Mon, 24 Nov 2014 07:29:03 -0800
8058148: MaxNodeLimit and LiveNodeCountInliningCutoff
Reviewed-by: kvn, roland
1 /*
2 * Copyright (c) 1997, 2013, 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 #ifndef SHARE_VM_OPTO_PHASEX_HPP
26 #define SHARE_VM_OPTO_PHASEX_HPP
28 #include "libadt/dict.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/resourceArea.hpp"
31 #include "opto/memnode.hpp"
32 #include "opto/node.hpp"
33 #include "opto/phase.hpp"
34 #include "opto/type.hpp"
36 class Compile;
37 class ConINode;
38 class ConLNode;
39 class Node;
40 class Type;
41 class PhaseTransform;
42 class PhaseGVN;
43 class PhaseIterGVN;
44 class PhaseCCP;
45 class PhasePeephole;
46 class PhaseRegAlloc;
49 //-----------------------------------------------------------------------------
50 // Expandable closed hash-table of nodes, initialized to NULL.
51 // Note that the constructor just zeros things
52 // Storage is reclaimed when the Arena's lifetime is over.
53 class NodeHash : public StackObj {
54 protected:
55 Arena *_a; // Arena to allocate in
56 uint _max; // Size of table (power of 2)
57 uint _inserts; // For grow and debug, count of hash_inserts
58 uint _insert_limit; // 'grow' when _inserts reaches _insert_limit
59 Node **_table; // Hash table of Node pointers
60 Node *_sentinel; // Replaces deleted entries in hash table
62 public:
63 NodeHash(uint est_max_size);
64 NodeHash(Arena *arena, uint est_max_size);
65 NodeHash(NodeHash *use_this_state);
66 #ifdef ASSERT
67 ~NodeHash(); // Unlock all nodes upon destruction of table.
68 void operator=(const NodeHash&); // Unlock all nodes upon replacement of table.
69 #endif
70 Node *hash_find(const Node*);// Find an equivalent version in hash table
71 Node *hash_find_insert(Node*);// If not in table insert else return found node
72 void hash_insert(Node*); // Insert into hash table
73 bool hash_delete(const Node*);// Replace with _sentinel in hash table
74 void check_grow() {
75 _inserts++;
76 if( _inserts == _insert_limit ) { grow(); }
77 assert( _inserts <= _insert_limit, "hash table overflow");
78 assert( _inserts < _max, "hash table overflow" );
79 }
80 static uint round_up(uint); // Round up to nearest power of 2
81 void grow(); // Grow _table to next power of 2 and rehash
82 // Return 75% of _max, rounded up.
83 uint insert_limit() const { return _max - (_max>>2); }
85 void clear(); // Set all entries to NULL, keep storage.
86 // Size of hash table
87 uint size() const { return _max; }
88 // Return Node* at index in table
89 Node *at(uint table_index) {
90 assert(table_index < _max, "Must be within table");
91 return _table[table_index];
92 }
94 void remove_useless_nodes(VectorSet &useful); // replace with sentinel
95 void replace_with(NodeHash* nh);
96 void check_no_speculative_types(); // Check no speculative part for type nodes in table
98 Node *sentinel() { return _sentinel; }
100 #ifndef PRODUCT
101 Node *find_index(uint idx); // For debugging
102 void dump(); // For debugging, dump statistics
103 #endif
104 uint _grows; // For debugging, count of table grow()s
105 uint _look_probes; // For debugging, count of hash probes
106 uint _lookup_hits; // For debugging, count of hash_finds
107 uint _lookup_misses; // For debugging, count of hash_finds
108 uint _insert_probes; // For debugging, count of hash probes
109 uint _delete_probes; // For debugging, count of hash probes for deletes
110 uint _delete_hits; // For debugging, count of hash probes for deletes
111 uint _delete_misses; // For debugging, count of hash probes for deletes
112 uint _total_inserts; // For debugging, total inserts into hash table
113 uint _total_insert_probes; // For debugging, total probes while inserting
114 };
117 //-----------------------------------------------------------------------------
118 // Map dense integer indices to Types. Uses classic doubling-array trick.
119 // Abstractly provides an infinite array of Type*'s, initialized to NULL.
120 // Note that the constructor just zeros things, and since I use Arena
121 // allocation I do not need a destructor to reclaim storage.
122 // Despite the general name, this class is customized for use by PhaseTransform.
123 class Type_Array : public StackObj {
124 Arena *_a; // Arena to allocate in
125 uint _max;
126 const Type **_types;
127 void grow( uint i ); // Grow array node to fit
128 const Type *operator[] ( uint i ) const // Lookup, or NULL for not mapped
129 { return (i<_max) ? _types[i] : (Type*)NULL; }
130 friend class PhaseTransform;
131 public:
132 Type_Array(Arena *a) : _a(a), _max(0), _types(0) {}
133 Type_Array(Type_Array *ta) : _a(ta->_a), _max(ta->_max), _types(ta->_types) { }
134 const Type *fast_lookup(uint i) const{assert(i<_max,"oob");return _types[i];}
135 // Extend the mapping: index i maps to Type *n.
136 void map( uint i, const Type *n ) { if( i>=_max ) grow(i); _types[i] = n; }
137 uint Size() const { return _max; }
138 #ifndef PRODUCT
139 void dump() const;
140 #endif
141 };
144 //------------------------------PhaseRemoveUseless-----------------------------
145 // Remove useless nodes from GVN hash-table, worklist, and graph
146 class PhaseRemoveUseless : public Phase {
147 protected:
148 Unique_Node_List _useful; // Nodes reachable from root
149 // list is allocated from current resource area
150 public:
151 PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist );
153 Unique_Node_List *get_useful() { return &_useful; }
154 };
157 //------------------------------PhaseTransform---------------------------------
158 // Phases that analyze, then transform. Constructing the Phase object does any
159 // global or slow analysis. The results are cached later for a fast
160 // transformation pass. When the Phase object is deleted the cached analysis
161 // results are deleted.
162 class PhaseTransform : public Phase {
163 protected:
164 Arena* _arena;
165 Node_Array _nodes; // Map old node indices to new nodes.
166 Type_Array _types; // Map old node indices to Types.
168 // ConNode caches:
169 enum { _icon_min = -1 * HeapWordSize,
170 _icon_max = 16 * HeapWordSize,
171 _lcon_min = _icon_min,
172 _lcon_max = _icon_max,
173 _zcon_max = (uint)T_CONFLICT
174 };
175 ConINode* _icons[_icon_max - _icon_min + 1]; // cached jint constant nodes
176 ConLNode* _lcons[_lcon_max - _lcon_min + 1]; // cached jlong constant nodes
177 ConNode* _zcons[_zcon_max + 1]; // cached is_zero_type nodes
178 void init_con_caches();
180 // Support both int and long caches because either might be an intptr_t,
181 // so they show up frequently in address computations.
183 public:
184 PhaseTransform( PhaseNumber pnum );
185 PhaseTransform( Arena *arena, PhaseNumber pnum );
186 PhaseTransform( PhaseTransform *phase, PhaseNumber pnum );
188 Arena* arena() { return _arena; }
189 Type_Array& types() { return _types; }
190 // _nodes is used in varying ways by subclasses, which define local accessors
192 public:
193 // Get a previously recorded type for the node n.
194 // This type must already have been recorded.
195 // If you want the type of a very new (untransformed) node,
196 // you must use type_or_null, and test the result for NULL.
197 const Type* type(const Node* n) const {
198 assert(n != NULL, "must not be null");
199 const Type* t = _types.fast_lookup(n->_idx);
200 assert(t != NULL, "must set before get");
201 return t;
202 }
203 // Get a previously recorded type for the node n,
204 // or else return NULL if there is none.
205 const Type* type_or_null(const Node* n) const {
206 return _types.fast_lookup(n->_idx);
207 }
208 // Record a type for a node.
209 void set_type(const Node* n, const Type *t) {
210 assert(t != NULL, "type must not be null");
211 _types.map(n->_idx, t);
212 }
213 // Record an initial type for a node, the node's bottom type.
214 void set_type_bottom(const Node* n) {
215 // Use this for initialization when bottom_type() (or better) is not handy.
216 // Usually the initialization shoudl be to n->Value(this) instead,
217 // or a hand-optimized value like Type::MEMORY or Type::CONTROL.
218 assert(_types[n->_idx] == NULL, "must set the initial type just once");
219 _types.map(n->_idx, n->bottom_type());
220 }
221 // Make sure the types array is big enough to record a size for the node n.
222 // (In product builds, we never want to do range checks on the types array!)
223 void ensure_type_or_null(const Node* n) {
224 if (n->_idx >= _types.Size())
225 _types.map(n->_idx, NULL); // Grow the types array as needed.
226 }
228 // Utility functions:
229 const TypeInt* find_int_type( Node* n);
230 const TypeLong* find_long_type(Node* n);
231 jint find_int_con( Node* n, jint value_if_unknown) {
232 const TypeInt* t = find_int_type(n);
233 return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown;
234 }
235 jlong find_long_con(Node* n, jlong value_if_unknown) {
236 const TypeLong* t = find_long_type(n);
237 return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown;
238 }
240 // Make an idealized constant, i.e., one of ConINode, ConPNode, ConFNode, etc.
241 // Same as transform(ConNode::make(t)).
242 ConNode* makecon(const Type* t);
243 virtual ConNode* uncached_makecon(const Type* t) // override in PhaseValues
244 { ShouldNotCallThis(); return NULL; }
246 // Fast int or long constant. Same as TypeInt::make(i) or TypeLong::make(l).
247 ConINode* intcon(jint i);
248 ConLNode* longcon(jlong l);
250 // Fast zero or null constant. Same as makecon(Type::get_zero_type(bt)).
251 ConNode* zerocon(BasicType bt);
253 // Return a node which computes the same function as this node, but
254 // in a faster or cheaper fashion.
255 virtual Node *transform( Node *n ) = 0;
257 // Return whether two Nodes are equivalent.
258 // Must not be recursive, since the recursive version is built from this.
259 // For pessimistic optimizations this is simply pointer equivalence.
260 bool eqv(const Node* n1, const Node* n2) const { return n1 == n2; }
262 // For pessimistic passes, the return type must monotonically narrow.
263 // For optimistic passes, the return type must monotonically widen.
264 // It is possible to get into a "death march" in either type of pass,
265 // where the types are continually moving but it will take 2**31 or
266 // more steps to converge. This doesn't happen on most normal loops.
267 //
268 // Here is an example of a deadly loop for an optimistic pass, along
269 // with a partial trace of inferred types:
270 // x = phi(0,x'); L: x' = x+1; if (x' >= 0) goto L;
271 // 0 1 join([0..max], 1)
272 // [0..1] [1..2] join([0..max], [1..2])
273 // [0..2] [1..3] join([0..max], [1..3])
274 // ... ... ...
275 // [0..max] [min]u[1..max] join([0..max], [min..max])
276 // [0..max] ==> fixpoint
277 // We would have proven, the hard way, that the iteration space is all
278 // non-negative ints, with the loop terminating due to 32-bit overflow.
279 //
280 // Here is the corresponding example for a pessimistic pass:
281 // x = phi(0,x'); L: x' = x-1; if (x' >= 0) goto L;
282 // int int join([0..max], int)
283 // [0..max] [-1..max-1] join([0..max], [-1..max-1])
284 // [0..max-1] [-1..max-2] join([0..max], [-1..max-2])
285 // ... ... ...
286 // [0..1] [-1..0] join([0..max], [-1..0])
287 // 0 -1 join([0..max], -1)
288 // 0 == fixpoint
289 // We would have proven, the hard way, that the iteration space is {0}.
290 // (Usually, other optimizations will make the "if (x >= 0)" fold up
291 // before we get into trouble. But not always.)
292 //
293 // It's a pleasant thing to observe that the pessimistic pass
294 // will make short work of the optimistic pass's deadly loop,
295 // and vice versa. That is a good example of the complementary
296 // purposes of the CCP (optimistic) vs. GVN (pessimistic) phases.
297 //
298 // In any case, only widen or narrow a few times before going to the
299 // correct flavor of top or bottom.
300 //
301 // This call only needs to be made once as the data flows around any
302 // given cycle. We do it at Phis, and nowhere else.
303 // The types presented are the new type of a phi (computed by PhiNode::Value)
304 // and the previously computed type, last time the phi was visited.
305 //
306 // The third argument is upper limit for the saturated value,
307 // if the phase wishes to widen the new_type.
308 // If the phase is narrowing, the old type provides a lower limit.
309 // Caller guarantees that old_type and new_type are no higher than limit_type.
310 virtual const Type* saturate(const Type* new_type, const Type* old_type,
311 const Type* limit_type) const
312 { ShouldNotCallThis(); return NULL; }
314 #ifndef PRODUCT
315 void dump_old2new_map() const;
316 void dump_new( uint new_lidx ) const;
317 void dump_types() const;
318 void dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl = true);
319 void dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited);
321 uint _count_progress; // For profiling, count transforms that make progress
322 void set_progress() { ++_count_progress; assert( allow_progress(),"No progress allowed during verification"); }
323 void clear_progress() { _count_progress = 0; }
324 uint made_progress() const { return _count_progress; }
326 uint _count_transforms; // For profiling, count transforms performed
327 void set_transforms() { ++_count_transforms; }
328 void clear_transforms() { _count_transforms = 0; }
329 uint made_transforms() const{ return _count_transforms; }
331 bool _allow_progress; // progress not allowed during verification pass
332 void set_allow_progress(bool allow) { _allow_progress = allow; }
333 bool allow_progress() { return _allow_progress; }
334 #endif
335 };
337 //------------------------------PhaseValues------------------------------------
338 // Phase infrastructure to support values
339 class PhaseValues : public PhaseTransform {
340 protected:
341 NodeHash _table; // Hash table for value-numbering
343 public:
344 PhaseValues( Arena *arena, uint est_max_size );
345 PhaseValues( PhaseValues *pt );
346 PhaseValues( PhaseValues *ptv, const char *dummy );
347 NOT_PRODUCT( ~PhaseValues(); )
348 virtual PhaseIterGVN *is_IterGVN() { return 0; }
350 // Some Ideal and other transforms delete --> modify --> insert values
351 bool hash_delete(Node *n) { return _table.hash_delete(n); }
352 void hash_insert(Node *n) { _table.hash_insert(n); }
353 Node *hash_find_insert(Node *n){ return _table.hash_find_insert(n); }
354 Node *hash_find(const Node *n) { return _table.hash_find(n); }
356 // Used after parsing to eliminate values that are no longer in program
357 void remove_useless_nodes(VectorSet &useful) {
358 _table.remove_useless_nodes(useful);
359 // this may invalidate cached cons so reset the cache
360 init_con_caches();
361 }
363 virtual ConNode* uncached_makecon(const Type* t); // override from PhaseTransform
365 virtual const Type* saturate(const Type* new_type, const Type* old_type,
366 const Type* limit_type) const
367 { return new_type; }
369 #ifndef PRODUCT
370 uint _count_new_values; // For profiling, count new values produced
371 void inc_new_values() { ++_count_new_values; }
372 void clear_new_values() { _count_new_values = 0; }
373 uint made_new_values() const { return _count_new_values; }
374 #endif
375 };
378 //------------------------------PhaseGVN---------------------------------------
379 // Phase for performing local, pessimistic GVN-style optimizations.
380 class PhaseGVN : public PhaseValues {
381 public:
382 PhaseGVN( Arena *arena, uint est_max_size ) : PhaseValues( arena, est_max_size ) {}
383 PhaseGVN( PhaseGVN *gvn ) : PhaseValues( gvn ) {}
384 PhaseGVN( PhaseGVN *gvn, const char *dummy ) : PhaseValues( gvn, dummy ) {}
386 // Return a node which computes the same function as this node, but
387 // in a faster or cheaper fashion.
388 Node *transform( Node *n );
389 Node *transform_no_reclaim( Node *n );
391 void replace_with(PhaseGVN* gvn) {
392 _table.replace_with(&gvn->_table);
393 _types = gvn->_types;
394 }
396 // Check for a simple dead loop when a data node references itself.
397 DEBUG_ONLY(void dead_loop_check(Node *n);)
398 };
400 //------------------------------PhaseIterGVN-----------------------------------
401 // Phase for iteratively performing local, pessimistic GVN-style optimizations.
402 // and ideal transformations on the graph.
403 class PhaseIterGVN : public PhaseGVN {
404 private:
405 bool _delay_transform; // When true simply register the node when calling transform
406 // instead of actually optimizing it
408 // Idealize old Node 'n' with respect to its inputs and its value
409 virtual Node *transform_old( Node *a_node );
411 // Subsume users of node 'old' into node 'nn'
412 void subsume_node( Node *old, Node *nn );
414 Node_Stack _stack; // Stack used to avoid recursion
416 protected:
418 // Idealize new Node 'n' with respect to its inputs and its value
419 virtual Node *transform( Node *a_node );
421 // Warm up hash table, type table and initial worklist
422 void init_worklist( Node *a_root );
424 virtual const Type* saturate(const Type* new_type, const Type* old_type,
425 const Type* limit_type) const;
426 // Usually returns new_type. Returns old_type if new_type is only a slight
427 // improvement, such that it would take many (>>10) steps to reach 2**32.
429 public:
430 PhaseIterGVN( PhaseIterGVN *igvn ); // Used by CCP constructor
431 PhaseIterGVN( PhaseGVN *gvn ); // Used after Parser
432 PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ); // Used after +VerifyOpto
434 virtual PhaseIterGVN *is_IterGVN() { return this; }
436 Unique_Node_List _worklist; // Iterative worklist
438 // Given def-use info and an initial worklist, apply Node::Ideal,
439 // Node::Value, Node::Identity, hash-based value numbering, Node::Ideal_DU
440 // and dominator info to a fixed point.
441 void optimize();
443 // Register a new node with the iter GVN pass without transforming it.
444 // Used when we need to restructure a Region/Phi area and all the Regions
445 // and Phis need to complete this one big transform before any other
446 // transforms can be triggered on the region.
447 // Optional 'orig' is an earlier version of this node.
448 // It is significant only for debugging and profiling.
449 Node* register_new_node_with_optimizer(Node* n, Node* orig = NULL);
451 // Kill a globally dead Node. All uses are also globally dead and are
452 // aggressively trimmed.
453 void remove_globally_dead_node( Node *dead );
455 // Kill all inputs to a dead node, recursively making more dead nodes.
456 // The Node must be dead locally, i.e., have no uses.
457 void remove_dead_node( Node *dead ) {
458 assert(dead->outcnt() == 0 && !dead->is_top(), "node must be dead");
459 remove_globally_dead_node(dead);
460 }
462 // Add users of 'n' to worklist
463 void add_users_to_worklist0( Node *n );
464 void add_users_to_worklist ( Node *n );
466 // Replace old node with new one.
467 void replace_node( Node *old, Node *nn ) {
468 add_users_to_worklist(old);
469 hash_delete(old); // Yank from hash before hacking edges
470 subsume_node(old, nn);
471 }
473 // Delayed node rehash: remove a node from the hash table and rehash it during
474 // next optimizing pass
475 void rehash_node_delayed(Node* n) {
476 hash_delete(n);
477 _worklist.push(n);
478 }
480 // Replace ith edge of "n" with "in"
481 void replace_input_of(Node* n, int i, Node* in) {
482 rehash_node_delayed(n);
483 n->set_req(i, in);
484 }
486 // Delete ith edge of "n"
487 void delete_input_of(Node* n, int i) {
488 rehash_node_delayed(n);
489 n->del_req(i);
490 }
492 bool delay_transform() const { return _delay_transform; }
494 void set_delay_transform(bool delay) {
495 _delay_transform = delay;
496 }
498 // Clone loop predicates. Defined in loopTransform.cpp.
499 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
500 // Create a new if below new_entry for the predicate to be cloned
501 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
502 Deoptimization::DeoptReason reason);
504 void remove_speculative_types();
505 void check_no_speculative_types() {
506 _table.check_no_speculative_types();
507 }
509 #ifndef PRODUCT
510 protected:
511 // Sub-quadratic implementation of VerifyIterativeGVN.
512 julong _verify_counter;
513 julong _verify_full_passes;
514 enum { _verify_window_size = 30 };
515 Node* _verify_window[_verify_window_size];
516 void verify_step(Node* n);
517 #endif
518 };
520 //------------------------------PhaseCCP---------------------------------------
521 // Phase for performing global Conditional Constant Propagation.
522 // Should be replaced with combined CCP & GVN someday.
523 class PhaseCCP : public PhaseIterGVN {
524 // Non-recursive. Use analysis to transform single Node.
525 virtual Node *transform_once( Node *n );
527 public:
528 PhaseCCP( PhaseIterGVN *igvn ); // Compute conditional constants
529 NOT_PRODUCT( ~PhaseCCP(); )
531 // Worklist algorithm identifies constants
532 void analyze();
533 // Recursive traversal of program. Used analysis to modify program.
534 virtual Node *transform( Node *n );
535 // Do any transformation after analysis
536 void do_transform();
538 virtual const Type* saturate(const Type* new_type, const Type* old_type,
539 const Type* limit_type) const;
540 // Returns new_type->widen(old_type), which increments the widen bits until
541 // giving up with TypeInt::INT or TypeLong::LONG.
542 // Result is clipped to limit_type if necessary.
544 #ifndef PRODUCT
545 static uint _total_invokes; // For profiling, count invocations
546 void inc_invokes() { ++PhaseCCP::_total_invokes; }
548 static uint _total_constants; // For profiling, count constants found
549 uint _count_constants;
550 void clear_constants() { _count_constants = 0; }
551 void inc_constants() { ++_count_constants; }
552 uint count_constants() const { return _count_constants; }
554 static void print_statistics();
555 #endif
556 };
559 //------------------------------PhasePeephole----------------------------------
560 // Phase for performing peephole optimizations on register allocated basic blocks.
561 class PhasePeephole : public PhaseTransform {
562 PhaseRegAlloc *_regalloc;
563 PhaseCFG &_cfg;
564 // Recursive traversal of program. Pure function is unused in this phase
565 virtual Node *transform( Node *n );
567 public:
568 PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg );
569 NOT_PRODUCT( ~PhasePeephole(); )
571 // Do any transformation after analysis
572 void do_transform();
574 #ifndef PRODUCT
575 static uint _total_peepholes; // For profiling, count peephole rules applied
576 uint _count_peepholes;
577 void clear_peepholes() { _count_peepholes = 0; }
578 void inc_peepholes() { ++_count_peepholes; }
579 uint count_peepholes() const { return _count_peepholes; }
581 static void print_statistics();
582 #endif
583 };
585 #endif // SHARE_VM_OPTO_PHASEX_HPP