Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 1997, 2011, 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 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/block.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/cfgnode.hpp"
30 #include "opto/connode.hpp"
31 #include "opto/idealGraphPrinter.hpp"
32 #include "opto/loopnode.hpp"
33 #include "opto/machnode.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/regalloc.hpp"
37 #include "opto/rootnode.hpp"
39 //=============================================================================
40 #define NODE_HASH_MINIMUM_SIZE 255
41 //------------------------------NodeHash---------------------------------------
42 NodeHash::NodeHash(uint est_max_size) :
43 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
44 _a(Thread::current()->resource_area()),
45 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ), // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
46 _inserts(0), _insert_limit( insert_limit() ),
47 _look_probes(0), _lookup_hits(0), _lookup_misses(0),
48 _total_insert_probes(0), _total_inserts(0),
49 _insert_probes(0), _grows(0) {
50 // _sentinel must be in the current node space
51 _sentinel = new (Compile::current(), 1) ProjNode(NULL, TypeFunc::Control);
52 memset(_table,0,sizeof(Node*)*_max);
53 }
55 //------------------------------NodeHash---------------------------------------
56 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
57 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
58 _a(arena),
59 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ),
60 _inserts(0), _insert_limit( insert_limit() ),
61 _look_probes(0), _lookup_hits(0), _lookup_misses(0),
62 _delete_probes(0), _delete_hits(0), _delete_misses(0),
63 _total_insert_probes(0), _total_inserts(0),
64 _insert_probes(0), _grows(0) {
65 // _sentinel must be in the current node space
66 _sentinel = new (Compile::current(), 1) ProjNode(NULL, TypeFunc::Control);
67 memset(_table,0,sizeof(Node*)*_max);
68 }
70 //------------------------------NodeHash---------------------------------------
71 NodeHash::NodeHash(NodeHash *nh) {
72 debug_only(_table = (Node**)badAddress); // interact correctly w/ operator=
73 // just copy in all the fields
74 *this = *nh;
75 // nh->_sentinel must be in the current node space
76 }
78 //------------------------------hash_find--------------------------------------
79 // Find in hash table
80 Node *NodeHash::hash_find( const Node *n ) {
81 // ((Node*)n)->set_hash( n->hash() );
82 uint hash = n->hash();
83 if (hash == Node::NO_HASH) {
84 debug_only( _lookup_misses++ );
85 return NULL;
86 }
87 uint key = hash & (_max-1);
88 uint stride = key | 0x01;
89 debug_only( _look_probes++ );
90 Node *k = _table[key]; // Get hashed value
91 if( !k ) { // ?Miss?
92 debug_only( _lookup_misses++ );
93 return NULL; // Miss!
94 }
96 int op = n->Opcode();
97 uint req = n->req();
98 while( 1 ) { // While probing hash table
99 if( k->req() == req && // Same count of inputs
100 k->Opcode() == op ) { // Same Opcode
101 for( uint i=0; i<req; i++ )
102 if( n->in(i)!=k->in(i)) // Different inputs?
103 goto collision; // "goto" is a speed hack...
104 if( n->cmp(*k) ) { // Check for any special bits
105 debug_only( _lookup_hits++ );
106 return k; // Hit!
107 }
108 }
109 collision:
110 debug_only( _look_probes++ );
111 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
112 k = _table[key]; // Get hashed value
113 if( !k ) { // ?Miss?
114 debug_only( _lookup_misses++ );
115 return NULL; // Miss!
116 }
117 }
118 ShouldNotReachHere();
119 return NULL;
120 }
122 //------------------------------hash_find_insert-------------------------------
123 // Find in hash table, insert if not already present
124 // Used to preserve unique entries in hash table
125 Node *NodeHash::hash_find_insert( Node *n ) {
126 // n->set_hash( );
127 uint hash = n->hash();
128 if (hash == Node::NO_HASH) {
129 debug_only( _lookup_misses++ );
130 return NULL;
131 }
132 uint key = hash & (_max-1);
133 uint stride = key | 0x01; // stride must be relatively prime to table siz
134 uint first_sentinel = 0; // replace a sentinel if seen.
135 debug_only( _look_probes++ );
136 Node *k = _table[key]; // Get hashed value
137 if( !k ) { // ?Miss?
138 debug_only( _lookup_misses++ );
139 _table[key] = n; // Insert into table!
140 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
141 check_grow(); // Grow table if insert hit limit
142 return NULL; // Miss!
143 }
144 else if( k == _sentinel ) {
145 first_sentinel = key; // Can insert here
146 }
148 int op = n->Opcode();
149 uint req = n->req();
150 while( 1 ) { // While probing hash table
151 if( k->req() == req && // Same count of inputs
152 k->Opcode() == op ) { // Same Opcode
153 for( uint i=0; i<req; i++ )
154 if( n->in(i)!=k->in(i)) // Different inputs?
155 goto collision; // "goto" is a speed hack...
156 if( n->cmp(*k) ) { // Check for any special bits
157 debug_only( _lookup_hits++ );
158 return k; // Hit!
159 }
160 }
161 collision:
162 debug_only( _look_probes++ );
163 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
164 k = _table[key]; // Get hashed value
165 if( !k ) { // ?Miss?
166 debug_only( _lookup_misses++ );
167 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
168 _table[key] = n; // Insert into table!
169 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
170 check_grow(); // Grow table if insert hit limit
171 return NULL; // Miss!
172 }
173 else if( first_sentinel == 0 && k == _sentinel ) {
174 first_sentinel = key; // Can insert here
175 }
177 }
178 ShouldNotReachHere();
179 return NULL;
180 }
182 //------------------------------hash_insert------------------------------------
183 // Insert into hash table
184 void NodeHash::hash_insert( Node *n ) {
185 // // "conflict" comments -- print nodes that conflict
186 // bool conflict = false;
187 // n->set_hash();
188 uint hash = n->hash();
189 if (hash == Node::NO_HASH) {
190 return;
191 }
192 check_grow();
193 uint key = hash & (_max-1);
194 uint stride = key | 0x01;
196 while( 1 ) { // While probing hash table
197 debug_only( _insert_probes++ );
198 Node *k = _table[key]; // Get hashed value
199 if( !k || (k == _sentinel) ) break; // Found a slot
200 assert( k != n, "already inserted" );
201 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
202 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
203 }
204 _table[key] = n; // Insert into table!
205 debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
206 // if( conflict ) { n->dump(); }
207 }
209 //------------------------------hash_delete------------------------------------
210 // Replace in hash table with sentinel
211 bool NodeHash::hash_delete( const Node *n ) {
212 Node *k;
213 uint hash = n->hash();
214 if (hash == Node::NO_HASH) {
215 debug_only( _delete_misses++ );
216 return false;
217 }
218 uint key = hash & (_max-1);
219 uint stride = key | 0x01;
220 debug_only( uint counter = 0; );
221 for( ; /* (k != NULL) && (k != _sentinel) */; ) {
222 debug_only( counter++ );
223 debug_only( _delete_probes++ );
224 k = _table[key]; // Get hashed value
225 if( !k ) { // Miss?
226 debug_only( _delete_misses++ );
227 #ifdef ASSERT
228 if( VerifyOpto ) {
229 for( uint i=0; i < _max; i++ )
230 assert( _table[i] != n, "changed edges with rehashing" );
231 }
232 #endif
233 return false; // Miss! Not in chain
234 }
235 else if( n == k ) {
236 debug_only( _delete_hits++ );
237 _table[key] = _sentinel; // Hit! Label as deleted entry
238 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
239 return true;
240 }
241 else {
242 // collision: move through table with prime offset
243 key = (key + stride/*7*/) & (_max-1);
244 assert( counter <= _insert_limit, "Cycle in hash-table");
245 }
246 }
247 ShouldNotReachHere();
248 return false;
249 }
251 //------------------------------round_up---------------------------------------
252 // Round up to nearest power of 2
253 uint NodeHash::round_up( uint x ) {
254 x += (x>>2); // Add 25% slop
255 if( x <16 ) return 16; // Small stuff
256 uint i=16;
257 while( i < x ) i <<= 1; // Double to fit
258 return i; // Return hash table size
259 }
261 //------------------------------grow-------------------------------------------
262 // Grow _table to next power of 2 and insert old entries
263 void NodeHash::grow() {
264 // Record old state
265 uint old_max = _max;
266 Node **old_table = _table;
267 // Construct new table with twice the space
268 _grows++;
269 _total_inserts += _inserts;
270 _total_insert_probes += _insert_probes;
271 _inserts = 0;
272 _insert_probes = 0;
273 _max = _max << 1;
274 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
275 memset(_table,0,sizeof(Node*)*_max);
276 _insert_limit = insert_limit();
277 // Insert old entries into the new table
278 for( uint i = 0; i < old_max; i++ ) {
279 Node *m = *old_table++;
280 if( !m || m == _sentinel ) continue;
281 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
282 hash_insert(m);
283 }
284 }
286 //------------------------------clear------------------------------------------
287 // Clear all entries in _table to NULL but keep storage
288 void NodeHash::clear() {
289 #ifdef ASSERT
290 // Unlock all nodes upon removal from table.
291 for (uint i = 0; i < _max; i++) {
292 Node* n = _table[i];
293 if (!n || n == _sentinel) continue;
294 n->exit_hash_lock();
295 }
296 #endif
298 memset( _table, 0, _max * sizeof(Node*) );
299 }
301 //-----------------------remove_useless_nodes----------------------------------
302 // Remove useless nodes from value table,
303 // implementation does not depend on hash function
304 void NodeHash::remove_useless_nodes(VectorSet &useful) {
306 // Dead nodes in the hash table inherited from GVN should not replace
307 // existing nodes, remove dead nodes.
308 uint max = size();
309 Node *sentinel_node = sentinel();
310 for( uint i = 0; i < max; ++i ) {
311 Node *n = at(i);
312 if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
313 debug_only(n->exit_hash_lock()); // Unlock the node when removed
314 _table[i] = sentinel_node; // Replace with placeholder
315 }
316 }
317 }
319 #ifndef PRODUCT
320 //------------------------------dump-------------------------------------------
321 // Dump statistics for the hash table
322 void NodeHash::dump() {
323 _total_inserts += _inserts;
324 _total_insert_probes += _insert_probes;
325 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
326 if (WizardMode) {
327 for (uint i=0; i<_max; i++) {
328 if (_table[i])
329 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
330 }
331 }
332 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
333 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
334 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
335 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
336 // sentinels increase lookup cost, but not insert cost
337 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
338 assert( _inserts+(_inserts>>3) < _max, "table too full" );
339 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
340 }
341 }
343 Node *NodeHash::find_index(uint idx) { // For debugging
344 // Find an entry by its index value
345 for( uint i = 0; i < _max; i++ ) {
346 Node *m = _table[i];
347 if( !m || m == _sentinel ) continue;
348 if( m->_idx == (uint)idx ) return m;
349 }
350 return NULL;
351 }
352 #endif
354 #ifdef ASSERT
355 NodeHash::~NodeHash() {
356 // Unlock all nodes upon destruction of table.
357 if (_table != (Node**)badAddress) clear();
358 }
360 void NodeHash::operator=(const NodeHash& nh) {
361 // Unlock all nodes upon replacement of table.
362 if (&nh == this) return;
363 if (_table != (Node**)badAddress) clear();
364 memcpy(this, &nh, sizeof(*this));
365 // Do not increment hash_lock counts again.
366 // Instead, be sure we never again use the source table.
367 ((NodeHash*)&nh)->_table = (Node**)badAddress;
368 }
371 #endif
374 //=============================================================================
375 //------------------------------PhaseRemoveUseless-----------------------------
376 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
377 PhaseRemoveUseless::PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist ) : Phase(Remove_Useless),
378 _useful(Thread::current()->resource_area()) {
380 // Implementation requires 'UseLoopSafepoints == true' and an edge from root
381 // to each SafePointNode at a backward branch. Inserted in add_safepoint().
382 if( !UseLoopSafepoints || !OptoRemoveUseless ) return;
384 // Identify nodes that are reachable from below, useful.
385 C->identify_useful_nodes(_useful);
387 // Remove all useless nodes from PhaseValues' recorded types
388 // Must be done before disconnecting nodes to preserve hash-table-invariant
389 gvn->remove_useless_nodes(_useful.member_set());
391 // Remove all useless nodes from future worklist
392 worklist->remove_useless_nodes(_useful.member_set());
394 // Disconnect 'useless' nodes that are adjacent to useful nodes
395 C->remove_useless_nodes(_useful);
397 // Remove edges from "root" to each SafePoint at a backward branch.
398 // They were inserted during parsing (see add_safepoint()) to make infinite
399 // loops without calls or exceptions visible to root, i.e., useful.
400 Node *root = C->root();
401 if( root != NULL ) {
402 for( uint i = root->req(); i < root->len(); ++i ) {
403 Node *n = root->in(i);
404 if( n != NULL && n->is_SafePoint() ) {
405 root->rm_prec(i);
406 --i;
407 }
408 }
409 }
410 }
413 //=============================================================================
414 //------------------------------PhaseTransform---------------------------------
415 PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
416 _arena(Thread::current()->resource_area()),
417 _nodes(_arena),
418 _types(_arena)
419 {
420 init_con_caches();
421 #ifndef PRODUCT
422 clear_progress();
423 clear_transforms();
424 set_allow_progress(true);
425 #endif
426 // Force allocation for currently existing nodes
427 _types.map(C->unique(), NULL);
428 }
430 //------------------------------PhaseTransform---------------------------------
431 PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
432 _arena(arena),
433 _nodes(arena),
434 _types(arena)
435 {
436 init_con_caches();
437 #ifndef PRODUCT
438 clear_progress();
439 clear_transforms();
440 set_allow_progress(true);
441 #endif
442 // Force allocation for currently existing nodes
443 _types.map(C->unique(), NULL);
444 }
446 //------------------------------PhaseTransform---------------------------------
447 // Initialize with previously generated type information
448 PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
449 _arena(pt->_arena),
450 _nodes(pt->_nodes),
451 _types(pt->_types)
452 {
453 init_con_caches();
454 #ifndef PRODUCT
455 clear_progress();
456 clear_transforms();
457 set_allow_progress(true);
458 #endif
459 }
461 void PhaseTransform::init_con_caches() {
462 memset(_icons,0,sizeof(_icons));
463 memset(_lcons,0,sizeof(_lcons));
464 memset(_zcons,0,sizeof(_zcons));
465 }
468 //--------------------------------find_int_type--------------------------------
469 const TypeInt* PhaseTransform::find_int_type(Node* n) {
470 if (n == NULL) return NULL;
471 // Call type_or_null(n) to determine node's type since we might be in
472 // parse phase and call n->Value() may return wrong type.
473 // (For example, a phi node at the beginning of loop parsing is not ready.)
474 const Type* t = type_or_null(n);
475 if (t == NULL) return NULL;
476 return t->isa_int();
477 }
480 //-------------------------------find_long_type--------------------------------
481 const TypeLong* PhaseTransform::find_long_type(Node* n) {
482 if (n == NULL) return NULL;
483 // (See comment above on type_or_null.)
484 const Type* t = type_or_null(n);
485 if (t == NULL) return NULL;
486 return t->isa_long();
487 }
490 #ifndef PRODUCT
491 void PhaseTransform::dump_old2new_map() const {
492 _nodes.dump();
493 }
495 void PhaseTransform::dump_new( uint nidx ) const {
496 for( uint i=0; i<_nodes.Size(); i++ )
497 if( _nodes[i] && _nodes[i]->_idx == nidx ) {
498 _nodes[i]->dump();
499 tty->cr();
500 tty->print_cr("Old index= %d",i);
501 return;
502 }
503 tty->print_cr("Node %d not found in the new indices", nidx);
504 }
506 //------------------------------dump_types-------------------------------------
507 void PhaseTransform::dump_types( ) const {
508 _types.dump();
509 }
511 //------------------------------dump_nodes_and_types---------------------------
512 void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) {
513 VectorSet visited(Thread::current()->resource_area());
514 dump_nodes_and_types_recur( root, depth, only_ctrl, visited );
515 }
517 //------------------------------dump_nodes_and_types_recur---------------------
518 void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
519 if( !n ) return;
520 if( depth == 0 ) return;
521 if( visited.test_set(n->_idx) ) return;
522 for( uint i=0; i<n->len(); i++ ) {
523 if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
524 dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
525 }
526 n->dump();
527 if (type_or_null(n) != NULL) {
528 tty->print(" "); type(n)->dump(); tty->cr();
529 }
530 }
532 #endif
535 //=============================================================================
536 //------------------------------PhaseValues------------------------------------
537 // Set minimum table size to "255"
538 PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) {
539 NOT_PRODUCT( clear_new_values(); )
540 }
542 //------------------------------PhaseValues------------------------------------
543 // Set minimum table size to "255"
544 PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ),
545 _table(&ptv->_table) {
546 NOT_PRODUCT( clear_new_values(); )
547 }
549 //------------------------------PhaseValues------------------------------------
550 // Used by +VerifyOpto. Clear out hash table but copy _types array.
551 PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ),
552 _table(ptv->arena(),ptv->_table.size()) {
553 NOT_PRODUCT( clear_new_values(); )
554 }
556 //------------------------------~PhaseValues-----------------------------------
557 #ifndef PRODUCT
558 PhaseValues::~PhaseValues() {
559 _table.dump();
561 // Statistics for value progress and efficiency
562 if( PrintCompilation && Verbose && WizardMode ) {
563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
565 if( made_transforms() != 0 ) {
566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
567 } else {
568 tty->cr();
569 }
570 }
571 }
572 #endif
574 //------------------------------makecon----------------------------------------
575 ConNode* PhaseTransform::makecon(const Type *t) {
576 assert(t->singleton(), "must be a constant");
577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
578 switch (t->base()) { // fast paths
579 case Type::Half:
580 case Type::Top: return (ConNode*) C->top();
581 case Type::Int: return intcon( t->is_int()->get_con() );
582 case Type::Long: return longcon( t->is_long()->get_con() );
583 }
584 if (t->is_zero_type())
585 return zerocon(t->basic_type());
586 return uncached_makecon(t);
587 }
589 //--------------------------uncached_makecon-----------------------------------
590 // Make an idealized constant - one of ConINode, ConPNode, etc.
591 ConNode* PhaseValues::uncached_makecon(const Type *t) {
592 assert(t->singleton(), "must be a constant");
593 ConNode* x = ConNode::make(C, t);
594 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
595 if (k == NULL) {
596 set_type(x, t); // Missed, provide type mapping
597 GrowableArray<Node_Notes*>* nna = C->node_note_array();
598 if (nna != NULL) {
599 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
600 loc->clear(); // do not put debug info on constants
601 }
602 } else {
603 x->destruct(); // Hit, destroy duplicate constant
604 x = k; // use existing constant
605 }
606 return x;
607 }
609 //------------------------------intcon-----------------------------------------
610 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
611 ConINode* PhaseTransform::intcon(int i) {
612 // Small integer? Check cache! Check that cached node is not dead
613 if (i >= _icon_min && i <= _icon_max) {
614 ConINode* icon = _icons[i-_icon_min];
615 if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
616 return icon;
617 }
618 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
619 assert(icon->is_Con(), "");
620 if (i >= _icon_min && i <= _icon_max)
621 _icons[i-_icon_min] = icon; // Cache small integers
622 return icon;
623 }
625 //------------------------------longcon----------------------------------------
626 // Fast long constant.
627 ConLNode* PhaseTransform::longcon(jlong l) {
628 // Small integer? Check cache! Check that cached node is not dead
629 if (l >= _lcon_min && l <= _lcon_max) {
630 ConLNode* lcon = _lcons[l-_lcon_min];
631 if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
632 return lcon;
633 }
634 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
635 assert(lcon->is_Con(), "");
636 if (l >= _lcon_min && l <= _lcon_max)
637 _lcons[l-_lcon_min] = lcon; // Cache small integers
638 return lcon;
639 }
641 //------------------------------zerocon-----------------------------------------
642 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
643 ConNode* PhaseTransform::zerocon(BasicType bt) {
644 assert((uint)bt <= _zcon_max, "domain check");
645 ConNode* zcon = _zcons[bt];
646 if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
647 return zcon;
648 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
649 _zcons[bt] = zcon;
650 return zcon;
651 }
655 //=============================================================================
656 //------------------------------transform--------------------------------------
657 // Return a node which computes the same function as this node, but in a
658 // faster or cheaper fashion.
659 Node *PhaseGVN::transform( Node *n ) {
660 return transform_no_reclaim(n);
661 }
663 //------------------------------transform--------------------------------------
664 // Return a node which computes the same function as this node, but
665 // in a faster or cheaper fashion.
666 Node *PhaseGVN::transform_no_reclaim( Node *n ) {
667 NOT_PRODUCT( set_transforms(); )
669 // Apply the Ideal call in a loop until it no longer applies
670 Node *k = n;
671 NOT_PRODUCT( uint loop_count = 0; )
672 while( 1 ) {
673 Node *i = k->Ideal(this, /*can_reshape=*/false);
674 if( !i ) break;
675 assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
676 k = i;
677 assert(loop_count++ < K, "infinite loop in PhaseGVN::transform");
678 }
679 NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } )
682 // If brand new node, make space in type array.
683 ensure_type_or_null(k);
685 // Since I just called 'Value' to compute the set of run-time values
686 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
687 // cache Value. Later requests for the local phase->type of this Node can
688 // use the cached Value instead of suffering with 'bottom_type'.
689 const Type *t = k->Value(this); // Get runtime Value set
690 assert(t != NULL, "value sanity");
691 if (type_or_null(k) != t) {
692 #ifndef PRODUCT
693 // Do not count initial visit to node as a transformation
694 if (type_or_null(k) == NULL) {
695 inc_new_values();
696 set_progress();
697 }
698 #endif
699 set_type(k, t);
700 // If k is a TypeNode, capture any more-precise type permanently into Node
701 k->raise_bottom_type(t);
702 }
704 if( t->singleton() && !k->is_Con() ) {
705 NOT_PRODUCT( set_progress(); )
706 return makecon(t); // Turn into a constant
707 }
709 // Now check for Identities
710 Node *i = k->Identity(this); // Look for a nearby replacement
711 if( i != k ) { // Found? Return replacement!
712 NOT_PRODUCT( set_progress(); )
713 return i;
714 }
716 // Global Value Numbering
717 i = hash_find_insert(k); // Insert if new
718 if( i && (i != k) ) {
719 // Return the pre-existing node
720 NOT_PRODUCT( set_progress(); )
721 return i;
722 }
724 // Return Idealized original
725 return k;
726 }
728 #ifdef ASSERT
729 //------------------------------dead_loop_check--------------------------------
730 // Check for a simple dead loop when a data node references itself directly
731 // or through an other data node excluding cons and phis.
732 void PhaseGVN::dead_loop_check( Node *n ) {
733 // Phi may reference itself in a loop
734 if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
735 // Do 2 levels check and only data inputs.
736 bool no_dead_loop = true;
737 uint cnt = n->req();
738 for (uint i = 1; i < cnt && no_dead_loop; i++) {
739 Node *in = n->in(i);
740 if (in == n) {
741 no_dead_loop = false;
742 } else if (in != NULL && !in->is_dead_loop_safe()) {
743 uint icnt = in->req();
744 for (uint j = 1; j < icnt && no_dead_loop; j++) {
745 if (in->in(j) == n || in->in(j) == in)
746 no_dead_loop = false;
747 }
748 }
749 }
750 if (!no_dead_loop) n->dump(3);
751 assert(no_dead_loop, "dead loop detected");
752 }
753 }
754 #endif
756 //=============================================================================
757 //------------------------------PhaseIterGVN-----------------------------------
758 // Initialize hash table to fresh and clean for +VerifyOpto
759 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ),
760 _stack(C->unique() >> 1),
761 _delay_transform(false) {
762 }
764 //------------------------------PhaseIterGVN-----------------------------------
765 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
766 PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn),
767 _worklist( igvn->_worklist ),
768 _stack( igvn->_stack ),
769 _delay_transform(igvn->_delay_transform)
770 {
771 }
773 //------------------------------PhaseIterGVN-----------------------------------
774 // Initialize with previous PhaseGVN info from Parser
775 PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn),
776 _worklist(*C->for_igvn()),
777 _stack(C->unique() >> 1),
778 _delay_transform(false)
779 {
780 uint max;
782 // Dead nodes in the hash table inherited from GVN were not treated as
783 // roots during def-use info creation; hence they represent an invisible
784 // use. Clear them out.
785 max = _table.size();
786 for( uint i = 0; i < max; ++i ) {
787 Node *n = _table.at(i);
788 if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
789 if( n->is_top() ) continue;
790 assert( false, "Parse::remove_useless_nodes missed this node");
791 hash_delete(n);
792 }
793 }
795 // Any Phis or Regions on the worklist probably had uses that could not
796 // make more progress because the uses were made while the Phis and Regions
797 // were in half-built states. Put all uses of Phis and Regions on worklist.
798 max = _worklist.size();
799 for( uint j = 0; j < max; j++ ) {
800 Node *n = _worklist.at(j);
801 uint uop = n->Opcode();
802 if( uop == Op_Phi || uop == Op_Region ||
803 n->is_Type() ||
804 n->is_Mem() )
805 add_users_to_worklist(n);
806 }
807 }
810 #ifndef PRODUCT
811 void PhaseIterGVN::verify_step(Node* n) {
812 _verify_window[_verify_counter % _verify_window_size] = n;
813 ++_verify_counter;
814 ResourceMark rm;
815 ResourceArea *area = Thread::current()->resource_area();
816 VectorSet old_space(area), new_space(area);
817 if (C->unique() < 1000 ||
818 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
819 ++_verify_full_passes;
820 Node::verify_recur(C->root(), -1, old_space, new_space);
821 }
822 const int verify_depth = 4;
823 for ( int i = 0; i < _verify_window_size; i++ ) {
824 Node* n = _verify_window[i];
825 if ( n == NULL ) continue;
826 if( n->in(0) == NodeSentinel ) { // xform_idom
827 _verify_window[i] = n->in(1);
828 --i; continue;
829 }
830 // Typical fanout is 1-2, so this call visits about 6 nodes.
831 Node::verify_recur(n, verify_depth, old_space, new_space);
832 }
833 }
834 #endif
837 //------------------------------init_worklist----------------------------------
838 // Initialize worklist for each node.
839 void PhaseIterGVN::init_worklist( Node *n ) {
840 if( _worklist.member(n) ) return;
841 _worklist.push(n);
842 uint cnt = n->req();
843 for( uint i =0 ; i < cnt; i++ ) {
844 Node *m = n->in(i);
845 if( m ) init_worklist(m);
846 }
847 }
849 //------------------------------optimize---------------------------------------
850 void PhaseIterGVN::optimize() {
851 debug_only(uint num_processed = 0;);
852 #ifndef PRODUCT
853 {
854 _verify_counter = 0;
855 _verify_full_passes = 0;
856 for ( int i = 0; i < _verify_window_size; i++ ) {
857 _verify_window[i] = NULL;
858 }
859 }
860 #endif
862 #ifdef ASSERT
863 Node* prev = NULL;
864 uint rep_cnt = 0;
865 #endif
866 uint loop_count = 0;
868 // Pull from worklist; transform node;
869 // If node has changed: update edge info and put uses on worklist.
870 while( _worklist.size() ) {
871 if (C->check_node_count(NodeLimitFudgeFactor * 2,
872 "out of nodes optimizing method")) {
873 return;
874 }
875 Node *n = _worklist.pop();
876 if (++loop_count >= K * C->unique()) {
877 debug_only(n->dump(4);)
878 assert(false, "infinite loop in PhaseIterGVN::optimize");
879 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
880 return;
881 }
882 #ifdef ASSERT
883 if (n == prev) {
884 if (++rep_cnt > 3) {
885 n->dump(4);
886 assert(false, "loop in Ideal transformation");
887 }
888 } else {
889 rep_cnt = 0;
890 }
891 prev = n;
892 #endif
893 if (TraceIterativeGVN && Verbose) {
894 tty->print(" Pop ");
895 NOT_PRODUCT( n->dump(); )
896 debug_only(if( (num_processed++ % 100) == 0 ) _worklist.print_set();)
897 }
899 if (n->outcnt() != 0) {
901 #ifndef PRODUCT
902 uint wlsize = _worklist.size();
903 const Type* oldtype = type_or_null(n);
904 #endif //PRODUCT
906 Node *nn = transform_old(n);
908 #ifndef PRODUCT
909 if (TraceIterativeGVN) {
910 const Type* newtype = type_or_null(n);
911 if (nn != n) {
912 // print old node
913 tty->print("< ");
914 if (oldtype != newtype && oldtype != NULL) {
915 oldtype->dump();
916 }
917 do { tty->print("\t"); } while (tty->position() < 16);
918 tty->print("<");
919 n->dump();
920 }
921 if (oldtype != newtype || nn != n) {
922 // print new node and/or new type
923 if (oldtype == NULL) {
924 tty->print("* ");
925 } else if (nn != n) {
926 tty->print("> ");
927 } else {
928 tty->print("= ");
929 }
930 if (newtype == NULL) {
931 tty->print("null");
932 } else {
933 newtype->dump();
934 }
935 do { tty->print("\t"); } while (tty->position() < 16);
936 nn->dump();
937 }
938 if (Verbose && wlsize < _worklist.size()) {
939 tty->print(" Push {");
940 while (wlsize != _worklist.size()) {
941 Node* pushed = _worklist.at(wlsize++);
942 tty->print(" %d", pushed->_idx);
943 }
944 tty->print_cr(" }");
945 }
946 }
947 if( VerifyIterativeGVN && nn != n ) {
948 verify_step((Node*) NULL); // ignore n, it might be subsumed
949 }
950 #endif
951 } else if (!n->is_top()) {
952 remove_dead_node(n);
953 }
954 }
956 #ifndef PRODUCT
957 C->verify_graph_edges();
958 if( VerifyOpto && allow_progress() ) {
959 // Must turn off allow_progress to enable assert and break recursion
960 C->root()->verify();
961 { // Check if any progress was missed using IterGVN
962 // Def-Use info enables transformations not attempted in wash-pass
963 // e.g. Region/Phi cleanup, ...
964 // Null-check elision -- may not have reached fixpoint
965 // do not propagate to dominated nodes
966 ResourceMark rm;
967 PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean!
968 // Fill worklist completely
969 igvn2.init_worklist(C->root());
971 igvn2.set_allow_progress(false);
972 igvn2.optimize();
973 igvn2.set_allow_progress(true);
974 }
975 }
976 if ( VerifyIterativeGVN && PrintOpto ) {
977 if ( _verify_counter == _verify_full_passes )
978 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
979 _verify_full_passes);
980 else
981 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
982 _verify_counter, _verify_full_passes);
983 }
984 #endif
985 }
988 //------------------register_new_node_with_optimizer---------------------------
989 // Register a new node with the optimizer. Update the types array, the def-use
990 // info. Put on worklist.
991 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
992 set_type_bottom(n);
993 _worklist.push(n);
994 if (orig != NULL) C->copy_node_notes_to(n, orig);
995 return n;
996 }
998 //------------------------------transform--------------------------------------
999 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
1000 Node *PhaseIterGVN::transform( Node *n ) {
1001 if (_delay_transform) {
1002 // Register the node but don't optimize for now
1003 register_new_node_with_optimizer(n);
1004 return n;
1005 }
1007 // If brand new node, make space in type array, and give it a type.
1008 ensure_type_or_null(n);
1009 if (type_or_null(n) == NULL) {
1010 set_type_bottom(n);
1011 }
1013 return transform_old(n);
1014 }
1016 //------------------------------transform_old----------------------------------
1017 Node *PhaseIterGVN::transform_old( Node *n ) {
1018 #ifndef PRODUCT
1019 debug_only(uint loop_count = 0;);
1020 set_transforms();
1021 #endif
1022 // Remove 'n' from hash table in case it gets modified
1023 _table.hash_delete(n);
1024 if( VerifyIterativeGVN ) {
1025 assert( !_table.find_index(n->_idx), "found duplicate entry in table");
1026 }
1028 // Apply the Ideal call in a loop until it no longer applies
1029 Node *k = n;
1030 DEBUG_ONLY(dead_loop_check(k);)
1031 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1032 Node *i = k->Ideal(this, /*can_reshape=*/true);
1033 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1034 #ifndef PRODUCT
1035 if( VerifyIterativeGVN )
1036 verify_step(k);
1037 if( i && VerifyOpto ) {
1038 if( !allow_progress() ) {
1039 if (i->is_Add() && i->outcnt() == 1) {
1040 // Switched input to left side because this is the only use
1041 } else if( i->is_If() && (i->in(0) == NULL) ) {
1042 // This IF is dead because it is dominated by an equivalent IF When
1043 // dominating if changed, info is not propagated sparsely to 'this'
1044 // Propagating this info further will spuriously identify other
1045 // progress.
1046 return i;
1047 } else
1048 set_progress();
1049 } else
1050 set_progress();
1051 }
1052 #endif
1054 while( i ) {
1055 #ifndef PRODUCT
1056 debug_only( if( loop_count >= K ) i->dump(4); )
1057 assert(loop_count < K, "infinite loop in PhaseIterGVN::transform");
1058 debug_only( loop_count++; )
1059 #endif
1060 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1061 // Made a change; put users of original Node on worklist
1062 add_users_to_worklist( k );
1063 // Replacing root of transform tree?
1064 if( k != i ) {
1065 // Make users of old Node now use new.
1066 subsume_node( k, i );
1067 k = i;
1068 }
1069 DEBUG_ONLY(dead_loop_check(k);)
1070 // Try idealizing again
1071 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1072 i = k->Ideal(this, /*can_reshape=*/true);
1073 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1074 #ifndef PRODUCT
1075 if( VerifyIterativeGVN )
1076 verify_step(k);
1077 if( i && VerifyOpto ) set_progress();
1078 #endif
1079 }
1081 // If brand new node, make space in type array.
1082 ensure_type_or_null(k);
1084 // See what kind of values 'k' takes on at runtime
1085 const Type *t = k->Value(this);
1086 assert(t != NULL, "value sanity");
1088 // Since I just called 'Value' to compute the set of run-time values
1089 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1090 // cache Value. Later requests for the local phase->type of this Node can
1091 // use the cached Value instead of suffering with 'bottom_type'.
1092 if (t != type_or_null(k)) {
1093 NOT_PRODUCT( set_progress(); )
1094 NOT_PRODUCT( inc_new_values();)
1095 set_type(k, t);
1096 // If k is a TypeNode, capture any more-precise type permanently into Node
1097 k->raise_bottom_type(t);
1098 // Move users of node to worklist
1099 add_users_to_worklist( k );
1100 }
1102 // If 'k' computes a constant, replace it with a constant
1103 if( t->singleton() && !k->is_Con() ) {
1104 NOT_PRODUCT( set_progress(); )
1105 Node *con = makecon(t); // Make a constant
1106 add_users_to_worklist( k );
1107 subsume_node( k, con ); // Everybody using k now uses con
1108 return con;
1109 }
1111 // Now check for Identities
1112 i = k->Identity(this); // Look for a nearby replacement
1113 if( i != k ) { // Found? Return replacement!
1114 NOT_PRODUCT( set_progress(); )
1115 add_users_to_worklist( k );
1116 subsume_node( k, i ); // Everybody using k now uses i
1117 return i;
1118 }
1120 // Global Value Numbering
1121 i = hash_find_insert(k); // Check for pre-existing node
1122 if( i && (i != k) ) {
1123 // Return the pre-existing node if it isn't dead
1124 NOT_PRODUCT( set_progress(); )
1125 add_users_to_worklist( k );
1126 subsume_node( k, i ); // Everybody using k now uses i
1127 return i;
1128 }
1130 // Return Idealized original
1131 return k;
1132 }
1134 //---------------------------------saturate------------------------------------
1135 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1136 const Type* limit_type) const {
1137 return new_type->narrow(old_type);
1138 }
1140 //------------------------------remove_globally_dead_node----------------------
1141 // Kill a globally dead Node. All uses are also globally dead and are
1142 // aggressively trimmed.
1143 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1144 enum DeleteProgress {
1145 PROCESS_INPUTS,
1146 PROCESS_OUTPUTS
1147 };
1148 assert(_stack.is_empty(), "not empty");
1149 _stack.push(dead, PROCESS_INPUTS);
1151 while (_stack.is_nonempty()) {
1152 dead = _stack.node();
1153 uint progress_state = _stack.index();
1154 assert(dead != C->root(), "killing root, eh?");
1155 assert(!dead->is_top(), "add check for top when pushing");
1156 NOT_PRODUCT( set_progress(); )
1157 if (progress_state == PROCESS_INPUTS) {
1158 // After following inputs, continue to outputs
1159 _stack.set_index(PROCESS_OUTPUTS);
1160 // Remove from iterative worklist
1161 _worklist.remove(dead);
1162 if (!dead->is_Con()) { // Don't kill cons but uses
1163 bool recurse = false;
1164 // Remove from hash table
1165 _table.hash_delete( dead );
1166 // Smash all inputs to 'dead', isolating him completely
1167 for( uint i = 0; i < dead->req(); i++ ) {
1168 Node *in = dead->in(i);
1169 if( in ) { // Points to something?
1170 dead->set_req(i,NULL); // Kill the edge
1171 if (in->outcnt() == 0 && in != C->top()) {// Made input go dead?
1172 _stack.push(in, PROCESS_INPUTS); // Recursively remove
1173 recurse = true;
1174 } else if (in->outcnt() == 1 &&
1175 in->has_special_unique_user()) {
1176 _worklist.push(in->unique_out());
1177 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1178 if( in->Opcode() == Op_Region )
1179 _worklist.push(in);
1180 else if( in->is_Store() ) {
1181 DUIterator_Fast imax, i = in->fast_outs(imax);
1182 _worklist.push(in->fast_out(i));
1183 i++;
1184 if(in->outcnt() == 2) {
1185 _worklist.push(in->fast_out(i));
1186 i++;
1187 }
1188 assert(!(i < imax), "sanity");
1189 }
1190 }
1191 }
1192 }
1194 if (dead->is_macro()) {
1195 C->remove_macro_node(dead);
1196 }
1198 if (recurse) {
1199 continue;
1200 }
1201 }
1202 }
1204 // Aggressively kill globally dead uses
1205 // (Rather than pushing all the outs at once, we push one at a time,
1206 // plus the parent to resume later, because of the indefinite number
1207 // of edge deletions per loop trip.)
1208 if (dead->outcnt() > 0) {
1209 // Recursively remove
1210 _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1211 } else {
1212 _stack.pop();
1213 }
1214 }
1215 }
1217 //------------------------------subsume_node-----------------------------------
1218 // Remove users from node 'old' and add them to node 'nn'.
1219 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1220 assert( old != hash_find(old), "should already been removed" );
1221 assert( old != C->top(), "cannot subsume top node");
1222 // Copy debug or profile information to the new version:
1223 C->copy_node_notes_to(nn, old);
1224 // Move users of node 'old' to node 'nn'
1225 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1226 Node* use = old->last_out(i); // for each use...
1227 // use might need re-hashing (but it won't if it's a new node)
1228 bool is_in_table = _table.hash_delete( use );
1229 // Update use-def info as well
1230 // We remove all occurrences of old within use->in,
1231 // so as to avoid rehashing any node more than once.
1232 // The hash table probe swamps any outer loop overhead.
1233 uint num_edges = 0;
1234 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1235 if (use->in(j) == old) {
1236 use->set_req(j, nn);
1237 ++num_edges;
1238 }
1239 }
1240 // Insert into GVN hash table if unique
1241 // If a duplicate, 'use' will be cleaned up when pulled off worklist
1242 if( is_in_table ) {
1243 hash_find_insert(use);
1244 }
1245 i -= num_edges; // we deleted 1 or more copies of this edge
1246 }
1248 // Smash all inputs to 'old', isolating him completely
1249 Node *temp = new (C, 1) Node(1);
1250 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
1251 remove_dead_node( old );
1252 temp->del_req(0); // Yank bogus edge
1253 #ifndef PRODUCT
1254 if( VerifyIterativeGVN ) {
1255 for ( int i = 0; i < _verify_window_size; i++ ) {
1256 if ( _verify_window[i] == old )
1257 _verify_window[i] = nn;
1258 }
1259 }
1260 #endif
1261 _worklist.remove(temp); // this can be necessary
1262 temp->destruct(); // reuse the _idx of this little guy
1263 }
1265 //------------------------------add_users_to_worklist--------------------------
1266 void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1267 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1268 _worklist.push(n->fast_out(i)); // Push on worklist
1269 }
1270 }
1272 void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1273 add_users_to_worklist0(n);
1275 // Move users of node to worklist
1276 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1277 Node* use = n->fast_out(i); // Get use
1279 if( use->is_Multi() || // Multi-definer? Push projs on worklist
1280 use->is_Store() ) // Enable store/load same address
1281 add_users_to_worklist0(use);
1283 // If we changed the receiver type to a call, we need to revisit
1284 // the Catch following the call. It's looking for a non-NULL
1285 // receiver to know when to enable the regular fall-through path
1286 // in addition to the NullPtrException path.
1287 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1288 Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control);
1289 if (p != NULL) {
1290 add_users_to_worklist0(p);
1291 }
1292 }
1294 if( use->is_Cmp() ) { // Enable CMP/BOOL optimization
1295 add_users_to_worklist(use); // Put Bool on worklist
1296 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1297 // phi merging either 0 or 1 onto the worklist
1298 if (use->outcnt() > 0) {
1299 Node* bol = use->raw_out(0);
1300 if (bol->outcnt() > 0) {
1301 Node* iff = bol->raw_out(0);
1302 if (iff->outcnt() == 2) {
1303 Node* ifproj0 = iff->raw_out(0);
1304 Node* ifproj1 = iff->raw_out(1);
1305 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1306 Node* region0 = ifproj0->raw_out(0);
1307 Node* region1 = ifproj1->raw_out(0);
1308 if( region0 == region1 )
1309 add_users_to_worklist0(region0);
1310 }
1311 }
1312 }
1313 }
1314 }
1316 uint use_op = use->Opcode();
1317 // If changed Cast input, check Phi users for simple cycles
1318 if( use->is_ConstraintCast() || use->is_CheckCastPP() ) {
1319 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1320 Node* u = use->fast_out(i2);
1321 if (u->is_Phi())
1322 _worklist.push(u);
1323 }
1324 }
1325 // If changed LShift inputs, check RShift users for useless sign-ext
1326 if( use_op == Op_LShiftI ) {
1327 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1328 Node* u = use->fast_out(i2);
1329 if (u->Opcode() == Op_RShiftI)
1330 _worklist.push(u);
1331 }
1332 }
1333 // If changed AddP inputs, check Stores for loop invariant
1334 if( use_op == Op_AddP ) {
1335 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1336 Node* u = use->fast_out(i2);
1337 if (u->is_Mem())
1338 _worklist.push(u);
1339 }
1340 }
1341 // If changed initialization activity, check dependent Stores
1342 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1343 InitializeNode* init = use->as_Allocate()->initialization();
1344 if (init != NULL) {
1345 Node* imem = init->proj_out(TypeFunc::Memory);
1346 if (imem != NULL) add_users_to_worklist0(imem);
1347 }
1348 }
1349 if (use_op == Op_Initialize) {
1350 Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory);
1351 if (imem != NULL) add_users_to_worklist0(imem);
1352 }
1353 }
1354 }
1356 //=============================================================================
1357 #ifndef PRODUCT
1358 uint PhaseCCP::_total_invokes = 0;
1359 uint PhaseCCP::_total_constants = 0;
1360 #endif
1361 //------------------------------PhaseCCP---------------------------------------
1362 // Conditional Constant Propagation, ala Wegman & Zadeck
1363 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1364 NOT_PRODUCT( clear_constants(); )
1365 assert( _worklist.size() == 0, "" );
1366 // Clear out _nodes from IterGVN. Must be clear to transform call.
1367 _nodes.clear(); // Clear out from IterGVN
1368 analyze();
1369 }
1371 #ifndef PRODUCT
1372 //------------------------------~PhaseCCP--------------------------------------
1373 PhaseCCP::~PhaseCCP() {
1374 inc_invokes();
1375 _total_constants += count_constants();
1376 }
1377 #endif
1380 #ifdef ASSERT
1381 static bool ccp_type_widens(const Type* t, const Type* t0) {
1382 assert(t->meet(t0) == t, "Not monotonic");
1383 switch (t->base() == t0->base() ? t->base() : Type::Top) {
1384 case Type::Int:
1385 assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1386 break;
1387 case Type::Long:
1388 assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1389 break;
1390 }
1391 return true;
1392 }
1393 #endif //ASSERT
1395 //------------------------------analyze----------------------------------------
1396 void PhaseCCP::analyze() {
1397 // Initialize all types to TOP, optimistic analysis
1398 for (int i = C->unique() - 1; i >= 0; i--) {
1399 _types.map(i,Type::TOP);
1400 }
1402 // Push root onto worklist
1403 Unique_Node_List worklist;
1404 worklist.push(C->root());
1406 // Pull from worklist; compute new value; push changes out.
1407 // This loop is the meat of CCP.
1408 while( worklist.size() ) {
1409 Node *n = worklist.pop();
1410 const Type *t = n->Value(this);
1411 if (t != type(n)) {
1412 assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1413 #ifndef PRODUCT
1414 if( TracePhaseCCP ) {
1415 t->dump();
1416 do { tty->print("\t"); } while (tty->position() < 16);
1417 n->dump();
1418 }
1419 #endif
1420 set_type(n, t);
1421 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1422 Node* m = n->fast_out(i); // Get user
1423 if( m->is_Region() ) { // New path to Region? Must recheck Phis too
1424 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1425 Node* p = m->fast_out(i2); // Propagate changes to uses
1426 if( p->bottom_type() != type(p) ) // If not already bottomed out
1427 worklist.push(p); // Propagate change to user
1428 }
1429 }
1430 // If we changed the receiver type to a call, we need to revisit
1431 // the Catch following the call. It's looking for a non-NULL
1432 // receiver to know when to enable the regular fall-through path
1433 // in addition to the NullPtrException path
1434 if (m->is_Call()) {
1435 for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1436 Node* p = m->fast_out(i2); // Propagate changes to uses
1437 if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1)
1438 worklist.push(p->unique_out());
1439 }
1440 }
1441 if( m->bottom_type() != type(m) ) // If not already bottomed out
1442 worklist.push(m); // Propagate change to user
1443 }
1444 }
1445 }
1446 }
1448 //------------------------------do_transform-----------------------------------
1449 // Top level driver for the recursive transformer
1450 void PhaseCCP::do_transform() {
1451 // Correct leaves of new-space Nodes; they point to old-space.
1452 C->set_root( transform(C->root())->as_Root() );
1453 assert( C->top(), "missing TOP node" );
1454 assert( C->root(), "missing root" );
1455 }
1457 //------------------------------transform--------------------------------------
1458 // Given a Node in old-space, clone him into new-space.
1459 // Convert any of his old-space children into new-space children.
1460 Node *PhaseCCP::transform( Node *n ) {
1461 Node *new_node = _nodes[n->_idx]; // Check for transformed node
1462 if( new_node != NULL )
1463 return new_node; // Been there, done that, return old answer
1464 new_node = transform_once(n); // Check for constant
1465 _nodes.map( n->_idx, new_node ); // Flag as having been cloned
1467 // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1468 GrowableArray <Node *> trstack(C->unique() >> 1);
1470 trstack.push(new_node); // Process children of cloned node
1471 while ( trstack.is_nonempty() ) {
1472 Node *clone = trstack.pop();
1473 uint cnt = clone->req();
1474 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
1475 Node *input = clone->in(i);
1476 if( input != NULL ) { // Ignore NULLs
1477 Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1478 if( new_input == NULL ) {
1479 new_input = transform_once(input); // Check for constant
1480 _nodes.map( input->_idx, new_input );// Flag as having been cloned
1481 trstack.push(new_input);
1482 }
1483 assert( new_input == clone->in(i), "insanity check");
1484 }
1485 }
1486 }
1487 return new_node;
1488 }
1491 //------------------------------transform_once---------------------------------
1492 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
1493 Node *PhaseCCP::transform_once( Node *n ) {
1494 const Type *t = type(n);
1495 // Constant? Use constant Node instead
1496 if( t->singleton() ) {
1497 Node *nn = n; // Default is to return the original constant
1498 if( t == Type::TOP ) {
1499 // cache my top node on the Compile instance
1500 if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1501 C->set_cached_top_node( ConNode::make(C, Type::TOP) );
1502 set_type(C->top(), Type::TOP);
1503 }
1504 nn = C->top();
1505 }
1506 if( !n->is_Con() ) {
1507 if( t != Type::TOP ) {
1508 nn = makecon(t); // ConNode::make(t);
1509 NOT_PRODUCT( inc_constants(); )
1510 } else if( n->is_Region() ) { // Unreachable region
1511 // Note: nn == C->top()
1512 n->set_req(0, NULL); // Cut selfreference
1513 // Eagerly remove dead phis to avoid phis copies creation.
1514 for (DUIterator i = n->outs(); n->has_out(i); i++) {
1515 Node* m = n->out(i);
1516 if( m->is_Phi() ) {
1517 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1518 replace_node(m, nn);
1519 --i; // deleted this phi; rescan starting with next position
1520 }
1521 }
1522 }
1523 replace_node(n,nn); // Update DefUse edges for new constant
1524 }
1525 return nn;
1526 }
1528 // If x is a TypeNode, capture any more-precise type permanently into Node
1529 if (t != n->bottom_type()) {
1530 hash_delete(n); // changing bottom type may force a rehash
1531 n->raise_bottom_type(t);
1532 _worklist.push(n); // n re-enters the hash table via the worklist
1533 }
1535 // Idealize graph using DU info. Must clone() into new-space.
1536 // DU info is generally used to show profitability, progress or safety
1537 // (but generally not needed for correctness).
1538 Node *nn = n->Ideal_DU_postCCP(this);
1540 // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
1541 switch( n->Opcode() ) {
1542 case Op_FastLock: // Revisit FastLocks for lock coarsening
1543 case Op_If:
1544 case Op_CountedLoopEnd:
1545 case Op_Region:
1546 case Op_Loop:
1547 case Op_CountedLoop:
1548 case Op_Conv2B:
1549 case Op_Opaque1:
1550 case Op_Opaque2:
1551 _worklist.push(n);
1552 break;
1553 default:
1554 break;
1555 }
1556 if( nn ) {
1557 _worklist.push(n);
1558 // Put users of 'n' onto worklist for second igvn transform
1559 add_users_to_worklist(n);
1560 return nn;
1561 }
1563 return n;
1564 }
1566 //---------------------------------saturate------------------------------------
1567 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
1568 const Type* limit_type) const {
1569 const Type* wide_type = new_type->widen(old_type, limit_type);
1570 if (wide_type != new_type) { // did we widen?
1571 // If so, we may have widened beyond the limit type. Clip it back down.
1572 new_type = wide_type->filter(limit_type);
1573 }
1574 return new_type;
1575 }
1577 //------------------------------print_statistics-------------------------------
1578 #ifndef PRODUCT
1579 void PhaseCCP::print_statistics() {
1580 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
1581 }
1582 #endif
1585 //=============================================================================
1586 #ifndef PRODUCT
1587 uint PhasePeephole::_total_peepholes = 0;
1588 #endif
1589 //------------------------------PhasePeephole----------------------------------
1590 // Conditional Constant Propagation, ala Wegman & Zadeck
1591 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
1592 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
1593 NOT_PRODUCT( clear_peepholes(); )
1594 }
1596 #ifndef PRODUCT
1597 //------------------------------~PhasePeephole---------------------------------
1598 PhasePeephole::~PhasePeephole() {
1599 _total_peepholes += count_peepholes();
1600 }
1601 #endif
1603 //------------------------------transform--------------------------------------
1604 Node *PhasePeephole::transform( Node *n ) {
1605 ShouldNotCallThis();
1606 return NULL;
1607 }
1609 //------------------------------do_transform-----------------------------------
1610 void PhasePeephole::do_transform() {
1611 bool method_name_not_printed = true;
1613 // Examine each basic block
1614 for( uint block_number = 1; block_number < _cfg._num_blocks; ++block_number ) {
1615 Block *block = _cfg._blocks[block_number];
1616 bool block_not_printed = true;
1618 // and each instruction within a block
1619 uint end_index = block->_nodes.size();
1620 // block->end_idx() not valid after PhaseRegAlloc
1621 for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
1622 Node *n = block->_nodes.at(instruction_index);
1623 if( n->is_Mach() ) {
1624 MachNode *m = n->as_Mach();
1625 int deleted_count = 0;
1626 // check for peephole opportunities
1627 MachNode *m2 = m->peephole( block, instruction_index, _regalloc, deleted_count, C );
1628 if( m2 != NULL ) {
1629 #ifndef PRODUCT
1630 if( PrintOptoPeephole ) {
1631 // Print method, first time only
1632 if( C->method() && method_name_not_printed ) {
1633 C->method()->print_short_name(); tty->cr();
1634 method_name_not_printed = false;
1635 }
1636 // Print this block
1637 if( Verbose && block_not_printed) {
1638 tty->print_cr("in block");
1639 block->dump();
1640 block_not_printed = false;
1641 }
1642 // Print instructions being deleted
1643 for( int i = (deleted_count - 1); i >= 0; --i ) {
1644 block->_nodes.at(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
1645 }
1646 tty->print_cr("replaced with");
1647 // Print new instruction
1648 m2->format(_regalloc);
1649 tty->print("\n\n");
1650 }
1651 #endif
1652 // Remove old nodes from basic block and update instruction_index
1653 // (old nodes still exist and may have edges pointing to them
1654 // as register allocation info is stored in the allocator using
1655 // the node index to live range mappings.)
1656 uint safe_instruction_index = (instruction_index - deleted_count);
1657 for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
1658 block->_nodes.remove( instruction_index );
1659 }
1660 // install new node after safe_instruction_index
1661 block->_nodes.insert( safe_instruction_index + 1, m2 );
1662 end_index = block->_nodes.size() - 1; // Recompute new block size
1663 NOT_PRODUCT( inc_peepholes(); )
1664 }
1665 }
1666 }
1667 }
1668 }
1670 //------------------------------print_statistics-------------------------------
1671 #ifndef PRODUCT
1672 void PhasePeephole::print_statistics() {
1673 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
1674 }
1675 #endif
1678 //=============================================================================
1679 //------------------------------set_req_X--------------------------------------
1680 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
1681 assert( is_not_dead(n), "can not use dead node");
1682 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
1683 Node *old = in(i);
1684 set_req(i, n);
1686 // old goes dead?
1687 if( old ) {
1688 switch (old->outcnt()) {
1689 case 0:
1690 // Put into the worklist to kill later. We do not kill it now because the
1691 // recursive kill will delete the current node (this) if dead-loop exists
1692 if (!old->is_top())
1693 igvn->_worklist.push( old );
1694 break;
1695 case 1:
1696 if( old->is_Store() || old->has_special_unique_user() )
1697 igvn->add_users_to_worklist( old );
1698 break;
1699 case 2:
1700 if( old->is_Store() )
1701 igvn->add_users_to_worklist( old );
1702 if( old->Opcode() == Op_Region )
1703 igvn->_worklist.push(old);
1704 break;
1705 case 3:
1706 if( old->Opcode() == Op_Region ) {
1707 igvn->_worklist.push(old);
1708 igvn->add_users_to_worklist( old );
1709 }
1710 break;
1711 default:
1712 break;
1713 }
1714 }
1716 }
1718 //-------------------------------replace_by-----------------------------------
1719 // Using def-use info, replace one node for another. Follow the def-use info
1720 // to all users of the OLD node. Then make all uses point to the NEW node.
1721 void Node::replace_by(Node *new_node) {
1722 assert(!is_top(), "top node has no DU info");
1723 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
1724 Node* use = last_out(i);
1725 uint uses_found = 0;
1726 for (uint j = 0; j < use->len(); j++) {
1727 if (use->in(j) == this) {
1728 if (j < use->req())
1729 use->set_req(j, new_node);
1730 else use->set_prec(j, new_node);
1731 uses_found++;
1732 }
1733 }
1734 i -= uses_found; // we deleted 1 or more copies of this edge
1735 }
1736 }
1738 //=============================================================================
1739 //-----------------------------------------------------------------------------
1740 void Type_Array::grow( uint i ) {
1741 if( !_max ) {
1742 _max = 1;
1743 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
1744 _types[0] = NULL;
1745 }
1746 uint old = _max;
1747 while( i >= _max ) _max <<= 1; // Double to fit
1748 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
1749 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
1750 }
1752 //------------------------------dump-------------------------------------------
1753 #ifndef PRODUCT
1754 void Type_Array::dump() const {
1755 uint max = Size();
1756 for( uint i = 0; i < max; i++ ) {
1757 if( _types[i] != NULL ) {
1758 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
1759 }
1760 }
1761 }
1762 #endif