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 // Remove edges from "root" to each SafePoint at a backward branch. 435 // They were inserted during parsing (see add_safepoint()) to make infinite 436 // loops without calls or exceptions visible to root, i.e., useful. 437 Node *root = C->root(); 438 if( root != NULL ) { 439 for( uint i = root->req(); i < root->len(); ++i ) { 440 Node *n = root->in(i); 441 if( n != NULL && n->is_SafePoint() ) { 442 root->rm_prec(i); 443 --i; 444 } 445 } 446 } 447 } 448 449 //============================================================================= 450 //------------------------------PhaseRenumberLive------------------------------ 451 // First, remove useless nodes (equivalent to identifying live nodes). 452 // Then, renumber live nodes. 453 // 454 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure. 455 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the 456 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique 457 // value in the range [0, x). 458 // 459 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is 460 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes). 461 // 462 // The PhaseRenumberLive phase updates two data structures with the new node IDs. 463 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be 464 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'. 465 // (2) Type information (the field PhaseGVN::_types) maps type information to each 466 // node ID. The mapping is updated to use the new node IDs as well. Updated type 467 // information is returned in PhaseGVN::_types. 468 // 469 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist. 470 // 471 // Other data structures used by the compiler are not updated. The hash table for value 472 // numbering (the field PhaseGVN::_table) is not updated because computing the hash 473 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either 474 // because it is empty wherever PhaseRenumberLive is used. 475 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn, 476 Unique_Node_List* worklist, Unique_Node_List* new_worklist, 477 PhaseNumber phase_num) : 478 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live) { 479 480 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place"); 481 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes"); 482 assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point"); 483 484 uint old_unique_count = C->unique(); 485 uint live_node_count = C->live_nodes(); 486 uint worklist_size = worklist->size(); 487 488 // Storage for the updated type information. 489 Type_Array new_type_array(C->comp_arena()); 490 491 // Iterate over the set of live nodes. 492 uint current_idx = 0; // The current new node ID. Incremented after every assignment. 493 for (uint i = 0; i < _useful.size(); i++) { 494 Node* n = _useful.at(i); 495 // Sanity check that fails if we ever decide to execute this phase after EA 496 assert(!n->is_Phi() || n->as_Phi()->inst_mem_id() == -1, "should not be linked to data Phi"); 497 const Type* type = gvn->type_or_null(n); 498 new_type_array.map(current_idx, type); 499 500 bool in_worklist = false; 501 if (worklist->member(n)) { 502 in_worklist = true; 503 } 504 505 n->set_idx(current_idx); // Update node ID. 506 507 if (in_worklist) { 508 new_worklist->push(n); 509 } 510 511 current_idx++; 512 } 513 514 assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist"); 515 assert(live_node_count == current_idx, "all live nodes must be processed"); 516 517 // Replace the compiler's type information with the updated type information. 518 gvn->replace_types(new_type_array); 519 520 // Update the unique node count of the compilation to the number of currently live nodes. 521 C->set_unique(live_node_count); 522 523 // Set the dead node count to 0 and reset dead node list. 524 C->reset_dead_node_list(); 525 } 526 527 528 //============================================================================= 529 //------------------------------PhaseTransform--------------------------------- 530 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 531 _arena(Thread::current()->resource_area()), 532 _nodes(_arena), 533 _types(_arena) 534 { 535 init_con_caches(); 536 #ifndef PRODUCT 537 clear_progress(); 538 clear_transforms(); 539 set_allow_progress(true); 540 #endif 541 // Force allocation for currently existing nodes 542 _types.map(C->unique(), NULL); 543 } 544 545 //------------------------------PhaseTransform--------------------------------- 546 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 547 _arena(arena), 548 _nodes(arena), 549 _types(arena) 550 { 551 init_con_caches(); 552 #ifndef PRODUCT 553 clear_progress(); 554 clear_transforms(); 555 set_allow_progress(true); 556 #endif 557 // Force allocation for currently existing nodes 558 _types.map(C->unique(), NULL); 559 } 560 561 //------------------------------PhaseTransform--------------------------------- 562 // Initialize with previously generated type information 563 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 564 _arena(pt->_arena), 565 _nodes(pt->_nodes), 566 _types(pt->_types) 567 { 568 init_con_caches(); 569 #ifndef PRODUCT 570 clear_progress(); 571 clear_transforms(); 572 set_allow_progress(true); 573 #endif 574 } 575 576 void PhaseTransform::init_con_caches() { 577 memset(_icons,0,sizeof(_icons)); 578 memset(_lcons,0,sizeof(_lcons)); 579 memset(_zcons,0,sizeof(_zcons)); 580 } 581 582 583 //--------------------------------find_int_type-------------------------------- 584 const TypeInt* PhaseTransform::find_int_type(Node* n) { 585 if (n == NULL) return NULL; 586 // Call type_or_null(n) to determine node's type since we might be in 587 // parse phase and call n->Value() may return wrong type. 588 // (For example, a phi node at the beginning of loop parsing is not ready.) 589 const Type* t = type_or_null(n); 590 if (t == NULL) return NULL; 591 return t->isa_int(); 592 } 593 594 595 //-------------------------------find_long_type-------------------------------- 596 const TypeLong* PhaseTransform::find_long_type(Node* n) { 597 if (n == NULL) return NULL; 598 // (See comment above on type_or_null.) 599 const Type* t = type_or_null(n); 600 if (t == NULL) return NULL; 601 return t->isa_long(); 602 } 603 604 605 #ifndef PRODUCT 606 void PhaseTransform::dump_old2new_map() const { 607 _nodes.dump(); 608 } 609 610 void PhaseTransform::dump_new( uint nidx ) const { 611 for( uint i=0; i<_nodes.Size(); i++ ) 612 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 613 _nodes[i]->dump(); 614 tty->cr(); 615 tty->print_cr("Old index= %d",i); 616 return; 617 } 618 tty->print_cr("Node %d not found in the new indices", nidx); 619 } 620 621 //------------------------------dump_types------------------------------------- 622 void PhaseTransform::dump_types( ) const { 623 _types.dump(); 624 } 625 626 //------------------------------dump_nodes_and_types--------------------------- 627 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 628 VectorSet visited(Thread::current()->resource_area()); 629 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 630 } 631 632 //------------------------------dump_nodes_and_types_recur--------------------- 633 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 634 if( !n ) return; 635 if( depth == 0 ) return; 636 if( visited.test_set(n->_idx) ) return; 637 for( uint i=0; i<n->len(); i++ ) { 638 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 639 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 640 } 641 n->dump(); 642 if (type_or_null(n) != NULL) { 643 tty->print(" "); type(n)->dump(); tty->cr(); 644 } 645 } 646 647 #endif 648 649 650 //============================================================================= 651 //------------------------------PhaseValues------------------------------------ 652 // Set minimum table size to "255" 653 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 654 NOT_PRODUCT( clear_new_values(); ) 655 } 656 657 //------------------------------PhaseValues------------------------------------ 658 // Set minimum table size to "255" 659 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 660 _table(&ptv->_table) { 661 NOT_PRODUCT( clear_new_values(); ) 662 } 663 664 //------------------------------PhaseValues------------------------------------ 665 // Used by +VerifyOpto. Clear out hash table but copy _types array. 666 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 667 _table(ptv->arena(),ptv->_table.size()) { 668 NOT_PRODUCT( clear_new_values(); ) 669 } 670 671 //------------------------------~PhaseValues----------------------------------- 672 #ifndef PRODUCT 673 PhaseValues::~PhaseValues() { 674 _table.dump(); 675 676 // Statistics for value progress and efficiency 677 if( PrintCompilation && Verbose && WizardMode ) { 678 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 679 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 680 if( made_transforms() != 0 ) { 681 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 682 } else { 683 tty->cr(); 684 } 685 } 686 } 687 #endif 688 689 //------------------------------makecon---------------------------------------- 690 ConNode* PhaseTransform::makecon(const Type *t) { 691 assert(t->singleton(), "must be a constant"); 692 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 693 switch (t->base()) { // fast paths 694 case Type::Half: 695 case Type::Top: return (ConNode*) C->top(); 696 case Type::Int: return intcon( t->is_int()->get_con() ); 697 case Type::Long: return longcon( t->is_long()->get_con() ); 698 default: break; 699 } 700 if (t->is_zero_type()) 701 return zerocon(t->basic_type()); 702 return uncached_makecon(t); 703 } 704 705 //--------------------------uncached_makecon----------------------------------- 706 // Make an idealized constant - one of ConINode, ConPNode, etc. 707 ConNode* PhaseValues::uncached_makecon(const Type *t) { 708 assert(t->singleton(), "must be a constant"); 709 ConNode* x = ConNode::make(t); 710 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 711 if (k == NULL) { 712 set_type(x, t); // Missed, provide type mapping 713 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 714 if (nna != NULL) { 715 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 716 loc->clear(); // do not put debug info on constants 717 } 718 } else { 719 x->destruct(); // Hit, destroy duplicate constant 720 x = k; // use existing constant 721 } 722 return x; 723 } 724 725 //------------------------------intcon----------------------------------------- 726 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 727 ConINode* PhaseTransform::intcon(jint i) { 728 // Small integer? Check cache! Check that cached node is not dead 729 if (i >= _icon_min && i <= _icon_max) { 730 ConINode* icon = _icons[i-_icon_min]; 731 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 732 return icon; 733 } 734 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 735 assert(icon->is_Con(), ""); 736 if (i >= _icon_min && i <= _icon_max) 737 _icons[i-_icon_min] = icon; // Cache small integers 738 return icon; 739 } 740 741 //------------------------------longcon---------------------------------------- 742 // Fast long constant. 743 ConLNode* PhaseTransform::longcon(jlong l) { 744 // Small integer? Check cache! Check that cached node is not dead 745 if (l >= _lcon_min && l <= _lcon_max) { 746 ConLNode* lcon = _lcons[l-_lcon_min]; 747 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 748 return lcon; 749 } 750 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 751 assert(lcon->is_Con(), ""); 752 if (l >= _lcon_min && l <= _lcon_max) 753 _lcons[l-_lcon_min] = lcon; // Cache small integers 754 return lcon; 755 } 756 757 //------------------------------zerocon----------------------------------------- 758 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 759 ConNode* PhaseTransform::zerocon(BasicType bt) { 760 assert((uint)bt <= _zcon_max, "domain check"); 761 ConNode* zcon = _zcons[bt]; 762 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 763 return zcon; 764 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 765 _zcons[bt] = zcon; 766 return zcon; 767 } 768 769 770 771 //============================================================================= 772 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) { 773 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape); 774 if (i == NULL) { 775 i = k->Ideal(this, can_reshape); 776 } 777 return i; 778 } 779 780 Node* PhaseGVN::apply_identity(Node* k) { 781 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->identity_node(this, k); 782 if (i == k) { 783 i = k->Identity(this); 784 } 785 return i; 786 } 787 788 //------------------------------transform-------------------------------------- 789 // Return a node which computes the same function as this node, but in a 790 // faster or cheaper fashion. 791 Node *PhaseGVN::transform( Node *n ) { 792 return transform_no_reclaim(n); 793 } 794 795 //------------------------------transform-------------------------------------- 796 // Return a node which computes the same function as this node, but 797 // in a faster or cheaper fashion. 798 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 799 NOT_PRODUCT( set_transforms(); ) 800 801 // Apply the Ideal call in a loop until it no longer applies 802 Node *k = n; 803 NOT_PRODUCT( uint loop_count = 0; ) 804 while( 1 ) { 805 Node *i = apply_ideal(k, /*can_reshape=*/false); 806 if( !i ) break; 807 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 808 k = i; 809 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 810 } 811 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 812 813 814 // If brand new node, make space in type array. 815 ensure_type_or_null(k); 816 817 // Since I just called 'Value' to compute the set of run-time values 818 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 819 // cache Value. Later requests for the local phase->type of this Node can 820 // use the cached Value instead of suffering with 'bottom_type'. 821 const Type *t = k->Value(this); // Get runtime Value set 822 assert(t != NULL, "value sanity"); 823 if (type_or_null(k) != t) { 824 #ifndef PRODUCT 825 // Do not count initial visit to node as a transformation 826 if (type_or_null(k) == NULL) { 827 inc_new_values(); 828 set_progress(); 829 } 830 #endif 831 set_type(k, t); 832 // If k is a TypeNode, capture any more-precise type permanently into Node 833 k->raise_bottom_type(t); 834 } 835 836 if( t->singleton() && !k->is_Con() ) { 837 NOT_PRODUCT( set_progress(); ) 838 return makecon(t); // Turn into a constant 839 } 840 841 // Now check for Identities 842 Node *i = apply_identity(k); // Look for a nearby replacement 843 if( i != k ) { // Found? Return replacement! 844 NOT_PRODUCT( set_progress(); ) 845 return i; 846 } 847 848 // Global Value Numbering 849 i = hash_find_insert(k); // Insert if new 850 if( i && (i != k) ) { 851 // Return the pre-existing node 852 NOT_PRODUCT( set_progress(); ) 853 return i; 854 } 855 856 // Return Idealized original 857 return k; 858 } 859 860 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 861 if (d->is_top() || n->is_top()) { 862 return false; 863 } 864 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 865 int i = 0; 866 while (d != n) { 867 n = IfNode::up_one_dom(n, linear_only); 868 i++; 869 if (n == NULL || i >= 10) { 870 return false; 871 } 872 } 873 return true; 874 } 875 876 #ifdef ASSERT 877 //------------------------------dead_loop_check-------------------------------- 878 // Check for a simple dead loop when a data node references itself directly 879 // or through an other data node excluding cons and phis. 880 void PhaseGVN::dead_loop_check( Node *n ) { 881 // Phi may reference itself in a loop 882 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 883 // Do 2 levels check and only data inputs. 884 bool no_dead_loop = true; 885 uint cnt = n->req(); 886 for (uint i = 1; i < cnt && no_dead_loop; i++) { 887 Node *in = n->in(i); 888 if (in == n) { 889 no_dead_loop = false; 890 } else if (in != NULL && !in->is_dead_loop_safe()) { 891 uint icnt = in->req(); 892 for (uint j = 1; j < icnt && no_dead_loop; j++) { 893 if (in->in(j) == n || in->in(j) == in) 894 no_dead_loop = false; 895 } 896 } 897 } 898 if (!no_dead_loop) n->dump(3); 899 assert(no_dead_loop, "dead loop detected"); 900 } 901 } 902 #endif 903 904 //============================================================================= 905 //------------------------------PhaseIterGVN----------------------------------- 906 // Initialize hash table to fresh and clean for +VerifyOpto 907 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), 908 _delay_transform(false), 909 _stack(C->live_nodes() >> 1), 910 _worklist( ) { 911 } 912 913 //------------------------------PhaseIterGVN----------------------------------- 914 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 915 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 916 _delay_transform(igvn->_delay_transform), 917 _stack( igvn->_stack ), 918 _worklist( igvn->_worklist ) 919 { 920 } 921 922 //------------------------------PhaseIterGVN----------------------------------- 923 // Initialize with previous PhaseGVN info from Parser 924 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 925 _delay_transform(false), 926 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that 927 // the constructor is used in incremental inlining, this consumes too much memory: 928 // _stack(C->live_nodes() >> 1), 929 // So, as a band-aid, we replace this by: 930 _stack(C->comp_arena(), 32), 931 _worklist(*C->for_igvn()) 932 { 933 uint max; 934 935 // Dead nodes in the hash table inherited from GVN were not treated as 936 // roots during def-use info creation; hence they represent an invisible 937 // use. Clear them out. 938 max = _table.size(); 939 for( uint i = 0; i < max; ++i ) { 940 Node *n = _table.at(i); 941 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 942 if( n->is_top() ) continue; 943 assert( false, "Parse::remove_useless_nodes missed this node"); 944 hash_delete(n); 945 } 946 } 947 948 // Any Phis or Regions on the worklist probably had uses that could not 949 // make more progress because the uses were made while the Phis and Regions 950 // were in half-built states. Put all uses of Phis and Regions on worklist. 951 max = _worklist.size(); 952 for( uint j = 0; j < max; j++ ) { 953 Node *n = _worklist.at(j); 954 uint uop = n->Opcode(); 955 if( uop == Op_Phi || uop == Op_Region || 956 n->is_Type() || 957 n->is_Mem() ) 958 add_users_to_worklist(n); 959 } 960 961 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 962 bs->add_users_to_worklist(&_worklist); 963 } 964 965 /** 966 * Initialize worklist for each node. 967 */ 968 void PhaseIterGVN::init_worklist(Node* first) { 969 Unique_Node_List to_process; 970 to_process.push(first); 971 972 while (to_process.size() > 0) { 973 Node* n = to_process.pop(); 974 if (!_worklist.member(n)) { 975 _worklist.push(n); 976 977 uint cnt = n->req(); 978 for(uint i = 0; i < cnt; i++) { 979 Node* m = n->in(i); 980 if (m != NULL) { 981 to_process.push(m); 982 } 983 } 984 } 985 } 986 } 987 988 #ifndef PRODUCT 989 void PhaseIterGVN::verify_step(Node* n) { 990 if (VerifyIterativeGVN) { 991 _verify_window[_verify_counter % _verify_window_size] = n; 992 ++_verify_counter; 993 ResourceMark rm; 994 ResourceArea* area = Thread::current()->resource_area(); 995 VectorSet old_space(area), new_space(area); 996 if (C->unique() < 1000 || 997 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 998 ++_verify_full_passes; 999 Node::verify_recur(C->root(), -1, old_space, new_space); 1000 } 1001 const int verify_depth = 4; 1002 for ( int i = 0; i < _verify_window_size; i++ ) { 1003 Node* n = _verify_window[i]; 1004 if ( n == NULL ) continue; 1005 if( n->in(0) == NodeSentinel ) { // xform_idom 1006 _verify_window[i] = n->in(1); 1007 --i; continue; 1008 } 1009 // Typical fanout is 1-2, so this call visits about 6 nodes. 1010 Node::verify_recur(n, verify_depth, old_space, new_space); 1011 } 1012 } 1013 } 1014 1015 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 1016 if (TraceIterativeGVN) { 1017 uint wlsize = _worklist.size(); 1018 const Type* newtype = type_or_null(n); 1019 if (nn != n) { 1020 // print old node 1021 tty->print("< "); 1022 if (oldtype != newtype && oldtype != NULL) { 1023 oldtype->dump(); 1024 } 1025 do { tty->print("\t"); } while (tty->position() < 16); 1026 tty->print("<"); 1027 n->dump(); 1028 } 1029 if (oldtype != newtype || nn != n) { 1030 // print new node and/or new type 1031 if (oldtype == NULL) { 1032 tty->print("* "); 1033 } else if (nn != n) { 1034 tty->print("> "); 1035 } else { 1036 tty->print("= "); 1037 } 1038 if (newtype == NULL) { 1039 tty->print("null"); 1040 } else { 1041 newtype->dump(); 1042 } 1043 do { tty->print("\t"); } while (tty->position() < 16); 1044 nn->dump(); 1045 } 1046 if (Verbose && wlsize < _worklist.size()) { 1047 tty->print(" Push {"); 1048 while (wlsize != _worklist.size()) { 1049 Node* pushed = _worklist.at(wlsize++); 1050 tty->print(" %d", pushed->_idx); 1051 } 1052 tty->print_cr(" }"); 1053 } 1054 if (nn != n) { 1055 // ignore n, it might be subsumed 1056 verify_step((Node*) NULL); 1057 } 1058 } 1059 } 1060 1061 void PhaseIterGVN::init_verifyPhaseIterGVN() { 1062 _verify_counter = 0; 1063 _verify_full_passes = 0; 1064 for (int i = 0; i < _verify_window_size; i++) { 1065 _verify_window[i] = NULL; 1066 } 1067 #ifdef ASSERT 1068 // Verify that all modified nodes are on _worklist 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() && !_worklist.member(n)) { 1073 n->dump(); 1074 assert(false, "modified node is not on IGVN._worklist"); 1075 } 1076 } 1077 #endif 1078 } 1079 1080 void PhaseIterGVN::verify_PhaseIterGVN() { 1081 #ifdef ASSERT 1082 // Verify nodes with changed inputs. 1083 Unique_Node_List* modified_list = C->modified_nodes(); 1084 while (modified_list != NULL && modified_list->size()) { 1085 Node* n = modified_list->pop(); 1086 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes 1087 n->dump(); 1088 assert(false, "modified node was not processed by IGVN.transform_old()"); 1089 } 1090 } 1091 #endif 1092 1093 C->verify_graph_edges(); 1094 if( VerifyOpto && allow_progress() ) { 1095 // Must turn off allow_progress to enable assert and break recursion 1096 C->root()->verify(); 1097 { // Check if any progress was missed using IterGVN 1098 // Def-Use info enables transformations not attempted in wash-pass 1099 // e.g. Region/Phi cleanup, ... 1100 // Null-check elision -- may not have reached fixpoint 1101 // do not propagate to dominated nodes 1102 ResourceMark rm; 1103 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 1104 // Fill worklist completely 1105 igvn2.init_worklist(C->root()); 1106 1107 igvn2.set_allow_progress(false); 1108 igvn2.optimize(); 1109 igvn2.set_allow_progress(true); 1110 } 1111 } 1112 if (VerifyIterativeGVN && PrintOpto) { 1113 if (_verify_counter == _verify_full_passes) { 1114 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 1115 (int) _verify_full_passes); 1116 } else { 1117 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 1118 (int) _verify_counter, (int) _verify_full_passes); 1119 } 1120 } 1121 1122 #ifdef ASSERT 1123 while (modified_list->size()) { 1124 Node* n = modified_list->pop(); 1125 n->dump(); 1126 assert(false, "VerifyIterativeGVN: new modified node was added"); 1127 } 1128 #endif 1129 } 1130 #endif /* PRODUCT */ 1131 1132 #ifdef ASSERT 1133 /** 1134 * Dumps information that can help to debug the problem. A debug 1135 * build fails with an assert. 1136 */ 1137 void PhaseIterGVN::dump_infinite_loop_info(Node* n) { 1138 n->dump(4); 1139 _worklist.dump(); 1140 assert(false, "infinite loop in PhaseIterGVN::optimize"); 1141 } 1142 1143 /** 1144 * Prints out information about IGVN if the 'verbose' option is used. 1145 */ 1146 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1147 if (TraceIterativeGVN && Verbose) { 1148 tty->print(" Pop "); 1149 n->dump(); 1150 if ((num_processed % 100) == 0) { 1151 _worklist.print_set(); 1152 } 1153 } 1154 } 1155 #endif /* ASSERT */ 1156 1157 void PhaseIterGVN::optimize() { 1158 DEBUG_ONLY(uint num_processed = 0;) 1159 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1160 1161 uint loop_count = 0; 1162 // Pull from worklist and transform the node. If the node has changed, 1163 // update edge info and put uses on worklist. 1164 while(_worklist.size()) { 1165 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1166 return; 1167 } 1168 Node* n = _worklist.pop(); 1169 if (++loop_count >= K * C->live_nodes()) { 1170 DEBUG_ONLY(dump_infinite_loop_info(n);) 1171 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1172 return; 1173 } 1174 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1175 if (n->outcnt() != 0) { 1176 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1177 // Do the transformation 1178 Node* nn = transform_old(n); 1179 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1180 } else if (!n->is_top()) { 1181 remove_dead_node(n); 1182 } 1183 } 1184 NOT_PRODUCT(verify_PhaseIterGVN();) 1185 } 1186 1187 1188 /** 1189 * Register a new node with the optimizer. Update the types array, the def-use 1190 * info. Put on worklist. 1191 */ 1192 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1193 set_type_bottom(n); 1194 _worklist.push(n); 1195 if (orig != NULL) C->copy_node_notes_to(n, orig); 1196 return n; 1197 } 1198 1199 //------------------------------transform-------------------------------------- 1200 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1201 Node *PhaseIterGVN::transform( Node *n ) { 1202 if (_delay_transform) { 1203 // Register the node but don't optimize for now 1204 register_new_node_with_optimizer(n); 1205 return n; 1206 } 1207 1208 // If brand new node, make space in type array, and give it a type. 1209 ensure_type_or_null(n); 1210 if (type_or_null(n) == NULL) { 1211 set_type_bottom(n); 1212 } 1213 1214 return transform_old(n); 1215 } 1216 1217 Node *PhaseIterGVN::transform_old(Node* n) { 1218 DEBUG_ONLY(uint loop_count = 0;); 1219 NOT_PRODUCT(set_transforms()); 1220 1221 // Remove 'n' from hash table in case it gets modified 1222 _table.hash_delete(n); 1223 if (VerifyIterativeGVN) { 1224 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1225 } 1226 1227 // Apply the Ideal call in a loop until it no longer applies 1228 Node* k = n; 1229 DEBUG_ONLY(dead_loop_check(k);) 1230 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1231 C->remove_modified_node(k); 1232 Node* i = apply_ideal(k, /*can_reshape=*/true); 1233 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1234 #ifndef PRODUCT 1235 verify_step(k); 1236 if (i && VerifyOpto ) { 1237 if (!allow_progress()) { 1238 if (i->is_Add() && (i->outcnt() == 1)) { 1239 // Switched input to left side because this is the only use 1240 } else if (i->is_If() && (i->in(0) == NULL)) { 1241 // This IF is dead because it is dominated by an equivalent IF When 1242 // dominating if changed, info is not propagated sparsely to 'this' 1243 // Propagating this info further will spuriously identify other 1244 // progress. 1245 return i; 1246 } else 1247 set_progress(); 1248 } else { 1249 set_progress(); 1250 } 1251 } 1252 #endif 1253 1254 while (i != NULL) { 1255 #ifdef ASSERT 1256 if (loop_count >= K) { 1257 dump_infinite_loop_info(i); 1258 } 1259 loop_count++; 1260 #endif 1261 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1262 // Made a change; put users of original Node on worklist 1263 add_users_to_worklist(k); 1264 // Replacing root of transform tree? 1265 if (k != i) { 1266 // Make users of old Node now use new. 1267 subsume_node(k, i); 1268 k = i; 1269 } 1270 DEBUG_ONLY(dead_loop_check(k);) 1271 // Try idealizing again 1272 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1273 C->remove_modified_node(k); 1274 i = apply_ideal(k, /*can_reshape=*/true); 1275 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1276 #ifndef PRODUCT 1277 verify_step(k); 1278 if (i && VerifyOpto) { 1279 set_progress(); 1280 } 1281 #endif 1282 } 1283 1284 // If brand new node, make space in type array. 1285 ensure_type_or_null(k); 1286 1287 // See what kind of values 'k' takes on at runtime 1288 const Type* t = k->Value(this); 1289 assert(t != NULL, "value sanity"); 1290 1291 // Since I just called 'Value' to compute the set of run-time values 1292 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1293 // cache Value. Later requests for the local phase->type of this Node can 1294 // use the cached Value instead of suffering with 'bottom_type'. 1295 if (type_or_null(k) != t) { 1296 #ifndef PRODUCT 1297 inc_new_values(); 1298 set_progress(); 1299 #endif 1300 set_type(k, t); 1301 // If k is a TypeNode, capture any more-precise type permanently into Node 1302 k->raise_bottom_type(t); 1303 // Move users of node to worklist 1304 add_users_to_worklist(k); 1305 } 1306 // If 'k' computes a constant, replace it with a constant 1307 if (t->singleton() && !k->is_Con()) { 1308 NOT_PRODUCT(set_progress();) 1309 Node* con = makecon(t); // Make a constant 1310 add_users_to_worklist(k); 1311 subsume_node(k, con); // Everybody using k now uses con 1312 return con; 1313 } 1314 1315 // Now check for Identities 1316 i = apply_identity(k); // Look for a nearby replacement 1317 if (i != k) { // Found? Return replacement! 1318 NOT_PRODUCT(set_progress();) 1319 add_users_to_worklist(k); 1320 subsume_node(k, i); // Everybody using k now uses i 1321 return i; 1322 } 1323 1324 // Global Value Numbering 1325 i = hash_find_insert(k); // Check for pre-existing node 1326 if (i && (i != k)) { 1327 // Return the pre-existing node if it isn't dead 1328 NOT_PRODUCT(set_progress();) 1329 add_users_to_worklist(k); 1330 subsume_node(k, i); // Everybody using k now uses i 1331 return i; 1332 } 1333 1334 // Return Idealized original 1335 return k; 1336 } 1337 1338 //---------------------------------saturate------------------------------------ 1339 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1340 const Type* limit_type) const { 1341 return new_type->narrow(old_type); 1342 } 1343 1344 //------------------------------remove_globally_dead_node---------------------- 1345 // Kill a globally dead Node. All uses are also globally dead and are 1346 // aggressively trimmed. 1347 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1348 enum DeleteProgress { 1349 PROCESS_INPUTS, 1350 PROCESS_OUTPUTS 1351 }; 1352 assert(_stack.is_empty(), "not empty"); 1353 _stack.push(dead, PROCESS_INPUTS); 1354 1355 while (_stack.is_nonempty()) { 1356 dead = _stack.node(); 1357 uint progress_state = _stack.index(); 1358 assert(dead != C->root(), "killing root, eh?"); 1359 assert(!dead->is_top(), "add check for top when pushing"); 1360 NOT_PRODUCT( set_progress(); ) 1361 if (progress_state == PROCESS_INPUTS) { 1362 // After following inputs, continue to outputs 1363 _stack.set_index(PROCESS_OUTPUTS); 1364 if (!dead->is_Con()) { // Don't kill cons but uses 1365 bool recurse = false; 1366 // Remove from hash table 1367 _table.hash_delete( dead ); 1368 // Smash all inputs to 'dead', isolating him completely 1369 for (uint i = 0; i < dead->req(); i++) { 1370 Node *in = dead->in(i); 1371 if (in != NULL && in != C->top()) { // Points to something? 1372 int nrep = dead->replace_edge(in, NULL); // Kill edges 1373 assert((nrep > 0), "sanity"); 1374 if (in->outcnt() == 0) { // Made input go dead? 1375 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1376 recurse = true; 1377 } else if (in->outcnt() == 1 && 1378 in->has_special_unique_user()) { 1379 _worklist.push(in->unique_out()); 1380 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1381 if (in->Opcode() == Op_Region) { 1382 _worklist.push(in); 1383 } else if (in->is_Store()) { 1384 DUIterator_Fast imax, i = in->fast_outs(imax); 1385 _worklist.push(in->fast_out(i)); 1386 i++; 1387 if (in->outcnt() == 2) { 1388 _worklist.push(in->fast_out(i)); 1389 i++; 1390 } 1391 assert(!(i < imax), "sanity"); 1392 } 1393 } else { 1394 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in); 1395 } 1396 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1397 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1398 // A Load that directly follows an InitializeNode is 1399 // going away. The Stores that follow are candidates 1400 // again to be captured by the InitializeNode. 1401 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1402 Node *n = in->fast_out(j); 1403 if (n->is_Store()) { 1404 _worklist.push(n); 1405 } 1406 } 1407 } 1408 } // if (in != NULL && in != C->top()) 1409 } // for (uint i = 0; i < dead->req(); i++) 1410 if (recurse) { 1411 continue; 1412 } 1413 } // if (!dead->is_Con()) 1414 } // if (progress_state == PROCESS_INPUTS) 1415 1416 // Aggressively kill globally dead uses 1417 // (Rather than pushing all the outs at once, we push one at a time, 1418 // plus the parent to resume later, because of the indefinite number 1419 // of edge deletions per loop trip.) 1420 if (dead->outcnt() > 0) { 1421 // Recursively remove output edges 1422 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1423 } else { 1424 // Finished disconnecting all input and output edges. 1425 _stack.pop(); 1426 // Remove dead node from iterative worklist 1427 _worklist.remove(dead); 1428 C->remove_modified_node(dead); 1429 // Constant node that has no out-edges and has only one in-edge from 1430 // root is usually dead. However, sometimes reshaping walk makes 1431 // it reachable by adding use edges. So, we will NOT count Con nodes 1432 // as dead to be conservative about the dead node count at any 1433 // given time. 1434 if (!dead->is_Con()) { 1435 C->record_dead_node(dead->_idx); 1436 } 1437 if (dead->is_macro()) { 1438 C->remove_macro_node(dead); 1439 } 1440 if (dead->is_expensive()) { 1441 C->remove_expensive_node(dead); 1442 } 1443 CastIINode* cast = dead->isa_CastII(); 1444 if (cast != NULL && cast->has_range_check()) { 1445 C->remove_range_check_cast(cast); 1446 } 1447 if (dead->Opcode() == Op_Opaque4) { 1448 C->remove_opaque4_node(dead); 1449 } 1450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1451 bs->unregister_potential_barrier_node(dead); 1452 } 1453 } // while (_stack.is_nonempty()) 1454 } 1455 1456 //------------------------------subsume_node----------------------------------- 1457 // Remove users from node 'old' and add them to node 'nn'. 1458 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1459 assert( old != hash_find(old), "should already been removed" ); 1460 assert( old != C->top(), "cannot subsume top node"); 1461 // Copy debug or profile information to the new version: 1462 C->copy_node_notes_to(nn, old); 1463 // Move users of node 'old' to node 'nn' 1464 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1465 Node* use = old->last_out(i); // for each use... 1466 // use might need re-hashing (but it won't if it's a new node) 1467 rehash_node_delayed(use); 1468 // Update use-def info as well 1469 // We remove all occurrences of old within use->in, 1470 // so as to avoid rehashing any node more than once. 1471 // The hash table probe swamps any outer loop overhead. 1472 uint num_edges = 0; 1473 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1474 if (use->in(j) == old) { 1475 use->set_req(j, nn); 1476 ++num_edges; 1477 } 1478 } 1479 i -= num_edges; // we deleted 1 or more copies of this edge 1480 } 1481 1482 // Search for instance field data PhiNodes in the same region pointing to the old 1483 // memory PhiNode and update their instance memory ids to point to the new node. 1484 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) { 1485 Node* region = old->in(0); 1486 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1487 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1488 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) { 1489 phi->set_inst_mem_id((int)nn->_idx); 1490 } 1491 } 1492 } 1493 1494 // Smash all inputs to 'old', isolating him completely 1495 Node *temp = new Node(1); 1496 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1497 remove_dead_node( old ); 1498 temp->del_req(0); // Yank bogus edge 1499 #ifndef PRODUCT 1500 if( VerifyIterativeGVN ) { 1501 for ( int i = 0; i < _verify_window_size; i++ ) { 1502 if ( _verify_window[i] == old ) 1503 _verify_window[i] = nn; 1504 } 1505 } 1506 #endif 1507 _worklist.remove(temp); // this can be necessary 1508 temp->destruct(); // reuse the _idx of this little guy 1509 } 1510 1511 //------------------------------add_users_to_worklist-------------------------- 1512 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1513 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1514 _worklist.push(n->fast_out(i)); // Push on worklist 1515 } 1516 } 1517 1518 // Return counted loop Phi if as a counted loop exit condition, cmp 1519 // compares the the induction variable with n 1520 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) { 1521 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1522 Node* bol = cmp->fast_out(i); 1523 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1524 Node* iff = bol->fast_out(i2); 1525 if (iff->is_CountedLoopEnd()) { 1526 CountedLoopEndNode* cle = iff->as_CountedLoopEnd(); 1527 if (cle->limit() == n) { 1528 PhiNode* phi = cle->phi(); 1529 if (phi != NULL) { 1530 return phi; 1531 } 1532 } 1533 } 1534 } 1535 } 1536 return NULL; 1537 } 1538 1539 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1540 add_users_to_worklist0(n); 1541 1542 // Move users of node to worklist 1543 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1544 Node* use = n->fast_out(i); // Get use 1545 1546 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1547 use->is_Store() ) // Enable store/load same address 1548 add_users_to_worklist0(use); 1549 1550 // If we changed the receiver type to a call, we need to revisit 1551 // the Catch following the call. It's looking for a non-NULL 1552 // receiver to know when to enable the regular fall-through path 1553 // in addition to the NullPtrException path. 1554 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1555 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control); 1556 if (p != NULL) { 1557 add_users_to_worklist0(p); 1558 } 1559 } 1560 1561 uint use_op = use->Opcode(); 1562 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1563 add_users_to_worklist(use); // Put Bool on worklist 1564 if (use->outcnt() > 0) { 1565 Node* bol = use->raw_out(0); 1566 if (bol->outcnt() > 0) { 1567 Node* iff = bol->raw_out(0); 1568 if (iff->outcnt() == 2) { 1569 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1570 // phi merging either 0 or 1 onto the worklist 1571 Node* ifproj0 = iff->raw_out(0); 1572 Node* ifproj1 = iff->raw_out(1); 1573 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1574 Node* region0 = ifproj0->raw_out(0); 1575 Node* region1 = ifproj1->raw_out(0); 1576 if( region0 == region1 ) 1577 add_users_to_worklist0(region0); 1578 } 1579 } 1580 } 1581 } 1582 if (use_op == Op_CmpI) { 1583 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n); 1584 if (phi != NULL) { 1585 // If an opaque node feeds into the limit condition of a 1586 // CountedLoop, we need to process the Phi node for the 1587 // induction variable when the opaque node is removed: 1588 // the range of values taken by the Phi is now known and 1589 // so its type is also known. 1590 _worklist.push(phi); 1591 } 1592 Node* in1 = use->in(1); 1593 for (uint i = 0; i < in1->outcnt(); i++) { 1594 if (in1->raw_out(i)->Opcode() == Op_CastII) { 1595 Node* castii = in1->raw_out(i); 1596 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) { 1597 Node* ifnode = castii->in(0)->in(0); 1598 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) { 1599 // Reprocess a CastII node that may depend on an 1600 // opaque node value when the opaque node is 1601 // removed. In case it carries a dependency we can do 1602 // a better job of computing its type. 1603 _worklist.push(castii); 1604 } 1605 } 1606 } 1607 } 1608 } 1609 } 1610 1611 // If changed Cast input, check Phi users for simple cycles 1612 if (use->is_ConstraintCast()) { 1613 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1614 Node* u = use->fast_out(i2); 1615 if (u->is_Phi()) 1616 _worklist.push(u); 1617 } 1618 } 1619 // If changed LShift inputs, check RShift users for useless sign-ext 1620 if( use_op == Op_LShiftI ) { 1621 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1622 Node* u = use->fast_out(i2); 1623 if (u->Opcode() == Op_RShiftI) 1624 _worklist.push(u); 1625 } 1626 } 1627 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1628 if (use_op == Op_AddI || use_op == Op_SubI) { 1629 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1630 Node* u = use->fast_out(i2); 1631 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1632 _worklist.push(u); 1633 } 1634 } 1635 } 1636 // If changed AddP inputs, check Stores for loop invariant 1637 if( use_op == Op_AddP ) { 1638 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1639 Node* u = use->fast_out(i2); 1640 if (u->is_Mem()) 1641 _worklist.push(u); 1642 } 1643 } 1644 // If changed initialization activity, check dependent Stores 1645 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1646 InitializeNode* init = use->as_Allocate()->initialization(); 1647 if (init != NULL) { 1648 Node* imem = init->proj_out_or_null(TypeFunc::Memory); 1649 if (imem != NULL) add_users_to_worklist0(imem); 1650 } 1651 } 1652 if (use_op == Op_Initialize) { 1653 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory); 1654 if (imem != NULL) add_users_to_worklist0(imem); 1655 } 1656 // Loading the java mirror from a Klass requires two loads and the type 1657 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1658 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1659 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1660 bool has_load_barriers = bs->has_load_barriers(); 1661 1662 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) { 1663 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1664 Node* u = use->fast_out(i2); 1665 const Type* ut = u->bottom_type(); 1666 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) { 1667 if (has_load_barriers) { 1668 // Search for load barriers behind the load 1669 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1670 Node* b = u->fast_out(i3); 1671 if (bs->is_gc_barrier_node(b)) { 1672 _worklist.push(b); 1673 } 1674 } 1675 } 1676 _worklist.push(u); 1677 } 1678 } 1679 } 1680 } 1681 } 1682 1683 /** 1684 * Remove the speculative part of all types that we know of 1685 */ 1686 void PhaseIterGVN::remove_speculative_types() { 1687 assert(UseTypeSpeculation, "speculation is off"); 1688 for (uint i = 0; i < _types.Size(); i++) { 1689 const Type* t = _types.fast_lookup(i); 1690 if (t != NULL) { 1691 _types.map(i, t->remove_speculative()); 1692 } 1693 } 1694 _table.check_no_speculative_types(); 1695 } 1696 1697 //============================================================================= 1698 #ifndef PRODUCT 1699 uint PhaseCCP::_total_invokes = 0; 1700 uint PhaseCCP::_total_constants = 0; 1701 #endif 1702 //------------------------------PhaseCCP--------------------------------------- 1703 // Conditional Constant Propagation, ala Wegman & Zadeck 1704 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1705 NOT_PRODUCT( clear_constants(); ) 1706 assert( _worklist.size() == 0, "" ); 1707 // Clear out _nodes from IterGVN. Must be clear to transform call. 1708 _nodes.clear(); // Clear out from IterGVN 1709 analyze(); 1710 } 1711 1712 #ifndef PRODUCT 1713 //------------------------------~PhaseCCP-------------------------------------- 1714 PhaseCCP::~PhaseCCP() { 1715 inc_invokes(); 1716 _total_constants += count_constants(); 1717 } 1718 #endif 1719 1720 1721 #ifdef ASSERT 1722 static bool ccp_type_widens(const Type* t, const Type* t0) { 1723 assert(t->meet(t0) == t, "Not monotonic"); 1724 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1725 case Type::Int: 1726 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1727 break; 1728 case Type::Long: 1729 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1730 break; 1731 default: 1732 break; 1733 } 1734 return true; 1735 } 1736 #endif //ASSERT 1737 1738 //------------------------------analyze---------------------------------------- 1739 void PhaseCCP::analyze() { 1740 // Initialize all types to TOP, optimistic analysis 1741 for (int i = C->unique() - 1; i >= 0; i--) { 1742 _types.map(i,Type::TOP); 1743 } 1744 1745 // Push root onto worklist 1746 Unique_Node_List worklist; 1747 worklist.push(C->root()); 1748 1749 // Pull from worklist; compute new value; push changes out. 1750 // This loop is the meat of CCP. 1751 while( worklist.size() ) { 1752 Node *n = worklist.pop(); 1753 const Type *t = n->Value(this); 1754 if (t != type(n)) { 1755 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1756 #ifndef PRODUCT 1757 if( TracePhaseCCP ) { 1758 t->dump(); 1759 do { tty->print("\t"); } while (tty->position() < 16); 1760 n->dump(); 1761 } 1762 #endif 1763 set_type(n, t); 1764 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1765 Node* m = n->fast_out(i); // Get user 1766 if (m->is_Region()) { // New path to Region? Must recheck Phis too 1767 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1768 Node* p = m->fast_out(i2); // Propagate changes to uses 1769 if (p->bottom_type() != type(p)) { // If not already bottomed out 1770 worklist.push(p); // Propagate change to user 1771 } 1772 } 1773 } 1774 // If we changed the receiver type to a call, we need to revisit 1775 // the Catch following the call. It's looking for a non-NULL 1776 // receiver to know when to enable the regular fall-through path 1777 // in addition to the NullPtrException path 1778 if (m->is_Call()) { 1779 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1780 Node* p = m->fast_out(i2); // Propagate changes to uses 1781 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) { 1782 worklist.push(p->unique_out()); 1783 } 1784 } 1785 } 1786 if (m->bottom_type() != type(m)) { // If not already bottomed out 1787 worklist.push(m); // Propagate change to user 1788 } 1789 1790 // CmpU nodes can get their type information from two nodes up in the 1791 // graph (instead of from the nodes immediately above). Make sure they 1792 // are added to the worklist if nodes they depend on are updated, since 1793 // they could be missed and get wrong types otherwise. 1794 uint m_op = m->Opcode(); 1795 if (m_op == Op_AddI || m_op == Op_SubI) { 1796 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1797 Node* p = m->fast_out(i2); // Propagate changes to uses 1798 if (p->Opcode() == Op_CmpU) { 1799 // Got a CmpU which might need the new type information from node n. 1800 if(p->bottom_type() != type(p)) { // If not already bottomed out 1801 worklist.push(p); // Propagate change to user 1802 } 1803 } 1804 } 1805 } 1806 // If n is used in a counted loop exit condition then the type 1807 // of the counted loop's Phi depends on the type of n. See 1808 // PhiNode::Value(). 1809 if (m_op == Op_CmpI) { 1810 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n); 1811 if (phi != NULL) { 1812 worklist.push(phi); 1813 } 1814 } 1815 // Loading the java mirror from a Klass requires two loads and the type 1816 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1817 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1818 bool has_load_barriers = bs->has_load_barriers(); 1819 1820 if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) { 1821 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1822 Node* u = m->fast_out(i2); 1823 const Type* ut = u->bottom_type(); 1824 if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) { 1825 if (has_load_barriers) { 1826 // Search for load barriers behind the load 1827 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1828 Node* b = u->fast_out(i3); 1829 if (bs->is_gc_barrier_node(b)) { 1830 _worklist.push(b); 1831 } 1832 } 1833 } 1834 worklist.push(u); 1835 } 1836 } 1837 } 1838 } 1839 } 1840 } 1841 } 1842 1843 //------------------------------do_transform----------------------------------- 1844 // Top level driver for the recursive transformer 1845 void PhaseCCP::do_transform() { 1846 // Correct leaves of new-space Nodes; they point to old-space. 1847 C->set_root( transform(C->root())->as_Root() ); 1848 assert( C->top(), "missing TOP node" ); 1849 assert( C->root(), "missing root" ); 1850 } 1851 1852 //------------------------------transform-------------------------------------- 1853 // Given a Node in old-space, clone him into new-space. 1854 // Convert any of his old-space children into new-space children. 1855 Node *PhaseCCP::transform( Node *n ) { 1856 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1857 if( new_node != NULL ) 1858 return new_node; // Been there, done that, return old answer 1859 new_node = transform_once(n); // Check for constant 1860 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1861 1862 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1863 GrowableArray <Node *> trstack(C->live_nodes() >> 1); 1864 1865 trstack.push(new_node); // Process children of cloned node 1866 while ( trstack.is_nonempty() ) { 1867 Node *clone = trstack.pop(); 1868 uint cnt = clone->req(); 1869 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1870 Node *input = clone->in(i); 1871 if( input != NULL ) { // Ignore NULLs 1872 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1873 if( new_input == NULL ) { 1874 new_input = transform_once(input); // Check for constant 1875 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1876 trstack.push(new_input); 1877 } 1878 assert( new_input == clone->in(i), "insanity check"); 1879 } 1880 } 1881 } 1882 return new_node; 1883 } 1884 1885 1886 //------------------------------transform_once--------------------------------- 1887 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1888 Node *PhaseCCP::transform_once( Node *n ) { 1889 const Type *t = type(n); 1890 // Constant? Use constant Node instead 1891 if( t->singleton() ) { 1892 Node *nn = n; // Default is to return the original constant 1893 if( t == Type::TOP ) { 1894 // cache my top node on the Compile instance 1895 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1896 C->set_cached_top_node(ConNode::make(Type::TOP)); 1897 set_type(C->top(), Type::TOP); 1898 } 1899 nn = C->top(); 1900 } 1901 if( !n->is_Con() ) { 1902 if( t != Type::TOP ) { 1903 nn = makecon(t); // ConNode::make(t); 1904 NOT_PRODUCT( inc_constants(); ) 1905 } else if( n->is_Region() ) { // Unreachable region 1906 // Note: nn == C->top() 1907 n->set_req(0, NULL); // Cut selfreference 1908 bool progress = true; 1909 uint max = n->outcnt(); 1910 DUIterator i; 1911 while (progress) { 1912 progress = false; 1913 // Eagerly remove dead phis to avoid phis copies creation. 1914 for (i = n->outs(); n->has_out(i); i++) { 1915 Node* m = n->out(i); 1916 if (m->is_Phi()) { 1917 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1918 replace_node(m, nn); 1919 if (max != n->outcnt()) { 1920 progress = true; 1921 i = n->refresh_out_pos(i); 1922 max = n->outcnt(); 1923 } 1924 } 1925 } 1926 } 1927 } 1928 replace_node(n,nn); // Update DefUse edges for new constant 1929 } 1930 return nn; 1931 } 1932 1933 // If x is a TypeNode, capture any more-precise type permanently into Node 1934 if (t != n->bottom_type()) { 1935 hash_delete(n); // changing bottom type may force a rehash 1936 n->raise_bottom_type(t); 1937 _worklist.push(n); // n re-enters the hash table via the worklist 1938 } 1939 1940 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1941 switch( n->Opcode() ) { 1942 case Op_FastLock: // Revisit FastLocks for lock coarsening 1943 case Op_If: 1944 case Op_CountedLoopEnd: 1945 case Op_Region: 1946 case Op_Loop: 1947 case Op_CountedLoop: 1948 case Op_Conv2B: 1949 case Op_Opaque1: 1950 case Op_Opaque2: 1951 _worklist.push(n); 1952 break; 1953 default: 1954 break; 1955 } 1956 1957 return n; 1958 } 1959 1960 //---------------------------------saturate------------------------------------ 1961 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1962 const Type* limit_type) const { 1963 const Type* wide_type = new_type->widen(old_type, limit_type); 1964 if (wide_type != new_type) { // did we widen? 1965 // If so, we may have widened beyond the limit type. Clip it back down. 1966 new_type = wide_type->filter(limit_type); 1967 } 1968 return new_type; 1969 } 1970 1971 //------------------------------print_statistics------------------------------- 1972 #ifndef PRODUCT 1973 void PhaseCCP::print_statistics() { 1974 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1975 } 1976 #endif 1977 1978 1979 //============================================================================= 1980 #ifndef PRODUCT 1981 uint PhasePeephole::_total_peepholes = 0; 1982 #endif 1983 //------------------------------PhasePeephole---------------------------------- 1984 // Conditional Constant Propagation, ala Wegman & Zadeck 1985 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1986 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1987 NOT_PRODUCT( clear_peepholes(); ) 1988 } 1989 1990 #ifndef PRODUCT 1991 //------------------------------~PhasePeephole--------------------------------- 1992 PhasePeephole::~PhasePeephole() { 1993 _total_peepholes += count_peepholes(); 1994 } 1995 #endif 1996 1997 //------------------------------transform-------------------------------------- 1998 Node *PhasePeephole::transform( Node *n ) { 1999 ShouldNotCallThis(); 2000 return NULL; 2001 } 2002 2003 //------------------------------do_transform----------------------------------- 2004 void PhasePeephole::do_transform() { 2005 bool method_name_not_printed = true; 2006 2007 // Examine each basic block 2008 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 2009 Block* block = _cfg.get_block(block_number); 2010 bool block_not_printed = true; 2011 2012 // and each instruction within a block 2013 uint end_index = block->number_of_nodes(); 2014 // block->end_idx() not valid after PhaseRegAlloc 2015 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 2016 Node *n = block->get_node(instruction_index); 2017 if( n->is_Mach() ) { 2018 MachNode *m = n->as_Mach(); 2019 int deleted_count = 0; 2020 // check for peephole opportunities 2021 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count); 2022 if( m2 != NULL ) { 2023 #ifndef PRODUCT 2024 if( PrintOptoPeephole ) { 2025 // Print method, first time only 2026 if( C->method() && method_name_not_printed ) { 2027 C->method()->print_short_name(); tty->cr(); 2028 method_name_not_printed = false; 2029 } 2030 // Print this block 2031 if( Verbose && block_not_printed) { 2032 tty->print_cr("in block"); 2033 block->dump(); 2034 block_not_printed = false; 2035 } 2036 // Print instructions being deleted 2037 for( int i = (deleted_count - 1); i >= 0; --i ) { 2038 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 2039 } 2040 tty->print_cr("replaced with"); 2041 // Print new instruction 2042 m2->format(_regalloc); 2043 tty->print("\n\n"); 2044 } 2045 #endif 2046 // Remove old nodes from basic block and update instruction_index 2047 // (old nodes still exist and may have edges pointing to them 2048 // as register allocation info is stored in the allocator using 2049 // the node index to live range mappings.) 2050 uint safe_instruction_index = (instruction_index - deleted_count); 2051 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 2052 block->remove_node( instruction_index ); 2053 } 2054 // install new node after safe_instruction_index 2055 block->insert_node(m2, safe_instruction_index + 1); 2056 end_index = block->number_of_nodes() - 1; // Recompute new block size 2057 NOT_PRODUCT( inc_peepholes(); ) 2058 } 2059 } 2060 } 2061 } 2062 } 2063 2064 //------------------------------print_statistics------------------------------- 2065 #ifndef PRODUCT 2066 void PhasePeephole::print_statistics() { 2067 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2068 } 2069 #endif 2070 2071 2072 //============================================================================= 2073 //------------------------------set_req_X-------------------------------------- 2074 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2075 assert( is_not_dead(n), "can not use dead node"); 2076 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2077 Node *old = in(i); 2078 set_req(i, n); 2079 2080 // old goes dead? 2081 if( old ) { 2082 switch (old->outcnt()) { 2083 case 0: 2084 // Put into the worklist to kill later. We do not kill it now because the 2085 // recursive kill will delete the current node (this) if dead-loop exists 2086 if (!old->is_top()) 2087 igvn->_worklist.push( old ); 2088 break; 2089 case 1: 2090 if( old->is_Store() || old->has_special_unique_user() ) 2091 igvn->add_users_to_worklist( old ); 2092 break; 2093 case 2: 2094 if( old->is_Store() ) 2095 igvn->add_users_to_worklist( old ); 2096 if( old->Opcode() == Op_Region ) 2097 igvn->_worklist.push(old); 2098 break; 2099 case 3: 2100 if( old->Opcode() == Op_Region ) { 2101 igvn->_worklist.push(old); 2102 igvn->add_users_to_worklist( old ); 2103 } 2104 break; 2105 default: 2106 break; 2107 } 2108 2109 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old); 2110 } 2111 2112 } 2113 2114 //-------------------------------replace_by----------------------------------- 2115 // Using def-use info, replace one node for another. Follow the def-use info 2116 // to all users of the OLD node. Then make all uses point to the NEW node. 2117 void Node::replace_by(Node *new_node) { 2118 assert(!is_top(), "top node has no DU info"); 2119 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2120 Node* use = last_out(i); 2121 uint uses_found = 0; 2122 for (uint j = 0; j < use->len(); j++) { 2123 if (use->in(j) == this) { 2124 if (j < use->req()) 2125 use->set_req(j, new_node); 2126 else use->set_prec(j, new_node); 2127 uses_found++; 2128 } 2129 } 2130 i -= uses_found; // we deleted 1 or more copies of this edge 2131 } 2132 } 2133 2134 //============================================================================= 2135 //----------------------------------------------------------------------------- 2136 void Type_Array::grow( uint i ) { 2137 if( !_max ) { 2138 _max = 1; 2139 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2140 _types[0] = NULL; 2141 } 2142 uint old = _max; 2143 while( i >= _max ) _max <<= 1; // Double to fit 2144 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2145 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2146 } 2147 2148 //------------------------------dump------------------------------------------- 2149 #ifndef PRODUCT 2150 void Type_Array::dump() const { 2151 uint max = Size(); 2152 for( uint i = 0; i < max; i++ ) { 2153 if( _types[i] != NULL ) { 2154 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2155 } 2156 } 2157 } 2158 #endif