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---------------------------------------
NodeHash(uint est_max_size)46 NodeHash::NodeHash(uint est_max_size) :
47   _a(Thread::current()->resource_area()),
48   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
49   _inserts(0), _insert_limit( insert_limit() ),
50   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
51 #ifndef PRODUCT
52   , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
53   _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
54    _total_inserts(0), _total_insert_probes(0)
55 #endif
56 {
57   // _sentinel must be in the current node space
58   _sentinel = new ProjNode(NULL, TypeFunc::Control);
59   memset(_table,0,sizeof(Node*)*_max);
60 }
61 
62 //------------------------------NodeHash---------------------------------------
NodeHash(Arena * arena,uint est_max_size)63 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
64   _a(arena),
65   _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
66   _inserts(0), _insert_limit( insert_limit() ),
67   _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
68 #ifndef PRODUCT
69   , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
70   _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
71    _total_inserts(0), _total_insert_probes(0)
72 #endif
73 {
74   // _sentinel must be in the current node space
75   _sentinel = new ProjNode(NULL, TypeFunc::Control);
76   memset(_table,0,sizeof(Node*)*_max);
77 }
78 
79 //------------------------------NodeHash---------------------------------------
NodeHash(NodeHash * nh)80 NodeHash::NodeHash(NodeHash *nh) {
81   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
82   // just copy in all the fields
83   *this = *nh;
84   // nh->_sentinel must be in the current node space
85 }
86 
replace_with(NodeHash * nh)87 void NodeHash::replace_with(NodeHash *nh) {
88   debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
89   // just copy in all the fields
90   *this = *nh;
91   // nh->_sentinel must be in the current node space
92 }
93 
94 //------------------------------hash_find--------------------------------------
95 // Find in hash table
hash_find(const Node * n)96 Node *NodeHash::hash_find( const Node *n ) {
97   // ((Node*)n)->set_hash( n->hash() );
98   uint hash = n->hash();
99   if (hash == Node::NO_HASH) {
100     NOT_PRODUCT( _lookup_misses++ );
101     return NULL;
102   }
103   uint key = hash & (_max-1);
104   uint stride = key | 0x01;
105   NOT_PRODUCT( _look_probes++ );
106   Node *k = _table[key];        // Get hashed value
107   if( !k ) {                    // ?Miss?
108     NOT_PRODUCT( _lookup_misses++ );
109     return NULL;                // Miss!
110   }
111 
112   int op = n->Opcode();
113   uint req = n->req();
114   while( 1 ) {                  // While probing hash table
115     if( k->req() == req &&      // Same count of inputs
116         k->Opcode() == op ) {   // Same Opcode
117       for( uint i=0; i<req; i++ )
118         if( n->in(i)!=k->in(i)) // Different inputs?
119           goto collision;       // "goto" is a speed hack...
120       if( n->cmp(*k) ) {        // Check for any special bits
121         NOT_PRODUCT( _lookup_hits++ );
122         return k;               // Hit!
123       }
124     }
125   collision:
126     NOT_PRODUCT( _look_probes++ );
127     key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
128     k = _table[key];            // Get hashed value
129     if( !k ) {                  // ?Miss?
130       NOT_PRODUCT( _lookup_misses++ );
131       return NULL;              // Miss!
132     }
133   }
134   ShouldNotReachHere();
135   return NULL;
136 }
137 
138 //------------------------------hash_find_insert-------------------------------
139 // Find in hash table, insert if not already present
140 // Used to preserve unique entries in hash table
hash_find_insert(Node * n)141 Node *NodeHash::hash_find_insert( Node *n ) {
142   // n->set_hash( );
143   uint hash = n->hash();
144   if (hash == Node::NO_HASH) {
145     NOT_PRODUCT( _lookup_misses++ );
146     return NULL;
147   }
148   uint key = hash & (_max-1);
149   uint stride = key | 0x01;     // stride must be relatively prime to table siz
150   uint first_sentinel = 0;      // replace a sentinel if seen.
151   NOT_PRODUCT( _look_probes++ );
152   Node *k = _table[key];        // Get hashed value
153   if( !k ) {                    // ?Miss?
154     NOT_PRODUCT( _lookup_misses++ );
155     _table[key] = n;            // Insert into table!
156     debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
157     check_grow();               // Grow table if insert hit limit
158     return NULL;                // Miss!
159   }
160   else if( k == _sentinel ) {
161     first_sentinel = key;      // Can insert here
162   }
163 
164   int op = n->Opcode();
165   uint req = n->req();
166   while( 1 ) {                  // While probing hash table
167     if( k->req() == req &&      // Same count of inputs
168         k->Opcode() == op ) {   // Same Opcode
169       for( uint i=0; i<req; i++ )
170         if( n->in(i)!=k->in(i)) // Different inputs?
171           goto collision;       // "goto" is a speed hack...
172       if( n->cmp(*k) ) {        // Check for any special bits
173         NOT_PRODUCT( _lookup_hits++ );
174         return k;               // Hit!
175       }
176     }
177   collision:
178     NOT_PRODUCT( _look_probes++ );
179     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
180     k = _table[key];            // Get hashed value
181     if( !k ) {                  // ?Miss?
182       NOT_PRODUCT( _lookup_misses++ );
183       key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
184       _table[key] = n;          // Insert into table!
185       debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
186       check_grow();             // Grow table if insert hit limit
187       return NULL;              // Miss!
188     }
189     else if( first_sentinel == 0 && k == _sentinel ) {
190       first_sentinel = key;    // Can insert here
191     }
192 
193   }
194   ShouldNotReachHere();
195   return NULL;
196 }
197 
198 //------------------------------hash_insert------------------------------------
199 // Insert into hash table
hash_insert(Node * n)200 void NodeHash::hash_insert( Node *n ) {
201   // // "conflict" comments -- print nodes that conflict
202   // bool conflict = false;
203   // n->set_hash();
204   uint hash = n->hash();
205   if (hash == Node::NO_HASH) {
206     return;
207   }
208   check_grow();
209   uint key = hash & (_max-1);
210   uint stride = key | 0x01;
211 
212   while( 1 ) {                  // While probing hash table
213     NOT_PRODUCT( _insert_probes++ );
214     Node *k = _table[key];      // Get hashed value
215     if( !k || (k == _sentinel) ) break;       // Found a slot
216     assert( k != n, "already inserted" );
217     // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print("  conflict: "); k->dump(); conflict = true; }
218     key = (key + stride) & (_max-1); // Stride through table w/ relative prime
219   }
220   _table[key] = n;              // Insert into table!
221   debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
222   // if( conflict ) { n->dump(); }
223 }
224 
225 //------------------------------hash_delete------------------------------------
226 // Replace in hash table with sentinel
hash_delete(const Node * n)227 bool NodeHash::hash_delete( const Node *n ) {
228   Node *k;
229   uint hash = n->hash();
230   if (hash == Node::NO_HASH) {
231     NOT_PRODUCT( _delete_misses++ );
232     return false;
233   }
234   uint key = hash & (_max-1);
235   uint stride = key | 0x01;
236   debug_only( uint counter = 0; );
237   for( ; /* (k != NULL) && (k != _sentinel) */; ) {
238     debug_only( counter++ );
239     NOT_PRODUCT( _delete_probes++ );
240     k = _table[key];            // Get hashed value
241     if( !k ) {                  // Miss?
242       NOT_PRODUCT( _delete_misses++ );
243 #ifdef ASSERT
244       if( VerifyOpto ) {
245         for( uint i=0; i < _max; i++ )
246           assert( _table[i] != n, "changed edges with rehashing" );
247       }
248 #endif
249       return false;             // Miss! Not in chain
250     }
251     else if( n == k ) {
252       NOT_PRODUCT( _delete_hits++ );
253       _table[key] = _sentinel;  // Hit! Label as deleted entry
254       debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
255       return true;
256     }
257     else {
258       // collision: move through table with prime offset
259       key = (key + stride/*7*/) & (_max-1);
260       assert( counter <= _insert_limit, "Cycle in hash-table");
261     }
262   }
263   ShouldNotReachHere();
264   return false;
265 }
266 
267 //------------------------------round_up---------------------------------------
268 // Round up to nearest power of 2
round_up(uint x)269 uint NodeHash::round_up( uint x ) {
270   x += (x>>2);                  // Add 25% slop
271   if( x <16 ) return 16;        // Small stuff
272   uint i=16;
273   while( i < x ) i <<= 1;       // Double to fit
274   return i;                     // Return hash table size
275 }
276 
277 //------------------------------grow-------------------------------------------
278 // Grow _table to next power of 2 and insert old entries
grow()279 void  NodeHash::grow() {
280   // Record old state
281   uint   old_max   = _max;
282   Node **old_table = _table;
283   // Construct new table with twice the space
284 #ifndef PRODUCT
285   _grows++;
286   _total_inserts       += _inserts;
287   _total_insert_probes += _insert_probes;
288   _insert_probes   = 0;
289 #endif
290   _inserts         = 0;
291   _max     = _max << 1;
292   _table   = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
293   memset(_table,0,sizeof(Node*)*_max);
294   _insert_limit = insert_limit();
295   // Insert old entries into the new table
296   for( uint i = 0; i < old_max; i++ ) {
297     Node *m = *old_table++;
298     if( !m || m == _sentinel ) continue;
299     debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
300     hash_insert(m);
301   }
302 }
303 
304 //------------------------------clear------------------------------------------
305 // Clear all entries in _table to NULL but keep storage
clear()306 void  NodeHash::clear() {
307 #ifdef ASSERT
308   // Unlock all nodes upon removal from table.
309   for (uint i = 0; i < _max; i++) {
310     Node* n = _table[i];
311     if (!n || n == _sentinel)  continue;
312     n->exit_hash_lock();
313   }
314 #endif
315 
316   memset( _table, 0, _max * sizeof(Node*) );
317 }
318 
319 //-----------------------remove_useless_nodes----------------------------------
320 // Remove useless nodes from value table,
321 // implementation does not depend on hash function
remove_useless_nodes(VectorSet & useful)322 void NodeHash::remove_useless_nodes(VectorSet &useful) {
323 
324   // Dead nodes in the hash table inherited from GVN should not replace
325   // existing nodes, remove dead nodes.
326   uint max = size();
327   Node *sentinel_node = sentinel();
328   for( uint i = 0; i < max; ++i ) {
329     Node *n = at(i);
330     if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
331       debug_only(n->exit_hash_lock()); // Unlock the node when removed
332       _table[i] = sentinel_node;       // Replace with placeholder
333     }
334   }
335 }
336 
337 
check_no_speculative_types()338 void NodeHash::check_no_speculative_types() {
339 #ifdef ASSERT
340   uint max = size();
341   Node *sentinel_node = sentinel();
342   for (uint i = 0; i < max; ++i) {
343     Node *n = at(i);
344     if(n != NULL && n != sentinel_node && n->is_Type() && n->outcnt() > 0) {
345       TypeNode* tn = n->as_Type();
346       const Type* t = tn->type();
347       const Type* t_no_spec = t->remove_speculative();
348       assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
349     }
350   }
351 #endif
352 }
353 
354 #ifndef PRODUCT
355 //------------------------------dump-------------------------------------------
356 // Dump statistics for the hash table
dump()357 void NodeHash::dump() {
358   _total_inserts       += _inserts;
359   _total_insert_probes += _insert_probes;
360   if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
361     if (WizardMode) {
362       for (uint i=0; i<_max; i++) {
363         if (_table[i])
364           tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
365       }
366     }
367     tty->print("\nGVN Hash stats:  %d grows to %d max_size\n", _grows, _max);
368     tty->print("  %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
369     tty->print("  %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
370     tty->print("  %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
371     // sentinels increase lookup cost, but not insert cost
372     assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
373     assert( _inserts+(_inserts>>3) < _max, "table too full" );
374     assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
375   }
376 }
377 
find_index(uint idx)378 Node *NodeHash::find_index(uint idx) { // For debugging
379   // Find an entry by its index value
380   for( uint i = 0; i < _max; i++ ) {
381     Node *m = _table[i];
382     if( !m || m == _sentinel ) continue;
383     if( m->_idx == (uint)idx ) return m;
384   }
385   return NULL;
386 }
387 #endif
388 
389 #ifdef ASSERT
~NodeHash()390 NodeHash::~NodeHash() {
391   // Unlock all nodes upon destruction of table.
392   if (_table != (Node**)badAddress)  clear();
393 }
394 
operator =(const NodeHash & nh)395 void NodeHash::operator=(const NodeHash& nh) {
396   // Unlock all nodes upon replacement of table.
397   if (&nh == this)  return;
398   if (_table != (Node**)badAddress)  clear();
399   memcpy((void*)this, (void*)&nh, sizeof(*this));
400   // Do not increment hash_lock counts again.
401   // Instead, be sure we never again use the source table.
402   ((NodeHash*)&nh)->_table = (Node**)badAddress;
403 }
404 
405 
406 #endif
407 
408 
409 //=============================================================================
410 //------------------------------PhaseRemoveUseless-----------------------------
411 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
PhaseRemoveUseless(PhaseGVN * gvn,Unique_Node_List * worklist,PhaseNumber phase_num)412 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN *gvn, Unique_Node_List *worklist, PhaseNumber phase_num) : Phase(phase_num),
413   _useful(Thread::current()->resource_area()) {
414 
415   // Implementation requires 'UseLoopSafepoints == true' and an edge from root
416   // to each SafePointNode at a backward branch.  Inserted in add_safepoint().
417   if( !UseLoopSafepoints || !OptoRemoveUseless ) return;
418 
419   // Identify nodes that are reachable from below, useful.
420   C->identify_useful_nodes(_useful);
421   // Update dead node list
422   C->update_dead_node_list(_useful);
423 
424   // Remove all useless nodes from PhaseValues' recorded types
425   // Must be done before disconnecting nodes to preserve hash-table-invariant
426   gvn->remove_useless_nodes(_useful.member_set());
427 
428   // Remove all useless nodes from future worklist
429   worklist->remove_useless_nodes(_useful.member_set());
430 
431   // Disconnect 'useless' nodes that are adjacent to useful nodes
432   C->remove_useless_nodes(_useful);
433 }
434 
435 //=============================================================================
436 //------------------------------PhaseRenumberLive------------------------------
437 // First, remove useless nodes (equivalent to identifying live nodes).
438 // Then, renumber live nodes.
439 //
440 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
441 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
442 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
443 // value in the range [0, x).
444 //
445 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
446 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
447 //
448 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
449 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be
450 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'.
451 // (2) Type information (the field PhaseGVN::_types) maps type information to each
452 // node ID. The mapping is updated to use the new node IDs as well. Updated type
453 // information is returned in PhaseGVN::_types.
454 //
455 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist.
456 //
457 // Other data structures used by the compiler are not updated. The hash table for value
458 // numbering (the field PhaseGVN::_table) is not updated because computing the hash
459 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either
460 // because it is empty wherever PhaseRenumberLive is used.
PhaseRenumberLive(PhaseGVN * gvn,Unique_Node_List * worklist,Unique_Node_List * new_worklist,PhaseNumber phase_num)461 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
462                                      Unique_Node_List* worklist, Unique_Node_List* new_worklist,
463                                      PhaseNumber phase_num) :
464   PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
465   _new_type_array(C->comp_arena()),
466   _old2new_map(C->unique(), C->unique(), -1),
467   _delayed(Thread::current()->resource_area()),
468   _is_pass_finished(false),
469   _live_node_count(C->live_nodes())
470 {
471   assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
472   assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
473   assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point");
474   assert(_delayed.size() == 0, "should be empty");
475 
476   uint worklist_size = worklist->size();
477 
478   // Iterate over the set of live nodes.
479   for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
480     Node* n = _useful.at(current_idx);
481 
482     bool in_worklist = false;
483     if (worklist->member(n)) {
484       in_worklist = true;
485     }
486 
487     const Type* type = gvn->type_or_null(n);
488     _new_type_array.map(current_idx, type);
489 
490     assert(_old2new_map.at(n->_idx) == -1, "already seen");
491     _old2new_map.at_put(n->_idx, current_idx);
492 
493     n->set_idx(current_idx); // Update node ID.
494 
495     if (in_worklist) {
496       new_worklist->push(n);
497     }
498 
499     if (update_embedded_ids(n) < 0) {
500       _delayed.push(n); // has embedded IDs; handle later
501     }
502   }
503 
504   assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist");
505   assert(_live_node_count == _useful.size(), "all live nodes must be processed");
506 
507   _is_pass_finished = true; // pass finished; safe to process delayed updates
508 
509   while (_delayed.size() > 0) {
510     Node* n = _delayed.pop();
511     int no_of_updates = update_embedded_ids(n);
512     assert(no_of_updates > 0, "should be updated");
513   }
514 
515   // Replace the compiler's type information with the updated type information.
516   gvn->replace_types(_new_type_array);
517 
518   // Update the unique node count of the compilation to the number of currently live nodes.
519   C->set_unique(_live_node_count);
520 
521   // Set the dead node count to 0 and reset dead node list.
522   C->reset_dead_node_list();
523 }
524 
new_index(int old_idx)525 int PhaseRenumberLive::new_index(int old_idx) {
526   assert(_is_pass_finished, "not finished");
527   if (_old2new_map.at(old_idx) == -1) { // absent
528     // Allocate a placeholder to preserve uniqueness
529     _old2new_map.at_put(old_idx, _live_node_count);
530     _live_node_count++;
531   }
532   return _old2new_map.at(old_idx);
533 }
534 
update_embedded_ids(Node * n)535 int PhaseRenumberLive::update_embedded_ids(Node* n) {
536   int no_of_updates = 0;
537   if (n->is_Phi()) {
538     PhiNode* phi = n->as_Phi();
539     if (phi->_inst_id != -1) {
540       if (!_is_pass_finished) {
541         return -1; // delay
542       }
543       int new_idx = new_index(phi->_inst_id);
544       assert(new_idx != -1, "");
545       phi->_inst_id = new_idx;
546       no_of_updates++;
547     }
548     if (phi->_inst_mem_id != -1) {
549       if (!_is_pass_finished) {
550         return -1; // delay
551       }
552       int new_idx = new_index(phi->_inst_mem_id);
553       assert(new_idx != -1, "");
554       phi->_inst_mem_id = new_idx;
555       no_of_updates++;
556     }
557   }
558 
559   const Type* type = _new_type_array.fast_lookup(n->_idx);
560   if (type != NULL && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
561     if (!_is_pass_finished) {
562         return -1; // delay
563     }
564     int old_idx = type->is_oopptr()->instance_id();
565     int new_idx = new_index(old_idx);
566     const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
567     _new_type_array.map(n->_idx, new_type);
568     no_of_updates++;
569   }
570 
571   return no_of_updates;
572 }
573 
574 //=============================================================================
575 //------------------------------PhaseTransform---------------------------------
PhaseTransform(PhaseNumber pnum)576 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
577   _arena(Thread::current()->resource_area()),
578   _nodes(_arena),
579   _types(_arena)
580 {
581   init_con_caches();
582 #ifndef PRODUCT
583   clear_progress();
584   clear_transforms();
585   set_allow_progress(true);
586 #endif
587   // Force allocation for currently existing nodes
588   _types.map(C->unique(), NULL);
589 }
590 
591 //------------------------------PhaseTransform---------------------------------
PhaseTransform(Arena * arena,PhaseNumber pnum)592 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
593   _arena(arena),
594   _nodes(arena),
595   _types(arena)
596 {
597   init_con_caches();
598 #ifndef PRODUCT
599   clear_progress();
600   clear_transforms();
601   set_allow_progress(true);
602 #endif
603   // Force allocation for currently existing nodes
604   _types.map(C->unique(), NULL);
605 }
606 
607 //------------------------------PhaseTransform---------------------------------
608 // Initialize with previously generated type information
PhaseTransform(PhaseTransform * pt,PhaseNumber pnum)609 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
610   _arena(pt->_arena),
611   _nodes(pt->_nodes),
612   _types(pt->_types)
613 {
614   init_con_caches();
615 #ifndef PRODUCT
616   clear_progress();
617   clear_transforms();
618   set_allow_progress(true);
619 #endif
620 }
621 
init_con_caches()622 void PhaseTransform::init_con_caches() {
623   memset(_icons,0,sizeof(_icons));
624   memset(_lcons,0,sizeof(_lcons));
625   memset(_zcons,0,sizeof(_zcons));
626 }
627 
628 
629 //--------------------------------find_int_type--------------------------------
find_int_type(Node * n)630 const TypeInt* PhaseTransform::find_int_type(Node* n) {
631   if (n == NULL)  return NULL;
632   // Call type_or_null(n) to determine node's type since we might be in
633   // parse phase and call n->Value() may return wrong type.
634   // (For example, a phi node at the beginning of loop parsing is not ready.)
635   const Type* t = type_or_null(n);
636   if (t == NULL)  return NULL;
637   return t->isa_int();
638 }
639 
640 
641 //-------------------------------find_long_type--------------------------------
find_long_type(Node * n)642 const TypeLong* PhaseTransform::find_long_type(Node* n) {
643   if (n == NULL)  return NULL;
644   // (See comment above on type_or_null.)
645   const Type* t = type_or_null(n);
646   if (t == NULL)  return NULL;
647   return t->isa_long();
648 }
649 
650 
651 #ifndef PRODUCT
dump_old2new_map() const652 void PhaseTransform::dump_old2new_map() const {
653   _nodes.dump();
654 }
655 
dump_new(uint nidx) const656 void PhaseTransform::dump_new( uint nidx ) const {
657   for( uint i=0; i<_nodes.Size(); i++ )
658     if( _nodes[i] && _nodes[i]->_idx == nidx ) {
659       _nodes[i]->dump();
660       tty->cr();
661       tty->print_cr("Old index= %d",i);
662       return;
663     }
664   tty->print_cr("Node %d not found in the new indices", nidx);
665 }
666 
667 //------------------------------dump_types-------------------------------------
dump_types() const668 void PhaseTransform::dump_types( ) const {
669   _types.dump();
670 }
671 
672 //------------------------------dump_nodes_and_types---------------------------
dump_nodes_and_types(const Node * root,uint depth,bool only_ctrl)673 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) {
674   VectorSet visited(Thread::current()->resource_area());
675   dump_nodes_and_types_recur( root, depth, only_ctrl, visited );
676 }
677 
678 //------------------------------dump_nodes_and_types_recur---------------------
dump_nodes_and_types_recur(const Node * n,uint depth,bool only_ctrl,VectorSet & visited)679 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
680   if( !n ) return;
681   if( depth == 0 ) return;
682   if( visited.test_set(n->_idx) ) return;
683   for( uint i=0; i<n->len(); i++ ) {
684     if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
685     dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
686   }
687   n->dump();
688   if (type_or_null(n) != NULL) {
689     tty->print("      "); type(n)->dump(); tty->cr();
690   }
691 }
692 
693 #endif
694 
695 
696 //=============================================================================
697 //------------------------------PhaseValues------------------------------------
698 // Set minimum table size to "255"
PhaseValues(Arena * arena,uint est_max_size)699 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) {
700   NOT_PRODUCT( clear_new_values(); )
701 }
702 
703 //------------------------------PhaseValues------------------------------------
704 // Set minimum table size to "255"
PhaseValues(PhaseValues * ptv)705 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ),
706   _table(&ptv->_table) {
707   NOT_PRODUCT( clear_new_values(); )
708 }
709 
710 //------------------------------PhaseValues------------------------------------
711 // Used by +VerifyOpto.  Clear out hash table but copy _types array.
PhaseValues(PhaseValues * ptv,const char * dummy)712 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ),
713   _table(ptv->arena(),ptv->_table.size()) {
714   NOT_PRODUCT( clear_new_values(); )
715 }
716 
717 //------------------------------~PhaseValues-----------------------------------
718 #ifndef PRODUCT
~PhaseValues()719 PhaseValues::~PhaseValues() {
720   _table.dump();
721 
722   // Statistics for value progress and efficiency
723   if( PrintCompilation && Verbose && WizardMode ) {
724     tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
725       is_IterGVN() ? "Iter" : "    ", C->unique(), made_progress(), made_transforms(), made_new_values());
726     if( made_transforms() != 0 ) {
727       tty->print_cr("  ratio %f", made_progress()/(float)made_transforms() );
728     } else {
729       tty->cr();
730     }
731   }
732 }
733 #endif
734 
735 //------------------------------makecon----------------------------------------
makecon(const Type * t)736 ConNode* PhaseTransform::makecon(const Type *t) {
737   assert(t->singleton(), "must be a constant");
738   assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
739   switch (t->base()) {  // fast paths
740   case Type::Half:
741   case Type::Top:  return (ConNode*) C->top();
742   case Type::Int:  return intcon( t->is_int()->get_con() );
743   case Type::Long: return longcon( t->is_long()->get_con() );
744   default:         break;
745   }
746   if (t->is_zero_type())
747     return zerocon(t->basic_type());
748   return uncached_makecon(t);
749 }
750 
751 //--------------------------uncached_makecon-----------------------------------
752 // Make an idealized constant - one of ConINode, ConPNode, etc.
uncached_makecon(const Type * t)753 ConNode* PhaseValues::uncached_makecon(const Type *t) {
754   assert(t->singleton(), "must be a constant");
755   ConNode* x = ConNode::make(t);
756   ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
757   if (k == NULL) {
758     set_type(x, t);             // Missed, provide type mapping
759     GrowableArray<Node_Notes*>* nna = C->node_note_array();
760     if (nna != NULL) {
761       Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
762       loc->clear(); // do not put debug info on constants
763     }
764   } else {
765     x->destruct();              // Hit, destroy duplicate constant
766     x = k;                      // use existing constant
767   }
768   return x;
769 }
770 
771 //------------------------------intcon-----------------------------------------
772 // Fast integer constant.  Same as "transform(new ConINode(TypeInt::make(i)))"
intcon(jint i)773 ConINode* PhaseTransform::intcon(jint i) {
774   // Small integer?  Check cache! Check that cached node is not dead
775   if (i >= _icon_min && i <= _icon_max) {
776     ConINode* icon = _icons[i-_icon_min];
777     if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
778       return icon;
779   }
780   ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
781   assert(icon->is_Con(), "");
782   if (i >= _icon_min && i <= _icon_max)
783     _icons[i-_icon_min] = icon;   // Cache small integers
784   return icon;
785 }
786 
787 //------------------------------longcon----------------------------------------
788 // Fast long constant.
longcon(jlong l)789 ConLNode* PhaseTransform::longcon(jlong l) {
790   // Small integer?  Check cache! Check that cached node is not dead
791   if (l >= _lcon_min && l <= _lcon_max) {
792     ConLNode* lcon = _lcons[l-_lcon_min];
793     if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
794       return lcon;
795   }
796   ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
797   assert(lcon->is_Con(), "");
798   if (l >= _lcon_min && l <= _lcon_max)
799     _lcons[l-_lcon_min] = lcon;      // Cache small integers
800   return lcon;
801 }
802 
803 //------------------------------zerocon-----------------------------------------
804 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
zerocon(BasicType bt)805 ConNode* PhaseTransform::zerocon(BasicType bt) {
806   assert((uint)bt <= _zcon_max, "domain check");
807   ConNode* zcon = _zcons[bt];
808   if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
809     return zcon;
810   zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
811   _zcons[bt] = zcon;
812   return zcon;
813 }
814 
815 
816 
817 //=============================================================================
apply_ideal(Node * k,bool can_reshape)818 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
819   Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
820   if (i == NULL) {
821     i = k->Ideal(this, can_reshape);
822   }
823   return i;
824 }
825 
apply_identity(Node * k)826 Node* PhaseGVN::apply_identity(Node* k) {
827   Node* i = BarrierSet::barrier_set()->barrier_set_c2()->identity_node(this, k);
828   if (i == k) {
829     i = k->Identity(this);
830   }
831   return i;
832 }
833 
834 //------------------------------transform--------------------------------------
835 // Return a node which computes the same function as this node, but in a
836 // faster or cheaper fashion.
transform(Node * n)837 Node *PhaseGVN::transform( Node *n ) {
838   return transform_no_reclaim(n);
839 }
840 
841 //------------------------------transform--------------------------------------
842 // Return a node which computes the same function as this node, but
843 // in a faster or cheaper fashion.
transform_no_reclaim(Node * n)844 Node *PhaseGVN::transform_no_reclaim( Node *n ) {
845   NOT_PRODUCT( set_transforms(); )
846 
847   // Apply the Ideal call in a loop until it no longer applies
848   Node *k = n;
849   NOT_PRODUCT( uint loop_count = 0; )
850   while( 1 ) {
851     Node *i = apply_ideal(k, /*can_reshape=*/false);
852     if( !i ) break;
853     assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
854     k = i;
855     assert(loop_count++ < K, "infinite loop in PhaseGVN::transform");
856   }
857   NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } )
858 
859 
860   // If brand new node, make space in type array.
861   ensure_type_or_null(k);
862 
863   // Since I just called 'Value' to compute the set of run-time values
864   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
865   // cache Value.  Later requests for the local phase->type of this Node can
866   // use the cached Value instead of suffering with 'bottom_type'.
867   const Type *t = k->Value(this); // Get runtime Value set
868   assert(t != NULL, "value sanity");
869   if (type_or_null(k) != t) {
870 #ifndef PRODUCT
871     // Do not count initial visit to node as a transformation
872     if (type_or_null(k) == NULL) {
873       inc_new_values();
874       set_progress();
875     }
876 #endif
877     set_type(k, t);
878     // If k is a TypeNode, capture any more-precise type permanently into Node
879     k->raise_bottom_type(t);
880   }
881 
882   if( t->singleton() && !k->is_Con() ) {
883     NOT_PRODUCT( set_progress(); )
884     return makecon(t);          // Turn into a constant
885   }
886 
887   // Now check for Identities
888   Node *i = apply_identity(k);  // Look for a nearby replacement
889   if( i != k ) {                // Found? Return replacement!
890     NOT_PRODUCT( set_progress(); )
891     return i;
892   }
893 
894   // Global Value Numbering
895   i = hash_find_insert(k);      // Insert if new
896   if( i && (i != k) ) {
897     // Return the pre-existing node
898     NOT_PRODUCT( set_progress(); )
899     return i;
900   }
901 
902   // Return Idealized original
903   return k;
904 }
905 
is_dominator_helper(Node * d,Node * n,bool linear_only)906 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
907   if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
908     return false;
909   }
910   if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
911     return false;
912   }
913   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
914   int i = 0;
915   while (d != n) {
916     n = IfNode::up_one_dom(n, linear_only);
917     i++;
918     if (n == NULL || i >= 10) {
919       return false;
920     }
921   }
922   return true;
923 }
924 
925 #ifdef ASSERT
926 //------------------------------dead_loop_check--------------------------------
927 // Check for a simple dead loop when a data node references itself directly
928 // or through an other data node excluding cons and phis.
dead_loop_check(Node * n)929 void PhaseGVN::dead_loop_check( Node *n ) {
930   // Phi may reference itself in a loop
931   if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
932     // Do 2 levels check and only data inputs.
933     bool no_dead_loop = true;
934     uint cnt = n->req();
935     for (uint i = 1; i < cnt && no_dead_loop; i++) {
936       Node *in = n->in(i);
937       if (in == n) {
938         no_dead_loop = false;
939       } else if (in != NULL && !in->is_dead_loop_safe()) {
940         uint icnt = in->req();
941         for (uint j = 1; j < icnt && no_dead_loop; j++) {
942           if (in->in(j) == n || in->in(j) == in)
943             no_dead_loop = false;
944         }
945       }
946     }
947     if (!no_dead_loop) n->dump(3);
948     assert(no_dead_loop, "dead loop detected");
949   }
950 }
951 #endif
952 
953 //=============================================================================
954 //------------------------------PhaseIterGVN-----------------------------------
955 // Initialize hash table to fresh and clean for +VerifyOpto
PhaseIterGVN(PhaseIterGVN * igvn,const char * dummy)956 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy),
957                                                                       _delay_transform(false),
958                                                                       _stack(C->live_nodes() >> 1),
959                                                                       _worklist( ) {
960 }
961 
962 //------------------------------PhaseIterGVN-----------------------------------
963 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
PhaseIterGVN(PhaseIterGVN * igvn)964 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn),
965                                                    _delay_transform(igvn->_delay_transform),
966                                                    _stack( igvn->_stack ),
967                                                    _worklist( igvn->_worklist )
968 {
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   uint max;
983 
984   // Dead nodes in the hash table inherited from GVN were not treated as
985   // roots during def-use info creation; hence they represent an invisible
986   // use.  Clear them out.
987   max = _table.size();
988   for( uint i = 0; i < max; ++i ) {
989     Node *n = _table.at(i);
990     if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
991       if( n->is_top() ) continue;
992       assert( false, "Parse::remove_useless_nodes missed this node");
993       hash_delete(n);
994     }
995   }
996 
997   // Any Phis or Regions on the worklist probably had uses that could not
998   // make more progress because the uses were made while the Phis and Regions
999   // were in half-built states.  Put all uses of Phis and Regions on worklist.
1000   max = _worklist.size();
1001   for( uint j = 0; j < max; j++ ) {
1002     Node *n = _worklist.at(j);
1003     uint uop = n->Opcode();
1004     if( uop == Op_Phi || uop == Op_Region ||
1005         n->is_Type() ||
1006         n->is_Mem() )
1007       add_users_to_worklist(n);
1008   }
1009 }
1010 
1011 /**
1012  * Initialize worklist for each node.
1013  */
init_worklist(Node * first)1014 void PhaseIterGVN::init_worklist(Node* first) {
1015   Unique_Node_List to_process;
1016   to_process.push(first);
1017 
1018   while (to_process.size() > 0) {
1019     Node* n = to_process.pop();
1020     if (!_worklist.member(n)) {
1021       _worklist.push(n);
1022 
1023       uint cnt = n->req();
1024       for(uint i = 0; i < cnt; i++) {
1025         Node* m = n->in(i);
1026         if (m != NULL) {
1027           to_process.push(m);
1028         }
1029       }
1030     }
1031   }
1032 }
1033 
1034 #ifndef PRODUCT
verify_step(Node * n)1035 void PhaseIterGVN::verify_step(Node* n) {
1036   if (VerifyIterativeGVN) {
1037     _verify_window[_verify_counter % _verify_window_size] = n;
1038     ++_verify_counter;
1039     ResourceMark rm;
1040     ResourceArea* area = Thread::current()->resource_area();
1041     VectorSet old_space(area), new_space(area);
1042     if (C->unique() < 1000 ||
1043         0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
1044       ++_verify_full_passes;
1045       Node::verify_recur(C->root(), -1, old_space, new_space);
1046     }
1047     const int verify_depth = 4;
1048     for ( int i = 0; i < _verify_window_size; i++ ) {
1049       Node* n = _verify_window[i];
1050       if ( n == NULL )  continue;
1051       if( n->in(0) == NodeSentinel ) {  // xform_idom
1052         _verify_window[i] = n->in(1);
1053         --i; continue;
1054       }
1055       // Typical fanout is 1-2, so this call visits about 6 nodes.
1056       Node::verify_recur(n, verify_depth, old_space, new_space);
1057     }
1058   }
1059 }
1060 
trace_PhaseIterGVN(Node * n,Node * nn,const Type * oldtype)1061 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
1062   if (TraceIterativeGVN) {
1063     uint wlsize = _worklist.size();
1064     const Type* newtype = type_or_null(n);
1065     if (nn != n) {
1066       // print old node
1067       tty->print("< ");
1068       if (oldtype != newtype && oldtype != NULL) {
1069         oldtype->dump();
1070       }
1071       do { tty->print("\t"); } while (tty->position() < 16);
1072       tty->print("<");
1073       n->dump();
1074     }
1075     if (oldtype != newtype || nn != n) {
1076       // print new node and/or new type
1077       if (oldtype == NULL) {
1078         tty->print("* ");
1079       } else if (nn != n) {
1080         tty->print("> ");
1081       } else {
1082         tty->print("= ");
1083       }
1084       if (newtype == NULL) {
1085         tty->print("null");
1086       } else {
1087         newtype->dump();
1088       }
1089       do { tty->print("\t"); } while (tty->position() < 16);
1090       nn->dump();
1091     }
1092     if (Verbose && wlsize < _worklist.size()) {
1093       tty->print("  Push {");
1094       while (wlsize != _worklist.size()) {
1095         Node* pushed = _worklist.at(wlsize++);
1096         tty->print(" %d", pushed->_idx);
1097       }
1098       tty->print_cr(" }");
1099     }
1100     if (nn != n) {
1101       // ignore n, it might be subsumed
1102       verify_step((Node*) NULL);
1103     }
1104   }
1105 }
1106 
init_verifyPhaseIterGVN()1107 void PhaseIterGVN::init_verifyPhaseIterGVN() {
1108   _verify_counter = 0;
1109   _verify_full_passes = 0;
1110   for (int i = 0; i < _verify_window_size; i++) {
1111     _verify_window[i] = NULL;
1112   }
1113 #ifdef ASSERT
1114   // Verify that all modified nodes are on _worklist
1115   Unique_Node_List* modified_list = C->modified_nodes();
1116   while (modified_list != NULL && modified_list->size()) {
1117     Node* n = modified_list->pop();
1118     if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) {
1119       n->dump();
1120       assert(false, "modified node is not on IGVN._worklist");
1121     }
1122   }
1123 #endif
1124 }
1125 
verify_PhaseIterGVN()1126 void PhaseIterGVN::verify_PhaseIterGVN() {
1127 #ifdef ASSERT
1128   // Verify nodes with changed inputs.
1129   Unique_Node_List* modified_list = C->modified_nodes();
1130   while (modified_list != NULL && modified_list->size()) {
1131     Node* n = modified_list->pop();
1132     if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes
1133       n->dump();
1134       assert(false, "modified node was not processed by IGVN.transform_old()");
1135     }
1136   }
1137 #endif
1138 
1139   C->verify_graph_edges();
1140   if( VerifyOpto && allow_progress() ) {
1141     // Must turn off allow_progress to enable assert and break recursion
1142     C->root()->verify();
1143     { // Check if any progress was missed using IterGVN
1144       // Def-Use info enables transformations not attempted in wash-pass
1145       // e.g. Region/Phi cleanup, ...
1146       // Null-check elision -- may not have reached fixpoint
1147       //                       do not propagate to dominated nodes
1148       ResourceMark rm;
1149       PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean!
1150       // Fill worklist completely
1151       igvn2.init_worklist(C->root());
1152 
1153       igvn2.set_allow_progress(false);
1154       igvn2.optimize();
1155       igvn2.set_allow_progress(true);
1156     }
1157   }
1158   if (VerifyIterativeGVN && PrintOpto) {
1159     if (_verify_counter == _verify_full_passes) {
1160       tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1161                     (int) _verify_full_passes);
1162     } else {
1163       tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1164                   (int) _verify_counter, (int) _verify_full_passes);
1165     }
1166   }
1167 
1168 #ifdef ASSERT
1169   while (modified_list->size()) {
1170     Node* n = modified_list->pop();
1171     n->dump();
1172     assert(false, "VerifyIterativeGVN: new modified node was added");
1173   }
1174 #endif
1175 }
1176 #endif /* PRODUCT */
1177 
1178 #ifdef ASSERT
1179 /**
1180  * Dumps information that can help to debug the problem. A debug
1181  * build fails with an assert.
1182  */
dump_infinite_loop_info(Node * n)1183 void PhaseIterGVN::dump_infinite_loop_info(Node* n) {
1184   n->dump(4);
1185   _worklist.dump();
1186   assert(false, "infinite loop in PhaseIterGVN::optimize");
1187 }
1188 
1189 /**
1190  * Prints out information about IGVN if the 'verbose' option is used.
1191  */
trace_PhaseIterGVN_verbose(Node * n,int num_processed)1192 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1193   if (TraceIterativeGVN && Verbose) {
1194     tty->print("  Pop ");
1195     n->dump();
1196     if ((num_processed % 100) == 0) {
1197       _worklist.print_set();
1198     }
1199   }
1200 }
1201 #endif /* ASSERT */
1202 
optimize()1203 void PhaseIterGVN::optimize() {
1204   DEBUG_ONLY(uint num_processed  = 0;)
1205   NOT_PRODUCT(init_verifyPhaseIterGVN();)
1206 
1207   uint loop_count = 0;
1208   // Pull from worklist and transform the node. If the node has changed,
1209   // update edge info and put uses on worklist.
1210   while(_worklist.size()) {
1211     if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1212       return;
1213     }
1214     Node* n  = _worklist.pop();
1215     if (++loop_count >= K * C->live_nodes()) {
1216       DEBUG_ONLY(dump_infinite_loop_info(n);)
1217       C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1218       return;
1219     }
1220     DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1221     if (n->outcnt() != 0) {
1222       NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1223       // Do the transformation
1224       Node* nn = transform_old(n);
1225       NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1226     } else if (!n->is_top()) {
1227       remove_dead_node(n);
1228     }
1229   }
1230   NOT_PRODUCT(verify_PhaseIterGVN();)
1231 }
1232 
1233 
1234 /**
1235  * Register a new node with the optimizer.  Update the types array, the def-use
1236  * info.  Put on worklist.
1237  */
register_new_node_with_optimizer(Node * n,Node * orig)1238 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1239   set_type_bottom(n);
1240   _worklist.push(n);
1241   if (orig != NULL)  C->copy_node_notes_to(n, orig);
1242   return n;
1243 }
1244 
1245 //------------------------------transform--------------------------------------
1246 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
transform(Node * n)1247 Node *PhaseIterGVN::transform( Node *n ) {
1248   if (_delay_transform) {
1249     // Register the node but don't optimize for now
1250     register_new_node_with_optimizer(n);
1251     return n;
1252   }
1253 
1254   // If brand new node, make space in type array, and give it a type.
1255   ensure_type_or_null(n);
1256   if (type_or_null(n) == NULL) {
1257     set_type_bottom(n);
1258   }
1259 
1260   return transform_old(n);
1261 }
1262 
transform_old(Node * n)1263 Node *PhaseIterGVN::transform_old(Node* n) {
1264   DEBUG_ONLY(uint loop_count = 0;);
1265   NOT_PRODUCT(set_transforms());
1266 
1267   // Remove 'n' from hash table in case it gets modified
1268   _table.hash_delete(n);
1269   if (VerifyIterativeGVN) {
1270    assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1271   }
1272 
1273   // Apply the Ideal call in a loop until it no longer applies
1274   Node* k = n;
1275   DEBUG_ONLY(dead_loop_check(k);)
1276   DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1277   C->remove_modified_node(k);
1278   Node* 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     if (!allow_progress()) {
1284       if (i->is_Add() && (i->outcnt() == 1)) {
1285         // Switched input to left side because this is the only use
1286       } else if (i->is_If() && (i->in(0) == NULL)) {
1287         // This IF is dead because it is dominated by an equivalent IF When
1288         // dominating if changed, info is not propagated sparsely to 'this'
1289         // Propagating this info further will spuriously identify other
1290         // progress.
1291         return i;
1292       } else
1293         set_progress();
1294     } else {
1295       set_progress();
1296     }
1297   }
1298 #endif
1299 
1300   while (i != NULL) {
1301 #ifdef ASSERT
1302     if (loop_count >= K) {
1303       dump_infinite_loop_info(i);
1304     }
1305     loop_count++;
1306 #endif
1307     assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1308     // Made a change; put users of original Node on worklist
1309     add_users_to_worklist(k);
1310     // Replacing root of transform tree?
1311     if (k != i) {
1312       // Make users of old Node now use new.
1313       subsume_node(k, i);
1314       k = i;
1315     }
1316     DEBUG_ONLY(dead_loop_check(k);)
1317     // Try idealizing again
1318     DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1319     C->remove_modified_node(k);
1320     i = apply_ideal(k, /*can_reshape=*/true);
1321     assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1322 #ifndef PRODUCT
1323     verify_step(k);
1324     if (i && VerifyOpto) {
1325       set_progress();
1326     }
1327 #endif
1328   }
1329 
1330   // If brand new node, make space in type array.
1331   ensure_type_or_null(k);
1332 
1333   // See what kind of values 'k' takes on at runtime
1334   const Type* t = k->Value(this);
1335   assert(t != NULL, "value sanity");
1336 
1337   // Since I just called 'Value' to compute the set of run-time values
1338   // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1339   // cache Value.  Later requests for the local phase->type of this Node can
1340   // use the cached Value instead of suffering with 'bottom_type'.
1341   if (type_or_null(k) != t) {
1342 #ifndef PRODUCT
1343     inc_new_values();
1344     set_progress();
1345 #endif
1346     set_type(k, t);
1347     // If k is a TypeNode, capture any more-precise type permanently into Node
1348     k->raise_bottom_type(t);
1349     // Move users of node to worklist
1350     add_users_to_worklist(k);
1351   }
1352   // If 'k' computes a constant, replace it with a constant
1353   if (t->singleton() && !k->is_Con()) {
1354     NOT_PRODUCT(set_progress();)
1355     Node* con = makecon(t);     // Make a constant
1356     add_users_to_worklist(k);
1357     subsume_node(k, con);       // Everybody using k now uses con
1358     return con;
1359   }
1360 
1361   // Now check for Identities
1362   i = apply_identity(k);      // Look for a nearby replacement
1363   if (i != k) {                // Found? Return replacement!
1364     NOT_PRODUCT(set_progress();)
1365     add_users_to_worklist(k);
1366     subsume_node(k, i);       // Everybody using k now uses i
1367     return i;
1368   }
1369 
1370   // Global Value Numbering
1371   i = hash_find_insert(k);      // Check for pre-existing node
1372   if (i && (i != k)) {
1373     // Return the pre-existing node if it isn't dead
1374     NOT_PRODUCT(set_progress();)
1375     add_users_to_worklist(k);
1376     subsume_node(k, i);       // Everybody using k now uses i
1377     return i;
1378   }
1379 
1380   // Return Idealized original
1381   return k;
1382 }
1383 
1384 //---------------------------------saturate------------------------------------
saturate(const Type * new_type,const Type * old_type,const Type * limit_type) const1385 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1386                                    const Type* limit_type) const {
1387   return new_type->narrow(old_type);
1388 }
1389 
1390 //------------------------------remove_globally_dead_node----------------------
1391 // Kill a globally dead Node.  All uses are also globally dead and are
1392 // aggressively trimmed.
remove_globally_dead_node(Node * dead)1393 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1394   enum DeleteProgress {
1395     PROCESS_INPUTS,
1396     PROCESS_OUTPUTS
1397   };
1398   assert(_stack.is_empty(), "not empty");
1399   _stack.push(dead, PROCESS_INPUTS);
1400 
1401   while (_stack.is_nonempty()) {
1402     dead = _stack.node();
1403     if (dead->Opcode() == Op_SafePoint) {
1404       dead->as_SafePoint()->disconnect_from_root(this);
1405     }
1406     uint progress_state = _stack.index();
1407     assert(dead != C->root(), "killing root, eh?");
1408     assert(!dead->is_top(), "add check for top when pushing");
1409     NOT_PRODUCT( set_progress(); )
1410     if (progress_state == PROCESS_INPUTS) {
1411       // After following inputs, continue to outputs
1412       _stack.set_index(PROCESS_OUTPUTS);
1413       if (!dead->is_Con()) { // Don't kill cons but uses
1414         bool recurse = false;
1415         // Remove from hash table
1416         _table.hash_delete( dead );
1417         // Smash all inputs to 'dead', isolating him completely
1418         for (uint i = 0; i < dead->req(); i++) {
1419           Node *in = dead->in(i);
1420           if (in != NULL && in != C->top()) {  // Points to something?
1421             int nrep = dead->replace_edge(in, NULL);  // Kill edges
1422             assert((nrep > 0), "sanity");
1423             if (in->outcnt() == 0) { // Made input go dead?
1424               _stack.push(in, PROCESS_INPUTS); // Recursively remove
1425               recurse = true;
1426             } else if (in->outcnt() == 1 &&
1427                        in->has_special_unique_user()) {
1428               _worklist.push(in->unique_out());
1429             } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1430               if (in->Opcode() == Op_Region) {
1431                 _worklist.push(in);
1432               } else if (in->is_Store()) {
1433                 DUIterator_Fast imax, i = in->fast_outs(imax);
1434                 _worklist.push(in->fast_out(i));
1435                 i++;
1436                 if (in->outcnt() == 2) {
1437                   _worklist.push(in->fast_out(i));
1438                   i++;
1439                 }
1440                 assert(!(i < imax), "sanity");
1441               }
1442             } else {
1443               BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
1444             }
1445             if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1446                 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1447               // A Load that directly follows an InitializeNode is
1448               // going away. The Stores that follow are candidates
1449               // again to be captured by the InitializeNode.
1450               for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1451                 Node *n = in->fast_out(j);
1452                 if (n->is_Store()) {
1453                   _worklist.push(n);
1454                 }
1455               }
1456             }
1457           } // if (in != NULL && in != C->top())
1458         } // for (uint i = 0; i < dead->req(); i++)
1459         if (recurse) {
1460           continue;
1461         }
1462       } // if (!dead->is_Con())
1463     } // if (progress_state == PROCESS_INPUTS)
1464 
1465     // Aggressively kill globally dead uses
1466     // (Rather than pushing all the outs at once, we push one at a time,
1467     // plus the parent to resume later, because of the indefinite number
1468     // of edge deletions per loop trip.)
1469     if (dead->outcnt() > 0) {
1470       // Recursively remove output edges
1471       _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1472     } else {
1473       // Finished disconnecting all input and output edges.
1474       _stack.pop();
1475       // Remove dead node from iterative worklist
1476       _worklist.remove(dead);
1477       C->remove_modified_node(dead);
1478       // Constant node that has no out-edges and has only one in-edge from
1479       // root is usually dead. However, sometimes reshaping walk makes
1480       // it reachable by adding use edges. So, we will NOT count Con nodes
1481       // as dead to be conservative about the dead node count at any
1482       // given time.
1483       if (!dead->is_Con()) {
1484         C->record_dead_node(dead->_idx);
1485       }
1486       if (dead->is_macro()) {
1487         C->remove_macro_node(dead);
1488       }
1489       if (dead->is_expensive()) {
1490         C->remove_expensive_node(dead);
1491       }
1492       CastIINode* cast = dead->isa_CastII();
1493       if (cast != NULL && cast->has_range_check()) {
1494         C->remove_range_check_cast(cast);
1495       }
1496       if (dead->Opcode() == Op_Opaque4) {
1497         C->remove_opaque4_node(dead);
1498       }
1499       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1500       bs->unregister_potential_barrier_node(dead);
1501     }
1502   } // while (_stack.is_nonempty())
1503 }
1504 
1505 //------------------------------subsume_node-----------------------------------
1506 // Remove users from node 'old' and add them to node 'nn'.
subsume_node(Node * old,Node * nn)1507 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1508   if (old->Opcode() == Op_SafePoint) {
1509     old->as_SafePoint()->disconnect_from_root(this);
1510   }
1511   assert( old != hash_find(old), "should already been removed" );
1512   assert( old != C->top(), "cannot subsume top node");
1513   // Copy debug or profile information to the new version:
1514   C->copy_node_notes_to(nn, old);
1515   // Move users of node 'old' to node 'nn'
1516   for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1517     Node* use = old->last_out(i);  // for each use...
1518     // use might need re-hashing (but it won't if it's a new node)
1519     rehash_node_delayed(use);
1520     // Update use-def info as well
1521     // We remove all occurrences of old within use->in,
1522     // so as to avoid rehashing any node more than once.
1523     // The hash table probe swamps any outer loop overhead.
1524     uint num_edges = 0;
1525     for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1526       if (use->in(j) == old) {
1527         use->set_req(j, nn);
1528         ++num_edges;
1529       }
1530     }
1531     i -= num_edges;    // we deleted 1 or more copies of this edge
1532   }
1533 
1534   // Search for instance field data PhiNodes in the same region pointing to the old
1535   // memory PhiNode and update their instance memory ids to point to the new node.
1536   if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) {
1537     Node* region = old->in(0);
1538     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1539       PhiNode* phi = region->fast_out(i)->isa_Phi();
1540       if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) {
1541         phi->set_inst_mem_id((int)nn->_idx);
1542       }
1543     }
1544   }
1545 
1546   // Smash all inputs to 'old', isolating him completely
1547   Node *temp = new Node(1);
1548   temp->init_req(0,nn);     // Add a use to nn to prevent him from dying
1549   remove_dead_node( old );
1550   temp->del_req(0);         // Yank bogus edge
1551 #ifndef PRODUCT
1552   if( VerifyIterativeGVN ) {
1553     for ( int i = 0; i < _verify_window_size; i++ ) {
1554       if ( _verify_window[i] == old )
1555         _verify_window[i] = nn;
1556     }
1557   }
1558 #endif
1559   _worklist.remove(temp);   // this can be necessary
1560   temp->destruct();         // reuse the _idx of this little guy
1561 }
1562 
1563 //------------------------------add_users_to_worklist--------------------------
add_users_to_worklist0(Node * n)1564 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1565   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1566     _worklist.push(n->fast_out(i));  // Push on worklist
1567   }
1568 }
1569 
1570 // Return counted loop Phi if as a counted loop exit condition, cmp
1571 // compares the the induction variable with n
countedloop_phi_from_cmp(CmpINode * cmp,Node * n)1572 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) {
1573   for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1574     Node* bol = cmp->fast_out(i);
1575     for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1576       Node* iff = bol->fast_out(i2);
1577       if (iff->is_CountedLoopEnd()) {
1578         CountedLoopEndNode* cle = iff->as_CountedLoopEnd();
1579         if (cle->limit() == n) {
1580           PhiNode* phi = cle->phi();
1581           if (phi != NULL) {
1582             return phi;
1583           }
1584         }
1585       }
1586     }
1587   }
1588   return NULL;
1589 }
1590 
add_users_to_worklist(Node * n)1591 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1592   add_users_to_worklist0(n);
1593 
1594   // Move users of node to worklist
1595   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1596     Node* use = n->fast_out(i); // Get use
1597 
1598     if( use->is_Multi() ||      // Multi-definer?  Push projs on worklist
1599         use->is_Store() )       // Enable store/load same address
1600       add_users_to_worklist0(use);
1601 
1602     // If we changed the receiver type to a call, we need to revisit
1603     // the Catch following the call.  It's looking for a non-NULL
1604     // receiver to know when to enable the regular fall-through path
1605     // in addition to the NullPtrException path.
1606     if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1607       Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1608       if (p != NULL) {
1609         add_users_to_worklist0(p);
1610       }
1611     }
1612 
1613     uint use_op = use->Opcode();
1614     if(use->is_Cmp()) {       // Enable CMP/BOOL optimization
1615       add_users_to_worklist(use); // Put Bool on worklist
1616       if (use->outcnt() > 0) {
1617         Node* bol = use->raw_out(0);
1618         if (bol->outcnt() > 0) {
1619           Node* iff = bol->raw_out(0);
1620           if (iff->outcnt() == 2) {
1621             // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1622             // phi merging either 0 or 1 onto the worklist
1623             Node* ifproj0 = iff->raw_out(0);
1624             Node* ifproj1 = iff->raw_out(1);
1625             if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1626               Node* region0 = ifproj0->raw_out(0);
1627               Node* region1 = ifproj1->raw_out(0);
1628               if( region0 == region1 )
1629                 add_users_to_worklist0(region0);
1630             }
1631           }
1632         }
1633       }
1634       if (use_op == Op_CmpI) {
1635         Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n);
1636         if (phi != NULL) {
1637           // If an opaque node feeds into the limit condition of a
1638           // CountedLoop, we need to process the Phi node for the
1639           // induction variable when the opaque node is removed:
1640           // the range of values taken by the Phi is now known and
1641           // so its type is also known.
1642           _worklist.push(phi);
1643         }
1644         Node* in1 = use->in(1);
1645         for (uint i = 0; i < in1->outcnt(); i++) {
1646           if (in1->raw_out(i)->Opcode() == Op_CastII) {
1647             Node* castii = in1->raw_out(i);
1648             if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) {
1649               Node* ifnode = castii->in(0)->in(0);
1650               if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) {
1651                 // Reprocess a CastII node that may depend on an
1652                 // opaque node value when the opaque node is
1653                 // removed. In case it carries a dependency we can do
1654                 // a better job of computing its type.
1655                 _worklist.push(castii);
1656               }
1657             }
1658           }
1659         }
1660       }
1661     }
1662 
1663     // If changed Cast input, check Phi users for simple cycles
1664     if (use->is_ConstraintCast()) {
1665       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1666         Node* u = use->fast_out(i2);
1667         if (u->is_Phi())
1668           _worklist.push(u);
1669       }
1670     }
1671     // If changed LShift inputs, check RShift users for useless sign-ext
1672     if( use_op == Op_LShiftI ) {
1673       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1674         Node* u = use->fast_out(i2);
1675         if (u->Opcode() == Op_RShiftI)
1676           _worklist.push(u);
1677       }
1678     }
1679     // If changed AddI/SubI inputs, check CmpU for range check optimization.
1680     if (use_op == Op_AddI || use_op == Op_SubI) {
1681       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1682         Node* u = use->fast_out(i2);
1683         if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1684           _worklist.push(u);
1685         }
1686       }
1687     }
1688     // If changed AddP inputs, check Stores for loop invariant
1689     if( use_op == Op_AddP ) {
1690       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1691         Node* u = use->fast_out(i2);
1692         if (u->is_Mem())
1693           _worklist.push(u);
1694       }
1695     }
1696     // If changed initialization activity, check dependent Stores
1697     if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1698       InitializeNode* init = use->as_Allocate()->initialization();
1699       if (init != NULL) {
1700         Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1701         if (imem != NULL)  add_users_to_worklist0(imem);
1702       }
1703     }
1704     if (use_op == Op_Initialize) {
1705       Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1706       if (imem != NULL)  add_users_to_worklist0(imem);
1707     }
1708     // Loading the java mirror from a Klass requires two loads and the type
1709     // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1710     //   LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
1711     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1712     bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1713 
1714     if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1715       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1716         Node* u = use->fast_out(i2);
1717         const Type* ut = u->bottom_type();
1718         if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1719           if (has_load_barrier_nodes) {
1720             // Search for load barriers behind the load
1721             for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1722               Node* b = u->fast_out(i3);
1723               if (bs->is_gc_barrier_node(b)) {
1724                 _worklist.push(b);
1725               }
1726             }
1727           }
1728           _worklist.push(u);
1729         }
1730       }
1731     }
1732 
1733     BarrierSet::barrier_set()->barrier_set_c2()->igvn_add_users_to_worklist(this, use);
1734   }
1735 }
1736 
1737 /**
1738  * Remove the speculative part of all types that we know of
1739  */
remove_speculative_types()1740 void PhaseIterGVN::remove_speculative_types()  {
1741   assert(UseTypeSpeculation, "speculation is off");
1742   for (uint i = 0; i < _types.Size(); i++)  {
1743     const Type* t = _types.fast_lookup(i);
1744     if (t != NULL) {
1745       _types.map(i, t->remove_speculative());
1746     }
1747   }
1748   _table.check_no_speculative_types();
1749 }
1750 
1751 //=============================================================================
1752 #ifndef PRODUCT
1753 uint PhaseCCP::_total_invokes   = 0;
1754 uint PhaseCCP::_total_constants = 0;
1755 #endif
1756 //------------------------------PhaseCCP---------------------------------------
1757 // Conditional Constant Propagation, ala Wegman & Zadeck
PhaseCCP(PhaseIterGVN * igvn)1758 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1759   NOT_PRODUCT( clear_constants(); )
1760   assert( _worklist.size() == 0, "" );
1761   // Clear out _nodes from IterGVN.  Must be clear to transform call.
1762   _nodes.clear();               // Clear out from IterGVN
1763   analyze();
1764 }
1765 
1766 #ifndef PRODUCT
1767 //------------------------------~PhaseCCP--------------------------------------
~PhaseCCP()1768 PhaseCCP::~PhaseCCP() {
1769   inc_invokes();
1770   _total_constants += count_constants();
1771 }
1772 #endif
1773 
1774 
1775 #ifdef ASSERT
ccp_type_widens(const Type * t,const Type * t0)1776 static bool ccp_type_widens(const Type* t, const Type* t0) {
1777   assert(t->meet(t0) == t, "Not monotonic");
1778   switch (t->base() == t0->base() ? t->base() : Type::Top) {
1779   case Type::Int:
1780     assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1781     break;
1782   case Type::Long:
1783     assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1784     break;
1785   default:
1786     break;
1787   }
1788   return true;
1789 }
1790 #endif //ASSERT
1791 
1792 //------------------------------analyze----------------------------------------
analyze()1793 void PhaseCCP::analyze() {
1794   // Initialize all types to TOP, optimistic analysis
1795   for (int i = C->unique() - 1; i >= 0; i--)  {
1796     _types.map(i,Type::TOP);
1797   }
1798 
1799   // Push root onto worklist
1800   Unique_Node_List worklist;
1801   worklist.push(C->root());
1802 
1803   // Pull from worklist; compute new value; push changes out.
1804   // This loop is the meat of CCP.
1805   while( worklist.size() ) {
1806     Node *n = worklist.pop();
1807     const Type *t = n->Value(this);
1808     if (t != type(n)) {
1809       assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1810 #ifndef PRODUCT
1811       if( TracePhaseCCP ) {
1812         t->dump();
1813         do { tty->print("\t"); } while (tty->position() < 16);
1814         n->dump();
1815       }
1816 #endif
1817       set_type(n, t);
1818       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1819         Node* m = n->fast_out(i);   // Get user
1820         if (m->is_Region()) {  // New path to Region?  Must recheck Phis too
1821           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1822             Node* p = m->fast_out(i2); // Propagate changes to uses
1823             if (p->bottom_type() != type(p)) { // If not already bottomed out
1824               worklist.push(p); // Propagate change to user
1825             }
1826           }
1827         }
1828         // If we changed the receiver type to a call, we need to revisit
1829         // the Catch following the call.  It's looking for a non-NULL
1830         // receiver to know when to enable the regular fall-through path
1831         // in addition to the NullPtrException path
1832         if (m->is_Call()) {
1833           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1834             Node* p = m->fast_out(i2);  // Propagate changes to uses
1835             if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control) {
1836               Node* catch_node = p->find_out_with(Op_Catch);
1837               if (catch_node != NULL) {
1838                 worklist.push(catch_node);
1839               }
1840             }
1841           }
1842         }
1843         if (m->bottom_type() != type(m)) { // If not already bottomed out
1844           worklist.push(m);     // Propagate change to user
1845         }
1846 
1847         // CmpU nodes can get their type information from two nodes up in the
1848         // graph (instead of from the nodes immediately above). Make sure they
1849         // are added to the worklist if nodes they depend on are updated, since
1850         // they could be missed and get wrong types otherwise.
1851         uint m_op = m->Opcode();
1852         if (m_op == Op_AddI || m_op == Op_SubI) {
1853           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1854             Node* p = m->fast_out(i2); // Propagate changes to uses
1855             if (p->Opcode() == Op_CmpU) {
1856               // Got a CmpU which might need the new type information from node n.
1857               if(p->bottom_type() != type(p)) { // If not already bottomed out
1858                 worklist.push(p); // Propagate change to user
1859               }
1860             }
1861           }
1862         }
1863         // If n is used in a counted loop exit condition then the type
1864         // of the counted loop's Phi depends on the type of n. See
1865         // PhiNode::Value().
1866         if (m_op == Op_CmpI) {
1867           PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n);
1868           if (phi != NULL) {
1869             worklist.push(phi);
1870           }
1871         }
1872         // Loading the java mirror from a Klass requires two loads and the type
1873         // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1874         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1875         bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1876 
1877         if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) {
1878           for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1879             Node* u = m->fast_out(i2);
1880             const Type* ut = u->bottom_type();
1881             if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) {
1882               if (has_load_barrier_nodes) {
1883                 // Search for load barriers behind the load
1884                 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1885                   Node* b = u->fast_out(i3);
1886                   if (bs->is_gc_barrier_node(b)) {
1887                     worklist.push(b);
1888                   }
1889                 }
1890               }
1891               worklist.push(u);
1892             }
1893           }
1894         }
1895 
1896         BarrierSet::barrier_set()->barrier_set_c2()->ccp_analyze(this, worklist, m);
1897       }
1898     }
1899   }
1900 }
1901 
1902 //------------------------------do_transform-----------------------------------
1903 // Top level driver for the recursive transformer
do_transform()1904 void PhaseCCP::do_transform() {
1905   // Correct leaves of new-space Nodes; they point to old-space.
1906   C->set_root( transform(C->root())->as_Root() );
1907   assert( C->top(),  "missing TOP node" );
1908   assert( C->root(), "missing root" );
1909 }
1910 
1911 //------------------------------transform--------------------------------------
1912 // Given a Node in old-space, clone him into new-space.
1913 // Convert any of his old-space children into new-space children.
transform(Node * n)1914 Node *PhaseCCP::transform( Node *n ) {
1915   Node *new_node = _nodes[n->_idx]; // Check for transformed node
1916   if( new_node != NULL )
1917     return new_node;                // Been there, done that, return old answer
1918   new_node = transform_once(n);     // Check for constant
1919   _nodes.map( n->_idx, new_node );  // Flag as having been cloned
1920 
1921   // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1922   GrowableArray <Node *> trstack(C->live_nodes() >> 1);
1923 
1924   trstack.push(new_node);           // Process children of cloned node
1925   while ( trstack.is_nonempty() ) {
1926     Node *clone = trstack.pop();
1927     uint cnt = clone->req();
1928     for( uint i = 0; i < cnt; i++ ) {          // For all inputs do
1929       Node *input = clone->in(i);
1930       if( input != NULL ) {                    // Ignore NULLs
1931         Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1932         if( new_input == NULL ) {
1933           new_input = transform_once(input);   // Check for constant
1934           _nodes.map( input->_idx, new_input );// Flag as having been cloned
1935           trstack.push(new_input);
1936         }
1937         assert( new_input == clone->in(i), "insanity check");
1938       }
1939     }
1940   }
1941   return new_node;
1942 }
1943 
1944 
1945 //------------------------------transform_once---------------------------------
1946 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
transform_once(Node * n)1947 Node *PhaseCCP::transform_once( Node *n ) {
1948   const Type *t = type(n);
1949   // Constant?  Use constant Node instead
1950   if( t->singleton() ) {
1951     Node *nn = n;               // Default is to return the original constant
1952     if( t == Type::TOP ) {
1953       // cache my top node on the Compile instance
1954       if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1955         C->set_cached_top_node(ConNode::make(Type::TOP));
1956         set_type(C->top(), Type::TOP);
1957       }
1958       nn = C->top();
1959     }
1960     if( !n->is_Con() ) {
1961       if( t != Type::TOP ) {
1962         nn = makecon(t);        // ConNode::make(t);
1963         NOT_PRODUCT( inc_constants(); )
1964       } else if( n->is_Region() ) { // Unreachable region
1965         // Note: nn == C->top()
1966         n->set_req(0, NULL);        // Cut selfreference
1967         bool progress = true;
1968         uint max = n->outcnt();
1969         DUIterator i;
1970         while (progress) {
1971           progress = false;
1972           // Eagerly remove dead phis to avoid phis copies creation.
1973           for (i = n->outs(); n->has_out(i); i++) {
1974             Node* m = n->out(i);
1975             if (m->is_Phi()) {
1976               assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1977               replace_node(m, nn);
1978               if (max != n->outcnt()) {
1979                 progress = true;
1980                 i = n->refresh_out_pos(i);
1981                 max = n->outcnt();
1982               }
1983             }
1984           }
1985         }
1986       }
1987       replace_node(n,nn);       // Update DefUse edges for new constant
1988     }
1989     return nn;
1990   }
1991 
1992   // If x is a TypeNode, capture any more-precise type permanently into Node
1993   if (t != n->bottom_type()) {
1994     hash_delete(n);             // changing bottom type may force a rehash
1995     n->raise_bottom_type(t);
1996     _worklist.push(n);          // n re-enters the hash table via the worklist
1997   }
1998 
1999   // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
2000   switch( n->Opcode() ) {
2001   case Op_FastLock:      // Revisit FastLocks for lock coarsening
2002   case Op_If:
2003   case Op_CountedLoopEnd:
2004   case Op_Region:
2005   case Op_Loop:
2006   case Op_CountedLoop:
2007   case Op_Conv2B:
2008   case Op_Opaque1:
2009   case Op_Opaque2:
2010     _worklist.push(n);
2011     break;
2012   default:
2013     break;
2014   }
2015 
2016   return  n;
2017 }
2018 
2019 //---------------------------------saturate------------------------------------
saturate(const Type * new_type,const Type * old_type,const Type * limit_type) const2020 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
2021                                const Type* limit_type) const {
2022   const Type* wide_type = new_type->widen(old_type, limit_type);
2023   if (wide_type != new_type) {          // did we widen?
2024     // If so, we may have widened beyond the limit type.  Clip it back down.
2025     new_type = wide_type->filter(limit_type);
2026   }
2027   return new_type;
2028 }
2029 
2030 //------------------------------print_statistics-------------------------------
2031 #ifndef PRODUCT
print_statistics()2032 void PhaseCCP::print_statistics() {
2033   tty->print_cr("CCP: %d  constants found: %d", _total_invokes, _total_constants);
2034 }
2035 #endif
2036 
2037 
2038 //=============================================================================
2039 #ifndef PRODUCT
2040 uint PhasePeephole::_total_peepholes = 0;
2041 #endif
2042 //------------------------------PhasePeephole----------------------------------
2043 // Conditional Constant Propagation, ala Wegman & Zadeck
PhasePeephole(PhaseRegAlloc * regalloc,PhaseCFG & cfg)2044 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
2045   : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
2046   NOT_PRODUCT( clear_peepholes(); )
2047 }
2048 
2049 #ifndef PRODUCT
2050 //------------------------------~PhasePeephole---------------------------------
~PhasePeephole()2051 PhasePeephole::~PhasePeephole() {
2052   _total_peepholes += count_peepholes();
2053 }
2054 #endif
2055 
2056 //------------------------------transform--------------------------------------
transform(Node * n)2057 Node *PhasePeephole::transform( Node *n ) {
2058   ShouldNotCallThis();
2059   return NULL;
2060 }
2061 
2062 //------------------------------do_transform-----------------------------------
do_transform()2063 void PhasePeephole::do_transform() {
2064   bool method_name_not_printed = true;
2065 
2066   // Examine each basic block
2067   for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
2068     Block* block = _cfg.get_block(block_number);
2069     bool block_not_printed = true;
2070 
2071     // and each instruction within a block
2072     uint end_index = block->number_of_nodes();
2073     // block->end_idx() not valid after PhaseRegAlloc
2074     for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
2075       Node     *n = block->get_node(instruction_index);
2076       if( n->is_Mach() ) {
2077         MachNode *m = n->as_Mach();
2078         int deleted_count = 0;
2079         // check for peephole opportunities
2080         MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count);
2081         if( m2 != NULL ) {
2082 #ifndef PRODUCT
2083           if( PrintOptoPeephole ) {
2084             // Print method, first time only
2085             if( C->method() && method_name_not_printed ) {
2086               C->method()->print_short_name(); tty->cr();
2087               method_name_not_printed = false;
2088             }
2089             // Print this block
2090             if( Verbose && block_not_printed) {
2091               tty->print_cr("in block");
2092               block->dump();
2093               block_not_printed = false;
2094             }
2095             // Print instructions being deleted
2096             for( int i = (deleted_count - 1); i >= 0; --i ) {
2097               block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
2098             }
2099             tty->print_cr("replaced with");
2100             // Print new instruction
2101             m2->format(_regalloc);
2102             tty->print("\n\n");
2103           }
2104 #endif
2105           // Remove old nodes from basic block and update instruction_index
2106           // (old nodes still exist and may have edges pointing to them
2107           //  as register allocation info is stored in the allocator using
2108           //  the node index to live range mappings.)
2109           uint safe_instruction_index = (instruction_index - deleted_count);
2110           for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
2111             block->remove_node( instruction_index );
2112           }
2113           // install new node after safe_instruction_index
2114           block->insert_node(m2, safe_instruction_index + 1);
2115           end_index = block->number_of_nodes() - 1; // Recompute new block size
2116           NOT_PRODUCT( inc_peepholes(); )
2117         }
2118       }
2119     }
2120   }
2121 }
2122 
2123 //------------------------------print_statistics-------------------------------
2124 #ifndef PRODUCT
print_statistics()2125 void PhasePeephole::print_statistics() {
2126   tty->print_cr("Peephole: peephole rules applied: %d",  _total_peepholes);
2127 }
2128 #endif
2129 
2130 
2131 //=============================================================================
2132 //------------------------------set_req_X--------------------------------------
set_req_X(uint i,Node * n,PhaseIterGVN * igvn)2133 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2134   assert( is_not_dead(n), "can not use dead node");
2135   assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2136   Node *old = in(i);
2137   set_req(i, n);
2138 
2139   // old goes dead?
2140   if( old ) {
2141     switch (old->outcnt()) {
2142     case 0:
2143       // Put into the worklist to kill later. We do not kill it now because the
2144       // recursive kill will delete the current node (this) if dead-loop exists
2145       if (!old->is_top())
2146         igvn->_worklist.push( old );
2147       break;
2148     case 1:
2149       if( old->is_Store() || old->has_special_unique_user() )
2150         igvn->add_users_to_worklist( old );
2151       break;
2152     case 2:
2153       if( old->is_Store() )
2154         igvn->add_users_to_worklist( old );
2155       if( old->Opcode() == Op_Region )
2156         igvn->_worklist.push(old);
2157       break;
2158     case 3:
2159       if( old->Opcode() == Op_Region ) {
2160         igvn->_worklist.push(old);
2161         igvn->add_users_to_worklist( old );
2162       }
2163       break;
2164     default:
2165       break;
2166     }
2167 
2168     BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
2169   }
2170 
2171 }
2172 
2173 //-------------------------------replace_by-----------------------------------
2174 // Using def-use info, replace one node for another.  Follow the def-use info
2175 // to all users of the OLD node.  Then make all uses point to the NEW node.
replace_by(Node * new_node)2176 void Node::replace_by(Node *new_node) {
2177   assert(!is_top(), "top node has no DU info");
2178   for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2179     Node* use = last_out(i);
2180     uint uses_found = 0;
2181     for (uint j = 0; j < use->len(); j++) {
2182       if (use->in(j) == this) {
2183         if (j < use->req())
2184               use->set_req(j, new_node);
2185         else  use->set_prec(j, new_node);
2186         uses_found++;
2187       }
2188     }
2189     i -= uses_found;    // we deleted 1 or more copies of this edge
2190   }
2191 }
2192 
2193 //=============================================================================
2194 //-----------------------------------------------------------------------------
grow(uint i)2195 void Type_Array::grow( uint i ) {
2196   if( !_max ) {
2197     _max = 1;
2198     _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2199     _types[0] = NULL;
2200   }
2201   uint old = _max;
2202   while( i >= _max ) _max <<= 1;        // Double to fit
2203   _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2204   memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2205 }
2206 
2207 //------------------------------dump-------------------------------------------
2208 #ifndef PRODUCT
dump() const2209 void Type_Array::dump() const {
2210   uint max = Size();
2211   for( uint i = 0; i < max; i++ ) {
2212     if( _types[i] != NULL ) {
2213       tty->print("  %d\t== ", i); _types[i]->dump(); tty->cr();
2214     }
2215   }
2216 }
2217 #endif
2218