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