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