phaseX.hpp revision 1472:c18cbe5936b8
1/* 2 * Copyright (c) 1997, 2009, 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 25class Compile; 26class ConINode; 27class ConLNode; 28class Node; 29class Type; 30class PhaseTransform; 31class PhaseGVN; 32class PhaseIterGVN; 33class PhaseCCP; 34class PhasePeephole; 35class PhaseRegAlloc; 36 37 38//----------------------------------------------------------------------------- 39// Expandable closed hash-table of nodes, initialized to NULL. 40// Note that the constructor just zeros things 41// Storage is reclaimed when the Arena's lifetime is over. 42class NodeHash : public StackObj { 43protected: 44 Arena *_a; // Arena to allocate in 45 uint _max; // Size of table (power of 2) 46 uint _inserts; // For grow and debug, count of hash_inserts 47 uint _insert_limit; // 'grow' when _inserts reaches _insert_limit 48 Node **_table; // Hash table of Node pointers 49 Node *_sentinel; // Replaces deleted entries in hash table 50 51public: 52 NodeHash(uint est_max_size); 53 NodeHash(Arena *arena, uint est_max_size); 54 NodeHash(NodeHash *use_this_state); 55#ifdef ASSERT 56 ~NodeHash(); // Unlock all nodes upon destruction of table. 57 void operator=(const NodeHash&); // Unlock all nodes upon replacement of table. 58#endif 59 Node *hash_find(const Node*);// Find an equivalent version in hash table 60 Node *hash_find_insert(Node*);// If not in table insert else return found node 61 void hash_insert(Node*); // Insert into hash table 62 bool hash_delete(const Node*);// Replace with _sentinel in hash table 63 void check_grow() { 64 _inserts++; 65 if( _inserts == _insert_limit ) { grow(); } 66 assert( _inserts <= _insert_limit, "hash table overflow"); 67 assert( _inserts < _max, "hash table overflow" ); 68 } 69 static uint round_up(uint); // Round up to nearest power of 2 70 void grow(); // Grow _table to next power of 2 and rehash 71 // Return 75% of _max, rounded up. 72 uint insert_limit() const { return _max - (_max>>2); } 73 74 void clear(); // Set all entries to NULL, keep storage. 75 // Size of hash table 76 uint size() const { return _max; } 77 // Return Node* at index in table 78 Node *at(uint table_index) { 79 assert(table_index < _max, "Must be within table"); 80 return _table[table_index]; 81 } 82 83 void remove_useless_nodes(VectorSet &useful); // replace with sentinel 84 85 Node *sentinel() { return _sentinel; } 86 87#ifndef PRODUCT 88 Node *find_index(uint idx); // For debugging 89 void dump(); // For debugging, dump statistics 90#endif 91 uint _grows; // For debugging, count of table grow()s 92 uint _look_probes; // For debugging, count of hash probes 93 uint _lookup_hits; // For debugging, count of hash_finds 94 uint _lookup_misses; // For debugging, count of hash_finds 95 uint _insert_probes; // For debugging, count of hash probes 96 uint _delete_probes; // For debugging, count of hash probes for deletes 97 uint _delete_hits; // For debugging, count of hash probes for deletes 98 uint _delete_misses; // For debugging, count of hash probes for deletes 99 uint _total_inserts; // For debugging, total inserts into hash table 100 uint _total_insert_probes; // For debugging, total probes while inserting 101}; 102 103 104//----------------------------------------------------------------------------- 105// Map dense integer indices to Types. Uses classic doubling-array trick. 106// Abstractly provides an infinite array of Type*'s, initialized to NULL. 107// Note that the constructor just zeros things, and since I use Arena 108// allocation I do not need a destructor to reclaim storage. 109// Despite the general name, this class is customized for use by PhaseTransform. 110class Type_Array : public StackObj { 111 Arena *_a; // Arena to allocate in 112 uint _max; 113 const Type **_types; 114 void grow( uint i ); // Grow array node to fit 115 const Type *operator[] ( uint i ) const // Lookup, or NULL for not mapped 116 { return (i<_max) ? _types[i] : (Type*)NULL; } 117 friend class PhaseTransform; 118public: 119 Type_Array(Arena *a) : _a(a), _max(0), _types(0) {} 120 Type_Array(Type_Array *ta) : _a(ta->_a), _max(ta->_max), _types(ta->_types) { } 121 const Type *fast_lookup(uint i) const{assert(i<_max,"oob");return _types[i];} 122 // Extend the mapping: index i maps to Type *n. 123 void map( uint i, const Type *n ) { if( i>=_max ) grow(i); _types[i] = n; } 124 uint Size() const { return _max; } 125#ifndef PRODUCT 126 void dump() const; 127#endif 128}; 129 130 131//------------------------------PhaseRemoveUseless----------------------------- 132// Remove useless nodes from GVN hash-table, worklist, and graph 133class PhaseRemoveUseless : public Phase { 134protected: 135 Unique_Node_List _useful; // Nodes reachable from root 136 // list is allocated from current resource area 137public: 138 PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist ); 139 140 Unique_Node_List *get_useful() { return &_useful; } 141}; 142 143 144//------------------------------PhaseTransform--------------------------------- 145// Phases that analyze, then transform. Constructing the Phase object does any 146// global or slow analysis. The results are cached later for a fast 147// transformation pass. When the Phase object is deleted the cached analysis 148// results are deleted. 149class PhaseTransform : public Phase { 150protected: 151 Arena* _arena; 152 Node_Array _nodes; // Map old node indices to new nodes. 153 Type_Array _types; // Map old node indices to Types. 154 155 // ConNode caches: 156 enum { _icon_min = -1 * HeapWordSize, 157 _icon_max = 16 * HeapWordSize, 158 _lcon_min = _icon_min, 159 _lcon_max = _icon_max, 160 _zcon_max = (uint)T_CONFLICT 161 }; 162 ConINode* _icons[_icon_max - _icon_min + 1]; // cached jint constant nodes 163 ConLNode* _lcons[_lcon_max - _lcon_min + 1]; // cached jlong constant nodes 164 ConNode* _zcons[_zcon_max + 1]; // cached is_zero_type nodes 165 void init_con_caches(); 166 167 // Support both int and long caches because either might be an intptr_t, 168 // so they show up frequently in address computations. 169 170public: 171 PhaseTransform( PhaseNumber pnum ); 172 PhaseTransform( Arena *arena, PhaseNumber pnum ); 173 PhaseTransform( PhaseTransform *phase, PhaseNumber pnum ); 174 175 Arena* arena() { return _arena; } 176 Type_Array& types() { return _types; } 177 // _nodes is used in varying ways by subclasses, which define local accessors 178 179public: 180 // Get a previously recorded type for the node n. 181 // This type must already have been recorded. 182 // If you want the type of a very new (untransformed) node, 183 // you must use type_or_null, and test the result for NULL. 184 const Type* type(const Node* n) const { 185 const Type* t = _types.fast_lookup(n->_idx); 186 assert(t != NULL, "must set before get"); 187 return t; 188 } 189 // Get a previously recorded type for the node n, 190 // or else return NULL if there is none. 191 const Type* type_or_null(const Node* n) const { 192 return _types.fast_lookup(n->_idx); 193 } 194 // Record a type for a node. 195 void set_type(const Node* n, const Type *t) { 196 assert(t != NULL, "type must not be null"); 197 _types.map(n->_idx, t); 198 } 199 // Record an initial type for a node, the node's bottom type. 200 void set_type_bottom(const Node* n) { 201 // Use this for initialization when bottom_type() (or better) is not handy. 202 // Usually the initialization shoudl be to n->Value(this) instead, 203 // or a hand-optimized value like Type::MEMORY or Type::CONTROL. 204 assert(_types[n->_idx] == NULL, "must set the initial type just once"); 205 _types.map(n->_idx, n->bottom_type()); 206 } 207 // Make sure the types array is big enough to record a size for the node n. 208 // (In product builds, we never want to do range checks on the types array!) 209 void ensure_type_or_null(const Node* n) { 210 if (n->_idx >= _types.Size()) 211 _types.map(n->_idx, NULL); // Grow the types array as needed. 212 } 213 214 // Utility functions: 215 const TypeInt* find_int_type( Node* n); 216 const TypeLong* find_long_type(Node* n); 217 jint find_int_con( Node* n, jint value_if_unknown) { 218 const TypeInt* t = find_int_type(n); 219 return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown; 220 } 221 jlong find_long_con(Node* n, jlong value_if_unknown) { 222 const TypeLong* t = find_long_type(n); 223 return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown; 224 } 225 226 // Make an idealized constant, i.e., one of ConINode, ConPNode, ConFNode, etc. 227 // Same as transform(ConNode::make(t)). 228 ConNode* makecon(const Type* t); 229 virtual ConNode* uncached_makecon(const Type* t) // override in PhaseValues 230 { ShouldNotCallThis(); return NULL; } 231 232 // Fast int or long constant. Same as TypeInt::make(i) or TypeLong::make(l). 233 ConINode* intcon(jint i); 234 ConLNode* longcon(jlong l); 235 236 // Fast zero or null constant. Same as makecon(Type::get_zero_type(bt)). 237 ConNode* zerocon(BasicType bt); 238 239 // Return a node which computes the same function as this node, but 240 // in a faster or cheaper fashion. 241 virtual Node *transform( Node *n ) = 0; 242 243 // Return whether two Nodes are equivalent. 244 // Must not be recursive, since the recursive version is built from this. 245 // For pessimistic optimizations this is simply pointer equivalence. 246 bool eqv(const Node* n1, const Node* n2) const { return n1 == n2; } 247 248 // Return whether two Nodes are equivalent, after stripping casting. 249 bool eqv_uncast(const Node* n1, const Node* n2) const { 250 return eqv(n1->uncast(), n2->uncast()); 251 } 252 253 // For pessimistic passes, the return type must monotonically narrow. 254 // For optimistic passes, the return type must monotonically widen. 255 // It is possible to get into a "death march" in either type of pass, 256 // where the types are continually moving but it will take 2**31 or 257 // more steps to converge. This doesn't happen on most normal loops. 258 // 259 // Here is an example of a deadly loop for an optimistic pass, along 260 // with a partial trace of inferred types: 261 // x = phi(0,x'); L: x' = x+1; if (x' >= 0) goto L; 262 // 0 1 join([0..max], 1) 263 // [0..1] [1..2] join([0..max], [1..2]) 264 // [0..2] [1..3] join([0..max], [1..3]) 265 // ... ... ... 266 // [0..max] [min]u[1..max] join([0..max], [min..max]) 267 // [0..max] ==> fixpoint 268 // We would have proven, the hard way, that the iteration space is all 269 // non-negative ints, with the loop terminating due to 32-bit overflow. 270 // 271 // Here is the corresponding example for a pessimistic pass: 272 // x = phi(0,x'); L: x' = x-1; if (x' >= 0) goto L; 273 // int int join([0..max], int) 274 // [0..max] [-1..max-1] join([0..max], [-1..max-1]) 275 // [0..max-1] [-1..max-2] join([0..max], [-1..max-2]) 276 // ... ... ... 277 // [0..1] [-1..0] join([0..max], [-1..0]) 278 // 0 -1 join([0..max], -1) 279 // 0 == fixpoint 280 // We would have proven, the hard way, that the iteration space is {0}. 281 // (Usually, other optimizations will make the "if (x >= 0)" fold up 282 // before we get into trouble. But not always.) 283 // 284 // It's a pleasant thing to observe that the pessimistic pass 285 // will make short work of the optimistic pass's deadly loop, 286 // and vice versa. That is a good example of the complementary 287 // purposes of the CCP (optimistic) vs. GVN (pessimistic) phases. 288 // 289 // In any case, only widen or narrow a few times before going to the 290 // correct flavor of top or bottom. 291 // 292 // This call only needs to be made once as the data flows around any 293 // given cycle. We do it at Phis, and nowhere else. 294 // The types presented are the new type of a phi (computed by PhiNode::Value) 295 // and the previously computed type, last time the phi was visited. 296 // 297 // The third argument is upper limit for the saturated value, 298 // if the phase wishes to widen the new_type. 299 // If the phase is narrowing, the old type provides a lower limit. 300 // Caller guarantees that old_type and new_type are no higher than limit_type. 301 virtual const Type* saturate(const Type* new_type, const Type* old_type, 302 const Type* limit_type) const 303 { ShouldNotCallThis(); return NULL; } 304 305#ifndef PRODUCT 306 void dump_old2new_map() const; 307 void dump_new( uint new_lidx ) const; 308 void dump_types() const; 309 void dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl = true); 310 void dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited); 311 312 uint _count_progress; // For profiling, count transforms that make progress 313 void set_progress() { ++_count_progress; assert( allow_progress(),"No progress allowed during verification"); } 314 void clear_progress() { _count_progress = 0; } 315 uint made_progress() const { return _count_progress; } 316 317 uint _count_transforms; // For profiling, count transforms performed 318 void set_transforms() { ++_count_transforms; } 319 void clear_transforms() { _count_transforms = 0; } 320 uint made_transforms() const{ return _count_transforms; } 321 322 bool _allow_progress; // progress not allowed during verification pass 323 void set_allow_progress(bool allow) { _allow_progress = allow; } 324 bool allow_progress() { return _allow_progress; } 325#endif 326}; 327 328//------------------------------PhaseValues------------------------------------ 329// Phase infrastructure to support values 330class PhaseValues : public PhaseTransform { 331protected: 332 NodeHash _table; // Hash table for value-numbering 333 334public: 335 PhaseValues( Arena *arena, uint est_max_size ); 336 PhaseValues( PhaseValues *pt ); 337 PhaseValues( PhaseValues *ptv, const char *dummy ); 338 NOT_PRODUCT( ~PhaseValues(); ) 339 virtual PhaseIterGVN *is_IterGVN() { return 0; } 340 341 // Some Ideal and other transforms delete --> modify --> insert values 342 bool hash_delete(Node *n) { return _table.hash_delete(n); } 343 void hash_insert(Node *n) { _table.hash_insert(n); } 344 Node *hash_find_insert(Node *n){ return _table.hash_find_insert(n); } 345 Node *hash_find(const Node *n) { return _table.hash_find(n); } 346 347 // Used after parsing to eliminate values that are no longer in program 348 void remove_useless_nodes(VectorSet &useful) { 349 _table.remove_useless_nodes(useful); 350 // this may invalidate cached cons so reset the cache 351 init_con_caches(); 352 } 353 354 virtual ConNode* uncached_makecon(const Type* t); // override from PhaseTransform 355 356 virtual const Type* saturate(const Type* new_type, const Type* old_type, 357 const Type* limit_type) const 358 { return new_type; } 359 360#ifndef PRODUCT 361 uint _count_new_values; // For profiling, count new values produced 362 void inc_new_values() { ++_count_new_values; } 363 void clear_new_values() { _count_new_values = 0; } 364 uint made_new_values() const { return _count_new_values; } 365#endif 366}; 367 368 369//------------------------------PhaseGVN--------------------------------------- 370// Phase for performing local, pessimistic GVN-style optimizations. 371class PhaseGVN : public PhaseValues { 372public: 373 PhaseGVN( Arena *arena, uint est_max_size ) : PhaseValues( arena, est_max_size ) {} 374 PhaseGVN( PhaseGVN *gvn ) : PhaseValues( gvn ) {} 375 PhaseGVN( PhaseGVN *gvn, const char *dummy ) : PhaseValues( gvn, dummy ) {} 376 377 // Return a node which computes the same function as this node, but 378 // in a faster or cheaper fashion. 379 Node *transform( Node *n ); 380 Node *transform_no_reclaim( Node *n ); 381 382 // Check for a simple dead loop when a data node references itself. 383 DEBUG_ONLY(void dead_loop_check(Node *n);) 384}; 385 386//------------------------------PhaseIterGVN----------------------------------- 387// Phase for iteratively performing local, pessimistic GVN-style optimizations. 388// and ideal transformations on the graph. 389class PhaseIterGVN : public PhaseGVN { 390 private: 391 bool _delay_transform; // When true simply register the node when calling transform 392 // instead of actually optimizing it 393 394 // Idealize old Node 'n' with respect to its inputs and its value 395 virtual Node *transform_old( Node *a_node ); 396protected: 397 398 // Idealize new Node 'n' with respect to its inputs and its value 399 virtual Node *transform( Node *a_node ); 400 401 // Warm up hash table, type table and initial worklist 402 void init_worklist( Node *a_root ); 403 404 virtual const Type* saturate(const Type* new_type, const Type* old_type, 405 const Type* limit_type) const; 406 // Usually returns new_type. Returns old_type if new_type is only a slight 407 // improvement, such that it would take many (>>10) steps to reach 2**32. 408 409public: 410 PhaseIterGVN( PhaseIterGVN *igvn ); // Used by CCP constructor 411 PhaseIterGVN( PhaseGVN *gvn ); // Used after Parser 412 PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ); // Used after +VerifyOpto 413 414 virtual PhaseIterGVN *is_IterGVN() { return this; } 415 416 Unique_Node_List _worklist; // Iterative worklist 417 418 // Given def-use info and an initial worklist, apply Node::Ideal, 419 // Node::Value, Node::Identity, hash-based value numbering, Node::Ideal_DU 420 // and dominator info to a fixed point. 421 void optimize(); 422 423 // Register a new node with the iter GVN pass without transforming it. 424 // Used when we need to restructure a Region/Phi area and all the Regions 425 // and Phis need to complete this one big transform before any other 426 // transforms can be triggered on the region. 427 // Optional 'orig' is an earlier version of this node. 428 // It is significant only for debugging and profiling. 429 Node* register_new_node_with_optimizer(Node* n, Node* orig = NULL); 430 431 // Kill a globally dead Node. It is allowed to have uses which are 432 // assumed dead and left 'in limbo'. 433 void remove_globally_dead_node( Node *dead ); 434 435 // Kill all inputs to a dead node, recursively making more dead nodes. 436 // The Node must be dead locally, i.e., have no uses. 437 void remove_dead_node( Node *dead ) { 438 assert(dead->outcnt() == 0 && !dead->is_top(), "node must be dead"); 439 remove_globally_dead_node(dead); 440 } 441 442 // Subsume users of node 'old' into node 'nn' 443 // If no Def-Use info existed for 'nn' it will after call. 444 void subsume_node( Node *old, Node *nn ); 445 446 // Add users of 'n' to worklist 447 void add_users_to_worklist0( Node *n ); 448 void add_users_to_worklist ( Node *n ); 449 450 // Replace old node with new one. 451 void replace_node( Node *old, Node *nn ) { 452 add_users_to_worklist(old); 453 hash_delete(old); 454 subsume_node(old, nn); 455 } 456 457 bool delay_transform() const { return _delay_transform; } 458 459 void set_delay_transform(bool delay) { 460 _delay_transform = delay; 461 } 462 463#ifndef PRODUCT 464protected: 465 // Sub-quadratic implementation of VerifyIterativeGVN. 466 unsigned long _verify_counter; 467 unsigned long _verify_full_passes; 468 enum { _verify_window_size = 30 }; 469 Node* _verify_window[_verify_window_size]; 470 void verify_step(Node* n); 471#endif 472}; 473 474//------------------------------PhaseCCP--------------------------------------- 475// Phase for performing global Conditional Constant Propagation. 476// Should be replaced with combined CCP & GVN someday. 477class PhaseCCP : public PhaseIterGVN { 478 // Non-recursive. Use analysis to transform single Node. 479 virtual Node *transform_once( Node *n ); 480 481public: 482 PhaseCCP( PhaseIterGVN *igvn ); // Compute conditional constants 483 NOT_PRODUCT( ~PhaseCCP(); ) 484 485 // Worklist algorithm identifies constants 486 void analyze(); 487 // Recursive traversal of program. Used analysis to modify program. 488 virtual Node *transform( Node *n ); 489 // Do any transformation after analysis 490 void do_transform(); 491 492 virtual const Type* saturate(const Type* new_type, const Type* old_type, 493 const Type* limit_type) const; 494 // Returns new_type->widen(old_type), which increments the widen bits until 495 // giving up with TypeInt::INT or TypeLong::LONG. 496 // Result is clipped to limit_type if necessary. 497 498#ifndef PRODUCT 499 static uint _total_invokes; // For profiling, count invocations 500 void inc_invokes() { ++PhaseCCP::_total_invokes; } 501 502 static uint _total_constants; // For profiling, count constants found 503 uint _count_constants; 504 void clear_constants() { _count_constants = 0; } 505 void inc_constants() { ++_count_constants; } 506 uint count_constants() const { return _count_constants; } 507 508 static void print_statistics(); 509#endif 510}; 511 512 513//------------------------------PhasePeephole---------------------------------- 514// Phase for performing peephole optimizations on register allocated basic blocks. 515class PhasePeephole : public PhaseTransform { 516 PhaseRegAlloc *_regalloc; 517 PhaseCFG &_cfg; 518 // Recursive traversal of program. Pure function is unused in this phase 519 virtual Node *transform( Node *n ); 520 521public: 522 PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ); 523 NOT_PRODUCT( ~PhasePeephole(); ) 524 525 // Do any transformation after analysis 526 void do_transform(); 527 528#ifndef PRODUCT 529 static uint _total_peepholes; // For profiling, count peephole rules applied 530 uint _count_peepholes; 531 void clear_peepholes() { _count_peepholes = 0; } 532 void inc_peepholes() { ++_count_peepholes; } 533 uint count_peepholes() const { return _count_peepholes; } 534 535 static void print_statistics(); 536#endif 537}; 538