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