1 /*
2 * Copyright (c) 1997, 2026, 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 "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/idealGraphPrinter.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/opcodes.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/regalloc.hpp"
41 #include "opto/rootnode.hpp"
42 #include "utilities/macros.hpp"
43 #include "utilities/powerOfTwo.hpp"
44
45 //=============================================================================
46 #define NODE_HASH_MINIMUM_SIZE 255
47
48 //------------------------------NodeHash---------------------------------------
49 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
50 _a(arena),
51 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
52 _inserts(0), _insert_limit( insert_limit() ),
53 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
54 #ifndef PRODUCT
55 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
56 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
57 _total_inserts(0), _total_insert_probes(0)
58 #endif
59 {
60 // _sentinel must be in the current node space
61 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
62 memset(_table,0,sizeof(Node*)*_max);
63 }
64
65 //------------------------------hash_find--------------------------------------
66 // Find in hash table
67 Node *NodeHash::hash_find( const Node *n ) {
68 // ((Node*)n)->set_hash( n->hash() );
69 uint hash = n->hash();
70 if (hash == Node::NO_HASH) {
71 NOT_PRODUCT( _lookup_misses++ );
72 return nullptr;
73 }
74 uint key = hash & (_max-1);
75 uint stride = key | 0x01;
76 NOT_PRODUCT( _look_probes++ );
77 Node *k = _table[key]; // Get hashed value
78 if( !k ) { // ?Miss?
79 NOT_PRODUCT( _lookup_misses++ );
80 return nullptr; // Miss!
81 }
82
83 int op = n->Opcode();
84 uint req = n->req();
85 while( 1 ) { // While probing hash table
86 if( k->req() == req && // Same count of inputs
87 k->Opcode() == op ) { // Same Opcode
88 for( uint i=0; i<req; i++ )
89 if( n->in(i)!=k->in(i)) // Different inputs?
90 goto collision; // "goto" is a speed hack...
91 if( n->cmp(*k) ) { // Check for any special bits
92 NOT_PRODUCT( _lookup_hits++ );
93 return k; // Hit!
94 }
95 }
96 collision:
97 NOT_PRODUCT( _look_probes++ );
98 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
99 k = _table[key]; // Get hashed value
100 if( !k ) { // ?Miss?
101 NOT_PRODUCT( _lookup_misses++ );
102 return nullptr; // Miss!
103 }
104 }
105 ShouldNotReachHere();
106 return nullptr;
107 }
108
109 //------------------------------hash_find_insert-------------------------------
110 // Find in hash table, insert if not already present
111 // Used to preserve unique entries in hash table
112 Node *NodeHash::hash_find_insert( Node *n ) {
113 // n->set_hash( );
114 uint hash = n->hash();
115 if (hash == Node::NO_HASH) {
116 NOT_PRODUCT( _lookup_misses++ );
117 return nullptr;
118 }
119 uint key = hash & (_max-1);
120 uint stride = key | 0x01; // stride must be relatively prime to table siz
121 uint first_sentinel = 0; // replace a sentinel if seen.
122 NOT_PRODUCT( _look_probes++ );
123 Node *k = _table[key]; // Get hashed value
124 if( !k ) { // ?Miss?
125 NOT_PRODUCT( _lookup_misses++ );
126 _table[key] = n; // Insert into table!
127 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
128 check_grow(); // Grow table if insert hit limit
129 return nullptr; // Miss!
130 }
131 else if( k == _sentinel ) {
132 first_sentinel = key; // Can insert here
133 }
134
135 int op = n->Opcode();
136 uint req = n->req();
137 while( 1 ) { // While probing hash table
138 if( k->req() == req && // Same count of inputs
139 k->Opcode() == op ) { // Same Opcode
140 for( uint i=0; i<req; i++ )
141 if( n->in(i)!=k->in(i)) // Different inputs?
142 goto collision; // "goto" is a speed hack...
143 if( n->cmp(*k) ) { // Check for any special bits
144 NOT_PRODUCT( _lookup_hits++ );
145 return k; // Hit!
146 }
147 }
148 collision:
149 NOT_PRODUCT( _look_probes++ );
150 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
151 k = _table[key]; // Get hashed value
152 if( !k ) { // ?Miss?
153 NOT_PRODUCT( _lookup_misses++ );
154 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
155 _table[key] = n; // Insert into table!
156 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
157 check_grow(); // Grow table if insert hit limit
158 return nullptr; // Miss!
159 }
160 else if( first_sentinel == 0 && k == _sentinel ) {
161 first_sentinel = key; // Can insert here
162 }
163
164 }
165 ShouldNotReachHere();
166 return nullptr;
167 }
168
169 //------------------------------hash_insert------------------------------------
170 // Insert into hash table
171 void NodeHash::hash_insert( Node *n ) {
172 // // "conflict" comments -- print nodes that conflict
173 // bool conflict = false;
174 // n->set_hash();
175 uint hash = n->hash();
176 if (hash == Node::NO_HASH) {
177 return;
178 }
179 check_grow();
180 uint key = hash & (_max-1);
181 uint stride = key | 0x01;
182
183 while( 1 ) { // While probing hash table
184 NOT_PRODUCT( _insert_probes++ );
185 Node *k = _table[key]; // Get hashed value
186 if( !k || (k == _sentinel) ) break; // Found a slot
187 assert( k != n, "already inserted" );
188 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
189 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
190 }
191 _table[key] = n; // Insert into table!
192 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
193 // if( conflict ) { n->dump(); }
194 }
195
196 //------------------------------hash_delete------------------------------------
197 // Replace in hash table with sentinel
198 bool NodeHash::hash_delete( const Node *n ) {
199 Node *k;
200 uint hash = n->hash();
201 if (hash == Node::NO_HASH) {
202 NOT_PRODUCT( _delete_misses++ );
203 return false;
204 }
205 uint key = hash & (_max-1);
206 uint stride = key | 0x01;
207 DEBUG_ONLY( uint counter = 0; );
208 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
209 DEBUG_ONLY( counter++ );
210 NOT_PRODUCT( _delete_probes++ );
211 k = _table[key]; // Get hashed value
212 if( !k ) { // Miss?
213 NOT_PRODUCT( _delete_misses++ );
214 return false; // Miss! Not in chain
215 }
216 else if( n == k ) {
217 NOT_PRODUCT( _delete_hits++ );
218 _table[key] = _sentinel; // Hit! Label as deleted entry
219 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
220 return true;
221 }
222 else {
223 // collision: move through table with prime offset
224 key = (key + stride/*7*/) & (_max-1);
225 assert( counter <= _insert_limit, "Cycle in hash-table");
226 }
227 }
228 ShouldNotReachHere();
229 return false;
230 }
231
232 //------------------------------round_up---------------------------------------
233 // Round up to nearest power of 2
234 uint NodeHash::round_up(uint x) {
235 x += (x >> 2); // Add 25% slop
236 return MAX2(16U, round_up_power_of_2(x));
237 }
238
239 //------------------------------grow-------------------------------------------
240 // Grow _table to next power of 2 and insert old entries
241 void NodeHash::grow() {
242 // Record old state
243 uint old_max = _max;
244 Node **old_table = _table;
245 // Construct new table with twice the space
246 #ifndef PRODUCT
247 _grows++;
248 _total_inserts += _inserts;
249 _total_insert_probes += _insert_probes;
250 _insert_probes = 0;
251 #endif
252 _inserts = 0;
253 _max = _max << 1;
254 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
255 memset(_table,0,sizeof(Node*)*_max);
256 _insert_limit = insert_limit();
257 // Insert old entries into the new table
258 for( uint i = 0; i < old_max; i++ ) {
259 Node *m = *old_table++;
260 if( !m || m == _sentinel ) continue;
261 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
262 hash_insert(m);
263 }
264 }
265
266 //------------------------------clear------------------------------------------
267 // Clear all entries in _table to null but keep storage
268 void NodeHash::clear() {
269 #ifdef ASSERT
270 // Unlock all nodes upon removal from table.
271 for (uint i = 0; i < _max; i++) {
272 Node* n = _table[i];
273 if (!n || n == _sentinel) continue;
274 n->exit_hash_lock();
275 }
276 #endif
277
278 memset( _table, 0, _max * sizeof(Node*) );
279 }
280
281 //-----------------------remove_useless_nodes----------------------------------
282 // Remove useless nodes from value table,
283 // implementation does not depend on hash function
284 void NodeHash::remove_useless_nodes(VectorSet &useful) {
285
286 // Dead nodes in the hash table inherited from GVN should not replace
287 // existing nodes, remove dead nodes.
288 uint max = size();
289 Node *sentinel_node = sentinel();
290 for( uint i = 0; i < max; ++i ) {
291 Node *n = at(i);
292 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
293 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
294 _table[i] = sentinel_node; // Replace with placeholder
295 }
296 }
297 }
298
299
300 void NodeHash::check_no_speculative_types() {
301 #ifdef ASSERT
302 uint max = size();
303 Unique_Node_List live_nodes;
304 Compile::current()->identify_useful_nodes(live_nodes);
305 Node *sentinel_node = sentinel();
306 for (uint i = 0; i < max; ++i) {
307 Node *n = at(i);
308 if (n != nullptr &&
309 n != sentinel_node &&
310 n->is_Type() &&
311 live_nodes.member(n)) {
312 TypeNode* tn = n->as_Type();
313 const Type* t = tn->type();
314 const Type* t_no_spec = t->remove_speculative();
315 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
316 }
317 }
318 #endif
319 }
320
321 #ifndef PRODUCT
322 //------------------------------dump-------------------------------------------
323 // Dump statistics for the hash table
324 void NodeHash::dump() {
325 _total_inserts += _inserts;
326 _total_insert_probes += _insert_probes;
327 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
328 if (WizardMode) {
329 for (uint i=0; i<_max; i++) {
330 if (_table[i])
331 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
332 }
333 }
334 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
335 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
336 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
337 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
338 // sentinels increase lookup cost, but not insert cost
339 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
340 assert( _inserts+(_inserts>>3) < _max, "table too full" );
341 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
342 }
343 }
344
345 Node *NodeHash::find_index(uint idx) { // For debugging
346 // Find an entry by its index value
347 for( uint i = 0; i < _max; i++ ) {
348 Node *m = _table[i];
349 if( !m || m == _sentinel ) continue;
350 if( m->_idx == (uint)idx ) return m;
351 }
352 return nullptr;
353 }
354 #endif
355
356 #ifdef ASSERT
357 NodeHash::~NodeHash() {
358 // Unlock all nodes upon destruction of table.
359 if (_table != (Node**)badAddress) clear();
360 }
361 #endif
362
363 // Add users of 'n' that match 'predicate' to worklist
364 template <class Predicate>
365 static void add_users_to_worklist_if(Unique_Node_List& worklist, const Node* n, Predicate predicate) {
366 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
367 Node* u = n->fast_out(i);
368 if (predicate(u)) {
369 worklist.push(u);
370 }
371 }
372 }
373
374 //=============================================================================
375 //------------------------------PhaseRemoveUseless-----------------------------
376 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
377 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
378 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
379 // Implementation requires an edge from root to each SafePointNode
380 // at a backward branch. Inserted in add_safepoint().
381
382 // Identify nodes that are reachable from below, useful.
383 C->identify_useful_nodes(_useful);
384 // Update dead node list
385 C->update_dead_node_list(_useful);
386
387 // Remove all useless nodes from PhaseValues' recorded types
388 // Must be done before disconnecting nodes to preserve hash-table-invariant
389 gvn->remove_useless_nodes(_useful.member_set());
390
391 // Remove all useless nodes from future worklist
392 worklist.remove_useless_nodes(_useful.member_set());
393
394 // Disconnect 'useless' nodes that are adjacent to useful nodes
395 C->disconnect_useless_nodes(_useful, worklist);
396 }
397
398 //=============================================================================
399 //------------------------------PhaseRenumberLive------------------------------
400 // First, remove useless nodes (equivalent to identifying live nodes).
401 // Then, renumber live nodes.
402 //
403 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
404 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
405 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
406 // value in the range [0, x).
407 //
408 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
409 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
410 //
411 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
412 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
413 // be processed by IGVN after removal of the useless nodes.
414 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
415 // the node's type. The mapping is updated to use the new node IDs as well. We
416 // create a new map, and swap it with the old one.
417 //
418 // Other data structures used by the compiler are not updated. The hash table for value
419 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
420 // computing the hash values is not based on node IDs.
421 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
422 Unique_Node_List& worklist,
423 PhaseNumber phase_num) :
424 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
425 _new_type_array(C->comp_arena()),
426 _old2new_map(C->unique(), C->unique(), -1),
427 _is_pass_finished(false),
428 _live_node_count(C->live_nodes())
429 {
430 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
431 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
432 assert(_delayed.size() == 0, "should be empty");
433 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
434 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
435
436 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
437 if (old_node_note_array != nullptr) {
438 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
439 new_size = MAX2(8, new_size);
440 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
441 C->grow_node_notes(C->node_note_array(), new_size);
442 }
443
444 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
445
446 // Iterate over the set of live nodes.
447 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
448 Node* n = _useful.at(current_idx);
449
450 const Type* type = gvn->type_or_null(n);
451 _new_type_array.map(current_idx, type);
452
453 assert(_old2new_map.at(n->_idx) == -1, "already seen");
454 _old2new_map.at_put(n->_idx, current_idx);
455
456 if (old_node_note_array != nullptr) {
457 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
458 C->set_node_notes_at(current_idx, nn);
459 }
460
461 n->set_idx(current_idx); // Update node ID.
462
463 if (update_embedded_ids(n) < 0) {
464 _delayed.push(n); // has embedded IDs; handle later
465 }
466 }
467
468 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
469 worklist.recompute_idx_set();
470
471 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
472
473 _is_pass_finished = true; // pass finished; safe to process delayed updates
474
475 while (_delayed.size() > 0) {
476 Node* n = _delayed.pop();
477 int no_of_updates = update_embedded_ids(n);
478 assert(no_of_updates > 0, "should be updated");
479 }
480
481 // Replace the compiler's type information with the updated type information.
482 gvn->types().swap(_new_type_array);
483
484 // Update the unique node count of the compilation to the number of currently live nodes.
485 C->set_unique(_live_node_count);
486
487 // Set the dead node count to 0 and reset dead node list.
488 C->reset_dead_node_list();
489 }
490
491 int PhaseRenumberLive::new_index(int old_idx) {
492 assert(_is_pass_finished, "not finished");
493 if (_old2new_map.at(old_idx) == -1) { // absent
494 // Allocate a placeholder to preserve uniqueness
495 _old2new_map.at_put(old_idx, _live_node_count);
496 _live_node_count++;
497 }
498 return _old2new_map.at(old_idx);
499 }
500
501 int PhaseRenumberLive::update_embedded_ids(Node* n) {
502 int no_of_updates = 0;
503 if (n->is_Phi()) {
504 PhiNode* phi = n->as_Phi();
505 if (phi->_inst_id != -1) {
506 if (!_is_pass_finished) {
507 return -1; // delay
508 }
509 int new_idx = new_index(phi->_inst_id);
510 assert(new_idx != -1, "");
511 phi->_inst_id = new_idx;
512 no_of_updates++;
513 }
514 if (phi->_inst_mem_id != -1) {
515 if (!_is_pass_finished) {
516 return -1; // delay
517 }
518 int new_idx = new_index(phi->_inst_mem_id);
519 assert(new_idx != -1, "");
520 phi->_inst_mem_id = new_idx;
521 no_of_updates++;
522 }
523 }
524
525 const Type* type = _new_type_array.fast_lookup(n->_idx);
526 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
527 if (!_is_pass_finished) {
528 return -1; // delay
529 }
530 int old_idx = type->is_oopptr()->instance_id();
531 int new_idx = new_index(old_idx);
532 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
533 _new_type_array.map(n->_idx, new_type);
534 no_of_updates++;
535 }
536
537 return no_of_updates;
538 }
539
540 void PhaseValues::init_con_caches() {
541 memset(_icons,0,sizeof(_icons));
542 memset(_lcons,0,sizeof(_lcons));
543 memset(_zcons,0,sizeof(_zcons));
544 }
545
546 PhaseIterGVN* PhaseValues::is_IterGVN() {
547 return (_phase == PhaseValuesType::iter_gvn || _phase == PhaseValuesType::ccp) ? static_cast<PhaseIterGVN*>(this) : nullptr;
548 }
549
550 //--------------------------------find_int_type--------------------------------
551 const TypeInt* PhaseValues::find_int_type(Node* n) {
552 if (n == nullptr) return nullptr;
553 // Call type_or_null(n) to determine node's type since we might be in
554 // parse phase and call n->Value() may return wrong type.
555 // (For example, a phi node at the beginning of loop parsing is not ready.)
556 const Type* t = type_or_null(n);
557 if (t == nullptr) return nullptr;
558 return t->isa_int();
559 }
560
561
562 //-------------------------------find_long_type--------------------------------
563 const TypeLong* PhaseValues::find_long_type(Node* n) {
564 if (n == nullptr) return nullptr;
565 // (See comment above on type_or_null.)
566 const Type* t = type_or_null(n);
567 if (t == nullptr) return nullptr;
568 return t->isa_long();
569 }
570
571 //------------------------------~PhaseValues-----------------------------------
572 #ifndef PRODUCT
573 PhaseValues::~PhaseValues() {
574 // Statistics for NodeHash
575 _table.dump();
576 // Statistics for value progress and efficiency
577 if( PrintCompilation && Verbose && WizardMode ) {
578 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
579 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
580 if( made_transforms() != 0 ) {
581 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
582 } else {
583 tty->cr();
584 }
585 }
586 }
587 #endif
588
589 //------------------------------makecon----------------------------------------
590 ConNode* PhaseValues::makecon(const Type* t) {
591 assert(t->singleton(), "must be a constant");
592 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
593 switch (t->base()) { // fast paths
594 case Type::Half:
595 case Type::Top: return (ConNode*) C->top();
596 case Type::Int: return intcon( t->is_int()->get_con() );
597 case Type::Long: return longcon( t->is_long()->get_con() );
598 default: break;
599 }
600 if (t->is_zero_type())
601 return zerocon(t->basic_type());
602 return uncached_makecon(t);
603 }
604
605 //--------------------------uncached_makecon-----------------------------------
606 // Make an idealized constant - one of ConINode, ConPNode, etc.
607 ConNode* PhaseValues::uncached_makecon(const Type *t) {
608 assert(t->singleton(), "must be a constant");
609 ConNode* x = ConNode::make(t);
610 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
611 if (k == nullptr) {
612 set_type(x, t); // Missed, provide type mapping
613 GrowableArray<Node_Notes*>* nna = C->node_note_array();
614 if (nna != nullptr) {
615 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
616 loc->clear(); // do not put debug info on constants
617 }
618 } else {
619 x->destruct(this); // Hit, destroy duplicate constant
620 x = k; // use existing constant
621 }
622 return x;
623 }
624
625 //------------------------------intcon-----------------------------------------
626 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
627 ConINode* PhaseValues::intcon(jint i) {
628 // Small integer? Check cache! Check that cached node is not dead
629 if (i >= _icon_min && i <= _icon_max) {
630 ConINode* icon = _icons[i-_icon_min];
631 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
632 return icon;
633 }
634 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
635 assert(icon->is_Con(), "");
636 if (i >= _icon_min && i <= _icon_max)
637 _icons[i-_icon_min] = icon; // Cache small integers
638 return icon;
639 }
640
641 //------------------------------longcon----------------------------------------
642 // Fast long constant.
643 ConLNode* PhaseValues::longcon(jlong l) {
644 // Small integer? Check cache! Check that cached node is not dead
645 if (l >= _lcon_min && l <= _lcon_max) {
646 ConLNode* lcon = _lcons[l-_lcon_min];
647 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
648 return lcon;
649 }
650 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
651 assert(lcon->is_Con(), "");
652 if (l >= _lcon_min && l <= _lcon_max)
653 _lcons[l-_lcon_min] = lcon; // Cache small integers
654 return lcon;
655 }
656 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
657 if (bt == T_INT) {
658 return intcon(checked_cast<jint>(l));
659 }
660 assert(bt == T_LONG, "not an integer");
661 return longcon(l);
662 }
663
664
665 //------------------------------zerocon-----------------------------------------
666 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
667 ConNode* PhaseValues::zerocon(BasicType bt) {
668 assert((uint)bt <= _zcon_max, "domain check");
669 ConNode* zcon = _zcons[bt];
670 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
671 return zcon;
672 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
673 _zcons[bt] = zcon;
674 return zcon;
675 }
676
677
678
679 //=============================================================================
680 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
681 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
682 if (i == nullptr) {
683 i = k->Ideal(this, can_reshape);
684 }
685 return i;
686 }
687
688 //------------------------------transform--------------------------------------
689 // Return a node which computes the same function as this node, but
690 // in a faster or cheaper fashion.
691 Node* PhaseGVN::transform(Node* n) {
692 NOT_PRODUCT( set_transforms(); )
693
694 // Apply the Ideal call in a loop until it no longer applies
695 Node* k = n;
696 Node* i = apply_ideal(k, /*can_reshape=*/false);
697 NOT_PRODUCT(uint loop_count = 1;)
698 while (i != nullptr) {
699 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
700 k = i;
701 #ifdef ASSERT
702 if (loop_count >= K + C->live_nodes()) {
703 dump_infinite_loop_info(i, "PhaseGVN::transform");
704 }
705 #endif
706 i = apply_ideal(k, /*can_reshape=*/false);
707 NOT_PRODUCT(loop_count++;)
708 }
709 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
710
711 // If brand new node, make space in type array.
712 ensure_type_or_null(k);
713
714 // Since I just called 'Value' to compute the set of run-time values
715 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
716 // cache Value. Later requests for the local phase->type of this Node can
717 // use the cached Value instead of suffering with 'bottom_type'.
718 const Type* t = k->Value(this); // Get runtime Value set
719 assert(t != nullptr, "value sanity");
720 if (type_or_null(k) != t) {
721 #ifndef PRODUCT
722 // Do not count initial visit to node as a transformation
723 if (type_or_null(k) == nullptr) {
724 inc_new_values();
725 set_progress();
726 }
727 #endif
728 set_type(k, t);
729 // If k is a TypeNode, capture any more-precise type permanently into Node
730 k->raise_bottom_type(t);
731 }
732
733 if (t->singleton() && !k->is_Con()) {
734 NOT_PRODUCT(set_progress();)
735 return makecon(t); // Turn into a constant
736 }
737
738 // Now check for Identities
739 i = k->Identity(this); // Look for a nearby replacement
740 if (i != k) { // Found? Return replacement!
741 NOT_PRODUCT(set_progress();)
742 return i;
743 }
744
745 // Global Value Numbering
746 i = hash_find_insert(k); // Insert if new
747 if (i && (i != k)) {
748 // Return the pre-existing node
749 NOT_PRODUCT(set_progress();)
750 return i;
751 }
752
753 // Return Idealized original
754 return k;
755 }
756
757 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
758 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
759 return false;
760 }
761 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
762 return false;
763 }
764 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
765 int i = 0;
766 while (d != n) {
767 n = IfNode::up_one_dom(n, linear_only);
768 i++;
769 if (n == nullptr || i >= 100) {
770 return false;
771 }
772 }
773 return true;
774 }
775
776 #ifdef ASSERT
777 //------------------------------dead_loop_check--------------------------------
778 // Check for a simple dead loop when a data node references itself directly
779 // or through an other data node excluding cons and phis.
780 void PhaseGVN::dead_loop_check(Node* n) {
781 // Phi may reference itself in a loop.
782 if (n == nullptr || n->is_dead_loop_safe() || n->is_CFG()) {
783 return;
784 }
785
786 // Do 2 levels check and only data inputs.
787 for (uint i = 1; i < n->req(); i++) {
788 Node* in = n->in(i);
789 if (in == n) {
790 n->dump_bfs(100, nullptr, "");
791 fatal("Dead loop detected, node references itself: %s (%d)",
792 n->Name(), n->_idx);
793 }
794
795 if (in == nullptr || in->is_dead_loop_safe()) {
796 continue;
797 }
798 for (uint j = 1; j < in->req(); j++) {
799 if (in->in(j) == n) {
800 n->dump_bfs(100, nullptr, "");
801 fatal("Dead loop detected, node input references current node: %s (%d) -> %s (%d)",
802 in->Name(), in->_idx, n->Name(), n->_idx);
803 }
804 if (in->in(j) == in) {
805 n->dump_bfs(100, nullptr, "");
806 fatal("Dead loop detected, node input references itself: %s (%d)",
807 in->Name(), in->_idx);
808 }
809 }
810 }
811 }
812
813
814 /**
815 * Dumps information that can help to debug the problem. A debug
816 * build fails with an assert.
817 */
818 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
819 n->dump(4);
820 assert(false, "infinite loop in %s", where);
821 }
822 #endif
823
824 //=============================================================================
825 //------------------------------PhaseIterGVN-----------------------------------
826 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
827 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
828 _worklist(*C->igvn_worklist())
829 {
830 _phase = PhaseValuesType::iter_gvn;
831 assert(&_worklist == &igvn->_worklist, "sanity");
832 }
833
834 //------------------------------PhaseIterGVN-----------------------------------
835 // Initialize from scratch
836 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
837 _worklist(*C->igvn_worklist())
838 {
839 _phase = PhaseValuesType::iter_gvn;
840 uint max;
841
842 // Dead nodes in the hash table inherited from GVN were not treated as
843 // roots during def-use info creation; hence they represent an invisible
844 // use. Clear them out.
845 max = _table.size();
846 for( uint i = 0; i < max; ++i ) {
847 Node *n = _table.at(i);
848 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
849 if( n->is_top() ) continue;
850 // If remove_useless_nodes() has run, we expect no such nodes left.
851 assert(false, "remove_useless_nodes missed this node");
852 hash_delete(n);
853 }
854 }
855
856 // Any Phis or Regions on the worklist probably had uses that could not
857 // make more progress because the uses were made while the Phis and Regions
858 // were in half-built states. Put all uses of Phis and Regions on worklist.
859 max = _worklist.size();
860 for( uint j = 0; j < max; j++ ) {
861 Node *n = _worklist.at(j);
862 uint uop = n->Opcode();
863 if( uop == Op_Phi || uop == Op_Region ||
864 n->is_Type() ||
865 n->is_Mem() )
866 add_users_to_worklist(n);
867 }
868 }
869
870 void PhaseIterGVN::shuffle_worklist() {
871 if (_worklist.size() < 2) return;
872 for (uint i = _worklist.size() - 1; i >= 1; i--) {
873 uint j = C->random() % (i + 1);
874 swap(_worklist.adr()[i], _worklist.adr()[j]);
875 }
876 }
877
878 #ifndef PRODUCT
879 void PhaseIterGVN::verify_step(Node* n) {
880 if (is_verify_def_use()) {
881 ResourceMark rm;
882 VectorSet visited;
883 Node_List worklist;
884
885 _verify_window[_verify_counter % _verify_window_size] = n;
886 ++_verify_counter;
887 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
888 ++_verify_full_passes;
889 worklist.push(C->root());
890 Node::verify(-1, visited, worklist);
891 return;
892 }
893 for (int i = 0; i < _verify_window_size; i++) {
894 Node* n = _verify_window[i];
895 if (n == nullptr) {
896 continue;
897 }
898 if (n->in(0) == NodeSentinel) { // xform_idom
899 _verify_window[i] = n->in(1);
900 --i;
901 continue;
902 }
903 // Typical fanout is 1-2, so this call visits about 6 nodes.
904 if (!visited.test_set(n->_idx)) {
905 worklist.push(n);
906 }
907 }
908 Node::verify(4, visited, worklist);
909 }
910 }
911
912 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
913 const Type* newtype = type_or_null(n);
914 if (nn != n || oldtype != newtype) {
915 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
916 }
917 if (TraceIterativeGVN) {
918 uint wlsize = _worklist.size();
919 if (nn != n) {
920 // print old node
921 tty->print("< ");
922 if (oldtype != newtype && oldtype != nullptr) {
923 oldtype->dump();
924 }
925 do { tty->print("\t"); } while (tty->position() < 16);
926 tty->print("<");
927 n->dump();
928 }
929 if (oldtype != newtype || nn != n) {
930 // print new node and/or new type
931 if (oldtype == nullptr) {
932 tty->print("* ");
933 } else if (nn != n) {
934 tty->print("> ");
935 } else {
936 tty->print("= ");
937 }
938 if (newtype == nullptr) {
939 tty->print("null");
940 } else {
941 newtype->dump();
942 }
943 do { tty->print("\t"); } while (tty->position() < 16);
944 nn->dump();
945 }
946 if (Verbose && wlsize < _worklist.size()) {
947 tty->print(" Push {");
948 while (wlsize != _worklist.size()) {
949 Node* pushed = _worklist.at(wlsize++);
950 tty->print(" %d", pushed->_idx);
951 }
952 tty->print_cr(" }");
953 }
954 if (nn != n) {
955 // ignore n, it might be subsumed
956 verify_step((Node*) nullptr);
957 }
958 }
959 }
960
961 void PhaseIterGVN::init_verifyPhaseIterGVN() {
962 _verify_counter = 0;
963 _verify_full_passes = 0;
964 for (int i = 0; i < _verify_window_size; i++) {
965 _verify_window[i] = nullptr;
966 }
967 #ifdef ASSERT
968 // Verify that all modified nodes are on _worklist
969 Unique_Node_List* modified_list = C->modified_nodes();
970 while (modified_list != nullptr && modified_list->size()) {
971 Node* n = modified_list->pop();
972 if (!n->is_Con() && !_worklist.member(n)) {
973 n->dump();
974 fatal("modified node is not on IGVN._worklist");
975 }
976 }
977 #endif
978 }
979
980 void PhaseIterGVN::verify_PhaseIterGVN() {
981 #ifdef ASSERT
982 // Verify nodes with changed inputs.
983 Unique_Node_List* modified_list = C->modified_nodes();
984 while (modified_list != nullptr && modified_list->size()) {
985 Node* n = modified_list->pop();
986 if (!n->is_Con()) { // skip Con nodes
987 n->dump();
988 fatal("modified node was not processed by IGVN.transform_old()");
989 }
990 }
991 #endif
992
993 C->verify_graph_edges();
994 if (is_verify_def_use() && PrintOpto) {
995 if (_verify_counter == _verify_full_passes) {
996 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
997 (int) _verify_full_passes);
998 } else {
999 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1000 (int) _verify_counter, (int) _verify_full_passes);
1001 }
1002 }
1003
1004 #ifdef ASSERT
1005 if (modified_list != nullptr) {
1006 while (modified_list->size() > 0) {
1007 Node* n = modified_list->pop();
1008 n->dump();
1009 assert(false, "VerifyIterativeGVN: new modified node was added");
1010 }
1011 }
1012
1013 verify_optimize();
1014 #endif
1015 }
1016 #endif /* PRODUCT */
1017
1018 #ifdef ASSERT
1019 /**
1020 * Dumps information that can help to debug the problem. A debug
1021 * build fails with an assert.
1022 */
1023 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1024 n->dump(4);
1025 _worklist.dump();
1026 assert(false, "infinite loop in %s", where);
1027 }
1028
1029 /**
1030 * Prints out information about IGVN if the 'verbose' option is used.
1031 */
1032 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1033 if (TraceIterativeGVN && Verbose) {
1034 tty->print(" Pop ");
1035 n->dump();
1036 if ((num_processed % 100) == 0) {
1037 _worklist.print_set();
1038 }
1039 }
1040 }
1041 #endif /* ASSERT */
1042
1043 void PhaseIterGVN::optimize() {
1044 DEBUG_ONLY(uint num_processed = 0;)
1045 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1046 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1047 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1048 if (StressIGVN) {
1049 shuffle_worklist();
1050 }
1051
1052 // The node count check in the loop below (check_node_count) assumes that we
1053 // increase the live node count with at most
1054 // max_live_nodes_increase_per_iteration in between checks. If this
1055 // assumption does not hold, there is a risk that we exceed the max node
1056 // limit in between checks and trigger an assert during node creation.
1057 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 3;
1058
1059 uint loop_count = 0;
1060 // Pull from worklist and transform the node. If the node has changed,
1061 // update edge info and put uses on worklist.
1062 while (_worklist.size() > 0) {
1063 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1064 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1065 return;
1066 }
1067 Node* n = _worklist.pop();
1068 if (loop_count >= K * C->live_nodes()) {
1069 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1070 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1071 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1072 return;
1073 }
1074 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1075 if (n->outcnt() != 0) {
1076 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1077 // Do the transformation
1078 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1079 Node* nn = transform_old(n);
1080 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1081 // Ensure we did not increase the live node count with more than
1082 // max_live_nodes_increase_per_iteration during the call to transform_old
1083 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1084 assert(increase < max_live_nodes_increase_per_iteration,
1085 "excessive live node increase in single iteration of IGVN: %d "
1086 "(should be at most %d)",
1087 increase, max_live_nodes_increase_per_iteration);
1088 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1089 } else if (!n->is_top()) {
1090 remove_dead_node(n);
1091 }
1092 loop_count++;
1093 }
1094 NOT_PRODUCT(verify_PhaseIterGVN();)
1095 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1096 }
1097
1098 #ifdef ASSERT
1099 void PhaseIterGVN::verify_optimize() {
1100 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1101
1102 if (is_verify_Value() ||
1103 is_verify_Ideal() ||
1104 is_verify_Identity() ||
1105 is_verify_invariants()) {
1106 ResourceMark rm;
1107 Unique_Node_List worklist;
1108 // BFS all nodes, starting at root
1109 worklist.push(C->root());
1110 for (uint j = 0; j < worklist.size(); ++j) {
1111 Node* n = worklist.at(j);
1112 // If we get an assert here, check why the reported node was not processed again in IGVN.
1113 // We should either make sure that this node is properly added back to the IGVN worklist
1114 // in PhaseIterGVN::add_users_to_worklist to update it again or add an exception
1115 // in the verification methods below if that is not possible for some reason (like Load nodes).
1116 if (is_verify_Value()) {
1117 verify_Value_for(n);
1118 }
1119 if (is_verify_Ideal()) {
1120 verify_Ideal_for(n, false);
1121 verify_Ideal_for(n, true);
1122 }
1123 if (is_verify_Identity()) {
1124 verify_Identity_for(n);
1125 }
1126 if (is_verify_invariants()) {
1127 verify_node_invariants_for(n);
1128 }
1129
1130 // traverse all inputs and outputs
1131 for (uint i = 0; i < n->req(); i++) {
1132 if (n->in(i) != nullptr) {
1133 worklist.push(n->in(i));
1134 }
1135 }
1136 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1137 worklist.push(n->fast_out(i));
1138 }
1139 }
1140 }
1141
1142 verify_empty_worklist(nullptr);
1143 }
1144
1145 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1146 // Verify that the igvn worklist is empty. If no optimization happened, then
1147 // nothing needs to be on the worklist.
1148 if (_worklist.size() == 0) { return; }
1149
1150 stringStream ss; // Print as a block without tty lock.
1151 for (uint j = 0; j < _worklist.size(); j++) {
1152 Node* n = _worklist.at(j);
1153 ss.print("igvn.worklist[%d] ", j);
1154 n->dump("\n", false, &ss);
1155 }
1156 if (_worklist.size() != 0 && node != nullptr) {
1157 ss.print_cr("Previously optimized:");
1158 node->dump("\n", false, &ss);
1159 }
1160 tty->print_cr("%s", ss.as_string());
1161 assert(false, "igvn worklist must still be empty after verify");
1162 }
1163
1164 // Check that type(n) == n->Value(), asserts if we have a failure.
1165 // We have a list of exceptions, see detailed comments in code.
1166 // (1) Integer "widen" changes, but the range is the same.
1167 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1168 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1169 void PhaseIterGVN::verify_Value_for(const Node* n, bool strict) {
1170 // If we assert inside type(n), because the type is still a null, then maybe
1171 // the node never went through gvn.transform, which would be a bug.
1172 const Type* told = type(n);
1173 const Type* tnew = n->Value(this);
1174 if (told == tnew) {
1175 return;
1176 }
1177 // Exception (1)
1178 // Integer "widen" changes, but range is the same.
1179 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1180 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1181 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1182 if (t0->lo_as_long() == t1->lo_as_long() &&
1183 t0->hi_as_long() == t1->hi_as_long()) {
1184 return; // ignore integer widen
1185 }
1186 }
1187 // Exception (2)
1188 // LoadNode performs deep traversals. Load is not notified for changes far away.
1189 if (!strict && n->is_Load() && !told->singleton()) {
1190 // MemNode::can_see_stored_value looks up through many memory nodes,
1191 // which means we would need to notify modifications from far up in
1192 // the inputs all the way down to the LoadNode. We don't do that.
1193 return;
1194 }
1195 // Exception (3)
1196 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1197 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1198 // SubNode::Value
1199 // CmpPNode::sub
1200 // MemNode::detect_ptr_independence
1201 // MemNode::all_controls_dominate
1202 // We find all controls of a pointer load, and see if they dominate the control of
1203 // an allocation. If they all dominate, we know the allocation is after (independent)
1204 // of the pointer load, and we can say the pointers are different. For this we call
1205 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1206 // control sub of the allocation. The problems is that sometimes dominates answers
1207 // false conservatively, and later it can determine that it is indeed true. Loops with
1208 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1209 // so the traversal becomes more powerful. This is difficult to remedy, we would have
1210 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1211 // after loop-opts, so that should take care of many of these cases.
1212 return;
1213 }
1214
1215 stringStream ss; // Print as a block without tty lock.
1216 ss.cr();
1217 ss.print_cr("Missed Value optimization:");
1218 n->dump_bfs(1, nullptr, "", &ss);
1219 ss.print_cr("Current type:");
1220 told->dump_on(&ss);
1221 ss.cr();
1222 ss.print_cr("Optimized type:");
1223 tnew->dump_on(&ss);
1224 ss.cr();
1225 tty->print_cr("%s", ss.as_string());
1226
1227 switch (_phase) {
1228 case PhaseValuesType::iter_gvn:
1229 assert(false, "Missed Value optimization opportunity in PhaseIterGVN for %s",n->Name());
1230 break;
1231 case PhaseValuesType::ccp:
1232 assert(false, "PhaseCCP not at fixpoint: analysis result may be unsound for %s", n->Name());
1233 break;
1234 default:
1235 assert(false, "Unexpected phase");
1236 break;
1237 }
1238 }
1239
1240 // Check that all Ideal optimizations that could be done were done.
1241 // Asserts if it found missed optimization opportunities or encountered unexpected changes, and
1242 // returns normally otherwise (no missed optimization, or skipped verification).
1243 void PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1244 // First, we check a list of exceptions, where we skip verification,
1245 // because there are known cases where Ideal can optimize after IGVN.
1246 // Some may be expected and cannot be fixed, and others should be fixed.
1247 switch (n->Opcode()) {
1248 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1249 // (see "Range-Check scan limit"). So, it is possible that something
1250 // is optimized in that input subgraph, and the RangeCheck was not
1251 // added to the worklist because it would be too expensive to walk
1252 // down the graph for 1000 nodes and put all on the worklist.
1253 //
1254 // Found with:
1255 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1256 case Op_RangeCheck:
1257 return;
1258
1259 // IfNode::Ideal does:
1260 // Node* prev_dom = search_identical(dist, igvn);
1261 // which means we seach up the CFG, traversing at most up to a distance.
1262 // If anything happens rather far away from the If, we may not put the If
1263 // back on the worklist.
1264 //
1265 // Found with:
1266 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1267 case Op_If:
1268 return;
1269
1270 // IfNode::simple_subsuming
1271 // Looks for dominating test that subsumes the current test.
1272 // Notification could be difficult because of larger distance.
1273 //
1274 // Found with:
1275 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1276 // -XX:VerifyIterativeGVN=1110
1277 case Op_CountedLoopEnd:
1278 return;
1279
1280 // LongCountedLoopEndNode::Ideal
1281 // Probably same issue as above.
1282 //
1283 // Found with:
1284 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1285 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1286 case Op_LongCountedLoopEnd:
1287 return;
1288
1289 // RegionNode::Ideal does "Skip around the useless IF diamond".
1290 // 245 IfTrue === 244
1291 // 258 If === 245 257
1292 // 259 IfTrue === 258 [[ 263 ]]
1293 // 260 IfFalse === 258 [[ 263 ]]
1294 // 263 Region === 263 260 259 [[ 263 268 ]]
1295 // to
1296 // 245 IfTrue === 244
1297 // 263 Region === 263 245 _ [[ 263 268 ]]
1298 //
1299 // "Useless" means that there is no code in either branch of the If.
1300 // I found a case where this was not done yet during IGVN.
1301 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1302 //
1303 // Found with:
1304 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1305 case Op_Region:
1306 return;
1307
1308 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1309 // that smaller idx are first. Tracking it back, it led me to
1310 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1311 //
1312 // Example:
1313 // Before PhaseIdealLoop::remix_address_expressions
1314 // 154 AddI === _ 12 144
1315 // After PhaseIdealLoop::remix_address_expressions
1316 // 154 AddI === _ 144 12
1317 // After AddNode::Ideal
1318 // 154 AddI === _ 12 144
1319 //
1320 // I suspect that the node should be added to the IGVN worklist after
1321 // PhaseIdealLoop::remix_address_expressions
1322 //
1323 // This is the only case I looked at, there may be others. Found like this:
1324 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1325 //
1326 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1327 //
1328 // Note: currently all of these fail also for other reasons, for example
1329 // because of "commute" doing the reordering with the phi below. Once
1330 // that is resolved, we can come back to this issue here.
1331 //
1332 // case Op_AddD:
1333 // case Op_AddI:
1334 // case Op_AddL:
1335 // case Op_AddF:
1336 // case Op_MulI:
1337 // case Op_MulL:
1338 // case Op_MulF:
1339 // case Op_MulD:
1340 // if (n->in(1)->_idx > n->in(2)->_idx) {
1341 // // Expect "commute" to revert this case.
1342 // return false;
1343 // }
1344 // break; // keep verifying
1345
1346 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1347 // Check for tight loop increments: Loop-phi of Add of loop-phi
1348 // It wants to take the phi into in(1):
1349 // 471 Phi === 435 38 390
1350 // 390 AddF === _ 471 391
1351 //
1352 // Other Associative operators are also affected equally.
1353 //
1354 // Investigate why this does not happen earlier during IGVN.
1355 //
1356 // Found with:
1357 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1358 // -XX:VerifyIterativeGVN=1110
1359 case Op_AddD:
1360 //case Op_AddI: // Also affected for other reasons, see case further down.
1361 //case Op_AddL: // Also affected for other reasons, see case further down.
1362 case Op_AddF:
1363 case Op_MulI:
1364 case Op_MulL:
1365 case Op_MulF:
1366 case Op_MulD:
1367 case Op_MinF:
1368 case Op_MinD:
1369 case Op_MaxF:
1370 case Op_MaxD:
1371 // XorINode::Ideal
1372 // Found with:
1373 // compiler/intrinsics/chacha/TestChaCha20.java
1374 // -XX:VerifyIterativeGVN=1110
1375 case Op_XorI:
1376 case Op_XorL:
1377 // It seems we may have similar issues with the HF cases.
1378 // Found with aarch64:
1379 // compiler/vectorization/TestFloat16VectorOperations.java
1380 // -XX:VerifyIterativeGVN=1110
1381 case Op_AddHF:
1382 case Op_MulHF:
1383 case Op_MaxHF:
1384 case Op_MinHF:
1385 return;
1386
1387 // In MulNode::Ideal the edges can be swapped to help value numbering:
1388 //
1389 // // We are OK if right is a constant, or right is a load and
1390 // // left is a non-constant.
1391 // if( !(t2->singleton() ||
1392 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1393 // if( t1->singleton() || // Left input is a constant?
1394 // // Otherwise, sort inputs (commutativity) to help value numbering.
1395 // (in(1)->_idx > in(2)->_idx) ) {
1396 // swap_edges(1, 2);
1397 //
1398 // Why was this not done earlier during IGVN?
1399 //
1400 // Found with:
1401 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1402 // -XX:VerifyIterativeGVN=1110
1403 case Op_AndI:
1404 // Same for AndL.
1405 // Found with:
1406 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1407 // -XX:VerifyIterativeGVN=1110
1408 case Op_AndL:
1409 return;
1410
1411 // SubLNode::Ideal does transform like:
1412 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1413 //
1414 // In IGVN before verification:
1415 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1416 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1417 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1418 // 8016 ConL === 0 [[ 8383 ]] #long:0
1419 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1420 //
1421 // And then in verification:
1422 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1423 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1424 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1425 // 8016 ConL === 0 [[ 8383 ]] #long:0
1426 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1427 //
1428 // So the form changed from:
1429 // c1 - (y + [8423 ConvI2L])
1430 // to
1431 // c1 - (y + -2)
1432 // but the SubL was not added to the IGVN worklist. Investigate why.
1433 // There could be other issues too.
1434 //
1435 // There seems to be a related AddL IGVN optimization that triggers
1436 // the same SubL optimization, so investigate that too.
1437 //
1438 // Found with:
1439 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1440 case Op_SubL:
1441 return;
1442
1443 // SubINode::Ideal does
1444 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1445 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1446 //
1447 // Investigate why this does not yet happen during IGVN.
1448 //
1449 // Found with:
1450 // test/hotspot/jtreg/compiler/c2/IVTest.java
1451 // -XX:VerifyIterativeGVN=1110
1452 case Op_SubI:
1453 return;
1454
1455 // AddNode::IdealIL does transform like:
1456 // Convert x + (con - y) into "(x - y) + con"
1457 //
1458 // In IGVN before verification:
1459 // 8382 ConvI2L
1460 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1461 // 8383 SubL === _ 8381 8382
1462 // 8168 ConvI2L
1463 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1464 //
1465 // And then in verification:
1466 // 8424 AddL
1467 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1468 // 8383 SubL === _ 8016 8424
1469 // 8168 ConvI2L
1470 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1471 //
1472 // So the form changed from:
1473 // x + (ConvI2L(0) - [8382 ConvI2L])
1474 // to
1475 // x + (0 - [8424 AddL])
1476 // but the AddL was not added to the IGVN worklist. Investigate why.
1477 // There could be other issues, too. For example with "commute", see above.
1478 //
1479 // Found with:
1480 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1481 case Op_AddL:
1482 return;
1483
1484 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1485 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1486 // record_for_cleanup(cmp, phase);
1487 // This verification code in the Ideal code creates new nodes, and checks
1488 // if they fold in unexpected ways. This means some nodes are created and
1489 // added to the worklist, even if the SubTypeCheck is not optimized. This
1490 // goes agains the assumption of the verification here, which assumes that
1491 // if the node is not optimized, then no new nodes should be created, and
1492 // also no nodes should be added to the worklist.
1493 // I see two options:
1494 // 1) forbid what verify_helper does, because for each Ideal call it
1495 // uses memory and that is suboptimal. But it is not clear how that
1496 // verification can be done otherwise.
1497 // 2) Special case the verification here. Probably the new nodes that
1498 // were just created are dead, i.e. they are not connected down to
1499 // root. We could verify that, and remove those nodes from the graph
1500 // by setting all their inputs to nullptr. And of course we would
1501 // have to remove those nodes from the worklist.
1502 // Maybe there are other options too, I did not dig much deeper yet.
1503 //
1504 // Found with:
1505 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1506 case Op_SubTypeCheck:
1507 return;
1508
1509 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1510 // to other nodes: Conv, Add, Sub, Mul, And ...
1511 //
1512 // 107 ConI === 0 [[ ... ]] #int:2
1513 // 84 LoadRange === _ 7 83
1514 // 50 ConI === 0 [[ ... ]] #int:0
1515 // 549 LoopLimit === _ 50 84 107
1516 //
1517 // I stepped backward, to see how the node was generated, and I found that it was
1518 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1519 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1520 // that, and it seems we want to skip lowering it until after loop-opts, but never
1521 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1522 // could be other issues too.
1523 //
1524 // Fond with:
1525 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1526 case Op_LoopLimit:
1527 return;
1528
1529 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1530 // "This optimization tries to find two or more inputs of phi with the same constant
1531 // value. It then splits them into a separate Phi, and according Region."
1532 //
1533 // Example:
1534 // 130 DecodeN === _ 129
1535 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1536 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1537 //
1538 // turns into:
1539 //
1540 // 50 ConP === 0 [[ 99 91 18 ]] #null
1541 // 130 DecodeN === _ 129 [[ 18 ]]
1542 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1543 //
1544 // We would have to investigate why this optimization does not happen during IGVN.
1545 // There could also be other issues - I did not investigate further yet.
1546 //
1547 // Found with:
1548 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1549 case Op_Phi:
1550 return;
1551
1552 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1553 // For examle "The allocated object does not escape".
1554 //
1555 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1556 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1557 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1558 // worklist when this change happens?
1559 //
1560 // Found with:
1561 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1562 case Op_MemBarStoreStore:
1563 return;
1564
1565 // ConvI2LNode::Ideal converts
1566 // 648 AddI === _ 583 645 [[ 661 ]]
1567 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1568 // into
1569 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1570 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1571 // 773 AddL === _ 771 772 [[ ]]
1572 //
1573 // We have to investigate why this does not happen during IGVN in this case.
1574 // There could also be other issues - I did not investigate further yet.
1575 //
1576 // Found with:
1577 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1578 case Op_ConvI2L:
1579 return;
1580
1581 // AddNode::IdealIL can do this transform (and similar other ones):
1582 // Convert "a*b+a*c into a*(b+c)
1583 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1584 // during IGVN? There was a mutation for one of the MulI, and only
1585 // after that the pattern was as needed for the optimization. The MulI
1586 // was added to the IGVN worklist, but not the AddI. This probably
1587 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1588 //
1589 // Found with:
1590 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1591 // -XX:VerifyIterativeGVN=1110
1592 case Op_AddI:
1593 return;
1594
1595 // ArrayCopyNode::Ideal
1596 // calls ArrayCopyNode::prepare_array_copy
1597 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1598 // is not expected, and we create a range int:0..-1
1599 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1600 // note: the type is already empty!
1601 // calls PhaseIterGVN::transform
1602 // calls PhaseIterGVN::transform_old
1603 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1604 // calls Unique_Node_List::push -> pushes TOP to worklist
1605 //
1606 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1607 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1608 //
1609 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1610 // return nullptr, indicating nothing happened?
1611 // Is it smart to do transform in Compile::constrained_convI2L, and then
1612 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1613 // Should we just allow TOP to land on the worklist, as an exception?
1614 //
1615 // Found with:
1616 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1617 // -XX:VerifyIterativeGVN=1110
1618 case Op_ArrayCopy:
1619 return;
1620
1621 // CastLLNode::Ideal
1622 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1623 //
1624 // Could be a notification issue, where updates inputs of CastLL do not notify
1625 // down through SubL to CastLL.
1626 //
1627 // Found With:
1628 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1629 // -XX:VerifyIterativeGVN=1110
1630 case Op_CastLL:
1631 return;
1632
1633 // Similar case happens to CastII
1634 //
1635 // Found With:
1636 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1637 // -XX:VerifyIterativeGVN=1110
1638 case Op_CastII:
1639 return;
1640
1641 // MaxLNode::Ideal
1642 // calls AddNode::Ideal
1643 // calls commute -> decides to swap edges
1644 //
1645 // Another notification issue, because we check inputs of inputs?
1646 // MaxL -> Phi -> Loop
1647 // MaxL -> Phi -> MaxL
1648 //
1649 // Found with:
1650 // compiler/c2/irTests/TestIfMinMax.java
1651 // -XX:VerifyIterativeGVN=1110
1652 case Op_MaxL:
1653 case Op_MinL:
1654 return;
1655
1656 // OrINode::Ideal
1657 // calls AddNode::Ideal
1658 // calls commute -> left is Load, right not -> commute.
1659 //
1660 // Not sure why notification does not work here, seems like
1661 // the depth is only 1, so it should work. Needs investigation.
1662 //
1663 // Found with:
1664 // compiler/codegen/TestCharVect2.java#id0
1665 // -XX:VerifyIterativeGVN=1110
1666 case Op_OrI:
1667 case Op_OrL:
1668 return;
1669
1670 // Bool -> constant folded to 1.
1671 // Issue with notification?
1672 //
1673 // Found with:
1674 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1675 // -XX:VerifyIterativeGVN=1110
1676 case Op_Bool:
1677 return;
1678
1679 // LShiftLNode::Ideal
1680 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1681 // Probably a notification issue.
1682 //
1683 // Found with:
1684 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1685 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1686 case Op_LShiftL:
1687 return;
1688
1689 // LShiftINode::Ideal
1690 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1691 // Could be issue with notification of inputs of inputs
1692 //
1693 // Side-note: should cases like these not be shared between
1694 // LShiftI and LShiftL?
1695 //
1696 // Found with:
1697 // compiler/escapeAnalysis/Test6689060.java
1698 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1699 case Op_LShiftI:
1700 return;
1701
1702 // AddPNode::Ideal seems to do set_req without removing lock first.
1703 // Found with various vector tests tier1-tier3.
1704 case Op_AddP:
1705 return;
1706
1707 // StrIndexOfNode::Ideal
1708 // Found in tier1-3.
1709 case Op_StrIndexOf:
1710 case Op_StrIndexOfChar:
1711 return;
1712
1713 // StrEqualsNode::Identity
1714 //
1715 // Found (linux x64 only?) with:
1716 // serviceability/sa/ClhsdbThreadContext.java
1717 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1718 // Note: The -XX:LockingMode option is not available anymore.
1719 case Op_StrEquals:
1720 return;
1721
1722 // AryEqNode::Ideal
1723 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1724 //
1725 // Found with:
1726 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1727 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1728 case Op_AryEq:
1729 return;
1730
1731 // MergeMemNode::Ideal
1732 // Found in tier1-3. Did not investigate further yet.
1733 case Op_MergeMem:
1734 return;
1735
1736 // CMoveINode::Ideal
1737 // Found in tier1-3. Did not investigate further yet.
1738 case Op_CMoveI:
1739 return;
1740
1741 // CmpPNode::Ideal calls isa_const_java_mirror
1742 // and generates new constant nodes, even if no progress is made.
1743 // We can probably rewrite this so that only types are generated.
1744 // It seems that object types are not hashed, we could investigate
1745 // if that is an option as well.
1746 //
1747 // Found with:
1748 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1749 case Op_CmpP:
1750 return;
1751
1752 // MinINode::Ideal
1753 // Did not investigate, but there are some patterns that might
1754 // need more notification.
1755 case Op_MinI:
1756 case Op_MaxI: // preemptively removed it as well.
1757 return;
1758 }
1759
1760 if (n->is_Load()) {
1761 // LoadNode::Ideal uses tries to find an earlier memory state, and
1762 // checks can_see_stored_value for it.
1763 //
1764 // Investigate why this was not already done during IGVN.
1765 // A similar issue happens with Identity.
1766 //
1767 // There seem to be other cases where loads go up some steps, like
1768 // LoadNode::Ideal going up 10x steps to find dominating load.
1769 //
1770 // Found with:
1771 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1772 // -XX:VerifyIterativeGVN=1110
1773 return;
1774 }
1775
1776 if (n->is_Store()) {
1777 // StoreNode::Ideal can do this:
1778 // // Capture an unaliased, unconditional, simple store into an initializer.
1779 // // Or, if it is independent of the allocation, hoist it above the allocation.
1780 // That replaces the Store with a MergeMem.
1781 //
1782 // We have to investigate why this does not happen during IGVN in this case.
1783 // There could also be other issues - I did not investigate further yet.
1784 //
1785 // Found with:
1786 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1787 return;
1788 }
1789
1790 if (n->is_Vector()) {
1791 // VectorNode::Ideal swaps edges, but only for ops
1792 // that are deemed commutable. But swap_edges
1793 // requires the hash to be invariant when the edges
1794 // are swapped, which is not implemented for these
1795 // vector nodes. This seems not to create any trouble
1796 // usually, but we can also get graphs where in the
1797 // end the nodes are not all commuted, so there is
1798 // definitively an issue here.
1799 //
1800 // Probably we have two options: kill the hash, or
1801 // properly make the hash commutation friendly.
1802 //
1803 // Found with:
1804 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1805 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1806 return;
1807 }
1808
1809 if (n->is_Region()) {
1810 // LoopNode::Ideal calls RegionNode::Ideal.
1811 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1812 // But I got an issue because RegionNode::optimize_trichotomy
1813 // then modifies another node, and pushes nodes to the worklist
1814 // Not sure if this is ok, modifying another node like that.
1815 // Maybe it is, then we need to look into what to do with
1816 // the nodes that are now on the worklist, maybe just clear
1817 // them out again. But maybe modifying other nodes like that
1818 // is also bad design. In the end, we return nullptr for
1819 // the current CountedLoop. But the extra nodes on the worklist
1820 // trip the asserts later on.
1821 //
1822 // Found with:
1823 // compiler/eliminateAutobox/TestShortBoxing.java
1824 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1825 return;
1826 }
1827
1828 if (n->is_CallJava()) {
1829 // CallStaticJavaNode::Ideal
1830 // Led to a crash:
1831 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1832 //
1833 // Did not investigate yet, could be a bug.
1834 // Or maybe it does not expect to be called during verification.
1835 //
1836 // Found with:
1837 // test/jdk/jdk/incubator/vector/VectorRuns.java
1838 // -XX:VerifyIterativeGVN=1110
1839
1840 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1841 // and possibly their subclasses.
1842 // During late inlining it can call CallJavaNode::register_for_late_inline
1843 // That means we do more rounds of late inlining, but might fail.
1844 // Then we do IGVN again, and register the node again for late inlining.
1845 // This creates an endless cycle. Everytime we try late inlining, we
1846 // are also creating more nodes, especially SafePoint and MergeMem.
1847 // These nodes are immediately rejected when the inlining fails in the
1848 // do_late_inline_check, but they still grow the memory, until we hit
1849 // the MemLimit and crash.
1850 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1851 // called repeatedly, and eventually we terminate. I fear this is not
1852 // a great assumption to make. We should investigate more.
1853 //
1854 // Found with:
1855 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1856 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1857 return;
1858 }
1859
1860 // The number of nodes shoud not increase.
1861 uint old_unique = C->unique();
1862 // The hash of a node should not change, this would indicate different inputs
1863 uint old_hash = n->hash();
1864 // Remove 'n' from hash table in case it gets modified. We want to avoid
1865 // hitting the "Need to remove from hash before changing edges" assert if
1866 // a change occurs. Instead, we would like to proceed with the optimization,
1867 // return and finally hit the assert in PhaseIterGVN::verify_optimize to get
1868 // a more meaningful message
1869 _table.hash_delete(n);
1870 Node* i = n->Ideal(this, can_reshape);
1871 // If there was no new Idealization, we are probably happy.
1872 if (i == nullptr) {
1873 if (old_unique < C->unique()) {
1874 stringStream ss; // Print as a block without tty lock.
1875 ss.cr();
1876 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1877 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1878 n->dump_bfs(1, nullptr, "", &ss);
1879 tty->print_cr("%s", ss.as_string());
1880 assert(false, "Unexpected new unused nodes from applying Ideal optimization on %s", n->Name());
1881 }
1882
1883 if (old_hash != n->hash()) {
1884 stringStream ss; // Print as a block without tty lock.
1885 ss.cr();
1886 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1887 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1888 n->dump_bfs(1, nullptr, "", &ss);
1889 tty->print_cr("%s", ss.as_string());
1890 assert(false, "Unexpected hash change from applying Ideal optimization on %s", n->Name());
1891 }
1892
1893 verify_empty_worklist(n);
1894
1895 // Everything is good.
1896 hash_find_insert(n);
1897 return;
1898 }
1899
1900 // We just saw a new Idealization which was not done during IGVN.
1901 stringStream ss; // Print as a block without tty lock.
1902 ss.cr();
1903 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1904 if (i == n) {
1905 ss.print_cr("The node was reshaped by Ideal.");
1906 } else {
1907 ss.print_cr("The node was replaced by Ideal.");
1908 ss.print_cr("Old node:");
1909 n->dump_bfs(1, nullptr, "", &ss);
1910 }
1911 ss.print_cr("The result after Ideal:");
1912 i->dump_bfs(1, nullptr, "", &ss);
1913 tty->print_cr("%s", ss.as_string());
1914
1915 assert(false, "Missed Ideal optimization opportunity in PhaseIterGVN for %s", n->Name());
1916 }
1917
1918 // Check that all Identity optimizations that could be done were done.
1919 // Asserts if it found missed optimization opportunities, and
1920 // returns normally otherwise (no missed optimization, or skipped verification).
1921 void PhaseIterGVN::verify_Identity_for(Node* n) {
1922 // First, we check a list of exceptions, where we skip verification,
1923 // because there are known cases where Ideal can optimize after IGVN.
1924 // Some may be expected and cannot be fixed, and others should be fixed.
1925 switch (n->Opcode()) {
1926 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1927 // after loopopts:
1928 // // Transforming long counted loops requires a safepoint node. Do not
1929 // // eliminate a safepoint until loop opts are over.
1930 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1931 //
1932 // I think the check for major_progress does delay it until after loopopts
1933 // but it does not ensure that the node is on the IGVN worklist after
1934 // loopopts. I think we should try to instead check for
1935 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1936 //
1937 // Found with:
1938 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1939 case Op_SafePoint:
1940 return;
1941
1942 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1943 // does not record any other memory splits.
1944 //
1945 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1946 // never got called during IGVN for this node, investigate why.
1947 //
1948 // Found with:
1949 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1950 case Op_MergeMem:
1951 return;
1952
1953 // ConstraintCastNode::Identity finds casts that are the same, except that
1954 // the control is "higher up", i.e. dominates. The call goes via
1955 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1956 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1957 // away from the cast, but within 100 idom steps, the cast may not be
1958 // put on the IGVN worklist any more.
1959 //
1960 // Found with:
1961 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1962 case Op_CastPP:
1963 case Op_CastII:
1964 case Op_CastLL:
1965 return;
1966
1967 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1968 // checks dominator, which may be changed, but too far up for notification
1969 // to work.
1970 //
1971 // Found with:
1972 // compiler/c2/irTests/TestSkeletonPredicates.java
1973 // -XX:VerifyIterativeGVN=1110
1974 case Op_CheckCastPP:
1975 return;
1976
1977 // In SubNode::Identity, we do:
1978 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1979 // In the example, the AddI had an input replaced, the AddI is
1980 // added to the IGVN worklist, but the SubI is one link further
1981 // down and is not added. I checked add_users_of_use_to_worklist
1982 // where I would expect the SubI would be added, and I cannot
1983 // find the pattern, only this one:
1984 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1985 //
1986 // Fix this "notification" issue and check if there are any other
1987 // issues.
1988 //
1989 // Found with:
1990 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1991 case Op_SubI:
1992 case Op_SubL:
1993 return;
1994
1995 // PhiNode::Identity checks for patterns like:
1996 // r = (x != con) ? x : con;
1997 // that can be constant folded to "x".
1998 //
1999 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
2000 // I suspect there was some earlier change to one of the inputs, but
2001 // not all relevant outputs were put on the IGVN worklist.
2002 //
2003 // Found with:
2004 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
2005 // -XX:VerifyIterativeGVN=1110
2006 case Op_Phi:
2007 return;
2008
2009 // ConvI2LNode::Identity does
2010 // convert I2L(L2I(x)) => x
2011 //
2012 // Investigate why this did not already happen during IGVN.
2013 //
2014 // Found with:
2015 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
2016 // -XX:VerifyIterativeGVN=1110
2017 case Op_ConvI2L:
2018 return;
2019
2020 // AbsINode::Identity
2021 // Not investigated yet.
2022 case Op_AbsI:
2023 return;
2024 }
2025
2026 if (n->is_Load()) {
2027 // LoadNode::Identity tries to look for an earlier store value via
2028 // can_see_stored_value. I found an example where this led to
2029 // an Allocation, where we could assume the value was still zero.
2030 // So the LoadN can be replaced with a zerocon.
2031 //
2032 // Investigate why this was not already done during IGVN.
2033 // A similar issue happens with Ideal.
2034 //
2035 // Found with:
2036 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2037 return;
2038 }
2039
2040 if (n->is_Store()) {
2041 // StoreNode::Identity
2042 // Not investigated, but found missing optimization for StoreI.
2043 // Looks like a StoreI is replaced with an InitializeNode.
2044 //
2045 // Found with:
2046 // applications/ctw/modules/java_base_2.java
2047 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2048 return;
2049 }
2050
2051 if (n->is_Vector()) {
2052 // Found with tier1-3. Not investigated yet.
2053 // The observed issue was with AndVNode::Identity
2054 return;
2055 }
2056
2057 Node* i = n->Identity(this);
2058 // If we cannot find any other Identity, we are happy.
2059 if (i == n) {
2060 verify_empty_worklist(n);
2061 return;
2062 }
2063
2064 // The verification just found a new Identity that was not found during IGVN.
2065 stringStream ss; // Print as a block without tty lock.
2066 ss.cr();
2067 ss.print_cr("Missed Identity optimization:");
2068 ss.print_cr("Old node:");
2069 n->dump_bfs(1, nullptr, "", &ss);
2070 ss.print_cr("New node:");
2071 i->dump_bfs(1, nullptr, "", &ss);
2072 tty->print_cr("%s", ss.as_string());
2073
2074 assert(false, "Missed Identity optimization opportunity in PhaseIterGVN for %s", n->Name());
2075 }
2076
2077 // Some other verifications that are not specific to a particular transformation.
2078 void PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2079 if (n->is_AddP()) {
2080 if (!n->as_AddP()->address_input_has_same_base()) {
2081 stringStream ss; // Print as a block without tty lock.
2082 ss.cr();
2083 ss.print_cr("Base pointers must match for AddP chain:");
2084 n->dump_bfs(2, nullptr, "", &ss);
2085 tty->print_cr("%s", ss.as_string());
2086
2087 assert(false, "Broken node invariant for %s", n->Name());
2088 }
2089 }
2090 }
2091 #endif
2092
2093 /**
2094 * Register a new node with the optimizer. Update the types array, the def-use
2095 * info. Put on worklist.
2096 */
2097 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2098 set_type_bottom(n);
2099 _worklist.push(n);
2100 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2101 return n;
2102 }
2103
2104 //------------------------------transform--------------------------------------
2105 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2106 Node *PhaseIterGVN::transform( Node *n ) {
2107 if (_delay_transform) {2108 // Register the node but don't optimize for now2109 register_new_node_with_optimizer(n);2110 return n;2111 }2112
2113 // If brand new node, make space in type array, and give it a type.
2114 ensure_type_or_null(n);
2115 if (type_or_null(n) == nullptr) {
2116 set_type_bottom(n);
2117 }
2118
2119 return transform_old(n);
2120 }
2121
2122 Node *PhaseIterGVN::transform_old(Node* n) {
2123 NOT_PRODUCT(set_transforms());
2124 // Remove 'n' from hash table in case it gets modified
2125 _table.hash_delete(n);
2126 #ifdef ASSERT
2127 if (is_verify_def_use()) {
2128 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2129 }
2130 #endif
2131
2132 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2133 if (n->is_Cmp()) {
2134 add_users_to_worklist(n);
2135 }
2136
2137 // Apply the Ideal call in a loop until it no longer applies
2138 Node* k = n;
2139 DEBUG_ONLY(dead_loop_check(k);)
2140 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2141 C->remove_modified_node(k);
2142 #ifndef PRODUCT
2143 uint hash_before = is_verify_Ideal_return() ? k->hash() : 0;
2144 #endif
2145 Node* i = apply_ideal(k, /*can_reshape=*/true);
2146 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2147 #ifndef PRODUCT
2148 if (is_verify_Ideal_return()) {
2149 assert(k->outcnt() == 0 || i != nullptr || hash_before == k->hash(), "hash changed after Ideal returned nullptr for %s", k->Name());
2150 }
2151 verify_step(k);
2152 #endif
2153
2154 DEBUG_ONLY(uint loop_count = 1;)
2155 while (i != nullptr) {
2156 #ifdef ASSERT
2157 if (loop_count >= K + C->live_nodes()) {
2158 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2159 }
2160 #endif
2161 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2162 // Made a change; put users of original Node on worklist
2163 add_users_to_worklist(k);
2164 // Replacing root of transform tree?
2165 if (k != i) {
2166 // Make users of old Node now use new.
2167 subsume_node(k, i);
2168 k = i;
2169 }
2170 DEBUG_ONLY(dead_loop_check(k);)
2171 // Try idealizing again
2172 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2173 C->remove_modified_node(k);
2174 #ifndef PRODUCT
2175 uint hash_before = is_verify_Ideal_return() ? k->hash() : 0;
2176 #endif
2177 i = apply_ideal(k, /*can_reshape=*/true);
2178 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2179 #ifndef PRODUCT
2180 if (is_verify_Ideal_return()) {
2181 assert(k->outcnt() == 0 || i != nullptr || hash_before == k->hash(), "hash changed after Ideal returned nullptr for %s", k->Name());
2182 }
2183 verify_step(k);
2184 #endif
2185 DEBUG_ONLY(loop_count++;)
2186 }
2187
2188 // If brand new node, make space in type array.
2189 ensure_type_or_null(k);
2190
2191 // See what kind of values 'k' takes on at runtime
2192 const Type* t = k->Value(this);
2193 assert(t != nullptr, "value sanity");
2194
2195 // Since I just called 'Value' to compute the set of run-time values
2196 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2197 // cache Value. Later requests for the local phase->type of this Node can
2198 // use the cached Value instead of suffering with 'bottom_type'.
2199 if (type_or_null(k) != t) {
2200 #ifndef PRODUCT
2201 inc_new_values();
2202 set_progress();
2203 #endif
2204 set_type(k, t);
2205 // If k is a TypeNode, capture any more-precise type permanently into Node
2206 k->raise_bottom_type(t);
2207 // Move users of node to worklist
2208 add_users_to_worklist(k);
2209 }
2210 // If 'k' computes a constant, replace it with a constant
2211 if (t->singleton() && !k->is_Con()) {
2212 NOT_PRODUCT(set_progress();)
2213 Node* con = makecon(t); // Make a constant
2214 add_users_to_worklist(k);
2215 subsume_node(k, con); // Everybody using k now uses con
2216 return con;
2217 }
2218
2219 // Now check for Identities
2220 i = k->Identity(this); // Look for a nearby replacement
2221 if (i != k) { // Found? Return replacement!
2222 NOT_PRODUCT(set_progress();)
2223 add_users_to_worklist(k);
2224 subsume_node(k, i); // Everybody using k now uses i
2225 return i;
2226 }
2227
2228 // Global Value Numbering
2229 i = hash_find_insert(k); // Check for pre-existing node
2230 if (i && (i != k)) {
2231 // Return the pre-existing node if it isn't dead
2232 NOT_PRODUCT(set_progress();)
2233 add_users_to_worklist(k);
2234 subsume_node(k, i); // Everybody using k now uses i
2235 return i;
2236 }
2237
2238 // Return Idealized original
2239 return k;
2240 }
2241
2242 //---------------------------------saturate------------------------------------
2243 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2244 const Type* limit_type) const {
2245 return new_type->narrow(old_type);
2246 }
2247
2248 //------------------------------remove_globally_dead_node----------------------
2249 // Kill a globally dead Node. All uses are also globally dead and are
2250 // aggressively trimmed.
2251 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2252 enum DeleteProgress {
2253 PROCESS_INPUTS,
2254 PROCESS_OUTPUTS
2255 };
2256 ResourceMark rm;
2257 Node_Stack stack(32);
2258 stack.push(dead, PROCESS_INPUTS);
2259
2260 while (stack.is_nonempty()) {
2261 dead = stack.node();
2262 if (dead->Opcode() == Op_SafePoint) {
2263 dead->as_SafePoint()->disconnect_from_root(this);
2264 }
2265 uint progress_state = stack.index();
2266 assert(dead != C->root(), "killing root, eh?");
2267 assert(!dead->is_top(), "add check for top when pushing");
2268 NOT_PRODUCT( set_progress(); )
2269 if (progress_state == PROCESS_INPUTS) {
2270 // After following inputs, continue to outputs
2271 stack.set_index(PROCESS_OUTPUTS);
2272 if (!dead->is_Con()) { // Don't kill cons but uses
2273 bool recurse = false;
2274 // Remove from hash table
2275 _table.hash_delete( dead );
2276 // Smash all inputs to 'dead', isolating him completely
2277 for (uint i = 0; i < dead->req(); i++) {
2278 Node *in = dead->in(i);
2279 if (in != nullptr && in != C->top()) { // Points to something?
2280 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2281 assert((nrep > 0), "sanity");
2282 if (in->outcnt() == 0) { // Made input go dead?
2283 stack.push(in, PROCESS_INPUTS); // Recursively remove
2284 recurse = true;
2285 } else if (in->outcnt() == 1 &&
2286 in->has_special_unique_user()) {
2287 _worklist.push(in->unique_out());
2288 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2289 if (in->Opcode() == Op_Region) {
2290 _worklist.push(in);
2291 } else if (in->is_Store()) {
2292 DUIterator_Fast imax, i = in->fast_outs(imax);
2293 _worklist.push(in->fast_out(i));
2294 i++;
2295 if (in->outcnt() == 2) {
2296 _worklist.push(in->fast_out(i));
2297 i++;
2298 }
2299 assert(!(i < imax), "sanity");
2300 }
2301 } else if (dead->is_data_proj_of_pure_function(in)) {
2302 _worklist.push(in);
2303 } else {
2304 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2305 }
2306 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2307 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2308 // A Load that directly follows an InitializeNode is
2309 // going away. The Stores that follow are candidates
2310 // again to be captured by the InitializeNode.
2311 add_users_to_worklist_if(_worklist, in, [](Node* n) { return n->is_Store(); });
2312 }
2313 } // if (in != nullptr && in != C->top())
2314 } // for (uint i = 0; i < dead->req(); i++)
2315 if (recurse) {
2316 continue;
2317 }
2318 } // if (!dead->is_Con())
2319 } // if (progress_state == PROCESS_INPUTS)
2320
2321 // Aggressively kill globally dead uses
2322 // (Rather than pushing all the outs at once, we push one at a time,
2323 // plus the parent to resume later, because of the indefinite number
2324 // of edge deletions per loop trip.)
2325 if (dead->outcnt() > 0) {
2326 // Recursively remove output edges
2327 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2328 } else {
2329 // Finished disconnecting all input and output edges.
2330 stack.pop();
2331 // Remove dead node from iterative worklist
2332 _worklist.remove(dead);
2333 C->remove_useless_node(dead);
2334 }
2335 } // while (stack.is_nonempty())
2336 }
2337
2338 //------------------------------subsume_node-----------------------------------
2339 // Remove users from node 'old' and add them to node 'nn'.
2340 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2341 if (old->Opcode() == Op_SafePoint) {
2342 old->as_SafePoint()->disconnect_from_root(this);
2343 }
2344 assert( old != hash_find(old), "should already been removed" );
2345 assert( old != C->top(), "cannot subsume top node");
2346 // Copy debug or profile information to the new version:
2347 C->copy_node_notes_to(nn, old);
2348 // Move users of node 'old' to node 'nn'
2349 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2350 Node* use = old->last_out(i); // for each use...
2351 // use might need re-hashing (but it won't if it's a new node)
2352 rehash_node_delayed(use);
2353 // Update use-def info as well
2354 // We remove all occurrences of old within use->in,
2355 // so as to avoid rehashing any node more than once.
2356 // The hash table probe swamps any outer loop overhead.
2357 uint num_edges = 0;
2358 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2359 if (use->in(j) == old) {
2360 use->set_req(j, nn);
2361 ++num_edges;
2362 }
2363 }
2364 i -= num_edges; // we deleted 1 or more copies of this edge
2365 }
2366
2367 // Search for instance field data PhiNodes in the same region pointing to the old
2368 // memory PhiNode and update their instance memory ids to point to the new node.
2369 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2370 Node* region = old->in(0);
2371 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2372 PhiNode* phi = region->fast_out(i)->isa_Phi();
2373 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2374 phi->set_inst_mem_id((int)nn->_idx);
2375 }
2376 }
2377 }
2378
2379 // Smash all inputs to 'old', isolating him completely
2380 Node *temp = new Node(1);
2381 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2382 remove_dead_node( old );
2383 temp->del_req(0); // Yank bogus edge
2384 if (nn != nullptr && nn->outcnt() == 0) {
2385 _worklist.push(nn);
2386 }
2387 #ifndef PRODUCT
2388 if (is_verify_def_use()) {
2389 for ( int i = 0; i < _verify_window_size; i++ ) {
2390 if ( _verify_window[i] == old )
2391 _verify_window[i] = nn;
2392 }
2393 }
2394 #endif
2395 temp->destruct(this); // reuse the _idx of this little guy
2396 }
2397
2398 //------------------------------add_users_to_worklist--------------------------
2399 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2400 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2401 worklist.push(n->fast_out(i)); // Push on worklist
2402 }
2403 }
2404
2405 // Return counted loop Phi if as a counted loop exit condition, cmp
2406 // compares the induction variable with n
2407 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2408 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2409 Node* bol = cmp->fast_out(i);
2410 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2411 Node* iff = bol->fast_out(i2);
2412 if (iff->is_BaseCountedLoopEnd()) {
2413 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2414 if (cle->limit() == n) {
2415 PhiNode* phi = cle->phi();
2416 if (phi != nullptr) {
2417 return phi;
2418 }
2419 }
2420 }
2421 }
2422 }
2423 return nullptr;
2424 }
2425
2426 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2427 add_users_to_worklist0(n, _worklist);
2428
2429 Unique_Node_List& worklist = _worklist;
2430 // Move users of node to worklist
2431 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2432 Node* use = n->fast_out(i); // Get use
2433 add_users_of_use_to_worklist(n, use, worklist);
2434 }
2435 }
2436
2437 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2438 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2439 use->is_Store() ) // Enable store/load same address
2440 add_users_to_worklist0(use, worklist);
2441
2442 // If we changed the receiver type to a call, we need to revisit
2443 // the Catch following the call. It's looking for a non-null
2444 // receiver to know when to enable the regular fall-through path
2445 // in addition to the NullPtrException path.
2446 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2447 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2448 if (p != nullptr) {
2449 add_users_to_worklist0(p, worklist);
2450 }
2451 }
2452
2453 uint use_op = use->Opcode();
2454 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2455 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2456 if (use->outcnt() > 0) {
2457 Node* bol = use->raw_out(0);
2458 if (bol->outcnt() > 0) {
2459 Node* iff = bol->raw_out(0);
2460 if (iff->outcnt() == 2) {
2461 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2462 // phi merging either 0 or 1 onto the worklist
2463 Node* ifproj0 = iff->raw_out(0);
2464 Node* ifproj1 = iff->raw_out(1);
2465 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2466 Node* region0 = ifproj0->raw_out(0);
2467 Node* region1 = ifproj1->raw_out(0);
2468 if( region0 == region1 )
2469 add_users_to_worklist0(region0, worklist);
2470 }
2471 }
2472 }
2473 }
2474 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2475 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2476 if (phi != nullptr) {
2477 // Input to the cmp of a loop exit check has changed, thus
2478 // the loop limit may have changed, which can then change the
2479 // range values of the trip-count Phi.
2480 worklist.push(phi);
2481 }
2482 }
2483 if (use_op == Op_CmpI) {
2484 Node* cmp = use;
2485 Node* in1 = cmp->in(1);
2486 Node* in2 = cmp->in(2);
2487 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2488 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2489 //
2490 // in1 in2 in1 in2
2491 // | | | |
2492 // +--- | --+ | |
2493 // | | | | |
2494 // CmpINode | CmpINode
2495 // | | |
2496 // BoolNode | BoolNode
2497 // | | OR |
2498 // IfNode | IfNode
2499 // | | |
2500 // IfProj | IfProj X
2501 // | | | |
2502 // CastIINode CastIINode
2503 //
2504 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2505 if (in1 == n) {
2506 // in1 modified -> could turn into X -> do traversal based on right pattern.
2507 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2508 Node* bol = cmp->fast_out(i2); // For each Bool
2509 if (bol->is_Bool()) {
2510 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2511 Node* iff = bol->fast_out(i3); // For each If
2512 if (iff->is_If()) {
2513 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2514 Node* if_proj = iff->fast_out(i4); // For each IfProj
2515 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2516 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2517 Node* castii = if_proj->fast_out(i5); // For each CastII
2518 if (castii->is_CastII() &&
2519 castii->as_CastII()->carry_dependency()) {
2520 worklist.push(castii);
2521 }
2522 }
2523 }
2524 }
2525 }
2526 }
2527 }
2528 } else {
2529 // Only in2 modified -> can assume X == in2 (left pattern).
2530 assert(n == in2, "only in2 modified");
2531 // Find all CastII with input in1.
2532 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2533 Node* castii = in1->fast_out(j);
2534 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2535 // Find If.
2536 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2537 Node* ifnode = castii->in(0)->in(0);
2538 // Check that if connects to the cmp
2539 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2540 worklist.push(castii);
2541 }
2542 }
2543 }
2544 }
2545 }
2546 }
2547 }
2548 }
2549
2550 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2551 // Patterns:
2552 // ConstraintCast+ -> Sub
2553 // ConstraintCast+ -> Phi
2554 // ConstraintCast+ -> Xor
2555 if (use->is_ConstraintCast()) {
2556 auto push_the_uses_to_worklist = [&](Node* n){
2557 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2558 worklist.push(n);
2559 }
2560 };
2561 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2562 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2563 }
2564 // If changed LShift inputs, check RShift/URShift users for
2565 // "(X << C) >> C" sign-ext and "(X << C) >>> C" zero-ext optimizations.
2566 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2567 add_users_to_worklist_if(worklist, use, [](Node* u) {
2568 return u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2569 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2570 });
2571 }
2572 // If changed LShift inputs, check And users for shift and mask (And) operation
2573 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2574 add_users_to_worklist_if(worklist, use, [](Node* u) {
2575 return u->Opcode() == Op_AndI || u->Opcode() == Op_AndL;
2576 });
2577 }
2578 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2579 if (use_op == Op_AddI || use_op == Op_SubI) {
2580 add_users_to_worklist_if(worklist, use, [](Node* u) {
2581 return u->Opcode() == Op_CmpU;
2582 });
2583 }
2584 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2585 if (use_op == Op_AndI || use_op == Op_AndL) {
2586 add_users_to_worklist_if(worklist, use, [](Node* u) {
2587 return u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2588 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2589 });
2590 }
2591 // Check for redundant conversion patterns:
2592 // ConvD2L->ConvL2D->ConvD2L
2593 // ConvF2I->ConvI2F->ConvF2I
2594 // ConvF2L->ConvL2F->ConvF2L
2595 // ConvI2F->ConvF2I->ConvI2F
2596 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2597 // are the only ones that are known to trigger missed optimizations otherwise
2598 if (use_op == Op_ConvL2D ||
2599 use_op == Op_ConvI2F ||
2600 use_op == Op_ConvL2F ||
2601 use_op == Op_ConvF2I) {
2602 add_users_to_worklist_if(worklist, use, [=](Node* u) {
2603 return (use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2604 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2605 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2606 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F);
2607 });
2608 }
2609 // ConvD2F::Ideal matches ConvD2F(SqrtD(ConvF2D(x))) => SqrtF(x).
2610 // Notify ConvD2F users of SqrtD when any input of the SqrtD changes.
2611 if (use_op == Op_SqrtD) {
2612 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_ConvD2F; });
2613 }
2614 // ConvF2HF::Ideal matches ConvF2HF(binopF(ConvHF2F(...))) => FP16BinOp(...).
2615 // Notify ConvF2HF users of float binary ops when any input changes.
2616 if (Float16NodeFactory::is_float32_binary_oper(use_op)) {
2617 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_ConvF2HF; });
2618 }
2619 // If changed AddP inputs:
2620 // - check Stores for loop invariant, and
2621 // - if the changed input is the offset, check constant-offset AddP users for
2622 // address expression flattening.
2623 if (use_op == Op_AddP) {
2624 bool offset_changed = n == use->in(AddPNode::Offset);
2625 add_users_to_worklist_if(worklist, use, [=](Node* u) {
2626 return u->is_Mem() ||
2627 (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con());
2628 });
2629 }
2630 // Check for "abs(0-x)" into "abs(x)" conversion
2631 if (use->is_Sub()) {
2632 add_users_to_worklist_if(worklist, use, [](Node* u) {
2633 return u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2634 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI;
2635 });
2636 }
2637 // Check for Max/Min(A, Max/Min(B, C)) where A == B or A == C
2638 if (use->is_MinMax()) {
2639 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->is_MinMax(); });
2640 }
2641 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2642 add_users_to_worklist0(proj, worklist);
2643 };
2644 // If changed initialization activity, check dependent Stores
2645 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2646 InitializeNode* init = use->as_Allocate()->initialization();
2647 if (init != nullptr) {
2648 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2649 }
2650 }
2651 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2652 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2653 // to guarantee the change is not missed.
2654 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2655 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2656 if (p != nullptr) {
2657 add_users_to_worklist0(p, worklist);
2658 }
2659 }
2660
2661 if (use_op == Op_Initialize) {
2662 InitializeNode* init = use->as_Initialize();
2663 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2664 }
2665 // Loading the java mirror from a Klass requires two loads and the type
2666 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2667 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2668 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2669 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2670
2671 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2672 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2673 Node* u = use->fast_out(i2);
2674 const Type* ut = u->bottom_type();
2675 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2676 if (has_load_barrier_nodes) {
2677 // Search for load barriers behind the load
2678 add_users_to_worklist_if(worklist, u, [&](Node* b) {
2679 return bs->is_gc_barrier_node(b);
2680 });
2681 }
2682 worklist.push(u);
2683 }
2684 }
2685 }
2686 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2687 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2688 if (use->outcnt() == 1) {
2689 Node* cmp = use->unique_out();
2690 worklist.push(cmp);
2691 }
2692 }
2693 // VectorMaskToLongNode::Ideal_MaskAll looks through VectorStoreMask
2694 // to fold constant masks.
2695 if (use_op == Op_VectorStoreMask) {
2696 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_VectorMaskToLong; });
2697 }
2698
2699 // From CastX2PNode::Ideal
2700 // CastX2P(AddX(x, y))
2701 // CastX2P(SubX(x, y))
2702 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2703 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_CastX2P; });
2704 }
2705
2706 /* AndNode has a special handling when one of the operands is a LShiftNode:
2707 * (LHS << s) & RHS
2708 * if RHS fits in less than s bits, the value of this expression is 0.
2709 * The difficulty is that there might be a conversion node (ConvI2L) between
2710 * the LShiftINode and the AndLNode, like so:
2711 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2712 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2713 * (see `AndIL_min_trailing_zeros`).
2714 *
2715 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2716 * is not enough: there might be no update happening there. We need to
2717 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2718 *
2719 * Moreover we can have ConstraintCasts in between. It may look like
2720 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2721 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2722 * So we must deal with that as well.
2723 */
2724 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2725 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2726 auto push_and_to_worklist = [&worklist](Node* n){
2727 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2728 worklist.push(n);
2729 }
2730 };
2731 use->visit_uses(push_and_to_worklist, is_boundary);
2732 }
2733
2734 // If changed Sub inputs, check Add for identity.
2735 // e.g., (x - y) + y -> x; x + (y - x) -> y.
2736 if (use_op == Op_SubI || use_op == Op_SubL) {
2737 const int add_op = (use_op == Op_SubI) ? Op_AddI : Op_AddL;
2738 add_users_to_worklist_if(worklist, use, [=](Node* u) { return u->Opcode() == add_op; });
2739 }
2740 }
2741
2742 /**
2743 * Remove the speculative part of all types that we know of
2744 */
2745 void PhaseIterGVN::remove_speculative_types() {
2746 assert(UseTypeSpeculation, "speculation is off");
2747 for (uint i = 0; i < _types.Size(); i++) {
2748 const Type* t = _types.fast_lookup(i);
2749 if (t != nullptr) {
2750 _types.map(i, t->remove_speculative());
2751 }
2752 }
2753 _table.check_no_speculative_types();
2754 }
2755
2756 //=============================================================================
2757 #ifndef PRODUCT
2758 uint PhaseCCP::_total_invokes = 0;
2759 uint PhaseCCP::_total_constants = 0;
2760 #endif
2761 //------------------------------PhaseCCP---------------------------------------
2762 // Conditional Constant Propagation, ala Wegman & Zadeck
2763 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2764 NOT_PRODUCT( clear_constants(); )
2765 assert( _worklist.size() == 0, "" );
2766 _phase = PhaseValuesType::ccp;
2767 analyze();
2768 }
2769
2770 #ifndef PRODUCT
2771 //------------------------------~PhaseCCP--------------------------------------
2772 PhaseCCP::~PhaseCCP() {
2773 inc_invokes();
2774 _total_constants += count_constants();
2775 }
2776 #endif
2777
2778
2779 #ifdef ASSERT
2780 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2781 if (tnew->meet(told) != tnew->remove_speculative()) {
2782 n->dump(1);
2783 tty->print("told = "); told->dump(); tty->cr();
2784 tty->print("tnew = "); tnew->dump(); tty->cr();
2785 fatal("Not monotonic");
2786 }
2787 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2788 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2789 }
2790 #endif //ASSERT
2791
2792 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2793 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2794 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2795 // or nodes even further down back to the worklist because their type could change as a result of the current type
2796 // change).
2797 void PhaseCCP::analyze() {
2798 // Initialize all types to TOP, optimistic analysis
2799 for (uint i = 0; i < C->unique(); i++) {
2800 _types.map(i, Type::TOP);
2801 }
2802
2803 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2804 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2805 // Compile is over. The local arena gets de-allocated at the end of its scope.
2806 ResourceArea local_arena(mtCompiler);
2807 Unique_Node_List worklist(&local_arena);
2808 Unique_Node_List worklist_revisit(&local_arena);
2809 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2810
2811 // Push root onto worklist
2812 worklist.push(C->root());
2813
2814 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2815 _root_and_safepoints.push(C->root());
2816
2817 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2818
2819 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2820 while (worklist.size() != 0) {
2821 Node* n = fetch_next_node(worklist);
2822 DEBUG_ONLY(worklist_verify.push(n);)
2823 if (needs_revisit(n)) {
2824 worklist_revisit.push(n);
2825 }
2826 if (n->is_SafePoint()) {
2827 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2828 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2829 _root_and_safepoints.push(n);
2830 }
2831 analyze_step(worklist, n);
2832 }
2833
2834 // More rounds to catch updates far in the graph.
2835 // Revisit nodes that might be able to refine their types at the end of the round.
2836 // If so, process these nodes. If there is remaining work, start another round.
2837 do {
2838 while (worklist.size() != 0) {
2839 Node* n = fetch_next_node(worklist);
2840 analyze_step(worklist, n);
2841 }
2842 for (uint t = 0; t < worklist_revisit.size(); t++) {
2843 Node* n = worklist_revisit.at(t);
2844 analyze_step(worklist, n);
2845 }
2846 } while (worklist.size() != 0);
2847
2848 DEBUG_ONLY(verify_analyze(worklist_verify);)
2849 }
2850
2851 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2852 const Type* new_type = n->Value(this);
2853 if (new_type != type(n)) {
2854 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2855 dump_type_and_node(n, new_type);
2856 set_type(n, new_type);
2857 push_child_nodes_to_worklist(worklist, n);
2858 }
2859 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2860 // Keep track of Type nodes to kill CFG paths that use Type
2861 // nodes that become dead.
2862 _maybe_top_type_nodes.push(n);
2863 }
2864 }
2865
2866 // Some nodes can refine their types due to type change somewhere deep
2867 // in the graph. We will need to revisit them before claiming convergence.
2868 // Add nodes here if particular *Node::Value is doing deep graph traversals
2869 // not handled by PhaseCCP::push_more_uses().
2870 bool PhaseCCP::needs_revisit(Node* n) const {
2871 // LoadNode performs deep traversals. Load is not notified for changes far away.
2872 if (n->is_Load()) {
2873 return true;
2874 }
2875 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2876 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2877 return true;
2878 }
2879 return false;
2880 }
2881
2882 #ifdef ASSERT
2883 // For every node n on verify list, check if type(n) == n->Value()
2884 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2885 // Therefore, we are verifying Value convergence strictly.
2886 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2887 while (worklist_verify.size()) {
2888 Node* n = worklist_verify.pop();
2889
2890 // An assert in verify_Value_for means that PhaseCCP is not at fixpoint
2891 // and that the analysis result may be unsound.
2892 // If this happens, check why the reported nodes were not processed again in CCP.
2893 // We should either make sure that these nodes are properly added back to the CCP worklist
2894 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2895 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2896 // them at the end of the round.
2897 verify_Value_for(n, true);
2898 }
2899 }
2900 #endif
2901
2902 // Fetch next node from worklist to be examined in this iteration.
2903 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2904 if (StressCCP) {
2905 return worklist.remove(C->random() % worklist.size());
2906 } else {
2907 return worklist.pop();
2908 }
2909 }
2910
2911 #ifndef PRODUCT
2912 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2913 if (TracePhaseCCP) {
2914 t->dump();
2915 do {
2916 tty->print("\t");
2917 } while (tty->position() < 16);
2918 n->dump();
2919 }
2920 }
2921 #endif
2922
2923 bool PhaseCCP::not_bottom_type(Node* n) const {
2924 return n->bottom_type() != type(n);
2925 }
2926
2927 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2928 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2929 // of the new type of 'n'. Push these nodes to the worklist.
2930 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2931 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2932 Node* use = n->fast_out(i);
2933 push_if_not_bottom_type(worklist, use);
2934 push_more_uses(worklist, n, use);
2935 }
2936 }
2937
2938 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2939 if (not_bottom_type(n)) {
2940 worklist.push(n);
2941 }
2942 }
2943
2944 // For some nodes, we need to propagate the type change to grandchildren or even further down.
2945 // Add them back to the worklist.
2946 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
2947 push_phis(worklist, use);
2948 push_catch(worklist, use);
2949 push_cmpu(worklist, use);
2950 push_counted_loop_phi(worklist, parent, use);
2951 push_loadp(worklist, use);
2952 push_and(worklist, parent, use);
2953 push_cast_ii(worklist, parent, use);
2954 push_opaque_zero_trip_guard(worklist, use);
2955 push_bool_with_cmpu_and_mask(worklist, use);
2956 }
2957
2958
2959 // We must recheck Phis too if use is a Region.
2960 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
2961 if (use->is_Region()) {
2962 add_users_to_worklist_if(worklist, use, [&](Node* u) {
2963 return not_bottom_type(u);
2964 });
2965 }
2966 }
2967
2968 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
2969 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
2970 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
2971 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
2972 if (use->is_Call()) {
2973 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2974 Node* proj = use->fast_out(i);
2975 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
2976 Node* catch_node = proj->find_out_with(Op_Catch);
2977 if (catch_node != nullptr) {
2978 worklist.push(catch_node);
2979 }
2980 }
2981 }
2982 }
2983 }
2984
2985 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
2986 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
2987 // and get wrong types otherwise.
2988 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
2989 uint use_op = use->Opcode();
2990 if (use_op == Op_AddI || use_op == Op_SubI) {
2991 // Got a CmpU or CmpU3 which might need the new type information from node n.
2992 add_users_to_worklist_if(worklist, use, [&](Node* u) {
2993 uint op = u->Opcode();
2994 return (op == Op_CmpU || op == Op_CmpU3) && not_bottom_type(u);
2995 });
2996 }
2997 }
2998
2999 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3000 // (1b): "(m & x) <u (m + 1))".
3001 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3002 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3003 //
3004 // m x m 1 (%%)
3005 // \ / \ /
3006 // AndI AddI
3007 // \ /
3008 // CmpU
3009 // |
3010 // Bool
3011 //
3012 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3013 uint use_op = use->Opcode();
3014 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3015 // Not "m & x" or "m + 1"
3016 return;
3017 }
3018 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3019 Node* cmpu = use->fast_out(i);
3020 if (cmpu->Opcode() == Op_CmpU) {
3021 push_bool_matching_case1b(worklist, cmpu);
3022 }
3023 }
3024 }
3025
3026 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3027 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3028 assert(cmpu->Opcode() == Op_CmpU, "must be");
3029 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3030 Node* bol = cmpu->fast_out(i);
3031 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3032 // Not a Bool with "<u"
3033 continue;
3034 }
3035 Node* andI = cmpu->in(1);
3036 Node* addI = cmpu->in(2);
3037 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3038 // Not "m & x" and "m + 1"
3039 continue;
3040 }
3041
3042 Node* m = addI->in(1);
3043 if (m == andI->in(1) || m == andI->in(2)) {
3044 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3045 push_if_not_bottom_type(worklist, bol);
3046 }
3047 }
3048 }
3049
3050 // 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'.
3051 // Seem PhiNode::Value().
3052 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3053 uint use_op = use->Opcode();
3054 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3055 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3056 if (phi != nullptr) {
3057 worklist.push(phi);
3058 }
3059 }
3060 }
3061
3062 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3063 // See LoadNode::Value().
3064 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3065 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
3066 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
3067
3068 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3069 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3070 Node* loadp = use->fast_out(i);
3071 const Type* ut = loadp->bottom_type();
3072 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3073 if (has_load_barrier_nodes) {
3074 // Search for load barriers behind the load
3075 push_load_barrier(worklist, barrier_set, loadp);
3076 }
3077 worklist.push(loadp);
3078 }
3079 }
3080 }
3081 }
3082
3083 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3084 add_users_to_worklist_if(worklist, use, [&](Node* u) {
3085 return barrier_set->is_gc_barrier_node(u);
3086 });
3087 }
3088
3089 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3090 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3091 // value changed.
3092 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3093 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3094 uint use_op = use->Opcode();
3095 if (
3096 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3097 (parent_type != nullptr && parent_type->is_con()) ||
3098 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3099 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3100
3101 auto push_and_uses_to_worklist = [&](Node* n) {
3102 uint opc = n->Opcode();
3103 if (opc == Op_AndI || opc == Op_AndL) {
3104 push_if_not_bottom_type(worklist, n);
3105 }
3106 };
3107 auto is_boundary = [](Node* n) {
3108 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3109 };
3110 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3111 }
3112 }
3113
3114 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3115 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3116 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3117 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3118 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3119 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3120 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3121 Node* other_cmp_input = use->in(1);
3122 add_users_to_worklist_if(worklist, other_cmp_input, [&](Node* u) {
3123 return u->is_CastII() && not_bottom_type(u);
3124 });
3125 }
3126 }
3127
3128 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3129 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3130 push_if_not_bottom_type(worklist, use->unique_out());
3131 }
3132 }
3133
3134 //------------------------------do_transform-----------------------------------
3135 // Top level driver for the recursive transformer
3136 void PhaseCCP::do_transform() {
3137 // Correct leaves of new-space Nodes; they point to old-space.
3138 C->set_root( transform(C->root())->as_Root() );
3139 assert( C->top(), "missing TOP node" );
3140 assert( C->root(), "missing root" );
3141 }
3142
3143 //------------------------------transform--------------------------------------
3144 // Given a Node in old-space, clone him into new-space.
3145 // Convert any of his old-space children into new-space children.
3146 Node *PhaseCCP::transform( Node *n ) {
3147 assert(n->is_Root(), "traversal must start at root");
3148 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3149
3150 ResourceMark rm;
3151 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3152 Node_List node_map;
3153 // Pre-allocate to avoid frequent realloc
3154 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3155 // track all visited nodes, so that we can remove the complement
3156 Unique_Node_List useful;
3157
3158 if (KillPathsReachableByDeadTypeNode) {
3159 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3160 Node* type_node = _maybe_top_type_nodes.at(i);
3161 if (type(type_node) == Type::TOP) {
3162 ResourceMark rm;
3163 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3164 }
3165 }
3166 } else {
3167 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3168 }
3169
3170 // Initialize the traversal.
3171 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3172 // 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
3173 // 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
3174 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3175 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3176 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3177 // and all safepoints.
3178 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3179 Node* nn = _root_and_safepoints.at(i);
3180 Node* new_node = node_map[nn->_idx];
3181 assert(new_node == nullptr, "");
3182 new_node = transform_once(nn); // Check for constant
3183 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3184 transform_stack.push(new_node); // Process children of cloned node
3185 useful.push(new_node);
3186 }
3187
3188 while (transform_stack.is_nonempty()) {
3189 Node* clone = transform_stack.pop();
3190 uint cnt = clone->req();
3191 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3192 Node *input = clone->in(i);
3193 if( input != nullptr ) { // Ignore nulls
3194 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3195 if( new_input == nullptr ) {
3196 new_input = transform_once(input); // Check for constant
3197 node_map.map( input->_idx, new_input );// Flag as having been cloned
3198 transform_stack.push(new_input); // Process children of cloned node
3199 useful.push(new_input);
3200 }
3201 assert( new_input == clone->in(i), "insanity check");
3202 }
3203 }
3204 }
3205
3206 // The above transformation might lead to subgraphs becoming unreachable from the
3207 // bottom while still being reachable from the top. As a result, nodes in that
3208 // subgraph are not transformed and their bottom types are not updated, leading to
3209 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3210 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3211 // because their address type is inconsistent. Therefore, we aggressively remove
3212 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3213 // to remove them anyway.
3214 if (C->cached_top_node()) {
3215 useful.push(C->cached_top_node());
3216 }
3217 C->update_dead_node_list(useful);
3218 remove_useless_nodes(useful.member_set());
3219 _worklist.remove_useless_nodes(useful.member_set());
3220 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3221
3222 Node* new_root = node_map[n->_idx];
3223 assert(new_root->is_Root(), "transformed root node must be a root node");
3224 return new_root;
3225 }
3226
3227 //------------------------------transform_once---------------------------------
3228 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3229 Node *PhaseCCP::transform_once( Node *n ) {
3230 const Type *t = type(n);
3231 // Constant? Use constant Node instead
3232 if( t->singleton() ) {
3233 Node *nn = n; // Default is to return the original constant
3234 if( t == Type::TOP ) {
3235 // cache my top node on the Compile instance
3236 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3237 C->set_cached_top_node(ConNode::make(Type::TOP));
3238 set_type(C->top(), Type::TOP);
3239 }
3240 nn = C->top();
3241 }
3242 if( !n->is_Con() ) {
3243 if( t != Type::TOP ) {
3244 nn = makecon(t); // ConNode::make(t);
3245 NOT_PRODUCT( inc_constants(); )
3246 } else if( n->is_Region() ) { // Unreachable region
3247 // Note: nn == C->top()
3248 n->set_req(0, nullptr); // Cut selfreference
3249 bool progress = true;
3250 uint max = n->outcnt();
3251 DUIterator i;
3252 while (progress) {
3253 progress = false;
3254 // Eagerly remove dead phis to avoid phis copies creation.
3255 for (i = n->outs(); n->has_out(i); i++) {
3256 Node* m = n->out(i);
3257 if (m->is_Phi()) {
3258 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3259 replace_node(m, nn);
3260 if (max != n->outcnt()) {
3261 progress = true;
3262 i = n->refresh_out_pos(i);
3263 max = n->outcnt();
3264 }
3265 }
3266 }
3267 }
3268 }
3269 replace_node(n,nn); // Update DefUse edges for new constant
3270 }
3271 return nn;
3272 }
3273
3274 // If x is a TypeNode, capture any more-precise type permanently into Node
3275 if (t != n->bottom_type()) {
3276 hash_delete(n); // changing bottom type may force a rehash
3277 n->raise_bottom_type(t);
3278 _worklist.push(n); // n re-enters the hash table via the worklist
3279 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3280 }
3281
3282 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3283 switch( n->Opcode() ) {
3284 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3285 case Op_CallDynamicJava:
3286 case Op_FastLock: // Revisit FastLocks for lock coarsening
3287 case Op_If:
3288 case Op_CountedLoopEnd:
3289 case Op_Region:
3290 case Op_Loop:
3291 case Op_CountedLoop:
3292 case Op_Conv2B:
3293 case Op_Opaque1:
3294 _worklist.push(n);
3295 break;
3296 default:
3297 break;
3298 }
3299
3300 return n;
3301 }
3302
3303 //---------------------------------saturate------------------------------------
3304 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3305 const Type* limit_type) const {
3306 const Type* wide_type = new_type->widen(old_type, limit_type);
3307 if (wide_type != new_type) { // did we widen?
3308 // If so, we may have widened beyond the limit type. Clip it back down.
3309 new_type = wide_type->filter(limit_type);
3310 }
3311 return new_type;
3312 }
3313
3314 //------------------------------print_statistics-------------------------------
3315 #ifndef PRODUCT
3316 void PhaseCCP::print_statistics() {
3317 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3318 }
3319 #endif
3320
3321
3322 //=============================================================================
3323 #ifndef PRODUCT
3324 uint PhasePeephole::_total_peepholes = 0;
3325 #endif
3326 //------------------------------PhasePeephole----------------------------------
3327 // Conditional Constant Propagation, ala Wegman & Zadeck
3328 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3329 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3330 NOT_PRODUCT( clear_peepholes(); )
3331 }
3332
3333 #ifndef PRODUCT
3334 //------------------------------~PhasePeephole---------------------------------
3335 PhasePeephole::~PhasePeephole() {
3336 _total_peepholes += count_peepholes();
3337 }
3338 #endif
3339
3340 //------------------------------transform--------------------------------------
3341 Node *PhasePeephole::transform( Node *n ) {
3342 ShouldNotCallThis();
3343 return nullptr;
3344 }
3345
3346 //------------------------------do_transform-----------------------------------
3347 void PhasePeephole::do_transform() {
3348 bool method_name_not_printed = true;
3349
3350 // Examine each basic block
3351 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3352 Block* block = _cfg.get_block(block_number);
3353 bool block_not_printed = true;
3354
3355 for (bool progress = true; progress;) {
3356 progress = false;
3357 // block->end_idx() not valid after PhaseRegAlloc
3358 uint end_index = block->number_of_nodes();
3359 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3360 Node *n = block->get_node(instruction_index);
3361 if( n->is_Mach() ) {
3362 MachNode *m = n->as_Mach();
3363 // check for peephole opportunities
3364 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3365 if( result != -1 ) {
3366 #ifndef PRODUCT
3367 if( PrintOptoPeephole ) {
3368 // Print method, first time only
3369 if( C->method() && method_name_not_printed ) {
3370 C->method()->print_short_name(); tty->cr();
3371 method_name_not_printed = false;
3372 }
3373 // Print this block
3374 if( Verbose && block_not_printed) {
3375 tty->print_cr("in block");
3376 block->dump();
3377 block_not_printed = false;
3378 }
3379 // Print the peephole number
3380 tty->print_cr("peephole number: %d", result);
3381 }
3382 inc_peepholes();
3383 #endif
3384 // Set progress, start again
3385 progress = true;
3386 break;
3387 }
3388 }
3389 }
3390 }
3391 }
3392 }
3393
3394 //------------------------------print_statistics-------------------------------
3395 #ifndef PRODUCT
3396 void PhasePeephole::print_statistics() {
3397 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3398 }
3399 #endif
3400
3401
3402 //=============================================================================
3403 //------------------------------set_req_X--------------------------------------
3404 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3405 assert( is_not_dead(n), "can not use dead node");
3406 #ifdef ASSERT
3407 if (igvn->hash_find(this) == this) {
3408 tty->print_cr("Need to remove from hash before changing edges");
3409 this->dump(1);
3410 tty->print_cr("Set at i = %d", i);
3411 n->dump();
3412 assert(false, "Need to remove from hash before changing edges");
3413 }
3414 #endif
3415 Node *old = in(i);
3416 set_req(i, n);
3417
3418 // old goes dead?
3419 if( old ) {
3420 switch (old->outcnt()) {
3421 case 0:
3422 // Put into the worklist to kill later. We do not kill it now because the
3423 // recursive kill will delete the current node (this) if dead-loop exists
3424 if (!old->is_top())
3425 igvn->_worklist.push( old );
3426 break;
3427 case 1:
3428 if( old->is_Store() || old->has_special_unique_user() )
3429 igvn->add_users_to_worklist( old );
3430 break;
3431 case 2:
3432 if( old->is_Store() )
3433 igvn->add_users_to_worklist( old );
3434 if( old->Opcode() == Op_Region )
3435 igvn->_worklist.push(old);
3436 break;
3437 case 3:
3438 if( old->Opcode() == Op_Region ) {
3439 igvn->_worklist.push(old);
3440 igvn->add_users_to_worklist( old );
3441 }
3442 break;
3443 default:
3444 break;
3445 }
3446
3447 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3448 }
3449 }
3450
3451 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3452 PhaseIterGVN* igvn = gvn->is_IterGVN();
3453 if (igvn == nullptr) {
3454 set_req(i, n);
3455 return;
3456 }
3457 set_req_X(i, n, igvn);
3458 }
3459
3460 //-------------------------------replace_by-----------------------------------
3461 // Using def-use info, replace one node for another. Follow the def-use info
3462 // to all users of the OLD node. Then make all uses point to the NEW node.
3463 void Node::replace_by(Node *new_node) {
3464 assert(!is_top(), "top node has no DU info");
3465 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3466 Node* use = last_out(i);
3467 uint uses_found = 0;
3468 for (uint j = 0; j < use->len(); j++) {
3469 if (use->in(j) == this) {
3470 if (j < use->req())
3471 use->set_req(j, new_node);
3472 else use->set_prec(j, new_node);
3473 uses_found++;
3474 }
3475 }
3476 i -= uses_found; // we deleted 1 or more copies of this edge
3477 }
3478 }
3479
3480 //=============================================================================
3481 //-----------------------------------------------------------------------------
3482 void Type_Array::grow( uint i ) {
3483 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3484 if( !_max ) {
3485 _max = 1;
3486 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3487 _types[0] = nullptr;
3488 }
3489 uint old = _max;
3490 _max = next_power_of_2(i);
3491 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3492 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3493 }
3494
3495 //------------------------------dump-------------------------------------------
3496 #ifndef PRODUCT
3497 void Type_Array::dump() const {
3498 uint max = Size();
3499 for( uint i = 0; i < max; i++ ) {
3500 if( _types[i] != nullptr ) {
3501 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3502 }
3503 }
3504 }
3505 #endif
--- EOF ---