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) { 465 466 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place"); 467 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes"); 468 assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point"); 469 470 uint old_unique_count = C->unique(); 471 uint live_node_count = C->live_nodes(); 472 uint worklist_size = worklist->size(); 473 474 // Storage for the updated type information. 475 Type_Array new_type_array(C->comp_arena()); 476 477 // Iterate over the set of live nodes. 478 uint current_idx = 0; // The current new node ID. Incremented after every assignment. 479 for (uint i = 0; i < _useful.size(); i++) { 480 Node* n = _useful.at(i); 481 // Sanity check that fails if we ever decide to execute this phase after EA 482 assert(!n->is_Phi() || n->as_Phi()->inst_mem_id() == -1, "should not be linked to data Phi"); 483 const Type* type = gvn->type_or_null(n); 484 new_type_array.map(current_idx, type); 485 486 bool in_worklist = false; 487 if (worklist->member(n)) { 488 in_worklist = true; 489 } 490 491 n->set_idx(current_idx); // Update node ID. 492 493 if (in_worklist) { 494 new_worklist->push(n); 495 } 496 497 current_idx++; 498 } 499 500 assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist"); 501 assert(live_node_count == current_idx, "all live nodes must be processed"); 502 503 // Replace the compiler's type information with the updated type information. 504 gvn->replace_types(new_type_array); 505 506 // Update the unique node count of the compilation to the number of currently live nodes. 507 C->set_unique(live_node_count); 508 509 // Set the dead node count to 0 and reset dead node list. 510 C->reset_dead_node_list(); 511 } 512 513 514 //============================================================================= 515 //------------------------------PhaseTransform--------------------------------- 516 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 517 _arena(Thread::current()->resource_area()), 518 _nodes(_arena), 519 _types(_arena) 520 { 521 init_con_caches(); 522 #ifndef PRODUCT 523 clear_progress(); 524 clear_transforms(); 525 set_allow_progress(true); 526 #endif 527 // Force allocation for currently existing nodes 528 _types.map(C->unique(), NULL); 529 } 530 531 //------------------------------PhaseTransform--------------------------------- 532 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 533 _arena(arena), 534 _nodes(arena), 535 _types(arena) 536 { 537 init_con_caches(); 538 #ifndef PRODUCT 539 clear_progress(); 540 clear_transforms(); 541 set_allow_progress(true); 542 #endif 543 // Force allocation for currently existing nodes 544 _types.map(C->unique(), NULL); 545 } 546 547 //------------------------------PhaseTransform--------------------------------- 548 // Initialize with previously generated type information 549 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 550 _arena(pt->_arena), 551 _nodes(pt->_nodes), 552 _types(pt->_types) 553 { 554 init_con_caches(); 555 #ifndef PRODUCT 556 clear_progress(); 557 clear_transforms(); 558 set_allow_progress(true); 559 #endif 560 } 561 562 void PhaseTransform::init_con_caches() { 563 memset(_icons,0,sizeof(_icons)); 564 memset(_lcons,0,sizeof(_lcons)); 565 memset(_zcons,0,sizeof(_zcons)); 566 } 567 568 569 //--------------------------------find_int_type-------------------------------- 570 const TypeInt* PhaseTransform::find_int_type(Node* n) { 571 if (n == NULL) return NULL; 572 // Call type_or_null(n) to determine node's type since we might be in 573 // parse phase and call n->Value() may return wrong type. 574 // (For example, a phi node at the beginning of loop parsing is not ready.) 575 const Type* t = type_or_null(n); 576 if (t == NULL) return NULL; 577 return t->isa_int(); 578 } 579 580 581 //-------------------------------find_long_type-------------------------------- 582 const TypeLong* PhaseTransform::find_long_type(Node* n) { 583 if (n == NULL) return NULL; 584 // (See comment above on type_or_null.) 585 const Type* t = type_or_null(n); 586 if (t == NULL) return NULL; 587 return t->isa_long(); 588 } 589 590 591 #ifndef PRODUCT 592 void PhaseTransform::dump_old2new_map() const { 593 _nodes.dump(); 594 } 595 596 void PhaseTransform::dump_new( uint nidx ) const { 597 for( uint i=0; i<_nodes.Size(); i++ ) 598 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 599 _nodes[i]->dump(); 600 tty->cr(); 601 tty->print_cr("Old index= %d",i); 602 return; 603 } 604 tty->print_cr("Node %d not found in the new indices", nidx); 605 } 606 607 //------------------------------dump_types------------------------------------- 608 void PhaseTransform::dump_types( ) const { 609 _types.dump(); 610 } 611 612 //------------------------------dump_nodes_and_types--------------------------- 613 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 614 VectorSet visited(Thread::current()->resource_area()); 615 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 616 } 617 618 //------------------------------dump_nodes_and_types_recur--------------------- 619 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 620 if( !n ) return; 621 if( depth == 0 ) return; 622 if( visited.test_set(n->_idx) ) return; 623 for( uint i=0; i<n->len(); i++ ) { 624 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 625 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 626 } 627 n->dump(); 628 if (type_or_null(n) != NULL) { 629 tty->print(" "); type(n)->dump(); tty->cr(); 630 } 631 } 632 633 #endif 634 635 636 //============================================================================= 637 //------------------------------PhaseValues------------------------------------ 638 // Set minimum table size to "255" 639 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 640 NOT_PRODUCT( clear_new_values(); ) 641 } 642 643 //------------------------------PhaseValues------------------------------------ 644 // Set minimum table size to "255" 645 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 646 _table(&ptv->_table) { 647 NOT_PRODUCT( clear_new_values(); ) 648 } 649 650 //------------------------------PhaseValues------------------------------------ 651 // Used by +VerifyOpto. Clear out hash table but copy _types array. 652 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 653 _table(ptv->arena(),ptv->_table.size()) { 654 NOT_PRODUCT( clear_new_values(); ) 655 } 656 657 //------------------------------~PhaseValues----------------------------------- 658 #ifndef PRODUCT 659 PhaseValues::~PhaseValues() { 660 _table.dump(); 661 662 // Statistics for value progress and efficiency 663 if( PrintCompilation && Verbose && WizardMode ) { 664 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 665 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 666 if( made_transforms() != 0 ) { 667 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 668 } else { 669 tty->cr(); 670 } 671 } 672 } 673 #endif 674 675 //------------------------------makecon---------------------------------------- 676 ConNode* PhaseTransform::makecon(const Type *t) { 677 assert(t->singleton(), "must be a constant"); 678 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 679 switch (t->base()) { // fast paths 680 case Type::Half: 681 case Type::Top: return (ConNode*) C->top(); 682 case Type::Int: return intcon( t->is_int()->get_con() ); 683 case Type::Long: return longcon( t->is_long()->get_con() ); 684 default: break; 685 } 686 if (t->is_zero_type()) 687 return zerocon(t->basic_type()); 688 return uncached_makecon(t); 689 } 690 691 //--------------------------uncached_makecon----------------------------------- 692 // Make an idealized constant - one of ConINode, ConPNode, etc. 693 ConNode* PhaseValues::uncached_makecon(const Type *t) { 694 assert(t->singleton(), "must be a constant"); 695 ConNode* x = ConNode::make(t); 696 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 697 if (k == NULL) { 698 set_type(x, t); // Missed, provide type mapping 699 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 700 if (nna != NULL) { 701 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 702 loc->clear(); // do not put debug info on constants 703 } 704 } else { 705 x->destruct(); // Hit, destroy duplicate constant 706 x = k; // use existing constant 707 } 708 return x; 709 } 710 711 //------------------------------intcon----------------------------------------- 712 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 713 ConINode* PhaseTransform::intcon(jint i) { 714 // Small integer? Check cache! Check that cached node is not dead 715 if (i >= _icon_min && i <= _icon_max) { 716 ConINode* icon = _icons[i-_icon_min]; 717 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 718 return icon; 719 } 720 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 721 assert(icon->is_Con(), ""); 722 if (i >= _icon_min && i <= _icon_max) 723 _icons[i-_icon_min] = icon; // Cache small integers 724 return icon; 725 } 726 727 //------------------------------longcon---------------------------------------- 728 // Fast long constant. 729 ConLNode* PhaseTransform::longcon(jlong l) { 730 // Small integer? Check cache! Check that cached node is not dead 731 if (l >= _lcon_min && l <= _lcon_max) { 732 ConLNode* lcon = _lcons[l-_lcon_min]; 733 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 734 return lcon; 735 } 736 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 737 assert(lcon->is_Con(), ""); 738 if (l >= _lcon_min && l <= _lcon_max) 739 _lcons[l-_lcon_min] = lcon; // Cache small integers 740 return lcon; 741 } 742 743 //------------------------------zerocon----------------------------------------- 744 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 745 ConNode* PhaseTransform::zerocon(BasicType bt) { 746 assert((uint)bt <= _zcon_max, "domain check"); 747 ConNode* zcon = _zcons[bt]; 748 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 749 return zcon; 750 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 751 _zcons[bt] = zcon; 752 return zcon; 753 } 754 755 756 757 //============================================================================= 758 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) { 759 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape); 760 if (i == NULL) { 761 i = k->Ideal(this, can_reshape); 762 } 763 return i; 764 } 765 766 Node* PhaseGVN::apply_identity(Node* k) { 767 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->identity_node(this, k); 768 if (i == k) { 769 i = k->Identity(this); 770 } 771 return i; 772 } 773 774 //------------------------------transform-------------------------------------- 775 // Return a node which computes the same function as this node, but in a 776 // faster or cheaper fashion. 777 Node *PhaseGVN::transform( Node *n ) { 778 return transform_no_reclaim(n); 779 } 780 781 //------------------------------transform-------------------------------------- 782 // Return a node which computes the same function as this node, but 783 // in a faster or cheaper fashion. 784 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 785 NOT_PRODUCT( set_transforms(); ) 786 787 // Apply the Ideal call in a loop until it no longer applies 788 Node *k = n; 789 NOT_PRODUCT( uint loop_count = 0; ) 790 while( 1 ) { 791 Node *i = apply_ideal(k, /*can_reshape=*/false); 792 if( !i ) break; 793 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 794 k = i; 795 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 796 } 797 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 798 799 800 // If brand new node, make space in type array. 801 ensure_type_or_null(k); 802 803 // Since I just called 'Value' to compute the set of run-time values 804 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 805 // cache Value. Later requests for the local phase->type of this Node can 806 // use the cached Value instead of suffering with 'bottom_type'. 807 const Type *t = k->Value(this); // Get runtime Value set 808 assert(t != NULL, "value sanity"); 809 if (type_or_null(k) != t) { 810 #ifndef PRODUCT 811 // Do not count initial visit to node as a transformation 812 if (type_or_null(k) == NULL) { 813 inc_new_values(); 814 set_progress(); 815 } 816 #endif 817 set_type(k, t); 818 // If k is a TypeNode, capture any more-precise type permanently into Node 819 k->raise_bottom_type(t); 820 } 821 822 if( t->singleton() && !k->is_Con() ) { 823 NOT_PRODUCT( set_progress(); ) 824 return makecon(t); // Turn into a constant 825 } 826 827 // Now check for Identities 828 Node *i = apply_identity(k); // Look for a nearby replacement 829 if( i != k ) { // Found? Return replacement! 830 NOT_PRODUCT( set_progress(); ) 831 return i; 832 } 833 834 // Global Value Numbering 835 i = hash_find_insert(k); // Insert if new 836 if( i && (i != k) ) { 837 // Return the pre-existing node 838 NOT_PRODUCT( set_progress(); ) 839 return i; 840 } 841 842 // Return Idealized original 843 return k; 844 } 845 846 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 847 if (d->is_top() || n->is_top()) { 848 return false; 849 } 850 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 851 int i = 0; 852 while (d != n) { 853 n = IfNode::up_one_dom(n, linear_only); 854 i++; 855 if (n == NULL || i >= 10) { 856 return false; 857 } 858 } 859 return true; 860 } 861 862 #ifdef ASSERT 863 //------------------------------dead_loop_check-------------------------------- 864 // Check for a simple dead loop when a data node references itself directly 865 // or through an other data node excluding cons and phis. 866 void PhaseGVN::dead_loop_check( Node *n ) { 867 // Phi may reference itself in a loop 868 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 869 // Do 2 levels check and only data inputs. 870 bool no_dead_loop = true; 871 uint cnt = n->req(); 872 for (uint i = 1; i < cnt && no_dead_loop; i++) { 873 Node *in = n->in(i); 874 if (in == n) { 875 no_dead_loop = false; 876 } else if (in != NULL && !in->is_dead_loop_safe()) { 877 uint icnt = in->req(); 878 for (uint j = 1; j < icnt && no_dead_loop; j++) { 879 if (in->in(j) == n || in->in(j) == in) 880 no_dead_loop = false; 881 } 882 } 883 } 884 if (!no_dead_loop) n->dump(3); 885 assert(no_dead_loop, "dead loop detected"); 886 } 887 } 888 #endif 889 890 //============================================================================= 891 //------------------------------PhaseIterGVN----------------------------------- 892 // Initialize hash table to fresh and clean for +VerifyOpto 893 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), 894 _delay_transform(false), 895 _stack(C->live_nodes() >> 1), 896 _worklist( ) { 897 } 898 899 //------------------------------PhaseIterGVN----------------------------------- 900 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 901 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 902 _delay_transform(igvn->_delay_transform), 903 _stack( igvn->_stack ), 904 _worklist( igvn->_worklist ) 905 { 906 } 907 908 //------------------------------PhaseIterGVN----------------------------------- 909 // Initialize with previous PhaseGVN info from Parser 910 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 911 _delay_transform(false), 912 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that 913 // the constructor is used in incremental inlining, this consumes too much memory: 914 // _stack(C->live_nodes() >> 1), 915 // So, as a band-aid, we replace this by: 916 _stack(C->comp_arena(), 32), 917 _worklist(*C->for_igvn()) 918 { 919 uint max; 920 921 // Dead nodes in the hash table inherited from GVN were not treated as 922 // roots during def-use info creation; hence they represent an invisible 923 // use. Clear them out. 924 max = _table.size(); 925 for( uint i = 0; i < max; ++i ) { 926 Node *n = _table.at(i); 927 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 928 if( n->is_top() ) continue; 929 assert( false, "Parse::remove_useless_nodes missed this node"); 930 hash_delete(n); 931 } 932 } 933 934 // Any Phis or Regions on the worklist probably had uses that could not 935 // make more progress because the uses were made while the Phis and Regions 936 // were in half-built states. Put all uses of Phis and Regions on worklist. 937 max = _worklist.size(); 938 for( uint j = 0; j < max; j++ ) { 939 Node *n = _worklist.at(j); 940 uint uop = n->Opcode(); 941 if( uop == Op_Phi || uop == Op_Region || 942 n->is_Type() || 943 n->is_Mem() ) 944 add_users_to_worklist(n); 945 } 946 947 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 948 bs->add_users_to_worklist(&_worklist); 949 } 950 951 /** 952 * Initialize worklist for each node. 953 */ 954 void PhaseIterGVN::init_worklist(Node* first) { 955 Unique_Node_List to_process; 956 to_process.push(first); 957 958 while (to_process.size() > 0) { 959 Node* n = to_process.pop(); 960 if (!_worklist.member(n)) { 961 _worklist.push(n); 962 963 uint cnt = n->req(); 964 for(uint i = 0; i < cnt; i++) { 965 Node* m = n->in(i); 966 if (m != NULL) { 967 to_process.push(m); 968 } 969 } 970 } 971 } 972 } 973 974 #ifndef PRODUCT 975 void PhaseIterGVN::verify_step(Node* n) { 976 if (VerifyIterativeGVN) { 977 _verify_window[_verify_counter % _verify_window_size] = n; 978 ++_verify_counter; 979 ResourceMark rm; 980 ResourceArea* area = Thread::current()->resource_area(); 981 VectorSet old_space(area), new_space(area); 982 if (C->unique() < 1000 || 983 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 984 ++_verify_full_passes; 985 Node::verify_recur(C->root(), -1, old_space, new_space); 986 } 987 const int verify_depth = 4; 988 for ( int i = 0; i < _verify_window_size; i++ ) { 989 Node* n = _verify_window[i]; 990 if ( n == NULL ) continue; 991 if( n->in(0) == NodeSentinel ) { // xform_idom 992 _verify_window[i] = n->in(1); 993 --i; continue; 994 } 995 // Typical fanout is 1-2, so this call visits about 6 nodes. 996 Node::verify_recur(n, verify_depth, old_space, new_space); 997 } 998 } 999 } 1000 1001 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 1002 if (TraceIterativeGVN) { 1003 uint wlsize = _worklist.size(); 1004 const Type* newtype = type_or_null(n); 1005 if (nn != n) { 1006 // print old node 1007 tty->print("< "); 1008 if (oldtype != newtype && oldtype != NULL) { 1009 oldtype->dump(); 1010 } 1011 do { tty->print("\t"); } while (tty->position() < 16); 1012 tty->print("<"); 1013 n->dump(); 1014 } 1015 if (oldtype != newtype || nn != n) { 1016 // print new node and/or new type 1017 if (oldtype == NULL) { 1018 tty->print("* "); 1019 } else if (nn != n) { 1020 tty->print("> "); 1021 } else { 1022 tty->print("= "); 1023 } 1024 if (newtype == NULL) { 1025 tty->print("null"); 1026 } else { 1027 newtype->dump(); 1028 } 1029 do { tty->print("\t"); } while (tty->position() < 16); 1030 nn->dump(); 1031 } 1032 if (Verbose && wlsize < _worklist.size()) { 1033 tty->print(" Push {"); 1034 while (wlsize != _worklist.size()) { 1035 Node* pushed = _worklist.at(wlsize++); 1036 tty->print(" %d", pushed->_idx); 1037 } 1038 tty->print_cr(" }"); 1039 } 1040 if (nn != n) { 1041 // ignore n, it might be subsumed 1042 verify_step((Node*) NULL); 1043 } 1044 } 1045 } 1046 1047 void PhaseIterGVN::init_verifyPhaseIterGVN() { 1048 _verify_counter = 0; 1049 _verify_full_passes = 0; 1050 for (int i = 0; i < _verify_window_size; i++) { 1051 _verify_window[i] = NULL; 1052 } 1053 #ifdef ASSERT 1054 // Verify that all modified nodes are on _worklist 1055 Unique_Node_List* modified_list = C->modified_nodes(); 1056 while (modified_list != NULL && modified_list->size()) { 1057 Node* n = modified_list->pop(); 1058 if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) { 1059 n->dump(); 1060 assert(false, "modified node is not on IGVN._worklist"); 1061 } 1062 } 1063 #endif 1064 } 1065 1066 void PhaseIterGVN::verify_PhaseIterGVN() { 1067 #ifdef ASSERT 1068 // Verify nodes with changed inputs. 1069 Unique_Node_List* modified_list = C->modified_nodes(); 1070 while (modified_list != NULL && modified_list->size()) { 1071 Node* n = modified_list->pop(); 1072 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes 1073 n->dump(); 1074 assert(false, "modified node was not processed by IGVN.transform_old()"); 1075 } 1076 } 1077 #endif 1078 1079 C->verify_graph_edges(); 1080 if( VerifyOpto && allow_progress() ) { 1081 // Must turn off allow_progress to enable assert and break recursion 1082 C->root()->verify(); 1083 { // Check if any progress was missed using IterGVN 1084 // Def-Use info enables transformations not attempted in wash-pass 1085 // e.g. Region/Phi cleanup, ... 1086 // Null-check elision -- may not have reached fixpoint 1087 // do not propagate to dominated nodes 1088 ResourceMark rm; 1089 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 1090 // Fill worklist completely 1091 igvn2.init_worklist(C->root()); 1092 1093 igvn2.set_allow_progress(false); 1094 igvn2.optimize(); 1095 igvn2.set_allow_progress(true); 1096 } 1097 } 1098 if (VerifyIterativeGVN && PrintOpto) { 1099 if (_verify_counter == _verify_full_passes) { 1100 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 1101 (int) _verify_full_passes); 1102 } else { 1103 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 1104 (int) _verify_counter, (int) _verify_full_passes); 1105 } 1106 } 1107 1108 #ifdef ASSERT 1109 while (modified_list->size()) { 1110 Node* n = modified_list->pop(); 1111 n->dump(); 1112 assert(false, "VerifyIterativeGVN: new modified node was added"); 1113 } 1114 #endif 1115 } 1116 #endif /* PRODUCT */ 1117 1118 #ifdef ASSERT 1119 /** 1120 * Dumps information that can help to debug the problem. A debug 1121 * build fails with an assert. 1122 */ 1123 void PhaseIterGVN::dump_infinite_loop_info(Node* n) { 1124 n->dump(4); 1125 _worklist.dump(); 1126 assert(false, "infinite loop in PhaseIterGVN::optimize"); 1127 } 1128 1129 /** 1130 * Prints out information about IGVN if the 'verbose' option is used. 1131 */ 1132 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1133 if (TraceIterativeGVN && Verbose) { 1134 tty->print(" Pop "); 1135 n->dump(); 1136 if ((num_processed % 100) == 0) { 1137 _worklist.print_set(); 1138 } 1139 } 1140 } 1141 #endif /* ASSERT */ 1142 1143 void PhaseIterGVN::optimize() { 1144 DEBUG_ONLY(uint num_processed = 0;) 1145 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1146 1147 uint loop_count = 0; 1148 // Pull from worklist and transform the node. If the node has changed, 1149 // update edge info and put uses on worklist. 1150 while(_worklist.size()) { 1151 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1152 return; 1153 } 1154 Node* n = _worklist.pop(); 1155 if (++loop_count >= K * C->live_nodes()) { 1156 DEBUG_ONLY(dump_infinite_loop_info(n);) 1157 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1158 return; 1159 } 1160 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1161 if (n->outcnt() != 0) { 1162 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1163 // Do the transformation 1164 Node* nn = transform_old(n); 1165 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1166 } else if (!n->is_top()) { 1167 remove_dead_node(n); 1168 } 1169 } 1170 NOT_PRODUCT(verify_PhaseIterGVN();) 1171 } 1172 1173 1174 /** 1175 * Register a new node with the optimizer. Update the types array, the def-use 1176 * info. Put on worklist. 1177 */ 1178 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1179 set_type_bottom(n); 1180 _worklist.push(n); 1181 if (orig != NULL) C->copy_node_notes_to(n, orig); 1182 return n; 1183 } 1184 1185 //------------------------------transform-------------------------------------- 1186 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1187 Node *PhaseIterGVN::transform( Node *n ) { 1188 if (_delay_transform) { 1189 // Register the node but don't optimize for now 1190 register_new_node_with_optimizer(n); 1191 return n; 1192 } 1193 1194 // If brand new node, make space in type array, and give it a type. 1195 ensure_type_or_null(n); 1196 if (type_or_null(n) == NULL) { 1197 set_type_bottom(n); 1198 } 1199 1200 return transform_old(n); 1201 } 1202 1203 Node *PhaseIterGVN::transform_old(Node* n) { 1204 DEBUG_ONLY(uint loop_count = 0;); 1205 NOT_PRODUCT(set_transforms()); 1206 1207 // Remove 'n' from hash table in case it gets modified 1208 _table.hash_delete(n); 1209 if (VerifyIterativeGVN) { 1210 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1211 } 1212 1213 // Apply the Ideal call in a loop until it no longer applies 1214 Node* k = n; 1215 DEBUG_ONLY(dead_loop_check(k);) 1216 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1217 C->remove_modified_node(k); 1218 Node* i = apply_ideal(k, /*can_reshape=*/true); 1219 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1220 #ifndef PRODUCT 1221 verify_step(k); 1222 if (i && VerifyOpto ) { 1223 if (!allow_progress()) { 1224 if (i->is_Add() && (i->outcnt() == 1)) { 1225 // Switched input to left side because this is the only use 1226 } else if (i->is_If() && (i->in(0) == NULL)) { 1227 // This IF is dead because it is dominated by an equivalent IF When 1228 // dominating if changed, info is not propagated sparsely to 'this' 1229 // Propagating this info further will spuriously identify other 1230 // progress. 1231 return i; 1232 } else 1233 set_progress(); 1234 } else { 1235 set_progress(); 1236 } 1237 } 1238 #endif 1239 1240 while (i != NULL) { 1241 #ifdef ASSERT 1242 if (loop_count >= K) { 1243 dump_infinite_loop_info(i); 1244 } 1245 loop_count++; 1246 #endif 1247 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1248 // Made a change; put users of original Node on worklist 1249 add_users_to_worklist(k); 1250 // Replacing root of transform tree? 1251 if (k != i) { 1252 // Make users of old Node now use new. 1253 subsume_node(k, i); 1254 k = i; 1255 } 1256 DEBUG_ONLY(dead_loop_check(k);) 1257 // Try idealizing again 1258 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1259 C->remove_modified_node(k); 1260 i = apply_ideal(k, /*can_reshape=*/true); 1261 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1262 #ifndef PRODUCT 1263 verify_step(k); 1264 if (i && VerifyOpto) { 1265 set_progress(); 1266 } 1267 #endif 1268 } 1269 1270 // If brand new node, make space in type array. 1271 ensure_type_or_null(k); 1272 1273 // See what kind of values 'k' takes on at runtime 1274 const Type* t = k->Value(this); 1275 assert(t != NULL, "value sanity"); 1276 1277 // Since I just called 'Value' to compute the set of run-time values 1278 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1279 // cache Value. Later requests for the local phase->type of this Node can 1280 // use the cached Value instead of suffering with 'bottom_type'. 1281 if (type_or_null(k) != t) { 1282 #ifndef PRODUCT 1283 inc_new_values(); 1284 set_progress(); 1285 #endif 1286 set_type(k, t); 1287 // If k is a TypeNode, capture any more-precise type permanently into Node 1288 k->raise_bottom_type(t); 1289 // Move users of node to worklist 1290 add_users_to_worklist(k); 1291 } 1292 // If 'k' computes a constant, replace it with a constant 1293 if (t->singleton() && !k->is_Con()) { 1294 NOT_PRODUCT(set_progress();) 1295 Node* con = makecon(t); // Make a constant 1296 add_users_to_worklist(k); 1297 subsume_node(k, con); // Everybody using k now uses con 1298 return con; 1299 } 1300 1301 // Now check for Identities 1302 i = apply_identity(k); // Look for a nearby replacement 1303 if (i != k) { // Found? Return replacement! 1304 NOT_PRODUCT(set_progress();) 1305 add_users_to_worklist(k); 1306 subsume_node(k, i); // Everybody using k now uses i 1307 return i; 1308 } 1309 1310 // Global Value Numbering 1311 i = hash_find_insert(k); // Check for pre-existing node 1312 if (i && (i != k)) { 1313 // Return the pre-existing node if it isn't dead 1314 NOT_PRODUCT(set_progress();) 1315 add_users_to_worklist(k); 1316 subsume_node(k, i); // Everybody using k now uses i 1317 return i; 1318 } 1319 1320 // Return Idealized original 1321 return k; 1322 } 1323 1324 //---------------------------------saturate------------------------------------ 1325 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1326 const Type* limit_type) const { 1327 return new_type->narrow(old_type); 1328 } 1329 1330 //------------------------------remove_globally_dead_node---------------------- 1331 // Kill a globally dead Node. All uses are also globally dead and are 1332 // aggressively trimmed. 1333 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1334 enum DeleteProgress { 1335 PROCESS_INPUTS, 1336 PROCESS_OUTPUTS 1337 }; 1338 assert(_stack.is_empty(), "not empty"); 1339 _stack.push(dead, PROCESS_INPUTS); 1340 1341 while (_stack.is_nonempty()) { 1342 dead = _stack.node(); 1343 if (dead->Opcode() == Op_SafePoint) { 1344 dead->as_SafePoint()->disconnect_from_root(this); 1345 } 1346 uint progress_state = _stack.index(); 1347 assert(dead != C->root(), "killing root, eh?"); 1348 assert(!dead->is_top(), "add check for top when pushing"); 1349 NOT_PRODUCT( set_progress(); ) 1350 if (progress_state == PROCESS_INPUTS) { 1351 // After following inputs, continue to outputs 1352 _stack.set_index(PROCESS_OUTPUTS); 1353 if (!dead->is_Con()) { // Don't kill cons but uses 1354 bool recurse = false; 1355 // Remove from hash table 1356 _table.hash_delete( dead ); 1357 // Smash all inputs to 'dead', isolating him completely 1358 for (uint i = 0; i < dead->req(); i++) { 1359 Node *in = dead->in(i); 1360 if (in != NULL && in != C->top()) { // Points to something? 1361 int nrep = dead->replace_edge(in, NULL); // Kill edges 1362 assert((nrep > 0), "sanity"); 1363 if (in->outcnt() == 0) { // Made input go dead? 1364 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1365 recurse = true; 1366 } else if (in->outcnt() == 1 && 1367 in->has_special_unique_user()) { 1368 _worklist.push(in->unique_out()); 1369 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1370 if (in->Opcode() == Op_Region) { 1371 _worklist.push(in); 1372 } else if (in->is_Store()) { 1373 DUIterator_Fast imax, i = in->fast_outs(imax); 1374 _worklist.push(in->fast_out(i)); 1375 i++; 1376 if (in->outcnt() == 2) { 1377 _worklist.push(in->fast_out(i)); 1378 i++; 1379 } 1380 assert(!(i < imax), "sanity"); 1381 } 1382 } else { 1383 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in); 1384 } 1385 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1386 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1387 // A Load that directly follows an InitializeNode is 1388 // going away. The Stores that follow are candidates 1389 // again to be captured by the InitializeNode. 1390 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1391 Node *n = in->fast_out(j); 1392 if (n->is_Store()) { 1393 _worklist.push(n); 1394 } 1395 } 1396 } 1397 } // if (in != NULL && in != C->top()) 1398 } // for (uint i = 0; i < dead->req(); i++) 1399 if (recurse) { 1400 continue; 1401 } 1402 } // if (!dead->is_Con()) 1403 } // if (progress_state == PROCESS_INPUTS) 1404 1405 // Aggressively kill globally dead uses 1406 // (Rather than pushing all the outs at once, we push one at a time, 1407 // plus the parent to resume later, because of the indefinite number 1408 // of edge deletions per loop trip.) 1409 if (dead->outcnt() > 0) { 1410 // Recursively remove output edges 1411 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1412 } else { 1413 // Finished disconnecting all input and output edges. 1414 _stack.pop(); 1415 // Remove dead node from iterative worklist 1416 _worklist.remove(dead); 1417 C->remove_modified_node(dead); 1418 // Constant node that has no out-edges and has only one in-edge from 1419 // root is usually dead. However, sometimes reshaping walk makes 1420 // it reachable by adding use edges. So, we will NOT count Con nodes 1421 // as dead to be conservative about the dead node count at any 1422 // given time. 1423 if (!dead->is_Con()) { 1424 C->record_dead_node(dead->_idx); 1425 } 1426 if (dead->is_macro()) { 1427 C->remove_macro_node(dead); 1428 } 1429 if (dead->is_expensive()) { 1430 C->remove_expensive_node(dead); 1431 } 1432 CastIINode* cast = dead->isa_CastII(); 1433 if (cast != NULL && cast->has_range_check()) { 1434 C->remove_range_check_cast(cast); 1435 } 1436 if (dead->Opcode() == Op_Opaque4) { 1437 C->remove_opaque4_node(dead); 1438 } 1439 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1440 bs->unregister_potential_barrier_node(dead); 1441 } 1442 } // while (_stack.is_nonempty()) 1443 } 1444 1445 //------------------------------subsume_node----------------------------------- 1446 // Remove users from node 'old' and add them to node 'nn'. 1447 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1448 if (old->Opcode() == Op_SafePoint) { 1449 old->as_SafePoint()->disconnect_from_root(this); 1450 } 1451 assert( old != hash_find(old), "should already been removed" ); 1452 assert( old != C->top(), "cannot subsume top node"); 1453 // Copy debug or profile information to the new version: 1454 C->copy_node_notes_to(nn, old); 1455 // Move users of node 'old' to node 'nn' 1456 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1457 Node* use = old->last_out(i); // for each use... 1458 // use might need re-hashing (but it won't if it's a new node) 1459 rehash_node_delayed(use); 1460 // Update use-def info as well 1461 // We remove all occurrences of old within use->in, 1462 // so as to avoid rehashing any node more than once. 1463 // The hash table probe swamps any outer loop overhead. 1464 uint num_edges = 0; 1465 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1466 if (use->in(j) == old) { 1467 use->set_req(j, nn); 1468 ++num_edges; 1469 } 1470 } 1471 i -= num_edges; // we deleted 1 or more copies of this edge 1472 } 1473 1474 // Search for instance field data PhiNodes in the same region pointing to the old 1475 // memory PhiNode and update their instance memory ids to point to the new node. 1476 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) { 1477 Node* region = old->in(0); 1478 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1479 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1480 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) { 1481 phi->set_inst_mem_id((int)nn->_idx); 1482 } 1483 } 1484 } 1485 1486 // Smash all inputs to 'old', isolating him completely 1487 Node *temp = new Node(1); 1488 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1489 remove_dead_node( old ); 1490 temp->del_req(0); // Yank bogus edge 1491 #ifndef PRODUCT 1492 if( VerifyIterativeGVN ) { 1493 for ( int i = 0; i < _verify_window_size; i++ ) { 1494 if ( _verify_window[i] == old ) 1495 _verify_window[i] = nn; 1496 } 1497 } 1498 #endif 1499 _worklist.remove(temp); // this can be necessary 1500 temp->destruct(); // reuse the _idx of this little guy 1501 } 1502 1503 //------------------------------add_users_to_worklist-------------------------- 1504 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1505 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1506 _worklist.push(n->fast_out(i)); // Push on worklist 1507 } 1508 } 1509 1510 // Return counted loop Phi if as a counted loop exit condition, cmp 1511 // compares the the induction variable with n 1512 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) { 1513 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1514 Node* bol = cmp->fast_out(i); 1515 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1516 Node* iff = bol->fast_out(i2); 1517 if (iff->is_CountedLoopEnd()) { 1518 CountedLoopEndNode* cle = iff->as_CountedLoopEnd(); 1519 if (cle->limit() == n) { 1520 PhiNode* phi = cle->phi(); 1521 if (phi != NULL) { 1522 return phi; 1523 } 1524 } 1525 } 1526 } 1527 } 1528 return NULL; 1529 } 1530 1531 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1532 add_users_to_worklist0(n); 1533 1534 // Move users of node to worklist 1535 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1536 Node* use = n->fast_out(i); // Get use 1537 1538 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1539 use->is_Store() ) // Enable store/load same address 1540 add_users_to_worklist0(use); 1541 1542 // If we changed the receiver type to a call, we need to revisit 1543 // the Catch following the call. It's looking for a non-NULL 1544 // receiver to know when to enable the regular fall-through path 1545 // in addition to the NullPtrException path. 1546 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1547 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control); 1548 if (p != NULL) { 1549 add_users_to_worklist0(p); 1550 } 1551 } 1552 1553 uint use_op = use->Opcode(); 1554 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1555 add_users_to_worklist(use); // Put Bool on worklist 1556 if (use->outcnt() > 0) { 1557 Node* bol = use->raw_out(0); 1558 if (bol->outcnt() > 0) { 1559 Node* iff = bol->raw_out(0); 1560 if (iff->outcnt() == 2) { 1561 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1562 // phi merging either 0 or 1 onto the worklist 1563 Node* ifproj0 = iff->raw_out(0); 1564 Node* ifproj1 = iff->raw_out(1); 1565 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1566 Node* region0 = ifproj0->raw_out(0); 1567 Node* region1 = ifproj1->raw_out(0); 1568 if( region0 == region1 ) 1569 add_users_to_worklist0(region0); 1570 } 1571 } 1572 } 1573 } 1574 if (use_op == Op_CmpI) { 1575 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n); 1576 if (phi != NULL) { 1577 // If an opaque node feeds into the limit condition of a 1578 // CountedLoop, we need to process the Phi node for the 1579 // induction variable when the opaque node is removed: 1580 // the range of values taken by the Phi is now known and 1581 // so its type is also known. 1582 _worklist.push(phi); 1583 } 1584 Node* in1 = use->in(1); 1585 for (uint i = 0; i < in1->outcnt(); i++) { 1586 if (in1->raw_out(i)->Opcode() == Op_CastII) { 1587 Node* castii = in1->raw_out(i); 1588 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) { 1589 Node* ifnode = castii->in(0)->in(0); 1590 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) { 1591 // Reprocess a CastII node that may depend on an 1592 // opaque node value when the opaque node is 1593 // removed. In case it carries a dependency we can do 1594 // a better job of computing its type. 1595 _worklist.push(castii); 1596 } 1597 } 1598 } 1599 } 1600 } 1601 } 1602 1603 // If changed Cast input, check Phi users for simple cycles 1604 if (use->is_ConstraintCast()) { 1605 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1606 Node* u = use->fast_out(i2); 1607 if (u->is_Phi()) 1608 _worklist.push(u); 1609 } 1610 } 1611 // If changed LShift inputs, check RShift users for useless sign-ext 1612 if( use_op == Op_LShiftI ) { 1613 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1614 Node* u = use->fast_out(i2); 1615 if (u->Opcode() == Op_RShiftI) 1616 _worklist.push(u); 1617 } 1618 } 1619 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1620 if (use_op == Op_AddI || use_op == Op_SubI) { 1621 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1622 Node* u = use->fast_out(i2); 1623 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1624 _worklist.push(u); 1625 } 1626 } 1627 } 1628 // If changed AddP inputs, check Stores for loop invariant 1629 if( use_op == Op_AddP ) { 1630 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1631 Node* u = use->fast_out(i2); 1632 if (u->is_Mem()) 1633 _worklist.push(u); 1634 } 1635 } 1636 // If changed initialization activity, check dependent Stores 1637 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1638 InitializeNode* init = use->as_Allocate()->initialization(); 1639 if (init != NULL) { 1640 Node* imem = init->proj_out_or_null(TypeFunc::Memory); 1641 if (imem != NULL) add_users_to_worklist0(imem); 1642 } 1643 } 1644 if (use_op == Op_Initialize) { 1645 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory); 1646 if (imem != NULL) add_users_to_worklist0(imem); 1647 } 1648 // Loading the java mirror from a Klass requires two loads and the type 1649 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1650 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1651 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1652 bool has_load_barriers = bs->has_load_barriers(); 1653 1654 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) { 1655 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1656 Node* u = use->fast_out(i2); 1657 const Type* ut = u->bottom_type(); 1658 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) { 1659 if (has_load_barriers) { 1660 // Search for load barriers behind the load 1661 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1662 Node* b = u->fast_out(i3); 1663 if (bs->is_gc_barrier_node(b)) { 1664 _worklist.push(b); 1665 } 1666 } 1667 } 1668 _worklist.push(u); 1669 } 1670 } 1671 } 1672 1673 BarrierSet::barrier_set()->barrier_set_c2()->igvn_add_users_to_worklist(this, use); 1674 } 1675 } 1676 1677 /** 1678 * Remove the speculative part of all types that we know of 1679 */ 1680 void PhaseIterGVN::remove_speculative_types() { 1681 assert(UseTypeSpeculation, "speculation is off"); 1682 for (uint i = 0; i < _types.Size(); i++) { 1683 const Type* t = _types.fast_lookup(i); 1684 if (t != NULL) { 1685 _types.map(i, t->remove_speculative()); 1686 } 1687 } 1688 _table.check_no_speculative_types(); 1689 } 1690 1691 //============================================================================= 1692 #ifndef PRODUCT 1693 uint PhaseCCP::_total_invokes = 0; 1694 uint PhaseCCP::_total_constants = 0; 1695 #endif 1696 //------------------------------PhaseCCP--------------------------------------- 1697 // Conditional Constant Propagation, ala Wegman & Zadeck 1698 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1699 NOT_PRODUCT( clear_constants(); ) 1700 assert( _worklist.size() == 0, "" ); 1701 // Clear out _nodes from IterGVN. Must be clear to transform call. 1702 _nodes.clear(); // Clear out from IterGVN 1703 analyze(); 1704 } 1705 1706 #ifndef PRODUCT 1707 //------------------------------~PhaseCCP-------------------------------------- 1708 PhaseCCP::~PhaseCCP() { 1709 inc_invokes(); 1710 _total_constants += count_constants(); 1711 } 1712 #endif 1713 1714 1715 #ifdef ASSERT 1716 static bool ccp_type_widens(const Type* t, const Type* t0) { 1717 assert(t->meet(t0) == t, "Not monotonic"); 1718 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1719 case Type::Int: 1720 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1721 break; 1722 case Type::Long: 1723 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1724 break; 1725 default: 1726 break; 1727 } 1728 return true; 1729 } 1730 #endif //ASSERT 1731 1732 //------------------------------analyze---------------------------------------- 1733 void PhaseCCP::analyze() { 1734 // Initialize all types to TOP, optimistic analysis 1735 for (int i = C->unique() - 1; i >= 0; i--) { 1736 _types.map(i,Type::TOP); 1737 } 1738 1739 // Push root onto worklist 1740 Unique_Node_List worklist; 1741 worklist.push(C->root()); 1742 1743 // Pull from worklist; compute new value; push changes out. 1744 // This loop is the meat of CCP. 1745 while( worklist.size() ) { 1746 Node *n = worklist.pop(); 1747 const Type *t = n->Value(this); 1748 if (t != type(n)) { 1749 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1750 #ifndef PRODUCT 1751 if( TracePhaseCCP ) { 1752 t->dump(); 1753 do { tty->print("\t"); } while (tty->position() < 16); 1754 n->dump(); 1755 } 1756 #endif 1757 set_type(n, t); 1758 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1759 Node* m = n->fast_out(i); // Get user 1760 if (m->is_Region()) { // New path to Region? Must recheck Phis too 1761 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1762 Node* p = m->fast_out(i2); // Propagate changes to uses 1763 if (p->bottom_type() != type(p)) { // If not already bottomed out 1764 worklist.push(p); // Propagate change to user 1765 } 1766 } 1767 } 1768 // If we changed the receiver type to a call, we need to revisit 1769 // the Catch following the call. It's looking for a non-NULL 1770 // receiver to know when to enable the regular fall-through path 1771 // in addition to the NullPtrException path 1772 if (m->is_Call()) { 1773 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1774 Node* p = m->fast_out(i2); // Propagate changes to uses 1775 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) { 1776 worklist.push(p->unique_out()); 1777 } 1778 } 1779 } 1780 if (m->bottom_type() != type(m)) { // If not already bottomed out 1781 worklist.push(m); // Propagate change to user 1782 } 1783 1784 // CmpU nodes can get their type information from two nodes up in the 1785 // graph (instead of from the nodes immediately above). Make sure they 1786 // are added to the worklist if nodes they depend on are updated, since 1787 // they could be missed and get wrong types otherwise. 1788 uint m_op = m->Opcode(); 1789 if (m_op == Op_AddI || m_op == Op_SubI) { 1790 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1791 Node* p = m->fast_out(i2); // Propagate changes to uses 1792 if (p->Opcode() == Op_CmpU) { 1793 // Got a CmpU which might need the new type information from node n. 1794 if(p->bottom_type() != type(p)) { // If not already bottomed out 1795 worklist.push(p); // Propagate change to user 1796 } 1797 } 1798 } 1799 } 1800 // If n is used in a counted loop exit condition then the type 1801 // of the counted loop's Phi depends on the type of n. See 1802 // PhiNode::Value(). 1803 if (m_op == Op_CmpI) { 1804 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n); 1805 if (phi != NULL) { 1806 worklist.push(phi); 1807 } 1808 } 1809 // Loading the java mirror from a Klass requires two loads and the type 1810 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1811 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1812 bool has_load_barriers = bs->has_load_barriers(); 1813 1814 if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) { 1815 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1816 Node* u = m->fast_out(i2); 1817 const Type* ut = u->bottom_type(); 1818 if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) { 1819 if (has_load_barriers) { 1820 // Search for load barriers behind the load 1821 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1822 Node* b = u->fast_out(i3); 1823 if (bs->is_gc_barrier_node(b)) { 1824 worklist.push(b); 1825 } 1826 } 1827 } 1828 worklist.push(u); 1829 } 1830 } 1831 } 1832 1833 BarrierSet::barrier_set()->barrier_set_c2()->ccp_analyze(this, worklist, m); 1834 } 1835 } 1836 } 1837 } 1838 1839 //------------------------------do_transform----------------------------------- 1840 // Top level driver for the recursive transformer 1841 void PhaseCCP::do_transform() { 1842 // Correct leaves of new-space Nodes; they point to old-space. 1843 C->set_root( transform(C->root())->as_Root() ); 1844 assert( C->top(), "missing TOP node" ); 1845 assert( C->root(), "missing root" ); 1846 } 1847 1848 //------------------------------transform-------------------------------------- 1849 // Given a Node in old-space, clone him into new-space. 1850 // Convert any of his old-space children into new-space children. 1851 Node *PhaseCCP::transform( Node *n ) { 1852 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1853 if( new_node != NULL ) 1854 return new_node; // Been there, done that, return old answer 1855 new_node = transform_once(n); // Check for constant 1856 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1857 1858 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1859 GrowableArray <Node *> trstack(C->live_nodes() >> 1); 1860 1861 trstack.push(new_node); // Process children of cloned node 1862 while ( trstack.is_nonempty() ) { 1863 Node *clone = trstack.pop(); 1864 uint cnt = clone->req(); 1865 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1866 Node *input = clone->in(i); 1867 if( input != NULL ) { // Ignore NULLs 1868 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1869 if( new_input == NULL ) { 1870 new_input = transform_once(input); // Check for constant 1871 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1872 trstack.push(new_input); 1873 } 1874 assert( new_input == clone->in(i), "insanity check"); 1875 } 1876 } 1877 } 1878 return new_node; 1879 } 1880 1881 1882 //------------------------------transform_once--------------------------------- 1883 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1884 Node *PhaseCCP::transform_once( Node *n ) { 1885 const Type *t = type(n); 1886 // Constant? Use constant Node instead 1887 if( t->singleton() ) { 1888 Node *nn = n; // Default is to return the original constant 1889 if( t == Type::TOP ) { 1890 // cache my top node on the Compile instance 1891 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1892 C->set_cached_top_node(ConNode::make(Type::TOP)); 1893 set_type(C->top(), Type::TOP); 1894 } 1895 nn = C->top(); 1896 } 1897 if( !n->is_Con() ) { 1898 if( t != Type::TOP ) { 1899 nn = makecon(t); // ConNode::make(t); 1900 NOT_PRODUCT( inc_constants(); ) 1901 } else if( n->is_Region() ) { // Unreachable region 1902 // Note: nn == C->top() 1903 n->set_req(0, NULL); // Cut selfreference 1904 bool progress = true; 1905 uint max = n->outcnt(); 1906 DUIterator i; 1907 while (progress) { 1908 progress = false; 1909 // Eagerly remove dead phis to avoid phis copies creation. 1910 for (i = n->outs(); n->has_out(i); i++) { 1911 Node* m = n->out(i); 1912 if (m->is_Phi()) { 1913 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1914 replace_node(m, nn); 1915 if (max != n->outcnt()) { 1916 progress = true; 1917 i = n->refresh_out_pos(i); 1918 max = n->outcnt(); 1919 } 1920 } 1921 } 1922 } 1923 } 1924 replace_node(n,nn); // Update DefUse edges for new constant 1925 } 1926 return nn; 1927 } 1928 1929 // If x is a TypeNode, capture any more-precise type permanently into Node 1930 if (t != n->bottom_type()) { 1931 hash_delete(n); // changing bottom type may force a rehash 1932 n->raise_bottom_type(t); 1933 _worklist.push(n); // n re-enters the hash table via the worklist 1934 } 1935 1936 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1937 switch( n->Opcode() ) { 1938 case Op_FastLock: // Revisit FastLocks for lock coarsening 1939 case Op_If: 1940 case Op_CountedLoopEnd: 1941 case Op_Region: 1942 case Op_Loop: 1943 case Op_CountedLoop: 1944 case Op_Conv2B: 1945 case Op_Opaque1: 1946 case Op_Opaque2: 1947 _worklist.push(n); 1948 break; 1949 default: 1950 break; 1951 } 1952 1953 return n; 1954 } 1955 1956 //---------------------------------saturate------------------------------------ 1957 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1958 const Type* limit_type) const { 1959 const Type* wide_type = new_type->widen(old_type, limit_type); 1960 if (wide_type != new_type) { // did we widen? 1961 // If so, we may have widened beyond the limit type. Clip it back down. 1962 new_type = wide_type->filter(limit_type); 1963 } 1964 return new_type; 1965 } 1966 1967 //------------------------------print_statistics------------------------------- 1968 #ifndef PRODUCT 1969 void PhaseCCP::print_statistics() { 1970 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1971 } 1972 #endif 1973 1974 1975 //============================================================================= 1976 #ifndef PRODUCT 1977 uint PhasePeephole::_total_peepholes = 0; 1978 #endif 1979 //------------------------------PhasePeephole---------------------------------- 1980 // Conditional Constant Propagation, ala Wegman & Zadeck 1981 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1982 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1983 NOT_PRODUCT( clear_peepholes(); ) 1984 } 1985 1986 #ifndef PRODUCT 1987 //------------------------------~PhasePeephole--------------------------------- 1988 PhasePeephole::~PhasePeephole() { 1989 _total_peepholes += count_peepholes(); 1990 } 1991 #endif 1992 1993 //------------------------------transform-------------------------------------- 1994 Node *PhasePeephole::transform( Node *n ) { 1995 ShouldNotCallThis(); 1996 return NULL; 1997 } 1998 1999 //------------------------------do_transform----------------------------------- 2000 void PhasePeephole::do_transform() { 2001 bool method_name_not_printed = true; 2002 2003 // Examine each basic block 2004 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 2005 Block* block = _cfg.get_block(block_number); 2006 bool block_not_printed = true; 2007 2008 // and each instruction within a block 2009 uint end_index = block->number_of_nodes(); 2010 // block->end_idx() not valid after PhaseRegAlloc 2011 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 2012 Node *n = block->get_node(instruction_index); 2013 if( n->is_Mach() ) { 2014 MachNode *m = n->as_Mach(); 2015 int deleted_count = 0; 2016 // check for peephole opportunities 2017 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count); 2018 if( m2 != NULL ) { 2019 #ifndef PRODUCT 2020 if( PrintOptoPeephole ) { 2021 // Print method, first time only 2022 if( C->method() && method_name_not_printed ) { 2023 C->method()->print_short_name(); tty->cr(); 2024 method_name_not_printed = false; 2025 } 2026 // Print this block 2027 if( Verbose && block_not_printed) { 2028 tty->print_cr("in block"); 2029 block->dump(); 2030 block_not_printed = false; 2031 } 2032 // Print instructions being deleted 2033 for( int i = (deleted_count - 1); i >= 0; --i ) { 2034 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 2035 } 2036 tty->print_cr("replaced with"); 2037 // Print new instruction 2038 m2->format(_regalloc); 2039 tty->print("\n\n"); 2040 } 2041 #endif 2042 // Remove old nodes from basic block and update instruction_index 2043 // (old nodes still exist and may have edges pointing to them 2044 // as register allocation info is stored in the allocator using 2045 // the node index to live range mappings.) 2046 uint safe_instruction_index = (instruction_index - deleted_count); 2047 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 2048 block->remove_node( instruction_index ); 2049 } 2050 // install new node after safe_instruction_index 2051 block->insert_node(m2, safe_instruction_index + 1); 2052 end_index = block->number_of_nodes() - 1; // Recompute new block size 2053 NOT_PRODUCT( inc_peepholes(); ) 2054 } 2055 } 2056 } 2057 } 2058 } 2059 2060 //------------------------------print_statistics------------------------------- 2061 #ifndef PRODUCT 2062 void PhasePeephole::print_statistics() { 2063 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2064 } 2065 #endif 2066 2067 2068 //============================================================================= 2069 //------------------------------set_req_X-------------------------------------- 2070 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2071 assert( is_not_dead(n), "can not use dead node"); 2072 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2073 Node *old = in(i); 2074 set_req(i, n); 2075 2076 // old goes dead? 2077 if( old ) { 2078 switch (old->outcnt()) { 2079 case 0: 2080 // Put into the worklist to kill later. We do not kill it now because the 2081 // recursive kill will delete the current node (this) if dead-loop exists 2082 if (!old->is_top()) 2083 igvn->_worklist.push( old ); 2084 break; 2085 case 1: 2086 if( old->is_Store() || old->has_special_unique_user() ) 2087 igvn->add_users_to_worklist( old ); 2088 break; 2089 case 2: 2090 if( old->is_Store() ) 2091 igvn->add_users_to_worklist( old ); 2092 if( old->Opcode() == Op_Region ) 2093 igvn->_worklist.push(old); 2094 break; 2095 case 3: 2096 if( old->Opcode() == Op_Region ) { 2097 igvn->_worklist.push(old); 2098 igvn->add_users_to_worklist( old ); 2099 } 2100 break; 2101 default: 2102 break; 2103 } 2104 2105 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old); 2106 } 2107 2108 } 2109 2110 //-------------------------------replace_by----------------------------------- 2111 // Using def-use info, replace one node for another. Follow the def-use info 2112 // to all users of the OLD node. Then make all uses point to the NEW node. 2113 void Node::replace_by(Node *new_node) { 2114 assert(!is_top(), "top node has no DU info"); 2115 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2116 Node* use = last_out(i); 2117 uint uses_found = 0; 2118 for (uint j = 0; j < use->len(); j++) { 2119 if (use->in(j) == this) { 2120 if (j < use->req()) 2121 use->set_req(j, new_node); 2122 else use->set_prec(j, new_node); 2123 uses_found++; 2124 } 2125 } 2126 i -= uses_found; // we deleted 1 or more copies of this edge 2127 } 2128 } 2129 2130 //============================================================================= 2131 //----------------------------------------------------------------------------- 2132 void Type_Array::grow( uint i ) { 2133 if( !_max ) { 2134 _max = 1; 2135 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2136 _types[0] = NULL; 2137 } 2138 uint old = _max; 2139 while( i >= _max ) _max <<= 1; // Double to fit 2140 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2141 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2142 } 2143 2144 //------------------------------dump------------------------------------------- 2145 #ifndef PRODUCT 2146 void Type_Array::dump() const { 2147 uint max = Size(); 2148 for( uint i = 0; i < max; i++ ) { 2149 if( _types[i] != NULL ) { 2150 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2151 } 2152 } 2153 } 2154 #endif