phaseX.cpp revision 5776:de6a9e811145
1/* 2 * Copyright (c) 1997, 2013, 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--------------------------------------- 42NodeHash::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--------------------------------------- 56NodeHash::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--------------------------------------- 71NodeHash::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 78void 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 87Node *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 132Node *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 191void 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 218bool 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 260uint 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 270void 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 295void 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 311void 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#ifndef PRODUCT 327//------------------------------dump------------------------------------------- 328// Dump statistics for the hash table 329void NodeHash::dump() { 330 _total_inserts += _inserts; 331 _total_insert_probes += _insert_probes; 332 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) { 333 if (WizardMode) { 334 for (uint i=0; i<_max; i++) { 335 if (_table[i]) 336 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx); 337 } 338 } 339 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max); 340 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0); 341 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses)); 342 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts); 343 // sentinels increase lookup cost, but not insert cost 344 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function"); 345 assert( _inserts+(_inserts>>3) < _max, "table too full" ); 346 assert( _inserts*3+100 >= _insert_probes, "bad hash function" ); 347 } 348} 349 350Node *NodeHash::find_index(uint idx) { // For debugging 351 // Find an entry by its index value 352 for( uint i = 0; i < _max; i++ ) { 353 Node *m = _table[i]; 354 if( !m || m == _sentinel ) continue; 355 if( m->_idx == (uint)idx ) return m; 356 } 357 return NULL; 358} 359#endif 360 361#ifdef ASSERT 362NodeHash::~NodeHash() { 363 // Unlock all nodes upon destruction of table. 364 if (_table != (Node**)badAddress) clear(); 365} 366 367void NodeHash::operator=(const NodeHash& nh) { 368 // Unlock all nodes upon replacement of table. 369 if (&nh == this) return; 370 if (_table != (Node**)badAddress) clear(); 371 memcpy(this, &nh, sizeof(*this)); 372 // Do not increment hash_lock counts again. 373 // Instead, be sure we never again use the source table. 374 ((NodeHash*)&nh)->_table = (Node**)badAddress; 375} 376 377 378#endif 379 380 381//============================================================================= 382//------------------------------PhaseRemoveUseless----------------------------- 383// 1) Use a breadthfirst walk to collect useful nodes reachable from root. 384PhaseRemoveUseless::PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist ) : Phase(Remove_Useless), 385 _useful(Thread::current()->resource_area()) { 386 387 // Implementation requires 'UseLoopSafepoints == true' and an edge from root 388 // to each SafePointNode at a backward branch. Inserted in add_safepoint(). 389 if( !UseLoopSafepoints || !OptoRemoveUseless ) return; 390 391 // Identify nodes that are reachable from below, useful. 392 C->identify_useful_nodes(_useful); 393 // Update dead node list 394 C->update_dead_node_list(_useful); 395 396 // Remove all useless nodes from PhaseValues' recorded types 397 // Must be done before disconnecting nodes to preserve hash-table-invariant 398 gvn->remove_useless_nodes(_useful.member_set()); 399 400 // Remove all useless nodes from future worklist 401 worklist->remove_useless_nodes(_useful.member_set()); 402 403 // Disconnect 'useless' nodes that are adjacent to useful nodes 404 C->remove_useless_nodes(_useful); 405 406 // Remove edges from "root" to each SafePoint at a backward branch. 407 // They were inserted during parsing (see add_safepoint()) to make infinite 408 // loops without calls or exceptions visible to root, i.e., useful. 409 Node *root = C->root(); 410 if( root != NULL ) { 411 for( uint i = root->req(); i < root->len(); ++i ) { 412 Node *n = root->in(i); 413 if( n != NULL && n->is_SafePoint() ) { 414 root->rm_prec(i); 415 --i; 416 } 417 } 418 } 419} 420 421 422//============================================================================= 423//------------------------------PhaseTransform--------------------------------- 424PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum), 425 _arena(Thread::current()->resource_area()), 426 _nodes(_arena), 427 _types(_arena) 428{ 429 init_con_caches(); 430#ifndef PRODUCT 431 clear_progress(); 432 clear_transforms(); 433 set_allow_progress(true); 434#endif 435 // Force allocation for currently existing nodes 436 _types.map(C->unique(), NULL); 437} 438 439//------------------------------PhaseTransform--------------------------------- 440PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum), 441 _arena(arena), 442 _nodes(arena), 443 _types(arena) 444{ 445 init_con_caches(); 446#ifndef PRODUCT 447 clear_progress(); 448 clear_transforms(); 449 set_allow_progress(true); 450#endif 451 // Force allocation for currently existing nodes 452 _types.map(C->unique(), NULL); 453} 454 455//------------------------------PhaseTransform--------------------------------- 456// Initialize with previously generated type information 457PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum), 458 _arena(pt->_arena), 459 _nodes(pt->_nodes), 460 _types(pt->_types) 461{ 462 init_con_caches(); 463#ifndef PRODUCT 464 clear_progress(); 465 clear_transforms(); 466 set_allow_progress(true); 467#endif 468} 469 470void PhaseTransform::init_con_caches() { 471 memset(_icons,0,sizeof(_icons)); 472 memset(_lcons,0,sizeof(_lcons)); 473 memset(_zcons,0,sizeof(_zcons)); 474} 475 476 477//--------------------------------find_int_type-------------------------------- 478const TypeInt* PhaseTransform::find_int_type(Node* n) { 479 if (n == NULL) return NULL; 480 // Call type_or_null(n) to determine node's type since we might be in 481 // parse phase and call n->Value() may return wrong type. 482 // (For example, a phi node at the beginning of loop parsing is not ready.) 483 const Type* t = type_or_null(n); 484 if (t == NULL) return NULL; 485 return t->isa_int(); 486} 487 488 489//-------------------------------find_long_type-------------------------------- 490const TypeLong* PhaseTransform::find_long_type(Node* n) { 491 if (n == NULL) return NULL; 492 // (See comment above on type_or_null.) 493 const Type* t = type_or_null(n); 494 if (t == NULL) return NULL; 495 return t->isa_long(); 496} 497 498 499#ifndef PRODUCT 500void PhaseTransform::dump_old2new_map() const { 501 _nodes.dump(); 502} 503 504void PhaseTransform::dump_new( uint nidx ) const { 505 for( uint i=0; i<_nodes.Size(); i++ ) 506 if( _nodes[i] && _nodes[i]->_idx == nidx ) { 507 _nodes[i]->dump(); 508 tty->cr(); 509 tty->print_cr("Old index= %d",i); 510 return; 511 } 512 tty->print_cr("Node %d not found in the new indices", nidx); 513} 514 515//------------------------------dump_types------------------------------------- 516void PhaseTransform::dump_types( ) const { 517 _types.dump(); 518} 519 520//------------------------------dump_nodes_and_types--------------------------- 521void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) { 522 VectorSet visited(Thread::current()->resource_area()); 523 dump_nodes_and_types_recur( root, depth, only_ctrl, visited ); 524} 525 526//------------------------------dump_nodes_and_types_recur--------------------- 527void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) { 528 if( !n ) return; 529 if( depth == 0 ) return; 530 if( visited.test_set(n->_idx) ) return; 531 for( uint i=0; i<n->len(); i++ ) { 532 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue; 533 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited ); 534 } 535 n->dump(); 536 if (type_or_null(n) != NULL) { 537 tty->print(" "); type(n)->dump(); tty->cr(); 538 } 539} 540 541#endif 542 543 544//============================================================================= 545//------------------------------PhaseValues------------------------------------ 546// Set minimum table size to "255" 547PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) { 548 NOT_PRODUCT( clear_new_values(); ) 549} 550 551//------------------------------PhaseValues------------------------------------ 552// Set minimum table size to "255" 553PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ), 554 _table(&ptv->_table) { 555 NOT_PRODUCT( clear_new_values(); ) 556} 557 558//------------------------------PhaseValues------------------------------------ 559// Used by +VerifyOpto. Clear out hash table but copy _types array. 560PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ), 561 _table(ptv->arena(),ptv->_table.size()) { 562 NOT_PRODUCT( clear_new_values(); ) 563} 564 565//------------------------------~PhaseValues----------------------------------- 566#ifndef PRODUCT 567PhaseValues::~PhaseValues() { 568 _table.dump(); 569 570 // Statistics for value progress and efficiency 571 if( PrintCompilation && Verbose && WizardMode ) { 572 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 573 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 574 if( made_transforms() != 0 ) { 575 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 576 } else { 577 tty->cr(); 578 } 579 } 580} 581#endif 582 583//------------------------------makecon---------------------------------------- 584ConNode* PhaseTransform::makecon(const Type *t) { 585 assert(t->singleton(), "must be a constant"); 586 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 587 switch (t->base()) { // fast paths 588 case Type::Half: 589 case Type::Top: return (ConNode*) C->top(); 590 case Type::Int: return intcon( t->is_int()->get_con() ); 591 case Type::Long: return longcon( t->is_long()->get_con() ); 592 } 593 if (t->is_zero_type()) 594 return zerocon(t->basic_type()); 595 return uncached_makecon(t); 596} 597 598//--------------------------uncached_makecon----------------------------------- 599// Make an idealized constant - one of ConINode, ConPNode, etc. 600ConNode* PhaseValues::uncached_makecon(const Type *t) { 601 assert(t->singleton(), "must be a constant"); 602 ConNode* x = ConNode::make(C, t); 603 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 604 if (k == NULL) { 605 set_type(x, t); // Missed, provide type mapping 606 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 607 if (nna != NULL) { 608 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 609 loc->clear(); // do not put debug info on constants 610 } 611 } else { 612 x->destruct(); // Hit, destroy duplicate constant 613 x = k; // use existing constant 614 } 615 return x; 616} 617 618//------------------------------intcon----------------------------------------- 619// Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 620ConINode* PhaseTransform::intcon(int i) { 621 // Small integer? Check cache! Check that cached node is not dead 622 if (i >= _icon_min && i <= _icon_max) { 623 ConINode* icon = _icons[i-_icon_min]; 624 if (icon != NULL && icon->in(TypeFunc::Control) != NULL) 625 return icon; 626 } 627 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 628 assert(icon->is_Con(), ""); 629 if (i >= _icon_min && i <= _icon_max) 630 _icons[i-_icon_min] = icon; // Cache small integers 631 return icon; 632} 633 634//------------------------------longcon---------------------------------------- 635// Fast long constant. 636ConLNode* PhaseTransform::longcon(jlong l) { 637 // Small integer? Check cache! Check that cached node is not dead 638 if (l >= _lcon_min && l <= _lcon_max) { 639 ConLNode* lcon = _lcons[l-_lcon_min]; 640 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL) 641 return lcon; 642 } 643 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 644 assert(lcon->is_Con(), ""); 645 if (l >= _lcon_min && l <= _lcon_max) 646 _lcons[l-_lcon_min] = lcon; // Cache small integers 647 return lcon; 648} 649 650//------------------------------zerocon----------------------------------------- 651// Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 652ConNode* PhaseTransform::zerocon(BasicType bt) { 653 assert((uint)bt <= _zcon_max, "domain check"); 654 ConNode* zcon = _zcons[bt]; 655 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL) 656 return zcon; 657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 658 _zcons[bt] = zcon; 659 return zcon; 660} 661 662 663 664//============================================================================= 665//------------------------------transform-------------------------------------- 666// Return a node which computes the same function as this node, but in a 667// faster or cheaper fashion. 668Node *PhaseGVN::transform( Node *n ) { 669 return transform_no_reclaim(n); 670} 671 672//------------------------------transform-------------------------------------- 673// Return a node which computes the same function as this node, but 674// in a faster or cheaper fashion. 675Node *PhaseGVN::transform_no_reclaim( Node *n ) { 676 NOT_PRODUCT( set_transforms(); ) 677 678 // Apply the Ideal call in a loop until it no longer applies 679 Node *k = n; 680 NOT_PRODUCT( uint loop_count = 0; ) 681 while( 1 ) { 682 Node *i = k->Ideal(this, /*can_reshape=*/false); 683 if( !i ) break; 684 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 685 k = i; 686 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform"); 687 } 688 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } ) 689 690 691 // If brand new node, make space in type array. 692 ensure_type_or_null(k); 693 694 // Since I just called 'Value' to compute the set of run-time values 695 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 696 // cache Value. Later requests for the local phase->type of this Node can 697 // use the cached Value instead of suffering with 'bottom_type'. 698 const Type *t = k->Value(this); // Get runtime Value set 699 assert(t != NULL, "value sanity"); 700 if (type_or_null(k) != t) { 701#ifndef PRODUCT 702 // Do not count initial visit to node as a transformation 703 if (type_or_null(k) == NULL) { 704 inc_new_values(); 705 set_progress(); 706 } 707#endif 708 set_type(k, t); 709 // If k is a TypeNode, capture any more-precise type permanently into Node 710 k->raise_bottom_type(t); 711 } 712 713 if( t->singleton() && !k->is_Con() ) { 714 NOT_PRODUCT( set_progress(); ) 715 return makecon(t); // Turn into a constant 716 } 717 718 // Now check for Identities 719 Node *i = k->Identity(this); // Look for a nearby replacement 720 if( i != k ) { // Found? Return replacement! 721 NOT_PRODUCT( set_progress(); ) 722 return i; 723 } 724 725 // Global Value Numbering 726 i = hash_find_insert(k); // Insert if new 727 if( i && (i != k) ) { 728 // Return the pre-existing node 729 NOT_PRODUCT( set_progress(); ) 730 return i; 731 } 732 733 // Return Idealized original 734 return k; 735} 736 737#ifdef ASSERT 738//------------------------------dead_loop_check-------------------------------- 739// Check for a simple dead loop when a data node references itself directly 740// or through an other data node excluding cons and phis. 741void PhaseGVN::dead_loop_check( Node *n ) { 742 // Phi may reference itself in a loop 743 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) { 744 // Do 2 levels check and only data inputs. 745 bool no_dead_loop = true; 746 uint cnt = n->req(); 747 for (uint i = 1; i < cnt && no_dead_loop; i++) { 748 Node *in = n->in(i); 749 if (in == n) { 750 no_dead_loop = false; 751 } else if (in != NULL && !in->is_dead_loop_safe()) { 752 uint icnt = in->req(); 753 for (uint j = 1; j < icnt && no_dead_loop; j++) { 754 if (in->in(j) == n || in->in(j) == in) 755 no_dead_loop = false; 756 } 757 } 758 } 759 if (!no_dead_loop) n->dump(3); 760 assert(no_dead_loop, "dead loop detected"); 761 } 762} 763#endif 764 765//============================================================================= 766//------------------------------PhaseIterGVN----------------------------------- 767// Initialize hash table to fresh and clean for +VerifyOpto 768PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ), 769 _stack(C->unique() >> 1), 770 _delay_transform(false) { 771} 772 773//------------------------------PhaseIterGVN----------------------------------- 774// Initialize with previous PhaseIterGVN info; used by PhaseCCP 775PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn), 776 _worklist( igvn->_worklist ), 777 _stack( igvn->_stack ), 778 _delay_transform(igvn->_delay_transform) 779{ 780} 781 782//------------------------------PhaseIterGVN----------------------------------- 783// Initialize with previous PhaseGVN info from Parser 784PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn), 785 _worklist(*C->for_igvn()), 786 _stack(C->unique() >> 1), 787 _delay_transform(false) 788{ 789 uint max; 790 791 // Dead nodes in the hash table inherited from GVN were not treated as 792 // roots during def-use info creation; hence they represent an invisible 793 // use. Clear them out. 794 max = _table.size(); 795 for( uint i = 0; i < max; ++i ) { 796 Node *n = _table.at(i); 797 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) { 798 if( n->is_top() ) continue; 799 assert( false, "Parse::remove_useless_nodes missed this node"); 800 hash_delete(n); 801 } 802 } 803 804 // Any Phis or Regions on the worklist probably had uses that could not 805 // make more progress because the uses were made while the Phis and Regions 806 // were in half-built states. Put all uses of Phis and Regions on worklist. 807 max = _worklist.size(); 808 for( uint j = 0; j < max; j++ ) { 809 Node *n = _worklist.at(j); 810 uint uop = n->Opcode(); 811 if( uop == Op_Phi || uop == Op_Region || 812 n->is_Type() || 813 n->is_Mem() ) 814 add_users_to_worklist(n); 815 } 816} 817 818 819#ifndef PRODUCT 820void PhaseIterGVN::verify_step(Node* n) { 821 _verify_window[_verify_counter % _verify_window_size] = n; 822 ++_verify_counter; 823 ResourceMark rm; 824 ResourceArea *area = Thread::current()->resource_area(); 825 VectorSet old_space(area), new_space(area); 826 if (C->unique() < 1000 || 827 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 828 ++_verify_full_passes; 829 Node::verify_recur(C->root(), -1, old_space, new_space); 830 } 831 const int verify_depth = 4; 832 for ( int i = 0; i < _verify_window_size; i++ ) { 833 Node* n = _verify_window[i]; 834 if ( n == NULL ) continue; 835 if( n->in(0) == NodeSentinel ) { // xform_idom 836 _verify_window[i] = n->in(1); 837 --i; continue; 838 } 839 // Typical fanout is 1-2, so this call visits about 6 nodes. 840 Node::verify_recur(n, verify_depth, old_space, new_space); 841 } 842} 843#endif 844 845 846//------------------------------init_worklist---------------------------------- 847// Initialize worklist for each node. 848void PhaseIterGVN::init_worklist( Node *n ) { 849 if( _worklist.member(n) ) return; 850 _worklist.push(n); 851 uint cnt = n->req(); 852 for( uint i =0 ; i < cnt; i++ ) { 853 Node *m = n->in(i); 854 if( m ) init_worklist(m); 855 } 856} 857 858//------------------------------optimize--------------------------------------- 859void PhaseIterGVN::optimize() { 860 debug_only(uint num_processed = 0;); 861#ifndef PRODUCT 862 { 863 _verify_counter = 0; 864 _verify_full_passes = 0; 865 for ( int i = 0; i < _verify_window_size; i++ ) { 866 _verify_window[i] = NULL; 867 } 868 } 869#endif 870 871#ifdef ASSERT 872 Node* prev = NULL; 873 uint rep_cnt = 0; 874#endif 875 uint loop_count = 0; 876 877 // Pull from worklist; transform node; 878 // If node has changed: update edge info and put uses on worklist. 879 while( _worklist.size() ) { 880 if (C->check_node_count(NodeLimitFudgeFactor * 2, 881 "out of nodes optimizing method")) { 882 return; 883 } 884 Node *n = _worklist.pop(); 885 if (++loop_count >= K * C->live_nodes()) { 886 debug_only(n->dump(4);) 887 assert(false, "infinite loop in PhaseIterGVN::optimize"); 888 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 889 return; 890 } 891#ifdef ASSERT 892 if (n == prev) { 893 if (++rep_cnt > 3) { 894 n->dump(4); 895 assert(false, "loop in Ideal transformation"); 896 } 897 } else { 898 rep_cnt = 0; 899 } 900 prev = n; 901#endif 902 if (TraceIterativeGVN && Verbose) { 903 tty->print(" Pop "); 904 NOT_PRODUCT( n->dump(); ) 905 debug_only(if( (num_processed++ % 100) == 0 ) _worklist.print_set();) 906 } 907 908 if (n->outcnt() != 0) { 909 910#ifndef PRODUCT 911 uint wlsize = _worklist.size(); 912 const Type* oldtype = type_or_null(n); 913#endif //PRODUCT 914 915 Node *nn = transform_old(n); 916 917#ifndef PRODUCT 918 if (TraceIterativeGVN) { 919 const Type* newtype = type_or_null(n); 920 if (nn != n) { 921 // print old node 922 tty->print("< "); 923 if (oldtype != newtype && oldtype != NULL) { 924 oldtype->dump(); 925 } 926 do { tty->print("\t"); } while (tty->position() < 16); 927 tty->print("<"); 928 n->dump(); 929 } 930 if (oldtype != newtype || nn != n) { 931 // print new node and/or new type 932 if (oldtype == NULL) { 933 tty->print("* "); 934 } else if (nn != n) { 935 tty->print("> "); 936 } else { 937 tty->print("= "); 938 } 939 if (newtype == NULL) { 940 tty->print("null"); 941 } else { 942 newtype->dump(); 943 } 944 do { tty->print("\t"); } while (tty->position() < 16); 945 nn->dump(); 946 } 947 if (Verbose && wlsize < _worklist.size()) { 948 tty->print(" Push {"); 949 while (wlsize != _worklist.size()) { 950 Node* pushed = _worklist.at(wlsize++); 951 tty->print(" %d", pushed->_idx); 952 } 953 tty->print_cr(" }"); 954 } 955 } 956 if( VerifyIterativeGVN && nn != n ) { 957 verify_step((Node*) NULL); // ignore n, it might be subsumed 958 } 959#endif 960 } else if (!n->is_top()) { 961 remove_dead_node(n); 962 } 963 } 964 965#ifndef PRODUCT 966 C->verify_graph_edges(); 967 if( VerifyOpto && allow_progress() ) { 968 // Must turn off allow_progress to enable assert and break recursion 969 C->root()->verify(); 970 { // Check if any progress was missed using IterGVN 971 // Def-Use info enables transformations not attempted in wash-pass 972 // e.g. Region/Phi cleanup, ... 973 // Null-check elision -- may not have reached fixpoint 974 // do not propagate to dominated nodes 975 ResourceMark rm; 976 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean! 977 // Fill worklist completely 978 igvn2.init_worklist(C->root()); 979 980 igvn2.set_allow_progress(false); 981 igvn2.optimize(); 982 igvn2.set_allow_progress(true); 983 } 984 } 985 if ( VerifyIterativeGVN && PrintOpto ) { 986 if ( _verify_counter == _verify_full_passes ) 987 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 988 _verify_full_passes); 989 else 990 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 991 _verify_counter, _verify_full_passes); 992 } 993#endif 994} 995 996 997//------------------register_new_node_with_optimizer--------------------------- 998// Register a new node with the optimizer. Update the types array, the def-use 999// info. Put on worklist. 1000Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1001 set_type_bottom(n); 1002 _worklist.push(n); 1003 if (orig != NULL) C->copy_node_notes_to(n, orig); 1004 return n; 1005} 1006 1007//------------------------------transform-------------------------------------- 1008// Non-recursive: idealize Node 'n' with respect to its inputs and its value 1009Node *PhaseIterGVN::transform( Node *n ) { 1010 if (_delay_transform) { 1011 // Register the node but don't optimize for now 1012 register_new_node_with_optimizer(n); 1013 return n; 1014 } 1015 1016 // If brand new node, make space in type array, and give it a type. 1017 ensure_type_or_null(n); 1018 if (type_or_null(n) == NULL) { 1019 set_type_bottom(n); 1020 } 1021 1022 return transform_old(n); 1023} 1024 1025//------------------------------transform_old---------------------------------- 1026Node *PhaseIterGVN::transform_old( Node *n ) { 1027#ifndef PRODUCT 1028 debug_only(uint loop_count = 0;); 1029 set_transforms(); 1030#endif 1031 // Remove 'n' from hash table in case it gets modified 1032 _table.hash_delete(n); 1033 if( VerifyIterativeGVN ) { 1034 assert( !_table.find_index(n->_idx), "found duplicate entry in table"); 1035 } 1036 1037 // Apply the Ideal call in a loop until it no longer applies 1038 Node *k = n; 1039 DEBUG_ONLY(dead_loop_check(k);) 1040 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1041 Node *i = k->Ideal(this, /*can_reshape=*/true); 1042 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1043#ifndef PRODUCT 1044 if( VerifyIterativeGVN ) 1045 verify_step(k); 1046 if( i && VerifyOpto ) { 1047 if( !allow_progress() ) { 1048 if (i->is_Add() && i->outcnt() == 1) { 1049 // Switched input to left side because this is the only use 1050 } else if( i->is_If() && (i->in(0) == NULL) ) { 1051 // This IF is dead because it is dominated by an equivalent IF When 1052 // dominating if changed, info is not propagated sparsely to 'this' 1053 // Propagating this info further will spuriously identify other 1054 // progress. 1055 return i; 1056 } else 1057 set_progress(); 1058 } else 1059 set_progress(); 1060 } 1061#endif 1062 1063 while( i ) { 1064#ifndef PRODUCT 1065 debug_only( if( loop_count >= K ) i->dump(4); ) 1066 assert(loop_count < K, "infinite loop in PhaseIterGVN::transform"); 1067 debug_only( loop_count++; ) 1068#endif 1069 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1070 // Made a change; put users of original Node on worklist 1071 add_users_to_worklist( k ); 1072 // Replacing root of transform tree? 1073 if( k != i ) { 1074 // Make users of old Node now use new. 1075 subsume_node( k, i ); 1076 k = i; 1077 } 1078 DEBUG_ONLY(dead_loop_check(k);) 1079 // Try idealizing again 1080 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1081 i = k->Ideal(this, /*can_reshape=*/true); 1082 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1083#ifndef PRODUCT 1084 if( VerifyIterativeGVN ) 1085 verify_step(k); 1086 if( i && VerifyOpto ) set_progress(); 1087#endif 1088 } 1089 1090 // If brand new node, make space in type array. 1091 ensure_type_or_null(k); 1092 1093 // See what kind of values 'k' takes on at runtime 1094 const Type *t = k->Value(this); 1095 assert(t != NULL, "value sanity"); 1096 1097 // Since I just called 'Value' to compute the set of run-time values 1098 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1099 // cache Value. Later requests for the local phase->type of this Node can 1100 // use the cached Value instead of suffering with 'bottom_type'. 1101 if (t != type_or_null(k)) { 1102 NOT_PRODUCT( set_progress(); ) 1103 NOT_PRODUCT( inc_new_values();) 1104 set_type(k, t); 1105 // If k is a TypeNode, capture any more-precise type permanently into Node 1106 k->raise_bottom_type(t); 1107 // Move users of node to worklist 1108 add_users_to_worklist( k ); 1109 } 1110 1111 // If 'k' computes a constant, replace it with a constant 1112 if( t->singleton() && !k->is_Con() ) { 1113 NOT_PRODUCT( set_progress(); ) 1114 Node *con = makecon(t); // Make a constant 1115 add_users_to_worklist( k ); 1116 subsume_node( k, con ); // Everybody using k now uses con 1117 return con; 1118 } 1119 1120 // Now check for Identities 1121 i = k->Identity(this); // Look for a nearby replacement 1122 if( i != k ) { // Found? Return replacement! 1123 NOT_PRODUCT( set_progress(); ) 1124 add_users_to_worklist( k ); 1125 subsume_node( k, i ); // Everybody using k now uses i 1126 return i; 1127 } 1128 1129 // Global Value Numbering 1130 i = hash_find_insert(k); // Check for pre-existing node 1131 if( i && (i != k) ) { 1132 // Return the pre-existing node if it isn't dead 1133 NOT_PRODUCT( set_progress(); ) 1134 add_users_to_worklist( k ); 1135 subsume_node( k, i ); // Everybody using k now uses i 1136 return i; 1137 } 1138 1139 // Return Idealized original 1140 return k; 1141} 1142 1143//---------------------------------saturate------------------------------------ 1144const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1145 const Type* limit_type) const { 1146 return new_type->narrow(old_type); 1147} 1148 1149//------------------------------remove_globally_dead_node---------------------- 1150// Kill a globally dead Node. All uses are also globally dead and are 1151// aggressively trimmed. 1152void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1153 enum DeleteProgress { 1154 PROCESS_INPUTS, 1155 PROCESS_OUTPUTS 1156 }; 1157 assert(_stack.is_empty(), "not empty"); 1158 _stack.push(dead, PROCESS_INPUTS); 1159 1160 while (_stack.is_nonempty()) { 1161 dead = _stack.node(); 1162 uint progress_state = _stack.index(); 1163 assert(dead != C->root(), "killing root, eh?"); 1164 assert(!dead->is_top(), "add check for top when pushing"); 1165 NOT_PRODUCT( set_progress(); ) 1166 if (progress_state == PROCESS_INPUTS) { 1167 // After following inputs, continue to outputs 1168 _stack.set_index(PROCESS_OUTPUTS); 1169 if (!dead->is_Con()) { // Don't kill cons but uses 1170 bool recurse = false; 1171 // Remove from hash table 1172 _table.hash_delete( dead ); 1173 // Smash all inputs to 'dead', isolating him completely 1174 for (uint i = 0; i < dead->req(); i++) { 1175 Node *in = dead->in(i); 1176 if (in != NULL && in != C->top()) { // Points to something? 1177 int nrep = dead->replace_edge(in, NULL); // Kill edges 1178 assert((nrep > 0), "sanity"); 1179 if (in->outcnt() == 0) { // Made input go dead? 1180 _stack.push(in, PROCESS_INPUTS); // Recursively remove 1181 recurse = true; 1182 } else if (in->outcnt() == 1 && 1183 in->has_special_unique_user()) { 1184 _worklist.push(in->unique_out()); 1185 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1186 if (in->Opcode() == Op_Region) { 1187 _worklist.push(in); 1188 } else if (in->is_Store()) { 1189 DUIterator_Fast imax, i = in->fast_outs(imax); 1190 _worklist.push(in->fast_out(i)); 1191 i++; 1192 if (in->outcnt() == 2) { 1193 _worklist.push(in->fast_out(i)); 1194 i++; 1195 } 1196 assert(!(i < imax), "sanity"); 1197 } 1198 } 1199 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1200 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) { 1201 // A Load that directly follows an InitializeNode is 1202 // going away. The Stores that follow are candidates 1203 // again to be captured by the InitializeNode. 1204 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1205 Node *n = in->fast_out(j); 1206 if (n->is_Store()) { 1207 _worklist.push(n); 1208 } 1209 } 1210 } 1211 } // if (in != NULL && in != C->top()) 1212 } // for (uint i = 0; i < dead->req(); i++) 1213 if (recurse) { 1214 continue; 1215 } 1216 } // if (!dead->is_Con()) 1217 } // if (progress_state == PROCESS_INPUTS) 1218 1219 // Aggressively kill globally dead uses 1220 // (Rather than pushing all the outs at once, we push one at a time, 1221 // plus the parent to resume later, because of the indefinite number 1222 // of edge deletions per loop trip.) 1223 if (dead->outcnt() > 0) { 1224 // Recursively remove output edges 1225 _stack.push(dead->raw_out(0), PROCESS_INPUTS); 1226 } else { 1227 // Finished disconnecting all input and output edges. 1228 _stack.pop(); 1229 // Remove dead node from iterative worklist 1230 _worklist.remove(dead); 1231 // Constant node that has no out-edges and has only one in-edge from 1232 // root is usually dead. However, sometimes reshaping walk makes 1233 // it reachable by adding use edges. So, we will NOT count Con nodes 1234 // as dead to be conservative about the dead node count at any 1235 // given time. 1236 if (!dead->is_Con()) { 1237 C->record_dead_node(dead->_idx); 1238 } 1239 if (dead->is_macro()) { 1240 C->remove_macro_node(dead); 1241 } 1242 if (dead->is_expensive()) { 1243 C->remove_expensive_node(dead); 1244 } 1245 } 1246 } // while (_stack.is_nonempty()) 1247} 1248 1249//------------------------------subsume_node----------------------------------- 1250// Remove users from node 'old' and add them to node 'nn'. 1251void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1252 assert( old != hash_find(old), "should already been removed" ); 1253 assert( old != C->top(), "cannot subsume top node"); 1254 // Copy debug or profile information to the new version: 1255 C->copy_node_notes_to(nn, old); 1256 // Move users of node 'old' to node 'nn' 1257 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1258 Node* use = old->last_out(i); // for each use... 1259 // use might need re-hashing (but it won't if it's a new node) 1260 bool is_in_table = _table.hash_delete( use ); 1261 // Update use-def info as well 1262 // We remove all occurrences of old within use->in, 1263 // so as to avoid rehashing any node more than once. 1264 // The hash table probe swamps any outer loop overhead. 1265 uint num_edges = 0; 1266 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1267 if (use->in(j) == old) { 1268 use->set_req(j, nn); 1269 ++num_edges; 1270 } 1271 } 1272 // Insert into GVN hash table if unique 1273 // If a duplicate, 'use' will be cleaned up when pulled off worklist 1274 if( is_in_table ) { 1275 hash_find_insert(use); 1276 } 1277 i -= num_edges; // we deleted 1 or more copies of this edge 1278 } 1279 1280 // Smash all inputs to 'old', isolating him completely 1281 Node *temp = new (C) Node(1); 1282 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1283 remove_dead_node( old ); 1284 temp->del_req(0); // Yank bogus edge 1285#ifndef PRODUCT 1286 if( VerifyIterativeGVN ) { 1287 for ( int i = 0; i < _verify_window_size; i++ ) { 1288 if ( _verify_window[i] == old ) 1289 _verify_window[i] = nn; 1290 } 1291 } 1292#endif 1293 _worklist.remove(temp); // this can be necessary 1294 temp->destruct(); // reuse the _idx of this little guy 1295} 1296 1297//------------------------------add_users_to_worklist-------------------------- 1298void PhaseIterGVN::add_users_to_worklist0( Node *n ) { 1299 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1300 _worklist.push(n->fast_out(i)); // Push on worklist 1301 } 1302} 1303 1304void PhaseIterGVN::add_users_to_worklist( Node *n ) { 1305 add_users_to_worklist0(n); 1306 1307 // Move users of node to worklist 1308 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1309 Node* use = n->fast_out(i); // Get use 1310 1311 if( use->is_Multi() || // Multi-definer? Push projs on worklist 1312 use->is_Store() ) // Enable store/load same address 1313 add_users_to_worklist0(use); 1314 1315 // If we changed the receiver type to a call, we need to revisit 1316 // the Catch following the call. It's looking for a non-NULL 1317 // receiver to know when to enable the regular fall-through path 1318 // in addition to the NullPtrException path. 1319 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1320 Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control); 1321 if (p != NULL) { 1322 add_users_to_worklist0(p); 1323 } 1324 } 1325 1326 if( use->is_Cmp() ) { // Enable CMP/BOOL optimization 1327 add_users_to_worklist(use); // Put Bool on worklist 1328 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1329 // phi merging either 0 or 1 onto the worklist 1330 if (use->outcnt() > 0) { 1331 Node* bol = use->raw_out(0); 1332 if (bol->outcnt() > 0) { 1333 Node* iff = bol->raw_out(0); 1334 if (iff->outcnt() == 2) { 1335 Node* ifproj0 = iff->raw_out(0); 1336 Node* ifproj1 = iff->raw_out(1); 1337 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1338 Node* region0 = ifproj0->raw_out(0); 1339 Node* region1 = ifproj1->raw_out(0); 1340 if( region0 == region1 ) 1341 add_users_to_worklist0(region0); 1342 } 1343 } 1344 } 1345 } 1346 } 1347 1348 uint use_op = use->Opcode(); 1349 // If changed Cast input, check Phi users for simple cycles 1350 if( use->is_ConstraintCast() || use->is_CheckCastPP() ) { 1351 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1352 Node* u = use->fast_out(i2); 1353 if (u->is_Phi()) 1354 _worklist.push(u); 1355 } 1356 } 1357 // If changed LShift inputs, check RShift users for useless sign-ext 1358 if( use_op == Op_LShiftI ) { 1359 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1360 Node* u = use->fast_out(i2); 1361 if (u->Opcode() == Op_RShiftI) 1362 _worklist.push(u); 1363 } 1364 } 1365 // If changed AddP inputs, check Stores for loop invariant 1366 if( use_op == Op_AddP ) { 1367 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1368 Node* u = use->fast_out(i2); 1369 if (u->is_Mem()) 1370 _worklist.push(u); 1371 } 1372 } 1373 // If changed initialization activity, check dependent Stores 1374 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1375 InitializeNode* init = use->as_Allocate()->initialization(); 1376 if (init != NULL) { 1377 Node* imem = init->proj_out(TypeFunc::Memory); 1378 if (imem != NULL) add_users_to_worklist0(imem); 1379 } 1380 } 1381 if (use_op == Op_Initialize) { 1382 Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory); 1383 if (imem != NULL) add_users_to_worklist0(imem); 1384 } 1385 } 1386} 1387 1388/** 1389 * Remove the speculative part of all types that we know of 1390 */ 1391void PhaseIterGVN::remove_speculative_types() { 1392 assert(UseTypeSpeculation, "speculation is off"); 1393 for (uint i = 0; i < _types.Size(); i++) { 1394 const Type* t = _types.fast_lookup(i); 1395 if (t != NULL && t->isa_oopptr()) { 1396 const TypeOopPtr* to = t->is_oopptr(); 1397 _types.map(i, to->remove_speculative()); 1398 } 1399 } 1400} 1401 1402//============================================================================= 1403#ifndef PRODUCT 1404uint PhaseCCP::_total_invokes = 0; 1405uint PhaseCCP::_total_constants = 0; 1406#endif 1407//------------------------------PhaseCCP--------------------------------------- 1408// Conditional Constant Propagation, ala Wegman & Zadeck 1409PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1410 NOT_PRODUCT( clear_constants(); ) 1411 assert( _worklist.size() == 0, "" ); 1412 // Clear out _nodes from IterGVN. Must be clear to transform call. 1413 _nodes.clear(); // Clear out from IterGVN 1414 analyze(); 1415} 1416 1417#ifndef PRODUCT 1418//------------------------------~PhaseCCP-------------------------------------- 1419PhaseCCP::~PhaseCCP() { 1420 inc_invokes(); 1421 _total_constants += count_constants(); 1422} 1423#endif 1424 1425 1426#ifdef ASSERT 1427static bool ccp_type_widens(const Type* t, const Type* t0) { 1428 assert(t->meet(t0) == t, "Not monotonic"); 1429 switch (t->base() == t0->base() ? t->base() : Type::Top) { 1430 case Type::Int: 1431 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases"); 1432 break; 1433 case Type::Long: 1434 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases"); 1435 break; 1436 } 1437 return true; 1438} 1439#endif //ASSERT 1440 1441//------------------------------analyze---------------------------------------- 1442void PhaseCCP::analyze() { 1443 // Initialize all types to TOP, optimistic analysis 1444 for (int i = C->unique() - 1; i >= 0; i--) { 1445 _types.map(i,Type::TOP); 1446 } 1447 1448 // Push root onto worklist 1449 Unique_Node_List worklist; 1450 worklist.push(C->root()); 1451 1452 // Pull from worklist; compute new value; push changes out. 1453 // This loop is the meat of CCP. 1454 while( worklist.size() ) { 1455 Node *n = worklist.pop(); 1456 const Type *t = n->Value(this); 1457 if (t != type(n)) { 1458 assert(ccp_type_widens(t, type(n)), "ccp type must widen"); 1459#ifndef PRODUCT 1460 if( TracePhaseCCP ) { 1461 t->dump(); 1462 do { tty->print("\t"); } while (tty->position() < 16); 1463 n->dump(); 1464 } 1465#endif 1466 set_type(n, t); 1467 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1468 Node* m = n->fast_out(i); // Get user 1469 if( m->is_Region() ) { // New path to Region? Must recheck Phis too 1470 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1471 Node* p = m->fast_out(i2); // Propagate changes to uses 1472 if( p->bottom_type() != type(p) ) // If not already bottomed out 1473 worklist.push(p); // Propagate change to user 1474 } 1475 } 1476 // If we changed the receiver type to a call, we need to revisit 1477 // the Catch following the call. It's looking for a non-NULL 1478 // receiver to know when to enable the regular fall-through path 1479 // in addition to the NullPtrException path 1480 if (m->is_Call()) { 1481 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) { 1482 Node* p = m->fast_out(i2); // Propagate changes to uses 1483 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) 1484 worklist.push(p->unique_out()); 1485 } 1486 } 1487 if( m->bottom_type() != type(m) ) // If not already bottomed out 1488 worklist.push(m); // Propagate change to user 1489 } 1490 } 1491 } 1492} 1493 1494//------------------------------do_transform----------------------------------- 1495// Top level driver for the recursive transformer 1496void PhaseCCP::do_transform() { 1497 // Correct leaves of new-space Nodes; they point to old-space. 1498 C->set_root( transform(C->root())->as_Root() ); 1499 assert( C->top(), "missing TOP node" ); 1500 assert( C->root(), "missing root" ); 1501} 1502 1503//------------------------------transform-------------------------------------- 1504// Given a Node in old-space, clone him into new-space. 1505// Convert any of his old-space children into new-space children. 1506Node *PhaseCCP::transform( Node *n ) { 1507 Node *new_node = _nodes[n->_idx]; // Check for transformed node 1508 if( new_node != NULL ) 1509 return new_node; // Been there, done that, return old answer 1510 new_node = transform_once(n); // Check for constant 1511 _nodes.map( n->_idx, new_node ); // Flag as having been cloned 1512 1513 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc 1514 GrowableArray <Node *> trstack(C->unique() >> 1); 1515 1516 trstack.push(new_node); // Process children of cloned node 1517 while ( trstack.is_nonempty() ) { 1518 Node *clone = trstack.pop(); 1519 uint cnt = clone->req(); 1520 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 1521 Node *input = clone->in(i); 1522 if( input != NULL ) { // Ignore NULLs 1523 Node *new_input = _nodes[input->_idx]; // Check for cloned input node 1524 if( new_input == NULL ) { 1525 new_input = transform_once(input); // Check for constant 1526 _nodes.map( input->_idx, new_input );// Flag as having been cloned 1527 trstack.push(new_input); 1528 } 1529 assert( new_input == clone->in(i), "insanity check"); 1530 } 1531 } 1532 } 1533 return new_node; 1534} 1535 1536 1537//------------------------------transform_once--------------------------------- 1538// For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 1539Node *PhaseCCP::transform_once( Node *n ) { 1540 const Type *t = type(n); 1541 // Constant? Use constant Node instead 1542 if( t->singleton() ) { 1543 Node *nn = n; // Default is to return the original constant 1544 if( t == Type::TOP ) { 1545 // cache my top node on the Compile instance 1546 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) { 1547 C->set_cached_top_node( ConNode::make(C, Type::TOP) ); 1548 set_type(C->top(), Type::TOP); 1549 } 1550 nn = C->top(); 1551 } 1552 if( !n->is_Con() ) { 1553 if( t != Type::TOP ) { 1554 nn = makecon(t); // ConNode::make(t); 1555 NOT_PRODUCT( inc_constants(); ) 1556 } else if( n->is_Region() ) { // Unreachable region 1557 // Note: nn == C->top() 1558 n->set_req(0, NULL); // Cut selfreference 1559 // Eagerly remove dead phis to avoid phis copies creation. 1560 for (DUIterator i = n->outs(); n->has_out(i); i++) { 1561 Node* m = n->out(i); 1562 if( m->is_Phi() ) { 1563 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 1564 replace_node(m, nn); 1565 --i; // deleted this phi; rescan starting with next position 1566 } 1567 } 1568 } 1569 replace_node(n,nn); // Update DefUse edges for new constant 1570 } 1571 return nn; 1572 } 1573 1574 // If x is a TypeNode, capture any more-precise type permanently into Node 1575 if (t != n->bottom_type()) { 1576 hash_delete(n); // changing bottom type may force a rehash 1577 n->raise_bottom_type(t); 1578 _worklist.push(n); // n re-enters the hash table via the worklist 1579 } 1580 1581 // Idealize graph using DU info. Must clone() into new-space. 1582 // DU info is generally used to show profitability, progress or safety 1583 // (but generally not needed for correctness). 1584 Node *nn = n->Ideal_DU_postCCP(this); 1585 1586 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks 1587 switch( n->Opcode() ) { 1588 case Op_FastLock: // Revisit FastLocks for lock coarsening 1589 case Op_If: 1590 case Op_CountedLoopEnd: 1591 case Op_Region: 1592 case Op_Loop: 1593 case Op_CountedLoop: 1594 case Op_Conv2B: 1595 case Op_Opaque1: 1596 case Op_Opaque2: 1597 _worklist.push(n); 1598 break; 1599 default: 1600 break; 1601 } 1602 if( nn ) { 1603 _worklist.push(n); 1604 // Put users of 'n' onto worklist for second igvn transform 1605 add_users_to_worklist(n); 1606 return nn; 1607 } 1608 1609 return n; 1610} 1611 1612//---------------------------------saturate------------------------------------ 1613const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 1614 const Type* limit_type) const { 1615 const Type* wide_type = new_type->widen(old_type, limit_type); 1616 if (wide_type != new_type) { // did we widen? 1617 // If so, we may have widened beyond the limit type. Clip it back down. 1618 new_type = wide_type->filter(limit_type); 1619 } 1620 return new_type; 1621} 1622 1623//------------------------------print_statistics------------------------------- 1624#ifndef PRODUCT 1625void PhaseCCP::print_statistics() { 1626 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 1627} 1628#endif 1629 1630 1631//============================================================================= 1632#ifndef PRODUCT 1633uint PhasePeephole::_total_peepholes = 0; 1634#endif 1635//------------------------------PhasePeephole---------------------------------- 1636// Conditional Constant Propagation, ala Wegman & Zadeck 1637PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 1638 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 1639 NOT_PRODUCT( clear_peepholes(); ) 1640} 1641 1642#ifndef PRODUCT 1643//------------------------------~PhasePeephole--------------------------------- 1644PhasePeephole::~PhasePeephole() { 1645 _total_peepholes += count_peepholes(); 1646} 1647#endif 1648 1649//------------------------------transform-------------------------------------- 1650Node *PhasePeephole::transform( Node *n ) { 1651 ShouldNotCallThis(); 1652 return NULL; 1653} 1654 1655//------------------------------do_transform----------------------------------- 1656void PhasePeephole::do_transform() { 1657 bool method_name_not_printed = true; 1658 1659 // Examine each basic block 1660 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 1661 Block* block = _cfg.get_block(block_number); 1662 bool block_not_printed = true; 1663 1664 // and each instruction within a block 1665 uint end_index = block->number_of_nodes(); 1666 // block->end_idx() not valid after PhaseRegAlloc 1667 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) { 1668 Node *n = block->get_node(instruction_index); 1669 if( n->is_Mach() ) { 1670 MachNode *m = n->as_Mach(); 1671 int deleted_count = 0; 1672 // check for peephole opportunities 1673 MachNode *m2 = m->peephole( block, instruction_index, _regalloc, deleted_count, C ); 1674 if( m2 != NULL ) { 1675#ifndef PRODUCT 1676 if( PrintOptoPeephole ) { 1677 // Print method, first time only 1678 if( C->method() && method_name_not_printed ) { 1679 C->method()->print_short_name(); tty->cr(); 1680 method_name_not_printed = false; 1681 } 1682 // Print this block 1683 if( Verbose && block_not_printed) { 1684 tty->print_cr("in block"); 1685 block->dump(); 1686 block_not_printed = false; 1687 } 1688 // Print instructions being deleted 1689 for( int i = (deleted_count - 1); i >= 0; --i ) { 1690 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr(); 1691 } 1692 tty->print_cr("replaced with"); 1693 // Print new instruction 1694 m2->format(_regalloc); 1695 tty->print("\n\n"); 1696 } 1697#endif 1698 // Remove old nodes from basic block and update instruction_index 1699 // (old nodes still exist and may have edges pointing to them 1700 // as register allocation info is stored in the allocator using 1701 // the node index to live range mappings.) 1702 uint safe_instruction_index = (instruction_index - deleted_count); 1703 for( ; (instruction_index > safe_instruction_index); --instruction_index ) { 1704 block->remove_node( instruction_index ); 1705 } 1706 // install new node after safe_instruction_index 1707 block->insert_node(m2, safe_instruction_index + 1); 1708 end_index = block->number_of_nodes() - 1; // Recompute new block size 1709 NOT_PRODUCT( inc_peepholes(); ) 1710 } 1711 } 1712 } 1713 } 1714} 1715 1716//------------------------------print_statistics------------------------------- 1717#ifndef PRODUCT 1718void PhasePeephole::print_statistics() { 1719 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 1720} 1721#endif 1722 1723 1724//============================================================================= 1725//------------------------------set_req_X-------------------------------------- 1726void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 1727 assert( is_not_dead(n), "can not use dead node"); 1728 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 1729 Node *old = in(i); 1730 set_req(i, n); 1731 1732 // old goes dead? 1733 if( old ) { 1734 switch (old->outcnt()) { 1735 case 0: 1736 // Put into the worklist to kill later. We do not kill it now because the 1737 // recursive kill will delete the current node (this) if dead-loop exists 1738 if (!old->is_top()) 1739 igvn->_worklist.push( old ); 1740 break; 1741 case 1: 1742 if( old->is_Store() || old->has_special_unique_user() ) 1743 igvn->add_users_to_worklist( old ); 1744 break; 1745 case 2: 1746 if( old->is_Store() ) 1747 igvn->add_users_to_worklist( old ); 1748 if( old->Opcode() == Op_Region ) 1749 igvn->_worklist.push(old); 1750 break; 1751 case 3: 1752 if( old->Opcode() == Op_Region ) { 1753 igvn->_worklist.push(old); 1754 igvn->add_users_to_worklist( old ); 1755 } 1756 break; 1757 default: 1758 break; 1759 } 1760 } 1761 1762} 1763 1764//-------------------------------replace_by----------------------------------- 1765// Using def-use info, replace one node for another. Follow the def-use info 1766// to all users of the OLD node. Then make all uses point to the NEW node. 1767void Node::replace_by(Node *new_node) { 1768 assert(!is_top(), "top node has no DU info"); 1769 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 1770 Node* use = last_out(i); 1771 uint uses_found = 0; 1772 for (uint j = 0; j < use->len(); j++) { 1773 if (use->in(j) == this) { 1774 if (j < use->req()) 1775 use->set_req(j, new_node); 1776 else use->set_prec(j, new_node); 1777 uses_found++; 1778 } 1779 } 1780 i -= uses_found; // we deleted 1 or more copies of this edge 1781 } 1782} 1783 1784//============================================================================= 1785//----------------------------------------------------------------------------- 1786void Type_Array::grow( uint i ) { 1787 if( !_max ) { 1788 _max = 1; 1789 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 1790 _types[0] = NULL; 1791 } 1792 uint old = _max; 1793 while( i >= _max ) _max <<= 1; // Double to fit 1794 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 1795 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 1796} 1797 1798//------------------------------dump------------------------------------------- 1799#ifndef PRODUCT 1800void Type_Array::dump() const { 1801 uint max = Size(); 1802 for( uint i = 0; i < max; i++ ) { 1803 if( _types[i] != NULL ) { 1804 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 1805 } 1806 } 1807} 1808#endif 1809