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 int 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 int 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 // Sanity check that fails if we ever decide to execute this phase after EA 495 assert(!n->is_Phi() || n->as_Phi()->inst_mem_id() == -1, "should not be linked to data Phi"); 496 const Type* type = gvn->type_or_null(n); 497 new_type_array.map(current_idx, type); 498 499 bool in_worklist = false; 500 if (worklist->member(n)) { 501 in_worklist = true; 502 } 503 504 n->set_idx(current_idx); // Update node ID. 505 506 if (in_worklist) { 507 new_worklist->push(n); 508 } 509 510 current_idx++; 511 } 512 513 assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist"); 514 assert(live_node_count == current_idx, "all live nodes must be processed"); 515 516 // Replace the compiler's type information with the updated type information. 517 gvn->replace_types(new_type_array); 518 519 // Update the unique node count of the compilation to the number of currently live nodes. 520 C->set_unique(live_node_count); 521 522 // Set the dead node count to 0 and reset dead node list. 523 C->reset_dead_node_list(); 524 } 525 526 527 //============================================================================= 528 //------------------------------PhaseTransform--------------------------------- 529 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 530 _arena(Thread::current()->resource_area()), 531 _nodes(_arena), 532 _types(_arena) 533 { 534 init_con_caches(); 535 #ifndef PRODUCT 536 clear_progress(); 537 clear_transforms(); 538 set_allow_progress(true); 539 #endif 540 // Force allocation for currently existing nodes 541 _types.map(C->unique(), NULL); 542 } 543 544 //------------------------------PhaseTransform--------------------------------- 545 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 546 _arena(arena), 547 _nodes(arena), 548 _types(arena) 549 { 550 init_con_caches(); 551 #ifndef PRODUCT 552 clear_progress(); 553 clear_transforms(); 554 set_allow_progress(true); 555 #endif 556 // Force allocation for currently existing nodes 557 _types.map(C->unique(), NULL); 558 } 559 560 //------------------------------PhaseTransform--------------------------------- 561 // Initialize with previously generated type information 562 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 563 _arena(pt->_arena), 564 _nodes(pt->_nodes), 565 _types(pt->_types) 566 { 567 init_con_caches(); 568 #ifndef PRODUCT 569 clear_progress(); 570 clear_transforms(); 571 set_allow_progress(true); 572 #endif 573 } 574 575 void PhaseTransform::init_con_caches() { 576 memset(_icons,0,sizeof(_icons)); 577 memset(_lcons,0,sizeof(_lcons)); 578 memset(_zcons,0,sizeof(_zcons)); 579 } 580 581 582 //--------------------------------find_int_type-------------------------------- 583 const TypeInt* PhaseTransform::find_int_type(Node* n) { 584 if (n == NULL) return NULL; 585 // Call type_or_null(n) to determine node's type since we might be in 586 // parse phase and call n->Value() may return wrong type. 587 // (For example, a phi node at the beginning of loop parsing is not ready.) 588 const Type* t = type_or_null(n); 589 if (t == NULL) return NULL; 590 return t->isa_int(); 591 } 592 593 594 //-------------------------------find_long_type-------------------------------- 595 const TypeLong* PhaseTransform::find_long_type(Node* n) { 596 if (n == NULL) return NULL; 597 // (See comment above on type_or_null.) 598 const Type* t = type_or_null(n); 599 if (t == NULL) return NULL; 600 return t->isa_long(); 601 } 602 603 604 #ifndef PRODUCT 605 void PhaseTransform::dump_old2new_map() const { 606 _nodes.dump(); 607 } 608 609 void PhaseTransform::dump_new( uint nidx ) const { 610 for( uint i=0; i<_nodes.Size(); i++ ) 611 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 612 _nodes[i]->dump(); 613 tty->cr(); 614 tty->print_cr("Old index= %d",i); 615 return; 616 } 617 tty->print_cr("Node %d not found in the new indices", nidx); 618 } 619 620 //------------------------------dump_types------------------------------------- 621 void PhaseTransform::dump_types( ) const { 622 _types.dump(); 623 } 624 625 //------------------------------dump_nodes_and_types--------------------------- 626 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 627 VectorSet visited(Thread::current()->resource_area()); 628 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 629 } 630 631 //------------------------------dump_nodes_and_types_recur--------------------- 632 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 633 if( !n ) return; 634 if( depth == 0 ) return; 635 if( visited.test_set(n->_idx) ) return; 636 for( uint i=0; i<n->len(); i++ ) { 637 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 638 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 639 } 640 n->dump(); 641 if (type_or_null(n) != NULL) { 642 tty->print(" "); type(n)->dump(); tty->cr(); 643 } 644 } 645 646 #endif 647 648 649 //============================================================================= 650 //------------------------------PhaseValues------------------------------------ 651 // Set minimum table size to "255" 652 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 653 NOT_PRODUCT( clear_new_values(); ) 654 } 655 656 //------------------------------PhaseValues------------------------------------ 657 // Set minimum table size to "255" 658 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 659 _table(&ptv->_table) { 660 NOT_PRODUCT( clear_new_values(); ) 661 } 662 663 //------------------------------PhaseValues------------------------------------ 664 // Used by +VerifyOpto. Clear out hash table but copy _types array. 665 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 666 _table(ptv->arena(),ptv->_table.size()) { 667 NOT_PRODUCT( clear_new_values(); ) 668 } 669 670 //------------------------------~PhaseValues----------------------------------- 671 #ifndef PRODUCT 672 PhaseValues::~PhaseValues() { 673 _table.dump(); 674 675 // Statistics for value progress and efficiency 676 if( PrintCompilation && Verbose && WizardMode ) { 677 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 678 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 679 if( made_transforms() != 0 ) { 680 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 681 } else { 682 tty->cr(); 683 } 684 } 685 } 686 #endif 687 688 //------------------------------makecon---------------------------------------- 689 ConNode* PhaseTransform::makecon(const Type *t) { 690 assert(t->singleton(), "must be a constant"); 691 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 692 switch (t->base()) { // fast paths 693 case Type::Half: 694 case Type::Top: return (ConNode*) C->top(); 695 case Type::Int: return intcon( t->is_int()->get_con() ); 696 case Type::Long: return longcon( t->is_long()->get_con() ); 697 } 698 if (t->is_zero_type()) 699 return zerocon(t->basic_type()); 700 return uncached_makecon(t); 701 } 702 703 //--------------------------uncached_makecon----------------------------------- 704 // Make an idealized constant - one of ConINode, ConPNode, etc. 705 ConNode* PhaseValues::uncached_makecon(const Type *t) { 706 assert(t->singleton(), "must be a constant"); 707 ConNode* x = ConNode::make(t); 708 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 709 if (k == NULL) { 710 set_type(x, t); // Missed, provide type mapping 711 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 712 if (nna != NULL) { 713 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 714 loc->clear(); // do not put debug info on constants 715 } 716 } else { 717 x->destruct(); // Hit, destroy duplicate constant 718 x = k; // use existing constant 719 } 720 return x; 721 } 722 723 //------------------------------intcon----------------------------------------- 724 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 725 ConINode* PhaseTransform::intcon(int i) { 726 // Small integer? Check cache! Check that cached node is not dead 727 if (i >= _icon_min && i <= _icon_max) { 728 ConINode* icon = _icons[i-_icon_min]; 729 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 730 return icon; 731 } 732 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 733 assert(icon->is_Con(), ""); 734 if (i >= _icon_min && i <= _icon_max) 735 _icons[i-_icon_min] = icon; // Cache small integers 736 return icon; 737 } 738 739 //------------------------------longcon---------------------------------------- 740 // Fast long constant. 741 ConLNode* PhaseTransform::longcon(jlong l) { 742 // Small integer? Check cache! Check that cached node is not dead 743 if (l >= _lcon_min && l <= _lcon_max) { 744 ConLNode* lcon = _lcons[l-_lcon_min]; 745 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 746 return lcon; 747 } 748 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 749 assert(lcon->is_Con(), ""); 750 if (l >= _lcon_min && l <= _lcon_max) 751 _lcons[l-_lcon_min] = lcon; // Cache small integers 752 return lcon; 753 } 754 755 //------------------------------zerocon----------------------------------------- 756 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 757 ConNode* PhaseTransform::zerocon(BasicType bt) { 758 assert((uint)bt <= _zcon_max, "domain check"); 759 ConNode* zcon = _zcons[bt]; 760 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 761 return zcon; 762 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 763 _zcons[bt] = zcon; 764 return zcon; 765 } 766 767 768 769 //============================================================================= 770 //------------------------------transform-------------------------------------- 771 // Return a node which computes the same function as this node, but in a 772 // faster or cheaper fashion. 773 Node *PhaseGVN::transform( Node *n ) { 774 return transform_no_reclaim(n); 775 } 776 777 //------------------------------transform-------------------------------------- 778 // Return a node which computes the same function as this node, but 779 // in a faster or cheaper fashion. 780 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 781 NOT_PRODUCT( set_transforms(); ) 782 783 // Apply the Ideal call in a loop until it no longer applies 784 Node *k = n; 785 NOT_PRODUCT( uint loop_count = 0; ) 786 while( 1 ) { 787 Node *i = k->Ideal(this, /*can_reshape=*/false); 788 if( !i ) break; 789 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 790 k = i; 791 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 792 } 793 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 794 795 796 // If brand new node, make space in type array. 797 ensure_type_or_null(k); 798 799 // Since I just called 'Value' to compute the set of run-time values 800 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 801 // cache Value. Later requests for the local phase->type of this Node can 802 // use the cached Value instead of suffering with 'bottom_type'. 803 const Type *t = k->Value(this); // Get runtime Value set 804 assert(t != NULL, "value sanity"); 805 if (type_or_null(k) != t) { 806 #ifndef PRODUCT 807 // Do not count initial visit to node as a transformation 808 if (type_or_null(k) == NULL) { 809 inc_new_values(); 810 set_progress(); 811 } 812 #endif 813 set_type(k, t); 814 // If k is a TypeNode, capture any more-precise type permanently into Node 815 k->raise_bottom_type(t); 816 } 817 818 if( t->singleton() && !k->is_Con() ) { 819 NOT_PRODUCT( set_progress(); ) 820 return makecon(t); // Turn into a constant 821 } 822 823 // Now check for Identities 824 Node *i = k->Identity(this); // Look for a nearby replacement 825 if( i != k ) { // Found? Return replacement! 826 NOT_PRODUCT( set_progress(); ) 827 return i; 828 } 829 830 // Global Value Numbering 831 i = hash_find_insert(k); // Insert if new 832 if( i && (i != k) ) { 833 // Return the pre-existing node 834 NOT_PRODUCT( set_progress(); ) 835 return i; 836 } 837 838 // Return Idealized original 839 return k; 840 } 841 842 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 843 if (d->is_top() || n->is_top()) { 844 return false; 845 } 846 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 847 int i = 0; 848 while (d != n) { 849 n = IfNode::up_one_dom(n, linear_only); 850 i++; 851 if (n == NULL || i >= 10) { 852 return false; 853 } 854 } 855 return true; 856 } 857 858 #ifdef ASSERT 859 //------------------------------dead_loop_check-------------------------------- 860 // Check for a simple dead loop when a data node references itself directly 861 // or through an other data node excluding cons and phis. 862 void PhaseGVN::dead_loop_check( Node *n ) { 863 // Phi may reference itself in a loop 864 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 865 // Do 2 levels check and only data inputs. 866 bool no_dead_loop = true; 867 uint cnt = n->req(); 868 for (uint i = 1; i < cnt && no_dead_loop; i++) { 869 Node *in = n->in(i); 870 if (in == n) { 871 no_dead_loop = false; 872 } else if (in != NULL && !in->is_dead_loop_safe()) { 873 uint icnt = in->req(); 874 for (uint j = 1; j < icnt && no_dead_loop; j++) { 875 if (in->in(j) == n || in->in(j) == in) 876 no_dead_loop = false; 877 } 878 } 879 } 880 if (!no_dead_loop) n->dump(3); 881 assert(no_dead_loop, "dead loop detected"); 882 } 883 } 884 #endif 885 886 //============================================================================= 887 //------------------------------PhaseIterGVN----------------------------------- 888 // Initialize hash table to fresh and clean for +VerifyOpto 889 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ), 890 _stack(C->live_nodes() >> 1), 891 _delay_transform(false) { 892 } 893 894 //------------------------------PhaseIterGVN----------------------------------- 895 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 896 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 897 _worklist( igvn->_worklist ), 898 _stack( igvn->_stack ), 899 _delay_transform(igvn->_delay_transform) 900 { 901 } 902 903 //------------------------------PhaseIterGVN----------------------------------- 904 // Initialize with previous PhaseGVN info from Parser 905 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 906 _worklist(*C->for_igvn()), 907 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that 908 // the constructor is used in incremental inlining, this consumes too much memory: 909 // _stack(C->live_nodes() >> 1), 910 // So, as a band-aid, we replace this by: 911 _stack(C->comp_arena(), 32), 912 _delay_transform(false) 913 { 914 uint max; 915 916 // Dead nodes in the hash table inherited from GVN were not treated as 917 // roots during def-use info creation; hence they represent an invisible 918 // use. Clear them out. 919 max = _table.size(); 920 for( uint i = 0; i < max; ++i ) { 921 Node *n = _table.at(i); 922 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 923 if( n->is_top() ) continue; 924 assert( false, "Parse::remove_useless_nodes missed this node"); 925 hash_delete(n); 926 } 927 } 928 929 // Any Phis or Regions on the worklist probably had uses that could not 930 // make more progress because the uses were made while the Phis and Regions 931 // were in half-built states. Put all uses of Phis and Regions on worklist. 932 max = _worklist.size(); 933 for( uint j = 0; j < max; j++ ) { 934 Node *n = _worklist.at(j); 935 uint uop = n->Opcode(); 936 if( uop == Op_Phi || uop == Op_Region || 937 n->is_Type() || 938 n->is_Mem() ) 939 add_users_to_worklist(n); 940 } 941 } 942 943 /** 944 * Initialize worklist for each node. 945 */ 946 void PhaseIterGVN::init_worklist(Node* first) { 947 Unique_Node_List to_process; 948 to_process.push(first); 949 950 while (to_process.size() > 0) { 951 Node* n = to_process.pop(); 952 if (!_worklist.member(n)) { 953 _worklist.push(n); 954 955 uint cnt = n->req(); 956 for(uint i = 0; i < cnt; i++) { 957 Node* m = n->in(i); 958 if (m != NULL) { 959 to_process.push(m); 960 } 961 } 962 } 963 } 964 } 965 966 #ifndef PRODUCT 967 void PhaseIterGVN::verify_step(Node* n) { 968 if (VerifyIterativeGVN) { 969 _verify_window[_verify_counter % _verify_window_size] = n; 970 ++_verify_counter; 971 ResourceMark rm; 972 ResourceArea* area = Thread::current()->resource_area(); 973 VectorSet old_space(area), new_space(area); 974 if (C->unique() < 1000 || 975 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 976 ++_verify_full_passes; 977 Node::verify_recur(C->root(), -1, old_space, new_space); 978 } 979 const int verify_depth = 4; 980 for ( int i = 0; i < _verify_window_size; i++ ) { 981 Node* n = _verify_window[i]; 982 if ( n == NULL ) continue; 983 if( n->in(0) == NodeSentinel ) { // xform_idom 984 _verify_window[i] = n->in(1); 985 --i; continue; 986 } 987 // Typical fanout is 1-2, so this call visits about 6 nodes. 988 Node::verify_recur(n, verify_depth, old_space, new_space); 989 } 990 } 991 } 992 993 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 994 if (TraceIterativeGVN) { 995 uint wlsize = _worklist.size(); 996 const Type* newtype = type_or_null(n); 997 if (nn != n) { 998 // print old node 999 tty->print("< "); 1000 if (oldtype != newtype && oldtype != NULL) { 1001 oldtype->dump(); 1002 } 1003 do { tty->print("\t"); } while (tty->position() < 16); 1004 tty->print("<"); 1005 n->dump(); 1006 } 1007 if (oldtype != newtype || nn != n) { 1008 // print new node and/or new type 1009 if (oldtype == NULL) { 1010 tty->print("* "); 1011 } else if (nn != n) { 1012 tty->print("> "); 1013 } else { 1014 tty->print("= "); 1015 } 1016 if (newtype == NULL) { 1017 tty->print("null"); 1018 } else { 1019 newtype->dump(); 1020 } 1021 do { tty->print("\t"); } while (tty->position() < 16); 1022 nn->dump(); 1023 } 1024 if (Verbose && wlsize < _worklist.size()) { 1025 tty->print(" Push {"); 1026 while (wlsize != _worklist.size()) { 1027 Node* pushed = _worklist.at(wlsize++); 1028 tty->print(" %d", pushed->_idx); 1029 } 1030 tty->print_cr(" }"); 1031 } 1032 if (nn != n) { 1033 // ignore n, it might be subsumed 1034 verify_step((Node*) NULL); 1035 } 1036 } 1037 } 1038 1039 void PhaseIterGVN::init_verifyPhaseIterGVN() { 1040 _verify_counter = 0; 1041 _verify_full_passes = 0; 1042 for (int i = 0; i < _verify_window_size; i++) { 1043 _verify_window[i] = NULL; 1044 } 1045 #ifdef ASSERT 1046 // Verify that all modified nodes are on _worklist 1047 Unique_Node_List* modified_list = C->modified_nodes(); 1048 while (modified_list != NULL && modified_list->size()) { 1049 Node* n = modified_list->pop(); 1050 if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) { 1051 n->dump(); 1052 assert(false, "modified node is not on IGVN._worklist"); 1053 } 1054 } 1055 #endif 1056 } 1057 1058 void PhaseIterGVN::verify_PhaseIterGVN() { 1059 #ifdef ASSERT 1060 // Verify nodes with changed inputs. 1061 Unique_Node_List* modified_list = C->modified_nodes(); 1062 while (modified_list != NULL && modified_list->size()) { 1063 Node* n = modified_list->pop(); 1064 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes 1065 n->dump(); 1066 assert(false, "modified node was not processed by IGVN.transform_old()"); 1067 } 1068 } 1069 #endif 1070 1071 C->verify_graph_edges(); 1072 if( VerifyOpto && allow_progress() ) { 1073 // Must turn off allow_progress to enable assert and break recursion 1074 C->root()->verify(); 1075 { // Check if any progress was missed using IterGVN 1076 // Def-Use info enables transformations not attempted in wash-pass 1077 // e.g. Region/Phi cleanup, ... 1078 // Null-check elision -- may not have reached fixpoint 1079 // do not propagate to dominated nodes 1080 ResourceMark rm; 1081 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 1082 // Fill worklist completely 1083 igvn2.init_worklist(C->root()); 1084 1085 igvn2.set_allow_progress(false); 1086 igvn2.optimize(); 1087 igvn2.set_allow_progress(true); 1088 } 1089 } 1090 if (VerifyIterativeGVN && PrintOpto) { 1091 if (_verify_counter == _verify_full_passes) { 1092 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 1093 (int) _verify_full_passes); 1094 } else { 1095 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 1096 (int) _verify_counter, (int) _verify_full_passes); 1097 } 1098 } 1099 1100 #ifdef ASSERT 1101 while (modified_list->size()) { 1102 Node* n = modified_list->pop(); 1103 n->dump(); 1104 assert(false, "VerifyIterativeGVN: new modified node was added"); 1105 } 1106 #endif 1107 } 1108 #endif /* PRODUCT */ 1109 1110 #ifdef ASSERT 1111 /** 1112 * Dumps information that can help to debug the problem. A debug 1113 * build fails with an assert. 1114 */ 1115 void PhaseIterGVN::dump_infinite_loop_info(Node* n) { 1116 n->dump(4); 1117 _worklist.dump(); 1118 assert(false, "infinite loop in PhaseIterGVN::optimize"); 1119 } 1120 1121 /** 1122 * Prints out information about IGVN if the 'verbose' option is used. 1123 */ 1124 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1125 if (TraceIterativeGVN && Verbose) { 1126 tty->print(" Pop "); 1127 n->dump(); 1128 if ((num_processed % 100) == 0) { 1129 _worklist.print_set(); 1130 } 1131 } 1132 } 1133 #endif /* ASSERT */ 1134 1135 void PhaseIterGVN::optimize() { 1136 DEBUG_ONLY(uint num_processed = 0;) 1137 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1138 1139 uint loop_count = 0; 1140 // Pull from worklist and transform the node. If the node has changed, 1141 // update edge info and put uses on worklist. 1142 while(_worklist.size()) { 1143 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1144 return; 1145 } 1146 Node* n = _worklist.pop(); 1147 if (++loop_count >= K * C->live_nodes()) { 1148 DEBUG_ONLY(dump_infinite_loop_info(n);) 1149 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1150 return; 1151 } 1152 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1153 if (n->outcnt() != 0) { 1154 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1155 // Do the transformation 1156 Node* nn = transform_old(n); 1157 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1158 } else if (!n->is_top()) { 1159 remove_dead_node(n); 1160 } 1161 } 1162 NOT_PRODUCT(verify_PhaseIterGVN();) 1163 } 1164 1165 1166 /** 1167 * Register a new node with the optimizer. Update the types array, the def-use 1168 * info. Put on worklist. 1169 */ 1170 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1171 set_type_bottom(n); 1172 _worklist.push(n); 1173 if (orig != NULL) C->copy_node_notes_to(n, orig); 1174 return n; 1175 } 1176 1177 //------------------------------transform-------------------------------------- 1178 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1179 Node *PhaseIterGVN::transform( Node *n ) { 1180 if (_delay_transform) { 1181 // Register the node but don't optimize for now 1182 register_new_node_with_optimizer(n); 1183 return n; 1184 } 1185 1186 // If brand new node, make space in type array, and give it a type. 1187 ensure_type_or_null(n); 1188 if (type_or_null(n) == NULL) { 1189 set_type_bottom(n); 1190 } 1191 1192 return transform_old(n); 1193 } 1194 1195 Node *PhaseIterGVN::transform_old(Node* n) { 1196 DEBUG_ONLY(uint loop_count = 0;); 1197 NOT_PRODUCT(set_transforms()); 1198 1199 // Remove 'n' from hash table in case it gets modified 1200 _table.hash_delete(n); 1201 if (VerifyIterativeGVN) { 1202 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1203 } 1204 1205 // Apply the Ideal call in a loop until it no longer applies 1206 Node* k = n; 1207 DEBUG_ONLY(dead_loop_check(k);) 1208 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1209 C->remove_modified_node(k); 1210 Node* i = k->Ideal(this, /*can_reshape=*/true); 1211 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1212 #ifndef PRODUCT 1213 verify_step(k); 1214 if (i && VerifyOpto ) { 1215 if (!allow_progress()) { 1216 if (i->is_Add() && (i->outcnt() == 1)) { 1217 // Switched input to left side because this is the only use 1218 } else if (i->is_If() && (i->in(0) == NULL)) { 1219 // This IF is dead because it is dominated by an equivalent IF When 1220 // dominating if changed, info is not propagated sparsely to 'this' 1221 // Propagating this info further will spuriously identify other 1222 // progress. 1223 return i; 1224 } else 1225 set_progress(); 1226 } else { 1227 set_progress(); 1228 } 1229 } 1230 #endif 1231 1232 while (i != NULL) { 1233 #ifdef ASSERT 1234 if (loop_count >= K) { 1235 dump_infinite_loop_info(i); 1236 } 1237 loop_count++; 1238 #endif 1239 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1240 // Made a change; put users of original Node on worklist 1241 add_users_to_worklist(k); 1242 // Replacing root of transform tree? 1243 if (k != i) { 1244 // Make users of old Node now use new. 1245 subsume_node(k, i); 1246 k = i; 1247 } 1248 DEBUG_ONLY(dead_loop_check(k);) 1249 // Try idealizing again 1250 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1251 C->remove_modified_node(k); 1252 i = k->Ideal(this, /*can_reshape=*/true); 1253 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1254 #ifndef PRODUCT 1255 verify_step(k); 1256 if (i && VerifyOpto) { 1257 set_progress(); 1258 } 1259 #endif 1260 } 1261 1262 // If brand new node, make space in type array. 1263 ensure_type_or_null(k); 1264 1265 // See what kind of values 'k' takes on at runtime 1266 const Type* t = k->Value(this); 1267 assert(t != NULL, "value sanity"); 1268 1269 // Since I just called 'Value' to compute the set of run-time values 1270 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1271 // cache Value. Later requests for the local phase->type of this Node can 1272 // use the cached Value instead of suffering with 'bottom_type'. 1273 if (type_or_null(k) != t) { 1274 #ifndef PRODUCT 1275 inc_new_values(); 1276 set_progress(); 1277 #endif 1278 set_type(k, t); 1279 // If k is a TypeNode, capture any more-precise type permanently into Node 1280 k->raise_bottom_type(t); 1281 // Move users of node to worklist 1282 add_users_to_worklist(k); 1283 } 1284 // If 'k' computes a constant, replace it with a constant 1285 if (t->singleton() && !k->is_Con()) { 1286 NOT_PRODUCT(set_progress();) 1287 Node* con = makecon(t); // Make a constant 1288 add_users_to_worklist(k); 1289 subsume_node(k, con); // Everybody using k now uses con 1290 return con; 1291 } 1292 1293 // Now check for Identities 1294 i = k->Identity(this); // Look for a nearby replacement 1295 if (i != k) { // Found? Return replacement! 1296 NOT_PRODUCT(set_progress();) 1297 add_users_to_worklist(k); 1298 subsume_node(k, i); // Everybody using k now uses i 1299 return i; 1300 } 1301 1302 // Global Value Numbering 1303 i = hash_find_insert(k); // Check for pre-existing node 1304 if (i && (i != k)) { 1305 // Return the pre-existing node if it isn't dead 1306 NOT_PRODUCT(set_progress();) 1307 add_users_to_worklist(k); 1308 subsume_node(k, i); // Everybody using k now uses i 1309 return i; 1310 } 1311 1312 // Return Idealized original 1313 return k; 1314 } 1315 1316 //---------------------------------saturate------------------------------------ 1317 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1318 const Type* limit_type) const { 1319 return new_type->narrow(old_type); 1320 } 1321 1322 //------------------------------remove_globally_dead_node---------------------- 1323 // Kill a globally dead Node. All uses are also globally dead and are 1324 // aggressively trimmed. 1325 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1326 enum DeleteProgress { 1327 PROCESS_INPUTS, 1328 PROCESS_OUTPUTS 1329 }; 1330 assert(_stack.is_empty(), "not empty"); 1331 _stack.push(dead, PROCESS_INPUTS); 1332 1333 while (_stack.is_nonempty()) { 1334 dead = _stack.node(); 1335 uint progress_state = _stack.index(); 1336 assert(dead != C->root(), "killing root, eh?"); 1337 assert(!dead->is_top(), "add check for top when pushing"); 1338 NOT_PRODUCT( set_progress(); ) 1339 if (progress_state == PROCESS_INPUTS) { 1340 // After following inputs, continue to outputs 1341 _stack.set_index(PROCESS_OUTPUTS); 1342 if (!dead->is_Con()) { // Don't kill cons but uses 1343 bool recurse = false; 1344 // Remove from hash table 1345 _table.hash_delete( dead ); 1346 // Smash all inputs to 'dead', isolating him completely 1347 for (uint i = 0; i < dead->req(); i++) { 1348 Node *in = dead->in(i); 1349 if (in != NULL && in != C->top()) { // Points to something? 1350 int nrep = dead->replace_edge(in, NULL); // Kill edges 1351 assert((nrep > 0), "sanity"); 1352 if (in->outcnt() == 0) { // Made input go dead? 1353 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1354 recurse = true; 1355 } else if (in->outcnt() == 1 && 1356 in->has_special_unique_user()) { 1357 _worklist.push(in->unique_out()); 1358 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1359 if (in->Opcode() == Op_Region) { 1360 _worklist.push(in); 1361 } else if (in->is_Store()) { 1362 DUIterator_Fast imax, i = in->fast_outs(imax); 1363 _worklist.push(in->fast_out(i)); 1364 i++; 1365 if (in->outcnt() == 2) { 1366 _worklist.push(in->fast_out(i)); 1367 i++; 1368 } 1369 assert(!(i < imax), "sanity"); 1370 } 1371 } 1372 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1373 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1374 // A Load that directly follows an InitializeNode is 1375 // going away. The Stores that follow are candidates 1376 // again to be captured by the InitializeNode. 1377 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1378 Node *n = in->fast_out(j); 1379 if (n->is_Store()) { 1380 _worklist.push(n); 1381 } 1382 } 1383 } 1384 } // if (in != NULL && in != C->top()) 1385 } // for (uint i = 0; i < dead->req(); i++) 1386 if (recurse) { 1387 continue; 1388 } 1389 } // if (!dead->is_Con()) 1390 } // if (progress_state == PROCESS_INPUTS) 1391 1392 // Aggressively kill globally dead uses 1393 // (Rather than pushing all the outs at once, we push one at a time, 1394 // plus the parent to resume later, because of the indefinite number 1395 // of edge deletions per loop trip.) 1396 if (dead->outcnt() > 0) { 1397 // Recursively remove output edges 1398 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1399 } else { 1400 // Finished disconnecting all input and output edges. 1401 _stack.pop(); 1402 // Remove dead node from iterative worklist 1403 _worklist.remove(dead); 1404 C->remove_modified_node(dead); 1405 // Constant node that has no out-edges and has only one in-edge from 1406 // root is usually dead. However, sometimes reshaping walk makes 1407 // it reachable by adding use edges. So, we will NOT count Con nodes 1408 // as dead to be conservative about the dead node count at any 1409 // given time. 1410 if (!dead->is_Con()) { 1411 C->record_dead_node(dead->_idx); 1412 } 1413 if (dead->is_macro()) { 1414 C->remove_macro_node(dead); 1415 } 1416 if (dead->is_expensive()) { 1417 C->remove_expensive_node(dead); 1418 } 1419 CastIINode* cast = dead->isa_CastII(); 1420 if (cast != NULL && cast->has_range_check()) { 1421 C->remove_range_check_cast(cast); 1422 } 1423 if (dead->is_ValueTypePtr()) { 1424 C->remove_value_type_ptr(dead->as_ValueTypePtr()); 1425 } 1426 } 1427 } // while (_stack.is_nonempty()) 1428 } 1429 1430 //------------------------------subsume_node----------------------------------- 1431 // Remove users from node 'old' and add them to node 'nn'. 1432 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1433 assert( old != hash_find(old), "should already been removed" ); 1434 assert( old != C->top(), "cannot subsume top node"); 1435 // Copy debug or profile information to the new version: 1436 C->copy_node_notes_to(nn, old); 1437 // Move users of node 'old' to node 'nn' 1438 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1439 Node* use = old->last_out(i); // for each use... 1440 // use might need re-hashing (but it won't if it's a new node) 1441 rehash_node_delayed(use); 1442 // Update use-def info as well 1443 // We remove all occurrences of old within use->in, 1444 // so as to avoid rehashing any node more than once. 1445 // The hash table probe swamps any outer loop overhead. 1446 uint num_edges = 0; 1447 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1448 if (use->in(j) == old) { 1449 use->set_req(j, nn); 1450 ++num_edges; 1451 } 1452 } 1453 i -= num_edges; // we deleted 1 or more copies of this edge 1454 } 1455 1456 // Search for instance field data PhiNodes in the same region pointing to the old 1457 // memory PhiNode and update their instance memory ids to point to the new node. 1458 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) { 1459 Node* region = old->in(0); 1460 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1461 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1462 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) { 1463 phi->set_inst_mem_id((int)nn->_idx); 1464 } 1465 } 1466 } 1467 1468 // Smash all inputs to 'old', isolating him completely 1469 Node *temp = new Node(1); 1470 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1471 remove_dead_node( old ); 1472 temp->del_req(0); // Yank bogus edge 1473 #ifndef PRODUCT 1474 if( VerifyIterativeGVN ) { 1475 for ( int i = 0; i < _verify_window_size; i++ ) { 1476 if ( _verify_window[i] == old ) 1477 _verify_window[i] = nn; 1478 } 1479 } 1480 #endif 1481 _worklist.remove(temp); // this can be necessary 1482 temp->destruct(); // reuse the _idx of this little guy 1483 } 1484 1485 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) { 1486 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1487 Node* u = n->fast_out(i); 1488 if (u != n) { 1489 rehash_node_delayed(u); 1490 int nb = u->replace_edge(n, m); 1491 --i, imax -= nb; 1492 } 1493 } 1494 } 1495 1496 //------------------------------add_users_to_worklist-------------------------- 1497 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1498 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1499 _worklist.push(n->fast_out(i)); // Push on worklist 1500 } 1501 } 1502 1503 // Return counted loop Phi if as a counted loop exit condition, cmp 1504 // compares the the induction variable with n 1505 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) { 1506 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1507 Node* bol = cmp->fast_out(i); 1508 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1509 Node* iff = bol->fast_out(i2); 1510 if (iff->is_CountedLoopEnd()) { 1511 CountedLoopEndNode* cle = iff->as_CountedLoopEnd(); 1512 if (cle->limit() == n) { 1513 PhiNode* phi = cle->phi(); 1514 if (phi != NULL) { 1515 return phi; 1516 } 1517 } 1518 } 1519 } 1520 } 1521 return NULL; 1522 } 1523 1524 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1525 add_users_to_worklist0(n); 1526 1527 // Move users of node to worklist 1528 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1529 Node* use = n->fast_out(i); // Get use 1530 1531 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1532 use->is_Store() ) // Enable store/load same address 1533 add_users_to_worklist0(use); 1534 1535 // If we changed the receiver type to a call, we need to revisit 1536 // the Catch following the call. It's looking for a non-NULL 1537 // receiver to know when to enable the regular fall-through path 1538 // in addition to the NullPtrException path. 1539 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1540 Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control); 1541 if (p != NULL) { 1542 add_users_to_worklist0(p); 1543 } 1544 } 1545 1546 uint use_op = use->Opcode(); 1547 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1548 add_users_to_worklist(use); // Put Bool on worklist 1549 if (use->outcnt() > 0) { 1550 Node* bol = use->raw_out(0); 1551 if (bol->outcnt() > 0) { 1552 Node* iff = bol->raw_out(0); 1553 if (iff->outcnt() == 2) { 1554 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1555 // phi merging either 0 or 1 onto the worklist 1556 Node* ifproj0 = iff->raw_out(0); 1557 Node* ifproj1 = iff->raw_out(1); 1558 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1559 Node* region0 = ifproj0->raw_out(0); 1560 Node* region1 = ifproj1->raw_out(0); 1561 if( region0 == region1 ) 1562 add_users_to_worklist0(region0); 1563 } 1564 } 1565 } 1566 } 1567 if (use_op == Op_CmpI) { 1568 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n); 1569 if (phi != NULL) { 1570 // If an opaque node feeds into the limit condition of a 1571 // CountedLoop, we need to process the Phi node for the 1572 // induction variable when the opaque node is removed: 1573 // the range of values taken by the Phi is now known and 1574 // so its type is also known. 1575 _worklist.push(phi); 1576 } 1577 Node* in1 = use->in(1); 1578 for (uint i = 0; i < in1->outcnt(); i++) { 1579 if (in1->raw_out(i)->Opcode() == Op_CastII) { 1580 Node* castii = in1->raw_out(i); 1581 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) { 1582 Node* ifnode = castii->in(0)->in(0); 1583 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) { 1584 // Reprocess a CastII node that may depend on an 1585 // opaque node value when the opaque node is 1586 // removed. In case it carries a dependency we can do 1587 // a better job of computing its type. 1588 _worklist.push(castii); 1589 } 1590 } 1591 } 1592 } 1593 } 1594 } 1595 1596 // If changed Cast input, check Phi users for simple cycles 1597 if (use->is_ConstraintCast()) { 1598 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1599 Node* u = use->fast_out(i2); 1600 if (u->is_Phi()) 1601 _worklist.push(u); 1602 } 1603 } 1604 // If changed LShift inputs, check RShift users for useless sign-ext 1605 if( use_op == Op_LShiftI ) { 1606 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1607 Node* u = use->fast_out(i2); 1608 if (u->Opcode() == Op_RShiftI) 1609 _worklist.push(u); 1610 } 1611 } 1612 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1613 if (use_op == Op_AddI || use_op == Op_SubI) { 1614 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1615 Node* u = use->fast_out(i2); 1616 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1617 _worklist.push(u); 1618 } 1619 } 1620 } 1621 // If changed AddP inputs, check Stores for loop invariant 1622 if( use_op == Op_AddP ) { 1623 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1624 Node* u = use->fast_out(i2); 1625 if (u->is_Mem()) 1626 _worklist.push(u); 1627 } 1628 } 1629 // If changed initialization activity, check dependent Stores 1630 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1631 InitializeNode* init = use->as_Allocate()->initialization(); 1632 if (init != NULL) { 1633 Node* imem = init->proj_out(TypeFunc::Memory); 1634 if (imem != NULL) add_users_to_worklist0(imem); 1635 } 1636 } 1637 if (use_op == Op_Initialize) { 1638 Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory); 1639 if (imem != NULL) add_users_to_worklist0(imem); 1640 } 1641 if (use_op == Op_CastP2X) { 1642 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1643 Node* u = use->fast_out(i2); 1644 if (u->Opcode() == Op_AndX) { 1645 _worklist.push(u); 1646 } 1647 } 1648 } 1649 } 1650 } 1651 1652 /** 1653 * Remove the speculative part of all types that we know of 1654 */ 1655 void PhaseIterGVN::remove_speculative_types() { 1656 assert(UseTypeSpeculation, "speculation is off"); 1657 for (uint i = 0; i < _types.Size(); i++) { 1658 const Type* t = _types.fast_lookup(i); 1659 if (t != NULL) { 1660 _types.map(i, t->remove_speculative()); 1661 } 1662 } 1663 _table.check_no_speculative_types(); 1664 } 1665 1666 //============================================================================= 1667 #ifndef PRODUCT 1668 uint PhaseCCP::_total_invokes = 0; 1669 uint PhaseCCP::_total_constants = 0; 1670 #endif 1671 //------------------------------PhaseCCP--------------------------------------- 1672 // Conditional Constant Propagation, ala Wegman & Zadeck 1673 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1674 NOT_PRODUCT( clear_constants(); ) 1675 assert( _worklist.size() == 0, "" ); 1676 // Clear out _nodes from IterGVN. Must be clear to transform call. 1677 _nodes.clear(); // Clear out from IterGVN 1678 analyze(); 1679 } 1680 1681 #ifndef PRODUCT 1682 //------------------------------~PhaseCCP-------------------------------------- 1683 PhaseCCP::~PhaseCCP() { 1684 inc_invokes(); 1685 _total_constants += count_constants(); 1686 } 1687 #endif 1688 1689 1690 #ifdef ASSERT 1691 static bool ccp_type_widens(const Type* t, const Type* t0) { 1692 assert(t->meet(t0) == t, "Not monotonic"); 1693 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1694 case Type::Int: 1695 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1696 break; 1697 case Type::Long: 1698 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1699 break; 1700 } 1701 return true; 1702 } 1703 #endif //ASSERT 1704 1705 //------------------------------analyze---------------------------------------- 1706 void PhaseCCP::analyze() { 1707 // Initialize all types to TOP, optimistic analysis 1708 for (int i = C->unique() - 1; i >= 0; i--) { 1709 _types.map(i,Type::TOP); 1710 } 1711 1712 // Push root onto worklist 1713 Unique_Node_List worklist; 1714 worklist.push(C->root()); 1715 1716 // Pull from worklist; compute new value; push changes out. 1717 // This loop is the meat of CCP. 1718 while( worklist.size() ) { 1719 Node *n = worklist.pop(); 1720 const Type *t = n->Value(this); 1721 if (t != type(n)) { 1722 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1723 #ifndef PRODUCT 1724 if( TracePhaseCCP ) { 1725 t->dump(); 1726 do { tty->print("\t"); } while (tty->position() < 16); 1727 n->dump(); 1728 } 1729 #endif 1730 set_type(n, t); 1731 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1732 Node* m = n->fast_out(i); // Get user 1733 if (m->is_Region()) { // New path to Region? Must recheck Phis too 1734 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1735 Node* p = m->fast_out(i2); // Propagate changes to uses 1736 if (p->bottom_type() != type(p)) { // If not already bottomed out 1737 worklist.push(p); // Propagate change to user 1738 } 1739 } 1740 } 1741 // If we changed the receiver type to a call, we need to revisit 1742 // the Catch following the call. It's looking for a non-NULL 1743 // receiver to know when to enable the regular fall-through path 1744 // in addition to the NullPtrException path 1745 if (m->is_Call()) { 1746 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1747 Node* p = m->fast_out(i2); // Propagate changes to uses 1748 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) { 1749 worklist.push(p->unique_out()); 1750 } 1751 } 1752 } 1753 if (m->bottom_type() != type(m)) { // If not already bottomed out 1754 worklist.push(m); // Propagate change to user 1755 } 1756 1757 // CmpU nodes can get their type information from two nodes up in the 1758 // graph (instead of from the nodes immediately above). Make sure they 1759 // are added to the worklist if nodes they depend on are updated, since 1760 // they could be missed and get wrong types otherwise. 1761 uint m_op = m->Opcode(); 1762 if (m_op == Op_AddI || m_op == Op_SubI) { 1763 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1764 Node* p = m->fast_out(i2); // Propagate changes to uses 1765 if (p->Opcode() == Op_CmpU) { 1766 // Got a CmpU which might need the new type information from node n. 1767 if(p->bottom_type() != type(p)) { // If not already bottomed out 1768 worklist.push(p); // Propagate change to user 1769 } 1770 } 1771 } 1772 } 1773 // If n is used in a counted loop exit condition then the type 1774 // of the counted loop's Phi depends on the type of n. See 1775 // PhiNode::Value(). 1776 if (m_op == Op_CmpI) { 1777 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n); 1778 if (phi != NULL) { 1779 worklist.push(phi); 1780 } 1781 } 1782 if (m_op == Op_CastP2X) { 1783 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1784 Node* u = m->fast_out(i2); 1785 if (u->Opcode() == Op_AndX) { 1786 worklist.push(u); 1787 } 1788 } 1789 } 1790 } 1791 } 1792 } 1793 } 1794 1795 //------------------------------do_transform----------------------------------- 1796 // Top level driver for the recursive transformer 1797 void PhaseCCP::do_transform() { 1798 // Correct leaves of new-space Nodes; they point to old-space. 1799 C->set_root( transform(C->root())->as_Root() ); 1800 assert( C->top(), "missing TOP node" ); 1801 assert( C->root(), "missing root" ); 1802 } 1803 1804 //------------------------------transform-------------------------------------- 1805 // Given a Node in old-space, clone him into new-space. 1806 // Convert any of his old-space children into new-space children. 1807 Node *PhaseCCP::transform( Node *n ) { 1808 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1809 if( new_node != NULL ) 1810 return new_node; // Been there, done that, return old answer 1811 new_node = transform_once(n); // Check for constant 1812 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1813 1814 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1815 GrowableArray <Node *> trstack(C->live_nodes() >> 1); 1816 1817 trstack.push(new_node); // Process children of cloned node 1818 while ( trstack.is_nonempty() ) { 1819 Node *clone = trstack.pop(); 1820 uint cnt = clone->req(); 1821 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1822 Node *input = clone->in(i); 1823 if( input != NULL ) { // Ignore NULLs 1824 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1825 if( new_input == NULL ) { 1826 new_input = transform_once(input); // Check for constant 1827 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1828 trstack.push(new_input); 1829 } 1830 assert( new_input == clone->in(i), "insanity check"); 1831 } 1832 } 1833 } 1834 return new_node; 1835 } 1836 1837 1838 //------------------------------transform_once--------------------------------- 1839 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1840 Node *PhaseCCP::transform_once( Node *n ) { 1841 const Type *t = type(n); 1842 // Constant? Use constant Node instead 1843 if( t->singleton() ) { 1844 Node *nn = n; // Default is to return the original constant 1845 if( t == Type::TOP ) { 1846 // cache my top node on the Compile instance 1847 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1848 C->set_cached_top_node(ConNode::make(Type::TOP)); 1849 set_type(C->top(), Type::TOP); 1850 } 1851 nn = C->top(); 1852 } 1853 if( !n->is_Con() ) { 1854 if( t != Type::TOP ) { 1855 nn = makecon(t); // ConNode::make(t); 1856 NOT_PRODUCT( inc_constants(); ) 1857 } else if( n->is_Region() ) { // Unreachable region 1858 // Note: nn == C->top() 1859 n->set_req(0, NULL); // Cut selfreference 1860 // Eagerly remove dead phis to avoid phis copies creation. 1861 for (DUIterator i = n->outs(); n->has_out(i); i++) { 1862 Node* m = n->out(i); 1863 if( m->is_Phi() ) { 1864 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1865 replace_node(m, nn); 1866 --i; // deleted this phi; rescan starting with next position 1867 } 1868 } 1869 } 1870 replace_node(n,nn); // Update DefUse edges for new constant 1871 } 1872 return nn; 1873 } 1874 1875 // If x is a TypeNode, capture any more-precise type permanently into Node 1876 if (t != n->bottom_type()) { 1877 hash_delete(n); // changing bottom type may force a rehash 1878 n->raise_bottom_type(t); 1879 _worklist.push(n); // n re-enters the hash table via the worklist 1880 } 1881 1882 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1883 switch( n->Opcode() ) { 1884 case Op_FastLock: // Revisit FastLocks for lock coarsening 1885 case Op_If: 1886 case Op_CountedLoopEnd: 1887 case Op_Region: 1888 case Op_Loop: 1889 case Op_CountedLoop: 1890 case Op_Conv2B: 1891 case Op_Opaque1: 1892 case Op_Opaque2: 1893 _worklist.push(n); 1894 break; 1895 default: 1896 break; 1897 } 1898 1899 return n; 1900 } 1901 1902 //---------------------------------saturate------------------------------------ 1903 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1904 const Type* limit_type) const { 1905 const Type* wide_type = new_type->widen(old_type, limit_type); 1906 if (wide_type != new_type) { // did we widen? 1907 // If so, we may have widened beyond the limit type. Clip it back down. 1908 new_type = wide_type->filter(limit_type); 1909 } 1910 return new_type; 1911 } 1912 1913 //------------------------------print_statistics------------------------------- 1914 #ifndef PRODUCT 1915 void PhaseCCP::print_statistics() { 1916 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1917 } 1918 #endif 1919 1920 1921 //============================================================================= 1922 #ifndef PRODUCT 1923 uint PhasePeephole::_total_peepholes = 0; 1924 #endif 1925 //------------------------------PhasePeephole---------------------------------- 1926 // Conditional Constant Propagation, ala Wegman & Zadeck 1927 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1928 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1929 NOT_PRODUCT( clear_peepholes(); ) 1930 } 1931 1932 #ifndef PRODUCT 1933 //------------------------------~PhasePeephole--------------------------------- 1934 PhasePeephole::~PhasePeephole() { 1935 _total_peepholes += count_peepholes(); 1936 } 1937 #endif 1938 1939 //------------------------------transform-------------------------------------- 1940 Node *PhasePeephole::transform( Node *n ) { 1941 ShouldNotCallThis(); 1942 return NULL; 1943 } 1944 1945 //------------------------------do_transform----------------------------------- 1946 void PhasePeephole::do_transform() { 1947 bool method_name_not_printed = true; 1948 1949 // Examine each basic block 1950 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 1951 Block* block = _cfg.get_block(block_number); 1952 bool block_not_printed = true; 1953 1954 // and each instruction within a block 1955 uint end_index = block->number_of_nodes(); 1956 // block->end_idx() not valid after PhaseRegAlloc 1957 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 1958 Node *n = block->get_node(instruction_index); 1959 if( n->is_Mach() ) { 1960 MachNode *m = n->as_Mach(); 1961 int deleted_count = 0; 1962 // check for peephole opportunities 1963 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count); 1964 if( m2 != NULL ) { 1965 #ifndef PRODUCT 1966 if( PrintOptoPeephole ) { 1967 // Print method, first time only 1968 if( C->method() && method_name_not_printed ) { 1969 C->method()->print_short_name(); tty->cr(); 1970 method_name_not_printed = false; 1971 } 1972 // Print this block 1973 if( Verbose && block_not_printed) { 1974 tty->print_cr("in block"); 1975 block->dump(); 1976 block_not_printed = false; 1977 } 1978 // Print instructions being deleted 1979 for( int i = (deleted_count - 1); i >= 0; --i ) { 1980 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 1981 } 1982 tty->print_cr("replaced with"); 1983 // Print new instruction 1984 m2->format(_regalloc); 1985 tty->print("\n\n"); 1986 } 1987 #endif 1988 // Remove old nodes from basic block and update instruction_index 1989 // (old nodes still exist and may have edges pointing to them 1990 // as register allocation info is stored in the allocator using 1991 // the node index to live range mappings.) 1992 uint safe_instruction_index = (instruction_index - deleted_count); 1993 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 1994 block->remove_node( instruction_index ); 1995 } 1996 // install new node after safe_instruction_index 1997 block->insert_node(m2, safe_instruction_index + 1); 1998 end_index = block->number_of_nodes() - 1; // Recompute new block size 1999 NOT_PRODUCT( inc_peepholes(); ) 2000 } 2001 } 2002 } 2003 } 2004 } 2005 2006 //------------------------------print_statistics------------------------------- 2007 #ifndef PRODUCT 2008 void PhasePeephole::print_statistics() { 2009 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2010 } 2011 #endif 2012 2013 2014 //============================================================================= 2015 //------------------------------set_req_X-------------------------------------- 2016 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2017 assert( is_not_dead(n), "can not use dead node"); 2018 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2019 Node *old = in(i); 2020 set_req(i, n); 2021 2022 // old goes dead? 2023 if( old ) { 2024 switch (old->outcnt()) { 2025 case 0: 2026 // Put into the worklist to kill later. We do not kill it now because the 2027 // recursive kill will delete the current node (this) if dead-loop exists 2028 if (!old->is_top()) 2029 igvn->_worklist.push( old ); 2030 break; 2031 case 1: 2032 if( old->is_Store() || old->has_special_unique_user() ) 2033 igvn->add_users_to_worklist( old ); 2034 break; 2035 case 2: 2036 if( old->is_Store() ) 2037 igvn->add_users_to_worklist( old ); 2038 if( old->Opcode() == Op_Region ) 2039 igvn->_worklist.push(old); 2040 break; 2041 case 3: 2042 if( old->Opcode() == Op_Region ) { 2043 igvn->_worklist.push(old); 2044 igvn->add_users_to_worklist( old ); 2045 } 2046 break; 2047 default: 2048 break; 2049 } 2050 } 2051 2052 } 2053 2054 //-------------------------------replace_by----------------------------------- 2055 // Using def-use info, replace one node for another. Follow the def-use info 2056 // to all users of the OLD node. Then make all uses point to the NEW node. 2057 void Node::replace_by(Node *new_node) { 2058 assert(!is_top(), "top node has no DU info"); 2059 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2060 Node* use = last_out(i); 2061 uint uses_found = 0; 2062 for (uint j = 0; j < use->len(); j++) { 2063 if (use->in(j) == this) { 2064 if (j < use->req()) 2065 use->set_req(j, new_node); 2066 else use->set_prec(j, new_node); 2067 uses_found++; 2068 } 2069 } 2070 i -= uses_found; // we deleted 1 or more copies of this edge 2071 } 2072 } 2073 2074 //============================================================================= 2075 //----------------------------------------------------------------------------- 2076 void Type_Array::grow( uint i ) { 2077 if( !_max ) { 2078 _max = 1; 2079 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2080 _types[0] = NULL; 2081 } 2082 uint old = _max; 2083 while( i >= _max ) _max <<= 1; // Double to fit 2084 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2085 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2086 } 2087 2088 //------------------------------dump------------------------------------------- 2089 #ifndef PRODUCT 2090 void Type_Array::dump() const { 2091 uint max = Size(); 2092 for( uint i = 0; i < max; i++ ) { 2093 if( _types[i] != NULL ) { 2094 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2095 } 2096 } 2097 } 2098 #endif