 _types.map(C->unique(), NULL); 605 } 606 607 //------------------------------PhaseTransform--------------------------------- 608 // Initialize with previously generated type information 609 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 610 _arena(pt->_arena), 611 _nodes(pt->_nodes), 612 _types(pt->_types) 613 { 614 init_con_caches(); 615 #ifndef PRODUCT 616 clear_progress(); 617 clear_transforms(); 618 set_allow_progress(true); 619 #endif 620 } 621 622 void PhaseTransform::init_con_caches() { 623 memset(_icons,0,sizeof(_icons)); 624 memset(_lcons,0,sizeof(_lcons)); 625 memset(_zcons,0,sizeof(_zcons)); 626 } 627 628 629 //--------------------------------find_int_type-------------------------------- 630 const TypeInt* PhaseTransform::find_int_type(Node* n) { 631 if (n == NULL) return NULL; 632 // Call type_or_null(n) to determine node's type since we might be in 633 // parse phase and call n->Value() may return wrong type. 634 // (For example, a phi node at the beginning of loop parsing is not ready.) 635 const Type* t = type_or_null(n); 636 if (t == NULL) return NULL; 637 return t->isa_int(); 638 } 639 640 641 //-------------------------------find_long_type-------------------------------- 642 const TypeLong* PhaseTransform::find_long_type(Node* n) { 643 if (n == NULL) return NULL; 644 // (See comment above on type_or_null.) 645 const Type* t = type_or_null(n); 646 if (t == NULL) return NULL; 647 return t->isa_long(); 648 } 649 650 651 #ifndef PRODUCT 652 void PhaseTransform::dump_old2new_map() const { 653 _nodes.dump(); 654 } 655 656 void PhaseTransform::dump_new( uint nidx ) const { 657 for( uint i=0; i<_nodes.Size(); i++ ) 658 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 659 _nodes[i]->dump(); 660 tty->cr(); 661 tty->print_cr("Old index= %d",i); 662 return; 663 } 664 tty->print_cr("Node %d not found in the new indices", nidx); 665 } 666 667 //------------------------------dump_types------------------------------------- 668 void PhaseTransform::dump_types( ) const { 669 _types.dump(); 670 } 671 672 //------------------------------dump_nodes_and_types--------------------------- 673 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 674 VectorSet visited(Thread::current()->resource_area()); 675 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 676 } 677 678 //------------------------------dump_nodes_and_types_recur--------------------- 679 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 680 if( !n ) return; 681 if( depth == 0 ) return; 682 if( visited.test_set(n->_idx) ) return; 683 for( uint i=0; i<n->len(); i++ ) { 684 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 685 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 686 } 687 n->dump(); 688 if (type_or_null(n) != NULL) { 689 tty->print(" "); type(n)->dump(); tty->cr(); 690 } 691 } 692 693 #endif 694 695 696 //============================================================================= 697 //------------------------------PhaseValues------------------------------------ 698 // Set minimum table size to "255" 699 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 700 NOT_PRODUCT( clear_new_values(); ) 701 } 702 703 //------------------------------PhaseValues------------------------------------ 704 // Set minimum table size to "255" 705 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 706 _table(&ptv->_table) { 707 NOT_PRODUCT( clear_new_values(); ) 708 } 709 710 //------------------------------PhaseValues------------------------------------ 711 // Used by +VerifyOpto. Clear out hash table but copy _types array. 712 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 713 _table(ptv->arena(),ptv->_table.size()) { 714 NOT_PRODUCT( clear_new_values(); ) 715 } 716 717 //------------------------------~PhaseValues----------------------------------- 718 #ifndef PRODUCT 719 PhaseValues::~PhaseValues() { 720 _table.dump(); 721 722 // Statistics for value progress and efficiency 723 if( PrintCompilation && Verbose && WizardMode ) { 724 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 725 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 726 if( made_transforms() != 0 ) { 727 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 728 } else { 729 tty->cr(); 730 } 731 } 732 } 733 #endif 734 735 //------------------------------makecon---------------------------------------- 736 ConNode* PhaseTransform::makecon(const Type *t) { 737 assert(t->singleton(), "must be a constant"); 738 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 739 switch (t->base()) { // fast paths 740 case Type::Half: 741 case Type::Top: return (ConNode*) C->top(); 742 case Type::Int: return intcon( t->is_int()->get_con() ); 743 case Type::Long: return longcon( t->is_long()->get_con() ); 744 default: break; 745 } 746 if (t->is_zero_type()) 747 return zerocon(t->basic_type()); 748 return uncached_makecon(t); 749 } 750 751 //--------------------------uncached_makecon----------------------------------- 752 // Make an idealized constant - one of ConINode, ConPNode, etc. 753 ConNode* PhaseValues::uncached_makecon(const Type *t) { 754 assert(t->singleton(), "must be a constant"); 755 ConNode* x = ConNode::make(t); 756 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 757 if (k == NULL) { 758 set_type(x, t); // Missed, provide type mapping 759 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 760 if (nna != NULL) { 761 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 762 loc->clear(); // do not put debug info on constants 763 } 764 } else { 765 x->destruct(); // Hit, destroy duplicate constant 766 x = k; // use existing constant 767 } 768 return x; 769 } 770 771 //------------------------------intcon----------------------------------------- 772 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 773 ConINode* PhaseTransform::intcon(jint i) { 774 // Small integer? Check cache! Check that cached node is not dead 775 if (i >= _icon_min && i <= _icon_max) { 776 ConINode* icon = _icons[i-_icon_min]; 777 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 778 return icon; 779 } 780 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 781 assert(icon->is_Con(), ""); 782 if (i >= _icon_min && i <= _icon_max) 783 _icons[i-_icon_min] = icon; // Cache small integers 784 return icon; 785 } 786 787 //------------------------------longcon---------------------------------------- 788 // Fast long constant. 789 ConLNode* PhaseTransform::longcon(jlong l) { 790 // Small integer? Check cache! Check that cached node is not dead 791 if (l >= _lcon_min && l <= _lcon_max) { 792 ConLNode* lcon = _lcons[l-_lcon_min]; 793 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 794 return lcon; 795 } 796 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 797 assert(lcon->is_Con(), ""); 798 if (l >= _lcon_min && l <= _lcon_max) 799 _lcons[l-_lcon_min] = lcon; // Cache small integers 800 return lcon; 801 } 802 803 //------------------------------zerocon----------------------------------------- 804 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 805 ConNode* PhaseTransform::zerocon(BasicType bt) { 806 assert((uint)bt <= _zcon_max, "domain check"); 807 ConNode* zcon = _zcons[bt]; 808 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 809 return zcon; 810 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 811 _zcons[bt] = zcon; 812 return zcon; 813 } 814 815 816 817 //============================================================================= 818 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) { 819 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape); 820 if (i == NULL) { 821 i = k->Ideal(this, can_reshape); 822 } 823 return i; 824 } 825 826 Node* PhaseGVN::apply_identity(Node* k) { 827 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->identity_node(this, k); 828 if (i == k) { 829 i = k->Identity(this); 830 } 831 return i; 832 } 833 834 //------------------------------transform-------------------------------------- 835 // Return a node which computes the same function as this node, but in a 836 // faster or cheaper fashion. 837 Node *PhaseGVN::transform( Node *n ) { 838 return transform_no_reclaim(n); 839 } 840 841 //------------------------------transform-------------------------------------- 842 // Return a node which computes the same function as this node, but 843 // in a faster or cheaper fashion. 844 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 845 NOT_PRODUCT( set_transforms(); ) 846 847 // Apply the Ideal call in a loop until it no longer applies 848 Node *k = n; 849 NOT_PRODUCT( uint loop_count = 0; ) 850 while( 1 ) { 851 Node *i = apply_ideal(k, /*can_reshape=*/false); 852 if( !i ) break; 853 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 854 k = i; 855 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 856 } 857 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 858 859 860 // If brand new node, make space in type array. 861 ensure_type_or_null(k); 862 863 // Since I just called 'Value' to compute the set of run-time values 864 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 865 // cache Value. Later requests for the local phase->type of this Node can 866 // use the cached Value instead of suffering with 'bottom_type'. 867 const Type *t = k->Value(this); // Get runtime Value set 868 assert(t != NULL, "value sanity"); 869 if (type_or_null(k) != t) { 870 #ifndef PRODUCT 871 // Do not count initial visit to node as a transformation 872 if (type_or_null(k) == NULL) { 873 inc_new_values(); 874 set_progress(); 875 } 876 #endif 877 set_type(k, t); 878 // If k is a TypeNode, capture any more-precise type permanently into Node 879 k->raise_bottom_type(t); 880 } 881 882 if( t->singleton() && !k->is_Con() ) { 883 NOT_PRODUCT( set_progress(); ) 884 return makecon(t); // Turn into a constant 885 } 886 887 // Now check for Identities 888 Node *i = apply_identity(k); // Look for a nearby replacement 889 if( i != k ) { // Found? Return replacement! 890 NOT_PRODUCT( set_progress(); ) 891 return i; 892 } 893 894 // Global Value Numbering 895 i = hash_find_insert(k); // Insert if new 896 if( i && (i != k) ) { 897 // Return the pre-existing node 898 NOT_PRODUCT( set_progress(); ) 899 return i; 900 } 901 902 // Return Idealized original 903 return k; 904 } 905 906 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 907 if (d->is_top() || n->is_top()) { 908 return false; 909 } 910 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 911 int i = 0; 912 while (d != n) { 913 n = IfNode::up_one_dom(n, linear_only); 914 i++; 915 if (n == NULL || i >= 10) { 916 return false; 917 } 918 } 919 return true; 920 } 921 922 #ifdef ASSERT 923 //------------------------------dead_loop_check-------------------------------- 924 // Check for a simple dead loop when a data node references itself directly 925 // or through an other data node excluding cons and phis. 926 void PhaseGVN::dead_loop_check( Node *n ) { 927 // Phi may reference itself in a loop 928 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 929 // Do 2 levels check and only data inputs. 930 bool no_dead_loop = true; 931 uint cnt = n->req(); 932 for (uint i = 1; i < cnt && no_dead_loop; i++) { 933 Node *in = n->in(i); 934 if (in == n) { 935 no_dead_loop = false; 936 } else if (in != NULL && !in->is_dead_loop_safe()) { 937 uint icnt = in->req(); 938 for (uint j = 1; j < icnt && no_dead_loop; j++) { 939 if (in->in(j) == n || in->in(j) == in) 940 no_dead_loop = false; 941 } 942 } 943 } 944 if (!no_dead_loop) n->dump(3); 945 assert(no_dead_loop, "dead loop detected"); 946 } 947 } 948 #endif 949 950 //============================================================================= 951 //------------------------------PhaseIterGVN----------------------------------- 952 // Initialize hash table to fresh and clean for +VerifyOpto 953 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), 954 _delay_transform(false), 955 _stack(C->live_nodes() >> 1), 956 _worklist( ) { 957 } 958 959 //------------------------------PhaseIterGVN----------------------------------- 960 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 961 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 962 _delay_transform(igvn->_delay_transform), 963 _stack( igvn->_stack ), 964 _worklist( igvn->_worklist ) 965 { 966 } 967 968 //------------------------------PhaseIterGVN----------------------------------- 969 // Initialize with previous PhaseGVN info from Parser 970 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 971 _delay_transform(false), 972 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that 973 // the constructor is used in incremental inlining, this consumes too much memory: 974 // _stack(C->live_nodes() >> 1), 975 // So, as a band-aid, we replace this by: 976 _stack(C->comp_arena(), 32), 977 _worklist(*C->for_igvn()) 978 { 979 uint max; 980 981 // Dead nodes in the hash table inherited from GVN were not treated as 982 // roots during def-use info creation; hence they represent an invisible 983 // use. Clear them out. 984 max = _table.size(); 985 for( uint i = 0; i < max; ++i ) { 986 Node *n = _table.at(i); 987 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 988 if( n->is_top() ) continue; 989 assert( false, "Parse::remove_useless_nodes missed this node"); 990 hash_delete(n); 991 } 992 } 993 994 // Any Phis or Regions on the worklist probably had uses that could not 995 // make more progress because the uses were made while the Phis and Regions 996 // were in half-built states. Put all uses of Phis and Regions on worklist. 997 max = _worklist.size(); 998 for( uint j = 0; j < max; j++ ) { 999 Node *n = _worklist.at(j); 1000 uint uop = n->Opcode(); 1001 if( uop == Op_Phi || uop == Op_Region || 1002 n->is_Type() || 1003 n->is_Mem() ) 1004 add_users_to_worklist(n); 1005 } 1006 } 1007 1008 /** 1009 * Initialize worklist for each node. 1010 */ 1011 void PhaseIterGVN::init_worklist(Node* first) { 1012 Unique_Node_List to_process; 1013 to_process.push(first); 1014 1015 while (to_process.size() > 0) { 1016 Node* n = to_process.pop(); 1017 if (!_worklist.member(n)) { 1018 _worklist.push(n); 1019 1020 uint cnt = n->req(); 1021 for(uint i = 0; i < cnt; i++) { 1022 Node* m = n->in(i); 1023 if (m != NULL) { 1024 to_process.push(m); 1025 } 1026 } 1027 } 1028 } 1029 } 1030 1031 #ifndef PRODUCT 1032 void PhaseIterGVN::verify_step(Node* n) { 1033 if (VerifyIterativeGVN) { 1034 _verify_window[_verify_counter % _verify_window_size] = n; 1035 ++_verify_counter; 1036 ResourceMark rm; 1037 ResourceArea* area = Thread::current()->resource_area(); 1038 VectorSet old_space(area), new_space(area); 1039 if (C->unique() < 1000 || 1040 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 1041 ++_verify_full_passes; 1042 Node::verify_recur(C->root(), -1, old_space, new_space); 1043 } 1044 const int verify_depth = 4; 1045 for ( int i = 0; i < _verify_window_size; i++ ) { 1046 Node* n = _verify_window[i]; 1047 if ( n == NULL ) continue; 1048 if( n->in(0) == NodeSentinel ) { // xform_idom 1049 _verify_window[i] = n->in(1); 1050 --i; continue; 1051 } 1052 // Typical fanout is 1-2, so this call visits about 6 nodes. 1053 Node::verify_recur(n, verify_depth, old_space, new_space); 1054 } 1055 } 1056 } 1057 1058 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 1059 if (TraceIterativeGVN) { 1060 uint wlsize = _worklist.size(); 1061 const Type* newtype = type_or_null(n); 1062 if (nn != n) { 1063 // print old node 1064 tty->print("< "); 1065 if (oldtype != newtype && oldtype != NULL) { 1066 oldtype->dump(); 1067 } 1068 do { tty->print("\t"); } while (tty->position() < 16); 1069 tty->print("<"); 1070 n->dump(); 1071 } 1072 if (oldtype != newtype || nn != n) { 1073 // print new node and/or new type 1074 if (oldtype == NULL) { 1075 tty->print("* "); 1076 } else if (nn != n) { 1077 tty->print("> "); 1078 } else { 1079 tty->print("= "); 1080 } 1081 if (newtype == NULL) { 1082 tty->print("null"); 1083 } else { 1084 newtype->dump(); 1085 } 1086 do { tty->print("\t"); } while (tty->position() < 16); 1087 nn->dump(); 1088 } 1089 if (Verbose && wlsize < _worklist.size()) { 1090 tty->print(" Push {"); 1091 while (wlsize != _worklist.size()) { 1092 Node* pushed = _worklist.at(wlsize++); 1093 tty->print(" %d", pushed->_idx); 1094 } 1095 tty->print_cr(" }"); 1096 } 1097 if (nn != n) { 1098 // ignore n, it might be subsumed 1099 verify_step((Node*) NULL); 1100 } 1101 } 1102 } 1103 1104 void PhaseIterGVN::init_verifyPhaseIterGVN() { 1105 _verify_counter = 0; 1106 _verify_full_passes = 0; 1107 for (int i = 0; i < _verify_window_size; i++) { 1108 _verify_window[i] = NULL; 1109 } 1110 #ifdef ASSERT 1111 // Verify that all modified nodes are on _worklist 1112 Unique_Node_List* modified_list = C->modified_nodes(); 1113 while (modified_list != NULL && modified_list->size()) { 1114 Node* n = modified_list->pop(); 1115 if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) { 1116 n->dump(); 1117 assert(false, "modified node is not on IGVN._worklist"); 1118 } 1119 } 1120 #endif 1121 } 1122 1123 void PhaseIterGVN::verify_PhaseIterGVN() { 1124 #ifdef ASSERT 1125 // Verify nodes with changed inputs. 1126 Unique_Node_List* modified_list = C->modified_nodes(); 1127 while (modified_list != NULL && modified_list->size()) { 1128 Node* n = modified_list->pop(); 1129 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes 1130 n->dump(); 1131 assert(false, "modified node was not processed by IGVN.transform_old()"); 1132 } 1133 } 1134 #endif 1135 1136 C->verify_graph_edges(); 1137 if( VerifyOpto && allow_progress() ) { 1138 // Must turn off allow_progress to enable assert and break recursion 1139 C->root()->verify(); 1140 { // Check if any progress was missed using IterGVN 1141 // Def-Use info enables transformations not attempted in wash-pass 1142 // e.g. Region/Phi cleanup, ... 1143 // Null-check elision -- may not have reached fixpoint 1144 // do not propagate to dominated nodes 1145 ResourceMark rm; 1146 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 1147 // Fill worklist completely 1148 igvn2.init_worklist(C->root()); 1149 1150 igvn2.set_allow_progress(false); 1151 igvn2.optimize(); 1152 igvn2.set_allow_progress(true); 1153 } 1154 } 1155 if (VerifyIterativeGVN && PrintOpto) { 1156 if (_verify_counter == _verify_full_passes) { 1157 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 1158 (int) _verify_full_passes); 1159 } else { 1160 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 1161 (int) _verify_counter, (int) _verify_full_passes); 1162 } 1163 } 1164 1165 #ifdef ASSERT 1166 while (modified_list->size()) { 1167 Node* n = modified_list->pop(); 1168 n->dump(); 1169 assert(false, "VerifyIterativeGVN: new modified node was added"); 1170 } 1171 #endif 1172 } 1173 #endif /* PRODUCT */ 1174 1175 #ifdef ASSERT 1176 /** 1177 * Dumps information that can help to debug the problem. A debug 1178 * build fails with an assert. 1179 */ 1180 void PhaseIterGVN::dump_infinite_loop_info(Node* n) { 1181 n->dump(4); 1182 _worklist.dump(); 1183 assert(false, "infinite loop in PhaseIterGVN::optimize"); 1184 } 1185 1186 /** 1187 * Prints out information about IGVN if the 'verbose' option is used. 1188 */ 1189 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1190 if (TraceIterativeGVN && Verbose) { 1191 tty->print(" Pop "); 1192 n->dump(); 1193 if ((num_processed % 100) == 0) { 1194 _worklist.print_set(); 1195 } 1196 } 1197 } 1198 #endif /* ASSERT */ 1199 1200 void PhaseIterGVN::optimize() { 1201 DEBUG_ONLY(uint num_processed = 0;) 1202 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1203 1204 uint loop_count = 0; 1205 // Pull from worklist and transform the node. If the node has changed, 1206 // update edge info and put uses on worklist. 1207 while(_worklist.size()) { 1208 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1209 return; 1210 } 1211 Node* n = _worklist.pop(); 1212 if (++loop_count >= K * C->live_nodes()) { 1213 DEBUG_ONLY(dump_infinite_loop_info(n);) 1214 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1215 return; 1216 } 1217 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1218 if (n->outcnt() != 0) { 1219 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1220 // Do the transformation 1221 Node* nn = transform_old(n); 1222 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1223 } else if (!n->is_top()) { 1224 remove_dead_node(n); 1225 } 1226 } 1227 NOT_PRODUCT(verify_PhaseIterGVN();) 1228 } 1229 1230 1231 /** 1232 * Register a new node with the optimizer. Update the types array, the def-use 1233 * info. Put on worklist. 1234 */ 1235 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1236 set_type_bottom(n); 1237 _worklist.push(n); 1238 if (orig != NULL) C->copy_node_notes_to(n, orig); 1239 return n; 1240 } 1241 1242 //------------------------------transform-------------------------------------- 1243 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1244 Node *PhaseIterGVN::transform( Node *n ) { 1245 if (_delay_transform) { 1246 // Register the node but don't optimize for now 1247 register_new_node_with_optimizer(n); 1248 return n; 1249 } 1250 1251 // If brand new node, make space in type array, and give it a type. 1252 ensure_type_or_null(n); 1253 if (type_or_null(n) == NULL) { 1254 set_type_bottom(n); 1255 } 1256 1257 return transform_old(n); 1258 } 1259 1260 Node *PhaseIterGVN::transform_old(Node* n) { 1261 DEBUG_ONLY(uint loop_count = 0;); 1262 NOT_PRODUCT(set_transforms()); 1263 1264 // Remove 'n' from hash table in case it gets modified 1265 _table.hash_delete(n); 1266 if (VerifyIterativeGVN) { 1267 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1268 } 1269 1270 // Apply the Ideal call in a loop until it no longer applies 1271 Node* k = n; 1272 DEBUG_ONLY(dead_loop_check(k);) 1273 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1274 C->remove_modified_node(k); 1275 Node* i = apply_ideal(k, /*can_reshape=*/true); 1276 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1277 #ifndef PRODUCT 1278 verify_step(k); 1279 if (i && VerifyOpto ) { 1280 if (!allow_progress()) { 1281 if (i->is_Add() && (i->outcnt() == 1)) { 1282 // Switched input to left side because this is the only use 1283 } else if (i->is_If() && (i->in(0) == NULL)) { 1284 // This IF is dead because it is dominated by an equivalent IF When 1285 // dominating if changed, info is not propagated sparsely to 'this' 1286 // Propagating this info further will spuriously identify other 1287 // progress. 1288 return i; 1289 } else 1290 set_progress(); 1291 } else { 1292 set_progress(); 1293 } 1294 } 1295 #endif 1296 1297 while (i != NULL) { 1298 #ifdef ASSERT 1299 if (loop_count >= K) { 1300 dump_infinite_loop_info(i); 1301 } 1302 loop_count++; 1303 #endif 1304 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1305 // Made a change; put users of original Node on worklist 1306 add_users_to_worklist(k); 1307 // Replacing root of transform tree? 1308 if (k != i) { 1309 // Make users of old Node now use new. 1310 subsume_node(k, i); 1311 k = i; 1312 } 1313 DEBUG_ONLY(dead_loop_check(k);) 1314 // Try idealizing again 1315 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1316 C->remove_modified_node(k); 1317 i = apply_ideal(k, /*can_reshape=*/true); 1318 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1319 #ifndef PRODUCT 1320 verify_step(k); 1321 if (i && VerifyOpto) { 1322 set_progress(); 1323 } 1324 #endif 1325 } 1326 1327 // If brand new node, make space in type array. 1328 ensure_type_or_null(k); 1329 1330 // See what kind of values 'k' takes on at runtime 1331 const Type* t = k->Value(this); 1332 assert(t != NULL, "value sanity"); 1333 1334 // Since I just called 'Value' to compute the set of run-time values 1335 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1336 // cache Value. Later requests for the local phase->type of this Node can 1337 // use the cached Value instead of suffering with 'bottom_type'. 1338 if (type_or_null(k) != t) { 1339 #ifndef PRODUCT 1340 inc_new_values(); 1341 set_progress(); 1342 #endif 1343 set_type(k, t); 1344 // If k is a TypeNode, capture any more-precise type permanently into Node 1345 k->raise_bottom_type(t); 1346 // Move users of node to worklist 1347 add_users_to_worklist(k); 1348 } 1349 // If 'k' computes a constant, replace it with a constant 1350 if (t->singleton() && !k->is_Con()) { 1351 NOT_PRODUCT(set_progress();) 1352 Node* con = makecon(t); // Make a constant 1353 add_users_to_worklist(k); 1354 subsume_node(k, con); // Everybody using k now uses con 1355 return con; 1356 } 1357 1358 // Now check for Identities 1359 i = apply_identity(k); // Look for a nearby replacement 1360 if (i != k) { // Found? Return replacement! 1361 NOT_PRODUCT(set_progress();) 1362 add_users_to_worklist(k); 1363 subsume_node(k, i); // Everybody using k now uses i 1364 return i; 1365 } 1366 1367 // Global Value Numbering 1368 i = hash_find_insert(k); // Check for pre-existing node 1369 if (i && (i != k)) { 1370 // Return the pre-existing node if it isn't dead 1371 NOT_PRODUCT(set_progress();) 1372 add_users_to_worklist(k); 1373 subsume_node(k, i); // Everybody using k now uses i 1374 return i; 1375 } 1376 1377 // Return Idealized original 1378 return k; 1379 } 1380 1381 //---------------------------------saturate------------------------------------ 1382 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1383 const Type* limit_type) const { 1384 return new_type->narrow(old_type); 1385 } 1386 1387 //------------------------------remove_globally_dead_node---------------------- 1388 // Kill a globally dead Node. All uses are also globally dead and are 1389 // aggressively trimmed. 1390 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1391 enum DeleteProgress { 1392 PROCESS_INPUTS, 1393 PROCESS_OUTPUTS 1394 }; 1395 assert(_stack.is_empty(), "not empty"); 1396 _stack.push(dead, PROCESS_INPUTS); 1397 1398 while (_stack.is_nonempty()) { 1399 dead = _stack.node(); 1400 if (dead->Opcode() == Op_SafePoint) { 1401 dead->as_SafePoint()->disconnect_from_root(this); 1402 } 1403 uint progress_state = _stack.index(); 1404 assert(dead != C->root(), "killing root, eh?"); 1405 assert(!dead->is_top(), "add check for top when pushing"); 1406 NOT_PRODUCT( set_progress(); ) 1407 if (progress_state == PROCESS_INPUTS) { 1408 // After following inputs, continue to outputs 1409 _stack.set_index(PROCESS_OUTPUTS); 1410 if (!dead->is_Con()) { // Don't kill cons but uses 1411 bool recurse = false; 1412 // Remove from hash table 1413 _table.hash_delete( dead ); 1414 // Smash all inputs to 'dead', isolating him completely 1415 for (uint i = 0; i < dead->req(); i++) { 1416 Node *in = dead->in(i); 1417 if (in != NULL && in != C->top()) { // Points to something? 1418 int nrep = dead->replace_edge(in, NULL); // Kill edges 1419 assert((nrep > 0), "sanity"); 1420 if (in->outcnt() == 0) { // Made input go dead? 1421 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1422 recurse = true; 1423 } else if (in->outcnt() == 1 && 1424 in->has_special_unique_user()) { 1425 _worklist.push(in->unique_out()); 1426 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1427 if (in->Opcode() == Op_Region) { 1428 _worklist.push(in); 1429 } else if (in->is_Store()) { 1430 DUIterator_Fast imax, i = in->fast_outs(imax); 1431 _worklist.push(in->fast_out(i)); 1432 i++; 1433 if (in->outcnt() == 2) { 1434 _worklist.push(in->fast_out(i)); 1435 i++; 1436 } 1437 assert(!(i < imax), "sanity"); 1438 } 1439 } else { 1440 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in); 1441 } 1442 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1443 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1444 // A Load that directly follows an InitializeNode is 1445 // going away. The Stores that follow are candidates 1446 // again to be captured by the InitializeNode. 1447 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1448 Node *n = in->fast_out(j); 1449 if (n->is_Store()) { 1450 _worklist.push(n); 1451 } 1452 } 1453 } 1454 } // if (in != NULL && in != C->top()) 1455 } // for (uint i = 0; i < dead->req(); i++) 1456 if (recurse) { 1457 continue; 1458 } 1459 } // if (!dead->is_Con()) 1460 } // if (progress_state == PROCESS_INPUTS) 1461 1462 // Aggressively kill globally dead uses 1463 // (Rather than pushing all the outs at once, we push one at a time, 1464 // plus the parent to resume later, because of the indefinite number 1465 // of edge deletions per loop trip.) 1466 if (dead->outcnt() > 0) { 1467 // Recursively remove output edges 1468 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1469 } else { 1470 // Finished disconnecting all input and output edges. 1471 _stack.pop(); 1472 // Remove dead node from iterative worklist 1473 _worklist.remove(dead); 1474 C->remove_modified_node(dead); 1475 // Constant node that has no out-edges and has only one in-edge from 1476 // root is usually dead. However, sometimes reshaping walk makes 1477 // it reachable by adding use edges. So, we will NOT count Con nodes 1478 // as dead to be conservative about the dead node count at any 1479 // given time. 1480 if (!dead->is_Con()) { 1481 C->record_dead_node(dead->_idx); 1482 } 1483 if (dead->is_macro()) { 1484 C->remove_macro_node(dead); 1485 } 1486 if (dead->is_expensive()) { 1487 C->remove_expensive_node(dead); 1488 } 1489 CastIINode* cast = dead->isa_CastII(); 1490 if (cast != NULL && cast->has_range_check()) { 1491 C->remove_range_check_cast(cast); 1492 } 1493 if (dead->Opcode() == Op_Opaque4) { 1494 C->remove_opaque4_node(dead); 1495 } 1496 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1497 bs->unregister_potential_barrier_node(dead); 1498 } 1499 } // while (_stack.is_nonempty()) 1500 } 1501 1502 //------------------------------subsume_node----------------------------------- 1503 // Remove users from node 'old' and add them to node 'nn'. 1504 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1505 if (old->Opcode() == Op_SafePoint) { 1506 old->as_SafePoint()->disconnect_from_root(this); 1507 } 1508 assert( old != hash_find(old), "should already been removed" ); 1509 assert( old != C->top(), "cannot subsume top node"); 1510 // Copy debug or profile information to the new version: 1511 C->copy_node_notes_to(nn, old); 1512 // Move users of node 'old' to node 'nn' 1513 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1514 Node* use = old->last_out(i); // for each use... 1515 // use might need re-hashing (but it won't if it's a new node) 1516 rehash_node_delayed(use); 1517 // Update use-def info as well 1518 // We remove all occurrences of old within use->in, 1519 // so as to avoid rehashing any node more than once. 1520 // The hash table probe swamps any outer loop overhead. 1521 uint num_edges = 0; 1522 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1523 if (use->in(j) == old) { 1524 use->set_req(j, nn); 1525 ++num_edges; 1526 } 1527 } 1528 i -= num_edges; // we deleted 1 or more copies of this edge 1529 } 1530 1531 // Search for instance field data PhiNodes in the same region pointing to the old 1532 // memory PhiNode and update their instance memory ids to point to the new node. 1533 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) { 1534 Node* region = old->in(0); 1535 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1536 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1537 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) { 1538 phi->set_inst_mem_id((int)nn->_idx); 1539 } 1540 } 1541 } 1542 1543 // Smash all inputs to 'old', isolating him completely 1544 Node *temp = new Node(1); 1545 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1546 remove_dead_node( old ); 1547 temp->del_req(0); // Yank bogus edge 1548 #ifndef PRODUCT 1549 if( VerifyIterativeGVN ) { 1550 for ( int i = 0; i < _verify_window_size; i++ ) { 1551 if ( _verify_window[i] == old ) 1552 _verify_window[i] = nn; 1553 } 1554 } 1555 #endif 1556 _worklist.remove(temp); // this can be necessary 1557 temp->destruct(); // reuse the _idx of this little guy 1558 } 1559 1560 //------------------------------add_users_to_worklist-------------------------- 1561 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1562 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1563 _worklist.push(n->fast_out(i)); // Push on worklist 1564 } 1565 } 1566 1567 // Return counted loop Phi if as a counted loop exit condition, cmp 1568 // compares the the induction variable with n 1569 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) { 1570 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1571 Node* bol = cmp->fast_out(i); 1572 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1573 Node* iff = bol->fast_out(i2); 1574 if (iff->is_CountedLoopEnd()) { 1575 CountedLoopEndNode* cle = iff->as_CountedLoopEnd(); 1576 if (cle->limit() == n) { 1577 PhiNode* phi = cle->phi(); 1578 if (phi != NULL) { 1579 return phi; 1580 } 1581 } 1582 } 1583 } 1584 } 1585 return NULL; 1586 } 1587 1588 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1589 add_users_to_worklist0(n); 1590 1591 // Move users of node to worklist 1592 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1593 Node* use = n->fast_out(i); // Get use 1594 1595 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1596 use->is_Store() ) // Enable store/load same address 1597 add_users_to_worklist0(use); 1598 1599 // If we changed the receiver type to a call, we need to revisit 1600 // the Catch following the call. It's looking for a non-NULL 1601 // receiver to know when to enable the regular fall-through path 1602 // in addition to the NullPtrException path. 1603 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1604 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control); 1605 if (p != NULL) { 1606 add_users_to_worklist0(p); 1607 } 1608 } 1609 1610 uint use_op = use->Opcode(); 1611 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1612 add_users_to_worklist(use); // Put Bool on worklist 1613 if (use->outcnt() > 0) { 1614 Node* bol = use->raw_out(0); 1615 if (bol->outcnt() > 0) { 1616 Node* iff = bol->raw_out(0); 1617 if (iff->outcnt() == 2) { 1618 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1619 // phi merging either 0 or 1 onto the worklist 1620 Node* ifproj0 = iff->raw_out(0); 1621 Node* ifproj1 = iff->raw_out(1); 1622 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1623 Node* region0 = ifproj0->raw_out(0); 1624 Node* region1 = ifproj1->raw_out(0); 1625 if( region0 == region1 ) 1626 add_users_to_worklist0(region0); 1627 } 1628 } 1629 } 1630 } 1631 if (use_op == Op_CmpI) { 1632 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n); 1633 if (phi != NULL) { 1634 // If an opaque node feeds into the limit condition of a 1635 // CountedLoop, we need to process the Phi node for the 1636 // induction variable when the opaque node is removed: 1637 // the range of values taken by the Phi is now known and 1638 // so its type is also known. 1639 _worklist.push(phi); 1640 } 1641 Node* in1 = use->in(1); 1642 for (uint i = 0; i < in1->outcnt(); i++) { 1643 if (in1->raw_out(i)->Opcode() == Op_CastII) { 1644 Node* castii = in1->raw_out(i); 1645 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) { 1646 Node* ifnode = castii->in(0)->in(0); 1647 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) { 1648 // Reprocess a CastII node that may depend on an 1649 // opaque node value when the opaque node is 1650 // removed. In case it carries a dependency we can do 1651 // a better job of computing its type. 1652 _worklist.push(castii); 1653 } 1654 } 1655 } 1656 } 1657 } 1658 } 1659 1660 // If changed Cast input, check Phi users for simple cycles 1661 if (use->is_ConstraintCast()) { 1662 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1663 Node* u = use->fast_out(i2); 1664 if (u->is_Phi()) 1665 _worklist.push(u); 1666 } 1667 } 1668 // If changed LShift inputs, check RShift users for useless sign-ext 1669 if( use_op == Op_LShiftI ) { 1670 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1671 Node* u = use->fast_out(i2); 1672 if (u->Opcode() == Op_RShiftI) 1673 _worklist.push(u); 1674 } 1675 } 1676 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1677 if (use_op == Op_AddI || use_op == Op_SubI) { 1678 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1679 Node* u = use->fast_out(i2); 1680 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1681 _worklist.push(u); 1682 } 1683 } 1684 } 1685 // If changed AddP inputs, check Stores for loop invariant 1686 if( use_op == Op_AddP ) { 1687 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1688 Node* u = use->fast_out(i2); 1689 if (u->is_Mem()) 1690 _worklist.push(u); 1691 } 1692 } 1693 // If changed initialization activity, check dependent Stores 1694 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1695 InitializeNode* init = use->as_Allocate()->initialization(); 1696 if (init != NULL) { 1697 Node* imem = init->proj_out_or_null(TypeFunc::Memory); 1698 if (imem != NULL) add_users_to_worklist0(imem); 1699 } 1700 } 1701 if (use_op == Op_Initialize) { 1702 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory); 1703 if (imem != NULL) add_users_to_worklist0(imem); 1704 } 1705 // Loading the java mirror from a Klass requires two loads and the type 1706 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1707 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1708 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1709 bool has_load_barriers = bs->has_load_barriers(); 1710 1711 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) { 1712 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1713 Node* u = use->fast_out(i2); 1714 const Type* ut = u->bottom_type(); 1715 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) { 1716 if (has_load_barriers) { 1717 // Search for load barriers behind the load 1718 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1719 Node* b = u->fast_out(i3); 1720 if (bs->is_gc_barrier_node(b)) { 1721 _worklist.push(b); 1722 } 1723 } 1724 } 1725 _worklist.push(u); 1726 } 1727 } 1728 } 1729 1730 BarrierSet::barrier_set()->barrier_set_c2()->igvn_add_users_to_worklist(this, use); 1731 } 1732 } 1733 1734 /** 1735 * Remove the speculative part of all types that we know of 1736 */ 1737 void PhaseIterGVN::remove_speculative_types() { 1738 assert(UseTypeSpeculation, "speculation is off"); 1739 for (uint i = 0; i < _types.Size(); i++) { 1740 const Type* t = _types.fast_lookup(i); 1741 if (t != NULL) { 1742 _types.map(i, t->remove_speculative()); 1743 } 1744 } 1745 _table.check_no_speculative_types(); 1746 } 1747 1748 //============================================================================= 1749 #ifndef PRODUCT 1750 uint PhaseCCP::_total_invokes = 0; 1751 uint PhaseCCP::_total_constants = 0; 1752 #endif 1753 //------------------------------PhaseCCP--------------------------------------- 1754 // Conditional Constant Propagation, ala Wegman & Zadeck 1755 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1756 NOT_PRODUCT( clear_constants(); ) 1757 assert( _worklist.size() == 0, "" ); 1758 // Clear out _nodes from IterGVN. Must be clear to transform call. 1759 _nodes.clear(); // Clear out from IterGVN 1760 analyze(); 1761 } 1762 1763 #ifndef PRODUCT 1764 //------------------------------~PhaseCCP-------------------------------------- 1765 PhaseCCP::~PhaseCCP() { 1766 inc_invokes(); 1767 _total_constants += count_constants(); 1768 } 1769 #endif 1770 1771 1772 #ifdef ASSERT 1773 static bool ccp_type_widens(const Type* t, const Type* t0) { 1774 assert(t->meet(t0) == t, "Not monotonic"); 1775 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1776 case Type::Int: 1777 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1778 break; 1779 case Type::Long: 1780 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1781 break; 1782 default: 1783 break; 1784 } 1785 return true; 1786 } 1787 #endif //ASSERT 1788 1789 //------------------------------analyze---------------------------------------- 1790 void PhaseCCP::analyze() { 1791 // Initialize all types to TOP, optimistic analysis 1792 for (int i = C->unique() - 1; i >= 0; i--) { 1793 _types.map(i,Type::TOP); 1794 } 1795 1796 // Push root onto worklist 1797 Unique_Node_List worklist; 1798 worklist.push(C->root()); 1799 1800 // Pull from worklist; compute new value; push changes out. 1801 // This loop is the meat of CCP. 1802 while( worklist.size() ) { 1803 Node *n = worklist.pop(); 1804 const Type *t = n->Value(this); 1805 if (t != type(n)) { 1806 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1807 #ifndef PRODUCT 1808 if( TracePhaseCCP ) { 1809 t->dump(); 1810 do { tty->print("\t"); } while (tty->position() < 16); 1811 n->dump(); 1812 } 1813 #endif 1814 set_type(n, t); 1815 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1816 Node* m = n->fast_out(i); // Get user 1817 if (m->is_Region()) { // New path to Region? Must recheck Phis too 1818 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1819 Node* p = m->fast_out(i2); // Propagate changes to uses 1820 if (p->bottom_type() != type(p)) { // If not already bottomed out 1821 worklist.push(p); // Propagate change to user 1822 } 1823 } 1824 } 1825 // If we changed the receiver type to a call, we need to revisit 1826 // the Catch following the call. It's looking for a non-NULL 1827 // receiver to know when to enable the regular fall-through path 1828 // in addition to the NullPtrException path 1829 if (m->is_Call()) { 1830 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1831 Node* p = m->fast_out(i2); // Propagate changes to uses 1832 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) { 1833 worklist.push(p->unique_out()); 1834 } 1835 } 1836 } 1837 if (m->bottom_type() != type(m)) { // If not already bottomed out 1838 worklist.push(m); // Propagate change to user 1839 } 1840 1841 // CmpU nodes can get their type information from two nodes up in the 1842 // graph (instead of from the nodes immediately above). Make sure they 1843 // are added to the worklist if nodes they depend on are updated, since 1844 // they could be missed and get wrong types otherwise. 1845 uint m_op = m->Opcode(); 1846 if (m_op == Op_AddI || m_op == Op_SubI) { 1847 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1848 Node* p = m->fast_out(i2); // Propagate changes to uses 1849 if (p->Opcode() == Op_CmpU) { 1850 // Got a CmpU which might need the new type information from node n. 1851 if(p->bottom_type() != type(p)) { // If not already bottomed out 1852 worklist.push(p); // Propagate change to user 1853 } 1854 } 1855 } 1856 } 1857 // If n is used in a counted loop exit condition then the type 1858 // of the counted loop's Phi depends on the type of n. See 1859 // PhiNode::Value(). 1860 if (m_op == Op_CmpI) { 1861 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n); 1862 if (phi != NULL) { 1863 worklist.push(phi); 1864 } 1865 } 1866 // Loading the java mirror from a Klass requires two loads and the type 1867 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1868 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1869 bool has_load_barriers = bs->has_load_barriers(); 1870 1871 if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) { 1872 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1873 Node* u = m->fast_out(i2); 1874 const Type* ut = u->bottom_type(); 1875 if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) { 1876 if (has_load_barriers) { 1877 // Search for load barriers behind the load 1878 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1879 Node* b = u->fast_out(i3); 1880 if (bs->is_gc_barrier_node(b)) { 1881 worklist.push(b); 1882 } 1883 } 1884 } 1885 worklist.push(u); 1886 } 1887 } 1888 } 1889 1890 BarrierSet::barrier_set()->barrier_set_c2()->ccp_analyze(this, worklist, m); 1891 } 1892 } 1893 } 1894 } 1895 1896 //------------------------------do_transform----------------------------------- 1897 // Top level driver for the recursive transformer 1898 void PhaseCCP::do_transform() { 1899 // Correct leaves of new-space Nodes; they point to old-space. 1900 C->set_root( transform(C->root())->as_Root() ); 1901 assert( C->top(), "missing TOP node" ); 1902 assert( C->root(), "missing root" ); 1903 } 1904 1905 //------------------------------transform-------------------------------------- 1906 // Given a Node in old-space, clone him into new-space. 1907 // Convert any of his old-space children into new-space children. 1908 Node *PhaseCCP::transform( Node *n ) { 1909 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1910 if( new_node != NULL ) 1911 return new_node; // Been there, done that, return old answer 1912 new_node = transform_once(n); // Check for constant 1913 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1914 1915 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1916 GrowableArray <Node *> trstack(C->live_nodes() >> 1); 1917 1918 trstack.push(new_node); // Process children of cloned node 1919 while ( trstack.is_nonempty() ) { 1920 Node *clone = trstack.pop(); 1921 uint cnt = clone->req(); 1922 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1923 Node *input = clone->in(i); 1924 if( input != NULL ) { // Ignore NULLs 1925 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1926 if( new_input == NULL ) { 1927 new_input = transform_once(input); // Check for constant 1928 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1929 trstack.push(new_input); 1930 } 1931 assert( new_input == clone->in(i), "insanity check"); 1932 } 1933 } 1934 } 1935 return new_node; 1936 } 1937 1938 1939 //------------------------------transform_once--------------------------------- 1940 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1941 Node *PhaseCCP::transform_once( Node *n ) { 1942 const Type *t = type(n); 1943 // Constant? Use constant Node instead 1944 if( t->singleton() ) { 1945 Node *nn = n; // Default is to return the original constant 1946 if( t == Type::TOP ) { 1947 // cache my top node on the Compile instance 1948 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1949 C->set_cached_top_node(ConNode::make(Type::TOP)); 1950 set_type(C->top(), Type::TOP); 1951 } 1952 nn = C->top(); 1953 } 1954 if( !n->is_Con() ) { 1955 if( t != Type::TOP ) { 1956 nn = makecon(t); // ConNode::make(t); 1957 NOT_PRODUCT( inc_constants(); ) 1958 } else if( n->is_Region() ) { // Unreachable region 1959 // Note: nn == C->top() 1960 n->set_req(0, NULL); // Cut selfreference 1961 bool progress = true; 1962 uint max = n->outcnt(); 1963 DUIterator i; 1964 while (progress) { 1965 progress = false; 1966 // Eagerly remove dead phis to avoid phis copies creation. 1967 for (i = n->outs(); n->has_out(i); i++) { 1968 Node* m = n->out(i); 1969 if (m->is_Phi()) { 1970 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1971 replace_node(m, nn); 1972 if (max != n->outcnt()) { 1973 progress = true; 1974 i = n->refresh_out_pos(i); 1975 max = n->outcnt(); 1976 } 1977 } 1978 } 1979 } 1980 } 1981 replace_node(n,nn); // Update DefUse edges for new constant 1982 } 1983 return nn; 1984 } 1985 1986 // If x is a TypeNode, capture any more-precise type permanently into Node 1987 if (t != n->bottom_type()) { 1988 hash_delete(n); // changing bottom type may force a rehash 1989 n->raise_bottom_type(t); 1990 _worklist.push(n); // n re-enters the hash table via the worklist 1991 } 1992 1993 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1994 switch( n->Opcode() ) { 1995 case Op_FastLock: // Revisit FastLocks for lock coarsening 1996 case Op_If: 1997 case Op_CountedLoopEnd: 1998 case Op_Region: 1999 case Op_Loop: 2000 case Op_CountedLoop: 2001 case Op_Conv2B: 2002 case Op_Opaque1: 2003 case Op_Opaque2: 2004 _worklist.push(n); 2005 break; 2006 default: 2007 break; 2008 } 2009 2010 return n; 2011 } 2012 2013 //---------------------------------saturate------------------------------------ 2014 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 2015 const Type* limit_type) const { 2016 const Type* wide_type = new_type->widen(old_type, limit_type); 2017 if (wide_type != new_type) { // did we widen? 2018 // If so, we may have widened beyond the limit type. Clip it back down. 2019 new_type = wide_type->filter(limit_type); 2020 } 2021 return new_type; 2022 } 2023 2024 //------------------------------print_statistics------------------------------- 2025 #ifndef PRODUCT 2026 void PhaseCCP::print_statistics() { 2027 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 2028 } 2029 #endif 2030 2031 2032 //============================================================================= 2033 #ifndef PRODUCT 2034 uint PhasePeephole::_total_peepholes = 0; 2035 #endif 2036 //------------------------------PhasePeephole---------------------------------- 2037 // Conditional Constant Propagation, ala Wegman & Zadeck 2038 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 2039 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 2040 NOT_PRODUCT( clear_peepholes(); ) 2041 } 2042 2043 #ifndef PRODUCT 2044 //------------------------------~PhasePeephole--------------------------------- 2045 PhasePeephole::~PhasePeephole() { 2046 _total_peepholes += count_peepholes(); 2047 } 2048 #endif 2049 2050 //------------------------------transform-------------------------------------- 2051 Node *PhasePeephole::transform( Node *n ) { 2052 ShouldNotCallThis(); 2053 return NULL; 2054 } 2055 2056 //------------------------------do_transform----------------------------------- 2057 void PhasePeephole::do_transform() { 2058 bool method_name_not_printed = true; 2059 2060 // Examine each basic block 2061 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 2062 Block* block = _cfg.get_block(block_number); 2063 bool block_not_printed = true; 2064 2065 // and each instruction within a block 2066 uint end_index = block->number_of_nodes(); 2067 // block->end_idx() not valid after PhaseRegAlloc 2068 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 2069 Node *n = block->get_node(instruction_index); 2070 if( n->is_Mach() ) { 2071 MachNode *m = n->as_Mach(); 2072 int deleted_count = 0; 2073 // check for peephole opportunities 2074 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count); 2075 if( m2 != NULL ) { 2076 #ifndef PRODUCT 2077 if( PrintOptoPeephole ) { 2078 // Print method, first time only 2079 if( C->method() && method_name_not_printed ) { 2080 C->method()->print_short_name(); tty->cr(); 2081 method_name_not_printed = false; 2082 } 2083 // Print this block 2084 if( Verbose && block_not_printed) { 2085 tty->print_cr("in block"); 2086 block->dump(); 2087 block_not_printed = false; 2088 } 2089 // Print instructions being deleted 2090 for( int i = (deleted_count - 1); i >= 0; --i ) { 2091 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 2092 } 2093 tty->print_cr("replaced with"); 2094 // Print new instruction 2095 m2->format(_regalloc); 2096 tty->print("\n\n"); 2097 } 2098 #endif 2099 // Remove old nodes from basic block and update instruction_index 2100 // (old nodes still exist and may have edges pointing to them 2101 // as register allocation info is stored in the allocator using 2102 // the node index to live range mappings.) 2103 uint safe_instruction_index = (instruction_index - deleted_count); 2104 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 2105 block->remove_node( instruction_index ); 2106 } 2107 // install new node after safe_instruction_index 2108 block->insert_node(m2, safe_instruction_index + 1); 2109 end_index = block->number_of_nodes() - 1; // Recompute new block size 2110 NOT_PRODUCT( inc_peepholes(); ) 2111 } 2112 } 2113 } 2114 } 2115 } 2116 2117 //------------------------------print_statistics------------------------------- 2118 #ifndef PRODUCT 2119 void PhasePeephole::print_statistics() { 2120 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2121 } 2122 #endif 2123 2124 2125 //============================================================================= 2126 //------------------------------set_req_X-------------------------------------- 2127 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2128 assert( is_not_dead(n), "can not use dead node"); 2129 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2130 Node *old = in(i); 2131 set_req(i, n); 2132 2133 // old goes dead? 2134 if( old ) { 2135 switch (old->outcnt()) { 2136 case 0: 2137 // Put into the worklist to kill later. We do not kill it now because the 2138 // recursive kill will delete the current node (this) if dead-loop exists 2139 if (!old->is_top()) 2140 igvn->_worklist.push( old ); 2141 break; 2142 case 1: 2143 if( old->is_Store() || old->has_special_unique_user() ) 2144 igvn->add_users_to_worklist( old ); 2145 break; 2146 case 2: 2147 if( old->is_Store() ) 2148 igvn->add_users_to_worklist( old ); 2149 if( old->Opcode() == Op_Region ) 2150 igvn->_worklist.push(old); 2151 break; 2152 case 3: 2153 if( old->Opcode() == Op_Region ) { 2154 igvn->_worklist.push(old); 2155 igvn->add_users_to_worklist( old ); 2156 } 2157 break; 2158 default: 2159 break; 2160 } 2161 2162 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old); 2163 } 2164 2165 } 2166 2167 //-------------------------------replace_by----------------------------------- 2168 // Using def-use info, replace one node for another. Follow the def-use info 2169 // to all users of the OLD node. Then make all uses point to the NEW node. 2170 void Node::replace_by(Node *new_node) { 2171 assert(!is_top(), "top node has no DU info"); 2172 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2173 Node* use = last_out(i); 2174 uint uses_found = 0; 2175 for (uint j = 0; j < use->len(); j++) { 2176 if (use->in(j) == this) { 2177 if (j < use->req()) 2178 use->set_req(j, new_node); 2179 else use->set_prec(j, new_node); 2180 uses_found++; 2181 } 2182 } 2183 i -= uses_found; // we deleted 1 or more copies of this edge 2184 } 2185 } 2186 2187 //============================================================================= 2188 //----------------------------------------------------------------------------- 2189 void Type_Array::grow( uint i ) { 2190 if( !_max ) { 2191 _max = 1; 2192 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2193 _types[0] = NULL; 2194 } 2195 uint old = _max; 2196 while( i >= _max ) _max <<= 1; // Double to fit 2197 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2198 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2199 } 2200 2201 //------------------------------dump------------------------------------------- 2202 #ifndef PRODUCT 2203 void Type_Array::dump() const { 2204 uint max = Size(); 2205 for( uint i = 0; i < max; i++ ) { 2206 if( _types[i] != NULL ) { 2207 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2208 } 2209 } 2210 } 2211 #endif