1 /* 2 * Copyright (c) 1997, 2011, 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(), 1) 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(), 1) 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 //------------------------------hash_find-------------------------------------- 79 // Find in hash table 80 Node *NodeHash::hash_find( const Node *n ) { 81 // ((Node*)n)->set_hash( n->hash() ); 82 uint hash = n->hash(); 83 if (hash == Node::NO_HASH) { 84 debug_only( _lookup_misses++ ); 85 return NULL; 86 } 87 uint key = hash & (_max-1); 88 uint stride = key | 0x01; 89 debug_only( _look_probes++ ); 90 Node *k = _table[key]; // Get hashed value 91 if( !k ) { // ?Miss? 92 debug_only( _lookup_misses++ ); 93 return NULL; // Miss! 94 } 95 96 int op = n->Opcode(); 97 uint req = n->req(); 98 while( 1 ) { // While probing hash table 99 if( k->req() == req && // Same count of inputs 100 k->Opcode() == op ) { // Same Opcode 101 for( uint i=0; i<req; i++ ) 102 if( n->in(i)!=k->in(i)) // Different inputs? 103 goto collision; // "goto" is a speed hack... 104 if( n->cmp(*k) ) { // Check for any special bits 105 debug_only( _lookup_hits++ ); 106 return k; // Hit! 107 } 108 } 109 collision: 110 debug_only( _look_probes++ ); 111 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime 112 k = _table[key]; // Get hashed value 113 if( !k ) { // ?Miss? 114 debug_only( _lookup_misses++ ); 115 return NULL; // Miss! 116 } 117 } 118 ShouldNotReachHere(); 119 return NULL; 120 } 121 122 //------------------------------hash_find_insert------------------------------- 123 // Find in hash table, insert if not already present 124 // Used to preserve unique entries in hash table 125 Node *NodeHash::hash_find_insert( Node *n ) { 126 // n->set_hash( ); 127 uint hash = n->hash(); 128 if (hash == Node::NO_HASH) { 129 debug_only( _lookup_misses++ ); 130 return NULL; 131 } 132 uint key = hash & (_max-1); 133 uint stride = key | 0x01; // stride must be relatively prime to table siz 134 uint first_sentinel = 0; // replace a sentinel if seen. 135 debug_only( _look_probes++ ); 136 Node *k = _table[key]; // Get hashed value 137 if( !k ) { // ?Miss? 138 debug_only( _lookup_misses++ ); 139 _table[key] = n; // Insert into table! 140 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 141 check_grow(); // Grow table if insert hit limit 142 return NULL; // Miss! 143 } 144 else if( k == _sentinel ) { 145 first_sentinel = key; // Can insert here 146 } 147 148 int op = n->Opcode(); 149 uint req = n->req(); 150 while( 1 ) { // While probing hash table 151 if( k->req() == req && // Same count of inputs 152 k->Opcode() == op ) { // Same Opcode 153 for( uint i=0; i<req; i++ ) 154 if( n->in(i)!=k->in(i)) // Different inputs? 155 goto collision; // "goto" is a speed hack... 156 if( n->cmp(*k) ) { // Check for any special bits 157 debug_only( _lookup_hits++ ); 158 return k; // Hit! 159 } 160 } 161 collision: 162 debug_only( _look_probes++ ); 163 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 164 k = _table[key]; // Get hashed value 165 if( !k ) { // ?Miss? 166 debug_only( _lookup_misses++ ); 167 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel? 168 _table[key] = n; // Insert into table! 169 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 170 check_grow(); // Grow table if insert hit limit 171 return NULL; // Miss! 172 } 173 else if( first_sentinel == 0 && k == _sentinel ) { 174 first_sentinel = key; // Can insert here 175 } 176 177 } 178 ShouldNotReachHere(); 179 return NULL; 180 } 181 182 //------------------------------hash_insert------------------------------------ 183 // Insert into hash table 184 void NodeHash::hash_insert( Node *n ) { 185 // // "conflict" comments -- print nodes that conflict 186 // bool conflict = false; 187 // n->set_hash(); 188 uint hash = n->hash(); 189 if (hash == Node::NO_HASH) { 190 return; 191 } 192 check_grow(); 193 uint key = hash & (_max-1); 194 uint stride = key | 0x01; 195 196 while( 1 ) { // While probing hash table 197 debug_only( _insert_probes++ ); 198 Node *k = _table[key]; // Get hashed value 199 if( !k || (k == _sentinel) ) break; // Found a slot 200 assert( k != n, "already inserted" ); 201 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; } 202 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 203 } 204 _table[key] = n; // Insert into table! 205 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 206 // if( conflict ) { n->dump(); } 207 } 208 209 //------------------------------hash_delete------------------------------------ 210 // Replace in hash table with sentinel 211 bool NodeHash::hash_delete( const Node *n ) { 212 Node *k; 213 uint hash = n->hash(); 214 if (hash == Node::NO_HASH) { 215 debug_only( _delete_misses++ ); 216 return false; 217 } 218 uint key = hash & (_max-1); 219 uint stride = key | 0x01; 220 debug_only( uint counter = 0; ); 221 for( ; /* (k != NULL) && (k != _sentinel) */; ) { 222 debug_only( counter++ ); 223 debug_only( _delete_probes++ ); 224 k = _table[key]; // Get hashed value 225 if( !k ) { // Miss? 226 debug_only( _delete_misses++ ); 227 #ifdef ASSERT 228 if( VerifyOpto ) { 229 for( uint i=0; i < _max; i++ ) 230 assert( _table[i] != n, "changed edges with rehashing" ); 231 } 232 #endif 233 return false; // Miss! Not in chain 234 } 235 else if( n == k ) { 236 debug_only( _delete_hits++ ); 237 _table[key] = _sentinel; // Hit! Label as deleted entry 238 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table. 239 return true; 240 } 241 else { 242 // collision: move through table with prime offset 243 key = (key + stride/*7*/) & (_max-1); 244 assert( counter <= _insert_limit, "Cycle in hash-table"); 245 } 246 } 247 ShouldNotReachHere(); 248 return false; 249 } 250 251 //------------------------------round_up--------------------------------------- 252 // Round up to nearest power of 2 253 uint NodeHash::round_up( uint x ) { 254 x += (x>>2); // Add 25% slop 255 if( x <16 ) return 16; // Small stuff 256 uint i=16; 257 while( i < x ) i <<= 1; // Double to fit 258 return i; // Return hash table size 259 } 260 261 //------------------------------grow------------------------------------------- 262 // Grow _table to next power of 2 and insert old entries 263 void NodeHash::grow() { 264 // Record old state 265 uint old_max = _max; 266 Node **old_table = _table; 267 // Construct new table with twice the space 268 _grows++; 269 _total_inserts += _inserts; 270 _total_insert_probes += _insert_probes; 271 _inserts = 0; 272 _insert_probes = 0; 273 _max = _max << 1; 274 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) ); 275 memset(_table,0,sizeof(Node*)*_max); 276 _insert_limit = insert_limit(); 277 // Insert old entries into the new table 278 for( uint i = 0; i < old_max; i++ ) { 279 Node *m = *old_table++; 280 if( !m || m == _sentinel ) continue; 281 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table. 282 hash_insert(m); 283 } 284 } 285 286 //------------------------------clear------------------------------------------ 287 // Clear all entries in _table to NULL but keep storage 288 void NodeHash::clear() { 289 #ifdef ASSERT 290 // Unlock all nodes upon removal from table. 291 for (uint i = 0; i < _max; i++) { 292 Node* n = _table[i]; 293 if (!n || n == _sentinel) continue; 294 n->exit_hash_lock(); 295 } 296 #endif 297 298 memset( _table, 0, _max * sizeof(Node*) ); 299 } 300 301 //-----------------------remove_useless_nodes---------------------------------- 302 // Remove useless nodes from value table, 303 // implementation does not depend on hash function 304 void NodeHash::remove_useless_nodes(VectorSet &useful) { 305 306 // Dead nodes in the hash table inherited from GVN should not replace 307 // existing nodes, remove dead nodes. 308 uint max = size(); 309 Node *sentinel_node = sentinel(); 310 for( uint i = 0; i < max; ++i ) { 311 Node *n = at(i); 312 if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) { 313 debug_only(n->exit_hash_lock()); // Unlock the node when removed 314 _table[i] = sentinel_node; // Replace with placeholder 315 } 316 } 317 } 318 319 #ifndef PRODUCT 320 //------------------------------dump------------------------------------------- 321 // Dump statistics for the hash table 322 void NodeHash::dump() { 323 _total_inserts += _inserts; 324 _total_insert_probes += _insert_probes; 325 if( PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0) ) { // PrintOptoGVN 326 if( PrintCompilation2 ) { 327 for( uint i=0; i<_max; i++ ) 328 if( _table[i] ) 329 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx); 330 } 331 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max); 332 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0); 333 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses)); 334 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts); 335 // sentinels increase lookup cost, but not insert cost 336 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function"); 337 assert( _inserts+(_inserts>>3) < _max, "table too full" ); 338 assert( _inserts*3+100 >= _insert_probes, "bad hash function" ); 339 } 340 } 341 342 Node *NodeHash::find_index(uint idx) { // For debugging 343 // Find an entry by its index value 344 for( uint i = 0; i < _max; i++ ) { 345 Node *m = _table[i]; 346 if( !m || m == _sentinel ) continue; 347 if( m->_idx == (uint)idx ) return m; 348 } 349 return NULL; 350 } 351 #endif 352 353 #ifdef ASSERT 354 NodeHash::~NodeHash() { 355 // Unlock all nodes upon destruction of table. 356 if (_table != (Node**)badAddress) clear(); 357 } 358 359 void NodeHash::operator=(const NodeHash& nh) { 360 // Unlock all nodes upon replacement of table. 361 if (&nh == this) return; 362 if (_table != (Node**)badAddress) clear(); 363 memcpy(this, &nh, sizeof(*this)); 364 // Do not increment hash_lock counts again. 365 // Instead, be sure we never again use the source table. 366 ((NodeHash*)&nh)->_table = (Node**)badAddress; 367 } 368 369 370 #endif 371 372 373 //============================================================================= 374 //------------------------------PhaseRemoveUseless----------------------------- 375 // 1) Use a breadthfirst walk to collect useful nodes reachable from root. 376 PhaseRemoveUseless::PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist ) : Phase(Remove_Useless), 377 _useful(Thread::current()->resource_area()) { 378 379 // Implementation requires 'UseLoopSafepoints == true' and an edge from root 380 // to each SafePointNode at a backward branch. Inserted in add_safepoint(). 381 if( !UseLoopSafepoints || !OptoRemoveUseless ) return; 382 383 // Identify nodes that are reachable from below, useful. 384 C->identify_useful_nodes(_useful); 385 386 // Remove all useless nodes from PhaseValues' recorded types 387 // Must be done before disconnecting nodes to preserve hash-table-invariant 388 gvn->remove_useless_nodes(_useful.member_set()); 389 390 // Remove all useless nodes from future worklist 391 worklist->remove_useless_nodes(_useful.member_set()); 392 393 // Disconnect 'useless' nodes that are adjacent to useful nodes 394 C->remove_useless_nodes(_useful); 395 396 // Remove edges from "root" to each SafePoint at a backward branch. 397 // They were inserted during parsing (see add_safepoint()) to make infinite 398 // loops without calls or exceptions visible to root, i.e., useful. 399 Node *root = C->root(); 400 if( root != NULL ) { 401 for( uint i = root->req(); i < root->len(); ++i ) { 402 Node *n = root->in(i); 403 if( n != NULL && n->is_SafePoint() ) { 404 root->rm_prec(i); 405 --i; 406 } 407 } 408 } 409 } 410 411 412 //============================================================================= 413 //------------------------------PhaseTransform--------------------------------- 414 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 415 _arena(Thread::current()->resource_area()), 416 _nodes(_arena), 417 _types(_arena) 418 { 419 init_con_caches(); 420 #ifndef PRODUCT 421 clear_progress(); 422 clear_transforms(); 423 set_allow_progress(true); 424 #endif 425 // Force allocation for currently existing nodes 426 _types.map(C->unique(), NULL); 427 } 428 429 //------------------------------PhaseTransform--------------------------------- 430 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 431 _arena(arena), 432 _nodes(arena), 433 _types(arena) 434 { 435 init_con_caches(); 436 #ifndef PRODUCT 437 clear_progress(); 438 clear_transforms(); 439 set_allow_progress(true); 440 #endif 441 // Force allocation for currently existing nodes 442 _types.map(C->unique(), NULL); 443 } 444 445 //------------------------------PhaseTransform--------------------------------- 446 // Initialize with previously generated type information 447 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 448 _arena(pt->_arena), 449 _nodes(pt->_nodes), 450 _types(pt->_types) 451 { 452 init_con_caches(); 453 #ifndef PRODUCT 454 clear_progress(); 455 clear_transforms(); 456 set_allow_progress(true); 457 #endif 458 } 459 460 void PhaseTransform::init_con_caches() { 461 memset(_icons,0,sizeof(_icons)); 462 memset(_lcons,0,sizeof(_lcons)); 463 memset(_zcons,0,sizeof(_zcons)); 464 } 465 466 467 //--------------------------------find_int_type-------------------------------- 468 const TypeInt* PhaseTransform::find_int_type(Node* n) { 469 if (n == NULL) return NULL; 470 // Call type_or_null(n) to determine node's type since we might be in 471 // parse phase and call n->Value() may return wrong type. 472 // (For example, a phi node at the beginning of loop parsing is not ready.) 473 const Type* t = type_or_null(n); 474 if (t == NULL) return NULL; 475 return t->isa_int(); 476 } 477 478 479 //-------------------------------find_long_type-------------------------------- 480 const TypeLong* PhaseTransform::find_long_type(Node* n) { 481 if (n == NULL) return NULL; 482 // (See comment above on type_or_null.) 483 const Type* t = type_or_null(n); 484 if (t == NULL) return NULL; 485 return t->isa_long(); 486 } 487 488 489 #ifndef PRODUCT 490 void PhaseTransform::dump_old2new_map() const { 491 _nodes.dump(); 492 } 493 494 void PhaseTransform::dump_new( uint nidx ) const { 495 for( uint i=0; i<_nodes.Size(); i++ ) 496 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 497 _nodes[i]->dump(); 498 tty->cr(); 499 tty->print_cr("Old index= %d",i); 500 return; 501 } 502 tty->print_cr("Node %d not found in the new indices", nidx); 503 } 504 505 //------------------------------dump_types------------------------------------- 506 void PhaseTransform::dump_types( ) const { 507 _types.dump(); 508 } 509 510 //------------------------------dump_nodes_and_types--------------------------- 511 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 512 VectorSet visited(Thread::current()->resource_area()); 513 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 514 } 515 516 //------------------------------dump_nodes_and_types_recur--------------------- 517 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 518 if( !n ) return; 519 if( depth == 0 ) return; 520 if( visited.test_set(n->_idx) ) return; 521 for( uint i=0; i<n->len(); i++ ) { 522 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 523 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 524 } 525 n->dump(); 526 if (type_or_null(n) != NULL) { 527 tty->print(" "); type(n)->dump(); tty->cr(); 528 } 529 } 530 531 #endif 532 533 534 //============================================================================= 535 //------------------------------PhaseValues------------------------------------ 536 // Set minimum table size to "255" 537 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 538 NOT_PRODUCT( clear_new_values(); ) 539 } 540 541 //------------------------------PhaseValues------------------------------------ 542 // Set minimum table size to "255" 543 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 544 _table(&ptv->_table) { 545 NOT_PRODUCT( clear_new_values(); ) 546 } 547 548 //------------------------------PhaseValues------------------------------------ 549 // Used by +VerifyOpto. Clear out hash table but copy _types array. 550 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 551 _table(ptv->arena(),ptv->_table.size()) { 552 NOT_PRODUCT( clear_new_values(); ) 553 } 554 555 //------------------------------~PhaseValues----------------------------------- 556 #ifndef PRODUCT 557 PhaseValues::~PhaseValues() { 558 _table.dump(); 559 560 // Statistics for value progress and efficiency 561 if( PrintCompilation && Verbose && WizardMode ) { 562 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 563 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 564 if( made_transforms() != 0 ) { 565 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 566 } else { 567 tty->cr(); 568 } 569 } 570 } 571 #endif 572 573 //------------------------------makecon---------------------------------------- 574 ConNode* PhaseTransform::makecon(const Type *t) { 575 assert(t->singleton(), "must be a constant"); 576 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 577 switch (t->base()) { // fast paths 578 case Type::Half: 579 case Type::Top: return (ConNode*) C->top(); 580 case Type::Int: return intcon( t->is_int()->get_con() ); 581 case Type::Long: return longcon( t->is_long()->get_con() ); 582 } 583 if (t->is_zero_type()) 584 return zerocon(t->basic_type()); 585 return uncached_makecon(t); 586 } 587 588 //--------------------------uncached_makecon----------------------------------- 589 // Make an idealized constant - one of ConINode, ConPNode, etc. 590 ConNode* PhaseValues::uncached_makecon(const Type *t) { 591 assert(t->singleton(), "must be a constant"); 592 ConNode* x = ConNode::make(C, t); 593 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 594 if (k == NULL) { 595 set_type(x, t); // Missed, provide type mapping 596 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 597 if (nna != NULL) { 598 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 599 loc->clear(); // do not put debug info on constants 600 } 601 } else { 602 x->destruct(); // Hit, destroy duplicate constant 603 x = k; // use existing constant 604 } 605 return x; 606 } 607 608 //------------------------------intcon----------------------------------------- 609 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 610 ConINode* PhaseTransform::intcon(int i) { 611 // Small integer? Check cache! Check that cached node is not dead 612 if (i >= _icon_min && i <= _icon_max) { 613 ConINode* icon = _icons[i-_icon_min]; 614 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 615 return icon; 616 } 617 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 618 assert(icon->is_Con(), ""); 619 if (i >= _icon_min && i <= _icon_max) 620 _icons[i-_icon_min] = icon; // Cache small integers 621 return icon; 622 } 623 624 //------------------------------longcon---------------------------------------- 625 // Fast long constant. 626 ConLNode* PhaseTransform::longcon(jlong l) { 627 // Small integer? Check cache! Check that cached node is not dead 628 if (l >= _lcon_min && l <= _lcon_max) { 629 ConLNode* lcon = _lcons[l-_lcon_min]; 630 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 631 return lcon; 632 } 633 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 634 assert(lcon->is_Con(), ""); 635 if (l >= _lcon_min && l <= _lcon_max) 636 _lcons[l-_lcon_min] = lcon; // Cache small integers 637 return lcon; 638 } 639 640 //------------------------------zerocon----------------------------------------- 641 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 642 ConNode* PhaseTransform::zerocon(BasicType bt) { 643 assert((uint)bt <= _zcon_max, "domain check"); 644 ConNode* zcon = _zcons[bt]; 645 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 646 return zcon; 647 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 648 _zcons[bt] = zcon; 649 return zcon; 650 } 651 652 653 654 //============================================================================= 655 //------------------------------transform-------------------------------------- 656 // Return a node which computes the same function as this node, but in a 657 // faster or cheaper fashion. 658 Node *PhaseGVN::transform( Node *n ) { 659 return transform_no_reclaim(n); 660 } 661 662 //------------------------------transform-------------------------------------- 663 // Return a node which computes the same function as this node, but 664 // in a faster or cheaper fashion. 665 Node *PhaseGVN::transform_no_reclaim( Node *n ) { 666 NOT_PRODUCT( set_transforms(); ) 667 668 // Apply the Ideal call in a loop until it no longer applies 669 Node *k = n; 670 NOT_PRODUCT( uint loop_count = 0; ) 671 while( 1 ) { 672 Node *i = k->Ideal(this, /*can_reshape=*/false); 673 if( !i ) break; 674 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 675 k = i; 676 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 677 } 678 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 679 680 681 // If brand new node, make space in type array. 682 ensure_type_or_null(k); 683 684 // Since I just called 'Value' to compute the set of run-time values 685 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 686 // cache Value. Later requests for the local phase->type of this Node can 687 // use the cached Value instead of suffering with 'bottom_type'. 688 const Type *t = k->Value(this); // Get runtime Value set 689 assert(t != NULL, "value sanity"); 690 if (type_or_null(k) != t) { 691 #ifndef PRODUCT 692 // Do not count initial visit to node as a transformation 693 if (type_or_null(k) == NULL) { 694 inc_new_values(); 695 set_progress(); 696 } 697 #endif 698 set_type(k, t); 699 // If k is a TypeNode, capture any more-precise type permanently into Node 700 k->raise_bottom_type(t); 701 } 702 703 if( t->singleton() && !k->is_Con() ) { 704 NOT_PRODUCT( set_progress(); ) 705 return makecon(t); // Turn into a constant 706 } 707 708 // Now check for Identities 709 Node *i = k->Identity(this); // Look for a nearby replacement 710 if( i != k ) { // Found? Return replacement! 711 NOT_PRODUCT( set_progress(); ) 712 return i; 713 } 714 715 // Global Value Numbering 716 i = hash_find_insert(k); // Insert if new 717 if( i && (i != k) ) { 718 // Return the pre-existing node 719 NOT_PRODUCT( set_progress(); ) 720 return i; 721 } 722 723 // Return Idealized original 724 return k; 725 } 726 727 #ifdef ASSERT 728 //------------------------------dead_loop_check-------------------------------- 729 // Check for a simple dead loop when a data node references itself directly 730 // or through an other data node excluding cons and phis. 731 void PhaseGVN::dead_loop_check( Node *n ) { 732 // Phi may reference itself in a loop 733 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 734 // Do 2 levels check and only data inputs. 735 bool no_dead_loop = true; 736 uint cnt = n->req(); 737 for (uint i = 1; i < cnt && no_dead_loop; i++) { 738 Node *in = n->in(i); 739 if (in == n) { 740 no_dead_loop = false; 741 } else if (in != NULL && !in->is_dead_loop_safe()) { 742 uint icnt = in->req(); 743 for (uint j = 1; j < icnt && no_dead_loop; j++) { 744 if (in->in(j) == n || in->in(j) == in) 745 no_dead_loop = false; 746 } 747 } 748 } 749 if (!no_dead_loop) n->dump(3); 750 assert(no_dead_loop, "dead loop detected"); 751 } 752 } 753 #endif 754 755 //============================================================================= 756 //------------------------------PhaseIterGVN----------------------------------- 757 // Initialize hash table to fresh and clean for +VerifyOpto 758 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ), 759 _delay_transform(false) { 760 } 761 762 //------------------------------PhaseIterGVN----------------------------------- 763 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 764 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 765 _worklist( igvn->_worklist ), 766 _delay_transform(igvn->_delay_transform) 767 { 768 } 769 770 //------------------------------PhaseIterGVN----------------------------------- 771 // Initialize with previous PhaseGVN info from Parser 772 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 773 _worklist(*C->for_igvn()), 774 _delay_transform(false) 775 { 776 uint max; 777 778 // Dead nodes in the hash table inherited from GVN were not treated as 779 // roots during def-use info creation; hence they represent an invisible 780 // use. Clear them out. 781 max = _table.size(); 782 for( uint i = 0; i < max; ++i ) { 783 Node *n = _table.at(i); 784 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 785 if( n->is_top() ) continue; 786 assert( false, "Parse::remove_useless_nodes missed this node"); 787 hash_delete(n); 788 } 789 } 790 791 // Any Phis or Regions on the worklist probably had uses that could not 792 // make more progress because the uses were made while the Phis and Regions 793 // were in half-built states. Put all uses of Phis and Regions on worklist. 794 max = _worklist.size(); 795 for( uint j = 0; j < max; j++ ) { 796 Node *n = _worklist.at(j); 797 uint uop = n->Opcode(); 798 if( uop == Op_Phi || uop == Op_Region || 799 n->is_Type() || 800 n->is_Mem() ) 801 add_users_to_worklist(n); 802 } 803 } 804 805 806 #ifndef PRODUCT 807 void PhaseIterGVN::verify_step(Node* n) { 808 _verify_window[_verify_counter % _verify_window_size] = n; 809 ++_verify_counter; 810 ResourceMark rm; 811 ResourceArea *area = Thread::current()->resource_area(); 812 VectorSet old_space(area), new_space(area); 813 if (C->unique() < 1000 || 814 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 815 ++_verify_full_passes; 816 Node::verify_recur(C->root(), -1, old_space, new_space); 817 } 818 const int verify_depth = 4; 819 for ( int i = 0; i < _verify_window_size; i++ ) { 820 Node* n = _verify_window[i]; 821 if ( n == NULL ) continue; 822 if( n->in(0) == NodeSentinel ) { // xform_idom 823 _verify_window[i] = n->in(1); 824 --i; continue; 825 } 826 // Typical fanout is 1-2, so this call visits about 6 nodes. 827 Node::verify_recur(n, verify_depth, old_space, new_space); 828 } 829 } 830 #endif 831 832 833 //------------------------------init_worklist---------------------------------- 834 // Initialize worklist for each node. 835 void PhaseIterGVN::init_worklist( Node *n ) { 836 if( _worklist.member(n) ) return; 837 _worklist.push(n); 838 uint cnt = n->req(); 839 for( uint i =0 ; i < cnt; i++ ) { 840 Node *m = n->in(i); 841 if( m ) init_worklist(m); 842 } 843 } 844 845 //------------------------------optimize--------------------------------------- 846 void PhaseIterGVN::optimize() { 847 debug_only(uint num_processed = 0;); 848 #ifndef PRODUCT 849 { 850 _verify_counter = 0; 851 _verify_full_passes = 0; 852 for ( int i = 0; i < _verify_window_size; i++ ) { 853 _verify_window[i] = NULL; 854 } 855 } 856 #endif 857 858 #ifdef ASSERT 859 Node* prev = NULL; 860 uint rep_cnt = 0; 861 #endif 862 uint loop_count = 0; 863 864 // Pull from worklist; transform node; 865 // If node has changed: update edge info and put uses on worklist. 866 while( _worklist.size() ) { 867 if (C->check_node_count(NodeLimitFudgeFactor * 2, 868 "out of nodes optimizing method")) { 869 return; 870 } 871 Node *n = _worklist.pop(); 872 if (++loop_count >= K * C->unique()) { 873 debug_only(n->dump(4);) 874 assert(false, "infinite loop in PhaseIterGVN::optimize"); 875 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 876 return; 877 } 878 #ifdef ASSERT 879 if (n == prev) { 880 if (++rep_cnt > 3) { 881 n->dump(4); 882 assert(false, "loop in Ideal transformation"); 883 } 884 } else { 885 rep_cnt = 0; 886 } 887 prev = n; 888 #endif 889 if (TraceIterativeGVN && Verbose) { 890 tty->print(" Pop "); 891 NOT_PRODUCT( n->dump(); ) 892 debug_only(if( (num_processed++ % 100) == 0 ) _worklist.print_set();) 893 } 894 895 if (n->outcnt() != 0) { 896 897 #ifndef PRODUCT 898 uint wlsize = _worklist.size(); 899 const Type* oldtype = type_or_null(n); 900 #endif //PRODUCT 901 902 Node *nn = transform_old(n); 903 904 #ifndef PRODUCT 905 if (TraceIterativeGVN) { 906 const Type* newtype = type_or_null(n); 907 if (nn != n) { 908 // print old node 909 tty->print("< "); 910 if (oldtype != newtype && oldtype != NULL) { 911 oldtype->dump(); 912 } 913 do { tty->print("\t"); } while (tty->position() < 16); 914 tty->print("<"); 915 n->dump(); 916 } 917 if (oldtype != newtype || nn != n) { 918 // print new node and/or new type 919 if (oldtype == NULL) { 920 tty->print("* "); 921 } else if (nn != n) { 922 tty->print("> "); 923 } else { 924 tty->print("= "); 925 } 926 if (newtype == NULL) { 927 tty->print("null"); 928 } else { 929 newtype->dump(); 930 } 931 do { tty->print("\t"); } while (tty->position() < 16); 932 nn->dump(); 933 } 934 if (Verbose && wlsize < _worklist.size()) { 935 tty->print(" Push {"); 936 while (wlsize != _worklist.size()) { 937 Node* pushed = _worklist.at(wlsize++); 938 tty->print(" %d", pushed->_idx); 939 } 940 tty->print_cr(" }"); 941 } 942 } 943 if( VerifyIterativeGVN && nn != n ) { 944 verify_step((Node*) NULL); // ignore n, it might be subsumed 945 } 946 #endif 947 } else if (!n->is_top()) { 948 remove_dead_node(n); 949 } 950 } 951 952 #ifndef PRODUCT 953 C->verify_graph_edges(); 954 if( VerifyOpto && allow_progress() ) { 955 // Must turn off allow_progress to enable assert and break recursion 956 C->root()->verify(); 957 { // Check if any progress was missed using IterGVN 958 // Def-Use info enables transformations not attempted in wash-pass 959 // e.g. Region/Phi cleanup, ... 960 // Null-check elision -- may not have reached fixpoint 961 // do not propagate to dominated nodes 962 ResourceMark rm; 963 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 964 // Fill worklist completely 965 igvn2.init_worklist(C->root()); 966 967 igvn2.set_allow_progress(false); 968 igvn2.optimize(); 969 igvn2.set_allow_progress(true); 970 } 971 } 972 if ( VerifyIterativeGVN && PrintOpto ) { 973 if ( _verify_counter == _verify_full_passes ) 974 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 975 _verify_full_passes); 976 else 977 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 978 _verify_counter, _verify_full_passes); 979 } 980 #endif 981 } 982 983 984 //------------------register_new_node_with_optimizer--------------------------- 985 // Register a new node with the optimizer. Update the types array, the def-use 986 // info. Put on worklist. 987 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 988 set_type_bottom(n); 989 _worklist.push(n); 990 if (orig != NULL) C->copy_node_notes_to(n, orig); 991 return n; 992 } 993 994 //------------------------------transform-------------------------------------- 995 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 996 Node *PhaseIterGVN::transform( Node *n ) { 997 if (_delay_transform) { 998 // Register the node but don't optimize for now 999 register_new_node_with_optimizer(n); 1000 return n; 1001 } 1002 1003 // If brand new node, make space in type array, and give it a type. 1004 ensure_type_or_null(n); 1005 if (type_or_null(n) == NULL) { 1006 set_type_bottom(n); 1007 } 1008 1009 return transform_old(n); 1010 } 1011 1012 //------------------------------transform_old---------------------------------- 1013 Node *PhaseIterGVN::transform_old( Node *n ) { 1014 #ifndef PRODUCT 1015 debug_only(uint loop_count = 0;); 1016 set_transforms(); 1017 #endif 1018 // Remove 'n' from hash table in case it gets modified 1019 _table.hash_delete(n); 1020 if( VerifyIterativeGVN ) { 1021 assert( !_table.find_index(n->_idx), "found duplicate entry in table"); 1022 } 1023 1024 // Apply the Ideal call in a loop until it no longer applies 1025 Node *k = n; 1026 DEBUG_ONLY(dead_loop_check(k);) 1027 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1028 Node *i = k->Ideal(this, /*can_reshape=*/true); 1029 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1030 #ifndef PRODUCT 1031 if( VerifyIterativeGVN ) 1032 verify_step(k); 1033 if( i && VerifyOpto ) { 1034 if( !allow_progress() ) { 1035 if (i->is_Add() && i->outcnt() == 1) { 1036 // Switched input to left side because this is the only use 1037 } else if( i->is_If() && (i->in(0) == NULL) ) { 1038 // This IF is dead because it is dominated by an equivalent IF When 1039 // dominating if changed, info is not propagated sparsely to 'this' 1040 // Propagating this info further will spuriously identify other 1041 // progress. 1042 return i; 1043 } else 1044 set_progress(); 1045 } else 1046 set_progress(); 1047 } 1048 #endif 1049 1050 while( i ) { 1051 #ifndef PRODUCT 1052 debug_only( if( loop_count >= K ) i->dump(4); ) 1053 assert(loop_count < K, "infinite loop in PhaseIterGVN::transform"); 1054 debug_only( loop_count++; ) 1055 #endif 1056 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1057 // Made a change; put users of original Node on worklist 1058 add_users_to_worklist( k ); 1059 // Replacing root of transform tree? 1060 if( k != i ) { 1061 // Make users of old Node now use new. 1062 subsume_node( k, i ); 1063 k = i; 1064 } 1065 DEBUG_ONLY(dead_loop_check(k);) 1066 // Try idealizing again 1067 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1068 i = k->Ideal(this, /*can_reshape=*/true); 1069 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1070 #ifndef PRODUCT 1071 if( VerifyIterativeGVN ) 1072 verify_step(k); 1073 if( i && VerifyOpto ) set_progress(); 1074 #endif 1075 } 1076 1077 // If brand new node, make space in type array. 1078 ensure_type_or_null(k); 1079 1080 // See what kind of values 'k' takes on at runtime 1081 const Type *t = k->Value(this); 1082 assert(t != NULL, "value sanity"); 1083 1084 // Since I just called 'Value' to compute the set of run-time values 1085 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1086 // cache Value. Later requests for the local phase->type of this Node can 1087 // use the cached Value instead of suffering with 'bottom_type'. 1088 if (t != type_or_null(k)) { 1089 NOT_PRODUCT( set_progress(); ) 1090 NOT_PRODUCT( inc_new_values();) 1091 set_type(k, t); 1092 // If k is a TypeNode, capture any more-precise type permanently into Node 1093 k->raise_bottom_type(t); 1094 // Move users of node to worklist 1095 add_users_to_worklist( k ); 1096 } 1097 1098 // If 'k' computes a constant, replace it with a constant 1099 if( t->singleton() && !k->is_Con() ) { 1100 NOT_PRODUCT( set_progress(); ) 1101 Node *con = makecon(t); // Make a constant 1102 add_users_to_worklist( k ); 1103 subsume_node( k, con ); // Everybody using k now uses con 1104 return con; 1105 } 1106 1107 // Now check for Identities 1108 i = k->Identity(this); // Look for a nearby replacement 1109 if( i != k ) { // Found? Return replacement! 1110 NOT_PRODUCT( set_progress(); ) 1111 add_users_to_worklist( k ); 1112 subsume_node( k, i ); // Everybody using k now uses i 1113 return i; 1114 } 1115 1116 // Global Value Numbering 1117 i = hash_find_insert(k); // Check for pre-existing node 1118 if( i && (i != k) ) { 1119 // Return the pre-existing node if it isn't dead 1120 NOT_PRODUCT( set_progress(); ) 1121 add_users_to_worklist( k ); 1122 subsume_node( k, i ); // Everybody using k now uses i 1123 return i; 1124 } 1125 1126 // Return Idealized original 1127 return k; 1128 } 1129 1130 //---------------------------------saturate------------------------------------ 1131 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1132 const Type* limit_type) const { 1133 return new_type->narrow(old_type); 1134 } 1135 1136 //------------------------------remove_globally_dead_node---------------------- 1137 // Kill a globally dead Node. All uses are also globally dead and are 1138 // aggressively trimmed. 1139 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1140 assert(dead != C->root(), "killing root, eh?"); 1141 if (dead->is_top()) return; 1142 NOT_PRODUCT( set_progress(); ) 1143 // Remove from iterative worklist 1144 _worklist.remove(dead); 1145 if (!dead->is_Con()) { // Don't kill cons but uses 1146 // Remove from hash table 1147 _table.hash_delete( dead ); 1148 // Smash all inputs to 'dead', isolating him completely 1149 for( uint i = 0; i < dead->req(); i++ ) { 1150 Node *in = dead->in(i); 1151 if( in ) { // Points to something? 1152 dead->set_req(i,NULL); // Kill the edge 1153 if (in->outcnt() == 0 && in != C->top()) {// Made input go dead? 1154 remove_dead_node(in); // Recursively remove 1155 } else if (in->outcnt() == 1 && 1156 in->has_special_unique_user()) { 1157 _worklist.push(in->unique_out()); 1158 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1159 if( in->Opcode() == Op_Region ) 1160 _worklist.push(in); 1161 else if( in->is_Store() ) { 1162 DUIterator_Fast imax, i = in->fast_outs(imax); 1163 _worklist.push(in->fast_out(i)); 1164 i++; 1165 if(in->outcnt() == 2) { 1166 _worklist.push(in->fast_out(i)); 1167 i++; 1168 } 1169 assert(!(i < imax), "sanity"); 1170 } 1171 } 1172 } 1173 } 1174 1175 if (dead->is_macro()) { 1176 C->remove_macro_node(dead); 1177 } 1178 } 1179 // Aggressively kill globally dead uses 1180 // (Cannot use DUIterator_Last because of the indefinite number 1181 // of edge deletions per loop trip.) 1182 while (dead->outcnt() > 0) { 1183 remove_globally_dead_node(dead->raw_out(0)); 1184 } 1185 } 1186 1187 //------------------------------subsume_node----------------------------------- 1188 // Remove users from node 'old' and add them to node 'nn'. 1189 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1190 assert( old != hash_find(old), "should already been removed" ); 1191 assert( old != C->top(), "cannot subsume top node"); 1192 // Copy debug or profile information to the new version: 1193 C->copy_node_notes_to(nn, old); 1194 // Move users of node 'old' to node 'nn' 1195 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1196 Node* use = old->last_out(i); // for each use... 1197 // use might need re-hashing (but it won't if it's a new node) 1198 bool is_in_table = _table.hash_delete( use ); 1199 // Update use-def info as well 1200 // We remove all occurrences of old within use->in, 1201 // so as to avoid rehashing any node more than once. 1202 // The hash table probe swamps any outer loop overhead. 1203 uint num_edges = 0; 1204 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1205 if (use->in(j) == old) { 1206 use->set_req(j, nn); 1207 ++num_edges; 1208 } 1209 } 1210 // Insert into GVN hash table if unique 1211 // If a duplicate, 'use' will be cleaned up when pulled off worklist 1212 if( is_in_table ) { 1213 hash_find_insert(use); 1214 } 1215 i -= num_edges; // we deleted 1 or more copies of this edge 1216 } 1217 1218 // Smash all inputs to 'old', isolating him completely 1219 Node *temp = new (C, 1) Node(1); 1220 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1221 remove_dead_node( old ); 1222 temp->del_req(0); // Yank bogus edge 1223 #ifndef PRODUCT 1224 if( VerifyIterativeGVN ) { 1225 for ( int i = 0; i < _verify_window_size; i++ ) { 1226 if ( _verify_window[i] == old ) 1227 _verify_window[i] = nn; 1228 } 1229 } 1230 #endif 1231 _worklist.remove(temp); // this can be necessary 1232 temp->destruct(); // reuse the _idx of this little guy 1233 } 1234 1235 //------------------------------add_users_to_worklist-------------------------- 1236 void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1237 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1238 _worklist.push(n->fast_out(i)); // Push on worklist 1239 } 1240 } 1241 1242 void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1243 add_users_to_worklist0(n); 1244 1245 // Move users of node to worklist 1246 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1247 Node* use = n->fast_out(i); // Get use 1248 1249 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1250 use->is_Store() ) // Enable store/load same address 1251 add_users_to_worklist0(use); 1252 1253 // If we changed the receiver type to a call, we need to revisit 1254 // the Catch following the call. It's looking for a non-NULL 1255 // receiver to know when to enable the regular fall-through path 1256 // in addition to the NullPtrException path. 1257 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1258 Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control); 1259 if (p != NULL) { 1260 add_users_to_worklist0(p); 1261 } 1262 } 1263 1264 if( use->is_Cmp() ) { // Enable CMP/BOOL optimization 1265 add_users_to_worklist(use); // Put Bool on worklist 1266 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1267 // phi merging either 0 or 1 onto the worklist 1268 if (use->outcnt() > 0) { 1269 Node* bol = use->raw_out(0); 1270 if (bol->outcnt() > 0) { 1271 Node* iff = bol->raw_out(0); 1272 if (iff->outcnt() == 2) { 1273 Node* ifproj0 = iff->raw_out(0); 1274 Node* ifproj1 = iff->raw_out(1); 1275 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1276 Node* region0 = ifproj0->raw_out(0); 1277 Node* region1 = ifproj1->raw_out(0); 1278 if( region0 == region1 ) 1279 add_users_to_worklist0(region0); 1280 } 1281 } 1282 } 1283 } 1284 } 1285 1286 uint use_op = use->Opcode(); 1287 // If changed Cast input, check Phi users for simple cycles 1288 if( use->is_ConstraintCast() || use->is_CheckCastPP() ) { 1289 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1290 Node* u = use->fast_out(i2); 1291 if (u->is_Phi()) 1292 _worklist.push(u); 1293 } 1294 } 1295 // If changed LShift inputs, check RShift users for useless sign-ext 1296 if( use_op == Op_LShiftI ) { 1297 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1298 Node* u = use->fast_out(i2); 1299 if (u->Opcode() == Op_RShiftI) 1300 _worklist.push(u); 1301 } 1302 } 1303 // If changed AddP inputs, check Stores for loop invariant 1304 if( use_op == Op_AddP ) { 1305 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1306 Node* u = use->fast_out(i2); 1307 if (u->is_Mem()) 1308 _worklist.push(u); 1309 } 1310 } 1311 // If changed initialization activity, check dependent Stores 1312 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1313 InitializeNode* init = use->as_Allocate()->initialization(); 1314 if (init != NULL) { 1315 Node* imem = init->proj_out(TypeFunc::Memory); 1316 if (imem != NULL) add_users_to_worklist0(imem); 1317 } 1318 } 1319 if (use_op == Op_Initialize) { 1320 Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory); 1321 if (imem != NULL) add_users_to_worklist0(imem); 1322 } 1323 } 1324 } 1325 1326 //============================================================================= 1327 #ifndef PRODUCT 1328 uint PhaseCCP::_total_invokes = 0; 1329 uint PhaseCCP::_total_constants = 0; 1330 #endif 1331 //------------------------------PhaseCCP--------------------------------------- 1332 // Conditional Constant Propagation, ala Wegman & Zadeck 1333 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1334 NOT_PRODUCT( clear_constants(); ) 1335 assert( _worklist.size() == 0, "" ); 1336 // Clear out _nodes from IterGVN. Must be clear to transform call. 1337 _nodes.clear(); // Clear out from IterGVN 1338 analyze(); 1339 } 1340 1341 #ifndef PRODUCT 1342 //------------------------------~PhaseCCP-------------------------------------- 1343 PhaseCCP::~PhaseCCP() { 1344 inc_invokes(); 1345 _total_constants += count_constants(); 1346 } 1347 #endif 1348 1349 1350 #ifdef ASSERT 1351 static bool ccp_type_widens(const Type* t, const Type* t0) { 1352 assert(t->meet(t0) == t, "Not monotonic"); 1353 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1354 case Type::Int: 1355 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1356 break; 1357 case Type::Long: 1358 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1359 break; 1360 } 1361 return true; 1362 } 1363 #endif //ASSERT 1364 1365 //------------------------------analyze---------------------------------------- 1366 void PhaseCCP::analyze() { 1367 // Initialize all types to TOP, optimistic analysis 1368 for (int i = C->unique() - 1; i >= 0; i--) { 1369 _types.map(i,Type::TOP); 1370 } 1371 1372 // Push root onto worklist 1373 Unique_Node_List worklist; 1374 worklist.push(C->root()); 1375 1376 // Pull from worklist; compute new value; push changes out. 1377 // This loop is the meat of CCP. 1378 while( worklist.size() ) { 1379 Node *n = worklist.pop(); 1380 const Type *t = n->Value(this); 1381 if (t != type(n)) { 1382 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1383 #ifndef PRODUCT 1384 if( TracePhaseCCP ) { 1385 t->dump(); 1386 do { tty->print("\t"); } while (tty->position() < 16); 1387 n->dump(); 1388 } 1389 #endif 1390 set_type(n, t); 1391 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1392 Node* m = n->fast_out(i); // Get user 1393 if( m->is_Region() ) { // New path to Region? Must recheck Phis too 1394 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1395 Node* p = m->fast_out(i2); // Propagate changes to uses 1396 if( p->bottom_type() != type(p) ) // If not already bottomed out 1397 worklist.push(p); // Propagate change to user 1398 } 1399 } 1400 // If we changed the receiver type to a call, we need to revisit 1401 // the Catch following the call. It's looking for a non-NULL 1402 // receiver to know when to enable the regular fall-through path 1403 // in addition to the NullPtrException path 1404 if (m->is_Call()) { 1405 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1406 Node* p = m->fast_out(i2); // Propagate changes to uses 1407 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) 1408 worklist.push(p->unique_out()); 1409 } 1410 } 1411 if( m->bottom_type() != type(m) ) // If not already bottomed out 1412 worklist.push(m); // Propagate change to user 1413 } 1414 } 1415 } 1416 } 1417 1418 //------------------------------do_transform----------------------------------- 1419 // Top level driver for the recursive transformer 1420 void PhaseCCP::do_transform() { 1421 // Correct leaves of new-space Nodes; they point to old-space. 1422 C->set_root( transform(C->root())->as_Root() ); 1423 assert( C->top(), "missing TOP node" ); 1424 assert( C->root(), "missing root" ); 1425 } 1426 1427 //------------------------------transform-------------------------------------- 1428 // Given a Node in old-space, clone him into new-space. 1429 // Convert any of his old-space children into new-space children. 1430 Node *PhaseCCP::transform( Node *n ) { 1431 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1432 if( new_node != NULL ) 1433 return new_node; // Been there, done that, return old answer 1434 new_node = transform_once(n); // Check for constant 1435 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1436 1437 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1438 GrowableArray <Node *> trstack(C->unique() >> 1); 1439 1440 trstack.push(new_node); // Process children of cloned node 1441 while ( trstack.is_nonempty() ) { 1442 Node *clone = trstack.pop(); 1443 uint cnt = clone->req(); 1444 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1445 Node *input = clone->in(i); 1446 if( input != NULL ) { // Ignore NULLs 1447 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1448 if( new_input == NULL ) { 1449 new_input = transform_once(input); // Check for constant 1450 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1451 trstack.push(new_input); 1452 } 1453 assert( new_input == clone->in(i), "insanity check"); 1454 } 1455 } 1456 } 1457 return new_node; 1458 } 1459 1460 1461 //------------------------------transform_once--------------------------------- 1462 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1463 Node *PhaseCCP::transform_once( Node *n ) { 1464 const Type *t = type(n); 1465 // Constant? Use constant Node instead 1466 if( t->singleton() ) { 1467 Node *nn = n; // Default is to return the original constant 1468 if( t == Type::TOP ) { 1469 // cache my top node on the Compile instance 1470 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1471 C->set_cached_top_node( ConNode::make(C, Type::TOP) ); 1472 set_type(C->top(), Type::TOP); 1473 } 1474 nn = C->top(); 1475 } 1476 if( !n->is_Con() ) { 1477 if( t != Type::TOP ) { 1478 nn = makecon(t); // ConNode::make(t); 1479 NOT_PRODUCT( inc_constants(); ) 1480 } else if( n->is_Region() ) { // Unreachable region 1481 // Note: nn == C->top() 1482 n->set_req(0, NULL); // Cut selfreference 1483 // Eagerly remove dead phis to avoid phis copies creation. 1484 for (DUIterator i = n->outs(); n->has_out(i); i++) { 1485 Node* m = n->out(i); 1486 if( m->is_Phi() ) { 1487 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1488 replace_node(m, nn); 1489 --i; // deleted this phi; rescan starting with next position 1490 } 1491 } 1492 } 1493 replace_node(n,nn); // Update DefUse edges for new constant 1494 } 1495 return nn; 1496 } 1497 1498 // If x is a TypeNode, capture any more-precise type permanently into Node 1499 if (t != n->bottom_type()) { 1500 hash_delete(n); // changing bottom type may force a rehash 1501 n->raise_bottom_type(t); 1502 _worklist.push(n); // n re-enters the hash table via the worklist 1503 } 1504 1505 // Idealize graph using DU info. Must clone() into new-space. 1506 // DU info is generally used to show profitability, progress or safety 1507 // (but generally not needed for correctness). 1508 Node *nn = n->Ideal_DU_postCCP(this); 1509 1510 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1511 switch( n->Opcode() ) { 1512 case Op_FastLock: // Revisit FastLocks for lock coarsening 1513 case Op_If: 1514 case Op_CountedLoopEnd: 1515 case Op_Region: 1516 case Op_Loop: 1517 case Op_CountedLoop: 1518 case Op_Conv2B: 1519 case Op_Opaque1: 1520 case Op_Opaque2: 1521 _worklist.push(n); 1522 break; 1523 default: 1524 break; 1525 } 1526 if( nn ) { 1527 _worklist.push(n); 1528 // Put users of 'n' onto worklist for second igvn transform 1529 add_users_to_worklist(n); 1530 return nn; 1531 } 1532 1533 return n; 1534 } 1535 1536 //---------------------------------saturate------------------------------------ 1537 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1538 const Type* limit_type) const { 1539 const Type* wide_type = new_type->widen(old_type, limit_type); 1540 if (wide_type != new_type) { // did we widen? 1541 // If so, we may have widened beyond the limit type. Clip it back down. 1542 new_type = wide_type->filter(limit_type); 1543 } 1544 return new_type; 1545 } 1546 1547 //------------------------------print_statistics------------------------------- 1548 #ifndef PRODUCT 1549 void PhaseCCP::print_statistics() { 1550 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1551 } 1552 #endif 1553 1554 1555 //============================================================================= 1556 #ifndef PRODUCT 1557 uint PhasePeephole::_total_peepholes = 0; 1558 #endif 1559 //------------------------------PhasePeephole---------------------------------- 1560 // Conditional Constant Propagation, ala Wegman & Zadeck 1561 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1562 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1563 NOT_PRODUCT( clear_peepholes(); ) 1564 } 1565 1566 #ifndef PRODUCT 1567 //------------------------------~PhasePeephole--------------------------------- 1568 PhasePeephole::~PhasePeephole() { 1569 _total_peepholes += count_peepholes(); 1570 } 1571 #endif 1572 1573 //------------------------------transform-------------------------------------- 1574 Node *PhasePeephole::transform( Node *n ) { 1575 ShouldNotCallThis(); 1576 return NULL; 1577 } 1578 1579 //------------------------------do_transform----------------------------------- 1580 void PhasePeephole::do_transform() { 1581 bool method_name_not_printed = true; 1582 1583 // Examine each basic block 1584 for( uint block_number = 1; block_number < _cfg._num_blocks; ++block_number ) { 1585 Block *block = _cfg._blocks[block_number]; 1586 bool block_not_printed = true; 1587 1588 // and each instruction within a block 1589 uint end_index = block->_nodes.size(); 1590 // block->end_idx() not valid after PhaseRegAlloc 1591 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 1592 Node *n = block->_nodes.at(instruction_index); 1593 if( n->is_Mach() ) { 1594 MachNode *m = n->as_Mach(); 1595 int deleted_count = 0; 1596 // check for peephole opportunities 1597 MachNode *m2 = m->peephole( block, instruction_index, _regalloc, deleted_count, C ); 1598 if( m2 != NULL ) { 1599 #ifndef PRODUCT 1600 if( PrintOptoPeephole ) { 1601 // Print method, first time only 1602 if( C->method() && method_name_not_printed ) { 1603 C->method()->print_short_name(); tty->cr(); 1604 method_name_not_printed = false; 1605 } 1606 // Print this block 1607 if( Verbose && block_not_printed) { 1608 tty->print_cr("in block"); 1609 block->dump(); 1610 block_not_printed = false; 1611 } 1612 // Print instructions being deleted 1613 for( int i = (deleted_count - 1); i >= 0; --i ) { 1614 block->_nodes.at(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 1615 } 1616 tty->print_cr("replaced with"); 1617 // Print new instruction 1618 m2->format(_regalloc); 1619 tty->print("\n\n"); 1620 } 1621 #endif 1622 // Remove old nodes from basic block and update instruction_index 1623 // (old nodes still exist and may have edges pointing to them 1624 // as register allocation info is stored in the allocator using 1625 // the node index to live range mappings.) 1626 uint safe_instruction_index = (instruction_index - deleted_count); 1627 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 1628 block->_nodes.remove( instruction_index ); 1629 } 1630 // install new node after safe_instruction_index 1631 block->_nodes.insert( safe_instruction_index + 1, m2 ); 1632 end_index = block->_nodes.size() - 1; // Recompute new block size 1633 NOT_PRODUCT( inc_peepholes(); ) 1634 } 1635 } 1636 } 1637 } 1638 } 1639 1640 //------------------------------print_statistics------------------------------- 1641 #ifndef PRODUCT 1642 void PhasePeephole::print_statistics() { 1643 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 1644 } 1645 #endif 1646 1647 1648 //============================================================================= 1649 //------------------------------set_req_X-------------------------------------- 1650 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 1651 assert( is_not_dead(n), "can not use dead node"); 1652 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 1653 Node *old = in(i); 1654 set_req(i, n); 1655 1656 // old goes dead? 1657 if( old ) { 1658 switch (old->outcnt()) { 1659 case 0: 1660 // Put into the worklist to kill later. We do not kill it now because the 1661 // recursive kill will delete the current node (this) if dead-loop exists 1662 if (!old->is_top()) 1663 igvn->_worklist.push( old ); 1664 break; 1665 case 1: 1666 if( old->is_Store() || old->has_special_unique_user() ) 1667 igvn->add_users_to_worklist( old ); 1668 break; 1669 case 2: 1670 if( old->is_Store() ) 1671 igvn->add_users_to_worklist( old ); 1672 if( old->Opcode() == Op_Region ) 1673 igvn->_worklist.push(old); 1674 break; 1675 case 3: 1676 if( old->Opcode() == Op_Region ) { 1677 igvn->_worklist.push(old); 1678 igvn->add_users_to_worklist( old ); 1679 } 1680 break; 1681 default: 1682 break; 1683 } 1684 } 1685 1686 } 1687 1688 //-------------------------------replace_by----------------------------------- 1689 // Using def-use info, replace one node for another. Follow the def-use info 1690 // to all users of the OLD node. Then make all uses point to the NEW node. 1691 void Node::replace_by(Node *new_node) { 1692 assert(!is_top(), "top node has no DU info"); 1693 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 1694 Node* use = last_out(i); 1695 uint uses_found = 0; 1696 for (uint j = 0; j < use->len(); j++) { 1697 if (use->in(j) == this) { 1698 if (j < use->req()) 1699 use->set_req(j, new_node); 1700 else use->set_prec(j, new_node); 1701 uses_found++; 1702 } 1703 } 1704 i -= uses_found; // we deleted 1 or more copies of this edge 1705 } 1706 } 1707 1708 //============================================================================= 1709 //----------------------------------------------------------------------------- 1710 void Type_Array::grow( uint i ) { 1711 if( !_max ) { 1712 _max = 1; 1713 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 1714 _types[0] = NULL; 1715 } 1716 uint old = _max; 1717 while( i >= _max ) _max <<= 1; // Double to fit 1718 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 1719 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 1720 } 1721 1722 //------------------------------dump------------------------------------------- 1723 #ifndef PRODUCT 1724 void Type_Array::dump() const { 1725 uint max = Size(); 1726 for( uint i = 0; i < max; i++ ) { 1727 if( _types[i] != NULL ) { 1728 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 1729 } 1730 } 1731 } 1732 #endif