1 /*
2 * Copyright (c) 1997, 2022, 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---------------------------------------
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---------------------------------------
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---------------------------------------
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
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
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
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
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
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
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
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
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
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
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
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
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
387 NodeHash::~NodeHash() {
388 // Unlock all nodes upon destruction of table.
389 if (_table != (Node**)badAddress) clear();
390 }
391
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.
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->disconnect_useless_nodes(_useful, worklist);
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.
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
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
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---------------------------------
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---------------------------------
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
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
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--------------------------------
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--------------------------------
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
649 void PhaseTransform::dump_old2new_map() const {
650 _nodes.dump();
651 }
652
653 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-------------------------------------
665 void PhaseTransform::dump_types( ) const {
666 _types.dump();
667 }
668
669 //------------------------------dump_nodes_and_types---------------------------
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---------------------
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"
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"
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
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----------------------------------------
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.
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)))"
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.
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 }
793 ConNode* PhaseTransform::integercon(jlong l, BasicType bt) {
794 if (bt == T_INT) {
795 return intcon(checked_cast<jint>(l));
796 }
797 assert(bt == T_LONG, "not an integer");
798 return longcon(l);
799 }
800
801
802 //------------------------------zerocon-----------------------------------------
803 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
804 ConNode* PhaseTransform::zerocon(BasicType bt) {
805 assert((uint)bt <= _zcon_max, "domain check");
806 ConNode* zcon = _zcons[bt];
807 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
808 return zcon;
809 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
810 _zcons[bt] = zcon;
811 return zcon;
812 }
813
814
815
816 //=============================================================================
817 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
818 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
819 if (i == NULL) {
820 i = k->Ideal(this, can_reshape);
821 }
822 return i;
823 }
824
825 //------------------------------transform--------------------------------------
826 // Return a node which computes the same function as this node, but in a
827 // faster or cheaper fashion.
828 Node *PhaseGVN::transform( Node *n ) {
829 return transform_no_reclaim(n);
830 }
831
832 //------------------------------transform--------------------------------------
833 // Return a node which computes the same function as this node, but
834 // in a faster or cheaper fashion.
835 Node *PhaseGVN::transform_no_reclaim(Node *n) {
836 NOT_PRODUCT( set_transforms(); )
837
838 // Apply the Ideal call in a loop until it no longer applies
839 Node* k = n;
840 Node* i = apply_ideal(k, /*can_reshape=*/false);
841 NOT_PRODUCT(uint loop_count = 1;)
842 while (i != NULL) {
843 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
844 k = i;
845 #ifdef ASSERT
846 if (loop_count >= K + C->live_nodes()) {
847 dump_infinite_loop_info(i, "PhaseGVN::transform_no_reclaim");
848 }
849 #endif
850 i = apply_ideal(k, /*can_reshape=*/false);
851 NOT_PRODUCT(loop_count++;)
852 }
853 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
854
855 // If brand new node, make space in type array.
856 ensure_type_or_null(k);
857
858 // Since I just called 'Value' to compute the set of run-time values
859 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
860 // cache Value. Later requests for the local phase->type of this Node can
861 // use the cached Value instead of suffering with 'bottom_type'.
862 const Type* t = k->Value(this); // Get runtime Value set
863 assert(t != NULL, "value sanity");
864 if (type_or_null(k) != t) {
865 #ifndef PRODUCT
866 // Do not count initial visit to node as a transformation
867 if (type_or_null(k) == NULL) {
868 inc_new_values();
869 set_progress();
870 }
871 #endif
872 set_type(k, t);
873 // If k is a TypeNode, capture any more-precise type permanently into Node
874 k->raise_bottom_type(t);
875 }
876
877 if (t->singleton() && !k->is_Con()) {
878 NOT_PRODUCT(set_progress();)
879 return makecon(t); // Turn into a constant
880 }
881
882 // Now check for Identities
883 i = k->Identity(this); // Look for a nearby replacement
884 if (i != k) { // Found? Return replacement!
885 NOT_PRODUCT(set_progress();)
886 return i;
887 }
888
889 // Global Value Numbering
890 i = hash_find_insert(k); // Insert if new
891 if (i && (i != k)) {
892 // Return the pre-existing node
893 NOT_PRODUCT(set_progress();)
894 return i;
895 }
896
897 // Return Idealized original
898 return k;
899 }
900
901 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
902 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
903 return false;
904 }
905 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
906 return false;
907 }
908 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
909 int i = 0;
910 while (d != n) {
911 n = IfNode::up_one_dom(n, linear_only);
912 i++;
913 if (n == NULL || i >= 100) {
914 return false;
915 }
916 }
917 return true;
918 }
919
920 #ifdef ASSERT
921 //------------------------------dead_loop_check--------------------------------
922 // Check for a simple dead loop when a data node references itself directly
923 // or through an other data node excluding cons and phis.
924 void PhaseGVN::dead_loop_check( Node *n ) {
925 // Phi may reference itself in a loop
926 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
927 // Do 2 levels check and only data inputs.
928 bool no_dead_loop = true;
929 uint cnt = n->req();
930 for (uint i = 1; i < cnt && no_dead_loop; i++) {
931 Node *in = n->in(i);
932 if (in == n) {
933 no_dead_loop = false;
934 } else if (in != NULL && !in->is_dead_loop_safe()) {
935 uint icnt = in->req();
936 for (uint j = 1; j < icnt && no_dead_loop; j++) {
937 if (in->in(j) == n || in->in(j) == in)
938 no_dead_loop = false;
939 }
940 }
941 }
942 if (!no_dead_loop) n->dump(3);
943 assert(no_dead_loop, "dead loop detected");
944 }
945 }
946
947
948 /**
949 * Dumps information that can help to debug the problem. A debug
950 * build fails with an assert.
951 */
952 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
953 n->dump(4);
954 assert(false, "infinite loop in %s", where);
955 }
956 #endif
957
958 //=============================================================================
959 //------------------------------PhaseIterGVN-----------------------------------
960 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
961 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : PhaseGVN(igvn),
962 _delay_transform(igvn->_delay_transform),
963 _stack(igvn->_stack ),
964 _worklist(igvn->_worklist)
965 {
966 _iterGVN = true;
967 }
968
969 //------------------------------PhaseIterGVN-----------------------------------
970 // Initialize with previous PhaseGVN info from Parser
971 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : PhaseGVN(gvn),
972 _delay_transform(false),
973 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that
974 // the constructor is used in incremental inlining, this consumes too much memory:
975 // _stack(C->live_nodes() >> 1),
976 // So, as a band-aid, we replace this by:
977 _stack(C->comp_arena(), 32),
978 _worklist(*C->for_igvn())
979 {
980 _iterGVN = true;
981 uint max;
982
983 // Dead nodes in the hash table inherited from GVN were not treated as
984 // roots during def-use info creation; hence they represent an invisible
985 // use. Clear them out.
986 max = _table.size();
987 for( uint i = 0; i < max; ++i ) {
988 Node *n = _table.at(i);
989 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
990 if( n->is_top() ) continue;
991 // If remove_useless_nodes() has run, we expect no such nodes left.
992 assert(false, "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 void PhaseIterGVN::shuffle_worklist() {
1012 if (_worklist.size() < 2) return;
1013 for (uint i = _worklist.size() - 1; i >= 1; i--) {
1014 uint j = C->random() % (i + 1);
1015 swap(_worklist.adr()[i], _worklist.adr()[j]);
1016 }
1017 }
1018
1019 #ifndef PRODUCT
1020 void PhaseIterGVN::verify_step(Node* n) {
1021 if (VerifyIterativeGVN) {
1022 ResourceMark rm;
1023 VectorSet visited;
1024 Node_List worklist;
1025
1026 _verify_window[_verify_counter % _verify_window_size] = n;
1027 ++_verify_counter;
1028 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
1029 ++_verify_full_passes;
1030 worklist.push(C->root());
1031 Node::verify(-1, visited, worklist);
1032 return;
1033 }
1034 for (int i = 0; i < _verify_window_size; i++) {
1035 Node* n = _verify_window[i];
1036 if (n == NULL) {
1037 continue;
1038 }
1039 if (n->in(0) == NodeSentinel) { // xform_idom
1040 _verify_window[i] = n->in(1);
1041 --i;
1042 continue;
1043 }
1044 // Typical fanout is 1-2, so this call visits about 6 nodes.
1045 if (!visited.test_set(n->_idx)) {
1046 worklist.push(n);
1047 }
1048 }
1049 Node::verify(4, visited, worklist);
1050 }
1051 }
1052
1053 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
1054 if (TraceIterativeGVN) {
1055 uint wlsize = _worklist.size();
1056 const Type* newtype = type_or_null(n);
1057 if (nn != n) {
1058 // print old node
1059 tty->print("< ");
1060 if (oldtype != newtype && oldtype != NULL) {
1061 oldtype->dump();
1062 }
1063 do { tty->print("\t"); } while (tty->position() < 16);
1064 tty->print("<");
1065 n->dump();
1066 }
1067 if (oldtype != newtype || nn != n) {
1068 // print new node and/or new type
1069 if (oldtype == NULL) {
1070 tty->print("* ");
1071 } else if (nn != n) {
1072 tty->print("> ");
1073 } else {
1074 tty->print("= ");
1075 }
1076 if (newtype == NULL) {
1077 tty->print("null");
1078 } else {
1079 newtype->dump();
1080 }
1081 do { tty->print("\t"); } while (tty->position() < 16);
1082 nn->dump();
1083 }
1084 if (Verbose && wlsize < _worklist.size()) {
1085 tty->print(" Push {");
1086 while (wlsize != _worklist.size()) {
1087 Node* pushed = _worklist.at(wlsize++);
1088 tty->print(" %d", pushed->_idx);
1089 }
1090 tty->print_cr(" }");
1091 }
1092 if (nn != n) {
1093 // ignore n, it might be subsumed
1094 verify_step((Node*) NULL);
1095 }
1096 }
1097 }
1098
1099 void PhaseIterGVN::init_verifyPhaseIterGVN() {
1100 _verify_counter = 0;
1101 _verify_full_passes = 0;
1102 for (int i = 0; i < _verify_window_size; i++) {
1103 _verify_window[i] = NULL;
1104 }
1105 #ifdef ASSERT
1106 // Verify that all modified nodes are on _worklist
1107 Unique_Node_List* modified_list = C->modified_nodes();
1108 while (modified_list != NULL && modified_list->size()) {
1109 Node* n = modified_list->pop();
1110 if (!n->is_Con() && !_worklist.member(n)) {
1111 n->dump();
1112 fatal("modified node is not on IGVN._worklist");
1113 }
1114 }
1115 #endif
1116 }
1117
1118 void PhaseIterGVN::verify_PhaseIterGVN() {
1119 #ifdef ASSERT
1120 // Verify nodes with changed inputs.
1121 Unique_Node_List* modified_list = C->modified_nodes();
1122 while (modified_list != NULL && modified_list->size()) {
1123 Node* n = modified_list->pop();
1124 if (!n->is_Con()) { // skip Con nodes
1125 n->dump();
1126 fatal("modified node was not processed by IGVN.transform_old()");
1127 }
1128 }
1129 #endif
1130
1131 C->verify_graph_edges();
1132 if (VerifyIterativeGVN && PrintOpto) {
1133 if (_verify_counter == _verify_full_passes) {
1134 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1135 (int) _verify_full_passes);
1136 } else {
1137 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1138 (int) _verify_counter, (int) _verify_full_passes);
1139 }
1140 }
1141
1142 #ifdef ASSERT
1143 if (modified_list != NULL) {
1144 while (modified_list->size() > 0) {
1145 Node* n = modified_list->pop();
1146 n->dump();
1147 assert(false, "VerifyIterativeGVN: new modified node was added");
1148 }
1149 }
1150 #endif
1151 }
1152 #endif /* PRODUCT */
1153
1154 #ifdef ASSERT
1155 /**
1156 * Dumps information that can help to debug the problem. A debug
1157 * build fails with an assert.
1158 */
1159 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1160 n->dump(4);
1161 _worklist.dump();
1162 assert(false, "infinite loop in %s", where);
1163 }
1164
1165 /**
1166 * Prints out information about IGVN if the 'verbose' option is used.
1167 */
1168 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1169 if (TraceIterativeGVN && Verbose) {
1170 tty->print(" Pop ");
1171 n->dump();
1172 if ((num_processed % 100) == 0) {
1173 _worklist.print_set();
1174 }
1175 }
1176 }
1177 #endif /* ASSERT */
1178
1179 void PhaseIterGVN::optimize() {
1180 DEBUG_ONLY(uint num_processed = 0;)
1181 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1182 if (StressIGVN) {
1183 shuffle_worklist();
1184 }
1185
1186 uint loop_count = 0;
1187 // Pull from worklist and transform the node. If the node has changed,
1188 // update edge info and put uses on worklist.
1189 while(_worklist.size()) {
1190 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1191 return;
1192 }
1193 Node* n = _worklist.pop();
1194 if (loop_count >= K * C->live_nodes()) {
1195 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1196 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1197 return;
1198 }
1199 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1200 if (n->outcnt() != 0) {
1201 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1202 // Do the transformation
1203 Node* nn = transform_old(n);
1204 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1205 } else if (!n->is_top()) {
1206 remove_dead_node(n);
1207 }
1208 loop_count++;
1209 }
1210 NOT_PRODUCT(verify_PhaseIterGVN();)
1211 }
1212
1213
1214 /**
1215 * Register a new node with the optimizer. Update the types array, the def-use
1216 * info. Put on worklist.
1217 */
1218 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1219 set_type_bottom(n);
1220 _worklist.push(n);
1221 if (orig != NULL) C->copy_node_notes_to(n, orig);
1222 return n;
1223 }
1224
1225 //------------------------------transform--------------------------------------
1226 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
1227 Node *PhaseIterGVN::transform( Node *n ) {
1228 if (_delay_transform) {
1229 // Register the node but don't optimize for now
1230 register_new_node_with_optimizer(n);
1231 return n;
1232 }
1233
1234 // If brand new node, make space in type array, and give it a type.
1235 ensure_type_or_null(n);
1236 if (type_or_null(n) == NULL) {
1237 set_type_bottom(n);
1238 }
1239
1240 return transform_old(n);
1241 }
1242
1243 Node *PhaseIterGVN::transform_old(Node* n) {
1244 NOT_PRODUCT(set_transforms());
1245 // Remove 'n' from hash table in case it gets modified
1246 _table.hash_delete(n);
1247 if (VerifyIterativeGVN) {
1248 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1249 }
1250
1251 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
1252 if (n->is_Cmp()) {
1253 add_users_to_worklist(n);
1254 }
1255
1256 // Apply the Ideal call in a loop until it no longer applies
1257 Node* k = n;
1258 DEBUG_ONLY(dead_loop_check(k);)
1259 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1260 C->remove_modified_node(k);
1261 Node* i = apply_ideal(k, /*can_reshape=*/true);
1262 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1263 #ifndef PRODUCT
1264 verify_step(k);
1265 #endif
1266
1267 DEBUG_ONLY(uint loop_count = 1;)
1268 while (i != NULL) {
1269 #ifdef ASSERT
1270 if (loop_count >= K + C->live_nodes()) {
1271 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
1272 }
1273 #endif
1274 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1275 // Made a change; put users of original Node on worklist
1276 add_users_to_worklist(k);
1277 // Replacing root of transform tree?
1278 if (k != i) {
1279 // Make users of old Node now use new.
1280 subsume_node(k, i);
1281 k = i;
1282 }
1283 DEBUG_ONLY(dead_loop_check(k);)
1284 // Try idealizing again
1285 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1286 C->remove_modified_node(k);
1287 i = apply_ideal(k, /*can_reshape=*/true);
1288 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1289 #ifndef PRODUCT
1290 verify_step(k);
1291 #endif
1292 DEBUG_ONLY(loop_count++;)
1293 }
1294
1295 // If brand new node, make space in type array.
1296 ensure_type_or_null(k);
1297
1298 // See what kind of values 'k' takes on at runtime
1299 const Type* t = k->Value(this);
1300 assert(t != NULL, "value sanity");
1301
1302 // Since I just called 'Value' to compute the set of run-time values
1303 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1304 // cache Value. Later requests for the local phase->type of this Node can
1305 // use the cached Value instead of suffering with 'bottom_type'.
1306 if (type_or_null(k) != t) {
1307 #ifndef PRODUCT
1308 inc_new_values();
1309 set_progress();
1310 #endif
1311 set_type(k, t);
1312 // If k is a TypeNode, capture any more-precise type permanently into Node
1313 k->raise_bottom_type(t);
1314 // Move users of node to worklist
1315 add_users_to_worklist(k);
1316 }
1317 // If 'k' computes a constant, replace it with a constant
1318 if (t->singleton() && !k->is_Con()) {
1319 NOT_PRODUCT(set_progress();)
1320 Node* con = makecon(t); // Make a constant
1321 add_users_to_worklist(k);
1322 subsume_node(k, con); // Everybody using k now uses con
1323 return con;
1324 }
1325
1326 // Now check for Identities
1327 i = k->Identity(this); // Look for a nearby replacement
1328 if (i != k) { // Found? Return replacement!
1329 NOT_PRODUCT(set_progress();)
1330 add_users_to_worklist(k);
1331 subsume_node(k, i); // Everybody using k now uses i
1332 return i;
1333 }
1334
1335 // Global Value Numbering
1336 i = hash_find_insert(k); // Check for pre-existing node
1337 if (i && (i != k)) {
1338 // Return the pre-existing node if it isn't dead
1339 NOT_PRODUCT(set_progress();)
1340 add_users_to_worklist(k);
1341 subsume_node(k, i); // Everybody using k now uses i
1342 return i;
1343 }
1344
1345 // Return Idealized original
1346 return k;
1347 }
1348
1349 //---------------------------------saturate------------------------------------
1350 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1351 const Type* limit_type) const {
1352 return new_type->narrow(old_type);
1353 }
1354
1355 //------------------------------remove_globally_dead_node----------------------
1356 // Kill a globally dead Node. All uses are also globally dead and are
1357 // aggressively trimmed.
1358 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1359 enum DeleteProgress {
1360 PROCESS_INPUTS,
1361 PROCESS_OUTPUTS
1362 };
1363 assert(_stack.is_empty(), "not empty");
1364 _stack.push(dead, PROCESS_INPUTS);
1365
1366 while (_stack.is_nonempty()) {
1367 dead = _stack.node();
1368 if (dead->Opcode() == Op_SafePoint) {
1369 dead->as_SafePoint()->disconnect_from_root(this);
1370 }
1371 uint progress_state = _stack.index();
1372 assert(dead != C->root(), "killing root, eh?");
1373 assert(!dead->is_top(), "add check for top when pushing");
1374 NOT_PRODUCT( set_progress(); )
1375 if (progress_state == PROCESS_INPUTS) {
1376 // After following inputs, continue to outputs
1377 _stack.set_index(PROCESS_OUTPUTS);
1378 if (!dead->is_Con()) { // Don't kill cons but uses
1379 bool recurse = false;
1380 // Remove from hash table
1381 _table.hash_delete( dead );
1382 // Smash all inputs to 'dead', isolating him completely
1383 for (uint i = 0; i < dead->req(); i++) {
1384 Node *in = dead->in(i);
1385 if (in != NULL && in != C->top()) { // Points to something?
1386 int nrep = dead->replace_edge(in, NULL, this); // Kill edges
1387 assert((nrep > 0), "sanity");
1388 if (in->outcnt() == 0) { // Made input go dead?
1389 _stack.push(in, PROCESS_INPUTS); // Recursively remove
1390 recurse = true;
1391 } else if (in->outcnt() == 1 &&
1392 in->has_special_unique_user()) {
1393 _worklist.push(in->unique_out());
1394 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1395 if (in->Opcode() == Op_Region) {
1396 _worklist.push(in);
1397 } else if (in->is_Store()) {
1398 DUIterator_Fast imax, i = in->fast_outs(imax);
1399 _worklist.push(in->fast_out(i));
1400 i++;
1401 if (in->outcnt() == 2) {
1402 _worklist.push(in->fast_out(i));
1403 i++;
1404 }
1405 assert(!(i < imax), "sanity");
1406 }
1407 } else {
1408 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
1409 }
1410 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1411 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1412 // A Load that directly follows an InitializeNode is
1413 // going away. The Stores that follow are candidates
1414 // again to be captured by the InitializeNode.
1415 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1416 Node *n = in->fast_out(j);
1417 if (n->is_Store()) {
1418 _worklist.push(n);
1419 }
1420 }
1421 }
1422 } // if (in != NULL && in != C->top())
1423 } // for (uint i = 0; i < dead->req(); i++)
1424 if (recurse) {
1425 continue;
1426 }
1427 } // if (!dead->is_Con())
1428 } // if (progress_state == PROCESS_INPUTS)
1429
1430 // Aggressively kill globally dead uses
1431 // (Rather than pushing all the outs at once, we push one at a time,
1432 // plus the parent to resume later, because of the indefinite number
1433 // of edge deletions per loop trip.)
1434 if (dead->outcnt() > 0) {
1435 // Recursively remove output edges
1436 _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1437 } else {
1438 // Finished disconnecting all input and output edges.
1439 _stack.pop();
1440 // Remove dead node from iterative worklist
1441 _worklist.remove(dead);
1442 C->remove_useless_node(dead);
1443 }
1444 } // while (_stack.is_nonempty())
1445 }
1446
1447 //------------------------------subsume_node-----------------------------------
1448 // Remove users from node 'old' and add them to node 'nn'.
1449 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1450 if (old->Opcode() == Op_SafePoint) {
1451 old->as_SafePoint()->disconnect_from_root(this);
1452 }
1453 assert( old != hash_find(old), "should already been removed" );
1454 assert( old != C->top(), "cannot subsume top node");
1455 // Copy debug or profile information to the new version:
1456 C->copy_node_notes_to(nn, old);
1457 // Move users of node 'old' to node 'nn'
1458 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1459 Node* use = old->last_out(i); // for each use...
1460 // use might need re-hashing (but it won't if it's a new node)
1461 rehash_node_delayed(use);
1462 // Update use-def info as well
1463 // We remove all occurrences of old within use->in,
1464 // so as to avoid rehashing any node more than once.
1465 // The hash table probe swamps any outer loop overhead.
1466 uint num_edges = 0;
1467 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1468 if (use->in(j) == old) {
1469 use->set_req(j, nn);
1470 ++num_edges;
1471 }
1472 }
1473 i -= num_edges; // we deleted 1 or more copies of this edge
1474 }
1475
1476 // Search for instance field data PhiNodes in the same region pointing to the old
1477 // memory PhiNode and update their instance memory ids to point to the new node.
1478 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) {
1479 Node* region = old->in(0);
1480 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1481 PhiNode* phi = region->fast_out(i)->isa_Phi();
1482 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) {
1483 phi->set_inst_mem_id((int)nn->_idx);
1484 }
1485 }
1486 }
1487
1488 // Smash all inputs to 'old', isolating him completely
1489 Node *temp = new Node(1);
1490 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
1491 remove_dead_node( old );
1492 temp->del_req(0); // Yank bogus edge
1493 if (nn != NULL && nn->outcnt() == 0) {
1494 _worklist.push(nn);
1495 }
1496 #ifndef PRODUCT
1497 if( VerifyIterativeGVN ) {
1498 for ( int i = 0; i < _verify_window_size; i++ ) {
1499 if ( _verify_window[i] == old )
1500 _verify_window[i] = nn;
1501 }
1502 }
1503 #endif
1504 temp->destruct(this); // reuse the _idx of this little guy
1505 }
1506
1507 //------------------------------add_users_to_worklist--------------------------
1508 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1509 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1510 _worklist.push(n->fast_out(i)); // Push on worklist
1511 }
1512 }
1513
1514 // Return counted loop Phi if as a counted loop exit condition, cmp
1515 // compares the induction variable with n
1516 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
1517 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1518 Node* bol = cmp->fast_out(i);
1519 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1520 Node* iff = bol->fast_out(i2);
1521 if (iff->is_BaseCountedLoopEnd()) {
1522 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
1523 if (cle->limit() == n) {
1524 PhiNode* phi = cle->phi();
1525 if (phi != NULL) {
1526 return phi;
1527 }
1528 }
1529 }
1530 }
1531 }
1532 return NULL;
1533 }
1534
1535 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1536 add_users_to_worklist0(n);
1537
1538 // Move users of node to worklist
1539 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1540 Node* use = n->fast_out(i); // Get use
1541
1542 if( use->is_Multi() || // Multi-definer? Push projs on worklist
1543 use->is_Store() ) // Enable store/load same address
1544 add_users_to_worklist0(use);
1545
1546 // If we changed the receiver type to a call, we need to revisit
1547 // the Catch following the call. It's looking for a non-NULL
1548 // receiver to know when to enable the regular fall-through path
1549 // in addition to the NullPtrException path.
1550 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1551 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1552 if (p != NULL) {
1553 add_users_to_worklist0(p);
1554 }
1555 }
1556
1557 uint use_op = use->Opcode();
1558 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
1559 add_users_to_worklist(use); // Put Bool on worklist
1560 if (use->outcnt() > 0) {
1561 Node* bol = use->raw_out(0);
1562 if (bol->outcnt() > 0) {
1563 Node* iff = bol->raw_out(0);
1564 if (iff->outcnt() == 2) {
1565 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1566 // phi merging either 0 or 1 onto the worklist
1567 Node* ifproj0 = iff->raw_out(0);
1568 Node* ifproj1 = iff->raw_out(1);
1569 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1570 Node* region0 = ifproj0->raw_out(0);
1571 Node* region1 = ifproj1->raw_out(0);
1572 if( region0 == region1 )
1573 add_users_to_worklist0(region0);
1574 }
1575 }
1576 }
1577 }
1578 if (use_op == Op_CmpI) {
1579 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n);
1580 if (phi != NULL) {
1581 // If an opaque node feeds into the limit condition of a
1582 // CountedLoop, we need to process the Phi node for the
1583 // induction variable when the opaque node is removed:
1584 // the range of values taken by the Phi is now known and
1585 // so its type is also known.
1586 _worklist.push(phi);
1587 }
1588 Node* in1 = use->in(1);
1589 for (uint i = 0; i < in1->outcnt(); i++) {
1590 if (in1->raw_out(i)->Opcode() == Op_CastII) {
1591 Node* castii = in1->raw_out(i);
1592 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) {
1593 Node* ifnode = castii->in(0)->in(0);
1594 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) {
1595 // Reprocess a CastII node that may depend on an
1596 // opaque node value when the opaque node is
1597 // removed. In case it carries a dependency we can do
1598 // a better job of computing its type.
1599 _worklist.push(castii);
1600 }
1601 }
1602 }
1603 }
1604 }
1605 }
1606
1607 // If changed Cast input, check Phi users for simple cycles
1608 if (use->is_ConstraintCast()) {
1609 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1610 Node* u = use->fast_out(i2);
1611 if (u->is_Phi())
1612 _worklist.push(u);
1613 }
1614 }
1615 // If changed LShift inputs, check RShift users for useless sign-ext
1616 if( use_op == Op_LShiftI ) {
1617 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1618 Node* u = use->fast_out(i2);
1619 if (u->Opcode() == Op_RShiftI)
1620 _worklist.push(u);
1621 }
1622 }
1623 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1624 if (use_op == Op_AddI || use_op == Op_SubI) {
1625 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1626 Node* u = use->fast_out(i2);
1627 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1628 _worklist.push(u);
1629 }
1630 }
1631 }
1632 // If changed AddP inputs, check Stores for loop invariant
1633 if( use_op == Op_AddP ) {
1634 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1635 Node* u = use->fast_out(i2);
1636 if (u->is_Mem())
1637 _worklist.push(u);
1638 }
1639 }
1640 // If changed initialization activity, check dependent Stores
1641 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1642 InitializeNode* init = use->as_Allocate()->initialization();
1643 if (init != NULL) {
1644 Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1645 if (imem != NULL) add_users_to_worklist0(imem);
1646 }
1647 }
1648 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
1649 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
1650 // to guarantee the change is not missed.
1651 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
1652 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
1653 if (p != NULL) {
1654 add_users_to_worklist0(p);
1655 }
1656 }
1657
1658 if (use_op == Op_Initialize) {
1659 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1660 if (imem != NULL) add_users_to_worklist0(imem);
1661 }
1662 // Loading the java mirror from a Klass requires two loads and the type
1663 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1664 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
1665 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1666 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1667
1668 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1669 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1670 Node* u = use->fast_out(i2);
1671 const Type* ut = u->bottom_type();
1672 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1673 if (has_load_barrier_nodes) {
1674 // Search for load barriers behind the load
1675 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1676 Node* b = u->fast_out(i3);
1677 if (bs->is_gc_barrier_node(b)) {
1678 _worklist.push(b);
1679 }
1680 }
1681 }
1682 _worklist.push(u);
1683 }
1684 }
1685 }
1686 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
1687 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
1688 if (use->outcnt() == 1) {
1689 Node* cmp = use->unique_out();
1690 _worklist.push(cmp);
1691 }
1692 }
1693 }
1694 }
1695
1696 /**
1697 * Remove the speculative part of all types that we know of
1698 */
1699 void PhaseIterGVN::remove_speculative_types() {
1700 assert(UseTypeSpeculation, "speculation is off");
1701 for (uint i = 0; i < _types.Size(); i++) {
1702 const Type* t = _types.fast_lookup(i);
1703 if (t != NULL) {
1704 _types.map(i, t->remove_speculative());
1705 }
1706 }
1707 _table.check_no_speculative_types();
1708 }
1709
1710 // Check if the type of a divisor of a Div or Mod node includes zero.
1711 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
1712 switch (n->Opcode()) {
1713 case Op_DivI:
1714 case Op_ModI: {
1715 // Type of divisor includes 0?
1716 if (type(n->in(2)) == Type::TOP) {
1717 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1718 return false;
1719 }
1720 const TypeInt* type_divisor = type(n->in(2))->is_int();
1721 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1722 }
1723 case Op_DivL:
1724 case Op_ModL: {
1725 // Type of divisor includes 0?
1726 if (type(n->in(2)) == Type::TOP) {
1727 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1728 return false;
1729 }
1730 const TypeLong* type_divisor = type(n->in(2))->is_long();
1731 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1732 }
1733 }
1734 return true;
1735 }
1736
1737 //=============================================================================
1738 #ifndef PRODUCT
1739 uint PhaseCCP::_total_invokes = 0;
1740 uint PhaseCCP::_total_constants = 0;
1741 #endif
1742 //------------------------------PhaseCCP---------------------------------------
1743 // Conditional Constant Propagation, ala Wegman & Zadeck
1744 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1745 NOT_PRODUCT( clear_constants(); )
1746 assert( _worklist.size() == 0, "" );
1747 // Clear out _nodes from IterGVN. Must be clear to transform call.
1748 _nodes.clear(); // Clear out from IterGVN
1749 analyze();
1750 }
1751
1752 #ifndef PRODUCT
1753 //------------------------------~PhaseCCP--------------------------------------
1754 PhaseCCP::~PhaseCCP() {
1755 inc_invokes();
1756 _total_constants += count_constants();
1757 }
1758 #endif
1759
1760
1761 #ifdef ASSERT
1762 static bool ccp_type_widens(const Type* t, const Type* t0) {
1763 assert(t->meet(t0) == t->remove_speculative(), "Not monotonic");
1764 switch (t->base() == t0->base() ? t->base() : Type::Top) {
1765 case Type::Int:
1766 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1767 break;
1768 case Type::Long:
1769 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1770 break;
1771 default:
1772 break;
1773 }
1774 return true;
1775 }
1776 #endif //ASSERT
1777
1778 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
1779 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
1780 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
1781 // or nodes even further down back to the worklist because their type could change as a result of the current type
1782 // change).
1783 void PhaseCCP::analyze() {
1784 // Initialize all types to TOP, optimistic analysis
1785 for (uint i = 0; i < C->unique(); i++) {
1786 _types.map(i, Type::TOP);
1787 }
1788
1789 // Push root onto worklist
1790 Unique_Node_List worklist;
1791 worklist.push(C->root());
1792 DEBUG_ONLY(Unique_Node_List worklist_verify;)
1793
1794 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
1795 _root_and_safepoints.push(C->root());
1796
1797 // Pull from worklist; compute new value; push changes out.
1798 // This loop is the meat of CCP.
1799 while (worklist.size() != 0) {
1800 Node* n = fetch_next_node(worklist);
1801 DEBUG_ONLY(worklist_verify.push(n);)
1802 if (n->is_SafePoint()) {
1803 // Make sure safepoints are processed by PhaseCCP::transform even if they are
1804 // not reachable from the bottom. Otherwise, infinite loops would be removed.
1805 _root_and_safepoints.push(n);
1806 }
1807 const Type* new_type = n->Value(this);
1808 if (new_type != type(n)) {
1809 assert(ccp_type_widens(new_type, type(n)), "ccp type must widen");
1810 dump_type_and_node(n, new_type);
1811 set_type(n, new_type);
1812 push_child_nodes_to_worklist(worklist, n);
1813 }
1814 }
1815 DEBUG_ONLY(verify_analyze(worklist_verify);)
1816 }
1817
1818 #ifdef ASSERT
1819 // For every node n on verify list, check if type(n) == n->Value()
1820 // We have a list of exceptions, see comments in code.
1821 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
1822 bool failure = false;
1823 while (worklist_verify.size()) {
1824 Node* n = worklist_verify.pop();
1825 const Type* told = type(n);
1826 const Type* tnew = n->Value(this);
1827 if (told != tnew) {
1828 // Check special cases that are ok
1829 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1830 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1831 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1832 if (t0->lo_as_long() == t1->lo_as_long() &&
1833 t0->hi_as_long() == t1->hi_as_long()) {
1834 continue; // ignore integer widen
1835 }
1836 }
1837 if (n->is_Load()) {
1838 // MemNode::can_see_stored_value looks up through many memory nodes,
1839 // which means we would need to notify modifications from far up in
1840 // the inputs all the way down to the LoadNode. We don't do that.
1841 continue;
1842 }
1843 tty->cr();
1844 tty->print_cr("Missed optimization (PhaseCCP):");
1845 n->dump_bfs(1, 0, "");
1846 tty->print_cr("Current type:");
1847 told->dump_on(tty);
1848 tty->cr();
1849 tty->print_cr("Optimized type:");
1850 tnew->dump_on(tty);
1851 tty->cr();
1852 failure = true;
1853 }
1854 }
1855 // If we get this assert, check why the reported nodes were not processed again in CCP.
1856 // We should either make sure that these nodes are properly added back to the CCP worklist
1857 // in PhaseCCP::push_child_nodes_to_worklist() to update their type or add an exception
1858 // in the verification code above if that is not possible for some reason (like Load nodes).
1859 assert(!failure, "Missed optimization opportunity in PhaseCCP");
1860 }
1861 #endif
1862
1863 // Fetch next node from worklist to be examined in this iteration.
1864 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
1865 if (StressCCP) {
1866 return worklist.remove(C->random() % worklist.size());
1867 } else {
1868 return worklist.pop();
1869 }
1870 }
1871
1872 #ifndef PRODUCT
1873 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
1874 if (TracePhaseCCP) {
1875 t->dump();
1876 do {
1877 tty->print("\t");
1878 } while (tty->position() < 16);
1879 n->dump();
1880 }
1881 }
1882 #endif
1883
1884 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
1885 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
1886 // of the new type of 'n'. Push these nodes to the worklist.
1887 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
1888 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1889 Node* use = n->fast_out(i);
1890 push_if_not_bottom_type(worklist, use);
1891 push_more_uses(worklist, n, use);
1892 }
1893 }
1894
1895 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
1896 if (n->bottom_type() != type(n)) {
1897 worklist.push(n);
1898 }
1899 }
1900
1901 // For some nodes, we need to propagate the type change to grandchildren or even further down.
1902 // Add them back to the worklist.
1903 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
1904 push_phis(worklist, use);
1905 push_catch(worklist, use);
1906 push_cmpu(worklist, use);
1907 push_counted_loop_phi(worklist, parent, use);
1908 push_loadp(worklist, use);
1909 push_and(worklist, parent, use);
1910 push_cast_ii(worklist, parent, use);
1911 push_opaque_zero_trip_guard(worklist, use);
1912 }
1913
1914
1915 // We must recheck Phis too if use is a Region.
1916 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
1917 if (use->is_Region()) {
1918 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1919 push_if_not_bottom_type(worklist, use->fast_out(i));
1920 }
1921 }
1922 }
1923
1924 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
1925 // non-NULL receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
1926 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
1927 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
1928 if (use->is_Call()) {
1929 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1930 Node* proj = use->fast_out(i);
1931 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
1932 Node* catch_node = proj->find_out_with(Op_Catch);
1933 if (catch_node != NULL) {
1934 worklist.push(catch_node);
1935 }
1936 }
1937 }
1938 }
1939 }
1940
1941 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
1942 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
1943 // and get wrong types otherwise.
1944 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
1945 uint use_op = use->Opcode();
1946 if (use_op == Op_AddI || use_op == Op_SubI) {
1947 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1948 Node* cmpu = use->fast_out(i);
1949 if (cmpu->Opcode() == Op_CmpU) {
1950 // Got a CmpU which might need the new type information from node n.
1951 push_if_not_bottom_type(worklist, cmpu);
1952 }
1953 }
1954 }
1955 }
1956
1957 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'.
1958 // Seem PhiNode::Value().
1959 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
1960 uint use_op = use->Opcode();
1961 if (use_op == Op_CmpI || use_op == Op_CmpL) {
1962 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
1963 if (phi != NULL) {
1964 worklist.push(phi);
1965 }
1966 }
1967 }
1968
1969 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
1970 // See LoadNode::Value().
1971 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
1972 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
1973 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
1974
1975 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1976 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1977 Node* loadp = use->fast_out(i);
1978 const Type* ut = loadp->bottom_type();
1979 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
1980 if (has_load_barrier_nodes) {
1981 // Search for load barriers behind the load
1982 push_load_barrier(worklist, barrier_set, loadp);
1983 }
1984 worklist.push(loadp);
1985 }
1986 }
1987 }
1988 }
1989
1990 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
1991 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1992 Node* barrier_node = use->fast_out(i);
1993 if (barrier_set->is_gc_barrier_node(barrier_node)) {
1994 worklist.push(barrier_node);
1995 }
1996 }
1997 }
1998
1999 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3 to zero if they are bitwise disjoint.
2000 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the shift value changed.
2001 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
2002 uint use_op = use->Opcode();
2003 if ((use_op == Op_LShiftI || use_op == Op_LShiftL)
2004 && use->in(2) == parent) { // is shift value (right-hand side of LShift)
2005 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2006 Node* and_node = use->fast_out(i);
2007 uint and_node_op = and_node->Opcode();
2008 if (and_node_op == Op_AndI || and_node_op == Op_AndL) {
2009 push_if_not_bottom_type(worklist, and_node);
2010 }
2011 }
2012 }
2013 }
2014
2015 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
2016 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
2017 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
2018 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
2019 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
2020 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
2021 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
2022 Node* other_cmp_input = use->in(1);
2023 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
2024 Node* cast_ii = other_cmp_input->fast_out(i);
2025 if (cast_ii->is_CastII()) {
2026 push_if_not_bottom_type(worklist, cast_ii);
2027 }
2028 }
2029 }
2030 }
2031
2032 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
2033 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2034 push_if_not_bottom_type(worklist, use->unique_out());
2035 }
2036 }
2037
2038 //------------------------------do_transform-----------------------------------
2039 // Top level driver for the recursive transformer
2040 void PhaseCCP::do_transform() {
2041 // Correct leaves of new-space Nodes; they point to old-space.
2042 C->set_root( transform(C->root())->as_Root() );
2043 assert( C->top(), "missing TOP node" );
2044 assert( C->root(), "missing root" );
2045 }
2046
2047 //------------------------------transform--------------------------------------
2048 // Given a Node in old-space, clone him into new-space.
2049 // Convert any of his old-space children into new-space children.
2050 Node *PhaseCCP::transform( Node *n ) {
2051 Node *new_node = _nodes[n->_idx]; // Check for transformed node
2052 if( new_node != NULL )
2053 return new_node; // Been there, done that, return old answer
2054
2055 assert(n->is_Root(), "traversal must start at root");
2056 assert(_root_and_safepoints.member(n), "root (n) must be in list");
2057
2058 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
2059 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
2060 Unique_Node_List useful; // track all visited nodes, so that we can remove the complement
2061
2062 // Initialize the traversal.
2063 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
2064 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven
2065 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is
2066 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
2067 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
2068 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
2069 // and all safepoints.
2070 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
2071 Node* nn = _root_and_safepoints.at(i);
2072 Node* new_node = _nodes[nn->_idx];
2073 assert(new_node == NULL, "");
2074 new_node = transform_once(nn); // Check for constant
2075 _nodes.map(nn->_idx, new_node); // Flag as having been cloned
2076 transform_stack.push(new_node); // Process children of cloned node
2077 useful.push(new_node);
2078 }
2079
2080 while (transform_stack.is_nonempty()) {
2081 Node* clone = transform_stack.pop();
2082 uint cnt = clone->req();
2083 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
2084 Node *input = clone->in(i);
2085 if( input != NULL ) { // Ignore NULLs
2086 Node *new_input = _nodes[input->_idx]; // Check for cloned input node
2087 if( new_input == NULL ) {
2088 new_input = transform_once(input); // Check for constant
2089 _nodes.map( input->_idx, new_input );// Flag as having been cloned
2090 transform_stack.push(new_input); // Process children of cloned node
2091 useful.push(new_input);
2092 }
2093 assert( new_input == clone->in(i), "insanity check");
2094 }
2095 }
2096 }
2097
2098 // The above transformation might lead to subgraphs becoming unreachable from the
2099 // bottom while still being reachable from the top. As a result, nodes in that
2100 // subgraph are not transformed and their bottom types are not updated, leading to
2101 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
2102 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
2103 // because their address type is inconsistent. Therefore, we aggressively remove
2104 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
2105 // to remove them anyway.
2106 if (C->cached_top_node()) {
2107 useful.push(C->cached_top_node());
2108 }
2109 C->update_dead_node_list(useful);
2110 remove_useless_nodes(useful.member_set());
2111 _worklist.remove_useless_nodes(useful.member_set());
2112 C->disconnect_useless_nodes(useful, &_worklist);
2113
2114 Node* new_root = _nodes[n->_idx];
2115 assert(new_root->is_Root(), "transformed root node must be a root node");
2116 return new_root;
2117 }
2118
2119
2120 //------------------------------transform_once---------------------------------
2121 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
2122 Node *PhaseCCP::transform_once( Node *n ) {
2123 const Type *t = type(n);
2124 // Constant? Use constant Node instead
2125 if( t->singleton() ) {
2126 Node *nn = n; // Default is to return the original constant
2127 if( t == Type::TOP ) {
2128 // cache my top node on the Compile instance
2129 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
2130 C->set_cached_top_node(ConNode::make(Type::TOP));
2131 set_type(C->top(), Type::TOP);
2132 }
2133 nn = C->top();
2134 }
2135 if( !n->is_Con() ) {
2136 if( t != Type::TOP ) {
2137 nn = makecon(t); // ConNode::make(t);
2138 NOT_PRODUCT( inc_constants(); )
2139 } else if( n->is_Region() ) { // Unreachable region
2140 // Note: nn == C->top()
2141 n->set_req(0, NULL); // Cut selfreference
2142 bool progress = true;
2143 uint max = n->outcnt();
2144 DUIterator i;
2145 while (progress) {
2146 progress = false;
2147 // Eagerly remove dead phis to avoid phis copies creation.
2148 for (i = n->outs(); n->has_out(i); i++) {
2149 Node* m = n->out(i);
2150 if (m->is_Phi()) {
2151 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
2152 replace_node(m, nn);
2153 if (max != n->outcnt()) {
2154 progress = true;
2155 i = n->refresh_out_pos(i);
2156 max = n->outcnt();
2157 }
2158 }
2159 }
2160 }
2161 }
2162 replace_node(n,nn); // Update DefUse edges for new constant
2163 }
2164 return nn;
2165 }
2166
2167 // If x is a TypeNode, capture any more-precise type permanently into Node
2168 if (t != n->bottom_type()) {
2169 hash_delete(n); // changing bottom type may force a rehash
2170 n->raise_bottom_type(t);
2171 _worklist.push(n); // n re-enters the hash table via the worklist
2172 }
2173
2174 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
2175 switch( n->Opcode() ) {
2176 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
2177 case Op_CallDynamicJava:
2178 case Op_FastLock: // Revisit FastLocks for lock coarsening
2179 case Op_If:
2180 case Op_CountedLoopEnd:
2181 case Op_Region:
2182 case Op_Loop:
2183 case Op_CountedLoop:
2184 case Op_Conv2B:
2185 case Op_Opaque1:
2186 _worklist.push(n);
2187 break;
2188 default:
2189 break;
2190 }
2191
2192 return n;
2193 }
2194
2195 //---------------------------------saturate------------------------------------
2196 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
2197 const Type* limit_type) const {
2198 const Type* wide_type = new_type->widen(old_type, limit_type);
2199 if (wide_type != new_type) { // did we widen?
2200 // If so, we may have widened beyond the limit type. Clip it back down.
2201 new_type = wide_type->filter(limit_type);
2202 }
2203 return new_type;
2204 }
2205
2206 //------------------------------print_statistics-------------------------------
2207 #ifndef PRODUCT
2208 void PhaseCCP::print_statistics() {
2209 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
2210 }
2211 #endif
2212
2213
2214 //=============================================================================
2215 #ifndef PRODUCT
2216 uint PhasePeephole::_total_peepholes = 0;
2217 #endif
2218 //------------------------------PhasePeephole----------------------------------
2219 // Conditional Constant Propagation, ala Wegman & Zadeck
2220 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
2221 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
2222 NOT_PRODUCT( clear_peepholes(); )
2223 }
2224
2225 #ifndef PRODUCT
2226 //------------------------------~PhasePeephole---------------------------------
2227 PhasePeephole::~PhasePeephole() {
2228 _total_peepholes += count_peepholes();
2229 }
2230 #endif
2231
2232 //------------------------------transform--------------------------------------
2233 Node *PhasePeephole::transform( Node *n ) {
2234 ShouldNotCallThis();
2235 return NULL;
2236 }
2237
2238 //------------------------------do_transform-----------------------------------
2239 void PhasePeephole::do_transform() {
2240 bool method_name_not_printed = true;
2241
2242 // Examine each basic block
2243 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
2244 Block* block = _cfg.get_block(block_number);
2245 bool block_not_printed = true;
2246
2247 for (bool progress = true; progress;) {
2248 progress = false;
2249 // block->end_idx() not valid after PhaseRegAlloc
2250 uint end_index = block->number_of_nodes();
2251 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
2252 Node *n = block->get_node(instruction_index);
2253 if( n->is_Mach() ) {
2254 MachNode *m = n->as_Mach();
2255 // check for peephole opportunities
2256 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
2257 if( result != -1 ) {
2258 #ifndef PRODUCT
2259 if( PrintOptoPeephole ) {
2260 // Print method, first time only
2261 if( C->method() && method_name_not_printed ) {
2262 C->method()->print_short_name(); tty->cr();
2263 method_name_not_printed = false;
2264 }
2265 // Print this block
2266 if( Verbose && block_not_printed) {
2267 tty->print_cr("in block");
2268 block->dump();
2269 block_not_printed = false;
2270 }
2271 // Print the peephole number
2272 tty->print_cr("peephole number: %d", result);
2273 }
2274 inc_peepholes();
2275 #endif
2276 // Set progress, start again
2277 progress = true;
2278 break;
2279 }
2280 }
2281 }
2282 }
2283 }
2284 }
2285
2286 //------------------------------print_statistics-------------------------------
2287 #ifndef PRODUCT
2288 void PhasePeephole::print_statistics() {
2289 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
2290 }
2291 #endif
2292
2293
2294 //=============================================================================
2295 //------------------------------set_req_X--------------------------------------
2296 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2297 assert( is_not_dead(n), "can not use dead node");
2298 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2299 Node *old = in(i);
2300 set_req(i, n);
2301
2302 // old goes dead?
2303 if( old ) {
2304 switch (old->outcnt()) {
2305 case 0:
2306 // Put into the worklist to kill later. We do not kill it now because the
2307 // recursive kill will delete the current node (this) if dead-loop exists
2308 if (!old->is_top())
2309 igvn->_worklist.push( old );
2310 break;
2311 case 1:
2312 if( old->is_Store() || old->has_special_unique_user() )
2313 igvn->add_users_to_worklist( old );
2314 break;
2315 case 2:
2316 if( old->is_Store() )
2317 igvn->add_users_to_worklist( old );
2318 if( old->Opcode() == Op_Region )
2319 igvn->_worklist.push(old);
2320 break;
2321 case 3:
2322 if( old->Opcode() == Op_Region ) {
2323 igvn->_worklist.push(old);
2324 igvn->add_users_to_worklist( old );
2325 }
2326 break;
2327 default:
2328 break;
2329 }
2330
2331 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
2332 }
2333 }
2334
2335 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
2336 PhaseIterGVN* igvn = gvn->is_IterGVN();
2337 if (igvn == NULL) {
2338 set_req(i, n);
2339 return;
2340 }
2341 set_req_X(i, n, igvn);
2342 }
2343
2344 //-------------------------------replace_by-----------------------------------
2345 // Using def-use info, replace one node for another. Follow the def-use info
2346 // to all users of the OLD node. Then make all uses point to the NEW node.
2347 void Node::replace_by(Node *new_node) {
2348 assert(!is_top(), "top node has no DU info");
2349 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2350 Node* use = last_out(i);
2351 uint uses_found = 0;
2352 for (uint j = 0; j < use->len(); j++) {
2353 if (use->in(j) == this) {
2354 if (j < use->req())
2355 use->set_req(j, new_node);
2356 else use->set_prec(j, new_node);
2357 uses_found++;
2358 }
2359 }
2360 i -= uses_found; // we deleted 1 or more copies of this edge
2361 }
2362 }
2363
2364 //=============================================================================
2365 //-----------------------------------------------------------------------------
2366 void Type_Array::grow( uint i ) {
2367 if( !_max ) {
2368 _max = 1;
2369 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2370 _types[0] = NULL;
2371 }
2372 uint old = _max;
2373 _max = next_power_of_2(i);
2374 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2375 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2376 }
2377
2378 //------------------------------dump-------------------------------------------
2379 #ifndef PRODUCT
2380 void Type_Array::dump() const {
2381 uint max = Size();
2382 for( uint i = 0; i < max; i++ ) {
2383 if( _types[i] != NULL ) {
2384 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
2385 }
2386 }
2387 }
2388 #endif