1 /* 2 * Copyright (c) 1998, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP 26 #define SHARE_VM_OPTO_LOOPNODE_HPP 27 28 #include "opto/cfgnode.hpp" 29 #include "opto/multnode.hpp" 30 #include "opto/phaseX.hpp" 31 #include "opto/subnode.hpp" 32 #include "opto/type.hpp" 33 34 class CmpNode; 35 class CountedLoopEndNode; 36 class CountedLoopNode; 37 class IdealLoopTree; 38 class LoopNode; 39 class Node; 40 class PhaseIdealLoop; 41 class VectorSet; 42 class Invariance; 43 struct small_cache; 44 45 // 46 // I D E A L I Z E D L O O P S 47 // 48 // Idealized loops are the set of loops I perform more interesting 49 // transformations on, beyond simple hoisting. 50 51 //------------------------------LoopNode--------------------------------------- 52 // Simple loop header. Fall in path on left, loop-back path on right. 53 class LoopNode : public RegionNode { 54 // Size is bigger to hold the flags. However, the flags do not change 55 // the semantics so it does not appear in the hash & cmp functions. 56 virtual uint size_of() const { return sizeof(*this); } 57 protected: 58 short _loop_flags; 59 // Names for flag bitfields 60 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 61 MainHasNoPreLoop=4, 62 HasExactTripCount=8, 63 InnerLoop=16, 64 PartialPeelLoop=32, 65 PartialPeelFailed=64 }; 66 char _unswitch_count; 67 enum { _unswitch_max=3 }; 68 69 public: 70 // Names for edge indices 71 enum { Self=0, EntryControl, LoopBackControl }; 72 73 int is_inner_loop() const { return _loop_flags & InnerLoop; } 74 void set_inner_loop() { _loop_flags |= InnerLoop; } 75 76 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 77 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 78 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 79 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 80 81 int unswitch_max() { return _unswitch_max; } 82 int unswitch_count() { return _unswitch_count; } 83 void set_unswitch_count(int val) { 84 assert (val <= unswitch_max(), "too many unswitches"); 85 _unswitch_count = val; 86 } 87 88 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { 89 init_class_id(Class_Loop); 90 init_req(EntryControl, entry); 91 init_req(LoopBackControl, backedge); 92 } 93 94 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 95 virtual int Opcode() const; 96 bool can_be_counted_loop(PhaseTransform* phase) const { 97 return req() == 3 && in(0) != NULL && 98 in(1) != NULL && phase->type(in(1)) != Type::TOP && 99 in(2) != NULL && phase->type(in(2)) != Type::TOP; 100 } 101 bool is_valid_counted_loop() const; 102 #ifndef PRODUCT 103 virtual void dump_spec(outputStream *st) const; 104 #endif 105 }; 106 107 //------------------------------Counted Loops---------------------------------- 108 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 109 // path (and maybe some other exit paths). The trip-counter exit is always 110 // last in the loop. The trip-counter have to stride by a constant; 111 // the exit value is also loop invariant. 112 113 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 114 // CountedLoopNode has the incoming loop control and the loop-back-control 115 // which is always the IfTrue before the matching CountedLoopEndNode. The 116 // CountedLoopEndNode has an incoming control (possibly not the 117 // CountedLoopNode if there is control flow in the loop), the post-increment 118 // trip-counter value, and the limit. The trip-counter value is always of 119 // the form (Op old-trip-counter stride). The old-trip-counter is produced 120 // by a Phi connected to the CountedLoopNode. The stride is constant. 121 // The Op is any commutable opcode, including Add, Mul, Xor. The 122 // CountedLoopEndNode also takes in the loop-invariant limit value. 123 124 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 125 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 126 // via the old-trip-counter from the Op node. 127 128 //------------------------------CountedLoopNode-------------------------------- 129 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 130 // inputs the incoming loop-start control and the loop-back control, so they 131 // act like RegionNodes. They also take in the initial trip counter, the 132 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 133 // produce a loop-body control and the trip counter value. Since 134 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 135 136 class CountedLoopNode : public LoopNode { 137 // Size is bigger to hold _main_idx. However, _main_idx does not change 138 // the semantics so it does not appear in the hash & cmp functions. 139 virtual uint size_of() const { return sizeof(*this); } 140 141 // For Pre- and Post-loops during debugging ONLY, this holds the index of 142 // the Main CountedLoop. Used to assert that we understand the graph shape. 143 node_idx_t _main_idx; 144 145 // Known trip count calculated by compute_exact_trip_count() 146 uint _trip_count; 147 148 // Expected trip count from profile data 149 float _profile_trip_cnt; 150 151 // Log2 of original loop bodies in unrolled loop 152 int _unrolled_count_log2; 153 154 // Node count prior to last unrolling - used to decide if 155 // unroll,optimize,unroll,optimize,... is making progress 156 int _node_count_before_unroll; 157 158 public: 159 CountedLoopNode( Node *entry, Node *backedge ) 160 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 161 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), 162 _node_count_before_unroll(0) { 163 init_class_id(Class_CountedLoop); 164 // Initialize _trip_count to the largest possible value. 165 // Will be reset (lower) if the loop's trip count is known. 166 } 167 168 virtual int Opcode() const; 169 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 170 171 Node *init_control() const { return in(EntryControl); } 172 Node *back_control() const { return in(LoopBackControl); } 173 CountedLoopEndNode *loopexit() const; 174 Node *init_trip() const; 175 Node *stride() const; 176 int stride_con() const; 177 bool stride_is_con() const; 178 Node *limit() const; 179 Node *incr() const; 180 Node *phi() const; 181 182 // Match increment with optional truncation 183 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); 184 185 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 186 // can run short a few iterations and may start a few iterations in. 187 // It will be RCE'd and unrolled and aligned. 188 189 // A following 'post' loop will run any remaining iterations. Used 190 // during Range Check Elimination, the 'post' loop will do any final 191 // iterations with full checks. Also used by Loop Unrolling, where 192 // the 'post' loop will do any epilog iterations needed. Basically, 193 // a 'post' loop can not profitably be further unrolled or RCE'd. 194 195 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 196 // it may do under-flow checks for RCE and may do alignment iterations 197 // so the following main loop 'knows' that it is striding down cache 198 // lines. 199 200 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 201 // Aligned, may be missing it's pre-loop. 202 int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 203 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 204 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 205 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 206 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 207 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 208 209 int main_idx() const { return _main_idx; } 210 211 212 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 213 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 214 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 215 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 216 217 void set_trip_count(uint tc) { _trip_count = tc; } 218 uint trip_count() { return _trip_count; } 219 220 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 221 void set_exact_trip_count(uint tc) { 222 _trip_count = tc; 223 _loop_flags |= HasExactTripCount; 224 } 225 void set_nonexact_trip_count() { 226 _loop_flags &= ~HasExactTripCount; 227 } 228 229 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 230 float profile_trip_cnt() { return _profile_trip_cnt; } 231 232 void double_unrolled_count() { _unrolled_count_log2++; } 233 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 234 235 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 236 int node_count_before_unroll() { return _node_count_before_unroll; } 237 238 #ifndef PRODUCT 239 virtual void dump_spec(outputStream *st) const; 240 #endif 241 }; 242 243 //------------------------------CountedLoopEndNode----------------------------- 244 // CountedLoopEndNodes end simple trip counted loops. They act much like 245 // IfNodes. 246 class CountedLoopEndNode : public IfNode { 247 public: 248 enum { TestControl, TestValue }; 249 250 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) 251 : IfNode( control, test, prob, cnt) { 252 init_class_id(Class_CountedLoopEnd); 253 } 254 virtual int Opcode() const; 255 256 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 257 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 258 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 259 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 260 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 261 int stride_con() const; 262 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } 263 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 264 PhiNode *phi() const { 265 Node *tmp = incr(); 266 if (tmp && tmp->req() == 3) { 267 Node* phi = tmp->in(1); 268 if (phi->is_Phi()) { 269 return phi->as_Phi(); 270 } 271 } 272 return NULL; 273 } 274 CountedLoopNode *loopnode() const { 275 // The CountedLoopNode that goes with this CountedLoopEndNode may 276 // have been optimized out by the IGVN so be cautious with the 277 // pattern matching on the graph 278 PhiNode* iv_phi = phi(); 279 if (iv_phi == NULL) { 280 return NULL; 281 } 282 Node *ln = iv_phi->in(0); 283 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) { 284 return (CountedLoopNode*)ln; 285 } 286 return NULL; 287 } 288 289 #ifndef PRODUCT 290 virtual void dump_spec(outputStream *st) const; 291 #endif 292 }; 293 294 295 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { 296 Node *bc = back_control(); 297 if( bc == NULL ) return NULL; 298 Node *le = bc->in(0); 299 if( le->Opcode() != Op_CountedLoopEnd ) 300 return NULL; 301 return (CountedLoopEndNode*)le; 302 } 303 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } 304 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } 305 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } 306 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } 307 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } 308 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } 309 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } 310 311 //------------------------------LoopLimitNode----------------------------- 312 // Counted Loop limit node which represents exact final iterator value: 313 // trip_count = (limit - init_trip + stride - 1)/stride 314 // final_value= trip_count * stride + init_trip. 315 // Use HW instructions to calculate it when it can overflow in integer. 316 // Note, final_value should fit into integer since counted loop has 317 // limit check: limit <= max_int-stride. 318 class LoopLimitNode : public Node { 319 enum { Init=1, Limit=2, Stride=3 }; 320 public: 321 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 322 // Put it on the Macro nodes list to optimize during macro nodes expansion. 323 init_flags(Flag_is_macro); 324 C->add_macro_node(this); 325 } 326 virtual int Opcode() const; 327 virtual const Type *bottom_type() const { return TypeInt::INT; } 328 virtual uint ideal_reg() const { return Op_RegI; } 329 virtual const Type *Value( PhaseTransform *phase ) const; 330 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 331 virtual Node *Identity( PhaseTransform *phase ); 332 }; 333 334 // -----------------------------IdealLoopTree---------------------------------- 335 class IdealLoopTree : public ResourceObj { 336 public: 337 IdealLoopTree *_parent; // Parent in loop tree 338 IdealLoopTree *_next; // Next sibling in loop tree 339 IdealLoopTree *_child; // First child in loop tree 340 341 // The head-tail backedge defines the loop. 342 // If tail is NULL then this loop has multiple backedges as part of the 343 // same loop. During cleanup I'll peel off the multiple backedges; merge 344 // them at the loop bottom and flow 1 real backedge into the loop. 345 Node *_head; // Head of loop 346 Node *_tail; // Tail of loop 347 inline Node *tail(); // Handle lazy update of _tail field 348 PhaseIdealLoop* _phase; 349 350 Node_List _body; // Loop body for inner loops 351 352 uint8 _nest; // Nesting depth 353 uint8 _irreducible:1, // True if irreducible 354 _has_call:1, // True if has call safepoint 355 _has_sfpt:1, // True if has non-call safepoint 356 _rce_candidate:1; // True if candidate for range check elimination 357 358 Node_List* _safepts; // List of safepoints in this loop 359 Node_List* _required_safept; // A inner loop cannot delete these safepts; 360 bool _allow_optimizations; // Allow loop optimizations 361 362 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 363 : _parent(0), _next(0), _child(0), 364 _head(head), _tail(tail), 365 _phase(phase), 366 _safepts(NULL), 367 _required_safept(NULL), 368 _allow_optimizations(true), 369 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) 370 { } 371 372 // Is 'l' a member of 'this'? 373 int is_member( const IdealLoopTree *l ) const; // Test for nested membership 374 375 // Set loop nesting depth. Accumulate has_call bits. 376 int set_nest( uint depth ); 377 378 // Split out multiple fall-in edges from the loop header. Move them to a 379 // private RegionNode before the loop. This becomes the loop landing pad. 380 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 381 382 // Split out the outermost loop from this shared header. 383 void split_outer_loop( PhaseIdealLoop *phase ); 384 385 // Merge all the backedges from the shared header into a private Region. 386 // Feed that region as the one backedge to this loop. 387 void merge_many_backedges( PhaseIdealLoop *phase ); 388 389 // Split shared headers and insert loop landing pads. 390 // Insert a LoopNode to replace the RegionNode. 391 // Returns TRUE if loop tree is structurally changed. 392 bool beautify_loops( PhaseIdealLoop *phase ); 393 394 // Perform optimization to use the loop predicates for null checks and range checks. 395 // Applies to any loop level (not just the innermost one) 396 bool loop_predication( PhaseIdealLoop *phase); 397 398 // Perform iteration-splitting on inner loops. Split iterations to 399 // avoid range checks or one-shot null checks. Returns false if the 400 // current round of loop opts should stop. 401 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 402 403 // Driver for various flavors of iteration splitting. Returns false 404 // if the current round of loop opts should stop. 405 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 406 407 // Given dominators, try to find loops with calls that must always be 408 // executed (call dominates loop tail). These loops do not need non-call 409 // safepoints (ncsfpt). 410 void check_safepts(VectorSet &visited, Node_List &stack); 411 412 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 413 // encountered. 414 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 415 416 // Remove safepoints from loop. Optionally keeping one. 417 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 418 419 // Convert to counted loops where possible 420 void counted_loop( PhaseIdealLoop *phase ); 421 422 // Check for Node being a loop-breaking test 423 Node *is_loop_exit(Node *iff) const; 424 425 // Returns true if ctrl is executed on every complete iteration 426 bool dominates_backedge(Node* ctrl); 427 428 // Remove simplistic dead code from loop body 429 void DCE_loop_body(); 430 431 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 432 // Replace with a 1-in-10 exit guess. 433 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 434 435 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 436 // Useful for unrolling loops with NO array accesses. 437 bool policy_peel_only( PhaseIdealLoop *phase ) const; 438 439 // Return TRUE or FALSE if the loop should be unswitched -- clone 440 // loop with an invariant test 441 bool policy_unswitching( PhaseIdealLoop *phase ) const; 442 443 // Micro-benchmark spamming. Remove empty loops. 444 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); 445 446 // Convert one iteration loop into normal code. 447 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase ); 448 449 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 450 // make some loop-invariant test (usually a null-check) happen before the 451 // loop. 452 bool policy_peeling( PhaseIdealLoop *phase ) const; 453 454 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 455 // known trip count in the counted loop node. 456 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; 457 458 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 459 // the loop is a CountedLoop and the body is small enough. 460 bool policy_unroll( PhaseIdealLoop *phase ) const; 461 462 // Return TRUE or FALSE if the loop should be range-check-eliminated. 463 // Gather a list of IF tests that are dominated by iteration splitting; 464 // also gather the end of the first split and the start of the 2nd split. 465 bool policy_range_check( PhaseIdealLoop *phase ) const; 466 467 // Return TRUE or FALSE if the loop should be cache-line aligned. 468 // Gather the expression that does the alignment. Note that only 469 // one array base can be aligned in a loop (unless the VM guarantees 470 // mutual alignment). Note that if we vectorize short memory ops 471 // into longer memory ops, we may want to increase alignment. 472 bool policy_align( PhaseIdealLoop *phase ) const; 473 474 // Return TRUE if "iff" is a range check. 475 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; 476 477 // Compute loop exact trip count if possible 478 void compute_exact_trip_count( PhaseIdealLoop *phase ); 479 480 // Compute loop trip count from profile data 481 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 482 483 // Reassociate invariant expressions. 484 void reassociate_invariants(PhaseIdealLoop *phase); 485 // Reassociate invariant add and subtract expressions. 486 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); 487 // Return nonzero index of invariant operand if invariant and variant 488 // are combined with an Add or Sub. Helper for reassociate_invariants. 489 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); 490 491 // Return true if n is invariant 492 bool is_invariant(Node* n) const; 493 494 // Put loop body on igvn work list 495 void record_for_igvn(); 496 497 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } 498 bool is_inner() { return is_loop() && _child == NULL; } 499 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } 500 501 #ifndef PRODUCT 502 void dump_head( ) const; // Dump loop head only 503 void dump() const; // Dump this loop recursively 504 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 505 #endif 506 507 }; 508 509 // -----------------------------PhaseIdealLoop--------------------------------- 510 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a 511 // loop tree. Drives the loop-based transformations on the ideal graph. 512 class PhaseIdealLoop : public PhaseTransform { 513 friend class IdealLoopTree; 514 friend class SuperWord; 515 // Pre-computed def-use info 516 PhaseIterGVN &_igvn; 517 518 // Head of loop tree 519 IdealLoopTree *_ltree_root; 520 521 // Array of pre-order numbers, plus post-visited bit. 522 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 523 // ODD for post-visited. Other bits are the pre-order number. 524 uint *_preorders; 525 uint _max_preorder; 526 527 const PhaseIdealLoop* _verify_me; 528 bool _verify_only; 529 530 // Allocate _preorders[] array 531 void allocate_preorders() { 532 _max_preorder = C->unique()+8; 533 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 534 memset(_preorders, 0, sizeof(uint) * _max_preorder); 535 } 536 537 // Allocate _preorders[] array 538 void reallocate_preorders() { 539 if ( _max_preorder < C->unique() ) { 540 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 541 _max_preorder = C->unique(); 542 } 543 memset(_preorders, 0, sizeof(uint) * _max_preorder); 544 } 545 546 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 547 // adds new nodes. 548 void check_grow_preorders( ) { 549 if ( _max_preorder < C->unique() ) { 550 uint newsize = _max_preorder<<1; // double size of array 551 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 552 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 553 _max_preorder = newsize; 554 } 555 } 556 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 557 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 558 // Pre-order numbers are written to the Nodes array as low-bit-set values. 559 void set_preorder_visited( Node *n, int pre_order ) { 560 assert( !is_visited( n ), "already set" ); 561 _preorders[n->_idx] = (pre_order<<1); 562 }; 563 // Return pre-order number. 564 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 565 566 // Check for being post-visited. 567 // Should be previsited already (checked with assert(is_visited(n))). 568 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 569 570 // Mark as post visited 571 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 572 573 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 574 // Returns true if "n" is a data node, false if it's a control node. 575 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 576 577 // clear out dead code after build_loop_late 578 Node_List _deadlist; 579 580 // Support for faster execution of get_late_ctrl()/dom_lca() 581 // when a node has many uses and dominator depth is deep. 582 Node_Array _dom_lca_tags; 583 void init_dom_lca_tags(); 584 void clear_dom_lca_tags(); 585 586 // Helper for debugging bad dominance relationships 587 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 588 589 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 590 591 // Inline wrapper for frequent cases: 592 // 1) only one use 593 // 2) a use is the same as the current LCA passed as 'n1' 594 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 595 assert( n->is_CFG(), "" ); 596 // Fast-path NULL lca 597 if( lca != NULL && lca != n ) { 598 assert( lca->is_CFG(), "" ); 599 // find LCA of all uses 600 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 601 } 602 return find_non_split_ctrl(n); 603 } 604 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 605 606 // Helper function for directing control inputs away from CFG split 607 // points. 608 Node *find_non_split_ctrl( Node *ctrl ) const { 609 if (ctrl != NULL) { 610 if (ctrl->is_MultiBranch()) { 611 ctrl = ctrl->in(0); 612 } 613 assert(ctrl->is_CFG(), "CFG"); 614 } 615 return ctrl; 616 } 617 618 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 619 620 public: 621 bool has_node( Node* n ) const { 622 guarantee(n != NULL, "No Node."); 623 return _nodes[n->_idx] != NULL; 624 } 625 // check if transform created new nodes that need _ctrl recorded 626 Node *get_late_ctrl( Node *n, Node *early ); 627 Node *get_early_ctrl( Node *n ); 628 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 629 void set_early_ctrl( Node *n ); 630 void set_subtree_ctrl( Node *root ); 631 void set_ctrl( Node *n, Node *ctrl ) { 632 assert( !has_node(n) || has_ctrl(n), "" ); 633 assert( ctrl->in(0), "cannot set dead control node" ); 634 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 635 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 636 } 637 // Set control and update loop membership 638 void set_ctrl_and_loop(Node* n, Node* ctrl) { 639 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 640 IdealLoopTree* new_loop = get_loop(ctrl); 641 if (old_loop != new_loop) { 642 if (old_loop->_child == NULL) old_loop->_body.yank(n); 643 if (new_loop->_child == NULL) new_loop->_body.push(n); 644 } 645 set_ctrl(n, ctrl); 646 } 647 // Control nodes can be replaced or subsumed. During this pass they 648 // get their replacement Node in slot 1. Instead of updating the block 649 // location of all Nodes in the subsumed block, we lazily do it. As we 650 // pull such a subsumed block out of the array, we write back the final 651 // correct block. 652 Node *get_ctrl( Node *i ) { 653 assert(has_node(i), ""); 654 Node *n = get_ctrl_no_update(i); 655 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 656 assert(has_node(i) && has_ctrl(i), ""); 657 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 658 return n; 659 } 660 // true if CFG node d dominates CFG node n 661 bool is_dominator(Node *d, Node *n); 662 // return get_ctrl for a data node and self(n) for a CFG node 663 Node* ctrl_or_self(Node* n) { 664 if (has_ctrl(n)) 665 return get_ctrl(n); 666 else { 667 assert (n->is_CFG(), "must be a CFG node"); 668 return n; 669 } 670 } 671 672 private: 673 Node *get_ctrl_no_update_helper(Node *i) const { 674 assert(has_ctrl(i), "should be control, not loop"); 675 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 676 } 677 678 Node *get_ctrl_no_update(Node *i) const { 679 assert( has_ctrl(i), "" ); 680 Node *n = get_ctrl_no_update_helper(i); 681 if (!n->in(0)) { 682 // Skip dead CFG nodes 683 do { 684 n = get_ctrl_no_update_helper(n); 685 } while (!n->in(0)); 686 n = find_non_split_ctrl(n); 687 } 688 return n; 689 } 690 691 // Check for loop being set 692 // "n" must be a control node. Returns true if "n" is known to be in a loop. 693 bool has_loop( Node *n ) const { 694 assert(!has_node(n) || !has_ctrl(n), ""); 695 return has_node(n); 696 } 697 // Set loop 698 void set_loop( Node *n, IdealLoopTree *loop ) { 699 _nodes.map(n->_idx, (Node*)loop); 700 } 701 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 702 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 703 // from old_node to new_node to support the lazy update. Reference 704 // replaces loop reference, since that is not needed for dead node. 705 public: 706 void lazy_update(Node *old_node, Node *new_node) { 707 assert(old_node != new_node, "no cycles please"); 708 // Re-use the side array slot for this node to provide the 709 // forwarding pointer. 710 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 711 } 712 void lazy_replace(Node *old_node, Node *new_node) { 713 _igvn.replace_node(old_node, new_node); 714 lazy_update(old_node, new_node); 715 } 716 717 private: 718 719 // Place 'n' in some loop nest, where 'n' is a CFG node 720 void build_loop_tree(); 721 int build_loop_tree_impl( Node *n, int pre_order ); 722 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 723 // loop tree, not the root. 724 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 725 726 // Place Data nodes in some loop nest 727 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 728 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 729 void build_loop_late_post ( Node* n ); 730 731 // Array of immediate dominance info for each CFG node indexed by node idx 732 private: 733 uint _idom_size; 734 Node **_idom; // Array of immediate dominators 735 uint *_dom_depth; // Used for fast LCA test 736 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 737 738 Node* idom_no_update(Node* d) const { 739 assert(d->_idx < _idom_size, "oob"); 740 Node* n = _idom[d->_idx]; 741 assert(n != NULL,"Bad immediate dominator info."); 742 while (n->in(0) == NULL) { // Skip dead CFG nodes 743 //n = n->in(1); 744 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 745 assert(n != NULL,"Bad immediate dominator info."); 746 } 747 return n; 748 } 749 Node *idom(Node* d) const { 750 uint didx = d->_idx; 751 Node *n = idom_no_update(d); 752 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 753 return n; 754 } 755 uint dom_depth(Node* d) const { 756 guarantee(d != NULL, "Null dominator info."); 757 guarantee(d->_idx < _idom_size, ""); 758 return _dom_depth[d->_idx]; 759 } 760 void set_idom(Node* d, Node* n, uint dom_depth); 761 // Locally compute IDOM using dom_lca call 762 Node *compute_idom( Node *region ) const; 763 // Recompute dom_depth 764 void recompute_dom_depth(); 765 766 // Is safept not required by an outer loop? 767 bool is_deleteable_safept(Node* sfpt); 768 769 // Replace parallel induction variable (parallel to trip counter) 770 void replace_parallel_iv(IdealLoopTree *loop); 771 772 // Perform verification that the graph is valid. 773 PhaseIdealLoop( PhaseIterGVN &igvn) : 774 PhaseTransform(Ideal_Loop), 775 _igvn(igvn), 776 _dom_lca_tags(arena()), // Thread::resource_area 777 _verify_me(NULL), 778 _verify_only(true) { 779 build_and_optimize(false, false); 780 } 781 782 // build the loop tree and perform any requested optimizations 783 void build_and_optimize(bool do_split_if, bool skip_loop_opts); 784 785 public: 786 // Dominators for the sea of nodes 787 void Dominators(); 788 Node *dom_lca( Node *n1, Node *n2 ) const { 789 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 790 } 791 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 792 793 // Compute the Ideal Node to Loop mapping 794 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) : 795 PhaseTransform(Ideal_Loop), 796 _igvn(igvn), 797 _dom_lca_tags(arena()), // Thread::resource_area 798 _verify_me(NULL), 799 _verify_only(false) { 800 build_and_optimize(do_split_ifs, skip_loop_opts); 801 } 802 803 // Verify that verify_me made the same decisions as a fresh run. 804 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : 805 PhaseTransform(Ideal_Loop), 806 _igvn(igvn), 807 _dom_lca_tags(arena()), // Thread::resource_area 808 _verify_me(verify_me), 809 _verify_only(false) { 810 build_and_optimize(false, false); 811 } 812 813 // Build and verify the loop tree without modifying the graph. This 814 // is useful to verify that all inputs properly dominate their uses. 815 static void verify(PhaseIterGVN& igvn) { 816 #ifdef ASSERT 817 PhaseIdealLoop v(igvn); 818 #endif 819 } 820 821 // True if the method has at least 1 irreducible loop 822 bool _has_irreducible_loops; 823 824 // Per-Node transform 825 virtual Node *transform( Node *a_node ) { return 0; } 826 827 bool is_counted_loop( Node *x, IdealLoopTree *loop ); 828 829 Node* exact_limit( IdealLoopTree *loop ); 830 831 // Return a post-walked LoopNode 832 IdealLoopTree *get_loop( Node *n ) const { 833 // Dead nodes have no loop, so return the top level loop instead 834 if (!has_node(n)) return _ltree_root; 835 assert(!has_ctrl(n), ""); 836 return (IdealLoopTree*)_nodes[n->_idx]; 837 } 838 839 // Is 'n' a (nested) member of 'loop'? 840 int is_member( const IdealLoopTree *loop, Node *n ) const { 841 return loop->is_member(get_loop(n)); } 842 843 // This is the basic building block of the loop optimizations. It clones an 844 // entire loop body. It makes an old_new loop body mapping; with this 845 // mapping you can find the new-loop equivalent to an old-loop node. All 846 // new-loop nodes are exactly equal to their old-loop counterparts, all 847 // edges are the same. All exits from the old-loop now have a RegionNode 848 // that merges the equivalent new-loop path. This is true even for the 849 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 850 // now come from (one or more) Phis that merge their new-loop equivalents. 851 // Parameter side_by_side_idom: 852 // When side_by_size_idom is NULL, the dominator tree is constructed for 853 // the clone loop to dominate the original. Used in construction of 854 // pre-main-post loop sequence. 855 // When nonnull, the clone and original are side-by-side, both are 856 // dominated by the passed in side_by_side_idom node. Used in 857 // construction of unswitched loops. 858 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 859 Node* side_by_side_idom = NULL); 860 861 // If we got the effect of peeling, either by actually peeling or by 862 // making a pre-loop which must execute at least once, we can remove 863 // all loop-invariant dominated tests in the main body. 864 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 865 866 // Generate code to do a loop peel for the given loop (and body). 867 // old_new is a temp array. 868 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 869 870 // Add pre and post loops around the given loop. These loops are used 871 // during RCE, unrolling and aligning loops. 872 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 873 // If Node n lives in the back_ctrl block, we clone a private version of n 874 // in preheader_ctrl block and return that, otherwise return n. 875 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 876 877 // Take steps to maximally unroll the loop. Peel any odd iterations, then 878 // unroll to do double iterations. The next round of major loop transforms 879 // will repeat till the doubled loop body does all remaining iterations in 1 880 // pass. 881 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 882 883 // Unroll the loop body one step - make each trip do 2 iterations. 884 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 885 886 // Return true if exp is a constant times an induction var 887 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); 888 889 // Return true if exp is a scaled induction var plus (or minus) constant 890 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); 891 892 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 893 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 894 Deoptimization::DeoptReason reason); 895 void register_control(Node* n, IdealLoopTree *loop, Node* pred); 896 897 // Clone loop predicates to cloned loops (peeled, unswitched) 898 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, 899 Deoptimization::DeoptReason reason, 900 PhaseIdealLoop* loop_phase, 901 PhaseIterGVN* igvn); 902 903 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, 904 bool clone_limit_check, 905 PhaseIdealLoop* loop_phase, 906 PhaseIterGVN* igvn); 907 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); 908 909 static Node* skip_loop_predicates(Node* entry); 910 911 // Find a good location to insert a predicate 912 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); 913 // Find a predicate 914 static Node* find_predicate(Node* entry); 915 // Construct a range check for a predicate if 916 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, 917 int scale, Node* offset, 918 Node* init, Node* limit, jint stride, 919 Node* range, bool upper, bool &overflow); 920 921 // Implementation of the loop predication to promote checks outside the loop 922 bool loop_predication_impl(IdealLoopTree *loop); 923 924 // Helper function to collect predicate for eliminating the useless ones 925 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); 926 void eliminate_useless_predicates(); 927 928 // Change the control input of expensive nodes to allow commoning by 929 // IGVN when it is guaranteed to not result in a more frequent 930 // execution of the expensive node. Return true if progress. 931 bool process_expensive_nodes(); 932 933 // Check whether node has become unreachable 934 bool is_node_unreachable(Node *n) const { 935 return !has_node(n) || n->is_unreachable(_igvn); 936 } 937 938 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 939 void do_range_check( IdealLoopTree *loop, Node_List &old_new ); 940 941 // Create a slow version of the loop by cloning the loop 942 // and inserting an if to select fast-slow versions. 943 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, 944 Node_List &old_new); 945 946 // Clone loop with an invariant test (that does not exit) and 947 // insert a clone of the test that selects which version to 948 // execute. 949 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 950 951 // Find candidate "if" for unswitching 952 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; 953 954 // Range Check Elimination uses this function! 955 // Constrain the main loop iterations so the affine function: 956 // low_limit <= scale_con * I + offset < upper_limit 957 // always holds true. That is, either increase the number of iterations in 958 // the pre-loop or the post-loop until the condition holds true in the main 959 // loop. Scale_con, offset and limit are all loop invariant. 960 void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); 961 // Helper function for add_constraint(). 962 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl ); 963 964 // Partially peel loop up through last_peel node. 965 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 966 967 // Create a scheduled list of nodes control dependent on ctrl set. 968 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 969 // Has a use in the vector set 970 bool has_use_in_set( Node* n, VectorSet& vset ); 971 // Has use internal to the vector set (ie. not in a phi at the loop head) 972 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 973 // clone "n" for uses that are outside of loop 974 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 975 // clone "n" for special uses that are in the not_peeled region 976 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 977 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 978 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 979 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 980 #ifdef ASSERT 981 // Validate the loop partition sets: peel and not_peel 982 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 983 // Ensure that uses outside of loop are of the right form 984 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 985 uint orig_exit_idx, uint clone_exit_idx); 986 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 987 #endif 988 989 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 990 int stride_of_possible_iv( Node* iff ); 991 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 992 // Return the (unique) control output node that's in the loop (if it exists.) 993 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 994 // Insert a signed compare loop exit cloned from an unsigned compare. 995 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 996 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 997 // Utility to register node "n" with PhaseIdealLoop 998 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 999 // Utility to create an if-projection 1000 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1001 // Force the iff control output to be the live_proj 1002 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1003 // Insert a region before an if projection 1004 RegionNode* insert_region_before_proj(ProjNode* proj); 1005 // Insert a new if before an if projection 1006 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1007 1008 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1009 // "Nearly" because all Nodes have been cloned from the original in the loop, 1010 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1011 // through the Phi recursively, and return a Bool. 1012 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); 1013 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); 1014 1015 1016 // Rework addressing expressions to get the most loop-invariant stuff 1017 // moved out. We'd like to do all associative operators, but it's especially 1018 // important (common) to do address expressions. 1019 Node *remix_address_expressions( Node *n ); 1020 1021 // Attempt to use a conditional move instead of a phi/branch 1022 Node *conditional_move( Node *n ); 1023 1024 // Reorganize offset computations to lower register pressure. 1025 // Mostly prevent loop-fallout uses of the pre-incremented trip counter 1026 // (which are then alive with the post-incremented trip counter 1027 // forcing an extra register move) 1028 void reorg_offsets( IdealLoopTree *loop ); 1029 1030 // Check for aggressive application of 'split-if' optimization, 1031 // using basic block level info. 1032 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); 1033 Node *split_if_with_blocks_pre ( Node *n ); 1034 void split_if_with_blocks_post( Node *n ); 1035 Node *has_local_phi_input( Node *n ); 1036 // Mark an IfNode as being dominated by a prior test, 1037 // without actually altering the CFG (and hence IDOM info). 1038 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); 1039 1040 // Split Node 'n' through merge point 1041 Node *split_thru_region( Node *n, Node *region ); 1042 // Split Node 'n' through merge point if there is enough win. 1043 Node *split_thru_phi( Node *n, Node *region, int policy ); 1044 // Found an If getting its condition-code input from a Phi in the 1045 // same block. Split thru the Region. 1046 void do_split_if( Node *iff ); 1047 1048 // Conversion of fill/copy patterns into intrisic versions 1049 bool do_intrinsify_fill(); 1050 bool intrinsify_fill(IdealLoopTree* lpt); 1051 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1052 Node*& shift, Node*& offset); 1053 1054 private: 1055 // Return a type based on condition control flow 1056 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1057 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 1058 // Helpers for filtered type 1059 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1060 1061 // Helper functions 1062 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1063 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1064 void handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ); 1065 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1066 void sink_use( Node *use, Node *post_loop ); 1067 Node *place_near_use( Node *useblock ) const; 1068 1069 bool _created_loop_node; 1070 public: 1071 void set_created_loop_node() { _created_loop_node = true; } 1072 bool created_loop_node() { return _created_loop_node; } 1073 void register_new_node( Node *n, Node *blk ); 1074 1075 #ifdef ASSERT 1076 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1077 #endif 1078 1079 #ifndef PRODUCT 1080 void dump( ) const; 1081 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; 1082 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; 1083 void verify() const; // Major slow :-) 1084 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; 1085 IdealLoopTree *get_loop_idx(Node* n) const { 1086 // Dead nodes have no loop, so return the top level loop instead 1087 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 1088 } 1089 // Print some stats 1090 static void print_statistics(); 1091 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1092 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1093 #endif 1094 }; 1095 1096 inline Node* IdealLoopTree::tail() { 1097 // Handle lazy update of _tail field 1098 Node *n = _tail; 1099 //while( !n->in(0) ) // Skip dead CFG nodes 1100 //n = n->in(1); 1101 if (n->in(0) == NULL) 1102 n = _phase->get_ctrl(n); 1103 _tail = n; 1104 return n; 1105 } 1106 1107 1108 // Iterate over the loop tree using a preorder, left-to-right traversal. 1109 // 1110 // Example that visits all counted loops from within PhaseIdealLoop 1111 // 1112 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1113 // IdealLoopTree* lpt = iter.current(); 1114 // if (!lpt->is_counted()) continue; 1115 // ... 1116 class LoopTreeIterator : public StackObj { 1117 private: 1118 IdealLoopTree* _root; 1119 IdealLoopTree* _curnt; 1120 1121 public: 1122 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1123 1124 bool done() { return _curnt == NULL; } // Finished iterating? 1125 1126 void next(); // Advance to next loop tree 1127 1128 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1129 }; 1130 1131 #endif // SHARE_VM_OPTO_LOOPNODE_HPP