1 /* 2 * Copyright (c) 1998, 2015, 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 CountedLoopReserveKit; 42 class VectorSet; 43 class Invariance; 44 struct small_cache; 45 46 // 47 // I D E A L I Z E D L O O P S 48 // 49 // Idealized loops are the set of loops I perform more interesting 50 // transformations on, beyond simple hoisting. 51 52 //------------------------------LoopNode--------------------------------------- 53 // Simple loop header. Fall in path on left, loop-back path on right. 54 class LoopNode : public RegionNode { 55 // Size is bigger to hold the flags. However, the flags do not change 56 // the semantics so it does not appear in the hash & cmp functions. 57 virtual uint size_of() const { return sizeof(*this); } 58 protected: 59 short _loop_flags; 60 // Names for flag bitfields 61 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 62 MainHasNoPreLoop=4, 63 HasExactTripCount=8, 64 InnerLoop=16, 65 PartialPeelLoop=32, 66 PartialPeelFailed=64, 67 HasReductions=128, 68 WasSlpAnalyzed=256, 69 PassedSlpAnalysis=512, 70 DoUnrollOnly=1024, 71 VectorizedLoop=2048, 72 HasAtomicPostLoop=4096 }; 73 char _unswitch_count; 74 enum { _unswitch_max=3 }; 75 76 public: 77 // Names for edge indices 78 enum { Self=0, EntryControl, LoopBackControl }; 79 80 int is_inner_loop() const { return _loop_flags & InnerLoop; } 81 void set_inner_loop() { _loop_flags |= InnerLoop; } 82 83 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 84 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 85 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 86 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 87 void mark_has_reductions() { _loop_flags |= HasReductions; } 88 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 89 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 90 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 91 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 92 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 93 94 int unswitch_max() { return _unswitch_max; } 95 int unswitch_count() { return _unswitch_count; } 96 void set_unswitch_count(int val) { 97 assert (val <= unswitch_max(), "too many unswitches"); 98 _unswitch_count = val; 99 } 100 101 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { 102 init_class_id(Class_Loop); 103 init_req(EntryControl, entry); 104 init_req(LoopBackControl, backedge); 105 } 106 107 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 108 virtual int Opcode() const; 109 bool can_be_counted_loop(PhaseTransform* phase) const { 110 return req() == 3 && in(0) != NULL && 111 in(1) != NULL && phase->type(in(1)) != Type::TOP && 112 in(2) != NULL && phase->type(in(2)) != Type::TOP; 113 } 114 bool is_valid_counted_loop() const; 115 #ifndef PRODUCT 116 virtual void dump_spec(outputStream *st) const; 117 #endif 118 }; 119 120 //------------------------------Counted Loops---------------------------------- 121 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 122 // path (and maybe some other exit paths). The trip-counter exit is always 123 // last in the loop. The trip-counter have to stride by a constant; 124 // the exit value is also loop invariant. 125 126 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 127 // CountedLoopNode has the incoming loop control and the loop-back-control 128 // which is always the IfTrue before the matching CountedLoopEndNode. The 129 // CountedLoopEndNode has an incoming control (possibly not the 130 // CountedLoopNode if there is control flow in the loop), the post-increment 131 // trip-counter value, and the limit. The trip-counter value is always of 132 // the form (Op old-trip-counter stride). The old-trip-counter is produced 133 // by a Phi connected to the CountedLoopNode. The stride is constant. 134 // The Op is any commutable opcode, including Add, Mul, Xor. The 135 // CountedLoopEndNode also takes in the loop-invariant limit value. 136 137 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 138 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 139 // via the old-trip-counter from the Op node. 140 141 //------------------------------CountedLoopNode-------------------------------- 142 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 143 // inputs the incoming loop-start control and the loop-back control, so they 144 // act like RegionNodes. They also take in the initial trip counter, the 145 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 146 // produce a loop-body control and the trip counter value. Since 147 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 148 149 class CountedLoopNode : public LoopNode { 150 // Size is bigger to hold _main_idx. However, _main_idx does not change 151 // the semantics so it does not appear in the hash & cmp functions. 152 virtual uint size_of() const { return sizeof(*this); } 153 154 // For Pre- and Post-loops during debugging ONLY, this holds the index of 155 // the Main CountedLoop. Used to assert that we understand the graph shape. 156 node_idx_t _main_idx; 157 158 // Known trip count calculated by compute_exact_trip_count() 159 uint _trip_count; 160 161 // Expected trip count from profile data 162 float _profile_trip_cnt; 163 164 // Log2 of original loop bodies in unrolled loop 165 int _unrolled_count_log2; 166 167 // Node count prior to last unrolling - used to decide if 168 // unroll,optimize,unroll,optimize,... is making progress 169 int _node_count_before_unroll; 170 171 // If slp analysis is performed we record the maximum 172 // vector mapped unroll factor here 173 int _slp_maximum_unroll_factor; 174 175 public: 176 CountedLoopNode( Node *entry, Node *backedge ) 177 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 178 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), 179 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) { 180 init_class_id(Class_CountedLoop); 181 // Initialize _trip_count to the largest possible value. 182 // Will be reset (lower) if the loop's trip count is known. 183 } 184 185 virtual int Opcode() const; 186 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 187 188 Node *init_control() const { return in(EntryControl); } 189 Node *back_control() const { return in(LoopBackControl); } 190 CountedLoopEndNode *loopexit() const; 191 Node *init_trip() const; 192 Node *stride() const; 193 int stride_con() const; 194 bool stride_is_con() const; 195 Node *limit() const; 196 Node *incr() const; 197 Node *phi() const; 198 199 // Match increment with optional truncation 200 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); 201 202 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 203 // can run short a few iterations and may start a few iterations in. 204 // It will be RCE'd and unrolled and aligned. 205 206 // A following 'post' loop will run any remaining iterations. Used 207 // during Range Check Elimination, the 'post' loop will do any final 208 // iterations with full checks. Also used by Loop Unrolling, where 209 // the 'post' loop will do any epilog iterations needed. Basically, 210 // a 'post' loop can not profitably be further unrolled or RCE'd. 211 212 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 213 // it may do under-flow checks for RCE and may do alignment iterations 214 // so the following main loop 'knows' that it is striding down cache 215 // lines. 216 217 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 218 // Aligned, may be missing it's pre-loop. 219 int is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 220 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 221 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 222 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 223 int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; } 224 int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 225 int has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 226 int do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 227 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 228 int is_vectorized_loop () const { return (_loop_flags & VectorizedLoop) == VectorizedLoop; } 229 int has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 230 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 231 232 int main_idx() const { return _main_idx; } 233 234 235 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 236 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 237 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 238 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 239 240 void set_trip_count(uint tc) { _trip_count = tc; } 241 uint trip_count() { return _trip_count; } 242 243 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 244 void set_exact_trip_count(uint tc) { 245 _trip_count = tc; 246 _loop_flags |= HasExactTripCount; 247 } 248 void set_nonexact_trip_count() { 249 _loop_flags &= ~HasExactTripCount; 250 } 251 void set_notpassed_slp() { 252 _loop_flags &= ~PassedSlpAnalysis; 253 } 254 255 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 256 float profile_trip_cnt() { return _profile_trip_cnt; } 257 258 void double_unrolled_count() { _unrolled_count_log2++; } 259 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 260 261 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 262 int node_count_before_unroll() { return _node_count_before_unroll; } 263 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 264 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 265 266 #ifndef PRODUCT 267 virtual void dump_spec(outputStream *st) const; 268 #endif 269 }; 270 271 //------------------------------CountedLoopEndNode----------------------------- 272 // CountedLoopEndNodes end simple trip counted loops. They act much like 273 // IfNodes. 274 class CountedLoopEndNode : public IfNode { 275 public: 276 enum { TestControl, TestValue }; 277 278 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) 279 : IfNode( control, test, prob, cnt) { 280 init_class_id(Class_CountedLoopEnd); 281 } 282 virtual int Opcode() const; 283 284 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 285 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 286 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 287 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 288 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 289 int stride_con() const; 290 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } 291 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 292 PhiNode *phi() const { 293 Node *tmp = incr(); 294 if (tmp && tmp->req() == 3) { 295 Node* phi = tmp->in(1); 296 if (phi->is_Phi()) { 297 return phi->as_Phi(); 298 } 299 } 300 return NULL; 301 } 302 CountedLoopNode *loopnode() const { 303 // The CountedLoopNode that goes with this CountedLoopEndNode may 304 // have been optimized out by the IGVN so be cautious with the 305 // pattern matching on the graph 306 PhiNode* iv_phi = phi(); 307 if (iv_phi == NULL) { 308 return NULL; 309 } 310 Node *ln = iv_phi->in(0); 311 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) { 312 return (CountedLoopNode*)ln; 313 } 314 return NULL; 315 } 316 317 #ifndef PRODUCT 318 virtual void dump_spec(outputStream *st) const; 319 #endif 320 }; 321 322 323 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { 324 Node *bc = back_control(); 325 if( bc == NULL ) return NULL; 326 Node *le = bc->in(0); 327 if( le->Opcode() != Op_CountedLoopEnd ) 328 return NULL; 329 return (CountedLoopEndNode*)le; 330 } 331 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } 332 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } 333 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } 334 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } 335 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } 336 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } 337 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } 338 339 //------------------------------LoopLimitNode----------------------------- 340 // Counted Loop limit node which represents exact final iterator value: 341 // trip_count = (limit - init_trip + stride - 1)/stride 342 // final_value= trip_count * stride + init_trip. 343 // Use HW instructions to calculate it when it can overflow in integer. 344 // Note, final_value should fit into integer since counted loop has 345 // limit check: limit <= max_int-stride. 346 class LoopLimitNode : public Node { 347 enum { Init=1, Limit=2, Stride=3 }; 348 public: 349 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 350 // Put it on the Macro nodes list to optimize during macro nodes expansion. 351 init_flags(Flag_is_macro); 352 C->add_macro_node(this); 353 } 354 virtual int Opcode() const; 355 virtual const Type *bottom_type() const { return TypeInt::INT; } 356 virtual uint ideal_reg() const { return Op_RegI; } 357 virtual const Type* Value(PhaseGVN* phase) const; 358 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 359 virtual Node* Identity(PhaseGVN* phase); 360 }; 361 362 // -----------------------------IdealLoopTree---------------------------------- 363 class IdealLoopTree : public ResourceObj { 364 public: 365 IdealLoopTree *_parent; // Parent in loop tree 366 IdealLoopTree *_next; // Next sibling in loop tree 367 IdealLoopTree *_child; // First child in loop tree 368 369 // The head-tail backedge defines the loop. 370 // If tail is NULL then this loop has multiple backedges as part of the 371 // same loop. During cleanup I'll peel off the multiple backedges; merge 372 // them at the loop bottom and flow 1 real backedge into the loop. 373 Node *_head; // Head of loop 374 Node *_tail; // Tail of loop 375 inline Node *tail(); // Handle lazy update of _tail field 376 PhaseIdealLoop* _phase; 377 int _local_loop_unroll_limit; 378 int _local_loop_unroll_factor; 379 380 Node_List _body; // Loop body for inner loops 381 382 uint8_t _nest; // Nesting depth 383 uint8_t _irreducible:1, // True if irreducible 384 _has_call:1, // True if has call safepoint 385 _has_sfpt:1, // True if has non-call safepoint 386 _rce_candidate:1; // True if candidate for range check elimination 387 388 Node_List* _safepts; // List of safepoints in this loop 389 Node_List* _required_safept; // A inner loop cannot delete these safepts; 390 bool _allow_optimizations; // Allow loop optimizations 391 392 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 393 : _parent(0), _next(0), _child(0), 394 _head(head), _tail(tail), 395 _phase(phase), 396 _safepts(NULL), 397 _required_safept(NULL), 398 _allow_optimizations(true), 399 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 400 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0) 401 { } 402 403 // Is 'l' a member of 'this'? 404 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 405 406 // Set loop nesting depth. Accumulate has_call bits. 407 int set_nest( uint depth ); 408 409 // Split out multiple fall-in edges from the loop header. Move them to a 410 // private RegionNode before the loop. This becomes the loop landing pad. 411 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 412 413 // Split out the outermost loop from this shared header. 414 void split_outer_loop( PhaseIdealLoop *phase ); 415 416 // Merge all the backedges from the shared header into a private Region. 417 // Feed that region as the one backedge to this loop. 418 void merge_many_backedges( PhaseIdealLoop *phase ); 419 420 // Split shared headers and insert loop landing pads. 421 // Insert a LoopNode to replace the RegionNode. 422 // Returns TRUE if loop tree is structurally changed. 423 bool beautify_loops( PhaseIdealLoop *phase ); 424 425 // Perform optimization to use the loop predicates for null checks and range checks. 426 // Applies to any loop level (not just the innermost one) 427 bool loop_predication( PhaseIdealLoop *phase); 428 429 // Perform iteration-splitting on inner loops. Split iterations to 430 // avoid range checks or one-shot null checks. Returns false if the 431 // current round of loop opts should stop. 432 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 433 434 // Driver for various flavors of iteration splitting. Returns false 435 // if the current round of loop opts should stop. 436 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 437 438 // Given dominators, try to find loops with calls that must always be 439 // executed (call dominates loop tail). These loops do not need non-call 440 // safepoints (ncsfpt). 441 void check_safepts(VectorSet &visited, Node_List &stack); 442 443 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 444 // encountered. 445 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 446 447 // Remove safepoints from loop. Optionally keeping one. 448 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 449 450 // Convert to counted loops where possible 451 void counted_loop( PhaseIdealLoop *phase ); 452 453 // Check for Node being a loop-breaking test 454 Node *is_loop_exit(Node *iff) const; 455 456 // Returns true if ctrl is executed on every complete iteration 457 bool dominates_backedge(Node* ctrl); 458 459 // Remove simplistic dead code from loop body 460 void DCE_loop_body(); 461 462 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 463 // Replace with a 1-in-10 exit guess. 464 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 465 466 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 467 // Useful for unrolling loops with NO array accesses. 468 bool policy_peel_only( PhaseIdealLoop *phase ) const; 469 470 // Return TRUE or FALSE if the loop should be unswitched -- clone 471 // loop with an invariant test 472 bool policy_unswitching( PhaseIdealLoop *phase ) const; 473 474 // Micro-benchmark spamming. Remove empty loops. 475 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); 476 477 // Convert one iteration loop into normal code. 478 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase ); 479 480 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 481 // make some loop-invariant test (usually a null-check) happen before the 482 // loop. 483 bool policy_peeling( PhaseIdealLoop *phase ) const; 484 485 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 486 // known trip count in the counted loop node. 487 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; 488 489 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 490 // the loop is a CountedLoop and the body is small enough. 491 bool policy_unroll(PhaseIdealLoop *phase); 492 493 // Loop analyses to map to a maximal superword unrolling for vectorization. 494 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 495 496 // Return TRUE or FALSE if the loop should be range-check-eliminated. 497 // Gather a list of IF tests that are dominated by iteration splitting; 498 // also gather the end of the first split and the start of the 2nd split. 499 bool policy_range_check( PhaseIdealLoop *phase ) const; 500 501 // Return TRUE or FALSE if the loop should be cache-line aligned. 502 // Gather the expression that does the alignment. Note that only 503 // one array base can be aligned in a loop (unless the VM guarantees 504 // mutual alignment). Note that if we vectorize short memory ops 505 // into longer memory ops, we may want to increase alignment. 506 bool policy_align( PhaseIdealLoop *phase ) const; 507 508 // Return TRUE if "iff" is a range check. 509 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; 510 511 // Compute loop exact trip count if possible 512 void compute_exact_trip_count( PhaseIdealLoop *phase ); 513 514 // Compute loop trip count from profile data 515 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 516 517 // Reassociate invariant expressions. 518 void reassociate_invariants(PhaseIdealLoop *phase); 519 // Reassociate invariant add and subtract expressions. 520 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); 521 // Return nonzero index of invariant operand if invariant and variant 522 // are combined with an Add or Sub. Helper for reassociate_invariants. 523 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); 524 525 // Return true if n is invariant 526 bool is_invariant(Node* n) const; 527 528 // Put loop body on igvn work list 529 void record_for_igvn(); 530 531 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } 532 bool is_inner() { return is_loop() && _child == NULL; } 533 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } 534 535 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 536 537 #ifndef PRODUCT 538 void dump_head( ) const; // Dump loop head only 539 void dump() const; // Dump this loop recursively 540 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 541 #endif 542 543 }; 544 545 // -----------------------------PhaseIdealLoop--------------------------------- 546 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a 547 // loop tree. Drives the loop-based transformations on the ideal graph. 548 class PhaseIdealLoop : public PhaseTransform { 549 friend class IdealLoopTree; 550 friend class SuperWord; 551 friend class CountedLoopReserveKit; 552 553 // Pre-computed def-use info 554 PhaseIterGVN &_igvn; 555 556 // Head of loop tree 557 IdealLoopTree *_ltree_root; 558 559 // Array of pre-order numbers, plus post-visited bit. 560 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 561 // ODD for post-visited. Other bits are the pre-order number. 562 uint *_preorders; 563 uint _max_preorder; 564 565 const PhaseIdealLoop* _verify_me; 566 bool _verify_only; 567 568 // Allocate _preorders[] array 569 void allocate_preorders() { 570 _max_preorder = C->unique()+8; 571 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 572 memset(_preorders, 0, sizeof(uint) * _max_preorder); 573 } 574 575 // Allocate _preorders[] array 576 void reallocate_preorders() { 577 if ( _max_preorder < C->unique() ) { 578 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 579 _max_preorder = C->unique(); 580 } 581 memset(_preorders, 0, sizeof(uint) * _max_preorder); 582 } 583 584 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 585 // adds new nodes. 586 void check_grow_preorders( ) { 587 if ( _max_preorder < C->unique() ) { 588 uint newsize = _max_preorder<<1; // double size of array 589 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 590 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 591 _max_preorder = newsize; 592 } 593 } 594 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 595 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 596 // Pre-order numbers are written to the Nodes array as low-bit-set values. 597 void set_preorder_visited( Node *n, int pre_order ) { 598 assert( !is_visited( n ), "already set" ); 599 _preorders[n->_idx] = (pre_order<<1); 600 }; 601 // Return pre-order number. 602 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 603 604 // Check for being post-visited. 605 // Should be previsited already (checked with assert(is_visited(n))). 606 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 607 608 // Mark as post visited 609 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 610 611 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 612 // Returns true if "n" is a data node, false if it's a control node. 613 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 614 615 // clear out dead code after build_loop_late 616 Node_List _deadlist; 617 618 // Support for faster execution of get_late_ctrl()/dom_lca() 619 // when a node has many uses and dominator depth is deep. 620 Node_Array _dom_lca_tags; 621 void init_dom_lca_tags(); 622 void clear_dom_lca_tags(); 623 624 // Helper for debugging bad dominance relationships 625 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 626 627 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 628 629 // Inline wrapper for frequent cases: 630 // 1) only one use 631 // 2) a use is the same as the current LCA passed as 'n1' 632 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 633 assert( n->is_CFG(), "" ); 634 // Fast-path NULL lca 635 if( lca != NULL && lca != n ) { 636 assert( lca->is_CFG(), "" ); 637 // find LCA of all uses 638 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 639 } 640 return find_non_split_ctrl(n); 641 } 642 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 643 644 // Helper function for directing control inputs away from CFG split 645 // points. 646 Node *find_non_split_ctrl( Node *ctrl ) const { 647 if (ctrl != NULL) { 648 if (ctrl->is_MultiBranch()) { 649 ctrl = ctrl->in(0); 650 } 651 assert(ctrl->is_CFG(), "CFG"); 652 } 653 return ctrl; 654 } 655 656 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 657 658 public: 659 bool has_node( Node* n ) const { 660 guarantee(n != NULL, "No Node."); 661 return _nodes[n->_idx] != NULL; 662 } 663 // check if transform created new nodes that need _ctrl recorded 664 Node *get_late_ctrl( Node *n, Node *early ); 665 Node *get_early_ctrl( Node *n ); 666 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 667 void set_early_ctrl( Node *n ); 668 void set_subtree_ctrl( Node *root ); 669 void set_ctrl( Node *n, Node *ctrl ) { 670 assert( !has_node(n) || has_ctrl(n), "" ); 671 assert( ctrl->in(0), "cannot set dead control node" ); 672 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 673 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 674 } 675 // Set control and update loop membership 676 void set_ctrl_and_loop(Node* n, Node* ctrl) { 677 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 678 IdealLoopTree* new_loop = get_loop(ctrl); 679 if (old_loop != new_loop) { 680 if (old_loop->_child == NULL) old_loop->_body.yank(n); 681 if (new_loop->_child == NULL) new_loop->_body.push(n); 682 } 683 set_ctrl(n, ctrl); 684 } 685 // Control nodes can be replaced or subsumed. During this pass they 686 // get their replacement Node in slot 1. Instead of updating the block 687 // location of all Nodes in the subsumed block, we lazily do it. As we 688 // pull such a subsumed block out of the array, we write back the final 689 // correct block. 690 Node *get_ctrl( Node *i ) { 691 assert(has_node(i), ""); 692 Node *n = get_ctrl_no_update(i); 693 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 694 assert(has_node(i) && has_ctrl(i), ""); 695 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 696 return n; 697 } 698 // true if CFG node d dominates CFG node n 699 bool is_dominator(Node *d, Node *n); 700 // return get_ctrl for a data node and self(n) for a CFG node 701 Node* ctrl_or_self(Node* n) { 702 if (has_ctrl(n)) 703 return get_ctrl(n); 704 else { 705 assert (n->is_CFG(), "must be a CFG node"); 706 return n; 707 } 708 } 709 710 private: 711 Node *get_ctrl_no_update_helper(Node *i) const { 712 assert(has_ctrl(i), "should be control, not loop"); 713 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 714 } 715 716 Node *get_ctrl_no_update(Node *i) const { 717 assert( has_ctrl(i), "" ); 718 Node *n = get_ctrl_no_update_helper(i); 719 if (!n->in(0)) { 720 // Skip dead CFG nodes 721 do { 722 n = get_ctrl_no_update_helper(n); 723 } while (!n->in(0)); 724 n = find_non_split_ctrl(n); 725 } 726 return n; 727 } 728 729 // Check for loop being set 730 // "n" must be a control node. Returns true if "n" is known to be in a loop. 731 bool has_loop( Node *n ) const { 732 assert(!has_node(n) || !has_ctrl(n), ""); 733 return has_node(n); 734 } 735 // Set loop 736 void set_loop( Node *n, IdealLoopTree *loop ) { 737 _nodes.map(n->_idx, (Node*)loop); 738 } 739 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 740 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 741 // from old_node to new_node to support the lazy update. Reference 742 // replaces loop reference, since that is not needed for dead node. 743 public: 744 void lazy_update(Node *old_node, Node *new_node) { 745 assert(old_node != new_node, "no cycles please"); 746 // Re-use the side array slot for this node to provide the 747 // forwarding pointer. 748 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 749 } 750 void lazy_replace(Node *old_node, Node *new_node) { 751 _igvn.replace_node(old_node, new_node); 752 lazy_update(old_node, new_node); 753 } 754 755 private: 756 757 // Place 'n' in some loop nest, where 'n' is a CFG node 758 void build_loop_tree(); 759 int build_loop_tree_impl( Node *n, int pre_order ); 760 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 761 // loop tree, not the root. 762 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 763 764 // Place Data nodes in some loop nest 765 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 766 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 767 void build_loop_late_post ( Node* n ); 768 769 // Array of immediate dominance info for each CFG node indexed by node idx 770 private: 771 uint _idom_size; 772 Node **_idom; // Array of immediate dominators 773 uint *_dom_depth; // Used for fast LCA test 774 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 775 776 Node* idom_no_update(Node* d) const { 777 assert(d->_idx < _idom_size, "oob"); 778 Node* n = _idom[d->_idx]; 779 assert(n != NULL,"Bad immediate dominator info."); 780 while (n->in(0) == NULL) { // Skip dead CFG nodes 781 //n = n->in(1); 782 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 783 assert(n != NULL,"Bad immediate dominator info."); 784 } 785 return n; 786 } 787 Node *idom(Node* d) const { 788 uint didx = d->_idx; 789 Node *n = idom_no_update(d); 790 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 791 return n; 792 } 793 uint dom_depth(Node* d) const { 794 guarantee(d != NULL, "Null dominator info."); 795 guarantee(d->_idx < _idom_size, ""); 796 return _dom_depth[d->_idx]; 797 } 798 void set_idom(Node* d, Node* n, uint dom_depth); 799 // Locally compute IDOM using dom_lca call 800 Node *compute_idom( Node *region ) const; 801 // Recompute dom_depth 802 void recompute_dom_depth(); 803 804 // Is safept not required by an outer loop? 805 bool is_deleteable_safept(Node* sfpt); 806 807 // Replace parallel induction variable (parallel to trip counter) 808 void replace_parallel_iv(IdealLoopTree *loop); 809 810 // Perform verification that the graph is valid. 811 PhaseIdealLoop( PhaseIterGVN &igvn) : 812 PhaseTransform(Ideal_Loop), 813 _igvn(igvn), 814 _dom_lca_tags(arena()), // Thread::resource_area 815 _verify_me(NULL), 816 _verify_only(true) { 817 build_and_optimize(false, false); 818 } 819 820 // build the loop tree and perform any requested optimizations 821 void build_and_optimize(bool do_split_if, bool skip_loop_opts); 822 823 public: 824 // Dominators for the sea of nodes 825 void Dominators(); 826 Node *dom_lca( Node *n1, Node *n2 ) const { 827 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 828 } 829 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 830 831 // Compute the Ideal Node to Loop mapping 832 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) : 833 PhaseTransform(Ideal_Loop), 834 _igvn(igvn), 835 _dom_lca_tags(arena()), // Thread::resource_area 836 _verify_me(NULL), 837 _verify_only(false) { 838 build_and_optimize(do_split_ifs, skip_loop_opts); 839 } 840 841 // Verify that verify_me made the same decisions as a fresh run. 842 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : 843 PhaseTransform(Ideal_Loop), 844 _igvn(igvn), 845 _dom_lca_tags(arena()), // Thread::resource_area 846 _verify_me(verify_me), 847 _verify_only(false) { 848 build_and_optimize(false, false); 849 } 850 851 // Build and verify the loop tree without modifying the graph. This 852 // is useful to verify that all inputs properly dominate their uses. 853 static void verify(PhaseIterGVN& igvn) { 854 #ifdef ASSERT 855 PhaseIdealLoop v(igvn); 856 #endif 857 } 858 859 // True if the method has at least 1 irreducible loop 860 bool _has_irreducible_loops; 861 862 // Per-Node transform 863 virtual Node *transform( Node *a_node ) { return 0; } 864 865 bool is_counted_loop( Node *x, IdealLoopTree *loop ); 866 867 Node* exact_limit( IdealLoopTree *loop ); 868 869 // Return a post-walked LoopNode 870 IdealLoopTree *get_loop( Node *n ) const { 871 // Dead nodes have no loop, so return the top level loop instead 872 if (!has_node(n)) return _ltree_root; 873 assert(!has_ctrl(n), ""); 874 return (IdealLoopTree*)_nodes[n->_idx]; 875 } 876 877 // Is 'n' a (nested) member of 'loop'? 878 int is_member( const IdealLoopTree *loop, Node *n ) const { 879 return loop->is_member(get_loop(n)); } 880 881 // This is the basic building block of the loop optimizations. It clones an 882 // entire loop body. It makes an old_new loop body mapping; with this 883 // mapping you can find the new-loop equivalent to an old-loop node. All 884 // new-loop nodes are exactly equal to their old-loop counterparts, all 885 // edges are the same. All exits from the old-loop now have a RegionNode 886 // that merges the equivalent new-loop path. This is true even for the 887 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 888 // now come from (one or more) Phis that merge their new-loop equivalents. 889 // Parameter side_by_side_idom: 890 // When side_by_size_idom is NULL, the dominator tree is constructed for 891 // the clone loop to dominate the original. Used in construction of 892 // pre-main-post loop sequence. 893 // When nonnull, the clone and original are side-by-side, both are 894 // dominated by the passed in side_by_side_idom node. Used in 895 // construction of unswitched loops. 896 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 897 Node* side_by_side_idom = NULL); 898 899 // If we got the effect of peeling, either by actually peeling or by 900 // making a pre-loop which must execute at least once, we can remove 901 // all loop-invariant dominated tests in the main body. 902 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 903 904 // Generate code to do a loop peel for the given loop (and body). 905 // old_new is a temp array. 906 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 907 908 // Add pre and post loops around the given loop. These loops are used 909 // during RCE, unrolling and aligning loops. 910 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 911 // Add a vector post loop between a vector main loop and the current post loop 912 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 913 // If Node n lives in the back_ctrl block, we clone a private version of n 914 // in preheader_ctrl block and return that, otherwise return n. 915 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 916 917 // Take steps to maximally unroll the loop. Peel any odd iterations, then 918 // unroll to do double iterations. The next round of major loop transforms 919 // will repeat till the doubled loop body does all remaining iterations in 1 920 // pass. 921 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 922 923 // Unroll the loop body one step - make each trip do 2 iterations. 924 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 925 926 // Mark vector reduction candidates before loop unrolling 927 void mark_reductions( IdealLoopTree *loop ); 928 929 // Return true if exp is a constant times an induction var 930 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); 931 932 // Return true if exp is a scaled induction var plus (or minus) constant 933 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); 934 935 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 936 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 937 Deoptimization::DeoptReason reason, 938 int opcode); 939 void register_control(Node* n, IdealLoopTree *loop, Node* pred); 940 941 // Clone loop predicates to cloned loops (peeled, unswitched) 942 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, 943 Deoptimization::DeoptReason reason, 944 PhaseIdealLoop* loop_phase, 945 PhaseIterGVN* igvn); 946 947 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, 948 bool clone_limit_check, 949 PhaseIdealLoop* loop_phase, 950 PhaseIterGVN* igvn); 951 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); 952 953 static Node* skip_loop_predicates(Node* entry); 954 955 // Find a good location to insert a predicate 956 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); 957 // Find a predicate 958 static Node* find_predicate(Node* entry); 959 // Construct a range check for a predicate if 960 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, 961 int scale, Node* offset, 962 Node* init, Node* limit, Node* stride, 963 Node* range, bool upper); 964 965 // Implementation of the loop predication to promote checks outside the loop 966 bool loop_predication_impl(IdealLoopTree *loop); 967 968 // Helper function to collect predicate for eliminating the useless ones 969 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); 970 void eliminate_useless_predicates(); 971 972 // Change the control input of expensive nodes to allow commoning by 973 // IGVN when it is guaranteed to not result in a more frequent 974 // execution of the expensive node. Return true if progress. 975 bool process_expensive_nodes(); 976 977 // Check whether node has become unreachable 978 bool is_node_unreachable(Node *n) const { 979 return !has_node(n) || n->is_unreachable(_igvn); 980 } 981 982 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 983 void do_range_check( IdealLoopTree *loop, Node_List &old_new ); 984 985 // Create a slow version of the loop by cloning the loop 986 // and inserting an if to select fast-slow versions. 987 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, 988 Node_List &old_new, 989 int opcode); 990 991 // Clone a loop and return the clone head (clone_loop_head). 992 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse, 993 // This routine was created for usage in CountedLoopReserveKit. 994 // 995 // int(1) -> If -> IfTrue -> original_loop_head 996 // | 997 // V 998 // IfFalse -> clone_loop_head (returned by function pointer) 999 // 1000 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk); 1001 // Clone loop with an invariant test (that does not exit) and 1002 // insert a clone of the test that selects which version to 1003 // execute. 1004 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 1005 1006 // Find candidate "if" for unswitching 1007 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; 1008 1009 // Range Check Elimination uses this function! 1010 // Constrain the main loop iterations so the affine function: 1011 // low_limit <= scale_con * I + offset < upper_limit 1012 // always holds true. That is, either increase the number of iterations in 1013 // the pre-loop or the post-loop until the condition holds true in the main 1014 // loop. Scale_con, offset and limit are all loop invariant. 1015 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 ); 1016 // Helper function for add_constraint(). 1017 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl ); 1018 1019 // Partially peel loop up through last_peel node. 1020 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1021 1022 // Create a scheduled list of nodes control dependent on ctrl set. 1023 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1024 // Has a use in the vector set 1025 bool has_use_in_set( Node* n, VectorSet& vset ); 1026 // Has use internal to the vector set (ie. not in a phi at the loop head) 1027 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1028 // clone "n" for uses that are outside of loop 1029 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1030 // clone "n" for special uses that are in the not_peeled region 1031 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1032 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1033 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1034 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1035 #ifdef ASSERT 1036 // Validate the loop partition sets: peel and not_peel 1037 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1038 // Ensure that uses outside of loop are of the right form 1039 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1040 uint orig_exit_idx, uint clone_exit_idx); 1041 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1042 #endif 1043 1044 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1045 int stride_of_possible_iv( Node* iff ); 1046 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1047 // Return the (unique) control output node that's in the loop (if it exists.) 1048 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1049 // Insert a signed compare loop exit cloned from an unsigned compare. 1050 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1051 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1052 // Utility to register node "n" with PhaseIdealLoop 1053 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 1054 // Utility to create an if-projection 1055 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1056 // Force the iff control output to be the live_proj 1057 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1058 // Insert a region before an if projection 1059 RegionNode* insert_region_before_proj(ProjNode* proj); 1060 // Insert a new if before an if projection 1061 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1062 1063 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1064 // "Nearly" because all Nodes have been cloned from the original in the loop, 1065 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1066 // through the Phi recursively, and return a Bool. 1067 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); 1068 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); 1069 1070 1071 // Rework addressing expressions to get the most loop-invariant stuff 1072 // moved out. We'd like to do all associative operators, but it's especially 1073 // important (common) to do address expressions. 1074 Node *remix_address_expressions( Node *n ); 1075 1076 // Attempt to use a conditional move instead of a phi/branch 1077 Node *conditional_move( Node *n ); 1078 1079 // Reorganize offset computations to lower register pressure. 1080 // Mostly prevent loop-fallout uses of the pre-incremented trip counter 1081 // (which are then alive with the post-incremented trip counter 1082 // forcing an extra register move) 1083 void reorg_offsets( IdealLoopTree *loop ); 1084 1085 // Check for aggressive application of 'split-if' optimization, 1086 // using basic block level info. 1087 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); 1088 Node *split_if_with_blocks_pre ( Node *n ); 1089 void split_if_with_blocks_post( Node *n ); 1090 Node *has_local_phi_input( Node *n ); 1091 // Mark an IfNode as being dominated by a prior test, 1092 // without actually altering the CFG (and hence IDOM info). 1093 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); 1094 1095 // Split Node 'n' through merge point 1096 Node *split_thru_region( Node *n, Node *region ); 1097 // Split Node 'n' through merge point if there is enough win. 1098 Node *split_thru_phi( Node *n, Node *region, int policy ); 1099 // Found an If getting its condition-code input from a Phi in the 1100 // same block. Split thru the Region. 1101 void do_split_if( Node *iff ); 1102 1103 // Conversion of fill/copy patterns into intrisic versions 1104 bool do_intrinsify_fill(); 1105 bool intrinsify_fill(IdealLoopTree* lpt); 1106 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1107 Node*& shift, Node*& offset); 1108 1109 private: 1110 // Return a type based on condition control flow 1111 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1112 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 1113 // Helpers for filtered type 1114 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1115 1116 // Helper functions 1117 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1118 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1119 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 ); 1120 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1121 void sink_use( Node *use, Node *post_loop ); 1122 Node *place_near_use( Node *useblock ) const; 1123 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1124 void try_move_store_after_loop(Node* n); 1125 bool identical_backtoback_ifs(Node *n); 1126 bool can_split_if(Node *n_ctrl); 1127 1128 bool _created_loop_node; 1129 public: 1130 void set_created_loop_node() { _created_loop_node = true; } 1131 bool created_loop_node() { return _created_loop_node; } 1132 void register_new_node( Node *n, Node *blk ); 1133 1134 #ifdef ASSERT 1135 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1136 #endif 1137 1138 #ifndef PRODUCT 1139 void dump( ) const; 1140 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; 1141 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; 1142 void verify() const; // Major slow :-) 1143 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; 1144 IdealLoopTree *get_loop_idx(Node* n) const { 1145 // Dead nodes have no loop, so return the top level loop instead 1146 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 1147 } 1148 // Print some stats 1149 static void print_statistics(); 1150 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1151 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1152 #endif 1153 }; 1154 1155 // This kit may be used for making of a reserved copy of a loop before this loop 1156 // goes under non-reversible changes. 1157 // 1158 // Function create_reserve() creates a reserved copy (clone) of the loop. 1159 // The reserved copy is created by calling 1160 // PhaseIdealLoop::create_reserve_version_of_loop - see there how 1161 // the original and reserved loops are connected in the outer graph. 1162 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'. 1163 // 1164 // By default the reserved copy (clone) of the loop is created as dead code - it is 1165 // dominated in the outer loop by this node chain: 1166 // intcon(1)->If->IfFalse->reserved_copy. 1167 // The original loop is dominated by the the same node chain but IfTrue projection: 1168 // intcon(0)->If->IfTrue->original_loop. 1169 // 1170 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve() 1171 // and the dtor, checks _use_new value. 1172 // If _use_new == false, it "switches" control to reserved copy of the loop 1173 // by simple replacing of node intcon(1) with node intcon(0). 1174 // 1175 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp). 1176 // 1177 // void CountedLoopReserveKit_example() 1178 // { 1179 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object 1180 // if (DoReserveCopy && !lrk.has_reserved()) { 1181 // return; //failed to create reserved loop copy 1182 // } 1183 // ... 1184 // //something is wrong, switch to original loop 1185 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch 1186 // ... 1187 // //everything worked ok, return with the newly modified loop 1188 // lrk.use_new(); 1189 // return; // ~CountedLoopReserveKit does nothing once use_new() was called 1190 // } 1191 // 1192 // Keep in mind, that by default if create_reserve() is not followed by use_new() 1193 // the dtor will "switch to the original" loop. 1194 // NOTE. You you modify outside of the original loop this class is no help. 1195 // 1196 class CountedLoopReserveKit { 1197 private: 1198 PhaseIdealLoop* _phase; 1199 IdealLoopTree* _lpt; 1200 LoopNode* _lp; 1201 IfNode* _iff; 1202 LoopNode* _lp_reserved; 1203 bool _has_reserved; 1204 bool _use_new; 1205 const bool _active; //may be set to false in ctor, then the object is dummy 1206 1207 public: 1208 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active); 1209 ~CountedLoopReserveKit(); 1210 void use_new() {_use_new = true;} 1211 void set_iff(IfNode* x) {_iff = x;} 1212 bool has_reserved() const { return _active && _has_reserved;} 1213 private: 1214 bool create_reserve(); 1215 };// class CountedLoopReserveKit 1216 1217 inline Node* IdealLoopTree::tail() { 1218 // Handle lazy update of _tail field 1219 Node *n = _tail; 1220 //while( !n->in(0) ) // Skip dead CFG nodes 1221 //n = n->in(1); 1222 if (n->in(0) == NULL) 1223 n = _phase->get_ctrl(n); 1224 _tail = n; 1225 return n; 1226 } 1227 1228 1229 // Iterate over the loop tree using a preorder, left-to-right traversal. 1230 // 1231 // Example that visits all counted loops from within PhaseIdealLoop 1232 // 1233 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1234 // IdealLoopTree* lpt = iter.current(); 1235 // if (!lpt->is_counted()) continue; 1236 // ... 1237 class LoopTreeIterator : public StackObj { 1238 private: 1239 IdealLoopTree* _root; 1240 IdealLoopTree* _curnt; 1241 1242 public: 1243 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1244 1245 bool done() { return _curnt == NULL; } // Finished iterating? 1246 1247 void next(); // Advance to next loop tree 1248 1249 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1250 }; 1251 1252 #endif // SHARE_VM_OPTO_LOOPNODE_HPP