1 /* 2 * Copyright (c) 1997, 2017, 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((void*)this, (void*)&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 default: break; 698 } 699 if (t->is_zero_type()) 700 return zerocon(t->basic_type()); 701 return uncached_makecon(t); 702 } 703 704 //--------------------------uncached_makecon----------------------------------- 705 // Make an idealized constant - one of ConINode, ConPNode, etc. 706 ConNode* PhaseValues::uncached_makecon(const Type *t) { 707 assert(t->singleton(), "must be a constant"); 708 ConNode* x = ConNode::make(t); 709 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 710 if (k == NULL) { 711 set_type(x, t); // Missed, provide type mapping 712 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 713 if (nna != NULL) { 714 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 715 loc->clear(); // do not put debug info on constants 716 } 717 } else { 718 x->destruct(); // Hit, destroy duplicate constant 719 x = k; // use existing constant 720 } 721 return x; 722 } 723 724 //------------------------------intcon----------------------------------------- 725 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 726 ConINode* PhaseTransform::intcon(int i) { 727 // Small integer? Check cache! Check that cached node is not dead 728 if (i >= _icon_min && i <= _icon_max) { 729 ConINode* icon = _icons[i-_icon_min]; 730 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 731 return icon; 732 } 733 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 734 assert(icon->is_Con(), ""); 735 if (i >= _icon_min && i <= _icon_max) 736 _icons[i-_icon_min] = icon; // Cache small integers 737 return icon; 738 } 739 740 //------------------------------longcon---------------------------------------- 741 // Fast long constant. 742 ConLNode* PhaseTransform::longcon(jlong l) { 743 // Small integer? Check cache! Check that cached node is not dead 744 if (l >= _lcon_min && l <= _lcon_max) { 745 ConLNode* lcon = _lcons[l-_lcon_min]; 746 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 747 return lcon; 748 } 749 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 750 assert(lcon->is_Con(), ""); 751 if (l >= _lcon_min && l <= _lcon_max) 752 _lcons[l-_lcon_min] = lcon; // Cache small integers 753 return lcon; 754 } 755 756 //------------------------------zerocon----------------------------------------- 757 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 758 ConNode* PhaseTransform::zerocon(BasicType bt) { 759 assert((uint)bt <= _zcon_max, "domain check"); 760 ConNode* zcon = _zcons[bt]; 761 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 762 return zcon; 763 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 764 _zcons[bt] = zcon; 765 return zcon; 766 } 767 768 769 770 //============================================================================= 771 //------------------------------transform-------------------------------------- 772 // Return a node which computes the same function as this node, but in a 773 // faster or cheaper fashion. 774 Node *PhaseGVN::transform( Node *n ) { 775 return transform_no_reclaim(n); 776 } 777 778 //------------------------------transform-------------------------------------- 779 // Return a node which computes the same function as this node, but 780 // in a faster or cheaper fashion. 781 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 782 NOT_PRODUCT( set_transforms(); ) 783 784 // Apply the Ideal call in a loop until it no longer applies 785 Node *k = n; 786 NOT_PRODUCT( uint loop_count = 0; ) 787 while( 1 ) { 788 Node *i = k->Ideal(this, /*can_reshape=*/false); 789 if( !i ) break; 790 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 791 k = i; 792 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 793 } 794 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 795 796 797 // If brand new node, make space in type array. 798 ensure_type_or_null(k); 799 800 // Since I just called 'Value' to compute the set of run-time values 801 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 802 // cache Value. Later requests for the local phase->type of this Node can 803 // use the cached Value instead of suffering with 'bottom_type'. 804 const Type *t = k->Value(this); // Get runtime Value set 805 assert(t != NULL, "value sanity"); 806 if (type_or_null(k) != t) { 807 #ifndef PRODUCT 808 // Do not count initial visit to node as a transformation 809 if (type_or_null(k) == NULL) { 810 inc_new_values(); 811 set_progress(); 812 } 813 #endif 814 set_type(k, t); 815 // If k is a TypeNode, capture any more-precise type permanently into Node 816 k->raise_bottom_type(t); 817 } 818 819 if( t->singleton() && !k->is_Con() ) { 820 NOT_PRODUCT( set_progress(); ) 821 return makecon(t); // Turn into a constant 822 } 823 824 // Now check for Identities 825 Node *i = k->Identity(this); // Look for a nearby replacement 826 if( i != k ) { // Found? Return replacement! 827 NOT_PRODUCT( set_progress(); ) 828 return i; 829 } 830 831 // Global Value Numbering 832 i = hash_find_insert(k); // Insert if new 833 if( i && (i != k) ) { 834 // Return the pre-existing node 835 NOT_PRODUCT( set_progress(); ) 836 return i; 837 } 838 839 // Return Idealized original 840 return k; 841 } 842 843 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 844 if (d->is_top() || n->is_top()) { 845 return false; 846 } 847 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 848 int i = 0; 849 while (d != n) { 850 n = IfNode::up_one_dom(n, linear_only); 851 i++; 852 if (n == NULL || i >= 10) { 853 return false; 854 } 855 } 856 return true; 857 } 858 859 #ifdef ASSERT 860 //------------------------------dead_loop_check-------------------------------- 861 // Check for a simple dead loop when a data node references itself directly 862 // or through an other data node excluding cons and phis. 863 void PhaseGVN::dead_loop_check( Node *n ) { 864 // Phi may reference itself in a loop 865 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 866 // Do 2 levels check and only data inputs. 867 bool no_dead_loop = true; 868 uint cnt = n->req(); 869 for (uint i = 1; i < cnt && no_dead_loop; i++) { 870 Node *in = n->in(i); 871 if (in == n) { 872 no_dead_loop = false; 873 } else if (in != NULL && !in->is_dead_loop_safe()) { 874 uint icnt = in->req(); 875 for (uint j = 1; j < icnt && no_dead_loop; j++) { 876 if (in->in(j) == n || in->in(j) == in) 877 no_dead_loop = false; 878 } 879 } 880 } 881 if (!no_dead_loop) n->dump(3); 882 assert(no_dead_loop, "dead loop detected"); 883 } 884 } 885 #endif 886 887 //============================================================================= 888 //------------------------------PhaseIterGVN----------------------------------- 889 // Initialize hash table to fresh and clean for +VerifyOpto 890 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ), 891 _stack(C->live_nodes() >> 1), 892 _delay_transform(false) { 893 } 894 895 //------------------------------PhaseIterGVN----------------------------------- 896 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 897 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 898 _worklist( igvn->_worklist ), 899 _stack( igvn->_stack ), 900 _delay_transform(igvn->_delay_transform) 901 { 902 } 903 904 //------------------------------PhaseIterGVN----------------------------------- 905 // Initialize with previous PhaseGVN info from Parser 906 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 907 _worklist(*C->for_igvn()), 908 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that 909 // the constructor is used in incremental inlining, this consumes too much memory: 910 // _stack(C->live_nodes() >> 1), 911 // So, as a band-aid, we replace this by: 912 _stack(C->comp_arena(), 32), 913 _delay_transform(false) 914 { 915 uint max; 916 917 // Dead nodes in the hash table inherited from GVN were not treated as 918 // roots during def-use info creation; hence they represent an invisible 919 // use. Clear them out. 920 max = _table.size(); 921 for( uint i = 0; i < max; ++i ) { 922 Node *n = _table.at(i); 923 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 924 if( n->is_top() ) continue; 925 assert( false, "Parse::remove_useless_nodes missed this node"); 926 hash_delete(n); 927 } 928 } 929 930 // Any Phis or Regions on the worklist probably had uses that could not 931 // make more progress because the uses were made while the Phis and Regions 932 // were in half-built states. Put all uses of Phis and Regions on worklist. 933 max = _worklist.size(); 934 for( uint j = 0; j < max; j++ ) { 935 Node *n = _worklist.at(j); 936 uint uop = n->Opcode(); 937 if( uop == Op_Phi || uop == Op_Region || 938 n->is_Type() || 939 n->is_Mem() ) 940 add_users_to_worklist(n); 941 } 942 } 943 944 /** 945 * Initialize worklist for each node. 946 */ 947 void PhaseIterGVN::init_worklist(Node* first) { 948 Unique_Node_List to_process; 949 to_process.push(first); 950 951 while (to_process.size() > 0) { 952 Node* n = to_process.pop(); 953 if (!_worklist.member(n)) { 954 _worklist.push(n); 955 956 uint cnt = n->req(); 957 for(uint i = 0; i < cnt; i++) { 958 Node* m = n->in(i); 959 if (m != NULL) { 960 to_process.push(m); 961 } 962 } 963 } 964 } 965 } 966 967 #ifndef PRODUCT 968 void PhaseIterGVN::verify_step(Node* n) { 969 if (VerifyIterativeGVN) { 970 _verify_window[_verify_counter % _verify_window_size] = n; 971 ++_verify_counter; 972 ResourceMark rm; 973 ResourceArea* area = Thread::current()->resource_area(); 974 VectorSet old_space(area), new_space(area); 975 if (C->unique() < 1000 || 976 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 977 ++_verify_full_passes; 978 Node::verify_recur(C->root(), -1, old_space, new_space); 979 } 980 const int verify_depth = 4; 981 for ( int i = 0; i < _verify_window_size; i++ ) { 982 Node* n = _verify_window[i]; 983 if ( n == NULL ) continue; 984 if( n->in(0) == NodeSentinel ) { // xform_idom 985 _verify_window[i] = n->in(1); 986 --i; continue; 987 } 988 // Typical fanout is 1-2, so this call visits about 6 nodes. 989 Node::verify_recur(n, verify_depth, old_space, new_space); 990 } 991 } 992 } 993 994 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 995 if (TraceIterativeGVN) { 996 uint wlsize = _worklist.size(); 997 const Type* newtype = type_or_null(n); 998 if (nn != n) { 999 // print old node 1000 tty->print("< "); 1001 if (oldtype != newtype && oldtype != NULL) { 1002 oldtype->dump(); 1003 } 1004 do { tty->print("\t"); } while (tty->position() < 16); 1005 tty->print("<"); 1006 n->dump(); 1007 } 1008 if (oldtype != newtype || nn != n) { 1009 // print new node and/or new type 1010 if (oldtype == NULL) { 1011 tty->print("* "); 1012 } else if (nn != n) { 1013 tty->print("> "); 1014 } else { 1015 tty->print("= "); 1016 } 1017 if (newtype == NULL) { 1018 tty->print("null"); 1019 } else { 1020 newtype->dump(); 1021 } 1022 do { tty->print("\t"); } while (tty->position() < 16); 1023 nn->dump(); 1024 } 1025 if (Verbose && wlsize < _worklist.size()) { 1026 tty->print(" Push {"); 1027 while (wlsize != _worklist.size()) { 1028 Node* pushed = _worklist.at(wlsize++); 1029 tty->print(" %d", pushed->_idx); 1030 } 1031 tty->print_cr(" }"); 1032 } 1033 if (nn != n) { 1034 // ignore n, it might be subsumed 1035 verify_step((Node*) NULL); 1036 } 1037 } 1038 } 1039 1040 void PhaseIterGVN::init_verifyPhaseIterGVN() { 1041 _verify_counter = 0; 1042 _verify_full_passes = 0; 1043 for (int i = 0; i < _verify_window_size; i++) { 1044 _verify_window[i] = NULL; 1045 } 1046 #ifdef ASSERT 1047 // Verify that all modified nodes are on _worklist 1048 Unique_Node_List* modified_list = C->modified_nodes(); 1049 while (modified_list != NULL && modified_list->size()) { 1050 Node* n = modified_list->pop(); 1051 if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) { 1052 n->dump(); 1053 assert(false, "modified node is not on IGVN._worklist"); 1054 } 1055 } 1056 #endif 1057 } 1058 1059 void PhaseIterGVN::verify_PhaseIterGVN() { 1060 #ifdef ASSERT 1061 // Verify nodes with changed inputs. 1062 Unique_Node_List* modified_list = C->modified_nodes(); 1063 while (modified_list != NULL && modified_list->size()) { 1064 Node* n = modified_list->pop(); 1065 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes 1066 n->dump(); 1067 assert(false, "modified node was not processed by IGVN.transform_old()"); 1068 } 1069 } 1070 #endif 1071 1072 C->verify_graph_edges(); 1073 if( VerifyOpto && allow_progress() ) { 1074 // Must turn off allow_progress to enable assert and break recursion 1075 C->root()->verify(); 1076 { // Check if any progress was missed using IterGVN 1077 // Def-Use info enables transformations not attempted in wash-pass 1078 // e.g. Region/Phi cleanup, ... 1079 // Null-check elision -- may not have reached fixpoint 1080 // do not propagate to dominated nodes 1081 ResourceMark rm; 1082 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 1083 // Fill worklist completely 1084 igvn2.init_worklist(C->root()); 1085 1086 igvn2.set_allow_progress(false); 1087 igvn2.optimize(); 1088 igvn2.set_allow_progress(true); 1089 } 1090 } 1091 if (VerifyIterativeGVN && PrintOpto) { 1092 if (_verify_counter == _verify_full_passes) { 1093 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 1094 (int) _verify_full_passes); 1095 } else { 1096 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 1097 (int) _verify_counter, (int) _verify_full_passes); 1098 } 1099 } 1100 1101 #ifdef ASSERT 1102 while (modified_list->size()) { 1103 Node* n = modified_list->pop(); 1104 n->dump(); 1105 assert(false, "VerifyIterativeGVN: new modified node was added"); 1106 } 1107 #endif 1108 } 1109 #endif /* PRODUCT */ 1110 1111 #ifdef ASSERT 1112 /** 1113 * Dumps information that can help to debug the problem. A debug 1114 * build fails with an assert. 1115 */ 1116 void PhaseIterGVN::dump_infinite_loop_info(Node* n) { 1117 n->dump(4); 1118 _worklist.dump(); 1119 assert(false, "infinite loop in PhaseIterGVN::optimize"); 1120 } 1121 1122 /** 1123 * Prints out information about IGVN if the 'verbose' option is used. 1124 */ 1125 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1126 if (TraceIterativeGVN && Verbose) { 1127 tty->print(" Pop "); 1128 n->dump(); 1129 if ((num_processed % 100) == 0) { 1130 _worklist.print_set(); 1131 } 1132 } 1133 } 1134 #endif /* ASSERT */ 1135 1136 void PhaseIterGVN::optimize() { 1137 DEBUG_ONLY(uint num_processed = 0;) 1138 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1139 1140 uint loop_count = 0; 1141 // Pull from worklist and transform the node. If the node has changed, 1142 // update edge info and put uses on worklist. 1143 while(_worklist.size()) { 1144 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1145 return; 1146 } 1147 Node* n = _worklist.pop(); 1148 if (++loop_count >= K * C->live_nodes()) { 1149 DEBUG_ONLY(dump_infinite_loop_info(n);) 1150 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1151 return; 1152 } 1153 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1154 if (n->outcnt() != 0) { 1155 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1156 // Do the transformation 1157 Node* nn = transform_old(n); 1158 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1159 } else if (!n->is_top()) { 1160 remove_dead_node(n); 1161 } 1162 } 1163 NOT_PRODUCT(verify_PhaseIterGVN();) 1164 } 1165 1166 1167 /** 1168 * Register a new node with the optimizer. Update the types array, the def-use 1169 * info. Put on worklist. 1170 */ 1171 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1172 set_type_bottom(n); 1173 _worklist.push(n); 1174 if (orig != NULL) C->copy_node_notes_to(n, orig); 1175 return n; 1176 } 1177 1178 //------------------------------transform-------------------------------------- 1179 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1180 Node *PhaseIterGVN::transform( Node *n ) { 1181 // If brand new node, make space in type array, and give it a type. 1182 ensure_type_or_null(n); 1183 if (type_or_null(n) == NULL) { 1184 set_type_bottom(n); 1185 } 1186 1187 if (_delay_transform) { 1188 // Add the node to the worklist but don't optimize for now 1189 _worklist.push(n); 1190 return n; 1191 } 1192 1193 return transform_old(n); 1194 } 1195 1196 Node *PhaseIterGVN::transform_old(Node* n) { 1197 DEBUG_ONLY(uint loop_count = 0;); 1198 NOT_PRODUCT(set_transforms()); 1199 1200 // Remove 'n' from hash table in case it gets modified 1201 _table.hash_delete(n); 1202 if (VerifyIterativeGVN) { 1203 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1204 } 1205 1206 // Apply the Ideal call in a loop until it no longer applies 1207 Node* k = n; 1208 DEBUG_ONLY(dead_loop_check(k);) 1209 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1210 C->remove_modified_node(k); 1211 Node* i = k->Ideal(this, /*can_reshape=*/true); 1212 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1213 #ifndef PRODUCT 1214 verify_step(k); 1215 if (i && VerifyOpto ) { 1216 if (!allow_progress()) { 1217 if (i->is_Add() && (i->outcnt() == 1)) { 1218 // Switched input to left side because this is the only use 1219 } else if (i->is_If() && (i->in(0) == NULL)) { 1220 // This IF is dead because it is dominated by an equivalent IF When 1221 // dominating if changed, info is not propagated sparsely to 'this' 1222 // Propagating this info further will spuriously identify other 1223 // progress. 1224 return i; 1225 } else 1226 set_progress(); 1227 } else { 1228 set_progress(); 1229 } 1230 } 1231 #endif 1232 1233 while (i != NULL) { 1234 #ifdef ASSERT 1235 if (loop_count >= K) { 1236 dump_infinite_loop_info(i); 1237 } 1238 loop_count++; 1239 #endif 1240 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1241 // Made a change; put users of original Node on worklist 1242 add_users_to_worklist(k); 1243 // Replacing root of transform tree? 1244 if (k != i) { 1245 // Make users of old Node now use new. 1246 subsume_node(k, i); 1247 k = i; 1248 } 1249 DEBUG_ONLY(dead_loop_check(k);) 1250 // Try idealizing again 1251 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1252 C->remove_modified_node(k); 1253 i = k->Ideal(this, /*can_reshape=*/true); 1254 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1255 #ifndef PRODUCT 1256 verify_step(k); 1257 if (i && VerifyOpto) { 1258 set_progress(); 1259 } 1260 #endif 1261 } 1262 1263 // If brand new node, make space in type array. 1264 ensure_type_or_null(k); 1265 1266 // See what kind of values 'k' takes on at runtime 1267 const Type* t = k->Value(this); 1268 assert(t != NULL, "value sanity"); 1269 1270 // Since I just called 'Value' to compute the set of run-time values 1271 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1272 // cache Value. Later requests for the local phase->type of this Node can 1273 // use the cached Value instead of suffering with 'bottom_type'. 1274 if (type_or_null(k) != t) { 1275 #ifndef PRODUCT 1276 inc_new_values(); 1277 set_progress(); 1278 #endif 1279 set_type(k, t); 1280 // If k is a TypeNode, capture any more-precise type permanently into Node 1281 k->raise_bottom_type(t); 1282 // Move users of node to worklist 1283 add_users_to_worklist(k); 1284 } 1285 // If 'k' computes a constant, replace it with a constant 1286 if (t->singleton() && !k->is_Con()) { 1287 NOT_PRODUCT(set_progress();) 1288 Node* con = makecon(t); // Make a constant 1289 add_users_to_worklist(k); 1290 subsume_node(k, con); // Everybody using k now uses con 1291 return con; 1292 } 1293 1294 // Now check for Identities 1295 i = k->Identity(this); // Look for a nearby replacement 1296 if (i != k) { // Found? Return replacement! 1297 NOT_PRODUCT(set_progress();) 1298 add_users_to_worklist(k); 1299 subsume_node(k, i); // Everybody using k now uses i 1300 return i; 1301 } 1302 1303 // Global Value Numbering 1304 i = hash_find_insert(k); // Check for pre-existing node 1305 if (i && (i != k)) { 1306 // Return the pre-existing node if it isn't dead 1307 NOT_PRODUCT(set_progress();) 1308 add_users_to_worklist(k); 1309 subsume_node(k, i); // Everybody using k now uses i 1310 return i; 1311 } 1312 1313 // Return Idealized original 1314 return k; 1315 } 1316 1317 //---------------------------------saturate------------------------------------ 1318 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1319 const Type* limit_type) const { 1320 return new_type->narrow(old_type); 1321 } 1322 1323 //------------------------------remove_globally_dead_node---------------------- 1324 // Kill a globally dead Node. All uses are also globally dead and are 1325 // aggressively trimmed. 1326 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1327 enum DeleteProgress { 1328 PROCESS_INPUTS, 1329 PROCESS_OUTPUTS 1330 }; 1331 assert(_stack.is_empty(), "not empty"); 1332 _stack.push(dead, PROCESS_INPUTS); 1333 1334 while (_stack.is_nonempty()) { 1335 dead = _stack.node(); 1336 uint progress_state = _stack.index(); 1337 assert(dead != C->root(), "killing root, eh?"); 1338 assert(!dead->is_top(), "add check for top when pushing"); 1339 NOT_PRODUCT( set_progress(); ) 1340 if (progress_state == PROCESS_INPUTS) { 1341 // After following inputs, continue to outputs 1342 _stack.set_index(PROCESS_OUTPUTS); 1343 if (!dead->is_Con()) { // Don't kill cons but uses 1344 bool recurse = false; 1345 // Remove from hash table 1346 _table.hash_delete( dead ); 1347 // Smash all inputs to 'dead', isolating him completely 1348 for (uint i = 0; i < dead->req(); i++) { 1349 Node *in = dead->in(i); 1350 if (in != NULL && in != C->top()) { // Points to something? 1351 int nrep = dead->replace_edge(in, NULL); // Kill edges 1352 assert((nrep > 0), "sanity"); 1353 if (in->outcnt() == 0) { // Made input go dead? 1354 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1355 recurse = true; 1356 } else if (in->outcnt() == 1 && 1357 in->has_special_unique_user()) { 1358 _worklist.push(in->unique_out()); 1359 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1360 if (in->Opcode() == Op_Region) { 1361 _worklist.push(in); 1362 } else if (in->is_Store()) { 1363 DUIterator_Fast imax, i = in->fast_outs(imax); 1364 _worklist.push(in->fast_out(i)); 1365 i++; 1366 if (in->outcnt() == 2) { 1367 _worklist.push(in->fast_out(i)); 1368 i++; 1369 } 1370 assert(!(i < imax), "sanity"); 1371 } 1372 } 1373 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1374 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1375 // A Load that directly follows an InitializeNode is 1376 // going away. The Stores that follow are candidates 1377 // again to be captured by the InitializeNode. 1378 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1379 Node *n = in->fast_out(j); 1380 if (n->is_Store()) { 1381 _worklist.push(n); 1382 } 1383 } 1384 } 1385 } // if (in != NULL && in != C->top()) 1386 } // for (uint i = 0; i < dead->req(); i++) 1387 if (recurse) { 1388 continue; 1389 } 1390 } // if (!dead->is_Con()) 1391 } // if (progress_state == PROCESS_INPUTS) 1392 1393 // Aggressively kill globally dead uses 1394 // (Rather than pushing all the outs at once, we push one at a time, 1395 // plus the parent to resume later, because of the indefinite number 1396 // of edge deletions per loop trip.) 1397 if (dead->outcnt() > 0) { 1398 // Recursively remove output edges 1399 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1400 } else { 1401 // Finished disconnecting all input and output edges. 1402 _stack.pop(); 1403 // Remove dead node from iterative worklist 1404 _worklist.remove(dead); 1405 C->remove_modified_node(dead); 1406 // Constant node that has no out-edges and has only one in-edge from 1407 // root is usually dead. However, sometimes reshaping walk makes 1408 // it reachable by adding use edges. So, we will NOT count Con nodes 1409 // as dead to be conservative about the dead node count at any 1410 // given time. 1411 if (!dead->is_Con()) { 1412 C->record_dead_node(dead->_idx); 1413 } 1414 if (dead->is_macro()) { 1415 C->remove_macro_node(dead); 1416 } 1417 if (dead->is_expensive()) { 1418 C->remove_expensive_node(dead); 1419 } 1420 CastIINode* cast = dead->isa_CastII(); 1421 if (cast != NULL && cast->has_range_check()) { 1422 C->remove_range_check_cast(cast); 1423 } 1424 if (dead->is_ValueTypePtr()) { 1425 C->remove_value_type_ptr(dead->as_ValueTypePtr()); 1426 } 1427 } 1428 } // while (_stack.is_nonempty()) 1429 } 1430 1431 //------------------------------subsume_node----------------------------------- 1432 // Remove users from node 'old' and add them to node 'nn'. 1433 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1434 assert( old != hash_find(old), "should already been removed" ); 1435 assert( old != C->top(), "cannot subsume top node"); 1436 // Copy debug or profile information to the new version: 1437 C->copy_node_notes_to(nn, old); 1438 // Move users of node 'old' to node 'nn' 1439 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1440 Node* use = old->last_out(i); // for each use... 1441 // use might need re-hashing (but it won't if it's a new node) 1442 rehash_node_delayed(use); 1443 // Update use-def info as well 1444 // We remove all occurrences of old within use->in, 1445 // so as to avoid rehashing any node more than once. 1446 // The hash table probe swamps any outer loop overhead. 1447 uint num_edges = 0; 1448 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1449 if (use->in(j) == old) { 1450 use->set_req(j, nn); 1451 ++num_edges; 1452 } 1453 } 1454 i -= num_edges; // we deleted 1 or more copies of this edge 1455 } 1456 1457 // Search for instance field data PhiNodes in the same region pointing to the old 1458 // memory PhiNode and update their instance memory ids to point to the new node. 1459 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) { 1460 Node* region = old->in(0); 1461 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1462 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1463 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) { 1464 phi->set_inst_mem_id((int)nn->_idx); 1465 } 1466 } 1467 } 1468 1469 // Smash all inputs to 'old', isolating him completely 1470 Node *temp = new Node(1); 1471 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1472 remove_dead_node( old ); 1473 temp->del_req(0); // Yank bogus edge 1474 #ifndef PRODUCT 1475 if( VerifyIterativeGVN ) { 1476 for ( int i = 0; i < _verify_window_size; i++ ) { 1477 if ( _verify_window[i] == old ) 1478 _verify_window[i] = nn; 1479 } 1480 } 1481 #endif 1482 _worklist.remove(temp); // this can be necessary 1483 temp->destruct(); // reuse the _idx of this little guy 1484 } 1485 1486 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) { 1487 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1488 Node* u = n->fast_out(i); 1489 if (u != n) { 1490 rehash_node_delayed(u); 1491 int nb = u->replace_edge(n, m); 1492 --i, imax -= nb; 1493 } 1494 } 1495 } 1496 1497 //------------------------------add_users_to_worklist-------------------------- 1498 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1499 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1500 _worklist.push(n->fast_out(i)); // Push on worklist 1501 } 1502 } 1503 1504 // Return counted loop Phi if as a counted loop exit condition, cmp 1505 // compares the the induction variable with n 1506 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) { 1507 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1508 Node* bol = cmp->fast_out(i); 1509 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1510 Node* iff = bol->fast_out(i2); 1511 if (iff->is_CountedLoopEnd()) { 1512 CountedLoopEndNode* cle = iff->as_CountedLoopEnd(); 1513 if (cle->limit() == n) { 1514 PhiNode* phi = cle->phi(); 1515 if (phi != NULL) { 1516 return phi; 1517 } 1518 } 1519 } 1520 } 1521 } 1522 return NULL; 1523 } 1524 1525 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1526 add_users_to_worklist0(n); 1527 1528 // Move users of node to worklist 1529 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1530 Node* use = n->fast_out(i); // Get use 1531 1532 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1533 use->is_Store() ) // Enable store/load same address 1534 add_users_to_worklist0(use); 1535 1536 // If we changed the receiver type to a call, we need to revisit 1537 // the Catch following the call. It's looking for a non-NULL 1538 // receiver to know when to enable the regular fall-through path 1539 // in addition to the NullPtrException path. 1540 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1541 Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control); 1542 if (p != NULL) { 1543 add_users_to_worklist0(p); 1544 } 1545 } 1546 1547 uint use_op = use->Opcode(); 1548 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1549 add_users_to_worklist(use); // Put Bool on worklist 1550 if (use->outcnt() > 0) { 1551 Node* bol = use->raw_out(0); 1552 if (bol->outcnt() > 0) { 1553 Node* iff = bol->raw_out(0); 1554 if (iff->outcnt() == 2) { 1555 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1556 // phi merging either 0 or 1 onto the worklist 1557 Node* ifproj0 = iff->raw_out(0); 1558 Node* ifproj1 = iff->raw_out(1); 1559 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1560 Node* region0 = ifproj0->raw_out(0); 1561 Node* region1 = ifproj1->raw_out(0); 1562 if( region0 == region1 ) 1563 add_users_to_worklist0(region0); 1564 } 1565 } 1566 } 1567 } 1568 if (use_op == Op_CmpI) { 1569 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n); 1570 if (phi != NULL) { 1571 // If an opaque node feeds into the limit condition of a 1572 // CountedLoop, we need to process the Phi node for the 1573 // induction variable when the opaque node is removed: 1574 // the range of values taken by the Phi is now known and 1575 // so its type is also known. 1576 _worklist.push(phi); 1577 } 1578 Node* in1 = use->in(1); 1579 for (uint i = 0; i < in1->outcnt(); i++) { 1580 if (in1->raw_out(i)->Opcode() == Op_CastII) { 1581 Node* castii = in1->raw_out(i); 1582 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) { 1583 Node* ifnode = castii->in(0)->in(0); 1584 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) { 1585 // Reprocess a CastII node that may depend on an 1586 // opaque node value when the opaque node is 1587 // removed. In case it carries a dependency we can do 1588 // a better job of computing its type. 1589 _worklist.push(castii); 1590 } 1591 } 1592 } 1593 } 1594 } 1595 } 1596 1597 // If changed Cast input, check Phi users for simple cycles 1598 if (use->is_ConstraintCast()) { 1599 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1600 Node* u = use->fast_out(i2); 1601 if (u->is_Phi()) 1602 _worklist.push(u); 1603 } 1604 } 1605 // If changed LShift inputs, check RShift users for useless sign-ext 1606 if( use_op == Op_LShiftI ) { 1607 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1608 Node* u = use->fast_out(i2); 1609 if (u->Opcode() == Op_RShiftI) 1610 _worklist.push(u); 1611 } 1612 } 1613 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1614 if (use_op == Op_AddI || use_op == Op_SubI) { 1615 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1616 Node* u = use->fast_out(i2); 1617 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1618 _worklist.push(u); 1619 } 1620 } 1621 } 1622 // If changed AddP inputs, check Stores for loop invariant 1623 if( use_op == Op_AddP ) { 1624 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1625 Node* u = use->fast_out(i2); 1626 if (u->is_Mem()) 1627 _worklist.push(u); 1628 } 1629 } 1630 // If changed initialization activity, check dependent Stores 1631 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1632 InitializeNode* init = use->as_Allocate()->initialization(); 1633 if (init != NULL) { 1634 Node* imem = init->proj_out(TypeFunc::Memory); 1635 if (imem != NULL) add_users_to_worklist0(imem); 1636 } 1637 } 1638 if (use_op == Op_Initialize) { 1639 Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory); 1640 if (imem != NULL) add_users_to_worklist0(imem); 1641 } 1642 if (use_op == Op_CastP2X) { 1643 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1644 Node* u = use->fast_out(i2); 1645 if (u->Opcode() == Op_AndX) { 1646 _worklist.push(u); 1647 } 1648 } 1649 } 1650 } 1651 } 1652 1653 /** 1654 * Remove the speculative part of all types that we know of 1655 */ 1656 void PhaseIterGVN::remove_speculative_types() { 1657 assert(UseTypeSpeculation, "speculation is off"); 1658 for (uint i = 0; i < _types.Size(); i++) { 1659 const Type* t = _types.fast_lookup(i); 1660 if (t != NULL) { 1661 _types.map(i, t->remove_speculative()); 1662 } 1663 } 1664 _table.check_no_speculative_types(); 1665 } 1666 1667 //============================================================================= 1668 #ifndef PRODUCT 1669 uint PhaseCCP::_total_invokes = 0; 1670 uint PhaseCCP::_total_constants = 0; 1671 #endif 1672 //------------------------------PhaseCCP--------------------------------------- 1673 // Conditional Constant Propagation, ala Wegman & Zadeck 1674 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1675 NOT_PRODUCT( clear_constants(); ) 1676 assert( _worklist.size() == 0, "" ); 1677 // Clear out _nodes from IterGVN. Must be clear to transform call. 1678 _nodes.clear(); // Clear out from IterGVN 1679 analyze(); 1680 } 1681 1682 #ifndef PRODUCT 1683 //------------------------------~PhaseCCP-------------------------------------- 1684 PhaseCCP::~PhaseCCP() { 1685 inc_invokes(); 1686 _total_constants += count_constants(); 1687 } 1688 #endif 1689 1690 1691 #ifdef ASSERT 1692 static bool ccp_type_widens(const Type* t, const Type* t0) { 1693 assert(t->meet(t0) == t, "Not monotonic"); 1694 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1695 case Type::Int: 1696 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1697 break; 1698 case Type::Long: 1699 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1700 break; 1701 default: 1702 break; 1703 } 1704 return true; 1705 } 1706 #endif //ASSERT 1707 1708 //------------------------------analyze---------------------------------------- 1709 void PhaseCCP::analyze() { 1710 // Initialize all types to TOP, optimistic analysis 1711 for (int i = C->unique() - 1; i >= 0; i--) { 1712 _types.map(i,Type::TOP); 1713 } 1714 1715 // Push root onto worklist 1716 Unique_Node_List worklist; 1717 worklist.push(C->root()); 1718 1719 // Pull from worklist; compute new value; push changes out. 1720 // This loop is the meat of CCP. 1721 while( worklist.size() ) { 1722 Node *n = worklist.pop(); 1723 const Type *t = n->Value(this); 1724 if (t != type(n)) { 1725 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1726 #ifndef PRODUCT 1727 if( TracePhaseCCP ) { 1728 t->dump(); 1729 do { tty->print("\t"); } while (tty->position() < 16); 1730 n->dump(); 1731 } 1732 #endif 1733 set_type(n, t); 1734 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1735 Node* m = n->fast_out(i); // Get user 1736 if (m->is_Region()) { // New path to Region? Must recheck Phis too 1737 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1738 Node* p = m->fast_out(i2); // Propagate changes to uses 1739 if (p->bottom_type() != type(p)) { // If not already bottomed out 1740 worklist.push(p); // Propagate change to user 1741 } 1742 } 1743 } 1744 // If we changed the receiver type to a call, we need to revisit 1745 // the Catch following the call. It's looking for a non-NULL 1746 // receiver to know when to enable the regular fall-through path 1747 // in addition to the NullPtrException path 1748 if (m->is_Call()) { 1749 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1750 Node* p = m->fast_out(i2); // Propagate changes to uses 1751 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) { 1752 worklist.push(p->unique_out()); 1753 } 1754 } 1755 } 1756 if (m->bottom_type() != type(m)) { // If not already bottomed out 1757 worklist.push(m); // Propagate change to user 1758 } 1759 1760 // CmpU nodes can get their type information from two nodes up in the 1761 // graph (instead of from the nodes immediately above). Make sure they 1762 // are added to the worklist if nodes they depend on are updated, since 1763 // they could be missed and get wrong types otherwise. 1764 uint m_op = m->Opcode(); 1765 if (m_op == Op_AddI || m_op == Op_SubI) { 1766 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1767 Node* p = m->fast_out(i2); // Propagate changes to uses 1768 if (p->Opcode() == Op_CmpU) { 1769 // Got a CmpU which might need the new type information from node n. 1770 if(p->bottom_type() != type(p)) { // If not already bottomed out 1771 worklist.push(p); // Propagate change to user 1772 } 1773 } 1774 } 1775 } 1776 // If n is used in a counted loop exit condition then the type 1777 // of the counted loop's Phi depends on the type of n. See 1778 // PhiNode::Value(). 1779 if (m_op == Op_CmpI) { 1780 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n); 1781 if (phi != NULL) { 1782 worklist.push(phi); 1783 } 1784 } 1785 if (m_op == Op_CastP2X) { 1786 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1787 Node* u = m->fast_out(i2); 1788 if (u->Opcode() == Op_AndX) { 1789 worklist.push(u); 1790 } 1791 } 1792 } 1793 } 1794 } 1795 } 1796 } 1797 1798 //------------------------------do_transform----------------------------------- 1799 // Top level driver for the recursive transformer 1800 void PhaseCCP::do_transform() { 1801 // Correct leaves of new-space Nodes; they point to old-space. 1802 C->set_root( transform(C->root())->as_Root() ); 1803 assert( C->top(), "missing TOP node" ); 1804 assert( C->root(), "missing root" ); 1805 } 1806 1807 //------------------------------transform-------------------------------------- 1808 // Given a Node in old-space, clone him into new-space. 1809 // Convert any of his old-space children into new-space children. 1810 Node *PhaseCCP::transform( Node *n ) { 1811 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1812 if( new_node != NULL ) 1813 return new_node; // Been there, done that, return old answer 1814 new_node = transform_once(n); // Check for constant 1815 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1816 1817 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1818 GrowableArray <Node *> trstack(C->live_nodes() >> 1); 1819 1820 trstack.push(new_node); // Process children of cloned node 1821 while ( trstack.is_nonempty() ) { 1822 Node *clone = trstack.pop(); 1823 uint cnt = clone->req(); 1824 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1825 Node *input = clone->in(i); 1826 if( input != NULL ) { // Ignore NULLs 1827 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1828 if( new_input == NULL ) { 1829 new_input = transform_once(input); // Check for constant 1830 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1831 trstack.push(new_input); 1832 } 1833 assert( new_input == clone->in(i), "insanity check"); 1834 } 1835 } 1836 } 1837 return new_node; 1838 } 1839 1840 1841 //------------------------------transform_once--------------------------------- 1842 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1843 Node *PhaseCCP::transform_once( Node *n ) { 1844 const Type *t = type(n); 1845 // Constant? Use constant Node instead 1846 if( t->singleton() ) { 1847 Node *nn = n; // Default is to return the original constant 1848 if( t == Type::TOP ) { 1849 // cache my top node on the Compile instance 1850 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1851 C->set_cached_top_node(ConNode::make(Type::TOP)); 1852 set_type(C->top(), Type::TOP); 1853 } 1854 nn = C->top(); 1855 } 1856 if( !n->is_Con() ) { 1857 if( t != Type::TOP ) { 1858 nn = makecon(t); // ConNode::make(t); 1859 NOT_PRODUCT( inc_constants(); ) 1860 } else if( n->is_Region() ) { // Unreachable region 1861 // Note: nn == C->top() 1862 n->set_req(0, NULL); // Cut selfreference 1863 // Eagerly remove dead phis to avoid phis copies creation. 1864 for (DUIterator i = n->outs(); n->has_out(i); i++) { 1865 Node* m = n->out(i); 1866 if( m->is_Phi() ) { 1867 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1868 replace_node(m, nn); 1869 --i; // deleted this phi; rescan starting with next position 1870 } 1871 } 1872 } 1873 replace_node(n,nn); // Update DefUse edges for new constant 1874 } 1875 return nn; 1876 } 1877 1878 // If x is a TypeNode, capture any more-precise type permanently into Node 1879 if (t != n->bottom_type()) { 1880 hash_delete(n); // changing bottom type may force a rehash 1881 n->raise_bottom_type(t); 1882 _worklist.push(n); // n re-enters the hash table via the worklist 1883 } 1884 1885 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1886 switch( n->Opcode() ) { 1887 case Op_FastLock: // Revisit FastLocks for lock coarsening 1888 case Op_If: 1889 case Op_CountedLoopEnd: 1890 case Op_Region: 1891 case Op_Loop: 1892 case Op_CountedLoop: 1893 case Op_Conv2B: 1894 case Op_Opaque1: 1895 case Op_Opaque2: 1896 _worklist.push(n); 1897 break; 1898 default: 1899 break; 1900 } 1901 1902 return n; 1903 } 1904 1905 //---------------------------------saturate------------------------------------ 1906 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1907 const Type* limit_type) const { 1908 const Type* wide_type = new_type->widen(old_type, limit_type); 1909 if (wide_type != new_type) { // did we widen? 1910 // If so, we may have widened beyond the limit type. Clip it back down. 1911 new_type = wide_type->filter(limit_type); 1912 } 1913 return new_type; 1914 } 1915 1916 //------------------------------print_statistics------------------------------- 1917 #ifndef PRODUCT 1918 void PhaseCCP::print_statistics() { 1919 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1920 } 1921 #endif 1922 1923 1924 //============================================================================= 1925 #ifndef PRODUCT 1926 uint PhasePeephole::_total_peepholes = 0; 1927 #endif 1928 //------------------------------PhasePeephole---------------------------------- 1929 // Conditional Constant Propagation, ala Wegman & Zadeck 1930 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1931 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1932 NOT_PRODUCT( clear_peepholes(); ) 1933 } 1934 1935 #ifndef PRODUCT 1936 //------------------------------~PhasePeephole--------------------------------- 1937 PhasePeephole::~PhasePeephole() { 1938 _total_peepholes += count_peepholes(); 1939 } 1940 #endif 1941 1942 //------------------------------transform-------------------------------------- 1943 Node *PhasePeephole::transform( Node *n ) { 1944 ShouldNotCallThis(); 1945 return NULL; 1946 } 1947 1948 //------------------------------do_transform----------------------------------- 1949 void PhasePeephole::do_transform() { 1950 bool method_name_not_printed = true; 1951 1952 // Examine each basic block 1953 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 1954 Block* block = _cfg.get_block(block_number); 1955 bool block_not_printed = true; 1956 1957 // and each instruction within a block 1958 uint end_index = block->number_of_nodes(); 1959 // block->end_idx() not valid after PhaseRegAlloc 1960 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 1961 Node *n = block->get_node(instruction_index); 1962 if( n->is_Mach() ) { 1963 MachNode *m = n->as_Mach(); 1964 int deleted_count = 0; 1965 // check for peephole opportunities 1966 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count); 1967 if( m2 != NULL ) { 1968 #ifndef PRODUCT 1969 if( PrintOptoPeephole ) { 1970 // Print method, first time only 1971 if( C->method() && method_name_not_printed ) { 1972 C->method()->print_short_name(); tty->cr(); 1973 method_name_not_printed = false; 1974 } 1975 // Print this block 1976 if( Verbose && block_not_printed) { 1977 tty->print_cr("in block"); 1978 block->dump(); 1979 block_not_printed = false; 1980 } 1981 // Print instructions being deleted 1982 for( int i = (deleted_count - 1); i >= 0; --i ) { 1983 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 1984 } 1985 tty->print_cr("replaced with"); 1986 // Print new instruction 1987 m2->format(_regalloc); 1988 tty->print("\n\n"); 1989 } 1990 #endif 1991 // Remove old nodes from basic block and update instruction_index 1992 // (old nodes still exist and may have edges pointing to them 1993 // as register allocation info is stored in the allocator using 1994 // the node index to live range mappings.) 1995 uint safe_instruction_index = (instruction_index - deleted_count); 1996 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 1997 block->remove_node( instruction_index ); 1998 } 1999 // install new node after safe_instruction_index 2000 block->insert_node(m2, safe_instruction_index + 1); 2001 end_index = block->number_of_nodes() - 1; // Recompute new block size 2002 NOT_PRODUCT( inc_peepholes(); ) 2003 } 2004 } 2005 } 2006 } 2007 } 2008 2009 //------------------------------print_statistics------------------------------- 2010 #ifndef PRODUCT 2011 void PhasePeephole::print_statistics() { 2012 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2013 } 2014 #endif 2015 2016 2017 //============================================================================= 2018 //------------------------------set_req_X-------------------------------------- 2019 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2020 assert( is_not_dead(n), "can not use dead node"); 2021 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2022 Node *old = in(i); 2023 set_req(i, n); 2024 2025 // old goes dead? 2026 if( old ) { 2027 switch (old->outcnt()) { 2028 case 0: 2029 // Put into the worklist to kill later. We do not kill it now because the 2030 // recursive kill will delete the current node (this) if dead-loop exists 2031 if (!old->is_top()) 2032 igvn->_worklist.push( old ); 2033 break; 2034 case 1: 2035 if( old->is_Store() || old->has_special_unique_user() ) 2036 igvn->add_users_to_worklist( old ); 2037 break; 2038 case 2: 2039 if( old->is_Store() ) 2040 igvn->add_users_to_worklist( old ); 2041 if( old->Opcode() == Op_Region ) 2042 igvn->_worklist.push(old); 2043 break; 2044 case 3: 2045 if( old->Opcode() == Op_Region ) { 2046 igvn->_worklist.push(old); 2047 igvn->add_users_to_worklist( old ); 2048 } 2049 break; 2050 default: 2051 break; 2052 } 2053 } 2054 2055 } 2056 2057 //-------------------------------replace_by----------------------------------- 2058 // Using def-use info, replace one node for another. Follow the def-use info 2059 // to all users of the OLD node. Then make all uses point to the NEW node. 2060 void Node::replace_by(Node *new_node) { 2061 assert(!is_top(), "top node has no DU info"); 2062 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2063 Node* use = last_out(i); 2064 uint uses_found = 0; 2065 for (uint j = 0; j < use->len(); j++) { 2066 if (use->in(j) == this) { 2067 if (j < use->req()) 2068 use->set_req(j, new_node); 2069 else use->set_prec(j, new_node); 2070 uses_found++; 2071 } 2072 } 2073 i -= uses_found; // we deleted 1 or more copies of this edge 2074 } 2075 } 2076 2077 //============================================================================= 2078 //----------------------------------------------------------------------------- 2079 void Type_Array::grow( uint i ) { 2080 if( !_max ) { 2081 _max = 1; 2082 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2083 _types[0] = NULL; 2084 } 2085 uint old = _max; 2086 while( i >= _max ) _max <<= 1; // Double to fit 2087 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2088 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2089 } 2090 2091 //------------------------------dump------------------------------------------- 2092 #ifndef PRODUCT 2093 void Type_Array::dump() const { 2094 uint max = Size(); 2095 for( uint i = 0; i < max; i++ ) { 2096 if( _types[i] != NULL ) { 2097 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2098 } 2099 } 2100 } 2101 #endif