1 /*
2 * Copyright (c) 1997, 2018, 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 "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.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
43 //=============================================================================
44 #define NODE_HASH_MINIMUM_SIZE 255
45 //------------------------------NodeHash---------------------------------------
46 NodeHash::NodeHash(uint est_max_size) :
47 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
48 _a(Thread::current()->resource_area()),
49 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ), // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
50 _inserts(0), _insert_limit( insert_limit() )
51 #ifndef PRODUCT
52 ,_look_probes(0), _lookup_hits(0), _lookup_misses(0),
53 _delete_probes(0), _delete_hits(0), _delete_misses(0),
54 _total_insert_probes(0), _total_inserts(0),
55 _insert_probes(0), _grows(0)
56 #endif
57 {
58 // _sentinel must be in the current node space
59 _sentinel = new ProjNode(NULL, TypeFunc::Control);
60 memset(_table,0,sizeof(Node*)*_max);
61 }
62
63 //------------------------------NodeHash---------------------------------------
64 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
65 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
66 _a(arena),
67 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ),
68 _inserts(0), _insert_limit( insert_limit() )
69 #ifndef PRODUCT
70 ,_look_probes(0), _lookup_hits(0), _lookup_misses(0),
71 _delete_probes(0), _delete_hits(0), _delete_misses(0),
72 _total_insert_probes(0), _total_inserts(0),
73 _insert_probes(0), _grows(0)
74 #endif
75 {
76 // _sentinel must be in the current node space
77 _sentinel = new ProjNode(NULL, TypeFunc::Control);
78 memset(_table,0,sizeof(Node*)*_max);
79 }
80
81 //------------------------------NodeHash---------------------------------------
82 NodeHash::NodeHash(NodeHash *nh) {
83 debug_only(_table = (Node**)badAddress); // interact correctly w/ operator=
84 // just copy in all the fields
85 *this = *nh;
86 // nh->_sentinel must be in the current node space
87 }
88
89 void NodeHash::replace_with(NodeHash *nh) {
90 debug_only(_table = (Node**)badAddress); // interact correctly w/ operator=
91 // just copy in all the fields
92 *this = *nh;
93 // nh->_sentinel must be in the current node space
94 }
95
96 //------------------------------hash_find--------------------------------------
97 // Find in hash table
98 Node *NodeHash::hash_find( const Node *n ) {
99 // ((Node*)n)->set_hash( n->hash() );
100 uint hash = n->hash();
101 if (hash == Node::NO_HASH) {
102 NOT_PRODUCT( _lookup_misses++ );
103 return NULL;
104 }
105 uint key = hash & (_max-1);
106 uint stride = key | 0x01;
107 NOT_PRODUCT( _look_probes++ );
108 Node *k = _table[key]; // Get hashed value
109 if( !k ) { // ?Miss?
110 NOT_PRODUCT( _lookup_misses++ );
111 return NULL; // Miss!
112 }
113
114 int op = n->Opcode();
115 uint req = n->req();
116 while( 1 ) { // While probing hash table
117 if( k->req() == req && // Same count of inputs
118 k->Opcode() == op ) { // Same Opcode
119 for( uint i=0; i<req; i++ )
120 if( n->in(i)!=k->in(i)) // Different inputs?
121 goto collision; // "goto" is a speed hack...
122 if( n->cmp(*k) ) { // Check for any special bits
123 NOT_PRODUCT( _lookup_hits++ );
124 return k; // Hit!
125 }
126 }
127 collision:
128 NOT_PRODUCT( _look_probes++ );
129 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
130 k = _table[key]; // Get hashed value
131 if( !k ) { // ?Miss?
132 NOT_PRODUCT( _lookup_misses++ );
133 return NULL; // Miss!
134 }
135 }
136 ShouldNotReachHere();
137 return NULL;
138 }
139
140 //------------------------------hash_find_insert-------------------------------
141 // Find in hash table, insert if not already present
142 // Used to preserve unique entries in hash table
143 Node *NodeHash::hash_find_insert( Node *n ) {
144 // n->set_hash( );
145 uint hash = n->hash();
146 if (hash == Node::NO_HASH) {
147 NOT_PRODUCT( _lookup_misses++ );
148 return NULL;
149 }
150 uint key = hash & (_max-1);
151 uint stride = key | 0x01; // stride must be relatively prime to table siz
152 uint first_sentinel = 0; // replace a sentinel if seen.
153 NOT_PRODUCT( _look_probes++ );
154 Node *k = _table[key]; // Get hashed value
155 if( !k ) { // ?Miss?
156 NOT_PRODUCT( _lookup_misses++ );
157 _table[key] = n; // Insert into table!
158 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
159 check_grow(); // Grow table if insert hit limit
160 return NULL; // Miss!
161 }
162 else if( k == _sentinel ) {
163 first_sentinel = key; // Can insert here
164 }
165
166 int op = n->Opcode();
167 uint req = n->req();
168 while( 1 ) { // While probing hash table
169 if( k->req() == req && // Same count of inputs
170 k->Opcode() == op ) { // Same Opcode
171 for( uint i=0; i<req; i++ )
172 if( n->in(i)!=k->in(i)) // Different inputs?
173 goto collision; // "goto" is a speed hack...
174 if( n->cmp(*k) ) { // Check for any special bits
175 NOT_PRODUCT( _lookup_hits++ );
176 return k; // Hit!
177 }
178 }
179 collision:
180 NOT_PRODUCT( _look_probes++ );
181 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
182 k = _table[key]; // Get hashed value
183 if( !k ) { // ?Miss?
184 NOT_PRODUCT( _lookup_misses++ );
185 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
186 _table[key] = n; // Insert into table!
187 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
188 check_grow(); // Grow table if insert hit limit
189 return NULL; // Miss!
190 }
191 else if( first_sentinel == 0 && k == _sentinel ) {
192 first_sentinel = key; // Can insert here
193 }
194
195 }
196 ShouldNotReachHere();
197 return NULL;
198 }
199
200 //------------------------------hash_insert------------------------------------
201 // Insert into hash table
202 void NodeHash::hash_insert( Node *n ) {
203 // // "conflict" comments -- print nodes that conflict
204 // bool conflict = false;
205 // n->set_hash();
206 uint hash = n->hash();
207 if (hash == Node::NO_HASH) {
208 return;
209 }
210 check_grow();
211 uint key = hash & (_max-1);
212 uint stride = key | 0x01;
213
214 while( 1 ) { // While probing hash table
215 NOT_PRODUCT( _insert_probes++ );
216 Node *k = _table[key]; // Get hashed value
217 if( !k || (k == _sentinel) ) break; // Found a slot
218 assert( k != n, "already inserted" );
219 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
220 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
221 }
222 _table[key] = n; // Insert into table!
223 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
224 // if( conflict ) { n->dump(); }
225 }
226
227 //------------------------------hash_delete------------------------------------
228 // Replace in hash table with sentinel
229 bool NodeHash::hash_delete( const Node *n ) {
230 Node *k;
231 uint hash = n->hash();
232 if (hash == Node::NO_HASH) {
233 NOT_PRODUCT( _delete_misses++ );
234 return false;
235 }
236 uint key = hash & (_max-1);
237 uint stride = key | 0x01;
238 debug_only( uint counter = 0; );
239 for( ; /* (k != NULL) && (k != _sentinel) */; ) {
240 debug_only( counter++ );
241 NOT_PRODUCT( _delete_probes++ );
242 k = _table[key]; // Get hashed value
243 if( !k ) { // Miss?
244 NOT_PRODUCT( _delete_misses++ );
245 #ifdef ASSERT
246 if( VerifyOpto ) {
247 for( uint i=0; i < _max; i++ )
248 assert( _table[i] != n, "changed edges with rehashing" );
249 }
250 #endif
251 return false; // Miss! Not in chain
252 }
253 else if( n == k ) {
254 NOT_PRODUCT( _delete_hits++ );
255 _table[key] = _sentinel; // Hit! Label as deleted entry
256 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
257 return true;
258 }
259 else {
260 // collision: move through table with prime offset
261 key = (key + stride/*7*/) & (_max-1);
262 assert( counter <= _insert_limit, "Cycle in hash-table");
263 }
264 }
265 ShouldNotReachHere();
266 return false;
267 }
268
269 //------------------------------round_up---------------------------------------
270 // Round up to nearest power of 2
271 uint NodeHash::round_up( uint x ) {
272 x += (x>>2); // Add 25% slop
273 if( x <16 ) return 16; // Small stuff
274 uint i=16;
275 while( i < x ) i <<= 1; // Double to fit
276 return i; // Return hash table size
277 }
278
279 //------------------------------grow-------------------------------------------
280 // Grow _table to next power of 2 and insert old entries
281 void NodeHash::grow() {
282 // Record old state
283 uint old_max = _max;
284 Node **old_table = _table;
285 // Construct new table with twice the space
286 #ifndef PRODUCT
287 _grows++;
288 _total_inserts += _inserts;
289 _total_insert_probes += _insert_probes;
290 _insert_probes = 0;
291 #endif
292 _inserts = 0;
293 _max = _max << 1;
294 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
295 memset(_table,0,sizeof(Node*)*_max);
296 _insert_limit = insert_limit();
297 // Insert old entries into the new table
298 for( uint i = 0; i < old_max; i++ ) {
299 Node *m = *old_table++;
300 if( !m || m == _sentinel ) continue;
301 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
302 hash_insert(m);
303 }
304 }
305
306 //------------------------------clear------------------------------------------
307 // Clear all entries in _table to NULL but keep storage
308 void NodeHash::clear() {
309 #ifdef ASSERT
310 // Unlock all nodes upon removal from table.
311 for (uint i = 0; i < _max; i++) {
312 Node* n = _table[i];
313 if (!n || n == _sentinel) continue;
314 n->exit_hash_lock();
315 }
316 #endif
317
318 memset( _table, 0, _max * sizeof(Node*) );
319 }
320
321 //-----------------------remove_useless_nodes----------------------------------
322 // Remove useless nodes from value table,
323 // implementation does not depend on hash function
324 void NodeHash::remove_useless_nodes(VectorSet &useful) {
325
326 // Dead nodes in the hash table inherited from GVN should not replace
327 // existing nodes, remove dead nodes.
328 uint max = size();
329 Node *sentinel_node = sentinel();
330 for( uint i = 0; i < max; ++i ) {
331 Node *n = at(i);
332 if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
333 debug_only(n->exit_hash_lock()); // Unlock the node when removed
334 _table[i] = sentinel_node; // Replace with placeholder
335 }
336 }
337 }
338
339
340 void NodeHash::check_no_speculative_types() {
341 #ifdef ASSERT
342 uint max = size();
343 Node *sentinel_node = sentinel();
344 for (uint i = 0; i < max; ++i) {
345 Node *n = at(i);
346 if(n != NULL && n != sentinel_node && n->is_Type() && n->outcnt() > 0) {
347 TypeNode* tn = n->as_Type();
348 const Type* t = tn->type();
349 const Type* t_no_spec = t->remove_speculative();
350 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
351 }
352 }
353 #endif
354 }
355
356 #ifndef PRODUCT
357 //------------------------------dump-------------------------------------------
358 // Dump statistics for the hash table
359 void NodeHash::dump() {
360 _total_inserts += _inserts;
361 _total_insert_probes += _insert_probes;
362 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
363 if (WizardMode) {
364 for (uint i=0; i<_max; i++) {
365 if (_table[i])
366 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
367 }
368 }
369 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
370 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
371 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
372 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
373 // sentinels increase lookup cost, but not insert cost
374 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
375 assert( _inserts+(_inserts>>3) < _max, "table too full" );
376 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
377 }
378 }
379
380 Node *NodeHash::find_index(uint idx) { // For debugging
381 // Find an entry by its index value
382 for( uint i = 0; i < _max; i++ ) {
383 Node *m = _table[i];
384 if( !m || m == _sentinel ) continue;
385 if( m->_idx == (uint)idx ) return m;
386 }
387 return NULL;
388 }
389 #endif
390
391 #ifdef ASSERT
392 NodeHash::~NodeHash() {
393 // Unlock all nodes upon destruction of table.
394 if (_table != (Node**)badAddress) clear();
395 }
396
397 void NodeHash::operator=(const NodeHash& nh) {
398 // Unlock all nodes upon replacement of table.
399 if (&nh == this) return;
400 if (_table != (Node**)badAddress) clear();
401 memcpy((void*)this, (void*)&nh, sizeof(*this));
402 // Do not increment hash_lock counts again.
403 // Instead, be sure we never again use the source table.
404 ((NodeHash*)&nh)->_table = (Node**)badAddress;
405 }
406
407
408 #endif
409
410
411 //=============================================================================
412 //------------------------------PhaseRemoveUseless-----------------------------
413 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
414 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN *gvn, Unique_Node_List *worklist, PhaseNumber phase_num) : Phase(phase_num),
415 _useful(Thread::current()->resource_area()) {
416
417 // Implementation requires 'UseLoopSafepoints == true' and an edge from root
418 // to each SafePointNode at a backward branch. Inserted in add_safepoint().
419 if( !UseLoopSafepoints || !OptoRemoveUseless ) return;
420
421 // Identify nodes that are reachable from below, useful.
422 C->identify_useful_nodes(_useful);
423 // Update dead node list
424 C->update_dead_node_list(_useful);
425
426 // Remove all useless nodes from PhaseValues' recorded types
427 // Must be done before disconnecting nodes to preserve hash-table-invariant
428 gvn->remove_useless_nodes(_useful.member_set());
429
430 // Remove all useless nodes from future worklist
431 worklist->remove_useless_nodes(_useful.member_set());
432
433 // Disconnect 'useless' nodes that are adjacent to useful nodes
434 C->remove_useless_nodes(_useful);
435
436 // Remove edges from "root" to each SafePoint at a backward branch.
437 // They were inserted during parsing (see add_safepoint()) to make infinite
438 // loops without calls or exceptions visible to root, i.e., useful.
439 Node *root = C->root();
440 if( root != NULL ) {
441 for( uint i = root->req(); i < root->len(); ++i ) {
442 Node *n = root->in(i);
443 if( n != NULL && n->is_SafePoint() ) {
444 root->rm_prec(i);
445 --i;
446 }
447 }
448 }
449 }
450
451 //=============================================================================
452 //------------------------------PhaseRenumberLive------------------------------
453 // First, remove useless nodes (equivalent to identifying live nodes).
454 // Then, renumber live nodes.
455 //
456 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
457 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
458 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
459 // value in the range [0, x).
460 //
461 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
462 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
463 //
464 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
465 // (1) The worklist is used by the PhaseIterGVN phase to identify nodes that must be
466 // processed. A new worklist (with the updated node IDs) is returned in 'new_worklist'.
467 // (2) Type information (the field PhaseGVN::_types) maps type information to each
468 // node ID. The mapping is updated to use the new node IDs as well. Updated type
469 // information is returned in PhaseGVN::_types.
470 //
471 // The PhaseRenumberLive phase does not preserve the order of elements in the worklist.
472 //
473 // Other data structures used by the compiler are not updated. The hash table for value
474 // numbering (the field PhaseGVN::_table) is not updated because computing the hash
475 // values is not based on node IDs. The field PhaseGVN::_nodes is not updated either
476 // because it is empty wherever PhaseRenumberLive is used.
477 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
478 Unique_Node_List* worklist, Unique_Node_List* new_worklist,
479 PhaseNumber phase_num) :
480 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live) {
481
482 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
483 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
484 assert(gvn->nodes_size() == 0, "GVN must not contain any nodes at this point");
485
486 uint old_unique_count = C->unique();
487 uint live_node_count = C->live_nodes();
488 uint worklist_size = worklist->size();
489
490 // Storage for the updated type information.
491 Type_Array new_type_array(C->comp_arena());
492
493 // Iterate over the set of live nodes.
494 uint current_idx = 0; // The current new node ID. Incremented after every assignment.
495 for (uint i = 0; i < _useful.size(); i++) {
496 Node* n = _useful.at(i);
497 // Sanity check that fails if we ever decide to execute this phase after EA
498 assert(!n->is_Phi() || n->as_Phi()->inst_mem_id() == -1, "should not be linked to data Phi");
499 const Type* type = gvn->type_or_null(n);
500 new_type_array.map(current_idx, type);
501
502 bool in_worklist = false;
503 if (worklist->member(n)) {
504 in_worklist = true;
505 }
506
507 n->set_idx(current_idx); // Update node ID.
508
509 if (in_worklist) {
510 new_worklist->push(n);
511 }
512
513 current_idx++;
514 }
515
516 assert(worklist_size == new_worklist->size(), "the new worklist must have the same size as the original worklist");
517 assert(live_node_count == current_idx, "all live nodes must be processed");
518
519 // Replace the compiler's type information with the updated type information.
520 gvn->replace_types(new_type_array);
521
522 // Update the unique node count of the compilation to the number of currently live nodes.
523 C->set_unique(live_node_count);
524
525 // Set the dead node count to 0 and reset dead node list.
526 C->reset_dead_node_list();
527 }
528
529
530 //=============================================================================
531 //------------------------------PhaseTransform---------------------------------
532 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
533 _arena(Thread::current()->resource_area()),
534 _nodes(_arena),
535 _types(_arena)
536 {
537 init_con_caches();
538 #ifndef PRODUCT
539 clear_progress();
540 clear_transforms();
541 set_allow_progress(true);
542 #endif
543 // Force allocation for currently existing nodes
544 _types.map(C->unique(), NULL);
545 }
546
547 //------------------------------PhaseTransform---------------------------------
548 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
549 _arena(arena),
550 _nodes(arena),
551 _types(arena)
552 {
553 init_con_caches();
554 #ifndef PRODUCT
555 clear_progress();
556 clear_transforms();
557 set_allow_progress(true);
558 #endif
559 // Force allocation for currently existing nodes
560 _types.map(C->unique(), NULL);
561 }
562
563 //------------------------------PhaseTransform---------------------------------
564 // Initialize with previously generated type information
565 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
566 _arena(pt->_arena),
567 _nodes(pt->_nodes),
568 _types(pt->_types)
569 {
570 init_con_caches();
571 #ifndef PRODUCT
572 clear_progress();
573 clear_transforms();
574 set_allow_progress(true);
575 #endif
576 }
577
578 void PhaseTransform::init_con_caches() {
579 memset(_icons,0,sizeof(_icons));
580 memset(_lcons,0,sizeof(_lcons));
581 memset(_zcons,0,sizeof(_zcons));
582 }
583
584
585 //--------------------------------find_int_type--------------------------------
586 const TypeInt* PhaseTransform::find_int_type(Node* n) {
587 if (n == NULL) return NULL;
588 // Call type_or_null(n) to determine node's type since we might be in
589 // parse phase and call n->Value() may return wrong type.
590 // (For example, a phi node at the beginning of loop parsing is not ready.)
591 const Type* t = type_or_null(n);
592 if (t == NULL) return NULL;
593 return t->isa_int();
594 }
595
596
597 //-------------------------------find_long_type--------------------------------
598 const TypeLong* PhaseTransform::find_long_type(Node* n) {
599 if (n == NULL) return NULL;
600 // (See comment above on type_or_null.)
601 const Type* t = type_or_null(n);
602 if (t == NULL) return NULL;
603 return t->isa_long();
604 }
605
606
607 #ifndef PRODUCT
608 void PhaseTransform::dump_old2new_map() const {
609 _nodes.dump();
610 }
611
612 void PhaseTransform::dump_new( uint nidx ) const {
613 for( uint i=0; i<_nodes.Size(); i++ )
614 if( _nodes[i] && _nodes[i]->_idx == nidx ) {
615 _nodes[i]->dump();
616 tty->cr();
617 tty->print_cr("Old index= %d",i);
618 return;
619 }
620 tty->print_cr("Node %d not found in the new indices", nidx);
621 }
622
623 //------------------------------dump_types-------------------------------------
624 void PhaseTransform::dump_types( ) const {
625 _types.dump();
626 }
627
628 //------------------------------dump_nodes_and_types---------------------------
629 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) {
630 VectorSet visited(Thread::current()->resource_area());
631 dump_nodes_and_types_recur( root, depth, only_ctrl, visited );
632 }
633
634 //------------------------------dump_nodes_and_types_recur---------------------
635 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
636 if( !n ) return;
637 if( depth == 0 ) return;
638 if( visited.test_set(n->_idx) ) return;
639 for( uint i=0; i<n->len(); i++ ) {
640 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
641 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
642 }
643 n->dump();
644 if (type_or_null(n) != NULL) {
645 tty->print(" "); type(n)->dump(); tty->cr();
646 }
647 }
648
649 #endif
650
651
652 //=============================================================================
653 //------------------------------PhaseValues------------------------------------
654 // Set minimum table size to "255"
655 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) {
656 NOT_PRODUCT( clear_new_values(); )
657 }
658
659 //------------------------------PhaseValues------------------------------------
660 // Set minimum table size to "255"
661 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ),
662 _table(&ptv->_table) {
663 NOT_PRODUCT( clear_new_values(); )
664 }
665
666 //------------------------------PhaseValues------------------------------------
667 // Used by +VerifyOpto. Clear out hash table but copy _types array.
668 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ),
669 _table(ptv->arena(),ptv->_table.size()) {
670 NOT_PRODUCT( clear_new_values(); )
671 }
672
673 //------------------------------~PhaseValues-----------------------------------
674 #ifndef PRODUCT
675 PhaseValues::~PhaseValues() {
676 _table.dump();
677
678 // Statistics for value progress and efficiency
679 if( PrintCompilation && Verbose && WizardMode ) {
680 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
681 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
682 if( made_transforms() != 0 ) {
683 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
684 } else {
685 tty->cr();
686 }
687 }
688 }
689 #endif
690
691 //------------------------------makecon----------------------------------------
692 ConNode* PhaseTransform::makecon(const Type *t) {
693 assert(t->singleton(), "must be a constant");
694 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
695 switch (t->base()) { // fast paths
696 case Type::Half:
697 case Type::Top: return (ConNode*) C->top();
698 case Type::Int: return intcon( t->is_int()->get_con() );
699 case Type::Long: return longcon( t->is_long()->get_con() );
700 default: break;
701 }
702 if (t->is_zero_type())
703 return zerocon(t->basic_type());
704 return uncached_makecon(t);
705 }
706
707 //--------------------------uncached_makecon-----------------------------------
708 // Make an idealized constant - one of ConINode, ConPNode, etc.
709 ConNode* PhaseValues::uncached_makecon(const Type *t) {
710 assert(t->singleton(), "must be a constant");
711 ConNode* x = ConNode::make(t);
712 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
713 if (k == NULL) {
714 set_type(x, t); // Missed, provide type mapping
715 GrowableArray<Node_Notes*>* nna = C->node_note_array();
716 if (nna != NULL) {
717 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
718 loc->clear(); // do not put debug info on constants
719 }
720 } else {
721 x->destruct(); // Hit, destroy duplicate constant
722 x = k; // use existing constant
723 }
724 return x;
725 }
726
727 //------------------------------intcon-----------------------------------------
728 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
729 ConINode* PhaseTransform::intcon(jint i) {
730 // Small integer? Check cache! Check that cached node is not dead
731 if (i >= _icon_min && i <= _icon_max) {
732 ConINode* icon = _icons[i-_icon_min];
733 if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
734 return icon;
735 }
736 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
737 assert(icon->is_Con(), "");
738 if (i >= _icon_min && i <= _icon_max)
739 _icons[i-_icon_min] = icon; // Cache small integers
740 return icon;
741 }
742
743 //------------------------------longcon----------------------------------------
744 // Fast long constant.
745 ConLNode* PhaseTransform::longcon(jlong l) {
746 // Small integer? Check cache! Check that cached node is not dead
747 if (l >= _lcon_min && l <= _lcon_max) {
748 ConLNode* lcon = _lcons[l-_lcon_min];
749 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
750 return lcon;
751 }
752 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
753 assert(lcon->is_Con(), "");
754 if (l >= _lcon_min && l <= _lcon_max)
755 _lcons[l-_lcon_min] = lcon; // Cache small integers
756 return lcon;
757 }
758
759 //------------------------------zerocon-----------------------------------------
760 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
761 ConNode* PhaseTransform::zerocon(BasicType bt) {
762 assert((uint)bt <= _zcon_max, "domain check");
763 ConNode* zcon = _zcons[bt];
764 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
765 return zcon;
766 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
767 _zcons[bt] = zcon;
768 return zcon;
769 }
770
771
772
773 //=============================================================================
774 //------------------------------transform--------------------------------------
775 // Return a node which computes the same function as this node, but in a
776 // faster or cheaper fashion.
777 Node *PhaseGVN::transform( Node *n ) {
778 return transform_no_reclaim(n);
779 }
780
781 //------------------------------transform--------------------------------------
782 // Return a node which computes the same function as this node, but
783 // in a faster or cheaper fashion.
784 Node *PhaseGVN::transform_no_reclaim( Node *n ) {
785 NOT_PRODUCT( set_transforms(); )
786
787 // Apply the Ideal call in a loop until it no longer applies
788 Node *k = n;
789 NOT_PRODUCT( uint loop_count = 0; )
790 while( 1 ) {
791 Node *i = k->Ideal(this, /*can_reshape=*/false);
792 if( !i ) break;
793 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
794 k = i;
795 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform");
796 }
797 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } )
798
799
800 // If brand new node, make space in type array.
801 ensure_type_or_null(k);
802
803 // Since I just called 'Value' to compute the set of run-time values
804 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
805 // cache Value. Later requests for the local phase->type of this Node can
806 // use the cached Value instead of suffering with 'bottom_type'.
807 const Type *t = k->Value(this); // Get runtime Value set
808 assert(t != NULL, "value sanity");
809 if (type_or_null(k) != t) {
810 #ifndef PRODUCT
811 // Do not count initial visit to node as a transformation
812 if (type_or_null(k) == NULL) {
813 inc_new_values();
814 set_progress();
815 }
816 #endif
817 set_type(k, t);
818 // If k is a TypeNode, capture any more-precise type permanently into Node
819 k->raise_bottom_type(t);
820 }
821
822 if( t->singleton() && !k->is_Con() ) {
823 NOT_PRODUCT( set_progress(); )
824 return makecon(t); // Turn into a constant
825 }
826
827 // Now check for Identities
828 Node *i = k->Identity(this); // Look for a nearby replacement
829 if( i != k ) { // Found? Return replacement!
830 NOT_PRODUCT( set_progress(); )
831 return i;
832 }
833
834 // Global Value Numbering
835 i = hash_find_insert(k); // Insert if new
836 if( i && (i != k) ) {
837 // Return the pre-existing node
838 NOT_PRODUCT( set_progress(); )
839 return i;
840 }
841
842 // Return Idealized original
843 return k;
844 }
845
846 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
847 if (d->is_top() || n->is_top()) {
848 return false;
849 }
850 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
851 int i = 0;
852 while (d != n) {
853 n = IfNode::up_one_dom(n, linear_only);
854 i++;
855 if (n == NULL || i >= 10) {
856 return false;
857 }
858 }
859 return true;
860 }
861
862 #ifdef ASSERT
863 //------------------------------dead_loop_check--------------------------------
864 // Check for a simple dead loop when a data node references itself directly
865 // or through an other data node excluding cons and phis.
866 void PhaseGVN::dead_loop_check( Node *n ) {
867 // Phi may reference itself in a loop
868 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
869 // Do 2 levels check and only data inputs.
870 bool no_dead_loop = true;
871 uint cnt = n->req();
872 for (uint i = 1; i < cnt && no_dead_loop; i++) {
873 Node *in = n->in(i);
874 if (in == n) {
875 no_dead_loop = false;
876 } else if (in != NULL && !in->is_dead_loop_safe()) {
877 uint icnt = in->req();
878 for (uint j = 1; j < icnt && no_dead_loop; j++) {
879 if (in->in(j) == n || in->in(j) == in)
880 no_dead_loop = false;
881 }
882 }
883 }
884 if (!no_dead_loop) n->dump(3);
885 assert(no_dead_loop, "dead loop detected");
886 }
887 }
888 #endif
889
890 //=============================================================================
891 //------------------------------PhaseIterGVN-----------------------------------
892 // Initialize hash table to fresh and clean for +VerifyOpto
893 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ),
894 _stack(C->live_nodes() >> 1),
895 _delay_transform(false) {
896 }
897
898 //------------------------------PhaseIterGVN-----------------------------------
899 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
900 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn),
901 _worklist( igvn->_worklist ),
902 _stack( igvn->_stack ),
903 _delay_transform(igvn->_delay_transform)
904 {
905 }
906
907 //------------------------------PhaseIterGVN-----------------------------------
908 // Initialize with previous PhaseGVN info from Parser
909 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn),
910 _worklist(*C->for_igvn()),
911 // TODO: Before incremental inlining it was allocated only once and it was fine. Now that
912 // the constructor is used in incremental inlining, this consumes too much memory:
913 // _stack(C->live_nodes() >> 1),
914 // So, as a band-aid, we replace this by:
915 _stack(C->comp_arena(), 32),
916 _delay_transform(false)
917 {
918 uint max;
919
920 // Dead nodes in the hash table inherited from GVN were not treated as
921 // roots during def-use info creation; hence they represent an invisible
922 // use. Clear them out.
923 max = _table.size();
924 for( uint i = 0; i < max; ++i ) {
925 Node *n = _table.at(i);
926 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
927 if( n->is_top() ) continue;
928 assert( false, "Parse::remove_useless_nodes missed this node");
929 hash_delete(n);
930 }
931 }
932
933 // Any Phis or Regions on the worklist probably had uses that could not
934 // make more progress because the uses were made while the Phis and Regions
935 // were in half-built states. Put all uses of Phis and Regions on worklist.
936 max = _worklist.size();
937 for( uint j = 0; j < max; j++ ) {
938 Node *n = _worklist.at(j);
939 uint uop = n->Opcode();
940 if( uop == Op_Phi || uop == Op_Region ||
941 n->is_Type() ||
942 n->is_Mem() )
943 add_users_to_worklist(n);
944 }
945
946 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
947 bs->add_users_to_worklist(&_worklist);
948 }
949
950 /**
951 * Initialize worklist for each node.
952 */
953 void PhaseIterGVN::init_worklist(Node* first) {
954 Unique_Node_List to_process;
955 to_process.push(first);
956
957 while (to_process.size() > 0) {
958 Node* n = to_process.pop();
959 if (!_worklist.member(n)) {
960 _worklist.push(n);
961
962 uint cnt = n->req();
963 for(uint i = 0; i < cnt; i++) {
964 Node* m = n->in(i);
965 if (m != NULL) {
966 to_process.push(m);
967 }
968 }
969 }
970 }
971 }
972
973 #ifndef PRODUCT
974 void PhaseIterGVN::verify_step(Node* n) {
975 if (VerifyIterativeGVN) {
976 _verify_window[_verify_counter % _verify_window_size] = n;
977 ++_verify_counter;
978 ResourceMark rm;
979 ResourceArea* area = Thread::current()->resource_area();
980 VectorSet old_space(area), new_space(area);
981 if (C->unique() < 1000 ||
982 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
983 ++_verify_full_passes;
984 Node::verify_recur(C->root(), -1, old_space, new_space);
985 }
986 const int verify_depth = 4;
987 for ( int i = 0; i < _verify_window_size; i++ ) {
988 Node* n = _verify_window[i];
989 if ( n == NULL ) continue;
990 if( n->in(0) == NodeSentinel ) { // xform_idom
991 _verify_window[i] = n->in(1);
992 --i; continue;
993 }
994 // Typical fanout is 1-2, so this call visits about 6 nodes.
995 Node::verify_recur(n, verify_depth, old_space, new_space);
996 }
997 }
998 }
999
1000 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
1001 if (TraceIterativeGVN) {
1002 uint wlsize = _worklist.size();
1003 const Type* newtype = type_or_null(n);
1004 if (nn != n) {
1005 // print old node
1006 tty->print("< ");
1007 if (oldtype != newtype && oldtype != NULL) {
1008 oldtype->dump();
1009 }
1010 do { tty->print("\t"); } while (tty->position() < 16);
1011 tty->print("<");
1012 n->dump();
1013 }
1014 if (oldtype != newtype || nn != n) {
1015 // print new node and/or new type
1016 if (oldtype == NULL) {
1017 tty->print("* ");
1018 } else if (nn != n) {
1019 tty->print("> ");
1020 } else {
1021 tty->print("= ");
1022 }
1023 if (newtype == NULL) {
1024 tty->print("null");
1025 } else {
1026 newtype->dump();
1027 }
1028 do { tty->print("\t"); } while (tty->position() < 16);
1029 nn->dump();
1030 }
1031 if (Verbose && wlsize < _worklist.size()) {
1032 tty->print(" Push {");
1033 while (wlsize != _worklist.size()) {
1034 Node* pushed = _worklist.at(wlsize++);
1035 tty->print(" %d", pushed->_idx);
1036 }
1037 tty->print_cr(" }");
1038 }
1039 if (nn != n) {
1040 // ignore n, it might be subsumed
1041 verify_step((Node*) NULL);
1042 }
1043 }
1044 }
1045
1046 void PhaseIterGVN::init_verifyPhaseIterGVN() {
1047 _verify_counter = 0;
1048 _verify_full_passes = 0;
1049 for (int i = 0; i < _verify_window_size; i++) {
1050 _verify_window[i] = NULL;
1051 }
1052 #ifdef ASSERT
1053 // Verify that all modified nodes are on _worklist
1054 Unique_Node_List* modified_list = C->modified_nodes();
1055 while (modified_list != NULL && modified_list->size()) {
1056 Node* n = modified_list->pop();
1057 if (n->outcnt() != 0 && !n->is_Con() && !_worklist.member(n)) {
1058 n->dump();
1059 assert(false, "modified node is not on IGVN._worklist");
1060 }
1061 }
1062 #endif
1063 }
1064
1065 void PhaseIterGVN::verify_PhaseIterGVN() {
1066 #ifdef ASSERT
1067 // Verify nodes with changed inputs.
1068 Unique_Node_List* modified_list = C->modified_nodes();
1069 while (modified_list != NULL && modified_list->size()) {
1070 Node* n = modified_list->pop();
1071 if (n->outcnt() != 0 && !n->is_Con()) { // skip dead and Con nodes
1072 n->dump();
1073 assert(false, "modified node was not processed by IGVN.transform_old()");
1074 }
1075 }
1076 #endif
1077
1078 C->verify_graph_edges();
1079 if( VerifyOpto && allow_progress() ) {
1080 // Must turn off allow_progress to enable assert and break recursion
1081 C->root()->verify();
1082 { // Check if any progress was missed using IterGVN
1083 // Def-Use info enables transformations not attempted in wash-pass
1084 // e.g. Region/Phi cleanup, ...
1085 // Null-check elision -- may not have reached fixpoint
1086 // do not propagate to dominated nodes
1087 ResourceMark rm;
1088 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean!
1089 // Fill worklist completely
1090 igvn2.init_worklist(C->root());
1091
1092 igvn2.set_allow_progress(false);
1093 igvn2.optimize();
1094 igvn2.set_allow_progress(true);
1095 }
1096 }
1097 if (VerifyIterativeGVN && PrintOpto) {
1098 if (_verify_counter == _verify_full_passes) {
1099 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1100 (int) _verify_full_passes);
1101 } else {
1102 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1103 (int) _verify_counter, (int) _verify_full_passes);
1104 }
1105 }
1106
1107 #ifdef ASSERT
1108 while (modified_list->size()) {
1109 Node* n = modified_list->pop();
1110 n->dump();
1111 assert(false, "VerifyIterativeGVN: new modified node was added");
1112 }
1113 #endif
1114 }
1115 #endif /* PRODUCT */
1116
1117 #ifdef ASSERT
1118 /**
1119 * Dumps information that can help to debug the problem. A debug
1120 * build fails with an assert.
1121 */
1122 void PhaseIterGVN::dump_infinite_loop_info(Node* n) {
1123 n->dump(4);
1124 _worklist.dump();
1125 assert(false, "infinite loop in PhaseIterGVN::optimize");
1126 }
1127
1128 /**
1129 * Prints out information about IGVN if the 'verbose' option is used.
1130 */
1131 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1132 if (TraceIterativeGVN && Verbose) {
1133 tty->print(" Pop ");
1134 n->dump();
1135 if ((num_processed % 100) == 0) {
1136 _worklist.print_set();
1137 }
1138 }
1139 }
1140 #endif /* ASSERT */
1141
1142 void PhaseIterGVN::optimize() {
1143 DEBUG_ONLY(uint num_processed = 0;)
1144 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1145
1146 uint loop_count = 0;
1147 // Pull from worklist and transform the node. If the node has changed,
1148 // update edge info and put uses on worklist.
1149 while(_worklist.size()) {
1150 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) {
1151 return;
1152 }
1153 Node* n = _worklist.pop();
1154 if (++loop_count >= K * C->live_nodes()) {
1155 DEBUG_ONLY(dump_infinite_loop_info(n);)
1156 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1157 return;
1158 }
1159 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1160 if (n->outcnt() != 0) {
1161 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1162 // Do the transformation
1163 Node* nn = transform_old(n);
1164 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1165 } else if (!n->is_top()) {
1166 remove_dead_node(n);
1167 }
1168 }
1169 NOT_PRODUCT(verify_PhaseIterGVN();)
1170 }
1171
1172
1173 /**
1174 * Register a new node with the optimizer. Update the types array, the def-use
1175 * info. Put on worklist.
1176 */
1177 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1178 set_type_bottom(n);
1179 _worklist.push(n);
1180 if (orig != NULL) C->copy_node_notes_to(n, orig);
1181 return n;
1182 }
1183
1184 //------------------------------transform--------------------------------------
1185 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
1186 Node *PhaseIterGVN::transform( Node *n ) {
1187 if (_delay_transform) {
1188 // Register the node but don't optimize for now
1189 register_new_node_with_optimizer(n);
1190 return n;
1191 }
1192
1193 // If brand new node, make space in type array, and give it a type.
1194 ensure_type_or_null(n);
1195 if (type_or_null(n) == NULL) {
1196 set_type_bottom(n);
1197 }
1198
1199 return transform_old(n);
1200 }
1201
1202 Node *PhaseIterGVN::transform_old(Node* n) {
1203 DEBUG_ONLY(uint loop_count = 0;);
1204 NOT_PRODUCT(set_transforms());
1205
1206 // Remove 'n' from hash table in case it gets modified
1207 _table.hash_delete(n);
1208 if (VerifyIterativeGVN) {
1209 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
1210 }
1211
1212 // Apply the Ideal call in a loop until it no longer applies
1213 Node* k = n;
1214 DEBUG_ONLY(dead_loop_check(k);)
1215 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1216 C->remove_modified_node(k);
1217 Node* i = k->Ideal(this, /*can_reshape=*/true);
1218 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1219 #ifndef PRODUCT
1220 verify_step(k);
1221 if (i && VerifyOpto ) {
1222 if (!allow_progress()) {
1223 if (i->is_Add() && (i->outcnt() == 1)) {
1224 // Switched input to left side because this is the only use
1225 } else if (i->is_If() && (i->in(0) == NULL)) {
1226 // This IF is dead because it is dominated by an equivalent IF When
1227 // dominating if changed, info is not propagated sparsely to 'this'
1228 // Propagating this info further will spuriously identify other
1229 // progress.
1230 return i;
1231 } else
1232 set_progress();
1233 } else {
1234 set_progress();
1235 }
1236 }
1237 #endif
1238
1239 while (i != NULL) {
1240 #ifdef ASSERT
1241 if (loop_count >= K) {
1242 dump_infinite_loop_info(i);
1243 }
1244 loop_count++;
1245 #endif
1246 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1247 // Made a change; put users of original Node on worklist
1248 add_users_to_worklist(k);
1249 // Replacing root of transform tree?
1250 if (k != i) {
1251 // Make users of old Node now use new.
1252 subsume_node(k, i);
1253 k = i;
1254 }
1255 DEBUG_ONLY(dead_loop_check(k);)
1256 // Try idealizing again
1257 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1258 C->remove_modified_node(k);
1259 i = k->Ideal(this, /*can_reshape=*/true);
1260 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
1261 #ifndef PRODUCT
1262 verify_step(k);
1263 if (i && VerifyOpto) {
1264 set_progress();
1265 }
1266 #endif
1267 }
1268
1269 // If brand new node, make space in type array.
1270 ensure_type_or_null(k);
1271
1272 // See what kind of values 'k' takes on at runtime
1273 const Type* t = k->Value(this);
1274 assert(t != NULL, "value sanity");
1275
1276 // Since I just called 'Value' to compute the set of run-time values
1277 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1278 // cache Value. Later requests for the local phase->type of this Node can
1279 // use the cached Value instead of suffering with 'bottom_type'.
1280 if (type_or_null(k) != t) {
1281 #ifndef PRODUCT
1282 inc_new_values();
1283 set_progress();
1284 #endif
1285 set_type(k, t);
1286 // If k is a TypeNode, capture any more-precise type permanently into Node
1287 k->raise_bottom_type(t);
1288 // Move users of node to worklist
1289 add_users_to_worklist(k);
1290 }
1291 // If 'k' computes a constant, replace it with a constant
1292 if (t->singleton() && !k->is_Con()) {
1293 NOT_PRODUCT(set_progress();)
1294 Node* con = makecon(t); // Make a constant
1295 add_users_to_worklist(k);
1296 subsume_node(k, con); // Everybody using k now uses con
1297 return con;
1298 }
1299
1300 // Now check for Identities
1301 i = k->Identity(this); // Look for a nearby replacement
1302 if (i != k) { // Found? Return replacement!
1303 NOT_PRODUCT(set_progress();)
1304 add_users_to_worklist(k);
1305 subsume_node(k, i); // Everybody using k now uses i
1306 return i;
1307 }
1308
1309 // Global Value Numbering
1310 i = hash_find_insert(k); // Check for pre-existing node
1311 if (i && (i != k)) {
1312 // Return the pre-existing node if it isn't dead
1313 NOT_PRODUCT(set_progress();)
1314 add_users_to_worklist(k);
1315 subsume_node(k, i); // Everybody using k now uses i
1316 return i;
1317 }
1318
1319 // Return Idealized original
1320 return k;
1321 }
1322
1323 //---------------------------------saturate------------------------------------
1324 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1325 const Type* limit_type) const {
1326 return new_type->narrow(old_type);
1327 }
1328
1329 //------------------------------remove_globally_dead_node----------------------
1330 // Kill a globally dead Node. All uses are also globally dead and are
1331 // aggressively trimmed.
1332 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1333 enum DeleteProgress {
1334 PROCESS_INPUTS,
1335 PROCESS_OUTPUTS
1336 };
1337 assert(_stack.is_empty(), "not empty");
1338 _stack.push(dead, PROCESS_INPUTS);
1339
1340 while (_stack.is_nonempty()) {
1341 dead = _stack.node();
1342 uint progress_state = _stack.index();
1343 assert(dead != C->root(), "killing root, eh?");
1344 assert(!dead->is_top(), "add check for top when pushing");
1345 NOT_PRODUCT( set_progress(); )
1346 if (progress_state == PROCESS_INPUTS) {
1347 // After following inputs, continue to outputs
1348 _stack.set_index(PROCESS_OUTPUTS);
1349 if (!dead->is_Con()) { // Don't kill cons but uses
1350 bool recurse = false;
1351 // Remove from hash table
1352 _table.hash_delete( dead );
1353 // Smash all inputs to 'dead', isolating him completely
1354 for (uint i = 0; i < dead->req(); i++) {
1355 Node *in = dead->in(i);
1356 if (in != NULL && in != C->top()) { // Points to something?
1357 int nrep = dead->replace_edge(in, NULL); // Kill edges
1358 assert((nrep > 0), "sanity");
1359 if (in->outcnt() == 0) { // Made input go dead?
1360 _stack.push(in, PROCESS_INPUTS); // Recursively remove
1361 recurse = true;
1362 } else if (in->outcnt() == 1 &&
1363 in->has_special_unique_user()) {
1364 _worklist.push(in->unique_out());
1365 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1366 if (in->Opcode() == Op_Region) {
1367 _worklist.push(in);
1368 } else if (in->is_Store()) {
1369 DUIterator_Fast imax, i = in->fast_outs(imax);
1370 _worklist.push(in->fast_out(i));
1371 i++;
1372 if (in->outcnt() == 2) {
1373 _worklist.push(in->fast_out(i));
1374 i++;
1375 }
1376 assert(!(i < imax), "sanity");
1377 }
1378 } else {
1379 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(_worklist, in);
1380 }
1381 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1382 in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1383 // A Load that directly follows an InitializeNode is
1384 // going away. The Stores that follow are candidates
1385 // again to be captured by the InitializeNode.
1386 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1387 Node *n = in->fast_out(j);
1388 if (n->is_Store()) {
1389 _worklist.push(n);
1390 }
1391 }
1392 }
1393 } // if (in != NULL && in != C->top())
1394 } // for (uint i = 0; i < dead->req(); i++)
1395 if (recurse) {
1396 continue;
1397 }
1398 } // if (!dead->is_Con())
1399 } // if (progress_state == PROCESS_INPUTS)
1400
1401 // Aggressively kill globally dead uses
1402 // (Rather than pushing all the outs at once, we push one at a time,
1403 // plus the parent to resume later, because of the indefinite number
1404 // of edge deletions per loop trip.)
1405 if (dead->outcnt() > 0) {
1406 // Recursively remove output edges
1407 _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1408 } else {
1409 // Finished disconnecting all input and output edges.
1410 _stack.pop();
1411 // Remove dead node from iterative worklist
1412 _worklist.remove(dead);
1413 C->remove_modified_node(dead);
1414 // Constant node that has no out-edges and has only one in-edge from
1415 // root is usually dead. However, sometimes reshaping walk makes
1416 // it reachable by adding use edges. So, we will NOT count Con nodes
1417 // as dead to be conservative about the dead node count at any
1418 // given time.
1419 if (!dead->is_Con()) {
1420 C->record_dead_node(dead->_idx);
1421 }
1422 if (dead->is_macro()) {
1423 C->remove_macro_node(dead);
1424 }
1425 if (dead->is_expensive()) {
1426 C->remove_expensive_node(dead);
1427 }
1428 CastIINode* cast = dead->isa_CastII();
1429 if (cast != NULL && cast->has_range_check()) {
1430 C->remove_range_check_cast(cast);
1431 }
1432 if (dead->Opcode() == Op_Opaque4) {
1433 C->remove_opaque4_node(dead);
1434 }
1435 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1436 bs->unregister_potential_barrier_node(dead);
1437 }
1438 } // while (_stack.is_nonempty())
1439 }
1440
1441 //------------------------------subsume_node-----------------------------------
1442 // Remove users from node 'old' and add them to node 'nn'.
1443 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1444 assert( old != hash_find(old), "should already been removed" );
1445 assert( old != C->top(), "cannot subsume top node");
1446 // Copy debug or profile information to the new version:
1447 C->copy_node_notes_to(nn, old);
1448 // Move users of node 'old' to node 'nn'
1449 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1450 Node* use = old->last_out(i); // for each use...
1451 // use might need re-hashing (but it won't if it's a new node)
1452 rehash_node_delayed(use);
1453 // Update use-def info as well
1454 // We remove all occurrences of old within use->in,
1455 // so as to avoid rehashing any node more than once.
1456 // The hash table probe swamps any outer loop overhead.
1457 uint num_edges = 0;
1458 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1459 if (use->in(j) == old) {
1460 use->set_req(j, nn);
1461 ++num_edges;
1462 }
1463 }
1464 i -= num_edges; // we deleted 1 or more copies of this edge
1465 }
1466
1467 // Search for instance field data PhiNodes in the same region pointing to the old
1468 // memory PhiNode and update their instance memory ids to point to the new node.
1469 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != NULL) {
1470 Node* region = old->in(0);
1471 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1472 PhiNode* phi = region->fast_out(i)->isa_Phi();
1473 if (phi != NULL && phi->inst_mem_id() == (int)old->_idx) {
1474 phi->set_inst_mem_id((int)nn->_idx);
1475 }
1476 }
1477 }
1478
1479 // Smash all inputs to 'old', isolating him completely
1480 Node *temp = new Node(1);
1481 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
1482 remove_dead_node( old );
1483 temp->del_req(0); // Yank bogus edge
1484 #ifndef PRODUCT
1485 if( VerifyIterativeGVN ) {
1486 for ( int i = 0; i < _verify_window_size; i++ ) {
1487 if ( _verify_window[i] == old )
1488 _verify_window[i] = nn;
1489 }
1490 }
1491 #endif
1492 _worklist.remove(temp); // this can be necessary
1493 temp->destruct(); // reuse the _idx of this little guy
1494 }
1495
1496 //------------------------------add_users_to_worklist--------------------------
1497 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1498 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1499 _worklist.push(n->fast_out(i)); // Push on worklist
1500 }
1501 }
1502
1503 // Return counted loop Phi if as a counted loop exit condition, cmp
1504 // compares the the induction variable with n
1505 static PhiNode* countedloop_phi_from_cmp(CmpINode* cmp, Node* n) {
1506 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
1507 Node* bol = cmp->fast_out(i);
1508 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
1509 Node* iff = bol->fast_out(i2);
1510 if (iff->is_CountedLoopEnd()) {
1511 CountedLoopEndNode* cle = iff->as_CountedLoopEnd();
1512 if (cle->limit() == n) {
1513 PhiNode* phi = cle->phi();
1514 if (phi != NULL) {
1515 return phi;
1516 }
1517 }
1518 }
1519 }
1520 }
1521 return NULL;
1522 }
1523
1524 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1525 add_users_to_worklist0(n);
1526
1527 // Move users of node to worklist
1528 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1529 Node* use = n->fast_out(i); // Get use
1530
1531 if( use->is_Multi() || // Multi-definer? Push projs on worklist
1532 use->is_Store() ) // Enable store/load same address
1533 add_users_to_worklist0(use);
1534
1535 // If we changed the receiver type to a call, we need to revisit
1536 // the Catch following the call. It's looking for a non-NULL
1537 // receiver to know when to enable the regular fall-through path
1538 // in addition to the NullPtrException path.
1539 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1540 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
1541 if (p != NULL) {
1542 add_users_to_worklist0(p);
1543 }
1544 }
1545
1546 uint use_op = use->Opcode();
1547 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
1548 add_users_to_worklist(use); // Put Bool on worklist
1549 if (use->outcnt() > 0) {
1550 Node* bol = use->raw_out(0);
1551 if (bol->outcnt() > 0) {
1552 Node* iff = bol->raw_out(0);
1553 if (iff->outcnt() == 2) {
1554 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1555 // phi merging either 0 or 1 onto the worklist
1556 Node* ifproj0 = iff->raw_out(0);
1557 Node* ifproj1 = iff->raw_out(1);
1558 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1559 Node* region0 = ifproj0->raw_out(0);
1560 Node* region1 = ifproj1->raw_out(0);
1561 if( region0 == region1 )
1562 add_users_to_worklist0(region0);
1563 }
1564 }
1565 }
1566 }
1567 if (use_op == Op_CmpI) {
1568 Node* phi = countedloop_phi_from_cmp((CmpINode*)use, n);
1569 if (phi != NULL) {
1570 // If an opaque node feeds into the limit condition of a
1571 // CountedLoop, we need to process the Phi node for the
1572 // induction variable when the opaque node is removed:
1573 // the range of values taken by the Phi is now known and
1574 // so its type is also known.
1575 _worklist.push(phi);
1576 }
1577 Node* in1 = use->in(1);
1578 for (uint i = 0; i < in1->outcnt(); i++) {
1579 if (in1->raw_out(i)->Opcode() == Op_CastII) {
1580 Node* castii = in1->raw_out(i);
1581 if (castii->in(0) != NULL && castii->in(0)->in(0) != NULL && castii->in(0)->in(0)->is_If()) {
1582 Node* ifnode = castii->in(0)->in(0);
1583 if (ifnode->in(1) != NULL && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == use) {
1584 // Reprocess a CastII node that may depend on an
1585 // opaque node value when the opaque node is
1586 // removed. In case it carries a dependency we can do
1587 // a better job of computing its type.
1588 _worklist.push(castii);
1589 }
1590 }
1591 }
1592 }
1593 }
1594 }
1595
1596 // If changed Cast input, check Phi users for simple cycles
1597 if (use->is_ConstraintCast()) {
1598 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1599 Node* u = use->fast_out(i2);
1600 if (u->is_Phi())
1601 _worklist.push(u);
1602 }
1603 }
1604 // If changed LShift inputs, check RShift users for useless sign-ext
1605 if( use_op == Op_LShiftI ) {
1606 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1607 Node* u = use->fast_out(i2);
1608 if (u->Opcode() == Op_RShiftI)
1609 _worklist.push(u);
1610 }
1611 }
1612 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1613 if (use_op == Op_AddI || use_op == Op_SubI) {
1614 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1615 Node* u = use->fast_out(i2);
1616 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1617 _worklist.push(u);
1618 }
1619 }
1620 }
1621 // If changed AddP inputs, check Stores for loop invariant
1622 if( use_op == Op_AddP ) {
1623 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1624 Node* u = use->fast_out(i2);
1625 if (u->is_Mem())
1626 _worklist.push(u);
1627 }
1628 }
1629 // If changed initialization activity, check dependent Stores
1630 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1631 InitializeNode* init = use->as_Allocate()->initialization();
1632 if (init != NULL) {
1633 Node* imem = init->proj_out_or_null(TypeFunc::Memory);
1634 if (imem != NULL) add_users_to_worklist0(imem);
1635 }
1636 }
1637 if (use_op == Op_Initialize) {
1638 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
1639 if (imem != NULL) add_users_to_worklist0(imem);
1640 }
1641 // Loading the java mirror from a Klass requires two loads and the type
1642 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1643 // If the code pattern requires a barrier for
1644 // mirror = ((OopHandle)mirror)->resolve();
1645 // this won't match.
1646 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
1647 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1648 Node* u = use->fast_out(i2);
1649 const Type* ut = u->bottom_type();
1650 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
1651 _worklist.push(u);
1652 }
1653 }
1654 }
1655 }
1656 }
1657
1658 /**
1659 * Remove the speculative part of all types that we know of
1660 */
1661 void PhaseIterGVN::remove_speculative_types() {
1662 assert(UseTypeSpeculation, "speculation is off");
1663 for (uint i = 0; i < _types.Size(); i++) {
1664 const Type* t = _types.fast_lookup(i);
1665 if (t != NULL) {
1666 _types.map(i, t->remove_speculative());
1667 }
1668 }
1669 _table.check_no_speculative_types();
1670 }
1671
1672 //=============================================================================
1673 #ifndef PRODUCT
1674 uint PhaseCCP::_total_invokes = 0;
1675 uint PhaseCCP::_total_constants = 0;
1676 #endif
1677 //------------------------------PhaseCCP---------------------------------------
1678 // Conditional Constant Propagation, ala Wegman & Zadeck
1679 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1680 NOT_PRODUCT( clear_constants(); )
1681 assert( _worklist.size() == 0, "" );
1682 // Clear out _nodes from IterGVN. Must be clear to transform call.
1683 _nodes.clear(); // Clear out from IterGVN
1684 analyze();
1685 }
1686
1687 #ifndef PRODUCT
1688 //------------------------------~PhaseCCP--------------------------------------
1689 PhaseCCP::~PhaseCCP() {
1690 inc_invokes();
1691 _total_constants += count_constants();
1692 }
1693 #endif
1694
1695
1696 #ifdef ASSERT
1697 static bool ccp_type_widens(const Type* t, const Type* t0) {
1698 assert(t->meet(t0) == t, "Not monotonic");
1699 switch (t->base() == t0->base() ? t->base() : Type::Top) {
1700 case Type::Int:
1701 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1702 break;
1703 case Type::Long:
1704 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1705 break;
1706 default:
1707 break;
1708 }
1709 return true;
1710 }
1711 #endif //ASSERT
1712
1713 //------------------------------analyze----------------------------------------
1714 void PhaseCCP::analyze() {
1715 // Initialize all types to TOP, optimistic analysis
1716 for (int i = C->unique() - 1; i >= 0; i--) {
1717 _types.map(i,Type::TOP);
1718 }
1719
1720 // Push root onto worklist
1721 Unique_Node_List worklist;
1722 worklist.push(C->root());
1723
1724 // Pull from worklist; compute new value; push changes out.
1725 // This loop is the meat of CCP.
1726 while( worklist.size() ) {
1727 Node *n = worklist.pop();
1728 const Type *t = n->Value(this);
1729 if (t != type(n)) {
1730 assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1731 #ifndef PRODUCT
1732 if( TracePhaseCCP ) {
1733 t->dump();
1734 do { tty->print("\t"); } while (tty->position() < 16);
1735 n->dump();
1736 }
1737 #endif
1738 set_type(n, t);
1739 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1740 Node* m = n->fast_out(i); // Get user
1741 if (m->is_Region()) { // New path to Region? Must recheck Phis too
1742 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1743 Node* p = m->fast_out(i2); // Propagate changes to uses
1744 if (p->bottom_type() != type(p)) { // If not already bottomed out
1745 worklist.push(p); // Propagate change to user
1746 }
1747 }
1748 }
1749 // If we changed the receiver type to a call, we need to revisit
1750 // the Catch following the call. It's looking for a non-NULL
1751 // receiver to know when to enable the regular fall-through path
1752 // in addition to the NullPtrException path
1753 if (m->is_Call()) {
1754 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1755 Node* p = m->fast_out(i2); // Propagate changes to uses
1756 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1) {
1757 worklist.push(p->unique_out());
1758 }
1759 }
1760 }
1761 if (m->bottom_type() != type(m)) { // If not already bottomed out
1762 worklist.push(m); // Propagate change to user
1763 }
1764
1765 // CmpU nodes can get their type information from two nodes up in the
1766 // graph (instead of from the nodes immediately above). Make sure they
1767 // are added to the worklist if nodes they depend on are updated, since
1768 // they could be missed and get wrong types otherwise.
1769 uint m_op = m->Opcode();
1770 if (m_op == Op_AddI || m_op == Op_SubI) {
1771 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1772 Node* p = m->fast_out(i2); // Propagate changes to uses
1773 if (p->Opcode() == Op_CmpU) {
1774 // Got a CmpU which might need the new type information from node n.
1775 if(p->bottom_type() != type(p)) { // If not already bottomed out
1776 worklist.push(p); // Propagate change to user
1777 }
1778 }
1779 }
1780 }
1781 // If n is used in a counted loop exit condition then the type
1782 // of the counted loop's Phi depends on the type of n. See
1783 // PhiNode::Value().
1784 if (m_op == Op_CmpI) {
1785 PhiNode* phi = countedloop_phi_from_cmp((CmpINode*)m, n);
1786 if (phi != NULL) {
1787 worklist.push(phi);
1788 }
1789 }
1790 // Loading the java mirror from a Klass requires two loads and the type
1791 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
1792 // If the code pattern requires a barrier for
1793 // mirror = ((OopHandle)mirror)->resolve();
1794 // this won't match.
1795 if (m_op == Op_LoadP && m->bottom_type()->isa_rawptr()) {
1796 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1797 Node* u = m->fast_out(i2);
1798 const Type* ut = u->bottom_type();
1799 if (u->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(u)) {
1800 worklist.push(u);
1801 }
1802 }
1803 }
1804 }
1805 }
1806 }
1807 }
1808
1809 //------------------------------do_transform-----------------------------------
1810 // Top level driver for the recursive transformer
1811 void PhaseCCP::do_transform() {
1812 // Correct leaves of new-space Nodes; they point to old-space.
1813 C->set_root( transform(C->root())->as_Root() );
1814 assert( C->top(), "missing TOP node" );
1815 assert( C->root(), "missing root" );
1816 }
1817
1818 //------------------------------transform--------------------------------------
1819 // Given a Node in old-space, clone him into new-space.
1820 // Convert any of his old-space children into new-space children.
1821 Node *PhaseCCP::transform( Node *n ) {
1822 Node *new_node = _nodes[n->_idx]; // Check for transformed node
1823 if( new_node != NULL )
1824 return new_node; // Been there, done that, return old answer
1825 new_node = transform_once(n); // Check for constant
1826 _nodes.map( n->_idx, new_node ); // Flag as having been cloned
1827
1828 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1829 GrowableArray <Node *> trstack(C->live_nodes() >> 1);
1830
1831 trstack.push(new_node); // Process children of cloned node
1832 while ( trstack.is_nonempty() ) {
1833 Node *clone = trstack.pop();
1834 uint cnt = clone->req();
1835 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
1836 Node *input = clone->in(i);
1837 if( input != NULL ) { // Ignore NULLs
1838 Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1839 if( new_input == NULL ) {
1840 new_input = transform_once(input); // Check for constant
1841 _nodes.map( input->_idx, new_input );// Flag as having been cloned
1842 trstack.push(new_input);
1843 }
1844 assert( new_input == clone->in(i), "insanity check");
1845 }
1846 }
1847 }
1848 return new_node;
1849 }
1850
1851
1852 //------------------------------transform_once---------------------------------
1853 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
1854 Node *PhaseCCP::transform_once( Node *n ) {
1855 const Type *t = type(n);
1856 // Constant? Use constant Node instead
1857 if( t->singleton() ) {
1858 Node *nn = n; // Default is to return the original constant
1859 if( t == Type::TOP ) {
1860 // cache my top node on the Compile instance
1861 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1862 C->set_cached_top_node(ConNode::make(Type::TOP));
1863 set_type(C->top(), Type::TOP);
1864 }
1865 nn = C->top();
1866 }
1867 if( !n->is_Con() ) {
1868 if( t != Type::TOP ) {
1869 nn = makecon(t); // ConNode::make(t);
1870 NOT_PRODUCT( inc_constants(); )
1871 } else if( n->is_Region() ) { // Unreachable region
1872 // Note: nn == C->top()
1873 n->set_req(0, NULL); // Cut selfreference
1874 // Eagerly remove dead phis to avoid phis copies creation.
1875 for (DUIterator i = n->outs(); n->has_out(i); i++) {
1876 Node* m = n->out(i);
1877 if( m->is_Phi() ) {
1878 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1879 replace_node(m, nn);
1880 --i; // deleted this phi; rescan starting with next position
1881 }
1882 }
1883 }
1884 replace_node(n,nn); // Update DefUse edges for new constant
1885 }
1886 return nn;
1887 }
1888
1889 // If x is a TypeNode, capture any more-precise type permanently into Node
1890 if (t != n->bottom_type()) {
1891 hash_delete(n); // changing bottom type may force a rehash
1892 n->raise_bottom_type(t);
1893 _worklist.push(n); // n re-enters the hash table via the worklist
1894 }
1895
1896 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
1897 switch( n->Opcode() ) {
1898 case Op_FastLock: // Revisit FastLocks for lock coarsening
1899 case Op_If:
1900 case Op_CountedLoopEnd:
1901 case Op_Region:
1902 case Op_Loop:
1903 case Op_CountedLoop:
1904 case Op_Conv2B:
1905 case Op_Opaque1:
1906 case Op_Opaque2:
1907 _worklist.push(n);
1908 break;
1909 default:
1910 break;
1911 }
1912
1913 return n;
1914 }
1915
1916 //---------------------------------saturate------------------------------------
1917 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
1918 const Type* limit_type) const {
1919 const Type* wide_type = new_type->widen(old_type, limit_type);
1920 if (wide_type != new_type) { // did we widen?
1921 // If so, we may have widened beyond the limit type. Clip it back down.
1922 new_type = wide_type->filter(limit_type);
1923 }
1924 return new_type;
1925 }
1926
1927 //------------------------------print_statistics-------------------------------
1928 #ifndef PRODUCT
1929 void PhaseCCP::print_statistics() {
1930 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
1931 }
1932 #endif
1933
1934
1935 //=============================================================================
1936 #ifndef PRODUCT
1937 uint PhasePeephole::_total_peepholes = 0;
1938 #endif
1939 //------------------------------PhasePeephole----------------------------------
1940 // Conditional Constant Propagation, ala Wegman & Zadeck
1941 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
1942 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
1943 NOT_PRODUCT( clear_peepholes(); )
1944 }
1945
1946 #ifndef PRODUCT
1947 //------------------------------~PhasePeephole---------------------------------
1948 PhasePeephole::~PhasePeephole() {
1949 _total_peepholes += count_peepholes();
1950 }
1951 #endif
1952
1953 //------------------------------transform--------------------------------------
1954 Node *PhasePeephole::transform( Node *n ) {
1955 ShouldNotCallThis();
1956 return NULL;
1957 }
1958
1959 //------------------------------do_transform-----------------------------------
1960 void PhasePeephole::do_transform() {
1961 bool method_name_not_printed = true;
1962
1963 // Examine each basic block
1964 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
1965 Block* block = _cfg.get_block(block_number);
1966 bool block_not_printed = true;
1967
1968 // and each instruction within a block
1969 uint end_index = block->number_of_nodes();
1970 // block->end_idx() not valid after PhaseRegAlloc
1971 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
1972 Node *n = block->get_node(instruction_index);
1973 if( n->is_Mach() ) {
1974 MachNode *m = n->as_Mach();
1975 int deleted_count = 0;
1976 // check for peephole opportunities
1977 MachNode *m2 = m->peephole(block, instruction_index, _regalloc, deleted_count);
1978 if( m2 != NULL ) {
1979 #ifndef PRODUCT
1980 if( PrintOptoPeephole ) {
1981 // Print method, first time only
1982 if( C->method() && method_name_not_printed ) {
1983 C->method()->print_short_name(); tty->cr();
1984 method_name_not_printed = false;
1985 }
1986 // Print this block
1987 if( Verbose && block_not_printed) {
1988 tty->print_cr("in block");
1989 block->dump();
1990 block_not_printed = false;
1991 }
1992 // Print instructions being deleted
1993 for( int i = (deleted_count - 1); i >= 0; --i ) {
1994 block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
1995 }
1996 tty->print_cr("replaced with");
1997 // Print new instruction
1998 m2->format(_regalloc);
1999 tty->print("\n\n");
2000 }
2001 #endif
2002 // Remove old nodes from basic block and update instruction_index
2003 // (old nodes still exist and may have edges pointing to them
2004 // as register allocation info is stored in the allocator using
2005 // the node index to live range mappings.)
2006 uint safe_instruction_index = (instruction_index - deleted_count);
2007 for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
2008 block->remove_node( instruction_index );
2009 }
2010 // install new node after safe_instruction_index
2011 block->insert_node(m2, safe_instruction_index + 1);
2012 end_index = block->number_of_nodes() - 1; // Recompute new block size
2013 NOT_PRODUCT( inc_peepholes(); )
2014 }
2015 }
2016 }
2017 }
2018 }
2019
2020 //------------------------------print_statistics-------------------------------
2021 #ifndef PRODUCT
2022 void PhasePeephole::print_statistics() {
2023 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
2024 }
2025 #endif
2026
2027
2028 //=============================================================================
2029 //------------------------------set_req_X--------------------------------------
2030 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
2031 assert( is_not_dead(n), "can not use dead node");
2032 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
2033 Node *old = in(i);
2034 set_req(i, n);
2035
2036 // old goes dead?
2037 if( old ) {
2038 switch (old->outcnt()) {
2039 case 0:
2040 // Put into the worklist to kill later. We do not kill it now because the
2041 // recursive kill will delete the current node (this) if dead-loop exists
2042 if (!old->is_top())
2043 igvn->_worklist.push( old );
2044 break;
2045 case 1:
2046 if( old->is_Store() || old->has_special_unique_user() )
2047 igvn->add_users_to_worklist( old );
2048 break;
2049 case 2:
2050 if( old->is_Store() )
2051 igvn->add_users_to_worklist( old );
2052 if( old->Opcode() == Op_Region )
2053 igvn->_worklist.push(old);
2054 break;
2055 case 3:
2056 if( old->Opcode() == Op_Region ) {
2057 igvn->_worklist.push(old);
2058 igvn->add_users_to_worklist( old );
2059 }
2060 break;
2061 default:
2062 break;
2063 }
2064 }
2065
2066 }
2067
2068 //-------------------------------replace_by-----------------------------------
2069 // Using def-use info, replace one node for another. Follow the def-use info
2070 // to all users of the OLD node. Then make all uses point to the NEW node.
2071 void Node::replace_by(Node *new_node) {
2072 assert(!is_top(), "top node has no DU info");
2073 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
2074 Node* use = last_out(i);
2075 uint uses_found = 0;
2076 for (uint j = 0; j < use->len(); j++) {
2077 if (use->in(j) == this) {
2078 if (j < use->req())
2079 use->set_req(j, new_node);
2080 else use->set_prec(j, new_node);
2081 uses_found++;
2082 }
2083 }
2084 i -= uses_found; // we deleted 1 or more copies of this edge
2085 }
2086 }
2087
2088 //=============================================================================
2089 //-----------------------------------------------------------------------------
2090 void Type_Array::grow( uint i ) {
2091 if( !_max ) {
2092 _max = 1;
2093 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
2094 _types[0] = NULL;
2095 }
2096 uint old = _max;
2097 while( i >= _max ) _max <<= 1; // Double to fit
2098 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
2099 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
2100 }
2101
2102 //------------------------------dump-------------------------------------------
2103 #ifndef PRODUCT
2104 void Type_Array::dump() const {
2105 uint max = Size();
2106 for( uint i = 0; i < max; i++ ) {
2107 if( _types[i] != NULL ) {
2108 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
2109 }
2110 }
2111 }
2112 #endif