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