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