1 /*
2 * Copyright (c) 1997, 2023, 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
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 debug_only( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 debug_only( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 debug_only(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 // Implementation requires an edge from root to each SafePointNode
368 // at a backward branch. Inserted in add_safepoint().
369
370 // Identify nodes that are reachable from below, useful.
371 C->identify_useful_nodes(_useful);
372 // Update dead node list
373 C->update_dead_node_list(_useful);
374
375 // Remove all useless nodes from PhaseValues' recorded types
376 // Must be done before disconnecting nodes to preserve hash-table-invariant
377 gvn->remove_useless_nodes(_useful.member_set());
378
379 // Remove all useless nodes from future worklist
380 worklist.remove_useless_nodes(_useful.member_set());
381
382 // Disconnect 'useless' nodes that are adjacent to useful nodes
383 C->disconnect_useless_nodes(_useful, worklist);
384 }
385
386 //=============================================================================
387 //------------------------------PhaseRenumberLive------------------------------
388 // First, remove useless nodes (equivalent to identifying live nodes).
389 // Then, renumber live nodes.
390 //
391 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
392 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
393 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
394 // value in the range [0, x).
395 //
396 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
397 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
398 //
399 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
400 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
401 // be processed by IGVN after removal of the useless nodes.
402 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
403 // the node's type. The mapping is updated to use the new node IDs as well. We
404 // create a new map, and swap it with the old one.
405 //
406 // Other data structures used by the compiler are not updated. The hash table for value
407 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
408 // computing the hash values is not based on node IDs.
409 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
410 Unique_Node_List& worklist,
411 PhaseNumber phase_num) :
412 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
413 _new_type_array(C->comp_arena()),
414 _old2new_map(C->unique(), C->unique(), -1),
415 _is_pass_finished(false),
416 _live_node_count(C->live_nodes())
417 {
418 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
419 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
420 assert(_delayed.size() == 0, "should be empty");
421 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
422 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
423
424 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
425 if (old_node_note_array != nullptr) {
426 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
427 new_size = MAX2(8, new_size);
428 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
429 C->grow_node_notes(C->node_note_array(), new_size);
430 }
431
432 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
433
434 // Iterate over the set of live nodes.
435 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
436 Node* n = _useful.at(current_idx);
437
438 const Type* type = gvn->type_or_null(n);
439 _new_type_array.map(current_idx, type);
440
441 assert(_old2new_map.at(n->_idx) == -1, "already seen");
442 _old2new_map.at_put(n->_idx, current_idx);
443
444 if (old_node_note_array != nullptr) {
445 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
446 C->set_node_notes_at(current_idx, nn);
447 }
448
449 n->set_idx(current_idx); // Update node ID.
450
451 if (update_embedded_ids(n) < 0) {
452 _delayed.push(n); // has embedded IDs; handle later
453 }
454 }
455
456 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
457 worklist.recompute_idx_set();
458
459 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
460
461 _is_pass_finished = true; // pass finished; safe to process delayed updates
462
463 while (_delayed.size() > 0) {
464 Node* n = _delayed.pop();
465 int no_of_updates = update_embedded_ids(n);
466 assert(no_of_updates > 0, "should be updated");
467 }
468
469 // Replace the compiler's type information with the updated type information.
470 gvn->types().swap(_new_type_array);
471
472 // Update the unique node count of the compilation to the number of currently live nodes.
473 C->set_unique(_live_node_count);
474
475 // Set the dead node count to 0 and reset dead node list.
476 C->reset_dead_node_list();
477 }
478
479 int PhaseRenumberLive::new_index(int old_idx) {
480 assert(_is_pass_finished, "not finished");
481 if (_old2new_map.at(old_idx) == -1) { // absent
482 // Allocate a placeholder to preserve uniqueness
483 _old2new_map.at_put(old_idx, _live_node_count);
484 _live_node_count++;
485 }
486 return _old2new_map.at(old_idx);
487 }
488
489 int PhaseRenumberLive::update_embedded_ids(Node* n) {
490 int no_of_updates = 0;
491 if (n->is_Phi()) {
492 PhiNode* phi = n->as_Phi();
493 if (phi->_inst_id != -1) {
494 if (!_is_pass_finished) {
495 return -1; // delay
496 }
497 int new_idx = new_index(phi->_inst_id);
498 assert(new_idx != -1, "");
499 phi->_inst_id = new_idx;
500 no_of_updates++;
501 }
502 if (phi->_inst_mem_id != -1) {
503 if (!_is_pass_finished) {
504 return -1; // delay
505 }
506 int new_idx = new_index(phi->_inst_mem_id);
507 assert(new_idx != -1, "");
508 phi->_inst_mem_id = new_idx;
509 no_of_updates++;
510 }
511 }
512
513 const Type* type = _new_type_array.fast_lookup(n->_idx);
514 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
515 if (!_is_pass_finished) {
516 return -1; // delay
517 }
518 int old_idx = type->is_oopptr()->instance_id();
519 int new_idx = new_index(old_idx);
520 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
521 _new_type_array.map(n->_idx, new_type);
522 no_of_updates++;
523 }
524
525 return no_of_updates;
526 }
527
528 void PhaseValues::init_con_caches() {
529 memset(_icons,0,sizeof(_icons));
530 memset(_lcons,0,sizeof(_lcons));
531 memset(_zcons,0,sizeof(_zcons));
532 }
533
534 //--------------------------------find_int_type--------------------------------
535 const TypeInt* PhaseValues::find_int_type(Node* n) {
536 if (n == nullptr) return nullptr;
537 // Call type_or_null(n) to determine node's type since we might be in
538 // parse phase and call n->Value() may return wrong type.
539 // (For example, a phi node at the beginning of loop parsing is not ready.)
540 const Type* t = type_or_null(n);
541 if (t == nullptr) return nullptr;
542 return t->isa_int();
543 }
544
545
546 //-------------------------------find_long_type--------------------------------
547 const TypeLong* PhaseValues::find_long_type(Node* n) {
548 if (n == nullptr) return nullptr;
549 // (See comment above on type_or_null.)
550 const Type* t = type_or_null(n);
551 if (t == nullptr) return nullptr;
552 return t->isa_long();
553 }
554
555 //------------------------------~PhaseValues-----------------------------------
556 #ifndef PRODUCT
557 PhaseValues::~PhaseValues() {
558 // Statistics for NodeHash
559 _table.dump();
560 // Statistics for value progress and efficiency
561 if( PrintCompilation && Verbose && WizardMode ) {
562 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
563 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
564 if( made_transforms() != 0 ) {
565 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
566 } else {
567 tty->cr();
568 }
569 }
570 }
571 #endif
572
573 //------------------------------makecon----------------------------------------
574 ConNode* PhaseValues::makecon(const Type* t) {
575 assert(t->singleton(), "must be a constant");
576 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
577 switch (t->base()) { // fast paths
578 case Type::Half:
579 case Type::Top: return (ConNode*) C->top();
580 case Type::Int: return intcon( t->is_int()->get_con() );
581 case Type::Long: return longcon( t->is_long()->get_con() );
582 default: break;
583 }
584 if (t->is_zero_type())
585 return zerocon(t->basic_type());
586 return uncached_makecon(t);
587 }
588
589 //--------------------------uncached_makecon-----------------------------------
590 // Make an idealized constant - one of ConINode, ConPNode, etc.
591 ConNode* PhaseValues::uncached_makecon(const Type *t) {
592 assert(t->singleton(), "must be a constant");
593 ConNode* x = ConNode::make(t);
594 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
595 if (k == nullptr) {
596 set_type(x, t); // Missed, provide type mapping
597 GrowableArray<Node_Notes*>* nna = C->node_note_array();
598 if (nna != nullptr) {
599 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
600 loc->clear(); // do not put debug info on constants
601 }
602 } else {
603 x->destruct(this); // Hit, destroy duplicate constant
604 x = k; // use existing constant
605 }
606 return x;
607 }
608
609 //------------------------------intcon-----------------------------------------
610 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
611 ConINode* PhaseValues::intcon(jint i) {
612 // Small integer? Check cache! Check that cached node is not dead
613 if (i >= _icon_min && i <= _icon_max) {
614 ConINode* icon = _icons[i-_icon_min];
615 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
616 return icon;
617 }
618 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
619 assert(icon->is_Con(), "");
620 if (i >= _icon_min && i <= _icon_max)
621 _icons[i-_icon_min] = icon; // Cache small integers
622 return icon;
623 }
624
625 //------------------------------longcon----------------------------------------
626 // Fast long constant.
627 ConLNode* PhaseValues::longcon(jlong l) {
628 // Small integer? Check cache! Check that cached node is not dead
629 if (l >= _lcon_min && l <= _lcon_max) {
630 ConLNode* lcon = _lcons[l-_lcon_min];
631 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
632 return lcon;
633 }
634 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
635 assert(lcon->is_Con(), "");
636 if (l >= _lcon_min && l <= _lcon_max)
637 _lcons[l-_lcon_min] = lcon; // Cache small integers
638 return lcon;
639 }
640 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
641 if (bt == T_INT) {
642 return intcon(checked_cast<jint>(l));
643 }
644 assert(bt == T_LONG, "not an integer");
645 return longcon(l);
646 }
647
648
649 //------------------------------zerocon-----------------------------------------
650 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
651 ConNode* PhaseValues::zerocon(BasicType bt) {
652 assert((uint)bt <= _zcon_max, "domain check");
653 ConNode* zcon = _zcons[bt];
654 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
655 return zcon;
656 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
657 _zcons[bt] = zcon;
658 return zcon;
659 }
660
661
662
663 //=============================================================================
664 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
665 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
666 if (i == nullptr) {
667 i = k->Ideal(this, can_reshape);
668 }
669 return i;
670 }
671
672 //------------------------------transform--------------------------------------
673 // Return a node which computes the same function as this node, but in a
674 // faster or cheaper fashion.
675 Node *PhaseGVN::transform( Node *n ) {
676 return transform_no_reclaim(n);
677 }
678
679 //------------------------------transform--------------------------------------
680 // Return a node which computes the same function as this node, but
681 // in a faster or cheaper fashion.
682 Node *PhaseGVN::transform_no_reclaim(Node *n) {
683 NOT_PRODUCT( set_transforms(); )
684
685 // Apply the Ideal call in a loop until it no longer applies
686 Node* k = n;
687 Node* i = apply_ideal(k, /*can_reshape=*/false);
688 NOT_PRODUCT(uint loop_count = 1;)
689 while (i != nullptr) {
690 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
691 k = i;
692 #ifdef ASSERT
693 if (loop_count >= K + C->live_nodes()) {
694 dump_infinite_loop_info(i, "PhaseGVN::transform_no_reclaim");
695 }
696 #endif
697 i = apply_ideal(k, /*can_reshape=*/false);
698 NOT_PRODUCT(loop_count++;)
699 }
700 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
701
702 // If brand new node, make space in type array.
703 ensure_type_or_null(k);
704
705 // Since I just called 'Value' to compute the set of run-time values
706 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
707 // cache Value. Later requests for the local phase->type of this Node can
708 // use the cached Value instead of suffering with 'bottom_type'.
709 const Type* t = k->Value(this); // Get runtime Value set
710 assert(t != nullptr, "value sanity");
711 if (type_or_null(k) != t) {
712 #ifndef PRODUCT
713 // Do not count initial visit to node as a transformation
714 if (type_or_null(k) == nullptr) {
715 inc_new_values();
716 set_progress();
717 }
718 #endif
719 set_type(k, t);
720 // If k is a TypeNode, capture any more-precise type permanently into Node
721 k->raise_bottom_type(t);
722 }
723
724 if (t->singleton() && !k->is_Con()) {
725 NOT_PRODUCT(set_progress();)
726 return makecon(t); // Turn into a constant
727 }
728
729 // Now check for Identities
730 i = k->Identity(this); // Look for a nearby replacement
731 if (i != k) { // Found? Return replacement!
732 NOT_PRODUCT(set_progress();)
733 return i;
734 }
735
736 // Global Value Numbering
737 i = hash_find_insert(k); // Insert if new
738 if (i && (i != k)) {
739 // Return the pre-existing node
740 NOT_PRODUCT(set_progress();)
741 return i;
742 }
743
744 // Return Idealized original
745 return k;
746 }
747
748 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
749 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
750 return false;
751 }
752 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
753 return false;
754 }
755 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
756 int i = 0;
757 while (d != n) {
758 n = IfNode::up_one_dom(n, linear_only);
759 i++;
760 if (n == nullptr || i >= 100) {
761 return false;
762 }
763 }
764 return true;
765 }
766
767 #ifdef ASSERT
768 //------------------------------dead_loop_check--------------------------------
769 // Check for a simple dead loop when a data node references itself directly
770 // or through an other data node excluding cons and phis.
771 void PhaseGVN::dead_loop_check( Node *n ) {
772 // Phi may reference itself in a loop
773 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) {
774 // Do 2 levels check and only data inputs.
775 bool no_dead_loop = true;
776 uint cnt = n->req();
777 for (uint i = 1; i < cnt && no_dead_loop; i++) {
778 Node *in = n->in(i);
779 if (in == n) {
780 no_dead_loop = false;
781 } else if (in != nullptr && !in->is_dead_loop_safe()) {
782 uint icnt = in->req();
783 for (uint j = 1; j < icnt && no_dead_loop; j++) {
784 if (in->in(j) == n || in->in(j) == in)
785 no_dead_loop = false;
786 }
787 }
788 }
789 if (!no_dead_loop) n->dump_bfs(100,0,"#");
790 assert(no_dead_loop, "dead loop detected");
791 }
792 }
793
794
795 /**
796 * Dumps information that can help to debug the problem. A debug
797 * build fails with an assert.
798 */
799 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
800 n->dump(4);
801 assert(false, "infinite loop in %s", where);
802 }
803 #endif
804
805 //=============================================================================
806 //------------------------------PhaseIterGVN-----------------------------------
807 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
808 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
809 _worklist(*C->igvn_worklist())
810 {
811 _iterGVN = true;
812 assert(&_worklist == &igvn->_worklist, "sanity");
813 }
814
815 //------------------------------PhaseIterGVN-----------------------------------
816 // Initialize with previous PhaseGVN info from Parser
817 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : _delay_transform(false),
818 _worklist(*C->igvn_worklist())
819 {
820 _iterGVN = true;
821 uint max;
822
823 // Dead nodes in the hash table inherited from GVN were not treated as
824 // roots during def-use info creation; hence they represent an invisible
825 // use. Clear them out.
826 max = _table.size();
827 for( uint i = 0; i < max; ++i ) {
828 Node *n = _table.at(i);
829 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
830 if( n->is_top() ) continue;
831 // If remove_useless_nodes() has run, we expect no such nodes left.
832 assert(false, "remove_useless_nodes missed this node");
833 hash_delete(n);
834 }
835 }
836
837 // Any Phis or Regions on the worklist probably had uses that could not
838 // make more progress because the uses were made while the Phis and Regions
839 // were in half-built states. Put all uses of Phis and Regions on worklist.
840 max = _worklist.size();
841 for( uint j = 0; j < max; j++ ) {
842 Node *n = _worklist.at(j);
843 uint uop = n->Opcode();
844 if( uop == Op_Phi || uop == Op_Region ||
845 n->is_Type() ||
846 n->is_Mem() )
847 add_users_to_worklist(n);
848 }
849 }
850
851 void PhaseIterGVN::shuffle_worklist() {
852 if (_worklist.size() < 2) return;
853 for (uint i = _worklist.size() - 1; i >= 1; i--) {
854 uint j = C->random() % (i + 1);
855 swap(_worklist.adr()[i], _worklist.adr()[j]);
856 }
857 }
858
859 #ifndef PRODUCT
860 void PhaseIterGVN::verify_step(Node* n) {
861 if (is_verify_def_use()) {
862 ResourceMark rm;
863 VectorSet visited;
864 Node_List worklist;
865
866 _verify_window[_verify_counter % _verify_window_size] = n;
867 ++_verify_counter;
868 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
869 ++_verify_full_passes;
870 worklist.push(C->root());
871 Node::verify(-1, visited, worklist);
872 return;
873 }
874 for (int i = 0; i < _verify_window_size; i++) {
875 Node* n = _verify_window[i];
876 if (n == nullptr) {
877 continue;
878 }
879 if (n->in(0) == NodeSentinel) { // xform_idom
880 _verify_window[i] = n->in(1);
881 --i;
882 continue;
883 }
884 // Typical fanout is 1-2, so this call visits about 6 nodes.
885 if (!visited.test_set(n->_idx)) {
886 worklist.push(n);
887 }
888 }
889 Node::verify(4, visited, worklist);
890 }
891 }
892
893 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
894 if (TraceIterativeGVN) {
895 uint wlsize = _worklist.size();
896 const Type* newtype = type_or_null(n);
897 if (nn != n) {
898 // print old node
899 tty->print("< ");
900 if (oldtype != newtype && oldtype != nullptr) {
901 oldtype->dump();
902 }
903 do { tty->print("\t"); } while (tty->position() < 16);
904 tty->print("<");
905 n->dump();
906 }
907 if (oldtype != newtype || nn != n) {
908 // print new node and/or new type
909 if (oldtype == nullptr) {
910 tty->print("* ");
911 } else if (nn != n) {
912 tty->print("> ");
913 } else {
914 tty->print("= ");
915 }
916 if (newtype == nullptr) {
917 tty->print("null");
918 } else {
919 newtype->dump();
920 }
921 do { tty->print("\t"); } while (tty->position() < 16);
922 nn->dump();
923 }
924 if (Verbose && wlsize < _worklist.size()) {
925 tty->print(" Push {");
926 while (wlsize != _worklist.size()) {
927 Node* pushed = _worklist.at(wlsize++);
928 tty->print(" %d", pushed->_idx);
929 }
930 tty->print_cr(" }");
931 }
932 if (nn != n) {
933 // ignore n, it might be subsumed
934 verify_step((Node*) nullptr);
935 }
936 }
937 }
938
939 void PhaseIterGVN::init_verifyPhaseIterGVN() {
940 _verify_counter = 0;
941 _verify_full_passes = 0;
942 for (int i = 0; i < _verify_window_size; i++) {
943 _verify_window[i] = nullptr;
944 }
945 #ifdef ASSERT
946 // Verify that all modified nodes are on _worklist
947 Unique_Node_List* modified_list = C->modified_nodes();
948 while (modified_list != nullptr && modified_list->size()) {
949 Node* n = modified_list->pop();
950 if (!n->is_Con() && !_worklist.member(n)) {
951 n->dump();
952 fatal("modified node is not on IGVN._worklist");
953 }
954 }
955 #endif
956 }
957
958 void PhaseIterGVN::verify_PhaseIterGVN() {
959 #ifdef ASSERT
960 // Verify nodes with changed inputs.
961 Unique_Node_List* modified_list = C->modified_nodes();
962 while (modified_list != nullptr && modified_list->size()) {
963 Node* n = modified_list->pop();
964 if (!n->is_Con()) { // skip Con nodes
965 n->dump();
966 fatal("modified node was not processed by IGVN.transform_old()");
967 }
968 }
969 #endif
970
971 C->verify_graph_edges();
972 if (is_verify_def_use() && PrintOpto) {
973 if (_verify_counter == _verify_full_passes) {
974 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
975 (int) _verify_full_passes);
976 } else {
977 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
978 (int) _verify_counter, (int) _verify_full_passes);
979 }
980 }
981
982 #ifdef ASSERT
983 if (modified_list != nullptr) {
984 while (modified_list->size() > 0) {
985 Node* n = modified_list->pop();
986 n->dump();
987 assert(false, "VerifyIterativeGVN: new modified node was added");
988 }
989 }
990
991 verify_optimize();
992 #endif
993 }
994 #endif /* PRODUCT */
995
996 #ifdef ASSERT
997 /**
998 * Dumps information that can help to debug the problem. A debug
999 * build fails with an assert.
1000 */
1001 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1002 n->dump(4);
1003 _worklist.dump();
1004 assert(false, "infinite loop in %s", where);
1005 }
1006
1007 /**
1008 * Prints out information about IGVN if the 'verbose' option is used.
1009 */
1010 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1011 if (TraceIterativeGVN && Verbose) {
1012 tty->print(" Pop ");
1013 n->dump();
1014 if ((num_processed % 100) == 0) {
1015 _worklist.print_set();
1016 }
1017 }
1018 }
1019 #endif /* ASSERT */
1020
1021 void PhaseIterGVN::optimize() {
1022 DEBUG_ONLY(uint num_processed = 0;)
1023 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1024 if (StressIGVN) {
1025 shuffle_worklist();
1026 }
1027
1028 uint loop_count = 0;
1029 // Pull from worklist and transform the node. If the node has changed,
1030 // update edge info and put uses on worklist.
1031 while(_worklist.size()) {
1032 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1033 return;
1034 }
1035 Node* n = _worklist.pop();
1036 if (loop_count >= K * C->live_nodes()) {
1037 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1038 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1039 return;
1040 }
1041 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1042 if (n->outcnt() != 0) {
1043 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1044 // Do the transformation
1045 Node* nn = transform_old(n);
1046 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1047 } else if (!n->is_top()) {
1048 remove_dead_node(n);
1049 }
1050 loop_count++;
1051 }
1052 NOT_PRODUCT(verify_PhaseIterGVN();)
1053 }
1054
1055 #ifdef ASSERT
1056 void PhaseIterGVN::verify_optimize() {
1057 if (is_verify_Value()) {
1058 ResourceMark rm;
1059 Unique_Node_List worklist;
1060 bool failure = false;
1061 // BFS all nodes, starting at root
1062 worklist.push(C->root());
1063 for (uint j = 0; j < worklist.size(); ++j) {
1064 Node* n = worklist.at(j);
1065 failure |= verify_node_value(n);
1066 // traverse all inputs and outputs
1067 for (uint i = 0; i < n->req(); i++) {
1068 if (n->in(i) != nullptr) {
1069 worklist.push(n->in(i));
1070 }
1071 }
1072 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1073 worklist.push(n->fast_out(i));
1074 }
1075 }
1076 // If we get this assert, check why the reported nodes were not processed again in IGVN.
1077 // We should either make sure that these nodes are properly added back to the IGVN worklist
1078 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception
1079 // in the verification code above if that is not possible for some reason (like Load nodes).
1080 assert(!failure, "Missed optimization opportunity in PhaseIterGVN");
1081 }
1082 }
1083
1084 // Check that type(n) == n->Value(), return true if we have a failure.
1085 // We have a list of exceptions, see detailed comments in code.
1086 // (1) Integer "widen" changes, but the range is the same.
1087 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1088 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1089 bool PhaseIterGVN::verify_node_value(Node* n) {
1090 // If we assert inside type(n), because the type is still a null, then maybe
1091 // the node never went through gvn.transform, which would be a bug.
1092 const Type* told = type(n);
1093 const Type* tnew = n->Value(this);
1094 if (told == tnew) {
1095 return false;
1096 }
1097 // Exception (1)
1098 // Integer "widen" changes, but range is the same.
1099 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1100 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1101 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1102 if (t0->lo_as_long() == t1->lo_as_long() &&
1103 t0->hi_as_long() == t1->hi_as_long()) {
1104 return false; // ignore integer widen
1105 }
1106 }
1107 // Exception (2)
1108 // LoadNode performs deep traversals. Load is not notified for changes far away.
1109 if (n->is_Load() && !told->singleton()) {
1110 // MemNode::can_see_stored_value looks up through many memory nodes,
1111 // which means we would need to notify modifications from far up in
1112 // the inputs all the way down to the LoadNode. We don't do that.
1113 return false;
1114 }
1115 // Exception (3)
1116 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1117 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1118 // SubNode::Value
1119 // CmpPNode::sub
1120 // MemNode::detect_ptr_independence
1121 // MemNode::all_controls_dominate
1122 // We find all controls of a pointer load, and see if they dominate the control of
1123 // an allocation. If they all dominate, we know the allocation is after (independent)
1124 // of the pointer load, and we can say the pointers are different. For this we call
1125 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1126 // control sub of the allocation. The problems is that sometimes dominates answers
1127 // false conservatively, and later it can determine that it is indeed true. Loops with
1128 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1129 // so the traversal becomes more powerful. This is difficult to remidy, we would have
1130 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1131 // after loop-opts, so that should take care of many of these cases.
1132 return false;
1133 }
1134 tty->cr();
1135 tty->print_cr("Missed Value optimization:");
1136 n->dump_bfs(1, 0, "");
1137 tty->print_cr("Current type:");
1138 told->dump_on(tty);
1139 tty->cr();
1140 tty->print_cr("Optimized type:");
1141 tnew->dump_on(tty);
1142 tty->cr();
1143 return true;
1144 }
1145 #endif
1146
1147 /**
1148 * Register a new node with the optimizer. Update the types array, the def-use
1149 * info. Put on worklist.
1150 */
1151 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1152 set_type_bottom(n);
1153 _worklist.push(n);
1154 if (orig != nullptr) C->copy_node_notes_to(n, orig);
1155 return n;
1156 }
1157
1158 //------------------------------transform--------------------------------------
1159 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
1160 Node *PhaseIterGVN::transform( Node *n ) {
1161 if (_delay_transform) {
1162 // Register the node but don't optimize for now
1163 register_new_node_with_optimizer(n);
1164 return n;
1165 }
1166
1167 // If brand new node, make space in type array, and give it a type.
1168 ensure_type_or_null(n);
1169 if (type_or_null(n) == nullptr) {
1170 set_type_bottom(n);
1171 }
1172
1173 return transform_old(n);
1174 }
1175
1176 Node *PhaseIterGVN::transform_old(Node* n) {
1177 NOT_PRODUCT(set_transforms());
1178 // Remove 'n' from hash table in case it gets modified
1179 _table.hash_delete(n);
1180 #ifdef ASSERT
1181 if (is_verify_def_use()) {
1182 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1183 }
1184 #endif
1185
1186 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
1187 if (n->is_Cmp()) {
1188 add_users_to_worklist(n);
1189 }
1190
1191 // Apply the Ideal call in a loop until it no longer applies
1192 Node* k = n;
1193 DEBUG_ONLY(dead_loop_check(k);)
1194 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1195 C->remove_modified_node(k);
1196 Node* i = apply_ideal(k, /*can_reshape=*/true);
1197 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1198 #ifndef PRODUCT
1199 verify_step(k);
1200 #endif
1201
1202 DEBUG_ONLY(uint loop_count = 1;)
1203 while (i != nullptr) {
1204 #ifdef ASSERT
1205 if (loop_count >= K + C->live_nodes()) {
1206 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
1207 }
1208 #endif
1209 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1210 // Made a change; put users of original Node on worklist
1211 add_users_to_worklist(k);
1212 // Replacing root of transform tree?
1213 if (k != i) {
1214 // Make users of old Node now use new.
1215 subsume_node(k, i);
1216 k = i;
1217 }
1218 DEBUG_ONLY(dead_loop_check(k);)
1219 // Try idealizing again
1220 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1221 C->remove_modified_node(k);
1222 i = apply_ideal(k, /*can_reshape=*/true);
1223 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1224 #ifndef PRODUCT
1225 verify_step(k);
1226 #endif
1227 DEBUG_ONLY(loop_count++;)
1228 }
1229
1230 // If brand new node, make space in type array.
1231 ensure_type_or_null(k);
1232
1233 // See what kind of values 'k' takes on at runtime
1234 const Type* t = k->Value(this);
1235 assert(t != nullptr, "value sanity");
1236
1237 // Since I just called 'Value' to compute the set of run-time values
1238 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1239 // cache Value. Later requests for the local phase->type of this Node can
1240 // use the cached Value instead of suffering with 'bottom_type'.
1241 if (type_or_null(k) != t) {
1242 #ifndef PRODUCT
1243 inc_new_values();
1244 set_progress();
1245 #endif
1246 set_type(k, t);
1247 // If k is a TypeNode, capture any more-precise type permanently into Node
1248 k->raise_bottom_type(t);
1249 // Move users of node to worklist
1250 add_users_to_worklist(k);
1251 }
1252 // If 'k' computes a constant, replace it with a constant
1253 if (t->singleton() && !k->is_Con()) {
1254 NOT_PRODUCT(set_progress();)
1255 Node* con = makecon(t); // Make a constant
1256 add_users_to_worklist(k);
1257 subsume_node(k, con); // Everybody using k now uses con
1258 return con;
1259 }
1260
1261 // Now check for Identities
1262 i = k->Identity(this); // Look for a nearby replacement
1263 if (i != k) { // Found? Return replacement!
1264 NOT_PRODUCT(set_progress();)
1265 add_users_to_worklist(k);
1266 subsume_node(k, i); // Everybody using k now uses i
1267 return i;
1268 }
1269
1270 // Global Value Numbering
1271 i = hash_find_insert(k); // Check for pre-existing node
1272 if (i && (i != k)) {
1273 // Return the pre-existing node if it isn't dead
1274 NOT_PRODUCT(set_progress();)
1275 add_users_to_worklist(k);
1276 subsume_node(k, i); // Everybody using k now uses i
1277 return i;
1278 }
1279
1280 // Return Idealized original
1281 return k;
1282 }
1283
1284 //---------------------------------saturate------------------------------------
1285 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1286 const Type* limit_type) const {
1287 return new_type->narrow(old_type);
1288 }
1289
1290 //------------------------------remove_globally_dead_node----------------------
1291 // Kill a globally dead Node. All uses are also globally dead and are
1292 // aggressively trimmed.
1293 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1294 enum DeleteProgress {
1295 PROCESS_INPUTS,
1296 PROCESS_OUTPUTS
1297 };
1298 ResourceMark rm;
1299 Node_Stack stack(32);
1300 stack.push(dead, PROCESS_INPUTS);
1301
1302 while (stack.is_nonempty()) {
1303 dead = stack.node();
1304 if (dead->Opcode() == Op_SafePoint) {
1305 dead->as_SafePoint()->disconnect_from_root(this);
1306 }
1307 uint progress_state = stack.index();
1308 assert(dead != C->root(), "killing root, eh?");
1309 assert(!dead->is_top(), "add check for top when pushing");
1310 NOT_PRODUCT( set_progress(); )
1311 if (progress_state == PROCESS_INPUTS) {
1312 // After following inputs, continue to outputs
1313 stack.set_index(PROCESS_OUTPUTS);
1314 if (!dead->is_Con()) { // Don't kill cons but uses
1315 bool recurse = false;
1316 // Remove from hash table
1317 _table.hash_delete( dead );
1318 // Smash all inputs to 'dead', isolating him completely
1319 for (uint i = 0; i < dead->req(); i++) {
1320 Node *in = dead->in(i);
1321 if (in != nullptr && in != C->top()) { // Points to something?
1322 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
1323 assert((nrep > 0), "sanity");
1324 if (in->outcnt() == 0) { // Made input go dead?
1325 stack.push(in, PROCESS_INPUTS); // Recursively remove
1326 recurse = true;
1327 } else if (in->outcnt() == 1 &&
1328 in->has_special_unique_user()) {
1329 _worklist.push(in->unique_out());
1330 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1331 if (in->Opcode() == Op_Region) {
1332 _worklist.push(in);
1333 } else if (in->is_Store()) {
1334 DUIterator_Fast imax, i = in->fast_outs(imax);
1335 _worklist.push(in->fast_out(i));
1336 i++;
1337 if (in->outcnt() == 2) {
1338 _worklist.push(in->fast_out(i));
1339 i++;
1340 }
1341 assert(!(i < imax), "sanity");
1342 }
1343 } else {
1344 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
1345 }
1346 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1347 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
1348 // A Load that directly follows an InitializeNode is
1349 // going away. The Stores that follow are candidates
1350 // again to be captured by the InitializeNode.
1351 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1352 Node *n = in->fast_out(j);
1353 if (n->is_Store()) {
1354 _worklist.push(n);
1355 }
1356 }
1357 }
1358 } // if (in != nullptr && in != C->top())
1359 } // for (uint i = 0; i < dead->req(); i++)
1360 if (recurse) {
1361 continue;
1362 }
1363 } // if (!dead->is_Con())
1364 } // if (progress_state == PROCESS_INPUTS)
1365
1366 // Aggressively kill globally dead uses
1367 // (Rather than pushing all the outs at once, we push one at a time,
1368 // plus the parent to resume later, because of the indefinite number
1369 // of edge deletions per loop trip.)
1370 if (dead->outcnt() > 0) {
1371 // Recursively remove output edges
1372 stack.push(dead->raw_out(0), PROCESS_INPUTS);
1373 } else {
1374 // Finished disconnecting all input and output edges.
1375 stack.pop();
1376 // Remove dead node from iterative worklist
1377 _worklist.remove(dead);
1378 C->remove_useless_node(dead);
1379 }
1380 } // while (stack.is_nonempty())
1381 }
1382
1383 //------------------------------subsume_node-----------------------------------
1384 // Remove users from node 'old' and add them to node 'nn'.
1385 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1386 if (old->Opcode() == Op_SafePoint) {
1387 old->as_SafePoint()->disconnect_from_root(this);
1388 }
1389 assert( old != hash_find(old), "should already been removed" );
1390 assert( old != C->top(), "cannot subsume top node");
1391 // Copy debug or profile information to the new version:
1392 C->copy_node_notes_to(nn, old);
1393 // Move users of node 'old' to node 'nn'
1394 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1395 Node* use = old->last_out(i); // for each use...
1396 // use might need re-hashing (but it won't if it's a new node)
1397 rehash_node_delayed(use);
1398 // Update use-def info as well
1399 // We remove all occurrences of old within use->in,
1400 // so as to avoid rehashing any node more than once.
1401 // The hash table probe swamps any outer loop overhead.
1402 uint num_edges = 0;
1403 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1404 if (use->in(j) == old) {
1405 use->set_req(j, nn);
1406 ++num_edges;
1407 }
1408 }
1409 i -= num_edges; // we deleted 1 or more copies of this edge
1410 }
1411
1412 // Search for instance field data PhiNodes in the same region pointing to the old
1413 // memory PhiNode and update their instance memory ids to point to the new node.
1414 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
1415 Node* region = old->in(0);
1416 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1417 PhiNode* phi = region->fast_out(i)->isa_Phi();
1418 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
1419 phi->set_inst_mem_id((int)nn->_idx);
1420 }
1421 }
1422 }
1423
1424 // Smash all inputs to 'old', isolating him completely
1425 Node *temp = new Node(1);
1426 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
1427 remove_dead_node( old );
1428 temp->del_req(0); // Yank bogus edge
1429 if (nn != nullptr && nn->outcnt() == 0) {
1430 _worklist.push(nn);
1431 }
1432 #ifndef PRODUCT
1433 if (is_verify_def_use()) {
1434 for ( int i = 0; i < _verify_window_size; i++ ) {
1435 if ( _verify_window[i] == old )
1436 _verify_window[i] = nn;
1437 }
1438 }
1439 #endif
1440 temp->destruct(this); // reuse the _idx of this little guy
1441 }
1442
1443 //------------------------------add_users_to_worklist--------------------------
1444 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1445 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1446 _worklist.push(n->fast_out(i)); // Push on worklist
1447 }
1448 }
1449
1450 // Return counted loop Phi if as a counted loop exit condition, cmp
1451 // compares the induction variable with n
1452 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
1453 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1454 Node* bol = cmp->fast_out(i);
1455 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1456 Node* iff = bol->fast_out(i2);
1457 if (iff->is_BaseCountedLoopEnd()) {
1458 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
1459 if (cle->limit() == n) {
1460 PhiNode* phi = cle->phi();
1461 if (phi != nullptr) {
1462 return phi;
1463 }
1464 }
1465 }
1466 }
1467 }
1468 return nullptr;
1469 }
1470
1471 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1472 add_users_to_worklist0(n);
1473
1474 // Move users of node to worklist
1475 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1476 Node* use = n->fast_out(i); // Get use
1477
1478 if( use->is_Multi() || // Multi-definer? Push projs on worklist
1479 use->is_Store() ) // Enable store/load same address
1480 add_users_to_worklist0(use);
1481
1482 // If we changed the receiver type to a call, we need to revisit
1483 // the Catch following the call. It's looking for a non-null
1484 // receiver to know when to enable the regular fall-through path
1485 // in addition to the NullPtrException path.
1486 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1487 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1488 if (p != nullptr) {
1489 add_users_to_worklist0(p);
1490 }
1491 }
1492
1493 uint use_op = use->Opcode();
1494 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
1495 add_users_to_worklist(use); // Put Bool on worklist
1496 if (use->outcnt() > 0) {
1497 Node* bol = use->raw_out(0);
1498 if (bol->outcnt() > 0) {
1499 Node* iff = bol->raw_out(0);
1500 if (iff->outcnt() == 2) {
1501 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1502 // phi merging either 0 or 1 onto the worklist
1503 Node* ifproj0 = iff->raw_out(0);
1504 Node* ifproj1 = iff->raw_out(1);
1505 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1506 Node* region0 = ifproj0->raw_out(0);
1507 Node* region1 = ifproj1->raw_out(0);
1508 if( region0 == region1 )
1509 add_users_to_worklist0(region0);
1510 }
1511 }
1512 }
1513 }
1514 if (use_op == Op_CmpI || use_op == Op_CmpL) {
1515 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
1516 if (phi != nullptr) {
1517 // Input to the cmp of a loop exit check has changed, thus
1518 // the loop limit may have changed, which can then change the
1519 // range values of the trip-count Phi.
1520 _worklist.push(phi);
1521 }
1522 }
1523 if (use_op == Op_CmpI) {
1524 Node* cmp = use;
1525 Node* in1 = cmp->in(1);
1526 Node* in2 = cmp->in(2);
1527 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
1528 // Must also notify if in1 is modified and possibly turns into X (right pattern).
1529 //
1530 // in1 in2 in1 in2
1531 // | | | |
1532 // +--- | --+ | |
1533 // | | | | |
1534 // CmpINode | CmpINode
1535 // | | |
1536 // BoolNode | BoolNode
1537 // | | OR |
1538 // IfNode | IfNode
1539 // | | |
1540 // IfProj | IfProj X
1541 // | | | |
1542 // CastIINode CastIINode
1543 //
1544 if (in1 != in2) { // if they are equal, the CmpI can fold them away
1545 if (in1 == n) {
1546 // in1 modified -> could turn into X -> do traversal based on right pattern.
1547 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
1548 Node* bol = cmp->fast_out(i2); // For each Bool
1549 if (bol->is_Bool()) {
1550 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
1551 Node* iff = bol->fast_out(i3); // For each If
1552 if (iff->is_If()) {
1553 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
1554 Node* if_proj = iff->fast_out(i4); // For each IfProj
1555 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
1556 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
1557 Node* castii = if_proj->fast_out(i5); // For each CastII
1558 if (castii->is_CastII() &&
1559 castii->as_CastII()->carry_dependency()) {
1560 _worklist.push(castii);
1561 }
1562 }
1563 }
1564 }
1565 }
1566 }
1567 }
1568 } else {
1569 // Only in2 modified -> can assume X == in2 (left pattern).
1570 assert(n == in2, "only in2 modified");
1571 // Find all CastII with input in1.
1572 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
1573 Node* castii = in1->fast_out(j);
1574 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
1575 // Find If.
1576 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
1577 Node* ifnode = castii->in(0)->in(0);
1578 // Check that if connects to the cmp
1579 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
1580 _worklist.push(castii);
1581 }
1582 }
1583 }
1584 }
1585 }
1586 }
1587 }
1588 }
1589
1590 // If changed Cast input, notify down for Phi and Sub - both do "uncast"
1591 // Patterns:
1592 // ConstraintCast+ -> Sub
1593 // ConstraintCast+ -> Phi
1594 if (use->is_ConstraintCast()) {
1595 auto push_phi_or_sub_uses_to_worklist = [&](Node* n){
1596 if (n->is_Phi() || n->is_Sub()) {
1597 _worklist.push(n);
1598 }
1599 };
1600 ConstraintCastNode::visit_uncasted_uses(use, push_phi_or_sub_uses_to_worklist);
1601 }
1602 // If changed LShift inputs, check RShift users for useless sign-ext
1603 if( use_op == Op_LShiftI ) {
1604 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1605 Node* u = use->fast_out(i2);
1606 if (u->Opcode() == Op_RShiftI)
1607 _worklist.push(u);
1608 }
1609 }
1610 // If changed LShift inputs, check And users for shift and mask (And) operation
1611 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
1612 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1613 Node* u = use->fast_out(i2);
1614 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
1615 _worklist.push(u);
1616 }
1617 }
1618 }
1619 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1620 if (use_op == Op_AddI || use_op == Op_SubI) {
1621 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1622 Node* u = use->fast_out(i2);
1623 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1624 _worklist.push(u);
1625 }
1626 }
1627 }
1628 // If changed AddP inputs, check Stores for loop invariant
1629 if( use_op == Op_AddP ) {
1630 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1631 Node* u = use->fast_out(i2);
1632 if (u->is_Mem())
1633 _worklist.push(u);
1634 }
1635 }
1636 // If changed initialization activity, check dependent Stores
1637 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1638 InitializeNode* init = use->as_Allocate()->initialization();
1639 if (init != nullptr) {
1640 Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1641 if (imem != nullptr) add_users_to_worklist0(imem);
1642 }
1643 }
1644 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
1645 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
1646 // to guarantee the change is not missed.
1647 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
1648 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
1649 if (p != nullptr) {
1650 add_users_to_worklist0(p);
1651 }
1652 }
1653
1654 if (use_op == Op_Initialize) {
1655 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1656 if (imem != nullptr) add_users_to_worklist0(imem);
1657 }
1658 // Loading the java mirror from a Klass requires two loads and the type
1659 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1660 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
1661 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1662 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
1663
1664 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1665 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1666 Node* u = use->fast_out(i2);
1667 const Type* ut = u->bottom_type();
1668 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1669 if (has_load_barrier_nodes) {
1670 // Search for load barriers behind the load
1671 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
1672 Node* b = u->fast_out(i3);
1673 if (bs->is_gc_barrier_node(b)) {
1674 _worklist.push(b);
1675 }
1676 }
1677 }
1678 _worklist.push(u);
1679 }
1680 }
1681 }
1682 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
1683 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
1684 if (use->outcnt() == 1) {
1685 Node* cmp = use->unique_out();
1686 _worklist.push(cmp);
1687 }
1688 }
1689 }
1690 }
1691
1692 /**
1693 * Remove the speculative part of all types that we know of
1694 */
1695 void PhaseIterGVN::remove_speculative_types() {
1696 assert(UseTypeSpeculation, "speculation is off");
1697 for (uint i = 0; i < _types.Size(); i++) {
1698 const Type* t = _types.fast_lookup(i);
1699 if (t != nullptr) {
1700 _types.map(i, t->remove_speculative());
1701 }
1702 }
1703 _table.check_no_speculative_types();
1704 }
1705
1706 // Check if the type of a divisor of a Div or Mod node includes zero.
1707 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
1708 switch (n->Opcode()) {
1709 case Op_DivI:
1710 case Op_ModI: {
1711 // Type of divisor includes 0?
1712 if (type(n->in(2)) == Type::TOP) {
1713 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1714 return false;
1715 }
1716 const TypeInt* type_divisor = type(n->in(2))->is_int();
1717 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1718 }
1719 case Op_DivL:
1720 case Op_ModL: {
1721 // Type of divisor includes 0?
1722 if (type(n->in(2)) == Type::TOP) {
1723 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
1724 return false;
1725 }
1726 const TypeLong* type_divisor = type(n->in(2))->is_long();
1727 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
1728 }
1729 }
1730 return true;
1731 }
1732
1733 //=============================================================================
1734 #ifndef PRODUCT
1735 uint PhaseCCP::_total_invokes = 0;
1736 uint PhaseCCP::_total_constants = 0;
1737 #endif
1738 //------------------------------PhaseCCP---------------------------------------
1739 // Conditional Constant Propagation, ala Wegman & Zadeck
1740 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1741 NOT_PRODUCT( clear_constants(); )
1742 assert( _worklist.size() == 0, "" );
1743 analyze();
1744 }
1745
1746 #ifndef PRODUCT
1747 //------------------------------~PhaseCCP--------------------------------------
1748 PhaseCCP::~PhaseCCP() {
1749 inc_invokes();
1750 _total_constants += count_constants();
1751 }
1752 #endif
1753
1754
1755 #ifdef ASSERT
1756 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
1757 if (tnew->meet(told) != tnew->remove_speculative()) {
1758 n->dump(1);
1759 tty->print("told = "); told->dump(); tty->cr();
1760 tty->print("tnew = "); tnew->dump(); tty->cr();
1761 fatal("Not monotonic");
1762 }
1763 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
1764 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
1765 }
1766 #endif //ASSERT
1767
1768 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
1769 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
1770 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
1771 // or nodes even further down back to the worklist because their type could change as a result of the current type
1772 // change).
1773 void PhaseCCP::analyze() {
1774 // Initialize all types to TOP, optimistic analysis
1775 for (uint i = 0; i < C->unique(); i++) {
1776 _types.map(i, Type::TOP);
1777 }
1778
1779 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
1780 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
1781 // Compile is over. The local arena gets de-allocated at the end of its scope.
1782 ResourceArea local_arena(mtCompiler);
1783 Unique_Node_List worklist(&local_arena);
1784 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
1785
1786 // Push root onto worklist
1787 worklist.push(C->root());
1788
1789 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
1790 _root_and_safepoints.push(C->root());
1791
1792 // Pull from worklist; compute new value; push changes out.
1793 // This loop is the meat of CCP.
1794 while (worklist.size() != 0) {
1795 Node* n = fetch_next_node(worklist);
1796 DEBUG_ONLY(worklist_verify.push(n);)
1797 if (n->is_SafePoint()) {
1798 // Make sure safepoints are processed by PhaseCCP::transform even if they are
1799 // not reachable from the bottom. Otherwise, infinite loops would be removed.
1800 _root_and_safepoints.push(n);
1801 }
1802 const Type* new_type = n->Value(this);
1803 if (new_type != type(n)) {
1804 DEBUG_ONLY(verify_type(n, new_type, type(n));)
1805 dump_type_and_node(n, new_type);
1806 set_type(n, new_type);
1807 push_child_nodes_to_worklist(worklist, n);
1808 }
1809 }
1810 DEBUG_ONLY(verify_analyze(worklist_verify);)
1811 }
1812
1813 #ifdef ASSERT
1814 // For every node n on verify list, check if type(n) == n->Value()
1815 // We have a list of exceptions, see comments in verify_node_value.
1816 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
1817 bool failure = false;
1818 while (worklist_verify.size()) {
1819 Node* n = worklist_verify.pop();
1820 failure |= verify_node_value(n);
1821 }
1822 // If we get this assert, check why the reported nodes were not processed again in CCP.
1823 // We should either make sure that these nodes are properly added back to the CCP worklist
1824 // in PhaseCCP::push_child_nodes_to_worklist() to update their type or add an exception
1825 // in the verification code above if that is not possible for some reason (like Load nodes).
1826 assert(!failure, "Missed optimization opportunity in PhaseCCP");
1827 }
1828 #endif
1829
1830 // Fetch next node from worklist to be examined in this iteration.
1831 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
1832 if (StressCCP) {
1833 return worklist.remove(C->random() % worklist.size());
1834 } else {
1835 return worklist.pop();
1836 }
1837 }
1838
1839 #ifndef PRODUCT
1840 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
1841 if (TracePhaseCCP) {
1842 t->dump();
1843 do {
1844 tty->print("\t");
1845 } while (tty->position() < 16);
1846 n->dump();
1847 }
1848 }
1849 #endif
1850
1851 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
1852 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
1853 // of the new type of 'n'. Push these nodes to the worklist.
1854 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
1855 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1856 Node* use = n->fast_out(i);
1857 push_if_not_bottom_type(worklist, use);
1858 push_more_uses(worklist, n, use);
1859 }
1860 }
1861
1862 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
1863 if (n->bottom_type() != type(n)) {
1864 worklist.push(n);
1865 }
1866 }
1867
1868 // For some nodes, we need to propagate the type change to grandchildren or even further down.
1869 // Add them back to the worklist.
1870 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
1871 push_phis(worklist, use);
1872 push_catch(worklist, use);
1873 push_cmpu(worklist, use);
1874 push_counted_loop_phi(worklist, parent, use);
1875 push_loadp(worklist, use);
1876 push_and(worklist, parent, use);
1877 push_cast_ii(worklist, parent, use);
1878 push_opaque_zero_trip_guard(worklist, use);
1879 }
1880
1881
1882 // We must recheck Phis too if use is a Region.
1883 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
1884 if (use->is_Region()) {
1885 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1886 push_if_not_bottom_type(worklist, use->fast_out(i));
1887 }
1888 }
1889 }
1890
1891 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
1892 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
1893 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
1894 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
1895 if (use->is_Call()) {
1896 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1897 Node* proj = use->fast_out(i);
1898 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
1899 Node* catch_node = proj->find_out_with(Op_Catch);
1900 if (catch_node != nullptr) {
1901 worklist.push(catch_node);
1902 }
1903 }
1904 }
1905 }
1906 }
1907
1908 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
1909 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
1910 // and get wrong types otherwise.
1911 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
1912 uint use_op = use->Opcode();
1913 if (use_op == Op_AddI || use_op == Op_SubI) {
1914 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1915 Node* cmpu = use->fast_out(i);
1916 if (cmpu->Opcode() == Op_CmpU) {
1917 // Got a CmpU which might need the new type information from node n.
1918 push_if_not_bottom_type(worklist, cmpu);
1919 }
1920 }
1921 }
1922 }
1923
1924 // 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'.
1925 // Seem PhiNode::Value().
1926 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
1927 uint use_op = use->Opcode();
1928 if (use_op == Op_CmpI || use_op == Op_CmpL) {
1929 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
1930 if (phi != nullptr) {
1931 worklist.push(phi);
1932 }
1933 }
1934 }
1935
1936 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
1937 // See LoadNode::Value().
1938 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
1939 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
1940 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
1941
1942 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1943 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1944 Node* loadp = use->fast_out(i);
1945 const Type* ut = loadp->bottom_type();
1946 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
1947 if (has_load_barrier_nodes) {
1948 // Search for load barriers behind the load
1949 push_load_barrier(worklist, barrier_set, loadp);
1950 }
1951 worklist.push(loadp);
1952 }
1953 }
1954 }
1955 }
1956
1957 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
1958 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1959 Node* barrier_node = use->fast_out(i);
1960 if (barrier_set->is_gc_barrier_node(barrier_node)) {
1961 worklist.push(barrier_node);
1962 }
1963 }
1964 }
1965
1966 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3 to zero if they are bitwise disjoint.
1967 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the shift value changed.
1968 // Pattern: parent -> LShift (use) -> ConstraintCast* -> And
1969 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
1970 uint use_op = use->Opcode();
1971 if ((use_op == Op_LShiftI || use_op == Op_LShiftL)
1972 && use->in(2) == parent) { // is shift value (right-hand side of LShift)
1973 auto push_and_uses_to_worklist = [&](Node* n){
1974 uint opc = n->Opcode();
1975 if (opc == Op_AndI || opc == Op_AndL) {
1976 push_if_not_bottom_type(worklist, n);
1977 }
1978 };
1979 ConstraintCastNode::visit_uncasted_uses(use, push_and_uses_to_worklist);
1980 }
1981 }
1982
1983 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
1984 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
1985 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
1986 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
1987 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
1988 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
1989 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
1990 Node* other_cmp_input = use->in(1);
1991 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
1992 Node* cast_ii = other_cmp_input->fast_out(i);
1993 if (cast_ii->is_CastII()) {
1994 push_if_not_bottom_type(worklist, cast_ii);
1995 }
1996 }
1997 }
1998 }
1999
2000 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
2001 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2002 push_if_not_bottom_type(worklist, use->unique_out());
2003 }
2004 }
2005
2006 //------------------------------do_transform-----------------------------------
2007 // Top level driver for the recursive transformer
2008 void PhaseCCP::do_transform() {
2009 // Correct leaves of new-space Nodes; they point to old-space.
2010 C->set_root( transform(C->root())->as_Root() );
2011 assert( C->top(), "missing TOP node" );
2012 assert( C->root(), "missing root" );
2013 }
2014
2015 //------------------------------transform--------------------------------------
2016 // Given a Node in old-space, clone him into new-space.
2017 // Convert any of his old-space children into new-space children.
2018 Node *PhaseCCP::transform( Node *n ) {
2019 assert(n->is_Root(), "traversal must start at root");
2020 assert(_root_and_safepoints.member(n), "root (n) must be in list");
2021
2022 ResourceMark rm;
2023 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
2024 Node_List node_map;
2025 // Pre-allocate to avoid frequent realloc
2026 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
2027 // track all visited nodes, so that we can remove the complement
2028 Unique_Node_List useful;
2029
2030 // Initialize the traversal.
2031 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
2032 // 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
2033 // 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
2034 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
2035 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
2036 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
2037 // and all safepoints.
2038 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
2039 Node* nn = _root_and_safepoints.at(i);
2040 Node* new_node = node_map[nn->_idx];
2041 assert(new_node == nullptr, "");
2042 new_node = transform_once(nn); // Check for constant
2043 node_map.map(nn->_idx, new_node); // Flag as having been cloned
2044 transform_stack.push(new_node); // Process children of cloned node
2045 useful.push(new_node);
2046 }
2047
2048 while (transform_stack.is_nonempty()) {
2049 Node* clone = transform_stack.pop();
2050 uint cnt = clone->req();
2051 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
2052 Node *input = clone->in(i);
2053 if( input != nullptr ) { // Ignore nulls
2054 Node *new_input = node_map[input->_idx]; // Check for cloned input node
2055 if( new_input == nullptr ) {
2056 new_input = transform_once(input); // Check for constant
2057 node_map.map( input->_idx, new_input );// Flag as having been cloned
2058 transform_stack.push(new_input); // Process children of cloned node
2059 useful.push(new_input);
2060 }
2061 assert( new_input == clone->in(i), "insanity check");
2062 }
2063 }
2064 }
2065
2066 // The above transformation might lead to subgraphs becoming unreachable from the
2067 // bottom while still being reachable from the top. As a result, nodes in that
2068 // subgraph are not transformed and their bottom types are not updated, leading to
2069 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
2070 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
2071 // because their address type is inconsistent. Therefore, we aggressively remove
2072 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
2073 // to remove them anyway.
2074 if (C->cached_top_node()) {
2075 useful.push(C->cached_top_node());
2076 }
2077 C->update_dead_node_list(useful);
2078 remove_useless_nodes(useful.member_set());
2079 _worklist.remove_useless_nodes(useful.member_set());
2080 C->disconnect_useless_nodes(useful, _worklist);
2081
2082 Node* new_root = node_map[n->_idx];
2083 assert(new_root->is_Root(), "transformed root node must be a root node");
2084 return new_root;
2085 }
2086
2087 //------------------------------transform_once---------------------------------
2088 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
2089 Node *PhaseCCP::transform_once( Node *n ) {
2090 const Type *t = type(n);
2091 // Constant? Use constant Node instead
2092 if( t->singleton() ) {
2093 Node *nn = n; // Default is to return the original constant
2094 if( t == Type::TOP ) {
2095 // cache my top node on the Compile instance
2096 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
2097 C->set_cached_top_node(ConNode::make(Type::TOP));
2098 set_type(C->top(), Type::TOP);
2099 }
2100 nn = C->top();
2101 }
2102 if( !n->is_Con() ) {
2103 if( t != Type::TOP ) {
2104 nn = makecon(t); // ConNode::make(t);
2105 NOT_PRODUCT( inc_constants(); )
2106 } else if( n->is_Region() ) { // Unreachable region
2107 // Note: nn == C->top()
2108 n->set_req(0, nullptr); // Cut selfreference
2109 bool progress = true;
2110 uint max = n->outcnt();
2111 DUIterator i;
2112 while (progress) {
2113 progress = false;
2114 // Eagerly remove dead phis to avoid phis copies creation.
2115 for (i = n->outs(); n->has_out(i); i++) {
2116 Node* m = n->out(i);
2117 if (m->is_Phi()) {
2118 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
2119 replace_node(m, nn);
2120 if (max != n->outcnt()) {
2121 progress = true;
2122 i = n->refresh_out_pos(i);
2123 max = n->outcnt();
2124 }
2125 }
2126 }
2127 }
2128 }
2129 replace_node(n,nn); // Update DefUse edges for new constant
2130 }
2131 return nn;
2132 }
2133
2134 // If x is a TypeNode, capture any more-precise type permanently into Node
2135 if (t != n->bottom_type()) {
2136 hash_delete(n); // changing bottom type may force a rehash
2137 n->raise_bottom_type(t);
2138 _worklist.push(n); // n re-enters the hash table via the worklist
2139 }
2140
2141 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
2142 switch( n->Opcode() ) {
2143 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
2144 case Op_CallDynamicJava:
2145 case Op_FastLock: // Revisit FastLocks for lock coarsening
2146 case Op_If:
2147 case Op_CountedLoopEnd:
2148 case Op_Region:
2149 case Op_Loop:
2150 case Op_CountedLoop:
2151 case Op_Conv2B:
2152 case Op_Opaque1:
2153 _worklist.push(n);
2154 break;
2155 default:
2156 break;
2157 }
2158
2159 return n;
2160 }
2161
2162 //---------------------------------saturate------------------------------------
2163 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
2164 const Type* limit_type) const {
2165 const Type* wide_type = new_type->widen(old_type, limit_type);
2166 if (wide_type != new_type) { // did we widen?
2167 // If so, we may have widened beyond the limit type. Clip it back down.
2168 new_type = wide_type->filter(limit_type);
2169 }
2170 return new_type;
2171 }
2172
2173 //------------------------------print_statistics-------------------------------
2174 #ifndef PRODUCT
2175 void PhaseCCP::print_statistics() {
2176 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
2177 }
2178 #endif
2179
2180
2181 //=============================================================================
2182 #ifndef PRODUCT
2183 uint PhasePeephole::_total_peepholes = 0;
2184 #endif
2185 //------------------------------PhasePeephole----------------------------------
2186 // Conditional Constant Propagation, ala Wegman & Zadeck
2187 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
2188 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
2189 NOT_PRODUCT( clear_peepholes(); )
2190 }
2191
2192 #ifndef PRODUCT
2193 //------------------------------~PhasePeephole---------------------------------
2194 PhasePeephole::~PhasePeephole() {
2195 _total_peepholes += count_peepholes();
2196 }
2197 #endif
2198
2199 //------------------------------transform--------------------------------------
2200 Node *PhasePeephole::transform( Node *n ) {
2201 ShouldNotCallThis();
2202 return nullptr;
2203 }
2204
2205 //------------------------------do_transform-----------------------------------
2206 void PhasePeephole::do_transform() {
2207 bool method_name_not_printed = true;
2208
2209 // Examine each basic block
2210 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
2211 Block* block = _cfg.get_block(block_number);
2212 bool block_not_printed = true;
2213
2214 for (bool progress = true; progress;) {
2215 progress = false;
2216 // block->end_idx() not valid after PhaseRegAlloc
2217 uint end_index = block->number_of_nodes();
2218 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
2219 Node *n = block->get_node(instruction_index);
2220 if( n->is_Mach() ) {
2221 MachNode *m = n->as_Mach();
2222 // check for peephole opportunities
2223 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
2224 if( result != -1 ) {
2225 #ifndef PRODUCT
2226 if( PrintOptoPeephole ) {
2227 // Print method, first time only
2228 if( C->method() && method_name_not_printed ) {
2229 C->method()->print_short_name(); tty->cr();
2230 method_name_not_printed = false;
2231 }
2232 // Print this block
2233 if( Verbose && block_not_printed) {
2234 tty->print_cr("in block");
2235 block->dump();
2236 block_not_printed = false;
2237 }
2238 // Print the peephole number
2239 tty->print_cr("peephole number: %d", result);
2240 }
2241 inc_peepholes();
2242 #endif
2243 // Set progress, start again
2244 progress = true;
2245 break;
2246 }
2247 }
2248 }
2249 }
2250 }
2251 }
2252
2253 //------------------------------print_statistics-------------------------------
2254 #ifndef PRODUCT
2255 void PhasePeephole::print_statistics() {
2256 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
2257 }
2258 #endif
2259
2260
2261 //=============================================================================
2262 //------------------------------set_req_X--------------------------------------
2263 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2264 assert( is_not_dead(n), "can not use dead node");
2265 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2266 Node *old = in(i);
2267 set_req(i, n);
2268
2269 // old goes dead?
2270 if( old ) {
2271 switch (old->outcnt()) {
2272 case 0:
2273 // Put into the worklist to kill later. We do not kill it now because the
2274 // recursive kill will delete the current node (this) if dead-loop exists
2275 if (!old->is_top())
2276 igvn->_worklist.push( old );
2277 break;
2278 case 1:
2279 if( old->is_Store() || old->has_special_unique_user() )
2280 igvn->add_users_to_worklist( old );
2281 break;
2282 case 2:
2283 if( old->is_Store() )
2284 igvn->add_users_to_worklist( old );
2285 if( old->Opcode() == Op_Region )
2286 igvn->_worklist.push(old);
2287 break;
2288 case 3:
2289 if( old->Opcode() == Op_Region ) {
2290 igvn->_worklist.push(old);
2291 igvn->add_users_to_worklist( old );
2292 }
2293 break;
2294 default:
2295 break;
2296 }
2297
2298 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
2299 }
2300 }
2301
2302 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
2303 PhaseIterGVN* igvn = gvn->is_IterGVN();
2304 if (igvn == nullptr) {
2305 set_req(i, n);
2306 return;
2307 }
2308 set_req_X(i, n, igvn);
2309 }
2310
2311 //-------------------------------replace_by-----------------------------------
2312 // Using def-use info, replace one node for another. Follow the def-use info
2313 // to all users of the OLD node. Then make all uses point to the NEW node.
2314 void Node::replace_by(Node *new_node) {
2315 assert(!is_top(), "top node has no DU info");
2316 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2317 Node* use = last_out(i);
2318 uint uses_found = 0;
2319 for (uint j = 0; j < use->len(); j++) {
2320 if (use->in(j) == this) {
2321 if (j < use->req())
2322 use->set_req(j, new_node);
2323 else use->set_prec(j, new_node);
2324 uses_found++;
2325 }
2326 }
2327 i -= uses_found; // we deleted 1 or more copies of this edge
2328 }
2329 }
2330
2331 //=============================================================================
2332 //-----------------------------------------------------------------------------
2333 void Type_Array::grow( uint i ) {
2334 if( !_max ) {
2335 _max = 1;
2336 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2337 _types[0] = nullptr;
2338 }
2339 uint old = _max;
2340 _max = next_power_of_2(i);
2341 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2342 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2343 }
2344
2345 //------------------------------dump-------------------------------------------
2346 #ifndef PRODUCT
2347 void Type_Array::dump() const {
2348 uint max = Size();
2349 for( uint i = 0; i < max; i++ ) {
2350 if( _types[i] != nullptr ) {
2351 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
2352 }
2353 }
2354 }
2355 #endif