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 PhaseIterGVN* PhaseValues::is_IterGVN() {
536 return (_phase == PhaseValuesType::iter_gvn || _phase == PhaseValuesType::ccp) ? static_cast<PhaseIterGVN*>(this) : nullptr;
537 }
538
539 //--------------------------------find_int_type--------------------------------
540 const TypeInt* PhaseValues::find_int_type(Node* n) {
541 if (n == nullptr) return nullptr;
542 // Call type_or_null(n) to determine node's type since we might be in
543 // parse phase and call n->Value() may return wrong type.
544 // (For example, a phi node at the beginning of loop parsing is not ready.)
545 const Type* t = type_or_null(n);
546 if (t == nullptr) return nullptr;
547 return t->isa_int();
548 }
549
550
551 //-------------------------------find_long_type--------------------------------
552 const TypeLong* PhaseValues::find_long_type(Node* n) {
553 if (n == nullptr) return nullptr;
554 // (See comment above on type_or_null.)
555 const Type* t = type_or_null(n);
556 if (t == nullptr) return nullptr;
557 return t->isa_long();
558 }
559
560 //------------------------------~PhaseValues-----------------------------------
561 #ifndef PRODUCT
562 PhaseValues::~PhaseValues() {
563 // Statistics for NodeHash
564 _table.dump();
565 // Statistics for value progress and efficiency
566 if( PrintCompilation && Verbose && WizardMode ) {
567 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
568 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
569 if( made_transforms() != 0 ) {
570 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
571 } else {
572 tty->cr();
573 }
574 }
575 }
576 #endif
577
578 //------------------------------makecon----------------------------------------
579 ConNode* PhaseValues::makecon(const Type* t) {
580 assert(t->singleton(), "must be a constant");
581 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
582 switch (t->base()) { // fast paths
583 case Type::Half:
584 case Type::Top: return (ConNode*) C->top();
585 case Type::Int: return intcon( t->is_int()->get_con() );
586 case Type::Long: return longcon( t->is_long()->get_con() );
587 default: break;
588 }
589 if (t->is_zero_type())
590 return zerocon(t->basic_type());
591 return uncached_makecon(t);
592 }
593
594 //--------------------------uncached_makecon-----------------------------------
595 // Make an idealized constant - one of ConINode, ConPNode, etc.
596 ConNode* PhaseValues::uncached_makecon(const Type *t) {
597 assert(t->singleton(), "must be a constant");
598 ConNode* x = ConNode::make(t);
599 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
600 if (k == nullptr) {
601 set_type(x, t); // Missed, provide type mapping
602 GrowableArray<Node_Notes*>* nna = C->node_note_array();
603 if (nna != nullptr) {
604 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
605 loc->clear(); // do not put debug info on constants
606 }
607 } else {
608 x->destruct(this); // Hit, destroy duplicate constant
609 x = k; // use existing constant
610 }
611 return x;
612 }
613
614 //------------------------------intcon-----------------------------------------
615 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
616 ConINode* PhaseValues::intcon(jint i) {
617 // Small integer? Check cache! Check that cached node is not dead
618 if (i >= _icon_min && i <= _icon_max) {
619 ConINode* icon = _icons[i-_icon_min];
620 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
621 return icon;
622 }
623 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
624 assert(icon->is_Con(), "");
625 if (i >= _icon_min && i <= _icon_max)
626 _icons[i-_icon_min] = icon; // Cache small integers
627 return icon;
628 }
629
630 //------------------------------longcon----------------------------------------
631 // Fast long constant.
632 ConLNode* PhaseValues::longcon(jlong l) {
633 // Small integer? Check cache! Check that cached node is not dead
634 if (l >= _lcon_min && l <= _lcon_max) {
635 ConLNode* lcon = _lcons[l-_lcon_min];
636 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
637 return lcon;
638 }
639 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
640 assert(lcon->is_Con(), "");
641 if (l >= _lcon_min && l <= _lcon_max)
642 _lcons[l-_lcon_min] = lcon; // Cache small integers
643 return lcon;
644 }
645 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
646 if (bt == T_INT) {
647 return intcon(checked_cast<jint>(l));
648 }
649 assert(bt == T_LONG, "not an integer");
650 return longcon(l);
651 }
652
653
654 //------------------------------zerocon-----------------------------------------
655 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
656 ConNode* PhaseValues::zerocon(BasicType bt) {
657 assert((uint)bt <= _zcon_max, "domain check");
658 ConNode* zcon = _zcons[bt];
659 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
660 return zcon;
661 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
662 _zcons[bt] = zcon;
663 return zcon;
664 }
665
666
667
668 //=============================================================================
669 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
670 return k->Ideal(this, can_reshape);
671 }
672
673 //------------------------------transform--------------------------------------
674 // Return a node which computes the same function as this node, but
675 // in a faster or cheaper fashion.
676 Node* PhaseGVN::transform(Node* n) {
677 NOT_PRODUCT( set_transforms(); )
678
679 // Apply the Ideal call in a loop until it no longer applies
680 Node* k = n;
681 Node* i = apply_ideal(k, /*can_reshape=*/false);
682 NOT_PRODUCT(uint loop_count = 1;)
683 while (i != nullptr) {
684 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
685 k = i;
686 #ifdef ASSERT
687 if (loop_count >= K + C->live_nodes()) {
688 dump_infinite_loop_info(i, "PhaseGVN::transform");
689 }
690 #endif
691 i = apply_ideal(k, /*can_reshape=*/false);
692 NOT_PRODUCT(loop_count++;)
693 }
694 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
695
696 // If brand new node, make space in type array.
697 ensure_type_or_null(k);
698
699 // Since I just called 'Value' to compute the set of run-time values
700 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
701 // cache Value. Later requests for the local phase->type of this Node can
702 // use the cached Value instead of suffering with 'bottom_type'.
703 const Type* t = k->Value(this); // Get runtime Value set
704 assert(t != nullptr, "value sanity");
705 if (type_or_null(k) != t) {
706 #ifndef PRODUCT
707 // Do not count initial visit to node as a transformation
708 if (type_or_null(k) == nullptr) {
709 inc_new_values();
710 set_progress();
711 }
712 #endif
713 set_type(k, t);
714 // If k is a TypeNode, capture any more-precise type permanently into Node
715 k->raise_bottom_type(t);
716 }
717
718 if (t->singleton() && !k->is_Con()) {
719 NOT_PRODUCT(set_progress();)
720 return makecon(t); // Turn into a constant
721 }
722
723 // Now check for Identities
724 i = k->Identity(this); // Look for a nearby replacement
725 if (i != k) { // Found? Return replacement!
726 NOT_PRODUCT(set_progress();)
727 return i;
728 }
729
730 // Global Value Numbering
731 i = hash_find_insert(k); // Insert if new
732 if (i && (i != k)) {
733 // Return the pre-existing node
734 NOT_PRODUCT(set_progress();)
735 return i;
736 }
737
738 // Return Idealized original
739 return k;
740 }
741
742 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
743 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
744 return false;
745 }
746 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
747 return false;
748 }
749 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
750 int i = 0;
751 while (d != n) {
752 n = IfNode::up_one_dom(n, linear_only);
753 i++;
754 if (n == nullptr || i >= 100) {
755 return false;
756 }
757 }
758 return true;
759 }
760
761 #ifdef ASSERT
762 //------------------------------dead_loop_check--------------------------------
763 // Check for a simple dead loop when a data node references itself directly
764 // or through an other data node excluding cons and phis.
765 void PhaseGVN::dead_loop_check(Node* n) {
766 // Phi may reference itself in a loop.
767 if (n == nullptr || n->is_dead_loop_safe() || n->is_CFG()) {
768 return;
769 }
770
771 // Do 2 levels check and only data inputs.
772 for (uint i = 1; i < n->req(); i++) {
773 Node* in = n->in(i);
774 if (in == n) {
775 n->dump_bfs(100, nullptr, "");
776 fatal("Dead loop detected, node references itself: %s (%d)",
777 n->Name(), n->_idx);
778 }
779
780 if (in == nullptr || in->is_dead_loop_safe()) {
781 continue;
782 }
783 for (uint j = 1; j < in->req(); j++) {
784 if (in->in(j) == n) {
785 n->dump_bfs(100, nullptr, "");
786 fatal("Dead loop detected, node input references current node: %s (%d) -> %s (%d)",
787 in->Name(), in->_idx, n->Name(), n->_idx);
788 }
789 if (in->in(j) == in) {
790 n->dump_bfs(100, nullptr, "");
791 fatal("Dead loop detected, node input references itself: %s (%d)",
792 in->Name(), in->_idx);
793 }
794 }
795 }
796 }
797
798
799 /**
800 * Dumps information that can help to debug the problem. A debug
801 * build fails with an assert.
802 */
803 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
804 n->dump(4);
805 assert(false, "infinite loop in %s", where);
806 }
807 #endif
808
809 //=============================================================================
810 //------------------------------PhaseIterGVN-----------------------------------
811 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
812 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
813 _worklist(*C->igvn_worklist())
814 {
815 _phase = PhaseValuesType::iter_gvn;
816 assert(&_worklist == &igvn->_worklist, "sanity");
817 }
818
819 //------------------------------PhaseIterGVN-----------------------------------
820 // Initialize from scratch
821 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
822 _worklist(*C->igvn_worklist())
823 {
824 _phase = PhaseValuesType::iter_gvn;
825 uint max;
826
827 // Dead nodes in the hash table inherited from GVN were not treated as
828 // roots during def-use info creation; hence they represent an invisible
829 // use. Clear them out.
830 max = _table.size();
831 for( uint i = 0; i < max; ++i ) {
832 Node *n = _table.at(i);
833 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
834 if( n->is_top() ) continue;
835 // If remove_useless_nodes() has run, we expect no such nodes left.
836 assert(false, "remove_useless_nodes missed this node");
837 hash_delete(n);
838 }
839 }
840
841 // Any Phis or Regions on the worklist probably had uses that could not
842 // make more progress because the uses were made while the Phis and Regions
843 // were in half-built states. Put all uses of Phis and Regions on worklist.
844 max = _worklist.size();
845 for( uint j = 0; j < max; j++ ) {
846 Node *n = _worklist.at(j);
847 uint uop = n->Opcode();
848 if( uop == Op_Phi || uop == Op_Region ||
849 n->is_Type() ||
850 n->is_Mem() )
851 add_users_to_worklist(n);
852 }
853 }
854
855 void PhaseIterGVN::shuffle_worklist() {
856 if (_worklist.size() < 2) return;
857 for (uint i = _worklist.size() - 1; i >= 1; i--) {
858 uint j = C->random() % (i + 1);
859 swap(_worklist.adr()[i], _worklist.adr()[j]);
860 }
861 }
862
863 #ifndef PRODUCT
864 void PhaseIterGVN::verify_step(Node* n) {
865 if (is_verify_def_use()) {
866 ResourceMark rm;
867 VectorSet visited;
868 Node_List worklist;
869
870 _verify_window[_verify_counter % _verify_window_size] = n;
871 ++_verify_counter;
872 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
873 ++_verify_full_passes;
874 worklist.push(C->root());
875 Node::verify(-1, visited, worklist);
876 return;
877 }
878 for (int i = 0; i < _verify_window_size; i++) {
879 Node* n = _verify_window[i];
880 if (n == nullptr) {
881 continue;
882 }
883 if (n->in(0) == NodeSentinel) { // xform_idom
884 _verify_window[i] = n->in(1);
885 --i;
886 continue;
887 }
888 // Typical fanout is 1-2, so this call visits about 6 nodes.
889 if (!visited.test_set(n->_idx)) {
890 worklist.push(n);
891 }
892 }
893 Node::verify(4, visited, worklist);
894 }
895 }
896
897 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
898 const Type* newtype = type_or_null(n);
899 if (nn != n || oldtype != newtype) {
900 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
901 }
902 if (TraceIterativeGVN) {
903 uint wlsize = _worklist.size();
904 if (nn != n) {
905 // print old node
906 tty->print("< ");
907 if (oldtype != newtype && oldtype != nullptr) {
908 oldtype->dump();
909 }
910 do { tty->print("\t"); } while (tty->position() < 16);
911 tty->print("<");
912 n->dump();
913 }
914 if (oldtype != newtype || nn != n) {
915 // print new node and/or new type
916 if (oldtype == nullptr) {
917 tty->print("* ");
918 } else if (nn != n) {
919 tty->print("> ");
920 } else {
921 tty->print("= ");
922 }
923 if (newtype == nullptr) {
924 tty->print("null");
925 } else {
926 newtype->dump();
927 }
928 do { tty->print("\t"); } while (tty->position() < 16);
929 nn->dump();
930 }
931 if (Verbose && wlsize < _worklist.size()) {
932 tty->print(" Push {");
933 while (wlsize != _worklist.size()) {
934 Node* pushed = _worklist.at(wlsize++);
935 tty->print(" %d", pushed->_idx);
936 }
937 tty->print_cr(" }");
938 }
939 if (nn != n) {
940 // ignore n, it might be subsumed
941 verify_step((Node*) nullptr);
942 }
943 }
944 }
945
946 void PhaseIterGVN::init_verifyPhaseIterGVN() {
947 _verify_counter = 0;
948 _verify_full_passes = 0;
949 for (int i = 0; i < _verify_window_size; i++) {
950 _verify_window[i] = nullptr;
951 }
952 #ifdef ASSERT
953 // Verify that all modified nodes are on _worklist
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() && !_worklist.member(n)) {
958 n->dump();
959 fatal("modified node is not on IGVN._worklist");
960 }
961 }
962 #endif
963 }
964
965 void PhaseIterGVN::verify_PhaseIterGVN() {
966 #ifdef ASSERT
967 // Verify nodes with changed inputs.
968 Unique_Node_List* modified_list = C->modified_nodes();
969 while (modified_list != nullptr && modified_list->size()) {
970 Node* n = modified_list->pop();
971 if (!n->is_Con()) { // skip Con nodes
972 n->dump();
973 fatal("modified node was not processed by IGVN.transform_old()");
974 }
975 }
976 #endif
977
978 C->verify_graph_edges();
979 if (is_verify_def_use() && PrintOpto) {
980 if (_verify_counter == _verify_full_passes) {
981 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
982 (int) _verify_full_passes);
983 } else {
984 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
985 (int) _verify_counter, (int) _verify_full_passes);
986 }
987 }
988
989 #ifdef ASSERT
990 if (modified_list != nullptr) {
991 while (modified_list->size() > 0) {
992 Node* n = modified_list->pop();
993 n->dump();
994 assert(false, "VerifyIterativeGVN: new modified node was added");
995 }
996 }
997
998 verify_optimize();
999 #endif
1000 }
1001 #endif /* PRODUCT */
1002
1003 #ifdef ASSERT
1004 /**
1005 * Dumps information that can help to debug the problem. A debug
1006 * build fails with an assert.
1007 */
1008 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1009 n->dump(4);
1010 _worklist.dump();
1011 assert(false, "infinite loop in %s", where);
1012 }
1013
1014 /**
1015 * Prints out information about IGVN if the 'verbose' option is used.
1016 */
1017 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1018 if (TraceIterativeGVN && Verbose) {
1019 tty->print(" Pop ");
1020 n->dump();
1021 if ((num_processed % 100) == 0) {
1022 _worklist.print_set();
1023 }
1024 }
1025 }
1026 #endif /* ASSERT */
1027
1028 void PhaseIterGVN::optimize() {
1029 DEBUG_ONLY(uint num_processed = 0;)
1030 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1031 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1032 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1033 if (StressIGVN) {
1034 shuffle_worklist();
1035 }
1036
1037 // The node count check in the loop below (check_node_count) assumes that we
1038 // increase the live node count with at most
1039 // max_live_nodes_increase_per_iteration in between checks. If this
1040 // assumption does not hold, there is a risk that we exceed the max node
1041 // limit in between checks and trigger an assert during node creation.
1042 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 3;
1043
1044 uint loop_count = 0;
1045 // Pull from worklist and transform the node. If the node has changed,
1046 // update edge info and put uses on worklist.
1047 while (_worklist.size() > 0) {
1048 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1049 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1050 return;
1051 }
1052 Node* n = _worklist.pop();
1053 if (loop_count >= K * C->live_nodes()) {
1054 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1055 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1056 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1057 return;
1058 }
1059 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1060 if (n->outcnt() != 0) {
1061 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1062 // Do the transformation
1063 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1064 Node* nn = transform_old(n);
1065 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1066 // Ensure we did not increase the live node count with more than
1067 // max_live_nodes_increase_per_iteration during the call to transform_old
1068 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1069 assert(increase < max_live_nodes_increase_per_iteration,
1070 "excessive live node increase in single iteration of IGVN: %d "
1071 "(should be at most %d)",
1072 increase, max_live_nodes_increase_per_iteration);
1073 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1074 } else if (!n->is_top()) {
1075 remove_dead_node(n);
1076 }
1077 loop_count++;
1078 }
1079 NOT_PRODUCT(verify_PhaseIterGVN();)
1080 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1081 }
1082
1083 #ifdef ASSERT
1084 void PhaseIterGVN::verify_optimize() {
1085 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1086
1087 if (is_verify_Value() ||
1088 is_verify_Ideal() ||
1089 is_verify_Identity() ||
1090 is_verify_invariants()) {
1091 ResourceMark rm;
1092 Unique_Node_List worklist;
1093 // BFS all nodes, starting at root
1094 worklist.push(C->root());
1095 for (uint j = 0; j < worklist.size(); ++j) {
1096 Node* n = worklist.at(j);
1097 // If we get an assert here, check why the reported node was not processed again in IGVN.
1098 // We should either make sure that this node is properly added back to the IGVN worklist
1099 // in PhaseIterGVN::add_users_to_worklist to update it again or add an exception
1100 // in the verification methods below if that is not possible for some reason (like Load nodes).
1101 if (is_verify_Value()) {
1102 verify_Value_for(n);
1103 }
1104 if (is_verify_Ideal()) {
1105 verify_Ideal_for(n, false);
1106 verify_Ideal_for(n, true);
1107 }
1108 if (is_verify_Identity()) {
1109 verify_Identity_for(n);
1110 }
1111 if (is_verify_invariants()) {
1112 verify_node_invariants_for(n);
1113 }
1114
1115 // traverse all inputs and outputs
1116 for (uint i = 0; i < n->req(); i++) {
1117 if (n->in(i) != nullptr) {
1118 worklist.push(n->in(i));
1119 }
1120 }
1121 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1122 worklist.push(n->fast_out(i));
1123 }
1124 }
1125 }
1126
1127 verify_empty_worklist(nullptr);
1128 }
1129
1130 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1131 // Verify that the igvn worklist is empty. If no optimization happened, then
1132 // nothing needs to be on the worklist.
1133 if (_worklist.size() == 0) { return; }
1134
1135 stringStream ss; // Print as a block without tty lock.
1136 for (uint j = 0; j < _worklist.size(); j++) {
1137 Node* n = _worklist.at(j);
1138 ss.print("igvn.worklist[%d] ", j);
1139 n->dump("\n", false, &ss);
1140 }
1141 if (_worklist.size() != 0 && node != nullptr) {
1142 ss.print_cr("Previously optimized:");
1143 node->dump("\n", false, &ss);
1144 }
1145 tty->print_cr("%s", ss.as_string());
1146 assert(false, "igvn worklist must still be empty after verify");
1147 }
1148
1149 // Check that type(n) == n->Value(), asserts if we have a failure.
1150 // We have a list of exceptions, see detailed comments in code.
1151 // (1) Integer "widen" changes, but the range is the same.
1152 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1153 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1154 void PhaseIterGVN::verify_Value_for(const Node* n, bool strict) {
1155 // If we assert inside type(n), because the type is still a null, then maybe
1156 // the node never went through gvn.transform, which would be a bug.
1157 const Type* told = type(n);
1158 const Type* tnew = n->Value(this);
1159 if (told == tnew) {
1160 return;
1161 }
1162 // Exception (1)
1163 // Integer "widen" changes, but range is the same.
1164 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1165 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1166 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1167 if (t0->lo_as_long() == t1->lo_as_long() &&
1168 t0->hi_as_long() == t1->hi_as_long()) {
1169 return; // ignore integer widen
1170 }
1171 }
1172 // Exception (2)
1173 // LoadNode performs deep traversals. Load is not notified for changes far away.
1174 if (!strict && n->is_Load() && !told->singleton()) {
1175 // MemNode::can_see_stored_value looks up through many memory nodes,
1176 // which means we would need to notify modifications from far up in
1177 // the inputs all the way down to the LoadNode. We don't do that.
1178 return;
1179 }
1180 // Exception (3)
1181 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1182 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1183 // SubNode::Value
1184 // CmpPNode::sub
1185 // MemNode::detect_ptr_independence
1186 // MemNode::all_controls_dominate
1187 // We find all controls of a pointer load, and see if they dominate the control of
1188 // an allocation. If they all dominate, we know the allocation is after (independent)
1189 // of the pointer load, and we can say the pointers are different. For this we call
1190 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1191 // control sub of the allocation. The problems is that sometimes dominates answers
1192 // false conservatively, and later it can determine that it is indeed true. Loops with
1193 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1194 // so the traversal becomes more powerful. This is difficult to remedy, we would have
1195 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1196 // after loop-opts, so that should take care of many of these cases.
1197 return;
1198 }
1199
1200 stringStream ss; // Print as a block without tty lock.
1201 ss.cr();
1202 ss.print_cr("Missed Value optimization:");
1203 n->dump_bfs(1, nullptr, "", &ss);
1204 ss.print_cr("Current type:");
1205 told->dump_on(&ss);
1206 ss.cr();
1207 ss.print_cr("Optimized type:");
1208 tnew->dump_on(&ss);
1209 ss.cr();
1210 tty->print_cr("%s", ss.as_string());
1211
1212 switch (_phase) {
1213 case PhaseValuesType::iter_gvn:
1214 assert(false, "Missed Value optimization opportunity in PhaseIterGVN for %s",n->Name());
1215 break;
1216 case PhaseValuesType::ccp:
1217 assert(false, "PhaseCCP not at fixpoint: analysis result may be unsound for %s", n->Name());
1218 break;
1219 default:
1220 assert(false, "Unexpected phase");
1221 break;
1222 }
1223 }
1224
1225 // Check that all Ideal optimizations that could be done were done.
1226 // Asserts if it found missed optimization opportunities or encountered unexpected changes, and
1227 // returns normally otherwise (no missed optimization, or skipped verification).
1228 void PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1229 // First, we check a list of exceptions, where we skip verification,
1230 // because there are known cases where Ideal can optimize after IGVN.
1231 // Some may be expected and cannot be fixed, and others should be fixed.
1232 switch (n->Opcode()) {
1233 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1234 // (see "Range-Check scan limit"). So, it is possible that something
1235 // is optimized in that input subgraph, and the RangeCheck was not
1236 // added to the worklist because it would be too expensive to walk
1237 // down the graph for 1000 nodes and put all on the worklist.
1238 //
1239 // Found with:
1240 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1241 case Op_RangeCheck:
1242 return;
1243
1244 // IfNode::Ideal does:
1245 // Node* prev_dom = search_identical(dist, igvn);
1246 // which means we seach up the CFG, traversing at most up to a distance.
1247 // If anything happens rather far away from the If, we may not put the If
1248 // back on the worklist.
1249 //
1250 // Found with:
1251 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1252 case Op_If:
1253 return;
1254
1255 // IfNode::simple_subsuming
1256 // Looks for dominating test that subsumes the current test.
1257 // Notification could be difficult because of larger distance.
1258 //
1259 // Found with:
1260 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1261 // -XX:VerifyIterativeGVN=1110
1262 case Op_CountedLoopEnd:
1263 return;
1264
1265 // LongCountedLoopEndNode::Ideal
1266 // Probably same issue as above.
1267 //
1268 // Found with:
1269 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1270 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1271 case Op_LongCountedLoopEnd:
1272 return;
1273
1274 // RegionNode::Ideal does "Skip around the useless IF diamond".
1275 // 245 IfTrue === 244
1276 // 258 If === 245 257
1277 // 259 IfTrue === 258 [[ 263 ]]
1278 // 260 IfFalse === 258 [[ 263 ]]
1279 // 263 Region === 263 260 259 [[ 263 268 ]]
1280 // to
1281 // 245 IfTrue === 244
1282 // 263 Region === 263 245 _ [[ 263 268 ]]
1283 //
1284 // "Useless" means that there is no code in either branch of the If.
1285 // I found a case where this was not done yet during IGVN.
1286 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1287 //
1288 // Found with:
1289 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1290 case Op_Region:
1291 return;
1292
1293 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1294 // that smaller idx are first. Tracking it back, it led me to
1295 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1296 //
1297 // Example:
1298 // Before PhaseIdealLoop::remix_address_expressions
1299 // 154 AddI === _ 12 144
1300 // After PhaseIdealLoop::remix_address_expressions
1301 // 154 AddI === _ 144 12
1302 // After AddNode::Ideal
1303 // 154 AddI === _ 12 144
1304 //
1305 // I suspect that the node should be added to the IGVN worklist after
1306 // PhaseIdealLoop::remix_address_expressions
1307 //
1308 // This is the only case I looked at, there may be others. Found like this:
1309 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1310 //
1311 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1312 //
1313 // Note: currently all of these fail also for other reasons, for example
1314 // because of "commute" doing the reordering with the phi below. Once
1315 // that is resolved, we can come back to this issue here.
1316 //
1317 // case Op_AddD:
1318 // case Op_AddI:
1319 // case Op_AddL:
1320 // case Op_AddF:
1321 // case Op_MulI:
1322 // case Op_MulL:
1323 // case Op_MulF:
1324 // case Op_MulD:
1325 // if (n->in(1)->_idx > n->in(2)->_idx) {
1326 // // Expect "commute" to revert this case.
1327 // return false;
1328 // }
1329 // break; // keep verifying
1330
1331 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1332 // Check for tight loop increments: Loop-phi of Add of loop-phi
1333 // It wants to take the phi into in(1):
1334 // 471 Phi === 435 38 390
1335 // 390 AddF === _ 471 391
1336 //
1337 // Other Associative operators are also affected equally.
1338 //
1339 // Investigate why this does not happen earlier during IGVN.
1340 //
1341 // Found with:
1342 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1343 // -XX:VerifyIterativeGVN=1110
1344 case Op_AddD:
1345 //case Op_AddI: // Also affected for other reasons, see case further down.
1346 //case Op_AddL: // Also affected for other reasons, see case further down.
1347 case Op_AddF:
1348 case Op_MulI:
1349 case Op_MulL:
1350 case Op_MulF:
1351 case Op_MulD:
1352 case Op_MinF:
1353 case Op_MinD:
1354 case Op_MaxF:
1355 case Op_MaxD:
1356 // XorINode::Ideal
1357 // Found with:
1358 // compiler/intrinsics/chacha/TestChaCha20.java
1359 // -XX:VerifyIterativeGVN=1110
1360 case Op_XorI:
1361 case Op_XorL:
1362 // It seems we may have similar issues with the HF cases.
1363 // Found with aarch64:
1364 // compiler/vectorization/TestFloat16VectorOperations.java
1365 // -XX:VerifyIterativeGVN=1110
1366 case Op_AddHF:
1367 case Op_MulHF:
1368 case Op_MaxHF:
1369 case Op_MinHF:
1370 return;
1371
1372 // In MulNode::Ideal the edges can be swapped to help value numbering:
1373 //
1374 // // We are OK if right is a constant, or right is a load and
1375 // // left is a non-constant.
1376 // if( !(t2->singleton() ||
1377 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1378 // if( t1->singleton() || // Left input is a constant?
1379 // // Otherwise, sort inputs (commutativity) to help value numbering.
1380 // (in(1)->_idx > in(2)->_idx) ) {
1381 // swap_edges(1, 2);
1382 //
1383 // Why was this not done earlier during IGVN?
1384 //
1385 // Found with:
1386 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1387 // -XX:VerifyIterativeGVN=1110
1388 case Op_AndI:
1389 // Same for AndL.
1390 // Found with:
1391 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1392 // -XX:VerifyIterativeGVN=1110
1393 case Op_AndL:
1394 return;
1395
1396 // SubLNode::Ideal does transform like:
1397 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1398 //
1399 // In IGVN before verification:
1400 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1401 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1402 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1403 // 8016 ConL === 0 [[ 8383 ]] #long:0
1404 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1405 //
1406 // And then in verification:
1407 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1408 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1409 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1410 // 8016 ConL === 0 [[ 8383 ]] #long:0
1411 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1412 //
1413 // So the form changed from:
1414 // c1 - (y + [8423 ConvI2L])
1415 // to
1416 // c1 - (y + -2)
1417 // but the SubL was not added to the IGVN worklist. Investigate why.
1418 // There could be other issues too.
1419 //
1420 // There seems to be a related AddL IGVN optimization that triggers
1421 // the same SubL optimization, so investigate that too.
1422 //
1423 // Found with:
1424 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1425 case Op_SubL:
1426 return;
1427
1428 // SubINode::Ideal does
1429 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1430 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1431 //
1432 // Investigate why this does not yet happen during IGVN.
1433 //
1434 // Found with:
1435 // test/hotspot/jtreg/compiler/c2/IVTest.java
1436 // -XX:VerifyIterativeGVN=1110
1437 case Op_SubI:
1438 return;
1439
1440 // AddNode::IdealIL does transform like:
1441 // Convert x + (con - y) into "(x - y) + con"
1442 //
1443 // In IGVN before verification:
1444 // 8382 ConvI2L
1445 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1446 // 8383 SubL === _ 8381 8382
1447 // 8168 ConvI2L
1448 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1449 //
1450 // And then in verification:
1451 // 8424 AddL
1452 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1453 // 8383 SubL === _ 8016 8424
1454 // 8168 ConvI2L
1455 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1456 //
1457 // So the form changed from:
1458 // x + (ConvI2L(0) - [8382 ConvI2L])
1459 // to
1460 // x + (0 - [8424 AddL])
1461 // but the AddL was not added to the IGVN worklist. Investigate why.
1462 // There could be other issues, too. For example with "commute", see above.
1463 //
1464 // Found with:
1465 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1466 case Op_AddL:
1467 return;
1468
1469 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1470 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1471 // record_for_cleanup(cmp, phase);
1472 // This verification code in the Ideal code creates new nodes, and checks
1473 // if they fold in unexpected ways. This means some nodes are created and
1474 // added to the worklist, even if the SubTypeCheck is not optimized. This
1475 // goes agains the assumption of the verification here, which assumes that
1476 // if the node is not optimized, then no new nodes should be created, and
1477 // also no nodes should be added to the worklist.
1478 // I see two options:
1479 // 1) forbid what verify_helper does, because for each Ideal call it
1480 // uses memory and that is suboptimal. But it is not clear how that
1481 // verification can be done otherwise.
1482 // 2) Special case the verification here. Probably the new nodes that
1483 // were just created are dead, i.e. they are not connected down to
1484 // root. We could verify that, and remove those nodes from the graph
1485 // by setting all their inputs to nullptr. And of course we would
1486 // have to remove those nodes from the worklist.
1487 // Maybe there are other options too, I did not dig much deeper yet.
1488 //
1489 // Found with:
1490 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1491 case Op_SubTypeCheck:
1492 return;
1493
1494 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1495 // to other nodes: Conv, Add, Sub, Mul, And ...
1496 //
1497 // 107 ConI === 0 [[ ... ]] #int:2
1498 // 84 LoadRange === _ 7 83
1499 // 50 ConI === 0 [[ ... ]] #int:0
1500 // 549 LoopLimit === _ 50 84 107
1501 //
1502 // I stepped backward, to see how the node was generated, and I found that it was
1503 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1504 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1505 // that, and it seems we want to skip lowering it until after loop-opts, but never
1506 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1507 // could be other issues too.
1508 //
1509 // Fond with:
1510 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1511 case Op_LoopLimit:
1512 return;
1513
1514 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1515 // "This optimization tries to find two or more inputs of phi with the same constant
1516 // value. It then splits them into a separate Phi, and according Region."
1517 //
1518 // Example:
1519 // 130 DecodeN === _ 129
1520 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1521 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1522 //
1523 // turns into:
1524 //
1525 // 50 ConP === 0 [[ 99 91 18 ]] #null
1526 // 130 DecodeN === _ 129 [[ 18 ]]
1527 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1528 //
1529 // We would have to investigate why this optimization does not happen during IGVN.
1530 // There could also be other issues - I did not investigate further yet.
1531 //
1532 // Found with:
1533 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1534 case Op_Phi:
1535 return;
1536
1537 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1538 // For examle "The allocated object does not escape".
1539 //
1540 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1541 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1542 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1543 // worklist when this change happens?
1544 //
1545 // Found with:
1546 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1547 case Op_MemBarStoreStore:
1548 return;
1549
1550 // ConvI2LNode::Ideal converts
1551 // 648 AddI === _ 583 645 [[ 661 ]]
1552 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1553 // into
1554 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1555 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1556 // 773 AddL === _ 771 772 [[ ]]
1557 //
1558 // We have to investigate why this does not happen during IGVN in this case.
1559 // There could also be other issues - I did not investigate further yet.
1560 //
1561 // Found with:
1562 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1563 case Op_ConvI2L:
1564 return;
1565
1566 // AddNode::IdealIL can do this transform (and similar other ones):
1567 // Convert "a*b+a*c into a*(b+c)
1568 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1569 // during IGVN? There was a mutation for one of the MulI, and only
1570 // after that the pattern was as needed for the optimization. The MulI
1571 // was added to the IGVN worklist, but not the AddI. This probably
1572 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1573 //
1574 // Found with:
1575 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1576 // -XX:VerifyIterativeGVN=1110
1577 case Op_AddI:
1578 return;
1579
1580 // ArrayCopyNode::Ideal
1581 // calls ArrayCopyNode::prepare_array_copy
1582 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1583 // is not expected, and we create a range int:0..-1
1584 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1585 // note: the type is already empty!
1586 // calls PhaseIterGVN::transform
1587 // calls PhaseIterGVN::transform_old
1588 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1589 // calls Unique_Node_List::push -> pushes TOP to worklist
1590 //
1591 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1592 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1593 //
1594 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1595 // return nullptr, indicating nothing happened?
1596 // Is it smart to do transform in Compile::constrained_convI2L, and then
1597 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1598 // Should we just allow TOP to land on the worklist, as an exception?
1599 //
1600 // Found with:
1601 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1602 // -XX:VerifyIterativeGVN=1110
1603 case Op_ArrayCopy:
1604 return;
1605
1606 // CastLLNode::Ideal
1607 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1608 //
1609 // Could be a notification issue, where updates inputs of CastLL do not notify
1610 // down through SubL to CastLL.
1611 //
1612 // Found With:
1613 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1614 // -XX:VerifyIterativeGVN=1110
1615 case Op_CastLL:
1616 return;
1617
1618 // Similar case happens to CastII
1619 //
1620 // Found With:
1621 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1622 // -XX:VerifyIterativeGVN=1110
1623 case Op_CastII:
1624 return;
1625
1626 // MaxLNode::Ideal
1627 // calls AddNode::Ideal
1628 // calls commute -> decides to swap edges
1629 //
1630 // Another notification issue, because we check inputs of inputs?
1631 // MaxL -> Phi -> Loop
1632 // MaxL -> Phi -> MaxL
1633 //
1634 // Found with:
1635 // compiler/c2/irTests/TestIfMinMax.java
1636 // -XX:VerifyIterativeGVN=1110
1637 case Op_MaxL:
1638 case Op_MinL:
1639 return;
1640
1641 // OrINode::Ideal
1642 // calls AddNode::Ideal
1643 // calls commute -> left is Load, right not -> commute.
1644 //
1645 // Not sure why notification does not work here, seems like
1646 // the depth is only 1, so it should work. Needs investigation.
1647 //
1648 // Found with:
1649 // compiler/codegen/TestCharVect2.java#id0
1650 // -XX:VerifyIterativeGVN=1110
1651 case Op_OrI:
1652 case Op_OrL:
1653 return;
1654
1655 // Bool -> constant folded to 1.
1656 // Issue with notification?
1657 //
1658 // Found with:
1659 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1660 // -XX:VerifyIterativeGVN=1110
1661 case Op_Bool:
1662 return;
1663
1664 // LShiftLNode::Ideal
1665 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1666 // Probably a notification issue.
1667 //
1668 // Found with:
1669 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1670 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1671 case Op_LShiftL:
1672 return;
1673
1674 // LShiftINode::Ideal
1675 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1676 // Could be issue with notification of inputs of inputs
1677 //
1678 // Side-note: should cases like these not be shared between
1679 // LShiftI and LShiftL?
1680 //
1681 // Found with:
1682 // compiler/escapeAnalysis/Test6689060.java
1683 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1684 case Op_LShiftI:
1685 return;
1686
1687 // AddPNode::Ideal seems to do set_req without removing lock first.
1688 // Found with various vector tests tier1-tier3.
1689 case Op_AddP:
1690 return;
1691
1692 // StrIndexOfNode::Ideal
1693 // Found in tier1-3.
1694 case Op_StrIndexOf:
1695 case Op_StrIndexOfChar:
1696 return;
1697
1698 // StrEqualsNode::Identity
1699 //
1700 // Found (linux x64 only?) with:
1701 // serviceability/sa/ClhsdbThreadContext.java
1702 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1703 // Note: The -XX:LockingMode option is not available anymore.
1704 case Op_StrEquals:
1705 return;
1706
1707 // AryEqNode::Ideal
1708 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1709 //
1710 // Found with:
1711 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1712 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1713 case Op_AryEq:
1714 return;
1715
1716 // MergeMemNode::Ideal
1717 // Found in tier1-3. Did not investigate further yet.
1718 case Op_MergeMem:
1719 return;
1720
1721 // URShiftINode::Ideal
1722 // Found in tier1-3. Did not investigate further yet.
1723 case Op_URShiftI:
1724 return;
1725
1726 // CMoveINode::Ideal
1727 // Found in tier1-3. Did not investigate further yet.
1728 case Op_CMoveI:
1729 return;
1730
1731 // CmpPNode::Ideal calls isa_const_java_mirror
1732 // and generates new constant nodes, even if no progress is made.
1733 // We can probably rewrite this so that only types are generated.
1734 // It seems that object types are not hashed, we could investigate
1735 // if that is an option as well.
1736 //
1737 // Found with:
1738 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1739 case Op_CmpP:
1740 return;
1741
1742 // MinINode::Ideal
1743 // Did not investigate, but there are some patterns that might
1744 // need more notification.
1745 case Op_MinI:
1746 case Op_MaxI: // preemptively removed it as well.
1747 return;
1748 }
1749
1750 if (n->is_Load()) {
1751 // LoadNode::Ideal uses tries to find an earlier memory state, and
1752 // checks can_see_stored_value for it.
1753 //
1754 // Investigate why this was not already done during IGVN.
1755 // A similar issue happens with Identity.
1756 //
1757 // There seem to be other cases where loads go up some steps, like
1758 // LoadNode::Ideal going up 10x steps to find dominating load.
1759 //
1760 // Found with:
1761 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1762 // -XX:VerifyIterativeGVN=1110
1763 return;
1764 }
1765
1766 if (n->is_Store()) {
1767 // StoreNode::Ideal can do this:
1768 // // Capture an unaliased, unconditional, simple store into an initializer.
1769 // // Or, if it is independent of the allocation, hoist it above the allocation.
1770 // That replaces the Store with a MergeMem.
1771 //
1772 // We have to investigate why this does not happen during IGVN in this case.
1773 // There could also be other issues - I did not investigate further yet.
1774 //
1775 // Found with:
1776 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1777 return;
1778 }
1779
1780 if (n->is_Vector()) {
1781 // VectorNode::Ideal swaps edges, but only for ops
1782 // that are deemed commutable. But swap_edges
1783 // requires the hash to be invariant when the edges
1784 // are swapped, which is not implemented for these
1785 // vector nodes. This seems not to create any trouble
1786 // usually, but we can also get graphs where in the
1787 // end the nodes are not all commuted, so there is
1788 // definitively an issue here.
1789 //
1790 // Probably we have two options: kill the hash, or
1791 // properly make the hash commutation friendly.
1792 //
1793 // Found with:
1794 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1795 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1796 return;
1797 }
1798
1799 if (n->is_Region()) {
1800 // LoopNode::Ideal calls RegionNode::Ideal.
1801 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1802 // But I got an issue because RegionNode::optimize_trichotomy
1803 // then modifies another node, and pushes nodes to the worklist
1804 // Not sure if this is ok, modifying another node like that.
1805 // Maybe it is, then we need to look into what to do with
1806 // the nodes that are now on the worklist, maybe just clear
1807 // them out again. But maybe modifying other nodes like that
1808 // is also bad design. In the end, we return nullptr for
1809 // the current CountedLoop. But the extra nodes on the worklist
1810 // trip the asserts later on.
1811 //
1812 // Found with:
1813 // compiler/eliminateAutobox/TestShortBoxing.java
1814 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1815 return;
1816 }
1817
1818 if (n->is_CallJava()) {
1819 // CallStaticJavaNode::Ideal
1820 // Led to a crash:
1821 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1822 //
1823 // Did not investigate yet, could be a bug.
1824 // Or maybe it does not expect to be called during verification.
1825 //
1826 // Found with:
1827 // test/jdk/jdk/incubator/vector/VectorRuns.java
1828 // -XX:VerifyIterativeGVN=1110
1829
1830 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1831 // and possibly their subclasses.
1832 // During late inlining it can call CallJavaNode::register_for_late_inline
1833 // That means we do more rounds of late inlining, but might fail.
1834 // Then we do IGVN again, and register the node again for late inlining.
1835 // This creates an endless cycle. Everytime we try late inlining, we
1836 // are also creating more nodes, especially SafePoint and MergeMem.
1837 // These nodes are immediately rejected when the inlining fails in the
1838 // do_late_inline_check, but they still grow the memory, until we hit
1839 // the MemLimit and crash.
1840 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1841 // called repeatedly, and eventually we terminate. I fear this is not
1842 // a great assumption to make. We should investigate more.
1843 //
1844 // Found with:
1845 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1846 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1847 return;
1848 }
1849
1850 // The number of nodes shoud not increase.
1851 uint old_unique = C->unique();
1852 // The hash of a node should not change, this would indicate different inputs
1853 uint old_hash = n->hash();
1854 // Remove 'n' from hash table in case it gets modified. We want to avoid
1855 // hitting the "Need to remove from hash before changing edges" assert if
1856 // a change occurs. Instead, we would like to proceed with the optimization,
1857 // return and finally hit the assert in PhaseIterGVN::verify_optimize to get
1858 // a more meaningful message
1859 _table.hash_delete(n);
1860 Node* i = n->Ideal(this, can_reshape);
1861 // If there was no new Idealization, we are probably happy.
1862 if (i == nullptr) {
1863 if (old_unique < C->unique()) {
1864 stringStream ss; // Print as a block without tty lock.
1865 ss.cr();
1866 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1867 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1868 n->dump_bfs(1, nullptr, "", &ss);
1869 tty->print_cr("%s", ss.as_string());
1870 assert(false, "Unexpected new unused nodes from applying Ideal optimization on %s", n->Name());
1871 }
1872
1873 if (old_hash != n->hash()) {
1874 stringStream ss; // Print as a block without tty lock.
1875 ss.cr();
1876 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1877 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1878 n->dump_bfs(1, nullptr, "", &ss);
1879 tty->print_cr("%s", ss.as_string());
1880 assert(false, "Unexpected hash change from applying Ideal optimization on %s", n->Name());
1881 }
1882
1883 verify_empty_worklist(n);
1884
1885 // Everything is good.
1886 hash_find_insert(n);
1887 return;
1888 }
1889
1890 // We just saw a new Idealization which was not done during IGVN.
1891 stringStream ss; // Print as a block without tty lock.
1892 ss.cr();
1893 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1894 if (i == n) {
1895 ss.print_cr("The node was reshaped by Ideal.");
1896 } else {
1897 ss.print_cr("The node was replaced by Ideal.");
1898 ss.print_cr("Old node:");
1899 n->dump_bfs(1, nullptr, "", &ss);
1900 }
1901 ss.print_cr("The result after Ideal:");
1902 i->dump_bfs(1, nullptr, "", &ss);
1903 tty->print_cr("%s", ss.as_string());
1904
1905 assert(false, "Missed Ideal optimization opportunity in PhaseIterGVN for %s", n->Name());
1906 }
1907
1908 // Check that all Identity optimizations that could be done were done.
1909 // Asserts if it found missed optimization opportunities, and
1910 // returns normally otherwise (no missed optimization, or skipped verification).
1911 void PhaseIterGVN::verify_Identity_for(Node* n) {
1912 // First, we check a list of exceptions, where we skip verification,
1913 // because there are known cases where Ideal can optimize after IGVN.
1914 // Some may be expected and cannot be fixed, and others should be fixed.
1915 switch (n->Opcode()) {
1916 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1917 // after loopopts:
1918 // // Transforming long counted loops requires a safepoint node. Do not
1919 // // eliminate a safepoint until loop opts are over.
1920 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1921 //
1922 // I think the check for major_progress does delay it until after loopopts
1923 // but it does not ensure that the node is on the IGVN worklist after
1924 // loopopts. I think we should try to instead check for
1925 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1926 //
1927 // Found with:
1928 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1929 case Op_SafePoint:
1930 return;
1931
1932 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1933 // does not record any other memory splits.
1934 //
1935 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1936 // never got called during IGVN for this node, investigate why.
1937 //
1938 // Found with:
1939 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1940 case Op_MergeMem:
1941 return;
1942
1943 // ConstraintCastNode::Identity finds casts that are the same, except that
1944 // the control is "higher up", i.e. dominates. The call goes via
1945 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1946 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1947 // away from the cast, but within 100 idom steps, the cast may not be
1948 // put on the IGVN worklist any more.
1949 //
1950 // Found with:
1951 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1952 case Op_CastPP:
1953 case Op_CastII:
1954 case Op_CastLL:
1955 return;
1956
1957 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1958 // checks dominator, which may be changed, but too far up for notification
1959 // to work.
1960 //
1961 // Found with:
1962 // compiler/c2/irTests/TestSkeletonPredicates.java
1963 // -XX:VerifyIterativeGVN=1110
1964 case Op_CheckCastPP:
1965 return;
1966
1967 // In SubNode::Identity, we do:
1968 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1969 // In the example, the AddI had an input replaced, the AddI is
1970 // added to the IGVN worklist, but the SubI is one link further
1971 // down and is not added. I checked add_users_of_use_to_worklist
1972 // where I would expect the SubI would be added, and I cannot
1973 // find the pattern, only this one:
1974 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1975 //
1976 // Fix this "notification" issue and check if there are any other
1977 // issues.
1978 //
1979 // Found with:
1980 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1981 case Op_SubI:
1982 case Op_SubL:
1983 return;
1984
1985 // PhiNode::Identity checks for patterns like:
1986 // r = (x != con) ? x : con;
1987 // that can be constant folded to "x".
1988 //
1989 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1990 // I suspect there was some earlier change to one of the inputs, but
1991 // not all relevant outputs were put on the IGVN worklist.
1992 //
1993 // Found with:
1994 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1995 // -XX:VerifyIterativeGVN=1110
1996 case Op_Phi:
1997 return;
1998
1999 // ConvI2LNode::Identity does
2000 // convert I2L(L2I(x)) => x
2001 //
2002 // Investigate why this did not already happen during IGVN.
2003 //
2004 // Found with:
2005 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
2006 // -XX:VerifyIterativeGVN=1110
2007 case Op_ConvI2L:
2008 return;
2009
2010 // MaxNode::find_identity_operation
2011 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
2012 // This can be a 2-hop search, so maybe notification is not
2013 // good enough.
2014 //
2015 // Found with:
2016 // compiler/codegen/TestBooleanVect.java
2017 // -XX:VerifyIterativeGVN=1110
2018 case Op_MaxL:
2019 case Op_MinL:
2020 case Op_MaxI:
2021 case Op_MinI:
2022 case Op_MaxF:
2023 case Op_MinF:
2024 case Op_MaxHF:
2025 case Op_MinHF:
2026 case Op_MaxD:
2027 case Op_MinD:
2028 return;
2029
2030
2031 // AddINode::Identity
2032 // Converts (x-y)+y to x
2033 // Could be issue with notification
2034 //
2035 // Turns out AddL does the same.
2036 //
2037 // Found with:
2038 // compiler/c2/Test6792161.java
2039 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2040 case Op_AddI:
2041 case Op_AddL:
2042 return;
2043
2044 // AbsINode::Identity
2045 // Not investigated yet.
2046 case Op_AbsI:
2047 return;
2048 }
2049
2050 if (n->is_Load()) {
2051 // LoadNode::Identity tries to look for an earlier store value via
2052 // can_see_stored_value. I found an example where this led to
2053 // an Allocation, where we could assume the value was still zero.
2054 // So the LoadN can be replaced with a zerocon.
2055 //
2056 // Investigate why this was not already done during IGVN.
2057 // A similar issue happens with Ideal.
2058 //
2059 // Found with:
2060 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2061 return;
2062 }
2063
2064 if (n->is_Store()) {
2065 // StoreNode::Identity
2066 // Not investigated, but found missing optimization for StoreI.
2067 // Looks like a StoreI is replaced with an InitializeNode.
2068 //
2069 // Found with:
2070 // applications/ctw/modules/java_base_2.java
2071 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2072 return;
2073 }
2074
2075 if (n->is_Vector()) {
2076 // Found with tier1-3. Not investigated yet.
2077 // The observed issue was with AndVNode::Identity
2078 return;
2079 }
2080
2081 Node* i = n->Identity(this);
2082 // If we cannot find any other Identity, we are happy.
2083 if (i == n) {
2084 verify_empty_worklist(n);
2085 return;
2086 }
2087
2088 // The verification just found a new Identity that was not found during IGVN.
2089 stringStream ss; // Print as a block without tty lock.
2090 ss.cr();
2091 ss.print_cr("Missed Identity optimization:");
2092 ss.print_cr("Old node:");
2093 n->dump_bfs(1, nullptr, "", &ss);
2094 ss.print_cr("New node:");
2095 i->dump_bfs(1, nullptr, "", &ss);
2096 tty->print_cr("%s", ss.as_string());
2097
2098 assert(false, "Missed Identity optimization opportunity in PhaseIterGVN for %s", n->Name());
2099 }
2100
2101 // Some other verifications that are not specific to a particular transformation.
2102 void PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2103 if (n->is_AddP()) {
2104 if (!n->as_AddP()->address_input_has_same_base()) {
2105 stringStream ss; // Print as a block without tty lock.
2106 ss.cr();
2107 ss.print_cr("Base pointers must match for AddP chain:");
2108 n->dump_bfs(2, nullptr, "", &ss);
2109 tty->print_cr("%s", ss.as_string());
2110
2111 assert(false, "Broken node invariant for %s", n->Name());
2112 }
2113 }
2114 }
2115 #endif
2116
2117 /**
2118 * Register a new node with the optimizer. Update the types array, the def-use
2119 * info. Put on worklist.
2120 */
2121 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2122 set_type_bottom(n);
2123 _worklist.push(n);
2124 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2125 return n;
2126 }
2127
2128 //------------------------------transform--------------------------------------
2129 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2130 Node *PhaseIterGVN::transform( Node *n ) {
2131 if (_delay_transform) {
2132 // Register the node but don't optimize for now
2133 register_new_node_with_optimizer(n);
2134 return n;
2135 }
2136
2137 // If brand new node, make space in type array, and give it a type.
2138 ensure_type_or_null(n);
2139 if (type_or_null(n) == nullptr) {
2140 set_type_bottom(n);
2141 }
2142
2143 return transform_old(n);
2144 }
2145
2146 Node *PhaseIterGVN::transform_old(Node* n) {
2147 NOT_PRODUCT(set_transforms());
2148 // Remove 'n' from hash table in case it gets modified
2149 _table.hash_delete(n);
2150 #ifdef ASSERT
2151 if (is_verify_def_use()) {
2152 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2153 }
2154 #endif
2155
2156 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2157 if (n->is_Cmp()) {
2158 add_users_to_worklist(n);
2159 }
2160
2161 // Apply the Ideal call in a loop until it no longer applies
2162 Node* k = n;
2163 DEBUG_ONLY(dead_loop_check(k);)
2164 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2165 C->remove_modified_node(k);
2166 Node* i = apply_ideal(k, /*can_reshape=*/true);
2167 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2168 #ifndef PRODUCT
2169 verify_step(k);
2170 #endif
2171
2172 DEBUG_ONLY(uint loop_count = 1;)
2173 while (i != nullptr) {
2174 #ifdef ASSERT
2175 if (loop_count >= K + C->live_nodes()) {
2176 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2177 }
2178 #endif
2179 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2180 // Made a change; put users of original Node on worklist
2181 add_users_to_worklist(k);
2182 // Replacing root of transform tree?
2183 if (k != i) {
2184 // Make users of old Node now use new.
2185 subsume_node(k, i);
2186 k = i;
2187 }
2188 DEBUG_ONLY(dead_loop_check(k);)
2189 // Try idealizing again
2190 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2191 C->remove_modified_node(k);
2192 i = apply_ideal(k, /*can_reshape=*/true);
2193 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2194 #ifndef PRODUCT
2195 verify_step(k);
2196 #endif
2197 DEBUG_ONLY(loop_count++;)
2198 }
2199
2200 // If brand new node, make space in type array.
2201 ensure_type_or_null(k);
2202
2203 // See what kind of values 'k' takes on at runtime
2204 const Type* t = k->Value(this);
2205 assert(t != nullptr, "value sanity");
2206
2207 // Since I just called 'Value' to compute the set of run-time values
2208 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2209 // cache Value. Later requests for the local phase->type of this Node can
2210 // use the cached Value instead of suffering with 'bottom_type'.
2211 if (type_or_null(k) != t) {
2212 #ifndef PRODUCT
2213 inc_new_values();
2214 set_progress();
2215 #endif
2216 set_type(k, t);
2217 // If k is a TypeNode, capture any more-precise type permanently into Node
2218 k->raise_bottom_type(t);
2219 // Move users of node to worklist
2220 add_users_to_worklist(k);
2221 }
2222 // If 'k' computes a constant, replace it with a constant
2223 if (t->singleton() && !k->is_Con()) {
2224 NOT_PRODUCT(set_progress();)
2225 Node* con = makecon(t); // Make a constant
2226 add_users_to_worklist(k);
2227 subsume_node(k, con); // Everybody using k now uses con
2228 return con;
2229 }
2230
2231 // Now check for Identities
2232 i = k->Identity(this); // Look for a nearby replacement
2233 if (i != k) { // Found? Return replacement!
2234 NOT_PRODUCT(set_progress();)
2235 add_users_to_worklist(k);
2236 subsume_node(k, i); // Everybody using k now uses i
2237 return i;
2238 }
2239
2240 // Global Value Numbering
2241 i = hash_find_insert(k); // Check for pre-existing node
2242 if (i && (i != k)) {
2243 // Return the pre-existing node if it isn't dead
2244 NOT_PRODUCT(set_progress();)
2245 add_users_to_worklist(k);
2246 subsume_node(k, i); // Everybody using k now uses i
2247 return i;
2248 }
2249
2250 // Return Idealized original
2251 return k;
2252 }
2253
2254 //---------------------------------saturate------------------------------------
2255 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2256 const Type* limit_type) const {
2257 return new_type->narrow(old_type);
2258 }
2259
2260 //------------------------------remove_globally_dead_node----------------------
2261 // Kill a globally dead Node. All uses are also globally dead and are
2262 // aggressively trimmed.
2263 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2264 enum DeleteProgress {
2265 PROCESS_INPUTS,
2266 PROCESS_OUTPUTS
2267 };
2268 ResourceMark rm;
2269 Node_Stack stack(32);
2270 stack.push(dead, PROCESS_INPUTS);
2271
2272 while (stack.is_nonempty()) {
2273 dead = stack.node();
2274 if (dead->Opcode() == Op_SafePoint) {
2275 dead->as_SafePoint()->disconnect_from_root(this);
2276 }
2277 uint progress_state = stack.index();
2278 assert(dead != C->root(), "killing root, eh?");
2279 assert(!dead->is_top(), "add check for top when pushing");
2280 NOT_PRODUCT( set_progress(); )
2281 if (progress_state == PROCESS_INPUTS) {
2282 // After following inputs, continue to outputs
2283 stack.set_index(PROCESS_OUTPUTS);
2284 if (!dead->is_Con()) { // Don't kill cons but uses
2285 bool recurse = false;
2286 // Remove from hash table
2287 _table.hash_delete( dead );
2288 // Smash all inputs to 'dead', isolating him completely
2289 for (uint i = 0; i < dead->req(); i++) {
2290 Node *in = dead->in(i);
2291 if (in != nullptr && in != C->top()) { // Points to something?
2292 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2293 assert((nrep > 0), "sanity");
2294 if (in->outcnt() == 0) { // Made input go dead?
2295 stack.push(in, PROCESS_INPUTS); // Recursively remove
2296 recurse = true;
2297 } else if (in->outcnt() == 1 &&
2298 in->has_special_unique_user()) {
2299 _worklist.push(in->unique_out());
2300 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2301 if (in->Opcode() == Op_Region) {
2302 _worklist.push(in);
2303 } else if (in->is_Store()) {
2304 DUIterator_Fast imax, i = in->fast_outs(imax);
2305 _worklist.push(in->fast_out(i));
2306 i++;
2307 if (in->outcnt() == 2) {
2308 _worklist.push(in->fast_out(i));
2309 i++;
2310 }
2311 assert(!(i < imax), "sanity");
2312 }
2313 } else if (dead->is_data_proj_of_pure_function(in)) {
2314 _worklist.push(in);
2315 }
2316 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2317 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2318 // A Load that directly follows an InitializeNode is
2319 // going away. The Stores that follow are candidates
2320 // again to be captured by the InitializeNode.
2321 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2322 Node *n = in->fast_out(j);
2323 if (n->is_Store()) {
2324 _worklist.push(n);
2325 }
2326 }
2327 }
2328 } // if (in != nullptr && in != C->top())
2329 } // for (uint i = 0; i < dead->req(); i++)
2330 if (recurse) {
2331 continue;
2332 }
2333 } // if (!dead->is_Con())
2334 } // if (progress_state == PROCESS_INPUTS)
2335
2336 // Aggressively kill globally dead uses
2337 // (Rather than pushing all the outs at once, we push one at a time,
2338 // plus the parent to resume later, because of the indefinite number
2339 // of edge deletions per loop trip.)
2340 if (dead->outcnt() > 0) {
2341 // Recursively remove output edges
2342 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2343 } else {
2344 // Finished disconnecting all input and output edges.
2345 stack.pop();
2346 // Remove dead node from iterative worklist
2347 _worklist.remove(dead);
2348 C->remove_useless_node(dead);
2349 }
2350 } // while (stack.is_nonempty())
2351 }
2352
2353 //------------------------------subsume_node-----------------------------------
2354 // Remove users from node 'old' and add them to node 'nn'.
2355 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2356 if (old->Opcode() == Op_SafePoint) {
2357 old->as_SafePoint()->disconnect_from_root(this);
2358 }
2359 assert( old != hash_find(old), "should already been removed" );
2360 assert( old != C->top(), "cannot subsume top node");
2361 // Copy debug or profile information to the new version:
2362 C->copy_node_notes_to(nn, old);
2363 // Move users of node 'old' to node 'nn'
2364 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2365 Node* use = old->last_out(i); // for each use...
2366 // use might need re-hashing (but it won't if it's a new node)
2367 rehash_node_delayed(use);
2368 // Update use-def info as well
2369 // We remove all occurrences of old within use->in,
2370 // so as to avoid rehashing any node more than once.
2371 // The hash table probe swamps any outer loop overhead.
2372 uint num_edges = 0;
2373 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2374 if (use->in(j) == old) {
2375 use->set_req(j, nn);
2376 ++num_edges;
2377 }
2378 }
2379 i -= num_edges; // we deleted 1 or more copies of this edge
2380 }
2381
2382 // Search for instance field data PhiNodes in the same region pointing to the old
2383 // memory PhiNode and update their instance memory ids to point to the new node.
2384 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2385 Node* region = old->in(0);
2386 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2387 PhiNode* phi = region->fast_out(i)->isa_Phi();
2388 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2389 phi->set_inst_mem_id((int)nn->_idx);
2390 }
2391 }
2392 }
2393
2394 // Smash all inputs to 'old', isolating him completely
2395 Node *temp = new Node(1);
2396 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2397 remove_dead_node( old );
2398 temp->del_req(0); // Yank bogus edge
2399 if (nn != nullptr && nn->outcnt() == 0) {
2400 _worklist.push(nn);
2401 }
2402 #ifndef PRODUCT
2403 if (is_verify_def_use()) {
2404 for ( int i = 0; i < _verify_window_size; i++ ) {
2405 if ( _verify_window[i] == old )
2406 _verify_window[i] = nn;
2407 }
2408 }
2409 #endif
2410 temp->destruct(this); // reuse the _idx of this little guy
2411 }
2412
2413 //------------------------------add_users_to_worklist--------------------------
2414 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2415 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2416 worklist.push(n->fast_out(i)); // Push on worklist
2417 }
2418 }
2419
2420 // Return counted loop Phi if as a counted loop exit condition, cmp
2421 // compares the induction variable with n
2422 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2423 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2424 Node* bol = cmp->fast_out(i);
2425 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2426 Node* iff = bol->fast_out(i2);
2427 if (iff->is_BaseCountedLoopEnd()) {
2428 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2429 if (cle->limit() == n) {
2430 PhiNode* phi = cle->phi();
2431 if (phi != nullptr) {
2432 return phi;
2433 }
2434 }
2435 }
2436 }
2437 }
2438 return nullptr;
2439 }
2440
2441 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2442 add_users_to_worklist0(n, _worklist);
2443
2444 Unique_Node_List& worklist = _worklist;
2445 // Move users of node to worklist
2446 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2447 Node* use = n->fast_out(i); // Get use
2448 add_users_of_use_to_worklist(n, use, worklist);
2449 }
2450 }
2451
2452 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2453 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2454 use->is_Store() ) // Enable store/load same address
2455 add_users_to_worklist0(use, worklist);
2456
2457 // If we changed the receiver type to a call, we need to revisit
2458 // the Catch following the call. It's looking for a non-null
2459 // receiver to know when to enable the regular fall-through path
2460 // in addition to the NullPtrException path.
2461 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2462 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2463 if (p != nullptr) {
2464 add_users_to_worklist0(p, worklist);
2465 }
2466 }
2467
2468 uint use_op = use->Opcode();
2469 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2470 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2471 if (use->outcnt() > 0) {
2472 Node* bol = use->raw_out(0);
2473 if (bol->outcnt() > 0) {
2474 Node* iff = bol->raw_out(0);
2475 if (iff->outcnt() == 2) {
2476 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2477 // phi merging either 0 or 1 onto the worklist
2478 Node* ifproj0 = iff->raw_out(0);
2479 Node* ifproj1 = iff->raw_out(1);
2480 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2481 Node* region0 = ifproj0->raw_out(0);
2482 Node* region1 = ifproj1->raw_out(0);
2483 if( region0 == region1 )
2484 add_users_to_worklist0(region0, worklist);
2485 }
2486 }
2487 }
2488 }
2489 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2490 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2491 if (phi != nullptr) {
2492 // Input to the cmp of a loop exit check has changed, thus
2493 // the loop limit may have changed, which can then change the
2494 // range values of the trip-count Phi.
2495 worklist.push(phi);
2496 }
2497 }
2498 if (use_op == Op_CmpI) {
2499 Node* cmp = use;
2500 Node* in1 = cmp->in(1);
2501 Node* in2 = cmp->in(2);
2502 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2503 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2504 //
2505 // in1 in2 in1 in2
2506 // | | | |
2507 // +--- | --+ | |
2508 // | | | | |
2509 // CmpINode | CmpINode
2510 // | | |
2511 // BoolNode | BoolNode
2512 // | | OR |
2513 // IfNode | IfNode
2514 // | | |
2515 // IfProj | IfProj X
2516 // | | | |
2517 // CastIINode CastIINode
2518 //
2519 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2520 if (in1 == n) {
2521 // in1 modified -> could turn into X -> do traversal based on right pattern.
2522 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2523 Node* bol = cmp->fast_out(i2); // For each Bool
2524 if (bol->is_Bool()) {
2525 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2526 Node* iff = bol->fast_out(i3); // For each If
2527 if (iff->is_If()) {
2528 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2529 Node* if_proj = iff->fast_out(i4); // For each IfProj
2530 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2531 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2532 Node* castii = if_proj->fast_out(i5); // For each CastII
2533 if (castii->is_CastII() &&
2534 castii->as_CastII()->carry_dependency()) {
2535 worklist.push(castii);
2536 }
2537 }
2538 }
2539 }
2540 }
2541 }
2542 }
2543 } else {
2544 // Only in2 modified -> can assume X == in2 (left pattern).
2545 assert(n == in2, "only in2 modified");
2546 // Find all CastII with input in1.
2547 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2548 Node* castii = in1->fast_out(j);
2549 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2550 // Find If.
2551 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2552 Node* ifnode = castii->in(0)->in(0);
2553 // Check that if connects to the cmp
2554 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2555 worklist.push(castii);
2556 }
2557 }
2558 }
2559 }
2560 }
2561 }
2562 }
2563 }
2564
2565 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2566 // Patterns:
2567 // ConstraintCast+ -> Sub
2568 // ConstraintCast+ -> Phi
2569 // ConstraintCast+ -> Xor
2570 if (use->is_ConstraintCast()) {
2571 auto push_the_uses_to_worklist = [&](Node* n){
2572 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2573 worklist.push(n);
2574 }
2575 };
2576 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2577 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2578 }
2579 // If changed LShift inputs, check RShift users for useless sign-ext
2580 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2581 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2582 Node* u = use->fast_out(i2);
2583 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2584 worklist.push(u);
2585 }
2586 }
2587 // If changed LShift inputs, check And users for shift and mask (And) operation
2588 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2589 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2590 Node* u = use->fast_out(i2);
2591 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2592 worklist.push(u);
2593 }
2594 }
2595 }
2596 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2597 if (use_op == Op_AddI || use_op == Op_SubI) {
2598 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2599 Node* u = use->fast_out(i2);
2600 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2601 worklist.push(u);
2602 }
2603 }
2604 }
2605 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2606 if (use_op == Op_AndI || use_op == Op_AndL) {
2607 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2608 Node* u = use->fast_out(i2);
2609 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2610 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL) {
2611 worklist.push(u);
2612 }
2613 }
2614 }
2615 // Check for redundant conversion patterns:
2616 // ConvD2L->ConvL2D->ConvD2L
2617 // ConvF2I->ConvI2F->ConvF2I
2618 // ConvF2L->ConvL2F->ConvF2L
2619 // ConvI2F->ConvF2I->ConvI2F
2620 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2621 // are the only ones that are known to trigger missed optimizations otherwise
2622 if (use_op == Op_ConvL2D ||
2623 use_op == Op_ConvI2F ||
2624 use_op == Op_ConvL2F ||
2625 use_op == Op_ConvF2I) {
2626 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2627 Node* u = use->fast_out(i2);
2628 if ((use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2629 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2630 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2631 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F)) {
2632 worklist.push(u);
2633 }
2634 }
2635 }
2636 // If changed AddP inputs:
2637 // - check Stores for loop invariant, and
2638 // - if the changed input is the offset, check constant-offset AddP users for
2639 // address expression flattening.
2640 if (use_op == Op_AddP) {
2641 bool offset_changed = n == use->in(AddPNode::Offset);
2642 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2643 Node* u = use->fast_out(i2);
2644 if (u->is_Mem()) {
2645 worklist.push(u);
2646 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2647 worklist.push(u);
2648 }
2649 }
2650 }
2651 // Check for "abs(0-x)" into "abs(x)" conversion
2652 if (use->is_Sub()) {
2653 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2654 Node* u = use->fast_out(i2);
2655 if (u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2656 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI) {
2657 worklist.push(u);
2658 }
2659 }
2660 }
2661 // Check for Max/Min(A, Max/Min(B, C)) where A == B or A == C
2662 if (use->is_MinMax()) {
2663 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2664 Node* u = use->fast_out(i2);
2665 if (u->Opcode() == use->Opcode()) {
2666 worklist.push(u);
2667 }
2668 }
2669 }
2670 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2671 add_users_to_worklist0(proj, worklist);
2672 };
2673 // If changed initialization activity, check dependent Stores
2674 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2675 InitializeNode* init = use->as_Allocate()->initialization();
2676 if (init != nullptr) {
2677 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2678 }
2679 }
2680 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2681 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2682 // to guarantee the change is not missed.
2683 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2684 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2685 if (p != nullptr) {
2686 add_users_to_worklist0(p, worklist);
2687 }
2688 }
2689
2690 if (use_op == Op_Initialize) {
2691 InitializeNode* init = use->as_Initialize();
2692 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2693 }
2694 // Loading the java mirror from a Klass requires two loads and the type
2695 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2696 // LoadP(LoadP(AddP(foo:Klass, #java_mirror)))
2697 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2698 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2699 Node* u = use->fast_out(i2);
2700 const Type* ut = u->bottom_type();
2701 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2702 worklist.push(u);
2703 }
2704 }
2705 }
2706 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2707 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2708 if (use->outcnt() == 1) {
2709 Node* cmp = use->unique_out();
2710 worklist.push(cmp);
2711 }
2712 }
2713
2714 // From CastX2PNode::Ideal
2715 // CastX2P(AddX(x, y))
2716 // CastX2P(SubX(x, y))
2717 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2718 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2719 Node* u = use->fast_out(i2);
2720 if (u->Opcode() == Op_CastX2P) {
2721 worklist.push(u);
2722 }
2723 }
2724 }
2725
2726 /* AndNode has a special handling when one of the operands is a LShiftNode:
2727 * (LHS << s) & RHS
2728 * if RHS fits in less than s bits, the value of this expression is 0.
2729 * The difficulty is that there might be a conversion node (ConvI2L) between
2730 * the LShiftINode and the AndLNode, like so:
2731 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2732 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2733 * (see `AndIL_min_trailing_zeros`).
2734 *
2735 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2736 * is not enough: there might be no update happening there. We need to
2737 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2738 *
2739 * Moreover we can have ConstraintCasts in between. It may look like
2740 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2741 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2742 * So we must deal with that as well.
2743 */
2744 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2745 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2746 auto push_and_to_worklist = [&worklist](Node* n){
2747 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2748 worklist.push(n);
2749 }
2750 };
2751 use->visit_uses(push_and_to_worklist, is_boundary);
2752 }
2753 }
2754
2755 /**
2756 * Remove the speculative part of all types that we know of
2757 */
2758 void PhaseIterGVN::remove_speculative_types() {
2759 assert(UseTypeSpeculation, "speculation is off");
2760 for (uint i = 0; i < _types.Size(); i++) {
2761 const Type* t = _types.fast_lookup(i);
2762 if (t != nullptr) {
2763 _types.map(i, t->remove_speculative());
2764 }
2765 }
2766 _table.check_no_speculative_types();
2767 }
2768
2769 // Check if the type of a divisor of a Div or Mod node includes zero.
2770 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2771 switch (n->Opcode()) {
2772 case Op_DivI:
2773 case Op_ModI:
2774 case Op_UDivI:
2775 case Op_UModI: {
2776 // Type of divisor includes 0?
2777 if (type(n->in(2)) == Type::TOP) {
2778 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2779 return false;
2780 }
2781 const TypeInt* type_divisor = type(n->in(2))->is_int();
2782 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2783 }
2784 case Op_DivL:
2785 case Op_ModL:
2786 case Op_UDivL:
2787 case Op_UModL: {
2788 // Type of divisor includes 0?
2789 if (type(n->in(2)) == Type::TOP) {
2790 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2791 return false;
2792 }
2793 const TypeLong* type_divisor = type(n->in(2))->is_long();
2794 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2795 }
2796 }
2797 return true;
2798 }
2799
2800 //=============================================================================
2801 #ifndef PRODUCT
2802 uint PhaseCCP::_total_invokes = 0;
2803 uint PhaseCCP::_total_constants = 0;
2804 #endif
2805 //------------------------------PhaseCCP---------------------------------------
2806 // Conditional Constant Propagation, ala Wegman & Zadeck
2807 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2808 NOT_PRODUCT( clear_constants(); )
2809 assert( _worklist.size() == 0, "" );
2810 _phase = PhaseValuesType::ccp;
2811 analyze();
2812 }
2813
2814 #ifndef PRODUCT
2815 //------------------------------~PhaseCCP--------------------------------------
2816 PhaseCCP::~PhaseCCP() {
2817 inc_invokes();
2818 _total_constants += count_constants();
2819 }
2820 #endif
2821
2822
2823 #ifdef ASSERT
2824 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2825 if (tnew->meet(told) != tnew->remove_speculative()) {
2826 n->dump(1);
2827 tty->print("told = "); told->dump(); tty->cr();
2828 tty->print("tnew = "); tnew->dump(); tty->cr();
2829 fatal("Not monotonic");
2830 }
2831 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2832 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2833 }
2834 #endif //ASSERT
2835
2836 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2837 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2838 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2839 // or nodes even further down back to the worklist because their type could change as a result of the current type
2840 // change).
2841 void PhaseCCP::analyze() {
2842 // Initialize all types to TOP, optimistic analysis
2843 for (uint i = 0; i < C->unique(); i++) {
2844 _types.map(i, Type::TOP);
2845 }
2846
2847 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2848 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2849 // Compile is over. The local arena gets de-allocated at the end of its scope.
2850 ResourceArea local_arena(mtCompiler);
2851 Unique_Node_List worklist(&local_arena);
2852 Unique_Node_List worklist_revisit(&local_arena);
2853 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2854
2855 // Push root onto worklist
2856 worklist.push(C->root());
2857
2858 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2859 _root_and_safepoints.push(C->root());
2860
2861 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2862
2863 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2864 while (worklist.size() != 0) {
2865 Node* n = fetch_next_node(worklist);
2866 DEBUG_ONLY(worklist_verify.push(n);)
2867 if (needs_revisit(n)) {
2868 worklist_revisit.push(n);
2869 }
2870 if (n->is_SafePoint()) {
2871 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2872 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2873 _root_and_safepoints.push(n);
2874 }
2875 analyze_step(worklist, n);
2876 }
2877
2878 // More rounds to catch updates far in the graph.
2879 // Revisit nodes that might be able to refine their types at the end of the round.
2880 // If so, process these nodes. If there is remaining work, start another round.
2881 do {
2882 while (worklist.size() != 0) {
2883 Node* n = fetch_next_node(worklist);
2884 analyze_step(worklist, n);
2885 }
2886 for (uint t = 0; t < worklist_revisit.size(); t++) {
2887 Node* n = worklist_revisit.at(t);
2888 analyze_step(worklist, n);
2889 }
2890 } while (worklist.size() != 0);
2891
2892 DEBUG_ONLY(verify_analyze(worklist_verify);)
2893 }
2894
2895 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2896 const Type* new_type = n->Value(this);
2897 if (new_type != type(n)) {
2898 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2899 dump_type_and_node(n, new_type);
2900 set_type(n, new_type);
2901 push_child_nodes_to_worklist(worklist, n);
2902 }
2903 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2904 // Keep track of Type nodes to kill CFG paths that use Type
2905 // nodes that become dead.
2906 _maybe_top_type_nodes.push(n);
2907 }
2908 }
2909
2910 // Some nodes can refine their types due to type change somewhere deep
2911 // in the graph. We will need to revisit them before claiming convergence.
2912 // Add nodes here if particular *Node::Value is doing deep graph traversals
2913 // not handled by PhaseCCP::push_more_uses().
2914 bool PhaseCCP::needs_revisit(Node* n) const {
2915 // LoadNode performs deep traversals. Load is not notified for changes far away.
2916 if (n->is_Load()) {
2917 return true;
2918 }
2919 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2920 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2921 return true;
2922 }
2923 return false;
2924 }
2925
2926 #ifdef ASSERT
2927 // For every node n on verify list, check if type(n) == n->Value()
2928 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2929 // Therefore, we are verifying Value convergence strictly.
2930 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2931 while (worklist_verify.size()) {
2932 Node* n = worklist_verify.pop();
2933
2934 // An assert in verify_Value_for means that PhaseCCP is not at fixpoint
2935 // and that the analysis result may be unsound.
2936 // If this happens, check why the reported nodes were not processed again in CCP.
2937 // We should either make sure that these nodes are properly added back to the CCP worklist
2938 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2939 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2940 // them at the end of the round.
2941 verify_Value_for(n, true);
2942 }
2943 }
2944 #endif
2945
2946 // Fetch next node from worklist to be examined in this iteration.
2947 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2948 if (StressCCP) {
2949 return worklist.remove(C->random() % worklist.size());
2950 } else {
2951 return worklist.pop();
2952 }
2953 }
2954
2955 #ifndef PRODUCT
2956 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2957 if (TracePhaseCCP) {
2958 t->dump();
2959 do {
2960 tty->print("\t");
2961 } while (tty->position() < 16);
2962 n->dump();
2963 }
2964 }
2965 #endif
2966
2967 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2968 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2969 // of the new type of 'n'. Push these nodes to the worklist.
2970 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2971 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2972 Node* use = n->fast_out(i);
2973 push_if_not_bottom_type(worklist, use);
2974 push_more_uses(worklist, n, use);
2975 }
2976 }
2977
2978 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2979 if (n->bottom_type() != type(n)) {
2980 worklist.push(n);
2981 }
2982 }
2983
2984 // For some nodes, we need to propagate the type change to grandchildren or even further down.
2985 // Add them back to the worklist.
2986 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
2987 push_phis(worklist, use);
2988 push_catch(worklist, use);
2989 push_cmpu(worklist, use);
2990 push_counted_loop_phi(worklist, parent, use);
2991 push_loadp(worklist, use);
2992 push_and(worklist, parent, use);
2993 push_cast_ii(worklist, parent, use);
2994 push_opaque_zero_trip_guard(worklist, use);
2995 push_bool_with_cmpu_and_mask(worklist, use);
2996 }
2997
2998
2999 // We must recheck Phis too if use is a Region.
3000 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
3001 if (use->is_Region()) {
3002 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3003 push_if_not_bottom_type(worklist, use->fast_out(i));
3004 }
3005 }
3006 }
3007
3008 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
3009 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
3010 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
3011 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
3012 if (use->is_Call()) {
3013 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3014 Node* proj = use->fast_out(i);
3015 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
3016 Node* catch_node = proj->find_out_with(Op_Catch);
3017 if (catch_node != nullptr) {
3018 worklist.push(catch_node);
3019 }
3020 }
3021 }
3022 }
3023 }
3024
3025 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
3026 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
3027 // and get wrong types otherwise.
3028 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
3029 uint use_op = use->Opcode();
3030 if (use_op == Op_AddI || use_op == Op_SubI) {
3031 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3032 Node* cmpu = use->fast_out(i);
3033 const uint cmpu_opcode = cmpu->Opcode();
3034 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
3035 // Got a CmpU or CmpU3 which might need the new type information from node n.
3036 push_if_not_bottom_type(worklist, cmpu);
3037 }
3038 }
3039 }
3040 }
3041
3042 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3043 // (1b): "(m & x) <u (m + 1))".
3044 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3045 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3046 //
3047 // m x m 1 (%%)
3048 // \ / \ /
3049 // AndI AddI
3050 // \ /
3051 // CmpU
3052 // |
3053 // Bool
3054 //
3055 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3056 uint use_op = use->Opcode();
3057 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3058 // Not "m & x" or "m + 1"
3059 return;
3060 }
3061 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3062 Node* cmpu = use->fast_out(i);
3063 if (cmpu->Opcode() == Op_CmpU) {
3064 push_bool_matching_case1b(worklist, cmpu);
3065 }
3066 }
3067 }
3068
3069 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3070 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3071 assert(cmpu->Opcode() == Op_CmpU, "must be");
3072 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3073 Node* bol = cmpu->fast_out(i);
3074 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3075 // Not a Bool with "<u"
3076 continue;
3077 }
3078 Node* andI = cmpu->in(1);
3079 Node* addI = cmpu->in(2);
3080 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3081 // Not "m & x" and "m + 1"
3082 continue;
3083 }
3084
3085 Node* m = addI->in(1);
3086 if (m == andI->in(1) || m == andI->in(2)) {
3087 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3088 push_if_not_bottom_type(worklist, bol);
3089 }
3090 }
3091 }
3092
3093 // 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'.
3094 // Seem PhiNode::Value().
3095 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3096 uint use_op = use->Opcode();
3097 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3098 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3099 if (phi != nullptr) {
3100 worklist.push(phi);
3101 }
3102 }
3103 }
3104
3105 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3106 // See LoadNode::Value().
3107 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3108 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3109 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3110 Node* loadp = use->fast_out(i);
3111 const Type* ut = loadp->bottom_type();
3112 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3113 worklist.push(loadp);
3114 }
3115 }
3116 }
3117 }
3118
3119 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3120 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3121 // value changed.
3122 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3123 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3124 uint use_op = use->Opcode();
3125 if (
3126 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3127 (parent_type != nullptr && parent_type->is_con()) ||
3128 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3129 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3130
3131 auto push_and_uses_to_worklist = [&](Node* n) {
3132 uint opc = n->Opcode();
3133 if (opc == Op_AndI || opc == Op_AndL) {
3134 push_if_not_bottom_type(worklist, n);
3135 }
3136 };
3137 auto is_boundary = [](Node* n) {
3138 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3139 };
3140 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3141 }
3142 }
3143
3144 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3145 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3146 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3147 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3148 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3149 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3150 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3151 Node* other_cmp_input = use->in(1);
3152 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3153 Node* cast_ii = other_cmp_input->fast_out(i);
3154 if (cast_ii->is_CastII()) {
3155 push_if_not_bottom_type(worklist, cast_ii);
3156 }
3157 }
3158 }
3159 }
3160
3161 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3162 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3163 push_if_not_bottom_type(worklist, use->unique_out());
3164 }
3165 }
3166
3167 //------------------------------do_transform-----------------------------------
3168 // Top level driver for the recursive transformer
3169 void PhaseCCP::do_transform() {
3170 // Correct leaves of new-space Nodes; they point to old-space.
3171 C->set_root( transform(C->root())->as_Root() );
3172 assert( C->top(), "missing TOP node" );
3173 assert( C->root(), "missing root" );
3174 }
3175
3176 //------------------------------transform--------------------------------------
3177 // Given a Node in old-space, clone him into new-space.
3178 // Convert any of his old-space children into new-space children.
3179 Node *PhaseCCP::transform( Node *n ) {
3180 assert(n->is_Root(), "traversal must start at root");
3181 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3182
3183 ResourceMark rm;
3184 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3185 Node_List node_map;
3186 // Pre-allocate to avoid frequent realloc
3187 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3188 // track all visited nodes, so that we can remove the complement
3189 Unique_Node_List useful;
3190
3191 if (KillPathsReachableByDeadTypeNode) {
3192 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3193 Node* type_node = _maybe_top_type_nodes.at(i);
3194 if (type(type_node) == Type::TOP) {
3195 ResourceMark rm;
3196 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3197 }
3198 }
3199 } else {
3200 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3201 }
3202
3203 // Initialize the traversal.
3204 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3205 // 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
3206 // 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
3207 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3208 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3209 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3210 // and all safepoints.
3211 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3212 Node* nn = _root_and_safepoints.at(i);
3213 Node* new_node = node_map[nn->_idx];
3214 assert(new_node == nullptr, "");
3215 new_node = transform_once(nn); // Check for constant
3216 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3217 transform_stack.push(new_node); // Process children of cloned node
3218 useful.push(new_node);
3219 }
3220
3221 while (transform_stack.is_nonempty()) {
3222 Node* clone = transform_stack.pop();
3223 uint cnt = clone->req();
3224 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3225 Node *input = clone->in(i);
3226 if( input != nullptr ) { // Ignore nulls
3227 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3228 if( new_input == nullptr ) {
3229 new_input = transform_once(input); // Check for constant
3230 node_map.map( input->_idx, new_input );// Flag as having been cloned
3231 transform_stack.push(new_input); // Process children of cloned node
3232 useful.push(new_input);
3233 }
3234 assert( new_input == clone->in(i), "insanity check");
3235 }
3236 }
3237 }
3238
3239 // The above transformation might lead to subgraphs becoming unreachable from the
3240 // bottom while still being reachable from the top. As a result, nodes in that
3241 // subgraph are not transformed and their bottom types are not updated, leading to
3242 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3243 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3244 // because their address type is inconsistent. Therefore, we aggressively remove
3245 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3246 // to remove them anyway.
3247 if (C->cached_top_node()) {
3248 useful.push(C->cached_top_node());
3249 }
3250 C->update_dead_node_list(useful);
3251 remove_useless_nodes(useful.member_set());
3252 _worklist.remove_useless_nodes(useful.member_set());
3253 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3254
3255 Node* new_root = node_map[n->_idx];
3256 assert(new_root->is_Root(), "transformed root node must be a root node");
3257 return new_root;
3258 }
3259
3260 //------------------------------transform_once---------------------------------
3261 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3262 Node *PhaseCCP::transform_once( Node *n ) {
3263 const Type *t = type(n);
3264 // Constant? Use constant Node instead
3265 if( t->singleton() ) {
3266 Node *nn = n; // Default is to return the original constant
3267 if( t == Type::TOP ) {
3268 // cache my top node on the Compile instance
3269 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3270 C->set_cached_top_node(ConNode::make(Type::TOP));
3271 set_type(C->top(), Type::TOP);
3272 }
3273 nn = C->top();
3274 }
3275 if( !n->is_Con() ) {
3276 if( t != Type::TOP ) {
3277 nn = makecon(t); // ConNode::make(t);
3278 NOT_PRODUCT( inc_constants(); )
3279 } else if( n->is_Region() ) { // Unreachable region
3280 // Note: nn == C->top()
3281 n->set_req(0, nullptr); // Cut selfreference
3282 bool progress = true;
3283 uint max = n->outcnt();
3284 DUIterator i;
3285 while (progress) {
3286 progress = false;
3287 // Eagerly remove dead phis to avoid phis copies creation.
3288 for (i = n->outs(); n->has_out(i); i++) {
3289 Node* m = n->out(i);
3290 if (m->is_Phi()) {
3291 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3292 replace_node(m, nn);
3293 if (max != n->outcnt()) {
3294 progress = true;
3295 i = n->refresh_out_pos(i);
3296 max = n->outcnt();
3297 }
3298 }
3299 }
3300 }
3301 }
3302 replace_node(n,nn); // Update DefUse edges for new constant
3303 }
3304 return nn;
3305 }
3306
3307 // If x is a TypeNode, capture any more-precise type permanently into Node
3308 if (t != n->bottom_type()) {
3309 hash_delete(n); // changing bottom type may force a rehash
3310 n->raise_bottom_type(t);
3311 _worklist.push(n); // n re-enters the hash table via the worklist
3312 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3313 }
3314
3315 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3316 switch( n->Opcode() ) {
3317 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3318 case Op_CallDynamicJava:
3319 case Op_FastLock: // Revisit FastLocks for lock coarsening
3320 case Op_If:
3321 case Op_CountedLoopEnd:
3322 case Op_Region:
3323 case Op_Loop:
3324 case Op_CountedLoop:
3325 case Op_Conv2B:
3326 case Op_Opaque1:
3327 _worklist.push(n);
3328 break;
3329 default:
3330 break;
3331 }
3332
3333 return n;
3334 }
3335
3336 //---------------------------------saturate------------------------------------
3337 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3338 const Type* limit_type) const {
3339 const Type* wide_type = new_type->widen(old_type, limit_type);
3340 if (wide_type != new_type) { // did we widen?
3341 // If so, we may have widened beyond the limit type. Clip it back down.
3342 new_type = wide_type->filter(limit_type);
3343 }
3344 return new_type;
3345 }
3346
3347 //------------------------------print_statistics-------------------------------
3348 #ifndef PRODUCT
3349 void PhaseCCP::print_statistics() {
3350 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3351 }
3352 #endif
3353
3354
3355 //=============================================================================
3356 #ifndef PRODUCT
3357 uint PhasePeephole::_total_peepholes = 0;
3358 #endif
3359 //------------------------------PhasePeephole----------------------------------
3360 // Conditional Constant Propagation, ala Wegman & Zadeck
3361 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3362 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3363 NOT_PRODUCT( clear_peepholes(); )
3364 }
3365
3366 #ifndef PRODUCT
3367 //------------------------------~PhasePeephole---------------------------------
3368 PhasePeephole::~PhasePeephole() {
3369 _total_peepholes += count_peepholes();
3370 }
3371 #endif
3372
3373 //------------------------------transform--------------------------------------
3374 Node *PhasePeephole::transform( Node *n ) {
3375 ShouldNotCallThis();
3376 return nullptr;
3377 }
3378
3379 //------------------------------do_transform-----------------------------------
3380 void PhasePeephole::do_transform() {
3381 bool method_name_not_printed = true;
3382
3383 // Examine each basic block
3384 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3385 Block* block = _cfg.get_block(block_number);
3386 bool block_not_printed = true;
3387
3388 for (bool progress = true; progress;) {
3389 progress = false;
3390 // block->end_idx() not valid after PhaseRegAlloc
3391 uint end_index = block->number_of_nodes();
3392 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3393 Node *n = block->get_node(instruction_index);
3394 if( n->is_Mach() ) {
3395 MachNode *m = n->as_Mach();
3396 // check for peephole opportunities
3397 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3398 if( result != -1 ) {
3399 #ifndef PRODUCT
3400 if( PrintOptoPeephole ) {
3401 // Print method, first time only
3402 if( C->method() && method_name_not_printed ) {
3403 C->method()->print_short_name(); tty->cr();
3404 method_name_not_printed = false;
3405 }
3406 // Print this block
3407 if( Verbose && block_not_printed) {
3408 tty->print_cr("in block");
3409 block->dump();
3410 block_not_printed = false;
3411 }
3412 // Print the peephole number
3413 tty->print_cr("peephole number: %d", result);
3414 }
3415 inc_peepholes();
3416 #endif
3417 // Set progress, start again
3418 progress = true;
3419 break;
3420 }
3421 }
3422 }
3423 }
3424 }
3425 }
3426
3427 //------------------------------print_statistics-------------------------------
3428 #ifndef PRODUCT
3429 void PhasePeephole::print_statistics() {
3430 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3431 }
3432 #endif
3433
3434
3435 //=============================================================================
3436 //------------------------------set_req_X--------------------------------------
3437 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3438 assert( is_not_dead(n), "can not use dead node");
3439 #ifdef ASSERT
3440 if (igvn->hash_find(this) == this) {
3441 tty->print_cr("Need to remove from hash before changing edges");
3442 this->dump(1);
3443 tty->print_cr("Set at i = %d", i);
3444 n->dump();
3445 assert(false, "Need to remove from hash before changing edges");
3446 }
3447 #endif
3448 Node *old = in(i);
3449 set_req(i, n);
3450
3451 // old goes dead?
3452 if( old ) {
3453 switch (old->outcnt()) {
3454 case 0:
3455 // Put into the worklist to kill later. We do not kill it now because the
3456 // recursive kill will delete the current node (this) if dead-loop exists
3457 if (!old->is_top())
3458 igvn->_worklist.push( old );
3459 break;
3460 case 1:
3461 if( old->is_Store() || old->has_special_unique_user() )
3462 igvn->add_users_to_worklist( old );
3463 break;
3464 case 2:
3465 if( old->is_Store() )
3466 igvn->add_users_to_worklist( old );
3467 if( old->Opcode() == Op_Region )
3468 igvn->_worklist.push(old);
3469 break;
3470 case 3:
3471 if( old->Opcode() == Op_Region ) {
3472 igvn->_worklist.push(old);
3473 igvn->add_users_to_worklist( old );
3474 }
3475 break;
3476 default:
3477 break;
3478 }
3479 }
3480 }
3481
3482 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3483 PhaseIterGVN* igvn = gvn->is_IterGVN();
3484 if (igvn == nullptr) {
3485 set_req(i, n);
3486 return;
3487 }
3488 set_req_X(i, n, igvn);
3489 }
3490
3491 //-------------------------------replace_by-----------------------------------
3492 // Using def-use info, replace one node for another. Follow the def-use info
3493 // to all users of the OLD node. Then make all uses point to the NEW node.
3494 void Node::replace_by(Node *new_node) {
3495 assert(!is_top(), "top node has no DU info");
3496 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3497 Node* use = last_out(i);
3498 uint uses_found = 0;
3499 for (uint j = 0; j < use->len(); j++) {
3500 if (use->in(j) == this) {
3501 if (j < use->req())
3502 use->set_req(j, new_node);
3503 else use->set_prec(j, new_node);
3504 uses_found++;
3505 }
3506 }
3507 i -= uses_found; // we deleted 1 or more copies of this edge
3508 }
3509 }
3510
3511 //=============================================================================
3512 //-----------------------------------------------------------------------------
3513 void Type_Array::grow( uint i ) {
3514 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3515 if( !_max ) {
3516 _max = 1;
3517 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3518 _types[0] = nullptr;
3519 }
3520 uint old = _max;
3521 _max = next_power_of_2(i);
3522 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3523 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3524 }
3525
3526 //------------------------------dump-------------------------------------------
3527 #ifndef PRODUCT
3528 void Type_Array::dump() const {
3529 uint max = Size();
3530 for( uint i = 0; i < max; i++ ) {
3531 if( _types[i] != nullptr ) {
3532 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3533 }
3534 }
3535 }
3536 #endif