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