1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE 255
46
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 DEBUG_ONLY( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 DEBUG_ONLY( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
368 // Implementation requires an edge from root to each SafePointNode
369 // at a backward branch. Inserted in add_safepoint().
370
371 // Identify nodes that are reachable from below, useful.
372 C->identify_useful_nodes(_useful);
373 // Update dead node list
374 C->update_dead_node_list(_useful);
375
376 // Remove all useless nodes from PhaseValues' recorded types
377 // Must be done before disconnecting nodes to preserve hash-table-invariant
378 gvn->remove_useless_nodes(_useful.member_set());
379
380 // Remove all useless nodes from future worklist
381 worklist.remove_useless_nodes(_useful.member_set());
382
383 // Disconnect 'useless' nodes that are adjacent to useful nodes
384 C->disconnect_useless_nodes(_useful, worklist);
385 }
386
387 //=============================================================================
388 //------------------------------PhaseRenumberLive------------------------------
389 // First, remove useless nodes (equivalent to identifying live nodes).
390 // Then, renumber live nodes.
391 //
392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
395 // value in the range [0, x).
396 //
397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
399 //
400 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
402 // be processed by IGVN after removal of the useless nodes.
403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
404 // the node's type. The mapping is updated to use the new node IDs as well. We
405 // create a new map, and swap it with the old one.
406 //
407 // Other data structures used by the compiler are not updated. The hash table for value
408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
409 // computing the hash values is not based on node IDs.
410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
411 Unique_Node_List& worklist,
412 PhaseNumber phase_num) :
413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
414 _new_type_array(C->comp_arena()),
415 _old2new_map(C->unique(), C->unique(), -1),
416 _is_pass_finished(false),
417 _live_node_count(C->live_nodes())
418 {
419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
421 assert(_delayed.size() == 0, "should be empty");
422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
423 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
424
425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
426 if (old_node_note_array != nullptr) {
427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
428 new_size = MAX2(8, new_size);
429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
430 C->grow_node_notes(C->node_note_array(), new_size);
431 }
432
433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
434
435 // Iterate over the set of live nodes.
436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
437 Node* n = _useful.at(current_idx);
438
439 const Type* type = gvn->type_or_null(n);
440 _new_type_array.map(current_idx, type);
441
442 assert(_old2new_map.at(n->_idx) == -1, "already seen");
443 _old2new_map.at_put(n->_idx, current_idx);
444
445 if (old_node_note_array != nullptr) {
446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
447 C->set_node_notes_at(current_idx, nn);
448 }
449
450 n->set_idx(current_idx); // Update node ID.
451
452 if (update_embedded_ids(n) < 0) {
453 _delayed.push(n); // has embedded IDs; handle later
454 }
455 }
456
457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
458 worklist.recompute_idx_set();
459
460 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
461
462 _is_pass_finished = true; // pass finished; safe to process delayed updates
463
464 while (_delayed.size() > 0) {
465 Node* n = _delayed.pop();
466 int no_of_updates = update_embedded_ids(n);
467 assert(no_of_updates > 0, "should be updated");
468 }
469
470 // Replace the compiler's type information with the updated type information.
471 gvn->types().swap(_new_type_array);
472
473 // Update the unique node count of the compilation to the number of currently live nodes.
474 C->set_unique(_live_node_count);
475
476 // Set the dead node count to 0 and reset dead node list.
477 C->reset_dead_node_list();
478 }
479
480 int PhaseRenumberLive::new_index(int old_idx) {
481 assert(_is_pass_finished, "not finished");
482 if (_old2new_map.at(old_idx) == -1) { // absent
483 // Allocate a placeholder to preserve uniqueness
484 _old2new_map.at_put(old_idx, _live_node_count);
485 _live_node_count++;
486 }
487 return _old2new_map.at(old_idx);
488 }
489
490 int PhaseRenumberLive::update_embedded_ids(Node* n) {
491 int no_of_updates = 0;
492 if (n->is_Phi()) {
493 PhiNode* phi = n->as_Phi();
494 if (phi->_inst_id != -1) {
495 if (!_is_pass_finished) {
496 return -1; // delay
497 }
498 int new_idx = new_index(phi->_inst_id);
499 assert(new_idx != -1, "");
500 phi->_inst_id = new_idx;
501 no_of_updates++;
502 }
503 if (phi->_inst_mem_id != -1) {
504 if (!_is_pass_finished) {
505 return -1; // delay
506 }
507 int new_idx = new_index(phi->_inst_mem_id);
508 assert(new_idx != -1, "");
509 phi->_inst_mem_id = new_idx;
510 no_of_updates++;
511 }
512 }
513
514 const Type* type = _new_type_array.fast_lookup(n->_idx);
515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
516 if (!_is_pass_finished) {
517 return -1; // delay
518 }
519 int old_idx = type->is_oopptr()->instance_id();
520 int new_idx = new_index(old_idx);
521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
522 _new_type_array.map(n->_idx, new_type);
523 no_of_updates++;
524 }
525
526 return no_of_updates;
527 }
528
529 void PhaseValues::init_con_caches() {
530 memset(_icons,0,sizeof(_icons));
531 memset(_lcons,0,sizeof(_lcons));
532 memset(_zcons,0,sizeof(_zcons));
533 }
534
535 PhaseIterGVN* PhaseValues::is_IterGVN() {
536 return (_phase == PhaseValuesType::iter_gvn || _phase == PhaseValuesType::ccp) ? static_cast<PhaseIterGVN*>(this) : nullptr;
537 }
538
539 //--------------------------------find_int_type--------------------------------
540 const TypeInt* PhaseValues::find_int_type(Node* n) {
541 if (n == nullptr) return nullptr;
542 // Call type_or_null(n) to determine node's type since we might be in
543 // parse phase and call n->Value() may return wrong type.
544 // (For example, a phi node at the beginning of loop parsing is not ready.)
545 const Type* t = type_or_null(n);
546 if (t == nullptr) return nullptr;
547 return t->isa_int();
548 }
549
550
551 //-------------------------------find_long_type--------------------------------
552 const TypeLong* PhaseValues::find_long_type(Node* n) {
553 if (n == nullptr) return nullptr;
554 // (See comment above on type_or_null.)
555 const Type* t = type_or_null(n);
556 if (t == nullptr) return nullptr;
557 return t->isa_long();
558 }
559
560 //------------------------------~PhaseValues-----------------------------------
561 #ifndef PRODUCT
562 PhaseValues::~PhaseValues() {
563 // Statistics for NodeHash
564 _table.dump();
565 // Statistics for value progress and efficiency
566 if( PrintCompilation && Verbose && WizardMode ) {
567 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
568 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
569 if( made_transforms() != 0 ) {
570 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
571 } else {
572 tty->cr();
573 }
574 }
575 }
576 #endif
577
578 //------------------------------makecon----------------------------------------
579 ConNode* PhaseValues::makecon(const Type* t) {
580 assert(t->singleton(), "must be a constant");
581 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
582 switch (t->base()) { // fast paths
583 case Type::Half:
584 case Type::Top: return (ConNode*) C->top();
585 case Type::Int: return intcon( t->is_int()->get_con() );
586 case Type::Long: return longcon( t->is_long()->get_con() );
587 default: break;
588 }
589 if (t->is_zero_type())
590 return zerocon(t->basic_type());
591 return uncached_makecon(t);
592 }
593
594 //--------------------------uncached_makecon-----------------------------------
595 // Make an idealized constant - one of ConINode, ConPNode, etc.
596 ConNode* PhaseValues::uncached_makecon(const Type *t) {
597 assert(t->singleton(), "must be a constant");
598 ConNode* x = ConNode::make(t);
599 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
600 if (k == nullptr) {
601 set_type(x, t); // Missed, provide type mapping
602 GrowableArray<Node_Notes*>* nna = C->node_note_array();
603 if (nna != nullptr) {
604 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
605 loc->clear(); // do not put debug info on constants
606 }
607 } else {
608 x->destruct(this); // Hit, destroy duplicate constant
609 x = k; // use existing constant
610 }
611 return x;
612 }
613
614 //------------------------------intcon-----------------------------------------
615 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
616 ConINode* PhaseValues::intcon(jint i) {
617 // Small integer? Check cache! Check that cached node is not dead
618 if (i >= _icon_min && i <= _icon_max) {
619 ConINode* icon = _icons[i-_icon_min];
620 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
621 return icon;
622 }
623 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
624 assert(icon->is_Con(), "");
625 if (i >= _icon_min && i <= _icon_max)
626 _icons[i-_icon_min] = icon; // Cache small integers
627 return icon;
628 }
629
630 //------------------------------longcon----------------------------------------
631 // Fast long constant.
632 ConLNode* PhaseValues::longcon(jlong l) {
633 // Small integer? Check cache! Check that cached node is not dead
634 if (l >= _lcon_min && l <= _lcon_max) {
635 ConLNode* lcon = _lcons[l-_lcon_min];
636 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
637 return lcon;
638 }
639 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
640 assert(lcon->is_Con(), "");
641 if (l >= _lcon_min && l <= _lcon_max)
642 _lcons[l-_lcon_min] = lcon; // Cache small integers
643 return lcon;
644 }
645 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
646 if (bt == T_INT) {
647 return intcon(checked_cast<jint>(l));
648 }
649 assert(bt == T_LONG, "not an integer");
650 return longcon(l);
651 }
652
653
654 //------------------------------zerocon-----------------------------------------
655 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
656 ConNode* PhaseValues::zerocon(BasicType bt) {
657 assert((uint)bt <= _zcon_max, "domain check");
658 ConNode* zcon = _zcons[bt];
659 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
660 return zcon;
661 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
662 _zcons[bt] = zcon;
663 return zcon;
664 }
665
666
667
668 //=============================================================================
669 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
670 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
671 if (i == nullptr) {
672 i = k->Ideal(this, can_reshape);
673 }
674 return i;
675 }
676
677 //------------------------------transform--------------------------------------
678 // Return a node which computes the same function as this node, but
679 // in a faster or cheaper fashion.
680 Node* PhaseGVN::transform(Node* n) {
681 NOT_PRODUCT( set_transforms(); )
682
683 // Apply the Ideal call in a loop until it no longer applies
684 Node* k = n;
685 Node* i = apply_ideal(k, /*can_reshape=*/false);
686 NOT_PRODUCT(uint loop_count = 1;)
687 while (i != nullptr) {
688 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
689 k = i;
690 #ifdef ASSERT
691 if (loop_count >= K + C->live_nodes()) {
692 dump_infinite_loop_info(i, "PhaseGVN::transform");
693 }
694 #endif
695 i = apply_ideal(k, /*can_reshape=*/false);
696 NOT_PRODUCT(loop_count++;)
697 }
698 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
699
700 // If brand new node, make space in type array.
701 ensure_type_or_null(k);
702
703 // Since I just called 'Value' to compute the set of run-time values
704 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
705 // cache Value. Later requests for the local phase->type of this Node can
706 // use the cached Value instead of suffering with 'bottom_type'.
707 const Type* t = k->Value(this); // Get runtime Value set
708 assert(t != nullptr, "value sanity");
709 if (type_or_null(k) != t) {
710 #ifndef PRODUCT
711 // Do not count initial visit to node as a transformation
712 if (type_or_null(k) == nullptr) {
713 inc_new_values();
714 set_progress();
715 }
716 #endif
717 set_type(k, t);
718 // If k is a TypeNode, capture any more-precise type permanently into Node
719 k->raise_bottom_type(t);
720 }
721
722 if (t->singleton() && !k->is_Con()) {
723 NOT_PRODUCT(set_progress();)
724 return makecon(t); // Turn into a constant
725 }
726
727 // Now check for Identities
728 i = k->Identity(this); // Look for a nearby replacement
729 if (i != k) { // Found? Return replacement!
730 NOT_PRODUCT(set_progress();)
731 return i;
732 }
733
734 // Global Value Numbering
735 i = hash_find_insert(k); // Insert if new
736 if (i && (i != k)) {
737 // Return the pre-existing node
738 NOT_PRODUCT(set_progress();)
739 return i;
740 }
741
742 // Return Idealized original
743 return k;
744 }
745
746 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
747 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
748 return false;
749 }
750 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
751 return false;
752 }
753 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
754 int i = 0;
755 while (d != n) {
756 n = IfNode::up_one_dom(n, linear_only);
757 i++;
758 if (n == nullptr || i >= 100) {
759 return false;
760 }
761 }
762 return true;
763 }
764
765 #ifdef ASSERT
766 //------------------------------dead_loop_check--------------------------------
767 // Check for a simple dead loop when a data node references itself directly
768 // or through an other data node excluding cons and phis.
769 void PhaseGVN::dead_loop_check(Node* n) {
770 // Phi may reference itself in a loop.
771 if (n == nullptr || n->is_dead_loop_safe() || n->is_CFG()) {
772 return;
773 }
774
775 // Do 2 levels check and only data inputs.
776 for (uint i = 1; i < n->req(); i++) {
777 Node* in = n->in(i);
778 if (in == n) {
779 n->dump_bfs(100, nullptr, "");
780 fatal("Dead loop detected, node references itself: %s (%d)",
781 n->Name(), n->_idx);
782 }
783
784 if (in == nullptr || in->is_dead_loop_safe()) {
785 continue;
786 }
787 for (uint j = 1; j < in->req(); j++) {
788 if (in->in(j) == n) {
789 n->dump_bfs(100, nullptr, "");
790 fatal("Dead loop detected, node input references current node: %s (%d) -> %s (%d)",
791 in->Name(), in->_idx, n->Name(), n->_idx);
792 }
793 if (in->in(j) == in) {
794 n->dump_bfs(100, nullptr, "");
795 fatal("Dead loop detected, node input references itself: %s (%d)",
796 in->Name(), in->_idx);
797 }
798 }
799 }
800 }
801
802
803 /**
804 * Dumps information that can help to debug the problem. A debug
805 * build fails with an assert.
806 */
807 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
808 n->dump(4);
809 assert(false, "infinite loop in %s", where);
810 }
811 #endif
812
813 //=============================================================================
814 //------------------------------PhaseIterGVN-----------------------------------
815 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
816 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
817 _worklist(*C->igvn_worklist())
818 {
819 _phase = PhaseValuesType::iter_gvn;
820 assert(&_worklist == &igvn->_worklist, "sanity");
821 }
822
823 //------------------------------PhaseIterGVN-----------------------------------
824 // Initialize from scratch
825 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
826 _worklist(*C->igvn_worklist())
827 {
828 _phase = PhaseValuesType::iter_gvn;
829 uint max;
830
831 // Dead nodes in the hash table inherited from GVN were not treated as
832 // roots during def-use info creation; hence they represent an invisible
833 // use. Clear them out.
834 max = _table.size();
835 for( uint i = 0; i < max; ++i ) {
836 Node *n = _table.at(i);
837 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
838 if( n->is_top() ) continue;
839 // If remove_useless_nodes() has run, we expect no such nodes left.
840 assert(false, "remove_useless_nodes missed this node");
841 hash_delete(n);
842 }
843 }
844
845 // Any Phis or Regions on the worklist probably had uses that could not
846 // make more progress because the uses were made while the Phis and Regions
847 // were in half-built states. Put all uses of Phis and Regions on worklist.
848 max = _worklist.size();
849 for( uint j = 0; j < max; j++ ) {
850 Node *n = _worklist.at(j);
851 uint uop = n->Opcode();
852 if( uop == Op_Phi || uop == Op_Region ||
853 n->is_Type() ||
854 n->is_Mem() )
855 add_users_to_worklist(n);
856 }
857 }
858
859 void PhaseIterGVN::shuffle_worklist() {
860 if (_worklist.size() < 2) return;
861 for (uint i = _worklist.size() - 1; i >= 1; i--) {
862 uint j = C->random() % (i + 1);
863 swap(_worklist.adr()[i], _worklist.adr()[j]);
864 }
865 }
866
867 #ifndef PRODUCT
868 void PhaseIterGVN::verify_step(Node* n) {
869 if (is_verify_def_use()) {
870 ResourceMark rm;
871 VectorSet visited;
872 Node_List worklist;
873
874 _verify_window[_verify_counter % _verify_window_size] = n;
875 ++_verify_counter;
876 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
877 ++_verify_full_passes;
878 worklist.push(C->root());
879 Node::verify(-1, visited, worklist);
880 return;
881 }
882 for (int i = 0; i < _verify_window_size; i++) {
883 Node* n = _verify_window[i];
884 if (n == nullptr) {
885 continue;
886 }
887 if (n->in(0) == NodeSentinel) { // xform_idom
888 _verify_window[i] = n->in(1);
889 --i;
890 continue;
891 }
892 // Typical fanout is 1-2, so this call visits about 6 nodes.
893 if (!visited.test_set(n->_idx)) {
894 worklist.push(n);
895 }
896 }
897 Node::verify(4, visited, worklist);
898 }
899 }
900
901 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
902 const Type* newtype = type_or_null(n);
903 if (nn != n || oldtype != newtype) {
904 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
905 }
906 if (TraceIterativeGVN) {
907 uint wlsize = _worklist.size();
908 if (nn != n) {
909 // print old node
910 tty->print("< ");
911 if (oldtype != newtype && oldtype != nullptr) {
912 oldtype->dump();
913 }
914 do { tty->print("\t"); } while (tty->position() < 16);
915 tty->print("<");
916 n->dump();
917 }
918 if (oldtype != newtype || nn != n) {
919 // print new node and/or new type
920 if (oldtype == nullptr) {
921 tty->print("* ");
922 } else if (nn != n) {
923 tty->print("> ");
924 } else {
925 tty->print("= ");
926 }
927 if (newtype == nullptr) {
928 tty->print("null");
929 } else {
930 newtype->dump();
931 }
932 do { tty->print("\t"); } while (tty->position() < 16);
933 nn->dump();
934 }
935 if (Verbose && wlsize < _worklist.size()) {
936 tty->print(" Push {");
937 while (wlsize != _worklist.size()) {
938 Node* pushed = _worklist.at(wlsize++);
939 tty->print(" %d", pushed->_idx);
940 }
941 tty->print_cr(" }");
942 }
943 if (nn != n) {
944 // ignore n, it might be subsumed
945 verify_step((Node*) nullptr);
946 }
947 }
948 }
949
950 void PhaseIterGVN::init_verifyPhaseIterGVN() {
951 _verify_counter = 0;
952 _verify_full_passes = 0;
953 for (int i = 0; i < _verify_window_size; i++) {
954 _verify_window[i] = nullptr;
955 }
956 #ifdef ASSERT
957 // Verify that all modified nodes are on _worklist
958 Unique_Node_List* modified_list = C->modified_nodes();
959 while (modified_list != nullptr && modified_list->size()) {
960 Node* n = modified_list->pop();
961 if (!n->is_Con() && !_worklist.member(n)) {
962 n->dump();
963 fatal("modified node is not on IGVN._worklist");
964 }
965 }
966 #endif
967 }
968
969 void PhaseIterGVN::verify_PhaseIterGVN() {
970 #ifdef ASSERT
971 // Verify nodes with changed inputs.
972 Unique_Node_List* modified_list = C->modified_nodes();
973 while (modified_list != nullptr && modified_list->size()) {
974 Node* n = modified_list->pop();
975 if (!n->is_Con()) { // skip Con nodes
976 n->dump();
977 fatal("modified node was not processed by IGVN.transform_old()");
978 }
979 }
980 #endif
981
982 C->verify_graph_edges();
983 if (is_verify_def_use() && PrintOpto) {
984 if (_verify_counter == _verify_full_passes) {
985 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
986 (int) _verify_full_passes);
987 } else {
988 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
989 (int) _verify_counter, (int) _verify_full_passes);
990 }
991 }
992
993 #ifdef ASSERT
994 if (modified_list != nullptr) {
995 while (modified_list->size() > 0) {
996 Node* n = modified_list->pop();
997 n->dump();
998 assert(false, "VerifyIterativeGVN: new modified node was added");
999 }
1000 }
1001
1002 verify_optimize();
1003 #endif
1004 }
1005 #endif /* PRODUCT */
1006
1007 #ifdef ASSERT
1008 /**
1009 * Dumps information that can help to debug the problem. A debug
1010 * build fails with an assert.
1011 */
1012 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1013 n->dump(4);
1014 _worklist.dump();
1015 assert(false, "infinite loop in %s", where);
1016 }
1017
1018 /**
1019 * Prints out information about IGVN if the 'verbose' option is used.
1020 */
1021 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1022 if (TraceIterativeGVN && Verbose) {
1023 tty->print(" Pop ");
1024 n->dump();
1025 if ((num_processed % 100) == 0) {
1026 _worklist.print_set();
1027 }
1028 }
1029 }
1030 #endif /* ASSERT */
1031
1032 void PhaseIterGVN::optimize() {
1033 DEBUG_ONLY(uint num_processed = 0;)
1034 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1035 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1036 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1037 if (StressIGVN) {
1038 shuffle_worklist();
1039 }
1040
1041 // The node count check in the loop below (check_node_count) assumes that we
1042 // increase the live node count with at most
1043 // max_live_nodes_increase_per_iteration in between checks. If this
1044 // assumption does not hold, there is a risk that we exceed the max node
1045 // limit in between checks and trigger an assert during node creation.
1046 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 3;
1047
1048 uint loop_count = 0;
1049 // Pull from worklist and transform the node. If the node has changed,
1050 // update edge info and put uses on worklist.
1051 while (_worklist.size() > 0) {
1052 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1053 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1054 return;
1055 }
1056 Node* n = _worklist.pop();
1057 if (loop_count >= K * C->live_nodes()) {
1058 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1059 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1060 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1061 return;
1062 }
1063 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1064 if (n->outcnt() != 0) {
1065 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1066 // Do the transformation
1067 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1068 Node* nn = transform_old(n);
1069 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1070 // Ensure we did not increase the live node count with more than
1071 // max_live_nodes_increase_per_iteration during the call to transform_old
1072 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1073 assert(increase < max_live_nodes_increase_per_iteration,
1074 "excessive live node increase in single iteration of IGVN: %d "
1075 "(should be at most %d)",
1076 increase, max_live_nodes_increase_per_iteration);
1077 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1078 } else if (!n->is_top()) {
1079 remove_dead_node(n);
1080 }
1081 loop_count++;
1082 }
1083 NOT_PRODUCT(verify_PhaseIterGVN();)
1084 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1085 }
1086
1087 #ifdef ASSERT
1088 void PhaseIterGVN::verify_optimize() {
1089 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1090
1091 if (is_verify_Value() ||
1092 is_verify_Ideal() ||
1093 is_verify_Identity() ||
1094 is_verify_invariants()) {
1095 ResourceMark rm;
1096 Unique_Node_List worklist;
1097 // BFS all nodes, starting at root
1098 worklist.push(C->root());
1099 for (uint j = 0; j < worklist.size(); ++j) {
1100 Node* n = worklist.at(j);
1101 // If we get an assert here, check why the reported node was not processed again in IGVN.
1102 // We should either make sure that this node is properly added back to the IGVN worklist
1103 // in PhaseIterGVN::add_users_to_worklist to update it again or add an exception
1104 // in the verification methods below if that is not possible for some reason (like Load nodes).
1105 if (is_verify_Value()) {
1106 verify_Value_for(n);
1107 }
1108 if (is_verify_Ideal()) {
1109 verify_Ideal_for(n, false);
1110 verify_Ideal_for(n, true);
1111 }
1112 if (is_verify_Identity()) {
1113 verify_Identity_for(n);
1114 }
1115 if (is_verify_invariants()) {
1116 verify_node_invariants_for(n);
1117 }
1118
1119 // traverse all inputs and outputs
1120 for (uint i = 0; i < n->req(); i++) {
1121 if (n->in(i) != nullptr) {
1122 worklist.push(n->in(i));
1123 }
1124 }
1125 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1126 worklist.push(n->fast_out(i));
1127 }
1128 }
1129 }
1130
1131 verify_empty_worklist(nullptr);
1132 }
1133
1134 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1135 // Verify that the igvn worklist is empty. If no optimization happened, then
1136 // nothing needs to be on the worklist.
1137 if (_worklist.size() == 0) { return; }
1138
1139 stringStream ss; // Print as a block without tty lock.
1140 for (uint j = 0; j < _worklist.size(); j++) {
1141 Node* n = _worklist.at(j);
1142 ss.print("igvn.worklist[%d] ", j);
1143 n->dump("\n", false, &ss);
1144 }
1145 if (_worklist.size() != 0 && node != nullptr) {
1146 ss.print_cr("Previously optimized:");
1147 node->dump("\n", false, &ss);
1148 }
1149 tty->print_cr("%s", ss.as_string());
1150 assert(false, "igvn worklist must still be empty after verify");
1151 }
1152
1153 // Check that type(n) == n->Value(), asserts if we have a failure.
1154 // We have a list of exceptions, see detailed comments in code.
1155 // (1) Integer "widen" changes, but the range is the same.
1156 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1157 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1158 void PhaseIterGVN::verify_Value_for(const Node* n, bool strict) {
1159 // If we assert inside type(n), because the type is still a null, then maybe
1160 // the node never went through gvn.transform, which would be a bug.
1161 const Type* told = type(n);
1162 const Type* tnew = n->Value(this);
1163 if (told == tnew) {
1164 return;
1165 }
1166 // Exception (1)
1167 // Integer "widen" changes, but range is the same.
1168 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1169 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1170 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1171 if (t0->lo_as_long() == t1->lo_as_long() &&
1172 t0->hi_as_long() == t1->hi_as_long()) {
1173 return; // ignore integer widen
1174 }
1175 }
1176 // Exception (2)
1177 // LoadNode performs deep traversals. Load is not notified for changes far away.
1178 if (!strict && n->is_Load() && !told->singleton()) {
1179 // MemNode::can_see_stored_value looks up through many memory nodes,
1180 // which means we would need to notify modifications from far up in
1181 // the inputs all the way down to the LoadNode. We don't do that.
1182 return;
1183 }
1184 // Exception (3)
1185 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1186 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1187 // SubNode::Value
1188 // CmpPNode::sub
1189 // MemNode::detect_ptr_independence
1190 // MemNode::all_controls_dominate
1191 // We find all controls of a pointer load, and see if they dominate the control of
1192 // an allocation. If they all dominate, we know the allocation is after (independent)
1193 // of the pointer load, and we can say the pointers are different. For this we call
1194 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1195 // control sub of the allocation. The problems is that sometimes dominates answers
1196 // false conservatively, and later it can determine that it is indeed true. Loops with
1197 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1198 // so the traversal becomes more powerful. This is difficult to remedy, we would have
1199 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1200 // after loop-opts, so that should take care of many of these cases.
1201 return;
1202 }
1203
1204 stringStream ss; // Print as a block without tty lock.
1205 ss.cr();
1206 ss.print_cr("Missed Value optimization:");
1207 n->dump_bfs(1, nullptr, "", &ss);
1208 ss.print_cr("Current type:");
1209 told->dump_on(&ss);
1210 ss.cr();
1211 ss.print_cr("Optimized type:");
1212 tnew->dump_on(&ss);
1213 ss.cr();
1214 tty->print_cr("%s", ss.as_string());
1215
1216 switch (_phase) {
1217 case PhaseValuesType::iter_gvn:
1218 assert(false, "Missed Value optimization opportunity in PhaseIterGVN for %s",n->Name());
1219 break;
1220 case PhaseValuesType::ccp:
1221 assert(false, "PhaseCCP not at fixpoint: analysis result may be unsound for %s", n->Name());
1222 break;
1223 default:
1224 assert(false, "Unexpected phase");
1225 break;
1226 }
1227 }
1228
1229 // Check that all Ideal optimizations that could be done were done.
1230 // Asserts if it found missed optimization opportunities or encountered unexpected changes, and
1231 // returns normally otherwise (no missed optimization, or skipped verification).
1232 void PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1233 // First, we check a list of exceptions, where we skip verification,
1234 // because there are known cases where Ideal can optimize after IGVN.
1235 // Some may be expected and cannot be fixed, and others should be fixed.
1236 switch (n->Opcode()) {
1237 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1238 // (see "Range-Check scan limit"). So, it is possible that something
1239 // is optimized in that input subgraph, and the RangeCheck was not
1240 // added to the worklist because it would be too expensive to walk
1241 // down the graph for 1000 nodes and put all on the worklist.
1242 //
1243 // Found with:
1244 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1245 case Op_RangeCheck:
1246 return;
1247
1248 // IfNode::Ideal does:
1249 // Node* prev_dom = search_identical(dist, igvn);
1250 // which means we seach up the CFG, traversing at most up to a distance.
1251 // If anything happens rather far away from the If, we may not put the If
1252 // back on the worklist.
1253 //
1254 // Found with:
1255 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1256 case Op_If:
1257 return;
1258
1259 // IfNode::simple_subsuming
1260 // Looks for dominating test that subsumes the current test.
1261 // Notification could be difficult because of larger distance.
1262 //
1263 // Found with:
1264 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1265 // -XX:VerifyIterativeGVN=1110
1266 case Op_CountedLoopEnd:
1267 return;
1268
1269 // LongCountedLoopEndNode::Ideal
1270 // Probably same issue as above.
1271 //
1272 // Found with:
1273 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1274 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1275 case Op_LongCountedLoopEnd:
1276 return;
1277
1278 // RegionNode::Ideal does "Skip around the useless IF diamond".
1279 // 245 IfTrue === 244
1280 // 258 If === 245 257
1281 // 259 IfTrue === 258 [[ 263 ]]
1282 // 260 IfFalse === 258 [[ 263 ]]
1283 // 263 Region === 263 260 259 [[ 263 268 ]]
1284 // to
1285 // 245 IfTrue === 244
1286 // 263 Region === 263 245 _ [[ 263 268 ]]
1287 //
1288 // "Useless" means that there is no code in either branch of the If.
1289 // I found a case where this was not done yet during IGVN.
1290 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1291 //
1292 // Found with:
1293 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1294 case Op_Region:
1295 return;
1296
1297 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1298 // that smaller idx are first. Tracking it back, it led me to
1299 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1300 //
1301 // Example:
1302 // Before PhaseIdealLoop::remix_address_expressions
1303 // 154 AddI === _ 12 144
1304 // After PhaseIdealLoop::remix_address_expressions
1305 // 154 AddI === _ 144 12
1306 // After AddNode::Ideal
1307 // 154 AddI === _ 12 144
1308 //
1309 // I suspect that the node should be added to the IGVN worklist after
1310 // PhaseIdealLoop::remix_address_expressions
1311 //
1312 // This is the only case I looked at, there may be others. Found like this:
1313 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1314 //
1315 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1316 //
1317 // Note: currently all of these fail also for other reasons, for example
1318 // because of "commute" doing the reordering with the phi below. Once
1319 // that is resolved, we can come back to this issue here.
1320 //
1321 // case Op_AddD:
1322 // case Op_AddI:
1323 // case Op_AddL:
1324 // case Op_AddF:
1325 // case Op_MulI:
1326 // case Op_MulL:
1327 // case Op_MulF:
1328 // case Op_MulD:
1329 // if (n->in(1)->_idx > n->in(2)->_idx) {
1330 // // Expect "commute" to revert this case.
1331 // return false;
1332 // }
1333 // break; // keep verifying
1334
1335 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1336 // Check for tight loop increments: Loop-phi of Add of loop-phi
1337 // It wants to take the phi into in(1):
1338 // 471 Phi === 435 38 390
1339 // 390 AddF === _ 471 391
1340 //
1341 // Other Associative operators are also affected equally.
1342 //
1343 // Investigate why this does not happen earlier during IGVN.
1344 //
1345 // Found with:
1346 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1347 // -XX:VerifyIterativeGVN=1110
1348 case Op_AddD:
1349 //case Op_AddI: // Also affected for other reasons, see case further down.
1350 //case Op_AddL: // Also affected for other reasons, see case further down.
1351 case Op_AddF:
1352 case Op_MulI:
1353 case Op_MulL:
1354 case Op_MulF:
1355 case Op_MulD:
1356 case Op_MinF:
1357 case Op_MinD:
1358 case Op_MaxF:
1359 case Op_MaxD:
1360 // XorINode::Ideal
1361 // Found with:
1362 // compiler/intrinsics/chacha/TestChaCha20.java
1363 // -XX:VerifyIterativeGVN=1110
1364 case Op_XorI:
1365 case Op_XorL:
1366 // It seems we may have similar issues with the HF cases.
1367 // Found with aarch64:
1368 // compiler/vectorization/TestFloat16VectorOperations.java
1369 // -XX:VerifyIterativeGVN=1110
1370 case Op_AddHF:
1371 case Op_MulHF:
1372 case Op_MaxHF:
1373 case Op_MinHF:
1374 return;
1375
1376 // In MulNode::Ideal the edges can be swapped to help value numbering:
1377 //
1378 // // We are OK if right is a constant, or right is a load and
1379 // // left is a non-constant.
1380 // if( !(t2->singleton() ||
1381 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1382 // if( t1->singleton() || // Left input is a constant?
1383 // // Otherwise, sort inputs (commutativity) to help value numbering.
1384 // (in(1)->_idx > in(2)->_idx) ) {
1385 // swap_edges(1, 2);
1386 //
1387 // Why was this not done earlier during IGVN?
1388 //
1389 // Found with:
1390 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1391 // -XX:VerifyIterativeGVN=1110
1392 case Op_AndI:
1393 // Same for AndL.
1394 // Found with:
1395 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1396 // -XX:VerifyIterativeGVN=1110
1397 case Op_AndL:
1398 return;
1399
1400 // SubLNode::Ideal does transform like:
1401 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1402 //
1403 // In IGVN before verification:
1404 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1405 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1406 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1407 // 8016 ConL === 0 [[ 8383 ]] #long:0
1408 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1409 //
1410 // And then in verification:
1411 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1412 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1413 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1414 // 8016 ConL === 0 [[ 8383 ]] #long:0
1415 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1416 //
1417 // So the form changed from:
1418 // c1 - (y + [8423 ConvI2L])
1419 // to
1420 // c1 - (y + -2)
1421 // but the SubL was not added to the IGVN worklist. Investigate why.
1422 // There could be other issues too.
1423 //
1424 // There seems to be a related AddL IGVN optimization that triggers
1425 // the same SubL optimization, so investigate that too.
1426 //
1427 // Found with:
1428 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1429 case Op_SubL:
1430 return;
1431
1432 // SubINode::Ideal does
1433 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1434 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1435 //
1436 // Investigate why this does not yet happen during IGVN.
1437 //
1438 // Found with:
1439 // test/hotspot/jtreg/compiler/c2/IVTest.java
1440 // -XX:VerifyIterativeGVN=1110
1441 case Op_SubI:
1442 return;
1443
1444 // AddNode::IdealIL does transform like:
1445 // Convert x + (con - y) into "(x - y) + con"
1446 //
1447 // In IGVN before verification:
1448 // 8382 ConvI2L
1449 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1450 // 8383 SubL === _ 8381 8382
1451 // 8168 ConvI2L
1452 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1453 //
1454 // And then in verification:
1455 // 8424 AddL
1456 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1457 // 8383 SubL === _ 8016 8424
1458 // 8168 ConvI2L
1459 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1460 //
1461 // So the form changed from:
1462 // x + (ConvI2L(0) - [8382 ConvI2L])
1463 // to
1464 // x + (0 - [8424 AddL])
1465 // but the AddL was not added to the IGVN worklist. Investigate why.
1466 // There could be other issues, too. For example with "commute", see above.
1467 //
1468 // Found with:
1469 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1470 case Op_AddL:
1471 return;
1472
1473 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1474 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1475 // record_for_cleanup(cmp, phase);
1476 // This verification code in the Ideal code creates new nodes, and checks
1477 // if they fold in unexpected ways. This means some nodes are created and
1478 // added to the worklist, even if the SubTypeCheck is not optimized. This
1479 // goes agains the assumption of the verification here, which assumes that
1480 // if the node is not optimized, then no new nodes should be created, and
1481 // also no nodes should be added to the worklist.
1482 // I see two options:
1483 // 1) forbid what verify_helper does, because for each Ideal call it
1484 // uses memory and that is suboptimal. But it is not clear how that
1485 // verification can be done otherwise.
1486 // 2) Special case the verification here. Probably the new nodes that
1487 // were just created are dead, i.e. they are not connected down to
1488 // root. We could verify that, and remove those nodes from the graph
1489 // by setting all their inputs to nullptr. And of course we would
1490 // have to remove those nodes from the worklist.
1491 // Maybe there are other options too, I did not dig much deeper yet.
1492 //
1493 // Found with:
1494 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1495 case Op_SubTypeCheck:
1496 return;
1497
1498 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1499 // to other nodes: Conv, Add, Sub, Mul, And ...
1500 //
1501 // 107 ConI === 0 [[ ... ]] #int:2
1502 // 84 LoadRange === _ 7 83
1503 // 50 ConI === 0 [[ ... ]] #int:0
1504 // 549 LoopLimit === _ 50 84 107
1505 //
1506 // I stepped backward, to see how the node was generated, and I found that it was
1507 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1508 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1509 // that, and it seems we want to skip lowering it until after loop-opts, but never
1510 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1511 // could be other issues too.
1512 //
1513 // Fond with:
1514 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1515 case Op_LoopLimit:
1516 return;
1517
1518 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1519 // "This optimization tries to find two or more inputs of phi with the same constant
1520 // value. It then splits them into a separate Phi, and according Region."
1521 //
1522 // Example:
1523 // 130 DecodeN === _ 129
1524 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1525 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1526 //
1527 // turns into:
1528 //
1529 // 50 ConP === 0 [[ 99 91 18 ]] #null
1530 // 130 DecodeN === _ 129 [[ 18 ]]
1531 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1532 //
1533 // We would have to investigate why this optimization does not happen during IGVN.
1534 // There could also be other issues - I did not investigate further yet.
1535 //
1536 // Found with:
1537 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1538 case Op_Phi:
1539 return;
1540
1541 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1542 // For examle "The allocated object does not escape".
1543 //
1544 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1545 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1546 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1547 // worklist when this change happens?
1548 //
1549 // Found with:
1550 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1551 case Op_MemBarStoreStore:
1552 return;
1553
1554 // ConvI2LNode::Ideal converts
1555 // 648 AddI === _ 583 645 [[ 661 ]]
1556 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1557 // into
1558 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1559 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1560 // 773 AddL === _ 771 772 [[ ]]
1561 //
1562 // We have to investigate why this does not happen during IGVN in this case.
1563 // There could also be other issues - I did not investigate further yet.
1564 //
1565 // Found with:
1566 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1567 case Op_ConvI2L:
1568 return;
1569
1570 // AddNode::IdealIL can do this transform (and similar other ones):
1571 // Convert "a*b+a*c into a*(b+c)
1572 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1573 // during IGVN? There was a mutation for one of the MulI, and only
1574 // after that the pattern was as needed for the optimization. The MulI
1575 // was added to the IGVN worklist, but not the AddI. This probably
1576 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1577 //
1578 // Found with:
1579 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1580 // -XX:VerifyIterativeGVN=1110
1581 case Op_AddI:
1582 return;
1583
1584 // ArrayCopyNode::Ideal
1585 // calls ArrayCopyNode::prepare_array_copy
1586 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1587 // is not expected, and we create a range int:0..-1
1588 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1589 // note: the type is already empty!
1590 // calls PhaseIterGVN::transform
1591 // calls PhaseIterGVN::transform_old
1592 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1593 // calls Unique_Node_List::push -> pushes TOP to worklist
1594 //
1595 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1596 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1597 //
1598 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1599 // return nullptr, indicating nothing happened?
1600 // Is it smart to do transform in Compile::constrained_convI2L, and then
1601 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1602 // Should we just allow TOP to land on the worklist, as an exception?
1603 //
1604 // Found with:
1605 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1606 // -XX:VerifyIterativeGVN=1110
1607 case Op_ArrayCopy:
1608 return;
1609
1610 // CastLLNode::Ideal
1611 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1612 //
1613 // Could be a notification issue, where updates inputs of CastLL do not notify
1614 // down through SubL to CastLL.
1615 //
1616 // Found With:
1617 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1618 // -XX:VerifyIterativeGVN=1110
1619 case Op_CastLL:
1620 return;
1621
1622 // Similar case happens to CastII
1623 //
1624 // Found With:
1625 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1626 // -XX:VerifyIterativeGVN=1110
1627 case Op_CastII:
1628 return;
1629
1630 // MaxLNode::Ideal
1631 // calls AddNode::Ideal
1632 // calls commute -> decides to swap edges
1633 //
1634 // Another notification issue, because we check inputs of inputs?
1635 // MaxL -> Phi -> Loop
1636 // MaxL -> Phi -> MaxL
1637 //
1638 // Found with:
1639 // compiler/c2/irTests/TestIfMinMax.java
1640 // -XX:VerifyIterativeGVN=1110
1641 case Op_MaxL:
1642 case Op_MinL:
1643 return;
1644
1645 // OrINode::Ideal
1646 // calls AddNode::Ideal
1647 // calls commute -> left is Load, right not -> commute.
1648 //
1649 // Not sure why notification does not work here, seems like
1650 // the depth is only 1, so it should work. Needs investigation.
1651 //
1652 // Found with:
1653 // compiler/codegen/TestCharVect2.java#id0
1654 // -XX:VerifyIterativeGVN=1110
1655 case Op_OrI:
1656 case Op_OrL:
1657 return;
1658
1659 // Bool -> constant folded to 1.
1660 // Issue with notification?
1661 //
1662 // Found with:
1663 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1664 // -XX:VerifyIterativeGVN=1110
1665 case Op_Bool:
1666 return;
1667
1668 // LShiftLNode::Ideal
1669 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1670 // Probably a notification issue.
1671 //
1672 // Found with:
1673 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1674 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1675 case Op_LShiftL:
1676 return;
1677
1678 // LShiftINode::Ideal
1679 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1680 // Could be issue with notification of inputs of inputs
1681 //
1682 // Side-note: should cases like these not be shared between
1683 // LShiftI and LShiftL?
1684 //
1685 // Found with:
1686 // compiler/escapeAnalysis/Test6689060.java
1687 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1688 case Op_LShiftI:
1689 return;
1690
1691 // AddPNode::Ideal seems to do set_req without removing lock first.
1692 // Found with various vector tests tier1-tier3.
1693 case Op_AddP:
1694 return;
1695
1696 // StrIndexOfNode::Ideal
1697 // Found in tier1-3.
1698 case Op_StrIndexOf:
1699 case Op_StrIndexOfChar:
1700 return;
1701
1702 // StrEqualsNode::Identity
1703 //
1704 // Found (linux x64 only?) with:
1705 // serviceability/sa/ClhsdbThreadContext.java
1706 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1707 // Note: The -XX:LockingMode option is not available anymore.
1708 case Op_StrEquals:
1709 return;
1710
1711 // AryEqNode::Ideal
1712 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1713 //
1714 // Found with:
1715 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1716 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1717 case Op_AryEq:
1718 return;
1719
1720 // MergeMemNode::Ideal
1721 // Found in tier1-3. Did not investigate further yet.
1722 case Op_MergeMem:
1723 return;
1724
1725 // URShiftINode::Ideal
1726 // Found in tier1-3. Did not investigate further yet.
1727 case Op_URShiftI:
1728 return;
1729
1730 // CMoveINode::Ideal
1731 // Found in tier1-3. Did not investigate further yet.
1732 case Op_CMoveI:
1733 return;
1734
1735 // CmpPNode::Ideal calls isa_const_java_mirror
1736 // and generates new constant nodes, even if no progress is made.
1737 // We can probably rewrite this so that only types are generated.
1738 // It seems that object types are not hashed, we could investigate
1739 // if that is an option as well.
1740 //
1741 // Found with:
1742 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1743 case Op_CmpP:
1744 return;
1745
1746 // MinINode::Ideal
1747 // Did not investigate, but there are some patterns that might
1748 // need more notification.
1749 case Op_MinI:
1750 case Op_MaxI: // preemptively removed it as well.
1751 return;
1752 }
1753
1754 if (n->is_Load()) {
1755 // LoadNode::Ideal uses tries to find an earlier memory state, and
1756 // checks can_see_stored_value for it.
1757 //
1758 // Investigate why this was not already done during IGVN.
1759 // A similar issue happens with Identity.
1760 //
1761 // There seem to be other cases where loads go up some steps, like
1762 // LoadNode::Ideal going up 10x steps to find dominating load.
1763 //
1764 // Found with:
1765 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1766 // -XX:VerifyIterativeGVN=1110
1767 return;
1768 }
1769
1770 if (n->is_Store()) {
1771 // StoreNode::Ideal can do this:
1772 // // Capture an unaliased, unconditional, simple store into an initializer.
1773 // // Or, if it is independent of the allocation, hoist it above the allocation.
1774 // That replaces the Store with a MergeMem.
1775 //
1776 // We have to investigate why this does not happen during IGVN in this case.
1777 // There could also be other issues - I did not investigate further yet.
1778 //
1779 // Found with:
1780 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1781 return;
1782 }
1783
1784 if (n->is_Vector()) {
1785 // VectorNode::Ideal swaps edges, but only for ops
1786 // that are deemed commutable. But swap_edges
1787 // requires the hash to be invariant when the edges
1788 // are swapped, which is not implemented for these
1789 // vector nodes. This seems not to create any trouble
1790 // usually, but we can also get graphs where in the
1791 // end the nodes are not all commuted, so there is
1792 // definitively an issue here.
1793 //
1794 // Probably we have two options: kill the hash, or
1795 // properly make the hash commutation friendly.
1796 //
1797 // Found with:
1798 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1799 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1800 return;
1801 }
1802
1803 if (n->is_Region()) {
1804 // LoopNode::Ideal calls RegionNode::Ideal.
1805 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1806 // But I got an issue because RegionNode::optimize_trichotomy
1807 // then modifies another node, and pushes nodes to the worklist
1808 // Not sure if this is ok, modifying another node like that.
1809 // Maybe it is, then we need to look into what to do with
1810 // the nodes that are now on the worklist, maybe just clear
1811 // them out again. But maybe modifying other nodes like that
1812 // is also bad design. In the end, we return nullptr for
1813 // the current CountedLoop. But the extra nodes on the worklist
1814 // trip the asserts later on.
1815 //
1816 // Found with:
1817 // compiler/eliminateAutobox/TestShortBoxing.java
1818 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1819 return;
1820 }
1821
1822 if (n->is_CallJava()) {
1823 // CallStaticJavaNode::Ideal
1824 // Led to a crash:
1825 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1826 //
1827 // Did not investigate yet, could be a bug.
1828 // Or maybe it does not expect to be called during verification.
1829 //
1830 // Found with:
1831 // test/jdk/jdk/incubator/vector/VectorRuns.java
1832 // -XX:VerifyIterativeGVN=1110
1833
1834 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1835 // and possibly their subclasses.
1836 // During late inlining it can call CallJavaNode::register_for_late_inline
1837 // That means we do more rounds of late inlining, but might fail.
1838 // Then we do IGVN again, and register the node again for late inlining.
1839 // This creates an endless cycle. Everytime we try late inlining, we
1840 // are also creating more nodes, especially SafePoint and MergeMem.
1841 // These nodes are immediately rejected when the inlining fails in the
1842 // do_late_inline_check, but they still grow the memory, until we hit
1843 // the MemLimit and crash.
1844 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1845 // called repeatedly, and eventually we terminate. I fear this is not
1846 // a great assumption to make. We should investigate more.
1847 //
1848 // Found with:
1849 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1850 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1851 return;
1852 }
1853
1854 // The number of nodes shoud not increase.
1855 uint old_unique = C->unique();
1856 // The hash of a node should not change, this would indicate different inputs
1857 uint old_hash = n->hash();
1858 // Remove 'n' from hash table in case it gets modified. We want to avoid
1859 // hitting the "Need to remove from hash before changing edges" assert if
1860 // a change occurs. Instead, we would like to proceed with the optimization,
1861 // return and finally hit the assert in PhaseIterGVN::verify_optimize to get
1862 // a more meaningful message
1863 _table.hash_delete(n);
1864 Node* i = n->Ideal(this, can_reshape);
1865 // If there was no new Idealization, we are probably happy.
1866 if (i == nullptr) {
1867 if (old_unique < C->unique()) {
1868 stringStream ss; // Print as a block without tty lock.
1869 ss.cr();
1870 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1871 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1872 n->dump_bfs(1, nullptr, "", &ss);
1873 tty->print_cr("%s", ss.as_string());
1874 assert(false, "Unexpected new unused nodes from applying Ideal optimization on %s", n->Name());
1875 }
1876
1877 if (old_hash != n->hash()) {
1878 stringStream ss; // Print as a block without tty lock.
1879 ss.cr();
1880 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1881 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1882 n->dump_bfs(1, nullptr, "", &ss);
1883 tty->print_cr("%s", ss.as_string());
1884 assert(false, "Unexpected hash change from applying Ideal optimization on %s", n->Name());
1885 }
1886
1887 verify_empty_worklist(n);
1888
1889 // Everything is good.
1890 hash_find_insert(n);
1891 return;
1892 }
1893
1894 // We just saw a new Idealization which was not done during IGVN.
1895 stringStream ss; // Print as a block without tty lock.
1896 ss.cr();
1897 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1898 if (i == n) {
1899 ss.print_cr("The node was reshaped by Ideal.");
1900 } else {
1901 ss.print_cr("The node was replaced by Ideal.");
1902 ss.print_cr("Old node:");
1903 n->dump_bfs(1, nullptr, "", &ss);
1904 }
1905 ss.print_cr("The result after Ideal:");
1906 i->dump_bfs(1, nullptr, "", &ss);
1907 tty->print_cr("%s", ss.as_string());
1908
1909 assert(false, "Missed Ideal optimization opportunity in PhaseIterGVN for %s", n->Name());
1910 }
1911
1912 // Check that all Identity optimizations that could be done were done.
1913 // Asserts if it found missed optimization opportunities, and
1914 // returns normally otherwise (no missed optimization, or skipped verification).
1915 void PhaseIterGVN::verify_Identity_for(Node* n) {
1916 // First, we check a list of exceptions, where we skip verification,
1917 // because there are known cases where Ideal can optimize after IGVN.
1918 // Some may be expected and cannot be fixed, and others should be fixed.
1919 switch (n->Opcode()) {
1920 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1921 // after loopopts:
1922 // // Transforming long counted loops requires a safepoint node. Do not
1923 // // eliminate a safepoint until loop opts are over.
1924 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1925 //
1926 // I think the check for major_progress does delay it until after loopopts
1927 // but it does not ensure that the node is on the IGVN worklist after
1928 // loopopts. I think we should try to instead check for
1929 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1930 //
1931 // Found with:
1932 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1933 case Op_SafePoint:
1934 return;
1935
1936 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1937 // does not record any other memory splits.
1938 //
1939 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1940 // never got called during IGVN for this node, investigate why.
1941 //
1942 // Found with:
1943 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1944 case Op_MergeMem:
1945 return;
1946
1947 // ConstraintCastNode::Identity finds casts that are the same, except that
1948 // the control is "higher up", i.e. dominates. The call goes via
1949 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1950 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1951 // away from the cast, but within 100 idom steps, the cast may not be
1952 // put on the IGVN worklist any more.
1953 //
1954 // Found with:
1955 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1956 case Op_CastPP:
1957 case Op_CastII:
1958 case Op_CastLL:
1959 return;
1960
1961 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1962 // checks dominator, which may be changed, but too far up for notification
1963 // to work.
1964 //
1965 // Found with:
1966 // compiler/c2/irTests/TestSkeletonPredicates.java
1967 // -XX:VerifyIterativeGVN=1110
1968 case Op_CheckCastPP:
1969 return;
1970
1971 // In SubNode::Identity, we do:
1972 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1973 // In the example, the AddI had an input replaced, the AddI is
1974 // added to the IGVN worklist, but the SubI is one link further
1975 // down and is not added. I checked add_users_of_use_to_worklist
1976 // where I would expect the SubI would be added, and I cannot
1977 // find the pattern, only this one:
1978 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1979 //
1980 // Fix this "notification" issue and check if there are any other
1981 // issues.
1982 //
1983 // Found with:
1984 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1985 case Op_SubI:
1986 case Op_SubL:
1987 return;
1988
1989 // PhiNode::Identity checks for patterns like:
1990 // r = (x != con) ? x : con;
1991 // that can be constant folded to "x".
1992 //
1993 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1994 // I suspect there was some earlier change to one of the inputs, but
1995 // not all relevant outputs were put on the IGVN worklist.
1996 //
1997 // Found with:
1998 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1999 // -XX:VerifyIterativeGVN=1110
2000 case Op_Phi:
2001 return;
2002
2003 // ConvI2LNode::Identity does
2004 // convert I2L(L2I(x)) => x
2005 //
2006 // Investigate why this did not already happen during IGVN.
2007 //
2008 // Found with:
2009 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
2010 // -XX:VerifyIterativeGVN=1110
2011 case Op_ConvI2L:
2012 return;
2013
2014 // MaxNode::find_identity_operation
2015 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
2016 // This can be a 2-hop search, so maybe notification is not
2017 // good enough.
2018 //
2019 // Found with:
2020 // compiler/codegen/TestBooleanVect.java
2021 // -XX:VerifyIterativeGVN=1110
2022 case Op_MaxL:
2023 case Op_MinL:
2024 case Op_MaxI:
2025 case Op_MinI:
2026 case Op_MaxF:
2027 case Op_MinF:
2028 case Op_MaxHF:
2029 case Op_MinHF:
2030 case Op_MaxD:
2031 case Op_MinD:
2032 return;
2033
2034
2035 // AddINode::Identity
2036 // Converts (x-y)+y to x
2037 // Could be issue with notification
2038 //
2039 // Turns out AddL does the same.
2040 //
2041 // Found with:
2042 // compiler/c2/Test6792161.java
2043 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2044 case Op_AddI:
2045 case Op_AddL:
2046 return;
2047
2048 // AbsINode::Identity
2049 // Not investigated yet.
2050 case Op_AbsI:
2051 return;
2052 }
2053
2054 if (n->is_Load()) {
2055 // LoadNode::Identity tries to look for an earlier store value via
2056 // can_see_stored_value. I found an example where this led to
2057 // an Allocation, where we could assume the value was still zero.
2058 // So the LoadN can be replaced with a zerocon.
2059 //
2060 // Investigate why this was not already done during IGVN.
2061 // A similar issue happens with Ideal.
2062 //
2063 // Found with:
2064 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2065 return;
2066 }
2067
2068 if (n->is_Store()) {
2069 // StoreNode::Identity
2070 // Not investigated, but found missing optimization for StoreI.
2071 // Looks like a StoreI is replaced with an InitializeNode.
2072 //
2073 // Found with:
2074 // applications/ctw/modules/java_base_2.java
2075 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2076 return;
2077 }
2078
2079 if (n->is_Vector()) {
2080 // Found with tier1-3. Not investigated yet.
2081 // The observed issue was with AndVNode::Identity
2082 return;
2083 }
2084
2085 Node* i = n->Identity(this);
2086 // If we cannot find any other Identity, we are happy.
2087 if (i == n) {
2088 verify_empty_worklist(n);
2089 return;
2090 }
2091
2092 // The verification just found a new Identity that was not found during IGVN.
2093 stringStream ss; // Print as a block without tty lock.
2094 ss.cr();
2095 ss.print_cr("Missed Identity optimization:");
2096 ss.print_cr("Old node:");
2097 n->dump_bfs(1, nullptr, "", &ss);
2098 ss.print_cr("New node:");
2099 i->dump_bfs(1, nullptr, "", &ss);
2100 tty->print_cr("%s", ss.as_string());
2101
2102 assert(false, "Missed Identity optimization opportunity in PhaseIterGVN for %s", n->Name());
2103 }
2104
2105 // Some other verifications that are not specific to a particular transformation.
2106 void PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2107 if (n->is_AddP()) {
2108 if (!n->as_AddP()->address_input_has_same_base()) {
2109 stringStream ss; // Print as a block without tty lock.
2110 ss.cr();
2111 ss.print_cr("Base pointers must match for AddP chain:");
2112 n->dump_bfs(2, nullptr, "", &ss);
2113 tty->print_cr("%s", ss.as_string());
2114
2115 assert(false, "Broken node invariant for %s", n->Name());
2116 }
2117 }
2118 }
2119 #endif
2120
2121 /**
2122 * Register a new node with the optimizer. Update the types array, the def-use
2123 * info. Put on worklist.
2124 */
2125 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2126 set_type_bottom(n);
2127 _worklist.push(n);
2128 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2129 return n;
2130 }
2131
2132 //------------------------------transform--------------------------------------
2133 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2134 Node *PhaseIterGVN::transform( Node *n ) {
2135 if (_delay_transform) {
2136 // Register the node but don't optimize for now
2137 register_new_node_with_optimizer(n);
2138 return n;
2139 }
2140
2141 // If brand new node, make space in type array, and give it a type.
2142 ensure_type_or_null(n);
2143 if (type_or_null(n) == nullptr) {
2144 set_type_bottom(n);
2145 }
2146
2147 return transform_old(n);
2148 }
2149
2150 Node *PhaseIterGVN::transform_old(Node* n) {
2151 NOT_PRODUCT(set_transforms());
2152 // Remove 'n' from hash table in case it gets modified
2153 _table.hash_delete(n);
2154 #ifdef ASSERT
2155 if (is_verify_def_use()) {
2156 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2157 }
2158 #endif
2159
2160 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2161 if (n->is_Cmp()) {
2162 add_users_to_worklist(n);
2163 }
2164
2165 // Apply the Ideal call in a loop until it no longer applies
2166 Node* k = n;
2167 DEBUG_ONLY(dead_loop_check(k);)
2168 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2169 C->remove_modified_node(k);
2170 Node* i = apply_ideal(k, /*can_reshape=*/true);
2171 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2172 #ifndef PRODUCT
2173 verify_step(k);
2174 #endif
2175
2176 DEBUG_ONLY(uint loop_count = 1;)
2177 while (i != nullptr) {
2178 #ifdef ASSERT
2179 if (loop_count >= K + C->live_nodes()) {
2180 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2181 }
2182 #endif
2183 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2184 // Made a change; put users of original Node on worklist
2185 add_users_to_worklist(k);
2186 // Replacing root of transform tree?
2187 if (k != i) {
2188 // Make users of old Node now use new.
2189 subsume_node(k, i);
2190 k = i;
2191 }
2192 DEBUG_ONLY(dead_loop_check(k);)
2193 // Try idealizing again
2194 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2195 C->remove_modified_node(k);
2196 i = apply_ideal(k, /*can_reshape=*/true);
2197 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2198 #ifndef PRODUCT
2199 verify_step(k);
2200 #endif
2201 DEBUG_ONLY(loop_count++;)
2202 }
2203
2204 // If brand new node, make space in type array.
2205 ensure_type_or_null(k);
2206
2207 // See what kind of values 'k' takes on at runtime
2208 const Type* t = k->Value(this);
2209 assert(t != nullptr, "value sanity");
2210
2211 // Since I just called 'Value' to compute the set of run-time values
2212 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2213 // cache Value. Later requests for the local phase->type of this Node can
2214 // use the cached Value instead of suffering with 'bottom_type'.
2215 if (type_or_null(k) != t) {
2216 #ifndef PRODUCT
2217 inc_new_values();
2218 set_progress();
2219 #endif
2220 set_type(k, t);
2221 // If k is a TypeNode, capture any more-precise type permanently into Node
2222 k->raise_bottom_type(t);
2223 // Move users of node to worklist
2224 add_users_to_worklist(k);
2225 }
2226 // If 'k' computes a constant, replace it with a constant
2227 if (t->singleton() && !k->is_Con()) {
2228 NOT_PRODUCT(set_progress();)
2229 Node* con = makecon(t); // Make a constant
2230 add_users_to_worklist(k);
2231 subsume_node(k, con); // Everybody using k now uses con
2232 return con;
2233 }
2234
2235 // Now check for Identities
2236 i = k->Identity(this); // Look for a nearby replacement
2237 if (i != k) { // Found? Return replacement!
2238 NOT_PRODUCT(set_progress();)
2239 add_users_to_worklist(k);
2240 subsume_node(k, i); // Everybody using k now uses i
2241 return i;
2242 }
2243
2244 // Global Value Numbering
2245 i = hash_find_insert(k); // Check for pre-existing node
2246 if (i && (i != k)) {
2247 // Return the pre-existing node if it isn't dead
2248 NOT_PRODUCT(set_progress();)
2249 add_users_to_worklist(k);
2250 subsume_node(k, i); // Everybody using k now uses i
2251 return i;
2252 }
2253
2254 // Return Idealized original
2255 return k;
2256 }
2257
2258 //---------------------------------saturate------------------------------------
2259 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2260 const Type* limit_type) const {
2261 return new_type->narrow(old_type);
2262 }
2263
2264 //------------------------------remove_globally_dead_node----------------------
2265 // Kill a globally dead Node. All uses are also globally dead and are
2266 // aggressively trimmed.
2267 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2268 enum DeleteProgress {
2269 PROCESS_INPUTS,
2270 PROCESS_OUTPUTS
2271 };
2272 ResourceMark rm;
2273 Node_Stack stack(32);
2274 stack.push(dead, PROCESS_INPUTS);
2275
2276 while (stack.is_nonempty()) {
2277 dead = stack.node();
2278 if (dead->Opcode() == Op_SafePoint) {
2279 dead->as_SafePoint()->disconnect_from_root(this);
2280 }
2281 uint progress_state = stack.index();
2282 assert(dead != C->root(), "killing root, eh?");
2283 assert(!dead->is_top(), "add check for top when pushing");
2284 NOT_PRODUCT( set_progress(); )
2285 if (progress_state == PROCESS_INPUTS) {
2286 // After following inputs, continue to outputs
2287 stack.set_index(PROCESS_OUTPUTS);
2288 if (!dead->is_Con()) { // Don't kill cons but uses
2289 bool recurse = false;
2290 // Remove from hash table
2291 _table.hash_delete( dead );
2292 // Smash all inputs to 'dead', isolating him completely
2293 for (uint i = 0; i < dead->req(); i++) {
2294 Node *in = dead->in(i);
2295 if (in != nullptr && in != C->top()) { // Points to something?
2296 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2297 assert((nrep > 0), "sanity");
2298 if (in->outcnt() == 0) { // Made input go dead?
2299 stack.push(in, PROCESS_INPUTS); // Recursively remove
2300 recurse = true;
2301 } else if (in->outcnt() == 1 &&
2302 in->has_special_unique_user()) {
2303 _worklist.push(in->unique_out());
2304 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2305 if (in->Opcode() == Op_Region) {
2306 _worklist.push(in);
2307 } else if (in->is_Store()) {
2308 DUIterator_Fast imax, i = in->fast_outs(imax);
2309 _worklist.push(in->fast_out(i));
2310 i++;
2311 if (in->outcnt() == 2) {
2312 _worklist.push(in->fast_out(i));
2313 i++;
2314 }
2315 assert(!(i < imax), "sanity");
2316 }
2317 } else if (dead->is_data_proj_of_pure_function(in)) {
2318 _worklist.push(in);
2319 } else {
2320 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2321 }
2322 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2323 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2324 // A Load that directly follows an InitializeNode is
2325 // going away. The Stores that follow are candidates
2326 // again to be captured by the InitializeNode.
2327 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2328 Node *n = in->fast_out(j);
2329 if (n->is_Store()) {
2330 _worklist.push(n);
2331 }
2332 }
2333 }
2334 } // if (in != nullptr && in != C->top())
2335 } // for (uint i = 0; i < dead->req(); i++)
2336 if (recurse) {
2337 continue;
2338 }
2339 } // if (!dead->is_Con())
2340 } // if (progress_state == PROCESS_INPUTS)
2341
2342 // Aggressively kill globally dead uses
2343 // (Rather than pushing all the outs at once, we push one at a time,
2344 // plus the parent to resume later, because of the indefinite number
2345 // of edge deletions per loop trip.)
2346 if (dead->outcnt() > 0) {
2347 // Recursively remove output edges
2348 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2349 } else {
2350 // Finished disconnecting all input and output edges.
2351 stack.pop();
2352 // Remove dead node from iterative worklist
2353 _worklist.remove(dead);
2354 C->remove_useless_node(dead);
2355 }
2356 } // while (stack.is_nonempty())
2357 }
2358
2359 //------------------------------subsume_node-----------------------------------
2360 // Remove users from node 'old' and add them to node 'nn'.
2361 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2362 if (old->Opcode() == Op_SafePoint) {
2363 old->as_SafePoint()->disconnect_from_root(this);
2364 }
2365 assert( old != hash_find(old), "should already been removed" );
2366 assert( old != C->top(), "cannot subsume top node");
2367 // Copy debug or profile information to the new version:
2368 C->copy_node_notes_to(nn, old);
2369 // Move users of node 'old' to node 'nn'
2370 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2371 Node* use = old->last_out(i); // for each use...
2372 // use might need re-hashing (but it won't if it's a new node)
2373 rehash_node_delayed(use);
2374 // Update use-def info as well
2375 // We remove all occurrences of old within use->in,
2376 // so as to avoid rehashing any node more than once.
2377 // The hash table probe swamps any outer loop overhead.
2378 uint num_edges = 0;
2379 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2380 if (use->in(j) == old) {
2381 use->set_req(j, nn);
2382 ++num_edges;
2383 }
2384 }
2385 i -= num_edges; // we deleted 1 or more copies of this edge
2386 }
2387
2388 // Search for instance field data PhiNodes in the same region pointing to the old
2389 // memory PhiNode and update their instance memory ids to point to the new node.
2390 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2391 Node* region = old->in(0);
2392 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2393 PhiNode* phi = region->fast_out(i)->isa_Phi();
2394 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2395 phi->set_inst_mem_id((int)nn->_idx);
2396 }
2397 }
2398 }
2399
2400 // Smash all inputs to 'old', isolating him completely
2401 Node *temp = new Node(1);
2402 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2403 remove_dead_node( old );
2404 temp->del_req(0); // Yank bogus edge
2405 if (nn != nullptr && nn->outcnt() == 0) {
2406 _worklist.push(nn);
2407 }
2408 #ifndef PRODUCT
2409 if (is_verify_def_use()) {
2410 for ( int i = 0; i < _verify_window_size; i++ ) {
2411 if ( _verify_window[i] == old )
2412 _verify_window[i] = nn;
2413 }
2414 }
2415 #endif
2416 temp->destruct(this); // reuse the _idx of this little guy
2417 }
2418
2419 //------------------------------add_users_to_worklist--------------------------
2420 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2421 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2422 worklist.push(n->fast_out(i)); // Push on worklist
2423 }
2424 }
2425
2426 // Return counted loop Phi if as a counted loop exit condition, cmp
2427 // compares the induction variable with n
2428 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2429 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2430 Node* bol = cmp->fast_out(i);
2431 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2432 Node* iff = bol->fast_out(i2);
2433 if (iff->is_BaseCountedLoopEnd()) {
2434 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2435 if (cle->limit() == n) {
2436 PhiNode* phi = cle->phi();
2437 if (phi != nullptr) {
2438 return phi;
2439 }
2440 }
2441 }
2442 }
2443 }
2444 return nullptr;
2445 }
2446
2447 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2448 add_users_to_worklist0(n, _worklist);
2449
2450 Unique_Node_List& worklist = _worklist;
2451 // Move users of node to worklist
2452 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2453 Node* use = n->fast_out(i); // Get use
2454 add_users_of_use_to_worklist(n, use, worklist);
2455 }
2456 }
2457
2458 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2459 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2460 use->is_Store() ) // Enable store/load same address
2461 add_users_to_worklist0(use, worklist);
2462
2463 // If we changed the receiver type to a call, we need to revisit
2464 // the Catch following the call. It's looking for a non-null
2465 // receiver to know when to enable the regular fall-through path
2466 // in addition to the NullPtrException path.
2467 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2468 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2469 if (p != nullptr) {
2470 add_users_to_worklist0(p, worklist);
2471 }
2472 }
2473
2474 uint use_op = use->Opcode();
2475 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2476 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2477 if (use->outcnt() > 0) {
2478 Node* bol = use->raw_out(0);
2479 if (bol->outcnt() > 0) {
2480 Node* iff = bol->raw_out(0);
2481 if (iff->outcnt() == 2) {
2482 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2483 // phi merging either 0 or 1 onto the worklist
2484 Node* ifproj0 = iff->raw_out(0);
2485 Node* ifproj1 = iff->raw_out(1);
2486 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2487 Node* region0 = ifproj0->raw_out(0);
2488 Node* region1 = ifproj1->raw_out(0);
2489 if( region0 == region1 )
2490 add_users_to_worklist0(region0, worklist);
2491 }
2492 }
2493 }
2494 }
2495 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2496 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2497 if (phi != nullptr) {
2498 // Input to the cmp of a loop exit check has changed, thus
2499 // the loop limit may have changed, which can then change the
2500 // range values of the trip-count Phi.
2501 worklist.push(phi);
2502 }
2503 }
2504 if (use_op == Op_CmpI) {
2505 Node* cmp = use;
2506 Node* in1 = cmp->in(1);
2507 Node* in2 = cmp->in(2);
2508 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2509 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2510 //
2511 // in1 in2 in1 in2
2512 // | | | |
2513 // +--- | --+ | |
2514 // | | | | |
2515 // CmpINode | CmpINode
2516 // | | |
2517 // BoolNode | BoolNode
2518 // | | OR |
2519 // IfNode | IfNode
2520 // | | |
2521 // IfProj | IfProj X
2522 // | | | |
2523 // CastIINode CastIINode
2524 //
2525 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2526 if (in1 == n) {
2527 // in1 modified -> could turn into X -> do traversal based on right pattern.
2528 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2529 Node* bol = cmp->fast_out(i2); // For each Bool
2530 if (bol->is_Bool()) {
2531 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2532 Node* iff = bol->fast_out(i3); // For each If
2533 if (iff->is_If()) {
2534 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2535 Node* if_proj = iff->fast_out(i4); // For each IfProj
2536 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2537 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2538 Node* castii = if_proj->fast_out(i5); // For each CastII
2539 if (castii->is_CastII() &&
2540 castii->as_CastII()->carry_dependency()) {
2541 worklist.push(castii);
2542 }
2543 }
2544 }
2545 }
2546 }
2547 }
2548 }
2549 } else {
2550 // Only in2 modified -> can assume X == in2 (left pattern).
2551 assert(n == in2, "only in2 modified");
2552 // Find all CastII with input in1.
2553 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2554 Node* castii = in1->fast_out(j);
2555 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2556 // Find If.
2557 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2558 Node* ifnode = castii->in(0)->in(0);
2559 // Check that if connects to the cmp
2560 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2561 worklist.push(castii);
2562 }
2563 }
2564 }
2565 }
2566 }
2567 }
2568 }
2569 }
2570
2571 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2572 // Patterns:
2573 // ConstraintCast+ -> Sub
2574 // ConstraintCast+ -> Phi
2575 // ConstraintCast+ -> Xor
2576 if (use->is_ConstraintCast()) {
2577 auto push_the_uses_to_worklist = [&](Node* n){
2578 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2579 worklist.push(n);
2580 }
2581 };
2582 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2583 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2584 }
2585 // If changed LShift inputs, check RShift users for useless sign-ext
2586 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2587 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2588 Node* u = use->fast_out(i2);
2589 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2590 worklist.push(u);
2591 }
2592 }
2593 // If changed LShift inputs, check And users for shift and mask (And) operation
2594 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2595 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2596 Node* u = use->fast_out(i2);
2597 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2598 worklist.push(u);
2599 }
2600 }
2601 }
2602 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2603 if (use_op == Op_AddI || use_op == Op_SubI) {
2604 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2605 Node* u = use->fast_out(i2);
2606 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2607 worklist.push(u);
2608 }
2609 }
2610 }
2611 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2612 if (use_op == Op_AndI || use_op == Op_AndL) {
2613 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2614 Node* u = use->fast_out(i2);
2615 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2616 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL) {
2617 worklist.push(u);
2618 }
2619 }
2620 }
2621 // Check for redundant conversion patterns:
2622 // ConvD2L->ConvL2D->ConvD2L
2623 // ConvF2I->ConvI2F->ConvF2I
2624 // ConvF2L->ConvL2F->ConvF2L
2625 // ConvI2F->ConvF2I->ConvI2F
2626 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2627 // are the only ones that are known to trigger missed optimizations otherwise
2628 if (use_op == Op_ConvL2D ||
2629 use_op == Op_ConvI2F ||
2630 use_op == Op_ConvL2F ||
2631 use_op == Op_ConvF2I) {
2632 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2633 Node* u = use->fast_out(i2);
2634 if ((use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2635 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2636 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2637 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F)) {
2638 worklist.push(u);
2639 }
2640 }
2641 }
2642 // If changed AddP inputs:
2643 // - check Stores for loop invariant, and
2644 // - if the changed input is the offset, check constant-offset AddP users for
2645 // address expression flattening.
2646 if (use_op == Op_AddP) {
2647 bool offset_changed = n == use->in(AddPNode::Offset);
2648 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2649 Node* u = use->fast_out(i2);
2650 if (u->is_Mem()) {
2651 worklist.push(u);
2652 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2653 worklist.push(u);
2654 }
2655 }
2656 }
2657 // Check for "abs(0-x)" into "abs(x)" conversion
2658 if (use->is_Sub()) {
2659 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2660 Node* u = use->fast_out(i2);
2661 if (u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2662 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI) {
2663 worklist.push(u);
2664 }
2665 }
2666 }
2667 // Check for Max/Min(A, Max/Min(B, C)) where A == B or A == C
2668 if (use->is_MinMax()) {
2669 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2670 Node* u = use->fast_out(i2);
2671 if (u->Opcode() == use->Opcode()) {
2672 worklist.push(u);
2673 }
2674 }
2675 }
2676 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2677 add_users_to_worklist0(proj, worklist);
2678 };
2679 // If changed initialization activity, check dependent Stores
2680 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2681 InitializeNode* init = use->as_Allocate()->initialization();
2682 if (init != nullptr) {
2683 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2684 }
2685 }
2686 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2687 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2688 // to guarantee the change is not missed.
2689 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2690 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2691 if (p != nullptr) {
2692 add_users_to_worklist0(p, worklist);
2693 }
2694 }
2695
2696 if (use_op == Op_Initialize) {
2697 InitializeNode* init = use->as_Initialize();
2698 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2699 }
2700 // Loading the java mirror from a Klass requires two loads and the type
2701 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2702 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2703 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2704 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2705
2706 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2707 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2708 Node* u = use->fast_out(i2);
2709 const Type* ut = u->bottom_type();
2710 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2711 if (has_load_barrier_nodes) {
2712 // Search for load barriers behind the load
2713 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2714 Node* b = u->fast_out(i3);
2715 if (bs->is_gc_barrier_node(b)) {
2716 worklist.push(b);
2717 }
2718 }
2719 }
2720 worklist.push(u);
2721 }
2722 }
2723 }
2724 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2725 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2726 if (use->outcnt() == 1) {
2727 Node* cmp = use->unique_out();
2728 worklist.push(cmp);
2729 }
2730 }
2731
2732 // From CastX2PNode::Ideal
2733 // CastX2P(AddX(x, y))
2734 // CastX2P(SubX(x, y))
2735 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2736 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2737 Node* u = use->fast_out(i2);
2738 if (u->Opcode() == Op_CastX2P) {
2739 worklist.push(u);
2740 }
2741 }
2742 }
2743
2744 /* AndNode has a special handling when one of the operands is a LShiftNode:
2745 * (LHS << s) & RHS
2746 * if RHS fits in less than s bits, the value of this expression is 0.
2747 * The difficulty is that there might be a conversion node (ConvI2L) between
2748 * the LShiftINode and the AndLNode, like so:
2749 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2750 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2751 * (see `AndIL_min_trailing_zeros`).
2752 *
2753 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2754 * is not enough: there might be no update happening there. We need to
2755 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2756 *
2757 * Moreover we can have ConstraintCasts in between. It may look like
2758 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2759 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2760 * So we must deal with that as well.
2761 */
2762 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2763 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2764 auto push_and_to_worklist = [&worklist](Node* n){
2765 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2766 worklist.push(n);
2767 }
2768 };
2769 use->visit_uses(push_and_to_worklist, is_boundary);
2770 }
2771 }
2772
2773 /**
2774 * Remove the speculative part of all types that we know of
2775 */
2776 void PhaseIterGVN::remove_speculative_types() {
2777 assert(UseTypeSpeculation, "speculation is off");
2778 for (uint i = 0; i < _types.Size(); i++) {
2779 const Type* t = _types.fast_lookup(i);
2780 if (t != nullptr) {
2781 _types.map(i, t->remove_speculative());
2782 }
2783 }
2784 _table.check_no_speculative_types();
2785 }
2786
2787 // Check if the type of a divisor of a Div or Mod node includes zero.
2788 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2789 switch (n->Opcode()) {
2790 case Op_DivI:
2791 case Op_ModI:
2792 case Op_UDivI:
2793 case Op_UModI: {
2794 // Type of divisor includes 0?
2795 if (type(n->in(2)) == Type::TOP) {
2796 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2797 return false;
2798 }
2799 const TypeInt* type_divisor = type(n->in(2))->is_int();
2800 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2801 }
2802 case Op_DivL:
2803 case Op_ModL:
2804 case Op_UDivL:
2805 case Op_UModL: {
2806 // Type of divisor includes 0?
2807 if (type(n->in(2)) == Type::TOP) {
2808 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2809 return false;
2810 }
2811 const TypeLong* type_divisor = type(n->in(2))->is_long();
2812 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2813 }
2814 }
2815 return true;
2816 }
2817
2818 //=============================================================================
2819 #ifndef PRODUCT
2820 uint PhaseCCP::_total_invokes = 0;
2821 uint PhaseCCP::_total_constants = 0;
2822 #endif
2823 //------------------------------PhaseCCP---------------------------------------
2824 // Conditional Constant Propagation, ala Wegman & Zadeck
2825 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2826 NOT_PRODUCT( clear_constants(); )
2827 assert( _worklist.size() == 0, "" );
2828 _phase = PhaseValuesType::ccp;
2829 analyze();
2830 }
2831
2832 #ifndef PRODUCT
2833 //------------------------------~PhaseCCP--------------------------------------
2834 PhaseCCP::~PhaseCCP() {
2835 inc_invokes();
2836 _total_constants += count_constants();
2837 }
2838 #endif
2839
2840
2841 #ifdef ASSERT
2842 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2843 if (tnew->meet(told) != tnew->remove_speculative()) {
2844 n->dump(1);
2845 tty->print("told = "); told->dump(); tty->cr();
2846 tty->print("tnew = "); tnew->dump(); tty->cr();
2847 fatal("Not monotonic");
2848 }
2849 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2850 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2851 }
2852 #endif //ASSERT
2853
2854 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2855 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2856 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2857 // or nodes even further down back to the worklist because their type could change as a result of the current type
2858 // change).
2859 void PhaseCCP::analyze() {
2860 // Initialize all types to TOP, optimistic analysis
2861 for (uint i = 0; i < C->unique(); i++) {
2862 _types.map(i, Type::TOP);
2863 }
2864
2865 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2866 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2867 // Compile is over. The local arena gets de-allocated at the end of its scope.
2868 ResourceArea local_arena(mtCompiler);
2869 Unique_Node_List worklist(&local_arena);
2870 Unique_Node_List worklist_revisit(&local_arena);
2871 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2872
2873 // Push root onto worklist
2874 worklist.push(C->root());
2875
2876 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2877 _root_and_safepoints.push(C->root());
2878
2879 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2880
2881 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2882 while (worklist.size() != 0) {
2883 Node* n = fetch_next_node(worklist);
2884 DEBUG_ONLY(worklist_verify.push(n);)
2885 if (needs_revisit(n)) {
2886 worklist_revisit.push(n);
2887 }
2888 if (n->is_SafePoint()) {
2889 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2890 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2891 _root_and_safepoints.push(n);
2892 }
2893 analyze_step(worklist, n);
2894 }
2895
2896 // More rounds to catch updates far in the graph.
2897 // Revisit nodes that might be able to refine their types at the end of the round.
2898 // If so, process these nodes. If there is remaining work, start another round.
2899 do {
2900 while (worklist.size() != 0) {
2901 Node* n = fetch_next_node(worklist);
2902 analyze_step(worklist, n);
2903 }
2904 for (uint t = 0; t < worklist_revisit.size(); t++) {
2905 Node* n = worklist_revisit.at(t);
2906 analyze_step(worklist, n);
2907 }
2908 } while (worklist.size() != 0);
2909
2910 DEBUG_ONLY(verify_analyze(worklist_verify);)
2911 }
2912
2913 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
2914 const Type* new_type = n->Value(this);
2915 if (new_type != type(n)) {
2916 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2917 dump_type_and_node(n, new_type);
2918 set_type(n, new_type);
2919 push_child_nodes_to_worklist(worklist, n);
2920 }
2921 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2922 // Keep track of Type nodes to kill CFG paths that use Type
2923 // nodes that become dead.
2924 _maybe_top_type_nodes.push(n);
2925 }
2926 }
2927
2928 // Some nodes can refine their types due to type change somewhere deep
2929 // in the graph. We will need to revisit them before claiming convergence.
2930 // Add nodes here if particular *Node::Value is doing deep graph traversals
2931 // not handled by PhaseCCP::push_more_uses().
2932 bool PhaseCCP::needs_revisit(Node* n) const {
2933 // LoadNode performs deep traversals. Load is not notified for changes far away.
2934 if (n->is_Load()) {
2935 return true;
2936 }
2937 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
2938 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
2939 return true;
2940 }
2941 return false;
2942 }
2943
2944 #ifdef ASSERT
2945 // For every node n on verify list, check if type(n) == n->Value()
2946 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
2947 // Therefore, we are verifying Value convergence strictly.
2948 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2949 while (worklist_verify.size()) {
2950 Node* n = worklist_verify.pop();
2951
2952 // An assert in verify_Value_for means that PhaseCCP is not at fixpoint
2953 // and that the analysis result may be unsound.
2954 // If this happens, check why the reported nodes were not processed again in CCP.
2955 // We should either make sure that these nodes are properly added back to the CCP worklist
2956 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
2957 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
2958 // them at the end of the round.
2959 verify_Value_for(n, true);
2960 }
2961 }
2962 #endif
2963
2964 // Fetch next node from worklist to be examined in this iteration.
2965 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2966 if (StressCCP) {
2967 return worklist.remove(C->random() % worklist.size());
2968 } else {
2969 return worklist.pop();
2970 }
2971 }
2972
2973 #ifndef PRODUCT
2974 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2975 if (TracePhaseCCP) {
2976 t->dump();
2977 do {
2978 tty->print("\t");
2979 } while (tty->position() < 16);
2980 n->dump();
2981 }
2982 }
2983 #endif
2984
2985 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2986 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2987 // of the new type of 'n'. Push these nodes to the worklist.
2988 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2989 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2990 Node* use = n->fast_out(i);
2991 push_if_not_bottom_type(worklist, use);
2992 push_more_uses(worklist, n, use);
2993 }
2994 }
2995
2996 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2997 if (n->bottom_type() != type(n)) {
2998 worklist.push(n);
2999 }
3000 }
3001
3002 // For some nodes, we need to propagate the type change to grandchildren or even further down.
3003 // Add them back to the worklist.
3004 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
3005 push_phis(worklist, use);
3006 push_catch(worklist, use);
3007 push_cmpu(worklist, use);
3008 push_counted_loop_phi(worklist, parent, use);
3009 push_loadp(worklist, use);
3010 push_and(worklist, parent, use);
3011 push_cast_ii(worklist, parent, use);
3012 push_opaque_zero_trip_guard(worklist, use);
3013 push_bool_with_cmpu_and_mask(worklist, use);
3014 }
3015
3016
3017 // We must recheck Phis too if use is a Region.
3018 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
3019 if (use->is_Region()) {
3020 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3021 push_if_not_bottom_type(worklist, use->fast_out(i));
3022 }
3023 }
3024 }
3025
3026 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
3027 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
3028 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
3029 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
3030 if (use->is_Call()) {
3031 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3032 Node* proj = use->fast_out(i);
3033 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
3034 Node* catch_node = proj->find_out_with(Op_Catch);
3035 if (catch_node != nullptr) {
3036 worklist.push(catch_node);
3037 }
3038 }
3039 }
3040 }
3041 }
3042
3043 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
3044 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
3045 // and get wrong types otherwise.
3046 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
3047 uint use_op = use->Opcode();
3048 if (use_op == Op_AddI || use_op == Op_SubI) {
3049 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3050 Node* cmpu = use->fast_out(i);
3051 const uint cmpu_opcode = cmpu->Opcode();
3052 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
3053 // Got a CmpU or CmpU3 which might need the new type information from node n.
3054 push_if_not_bottom_type(worklist, cmpu);
3055 }
3056 }
3057 }
3058 }
3059
3060 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3061 // (1b): "(m & x) <u (m + 1))".
3062 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3063 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3064 //
3065 // m x m 1 (%%)
3066 // \ / \ /
3067 // AndI AddI
3068 // \ /
3069 // CmpU
3070 // |
3071 // Bool
3072 //
3073 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3074 uint use_op = use->Opcode();
3075 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3076 // Not "m & x" or "m + 1"
3077 return;
3078 }
3079 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3080 Node* cmpu = use->fast_out(i);
3081 if (cmpu->Opcode() == Op_CmpU) {
3082 push_bool_matching_case1b(worklist, cmpu);
3083 }
3084 }
3085 }
3086
3087 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3088 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3089 assert(cmpu->Opcode() == Op_CmpU, "must be");
3090 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3091 Node* bol = cmpu->fast_out(i);
3092 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3093 // Not a Bool with "<u"
3094 continue;
3095 }
3096 Node* andI = cmpu->in(1);
3097 Node* addI = cmpu->in(2);
3098 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3099 // Not "m & x" and "m + 1"
3100 continue;
3101 }
3102
3103 Node* m = addI->in(1);
3104 if (m == andI->in(1) || m == andI->in(2)) {
3105 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3106 push_if_not_bottom_type(worklist, bol);
3107 }
3108 }
3109 }
3110
3111 // 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'.
3112 // Seem PhiNode::Value().
3113 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3114 uint use_op = use->Opcode();
3115 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3116 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3117 if (phi != nullptr) {
3118 worklist.push(phi);
3119 }
3120 }
3121 }
3122
3123 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3124 // See LoadNode::Value().
3125 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3126 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
3127 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
3128
3129 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3130 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3131 Node* loadp = use->fast_out(i);
3132 const Type* ut = loadp->bottom_type();
3133 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3134 if (has_load_barrier_nodes) {
3135 // Search for load barriers behind the load
3136 push_load_barrier(worklist, barrier_set, loadp);
3137 }
3138 worklist.push(loadp);
3139 }
3140 }
3141 }
3142 }
3143
3144 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3145 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3146 Node* barrier_node = use->fast_out(i);
3147 if (barrier_set->is_gc_barrier_node(barrier_node)) {
3148 worklist.push(barrier_node);
3149 }
3150 }
3151 }
3152
3153 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3154 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3155 // value changed.
3156 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3157 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3158 uint use_op = use->Opcode();
3159 if (
3160 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3161 (parent_type != nullptr && parent_type->is_con()) ||
3162 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3163 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3164
3165 auto push_and_uses_to_worklist = [&](Node* n) {
3166 uint opc = n->Opcode();
3167 if (opc == Op_AndI || opc == Op_AndL) {
3168 push_if_not_bottom_type(worklist, n);
3169 }
3170 };
3171 auto is_boundary = [](Node* n) {
3172 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3173 };
3174 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3175 }
3176 }
3177
3178 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3179 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3180 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3181 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3182 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3183 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3184 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3185 Node* other_cmp_input = use->in(1);
3186 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3187 Node* cast_ii = other_cmp_input->fast_out(i);
3188 if (cast_ii->is_CastII()) {
3189 push_if_not_bottom_type(worklist, cast_ii);
3190 }
3191 }
3192 }
3193 }
3194
3195 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3196 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3197 push_if_not_bottom_type(worklist, use->unique_out());
3198 }
3199 }
3200
3201 //------------------------------do_transform-----------------------------------
3202 // Top level driver for the recursive transformer
3203 void PhaseCCP::do_transform() {
3204 // Correct leaves of new-space Nodes; they point to old-space.
3205 C->set_root( transform(C->root())->as_Root() );
3206 assert( C->top(), "missing TOP node" );
3207 assert( C->root(), "missing root" );
3208 }
3209
3210 //------------------------------transform--------------------------------------
3211 // Given a Node in old-space, clone him into new-space.
3212 // Convert any of his old-space children into new-space children.
3213 Node *PhaseCCP::transform( Node *n ) {
3214 assert(n->is_Root(), "traversal must start at root");
3215 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3216
3217 ResourceMark rm;
3218 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3219 Node_List node_map;
3220 // Pre-allocate to avoid frequent realloc
3221 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3222 // track all visited nodes, so that we can remove the complement
3223 Unique_Node_List useful;
3224
3225 if (KillPathsReachableByDeadTypeNode) {
3226 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3227 Node* type_node = _maybe_top_type_nodes.at(i);
3228 if (type(type_node) == Type::TOP) {
3229 ResourceMark rm;
3230 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3231 }
3232 }
3233 } else {
3234 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3235 }
3236
3237 // Initialize the traversal.
3238 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3239 // 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
3240 // 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
3241 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3242 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3243 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3244 // and all safepoints.
3245 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3246 Node* nn = _root_and_safepoints.at(i);
3247 Node* new_node = node_map[nn->_idx];
3248 assert(new_node == nullptr, "");
3249 new_node = transform_once(nn); // Check for constant
3250 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3251 transform_stack.push(new_node); // Process children of cloned node
3252 useful.push(new_node);
3253 }
3254
3255 while (transform_stack.is_nonempty()) {
3256 Node* clone = transform_stack.pop();
3257 uint cnt = clone->req();
3258 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3259 Node *input = clone->in(i);
3260 if( input != nullptr ) { // Ignore nulls
3261 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3262 if( new_input == nullptr ) {
3263 new_input = transform_once(input); // Check for constant
3264 node_map.map( input->_idx, new_input );// Flag as having been cloned
3265 transform_stack.push(new_input); // Process children of cloned node
3266 useful.push(new_input);
3267 }
3268 assert( new_input == clone->in(i), "insanity check");
3269 }
3270 }
3271 }
3272
3273 // The above transformation might lead to subgraphs becoming unreachable from the
3274 // bottom while still being reachable from the top. As a result, nodes in that
3275 // subgraph are not transformed and their bottom types are not updated, leading to
3276 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3277 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3278 // because their address type is inconsistent. Therefore, we aggressively remove
3279 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3280 // to remove them anyway.
3281 if (C->cached_top_node()) {
3282 useful.push(C->cached_top_node());
3283 }
3284 C->update_dead_node_list(useful);
3285 remove_useless_nodes(useful.member_set());
3286 _worklist.remove_useless_nodes(useful.member_set());
3287 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3288
3289 Node* new_root = node_map[n->_idx];
3290 assert(new_root->is_Root(), "transformed root node must be a root node");
3291 return new_root;
3292 }
3293
3294 //------------------------------transform_once---------------------------------
3295 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3296 Node *PhaseCCP::transform_once( Node *n ) {
3297 const Type *t = type(n);
3298 // Constant? Use constant Node instead
3299 if( t->singleton() ) {
3300 Node *nn = n; // Default is to return the original constant
3301 if( t == Type::TOP ) {
3302 // cache my top node on the Compile instance
3303 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3304 C->set_cached_top_node(ConNode::make(Type::TOP));
3305 set_type(C->top(), Type::TOP);
3306 }
3307 nn = C->top();
3308 }
3309 if( !n->is_Con() ) {
3310 if( t != Type::TOP ) {
3311 nn = makecon(t); // ConNode::make(t);
3312 NOT_PRODUCT( inc_constants(); )
3313 } else if( n->is_Region() ) { // Unreachable region
3314 // Note: nn == C->top()
3315 n->set_req(0, nullptr); // Cut selfreference
3316 bool progress = true;
3317 uint max = n->outcnt();
3318 DUIterator i;
3319 while (progress) {
3320 progress = false;
3321 // Eagerly remove dead phis to avoid phis copies creation.
3322 for (i = n->outs(); n->has_out(i); i++) {
3323 Node* m = n->out(i);
3324 if (m->is_Phi()) {
3325 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3326 replace_node(m, nn);
3327 if (max != n->outcnt()) {
3328 progress = true;
3329 i = n->refresh_out_pos(i);
3330 max = n->outcnt();
3331 }
3332 }
3333 }
3334 }
3335 }
3336 replace_node(n,nn); // Update DefUse edges for new constant
3337 }
3338 return nn;
3339 }
3340
3341 // If x is a TypeNode, capture any more-precise type permanently into Node
3342 if (t != n->bottom_type()) {
3343 hash_delete(n); // changing bottom type may force a rehash
3344 n->raise_bottom_type(t);
3345 _worklist.push(n); // n re-enters the hash table via the worklist
3346 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3347 }
3348
3349 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3350 switch( n->Opcode() ) {
3351 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3352 case Op_CallDynamicJava:
3353 case Op_FastLock: // Revisit FastLocks for lock coarsening
3354 case Op_If:
3355 case Op_CountedLoopEnd:
3356 case Op_Region:
3357 case Op_Loop:
3358 case Op_CountedLoop:
3359 case Op_Conv2B:
3360 case Op_Opaque1:
3361 _worklist.push(n);
3362 break;
3363 default:
3364 break;
3365 }
3366
3367 return n;
3368 }
3369
3370 //---------------------------------saturate------------------------------------
3371 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3372 const Type* limit_type) const {
3373 const Type* wide_type = new_type->widen(old_type, limit_type);
3374 if (wide_type != new_type) { // did we widen?
3375 // If so, we may have widened beyond the limit type. Clip it back down.
3376 new_type = wide_type->filter(limit_type);
3377 }
3378 return new_type;
3379 }
3380
3381 //------------------------------print_statistics-------------------------------
3382 #ifndef PRODUCT
3383 void PhaseCCP::print_statistics() {
3384 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3385 }
3386 #endif
3387
3388
3389 //=============================================================================
3390 #ifndef PRODUCT
3391 uint PhasePeephole::_total_peepholes = 0;
3392 #endif
3393 //------------------------------PhasePeephole----------------------------------
3394 // Conditional Constant Propagation, ala Wegman & Zadeck
3395 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3396 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3397 NOT_PRODUCT( clear_peepholes(); )
3398 }
3399
3400 #ifndef PRODUCT
3401 //------------------------------~PhasePeephole---------------------------------
3402 PhasePeephole::~PhasePeephole() {
3403 _total_peepholes += count_peepholes();
3404 }
3405 #endif
3406
3407 //------------------------------transform--------------------------------------
3408 Node *PhasePeephole::transform( Node *n ) {
3409 ShouldNotCallThis();
3410 return nullptr;
3411 }
3412
3413 //------------------------------do_transform-----------------------------------
3414 void PhasePeephole::do_transform() {
3415 bool method_name_not_printed = true;
3416
3417 // Examine each basic block
3418 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3419 Block* block = _cfg.get_block(block_number);
3420 bool block_not_printed = true;
3421
3422 for (bool progress = true; progress;) {
3423 progress = false;
3424 // block->end_idx() not valid after PhaseRegAlloc
3425 uint end_index = block->number_of_nodes();
3426 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3427 Node *n = block->get_node(instruction_index);
3428 if( n->is_Mach() ) {
3429 MachNode *m = n->as_Mach();
3430 // check for peephole opportunities
3431 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3432 if( result != -1 ) {
3433 #ifndef PRODUCT
3434 if( PrintOptoPeephole ) {
3435 // Print method, first time only
3436 if( C->method() && method_name_not_printed ) {
3437 C->method()->print_short_name(); tty->cr();
3438 method_name_not_printed = false;
3439 }
3440 // Print this block
3441 if( Verbose && block_not_printed) {
3442 tty->print_cr("in block");
3443 block->dump();
3444 block_not_printed = false;
3445 }
3446 // Print the peephole number
3447 tty->print_cr("peephole number: %d", result);
3448 }
3449 inc_peepholes();
3450 #endif
3451 // Set progress, start again
3452 progress = true;
3453 break;
3454 }
3455 }
3456 }
3457 }
3458 }
3459 }
3460
3461 //------------------------------print_statistics-------------------------------
3462 #ifndef PRODUCT
3463 void PhasePeephole::print_statistics() {
3464 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3465 }
3466 #endif
3467
3468
3469 //=============================================================================
3470 //------------------------------set_req_X--------------------------------------
3471 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3472 assert( is_not_dead(n), "can not use dead node");
3473 #ifdef ASSERT
3474 if (igvn->hash_find(this) == this) {
3475 tty->print_cr("Need to remove from hash before changing edges");
3476 this->dump(1);
3477 tty->print_cr("Set at i = %d", i);
3478 n->dump();
3479 assert(false, "Need to remove from hash before changing edges");
3480 }
3481 #endif
3482 Node *old = in(i);
3483 set_req(i, n);
3484
3485 // old goes dead?
3486 if( old ) {
3487 switch (old->outcnt()) {
3488 case 0:
3489 // Put into the worklist to kill later. We do not kill it now because the
3490 // recursive kill will delete the current node (this) if dead-loop exists
3491 if (!old->is_top())
3492 igvn->_worklist.push( old );
3493 break;
3494 case 1:
3495 if( old->is_Store() || old->has_special_unique_user() )
3496 igvn->add_users_to_worklist( old );
3497 break;
3498 case 2:
3499 if( old->is_Store() )
3500 igvn->add_users_to_worklist( old );
3501 if( old->Opcode() == Op_Region )
3502 igvn->_worklist.push(old);
3503 break;
3504 case 3:
3505 if( old->Opcode() == Op_Region ) {
3506 igvn->_worklist.push(old);
3507 igvn->add_users_to_worklist( old );
3508 }
3509 break;
3510 default:
3511 break;
3512 }
3513
3514 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3515 }
3516 }
3517
3518 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3519 PhaseIterGVN* igvn = gvn->is_IterGVN();
3520 if (igvn == nullptr) {
3521 set_req(i, n);
3522 return;
3523 }
3524 set_req_X(i, n, igvn);
3525 }
3526
3527 //-------------------------------replace_by-----------------------------------
3528 // Using def-use info, replace one node for another. Follow the def-use info
3529 // to all users of the OLD node. Then make all uses point to the NEW node.
3530 void Node::replace_by(Node *new_node) {
3531 assert(!is_top(), "top node has no DU info");
3532 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3533 Node* use = last_out(i);
3534 uint uses_found = 0;
3535 for (uint j = 0; j < use->len(); j++) {
3536 if (use->in(j) == this) {
3537 if (j < use->req())
3538 use->set_req(j, new_node);
3539 else use->set_prec(j, new_node);
3540 uses_found++;
3541 }
3542 }
3543 i -= uses_found; // we deleted 1 or more copies of this edge
3544 }
3545 }
3546
3547 //=============================================================================
3548 //-----------------------------------------------------------------------------
3549 void Type_Array::grow( uint i ) {
3550 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3551 if( !_max ) {
3552 _max = 1;
3553 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3554 _types[0] = nullptr;
3555 }
3556 uint old = _max;
3557 _max = next_power_of_2(i);
3558 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3559 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3560 }
3561
3562 //------------------------------dump-------------------------------------------
3563 #ifndef PRODUCT
3564 void Type_Array::dump() const {
3565 uint max = Size();
3566 for( uint i = 0; i < max; i++ ) {
3567 if( _types[i] != nullptr ) {
3568 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3569 }
3570 }
3571 }
3572 #endif