1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE 255
46
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 DEBUG_ONLY( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 DEBUG_ONLY( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
368 // Implementation requires an edge from root to each SafePointNode
369 // at a backward branch. Inserted in add_safepoint().
370
371 // Identify nodes that are reachable from below, useful.
372 C->identify_useful_nodes(_useful);
373 // Update dead node list
374 C->update_dead_node_list(_useful);
375
376 // Remove all useless nodes from PhaseValues' recorded types
377 // Must be done before disconnecting nodes to preserve hash-table-invariant
378 gvn->remove_useless_nodes(_useful.member_set());
379
380 // Remove all useless nodes from future worklist
381 worklist.remove_useless_nodes(_useful.member_set());
382
383 // Disconnect 'useless' nodes that are adjacent to useful nodes
384 C->disconnect_useless_nodes(_useful, worklist);
385 }
386
387 //=============================================================================
388 //------------------------------PhaseRenumberLive------------------------------
389 // First, remove useless nodes (equivalent to identifying live nodes).
390 // Then, renumber live nodes.
391 //
392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
395 // value in the range [0, x).
396 //
397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
399 //
400 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
402 // be processed by IGVN after removal of the useless nodes.
403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
404 // the node's type. The mapping is updated to use the new node IDs as well. We
405 // create a new map, and swap it with the old one.
406 //
407 // Other data structures used by the compiler are not updated. The hash table for value
408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
409 // computing the hash values is not based on node IDs.
410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
411 Unique_Node_List& worklist,
412 PhaseNumber phase_num) :
413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
414 _new_type_array(C->comp_arena()),
415 _old2new_map(C->unique(), C->unique(), -1),
416 _is_pass_finished(false),
417 _live_node_count(C->live_nodes())
418 {
419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
421 assert(_delayed.size() == 0, "should be empty");
422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
423 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
424
425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
426 if (old_node_note_array != nullptr) {
427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
428 new_size = MAX2(8, new_size);
429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
430 C->grow_node_notes(C->node_note_array(), new_size);
431 }
432
433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
434
435 // Iterate over the set of live nodes.
436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
437 Node* n = _useful.at(current_idx);
438
439 const Type* type = gvn->type_or_null(n);
440 _new_type_array.map(current_idx, type);
441
442 assert(_old2new_map.at(n->_idx) == -1, "already seen");
443 _old2new_map.at_put(n->_idx, current_idx);
444
445 if (old_node_note_array != nullptr) {
446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
447 C->set_node_notes_at(current_idx, nn);
448 }
449
450 n->set_idx(current_idx); // Update node ID.
451
452 if (update_embedded_ids(n) < 0) {
453 _delayed.push(n); // has embedded IDs; handle later
454 }
455 }
456
457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
458 worklist.recompute_idx_set();
459
460 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
461
462 _is_pass_finished = true; // pass finished; safe to process delayed updates
463
464 while (_delayed.size() > 0) {
465 Node* n = _delayed.pop();
466 int no_of_updates = update_embedded_ids(n);
467 assert(no_of_updates > 0, "should be updated");
468 }
469
470 // Replace the compiler's type information with the updated type information.
471 gvn->types().swap(_new_type_array);
472
473 // Update the unique node count of the compilation to the number of currently live nodes.
474 C->set_unique(_live_node_count);
475
476 // Set the dead node count to 0 and reset dead node list.
477 C->reset_dead_node_list();
478 }
479
480 int PhaseRenumberLive::new_index(int old_idx) {
481 assert(_is_pass_finished, "not finished");
482 if (_old2new_map.at(old_idx) == -1) { // absent
483 // Allocate a placeholder to preserve uniqueness
484 _old2new_map.at_put(old_idx, _live_node_count);
485 _live_node_count++;
486 }
487 return _old2new_map.at(old_idx);
488 }
489
490 int PhaseRenumberLive::update_embedded_ids(Node* n) {
491 int no_of_updates = 0;
492 if (n->is_Phi()) {
493 PhiNode* phi = n->as_Phi();
494 if (phi->_inst_id != -1) {
495 if (!_is_pass_finished) {
496 return -1; // delay
497 }
498 int new_idx = new_index(phi->_inst_id);
499 assert(new_idx != -1, "");
500 phi->_inst_id = new_idx;
501 no_of_updates++;
502 }
503 if (phi->_inst_mem_id != -1) {
504 if (!_is_pass_finished) {
505 return -1; // delay
506 }
507 int new_idx = new_index(phi->_inst_mem_id);
508 assert(new_idx != -1, "");
509 phi->_inst_mem_id = new_idx;
510 no_of_updates++;
511 }
512 }
513
514 const Type* type = _new_type_array.fast_lookup(n->_idx);
515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
516 if (!_is_pass_finished) {
517 return -1; // delay
518 }
519 int old_idx = type->is_oopptr()->instance_id();
520 int new_idx = new_index(old_idx);
521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
522 _new_type_array.map(n->_idx, new_type);
523 no_of_updates++;
524 }
525
526 return no_of_updates;
527 }
528
529 void PhaseValues::init_con_caches() {
530 memset(_icons,0,sizeof(_icons));
531 memset(_lcons,0,sizeof(_lcons));
532 memset(_zcons,0,sizeof(_zcons));
533 }
534
535 //--------------------------------find_int_type--------------------------------
536 const TypeInt* PhaseValues::find_int_type(Node* n) {
537 if (n == nullptr) return nullptr;
538 // Call type_or_null(n) to determine node's type since we might be in
539 // parse phase and call n->Value() may return wrong type.
540 // (For example, a phi node at the beginning of loop parsing is not ready.)
541 const Type* t = type_or_null(n);
542 if (t == nullptr) return nullptr;
543 return t->isa_int();
544 }
545
546
547 //-------------------------------find_long_type--------------------------------
548 const TypeLong* PhaseValues::find_long_type(Node* n) {
549 if (n == nullptr) return nullptr;
550 // (See comment above on type_or_null.)
551 const Type* t = type_or_null(n);
552 if (t == nullptr) return nullptr;
553 return t->isa_long();
554 }
555
556 //------------------------------~PhaseValues-----------------------------------
557 #ifndef PRODUCT
558 PhaseValues::~PhaseValues() {
559 // Statistics for NodeHash
560 _table.dump();
561 // Statistics for value progress and efficiency
562 if( PrintCompilation && Verbose && WizardMode ) {
563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
565 if( made_transforms() != 0 ) {
566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
567 } else {
568 tty->cr();
569 }
570 }
571 }
572 #endif
573
574 //------------------------------makecon----------------------------------------
575 ConNode* PhaseValues::makecon(const Type* t) {
576 assert(t->singleton(), "must be a constant");
577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
578 switch (t->base()) { // fast paths
579 case Type::Half:
580 case Type::Top: return (ConNode*) C->top();
581 case Type::Int: return intcon( t->is_int()->get_con() );
582 case Type::Long: return longcon( t->is_long()->get_con() );
583 default: break;
584 }
585 if (t->is_zero_type())
586 return zerocon(t->basic_type());
587 return uncached_makecon(t);
588 }
589
590 //--------------------------uncached_makecon-----------------------------------
591 // Make an idealized constant - one of ConINode, ConPNode, etc.
592 ConNode* PhaseValues::uncached_makecon(const Type *t) {
593 assert(t->singleton(), "must be a constant");
594 ConNode* x = ConNode::make(t);
595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
596 if (k == nullptr) {
597 set_type(x, t); // Missed, provide type mapping
598 GrowableArray<Node_Notes*>* nna = C->node_note_array();
599 if (nna != nullptr) {
600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
601 loc->clear(); // do not put debug info on constants
602 }
603 } else {
604 x->destruct(this); // Hit, destroy duplicate constant
605 x = k; // use existing constant
606 }
607 return x;
608 }
609
610 //------------------------------intcon-----------------------------------------
611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
612 ConINode* PhaseValues::intcon(jint i) {
613 // Small integer? Check cache! Check that cached node is not dead
614 if (i >= _icon_min && i <= _icon_max) {
615 ConINode* icon = _icons[i-_icon_min];
616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
617 return icon;
618 }
619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
620 assert(icon->is_Con(), "");
621 if (i >= _icon_min && i <= _icon_max)
622 _icons[i-_icon_min] = icon; // Cache small integers
623 return icon;
624 }
625
626 //------------------------------longcon----------------------------------------
627 // Fast long constant.
628 ConLNode* PhaseValues::longcon(jlong l) {
629 // Small integer? Check cache! Check that cached node is not dead
630 if (l >= _lcon_min && l <= _lcon_max) {
631 ConLNode* lcon = _lcons[l-_lcon_min];
632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
633 return lcon;
634 }
635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
636 assert(lcon->is_Con(), "");
637 if (l >= _lcon_min && l <= _lcon_max)
638 _lcons[l-_lcon_min] = lcon; // Cache small integers
639 return lcon;
640 }
641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
642 if (bt == T_INT) {
643 return intcon(checked_cast<jint>(l));
644 }
645 assert(bt == T_LONG, "not an integer");
646 return longcon(l);
647 }
648
649
650 //------------------------------zerocon-----------------------------------------
651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
652 ConNode* PhaseValues::zerocon(BasicType bt) {
653 assert((uint)bt <= _zcon_max, "domain check");
654 ConNode* zcon = _zcons[bt];
655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
656 return zcon;
657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
658 _zcons[bt] = zcon;
659 return zcon;
660 }
661
662
663
664 //=============================================================================
665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
666 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
667 if (i == nullptr) {
668 i = k->Ideal(this, can_reshape);
669 }
670 return i;
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 _iterGVN = true;
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 _iterGVN = true;
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 bool failure = false;
1094 // BFS all nodes, starting at root
1095 worklist.push(C->root());
1096 for (uint j = 0; j < worklist.size(); ++j) {
1097 Node* n = worklist.at(j);
1098 if (is_verify_Value()) { failure |= verify_Value_for(n); }
1099 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, false); }
1100 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, true); }
1101 if (is_verify_Identity()) { failure |= verify_Identity_for(n); }
1102 if (is_verify_invariants()) { failure |= verify_node_invariants_for(n); }
1103 // traverse all inputs and outputs
1104 for (uint i = 0; i < n->req(); i++) {
1105 if (n->in(i) != nullptr) {
1106 worklist.push(n->in(i));
1107 }
1108 }
1109 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1110 worklist.push(n->fast_out(i));
1111 }
1112 }
1113 // If we get this assert, check why the reported nodes were not processed again in IGVN.
1114 // We should either make sure that these nodes are properly added back to the IGVN worklist
1115 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception
1116 // in the verification code above if that is not possible for some reason (like Load nodes).
1117 assert(!failure, "Missed optimization opportunity/broken graph in PhaseIterGVN");
1118 }
1119
1120 verify_empty_worklist(nullptr);
1121 }
1122
1123 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1124 // Verify that the igvn worklist is empty. If no optimization happened, then
1125 // nothing needs to be on the worklist.
1126 if (_worklist.size() == 0) { return; }
1127
1128 stringStream ss; // Print as a block without tty lock.
1129 for (uint j = 0; j < _worklist.size(); j++) {
1130 Node* n = _worklist.at(j);
1131 ss.print("igvn.worklist[%d] ", j);
1132 n->dump("\n", false, &ss);
1133 }
1134 if (_worklist.size() != 0 && node != nullptr) {
1135 ss.print_cr("Previously optimized:");
1136 node->dump("\n", false, &ss);
1137 }
1138 tty->print_cr("%s", ss.as_string());
1139 assert(false, "igvn worklist must still be empty after verify");
1140 }
1141
1142 // Check that type(n) == n->Value(), return true if we have a failure.
1143 // We have a list of exceptions, see detailed comments in code.
1144 // (1) Integer "widen" changes, but the range is the same.
1145 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1146 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1147 bool PhaseIterGVN::verify_Value_for(Node* n, bool strict) {
1148 // If we assert inside type(n), because the type is still a null, then maybe
1149 // the node never went through gvn.transform, which would be a bug.
1150 const Type* told = type(n);
1151 const Type* tnew = n->Value(this);
1152 if (told == tnew) {
1153 return false;
1154 }
1155 // Exception (1)
1156 // Integer "widen" changes, but range is the same.
1157 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1158 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1159 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1160 if (t0->lo_as_long() == t1->lo_as_long() &&
1161 t0->hi_as_long() == t1->hi_as_long()) {
1162 return false; // ignore integer widen
1163 }
1164 }
1165 // Exception (2)
1166 // LoadNode performs deep traversals. Load is not notified for changes far away.
1167 if (!strict && n->is_Load() && !told->singleton()) {
1168 // MemNode::can_see_stored_value looks up through many memory nodes,
1169 // which means we would need to notify modifications from far up in
1170 // the inputs all the way down to the LoadNode. We don't do that.
1171 return false;
1172 }
1173 // Exception (3)
1174 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1175 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1176 // SubNode::Value
1177 // CmpPNode::sub
1178 // MemNode::detect_ptr_independence
1179 // MemNode::all_controls_dominate
1180 // We find all controls of a pointer load, and see if they dominate the control of
1181 // an allocation. If they all dominate, we know the allocation is after (independent)
1182 // of the pointer load, and we can say the pointers are different. For this we call
1183 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1184 // control sub of the allocation. The problems is that sometimes dominates answers
1185 // false conservatively, and later it can determine that it is indeed true. Loops with
1186 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1187 // so the traversal becomes more powerful. This is difficult to remidy, we would have
1188 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1189 // after loop-opts, so that should take care of many of these cases.
1190 return false;
1191 }
1192
1193 stringStream ss; // Print as a block without tty lock.
1194 ss.cr();
1195 ss.print_cr("Missed Value optimization:");
1196 n->dump_bfs(1, nullptr, "", &ss);
1197 ss.print_cr("Current type:");
1198 told->dump_on(&ss);
1199 ss.cr();
1200 ss.print_cr("Optimized type:");
1201 tnew->dump_on(&ss);
1202 ss.cr();
1203 tty->print_cr("%s", ss.as_string());
1204 return true;
1205 }
1206
1207 // Check that all Ideal optimizations that could be done were done.
1208 // Returns true if it found missed optimization opportunities and
1209 // false otherwise (no missed optimization, or skipped verification).
1210 bool PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1211 // First, we check a list of exceptions, where we skip verification,
1212 // because there are known cases where Ideal can optimize after IGVN.
1213 // Some may be expected and cannot be fixed, and others should be fixed.
1214 switch (n->Opcode()) {
1215 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1216 // (see "Range-Check scan limit"). So, it is possible that something
1217 // is optimized in that input subgraph, and the RangeCheck was not
1218 // added to the worklist because it would be too expensive to walk
1219 // down the graph for 1000 nodes and put all on the worklist.
1220 //
1221 // Found with:
1222 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1223 case Op_RangeCheck:
1224 return false;
1225
1226 // IfNode::Ideal does:
1227 // Node* prev_dom = search_identical(dist, igvn);
1228 // which means we seach up the CFG, traversing at most up to a distance.
1229 // If anything happens rather far away from the If, we may not put the If
1230 // back on the worklist.
1231 //
1232 // Found with:
1233 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1234 case Op_If:
1235 return false;
1236
1237 // IfNode::simple_subsuming
1238 // Looks for dominating test that subsumes the current test.
1239 // Notification could be difficult because of larger distance.
1240 //
1241 // Found with:
1242 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1243 // -XX:VerifyIterativeGVN=1110
1244 case Op_CountedLoopEnd:
1245 return false;
1246
1247 // LongCountedLoopEndNode::Ideal
1248 // Probably same issue as above.
1249 //
1250 // Found with:
1251 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1252 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1253 case Op_LongCountedLoopEnd:
1254 return false;
1255
1256 // RegionNode::Ideal does "Skip around the useless IF diamond".
1257 // 245 IfTrue === 244
1258 // 258 If === 245 257
1259 // 259 IfTrue === 258 [[ 263 ]]
1260 // 260 IfFalse === 258 [[ 263 ]]
1261 // 263 Region === 263 260 259 [[ 263 268 ]]
1262 // to
1263 // 245 IfTrue === 244
1264 // 263 Region === 263 245 _ [[ 263 268 ]]
1265 //
1266 // "Useless" means that there is no code in either branch of the If.
1267 // I found a case where this was not done yet during IGVN.
1268 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1269 //
1270 // Found with:
1271 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1272 case Op_Region:
1273 return false;
1274
1275 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1276 // that smaller idx are first. Tracking it back, it led me to
1277 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1278 //
1279 // Example:
1280 // Before PhaseIdealLoop::remix_address_expressions
1281 // 154 AddI === _ 12 144
1282 // After PhaseIdealLoop::remix_address_expressions
1283 // 154 AddI === _ 144 12
1284 // After AddNode::Ideal
1285 // 154 AddI === _ 12 144
1286 //
1287 // I suspect that the node should be added to the IGVN worklist after
1288 // PhaseIdealLoop::remix_address_expressions
1289 //
1290 // This is the only case I looked at, there may be others. Found like this:
1291 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1292 //
1293 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1294 //
1295 // Note: currently all of these fail also for other reasons, for example
1296 // because of "commute" doing the reordering with the phi below. Once
1297 // that is resolved, we can come back to this issue here.
1298 //
1299 // case Op_AddD:
1300 // case Op_AddI:
1301 // case Op_AddL:
1302 // case Op_AddF:
1303 // case Op_MulI:
1304 // case Op_MulL:
1305 // case Op_MulF:
1306 // case Op_MulD:
1307 // if (n->in(1)->_idx > n->in(2)->_idx) {
1308 // // Expect "commute" to revert this case.
1309 // return false;
1310 // }
1311 // break; // keep verifying
1312
1313 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1314 // Check for tight loop increments: Loop-phi of Add of loop-phi
1315 // It wants to take the phi into in(1):
1316 // 471 Phi === 435 38 390
1317 // 390 AddF === _ 471 391
1318 //
1319 // Other Associative operators are also affected equally.
1320 //
1321 // Investigate why this does not happen earlier during IGVN.
1322 //
1323 // Found with:
1324 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1325 // -XX:VerifyIterativeGVN=1110
1326 case Op_AddD:
1327 //case Op_AddI: // Also affected for other reasons, see case further down.
1328 //case Op_AddL: // Also affected for other reasons, see case further down.
1329 case Op_AddF:
1330 case Op_MulI:
1331 case Op_MulL:
1332 case Op_MulF:
1333 case Op_MulD:
1334 case Op_MinF:
1335 case Op_MinD:
1336 case Op_MaxF:
1337 case Op_MaxD:
1338 // XorINode::Ideal
1339 // Found with:
1340 // compiler/intrinsics/chacha/TestChaCha20.java
1341 // -XX:VerifyIterativeGVN=1110
1342 case Op_XorI:
1343 case Op_XorL:
1344 // It seems we may have similar issues with the HF cases.
1345 // Found with aarch64:
1346 // compiler/vectorization/TestFloat16VectorOperations.java
1347 // -XX:VerifyIterativeGVN=1110
1348 case Op_AddHF:
1349 case Op_MulHF:
1350 case Op_MaxHF:
1351 case Op_MinHF:
1352 return false;
1353
1354 // In MulNode::Ideal the edges can be swapped to help value numbering:
1355 //
1356 // // We are OK if right is a constant, or right is a load and
1357 // // left is a non-constant.
1358 // if( !(t2->singleton() ||
1359 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1360 // if( t1->singleton() || // Left input is a constant?
1361 // // Otherwise, sort inputs (commutativity) to help value numbering.
1362 // (in(1)->_idx > in(2)->_idx) ) {
1363 // swap_edges(1, 2);
1364 //
1365 // Why was this not done earlier during IGVN?
1366 //
1367 // Found with:
1368 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1369 // -XX:VerifyIterativeGVN=1110
1370 case Op_AndI:
1371 // Same for AndL.
1372 // Found with:
1373 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1374 // -XX:VerifyIterativeGVN=1110
1375 case Op_AndL:
1376 return false;
1377
1378 // SubLNode::Ideal does transform like:
1379 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1380 //
1381 // In IGVN before verification:
1382 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1383 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1384 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1385 // 8016 ConL === 0 [[ 8383 ]] #long:0
1386 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1387 //
1388 // And then in verification:
1389 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1390 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1391 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1392 // 8016 ConL === 0 [[ 8383 ]] #long:0
1393 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1394 //
1395 // So the form changed from:
1396 // c1 - (y + [8423 ConvI2L])
1397 // to
1398 // c1 - (y + -2)
1399 // but the SubL was not added to the IGVN worklist. Investigate why.
1400 // There could be other issues too.
1401 //
1402 // There seems to be a related AddL IGVN optimization that triggers
1403 // the same SubL optimization, so investigate that too.
1404 //
1405 // Found with:
1406 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1407 case Op_SubL:
1408 return false;
1409
1410 // SubINode::Ideal does
1411 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1412 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1413 //
1414 // Investigate why this does not yet happen during IGVN.
1415 //
1416 // Found with:
1417 // test/hotspot/jtreg/compiler/c2/IVTest.java
1418 // -XX:VerifyIterativeGVN=1110
1419 case Op_SubI:
1420 return false;
1421
1422 // AddNode::IdealIL does transform like:
1423 // Convert x + (con - y) into "(x - y) + con"
1424 //
1425 // In IGVN before verification:
1426 // 8382 ConvI2L
1427 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1428 // 8383 SubL === _ 8381 8382
1429 // 8168 ConvI2L
1430 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1431 //
1432 // And then in verification:
1433 // 8424 AddL
1434 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1435 // 8383 SubL === _ 8016 8424
1436 // 8168 ConvI2L
1437 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1438 //
1439 // So the form changed from:
1440 // x + (ConvI2L(0) - [8382 ConvI2L])
1441 // to
1442 // x + (0 - [8424 AddL])
1443 // but the AddL was not added to the IGVN worklist. Investigate why.
1444 // There could be other issues, too. For example with "commute", see above.
1445 //
1446 // Found with:
1447 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1448 case Op_AddL:
1449 return false;
1450
1451 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1452 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1453 // record_for_cleanup(cmp, phase);
1454 // This verification code in the Ideal code creates new nodes, and checks
1455 // if they fold in unexpected ways. This means some nodes are created and
1456 // added to the worklist, even if the SubTypeCheck is not optimized. This
1457 // goes agains the assumption of the verification here, which assumes that
1458 // if the node is not optimized, then no new nodes should be created, and
1459 // also no nodes should be added to the worklist.
1460 // I see two options:
1461 // 1) forbid what verify_helper does, because for each Ideal call it
1462 // uses memory and that is suboptimal. But it is not clear how that
1463 // verification can be done otherwise.
1464 // 2) Special case the verification here. Probably the new nodes that
1465 // were just created are dead, i.e. they are not connected down to
1466 // root. We could verify that, and remove those nodes from the graph
1467 // by setting all their inputs to nullptr. And of course we would
1468 // have to remove those nodes from the worklist.
1469 // Maybe there are other options too, I did not dig much deeper yet.
1470 //
1471 // Found with:
1472 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1473 case Op_SubTypeCheck:
1474 return false;
1475
1476 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1477 // to other nodes: Conv, Add, Sub, Mul, And ...
1478 //
1479 // 107 ConI === 0 [[ ... ]] #int:2
1480 // 84 LoadRange === _ 7 83
1481 // 50 ConI === 0 [[ ... ]] #int:0
1482 // 549 LoopLimit === _ 50 84 107
1483 //
1484 // I stepped backward, to see how the node was generated, and I found that it was
1485 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1486 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1487 // that, and it seems we want to skip lowering it until after loop-opts, but never
1488 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1489 // could be other issues too.
1490 //
1491 // Fond with:
1492 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1493 case Op_LoopLimit:
1494 return false;
1495
1496 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1497 // "This optimization tries to find two or more inputs of phi with the same constant
1498 // value. It then splits them into a separate Phi, and according Region."
1499 //
1500 // Example:
1501 // 130 DecodeN === _ 129
1502 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1503 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1504 //
1505 // turns into:
1506 //
1507 // 50 ConP === 0 [[ 99 91 18 ]] #null
1508 // 130 DecodeN === _ 129 [[ 18 ]]
1509 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1510 //
1511 // We would have to investigate why this optimization does not happen during IGVN.
1512 // There could also be other issues - I did not investigate further yet.
1513 //
1514 // Found with:
1515 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1516 case Op_Phi:
1517 return false;
1518
1519 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1520 // For examle "The allocated object does not escape".
1521 //
1522 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1523 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1524 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1525 // worklist when this change happens?
1526 //
1527 // Found with:
1528 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1529 case Op_MemBarStoreStore:
1530 return false;
1531
1532 // ConvI2LNode::Ideal converts
1533 // 648 AddI === _ 583 645 [[ 661 ]]
1534 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1535 // into
1536 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1537 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1538 // 773 AddL === _ 771 772 [[ ]]
1539 //
1540 // We have to investigate why this does not happen during IGVN in this case.
1541 // There could also be other issues - I did not investigate further yet.
1542 //
1543 // Found with:
1544 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1545 case Op_ConvI2L:
1546 return false;
1547
1548 // AddNode::IdealIL can do this transform (and similar other ones):
1549 // Convert "a*b+a*c into a*(b+c)
1550 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1551 // during IGVN? There was a mutation for one of the MulI, and only
1552 // after that the pattern was as needed for the optimization. The MulI
1553 // was added to the IGVN worklist, but not the AddI. This probably
1554 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1555 //
1556 // Found with:
1557 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1558 // -XX:VerifyIterativeGVN=1110
1559 case Op_AddI:
1560 return false;
1561
1562 // ArrayCopyNode::Ideal
1563 // calls ArrayCopyNode::prepare_array_copy
1564 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1565 // is not expected, and we create a range int:0..-1
1566 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1567 // note: the type is already empty!
1568 // calls PhaseIterGVN::transform
1569 // calls PhaseIterGVN::transform_old
1570 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1571 // calls Unique_Node_List::push -> pushes TOP to worklist
1572 //
1573 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1574 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1575 //
1576 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1577 // return nullptr, indicating nothing happened?
1578 // Is it smart to do transform in Compile::constrained_convI2L, and then
1579 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1580 // Should we just allow TOP to land on the worklist, as an exception?
1581 //
1582 // Found with:
1583 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1584 // -XX:VerifyIterativeGVN=1110
1585 case Op_ArrayCopy:
1586 return false;
1587
1588 // CastLLNode::Ideal
1589 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1590 //
1591 // Could be a notification issue, where updates inputs of CastLL do not notify
1592 // down through SubL to CastLL.
1593 //
1594 // Found With:
1595 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1596 // -XX:VerifyIterativeGVN=1110
1597 case Op_CastLL:
1598 return false;
1599
1600 // Similar case happens to CastII
1601 //
1602 // Found With:
1603 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1604 // -XX:VerifyIterativeGVN=1110
1605 case Op_CastII:
1606 return false;
1607
1608 // MaxLNode::Ideal
1609 // calls AddNode::Ideal
1610 // calls commute -> decides to swap edges
1611 //
1612 // Another notification issue, because we check inputs of inputs?
1613 // MaxL -> Phi -> Loop
1614 // MaxL -> Phi -> MaxL
1615 //
1616 // Found with:
1617 // compiler/c2/irTests/TestIfMinMax.java
1618 // -XX:VerifyIterativeGVN=1110
1619 case Op_MaxL:
1620 case Op_MinL:
1621 return false;
1622
1623 // OrINode::Ideal
1624 // calls AddNode::Ideal
1625 // calls commute -> left is Load, right not -> commute.
1626 //
1627 // Not sure why notification does not work here, seems like
1628 // the depth is only 1, so it should work. Needs investigation.
1629 //
1630 // Found with:
1631 // compiler/codegen/TestCharVect2.java#id0
1632 // -XX:VerifyIterativeGVN=1110
1633 case Op_OrI:
1634 case Op_OrL:
1635 return false;
1636
1637 // Bool -> constant folded to 1.
1638 // Issue with notification?
1639 //
1640 // Found with:
1641 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1642 // -XX:VerifyIterativeGVN=1110
1643 case Op_Bool:
1644 return false;
1645
1646 // LShiftLNode::Ideal
1647 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1648 // Probably a notification issue.
1649 //
1650 // Found with:
1651 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1652 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1653 case Op_LShiftL:
1654 return false;
1655
1656 // LShiftINode::Ideal
1657 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1658 // Could be issue with notification of inputs of inputs
1659 //
1660 // Side-note: should cases like these not be shared between
1661 // LShiftI and LShiftL?
1662 //
1663 // Found with:
1664 // compiler/escapeAnalysis/Test6689060.java
1665 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1666 case Op_LShiftI:
1667 return false;
1668
1669 // AddPNode::Ideal seems to do set_req without removing lock first.
1670 // Found with various vector tests tier1-tier3.
1671 case Op_AddP:
1672 return false;
1673
1674 // StrIndexOfNode::Ideal
1675 // Found in tier1-3.
1676 case Op_StrIndexOf:
1677 case Op_StrIndexOfChar:
1678 return false;
1679
1680 // StrEqualsNode::Identity
1681 //
1682 // Found (linux x64 only?) with:
1683 // serviceability/sa/ClhsdbThreadContext.java
1684 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1685 // Note: The -XX:LockingMode option is not available anymore.
1686 case Op_StrEquals:
1687 return false;
1688
1689 // AryEqNode::Ideal
1690 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1691 //
1692 // Found with:
1693 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1694 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1695 case Op_AryEq:
1696 return false;
1697
1698 // MergeMemNode::Ideal
1699 // Found in tier1-3. Did not investigate further yet.
1700 case Op_MergeMem:
1701 return false;
1702
1703 // URShiftINode::Ideal
1704 // Found in tier1-3. Did not investigate further yet.
1705 case Op_URShiftI:
1706 return false;
1707
1708 // CMoveINode::Ideal
1709 // Found in tier1-3. Did not investigate further yet.
1710 case Op_CMoveI:
1711 return false;
1712
1713 // CmpPNode::Ideal calls isa_const_java_mirror
1714 // and generates new constant nodes, even if no progress is made.
1715 // We can probably rewrite this so that only types are generated.
1716 // It seems that object types are not hashed, we could investigate
1717 // if that is an option as well.
1718 //
1719 // Found with:
1720 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1721 case Op_CmpP:
1722 return false;
1723
1724 // MinINode::Ideal
1725 // Did not investigate, but there are some patterns that might
1726 // need more notification.
1727 case Op_MinI:
1728 case Op_MaxI: // preemptively removed it as well.
1729 return false;
1730 }
1731
1732 if (n->is_Load()) {
1733 // LoadNode::Ideal uses tries to find an earlier memory state, and
1734 // checks can_see_stored_value for it.
1735 //
1736 // Investigate why this was not already done during IGVN.
1737 // A similar issue happens with Identity.
1738 //
1739 // There seem to be other cases where loads go up some steps, like
1740 // LoadNode::Ideal going up 10x steps to find dominating load.
1741 //
1742 // Found with:
1743 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1744 // -XX:VerifyIterativeGVN=1110
1745 return false;
1746 }
1747
1748 if (n->is_Store()) {
1749 // StoreNode::Ideal can do this:
1750 // // Capture an unaliased, unconditional, simple store into an initializer.
1751 // // Or, if it is independent of the allocation, hoist it above the allocation.
1752 // That replaces the Store with a MergeMem.
1753 //
1754 // We have to investigate why this does not happen during IGVN in this case.
1755 // There could also be other issues - I did not investigate further yet.
1756 //
1757 // Found with:
1758 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1759 return false;
1760 }
1761
1762 if (n->is_Vector()) {
1763 // VectorNode::Ideal swaps edges, but only for ops
1764 // that are deemed commutable. But swap_edges
1765 // requires the hash to be invariant when the edges
1766 // are swapped, which is not implemented for these
1767 // vector nodes. This seems not to create any trouble
1768 // usually, but we can also get graphs where in the
1769 // end the nodes are not all commuted, so there is
1770 // definitively an issue here.
1771 //
1772 // Probably we have two options: kill the hash, or
1773 // properly make the hash commutation friendly.
1774 //
1775 // Found with:
1776 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1777 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1778 return false;
1779 }
1780
1781 if (n->is_Region()) {
1782 // LoopNode::Ideal calls RegionNode::Ideal.
1783 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1784 // But I got an issue because RegionNode::optimize_trichotomy
1785 // then modifies another node, and pushes nodes to the worklist
1786 // Not sure if this is ok, modifying another node like that.
1787 // Maybe it is, then we need to look into what to do with
1788 // the nodes that are now on the worklist, maybe just clear
1789 // them out again. But maybe modifying other nodes like that
1790 // is also bad design. In the end, we return nullptr for
1791 // the current CountedLoop. But the extra nodes on the worklist
1792 // trip the asserts later on.
1793 //
1794 // Found with:
1795 // compiler/eliminateAutobox/TestShortBoxing.java
1796 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1797 return false;
1798 }
1799
1800 if (n->is_CallJava()) {
1801 // CallStaticJavaNode::Ideal
1802 // Led to a crash:
1803 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1804 //
1805 // Did not investigate yet, could be a bug.
1806 // Or maybe it does not expect to be called during verification.
1807 //
1808 // Found with:
1809 // test/jdk/jdk/incubator/vector/VectorRuns.java
1810 // -XX:VerifyIterativeGVN=1110
1811
1812 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1813 // and possibly their subclasses.
1814 // During late inlining it can call CallJavaNode::register_for_late_inline
1815 // That means we do more rounds of late inlining, but might fail.
1816 // Then we do IGVN again, and register the node again for late inlining.
1817 // This creates an endless cycle. Everytime we try late inlining, we
1818 // are also creating more nodes, especially SafePoint and MergeMem.
1819 // These nodes are immediately rejected when the inlining fails in the
1820 // do_late_inline_check, but they still grow the memory, until we hit
1821 // the MemLimit and crash.
1822 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1823 // called repeatedly, and eventually we terminate. I fear this is not
1824 // a great assumption to make. We should investigate more.
1825 //
1826 // Found with:
1827 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1828 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1829 return false;
1830 }
1831
1832 // The number of nodes shoud not increase.
1833 uint old_unique = C->unique();
1834 // The hash of a node should not change, this would indicate different inputs
1835 uint old_hash = n->hash();
1836 Node* i = n->Ideal(this, can_reshape);
1837 // If there was no new Idealization, we are probably happy.
1838 if (i == nullptr) {
1839 if (old_unique < C->unique()) {
1840 stringStream ss; // Print as a block without tty lock.
1841 ss.cr();
1842 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1843 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1844 n->dump_bfs(1, nullptr, "", &ss);
1845 tty->print_cr("%s", ss.as_string());
1846 return true;
1847 }
1848
1849 if (old_hash != n->hash()) {
1850 stringStream ss; // Print as a block without tty lock.
1851 ss.cr();
1852 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1853 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1854 n->dump_bfs(1, nullptr, "", &ss);
1855 tty->print_cr("%s", ss.as_string());
1856 return true;
1857 }
1858
1859 verify_empty_worklist(n);
1860
1861 // Everything is good.
1862 return false;
1863 }
1864
1865 // We just saw a new Idealization which was not done during IGVN.
1866 stringStream ss; // Print as a block without tty lock.
1867 ss.cr();
1868 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1869 if (i == n) {
1870 ss.print_cr("The node was reshaped by Ideal.");
1871 } else {
1872 ss.print_cr("The node was replaced by Ideal.");
1873 ss.print_cr("Old node:");
1874 n->dump_bfs(1, nullptr, "", &ss);
1875 }
1876 ss.print_cr("The result after Ideal:");
1877 i->dump_bfs(1, nullptr, "", &ss);
1878 tty->print_cr("%s", ss.as_string());
1879 return true;
1880 }
1881
1882 // Check that all Identity optimizations that could be done were done.
1883 // Returns true if it found missed optimization opportunities and
1884 // false otherwise (no missed optimization, or skipped verification).
1885 bool PhaseIterGVN::verify_Identity_for(Node* n) {
1886 // First, we check a list of exceptions, where we skip verification,
1887 // because there are known cases where Ideal can optimize after IGVN.
1888 // Some may be expected and cannot be fixed, and others should be fixed.
1889 switch (n->Opcode()) {
1890 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1891 // after loopopts:
1892 // // Transforming long counted loops requires a safepoint node. Do not
1893 // // eliminate a safepoint until loop opts are over.
1894 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1895 //
1896 // I think the check for major_progress does delay it until after loopopts
1897 // but it does not ensure that the node is on the IGVN worklist after
1898 // loopopts. I think we should try to instead check for
1899 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1900 //
1901 // Found with:
1902 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1903 case Op_SafePoint:
1904 return false;
1905
1906 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1907 // does not record any other memory splits.
1908 //
1909 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1910 // never got called during IGVN for this node, investigate why.
1911 //
1912 // Found with:
1913 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1914 case Op_MergeMem:
1915 return false;
1916
1917 // ConstraintCastNode::Identity finds casts that are the same, except that
1918 // the control is "higher up", i.e. dominates. The call goes via
1919 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1920 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1921 // away from the cast, but within 100 idom steps, the cast may not be
1922 // put on the IGVN worklist any more.
1923 //
1924 // Found with:
1925 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1926 case Op_CastPP:
1927 case Op_CastII:
1928 case Op_CastLL:
1929 return false;
1930
1931 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1932 // checks dominator, which may be changed, but too far up for notification
1933 // to work.
1934 //
1935 // Found with:
1936 // compiler/c2/irTests/TestSkeletonPredicates.java
1937 // -XX:VerifyIterativeGVN=1110
1938 case Op_CheckCastPP:
1939 return false;
1940
1941 // In SubNode::Identity, we do:
1942 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1943 // In the example, the AddI had an input replaced, the AddI is
1944 // added to the IGVN worklist, but the SubI is one link further
1945 // down and is not added. I checked add_users_of_use_to_worklist
1946 // where I would expect the SubI would be added, and I cannot
1947 // find the pattern, only this one:
1948 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1949 //
1950 // Fix this "notification" issue and check if there are any other
1951 // issues.
1952 //
1953 // Found with:
1954 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1955 case Op_SubI:
1956 case Op_SubL:
1957 return false;
1958
1959 // PhiNode::Identity checks for patterns like:
1960 // r = (x != con) ? x : con;
1961 // that can be constant folded to "x".
1962 //
1963 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1964 // I suspect there was some earlier change to one of the inputs, but
1965 // not all relevant outputs were put on the IGVN worklist.
1966 //
1967 // Found with:
1968 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1969 // -XX:VerifyIterativeGVN=1110
1970 case Op_Phi:
1971 return false;
1972
1973 // ConvI2LNode::Identity does
1974 // convert I2L(L2I(x)) => x
1975 //
1976 // Investigate why this did not already happen during IGVN.
1977 //
1978 // Found with:
1979 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
1980 // -XX:VerifyIterativeGVN=1110
1981 case Op_ConvI2L:
1982 return false;
1983
1984 // MaxNode::find_identity_operation
1985 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
1986 // This can be a 2-hop search, so maybe notification is not
1987 // good enough.
1988 //
1989 // Found with:
1990 // compiler/codegen/TestBooleanVect.java
1991 // -XX:VerifyIterativeGVN=1110
1992 case Op_MaxL:
1993 case Op_MinL:
1994 case Op_MaxI:
1995 case Op_MinI:
1996 case Op_MaxF:
1997 case Op_MinF:
1998 case Op_MaxHF:
1999 case Op_MinHF:
2000 case Op_MaxD:
2001 case Op_MinD:
2002 return false;
2003
2004
2005 // AddINode::Identity
2006 // Converts (x-y)+y to x
2007 // Could be issue with notification
2008 //
2009 // Turns out AddL does the same.
2010 //
2011 // Found with:
2012 // compiler/c2/Test6792161.java
2013 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2014 case Op_AddI:
2015 case Op_AddL:
2016 return false;
2017
2018 // AbsINode::Identity
2019 // Not investigated yet.
2020 case Op_AbsI:
2021 return false;
2022 }
2023
2024 if (n->is_Load()) {
2025 // LoadNode::Identity tries to look for an earlier store value via
2026 // can_see_stored_value. I found an example where this led to
2027 // an Allocation, where we could assume the value was still zero.
2028 // So the LoadN can be replaced with a zerocon.
2029 //
2030 // Investigate why this was not already done during IGVN.
2031 // A similar issue happens with Ideal.
2032 //
2033 // Found with:
2034 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2035 return false;
2036 }
2037
2038 if (n->is_Store()) {
2039 // StoreNode::Identity
2040 // Not investigated, but found missing optimization for StoreI.
2041 // Looks like a StoreI is replaced with an InitializeNode.
2042 //
2043 // Found with:
2044 // applications/ctw/modules/java_base_2.java
2045 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2046 return false;
2047 }
2048
2049 if (n->is_Vector()) {
2050 // Found with tier1-3. Not investigated yet.
2051 // The observed issue was with AndVNode::Identity
2052 return false;
2053 }
2054
2055 Node* i = n->Identity(this);
2056 // If we cannot find any other Identity, we are happy.
2057 if (i == n) {
2058 verify_empty_worklist(n);
2059 return false;
2060 }
2061
2062 // The verification just found a new Identity that was not found during IGVN.
2063 stringStream ss; // Print as a block without tty lock.
2064 ss.cr();
2065 ss.print_cr("Missed Identity optimization:");
2066 ss.print_cr("Old node:");
2067 n->dump_bfs(1, nullptr, "", &ss);
2068 ss.print_cr("New node:");
2069 i->dump_bfs(1, nullptr, "", &ss);
2070 tty->print_cr("%s", ss.as_string());
2071 return true;
2072 }
2073
2074 // Some other verifications that are not specific to a particular transformation.
2075 bool PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2076 if (n->is_AddP()) {
2077 if (!n->as_AddP()->address_input_has_same_base()) {
2078 stringStream ss; // Print as a block without tty lock.
2079 ss.cr();
2080 ss.print_cr("Base pointers must match for AddP chain:");
2081 n->dump_bfs(2, nullptr, "", &ss);
2082 tty->print_cr("%s", ss.as_string());
2083 return true;
2084 }
2085 }
2086 return false;
2087 }
2088 #endif
2089
2090 /**
2091 * Register a new node with the optimizer. Update the types array, the def-use
2092 * info. Put on worklist.
2093 */
2094 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2095 set_type_bottom(n);
2096 _worklist.push(n);
2097 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2098 return n;
2099 }
2100
2101 //------------------------------transform--------------------------------------
2102 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2103 Node *PhaseIterGVN::transform( Node *n ) {
2104 if (_delay_transform) {2105 // Register the node but don't optimize for now2106 register_new_node_with_optimizer(n);2107 return n;2108 }2109
2110 // If brand new node, make space in type array, and give it a type.
2111 ensure_type_or_null(n);
2112 if (type_or_null(n) == nullptr) {
2113 set_type_bottom(n);
2114 }
2115
2116 return transform_old(n);
2117 }
2118
2119 Node *PhaseIterGVN::transform_old(Node* n) {
2120 NOT_PRODUCT(set_transforms());
2121 // Remove 'n' from hash table in case it gets modified
2122 _table.hash_delete(n);
2123 #ifdef ASSERT
2124 if (is_verify_def_use()) {
2125 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2126 }
2127 #endif
2128
2129 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2130 if (n->is_Cmp()) {
2131 add_users_to_worklist(n);
2132 }
2133
2134 // Apply the Ideal call in a loop until it no longer applies
2135 Node* k = n;
2136 DEBUG_ONLY(dead_loop_check(k);)
2137 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2138 C->remove_modified_node(k);
2139 Node* i = apply_ideal(k, /*can_reshape=*/true);
2140 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2141 #ifndef PRODUCT
2142 verify_step(k);
2143 #endif
2144
2145 DEBUG_ONLY(uint loop_count = 1;)
2146 while (i != nullptr) {
2147 #ifdef ASSERT
2148 if (loop_count >= K + C->live_nodes()) {
2149 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2150 }
2151 #endif
2152 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2153 // Made a change; put users of original Node on worklist
2154 add_users_to_worklist(k);
2155 // Replacing root of transform tree?
2156 if (k != i) {
2157 // Make users of old Node now use new.
2158 subsume_node(k, i);
2159 k = i;
2160 }
2161 DEBUG_ONLY(dead_loop_check(k);)
2162 // Try idealizing again
2163 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2164 C->remove_modified_node(k);
2165 i = apply_ideal(k, /*can_reshape=*/true);
2166 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2167 #ifndef PRODUCT
2168 verify_step(k);
2169 #endif
2170 DEBUG_ONLY(loop_count++;)
2171 }
2172
2173 // If brand new node, make space in type array.
2174 ensure_type_or_null(k);
2175
2176 // See what kind of values 'k' takes on at runtime
2177 const Type* t = k->Value(this);
2178 assert(t != nullptr, "value sanity");
2179
2180 // Since I just called 'Value' to compute the set of run-time values
2181 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2182 // cache Value. Later requests for the local phase->type of this Node can
2183 // use the cached Value instead of suffering with 'bottom_type'.
2184 if (type_or_null(k) != t) {
2185 #ifndef PRODUCT
2186 inc_new_values();
2187 set_progress();
2188 #endif
2189 set_type(k, t);
2190 // If k is a TypeNode, capture any more-precise type permanently into Node
2191 k->raise_bottom_type(t);
2192 // Move users of node to worklist
2193 add_users_to_worklist(k);
2194 }
2195 // If 'k' computes a constant, replace it with a constant
2196 if (t->singleton() && !k->is_Con()) {
2197 NOT_PRODUCT(set_progress();)
2198 Node* con = makecon(t); // Make a constant
2199 add_users_to_worklist(k);
2200 subsume_node(k, con); // Everybody using k now uses con
2201 return con;
2202 }
2203
2204 // Now check for Identities
2205 i = k->Identity(this); // Look for a nearby replacement
2206 if (i != k) { // Found? Return replacement!
2207 NOT_PRODUCT(set_progress();)
2208 add_users_to_worklist(k);
2209 subsume_node(k, i); // Everybody using k now uses i
2210 return i;
2211 }
2212
2213 // Global Value Numbering
2214 i = hash_find_insert(k); // Check for pre-existing node
2215 if (i && (i != k)) {
2216 // Return the pre-existing node if it isn't dead
2217 NOT_PRODUCT(set_progress();)
2218 add_users_to_worklist(k);
2219 subsume_node(k, i); // Everybody using k now uses i
2220 return i;
2221 }
2222
2223 // Return Idealized original
2224 return k;
2225 }
2226
2227 //---------------------------------saturate------------------------------------
2228 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2229 const Type* limit_type) const {
2230 return new_type->narrow(old_type);
2231 }
2232
2233 //------------------------------remove_globally_dead_node----------------------
2234 // Kill a globally dead Node. All uses are also globally dead and are
2235 // aggressively trimmed.
2236 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2237 enum DeleteProgress {
2238 PROCESS_INPUTS,
2239 PROCESS_OUTPUTS
2240 };
2241 ResourceMark rm;
2242 Node_Stack stack(32);
2243 stack.push(dead, PROCESS_INPUTS);
2244
2245 while (stack.is_nonempty()) {
2246 dead = stack.node();
2247 if (dead->Opcode() == Op_SafePoint) {
2248 dead->as_SafePoint()->disconnect_from_root(this);
2249 }
2250 uint progress_state = stack.index();
2251 assert(dead != C->root(), "killing root, eh?");
2252 assert(!dead->is_top(), "add check for top when pushing");
2253 NOT_PRODUCT( set_progress(); )
2254 if (progress_state == PROCESS_INPUTS) {
2255 // After following inputs, continue to outputs
2256 stack.set_index(PROCESS_OUTPUTS);
2257 if (!dead->is_Con()) { // Don't kill cons but uses
2258 bool recurse = false;
2259 // Remove from hash table
2260 _table.hash_delete( dead );
2261 // Smash all inputs to 'dead', isolating him completely
2262 for (uint i = 0; i < dead->req(); i++) {
2263 Node *in = dead->in(i);
2264 if (in != nullptr && in != C->top()) { // Points to something?
2265 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2266 assert((nrep > 0), "sanity");
2267 if (in->outcnt() == 0) { // Made input go dead?
2268 stack.push(in, PROCESS_INPUTS); // Recursively remove
2269 recurse = true;
2270 } else if (in->outcnt() == 1 &&
2271 in->has_special_unique_user()) {
2272 _worklist.push(in->unique_out());
2273 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2274 if (in->Opcode() == Op_Region) {
2275 _worklist.push(in);
2276 } else if (in->is_Store()) {
2277 DUIterator_Fast imax, i = in->fast_outs(imax);
2278 _worklist.push(in->fast_out(i));
2279 i++;
2280 if (in->outcnt() == 2) {
2281 _worklist.push(in->fast_out(i));
2282 i++;
2283 }
2284 assert(!(i < imax), "sanity");
2285 }
2286 } else if (dead->is_data_proj_of_pure_function(in)) {
2287 _worklist.push(in);
2288 } else {
2289 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2290 }
2291 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2292 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2293 // A Load that directly follows an InitializeNode is
2294 // going away. The Stores that follow are candidates
2295 // again to be captured by the InitializeNode.
2296 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2297 Node *n = in->fast_out(j);
2298 if (n->is_Store()) {
2299 _worklist.push(n);
2300 }
2301 }
2302 }
2303 } // if (in != nullptr && in != C->top())
2304 } // for (uint i = 0; i < dead->req(); i++)
2305 if (recurse) {
2306 continue;
2307 }
2308 } // if (!dead->is_Con())
2309 } // if (progress_state == PROCESS_INPUTS)
2310
2311 // Aggressively kill globally dead uses
2312 // (Rather than pushing all the outs at once, we push one at a time,
2313 // plus the parent to resume later, because of the indefinite number
2314 // of edge deletions per loop trip.)
2315 if (dead->outcnt() > 0) {
2316 // Recursively remove output edges
2317 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2318 } else {
2319 // Finished disconnecting all input and output edges.
2320 stack.pop();
2321 // Remove dead node from iterative worklist
2322 _worklist.remove(dead);
2323 C->remove_useless_node(dead);
2324 }
2325 } // while (stack.is_nonempty())
2326 }
2327
2328 //------------------------------subsume_node-----------------------------------
2329 // Remove users from node 'old' and add them to node 'nn'.
2330 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2331 if (old->Opcode() == Op_SafePoint) {
2332 old->as_SafePoint()->disconnect_from_root(this);
2333 }
2334 assert( old != hash_find(old), "should already been removed" );
2335 assert( old != C->top(), "cannot subsume top node");
2336 // Copy debug or profile information to the new version:
2337 C->copy_node_notes_to(nn, old);
2338 // Move users of node 'old' to node 'nn'
2339 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2340 Node* use = old->last_out(i); // for each use...
2341 // use might need re-hashing (but it won't if it's a new node)
2342 rehash_node_delayed(use);
2343 // Update use-def info as well
2344 // We remove all occurrences of old within use->in,
2345 // so as to avoid rehashing any node more than once.
2346 // The hash table probe swamps any outer loop overhead.
2347 uint num_edges = 0;
2348 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2349 if (use->in(j) == old) {
2350 use->set_req(j, nn);
2351 ++num_edges;
2352 }
2353 }
2354 i -= num_edges; // we deleted 1 or more copies of this edge
2355 }
2356
2357 // Search for instance field data PhiNodes in the same region pointing to the old
2358 // memory PhiNode and update their instance memory ids to point to the new node.
2359 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2360 Node* region = old->in(0);
2361 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2362 PhiNode* phi = region->fast_out(i)->isa_Phi();
2363 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2364 phi->set_inst_mem_id((int)nn->_idx);
2365 }
2366 }
2367 }
2368
2369 // Smash all inputs to 'old', isolating him completely
2370 Node *temp = new Node(1);
2371 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2372 remove_dead_node( old );
2373 temp->del_req(0); // Yank bogus edge
2374 if (nn != nullptr && nn->outcnt() == 0) {
2375 _worklist.push(nn);
2376 }
2377 #ifndef PRODUCT
2378 if (is_verify_def_use()) {
2379 for ( int i = 0; i < _verify_window_size; i++ ) {
2380 if ( _verify_window[i] == old )
2381 _verify_window[i] = nn;
2382 }
2383 }
2384 #endif
2385 temp->destruct(this); // reuse the _idx of this little guy
2386 }
2387
2388 //------------------------------add_users_to_worklist--------------------------
2389 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2390 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2391 worklist.push(n->fast_out(i)); // Push on worklist
2392 }
2393 }
2394
2395 // Return counted loop Phi if as a counted loop exit condition, cmp
2396 // compares the induction variable with n
2397 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2398 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2399 Node* bol = cmp->fast_out(i);
2400 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2401 Node* iff = bol->fast_out(i2);
2402 if (iff->is_BaseCountedLoopEnd()) {
2403 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2404 if (cle->limit() == n) {
2405 PhiNode* phi = cle->phi();
2406 if (phi != nullptr) {
2407 return phi;
2408 }
2409 }
2410 }
2411 }
2412 }
2413 return nullptr;
2414 }
2415
2416 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2417 add_users_to_worklist0(n, _worklist);
2418
2419 Unique_Node_List& worklist = _worklist;
2420 // Move users of node to worklist
2421 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2422 Node* use = n->fast_out(i); // Get use
2423 add_users_of_use_to_worklist(n, use, worklist);
2424 }
2425 }
2426
2427 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2428 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2429 use->is_Store() ) // Enable store/load same address
2430 add_users_to_worklist0(use, worklist);
2431
2432 // If we changed the receiver type to a call, we need to revisit
2433 // the Catch following the call. It's looking for a non-null
2434 // receiver to know when to enable the regular fall-through path
2435 // in addition to the NullPtrException path.
2436 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2437 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2438 if (p != nullptr) {
2439 add_users_to_worklist0(p, worklist);
2440 }
2441 }
2442
2443 uint use_op = use->Opcode();
2444 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2445 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2446 if (use->outcnt() > 0) {
2447 Node* bol = use->raw_out(0);
2448 if (bol->outcnt() > 0) {
2449 Node* iff = bol->raw_out(0);
2450 if (iff->outcnt() == 2) {
2451 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2452 // phi merging either 0 or 1 onto the worklist
2453 Node* ifproj0 = iff->raw_out(0);
2454 Node* ifproj1 = iff->raw_out(1);
2455 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2456 Node* region0 = ifproj0->raw_out(0);
2457 Node* region1 = ifproj1->raw_out(0);
2458 if( region0 == region1 )
2459 add_users_to_worklist0(region0, worklist);
2460 }
2461 }
2462 }
2463 }
2464 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2465 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2466 if (phi != nullptr) {
2467 // Input to the cmp of a loop exit check has changed, thus
2468 // the loop limit may have changed, which can then change the
2469 // range values of the trip-count Phi.
2470 worklist.push(phi);
2471 }
2472 }
2473 if (use_op == Op_CmpI) {
2474 Node* cmp = use;
2475 Node* in1 = cmp->in(1);
2476 Node* in2 = cmp->in(2);
2477 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2478 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2479 //
2480 // in1 in2 in1 in2
2481 // | | | |
2482 // +--- | --+ | |
2483 // | | | | |
2484 // CmpINode | CmpINode
2485 // | | |
2486 // BoolNode | BoolNode
2487 // | | OR |
2488 // IfNode | IfNode
2489 // | | |
2490 // IfProj | IfProj X
2491 // | | | |
2492 // CastIINode CastIINode
2493 //
2494 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2495 if (in1 == n) {
2496 // in1 modified -> could turn into X -> do traversal based on right pattern.
2497 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2498 Node* bol = cmp->fast_out(i2); // For each Bool
2499 if (bol->is_Bool()) {
2500 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2501 Node* iff = bol->fast_out(i3); // For each If
2502 if (iff->is_If()) {
2503 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2504 Node* if_proj = iff->fast_out(i4); // For each IfProj
2505 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2506 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2507 Node* castii = if_proj->fast_out(i5); // For each CastII
2508 if (castii->is_CastII() &&
2509 castii->as_CastII()->carry_dependency()) {
2510 worklist.push(castii);
2511 }
2512 }
2513 }
2514 }
2515 }
2516 }
2517 }
2518 } else {
2519 // Only in2 modified -> can assume X == in2 (left pattern).
2520 assert(n == in2, "only in2 modified");
2521 // Find all CastII with input in1.
2522 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2523 Node* castii = in1->fast_out(j);
2524 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2525 // Find If.
2526 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2527 Node* ifnode = castii->in(0)->in(0);
2528 // Check that if connects to the cmp
2529 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2530 worklist.push(castii);
2531 }
2532 }
2533 }
2534 }
2535 }
2536 }
2537 }
2538 }
2539
2540 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2541 // Patterns:
2542 // ConstraintCast+ -> Sub
2543 // ConstraintCast+ -> Phi
2544 // ConstraintCast+ -> Xor
2545 if (use->is_ConstraintCast()) {
2546 auto push_the_uses_to_worklist = [&](Node* n){
2547 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2548 worklist.push(n);
2549 }
2550 };
2551 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2552 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2553 }
2554 // If changed LShift inputs, check RShift users for useless sign-ext
2555 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2556 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2557 Node* u = use->fast_out(i2);
2558 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2559 worklist.push(u);
2560 }
2561 }
2562 // If changed LShift inputs, check And users for shift and mask (And) operation
2563 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2564 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2565 Node* u = use->fast_out(i2);
2566 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2567 worklist.push(u);
2568 }
2569 }
2570 }
2571 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2572 if (use_op == Op_AddI || use_op == Op_SubI) {
2573 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2574 Node* u = use->fast_out(i2);
2575 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2576 worklist.push(u);
2577 }
2578 }
2579 }
2580 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2581 if (use_op == Op_AndI || use_op == Op_AndL) {
2582 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2583 Node* u = use->fast_out(i2);
2584 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2585 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL) {
2586 worklist.push(u);
2587 }
2588 }
2589 }
2590 // Check for redundant conversion patterns:
2591 // ConvD2L->ConvL2D->ConvD2L
2592 // ConvF2I->ConvI2F->ConvF2I
2593 // ConvF2L->ConvL2F->ConvF2L
2594 // ConvI2F->ConvF2I->ConvI2F
2595 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2596 // are the only ones that are known to trigger missed optimizations otherwise
2597 if (use_op == Op_ConvL2D ||
2598 use_op == Op_ConvI2F ||
2599 use_op == Op_ConvL2F ||
2600 use_op == Op_ConvF2I) {
2601 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2602 Node* u = use->fast_out(i2);
2603 if ((use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2604 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2605 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2606 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F)) {
2607 worklist.push(u);
2608 }
2609 }
2610 }
2611 // If changed AddP inputs:
2612 // - check Stores for loop invariant, and
2613 // - if the changed input is the offset, check constant-offset AddP users for
2614 // address expression flattening.
2615 if (use_op == Op_AddP) {
2616 bool offset_changed = n == use->in(AddPNode::Offset);
2617 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2618 Node* u = use->fast_out(i2);
2619 if (u->is_Mem()) {
2620 worklist.push(u);
2621 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2622 worklist.push(u);
2623 }
2624 }
2625 }
2626 // Check for "abs(0-x)" into "abs(x)" conversion
2627 if (use->is_Sub()) {
2628 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2629 Node* u = use->fast_out(i2);
2630 if (u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2631 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI) {
2632 worklist.push(u);
2633 }
2634 }
2635 }
2636 // Check for Max/Min(A, Max/Min(B, C)) where A == B or A == C
2637 if (use->is_MinMax()) {
2638 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2639 Node* u = use->fast_out(i2);
2640 if (u->Opcode() == use->Opcode()) {
2641 worklist.push(u);
2642 }
2643 }
2644 }
2645 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2646 add_users_to_worklist0(proj, worklist);
2647 };
2648 // If changed initialization activity, check dependent Stores
2649 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2650 InitializeNode* init = use->as_Allocate()->initialization();
2651 if (init != nullptr) {
2652 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2653 }
2654 }
2655 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2656 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2657 // to guarantee the change is not missed.
2658 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2659 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2660 if (p != nullptr) {
2661 add_users_to_worklist0(p, worklist);
2662 }
2663 }
2664
2665 if (use_op == Op_Initialize) {
2666 InitializeNode* init = use->as_Initialize();
2667 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2668 }
2669 // Loading the java mirror from a Klass requires two loads and the type
2670 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2671 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2672 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2673 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2674
2675 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2676 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2677 Node* u = use->fast_out(i2);
2678 const Type* ut = u->bottom_type();
2679 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2680 if (has_load_barrier_nodes) {
2681 // Search for load barriers behind the load
2682 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2683 Node* b = u->fast_out(i3);
2684 if (bs->is_gc_barrier_node(b)) {
2685 worklist.push(b);
2686 }
2687 }
2688 }
2689 worklist.push(u);
2690 }
2691 }
2692 }
2693 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2694 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2695 if (use->outcnt() == 1) {
2696 Node* cmp = use->unique_out();
2697 worklist.push(cmp);
2698 }
2699 }
2700
2701 // From CastX2PNode::Ideal
2702 // CastX2P(AddX(x, y))
2703 // CastX2P(SubX(x, y))
2704 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2705 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2706 Node* u = use->fast_out(i2);
2707 if (u->Opcode() == Op_CastX2P) {
2708 worklist.push(u);
2709 }
2710 }
2711 }
2712
2713 /* AndNode has a special handling when one of the operands is a LShiftNode:
2714 * (LHS << s) & RHS
2715 * if RHS fits in less than s bits, the value of this expression is 0.
2716 * The difficulty is that there might be a conversion node (ConvI2L) between
2717 * the LShiftINode and the AndLNode, like so:
2718 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2719 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2720 * (see `AndIL_min_trailing_zeros`).
2721 *
2722 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2723 * is not enough: there might be no update happening there. We need to
2724 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2725 *
2726 * Moreover we can have ConstraintCasts in between. It may look like
2727 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2728 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2729 * So we must deal with that as well.
2730 */
2731 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2732 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2733 auto push_and_to_worklist = [&worklist](Node* n){
2734 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2735 worklist.push(n);
2736 }
2737 };
2738 use->visit_uses(push_and_to_worklist, is_boundary);
2739 }
2740 }
2741
2742 /**
2743 * Remove the speculative part of all types that we know of
2744 */
2745 void PhaseIterGVN::remove_speculative_types() {
2746 assert(UseTypeSpeculation, "speculation is off");
2747 for (uint i = 0; i < _types.Size(); i++) {
2748 const Type* t = _types.fast_lookup(i);
2749 if (t != nullptr) {
2750 _types.map(i, t->remove_speculative());
2751 }
2752 }
2753 _table.check_no_speculative_types();
2754 }
2755
2756 // Check if the type of a divisor of a Div or Mod node includes zero.
2757 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2758 switch (n->Opcode()) {
2759 case Op_DivI:
2760 case Op_ModI:
2761 case Op_UDivI:
2762 case Op_UModI: {
2763 // Type of divisor includes 0?
2764 if (type(n->in(2)) == Type::TOP) {
2765 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2766 return false;
2767 }
2768 const TypeInt* type_divisor = type(n->in(2))->is_int();
2769 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2770 }
2771 case Op_DivL:
2772 case Op_ModL:
2773 case Op_UDivL:
2774 case Op_UModL: {
2775 // Type of divisor includes 0?
2776 if (type(n->in(2)) == Type::TOP) {
2777 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2778 return false;
2779 }
2780 const TypeLong* type_divisor = type(n->in(2))->is_long();
2781 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2782 }
2783 }
2784 return true;
2785 }
2786
2787 //=============================================================================
2788 #ifndef PRODUCT
2789 uint PhaseCCP::_total_invokes = 0;
2790 uint PhaseCCP::_total_constants = 0;
2791 #endif
2792 //------------------------------PhaseCCP---------------------------------------
2793 // Conditional Constant Propagation, ala Wegman & Zadeck
2794 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2795 NOT_PRODUCT( clear_constants(); )
2796 assert( _worklist.size() == 0, "" );
2797 analyze();
2798 }
2799
2800 #ifndef PRODUCT
2801 //------------------------------~PhaseCCP--------------------------------------
2802 PhaseCCP::~PhaseCCP() {
2803 inc_invokes();
2804 _total_constants += count_constants();
2805 }
2806 #endif
2807
2808
2809 #ifdef ASSERT
2810 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2811 if (tnew->meet(told) != tnew->remove_speculative()) {
2812 n->dump(1);
2813 tty->print("told = "); told->dump(); tty->cr();
2814 tty->print("tnew = "); tnew->dump(); tty->cr();
2815 fatal("Not monotonic");
2816 }
2817 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2818 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2819 }
2820 #endif //ASSERT
2821
2822 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2823 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2824 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2825 // or nodes even further down back to the worklist because their type could change as a result of the current type
2826 // change).
2827 void PhaseCCP::analyze() {
2828 // Initialize all types to TOP, optimistic analysis
2829 for (uint i = 0; i < C->unique(); i++) {
2830 _types.map(i, Type::TOP);
2831 }
2832
2833 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2834 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2835 // Compile is over. The local arena gets de-allocated at the end of its scope.
2836 ResourceArea local_arena(mtCompiler);
2837 Unique_Node_List worklist(&local_arena);
2838 Unique_Node_List worklist_revisit(&local_arena);
2839 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2840
2841 // Push root onto worklist
2842 worklist.push(C->root());
2843
2844 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2845 _root_and_safepoints.push(C->root());
2846
2847 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2848
2849 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2850 while (worklist.size() != 0) {
2851 Node* n = fetch_next_node(worklist);
2852 DEBUG_ONLY(worklist_verify.push(n);)
2853 if (needs_revisit(n)) {
2854 worklist_revisit.push(n);
2855 }
2856 if (n->is_SafePoint()) {
2857 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2858 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2859 _root_and_safepoints.push(n);
2860 }
2861 analyze_step(worklist, n);
2862 }
2863
2864 // More rounds to catch updates far in the graph.
2865 // Revisit nodes that might be able to refine their types at the end of the round.
2866 // If so, process these nodes. If there is remaining work, start another round.
2867 do {
2868 while (worklist.size() != 0) {
2869 Node* n = fetch_next_node(worklist);
2870 analyze_step(worklist, n);
2871 }
2872 for (uint t = 0; t < worklist_revisit.size(); t++) {
2873 Node* n = worklist_revisit.at(t);
2874 analyze_step(worklist, n);
2875 }
2876 } while (worklist.size() != 0);
2877
2878 DEBUG_ONLY(verify_analyze(worklist_verify);)
2879 }
2880
2881 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2882 const Type* new_type = n->Value(this);
2883 if (new_type != type(n)) {
2884 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2885 dump_type_and_node(n, new_type);
2886 set_type(n, new_type);
2887 push_child_nodes_to_worklist(worklist, n);
2888 }
2889 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2890 // Keep track of Type nodes to kill CFG paths that use Type
2891 // nodes that become dead.
2892 _maybe_top_type_nodes.push(n);
2893 }
2894 }
2895
2896 // Some nodes can refine their types due to type change somewhere deep
2897 // in the graph. We will need to revisit them before claiming convergence.
2898 // Add nodes here if particular *Node::Value is doing deep graph traversals
2899 // not handled by PhaseCCP::push_more_uses().
2900 bool PhaseCCP::needs_revisit(Node* n) const {
2901 // LoadNode performs deep traversals. Load is not notified for changes far away.
2902 if (n->is_Load()) {
2903 return true;
2904 }
2905 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2906 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2907 return true;
2908 }
2909 return false;
2910 }
2911
2912 #ifdef ASSERT
2913 // For every node n on verify list, check if type(n) == n->Value()
2914 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2915 // Therefore, we are verifying Value convergence strictly.
2916 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2917 bool failure = false;
2918 while (worklist_verify.size()) {
2919 Node* n = worklist_verify.pop();
2920 failure |= verify_Value_for(n, /* strict = */ true);
2921 }
2922 // If we get this assert, check why the reported nodes were not processed again in CCP.
2923 // We should either make sure that these nodes are properly added back to the CCP worklist
2924 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2925 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2926 // them at the end of the round.
2927 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound.");
2928 }
2929 #endif
2930
2931 // Fetch next node from worklist to be examined in this iteration.
2932 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2933 if (StressCCP) {
2934 return worklist.remove(C->random() % worklist.size());
2935 } else {
2936 return worklist.pop();
2937 }
2938 }
2939
2940 #ifndef PRODUCT
2941 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2942 if (TracePhaseCCP) {
2943 t->dump();
2944 do {
2945 tty->print("\t");
2946 } while (tty->position() < 16);
2947 n->dump();
2948 }
2949 }
2950 #endif
2951
2952 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2953 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2954 // of the new type of 'n'. Push these nodes to the worklist.
2955 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2956 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2957 Node* use = n->fast_out(i);
2958 push_if_not_bottom_type(worklist, use);
2959 push_more_uses(worklist, n, use);
2960 }
2961 }
2962
2963 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2964 if (n->bottom_type() != type(n)) {
2965 worklist.push(n);
2966 }
2967 }
2968
2969 // For some nodes, we need to propagate the type change to grandchildren or even further down.
2970 // Add them back to the worklist.
2971 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
2972 push_phis(worklist, use);
2973 push_catch(worklist, use);
2974 push_cmpu(worklist, use);
2975 push_counted_loop_phi(worklist, parent, use);
2976 push_loadp(worklist, use);
2977 push_and(worklist, parent, use);
2978 push_cast_ii(worklist, parent, use);
2979 push_opaque_zero_trip_guard(worklist, use);
2980 push_bool_with_cmpu_and_mask(worklist, use);
2981 }
2982
2983
2984 // We must recheck Phis too if use is a Region.
2985 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
2986 if (use->is_Region()) {
2987 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2988 push_if_not_bottom_type(worklist, use->fast_out(i));
2989 }
2990 }
2991 }
2992
2993 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
2994 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
2995 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
2996 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
2997 if (use->is_Call()) {
2998 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2999 Node* proj = use->fast_out(i);
3000 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
3001 Node* catch_node = proj->find_out_with(Op_Catch);
3002 if (catch_node != nullptr) {
3003 worklist.push(catch_node);
3004 }
3005 }
3006 }
3007 }
3008 }
3009
3010 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
3011 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
3012 // and get wrong types otherwise.
3013 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
3014 uint use_op = use->Opcode();
3015 if (use_op == Op_AddI || use_op == Op_SubI) {
3016 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3017 Node* cmpu = use->fast_out(i);
3018 const uint cmpu_opcode = cmpu->Opcode();
3019 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
3020 // Got a CmpU or CmpU3 which might need the new type information from node n.
3021 push_if_not_bottom_type(worklist, cmpu);
3022 }
3023 }
3024 }
3025 }
3026
3027 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3028 // (1b): "(m & x) <u (m + 1))".
3029 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3030 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3031 //
3032 // m x m 1 (%%)
3033 // \ / \ /
3034 // AndI AddI
3035 // \ /
3036 // CmpU
3037 // |
3038 // Bool
3039 //
3040 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3041 uint use_op = use->Opcode();
3042 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3043 // Not "m & x" or "m + 1"
3044 return;
3045 }
3046 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3047 Node* cmpu = use->fast_out(i);
3048 if (cmpu->Opcode() == Op_CmpU) {
3049 push_bool_matching_case1b(worklist, cmpu);
3050 }
3051 }
3052 }
3053
3054 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3055 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3056 assert(cmpu->Opcode() == Op_CmpU, "must be");
3057 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3058 Node* bol = cmpu->fast_out(i);
3059 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3060 // Not a Bool with "<u"
3061 continue;
3062 }
3063 Node* andI = cmpu->in(1);
3064 Node* addI = cmpu->in(2);
3065 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3066 // Not "m & x" and "m + 1"
3067 continue;
3068 }
3069
3070 Node* m = addI->in(1);
3071 if (m == andI->in(1) || m == andI->in(2)) {
3072 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3073 push_if_not_bottom_type(worklist, bol);
3074 }
3075 }
3076 }
3077
3078 // 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'.
3079 // Seem PhiNode::Value().
3080 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3081 uint use_op = use->Opcode();
3082 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3083 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3084 if (phi != nullptr) {
3085 worklist.push(phi);
3086 }
3087 }
3088 }
3089
3090 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3091 // See LoadNode::Value().
3092 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3093 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
3094 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
3095
3096 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3097 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3098 Node* loadp = use->fast_out(i);
3099 const Type* ut = loadp->bottom_type();
3100 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3101 if (has_load_barrier_nodes) {
3102 // Search for load barriers behind the load
3103 push_load_barrier(worklist, barrier_set, loadp);
3104 }
3105 worklist.push(loadp);
3106 }
3107 }
3108 }
3109 }
3110
3111 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3112 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3113 Node* barrier_node = use->fast_out(i);
3114 if (barrier_set->is_gc_barrier_node(barrier_node)) {
3115 worklist.push(barrier_node);
3116 }
3117 }
3118 }
3119
3120 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3121 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3122 // value changed.
3123 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3124 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3125 uint use_op = use->Opcode();
3126 if (
3127 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3128 (parent_type != nullptr && parent_type->is_con()) ||
3129 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3130 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3131
3132 auto push_and_uses_to_worklist = [&](Node* n) {
3133 uint opc = n->Opcode();
3134 if (opc == Op_AndI || opc == Op_AndL) {
3135 push_if_not_bottom_type(worklist, n);
3136 }
3137 };
3138 auto is_boundary = [](Node* n) {
3139 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3140 };
3141 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3142 }
3143 }
3144
3145 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3146 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3147 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3148 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3149 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3150 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3151 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3152 Node* other_cmp_input = use->in(1);
3153 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3154 Node* cast_ii = other_cmp_input->fast_out(i);
3155 if (cast_ii->is_CastII()) {
3156 push_if_not_bottom_type(worklist, cast_ii);
3157 }
3158 }
3159 }
3160 }
3161
3162 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3163 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3164 push_if_not_bottom_type(worklist, use->unique_out());
3165 }
3166 }
3167
3168 //------------------------------do_transform-----------------------------------
3169 // Top level driver for the recursive transformer
3170 void PhaseCCP::do_transform() {
3171 // Correct leaves of new-space Nodes; they point to old-space.
3172 C->set_root( transform(C->root())->as_Root() );
3173 assert( C->top(), "missing TOP node" );
3174 assert( C->root(), "missing root" );
3175 }
3176
3177 //------------------------------transform--------------------------------------
3178 // Given a Node in old-space, clone him into new-space.
3179 // Convert any of his old-space children into new-space children.
3180 Node *PhaseCCP::transform( Node *n ) {
3181 assert(n->is_Root(), "traversal must start at root");
3182 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3183
3184 ResourceMark rm;
3185 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3186 Node_List node_map;
3187 // Pre-allocate to avoid frequent realloc
3188 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3189 // track all visited nodes, so that we can remove the complement
3190 Unique_Node_List useful;
3191
3192 if (KillPathsReachableByDeadTypeNode) {
3193 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3194 Node* type_node = _maybe_top_type_nodes.at(i);
3195 if (type(type_node) == Type::TOP) {
3196 ResourceMark rm;
3197 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3198 }
3199 }
3200 } else {
3201 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3202 }
3203
3204 // Initialize the traversal.
3205 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3206 // 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
3207 // 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
3208 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3209 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3210 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3211 // and all safepoints.
3212 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3213 Node* nn = _root_and_safepoints.at(i);
3214 Node* new_node = node_map[nn->_idx];
3215 assert(new_node == nullptr, "");
3216 new_node = transform_once(nn); // Check for constant
3217 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3218 transform_stack.push(new_node); // Process children of cloned node
3219 useful.push(new_node);
3220 }
3221
3222 while (transform_stack.is_nonempty()) {
3223 Node* clone = transform_stack.pop();
3224 uint cnt = clone->req();
3225 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3226 Node *input = clone->in(i);
3227 if( input != nullptr ) { // Ignore nulls
3228 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3229 if( new_input == nullptr ) {
3230 new_input = transform_once(input); // Check for constant
3231 node_map.map( input->_idx, new_input );// Flag as having been cloned
3232 transform_stack.push(new_input); // Process children of cloned node
3233 useful.push(new_input);
3234 }
3235 assert( new_input == clone->in(i), "insanity check");
3236 }
3237 }
3238 }
3239
3240 // The above transformation might lead to subgraphs becoming unreachable from the
3241 // bottom while still being reachable from the top. As a result, nodes in that
3242 // subgraph are not transformed and their bottom types are not updated, leading to
3243 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3244 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3245 // because their address type is inconsistent. Therefore, we aggressively remove
3246 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3247 // to remove them anyway.
3248 if (C->cached_top_node()) {
3249 useful.push(C->cached_top_node());
3250 }
3251 C->update_dead_node_list(useful);
3252 remove_useless_nodes(useful.member_set());
3253 _worklist.remove_useless_nodes(useful.member_set());
3254 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3255
3256 Node* new_root = node_map[n->_idx];
3257 assert(new_root->is_Root(), "transformed root node must be a root node");
3258 return new_root;
3259 }
3260
3261 //------------------------------transform_once---------------------------------
3262 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3263 Node *PhaseCCP::transform_once( Node *n ) {
3264 const Type *t = type(n);
3265 // Constant? Use constant Node instead
3266 if( t->singleton() ) {
3267 Node *nn = n; // Default is to return the original constant
3268 if( t == Type::TOP ) {
3269 // cache my top node on the Compile instance
3270 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3271 C->set_cached_top_node(ConNode::make(Type::TOP));
3272 set_type(C->top(), Type::TOP);
3273 }
3274 nn = C->top();
3275 }
3276 if( !n->is_Con() ) {
3277 if( t != Type::TOP ) {
3278 nn = makecon(t); // ConNode::make(t);
3279 NOT_PRODUCT( inc_constants(); )
3280 } else if( n->is_Region() ) { // Unreachable region
3281 // Note: nn == C->top()
3282 n->set_req(0, nullptr); // Cut selfreference
3283 bool progress = true;
3284 uint max = n->outcnt();
3285 DUIterator i;
3286 while (progress) {
3287 progress = false;
3288 // Eagerly remove dead phis to avoid phis copies creation.
3289 for (i = n->outs(); n->has_out(i); i++) {
3290 Node* m = n->out(i);
3291 if (m->is_Phi()) {
3292 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3293 replace_node(m, nn);
3294 if (max != n->outcnt()) {
3295 progress = true;
3296 i = n->refresh_out_pos(i);
3297 max = n->outcnt();
3298 }
3299 }
3300 }
3301 }
3302 }
3303 replace_node(n,nn); // Update DefUse edges for new constant
3304 }
3305 return nn;
3306 }
3307
3308 // If x is a TypeNode, capture any more-precise type permanently into Node
3309 if (t != n->bottom_type()) {
3310 hash_delete(n); // changing bottom type may force a rehash
3311 n->raise_bottom_type(t);
3312 _worklist.push(n); // n re-enters the hash table via the worklist
3313 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3314 }
3315
3316 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3317 switch( n->Opcode() ) {
3318 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3319 case Op_CallDynamicJava:
3320 case Op_FastLock: // Revisit FastLocks for lock coarsening
3321 case Op_If:
3322 case Op_CountedLoopEnd:
3323 case Op_Region:
3324 case Op_Loop:
3325 case Op_CountedLoop:
3326 case Op_Conv2B:
3327 case Op_Opaque1:
3328 _worklist.push(n);
3329 break;
3330 default:
3331 break;
3332 }
3333
3334 return n;
3335 }
3336
3337 //---------------------------------saturate------------------------------------
3338 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3339 const Type* limit_type) const {
3340 const Type* wide_type = new_type->widen(old_type, limit_type);
3341 if (wide_type != new_type) { // did we widen?
3342 // If so, we may have widened beyond the limit type. Clip it back down.
3343 new_type = wide_type->filter(limit_type);
3344 }
3345 return new_type;
3346 }
3347
3348 //------------------------------print_statistics-------------------------------
3349 #ifndef PRODUCT
3350 void PhaseCCP::print_statistics() {
3351 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3352 }
3353 #endif
3354
3355
3356 //=============================================================================
3357 #ifndef PRODUCT
3358 uint PhasePeephole::_total_peepholes = 0;
3359 #endif
3360 //------------------------------PhasePeephole----------------------------------
3361 // Conditional Constant Propagation, ala Wegman & Zadeck
3362 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3363 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3364 NOT_PRODUCT( clear_peepholes(); )
3365 }
3366
3367 #ifndef PRODUCT
3368 //------------------------------~PhasePeephole---------------------------------
3369 PhasePeephole::~PhasePeephole() {
3370 _total_peepholes += count_peepholes();
3371 }
3372 #endif
3373
3374 //------------------------------transform--------------------------------------
3375 Node *PhasePeephole::transform( Node *n ) {
3376 ShouldNotCallThis();
3377 return nullptr;
3378 }
3379
3380 //------------------------------do_transform-----------------------------------
3381 void PhasePeephole::do_transform() {
3382 bool method_name_not_printed = true;
3383
3384 // Examine each basic block
3385 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3386 Block* block = _cfg.get_block(block_number);
3387 bool block_not_printed = true;
3388
3389 for (bool progress = true; progress;) {
3390 progress = false;
3391 // block->end_idx() not valid after PhaseRegAlloc
3392 uint end_index = block->number_of_nodes();
3393 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3394 Node *n = block->get_node(instruction_index);
3395 if( n->is_Mach() ) {
3396 MachNode *m = n->as_Mach();
3397 // check for peephole opportunities
3398 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3399 if( result != -1 ) {
3400 #ifndef PRODUCT
3401 if( PrintOptoPeephole ) {
3402 // Print method, first time only
3403 if( C->method() && method_name_not_printed ) {
3404 C->method()->print_short_name(); tty->cr();
3405 method_name_not_printed = false;
3406 }
3407 // Print this block
3408 if( Verbose && block_not_printed) {
3409 tty->print_cr("in block");
3410 block->dump();
3411 block_not_printed = false;
3412 }
3413 // Print the peephole number
3414 tty->print_cr("peephole number: %d", result);
3415 }
3416 inc_peepholes();
3417 #endif
3418 // Set progress, start again
3419 progress = true;
3420 break;
3421 }
3422 }
3423 }
3424 }
3425 }
3426 }
3427
3428 //------------------------------print_statistics-------------------------------
3429 #ifndef PRODUCT
3430 void PhasePeephole::print_statistics() {
3431 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3432 }
3433 #endif
3434
3435
3436 //=============================================================================
3437 //------------------------------set_req_X--------------------------------------
3438 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3439 assert( is_not_dead(n), "can not use dead node");
3440 #ifdef ASSERT
3441 if (igvn->hash_find(this) == this) {
3442 tty->print_cr("Need to remove from hash before changing edges");
3443 this->dump(1);
3444 tty->print_cr("Set at i = %d", i);
3445 n->dump();
3446 assert(false, "Need to remove from hash before changing edges");
3447 }
3448 #endif
3449 Node *old = in(i);
3450 set_req(i, n);
3451
3452 // old goes dead?
3453 if( old ) {
3454 switch (old->outcnt()) {
3455 case 0:
3456 // Put into the worklist to kill later. We do not kill it now because the
3457 // recursive kill will delete the current node (this) if dead-loop exists
3458 if (!old->is_top())
3459 igvn->_worklist.push( old );
3460 break;
3461 case 1:
3462 if( old->is_Store() || old->has_special_unique_user() )
3463 igvn->add_users_to_worklist( old );
3464 break;
3465 case 2:
3466 if( old->is_Store() )
3467 igvn->add_users_to_worklist( old );
3468 if( old->Opcode() == Op_Region )
3469 igvn->_worklist.push(old);
3470 break;
3471 case 3:
3472 if( old->Opcode() == Op_Region ) {
3473 igvn->_worklist.push(old);
3474 igvn->add_users_to_worklist( old );
3475 }
3476 break;
3477 default:
3478 break;
3479 }
3480
3481 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3482 }
3483 }
3484
3485 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3486 PhaseIterGVN* igvn = gvn->is_IterGVN();
3487 if (igvn == nullptr) {
3488 set_req(i, n);
3489 return;
3490 }
3491 set_req_X(i, n, igvn);
3492 }
3493
3494 //-------------------------------replace_by-----------------------------------
3495 // Using def-use info, replace one node for another. Follow the def-use info
3496 // to all users of the OLD node. Then make all uses point to the NEW node.
3497 void Node::replace_by(Node *new_node) {
3498 assert(!is_top(), "top node has no DU info");
3499 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3500 Node* use = last_out(i);
3501 uint uses_found = 0;
3502 for (uint j = 0; j < use->len(); j++) {
3503 if (use->in(j) == this) {
3504 if (j < use->req())
3505 use->set_req(j, new_node);
3506 else use->set_prec(j, new_node);
3507 uses_found++;
3508 }
3509 }
3510 i -= uses_found; // we deleted 1 or more copies of this edge
3511 }
3512 }
3513
3514 //=============================================================================
3515 //-----------------------------------------------------------------------------
3516 void Type_Array::grow( uint i ) {
3517 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3518 if( !_max ) {
3519 _max = 1;
3520 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3521 _types[0] = nullptr;
3522 }
3523 uint old = _max;
3524 _max = next_power_of_2(i);
3525 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3526 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3527 }
3528
3529 //------------------------------dump-------------------------------------------
3530 #ifndef PRODUCT
3531 void Type_Array::dump() const {
3532 uint max = Size();
3533 for( uint i = 0; i < max; i++ ) {
3534 if( _types[i] != nullptr ) {
3535 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3536 }
3537 }
3538 }
3539 #endif
--- EOF ---