1 /*
   2  * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp"
  27 #include "gc/shared/c2/barrierSetC2.hpp"
  28 #include "memory/allocation.inline.hpp"
  29 #include "memory/resourceArea.hpp"
  30 #include "opto/block.hpp"
  31 #include "opto/callnode.hpp"
  32 #include "opto/castnode.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/idealGraphPrinter.hpp"
  35 #include "opto/loopnode.hpp"
  36 #include "opto/machnode.hpp"
  37 #include "opto/opcodes.hpp"
  38 #include "opto/phaseX.hpp"
  39 #include "opto/regalloc.hpp"
  40 #include "opto/rootnode.hpp"
  41 #include "utilities/macros.hpp"
  42 
  43 //=============================================================================
  44 #define NODE_HASH_MINIMUM_SIZE    255
  45 //------------------------------NodeHash---------------------------------------
  46 NodeHash::NodeHash(uint est_max_size) :
  47   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
  48   _a(Thread::current()->resource_area()),
  49   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ), // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
  50   _inserts(0), _insert_limit( insert_limit() )
  51 #ifndef PRODUCT
  52   ,_look_probes(0), _lookup_hits(0), _lookup_misses(0),
  53   _delete_probes(0), _delete_hits(0), _delete_misses(0),
  54   _total_insert_probes(0), _total_inserts(0),
  55   _insert_probes(0), _grows(0)
  56 #endif
  57 {
  58   // _sentinel must be in the current node space
  59   _sentinel = new ProjNode(NULL, TypeFunc::Control);
  60   memset(_table,0,sizeof(Node*)*_max);
  61 }
  62 
  63 //------------------------------NodeHash---------------------------------------
  64 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
  65   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
  66   _a(arena),
  67   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ),
  68   _inserts(0), _insert_limit( insert_limit() )
  69 #ifndef PRODUCT
  70   ,_look_probes(0), _lookup_hits(0), _lookup_misses(0),
  71   _delete_probes(0), _delete_hits(0), _delete_misses(0),
  72   _total_insert_probes(0), _total_inserts(0),
  73   _insert_probes(0), _grows(0)
  74 #endif
  75 {
  76   // _sentinel must be in the current node space
  77   _sentinel = new ProjNode(NULL, TypeFunc::Control);
  78   memset(_table,0,sizeof(Node*)*_max);
  79 }
  80 
  81 //------------------------------NodeHash---------------------------------------
  82 NodeHash::NodeHash(NodeHash *nh) {
  83   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
  84   // just copy in all the fields
  85   *this = *nh;
  86   // nh->_sentinel must be in the current node space
  87 }
  88 
  89 void NodeHash::replace_with(NodeHash *nh) {
  90   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
  91   // just copy in all the fields
  92   *this = *nh;
  93   // nh->_sentinel must be in the current node space
  94 }
  95 
  96 //------------------------------hash_find--------------------------------------
  97 // Find in hash table
  98 Node *NodeHash::hash_find( const Node *n ) {
  99   // ((Node*)n)->set_hash( n->hash() );
 100   uint hash = n->hash();
 101   if (hash == Node::NO_HASH) {
 102     NOT_PRODUCT( _lookup_misses++ );
 103     return NULL;
 104   }
 105   uint key = hash & (_max-1);
 106   uint stride = key | 0x01;
 107   NOT_PRODUCT( _look_probes++ );
 108   Node *k = _table[key];        // Get hashed value
 109   if( !k ) {                    // ?Miss?
 110     NOT_PRODUCT( _lookup_misses++ );
 111     return NULL;                // Miss!
 112   }
 113 
 114   int op = n->Opcode();
 115   uint req = n->req();
 116   while( 1 ) {                  // While probing hash table
 117     if( k->req() == req &&      // Same count of inputs
 118         k->Opcode() == op ) {   // Same Opcode
 119       for( uint i=0; i<req; i++ )
 120         if( n->in(i)!=k->in(i)) // Different inputs?
 121           goto collision;       // "goto" is a speed hack...
 122       if( n->cmp(*k) ) {        // Check for any special bits
 123         NOT_PRODUCT( _lookup_hits++ );
 124         return k;               // Hit!
 125       }
 126     }
 127   collision:
 128     NOT_PRODUCT( _look_probes++ );
 129     key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
 130     k = _table[key];            // Get hashed value
 131     if( !k ) {                  // ?Miss?
 132       NOT_PRODUCT( _lookup_misses++ );
 133       return NULL;              // Miss!
 134     }
 135   }
 136   ShouldNotReachHere();
 137   return NULL;
 138 }
 139 
 140 //------------------------------hash_find_insert-------------------------------
 141 // Find in hash table, insert if not already present
 142 // Used to preserve unique entries in hash table
 143 Node *NodeHash::hash_find_insert( Node *n ) {
 144   // n->set_hash( );
 145   uint hash = n->hash();
 146   if (hash == Node::NO_HASH) {
 147     NOT_PRODUCT( _lookup_misses++ );
 148     return NULL;
 149   }
 150   uint key = hash & (_max-1);
 151   uint stride = key | 0x01;     // stride must be relatively prime to table siz
 152   uint first_sentinel = 0;      // replace a sentinel if seen.
 153   NOT_PRODUCT( _look_probes++ );
 154   Node *k = _table[key];        // Get hashed value
 155   if( !k ) {                    // ?Miss?
 156     NOT_PRODUCT( _lookup_misses++ );
 157     _table[key] = n;            // Insert into table!
 158     debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
 159     check_grow();               // Grow table if insert hit limit
 160     return NULL;                // Miss!
 161   }
 162   else if( k == _sentinel ) {
 163     first_sentinel = key;      // Can insert here
 164   }
 165 
 166   int op = n->Opcode();
 167   uint req = n->req();
 168   while( 1 ) {                  // While probing hash table
 169     if( k->req() == req &&      // Same count of inputs
 170         k->Opcode() == op ) {   // Same Opcode
 171       for( uint i=0; i<req; i++ )
 172         if( n->in(i)!=k->in(i)) // Different inputs?
 173           goto collision;       // "goto" is a speed hack...
 174       if( n->cmp(*k) ) {        // Check for any special bits
 175         NOT_PRODUCT( _lookup_hits++ );
 176         return k;               // Hit!
 177       }
 178     }
 179   collision:
 180     NOT_PRODUCT( _look_probes++ );
 181     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
 182     k = _table[key];            // Get hashed value
 183     if( !k ) {                  // ?Miss?
 184       NOT_PRODUCT( _lookup_misses++ );
 185       key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
 186       _table[key] = n;          // Insert into table!
 187       debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
 188       check_grow();             // Grow table if insert hit limit
 189       return NULL;              // Miss!
 190     }
 191     else if( first_sentinel == 0 && k == _sentinel ) {
 192       first_sentinel = key;    // Can insert here
 193     }
 194 
 195   }
 196   ShouldNotReachHere();
 197   return NULL;
 198 }
 199 
 200 //------------------------------hash_insert------------------------------------
 201 // Insert into hash table
 202 void NodeHash::hash_insert( Node *n ) {
 203   // // "conflict" comments -- print nodes that conflict
 204   // bool conflict = false;
 205   // n->set_hash();
 206   uint hash = n->hash();
 207   if (hash == Node::NO_HASH) {
 208     return;
 209   }
 210   check_grow();
 211   uint key = hash & (_max-1);
 212   uint stride = key | 0x01;
 213 
 214   while( 1 ) {                  // While probing hash table
 215     NOT_PRODUCT( _insert_probes++ );
 216     Node *k = _table[key];      // Get hashed value
 217     if( !k || (k == _sentinel) ) break;       // Found a slot
 218     assert( k != n, "already inserted" );
 219     // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print("  conflict: "); k->dump(); conflict = true; }
 220     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
 221   }
 222   _table[key] = n;              // Insert into table!
 223   debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
 224   // if( conflict ) { n->dump(); }
 225 }
 226 
 227 //------------------------------hash_delete------------------------------------
 228 // Replace in hash table with sentinel
 229 bool NodeHash::hash_delete( const Node *n ) {
 230   Node *k;
 231   uint hash = n->hash();
 232   if (hash == Node::NO_HASH) {
 233     NOT_PRODUCT( _delete_misses++ );
 234     return false;
 235   }
 236   uint key = hash & (_max-1);
 237   uint stride = key | 0x01;
 238   debug_only( uint counter = 0; );
 239   for( ; /* (k != NULL) && (k != _sentinel) */; ) {
 240     debug_only( counter++ );
 241     NOT_PRODUCT( _delete_probes++ );
 242     k = _table[key];            // Get hashed value
 243     if( !k ) {                  // Miss?
 244       NOT_PRODUCT( _delete_misses++ );
 245 #ifdef ASSERT
 246       if( VerifyOpto ) {
 247         for( uint i=0; i < _max; i++ )
 248           assert( _table[i] != n, "changed edges with rehashing" );
 249       }
 250 #endif
 251       return false;             // Miss! Not in chain
 252     }
 253     else if( n == k ) {
 254       NOT_PRODUCT( _delete_hits++ );
 255       _table[key] = _sentinel;  // Hit! Label as deleted entry
 256       debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
 257       return true;
 258     }
 259     else {
 260       // collision: move through table with prime offset
 261       key = (key + stride/*7*/) & (_max-1);
 262       assert( counter <= _insert_limit, "Cycle in hash-table");
 263     }
 264   }
 265   ShouldNotReachHere();
 266   return false;
 267 }
 268 
 269 //------------------------------round_up---------------------------------------
 270 // Round up to nearest power of 2
 271 uint NodeHash::round_up( uint x ) {
 272   x += (x>>2);                  // Add 25% slop
 273   if( x <16 ) return 16;        // Small stuff
 274   uint i=16;
 275   while( i < x ) i <<= 1;       // Double to fit
 276   return i;                     // Return hash table size
 277 }
 278 
 279 //------------------------------grow-------------------------------------------
 280 // Grow _table to next power of 2 and insert old entries
 281 void  NodeHash::grow() {
 282   // Record old state
 283   uint   old_max   = _max;
 284   Node **old_table = _table;
 285   // Construct new table with twice the space
 286 #ifndef PRODUCT
 287   _grows++;
 288   _total_inserts       += _inserts;
 289   _total_insert_probes += _insert_probes;
 290   _insert_probes   = 0;
 291 #endif
 292   _inserts         = 0;
 293   _max     = _max << 1;
 294   _table   = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
 295   memset(_table,0,sizeof(Node*)*_max);
 296   _insert_limit = insert_limit();
 297   // Insert old entries into the new table
 298   for( uint i = 0; i < old_max; i++ ) {
 299     Node *m = *old_table++;
 300     if( !m || m == _sentinel ) continue;
 301     debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
 302     hash_insert(m);
 303   }
 304 }
 305 
 306 //------------------------------clear------------------------------------------
 307 // Clear all entries in _table to NULL but keep storage
 308 void  NodeHash::clear() {
 309 #ifdef ASSERT
 310   // Unlock all nodes upon removal from table.
 311   for (uint i = 0; i < _max; i++) {
 312     Node* n = _table[i];
 313     if (!n || n == _sentinel)  continue;
 314     n->exit_hash_lock();
 315   }
 316 #endif
 317 
 318   memset( _table, 0, _max * sizeof(Node*) );
 319 }
 320 
 321 //-----------------------remove_useless_nodes----------------------------------
 322 // Remove useless nodes from value table,
 323 // implementation does not depend on hash function
 324 void NodeHash::remove_useless_nodes(VectorSet &useful) {
 325 
 326   // Dead nodes in the hash table inherited from GVN should not replace
 327   // existing nodes, remove dead nodes.
 328   uint max = size();
 329   Node *sentinel_node = sentinel();
 330   for( uint i = 0; i < max; ++i ) {
 331     Node *n = at(i);
 332     if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
 333       debug_only(n->exit_hash_lock()); // Unlock the node when removed
 334       _table[i] = sentinel_node;       // Replace with placeholder
 335     }
 336   }
 337 }
 338 
 339 
 340 void NodeHash::check_no_speculative_types() {
 341 #ifdef ASSERT
 342   uint max = size();
 343   Node *sentinel_node = sentinel();
 344   for (uint i = 0; i < max; ++i) {
 345     Node *n = at(i);
 346     if(n != NULL && n != sentinel_node && n->is_Type() && n->outcnt() > 0) {
 347       TypeNode* tn = n->as_Type();
 348       const Type* t = tn->type();
 349       const Type* t_no_spec = t->remove_speculative();
 350       assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
 351     }
 352   }
 353 #endif
 354 }
 355 
 356 #ifndef PRODUCT
 357 //------------------------------dump-------------------------------------------
 358 // Dump statistics for the hash table
 359 void NodeHash::dump() {
 360   _total_inserts       += _inserts;
 361   _total_insert_probes += _insert_probes;
 362   if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
 363     if (WizardMode) {
 364       for (uint i=0; i<_max; i++) {
 365         if (_table[i])
 366           tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
 367       }
 368     }
 369     tty->print("\nGVN Hash stats:  %d grows to %d max_size\n", _grows, _max);
 370     tty->print("  %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
 371     tty->print("  %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
 372     tty->print("  %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
 373     // sentinels increase lookup cost, but not insert cost
 374     assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
 375     assert( _inserts+(_inserts>>3) < _max, "table too full" );
 376     assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
 377   }
 378 }
 379 
 380 Node *NodeHash::find_index(uint idx) { // For debugging
 381   // Find an entry by its index value
 382   for( uint i = 0; i < _max; i++ ) {
 383     Node *m = _table[i];
 384     if( !m || m == _sentinel ) continue;
 385     if( m->_idx == (uint)idx ) return m;
 386   }
 387   return NULL;
 388 }
 389 #endif
 390 
 391 #ifdef ASSERT
 392 NodeHash::~NodeHash() {
 393   // Unlock all nodes upon destruction of table.
 394   if (_table != (Node**)badAddress)  clear();
 395 }
 396 
 397 void NodeHash::operator=(const NodeHash& nh) {
 398   // Unlock all nodes upon replacement of table.
 399   if (&nh == this)  return;
 400   if (_table != (Node**)badAddress)  clear();
 401   memcpy((void*)this, (void*)&nh, sizeof(*this));
 402   // Do not increment hash_lock counts again.
 403   // Instead, be sure we never again use the source table.
 404   ((NodeHash*)&nh)->_table = (Node**)badAddress;
 405 }
 406 
 407 
 408 #endif
 409 
 410 
 411 //=============================================================================
 412 //------------------------------PhaseRemoveUseless-----------------------------
 413 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
 414 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN *gvn, Unique_Node_List *worklist, PhaseNumber phase_num) : Phase(phase_num),
 415   _useful(Thread::current()->resource_area()) {
 416 
 417   // Implementation requires 'UseLoopSafepoints == true' and an edge from root
 418   // to each SafePointNode at a backward branch.  Inserted in add_safepoint().
 419   if( !UseLoopSafepoints || !OptoRemoveUseless ) return;
 420 
 421   // Identify nodes that are reachable from below, useful.
 422   C->identify_useful_nodes(_useful);
 423   // Update dead node list
 424   C->update_dead_node_list(_useful);
 425 
 426   // Remove all useless nodes from PhaseValues' recorded types
 427   // Must be done before disconnecting nodes to preserve hash-table-invariant
 428   gvn->remove_useless_nodes(_useful.member_set());
 429 
 430   // Remove all useless nodes from future worklist
 431   worklist->remove_useless_nodes(_useful.member_set());
 432 
 433   // Disconnect 'useless' nodes that are adjacent to useful nodes
 434   C->remove_useless_nodes(_useful);
 435 
 436   // Remove edges from "root" to each SafePoint at a backward branch.
 437   // They were inserted during parsing (see add_safepoint()) to make infinite
 438   // loops without calls or exceptions visible to root, i.e., useful.
 439   Node *root = C->root();
 440   if( root != NULL ) {
 441     for( uint i = root->req(); i < root->len(); ++i ) {
 442       Node *n = root->in(i);
 443       if( n != NULL && n->is_SafePoint() ) {
 444         root->rm_prec(i);
 445         --i;
 446       }
 447     }
 448   }
 449 }
 450 
 451 //=============================================================================
 452 //------------------------------PhaseRenumberLive------------------------------
 453 // First, remove useless nodes (equivalent to identifying live nodes).
 454 // Then, renumber live nodes.
 455 //
 456 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
 457 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
 458 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
 459 // value in the range [0, x).
 460 //
 461 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
 462 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
 463 //
 464 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
 465 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be
 466 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'.
 467 // (2) Type information (the field PhaseGVN::_types) maps type information to each
 468 // node ID. The mapping is updated to use the new node IDs as well. Updated type
 469 // information is returned in PhaseGVN::_types.
 470 //
 471 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist.
 472 //
 473 // Other data structures used by the compiler are not updated. The hash table for value
 474 // numbering (the field PhaseGVN::_table) is not updated because computing the hash
 475 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either
 476 // because it is empty wherever PhaseRenumberLive is used.
 477 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
 478                                      Unique_Node_List* worklist, Unique_Node_List* new_worklist,
 479                                      PhaseNumber phase_num) :
 480   PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live) {
 481 
 482   assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
 483   assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
 484   assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point");
 485 
 486   uint old_unique_count = C->unique();
 487   uint live_node_count = C->live_nodes();
 488   uint worklist_size = worklist->size();
 489 
 490   // Storage for the updated type information.
 491   Type_Array new_type_array(C->comp_arena());
 492 
 493   // Iterate over the set of live nodes.
 494   uint current_idx = 0; // The current new node ID. Incremented after every assignment.
 495   for (uint i = 0; i < _useful.size(); i++) {
 496     Node* n = _useful.at(i);
 497     // Sanity check that fails if we ever decide to execute this phase after EA
 498     assert(!n->is_Phi() || n->as_Phi()->inst_mem_id() == -1, "should not be linked to data Phi");
 499     const Type* type = gvn->type_or_null(n);
 500     new_type_array.map(current_idx, type);
 501 
 502     bool in_worklist = false;
 503     if (worklist->member(n)) {
 504       in_worklist = true;
 505     }
 506 
 507     n->set_idx(current_idx); // Update node ID.
 508 
 509     if (in_worklist) {
 510       new_worklist->push(n);
 511     }
 512 
 513     current_idx++;
 514   }
 515 
 516   assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist");
 517   assert(live_node_count == current_idx, "all live nodes must be processed");
 518 
 519   // Replace the compiler's type information with the updated type information.
 520   gvn->replace_types(new_type_array);
 521 
 522   // Update the unique node count of the compilation to the number of currently live nodes.
 523   C->set_unique(live_node_count);
 524 
 525   // Set the dead node count to 0 and reset dead node list.
 526   C->reset_dead_node_list();
 527 }
 528 
 529 
 530 //=============================================================================
 531 //------------------------------PhaseTransform---------------------------------
 532 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
 533   _arena(Thread::current()->resource_area()),
 534   _nodes(_arena),
 535   _types(_arena)
 536 {
 537   init_con_caches();
 538 #ifndef PRODUCT
 539   clear_progress();
 540   clear_transforms();
 541   set_allow_progress(true);
 542 #endif
 543   // Force allocation for currently existing nodes
 544   _types.map(C->unique(), NULL);
 545 }
 546 
 547 //------------------------------PhaseTransform---------------------------------
 548 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
 549   _arena(arena),
 550   _nodes(arena),
 551   _types(arena)
 552 {
 553   init_con_caches();
 554 #ifndef PRODUCT
 555   clear_progress();
 556   clear_transforms();
 557   set_allow_progress(true);
 558 #endif
 559   // Force allocation for currently existing nodes
 560   _types.map(C->unique(), NULL);
 561 }
 562 
 563 //------------------------------PhaseTransform---------------------------------
 564 // Initialize with previously generated type information
 565 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
 566   _arena(pt->_arena),
 567   _nodes(pt->_nodes),
 568   _types(pt->_types)
 569 {
 570   init_con_caches();
 571 #ifndef PRODUCT
 572   clear_progress();
 573   clear_transforms();
 574   set_allow_progress(true);
 575 #endif
 576 }
 577 
 578 void PhaseTransform::init_con_caches() {
 579   memset(_icons,0,sizeof(_icons));
 580   memset(_lcons,0,sizeof(_lcons));
 581   memset(_zcons,0,sizeof(_zcons));
 582 }
 583 
 584 
 585 //--------------------------------find_int_type--------------------------------
 586 const TypeInt* PhaseTransform::find_int_type(Node* n) {
 587   if (n == NULL)  return NULL;
 588   // Call type_or_null(n) to determine node's type since we might be in
 589   // parse phase and call n->Value() may return wrong type.
 590   // (For example, a phi node at the beginning of loop parsing is not ready.)
 591   const Type* t = type_or_null(n);
 592   if (t == NULL)  return NULL;
 593   return t->isa_int();
 594 }
 595 
 596 
 597 //-------------------------------find_long_type--------------------------------
 598 const TypeLong* PhaseTransform::find_long_type(Node* n) {
 599   if (n == NULL)  return NULL;
 600   // (See comment above on type_or_null.)
 601   const Type* t = type_or_null(n);
 602   if (t == NULL)  return NULL;
 603   return t->isa_long();
 604 }
 605 
 606 
 607 #ifndef PRODUCT
 608 void PhaseTransform::dump_old2new_map() const {
 609   _nodes.dump();
 610 }
 611 
 612 void PhaseTransform::dump_new( uint nidx ) const {
 613   for( uint i=0; i<_nodes.Size(); i++ )
 614     if( _nodes[i] && _nodes[i]->_idx == nidx ) {
 615       _nodes[i]->dump();
 616       tty->cr();
 617       tty->print_cr("Old index= %d",i);
 618       return;
 619     }
 620   tty->print_cr("Node %d not found in the new indices", nidx);
 621 }
 622 
 623 //------------------------------dump_types-------------------------------------
 624 void PhaseTransform::dump_types( ) const {
 625   _types.dump();
 626 }
 627 
 628 //------------------------------dump_nodes_and_types---------------------------
 629 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) {
 630   VectorSet visited(Thread::current()->resource_area());
 631   dump_nodes_and_types_recur( root, depth, only_ctrl, visited );
 632 }
 633 
 634 //------------------------------dump_nodes_and_types_recur---------------------
 635 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
 636   if( !n ) return;
 637   if( depth == 0 ) return;
 638   if( visited.test_set(n->_idx) ) return;
 639   for( uint i=0; i<n->len(); i++ ) {
 640     if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
 641     dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
 642   }
 643   n->dump();
 644   if (type_or_null(n) != NULL) {
 645     tty->print("      "); type(n)->dump(); tty->cr();
 646   }
 647 }
 648 
 649 #endif
 650 
 651 
 652 //=============================================================================
 653 //------------------------------PhaseValues------------------------------------
 654 // Set minimum table size to "255"
 655 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) {
 656   NOT_PRODUCT( clear_new_values(); )
 657 }
 658 
 659 //------------------------------PhaseValues------------------------------------
 660 // Set minimum table size to "255"
 661 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ),
 662   _table(&ptv->_table) {
 663   NOT_PRODUCT( clear_new_values(); )
 664 }
 665 
 666 //------------------------------PhaseValues------------------------------------
 667 // Used by +VerifyOpto.  Clear out hash table but copy _types array.
 668 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ),
 669   _table(ptv->arena(),ptv->_table.size()) {
 670   NOT_PRODUCT( clear_new_values(); )
 671 }
 672 
 673 //------------------------------~PhaseValues-----------------------------------
 674 #ifndef PRODUCT
 675 PhaseValues::~PhaseValues() {
 676   _table.dump();
 677 
 678   // Statistics for value progress and efficiency
 679   if( PrintCompilation && Verbose && WizardMode ) {
 680     tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
 681       is_IterGVN() ? "Iter" : "    ", C->unique(), made_progress(), made_transforms(), made_new_values());
 682     if( made_transforms() != 0 ) {
 683       tty->print_cr("  ratio %f", made_progress()/(float)made_transforms() );
 684     } else {
 685       tty->cr();
 686     }
 687   }
 688 }
 689 #endif
 690 
 691 //------------------------------makecon----------------------------------------
 692 ConNode* PhaseTransform::makecon(const Type *t) {
 693   assert(t->singleton(), "must be a constant");
 694   assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
 695   switch (t->base()) {  // fast paths
 696   case Type::Half:
 697   case Type::Top:  return (ConNode*) C->top();
 698   case Type::Int:  return intcon( t->is_int()->get_con() );
 699   case Type::Long: return longcon( t->is_long()->get_con() );
 700   default:         break;
 701   }
 702   if (t->is_zero_type())
 703     return zerocon(t->basic_type());
 704   return uncached_makecon(t);
 705 }
 706 
 707 //--------------------------uncached_makecon-----------------------------------
 708 // Make an idealized constant - one of ConINode, ConPNode, etc.
 709 ConNode* PhaseValues::uncached_makecon(const Type *t) {
 710   assert(t->singleton(), "must be a constant");
 711   ConNode* x = ConNode::make(t);
 712   ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
 713   if (k == NULL) {
 714     set_type(x, t);             // Missed, provide type mapping
 715     GrowableArray<Node_Notes*>* nna = C->node_note_array();
 716     if (nna != NULL) {
 717       Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
 718       loc->clear(); // do not put debug info on constants
 719     }
 720   } else {
 721     x->destruct();              // Hit, destroy duplicate constant
 722     x = k;                      // use existing constant
 723   }
 724   return x;
 725 }
 726 
 727 //------------------------------intcon-----------------------------------------
 728 // Fast integer constant.  Same as "transform(new ConINode(TypeInt::make(i)))"
 729 ConINode* PhaseTransform::intcon(jint i) {
 730   // Small integer?  Check cache! Check that cached node is not dead
 731   if (i >= _icon_min && i <= _icon_max) {
 732     ConINode* icon = _icons[i-_icon_min];
 733     if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
 734       return icon;
 735   }
 736   ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
 737   assert(icon->is_Con(), "");
 738   if (i >= _icon_min && i <= _icon_max)
 739     _icons[i-_icon_min] = icon;   // Cache small integers
 740   return icon;
 741 }
 742 
 743 //------------------------------longcon----------------------------------------
 744 // Fast long constant.
 745 ConLNode* PhaseTransform::longcon(jlong l) {
 746   // Small integer?  Check cache! Check that cached node is not dead
 747   if (l >= _lcon_min && l <= _lcon_max) {
 748     ConLNode* lcon = _lcons[l-_lcon_min];
 749     if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
 750       return lcon;
 751   }
 752   ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
 753   assert(lcon->is_Con(), "");
 754   if (l >= _lcon_min && l <= _lcon_max)
 755     _lcons[l-_lcon_min] = lcon;      // Cache small integers
 756   return lcon;
 757 }
 758 
 759 //------------------------------zerocon-----------------------------------------
 760 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
 761 ConNode* PhaseTransform::zerocon(BasicType bt) {
 762   assert((uint)bt <= _zcon_max, "domain check");
 763   ConNode* zcon = _zcons[bt];
 764   if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
 765     return zcon;
 766   zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
 767   _zcons[bt] = zcon;
 768   return zcon;
 769 }
 770 
 771 
 772 
 773 //=============================================================================
 774 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
 775   Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
 776   if (i == NULL) {
 777     i = k->Ideal(this, can_reshape);
 778   }
 779   return i;
 780 }
 781 
 782 Node* PhaseGVN::apply_identity(Node* k) {
 783   Node* i = BarrierSet::barrier_set()->barrier_set_c2()->identity_node(this, k);
 784   if (i == k) {
 785     i = k->Identity(this);
 786   }
 787   return i;
 788 }
 789 
 790 //------------------------------transform--------------------------------------
 791 // Return a node which computes the same function as this node, but in a
 792 // faster or cheaper fashion.
 793 Node *PhaseGVN::transform( Node *n ) {
 794   return transform_no_reclaim(n);
 795 }
 796 
 797 //------------------------------transform--------------------------------------
 798 // Return a node which computes the same function as this node, but
 799 // in a faster or cheaper fashion.
 800 Node *PhaseGVN::transform_no_reclaim( Node *n ) {
 801   NOT_PRODUCT( set_transforms(); )
 802 
 803   // Apply the Ideal call in a loop until it no longer applies
 804   Node *k = n;
 805   NOT_PRODUCT( uint loop_count = 0; )
 806   while( 1 ) {
 807     Node *i = apply_ideal(k, /*can_reshape=*/false);
 808     if( !i ) break;
 809     assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
 810     k = i;
 811     assert(loop_count++ < K, "infinite loop in PhaseGVN::transform");
 812   }
 813   NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } )
 814 
 815 
 816   // If brand new node, make space in type array.
 817   ensure_type_or_null(k);
 818 
 819   // Since I just called 'Value' to compute the set of run-time values
 820   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
 821   // cache Value.  Later requests for the local phase->type of this Node can
 822   // use the cached Value instead of suffering with 'bottom_type'.
 823   const Type *t = k->Value(this); // Get runtime Value set
 824   assert(t != NULL, "value sanity");
 825   if (type_or_null(k) != t) {
 826 #ifndef PRODUCT
 827     // Do not count initial visit to node as a transformation
 828     if (type_or_null(k) == NULL) {
 829       inc_new_values();
 830       set_progress();
 831     }
 832 #endif
 833     set_type(k, t);
 834     // If k is a TypeNode, capture any more-precise type permanently into Node
 835     k->raise_bottom_type(t);
 836   }
 837 
 838   if( t->singleton() && !k->is_Con() ) {
 839     NOT_PRODUCT( set_progress(); )
 840     return makecon(t);          // Turn into a constant
 841   }
 842 
 843   // Now check for Identities
 844   Node *i = apply_identity(k);  // Look for a nearby replacement
 845   if( i != k ) {                // Found? Return replacement!
 846     NOT_PRODUCT( set_progress(); )
 847     return i;
 848   }
 849 
 850   // Global Value Numbering
 851   i = hash_find_insert(k);      // Insert if new
 852   if( i && (i != k) ) {
 853     // Return the pre-existing node
 854     NOT_PRODUCT( set_progress(); )
 855     return i;
 856   }
 857 
 858   // Return Idealized original
 859   return k;
 860 }
 861 
 862 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
 863   if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
 864     return false;
 865   }
 866   if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
 867     return false;
 868   }
 869   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
 870   int i = 0;
 871   while (d != n) {
 872     n = IfNode::up_one_dom(n, linear_only);
 873     i++;
 874     if (n == NULL || i >= 10) {
 875       return false;
 876     }
 877   }
 878   return true;
 879 }
 880 
 881 #ifdef ASSERT
 882 //------------------------------dead_loop_check--------------------------------
 883 // Check for a simple dead loop when a data node references itself directly
 884 // or through an other data node excluding cons and phis.
 885 void PhaseGVN::dead_loop_check( Node *n ) {
 886   // Phi may reference itself in a loop
 887   if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
 888     // Do 2 levels check and only data inputs.
 889     bool no_dead_loop = true;
 890     uint cnt = n->req();
 891     for (uint i = 1; i < cnt && no_dead_loop; i++) {
 892       Node *in = n->in(i);
 893       if (in == n) {
 894         no_dead_loop = false;
 895       } else if (in != NULL && !in->is_dead_loop_safe()) {
 896         uint icnt = in->req();
 897         for (uint j = 1; j < icnt && no_dead_loop; j++) {
 898           if (in->in(j) == n || in->in(j) == in)
 899             no_dead_loop = false;
 900         }
 901       }
 902     }
 903     if (!no_dead_loop) n->dump(3);
 904     assert(no_dead_loop, "dead loop detected");
 905   }
 906 }
 907 #endif
 908 
 909 //=============================================================================
 910 //------------------------------PhaseIterGVN-----------------------------------
 911 // Initialize hash table to fresh and clean for +VerifyOpto
 912 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ),
 913                                                                       _stack(C->live_nodes() >> 1),
 914                                                                       _delay_transform(false) {
 915 }
 916 
 917 //------------------------------PhaseIterGVN-----------------------------------
 918 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
 919 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn),
 920                                                    _worklist( igvn->_worklist ),
 921                                                    _stack( igvn->_stack ),
 922                                                    _delay_transform(igvn->_delay_transform)
 923 {
 924 }
 925 
 926 //------------------------------PhaseIterGVN-----------------------------------
 927 // Initialize with previous PhaseGVN info from Parser
 928 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn),
 929                                               _worklist(*C->for_igvn()),
 930 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that
 931 //       the constructor is used in incremental inlining, this consumes too much memory:
 932 //                                            _stack(C->live_nodes() >> 1),
 933 //       So, as a band-aid, we replace this by:
 934                                               _stack(C->comp_arena(), 32),
 935                                               _delay_transform(false)
 936 {
 937   uint max;
 938 
 939   // Dead nodes in the hash table inherited from GVN were not treated as
 940   // roots during def-use info creation; hence they represent an invisible
 941   // use.  Clear them out.
 942   max = _table.size();
 943   for( uint i = 0; i < max; ++i ) {
 944     Node *n = _table.at(i);
 945     if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
 946       if( n->is_top() ) continue;
 947       assert( false, "Parse::remove_useless_nodes missed this node");
 948       hash_delete(n);
 949     }
 950   }
 951 
 952   // Any Phis or Regions on the worklist probably had uses that could not
 953   // make more progress because the uses were made while the Phis and Regions
 954   // were in half-built states.  Put all uses of Phis and Regions on worklist.
 955   max = _worklist.size();
 956   for( uint j = 0; j < max; j++ ) {
 957     Node *n = _worklist.at(j);
 958     uint uop = n->Opcode();
 959     if( uop == Op_Phi || uop == Op_Region ||
 960         n->is_Type() ||
 961         n->is_Mem() )
 962       add_users_to_worklist(n);
 963   }
 964 
 965   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 966   bs->add_users_to_worklist(&_worklist);
 967 }
 968 
 969 /**
 970  * Initialize worklist for each node.
 971  */
 972 void PhaseIterGVN::init_worklist(Node* first) {
 973   Unique_Node_List to_process;
 974   to_process.push(first);
 975 
 976   while (to_process.size() > 0) {
 977     Node* n = to_process.pop();
 978     if (!_worklist.member(n)) {
 979       _worklist.push(n);
 980 
 981       uint cnt = n->req();
 982       for(uint i = 0; i < cnt; i++) {
 983         Node* m = n->in(i);
 984         if (m != NULL) {
 985           to_process.push(m);
 986         }
 987       }
 988     }
 989   }
 990 }
 991 
 992 #ifndef PRODUCT
 993 void PhaseIterGVN::verify_step(Node* n) {
 994   if (VerifyIterativeGVN) {
 995     _verify_window[_verify_counter % _verify_window_size] = n;
 996     ++_verify_counter;
 997     ResourceMark rm;
 998     ResourceArea* area = Thread::current()->resource_area();
 999     VectorSet old_space(area), new_space(area);
1000     if (C->unique() < 1000 ||
1001         0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
1002       ++_verify_full_passes;
1003       Node::verify_recur(C->root(), -1, old_space, new_space);
1004     }
1005     const int verify_depth = 4;
1006     for ( int i = 0; i < _verify_window_size; i++ ) {
1007       Node* n = _verify_window[i];
1008       if ( n == NULL )  continue;
1009       if( n->in(0) == NodeSentinel ) {  // xform_idom
1010         _verify_window[i] = n->in(1);
1011         --i; continue;
1012       }
1013       // Typical fanout is 1-2, so this call visits about 6 nodes.
1014       Node::verify_recur(n, verify_depth, old_space, new_space);
1015     }
1016   }
1017 }
1018 
1019 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
1020   if (TraceIterativeGVN) {
1021     uint wlsize = _worklist.size();
1022     const Type* newtype = type_or_null(n);
1023     if (nn != n) {
1024       // print old node
1025       tty->print("< ");
1026       if (oldtype != newtype && oldtype != NULL) {
1027         oldtype->dump();
1028       }
1029       do { tty->print("\t"); } while (tty->position() < 16);
1030       tty->print("<");
1031       n->dump();
1032     }
1033     if (oldtype != newtype || nn != n) {
1034       // print new node and/or new type
1035       if (oldtype == NULL) {
1036         tty->print("* ");
1037       } else if (nn != n) {
1038         tty->print("> ");
1039       } else {
1040         tty->print("= ");
1041       }
1042       if (newtype == NULL) {
1043         tty->print("null");
1044       } else {
1045         newtype->dump();
1046       }
1047       do { tty->print("\t"); } while (tty->position() < 16);
1048       nn->dump();
1049     }
1050     if (Verbose && wlsize < _worklist.size()) {
1051       tty->print("  Push {");
1052       while (wlsize != _worklist.size()) {
1053         Node* pushed = _worklist.at(wlsize++);
1054         tty->print(" %d", pushed->_idx);
1055       }
1056       tty->print_cr(" }");
1057     }
1058     if (nn != n) {
1059       // ignore n, it might be subsumed
1060       verify_step((Node*) NULL);
1061     }
1062   }
1063 }
1064 
1065 void PhaseIterGVN::init_verifyPhaseIterGVN() {
1066   _verify_counter = 0;
1067   _verify_full_passes = 0;
1068   for (int i = 0; i < _verify_window_size; i++) {
1069     _verify_window[i] = NULL;
1070   }
1071 #ifdef ASSERT
1072   // Verify that all modified nodes are on _worklist
1073   Unique_Node_List* modified_list = C->modified_nodes();
1074   while (modified_list != NULL && modified_list->size()) {
1075     Node* n = modified_list->pop();
1076     if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) {
1077       n->dump();
1078       assert(false, "modified node is not on IGVN._worklist");
1079     }
1080   }
1081 #endif
1082 }
1083 
1084 void PhaseIterGVN::verify_PhaseIterGVN() {
1085 #ifdef ASSERT
1086   // Verify nodes with changed inputs.
1087   Unique_Node_List* modified_list = C->modified_nodes();
1088   while (modified_list != NULL && modified_list->size()) {
1089     Node* n = modified_list->pop();
1090     if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes
1091       n->dump();
1092       assert(false, "modified node was not processed by IGVN.transform_old()");
1093     }
1094   }
1095 #endif
1096 
1097   C->verify_graph_edges();
1098   if( VerifyOpto && allow_progress() ) {
1099     // Must turn off allow_progress to enable assert and break recursion
1100     C->root()->verify();
1101     { // Check if any progress was missed using IterGVN
1102       // Def-Use info enables transformations not attempted in wash-pass
1103       // e.g. Region/Phi cleanup, ...
1104       // Null-check elision -- may not have reached fixpoint
1105       //                       do not propagate to dominated nodes
1106       ResourceMark rm;
1107       PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean!
1108       // Fill worklist completely
1109       igvn2.init_worklist(C->root());
1110 
1111       igvn2.set_allow_progress(false);
1112       igvn2.optimize();
1113       igvn2.set_allow_progress(true);
1114     }
1115   }
1116   if (VerifyIterativeGVN && PrintOpto) {
1117     if (_verify_counter == _verify_full_passes) {
1118       tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1119                     (int) _verify_full_passes);
1120     } else {
1121       tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1122                   (int) _verify_counter, (int) _verify_full_passes);
1123     }
1124   }
1125 
1126 #ifdef ASSERT
1127   while (modified_list->size()) {
1128     Node* n = modified_list->pop();
1129     n->dump();
1130     assert(false, "VerifyIterativeGVN: new modified node was added");
1131   }
1132 #endif
1133 }
1134 #endif /* PRODUCT */
1135 
1136 #ifdef ASSERT
1137 /**
1138  * Dumps information that can help to debug the problem. A debug
1139  * build fails with an assert.
1140  */
1141 void PhaseIterGVN::dump_infinite_loop_info(Node* n) {
1142   n->dump(4);
1143   _worklist.dump();
1144   assert(false, "infinite loop in PhaseIterGVN::optimize");
1145 }
1146 
1147 /**
1148  * Prints out information about IGVN if the 'verbose' option is used.
1149  */
1150 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1151   if (TraceIterativeGVN && Verbose) {
1152     tty->print("  Pop ");
1153     n->dump();
1154     if ((num_processed % 100) == 0) {
1155       _worklist.print_set();
1156     }
1157   }
1158 }
1159 #endif /* ASSERT */
1160 
1161 void PhaseIterGVN::optimize() {
1162   DEBUG_ONLY(uint num_processed  = 0;)
1163   NOT_PRODUCT(init_verifyPhaseIterGVN();)
1164 
1165   uint loop_count = 0;
1166   // Pull from worklist and transform the node. If the node has changed,
1167   // update edge info and put uses on worklist.
1168   while(_worklist.size()) {
1169     if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1170       return;
1171     }
1172     Node* n  = _worklist.pop();
1173     if (++loop_count >= K * C->live_nodes()) {
1174       DEBUG_ONLY(dump_infinite_loop_info(n);)
1175       C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1176       return;
1177     }
1178     DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1179     if (n->outcnt() != 0) {
1180       NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1181       // Do the transformation
1182       Node* nn = transform_old(n);
1183       NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1184     } else if (!n->is_top()) {
1185       remove_dead_node(n);
1186     }
1187   }
1188   NOT_PRODUCT(verify_PhaseIterGVN();)
1189 }
1190 
1191 
1192 /**
1193  * Register a new node with the optimizer.  Update the types array, the def-use
1194  * info.  Put on worklist.
1195  */
1196 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1197   set_type_bottom(n);
1198   _worklist.push(n);
1199   if (orig != NULL)  C->copy_node_notes_to(n, orig);
1200   return n;
1201 }
1202 
1203 //------------------------------transform--------------------------------------
1204 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
1205 Node *PhaseIterGVN::transform( Node *n ) {
1206   if (_delay_transform) {
1207     // Register the node but don't optimize for now
1208     register_new_node_with_optimizer(n);
1209     return n;
1210   }
1211 
1212   // If brand new node, make space in type array, and give it a type.
1213   ensure_type_or_null(n);
1214   if (type_or_null(n) == NULL) {
1215     set_type_bottom(n);
1216   }
1217 
1218   return transform_old(n);
1219 }
1220 
1221 Node *PhaseIterGVN::transform_old(Node* n) {
1222   DEBUG_ONLY(uint loop_count = 0;);
1223   NOT_PRODUCT(set_transforms());
1224 
1225   // Remove 'n' from hash table in case it gets modified
1226   _table.hash_delete(n);
1227   if (VerifyIterativeGVN) {
1228    assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1229   }
1230 
1231   // Apply the Ideal call in a loop until it no longer applies
1232   Node* k = n;
1233   DEBUG_ONLY(dead_loop_check(k);)
1234   DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1235   C->remove_modified_node(k);
1236   Node* i = apply_ideal(k, /*can_reshape=*/true);
1237   assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1238 #ifndef PRODUCT
1239   verify_step(k);
1240   if (i && VerifyOpto ) {
1241     if (!allow_progress()) {
1242       if (i->is_Add() && (i->outcnt() == 1)) {
1243         // Switched input to left side because this is the only use
1244       } else if (i->is_If() && (i->in(0) == NULL)) {
1245         // This IF is dead because it is dominated by an equivalent IF When
1246         // dominating if changed, info is not propagated sparsely to 'this'
1247         // Propagating this info further will spuriously identify other
1248         // progress.
1249         return i;
1250       } else
1251         set_progress();
1252     } else {
1253       set_progress();
1254     }
1255   }
1256 #endif
1257 
1258   while (i != NULL) {
1259 #ifdef ASSERT
1260     if (loop_count >= K) {
1261       dump_infinite_loop_info(i);
1262     }
1263     loop_count++;
1264 #endif
1265     assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1266     // Made a change; put users of original Node on worklist
1267     add_users_to_worklist(k);
1268     // Replacing root of transform tree?
1269     if (k != i) {
1270       // Make users of old Node now use new.
1271       subsume_node(k, i);
1272       k = i;
1273     }
1274     DEBUG_ONLY(dead_loop_check(k);)
1275     // Try idealizing again
1276     DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1277     C->remove_modified_node(k);
1278     i = apply_ideal(k, /*can_reshape=*/true);
1279     assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1280 #ifndef PRODUCT
1281     verify_step(k);
1282     if (i && VerifyOpto) {
1283       set_progress();
1284     }
1285 #endif
1286   }
1287 
1288   // If brand new node, make space in type array.
1289   ensure_type_or_null(k);
1290 
1291   // See what kind of values 'k' takes on at runtime
1292   const Type* t = k->Value(this);
1293   assert(t != NULL, "value sanity");
1294 
1295   // Since I just called 'Value' to compute the set of run-time values
1296   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1297   // cache Value.  Later requests for the local phase->type of this Node can
1298   // use the cached Value instead of suffering with 'bottom_type'.
1299   if (type_or_null(k) != t) {
1300 #ifndef PRODUCT
1301     inc_new_values();
1302     set_progress();
1303 #endif
1304     set_type(k, t);
1305     // If k is a TypeNode, capture any more-precise type permanently into Node
1306     k->raise_bottom_type(t);
1307     // Move users of node to worklist
1308     add_users_to_worklist(k);
1309   }
1310   // If 'k' computes a constant, replace it with a constant
1311   if (t->singleton() && !k->is_Con()) {
1312     NOT_PRODUCT(set_progress();)
1313     Node* con = makecon(t);     // Make a constant
1314     add_users_to_worklist(k);
1315     subsume_node(k, con);       // Everybody using k now uses con
1316     return con;
1317   }
1318 
1319   // Now check for Identities
1320   i = apply_identity(k);      // Look for a nearby replacement
1321   if (i != k) {                // Found? Return replacement!
1322     NOT_PRODUCT(set_progress();)
1323     add_users_to_worklist(k);
1324     subsume_node(k, i);       // Everybody using k now uses i
1325     return i;
1326   }
1327 
1328   // Global Value Numbering
1329   i = hash_find_insert(k);      // Check for pre-existing node
1330   if (i && (i != k)) {
1331     // Return the pre-existing node if it isn't dead
1332     NOT_PRODUCT(set_progress();)
1333     add_users_to_worklist(k);
1334     subsume_node(k, i);       // Everybody using k now uses i
1335     return i;
1336   }
1337 
1338   // Return Idealized original
1339   return k;
1340 }
1341 
1342 //---------------------------------saturate------------------------------------
1343 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1344                                    const Type* limit_type) const {
1345   return new_type->narrow(old_type);
1346 }
1347 
1348 //------------------------------remove_globally_dead_node----------------------
1349 // Kill a globally dead Node.  All uses are also globally dead and are
1350 // aggressively trimmed.
1351 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1352   enum DeleteProgress {
1353     PROCESS_INPUTS,
1354     PROCESS_OUTPUTS
1355   };
1356   assert(_stack.is_empty(), "not empty");
1357   _stack.push(dead, PROCESS_INPUTS);
1358 
1359   while (_stack.is_nonempty()) {
1360     dead = _stack.node();
1361     uint progress_state = _stack.index();
1362     assert(dead != C->root(), "killing root, eh?");
1363     assert(!dead->is_top(), "add check for top when pushing");
1364     NOT_PRODUCT( set_progress(); )
1365     if (progress_state == PROCESS_INPUTS) {
1366       // After following inputs, continue to outputs
1367       _stack.set_index(PROCESS_OUTPUTS);
1368       if (!dead->is_Con()) { // Don't kill cons but uses
1369         bool recurse = false;
1370         // Remove from hash table
1371         _table.hash_delete( dead );
1372         // Smash all inputs to 'dead', isolating him completely
1373         for (uint i = 0; i < dead->req(); i++) {
1374           Node *in = dead->in(i);
1375           if (in != NULL && in != C->top()) {  // Points to something?
1376             int nrep = dead->replace_edge(in, NULL);  // Kill edges
1377             assert((nrep > 0), "sanity");
1378             if (in->outcnt() == 0) { // Made input go dead?
1379               _stack.push(in, PROCESS_INPUTS); // Recursively remove
1380               recurse = true;
1381             } else if (in->outcnt() == 1 &&
1382                        in->has_special_unique_user()) {
1383               _worklist.push(in->unique_out());
1384             } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1385               if (in->Opcode() == Op_Region) {
1386                 _worklist.push(in);
1387               } else if (in->is_Store()) {
1388                 DUIterator_Fast imax, i = in->fast_outs(imax);
1389                 _worklist.push(in->fast_out(i));
1390                 i++;
1391                 if (in->outcnt() == 2) {
1392                   _worklist.push(in->fast_out(i));
1393                   i++;
1394                 }
1395                 assert(!(i < imax), "sanity");
1396               }
1397             } else {
1398               BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(_worklist, in);
1399             }
1400             if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1401                 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1402               // A Load that directly follows an InitializeNode is
1403               // going away. The Stores that follow are candidates
1404               // again to be captured by the InitializeNode.
1405               for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1406                 Node *n = in->fast_out(j);
1407                 if (n->is_Store()) {
1408                   _worklist.push(n);
1409                 }
1410               }
1411             }
1412           } // if (in != NULL && in != C->top())
1413         } // for (uint i = 0; i < dead->req(); i++)
1414         if (recurse) {
1415           continue;
1416         }
1417       } // if (!dead->is_Con())
1418     } // if (progress_state == PROCESS_INPUTS)
1419 
1420     // Aggressively kill globally dead uses
1421     // (Rather than pushing all the outs at once, we push one at a time,
1422     // plus the parent to resume later, because of the indefinite number
1423     // of edge deletions per loop trip.)
1424     if (dead->outcnt() > 0) {
1425       // Recursively remove output edges
1426       _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1427     } else {
1428       // Finished disconnecting all input and output edges.
1429       _stack.pop();
1430       // Remove dead node from iterative worklist
1431       _worklist.remove(dead);
1432       C->remove_modified_node(dead);
1433       // Constant node that has no out-edges and has only one in-edge from
1434       // root is usually dead. However, sometimes reshaping walk makes
1435       // it reachable by adding use edges. So, we will NOT count Con nodes
1436       // as dead to be conservative about the dead node count at any
1437       // given time.
1438       if (!dead->is_Con()) {
1439         C->record_dead_node(dead->_idx);
1440       }
1441       if (dead->is_macro()) {
1442         C->remove_macro_node(dead);
1443       }
1444       if (dead->is_expensive()) {
1445         C->remove_expensive_node(dead);
1446       }
1447       CastIINode* cast = dead->isa_CastII();
1448       if (cast != NULL && cast->has_range_check()) {
1449         C->remove_range_check_cast(cast);
1450       }
1451       if (dead->Opcode() == Op_Opaque4) {
1452         C->remove_opaque4_node(dead);
1453       }
1454       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1455       bs->unregister_potential_barrier_node(dead);
1456     }
1457   } // while (_stack.is_nonempty())
1458 }
1459 
1460 //------------------------------subsume_node-----------------------------------
1461 // Remove users from node 'old' and add them to node 'nn'.
1462 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1463   assert( old != hash_find(old), "should already been removed" );
1464   assert( old != C->top(), "cannot subsume top node");
1465   // Copy debug or profile information to the new version:
1466   C->copy_node_notes_to(nn, old);
1467   // Move users of node 'old' to node 'nn'
1468   for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1469     Node* use = old->last_out(i);  // for each use...
1470     // use might need re-hashing (but it won't if it's a new node)
1471     rehash_node_delayed(use);
1472     // Update use-def info as well
1473     // We remove all occurrences of old within use->in,
1474     // so as to avoid rehashing any node more than once.
1475     // The hash table probe swamps any outer loop overhead.
1476     uint num_edges = 0;
1477     for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1478       if (use->in(j) == old) {
1479         use->set_req(j, nn);
1480         ++num_edges;
1481       }
1482     }
1483     i -= num_edges;    // we deleted 1 or more copies of this edge
1484   }
1485 
1486   // Search for instance field data PhiNodes in the same region pointing to the old
1487   // memory PhiNode and update their instance memory ids to point to the new node.
1488   if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) {
1489     Node* region = old->in(0);
1490     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1491       PhiNode* phi = region->fast_out(i)->isa_Phi();
1492       if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) {
1493         phi->set_inst_mem_id((int)nn->_idx);
1494       }
1495     }
1496   }
1497 
1498   // Smash all inputs to 'old', isolating him completely
1499   Node *temp = new Node(1);
1500   temp->init_req(0,nn);     // Add a use to nn to prevent him from dying
1501   remove_dead_node( old );
1502   temp->del_req(0);         // Yank bogus edge
1503 #ifndef PRODUCT
1504   if( VerifyIterativeGVN ) {
1505     for ( int i = 0; i < _verify_window_size; i++ ) {
1506       if ( _verify_window[i] == old )
1507         _verify_window[i] = nn;
1508     }
1509   }
1510 #endif
1511   _worklist.remove(temp);   // this can be necessary
1512   temp->destruct();         // reuse the _idx of this little guy
1513 }
1514 
1515 //------------------------------add_users_to_worklist--------------------------
1516 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1517   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1518     _worklist.push(n->fast_out(i));  // Push on worklist
1519   }
1520 }
1521 
1522 // Return counted loop Phi if as a counted loop exit condition, cmp
1523 // compares the the induction variable with n
1524 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) {
1525   for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1526     Node* bol = cmp->fast_out(i);
1527     for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1528       Node* iff = bol->fast_out(i2);
1529       if (iff->is_CountedLoopEnd()) {
1530         CountedLoopEndNode* cle = iff->as_CountedLoopEnd();
1531         if (cle->limit() == n) {
1532           PhiNode* phi = cle->phi();
1533           if (phi != NULL) {
1534             return phi;
1535           }
1536         }
1537       }
1538     }
1539   }
1540   return NULL;
1541 }
1542 
1543 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1544   add_users_to_worklist0(n);
1545 
1546   // Move users of node to worklist
1547   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1548     Node* use = n->fast_out(i); // Get use
1549 
1550     if( use->is_Multi() ||      // Multi-definer?  Push projs on worklist
1551         use->is_Store() )       // Enable store/load same address
1552       add_users_to_worklist0(use);
1553 
1554     // If we changed the receiver type to a call, we need to revisit
1555     // the Catch following the call.  It's looking for a non-NULL
1556     // receiver to know when to enable the regular fall-through path
1557     // in addition to the NullPtrException path.
1558     if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1559       Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1560       if (p != NULL) {
1561         add_users_to_worklist0(p);
1562       }
1563     }
1564 
1565     uint use_op = use->Opcode();
1566     if(use->is_Cmp()) {       // Enable CMP/BOOL optimization
1567       add_users_to_worklist(use); // Put Bool on worklist
1568       if (use->outcnt() > 0) {
1569         Node* bol = use->raw_out(0);
1570         if (bol->outcnt() > 0) {
1571           Node* iff = bol->raw_out(0);
1572           if (iff->outcnt() == 2) {
1573             // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1574             // phi merging either 0 or 1 onto the worklist
1575             Node* ifproj0 = iff->raw_out(0);
1576             Node* ifproj1 = iff->raw_out(1);
1577             if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1578               Node* region0 = ifproj0->raw_out(0);
1579               Node* region1 = ifproj1->raw_out(0);
1580               if( region0 == region1 )
1581                 add_users_to_worklist0(region0);
1582             }
1583           }
1584         }
1585       }
1586       if (use_op == Op_CmpI) {
1587         Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n);
1588         if (phi != NULL) {
1589           // If an opaque node feeds into the limit condition of a
1590           // CountedLoop, we need to process the Phi node for the
1591           // induction variable when the opaque node is removed:
1592           // the range of values taken by the Phi is now known and
1593           // so its type is also known.
1594           _worklist.push(phi);
1595         }
1596         Node* in1 = use->in(1);
1597         for (uint i = 0; i < in1->outcnt(); i++) {
1598           if (in1->raw_out(i)->Opcode() == Op_CastII) {
1599             Node* castii = in1->raw_out(i);
1600             if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) {
1601               Node* ifnode = castii->in(0)->in(0);
1602               if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) {
1603                 // Reprocess a CastII node that may depend on an
1604                 // opaque node value when the opaque node is
1605                 // removed. In case it carries a dependency we can do
1606                 // a better job of computing its type.
1607                 _worklist.push(castii);
1608               }
1609             }
1610           }
1611         }
1612       }
1613     }
1614 
1615     // If changed Cast input, check Phi users for simple cycles
1616     if (use->is_ConstraintCast()) {
1617       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1618         Node* u = use->fast_out(i2);
1619         if (u->is_Phi())
1620           _worklist.push(u);
1621       }
1622     }
1623     // If changed LShift inputs, check RShift users for useless sign-ext
1624     if( use_op == Op_LShiftI ) {
1625       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1626         Node* u = use->fast_out(i2);
1627         if (u->Opcode() == Op_RShiftI)
1628           _worklist.push(u);
1629       }
1630     }
1631     // If changed AddI/SubI inputs, check CmpU for range check optimization.
1632     if (use_op == Op_AddI || use_op == Op_SubI) {
1633       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1634         Node* u = use->fast_out(i2);
1635         if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1636           _worklist.push(u);
1637         }
1638       }
1639     }
1640     // If changed AddP inputs, check Stores for loop invariant
1641     if( use_op == Op_AddP ) {
1642       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1643         Node* u = use->fast_out(i2);
1644         if (u->is_Mem())
1645           _worklist.push(u);
1646       }
1647     }
1648     // If changed initialization activity, check dependent Stores
1649     if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1650       InitializeNode* init = use->as_Allocate()->initialization();
1651       if (init != NULL) {
1652         Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1653         if (imem != NULL)  add_users_to_worklist0(imem);
1654       }
1655     }
1656     if (use_op == Op_Initialize) {
1657       Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1658       if (imem != NULL)  add_users_to_worklist0(imem);
1659     }
1660     // Loading the java mirror from a Klass requires two loads and the type
1661     // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1662     //   LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
1663     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1664     bool has_load_barriers = bs->has_load_barriers();
1665 
1666     if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1667       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1668         Node* u = use->fast_out(i2);
1669         const Type* ut = u->bottom_type();
1670         if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1671           if (has_load_barriers) {
1672             // Search for load barriers behind the load
1673             for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1674               Node* b = u->fast_out(i3);
1675               if (bs->is_gc_barrier_node(b)) {
1676                 _worklist.push(b);
1677               }
1678             }
1679           }
1680           _worklist.push(u);
1681         }
1682       }
1683     }
1684 
1685     // TODO: Needed after the block above?
1686     if (use->is_ShenandoahBarrier()) {
1687       Node* cmp = use->find_out_with(Op_CmpP);
1688       if (cmp != NULL) {
1689         _worklist.push(cmp);
1690       }
1691     }
1692   }
1693 }
1694 
1695 /**
1696  * Remove the speculative part of all types that we know of
1697  */
1698 void PhaseIterGVN::remove_speculative_types()  {
1699   assert(UseTypeSpeculation, "speculation is off");
1700   for (uint i = 0; i < _types.Size(); i++)  {
1701     const Type* t = _types.fast_lookup(i);
1702     if (t != NULL) {
1703       _types.map(i, t->remove_speculative());
1704     }
1705   }
1706   _table.check_no_speculative_types();
1707 }
1708 
1709 //=============================================================================
1710 #ifndef PRODUCT
1711 uint PhaseCCP::_total_invokes   = 0;
1712 uint PhaseCCP::_total_constants = 0;
1713 #endif
1714 //------------------------------PhaseCCP---------------------------------------
1715 // Conditional Constant Propagation, ala Wegman & Zadeck
1716 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1717   NOT_PRODUCT( clear_constants(); )
1718   assert( _worklist.size() == 0, "" );
1719   // Clear out _nodes from IterGVN.  Must be clear to transform call.
1720   _nodes.clear();               // Clear out from IterGVN
1721   analyze();
1722 }
1723 
1724 #ifndef PRODUCT
1725 //------------------------------~PhaseCCP--------------------------------------
1726 PhaseCCP::~PhaseCCP() {
1727   inc_invokes();
1728   _total_constants += count_constants();
1729 }
1730 #endif
1731 
1732 
1733 #ifdef ASSERT
1734 static bool ccp_type_widens(const Type* t, const Type* t0) {
1735   assert(t->meet(t0) == t, "Not monotonic");
1736   switch (t->base() == t0->base() ? t->base() : Type::Top) {
1737   case Type::Int:
1738     assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1739     break;
1740   case Type::Long:
1741     assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1742     break;
1743   default:
1744     break;
1745   }
1746   return true;
1747 }
1748 #endif //ASSERT
1749 
1750 //------------------------------analyze----------------------------------------
1751 void PhaseCCP::analyze() {
1752   // Initialize all types to TOP, optimistic analysis
1753   for (int i = C->unique() - 1; i >= 0; i--)  {
1754     _types.map(i,Type::TOP);
1755   }
1756 
1757   // Push root onto worklist
1758   Unique_Node_List worklist;
1759   worklist.push(C->root());
1760 
1761   // Pull from worklist; compute new value; push changes out.
1762   // This loop is the meat of CCP.
1763   while( worklist.size() ) {
1764     Node *n = worklist.pop();
1765     const Type *t = n->Value(this);
1766     if (t != type(n)) {
1767       assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1768 #ifndef PRODUCT
1769       if( TracePhaseCCP ) {
1770         t->dump();
1771         do { tty->print("\t"); } while (tty->position() < 16);
1772         n->dump();
1773       }
1774 #endif
1775       set_type(n, t);
1776       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1777         Node* m = n->fast_out(i);   // Get user
1778         if (m->is_Region()) {  // New path to Region?  Must recheck Phis too
1779           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1780             Node* p = m->fast_out(i2); // Propagate changes to uses
1781             if (p->bottom_type() != type(p)) { // If not already bottomed out
1782               worklist.push(p); // Propagate change to user
1783             }
1784           }
1785         }
1786         // If we changed the receiver type to a call, we need to revisit
1787         // the Catch following the call.  It's looking for a non-NULL
1788         // receiver to know when to enable the regular fall-through path
1789         // in addition to the NullPtrException path
1790         if (m->is_Call()) {
1791           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1792             Node* p = m->fast_out(i2);  // Propagate changes to uses
1793             if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) {
1794               worklist.push(p->unique_out());
1795             }
1796           }
1797         }
1798         if (m->bottom_type() != type(m)) { // If not already bottomed out
1799           worklist.push(m);     // Propagate change to user
1800         }
1801 
1802         // CmpU nodes can get their type information from two nodes up in the
1803         // graph (instead of from the nodes immediately above). Make sure they
1804         // are added to the worklist if nodes they depend on are updated, since
1805         // they could be missed and get wrong types otherwise.
1806         uint m_op = m->Opcode();
1807         if (m_op == Op_AddI || m_op == Op_SubI) {
1808           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1809             Node* p = m->fast_out(i2); // Propagate changes to uses
1810             if (p->Opcode() == Op_CmpU) {
1811               // Got a CmpU which might need the new type information from node n.
1812               if(p->bottom_type() != type(p)) { // If not already bottomed out
1813                 worklist.push(p); // Propagate change to user
1814               }
1815             }
1816           }
1817         }
1818         // If n is used in a counted loop exit condition then the type
1819         // of the counted loop's Phi depends on the type of n. See
1820         // PhiNode::Value().
1821         if (m_op == Op_CmpI) {
1822           PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n);
1823           if (phi != NULL) {
1824             worklist.push(phi);
1825           }
1826         }
1827         // Loading the java mirror from a Klass requires two loads and the type
1828         // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1829         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1830         bool has_load_barriers = bs->has_load_barriers();
1831 
1832         if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) {
1833           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1834             Node* u = m->fast_out(i2);
1835             const Type* ut = u->bottom_type();
1836             if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) {
1837               if (has_load_barriers) {
1838                 // Search for load barriers behind the load
1839                 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1840                   Node* b = u->fast_out(i3);
1841                   if (bs->is_gc_barrier_node(b)) {
1842                     _worklist.push(b);
1843                   }
1844                 }
1845               }
1846               worklist.push(u);
1847             }
1848           }
1849         }
1850       }
1851     }
1852   }
1853 }
1854 
1855 //------------------------------do_transform-----------------------------------
1856 // Top level driver for the recursive transformer
1857 void PhaseCCP::do_transform() {
1858   // Correct leaves of new-space Nodes; they point to old-space.
1859   C->set_root( transform(C->root())->as_Root() );
1860   assert( C->top(),  "missing TOP node" );
1861   assert( C->root(), "missing root" );
1862 }
1863 
1864 //------------------------------transform--------------------------------------
1865 // Given a Node in old-space, clone him into new-space.
1866 // Convert any of his old-space children into new-space children.
1867 Node *PhaseCCP::transform( Node *n ) {
1868   Node *new_node = _nodes[n->_idx]; // Check for transformed node
1869   if( new_node != NULL )
1870     return new_node;                // Been there, done that, return old answer
1871   new_node = transform_once(n);     // Check for constant
1872   _nodes.map( n->_idx, new_node );  // Flag as having been cloned
1873 
1874   // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1875   GrowableArray <Node *> trstack(C->live_nodes() >> 1);
1876 
1877   trstack.push(new_node);           // Process children of cloned node
1878   while ( trstack.is_nonempty() ) {
1879     Node *clone = trstack.pop();
1880     uint cnt = clone->req();
1881     for( uint i = 0; i < cnt; i++ ) {          // For all inputs do
1882       Node *input = clone->in(i);
1883       if( input != NULL ) {                    // Ignore NULLs
1884         Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1885         if( new_input == NULL ) {
1886           new_input = transform_once(input);   // Check for constant
1887           _nodes.map( input->_idx, new_input );// Flag as having been cloned
1888           trstack.push(new_input);
1889         }
1890         assert( new_input == clone->in(i), "insanity check");
1891       }
1892     }
1893   }
1894   return new_node;
1895 }
1896 
1897 
1898 //------------------------------transform_once---------------------------------
1899 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
1900 Node *PhaseCCP::transform_once( Node *n ) {
1901   const Type *t = type(n);
1902   // Constant?  Use constant Node instead
1903   if( t->singleton() ) {
1904     Node *nn = n;               // Default is to return the original constant
1905     if( t == Type::TOP ) {
1906       // cache my top node on the Compile instance
1907       if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1908         C->set_cached_top_node(ConNode::make(Type::TOP));
1909         set_type(C->top(), Type::TOP);
1910       }
1911       nn = C->top();
1912     }
1913     if( !n->is_Con() ) {
1914       if( t != Type::TOP ) {
1915         nn = makecon(t);        // ConNode::make(t);
1916         NOT_PRODUCT( inc_constants(); )
1917       } else if( n->is_Region() ) { // Unreachable region
1918         // Note: nn == C->top()
1919         n->set_req(0, NULL);        // Cut selfreference
1920         bool progress = true;
1921         uint max = n->outcnt();
1922         DUIterator i;
1923         while (progress) {
1924           progress = false;
1925           // Eagerly remove dead phis to avoid phis copies creation.
1926           for (i = n->outs(); n->has_out(i); i++) {
1927             Node* m = n->out(i);
1928             if (m->is_Phi()) {
1929               assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1930               replace_node(m, nn);
1931               if (max != n->outcnt()) {
1932                 progress = true;
1933                 i = n->refresh_out_pos(i);
1934                 max = n->outcnt();
1935               }
1936             }
1937           }
1938         }
1939       }
1940       replace_node(n,nn);       // Update DefUse edges for new constant
1941     }
1942     return nn;
1943   }
1944 
1945   // If x is a TypeNode, capture any more-precise type permanently into Node
1946   if (t != n->bottom_type()) {
1947     hash_delete(n);             // changing bottom type may force a rehash
1948     n->raise_bottom_type(t);
1949     _worklist.push(n);          // n re-enters the hash table via the worklist
1950   }
1951 
1952   // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
1953   switch( n->Opcode() ) {
1954   case Op_FastLock:      // Revisit FastLocks for lock coarsening
1955   case Op_If:
1956   case Op_CountedLoopEnd:
1957   case Op_Region:
1958   case Op_Loop:
1959   case Op_CountedLoop:
1960   case Op_Conv2B:
1961   case Op_Opaque1:
1962   case Op_Opaque2:
1963     _worklist.push(n);
1964     break;
1965   default:
1966     break;
1967   }
1968 
1969   return  n;
1970 }
1971 
1972 //---------------------------------saturate------------------------------------
1973 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
1974                                const Type* limit_type) const {
1975   const Type* wide_type = new_type->widen(old_type, limit_type);
1976   if (wide_type != new_type) {          // did we widen?
1977     // If so, we may have widened beyond the limit type.  Clip it back down.
1978     new_type = wide_type->filter(limit_type);
1979   }
1980   return new_type;
1981 }
1982 
1983 //------------------------------print_statistics-------------------------------
1984 #ifndef PRODUCT
1985 void PhaseCCP::print_statistics() {
1986   tty->print_cr("CCP: %d  constants found: %d", _total_invokes, _total_constants);
1987 }
1988 #endif
1989 
1990 
1991 //=============================================================================
1992 #ifndef PRODUCT
1993 uint PhasePeephole::_total_peepholes = 0;
1994 #endif
1995 //------------------------------PhasePeephole----------------------------------
1996 // Conditional Constant Propagation, ala Wegman & Zadeck
1997 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
1998   : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
1999   NOT_PRODUCT( clear_peepholes(); )
2000 }
2001 
2002 #ifndef PRODUCT
2003 //------------------------------~PhasePeephole---------------------------------
2004 PhasePeephole::~PhasePeephole() {
2005   _total_peepholes += count_peepholes();
2006 }
2007 #endif
2008 
2009 //------------------------------transform--------------------------------------
2010 Node *PhasePeephole::transform( Node *n ) {
2011   ShouldNotCallThis();
2012   return NULL;
2013 }
2014 
2015 //------------------------------do_transform-----------------------------------
2016 void PhasePeephole::do_transform() {
2017   bool method_name_not_printed = true;
2018 
2019   // Examine each basic block
2020   for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
2021     Block* block = _cfg.get_block(block_number);
2022     bool block_not_printed = true;
2023 
2024     // and each instruction within a block
2025     uint end_index = block->number_of_nodes();
2026     // block->end_idx() not valid after PhaseRegAlloc
2027     for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
2028       Node     *n = block->get_node(instruction_index);
2029       if( n->is_Mach() ) {
2030         MachNode *m = n->as_Mach();
2031         int deleted_count = 0;
2032         // check for peephole opportunities
2033         MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count);
2034         if( m2 != NULL ) {
2035 #ifndef PRODUCT
2036           if( PrintOptoPeephole ) {
2037             // Print method, first time only
2038             if( C->method() && method_name_not_printed ) {
2039               C->method()->print_short_name(); tty->cr();
2040               method_name_not_printed = false;
2041             }
2042             // Print this block
2043             if( Verbose && block_not_printed) {
2044               tty->print_cr("in block");
2045               block->dump();
2046               block_not_printed = false;
2047             }
2048             // Print instructions being deleted
2049             for( int i = (deleted_count - 1); i >= 0; --i ) {
2050               block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
2051             }
2052             tty->print_cr("replaced with");
2053             // Print new instruction
2054             m2->format(_regalloc);
2055             tty->print("\n\n");
2056           }
2057 #endif
2058           // Remove old nodes from basic block and update instruction_index
2059           // (old nodes still exist and may have edges pointing to them
2060           //  as register allocation info is stored in the allocator using
2061           //  the node index to live range mappings.)
2062           uint safe_instruction_index = (instruction_index - deleted_count);
2063           for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
2064             block->remove_node( instruction_index );
2065           }
2066           // install new node after safe_instruction_index
2067           block->insert_node(m2, safe_instruction_index + 1);
2068           end_index = block->number_of_nodes() - 1; // Recompute new block size
2069           NOT_PRODUCT( inc_peepholes(); )
2070         }
2071       }
2072     }
2073   }
2074 }
2075 
2076 //------------------------------print_statistics-------------------------------
2077 #ifndef PRODUCT
2078 void PhasePeephole::print_statistics() {
2079   tty->print_cr("Peephole: peephole rules applied: %d",  _total_peepholes);
2080 }
2081 #endif
2082 
2083 
2084 //=============================================================================
2085 //------------------------------set_req_X--------------------------------------
2086 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2087   assert( is_not_dead(n), "can not use dead node");
2088   assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2089   Node *old = in(i);
2090   set_req(i, n);
2091 
2092   // old goes dead?
2093   if( old ) {
2094     switch (old->outcnt()) {
2095     case 0:
2096       // Put into the worklist to kill later. We do not kill it now because the
2097       // recursive kill will delete the current node (this) if dead-loop exists
2098       if (!old->is_top())
2099         igvn->_worklist.push( old );
2100       break;
2101     case 1:
2102       if( old->is_Store() || old->has_special_unique_user() )
2103         igvn->add_users_to_worklist( old );
2104       break;
2105     case 2:
2106       if( old->is_Store() )
2107         igvn->add_users_to_worklist( old );
2108       if( old->Opcode() == Op_Region )
2109         igvn->_worklist.push(old);
2110       break;
2111     case 3:
2112       if( old->Opcode() == Op_Region ) {
2113         igvn->_worklist.push(old);
2114         igvn->add_users_to_worklist( old );
2115       }
2116       break;
2117     default:
2118       break;
2119     }
2120     if (UseShenandoahGC) {
2121       // TODO: Should we call this for ZGC as well?
2122       BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn->_worklist, old);
2123     }
2124   }
2125 
2126 }
2127 
2128 //-------------------------------replace_by-----------------------------------
2129 // Using def-use info, replace one node for another.  Follow the def-use info
2130 // to all users of the OLD node.  Then make all uses point to the NEW node.
2131 void Node::replace_by(Node *new_node) {
2132   assert(!is_top(), "top node has no DU info");
2133   for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2134     Node* use = last_out(i);
2135     uint uses_found = 0;
2136     for (uint j = 0; j < use->len(); j++) {
2137       if (use->in(j) == this) {
2138         if (j < use->req())
2139               use->set_req(j, new_node);
2140         else  use->set_prec(j, new_node);
2141         uses_found++;
2142       }
2143     }
2144     i -= uses_found;    // we deleted 1 or more copies of this edge
2145   }
2146 }
2147 
2148 //=============================================================================
2149 //-----------------------------------------------------------------------------
2150 void Type_Array::grow( uint i ) {
2151   if( !_max ) {
2152     _max = 1;
2153     _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2154     _types[0] = NULL;
2155   }
2156   uint old = _max;
2157   while( i >= _max ) _max <<= 1;        // Double to fit
2158   _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2159   memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2160 }
2161 
2162 //------------------------------dump-------------------------------------------
2163 #ifndef PRODUCT
2164 void Type_Array::dump() const {
2165   uint max = Size();
2166   for( uint i = 0; i < max; i++ ) {
2167     if( _types[i] != NULL ) {
2168       tty->print("  %d\t== ", i); _types[i]->dump(); tty->cr();
2169     }
2170   }
2171 }
2172 #endif