1 /*
2  * Copyright (c) 1997, 2021, 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 #include "utilities/powerOfTwo.hpp"
43 
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE    255
46 //------------------------------NodeHash---------------------------------------
NodeHash(uint est_max_size)47 NodeHash::NodeHash(uint est_max_size) :
48   _a(Thread::current()->resource_area()),
49   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
50   _inserts(0), _insert_limit( insert_limit() ),
51   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
52 #ifndef PRODUCT
53   , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
54   _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
55    _total_inserts(0), _total_insert_probes(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---------------------------------------
NodeHash(Arena * arena,uint est_max_size)64 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
65   _a(arena),
66   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
67   _inserts(0), _insert_limit( insert_limit() ),
68   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
69 #ifndef PRODUCT
70   , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
71   _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
72    _total_inserts(0), _total_insert_probes(0)
73 #endif
74 {
75   // _sentinel must be in the current node space
76   _sentinel = new ProjNode(NULL, TypeFunc::Control);
77   memset(_table,0,sizeof(Node*)*_max);
78 }
79 
80 //------------------------------NodeHash---------------------------------------
NodeHash(NodeHash * nh)81 NodeHash::NodeHash(NodeHash *nh) {
82   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
83   // just copy in all the fields
84   *this = *nh;
85   // nh->_sentinel must be in the current node space
86 }
87 
replace_with(NodeHash * nh)88 void NodeHash::replace_with(NodeHash *nh) {
89   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
90   // just copy in all the fields
91   *this = *nh;
92   // nh->_sentinel must be in the current node space
93 }
94 
95 //------------------------------hash_find--------------------------------------
96 // Find in hash table
hash_find(const Node * n)97 Node *NodeHash::hash_find( const Node *n ) {
98   // ((Node*)n)->set_hash( n->hash() );
99   uint hash = n->hash();
100   if (hash == Node::NO_HASH) {
101     NOT_PRODUCT( _lookup_misses++ );
102     return NULL;
103   }
104   uint key = hash & (_max-1);
105   uint stride = key | 0x01;
106   NOT_PRODUCT( _look_probes++ );
107   Node *k = _table[key];        // Get hashed value
108   if( !k ) {                    // ?Miss?
109     NOT_PRODUCT( _lookup_misses++ );
110     return NULL;                // Miss!
111   }
112 
113   int op = n->Opcode();
114   uint req = n->req();
115   while( 1 ) {                  // While probing hash table
116     if( k->req() == req &&      // Same count of inputs
117         k->Opcode() == op ) {   // Same Opcode
118       for( uint i=0; i<req; i++ )
119         if( n->in(i)!=k->in(i)) // Different inputs?
120           goto collision;       // "goto" is a speed hack...
121       if( n->cmp(*k) ) {        // Check for any special bits
122         NOT_PRODUCT( _lookup_hits++ );
123         return k;               // Hit!
124       }
125     }
126   collision:
127     NOT_PRODUCT( _look_probes++ );
128     key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
129     k = _table[key];            // Get hashed value
130     if( !k ) {                  // ?Miss?
131       NOT_PRODUCT( _lookup_misses++ );
132       return NULL;              // Miss!
133     }
134   }
135   ShouldNotReachHere();
136   return NULL;
137 }
138 
139 //------------------------------hash_find_insert-------------------------------
140 // Find in hash table, insert if not already present
141 // Used to preserve unique entries in hash table
hash_find_insert(Node * n)142 Node *NodeHash::hash_find_insert( Node *n ) {
143   // n->set_hash( );
144   uint hash = n->hash();
145   if (hash == Node::NO_HASH) {
146     NOT_PRODUCT( _lookup_misses++ );
147     return NULL;
148   }
149   uint key = hash & (_max-1);
150   uint stride = key | 0x01;     // stride must be relatively prime to table siz
151   uint first_sentinel = 0;      // replace a sentinel if seen.
152   NOT_PRODUCT( _look_probes++ );
153   Node *k = _table[key];        // Get hashed value
154   if( !k ) {                    // ?Miss?
155     NOT_PRODUCT( _lookup_misses++ );
156     _table[key] = n;            // Insert into table!
157     debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
158     check_grow();               // Grow table if insert hit limit
159     return NULL;                // Miss!
160   }
161   else if( k == _sentinel ) {
162     first_sentinel = key;      // Can insert here
163   }
164 
165   int op = n->Opcode();
166   uint req = n->req();
167   while( 1 ) {                  // While probing hash table
168     if( k->req() == req &&      // Same count of inputs
169         k->Opcode() == op ) {   // Same Opcode
170       for( uint i=0; i<req; i++ )
171         if( n->in(i)!=k->in(i)) // Different inputs?
172           goto collision;       // "goto" is a speed hack...
173       if( n->cmp(*k) ) {        // Check for any special bits
174         NOT_PRODUCT( _lookup_hits++ );
175         return k;               // Hit!
176       }
177     }
178   collision:
179     NOT_PRODUCT( _look_probes++ );
180     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
181     k = _table[key];            // Get hashed value
182     if( !k ) {                  // ?Miss?
183       NOT_PRODUCT( _lookup_misses++ );
184       key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
185       _table[key] = n;          // Insert into table!
186       debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
187       check_grow();             // Grow table if insert hit limit
188       return NULL;              // Miss!
189     }
190     else if( first_sentinel == 0 && k == _sentinel ) {
191       first_sentinel = key;    // Can insert here
192     }
193 
194   }
195   ShouldNotReachHere();
196   return NULL;
197 }
198 
199 //------------------------------hash_insert------------------------------------
200 // Insert into hash table
hash_insert(Node * n)201 void NodeHash::hash_insert( Node *n ) {
202   // // "conflict" comments -- print nodes that conflict
203   // bool conflict = false;
204   // n->set_hash();
205   uint hash = n->hash();
206   if (hash == Node::NO_HASH) {
207     return;
208   }
209   check_grow();
210   uint key = hash & (_max-1);
211   uint stride = key | 0x01;
212 
213   while( 1 ) {                  // While probing hash table
214     NOT_PRODUCT( _insert_probes++ );
215     Node *k = _table[key];      // Get hashed value
216     if( !k || (k == _sentinel) ) break;       // Found a slot
217     assert( k != n, "already inserted" );
218     // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print("  conflict: "); k->dump(); conflict = true; }
219     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
220   }
221   _table[key] = n;              // Insert into table!
222   debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
223   // if( conflict ) { n->dump(); }
224 }
225 
226 //------------------------------hash_delete------------------------------------
227 // Replace in hash table with sentinel
hash_delete(const Node * n)228 bool NodeHash::hash_delete( const Node *n ) {
229   Node *k;
230   uint hash = n->hash();
231   if (hash == Node::NO_HASH) {
232     NOT_PRODUCT( _delete_misses++ );
233     return false;
234   }
235   uint key = hash & (_max-1);
236   uint stride = key | 0x01;
237   debug_only( uint counter = 0; );
238   for( ; /* (k != NULL) && (k != _sentinel) */; ) {
239     debug_only( counter++ );
240     NOT_PRODUCT( _delete_probes++ );
241     k = _table[key];            // Get hashed value
242     if( !k ) {                  // Miss?
243       NOT_PRODUCT( _delete_misses++ );
244       return false;             // Miss! Not in chain
245     }
246     else if( n == k ) {
247       NOT_PRODUCT( _delete_hits++ );
248       _table[key] = _sentinel;  // Hit! Label as deleted entry
249       debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
250       return true;
251     }
252     else {
253       // collision: move through table with prime offset
254       key = (key + stride/*7*/) & (_max-1);
255       assert( counter <= _insert_limit, "Cycle in hash-table");
256     }
257   }
258   ShouldNotReachHere();
259   return false;
260 }
261 
262 //------------------------------round_up---------------------------------------
263 // Round up to nearest power of 2
round_up(uint x)264 uint NodeHash::round_up(uint x) {
265   x += (x >> 2);                  // Add 25% slop
266   return MAX2(16U, round_up_power_of_2(x));
267 }
268 
269 //------------------------------grow-------------------------------------------
270 // Grow _table to next power of 2 and insert old entries
grow()271 void  NodeHash::grow() {
272   // Record old state
273   uint   old_max   = _max;
274   Node **old_table = _table;
275   // Construct new table with twice the space
276 #ifndef PRODUCT
277   _grows++;
278   _total_inserts       += _inserts;
279   _total_insert_probes += _insert_probes;
280   _insert_probes   = 0;
281 #endif
282   _inserts         = 0;
283   _max     = _max << 1;
284   _table   = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
285   memset(_table,0,sizeof(Node*)*_max);
286   _insert_limit = insert_limit();
287   // Insert old entries into the new table
288   for( uint i = 0; i < old_max; i++ ) {
289     Node *m = *old_table++;
290     if( !m || m == _sentinel ) continue;
291     debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
292     hash_insert(m);
293   }
294 }
295 
296 //------------------------------clear------------------------------------------
297 // Clear all entries in _table to NULL but keep storage
clear()298 void  NodeHash::clear() {
299 #ifdef ASSERT
300   // Unlock all nodes upon removal from table.
301   for (uint i = 0; i < _max; i++) {
302     Node* n = _table[i];
303     if (!n || n == _sentinel)  continue;
304     n->exit_hash_lock();
305   }
306 #endif
307 
308   memset( _table, 0, _max * sizeof(Node*) );
309 }
310 
311 //-----------------------remove_useless_nodes----------------------------------
312 // Remove useless nodes from value table,
313 // implementation does not depend on hash function
remove_useless_nodes(VectorSet & useful)314 void NodeHash::remove_useless_nodes(VectorSet &useful) {
315 
316   // Dead nodes in the hash table inherited from GVN should not replace
317   // existing nodes, remove dead nodes.
318   uint max = size();
319   Node *sentinel_node = sentinel();
320   for( uint i = 0; i < max; ++i ) {
321     Node *n = at(i);
322     if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
323       debug_only(n->exit_hash_lock()); // Unlock the node when removed
324       _table[i] = sentinel_node;       // Replace with placeholder
325     }
326   }
327 }
328 
329 
check_no_speculative_types()330 void NodeHash::check_no_speculative_types() {
331 #ifdef ASSERT
332   uint max = size();
333   Unique_Node_List live_nodes;
334   Compile::current()->identify_useful_nodes(live_nodes);
335   Node *sentinel_node = sentinel();
336   for (uint i = 0; i < max; ++i) {
337     Node *n = at(i);
338     if (n != NULL &&
339         n != sentinel_node &&
340         n->is_Type() &&
341         live_nodes.member(n)) {
342       TypeNode* tn = n->as_Type();
343       const Type* t = tn->type();
344       const Type* t_no_spec = t->remove_speculative();
345       assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
346     }
347   }
348 #endif
349 }
350 
351 #ifndef PRODUCT
352 //------------------------------dump-------------------------------------------
353 // Dump statistics for the hash table
dump()354 void NodeHash::dump() {
355   _total_inserts       += _inserts;
356   _total_insert_probes += _insert_probes;
357   if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
358     if (WizardMode) {
359       for (uint i=0; i<_max; i++) {
360         if (_table[i])
361           tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
362       }
363     }
364     tty->print("\nGVN Hash stats:  %d grows to %d max_size\n", _grows, _max);
365     tty->print("  %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
366     tty->print("  %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
367     tty->print("  %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
368     // sentinels increase lookup cost, but not insert cost
369     assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
370     assert( _inserts+(_inserts>>3) < _max, "table too full" );
371     assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
372   }
373 }
374 
find_index(uint idx)375 Node *NodeHash::find_index(uint idx) { // For debugging
376   // Find an entry by its index value
377   for( uint i = 0; i < _max; i++ ) {
378     Node *m = _table[i];
379     if( !m || m == _sentinel ) continue;
380     if( m->_idx == (uint)idx ) return m;
381   }
382   return NULL;
383 }
384 #endif
385 
386 #ifdef ASSERT
~NodeHash()387 NodeHash::~NodeHash() {
388   // Unlock all nodes upon destruction of table.
389   if (_table != (Node**)badAddress)  clear();
390 }
391 
operator =(const NodeHash & nh)392 void NodeHash::operator=(const NodeHash& nh) {
393   // Unlock all nodes upon replacement of table.
394   if (&nh == this)  return;
395   if (_table != (Node**)badAddress)  clear();
396   memcpy((void*)this, (void*)&nh, sizeof(*this));
397   // Do not increment hash_lock counts again.
398   // Instead, be sure we never again use the source table.
399   ((NodeHash*)&nh)->_table = (Node**)badAddress;
400 }
401 
402 
403 #endif
404 
405 
406 //=============================================================================
407 //------------------------------PhaseRemoveUseless-----------------------------
408 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
PhaseRemoveUseless(PhaseGVN * gvn,Unique_Node_List * worklist,PhaseNumber phase_num)409 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List* worklist, PhaseNumber phase_num) : Phase(phase_num) {
410   // Implementation requires an edge from root to each SafePointNode
411   // at a backward branch. Inserted in add_safepoint().
412 
413   // Identify nodes that are reachable from below, useful.
414   C->identify_useful_nodes(_useful);
415   // Update dead node list
416   C->update_dead_node_list(_useful);
417 
418   // Remove all useless nodes from PhaseValues' recorded types
419   // Must be done before disconnecting nodes to preserve hash-table-invariant
420   gvn->remove_useless_nodes(_useful.member_set());
421 
422   // Remove all useless nodes from future worklist
423   worklist->remove_useless_nodes(_useful.member_set());
424 
425   // Disconnect 'useless' nodes that are adjacent to useful nodes
426   C->remove_useless_nodes(_useful);
427 }
428 
429 //=============================================================================
430 //------------------------------PhaseRenumberLive------------------------------
431 // First, remove useless nodes (equivalent to identifying live nodes).
432 // Then, renumber live nodes.
433 //
434 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
435 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
436 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
437 // value in the range [0, x).
438 //
439 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
440 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
441 //
442 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
443 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be
444 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'.
445 // 'worklist' is cleared upon returning.
446 // (2) Type information (the field PhaseGVN::_types) maps type information to each
447 // node ID. The mapping is updated to use the new node IDs as well. Updated type
448 // information is returned in PhaseGVN::_types.
449 //
450 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist.
451 //
452 // Other data structures used by the compiler are not updated. The hash table for value
453 // numbering (the field PhaseGVN::_table) is not updated because computing the hash
454 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either
455 // because it is empty wherever PhaseRenumberLive is used.
PhaseRenumberLive(PhaseGVN * gvn,Unique_Node_List * worklist,Unique_Node_List * new_worklist,PhaseNumber phase_num)456 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
457                                      Unique_Node_List* worklist, Unique_Node_List* new_worklist,
458                                      PhaseNumber phase_num) :
459   PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
460   _new_type_array(C->comp_arena()),
461   _old2new_map(C->unique(), C->unique(), -1),
462   _is_pass_finished(false),
463   _live_node_count(C->live_nodes())
464 {
465   assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
466   assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
467   assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point");
468   assert(_delayed.size() == 0, "should be empty");
469 
470   uint worklist_size = worklist->size();
471 
472   // Iterate over the set of live nodes.
473   for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
474     Node* n = _useful.at(current_idx);
475 
476     bool in_worklist = false;
477     if (worklist->member(n)) {
478       in_worklist = true;
479     }
480 
481     const Type* type = gvn->type_or_null(n);
482     _new_type_array.map(current_idx, type);
483 
484     assert(_old2new_map.at(n->_idx) == -1, "already seen");
485     _old2new_map.at_put(n->_idx, current_idx);
486 
487     n->set_idx(current_idx); // Update node ID.
488 
489     if (in_worklist) {
490       new_worklist->push(n);
491     }
492 
493     if (update_embedded_ids(n) < 0) {
494       _delayed.push(n); // has embedded IDs; handle later
495     }
496   }
497 
498   assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist");
499   assert(_live_node_count == _useful.size(), "all live nodes must be processed");
500 
501   _is_pass_finished = true; // pass finished; safe to process delayed updates
502 
503   while (_delayed.size() > 0) {
504     Node* n = _delayed.pop();
505     int no_of_updates = update_embedded_ids(n);
506     assert(no_of_updates > 0, "should be updated");
507   }
508 
509   // Replace the compiler's type information with the updated type information.
510   gvn->replace_types(_new_type_array);
511 
512   // Update the unique node count of the compilation to the number of currently live nodes.
513   C->set_unique(_live_node_count);
514 
515   // Set the dead node count to 0 and reset dead node list.
516   C->reset_dead_node_list();
517 
518   // Clear the original worklist
519   worklist->clear();
520 }
521 
new_index(int old_idx)522 int PhaseRenumberLive::new_index(int old_idx) {
523   assert(_is_pass_finished, "not finished");
524   if (_old2new_map.at(old_idx) == -1) { // absent
525     // Allocate a placeholder to preserve uniqueness
526     _old2new_map.at_put(old_idx, _live_node_count);
527     _live_node_count++;
528   }
529   return _old2new_map.at(old_idx);
530 }
531 
update_embedded_ids(Node * n)532 int PhaseRenumberLive::update_embedded_ids(Node* n) {
533   int no_of_updates = 0;
534   if (n->is_Phi()) {
535     PhiNode* phi = n->as_Phi();
536     if (phi->_inst_id != -1) {
537       if (!_is_pass_finished) {
538         return -1; // delay
539       }
540       int new_idx = new_index(phi->_inst_id);
541       assert(new_idx != -1, "");
542       phi->_inst_id = new_idx;
543       no_of_updates++;
544     }
545     if (phi->_inst_mem_id != -1) {
546       if (!_is_pass_finished) {
547         return -1; // delay
548       }
549       int new_idx = new_index(phi->_inst_mem_id);
550       assert(new_idx != -1, "");
551       phi->_inst_mem_id = new_idx;
552       no_of_updates++;
553     }
554   }
555 
556   const Type* type = _new_type_array.fast_lookup(n->_idx);
557   if (type != NULL && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
558     if (!_is_pass_finished) {
559         return -1; // delay
560     }
561     int old_idx = type->is_oopptr()->instance_id();
562     int new_idx = new_index(old_idx);
563     const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
564     _new_type_array.map(n->_idx, new_type);
565     no_of_updates++;
566   }
567 
568   return no_of_updates;
569 }
570 
571 //=============================================================================
572 //------------------------------PhaseTransform---------------------------------
PhaseTransform(PhaseNumber pnum)573 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
574   _arena(Thread::current()->resource_area()),
575   _nodes(_arena),
576   _types(_arena)
577 {
578   init_con_caches();
579 #ifndef PRODUCT
580   clear_progress();
581   clear_transforms();
582   set_allow_progress(true);
583 #endif
584   // Force allocation for currently existing nodes
585   _types.map(C->unique(), NULL);
586 }
587 
588 //------------------------------PhaseTransform---------------------------------
PhaseTransform(Arena * arena,PhaseNumber pnum)589 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
590   _arena(arena),
591   _nodes(arena),
592   _types(arena)
593 {
594   init_con_caches();
595 #ifndef PRODUCT
596   clear_progress();
597   clear_transforms();
598   set_allow_progress(true);
599 #endif
600   // Force allocation for currently existing nodes
601   _types.map(C->unique(), NULL);
602 }
603 
604 //------------------------------PhaseTransform---------------------------------
605 // Initialize with previously generated type information
PhaseTransform(PhaseTransform * pt,PhaseNumber pnum)606 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
607   _arena(pt->_arena),
608   _nodes(pt->_nodes),
609   _types(pt->_types)
610 {
611   init_con_caches();
612 #ifndef PRODUCT
613   clear_progress();
614   clear_transforms();
615   set_allow_progress(true);
616 #endif
617 }
618 
init_con_caches()619 void PhaseTransform::init_con_caches() {
620   memset(_icons,0,sizeof(_icons));
621   memset(_lcons,0,sizeof(_lcons));
622   memset(_zcons,0,sizeof(_zcons));
623 }
624 
625 
626 //--------------------------------find_int_type--------------------------------
find_int_type(Node * n)627 const TypeInt* PhaseTransform::find_int_type(Node* n) {
628   if (n == NULL)  return NULL;
629   // Call type_or_null(n) to determine node's type since we might be in
630   // parse phase and call n->Value() may return wrong type.
631   // (For example, a phi node at the beginning of loop parsing is not ready.)
632   const Type* t = type_or_null(n);
633   if (t == NULL)  return NULL;
634   return t->isa_int();
635 }
636 
637 
638 //-------------------------------find_long_type--------------------------------
find_long_type(Node * n)639 const TypeLong* PhaseTransform::find_long_type(Node* n) {
640   if (n == NULL)  return NULL;
641   // (See comment above on type_or_null.)
642   const Type* t = type_or_null(n);
643   if (t == NULL)  return NULL;
644   return t->isa_long();
645 }
646 
647 
648 #ifndef PRODUCT
dump_old2new_map() const649 void PhaseTransform::dump_old2new_map() const {
650   _nodes.dump();
651 }
652 
dump_new(uint nidx) const653 void PhaseTransform::dump_new( uint nidx ) const {
654   for( uint i=0; i<_nodes.Size(); i++ )
655     if( _nodes[i] && _nodes[i]->_idx == nidx ) {
656       _nodes[i]->dump();
657       tty->cr();
658       tty->print_cr("Old index= %d",i);
659       return;
660     }
661   tty->print_cr("Node %d not found in the new indices", nidx);
662 }
663 
664 //------------------------------dump_types-------------------------------------
dump_types() const665 void PhaseTransform::dump_types( ) const {
666   _types.dump();
667 }
668 
669 //------------------------------dump_nodes_and_types---------------------------
dump_nodes_and_types(const Node * root,uint depth,bool only_ctrl)670 void PhaseTransform::dump_nodes_and_types(const Node* root, uint depth, bool only_ctrl) {
671   VectorSet visited;
672   dump_nodes_and_types_recur(root, depth, only_ctrl, visited);
673 }
674 
675 //------------------------------dump_nodes_and_types_recur---------------------
dump_nodes_and_types_recur(const Node * n,uint depth,bool only_ctrl,VectorSet & visited)676 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
677   if( !n ) return;
678   if( depth == 0 ) return;
679   if( visited.test_set(n->_idx) ) return;
680   for( uint i=0; i<n->len(); i++ ) {
681     if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
682     dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
683   }
684   n->dump();
685   if (type_or_null(n) != NULL) {
686     tty->print("      "); type(n)->dump(); tty->cr();
687   }
688 }
689 
690 #endif
691 
692 
693 //=============================================================================
694 //------------------------------PhaseValues------------------------------------
695 // Set minimum table size to "255"
PhaseValues(Arena * arena,uint est_max_size)696 PhaseValues::PhaseValues( Arena *arena, uint est_max_size )
697   : PhaseTransform(arena, GVN), _table(arena, est_max_size), _iterGVN(false) {
698   NOT_PRODUCT( clear_new_values(); )
699 }
700 
701 //------------------------------PhaseValues------------------------------------
702 // Set minimum table size to "255"
PhaseValues(PhaseValues * ptv)703 PhaseValues::PhaseValues(PhaseValues* ptv)
704   : PhaseTransform(ptv, GVN), _table(&ptv->_table), _iterGVN(false) {
705   NOT_PRODUCT( clear_new_values(); )
706 }
707 
708 //------------------------------~PhaseValues-----------------------------------
709 #ifndef PRODUCT
~PhaseValues()710 PhaseValues::~PhaseValues() {
711   _table.dump();
712 
713   // Statistics for value progress and efficiency
714   if( PrintCompilation && Verbose && WizardMode ) {
715     tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
716       is_IterGVN() ? "Iter" : "    ", C->unique(), made_progress(), made_transforms(), made_new_values());
717     if( made_transforms() != 0 ) {
718       tty->print_cr("  ratio %f", made_progress()/(float)made_transforms() );
719     } else {
720       tty->cr();
721     }
722   }
723 }
724 #endif
725 
726 //------------------------------makecon----------------------------------------
makecon(const Type * t)727 ConNode* PhaseTransform::makecon(const Type *t) {
728   assert(t->singleton(), "must be a constant");
729   assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
730   switch (t->base()) {  // fast paths
731   case Type::Half:
732   case Type::Top:  return (ConNode*) C->top();
733   case Type::Int:  return intcon( t->is_int()->get_con() );
734   case Type::Long: return longcon( t->is_long()->get_con() );
735   default:         break;
736   }
737   if (t->is_zero_type())
738     return zerocon(t->basic_type());
739   return uncached_makecon(t);
740 }
741 
742 //--------------------------uncached_makecon-----------------------------------
743 // Make an idealized constant - one of ConINode, ConPNode, etc.
uncached_makecon(const Type * t)744 ConNode* PhaseValues::uncached_makecon(const Type *t) {
745   assert(t->singleton(), "must be a constant");
746   ConNode* x = ConNode::make(t);
747   ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
748   if (k == NULL) {
749     set_type(x, t);             // Missed, provide type mapping
750     GrowableArray<Node_Notes*>* nna = C->node_note_array();
751     if (nna != NULL) {
752       Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
753       loc->clear(); // do not put debug info on constants
754     }
755   } else {
756     x->destruct(this);          // Hit, destroy duplicate constant
757     x = k;                      // use existing constant
758   }
759   return x;
760 }
761 
762 //------------------------------intcon-----------------------------------------
763 // Fast integer constant.  Same as "transform(new ConINode(TypeInt::make(i)))"
intcon(jint i)764 ConINode* PhaseTransform::intcon(jint i) {
765   // Small integer?  Check cache! Check that cached node is not dead
766   if (i >= _icon_min && i <= _icon_max) {
767     ConINode* icon = _icons[i-_icon_min];
768     if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
769       return icon;
770   }
771   ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
772   assert(icon->is_Con(), "");
773   if (i >= _icon_min && i <= _icon_max)
774     _icons[i-_icon_min] = icon;   // Cache small integers
775   return icon;
776 }
777 
778 //------------------------------longcon----------------------------------------
779 // Fast long constant.
longcon(jlong l)780 ConLNode* PhaseTransform::longcon(jlong l) {
781   // Small integer?  Check cache! Check that cached node is not dead
782   if (l >= _lcon_min && l <= _lcon_max) {
783     ConLNode* lcon = _lcons[l-_lcon_min];
784     if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
785       return lcon;
786   }
787   ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
788   assert(lcon->is_Con(), "");
789   if (l >= _lcon_min && l <= _lcon_max)
790     _lcons[l-_lcon_min] = lcon;      // Cache small integers
791   return lcon;
792 }
integercon(jlong l,BasicType bt)793 ConNode* PhaseTransform::integercon(jlong l, BasicType bt) {
794   if (bt == T_INT) {
795     jint int_con = (jint)l;
796     assert(((long)int_con) == l, "not an int");
797     return intcon(int_con);
798   }
799   assert(bt == T_LONG, "not an integer");
800   return longcon(l);
801 }
802 
803 
804 //------------------------------zerocon-----------------------------------------
805 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
zerocon(BasicType bt)806 ConNode* PhaseTransform::zerocon(BasicType bt) {
807   assert((uint)bt <= _zcon_max, "domain check");
808   ConNode* zcon = _zcons[bt];
809   if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
810     return zcon;
811   zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
812   _zcons[bt] = zcon;
813   return zcon;
814 }
815 
816 
817 
818 //=============================================================================
apply_ideal(Node * k,bool can_reshape)819 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
820   Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
821   if (i == NULL) {
822     i = k->Ideal(this, can_reshape);
823   }
824   return i;
825 }
826 
827 //------------------------------transform--------------------------------------
828 // Return a node which computes the same function as this node, but in a
829 // faster or cheaper fashion.
transform(Node * n)830 Node *PhaseGVN::transform( Node *n ) {
831   return transform_no_reclaim(n);
832 }
833 
834 //------------------------------transform--------------------------------------
835 // Return a node which computes the same function as this node, but
836 // in a faster or cheaper fashion.
transform_no_reclaim(Node * n)837 Node *PhaseGVN::transform_no_reclaim(Node *n) {
838   NOT_PRODUCT( set_transforms(); )
839 
840   // Apply the Ideal call in a loop until it no longer applies
841   Node* k = n;
842   Node* i = apply_ideal(k, /*can_reshape=*/false);
843   NOT_PRODUCT(uint loop_count = 1;)
844   while (i != NULL) {
845     assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
846     k = i;
847 #ifdef ASSERT
848     if (loop_count >= K + C->live_nodes()) {
849       dump_infinite_loop_info(i, "PhaseGVN::transform_no_reclaim");
850     }
851 #endif
852     i = apply_ideal(k, /*can_reshape=*/false);
853     NOT_PRODUCT(loop_count++;)
854   }
855   NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
856 
857   // If brand new node, make space in type array.
858   ensure_type_or_null(k);
859 
860   // Since I just called 'Value' to compute the set of run-time values
861   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
862   // cache Value.  Later requests for the local phase->type of this Node can
863   // use the cached Value instead of suffering with 'bottom_type'.
864   const Type* t = k->Value(this); // Get runtime Value set
865   assert(t != NULL, "value sanity");
866   if (type_or_null(k) != t) {
867 #ifndef PRODUCT
868     // Do not count initial visit to node as a transformation
869     if (type_or_null(k) == NULL) {
870       inc_new_values();
871       set_progress();
872     }
873 #endif
874     set_type(k, t);
875     // If k is a TypeNode, capture any more-precise type permanently into Node
876     k->raise_bottom_type(t);
877   }
878 
879   if (t->singleton() && !k->is_Con()) {
880     NOT_PRODUCT(set_progress();)
881     return makecon(t);          // Turn into a constant
882   }
883 
884   // Now check for Identities
885   i = k->Identity(this);        // Look for a nearby replacement
886   if (i != k) {                 // Found? Return replacement!
887     NOT_PRODUCT(set_progress();)
888     return i;
889   }
890 
891   // Global Value Numbering
892   i = hash_find_insert(k);      // Insert if new
893   if (i && (i != k)) {
894     // Return the pre-existing node
895     NOT_PRODUCT(set_progress();)
896     return i;
897   }
898 
899   // Return Idealized original
900   return k;
901 }
902 
is_dominator_helper(Node * d,Node * n,bool linear_only)903 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
904   if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
905     return false;
906   }
907   if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
908     return false;
909   }
910   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
911   int i = 0;
912   while (d != n) {
913     n = IfNode::up_one_dom(n, linear_only);
914     i++;
915     if (n == NULL || i >= 100) {
916       return false;
917     }
918   }
919   return true;
920 }
921 
922 #ifdef ASSERT
923 //------------------------------dead_loop_check--------------------------------
924 // Check for a simple dead loop when a data node references itself directly
925 // or through an other data node excluding cons and phis.
dead_loop_check(Node * n)926 void PhaseGVN::dead_loop_check( Node *n ) {
927   // Phi may reference itself in a loop
928   if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
929     // Do 2 levels check and only data inputs.
930     bool no_dead_loop = true;
931     uint cnt = n->req();
932     for (uint i = 1; i < cnt && no_dead_loop; i++) {
933       Node *in = n->in(i);
934       if (in == n) {
935         no_dead_loop = false;
936       } else if (in != NULL && !in->is_dead_loop_safe()) {
937         uint icnt = in->req();
938         for (uint j = 1; j < icnt && no_dead_loop; j++) {
939           if (in->in(j) == n || in->in(j) == in)
940             no_dead_loop = false;
941         }
942       }
943     }
944     if (!no_dead_loop) n->dump(3);
945     assert(no_dead_loop, "dead loop detected");
946   }
947 }
948 
949 
950 /**
951  * Dumps information that can help to debug the problem. A debug
952  * build fails with an assert.
953  */
dump_infinite_loop_info(Node * n,const char * where)954 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
955   n->dump(4);
956   assert(false, "infinite loop in %s", where);
957 }
958 #endif
959 
960 //=============================================================================
961 //------------------------------PhaseIterGVN-----------------------------------
962 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
PhaseIterGVN(PhaseIterGVN * igvn)963 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : PhaseGVN(igvn),
964                                                  _delay_transform(igvn->_delay_transform),
965                                                  _stack(igvn->_stack ),
966                                                  _worklist(igvn->_worklist)
967 {
968   _iterGVN = true;
969 }
970 
971 //------------------------------PhaseIterGVN-----------------------------------
972 // Initialize with previous PhaseGVN info from Parser
PhaseIterGVN(PhaseGVN * gvn)973 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : PhaseGVN(gvn),
974                                             _delay_transform(false),
975 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that
976 //       the constructor is used in incremental inlining, this consumes too much memory:
977 //                                            _stack(C->live_nodes() >> 1),
978 //       So, as a band-aid, we replace this by:
979                                             _stack(C->comp_arena(), 32),
980                                             _worklist(*C->for_igvn())
981 {
982   _iterGVN = true;
983   uint max;
984 
985   // Dead nodes in the hash table inherited from GVN were not treated as
986   // roots during def-use info creation; hence they represent an invisible
987   // use.  Clear them out.
988   max = _table.size();
989   for( uint i = 0; i < max; ++i ) {
990     Node *n = _table.at(i);
991     if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
992       if( n->is_top() ) continue;
993       // If remove_useless_nodes() has run, we expect no such nodes left.
994       assert(false, "remove_useless_nodes missed this node");
995       hash_delete(n);
996     }
997   }
998 
999   // Any Phis or Regions on the worklist probably had uses that could not
1000   // make more progress because the uses were made while the Phis and Regions
1001   // were in half-built states.  Put all uses of Phis and Regions on worklist.
1002   max = _worklist.size();
1003   for( uint j = 0; j < max; j++ ) {
1004     Node *n = _worklist.at(j);
1005     uint uop = n->Opcode();
1006     if( uop == Op_Phi || uop == Op_Region ||
1007         n->is_Type() ||
1008         n->is_Mem() )
1009       add_users_to_worklist(n);
1010   }
1011 }
1012 
shuffle_worklist()1013 void PhaseIterGVN::shuffle_worklist() {
1014   if (_worklist.size() < 2) return;
1015   for (uint i = _worklist.size() - 1; i >= 1; i--) {
1016     uint j = C->random() % (i + 1);
1017     swap(_worklist.adr()[i], _worklist.adr()[j]);
1018   }
1019 }
1020 
1021 #ifndef PRODUCT
verify_step(Node * n)1022 void PhaseIterGVN::verify_step(Node* n) {
1023   if (VerifyIterativeGVN) {
1024     ResourceMark rm;
1025     VectorSet visited;
1026     Node_List worklist;
1027 
1028     _verify_window[_verify_counter % _verify_window_size] = n;
1029     ++_verify_counter;
1030     if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
1031       ++_verify_full_passes;
1032       worklist.push(C->root());
1033       Node::verify(-1, visited, worklist);
1034       return;
1035     }
1036     for (int i = 0; i < _verify_window_size; i++) {
1037       Node* n = _verify_window[i];
1038       if (n == NULL) {
1039         continue;
1040       }
1041       if (n->in(0) == NodeSentinel) { // xform_idom
1042         _verify_window[i] = n->in(1);
1043         --i;
1044         continue;
1045       }
1046       // Typical fanout is 1-2, so this call visits about 6 nodes.
1047       if (!visited.test_set(n->_idx)) {
1048         worklist.push(n);
1049       }
1050     }
1051     Node::verify(4, visited, worklist);
1052   }
1053 }
1054 
trace_PhaseIterGVN(Node * n,Node * nn,const Type * oldtype)1055 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
1056   if (TraceIterativeGVN) {
1057     uint wlsize = _worklist.size();
1058     const Type* newtype = type_or_null(n);
1059     if (nn != n) {
1060       // print old node
1061       tty->print("< ");
1062       if (oldtype != newtype && oldtype != NULL) {
1063         oldtype->dump();
1064       }
1065       do { tty->print("\t"); } while (tty->position() < 16);
1066       tty->print("<");
1067       n->dump();
1068     }
1069     if (oldtype != newtype || nn != n) {
1070       // print new node and/or new type
1071       if (oldtype == NULL) {
1072         tty->print("* ");
1073       } else if (nn != n) {
1074         tty->print("> ");
1075       } else {
1076         tty->print("= ");
1077       }
1078       if (newtype == NULL) {
1079         tty->print("null");
1080       } else {
1081         newtype->dump();
1082       }
1083       do { tty->print("\t"); } while (tty->position() < 16);
1084       nn->dump();
1085     }
1086     if (Verbose && wlsize < _worklist.size()) {
1087       tty->print("  Push {");
1088       while (wlsize != _worklist.size()) {
1089         Node* pushed = _worklist.at(wlsize++);
1090         tty->print(" %d", pushed->_idx);
1091       }
1092       tty->print_cr(" }");
1093     }
1094     if (nn != n) {
1095       // ignore n, it might be subsumed
1096       verify_step((Node*) NULL);
1097     }
1098   }
1099 }
1100 
init_verifyPhaseIterGVN()1101 void PhaseIterGVN::init_verifyPhaseIterGVN() {
1102   _verify_counter = 0;
1103   _verify_full_passes = 0;
1104   for (int i = 0; i < _verify_window_size; i++) {
1105     _verify_window[i] = NULL;
1106   }
1107 #ifdef ASSERT
1108   // Verify that all modified nodes are on _worklist
1109   Unique_Node_List* modified_list = C->modified_nodes();
1110   while (modified_list != NULL && modified_list->size()) {
1111     Node* n = modified_list->pop();
1112     if (!n->is_Con() && !_worklist.member(n)) {
1113       n->dump();
1114       fatal("modified node is not on IGVN._worklist");
1115     }
1116   }
1117 #endif
1118 }
1119 
verify_PhaseIterGVN()1120 void PhaseIterGVN::verify_PhaseIterGVN() {
1121 #ifdef ASSERT
1122   // Verify nodes with changed inputs.
1123   Unique_Node_List* modified_list = C->modified_nodes();
1124   while (modified_list != NULL && modified_list->size()) {
1125     Node* n = modified_list->pop();
1126     if (!n->is_Con()) { // skip Con nodes
1127       n->dump();
1128       fatal("modified node was not processed by IGVN.transform_old()");
1129     }
1130   }
1131 #endif
1132 
1133   C->verify_graph_edges();
1134   if (VerifyIterativeGVN && PrintOpto) {
1135     if (_verify_counter == _verify_full_passes) {
1136       tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1137                     (int) _verify_full_passes);
1138     } else {
1139       tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1140                   (int) _verify_counter, (int) _verify_full_passes);
1141     }
1142   }
1143 
1144 #ifdef ASSERT
1145   if (modified_list != NULL) {
1146     while (modified_list->size() > 0) {
1147       Node* n = modified_list->pop();
1148       n->dump();
1149       assert(false, "VerifyIterativeGVN: new modified node was added");
1150     }
1151   }
1152 #endif
1153 }
1154 #endif /* PRODUCT */
1155 
1156 #ifdef ASSERT
1157 /**
1158  * Dumps information that can help to debug the problem. A debug
1159  * build fails with an assert.
1160  */
dump_infinite_loop_info(Node * n,const char * where)1161 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1162   n->dump(4);
1163   _worklist.dump();
1164   assert(false, "infinite loop in %s", where);
1165 }
1166 
1167 /**
1168  * Prints out information about IGVN if the 'verbose' option is used.
1169  */
trace_PhaseIterGVN_verbose(Node * n,int num_processed)1170 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1171   if (TraceIterativeGVN && Verbose) {
1172     tty->print("  Pop ");
1173     n->dump();
1174     if ((num_processed % 100) == 0) {
1175       _worklist.print_set();
1176     }
1177   }
1178 }
1179 #endif /* ASSERT */
1180 
optimize()1181 void PhaseIterGVN::optimize() {
1182   DEBUG_ONLY(uint num_processed  = 0;)
1183   NOT_PRODUCT(init_verifyPhaseIterGVN();)
1184   if (StressIGVN) {
1185     shuffle_worklist();
1186   }
1187 
1188   uint loop_count = 0;
1189   // Pull from worklist and transform the node. If the node has changed,
1190   // update edge info and put uses on worklist.
1191   while(_worklist.size()) {
1192     if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1193       return;
1194     }
1195     Node* n  = _worklist.pop();
1196     if (loop_count >= K * C->live_nodes()) {
1197       DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1198       C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1199       return;
1200     }
1201     DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1202     if (n->outcnt() != 0) {
1203       NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1204       // Do the transformation
1205       Node* nn = transform_old(n);
1206       NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1207     } else if (!n->is_top()) {
1208       remove_dead_node(n);
1209     }
1210     loop_count++;
1211   }
1212   NOT_PRODUCT(verify_PhaseIterGVN();)
1213 }
1214 
1215 
1216 /**
1217  * Register a new node with the optimizer.  Update the types array, the def-use
1218  * info.  Put on worklist.
1219  */
register_new_node_with_optimizer(Node * n,Node * orig)1220 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1221   set_type_bottom(n);
1222   _worklist.push(n);
1223   if (orig != NULL)  C->copy_node_notes_to(n, orig);
1224   return n;
1225 }
1226 
1227 //------------------------------transform--------------------------------------
1228 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
transform(Node * n)1229 Node *PhaseIterGVN::transform( Node *n ) {
1230   if (_delay_transform) {
1231     // Register the node but don't optimize for now
1232     register_new_node_with_optimizer(n);
1233     return n;
1234   }
1235 
1236   // If brand new node, make space in type array, and give it a type.
1237   ensure_type_or_null(n);
1238   if (type_or_null(n) == NULL) {
1239     set_type_bottom(n);
1240   }
1241 
1242   return transform_old(n);
1243 }
1244 
transform_old(Node * n)1245 Node *PhaseIterGVN::transform_old(Node* n) {
1246   NOT_PRODUCT(set_transforms());
1247   // Remove 'n' from hash table in case it gets modified
1248   _table.hash_delete(n);
1249   if (VerifyIterativeGVN) {
1250     assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1251   }
1252 
1253   // Apply the Ideal call in a loop until it no longer applies
1254   Node* k = n;
1255   DEBUG_ONLY(dead_loop_check(k);)
1256   DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1257   C->remove_modified_node(k);
1258   Node* i = apply_ideal(k, /*can_reshape=*/true);
1259   assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1260 #ifndef PRODUCT
1261   verify_step(k);
1262 #endif
1263 
1264   DEBUG_ONLY(uint loop_count = 1;)
1265   while (i != NULL) {
1266 #ifdef ASSERT
1267     if (loop_count >= K + C->live_nodes()) {
1268       dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
1269     }
1270 #endif
1271     assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1272     // Made a change; put users of original Node on worklist
1273     add_users_to_worklist(k);
1274     // Replacing root of transform tree?
1275     if (k != i) {
1276       // Make users of old Node now use new.
1277       subsume_node(k, i);
1278       k = i;
1279     }
1280     DEBUG_ONLY(dead_loop_check(k);)
1281     // Try idealizing again
1282     DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1283     C->remove_modified_node(k);
1284     i = apply_ideal(k, /*can_reshape=*/true);
1285     assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1286 #ifndef PRODUCT
1287     verify_step(k);
1288 #endif
1289     DEBUG_ONLY(loop_count++;)
1290   }
1291 
1292   // If brand new node, make space in type array.
1293   ensure_type_or_null(k);
1294 
1295   // See what kind of values 'k' takes on at runtime
1296   const Type* t = k->Value(this);
1297   assert(t != NULL, "value sanity");
1298 
1299   // Since I just called 'Value' to compute the set of run-time values
1300   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1301   // cache Value.  Later requests for the local phase->type of this Node can
1302   // use the cached Value instead of suffering with 'bottom_type'.
1303   if (type_or_null(k) != t) {
1304 #ifndef PRODUCT
1305     inc_new_values();
1306     set_progress();
1307 #endif
1308     set_type(k, t);
1309     // If k is a TypeNode, capture any more-precise type permanently into Node
1310     k->raise_bottom_type(t);
1311     // Move users of node to worklist
1312     add_users_to_worklist(k);
1313   }
1314   // If 'k' computes a constant, replace it with a constant
1315   if (t->singleton() && !k->is_Con()) {
1316     NOT_PRODUCT(set_progress();)
1317     Node* con = makecon(t);     // Make a constant
1318     add_users_to_worklist(k);
1319     subsume_node(k, con);       // Everybody using k now uses con
1320     return con;
1321   }
1322 
1323   // Now check for Identities
1324   i = k->Identity(this);      // Look for a nearby replacement
1325   if (i != k) {                // Found? Return replacement!
1326     NOT_PRODUCT(set_progress();)
1327     add_users_to_worklist(k);
1328     subsume_node(k, i);       // Everybody using k now uses i
1329     return i;
1330   }
1331 
1332   // Global Value Numbering
1333   i = hash_find_insert(k);      // Check for pre-existing node
1334   if (i && (i != k)) {
1335     // Return the pre-existing node if it isn't dead
1336     NOT_PRODUCT(set_progress();)
1337     add_users_to_worklist(k);
1338     subsume_node(k, i);       // Everybody using k now uses i
1339     return i;
1340   }
1341 
1342   // Return Idealized original
1343   return k;
1344 }
1345 
1346 //---------------------------------saturate------------------------------------
saturate(const Type * new_type,const Type * old_type,const Type * limit_type) const1347 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1348                                    const Type* limit_type) const {
1349   return new_type->narrow(old_type);
1350 }
1351 
1352 //------------------------------remove_globally_dead_node----------------------
1353 // Kill a globally dead Node.  All uses are also globally dead and are
1354 // aggressively trimmed.
remove_globally_dead_node(Node * dead)1355 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1356   enum DeleteProgress {
1357     PROCESS_INPUTS,
1358     PROCESS_OUTPUTS
1359   };
1360   assert(_stack.is_empty(), "not empty");
1361   _stack.push(dead, PROCESS_INPUTS);
1362 
1363   while (_stack.is_nonempty()) {
1364     dead = _stack.node();
1365     if (dead->Opcode() == Op_SafePoint) {
1366       dead->as_SafePoint()->disconnect_from_root(this);
1367     }
1368     uint progress_state = _stack.index();
1369     assert(dead != C->root(), "killing root, eh?");
1370     assert(!dead->is_top(), "add check for top when pushing");
1371     NOT_PRODUCT( set_progress(); )
1372     if (progress_state == PROCESS_INPUTS) {
1373       // After following inputs, continue to outputs
1374       _stack.set_index(PROCESS_OUTPUTS);
1375       if (!dead->is_Con()) { // Don't kill cons but uses
1376         bool recurse = false;
1377         // Remove from hash table
1378         _table.hash_delete( dead );
1379         // Smash all inputs to 'dead', isolating him completely
1380         for (uint i = 0; i < dead->req(); i++) {
1381           Node *in = dead->in(i);
1382           if (in != NULL && in != C->top()) {  // Points to something?
1383             int nrep = dead->replace_edge(in, NULL, this);  // Kill edges
1384             assert((nrep > 0), "sanity");
1385             if (in->outcnt() == 0) { // Made input go dead?
1386               _stack.push(in, PROCESS_INPUTS); // Recursively remove
1387               recurse = true;
1388             } else if (in->outcnt() == 1 &&
1389                        in->has_special_unique_user()) {
1390               _worklist.push(in->unique_out());
1391             } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1392               if (in->Opcode() == Op_Region) {
1393                 _worklist.push(in);
1394               } else if (in->is_Store()) {
1395                 DUIterator_Fast imax, i = in->fast_outs(imax);
1396                 _worklist.push(in->fast_out(i));
1397                 i++;
1398                 if (in->outcnt() == 2) {
1399                   _worklist.push(in->fast_out(i));
1400                   i++;
1401                 }
1402                 assert(!(i < imax), "sanity");
1403               }
1404             } else {
1405               BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
1406             }
1407             if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1408                 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1409               // A Load that directly follows an InitializeNode is
1410               // going away. The Stores that follow are candidates
1411               // again to be captured by the InitializeNode.
1412               for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1413                 Node *n = in->fast_out(j);
1414                 if (n->is_Store()) {
1415                   _worklist.push(n);
1416                 }
1417               }
1418             }
1419           } // if (in != NULL && in != C->top())
1420         } // for (uint i = 0; i < dead->req(); i++)
1421         if (recurse) {
1422           continue;
1423         }
1424       } // if (!dead->is_Con())
1425     } // if (progress_state == PROCESS_INPUTS)
1426 
1427     // Aggressively kill globally dead uses
1428     // (Rather than pushing all the outs at once, we push one at a time,
1429     // plus the parent to resume later, because of the indefinite number
1430     // of edge deletions per loop trip.)
1431     if (dead->outcnt() > 0) {
1432       // Recursively remove output edges
1433       _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1434     } else {
1435       // Finished disconnecting all input and output edges.
1436       _stack.pop();
1437       // Remove dead node from iterative worklist
1438       _worklist.remove(dead);
1439       C->remove_useless_node(dead);
1440     }
1441   } // while (_stack.is_nonempty())
1442 }
1443 
1444 //------------------------------subsume_node-----------------------------------
1445 // Remove users from node 'old' and add them to node 'nn'.
subsume_node(Node * old,Node * nn)1446 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1447   if (old->Opcode() == Op_SafePoint) {
1448     old->as_SafePoint()->disconnect_from_root(this);
1449   }
1450   assert( old != hash_find(old), "should already been removed" );
1451   assert( old != C->top(), "cannot subsume top node");
1452   // Copy debug or profile information to the new version:
1453   C->copy_node_notes_to(nn, old);
1454   // Move users of node 'old' to node 'nn'
1455   for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1456     Node* use = old->last_out(i);  // for each use...
1457     // use might need re-hashing (but it won't if it's a new node)
1458     rehash_node_delayed(use);
1459     // Update use-def info as well
1460     // We remove all occurrences of old within use->in,
1461     // so as to avoid rehashing any node more than once.
1462     // The hash table probe swamps any outer loop overhead.
1463     uint num_edges = 0;
1464     for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1465       if (use->in(j) == old) {
1466         use->set_req(j, nn);
1467         ++num_edges;
1468       }
1469     }
1470     i -= num_edges;    // we deleted 1 or more copies of this edge
1471   }
1472 
1473   // Search for instance field data PhiNodes in the same region pointing to the old
1474   // memory PhiNode and update their instance memory ids to point to the new node.
1475   if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) {
1476     Node* region = old->in(0);
1477     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1478       PhiNode* phi = region->fast_out(i)->isa_Phi();
1479       if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) {
1480         phi->set_inst_mem_id((int)nn->_idx);
1481       }
1482     }
1483   }
1484 
1485   // Smash all inputs to 'old', isolating him completely
1486   Node *temp = new Node(1);
1487   temp->init_req(0,nn);     // Add a use to nn to prevent him from dying
1488   remove_dead_node( old );
1489   temp->del_req(0);         // Yank bogus edge
1490   if (nn != NULL && nn->outcnt() == 0) {
1491     _worklist.push(nn);
1492   }
1493 #ifndef PRODUCT
1494   if( VerifyIterativeGVN ) {
1495     for ( int i = 0; i < _verify_window_size; i++ ) {
1496       if ( _verify_window[i] == old )
1497         _verify_window[i] = nn;
1498     }
1499   }
1500 #endif
1501   temp->destruct(this);     // reuse the _idx of this little guy
1502 }
1503 
1504 //------------------------------add_users_to_worklist--------------------------
add_users_to_worklist0(Node * n)1505 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1506   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1507     _worklist.push(n->fast_out(i));  // Push on worklist
1508   }
1509 }
1510 
1511 // Return counted loop Phi if as a counted loop exit condition, cmp
1512 // compares the the induction variable with n
countedloop_phi_from_cmp(CmpINode * cmp,Node * n)1513 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) {
1514   for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1515     Node* bol = cmp->fast_out(i);
1516     for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1517       Node* iff = bol->fast_out(i2);
1518       if (iff->is_CountedLoopEnd()) {
1519         CountedLoopEndNode* cle = iff->as_CountedLoopEnd();
1520         if (cle->limit() == n) {
1521           PhiNode* phi = cle->phi();
1522           if (phi != NULL) {
1523             return phi;
1524           }
1525         }
1526       }
1527     }
1528   }
1529   return NULL;
1530 }
1531 
add_users_to_worklist(Node * n)1532 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1533   add_users_to_worklist0(n);
1534 
1535   // Move users of node to worklist
1536   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1537     Node* use = n->fast_out(i); // Get use
1538 
1539     if( use->is_Multi() ||      // Multi-definer?  Push projs on worklist
1540         use->is_Store() )       // Enable store/load same address
1541       add_users_to_worklist0(use);
1542 
1543     // If we changed the receiver type to a call, we need to revisit
1544     // the Catch following the call.  It's looking for a non-NULL
1545     // receiver to know when to enable the regular fall-through path
1546     // in addition to the NullPtrException path.
1547     if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1548       Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1549       if (p != NULL) {
1550         add_users_to_worklist0(p);
1551       }
1552     }
1553 
1554     uint use_op = use->Opcode();
1555     if(use->is_Cmp()) {       // Enable CMP/BOOL optimization
1556       add_users_to_worklist(use); // Put Bool on worklist
1557       if (use->outcnt() > 0) {
1558         Node* bol = use->raw_out(0);
1559         if (bol->outcnt() > 0) {
1560           Node* iff = bol->raw_out(0);
1561           if (iff->outcnt() == 2) {
1562             // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1563             // phi merging either 0 or 1 onto the worklist
1564             Node* ifproj0 = iff->raw_out(0);
1565             Node* ifproj1 = iff->raw_out(1);
1566             if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1567               Node* region0 = ifproj0->raw_out(0);
1568               Node* region1 = ifproj1->raw_out(0);
1569               if( region0 == region1 )
1570                 add_users_to_worklist0(region0);
1571             }
1572           }
1573         }
1574       }
1575       if (use_op == Op_CmpI) {
1576         Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n);
1577         if (phi != NULL) {
1578           // If an opaque node feeds into the limit condition of a
1579           // CountedLoop, we need to process the Phi node for the
1580           // induction variable when the opaque node is removed:
1581           // the range of values taken by the Phi is now known and
1582           // so its type is also known.
1583           _worklist.push(phi);
1584         }
1585         Node* in1 = use->in(1);
1586         for (uint i = 0; i < in1->outcnt(); i++) {
1587           if (in1->raw_out(i)->Opcode() == Op_CastII) {
1588             Node* castii = in1->raw_out(i);
1589             if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) {
1590               Node* ifnode = castii->in(0)->in(0);
1591               if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) {
1592                 // Reprocess a CastII node that may depend on an
1593                 // opaque node value when the opaque node is
1594                 // removed. In case it carries a dependency we can do
1595                 // a better job of computing its type.
1596                 _worklist.push(castii);
1597               }
1598             }
1599           }
1600         }
1601       }
1602     }
1603 
1604     // If changed Cast input, check Phi users for simple cycles
1605     if (use->is_ConstraintCast()) {
1606       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1607         Node* u = use->fast_out(i2);
1608         if (u->is_Phi())
1609           _worklist.push(u);
1610       }
1611     }
1612     // If changed LShift inputs, check RShift users for useless sign-ext
1613     if( use_op == Op_LShiftI ) {
1614       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1615         Node* u = use->fast_out(i2);
1616         if (u->Opcode() == Op_RShiftI)
1617           _worklist.push(u);
1618       }
1619     }
1620     // If changed AddI/SubI inputs, check CmpU for range check optimization.
1621     if (use_op == Op_AddI || use_op == Op_SubI) {
1622       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1623         Node* u = use->fast_out(i2);
1624         if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1625           _worklist.push(u);
1626         }
1627       }
1628     }
1629     // If changed AddP inputs, check Stores for loop invariant
1630     if( use_op == Op_AddP ) {
1631       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1632         Node* u = use->fast_out(i2);
1633         if (u->is_Mem())
1634           _worklist.push(u);
1635       }
1636     }
1637     // If changed initialization activity, check dependent Stores
1638     if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1639       InitializeNode* init = use->as_Allocate()->initialization();
1640       if (init != NULL) {
1641         Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1642         if (imem != NULL)  add_users_to_worklist0(imem);
1643       }
1644     }
1645     if (use_op == Op_Initialize) {
1646       Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1647       if (imem != NULL)  add_users_to_worklist0(imem);
1648     }
1649     // Loading the java mirror from a Klass requires two loads and the type
1650     // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1651     //   LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
1652     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1653     bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1654 
1655     if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1656       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1657         Node* u = use->fast_out(i2);
1658         const Type* ut = u->bottom_type();
1659         if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1660           if (has_load_barrier_nodes) {
1661             // Search for load barriers behind the load
1662             for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1663               Node* b = u->fast_out(i3);
1664               if (bs->is_gc_barrier_node(b)) {
1665                 _worklist.push(b);
1666               }
1667             }
1668           }
1669           _worklist.push(u);
1670         }
1671       }
1672     }
1673   }
1674 }
1675 
1676 /**
1677  * Remove the speculative part of all types that we know of
1678  */
remove_speculative_types()1679 void PhaseIterGVN::remove_speculative_types()  {
1680   assert(UseTypeSpeculation, "speculation is off");
1681   for (uint i = 0; i < _types.Size(); i++)  {
1682     const Type* t = _types.fast_lookup(i);
1683     if (t != NULL) {
1684       _types.map(i, t->remove_speculative());
1685     }
1686   }
1687   _table.check_no_speculative_types();
1688 }
1689 
1690 // Check if the type of a divisor of a Div or Mod node includes zero.
no_dependent_zero_check(Node * n) const1691 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
1692   switch (n->Opcode()) {
1693     case Op_DivI:
1694     case Op_ModI: {
1695       // Type of divisor includes 0?
1696       if (n->in(2)->is_top()) {
1697         // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1698         return false;
1699       }
1700       const TypeInt* type_divisor = type(n->in(2))->is_int();
1701       return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1702     }
1703     case Op_DivL:
1704     case Op_ModL: {
1705       // Type of divisor includes 0?
1706       if (n->in(2)->is_top()) {
1707         // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1708         return false;
1709       }
1710       const TypeLong* type_divisor = type(n->in(2))->is_long();
1711       return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1712     }
1713   }
1714   return true;
1715 }
1716 
1717 //=============================================================================
1718 #ifndef PRODUCT
1719 uint PhaseCCP::_total_invokes   = 0;
1720 uint PhaseCCP::_total_constants = 0;
1721 #endif
1722 //------------------------------PhaseCCP---------------------------------------
1723 // Conditional Constant Propagation, ala Wegman & Zadeck
PhaseCCP(PhaseIterGVN * igvn)1724 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1725   NOT_PRODUCT( clear_constants(); )
1726   assert( _worklist.size() == 0, "" );
1727   // Clear out _nodes from IterGVN.  Must be clear to transform call.
1728   _nodes.clear();               // Clear out from IterGVN
1729   analyze();
1730 }
1731 
1732 #ifndef PRODUCT
1733 //------------------------------~PhaseCCP--------------------------------------
~PhaseCCP()1734 PhaseCCP::~PhaseCCP() {
1735   inc_invokes();
1736   _total_constants += count_constants();
1737 }
1738 #endif
1739 
1740 
1741 #ifdef ASSERT
ccp_type_widens(const Type * t,const Type * t0)1742 static bool ccp_type_widens(const Type* t, const Type* t0) {
1743   assert(t->meet(t0) == t, "Not monotonic");
1744   switch (t->base() == t0->base() ? t->base() : Type::Top) {
1745   case Type::Int:
1746     assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1747     break;
1748   case Type::Long:
1749     assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1750     break;
1751   default:
1752     break;
1753   }
1754   return true;
1755 }
1756 #endif //ASSERT
1757 
1758 //------------------------------analyze----------------------------------------
analyze()1759 void PhaseCCP::analyze() {
1760   // Initialize all types to TOP, optimistic analysis
1761   for (int i = C->unique() - 1; i >= 0; i--)  {
1762     _types.map(i,Type::TOP);
1763   }
1764 
1765   // Push root onto worklist
1766   Unique_Node_List worklist;
1767   worklist.push(C->root());
1768 
1769   // Pull from worklist; compute new value; push changes out.
1770   // This loop is the meat of CCP.
1771   while( worklist.size() ) {
1772     Node* n; // Node to be examined in this iteration
1773     if (StressCCP) {
1774       n = worklist.remove(C->random() % worklist.size());
1775     } else {
1776       n = worklist.pop();
1777     }
1778     const Type *t = n->Value(this);
1779     if (t != type(n)) {
1780       assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1781 #ifndef PRODUCT
1782       if( TracePhaseCCP ) {
1783         t->dump();
1784         do { tty->print("\t"); } while (tty->position() < 16);
1785         n->dump();
1786       }
1787 #endif
1788       set_type(n, t);
1789       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1790         Node* m = n->fast_out(i);   // Get user
1791         if (m->is_Region()) {  // New path to Region?  Must recheck Phis too
1792           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1793             Node* p = m->fast_out(i2); // Propagate changes to uses
1794             if (p->bottom_type() != type(p)) { // If not already bottomed out
1795               worklist.push(p); // Propagate change to user
1796             }
1797           }
1798         }
1799         // If we changed the receiver type to a call, we need to revisit
1800         // the Catch following the call.  It's looking for a non-NULL
1801         // receiver to know when to enable the regular fall-through path
1802         // in addition to the NullPtrException path
1803         if (m->is_Call()) {
1804           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1805             Node* p = m->fast_out(i2);  // Propagate changes to uses
1806             if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control) {
1807               Node* catch_node = p->find_out_with(Op_Catch);
1808               if (catch_node != NULL) {
1809                 worklist.push(catch_node);
1810               }
1811             }
1812           }
1813         }
1814         if (m->bottom_type() != type(m)) { // If not already bottomed out
1815           worklist.push(m);     // Propagate change to user
1816         }
1817 
1818         // CmpU nodes can get their type information from two nodes up in the
1819         // graph (instead of from the nodes immediately above). Make sure they
1820         // are added to the worklist if nodes they depend on are updated, since
1821         // they could be missed and get wrong types otherwise.
1822         uint m_op = m->Opcode();
1823         if (m_op == Op_AddI || m_op == Op_SubI) {
1824           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1825             Node* p = m->fast_out(i2); // Propagate changes to uses
1826             if (p->Opcode() == Op_CmpU) {
1827               // Got a CmpU which might need the new type information from node n.
1828               if(p->bottom_type() != type(p)) { // If not already bottomed out
1829                 worklist.push(p); // Propagate change to user
1830               }
1831             }
1832           }
1833         }
1834         // If n is used in a counted loop exit condition then the type
1835         // of the counted loop's Phi depends on the type of n. See
1836         // PhiNode::Value().
1837         if (m_op == Op_CmpI) {
1838           PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n);
1839           if (phi != NULL) {
1840             worklist.push(phi);
1841           }
1842         }
1843         // Loading the java mirror from a Klass requires two loads and the type
1844         // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1845         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1846         bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1847 
1848         if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) {
1849           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1850             Node* u = m->fast_out(i2);
1851             const Type* ut = u->bottom_type();
1852             if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) {
1853               if (has_load_barrier_nodes) {
1854                 // Search for load barriers behind the load
1855                 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1856                   Node* b = u->fast_out(i3);
1857                   if (bs->is_gc_barrier_node(b)) {
1858                     worklist.push(b);
1859                   }
1860                 }
1861               }
1862               worklist.push(u);
1863             }
1864           }
1865         }
1866       }
1867     }
1868   }
1869 }
1870 
1871 //------------------------------do_transform-----------------------------------
1872 // Top level driver for the recursive transformer
do_transform()1873 void PhaseCCP::do_transform() {
1874   // Correct leaves of new-space Nodes; they point to old-space.
1875   C->set_root( transform(C->root())->as_Root() );
1876   assert( C->top(),  "missing TOP node" );
1877   assert( C->root(), "missing root" );
1878 }
1879 
1880 //------------------------------transform--------------------------------------
1881 // Given a Node in old-space, clone him into new-space.
1882 // Convert any of his old-space children into new-space children.
transform(Node * n)1883 Node *PhaseCCP::transform( Node *n ) {
1884   Node *new_node = _nodes[n->_idx]; // Check for transformed node
1885   if( new_node != NULL )
1886     return new_node;                // Been there, done that, return old answer
1887   new_node = transform_once(n);     // Check for constant
1888   _nodes.map( n->_idx, new_node );  // Flag as having been cloned
1889 
1890   // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1891   GrowableArray <Node *> trstack(C->live_nodes() >> 1);
1892 
1893   trstack.push(new_node);           // Process children of cloned node
1894   while ( trstack.is_nonempty() ) {
1895     Node *clone = trstack.pop();
1896     uint cnt = clone->req();
1897     for( uint i = 0; i < cnt; i++ ) {          // For all inputs do
1898       Node *input = clone->in(i);
1899       if( input != NULL ) {                    // Ignore NULLs
1900         Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1901         if( new_input == NULL ) {
1902           new_input = transform_once(input);   // Check for constant
1903           _nodes.map( input->_idx, new_input );// Flag as having been cloned
1904           trstack.push(new_input);
1905         }
1906         assert( new_input == clone->in(i), "insanity check");
1907       }
1908     }
1909   }
1910   return new_node;
1911 }
1912 
1913 
1914 //------------------------------transform_once---------------------------------
1915 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
transform_once(Node * n)1916 Node *PhaseCCP::transform_once( Node *n ) {
1917   const Type *t = type(n);
1918   // Constant?  Use constant Node instead
1919   if( t->singleton() ) {
1920     Node *nn = n;               // Default is to return the original constant
1921     if( t == Type::TOP ) {
1922       // cache my top node on the Compile instance
1923       if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1924         C->set_cached_top_node(ConNode::make(Type::TOP));
1925         set_type(C->top(), Type::TOP);
1926       }
1927       nn = C->top();
1928     }
1929     if( !n->is_Con() ) {
1930       if( t != Type::TOP ) {
1931         nn = makecon(t);        // ConNode::make(t);
1932         NOT_PRODUCT( inc_constants(); )
1933       } else if( n->is_Region() ) { // Unreachable region
1934         // Note: nn == C->top()
1935         n->set_req(0, NULL);        // Cut selfreference
1936         bool progress = true;
1937         uint max = n->outcnt();
1938         DUIterator i;
1939         while (progress) {
1940           progress = false;
1941           // Eagerly remove dead phis to avoid phis copies creation.
1942           for (i = n->outs(); n->has_out(i); i++) {
1943             Node* m = n->out(i);
1944             if (m->is_Phi()) {
1945               assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1946               replace_node(m, nn);
1947               if (max != n->outcnt()) {
1948                 progress = true;
1949                 i = n->refresh_out_pos(i);
1950                 max = n->outcnt();
1951               }
1952             }
1953           }
1954         }
1955       }
1956       replace_node(n,nn);       // Update DefUse edges for new constant
1957     }
1958     return nn;
1959   }
1960 
1961   // If x is a TypeNode, capture any more-precise type permanently into Node
1962   if (t != n->bottom_type()) {
1963     hash_delete(n);             // changing bottom type may force a rehash
1964     n->raise_bottom_type(t);
1965     _worklist.push(n);          // n re-enters the hash table via the worklist
1966   }
1967 
1968   // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
1969   switch( n->Opcode() ) {
1970   case Op_FastLock:      // Revisit FastLocks for lock coarsening
1971   case Op_If:
1972   case Op_CountedLoopEnd:
1973   case Op_Region:
1974   case Op_Loop:
1975   case Op_CountedLoop:
1976   case Op_Conv2B:
1977   case Op_Opaque1:
1978   case Op_Opaque2:
1979     _worklist.push(n);
1980     break;
1981   default:
1982     break;
1983   }
1984 
1985   return  n;
1986 }
1987 
1988 //---------------------------------saturate------------------------------------
saturate(const Type * new_type,const Type * old_type,const Type * limit_type) const1989 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
1990                                const Type* limit_type) const {
1991   const Type* wide_type = new_type->widen(old_type, limit_type);
1992   if (wide_type != new_type) {          // did we widen?
1993     // If so, we may have widened beyond the limit type.  Clip it back down.
1994     new_type = wide_type->filter(limit_type);
1995   }
1996   return new_type;
1997 }
1998 
1999 //------------------------------print_statistics-------------------------------
2000 #ifndef PRODUCT
print_statistics()2001 void PhaseCCP::print_statistics() {
2002   tty->print_cr("CCP: %d  constants found: %d", _total_invokes, _total_constants);
2003 }
2004 #endif
2005 
2006 
2007 //=============================================================================
2008 #ifndef PRODUCT
2009 uint PhasePeephole::_total_peepholes = 0;
2010 #endif
2011 //------------------------------PhasePeephole----------------------------------
2012 // Conditional Constant Propagation, ala Wegman & Zadeck
PhasePeephole(PhaseRegAlloc * regalloc,PhaseCFG & cfg)2013 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
2014   : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
2015   NOT_PRODUCT( clear_peepholes(); )
2016 }
2017 
2018 #ifndef PRODUCT
2019 //------------------------------~PhasePeephole---------------------------------
~PhasePeephole()2020 PhasePeephole::~PhasePeephole() {
2021   _total_peepholes += count_peepholes();
2022 }
2023 #endif
2024 
2025 //------------------------------transform--------------------------------------
transform(Node * n)2026 Node *PhasePeephole::transform( Node *n ) {
2027   ShouldNotCallThis();
2028   return NULL;
2029 }
2030 
2031 //------------------------------do_transform-----------------------------------
do_transform()2032 void PhasePeephole::do_transform() {
2033   bool method_name_not_printed = true;
2034 
2035   // Examine each basic block
2036   for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
2037     Block* block = _cfg.get_block(block_number);
2038     bool block_not_printed = true;
2039 
2040     // and each instruction within a block
2041     uint end_index = block->number_of_nodes();
2042     // block->end_idx() not valid after PhaseRegAlloc
2043     for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
2044       Node     *n = block->get_node(instruction_index);
2045       if( n->is_Mach() ) {
2046         MachNode *m = n->as_Mach();
2047         int deleted_count = 0;
2048         // check for peephole opportunities
2049         MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count);
2050         if( m2 != NULL ) {
2051 #ifndef PRODUCT
2052           if( PrintOptoPeephole ) {
2053             // Print method, first time only
2054             if( C->method() && method_name_not_printed ) {
2055               C->method()->print_short_name(); tty->cr();
2056               method_name_not_printed = false;
2057             }
2058             // Print this block
2059             if( Verbose && block_not_printed) {
2060               tty->print_cr("in block");
2061               block->dump();
2062               block_not_printed = false;
2063             }
2064             // Print instructions being deleted
2065             for( int i = (deleted_count - 1); i >= 0; --i ) {
2066               block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
2067             }
2068             tty->print_cr("replaced with");
2069             // Print new instruction
2070             m2->format(_regalloc);
2071             tty->print("\n\n");
2072           }
2073 #endif
2074           // Remove old nodes from basic block and update instruction_index
2075           // (old nodes still exist and may have edges pointing to them
2076           //  as register allocation info is stored in the allocator using
2077           //  the node index to live range mappings.)
2078           uint safe_instruction_index = (instruction_index - deleted_count);
2079           for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
2080             block->remove_node( instruction_index );
2081           }
2082           // install new node after safe_instruction_index
2083           block->insert_node(m2, safe_instruction_index + 1);
2084           end_index = block->number_of_nodes() - 1; // Recompute new block size
2085           NOT_PRODUCT( inc_peepholes(); )
2086         }
2087       }
2088     }
2089   }
2090 }
2091 
2092 //------------------------------print_statistics-------------------------------
2093 #ifndef PRODUCT
print_statistics()2094 void PhasePeephole::print_statistics() {
2095   tty->print_cr("Peephole: peephole rules applied: %d",  _total_peepholes);
2096 }
2097 #endif
2098 
2099 
2100 //=============================================================================
2101 //------------------------------set_req_X--------------------------------------
set_req_X(uint i,Node * n,PhaseIterGVN * igvn)2102 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2103   assert( is_not_dead(n), "can not use dead node");
2104   assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2105   Node *old = in(i);
2106   set_req(i, n);
2107 
2108   // old goes dead?
2109   if( old ) {
2110     switch (old->outcnt()) {
2111     case 0:
2112       // Put into the worklist to kill later. We do not kill it now because the
2113       // recursive kill will delete the current node (this) if dead-loop exists
2114       if (!old->is_top())
2115         igvn->_worklist.push( old );
2116       break;
2117     case 1:
2118       if( old->is_Store() || old->has_special_unique_user() )
2119         igvn->add_users_to_worklist( old );
2120       break;
2121     case 2:
2122       if( old->is_Store() )
2123         igvn->add_users_to_worklist( old );
2124       if( old->Opcode() == Op_Region )
2125         igvn->_worklist.push(old);
2126       break;
2127     case 3:
2128       if( old->Opcode() == Op_Region ) {
2129         igvn->_worklist.push(old);
2130         igvn->add_users_to_worklist( old );
2131       }
2132       break;
2133     default:
2134       break;
2135     }
2136 
2137     BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
2138   }
2139 }
2140 
set_req_X(uint i,Node * n,PhaseGVN * gvn)2141 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
2142   PhaseIterGVN* igvn = gvn->is_IterGVN();
2143   if (igvn == NULL) {
2144     set_req(i, n);
2145     return;
2146   }
2147   set_req_X(i, n, igvn);
2148 }
2149 
2150 //-------------------------------replace_by-----------------------------------
2151 // Using def-use info, replace one node for another.  Follow the def-use info
2152 // to all users of the OLD node.  Then make all uses point to the NEW node.
replace_by(Node * new_node)2153 void Node::replace_by(Node *new_node) {
2154   assert(!is_top(), "top node has no DU info");
2155   for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2156     Node* use = last_out(i);
2157     uint uses_found = 0;
2158     for (uint j = 0; j < use->len(); j++) {
2159       if (use->in(j) == this) {
2160         if (j < use->req())
2161               use->set_req(j, new_node);
2162         else  use->set_prec(j, new_node);
2163         uses_found++;
2164       }
2165     }
2166     i -= uses_found;    // we deleted 1 or more copies of this edge
2167   }
2168 }
2169 
2170 //=============================================================================
2171 //-----------------------------------------------------------------------------
grow(uint i)2172 void Type_Array::grow( uint i ) {
2173   if( !_max ) {
2174     _max = 1;
2175     _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2176     _types[0] = NULL;
2177   }
2178   uint old = _max;
2179   _max = next_power_of_2(i);
2180   _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2181   memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2182 }
2183 
2184 //------------------------------dump-------------------------------------------
2185 #ifndef PRODUCT
dump() const2186 void Type_Array::dump() const {
2187   uint max = Size();
2188   for( uint i = 0; i < max; i++ ) {
2189     if( _types[i] != NULL ) {
2190       tty->print("  %d\t== ", i); _types[i]->dump(); tty->cr();
2191     }
2192   }
2193 }
2194 #endif
2195