1 /*
2 * Copyright (c) 1997, 2025, 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/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE 255
46
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 DEBUG_ONLY( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 DEBUG_ONLY( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
368 // Implementation requires an edge from root to each SafePointNode
369 // at a backward branch. Inserted in add_safepoint().
370
371 // Identify nodes that are reachable from below, useful.
372 C->identify_useful_nodes(_useful);
373 // Update dead node list
374 C->update_dead_node_list(_useful);
375
376 // Remove all useless nodes from PhaseValues' recorded types
377 // Must be done before disconnecting nodes to preserve hash-table-invariant
378 gvn->remove_useless_nodes(_useful.member_set());
379
380 // Remove all useless nodes from future worklist
381 worklist.remove_useless_nodes(_useful.member_set());
382
383 // Disconnect 'useless' nodes that are adjacent to useful nodes
384 C->disconnect_useless_nodes(_useful, worklist);
385 }
386
387 //=============================================================================
388 //------------------------------PhaseRenumberLive------------------------------
389 // First, remove useless nodes (equivalent to identifying live nodes).
390 // Then, renumber live nodes.
391 //
392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
395 // value in the range [0, x).
396 //
397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
399 //
400 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
402 // be processed by IGVN after removal of the useless nodes.
403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
404 // the node's type. The mapping is updated to use the new node IDs as well. We
405 // create a new map, and swap it with the old one.
406 //
407 // Other data structures used by the compiler are not updated. The hash table for value
408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
409 // computing the hash values is not based on node IDs.
410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
411 Unique_Node_List& worklist,
412 PhaseNumber phase_num) :
413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
414 _new_type_array(C->comp_arena()),
415 _old2new_map(C->unique(), C->unique(), -1),
416 _is_pass_finished(false),
417 _live_node_count(C->live_nodes())
418 {
419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
421 assert(_delayed.size() == 0, "should be empty");
422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
423 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
424
425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
426 if (old_node_note_array != nullptr) {
427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
428 new_size = MAX2(8, new_size);
429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
430 C->grow_node_notes(C->node_note_array(), new_size);
431 }
432
433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
434
435 // Iterate over the set of live nodes.
436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
437 Node* n = _useful.at(current_idx);
438
439 const Type* type = gvn->type_or_null(n);
440 _new_type_array.map(current_idx, type);
441
442 assert(_old2new_map.at(n->_idx) == -1, "already seen");
443 _old2new_map.at_put(n->_idx, current_idx);
444
445 if (old_node_note_array != nullptr) {
446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
447 C->set_node_notes_at(current_idx, nn);
448 }
449
450 n->set_idx(current_idx); // Update node ID.
451
452 if (update_embedded_ids(n) < 0) {
453 _delayed.push(n); // has embedded IDs; handle later
454 }
455 }
456
457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
458 worklist.recompute_idx_set();
459
460 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
461
462 _is_pass_finished = true; // pass finished; safe to process delayed updates
463
464 while (_delayed.size() > 0) {
465 Node* n = _delayed.pop();
466 int no_of_updates = update_embedded_ids(n);
467 assert(no_of_updates > 0, "should be updated");
468 }
469
470 // Replace the compiler's type information with the updated type information.
471 gvn->types().swap(_new_type_array);
472
473 // Update the unique node count of the compilation to the number of currently live nodes.
474 C->set_unique(_live_node_count);
475
476 // Set the dead node count to 0 and reset dead node list.
477 C->reset_dead_node_list();
478 }
479
480 int PhaseRenumberLive::new_index(int old_idx) {
481 assert(_is_pass_finished, "not finished");
482 if (_old2new_map.at(old_idx) == -1) { // absent
483 // Allocate a placeholder to preserve uniqueness
484 _old2new_map.at_put(old_idx, _live_node_count);
485 _live_node_count++;
486 }
487 return _old2new_map.at(old_idx);
488 }
489
490 int PhaseRenumberLive::update_embedded_ids(Node* n) {
491 int no_of_updates = 0;
492 if (n->is_Phi()) {
493 PhiNode* phi = n->as_Phi();
494 if (phi->_inst_id != -1) {
495 if (!_is_pass_finished) {
496 return -1; // delay
497 }
498 int new_idx = new_index(phi->_inst_id);
499 assert(new_idx != -1, "");
500 phi->_inst_id = new_idx;
501 no_of_updates++;
502 }
503 if (phi->_inst_mem_id != -1) {
504 if (!_is_pass_finished) {
505 return -1; // delay
506 }
507 int new_idx = new_index(phi->_inst_mem_id);
508 assert(new_idx != -1, "");
509 phi->_inst_mem_id = new_idx;
510 no_of_updates++;
511 }
512 }
513
514 const Type* type = _new_type_array.fast_lookup(n->_idx);
515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
516 if (!_is_pass_finished) {
517 return -1; // delay
518 }
519 int old_idx = type->is_oopptr()->instance_id();
520 int new_idx = new_index(old_idx);
521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
522 _new_type_array.map(n->_idx, new_type);
523 no_of_updates++;
524 }
525
526 return no_of_updates;
527 }
528
529 void PhaseValues::init_con_caches() {
530 memset(_icons,0,sizeof(_icons));
531 memset(_lcons,0,sizeof(_lcons));
532 memset(_zcons,0,sizeof(_zcons));
533 }
534
535 //--------------------------------find_int_type--------------------------------
536 const TypeInt* PhaseValues::find_int_type(Node* n) {
537 if (n == nullptr) return nullptr;
538 // Call type_or_null(n) to determine node's type since we might be in
539 // parse phase and call n->Value() may return wrong type.
540 // (For example, a phi node at the beginning of loop parsing is not ready.)
541 const Type* t = type_or_null(n);
542 if (t == nullptr) return nullptr;
543 return t->isa_int();
544 }
545
546
547 //-------------------------------find_long_type--------------------------------
548 const TypeLong* PhaseValues::find_long_type(Node* n) {
549 if (n == nullptr) return nullptr;
550 // (See comment above on type_or_null.)
551 const Type* t = type_or_null(n);
552 if (t == nullptr) return nullptr;
553 return t->isa_long();
554 }
555
556 //------------------------------~PhaseValues-----------------------------------
557 #ifndef PRODUCT
558 PhaseValues::~PhaseValues() {
559 // Statistics for NodeHash
560 _table.dump();
561 // Statistics for value progress and efficiency
562 if( PrintCompilation && Verbose && WizardMode ) {
563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
565 if( made_transforms() != 0 ) {
566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
567 } else {
568 tty->cr();
569 }
570 }
571 }
572 #endif
573
574 //------------------------------makecon----------------------------------------
575 ConNode* PhaseValues::makecon(const Type* t) {
576 assert(t->singleton(), "must be a constant");
577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
578 switch (t->base()) { // fast paths
579 case Type::Half:
580 case Type::Top: return (ConNode*) C->top();
581 case Type::Int: return intcon( t->is_int()->get_con() );
582 case Type::Long: return longcon( t->is_long()->get_con() );
583 default: break;
584 }
585 if (t->is_zero_type())
586 return zerocon(t->basic_type());
587 return uncached_makecon(t);
588 }
589
590 //--------------------------uncached_makecon-----------------------------------
591 // Make an idealized constant - one of ConINode, ConPNode, etc.
592 ConNode* PhaseValues::uncached_makecon(const Type *t) {
593 assert(t->singleton(), "must be a constant");
594 ConNode* x = ConNode::make(t);
595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
596 if (k == nullptr) {
597 set_type(x, t); // Missed, provide type mapping
598 GrowableArray<Node_Notes*>* nna = C->node_note_array();
599 if (nna != nullptr) {
600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
601 loc->clear(); // do not put debug info on constants
602 }
603 } else {
604 x->destruct(this); // Hit, destroy duplicate constant
605 x = k; // use existing constant
606 }
607 return x;
608 }
609
610 //------------------------------intcon-----------------------------------------
611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
612 ConINode* PhaseValues::intcon(jint i) {
613 // Small integer? Check cache! Check that cached node is not dead
614 if (i >= _icon_min && i <= _icon_max) {
615 ConINode* icon = _icons[i-_icon_min];
616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
617 return icon;
618 }
619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
620 assert(icon->is_Con(), "");
621 if (i >= _icon_min && i <= _icon_max)
622 _icons[i-_icon_min] = icon; // Cache small integers
623 return icon;
624 }
625
626 //------------------------------longcon----------------------------------------
627 // Fast long constant.
628 ConLNode* PhaseValues::longcon(jlong l) {
629 // Small integer? Check cache! Check that cached node is not dead
630 if (l >= _lcon_min && l <= _lcon_max) {
631 ConLNode* lcon = _lcons[l-_lcon_min];
632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
633 return lcon;
634 }
635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
636 assert(lcon->is_Con(), "");
637 if (l >= _lcon_min && l <= _lcon_max)
638 _lcons[l-_lcon_min] = lcon; // Cache small integers
639 return lcon;
640 }
641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
642 if (bt == T_INT) {
643 return intcon(checked_cast<jint>(l));
644 }
645 assert(bt == T_LONG, "not an integer");
646 return longcon(l);
647 }
648
649
650 //------------------------------zerocon-----------------------------------------
651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
652 ConNode* PhaseValues::zerocon(BasicType bt) {
653 assert((uint)bt <= _zcon_max, "domain check");
654 ConNode* zcon = _zcons[bt];
655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
656 return zcon;
657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
658 _zcons[bt] = zcon;
659 return zcon;
660 }
661
662
663
664 //=============================================================================
665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
666 return k->Ideal(this, can_reshape);
667 }
668
669 //------------------------------transform--------------------------------------
670 // Return a node which computes the same function as this node, but
671 // in a faster or cheaper fashion.
672 Node* PhaseGVN::transform(Node* n) {
673 NOT_PRODUCT( set_transforms(); )
674
675 // Apply the Ideal call in a loop until it no longer applies
676 Node* k = n;
677 Node* i = apply_ideal(k, /*can_reshape=*/false);
678 NOT_PRODUCT(uint loop_count = 1;)
679 while (i != nullptr) {
680 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
681 k = i;
682 #ifdef ASSERT
683 if (loop_count >= K + C->live_nodes()) {
684 dump_infinite_loop_info(i, "PhaseGVN::transform");
685 }
686 #endif
687 i = apply_ideal(k, /*can_reshape=*/false);
688 NOT_PRODUCT(loop_count++;)
689 }
690 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
691
692 // If brand new node, make space in type array.
693 ensure_type_or_null(k);
694
695 // Since I just called 'Value' to compute the set of run-time values
696 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
697 // cache Value. Later requests for the local phase->type of this Node can
698 // use the cached Value instead of suffering with 'bottom_type'.
699 const Type* t = k->Value(this); // Get runtime Value set
700 assert(t != nullptr, "value sanity");
701 if (type_or_null(k) != t) {
702 #ifndef PRODUCT
703 // Do not count initial visit to node as a transformation
704 if (type_or_null(k) == nullptr) {
705 inc_new_values();
706 set_progress();
707 }
708 #endif
709 set_type(k, t);
710 // If k is a TypeNode, capture any more-precise type permanently into Node
711 k->raise_bottom_type(t);
712 }
713
714 if (t->singleton() && !k->is_Con()) {
715 NOT_PRODUCT(set_progress();)
716 return makecon(t); // Turn into a constant
717 }
718
719 // Now check for Identities
720 i = k->Identity(this); // Look for a nearby replacement
721 if (i != k) { // Found? Return replacement!
722 NOT_PRODUCT(set_progress();)
723 return i;
724 }
725
726 // Global Value Numbering
727 i = hash_find_insert(k); // Insert if new
728 if (i && (i != k)) {
729 // Return the pre-existing node
730 NOT_PRODUCT(set_progress();)
731 return i;
732 }
733
734 // Return Idealized original
735 return k;
736 }
737
738 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
739 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
740 return false;
741 }
742 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
743 return false;
744 }
745 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
746 int i = 0;
747 while (d != n) {
748 n = IfNode::up_one_dom(n, linear_only);
749 i++;
750 if (n == nullptr || i >= 100) {
751 return false;
752 }
753 }
754 return true;
755 }
756
757 #ifdef ASSERT
758 //------------------------------dead_loop_check--------------------------------
759 // Check for a simple dead loop when a data node references itself directly
760 // or through an other data node excluding cons and phis.
761 void PhaseGVN::dead_loop_check( Node *n ) {
762 // Phi may reference itself in a loop
763 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) {
764 // Do 2 levels check and only data inputs.
765 bool no_dead_loop = true;
766 uint cnt = n->req();
767 for (uint i = 1; i < cnt && no_dead_loop; i++) {
768 Node *in = n->in(i);
769 if (in == n) {
770 no_dead_loop = false;
771 } else if (in != nullptr && !in->is_dead_loop_safe()) {
772 uint icnt = in->req();
773 for (uint j = 1; j < icnt && no_dead_loop; j++) {
774 if (in->in(j) == n || in->in(j) == in)
775 no_dead_loop = false;
776 }
777 }
778 }
779 if (!no_dead_loop) { n->dump_bfs(100, nullptr, ""); }
780 assert(no_dead_loop, "dead loop detected");
781 }
782 }
783
784
785 /**
786 * Dumps information that can help to debug the problem. A debug
787 * build fails with an assert.
788 */
789 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
790 n->dump(4);
791 assert(false, "infinite loop in %s", where);
792 }
793 #endif
794
795 //=============================================================================
796 //------------------------------PhaseIterGVN-----------------------------------
797 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
798 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
799 _worklist(*C->igvn_worklist())
800 {
801 _iterGVN = true;
802 assert(&_worklist == &igvn->_worklist, "sanity");
803 }
804
805 //------------------------------PhaseIterGVN-----------------------------------
806 // Initialize from scratch
807 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
808 _worklist(*C->igvn_worklist())
809 {
810 _iterGVN = true;
811 uint max;
812
813 // Dead nodes in the hash table inherited from GVN were not treated as
814 // roots during def-use info creation; hence they represent an invisible
815 // use. Clear them out.
816 max = _table.size();
817 for( uint i = 0; i < max; ++i ) {
818 Node *n = _table.at(i);
819 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
820 if( n->is_top() ) continue;
821 // If remove_useless_nodes() has run, we expect no such nodes left.
822 assert(false, "remove_useless_nodes missed this node");
823 hash_delete(n);
824 }
825 }
826
827 // Any Phis or Regions on the worklist probably had uses that could not
828 // make more progress because the uses were made while the Phis and Regions
829 // were in half-built states. Put all uses of Phis and Regions on worklist.
830 max = _worklist.size();
831 for( uint j = 0; j < max; j++ ) {
832 Node *n = _worklist.at(j);
833 uint uop = n->Opcode();
834 if( uop == Op_Phi || uop == Op_Region ||
835 n->is_Type() ||
836 n->is_Mem() )
837 add_users_to_worklist(n);
838 }
839 }
840
841 void PhaseIterGVN::shuffle_worklist() {
842 if (_worklist.size() < 2) return;
843 for (uint i = _worklist.size() - 1; i >= 1; i--) {
844 uint j = C->random() % (i + 1);
845 swap(_worklist.adr()[i], _worklist.adr()[j]);
846 }
847 }
848
849 #ifndef PRODUCT
850 void PhaseIterGVN::verify_step(Node* n) {
851 if (is_verify_def_use()) {
852 ResourceMark rm;
853 VectorSet visited;
854 Node_List worklist;
855
856 _verify_window[_verify_counter % _verify_window_size] = n;
857 ++_verify_counter;
858 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
859 ++_verify_full_passes;
860 worklist.push(C->root());
861 Node::verify(-1, visited, worklist);
862 return;
863 }
864 for (int i = 0; i < _verify_window_size; i++) {
865 Node* n = _verify_window[i];
866 if (n == nullptr) {
867 continue;
868 }
869 if (n->in(0) == NodeSentinel) { // xform_idom
870 _verify_window[i] = n->in(1);
871 --i;
872 continue;
873 }
874 // Typical fanout is 1-2, so this call visits about 6 nodes.
875 if (!visited.test_set(n->_idx)) {
876 worklist.push(n);
877 }
878 }
879 Node::verify(4, visited, worklist);
880 }
881 }
882
883 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
884 const Type* newtype = type_or_null(n);
885 if (nn != n || oldtype != newtype) {
886 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
887 }
888 if (TraceIterativeGVN) {
889 uint wlsize = _worklist.size();
890 if (nn != n) {
891 // print old node
892 tty->print("< ");
893 if (oldtype != newtype && oldtype != nullptr) {
894 oldtype->dump();
895 }
896 do { tty->print("\t"); } while (tty->position() < 16);
897 tty->print("<");
898 n->dump();
899 }
900 if (oldtype != newtype || nn != n) {
901 // print new node and/or new type
902 if (oldtype == nullptr) {
903 tty->print("* ");
904 } else if (nn != n) {
905 tty->print("> ");
906 } else {
907 tty->print("= ");
908 }
909 if (newtype == nullptr) {
910 tty->print("null");
911 } else {
912 newtype->dump();
913 }
914 do { tty->print("\t"); } while (tty->position() < 16);
915 nn->dump();
916 }
917 if (Verbose && wlsize < _worklist.size()) {
918 tty->print(" Push {");
919 while (wlsize != _worklist.size()) {
920 Node* pushed = _worklist.at(wlsize++);
921 tty->print(" %d", pushed->_idx);
922 }
923 tty->print_cr(" }");
924 }
925 if (nn != n) {
926 // ignore n, it might be subsumed
927 verify_step((Node*) nullptr);
928 }
929 }
930 }
931
932 void PhaseIterGVN::init_verifyPhaseIterGVN() {
933 _verify_counter = 0;
934 _verify_full_passes = 0;
935 for (int i = 0; i < _verify_window_size; i++) {
936 _verify_window[i] = nullptr;
937 }
938 #ifdef ASSERT
939 // Verify that all modified nodes are on _worklist
940 Unique_Node_List* modified_list = C->modified_nodes();
941 while (modified_list != nullptr && modified_list->size()) {
942 Node* n = modified_list->pop();
943 if (!n->is_Con() && !_worklist.member(n)) {
944 n->dump();
945 fatal("modified node is not on IGVN._worklist");
946 }
947 }
948 #endif
949 }
950
951 void PhaseIterGVN::verify_PhaseIterGVN() {
952 #ifdef ASSERT
953 // Verify nodes with changed inputs.
954 Unique_Node_List* modified_list = C->modified_nodes();
955 while (modified_list != nullptr && modified_list->size()) {
956 Node* n = modified_list->pop();
957 if (!n->is_Con()) { // skip Con nodes
958 n->dump();
959 fatal("modified node was not processed by IGVN.transform_old()");
960 }
961 }
962 #endif
963
964 C->verify_graph_edges();
965 if (is_verify_def_use() && PrintOpto) {
966 if (_verify_counter == _verify_full_passes) {
967 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
968 (int) _verify_full_passes);
969 } else {
970 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
971 (int) _verify_counter, (int) _verify_full_passes);
972 }
973 }
974
975 #ifdef ASSERT
976 if (modified_list != nullptr) {
977 while (modified_list->size() > 0) {
978 Node* n = modified_list->pop();
979 n->dump();
980 assert(false, "VerifyIterativeGVN: new modified node was added");
981 }
982 }
983
984 verify_optimize();
985 #endif
986 }
987 #endif /* PRODUCT */
988
989 #ifdef ASSERT
990 /**
991 * Dumps information that can help to debug the problem. A debug
992 * build fails with an assert.
993 */
994 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
995 n->dump(4);
996 _worklist.dump();
997 assert(false, "infinite loop in %s", where);
998 }
999
1000 /**
1001 * Prints out information about IGVN if the 'verbose' option is used.
1002 */
1003 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1004 if (TraceIterativeGVN && Verbose) {
1005 tty->print(" Pop ");
1006 n->dump();
1007 if ((num_processed % 100) == 0) {
1008 _worklist.print_set();
1009 }
1010 }
1011 }
1012 #endif /* ASSERT */
1013
1014 void PhaseIterGVN::optimize() {
1015 DEBUG_ONLY(uint num_processed = 0;)
1016 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1017 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1018 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1019 if (StressIGVN) {
1020 shuffle_worklist();
1021 }
1022
1023 // The node count check in the loop below (check_node_count) assumes that we
1024 // increase the live node count with at most
1025 // max_live_nodes_increase_per_iteration in between checks. If this
1026 // assumption does not hold, there is a risk that we exceed the max node
1027 // limit in between checks and trigger an assert during node creation.
1028 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 3;
1029
1030 uint loop_count = 0;
1031 // Pull from worklist and transform the node. If the node has changed,
1032 // update edge info and put uses on worklist.
1033 while (_worklist.size() > 0) {
1034 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1035 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1036 return;
1037 }
1038 Node* n = _worklist.pop();
1039 if (loop_count >= K * C->live_nodes()) {
1040 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1041 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1042 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1043 return;
1044 }
1045 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1046 if (n->outcnt() != 0) {
1047 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1048 // Do the transformation
1049 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1050 Node* nn = transform_old(n);
1051 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1052 // Ensure we did not increase the live node count with more than
1053 // max_live_nodes_increase_per_iteration during the call to transform_old
1054 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1055 assert(increase < max_live_nodes_increase_per_iteration,
1056 "excessive live node increase in single iteration of IGVN: %d "
1057 "(should be at most %d)",
1058 increase, max_live_nodes_increase_per_iteration);
1059 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1060 } else if (!n->is_top()) {
1061 remove_dead_node(n);
1062 }
1063 loop_count++;
1064 }
1065 NOT_PRODUCT(verify_PhaseIterGVN();)
1066 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1067 }
1068
1069 #ifdef ASSERT
1070 void PhaseIterGVN::verify_optimize() {
1071 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1072
1073 if (is_verify_Value() ||
1074 is_verify_Ideal() ||
1075 is_verify_Identity() ||
1076 is_verify_invariants()) {
1077 ResourceMark rm;
1078 Unique_Node_List worklist;
1079 bool failure = false;
1080 // BFS all nodes, starting at root
1081 worklist.push(C->root());
1082 for (uint j = 0; j < worklist.size(); ++j) {
1083 Node* n = worklist.at(j);
1084 if (is_verify_Value()) { failure |= verify_Value_for(n); }
1085 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, false); }
1086 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, true); }
1087 if (is_verify_Identity()) { failure |= verify_Identity_for(n); }
1088 if (is_verify_invariants()) { failure |= verify_node_invariants_for(n); }
1089 // traverse all inputs and outputs
1090 for (uint i = 0; i < n->req(); i++) {
1091 if (n->in(i) != nullptr) {
1092 worklist.push(n->in(i));
1093 }
1094 }
1095 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1096 worklist.push(n->fast_out(i));
1097 }
1098 }
1099 // If we get this assert, check why the reported nodes were not processed again in IGVN.
1100 // We should either make sure that these nodes are properly added back to the IGVN worklist
1101 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception
1102 // in the verification code above if that is not possible for some reason (like Load nodes).
1103 assert(!failure, "Missed optimization opportunity/broken graph in PhaseIterGVN");
1104 }
1105
1106 verify_empty_worklist(nullptr);
1107 }
1108
1109 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1110 // Verify that the igvn worklist is empty. If no optimization happened, then
1111 // nothing needs to be on the worklist.
1112 if (_worklist.size() == 0) { return; }
1113
1114 stringStream ss; // Print as a block without tty lock.
1115 for (uint j = 0; j < _worklist.size(); j++) {
1116 Node* n = _worklist.at(j);
1117 ss.print("igvn.worklist[%d] ", j);
1118 n->dump("\n", false, &ss);
1119 }
1120 if (_worklist.size() != 0 && node != nullptr) {
1121 ss.print_cr("Previously optimized:");
1122 node->dump("\n", false, &ss);
1123 }
1124 tty->print_cr("%s", ss.as_string());
1125 assert(false, "igvn worklist must still be empty after verify");
1126 }
1127
1128 // Check that type(n) == n->Value(), return true if we have a failure.
1129 // We have a list of exceptions, see detailed comments in code.
1130 // (1) Integer "widen" changes, but the range is the same.
1131 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1132 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1133 bool PhaseIterGVN::verify_Value_for(Node* n, bool strict) {
1134 // If we assert inside type(n), because the type is still a null, then maybe
1135 // the node never went through gvn.transform, which would be a bug.
1136 const Type* told = type(n);
1137 const Type* tnew = n->Value(this);
1138 if (told == tnew) {
1139 return false;
1140 }
1141 // Exception (1)
1142 // Integer "widen" changes, but range is the same.
1143 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1144 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1145 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1146 if (t0->lo_as_long() == t1->lo_as_long() &&
1147 t0->hi_as_long() == t1->hi_as_long()) {
1148 return false; // ignore integer widen
1149 }
1150 }
1151 // Exception (2)
1152 // LoadNode performs deep traversals. Load is not notified for changes far away.
1153 if (!strict && n->is_Load() && !told->singleton()) {
1154 // MemNode::can_see_stored_value looks up through many memory nodes,
1155 // which means we would need to notify modifications from far up in
1156 // the inputs all the way down to the LoadNode. We don't do that.
1157 return false;
1158 }
1159 // Exception (3)
1160 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1161 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1162 // SubNode::Value
1163 // CmpPNode::sub
1164 // MemNode::detect_ptr_independence
1165 // MemNode::all_controls_dominate
1166 // We find all controls of a pointer load, and see if they dominate the control of
1167 // an allocation. If they all dominate, we know the allocation is after (independent)
1168 // of the pointer load, and we can say the pointers are different. For this we call
1169 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1170 // control sub of the allocation. The problems is that sometimes dominates answers
1171 // false conservatively, and later it can determine that it is indeed true. Loops with
1172 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1173 // so the traversal becomes more powerful. This is difficult to remidy, we would have
1174 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1175 // after loop-opts, so that should take care of many of these cases.
1176 return false;
1177 }
1178
1179 stringStream ss; // Print as a block without tty lock.
1180 ss.cr();
1181 ss.print_cr("Missed Value optimization:");
1182 n->dump_bfs(1, nullptr, "", &ss);
1183 ss.print_cr("Current type:");
1184 told->dump_on(&ss);
1185 ss.cr();
1186 ss.print_cr("Optimized type:");
1187 tnew->dump_on(&ss);
1188 ss.cr();
1189 tty->print_cr("%s", ss.as_string());
1190 return true;
1191 }
1192
1193 // Check that all Ideal optimizations that could be done were done.
1194 // Returns true if it found missed optimization opportunities and
1195 // false otherwise (no missed optimization, or skipped verification).
1196 bool PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1197 // First, we check a list of exceptions, where we skip verification,
1198 // because there are known cases where Ideal can optimize after IGVN.
1199 // Some may be expected and cannot be fixed, and others should be fixed.
1200 switch (n->Opcode()) {
1201 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1202 // (see "Range-Check scan limit"). So, it is possible that something
1203 // is optimized in that input subgraph, and the RangeCheck was not
1204 // added to the worklist because it would be too expensive to walk
1205 // down the graph for 1000 nodes and put all on the worklist.
1206 //
1207 // Found with:
1208 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1209 case Op_RangeCheck:
1210 return false;
1211
1212 // IfNode::Ideal does:
1213 // Node* prev_dom = search_identical(dist, igvn);
1214 // which means we seach up the CFG, traversing at most up to a distance.
1215 // If anything happens rather far away from the If, we may not put the If
1216 // back on the worklist.
1217 //
1218 // Found with:
1219 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1220 case Op_If:
1221 return false;
1222
1223 // IfNode::simple_subsuming
1224 // Looks for dominating test that subsumes the current test.
1225 // Notification could be difficult because of larger distance.
1226 //
1227 // Found with:
1228 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1229 // -XX:VerifyIterativeGVN=1110
1230 case Op_CountedLoopEnd:
1231 return false;
1232
1233 // LongCountedLoopEndNode::Ideal
1234 // Probably same issue as above.
1235 //
1236 // Found with:
1237 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1238 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1239 case Op_LongCountedLoopEnd:
1240 return false;
1241
1242 // RegionNode::Ideal does "Skip around the useless IF diamond".
1243 // 245 IfTrue === 244
1244 // 258 If === 245 257
1245 // 259 IfTrue === 258 [[ 263 ]]
1246 // 260 IfFalse === 258 [[ 263 ]]
1247 // 263 Region === 263 260 259 [[ 263 268 ]]
1248 // to
1249 // 245 IfTrue === 244
1250 // 263 Region === 263 245 _ [[ 263 268 ]]
1251 //
1252 // "Useless" means that there is no code in either branch of the If.
1253 // I found a case where this was not done yet during IGVN.
1254 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1255 //
1256 // Found with:
1257 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1258 case Op_Region:
1259 return false;
1260
1261 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1262 // that smaller idx are first. Tracking it back, it led me to
1263 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1264 //
1265 // Example:
1266 // Before PhaseIdealLoop::remix_address_expressions
1267 // 154 AddI === _ 12 144
1268 // After PhaseIdealLoop::remix_address_expressions
1269 // 154 AddI === _ 144 12
1270 // After AddNode::Ideal
1271 // 154 AddI === _ 12 144
1272 //
1273 // I suspect that the node should be added to the IGVN worklist after
1274 // PhaseIdealLoop::remix_address_expressions
1275 //
1276 // This is the only case I looked at, there may be others. Found like this:
1277 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1278 //
1279 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1280 //
1281 // Note: currently all of these fail also for other reasons, for example
1282 // because of "commute" doing the reordering with the phi below. Once
1283 // that is resolved, we can come back to this issue here.
1284 //
1285 // case Op_AddD:
1286 // case Op_AddI:
1287 // case Op_AddL:
1288 // case Op_AddF:
1289 // case Op_MulI:
1290 // case Op_MulL:
1291 // case Op_MulF:
1292 // case Op_MulD:
1293 // if (n->in(1)->_idx > n->in(2)->_idx) {
1294 // // Expect "commute" to revert this case.
1295 // return false;
1296 // }
1297 // break; // keep verifying
1298
1299 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1300 // Check for tight loop increments: Loop-phi of Add of loop-phi
1301 // It wants to take the phi into in(1):
1302 // 471 Phi === 435 38 390
1303 // 390 AddF === _ 471 391
1304 //
1305 // Other Associative operators are also affected equally.
1306 //
1307 // Investigate why this does not happen earlier during IGVN.
1308 //
1309 // Found with:
1310 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1311 // -XX:VerifyIterativeGVN=1110
1312 case Op_AddD:
1313 //case Op_AddI: // Also affected for other reasons, see case further down.
1314 //case Op_AddL: // Also affected for other reasons, see case further down.
1315 case Op_AddF:
1316 case Op_MulI:
1317 case Op_MulL:
1318 case Op_MulF:
1319 case Op_MulD:
1320 case Op_MinF:
1321 case Op_MinD:
1322 case Op_MaxF:
1323 case Op_MaxD:
1324 // XorINode::Ideal
1325 // Found with:
1326 // compiler/intrinsics/chacha/TestChaCha20.java
1327 // -XX:VerifyIterativeGVN=1110
1328 case Op_XorI:
1329 case Op_XorL:
1330 // It seems we may have similar issues with the HF cases.
1331 // Found with aarch64:
1332 // compiler/vectorization/TestFloat16VectorOperations.java
1333 // -XX:VerifyIterativeGVN=1110
1334 case Op_AddHF:
1335 case Op_MulHF:
1336 case Op_MaxHF:
1337 case Op_MinHF:
1338 return false;
1339
1340 // In MulNode::Ideal the edges can be swapped to help value numbering:
1341 //
1342 // // We are OK if right is a constant, or right is a load and
1343 // // left is a non-constant.
1344 // if( !(t2->singleton() ||
1345 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1346 // if( t1->singleton() || // Left input is a constant?
1347 // // Otherwise, sort inputs (commutativity) to help value numbering.
1348 // (in(1)->_idx > in(2)->_idx) ) {
1349 // swap_edges(1, 2);
1350 //
1351 // Why was this not done earlier during IGVN?
1352 //
1353 // Found with:
1354 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1355 // -XX:VerifyIterativeGVN=1110
1356 case Op_AndI:
1357 // Same for AndL.
1358 // Found with:
1359 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1360 // -XX:VerifyIterativeGVN=1110
1361 case Op_AndL:
1362 return false;
1363
1364 // SubLNode::Ideal does transform like:
1365 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1366 //
1367 // In IGVN before verification:
1368 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1369 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1370 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1371 // 8016 ConL === 0 [[ 8383 ]] #long:0
1372 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1373 //
1374 // And then in verification:
1375 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1376 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1377 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1378 // 8016 ConL === 0 [[ 8383 ]] #long:0
1379 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1380 //
1381 // So the form changed from:
1382 // c1 - (y + [8423 ConvI2L])
1383 // to
1384 // c1 - (y + -2)
1385 // but the SubL was not added to the IGVN worklist. Investigate why.
1386 // There could be other issues too.
1387 //
1388 // There seems to be a related AddL IGVN optimization that triggers
1389 // the same SubL optimization, so investigate that too.
1390 //
1391 // Found with:
1392 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1393 case Op_SubL:
1394 return false;
1395
1396 // SubINode::Ideal does
1397 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1398 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1399 //
1400 // Investigate why this does not yet happen during IGVN.
1401 //
1402 // Found with:
1403 // test/hotspot/jtreg/compiler/c2/IVTest.java
1404 // -XX:VerifyIterativeGVN=1110
1405 case Op_SubI:
1406 return false;
1407
1408 // AddNode::IdealIL does transform like:
1409 // Convert x + (con - y) into "(x - y) + con"
1410 //
1411 // In IGVN before verification:
1412 // 8382 ConvI2L
1413 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1414 // 8383 SubL === _ 8381 8382
1415 // 8168 ConvI2L
1416 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1417 //
1418 // And then in verification:
1419 // 8424 AddL
1420 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1421 // 8383 SubL === _ 8016 8424
1422 // 8168 ConvI2L
1423 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1424 //
1425 // So the form changed from:
1426 // x + (ConvI2L(0) - [8382 ConvI2L])
1427 // to
1428 // x + (0 - [8424 AddL])
1429 // but the AddL was not added to the IGVN worklist. Investigate why.
1430 // There could be other issues, too. For example with "commute", see above.
1431 //
1432 // Found with:
1433 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1434 case Op_AddL:
1435 return false;
1436
1437 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1438 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1439 // record_for_cleanup(cmp, phase);
1440 // This verification code in the Ideal code creates new nodes, and checks
1441 // if they fold in unexpected ways. This means some nodes are created and
1442 // added to the worklist, even if the SubTypeCheck is not optimized. This
1443 // goes agains the assumption of the verification here, which assumes that
1444 // if the node is not optimized, then no new nodes should be created, and
1445 // also no nodes should be added to the worklist.
1446 // I see two options:
1447 // 1) forbid what verify_helper does, because for each Ideal call it
1448 // uses memory and that is suboptimal. But it is not clear how that
1449 // verification can be done otherwise.
1450 // 2) Special case the verification here. Probably the new nodes that
1451 // were just created are dead, i.e. they are not connected down to
1452 // root. We could verify that, and remove those nodes from the graph
1453 // by setting all their inputs to nullptr. And of course we would
1454 // have to remove those nodes from the worklist.
1455 // Maybe there are other options too, I did not dig much deeper yet.
1456 //
1457 // Found with:
1458 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1459 case Op_SubTypeCheck:
1460 return false;
1461
1462 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1463 // to other nodes: Conv, Add, Sub, Mul, And ...
1464 //
1465 // 107 ConI === 0 [[ ... ]] #int:2
1466 // 84 LoadRange === _ 7 83
1467 // 50 ConI === 0 [[ ... ]] #int:0
1468 // 549 LoopLimit === _ 50 84 107
1469 //
1470 // I stepped backward, to see how the node was generated, and I found that it was
1471 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1472 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1473 // that, and it seems we want to skip lowering it until after loop-opts, but never
1474 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1475 // could be other issues too.
1476 //
1477 // Fond with:
1478 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1479 case Op_LoopLimit:
1480 return false;
1481
1482 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1483 // "This optimization tries to find two or more inputs of phi with the same constant
1484 // value. It then splits them into a separate Phi, and according Region."
1485 //
1486 // Example:
1487 // 130 DecodeN === _ 129
1488 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1489 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1490 //
1491 // turns into:
1492 //
1493 // 50 ConP === 0 [[ 99 91 18 ]] #null
1494 // 130 DecodeN === _ 129 [[ 18 ]]
1495 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1496 //
1497 // We would have to investigate why this optimization does not happen during IGVN.
1498 // There could also be other issues - I did not investigate further yet.
1499 //
1500 // Found with:
1501 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1502 case Op_Phi:
1503 return false;
1504
1505 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1506 // For examle "The allocated object does not escape".
1507 //
1508 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1509 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1510 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1511 // worklist when this change happens?
1512 //
1513 // Found with:
1514 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1515 case Op_MemBarStoreStore:
1516 return false;
1517
1518 // ConvI2LNode::Ideal converts
1519 // 648 AddI === _ 583 645 [[ 661 ]]
1520 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1521 // into
1522 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1523 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1524 // 773 AddL === _ 771 772 [[ ]]
1525 //
1526 // We have to investigate why this does not happen during IGVN in this case.
1527 // There could also be other issues - I did not investigate further yet.
1528 //
1529 // Found with:
1530 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1531 case Op_ConvI2L:
1532 return false;
1533
1534 // AddNode::IdealIL can do this transform (and similar other ones):
1535 // Convert "a*b+a*c into a*(b+c)
1536 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1537 // during IGVN? There was a mutation for one of the MulI, and only
1538 // after that the pattern was as needed for the optimization. The MulI
1539 // was added to the IGVN worklist, but not the AddI. This probably
1540 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1541 //
1542 // Found with:
1543 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1544 // -XX:VerifyIterativeGVN=1110
1545 case Op_AddI:
1546 return false;
1547
1548 // ArrayCopyNode::Ideal
1549 // calls ArrayCopyNode::prepare_array_copy
1550 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1551 // is not expected, and we create a range int:0..-1
1552 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1553 // note: the type is already empty!
1554 // calls PhaseIterGVN::transform
1555 // calls PhaseIterGVN::transform_old
1556 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1557 // calls Unique_Node_List::push -> pushes TOP to worklist
1558 //
1559 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1560 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1561 //
1562 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1563 // return nullptr, indicating nothing happened?
1564 // Is it smart to do transform in Compile::constrained_convI2L, and then
1565 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1566 // Should we just allow TOP to land on the worklist, as an exception?
1567 //
1568 // Found with:
1569 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1570 // -XX:VerifyIterativeGVN=1110
1571 case Op_ArrayCopy:
1572 return false;
1573
1574 // CastLLNode::Ideal
1575 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1576 //
1577 // Could be a notification issue, where updates inputs of CastLL do not notify
1578 // down through SubL to CastLL.
1579 //
1580 // Found With:
1581 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1582 // -XX:VerifyIterativeGVN=1110
1583 case Op_CastLL:
1584 return false;
1585
1586 // Similar case happens to CastII
1587 //
1588 // Found With:
1589 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1590 // -XX:VerifyIterativeGVN=1110
1591 case Op_CastII:
1592 return false;
1593
1594 // MaxLNode::Ideal
1595 // calls AddNode::Ideal
1596 // calls commute -> decides to swap edges
1597 //
1598 // Another notification issue, because we check inputs of inputs?
1599 // MaxL -> Phi -> Loop
1600 // MaxL -> Phi -> MaxL
1601 //
1602 // Found with:
1603 // compiler/c2/irTests/TestIfMinMax.java
1604 // -XX:VerifyIterativeGVN=1110
1605 case Op_MaxL:
1606 case Op_MinL:
1607 return false;
1608
1609 // OrINode::Ideal
1610 // calls AddNode::Ideal
1611 // calls commute -> left is Load, right not -> commute.
1612 //
1613 // Not sure why notification does not work here, seems like
1614 // the depth is only 1, so it should work. Needs investigation.
1615 //
1616 // Found with:
1617 // compiler/codegen/TestCharVect2.java#id0
1618 // -XX:VerifyIterativeGVN=1110
1619 case Op_OrI:
1620 case Op_OrL:
1621 return false;
1622
1623 // Bool -> constant folded to 1.
1624 // Issue with notification?
1625 //
1626 // Found with:
1627 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1628 // -XX:VerifyIterativeGVN=1110
1629 case Op_Bool:
1630 return false;
1631
1632 // LShiftLNode::Ideal
1633 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1634 // Probably a notification issue.
1635 //
1636 // Found with:
1637 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1638 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1639 case Op_LShiftL:
1640 return false;
1641
1642 // LShiftINode::Ideal
1643 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1644 // Could be issue with notification of inputs of inputs
1645 //
1646 // Side-note: should cases like these not be shared between
1647 // LShiftI and LShiftL?
1648 //
1649 // Found with:
1650 // compiler/escapeAnalysis/Test6689060.java
1651 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1652 case Op_LShiftI:
1653 return false;
1654
1655 // AddPNode::Ideal seems to do set_req without removing lock first.
1656 // Found with various vector tests tier1-tier3.
1657 case Op_AddP:
1658 return false;
1659
1660 // StrIndexOfNode::Ideal
1661 // Found in tier1-3.
1662 case Op_StrIndexOf:
1663 case Op_StrIndexOfChar:
1664 return false;
1665
1666 // StrEqualsNode::Identity
1667 //
1668 // Found (linux x64 only?) with:
1669 // serviceability/sa/ClhsdbThreadContext.java
1670 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1671 // Note: The -XX:LockingMode option is not available anymore.
1672 case Op_StrEquals:
1673 return false;
1674
1675 // AryEqNode::Ideal
1676 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1677 //
1678 // Found with:
1679 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1680 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1681 case Op_AryEq:
1682 return false;
1683
1684 // MergeMemNode::Ideal
1685 // Found in tier1-3. Did not investigate further yet.
1686 case Op_MergeMem:
1687 return false;
1688
1689 // URShiftINode::Ideal
1690 // Found in tier1-3. Did not investigate further yet.
1691 case Op_URShiftI:
1692 return false;
1693
1694 // CMoveINode::Ideal
1695 // Found in tier1-3. Did not investigate further yet.
1696 case Op_CMoveI:
1697 return false;
1698
1699 // CmpPNode::Ideal calls isa_const_java_mirror
1700 // and generates new constant nodes, even if no progress is made.
1701 // We can probably rewrite this so that only types are generated.
1702 // It seems that object types are not hashed, we could investigate
1703 // if that is an option as well.
1704 //
1705 // Found with:
1706 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1707 case Op_CmpP:
1708 return false;
1709
1710 // MinINode::Ideal
1711 // Did not investigate, but there are some patterns that might
1712 // need more notification.
1713 case Op_MinI:
1714 case Op_MaxI: // preemptively removed it as well.
1715 return false;
1716 }
1717
1718 if (n->is_Load()) {
1719 // LoadNode::Ideal uses tries to find an earlier memory state, and
1720 // checks can_see_stored_value for it.
1721 //
1722 // Investigate why this was not already done during IGVN.
1723 // A similar issue happens with Identity.
1724 //
1725 // There seem to be other cases where loads go up some steps, like
1726 // LoadNode::Ideal going up 10x steps to find dominating load.
1727 //
1728 // Found with:
1729 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1730 // -XX:VerifyIterativeGVN=1110
1731 return false;
1732 }
1733
1734 if (n->is_Store()) {
1735 // StoreNode::Ideal can do this:
1736 // // Capture an unaliased, unconditional, simple store into an initializer.
1737 // // Or, if it is independent of the allocation, hoist it above the allocation.
1738 // That replaces the Store with a MergeMem.
1739 //
1740 // We have to investigate why this does not happen during IGVN in this case.
1741 // There could also be other issues - I did not investigate further yet.
1742 //
1743 // Found with:
1744 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1745 return false;
1746 }
1747
1748 if (n->is_Vector()) {
1749 // VectorNode::Ideal swaps edges, but only for ops
1750 // that are deemed commutable. But swap_edges
1751 // requires the hash to be invariant when the edges
1752 // are swapped, which is not implemented for these
1753 // vector nodes. This seems not to create any trouble
1754 // usually, but we can also get graphs where in the
1755 // end the nodes are not all commuted, so there is
1756 // definitively an issue here.
1757 //
1758 // Probably we have two options: kill the hash, or
1759 // properly make the hash commutation friendly.
1760 //
1761 // Found with:
1762 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1763 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1764 return false;
1765 }
1766
1767 if (n->is_Region()) {
1768 // LoopNode::Ideal calls RegionNode::Ideal.
1769 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1770 // But I got an issue because RegionNode::optimize_trichotomy
1771 // then modifies another node, and pushes nodes to the worklist
1772 // Not sure if this is ok, modifying another node like that.
1773 // Maybe it is, then we need to look into what to do with
1774 // the nodes that are now on the worklist, maybe just clear
1775 // them out again. But maybe modifying other nodes like that
1776 // is also bad design. In the end, we return nullptr for
1777 // the current CountedLoop. But the extra nodes on the worklist
1778 // trip the asserts later on.
1779 //
1780 // Found with:
1781 // compiler/eliminateAutobox/TestShortBoxing.java
1782 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1783 return false;
1784 }
1785
1786 if (n->is_CallJava()) {
1787 // CallStaticJavaNode::Ideal
1788 // Led to a crash:
1789 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1790 //
1791 // Did not investigate yet, could be a bug.
1792 // Or maybe it does not expect to be called during verification.
1793 //
1794 // Found with:
1795 // test/jdk/jdk/incubator/vector/VectorRuns.java
1796 // -XX:VerifyIterativeGVN=1110
1797
1798 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1799 // and possibly their subclasses.
1800 // During late inlining it can call CallJavaNode::register_for_late_inline
1801 // That means we do more rounds of late inlining, but might fail.
1802 // Then we do IGVN again, and register the node again for late inlining.
1803 // This creates an endless cycle. Everytime we try late inlining, we
1804 // are also creating more nodes, especially SafePoint and MergeMem.
1805 // These nodes are immediately rejected when the inlining fails in the
1806 // do_late_inline_check, but they still grow the memory, until we hit
1807 // the MemLimit and crash.
1808 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1809 // called repeatedly, and eventually we terminate. I fear this is not
1810 // a great assumption to make. We should investigate more.
1811 //
1812 // Found with:
1813 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1814 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1815 return false;
1816 }
1817
1818 // The number of nodes shoud not increase.
1819 uint old_unique = C->unique();
1820 // The hash of a node should not change, this would indicate different inputs
1821 uint old_hash = n->hash();
1822 Node* i = n->Ideal(this, can_reshape);
1823 // If there was no new Idealization, we are probably happy.
1824 if (i == nullptr) {
1825 if (old_unique < C->unique()) {
1826 stringStream ss; // Print as a block without tty lock.
1827 ss.cr();
1828 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1829 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1830 n->dump_bfs(1, nullptr, "", &ss);
1831 tty->print_cr("%s", ss.as_string());
1832 return true;
1833 }
1834
1835 if (old_hash != n->hash()) {
1836 stringStream ss; // Print as a block without tty lock.
1837 ss.cr();
1838 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1839 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1840 n->dump_bfs(1, nullptr, "", &ss);
1841 tty->print_cr("%s", ss.as_string());
1842 return true;
1843 }
1844
1845 verify_empty_worklist(n);
1846
1847 // Everything is good.
1848 return false;
1849 }
1850
1851 // We just saw a new Idealization which was not done during IGVN.
1852 stringStream ss; // Print as a block without tty lock.
1853 ss.cr();
1854 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1855 if (i == n) {
1856 ss.print_cr("The node was reshaped by Ideal.");
1857 } else {
1858 ss.print_cr("The node was replaced by Ideal.");
1859 ss.print_cr("Old node:");
1860 n->dump_bfs(1, nullptr, "", &ss);
1861 }
1862 ss.print_cr("The result after Ideal:");
1863 i->dump_bfs(1, nullptr, "", &ss);
1864 tty->print_cr("%s", ss.as_string());
1865 return true;
1866 }
1867
1868 // Check that all Identity optimizations that could be done were done.
1869 // Returns true if it found missed optimization opportunities and
1870 // false otherwise (no missed optimization, or skipped verification).
1871 bool PhaseIterGVN::verify_Identity_for(Node* n) {
1872 // First, we check a list of exceptions, where we skip verification,
1873 // because there are known cases where Ideal can optimize after IGVN.
1874 // Some may be expected and cannot be fixed, and others should be fixed.
1875 switch (n->Opcode()) {
1876 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1877 // after loopopts:
1878 // // Transforming long counted loops requires a safepoint node. Do not
1879 // // eliminate a safepoint until loop opts are over.
1880 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1881 //
1882 // I think the check for major_progress does delay it until after loopopts
1883 // but it does not ensure that the node is on the IGVN worklist after
1884 // loopopts. I think we should try to instead check for
1885 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1886 //
1887 // Found with:
1888 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1889 case Op_SafePoint:
1890 return false;
1891
1892 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1893 // does not record any other memory splits.
1894 //
1895 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1896 // never got called during IGVN for this node, investigate why.
1897 //
1898 // Found with:
1899 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1900 case Op_MergeMem:
1901 return false;
1902
1903 // ConstraintCastNode::Identity finds casts that are the same, except that
1904 // the control is "higher up", i.e. dominates. The call goes via
1905 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1906 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1907 // away from the cast, but within 100 idom steps, the cast may not be
1908 // put on the IGVN worklist any more.
1909 //
1910 // Found with:
1911 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1912 case Op_CastPP:
1913 case Op_CastII:
1914 case Op_CastLL:
1915 return false;
1916
1917 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1918 // checks dominator, which may be changed, but too far up for notification
1919 // to work.
1920 //
1921 // Found with:
1922 // compiler/c2/irTests/TestSkeletonPredicates.java
1923 // -XX:VerifyIterativeGVN=1110
1924 case Op_CheckCastPP:
1925 return false;
1926
1927 // In SubNode::Identity, we do:
1928 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1929 // In the example, the AddI had an input replaced, the AddI is
1930 // added to the IGVN worklist, but the SubI is one link further
1931 // down and is not added. I checked add_users_of_use_to_worklist
1932 // where I would expect the SubI would be added, and I cannot
1933 // find the pattern, only this one:
1934 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1935 //
1936 // Fix this "notification" issue and check if there are any other
1937 // issues.
1938 //
1939 // Found with:
1940 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1941 case Op_SubI:
1942 case Op_SubL:
1943 return false;
1944
1945 // PhiNode::Identity checks for patterns like:
1946 // r = (x != con) ? x : con;
1947 // that can be constant folded to "x".
1948 //
1949 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1950 // I suspect there was some earlier change to one of the inputs, but
1951 // not all relevant outputs were put on the IGVN worklist.
1952 //
1953 // Found with:
1954 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1955 // -XX:VerifyIterativeGVN=1110
1956 case Op_Phi:
1957 return false;
1958
1959 // ConvI2LNode::Identity does
1960 // convert I2L(L2I(x)) => x
1961 //
1962 // Investigate why this did not already happen during IGVN.
1963 //
1964 // Found with:
1965 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
1966 // -XX:VerifyIterativeGVN=1110
1967 case Op_ConvI2L:
1968 return false;
1969
1970 // MaxNode::find_identity_operation
1971 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
1972 // This can be a 2-hop search, so maybe notification is not
1973 // good enough.
1974 //
1975 // Found with:
1976 // compiler/codegen/TestBooleanVect.java
1977 // -XX:VerifyIterativeGVN=1110
1978 case Op_MaxL:
1979 case Op_MinL:
1980 case Op_MaxI:
1981 case Op_MinI:
1982 case Op_MaxF:
1983 case Op_MinF:
1984 case Op_MaxHF:
1985 case Op_MinHF:
1986 case Op_MaxD:
1987 case Op_MinD:
1988 return false;
1989
1990
1991 // AddINode::Identity
1992 // Converts (x-y)+y to x
1993 // Could be issue with notification
1994 //
1995 // Turns out AddL does the same.
1996 //
1997 // Found with:
1998 // compiler/c2/Test6792161.java
1999 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2000 case Op_AddI:
2001 case Op_AddL:
2002 return false;
2003
2004 // AbsINode::Identity
2005 // Not investigated yet.
2006 case Op_AbsI:
2007 return false;
2008 }
2009
2010 if (n->is_Load()) {
2011 // LoadNode::Identity tries to look for an earlier store value via
2012 // can_see_stored_value. I found an example where this led to
2013 // an Allocation, where we could assume the value was still zero.
2014 // So the LoadN can be replaced with a zerocon.
2015 //
2016 // Investigate why this was not already done during IGVN.
2017 // A similar issue happens with Ideal.
2018 //
2019 // Found with:
2020 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2021 return false;
2022 }
2023
2024 if (n->is_Store()) {
2025 // StoreNode::Identity
2026 // Not investigated, but found missing optimization for StoreI.
2027 // Looks like a StoreI is replaced with an InitializeNode.
2028 //
2029 // Found with:
2030 // applications/ctw/modules/java_base_2.java
2031 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2032 return false;
2033 }
2034
2035 if (n->is_Vector()) {
2036 // Found with tier1-3. Not investigated yet.
2037 // The observed issue was with AndVNode::Identity
2038 return false;
2039 }
2040
2041 Node* i = n->Identity(this);
2042 // If we cannot find any other Identity, we are happy.
2043 if (i == n) {
2044 verify_empty_worklist(n);
2045 return false;
2046 }
2047
2048 // The verification just found a new Identity that was not found during IGVN.
2049 stringStream ss; // Print as a block without tty lock.
2050 ss.cr();
2051 ss.print_cr("Missed Identity optimization:");
2052 ss.print_cr("Old node:");
2053 n->dump_bfs(1, nullptr, "", &ss);
2054 ss.print_cr("New node:");
2055 i->dump_bfs(1, nullptr, "", &ss);
2056 tty->print_cr("%s", ss.as_string());
2057 return true;
2058 }
2059
2060 // Some other verifications that are not specific to a particular transformation.
2061 bool PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2062 if (n->is_AddP()) {
2063 if (!n->as_AddP()->address_input_has_same_base()) {
2064 stringStream ss; // Print as a block without tty lock.
2065 ss.cr();
2066 ss.print_cr("Base pointers must match for AddP chain:");
2067 n->dump_bfs(2, nullptr, "", &ss);
2068 tty->print_cr("%s", ss.as_string());
2069 return true;
2070 }
2071 }
2072 return false;
2073 }
2074 #endif
2075
2076 /**
2077 * Register a new node with the optimizer. Update the types array, the def-use
2078 * info. Put on worklist.
2079 */
2080 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2081 set_type_bottom(n);
2082 _worklist.push(n);
2083 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2084 return n;
2085 }
2086
2087 //------------------------------transform--------------------------------------
2088 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2089 Node *PhaseIterGVN::transform( Node *n ) {
2090 if (_delay_transform) {
2091 // Register the node but don't optimize for now
2092 register_new_node_with_optimizer(n);
2093 return n;
2094 }
2095
2096 // If brand new node, make space in type array, and give it a type.
2097 ensure_type_or_null(n);
2098 if (type_or_null(n) == nullptr) {
2099 set_type_bottom(n);
2100 }
2101
2102 return transform_old(n);
2103 }
2104
2105 Node *PhaseIterGVN::transform_old(Node* n) {
2106 NOT_PRODUCT(set_transforms());
2107 // Remove 'n' from hash table in case it gets modified
2108 _table.hash_delete(n);
2109 #ifdef ASSERT
2110 if (is_verify_def_use()) {
2111 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2112 }
2113 #endif
2114
2115 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2116 if (n->is_Cmp()) {
2117 add_users_to_worklist(n);
2118 }
2119
2120 // Apply the Ideal call in a loop until it no longer applies
2121 Node* k = n;
2122 DEBUG_ONLY(dead_loop_check(k);)
2123 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2124 C->remove_modified_node(k);
2125 Node* i = apply_ideal(k, /*can_reshape=*/true);
2126 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2127 #ifndef PRODUCT
2128 verify_step(k);
2129 #endif
2130
2131 DEBUG_ONLY(uint loop_count = 1;)
2132 while (i != nullptr) {
2133 #ifdef ASSERT
2134 if (loop_count >= K + C->live_nodes()) {
2135 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2136 }
2137 #endif
2138 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2139 // Made a change; put users of original Node on worklist
2140 add_users_to_worklist(k);
2141 // Replacing root of transform tree?
2142 if (k != i) {
2143 // Make users of old Node now use new.
2144 subsume_node(k, i);
2145 k = i;
2146 }
2147 DEBUG_ONLY(dead_loop_check(k);)
2148 // Try idealizing again
2149 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2150 C->remove_modified_node(k);
2151 i = apply_ideal(k, /*can_reshape=*/true);
2152 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2153 #ifndef PRODUCT
2154 verify_step(k);
2155 #endif
2156 DEBUG_ONLY(loop_count++;)
2157 }
2158
2159 // If brand new node, make space in type array.
2160 ensure_type_or_null(k);
2161
2162 // See what kind of values 'k' takes on at runtime
2163 const Type* t = k->Value(this);
2164 assert(t != nullptr, "value sanity");
2165
2166 // Since I just called 'Value' to compute the set of run-time values
2167 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2168 // cache Value. Later requests for the local phase->type of this Node can
2169 // use the cached Value instead of suffering with 'bottom_type'.
2170 if (type_or_null(k) != t) {
2171 #ifndef PRODUCT
2172 inc_new_values();
2173 set_progress();
2174 #endif
2175 set_type(k, t);
2176 // If k is a TypeNode, capture any more-precise type permanently into Node
2177 k->raise_bottom_type(t);
2178 // Move users of node to worklist
2179 add_users_to_worklist(k);
2180 }
2181 // If 'k' computes a constant, replace it with a constant
2182 if (t->singleton() && !k->is_Con()) {
2183 NOT_PRODUCT(set_progress();)
2184 Node* con = makecon(t); // Make a constant
2185 add_users_to_worklist(k);
2186 subsume_node(k, con); // Everybody using k now uses con
2187 return con;
2188 }
2189
2190 // Now check for Identities
2191 i = k->Identity(this); // Look for a nearby replacement
2192 if (i != k) { // Found? Return replacement!
2193 NOT_PRODUCT(set_progress();)
2194 add_users_to_worklist(k);
2195 subsume_node(k, i); // Everybody using k now uses i
2196 return i;
2197 }
2198
2199 // Global Value Numbering
2200 i = hash_find_insert(k); // Check for pre-existing node
2201 if (i && (i != k)) {
2202 // Return the pre-existing node if it isn't dead
2203 NOT_PRODUCT(set_progress();)
2204 add_users_to_worklist(k);
2205 subsume_node(k, i); // Everybody using k now uses i
2206 return i;
2207 }
2208
2209 // Return Idealized original
2210 return k;
2211 }
2212
2213 //---------------------------------saturate------------------------------------
2214 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2215 const Type* limit_type) const {
2216 return new_type->narrow(old_type);
2217 }
2218
2219 //------------------------------remove_globally_dead_node----------------------
2220 // Kill a globally dead Node. All uses are also globally dead and are
2221 // aggressively trimmed.
2222 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2223 enum DeleteProgress {
2224 PROCESS_INPUTS,
2225 PROCESS_OUTPUTS
2226 };
2227 ResourceMark rm;
2228 Node_Stack stack(32);
2229 stack.push(dead, PROCESS_INPUTS);
2230
2231 while (stack.is_nonempty()) {
2232 dead = stack.node();
2233 if (dead->Opcode() == Op_SafePoint) {
2234 dead->as_SafePoint()->disconnect_from_root(this);
2235 }
2236 uint progress_state = stack.index();
2237 assert(dead != C->root(), "killing root, eh?");
2238 assert(!dead->is_top(), "add check for top when pushing");
2239 NOT_PRODUCT( set_progress(); )
2240 if (progress_state == PROCESS_INPUTS) {
2241 // After following inputs, continue to outputs
2242 stack.set_index(PROCESS_OUTPUTS);
2243 if (!dead->is_Con()) { // Don't kill cons but uses
2244 bool recurse = false;
2245 // Remove from hash table
2246 _table.hash_delete( dead );
2247 // Smash all inputs to 'dead', isolating him completely
2248 for (uint i = 0; i < dead->req(); i++) {
2249 Node *in = dead->in(i);
2250 if (in != nullptr && in != C->top()) { // Points to something?
2251 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2252 assert((nrep > 0), "sanity");
2253 if (in->outcnt() == 0) { // Made input go dead?
2254 stack.push(in, PROCESS_INPUTS); // Recursively remove
2255 recurse = true;
2256 } else if (in->outcnt() == 1 &&
2257 in->has_special_unique_user()) {
2258 _worklist.push(in->unique_out());
2259 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2260 if (in->Opcode() == Op_Region) {
2261 _worklist.push(in);
2262 } else if (in->is_Store()) {
2263 DUIterator_Fast imax, i = in->fast_outs(imax);
2264 _worklist.push(in->fast_out(i));
2265 i++;
2266 if (in->outcnt() == 2) {
2267 _worklist.push(in->fast_out(i));
2268 i++;
2269 }
2270 assert(!(i < imax), "sanity");
2271 }
2272 } else if (dead->is_data_proj_of_pure_function(in)) {
2273 _worklist.push(in);
2274 }
2275 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2276 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2277 // A Load that directly follows an InitializeNode is
2278 // going away. The Stores that follow are candidates
2279 // again to be captured by the InitializeNode.
2280 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2281 Node *n = in->fast_out(j);
2282 if (n->is_Store()) {
2283 _worklist.push(n);
2284 }
2285 }
2286 }
2287 } // if (in != nullptr && in != C->top())
2288 } // for (uint i = 0; i < dead->req(); i++)
2289 if (recurse) {
2290 continue;
2291 }
2292 } // if (!dead->is_Con())
2293 } // if (progress_state == PROCESS_INPUTS)
2294
2295 // Aggressively kill globally dead uses
2296 // (Rather than pushing all the outs at once, we push one at a time,
2297 // plus the parent to resume later, because of the indefinite number
2298 // of edge deletions per loop trip.)
2299 if (dead->outcnt() > 0) {
2300 // Recursively remove output edges
2301 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2302 } else {
2303 // Finished disconnecting all input and output edges.
2304 stack.pop();
2305 // Remove dead node from iterative worklist
2306 _worklist.remove(dead);
2307 C->remove_useless_node(dead);
2308 }
2309 } // while (stack.is_nonempty())
2310 }
2311
2312 //------------------------------subsume_node-----------------------------------
2313 // Remove users from node 'old' and add them to node 'nn'.
2314 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2315 if (old->Opcode() == Op_SafePoint) {
2316 old->as_SafePoint()->disconnect_from_root(this);
2317 }
2318 assert( old != hash_find(old), "should already been removed" );
2319 assert( old != C->top(), "cannot subsume top node");
2320 // Copy debug or profile information to the new version:
2321 C->copy_node_notes_to(nn, old);
2322 // Move users of node 'old' to node 'nn'
2323 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2324 Node* use = old->last_out(i); // for each use...
2325 // use might need re-hashing (but it won't if it's a new node)
2326 rehash_node_delayed(use);
2327 // Update use-def info as well
2328 // We remove all occurrences of old within use->in,
2329 // so as to avoid rehashing any node more than once.
2330 // The hash table probe swamps any outer loop overhead.
2331 uint num_edges = 0;
2332 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2333 if (use->in(j) == old) {
2334 use->set_req(j, nn);
2335 ++num_edges;
2336 }
2337 }
2338 i -= num_edges; // we deleted 1 or more copies of this edge
2339 }
2340
2341 // Search for instance field data PhiNodes in the same region pointing to the old
2342 // memory PhiNode and update their instance memory ids to point to the new node.
2343 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2344 Node* region = old->in(0);
2345 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2346 PhiNode* phi = region->fast_out(i)->isa_Phi();
2347 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2348 phi->set_inst_mem_id((int)nn->_idx);
2349 }
2350 }
2351 }
2352
2353 // Smash all inputs to 'old', isolating him completely
2354 Node *temp = new Node(1);
2355 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2356 remove_dead_node( old );
2357 temp->del_req(0); // Yank bogus edge
2358 if (nn != nullptr && nn->outcnt() == 0) {
2359 _worklist.push(nn);
2360 }
2361 #ifndef PRODUCT
2362 if (is_verify_def_use()) {
2363 for ( int i = 0; i < _verify_window_size; i++ ) {
2364 if ( _verify_window[i] == old )
2365 _verify_window[i] = nn;
2366 }
2367 }
2368 #endif
2369 temp->destruct(this); // reuse the _idx of this little guy
2370 }
2371
2372 //------------------------------add_users_to_worklist--------------------------
2373 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2374 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2375 worklist.push(n->fast_out(i)); // Push on worklist
2376 }
2377 }
2378
2379 // Return counted loop Phi if as a counted loop exit condition, cmp
2380 // compares the induction variable with n
2381 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2382 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2383 Node* bol = cmp->fast_out(i);
2384 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2385 Node* iff = bol->fast_out(i2);
2386 if (iff->is_BaseCountedLoopEnd()) {
2387 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2388 if (cle->limit() == n) {
2389 PhiNode* phi = cle->phi();
2390 if (phi != nullptr) {
2391 return phi;
2392 }
2393 }
2394 }
2395 }
2396 }
2397 return nullptr;
2398 }
2399
2400 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2401 add_users_to_worklist0(n, _worklist);
2402
2403 Unique_Node_List& worklist = _worklist;
2404 // Move users of node to worklist
2405 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2406 Node* use = n->fast_out(i); // Get use
2407 add_users_of_use_to_worklist(n, use, worklist);
2408 }
2409 }
2410
2411 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2412 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2413 use->is_Store() ) // Enable store/load same address
2414 add_users_to_worklist0(use, worklist);
2415
2416 // If we changed the receiver type to a call, we need to revisit
2417 // the Catch following the call. It's looking for a non-null
2418 // receiver to know when to enable the regular fall-through path
2419 // in addition to the NullPtrException path.
2420 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2421 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2422 if (p != nullptr) {
2423 add_users_to_worklist0(p, worklist);
2424 }
2425 }
2426
2427 uint use_op = use->Opcode();
2428 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2429 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2430 if (use->outcnt() > 0) {
2431 Node* bol = use->raw_out(0);
2432 if (bol->outcnt() > 0) {
2433 Node* iff = bol->raw_out(0);
2434 if (iff->outcnt() == 2) {
2435 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2436 // phi merging either 0 or 1 onto the worklist
2437 Node* ifproj0 = iff->raw_out(0);
2438 Node* ifproj1 = iff->raw_out(1);
2439 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2440 Node* region0 = ifproj0->raw_out(0);
2441 Node* region1 = ifproj1->raw_out(0);
2442 if( region0 == region1 )
2443 add_users_to_worklist0(region0, worklist);
2444 }
2445 }
2446 }
2447 }
2448 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2449 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2450 if (phi != nullptr) {
2451 // Input to the cmp of a loop exit check has changed, thus
2452 // the loop limit may have changed, which can then change the
2453 // range values of the trip-count Phi.
2454 worklist.push(phi);
2455 }
2456 }
2457 if (use_op == Op_CmpI) {
2458 Node* cmp = use;
2459 Node* in1 = cmp->in(1);
2460 Node* in2 = cmp->in(2);
2461 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2462 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2463 //
2464 // in1 in2 in1 in2
2465 // | | | |
2466 // +--- | --+ | |
2467 // | | | | |
2468 // CmpINode | CmpINode
2469 // | | |
2470 // BoolNode | BoolNode
2471 // | | OR |
2472 // IfNode | IfNode
2473 // | | |
2474 // IfProj | IfProj X
2475 // | | | |
2476 // CastIINode CastIINode
2477 //
2478 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2479 if (in1 == n) {
2480 // in1 modified -> could turn into X -> do traversal based on right pattern.
2481 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2482 Node* bol = cmp->fast_out(i2); // For each Bool
2483 if (bol->is_Bool()) {
2484 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2485 Node* iff = bol->fast_out(i3); // For each If
2486 if (iff->is_If()) {
2487 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2488 Node* if_proj = iff->fast_out(i4); // For each IfProj
2489 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2490 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2491 Node* castii = if_proj->fast_out(i5); // For each CastII
2492 if (castii->is_CastII() &&
2493 castii->as_CastII()->carry_dependency()) {
2494 worklist.push(castii);
2495 }
2496 }
2497 }
2498 }
2499 }
2500 }
2501 }
2502 } else {
2503 // Only in2 modified -> can assume X == in2 (left pattern).
2504 assert(n == in2, "only in2 modified");
2505 // Find all CastII with input in1.
2506 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2507 Node* castii = in1->fast_out(j);
2508 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2509 // Find If.
2510 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2511 Node* ifnode = castii->in(0)->in(0);
2512 // Check that if connects to the cmp
2513 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2514 worklist.push(castii);
2515 }
2516 }
2517 }
2518 }
2519 }
2520 }
2521 }
2522 }
2523
2524 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2525 // Patterns:
2526 // ConstraintCast+ -> Sub
2527 // ConstraintCast+ -> Phi
2528 // ConstraintCast+ -> Xor
2529 if (use->is_ConstraintCast()) {
2530 auto push_the_uses_to_worklist = [&](Node* n){
2531 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2532 worklist.push(n);
2533 }
2534 };
2535 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2536 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2537 }
2538 // If changed LShift inputs, check RShift users for useless sign-ext
2539 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2540 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2541 Node* u = use->fast_out(i2);
2542 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2543 worklist.push(u);
2544 }
2545 }
2546 // If changed LShift inputs, check And users for shift and mask (And) operation
2547 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2548 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2549 Node* u = use->fast_out(i2);
2550 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2551 worklist.push(u);
2552 }
2553 }
2554 }
2555 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2556 if (use_op == Op_AddI || use_op == Op_SubI) {
2557 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2558 Node* u = use->fast_out(i2);
2559 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2560 worklist.push(u);
2561 }
2562 }
2563 }
2564 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2565 if (use_op == Op_AndI || use_op == Op_AndL) {
2566 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2567 Node* u = use->fast_out(i2);
2568 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2569 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL) {
2570 worklist.push(u);
2571 }
2572 }
2573 }
2574 // Check for redundant conversion patterns:
2575 // ConvD2L->ConvL2D->ConvD2L
2576 // ConvF2I->ConvI2F->ConvF2I
2577 // ConvF2L->ConvL2F->ConvF2L
2578 // ConvI2F->ConvF2I->ConvI2F
2579 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2580 // are the only ones that are known to trigger missed optimizations otherwise
2581 if (use_op == Op_ConvL2D ||
2582 use_op == Op_ConvI2F ||
2583 use_op == Op_ConvL2F ||
2584 use_op == Op_ConvF2I) {
2585 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2586 Node* u = use->fast_out(i2);
2587 if ((use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2588 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2589 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2590 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F)) {
2591 worklist.push(u);
2592 }
2593 }
2594 }
2595 // If changed AddP inputs:
2596 // - check Stores for loop invariant, and
2597 // - if the changed input is the offset, check constant-offset AddP users for
2598 // address expression flattening.
2599 if (use_op == Op_AddP) {
2600 bool offset_changed = n == use->in(AddPNode::Offset);
2601 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2602 Node* u = use->fast_out(i2);
2603 if (u->is_Mem()) {
2604 worklist.push(u);
2605 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2606 worklist.push(u);
2607 }
2608 }
2609 }
2610 // Check for "abs(0-x)" into "abs(x)" conversion
2611 if (use->is_Sub()) {
2612 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2613 Node* u = use->fast_out(i2);
2614 if (u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2615 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI) {
2616 worklist.push(u);
2617 }
2618 }
2619 }
2620 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2621 add_users_to_worklist0(proj, worklist);
2622 };
2623 // If changed initialization activity, check dependent Stores
2624 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2625 InitializeNode* init = use->as_Allocate()->initialization();
2626 if (init != nullptr) {
2627 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2628 }
2629 }
2630 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2631 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2632 // to guarantee the change is not missed.
2633 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2634 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2635 if (p != nullptr) {
2636 add_users_to_worklist0(p, worklist);
2637 }
2638 }
2639
2640 if (use_op == Op_Initialize) {
2641 InitializeNode* init = use->as_Initialize();
2642 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2643 }
2644 // Loading the java mirror from a Klass requires two loads and the type
2645 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2646 // LoadP(LoadP(AddP(foo:Klass, #java_mirror)))
2647 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2648 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2649 Node* u = use->fast_out(i2);
2650 const Type* ut = u->bottom_type();
2651 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2652 worklist.push(u);
2653 }
2654 }
2655 }
2656 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2657 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2658 if (use->outcnt() == 1) {
2659 Node* cmp = use->unique_out();
2660 worklist.push(cmp);
2661 }
2662 }
2663
2664 // From CastX2PNode::Ideal
2665 // CastX2P(AddX(x, y))
2666 // CastX2P(SubX(x, y))
2667 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2668 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2669 Node* u = use->fast_out(i2);
2670 if (u->Opcode() == Op_CastX2P) {
2671 worklist.push(u);
2672 }
2673 }
2674 }
2675
2676 /* AndNode has a special handling when one of the operands is a LShiftNode:
2677 * (LHS << s) & RHS
2678 * if RHS fits in less than s bits, the value of this expression is 0.
2679 * The difficulty is that there might be a conversion node (ConvI2L) between
2680 * the LShiftINode and the AndLNode, like so:
2681 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2682 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2683 * (see `AndIL_min_trailing_zeros`).
2684 *
2685 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2686 * is not enough: there might be no update happening there. We need to
2687 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2688 *
2689 * Moreover we can have ConstraintCasts in between. It may look like
2690 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2691 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2692 * So we must deal with that as well.
2693 */
2694 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2695 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2696 auto push_and_to_worklist = [&worklist](Node* n){
2697 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2698 worklist.push(n);
2699 }
2700 };
2701 use->visit_uses(push_and_to_worklist, is_boundary);
2702 }
2703 }
2704
2705 /**
2706 * Remove the speculative part of all types that we know of
2707 */
2708 void PhaseIterGVN::remove_speculative_types() {
2709 assert(UseTypeSpeculation, "speculation is off");
2710 for (uint i = 0; i < _types.Size(); i++) {
2711 const Type* t = _types.fast_lookup(i);
2712 if (t != nullptr) {
2713 _types.map(i, t->remove_speculative());
2714 }
2715 }
2716 _table.check_no_speculative_types();
2717 }
2718
2719 // Check if the type of a divisor of a Div or Mod node includes zero.
2720 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2721 switch (n->Opcode()) {
2722 case Op_DivI:
2723 case Op_ModI:
2724 case Op_UDivI:
2725 case Op_UModI: {
2726 // Type of divisor includes 0?
2727 if (type(n->in(2)) == Type::TOP) {
2728 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2729 return false;
2730 }
2731 const TypeInt* type_divisor = type(n->in(2))->is_int();
2732 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2733 }
2734 case Op_DivL:
2735 case Op_ModL:
2736 case Op_UDivL:
2737 case Op_UModL: {
2738 // Type of divisor includes 0?
2739 if (type(n->in(2)) == Type::TOP) {
2740 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2741 return false;
2742 }
2743 const TypeLong* type_divisor = type(n->in(2))->is_long();
2744 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2745 }
2746 }
2747 return true;
2748 }
2749
2750 //=============================================================================
2751 #ifndef PRODUCT
2752 uint PhaseCCP::_total_invokes = 0;
2753 uint PhaseCCP::_total_constants = 0;
2754 #endif
2755 //------------------------------PhaseCCP---------------------------------------
2756 // Conditional Constant Propagation, ala Wegman & Zadeck
2757 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2758 NOT_PRODUCT( clear_constants(); )
2759 assert( _worklist.size() == 0, "" );
2760 analyze();
2761 }
2762
2763 #ifndef PRODUCT
2764 //------------------------------~PhaseCCP--------------------------------------
2765 PhaseCCP::~PhaseCCP() {
2766 inc_invokes();
2767 _total_constants += count_constants();
2768 }
2769 #endif
2770
2771
2772 #ifdef ASSERT
2773 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2774 if (tnew->meet(told) != tnew->remove_speculative()) {
2775 n->dump(1);
2776 tty->print("told = "); told->dump(); tty->cr();
2777 tty->print("tnew = "); tnew->dump(); tty->cr();
2778 fatal("Not monotonic");
2779 }
2780 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2781 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2782 }
2783 #endif //ASSERT
2784
2785 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2786 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2787 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2788 // or nodes even further down back to the worklist because their type could change as a result of the current type
2789 // change).
2790 void PhaseCCP::analyze() {
2791 // Initialize all types to TOP, optimistic analysis
2792 for (uint i = 0; i < C->unique(); i++) {
2793 _types.map(i, Type::TOP);
2794 }
2795
2796 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2797 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2798 // Compile is over. The local arena gets de-allocated at the end of its scope.
2799 ResourceArea local_arena(mtCompiler);
2800 Unique_Node_List worklist(&local_arena);
2801 Unique_Node_List worklist_revisit(&local_arena);
2802 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2803
2804 // Push root onto worklist
2805 worklist.push(C->root());
2806
2807 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2808 _root_and_safepoints.push(C->root());
2809
2810 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2811
2812 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2813 while (worklist.size() != 0) {
2814 Node* n = fetch_next_node(worklist);
2815 DEBUG_ONLY(worklist_verify.push(n);)
2816 if (needs_revisit(n)) {
2817 worklist_revisit.push(n);
2818 }
2819 if (n->is_SafePoint()) {
2820 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2821 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2822 _root_and_safepoints.push(n);
2823 }
2824 analyze_step(worklist, n);
2825 }
2826
2827 // More rounds to catch updates far in the graph.
2828 // Revisit nodes that might be able to refine their types at the end of the round.
2829 // If so, process these nodes. If there is remaining work, start another round.
2830 do {
2831 while (worklist.size() != 0) {
2832 Node* n = fetch_next_node(worklist);
2833 analyze_step(worklist, n);
2834 }
2835 for (uint t = 0; t < worklist_revisit.size(); t++) {
2836 Node* n = worklist_revisit.at(t);
2837 analyze_step(worklist, n);
2838 }
2839 } while (worklist.size() != 0);
2840
2841 DEBUG_ONLY(verify_analyze(worklist_verify);)
2842 }
2843
2844 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2845 const Type* new_type = n->Value(this);
2846 if (new_type != type(n)) {
2847 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2848 dump_type_and_node(n, new_type);
2849 set_type(n, new_type);
2850 push_child_nodes_to_worklist(worklist, n);
2851 }
2852 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2853 // Keep track of Type nodes to kill CFG paths that use Type
2854 // nodes that become dead.
2855 _maybe_top_type_nodes.push(n);
2856 }
2857 }
2858
2859 // Some nodes can refine their types due to type change somewhere deep
2860 // in the graph. We will need to revisit them before claiming convergence.
2861 // Add nodes here if particular *Node::Value is doing deep graph traversals
2862 // not handled by PhaseCCP::push_more_uses().
2863 bool PhaseCCP::needs_revisit(Node* n) const {
2864 // LoadNode performs deep traversals. Load is not notified for changes far away.
2865 if (n->is_Load()) {
2866 return true;
2867 }
2868 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2869 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2870 return true;
2871 }
2872 return false;
2873 }
2874
2875 #ifdef ASSERT
2876 // For every node n on verify list, check if type(n) == n->Value()
2877 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2878 // Therefore, we are verifying Value convergence strictly.
2879 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2880 bool failure = false;
2881 while (worklist_verify.size()) {
2882 Node* n = worklist_verify.pop();
2883 failure |= verify_Value_for(n, /* strict = */ true);
2884 }
2885 // If we get this assert, check why the reported nodes were not processed again in CCP.
2886 // We should either make sure that these nodes are properly added back to the CCP worklist
2887 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2888 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2889 // them at the end of the round.
2890 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound.");
2891 }
2892 #endif
2893
2894 // Fetch next node from worklist to be examined in this iteration.
2895 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2896 if (StressCCP) {
2897 return worklist.remove(C->random() % worklist.size());
2898 } else {
2899 return worklist.pop();
2900 }
2901 }
2902
2903 #ifndef PRODUCT
2904 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2905 if (TracePhaseCCP) {
2906 t->dump();
2907 do {
2908 tty->print("\t");
2909 } while (tty->position() < 16);
2910 n->dump();
2911 }
2912 }
2913 #endif
2914
2915 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2916 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2917 // of the new type of 'n'. Push these nodes to the worklist.
2918 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2919 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2920 Node* use = n->fast_out(i);
2921 push_if_not_bottom_type(worklist, use);
2922 push_more_uses(worklist, n, use);
2923 }
2924 }
2925
2926 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2927 if (n->bottom_type() != type(n)) {
2928 worklist.push(n);
2929 }
2930 }
2931
2932 // For some nodes, we need to propagate the type change to grandchildren or even further down.
2933 // Add them back to the worklist.
2934 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
2935 push_phis(worklist, use);
2936 push_catch(worklist, use);
2937 push_cmpu(worklist, use);
2938 push_counted_loop_phi(worklist, parent, use);
2939 push_loadp(worklist, use);
2940 push_and(worklist, parent, use);
2941 push_cast_ii(worklist, parent, use);
2942 push_opaque_zero_trip_guard(worklist, use);
2943 push_bool_with_cmpu_and_mask(worklist, use);
2944 }
2945
2946
2947 // We must recheck Phis too if use is a Region.
2948 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
2949 if (use->is_Region()) {
2950 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2951 push_if_not_bottom_type(worklist, use->fast_out(i));
2952 }
2953 }
2954 }
2955
2956 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
2957 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
2958 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
2959 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
2960 if (use->is_Call()) {
2961 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2962 Node* proj = use->fast_out(i);
2963 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
2964 Node* catch_node = proj->find_out_with(Op_Catch);
2965 if (catch_node != nullptr) {
2966 worklist.push(catch_node);
2967 }
2968 }
2969 }
2970 }
2971 }
2972
2973 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
2974 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
2975 // and get wrong types otherwise.
2976 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
2977 uint use_op = use->Opcode();
2978 if (use_op == Op_AddI || use_op == Op_SubI) {
2979 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2980 Node* cmpu = use->fast_out(i);
2981 const uint cmpu_opcode = cmpu->Opcode();
2982 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
2983 // Got a CmpU or CmpU3 which might need the new type information from node n.
2984 push_if_not_bottom_type(worklist, cmpu);
2985 }
2986 }
2987 }
2988 }
2989
2990 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
2991 // (1b): "(m & x) <u (m + 1))".
2992 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
2993 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
2994 //
2995 // m x m 1 (%%)
2996 // \ / \ /
2997 // AndI AddI
2998 // \ /
2999 // CmpU
3000 // |
3001 // Bool
3002 //
3003 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3004 uint use_op = use->Opcode();
3005 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3006 // Not "m & x" or "m + 1"
3007 return;
3008 }
3009 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3010 Node* cmpu = use->fast_out(i);
3011 if (cmpu->Opcode() == Op_CmpU) {
3012 push_bool_matching_case1b(worklist, cmpu);
3013 }
3014 }
3015 }
3016
3017 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3018 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3019 assert(cmpu->Opcode() == Op_CmpU, "must be");
3020 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3021 Node* bol = cmpu->fast_out(i);
3022 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3023 // Not a Bool with "<u"
3024 continue;
3025 }
3026 Node* andI = cmpu->in(1);
3027 Node* addI = cmpu->in(2);
3028 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3029 // Not "m & x" and "m + 1"
3030 continue;
3031 }
3032
3033 Node* m = addI->in(1);
3034 if (m == andI->in(1) || m == andI->in(2)) {
3035 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3036 push_if_not_bottom_type(worklist, bol);
3037 }
3038 }
3039 }
3040
3041 // 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'.
3042 // Seem PhiNode::Value().
3043 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3044 uint use_op = use->Opcode();
3045 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3046 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3047 if (phi != nullptr) {
3048 worklist.push(phi);
3049 }
3050 }
3051 }
3052
3053 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3054 // See LoadNode::Value().
3055 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3056 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3057 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3058 Node* loadp = use->fast_out(i);
3059 const Type* ut = loadp->bottom_type();
3060 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3061 worklist.push(loadp);
3062 }
3063 }
3064 }
3065 }
3066
3067 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3068 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3069 Node* barrier_node = use->fast_out(i);
3070 if (barrier_set->is_gc_barrier_node(barrier_node)) {
3071 worklist.push(barrier_node);
3072 }
3073 }
3074 }
3075
3076 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3077 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3078 // value changed.
3079 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3080 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3081 uint use_op = use->Opcode();
3082 if (
3083 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3084 (parent_type != nullptr && parent_type->is_con()) ||
3085 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3086 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3087
3088 auto push_and_uses_to_worklist = [&](Node* n) {
3089 uint opc = n->Opcode();
3090 if (opc == Op_AndI || opc == Op_AndL) {
3091 push_if_not_bottom_type(worklist, n);
3092 }
3093 };
3094 auto is_boundary = [](Node* n) {
3095 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3096 };
3097 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3098 }
3099 }
3100
3101 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3102 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3103 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3104 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3105 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3106 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3107 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3108 Node* other_cmp_input = use->in(1);
3109 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3110 Node* cast_ii = other_cmp_input->fast_out(i);
3111 if (cast_ii->is_CastII()) {
3112 push_if_not_bottom_type(worklist, cast_ii);
3113 }
3114 }
3115 }
3116 }
3117
3118 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3119 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3120 push_if_not_bottom_type(worklist, use->unique_out());
3121 }
3122 }
3123
3124 //------------------------------do_transform-----------------------------------
3125 // Top level driver for the recursive transformer
3126 void PhaseCCP::do_transform() {
3127 // Correct leaves of new-space Nodes; they point to old-space.
3128 C->set_root( transform(C->root())->as_Root() );
3129 assert( C->top(), "missing TOP node" );
3130 assert( C->root(), "missing root" );
3131 }
3132
3133 //------------------------------transform--------------------------------------
3134 // Given a Node in old-space, clone him into new-space.
3135 // Convert any of his old-space children into new-space children.
3136 Node *PhaseCCP::transform( Node *n ) {
3137 assert(n->is_Root(), "traversal must start at root");
3138 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3139
3140 ResourceMark rm;
3141 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3142 Node_List node_map;
3143 // Pre-allocate to avoid frequent realloc
3144 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3145 // track all visited nodes, so that we can remove the complement
3146 Unique_Node_List useful;
3147
3148 if (KillPathsReachableByDeadTypeNode) {
3149 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3150 Node* type_node = _maybe_top_type_nodes.at(i);
3151 if (type(type_node) == Type::TOP) {
3152 ResourceMark rm;
3153 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3154 }
3155 }
3156 } else {
3157 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3158 }
3159
3160 // Initialize the traversal.
3161 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3162 // 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
3163 // 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
3164 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3165 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3166 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3167 // and all safepoints.
3168 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3169 Node* nn = _root_and_safepoints.at(i);
3170 Node* new_node = node_map[nn->_idx];
3171 assert(new_node == nullptr, "");
3172 new_node = transform_once(nn); // Check for constant
3173 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3174 transform_stack.push(new_node); // Process children of cloned node
3175 useful.push(new_node);
3176 }
3177
3178 while (transform_stack.is_nonempty()) {
3179 Node* clone = transform_stack.pop();
3180 uint cnt = clone->req();
3181 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3182 Node *input = clone->in(i);
3183 if( input != nullptr ) { // Ignore nulls
3184 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3185 if( new_input == nullptr ) {
3186 new_input = transform_once(input); // Check for constant
3187 node_map.map( input->_idx, new_input );// Flag as having been cloned
3188 transform_stack.push(new_input); // Process children of cloned node
3189 useful.push(new_input);
3190 }
3191 assert( new_input == clone->in(i), "insanity check");
3192 }
3193 }
3194 }
3195
3196 // The above transformation might lead to subgraphs becoming unreachable from the
3197 // bottom while still being reachable from the top. As a result, nodes in that
3198 // subgraph are not transformed and their bottom types are not updated, leading to
3199 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3200 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3201 // because their address type is inconsistent. Therefore, we aggressively remove
3202 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3203 // to remove them anyway.
3204 if (C->cached_top_node()) {
3205 useful.push(C->cached_top_node());
3206 }
3207 C->update_dead_node_list(useful);
3208 remove_useless_nodes(useful.member_set());
3209 _worklist.remove_useless_nodes(useful.member_set());
3210 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3211
3212 Node* new_root = node_map[n->_idx];
3213 assert(new_root->is_Root(), "transformed root node must be a root node");
3214 return new_root;
3215 }
3216
3217 //------------------------------transform_once---------------------------------
3218 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3219 Node *PhaseCCP::transform_once( Node *n ) {
3220 const Type *t = type(n);
3221 // Constant? Use constant Node instead
3222 if( t->singleton() ) {
3223 Node *nn = n; // Default is to return the original constant
3224 if( t == Type::TOP ) {
3225 // cache my top node on the Compile instance
3226 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3227 C->set_cached_top_node(ConNode::make(Type::TOP));
3228 set_type(C->top(), Type::TOP);
3229 }
3230 nn = C->top();
3231 }
3232 if( !n->is_Con() ) {
3233 if( t != Type::TOP ) {
3234 nn = makecon(t); // ConNode::make(t);
3235 NOT_PRODUCT( inc_constants(); )
3236 } else if( n->is_Region() ) { // Unreachable region
3237 // Note: nn == C->top()
3238 n->set_req(0, nullptr); // Cut selfreference
3239 bool progress = true;
3240 uint max = n->outcnt();
3241 DUIterator i;
3242 while (progress) {
3243 progress = false;
3244 // Eagerly remove dead phis to avoid phis copies creation.
3245 for (i = n->outs(); n->has_out(i); i++) {
3246 Node* m = n->out(i);
3247 if (m->is_Phi()) {
3248 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3249 replace_node(m, nn);
3250 if (max != n->outcnt()) {
3251 progress = true;
3252 i = n->refresh_out_pos(i);
3253 max = n->outcnt();
3254 }
3255 }
3256 }
3257 }
3258 }
3259 replace_node(n,nn); // Update DefUse edges for new constant
3260 }
3261 return nn;
3262 }
3263
3264 // If x is a TypeNode, capture any more-precise type permanently into Node
3265 if (t != n->bottom_type()) {
3266 hash_delete(n); // changing bottom type may force a rehash
3267 n->raise_bottom_type(t);
3268 _worklist.push(n); // n re-enters the hash table via the worklist
3269 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3270 }
3271
3272 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3273 switch( n->Opcode() ) {
3274 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3275 case Op_CallDynamicJava:
3276 case Op_FastLock: // Revisit FastLocks for lock coarsening
3277 case Op_If:
3278 case Op_CountedLoopEnd:
3279 case Op_Region:
3280 case Op_Loop:
3281 case Op_CountedLoop:
3282 case Op_Conv2B:
3283 case Op_Opaque1:
3284 _worklist.push(n);
3285 break;
3286 default:
3287 break;
3288 }
3289
3290 return n;
3291 }
3292
3293 //---------------------------------saturate------------------------------------
3294 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3295 const Type* limit_type) const {
3296 const Type* wide_type = new_type->widen(old_type, limit_type);
3297 if (wide_type != new_type) { // did we widen?
3298 // If so, we may have widened beyond the limit type. Clip it back down.
3299 new_type = wide_type->filter(limit_type);
3300 }
3301 return new_type;
3302 }
3303
3304 //------------------------------print_statistics-------------------------------
3305 #ifndef PRODUCT
3306 void PhaseCCP::print_statistics() {
3307 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3308 }
3309 #endif
3310
3311
3312 //=============================================================================
3313 #ifndef PRODUCT
3314 uint PhasePeephole::_total_peepholes = 0;
3315 #endif
3316 //------------------------------PhasePeephole----------------------------------
3317 // Conditional Constant Propagation, ala Wegman & Zadeck
3318 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3319 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3320 NOT_PRODUCT( clear_peepholes(); )
3321 }
3322
3323 #ifndef PRODUCT
3324 //------------------------------~PhasePeephole---------------------------------
3325 PhasePeephole::~PhasePeephole() {
3326 _total_peepholes += count_peepholes();
3327 }
3328 #endif
3329
3330 //------------------------------transform--------------------------------------
3331 Node *PhasePeephole::transform( Node *n ) {
3332 ShouldNotCallThis();
3333 return nullptr;
3334 }
3335
3336 //------------------------------do_transform-----------------------------------
3337 void PhasePeephole::do_transform() {
3338 bool method_name_not_printed = true;
3339
3340 // Examine each basic block
3341 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3342 Block* block = _cfg.get_block(block_number);
3343 bool block_not_printed = true;
3344
3345 for (bool progress = true; progress;) {
3346 progress = false;
3347 // block->end_idx() not valid after PhaseRegAlloc
3348 uint end_index = block->number_of_nodes();
3349 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3350 Node *n = block->get_node(instruction_index);
3351 if( n->is_Mach() ) {
3352 MachNode *m = n->as_Mach();
3353 // check for peephole opportunities
3354 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3355 if( result != -1 ) {
3356 #ifndef PRODUCT
3357 if( PrintOptoPeephole ) {
3358 // Print method, first time only
3359 if( C->method() && method_name_not_printed ) {
3360 C->method()->print_short_name(); tty->cr();
3361 method_name_not_printed = false;
3362 }
3363 // Print this block
3364 if( Verbose && block_not_printed) {
3365 tty->print_cr("in block");
3366 block->dump();
3367 block_not_printed = false;
3368 }
3369 // Print the peephole number
3370 tty->print_cr("peephole number: %d", result);
3371 }
3372 inc_peepholes();
3373 #endif
3374 // Set progress, start again
3375 progress = true;
3376 break;
3377 }
3378 }
3379 }
3380 }
3381 }
3382 }
3383
3384 //------------------------------print_statistics-------------------------------
3385 #ifndef PRODUCT
3386 void PhasePeephole::print_statistics() {
3387 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3388 }
3389 #endif
3390
3391
3392 //=============================================================================
3393 //------------------------------set_req_X--------------------------------------
3394 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3395 assert( is_not_dead(n), "can not use dead node");
3396 #ifdef ASSERT
3397 if (igvn->hash_find(this) == this) {
3398 tty->print_cr("Need to remove from hash before changing edges");
3399 this->dump(1);
3400 tty->print_cr("Set at i = %d", i);
3401 n->dump();
3402 assert(false, "Need to remove from hash before changing edges");
3403 }
3404 #endif
3405 Node *old = in(i);
3406 set_req(i, n);
3407
3408 // old goes dead?
3409 if( old ) {
3410 switch (old->outcnt()) {
3411 case 0:
3412 // Put into the worklist to kill later. We do not kill it now because the
3413 // recursive kill will delete the current node (this) if dead-loop exists
3414 if (!old->is_top())
3415 igvn->_worklist.push( old );
3416 break;
3417 case 1:
3418 if( old->is_Store() || old->has_special_unique_user() )
3419 igvn->add_users_to_worklist( old );
3420 break;
3421 case 2:
3422 if( old->is_Store() )
3423 igvn->add_users_to_worklist( old );
3424 if( old->Opcode() == Op_Region )
3425 igvn->_worklist.push(old);
3426 break;
3427 case 3:
3428 if( old->Opcode() == Op_Region ) {
3429 igvn->_worklist.push(old);
3430 igvn->add_users_to_worklist( old );
3431 }
3432 break;
3433 default:
3434 break;
3435 }
3436 }
3437 }
3438
3439 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3440 PhaseIterGVN* igvn = gvn->is_IterGVN();
3441 if (igvn == nullptr) {
3442 set_req(i, n);
3443 return;
3444 }
3445 set_req_X(i, n, igvn);
3446 }
3447
3448 //-------------------------------replace_by-----------------------------------
3449 // Using def-use info, replace one node for another. Follow the def-use info
3450 // to all users of the OLD node. Then make all uses point to the NEW node.
3451 void Node::replace_by(Node *new_node) {
3452 assert(!is_top(), "top node has no DU info");
3453 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3454 Node* use = last_out(i);
3455 uint uses_found = 0;
3456 for (uint j = 0; j < use->len(); j++) {
3457 if (use->in(j) == this) {
3458 if (j < use->req())
3459 use->set_req(j, new_node);
3460 else use->set_prec(j, new_node);
3461 uses_found++;
3462 }
3463 }
3464 i -= uses_found; // we deleted 1 or more copies of this edge
3465 }
3466 }
3467
3468 //=============================================================================
3469 //-----------------------------------------------------------------------------
3470 void Type_Array::grow( uint i ) {
3471 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3472 if( !_max ) {
3473 _max = 1;
3474 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3475 _types[0] = nullptr;
3476 }
3477 uint old = _max;
3478 _max = next_power_of_2(i);
3479 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3480 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3481 }
3482
3483 //------------------------------dump-------------------------------------------
3484 #ifndef PRODUCT
3485 void Type_Array::dump() const {
3486 uint max = Size();
3487 for( uint i = 0; i < max; i++ ) {
3488 if( _types[i] != nullptr ) {
3489 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3490 }
3491 }
3492 }
3493 #endif