1 /* 2 * Copyright (c) 1998, 2019, 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_OPTO_LOOPNODE_HPP 26 #define SHARE_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 OuterStripMinedLoopEndNode; 41 class PathFrequency; 42 class PhaseIdealLoop; 43 class CountedLoopReserveKit; 44 class VectorSet; 45 class Invariance; 46 struct small_cache; 47 48 // 49 // I D E A L I Z E D L O O P S 50 // 51 // Idealized loops are the set of loops I perform more interesting 52 // transformations on, beyond simple hoisting. 53 54 //------------------------------LoopNode--------------------------------------- 55 // Simple loop header. Fall in path on left, loop-back path on right. 56 class LoopNode : public RegionNode { 57 // Size is bigger to hold the flags. However, the flags do not change 58 // the semantics so it does not appear in the hash & cmp functions. 59 virtual uint size_of() const { return sizeof(*this); } 60 protected: 61 uint _loop_flags; 62 // Names for flag bitfields 63 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 64 MainHasNoPreLoop=4, 65 HasExactTripCount=8, 66 InnerLoop=16, 67 PartialPeelLoop=32, 68 PartialPeelFailed=64, 69 HasReductions=128, 70 WasSlpAnalyzed=256, 71 PassedSlpAnalysis=512, 72 DoUnrollOnly=1024, 73 VectorizedLoop=2048, 74 HasAtomicPostLoop=4096, 75 HasRangeChecks=8192, 76 IsMultiversioned=16384, 77 StripMined=32768, 78 SubwordLoop=65536, 79 ProfileTripFailed=131072, 80 FlattenedArrays=262144}; 81 char _unswitch_count; 82 enum { _unswitch_max=3 }; 83 char _postloop_flags; 84 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 }; 85 86 // Expected trip count from profile data 87 float _profile_trip_cnt; 88 89 public: 90 // Names for edge indices 91 enum { Self=0, EntryControl, LoopBackControl }; 92 93 bool is_inner_loop() const { return _loop_flags & InnerLoop; } 94 void set_inner_loop() { _loop_flags |= InnerLoop; } 95 96 bool range_checks_present() const { return _loop_flags & HasRangeChecks; } 97 bool is_multiversioned() const { return _loop_flags & IsMultiversioned; } 98 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } 99 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 100 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 101 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 102 bool is_strip_mined() const { return _loop_flags & StripMined; } 103 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } 104 bool is_subword_loop() const { return _loop_flags & SubwordLoop; } 105 bool is_flattened_arrays() const { return _loop_flags & FlattenedArrays; } 106 107 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 108 void mark_has_reductions() { _loop_flags |= HasReductions; } 109 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 110 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 111 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 112 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 113 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 114 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; } 115 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; } 116 void mark_strip_mined() { _loop_flags |= StripMined; } 117 void clear_strip_mined() { _loop_flags &= ~StripMined; } 118 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } 119 void mark_subword_loop() { _loop_flags |= SubwordLoop; } 120 void mark_flattened_arrays() { _loop_flags |= FlattenedArrays; } 121 122 int unswitch_max() { return _unswitch_max; } 123 int unswitch_count() { return _unswitch_count; } 124 125 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; } 126 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; } 127 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; } 128 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; } 129 130 void set_unswitch_count(int val) { 131 assert (val <= unswitch_max(), "too many unswitches"); 132 _unswitch_count = val; 133 } 134 135 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 136 float profile_trip_cnt() { return _profile_trip_cnt; } 137 138 LoopNode(Node *entry, Node *backedge) 139 : RegionNode(3), _loop_flags(0), _unswitch_count(0), 140 _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) { 141 init_class_id(Class_Loop); 142 init_req(EntryControl, entry); 143 init_req(LoopBackControl, backedge); 144 } 145 146 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 147 virtual int Opcode() const; 148 bool can_be_counted_loop(PhaseTransform* phase) const { 149 return req() == 3 && in(0) != NULL && 150 in(1) != NULL && phase->type(in(1)) != Type::TOP && 151 in(2) != NULL && phase->type(in(2)) != Type::TOP; 152 } 153 bool is_valid_counted_loop() const; 154 #ifndef PRODUCT 155 virtual void dump_spec(outputStream *st) const; 156 #endif 157 158 void verify_strip_mined(int expect_skeleton) const; 159 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } 160 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; } 161 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; } 162 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; } 163 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; } 164 }; 165 166 //------------------------------Counted Loops---------------------------------- 167 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 168 // path (and maybe some other exit paths). The trip-counter exit is always 169 // last in the loop. The trip-counter have to stride by a constant; 170 // the exit value is also loop invariant. 171 172 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 173 // CountedLoopNode has the incoming loop control and the loop-back-control 174 // which is always the IfTrue before the matching CountedLoopEndNode. The 175 // CountedLoopEndNode has an incoming control (possibly not the 176 // CountedLoopNode if there is control flow in the loop), the post-increment 177 // trip-counter value, and the limit. The trip-counter value is always of 178 // the form (Op old-trip-counter stride). The old-trip-counter is produced 179 // by a Phi connected to the CountedLoopNode. The stride is constant. 180 // The Op is any commutable opcode, including Add, Mul, Xor. The 181 // CountedLoopEndNode also takes in the loop-invariant limit value. 182 183 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 184 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 185 // via the old-trip-counter from the Op node. 186 187 //------------------------------CountedLoopNode-------------------------------- 188 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 189 // inputs the incoming loop-start control and the loop-back control, so they 190 // act like RegionNodes. They also take in the initial trip counter, the 191 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 192 // produce a loop-body control and the trip counter value. Since 193 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 194 195 class CountedLoopNode : public LoopNode { 196 // Size is bigger to hold _main_idx. However, _main_idx does not change 197 // the semantics so it does not appear in the hash & cmp functions. 198 virtual uint size_of() const { return sizeof(*this); } 199 200 // For Pre- and Post-loops during debugging ONLY, this holds the index of 201 // the Main CountedLoop. Used to assert that we understand the graph shape. 202 node_idx_t _main_idx; 203 204 // Known trip count calculated by compute_exact_trip_count() 205 uint _trip_count; 206 207 // Log2 of original loop bodies in unrolled loop 208 int _unrolled_count_log2; 209 210 // Node count prior to last unrolling - used to decide if 211 // unroll,optimize,unroll,optimize,... is making progress 212 int _node_count_before_unroll; 213 214 // If slp analysis is performed we record the maximum 215 // vector mapped unroll factor here 216 int _slp_maximum_unroll_factor; 217 218 public: 219 CountedLoopNode( Node *entry, Node *backedge ) 220 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 221 _unrolled_count_log2(0), _node_count_before_unroll(0), 222 _slp_maximum_unroll_factor(0) { 223 init_class_id(Class_CountedLoop); 224 // Initialize _trip_count to the largest possible value. 225 // Will be reset (lower) if the loop's trip count is known. 226 } 227 228 virtual int Opcode() const; 229 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 230 231 Node *init_control() const { return in(EntryControl); } 232 Node *back_control() const { return in(LoopBackControl); } 233 CountedLoopEndNode *loopexit_or_null() const; 234 CountedLoopEndNode *loopexit() const; 235 Node *init_trip() const; 236 Node *stride() const; 237 int stride_con() const; 238 bool stride_is_con() const; 239 Node *limit() const; 240 Node *incr() const; 241 Node *phi() const; 242 243 // Match increment with optional truncation 244 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); 245 246 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 247 // can run short a few iterations and may start a few iterations in. 248 // It will be RCE'd and unrolled and aligned. 249 250 // A following 'post' loop will run any remaining iterations. Used 251 // during Range Check Elimination, the 'post' loop will do any final 252 // iterations with full checks. Also used by Loop Unrolling, where 253 // the 'post' loop will do any epilog iterations needed. Basically, 254 // a 'post' loop can not profitably be further unrolled or RCE'd. 255 256 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 257 // it may do under-flow checks for RCE and may do alignment iterations 258 // so the following main loop 'knows' that it is striding down cache 259 // lines. 260 261 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 262 // Aligned, may be missing it's pre-loop. 263 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 264 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 265 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 266 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 267 bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; } 268 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 269 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 270 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 271 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 272 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 273 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 274 275 int main_idx() const { return _main_idx; } 276 277 278 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 279 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 280 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 281 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 282 283 void set_trip_count(uint tc) { _trip_count = tc; } 284 uint trip_count() { return _trip_count; } 285 286 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 287 void set_exact_trip_count(uint tc) { 288 _trip_count = tc; 289 _loop_flags |= HasExactTripCount; 290 } 291 void set_nonexact_trip_count() { 292 _loop_flags &= ~HasExactTripCount; 293 } 294 void set_notpassed_slp() { 295 _loop_flags &= ~PassedSlpAnalysis; 296 } 297 298 void double_unrolled_count() { _unrolled_count_log2++; } 299 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 300 301 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 302 int node_count_before_unroll() { return _node_count_before_unroll; } 303 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 304 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 305 306 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1); 307 OuterStripMinedLoopNode* outer_loop() const; 308 virtual IfTrueNode* outer_loop_tail() const; 309 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 310 virtual IfFalseNode* outer_loop_exit() const; 311 virtual SafePointNode* outer_safepoint() const; 312 313 // If this is a main loop in a pre/main/post loop nest, walk over 314 // the predicates that were inserted by 315 // duplicate_predicates()/add_range_check_predicate() 316 static Node* skip_predicates_from_entry(Node* ctrl); 317 Node* skip_predicates(); 318 319 #ifndef PRODUCT 320 virtual void dump_spec(outputStream *st) const; 321 #endif 322 }; 323 324 //------------------------------CountedLoopEndNode----------------------------- 325 // CountedLoopEndNodes end simple trip counted loops. They act much like 326 // IfNodes. 327 class CountedLoopEndNode : public IfNode { 328 public: 329 enum { TestControl, TestValue }; 330 331 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) 332 : IfNode( control, test, prob, cnt) { 333 init_class_id(Class_CountedLoopEnd); 334 } 335 virtual int Opcode() const; 336 337 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 338 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 339 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 340 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 341 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 342 int stride_con() const; 343 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } 344 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 345 PhiNode *phi() const { 346 Node *tmp = incr(); 347 if (tmp && tmp->req() == 3) { 348 Node* phi = tmp->in(1); 349 if (phi->is_Phi()) { 350 return phi->as_Phi(); 351 } 352 } 353 return NULL; 354 } 355 CountedLoopNode *loopnode() const { 356 // The CountedLoopNode that goes with this CountedLoopEndNode may 357 // have been optimized out by the IGVN so be cautious with the 358 // pattern matching on the graph 359 PhiNode* iv_phi = phi(); 360 if (iv_phi == NULL) { 361 return NULL; 362 } 363 Node *ln = iv_phi->in(0); 364 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) { 365 return (CountedLoopNode*)ln; 366 } 367 return NULL; 368 } 369 370 #ifndef PRODUCT 371 virtual void dump_spec(outputStream *st) const; 372 #endif 373 }; 374 375 376 inline CountedLoopEndNode* CountedLoopNode::loopexit_or_null() const { 377 Node* bctrl = back_control(); 378 if (bctrl == NULL) return NULL; 379 380 Node* lexit = bctrl->in(0); 381 return (CountedLoopEndNode*) 382 (lexit->Opcode() == Op_CountedLoopEnd ? lexit : NULL); 383 } 384 385 inline CountedLoopEndNode* CountedLoopNode::loopexit() const { 386 CountedLoopEndNode* cle = loopexit_or_null(); 387 assert(cle != NULL, "loopexit is NULL"); 388 return cle; 389 } 390 391 inline Node* CountedLoopNode::init_trip() const { 392 CountedLoopEndNode* cle = loopexit_or_null(); 393 return cle != NULL ? cle->init_trip() : NULL; 394 } 395 inline Node* CountedLoopNode::stride() const { 396 CountedLoopEndNode* cle = loopexit_or_null(); 397 return cle != NULL ? cle->stride() : NULL; 398 } 399 inline int CountedLoopNode::stride_con() const { 400 CountedLoopEndNode* cle = loopexit_or_null(); 401 return cle != NULL ? cle->stride_con() : 0; 402 } 403 inline bool CountedLoopNode::stride_is_con() const { 404 CountedLoopEndNode* cle = loopexit_or_null(); 405 return cle != NULL && cle->stride_is_con(); 406 } 407 inline Node* CountedLoopNode::limit() const { 408 CountedLoopEndNode* cle = loopexit_or_null(); 409 return cle != NULL ? cle->limit() : NULL; 410 } 411 inline Node* CountedLoopNode::incr() const { 412 CountedLoopEndNode* cle = loopexit_or_null(); 413 return cle != NULL ? cle->incr() : NULL; 414 } 415 inline Node* CountedLoopNode::phi() const { 416 CountedLoopEndNode* cle = loopexit_or_null(); 417 return cle != NULL ? cle->phi() : NULL; 418 } 419 420 //------------------------------LoopLimitNode----------------------------- 421 // Counted Loop limit node which represents exact final iterator value: 422 // trip_count = (limit - init_trip + stride - 1)/stride 423 // final_value= trip_count * stride + init_trip. 424 // Use HW instructions to calculate it when it can overflow in integer. 425 // Note, final_value should fit into integer since counted loop has 426 // limit check: limit <= max_int-stride. 427 class LoopLimitNode : public Node { 428 enum { Init=1, Limit=2, Stride=3 }; 429 public: 430 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 431 // Put it on the Macro nodes list to optimize during macro nodes expansion. 432 init_flags(Flag_is_macro); 433 C->add_macro_node(this); 434 } 435 virtual int Opcode() const; 436 virtual const Type *bottom_type() const { return TypeInt::INT; } 437 virtual uint ideal_reg() const { return Op_RegI; } 438 virtual const Type* Value(PhaseGVN* phase) const; 439 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 440 virtual Node* Identity(PhaseGVN* phase); 441 }; 442 443 // Support for strip mining 444 class OuterStripMinedLoopNode : public LoopNode { 445 private: 446 CountedLoopNode* inner_loop() const; 447 public: 448 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) 449 : LoopNode(entry, backedge) { 450 init_class_id(Class_OuterStripMinedLoop); 451 init_flags(Flag_is_macro); 452 C->add_macro_node(this); 453 } 454 455 virtual int Opcode() const; 456 457 virtual IfTrueNode* outer_loop_tail() const; 458 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 459 virtual IfFalseNode* outer_loop_exit() const; 460 virtual SafePointNode* outer_safepoint() const; 461 void adjust_strip_mined_loop(PhaseIterGVN* igvn); 462 }; 463 464 class OuterStripMinedLoopEndNode : public IfNode { 465 public: 466 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) 467 : IfNode(control, test, prob, cnt) { 468 init_class_id(Class_OuterStripMinedLoopEnd); 469 } 470 471 virtual int Opcode() const; 472 473 virtual const Type* Value(PhaseGVN* phase) const; 474 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 475 }; 476 477 // -----------------------------IdealLoopTree---------------------------------- 478 class IdealLoopTree : public ResourceObj { 479 public: 480 IdealLoopTree *_parent; // Parent in loop tree 481 IdealLoopTree *_next; // Next sibling in loop tree 482 IdealLoopTree *_child; // First child in loop tree 483 484 // The head-tail backedge defines the loop. 485 // If a loop has multiple backedges, this is addressed during cleanup where 486 // we peel off the multiple backedges, merging all edges at the bottom and 487 // ensuring that one proper backedge flow into the loop. 488 Node *_head; // Head of loop 489 Node *_tail; // Tail of loop 490 inline Node *tail(); // Handle lazy update of _tail field 491 PhaseIdealLoop* _phase; 492 int _local_loop_unroll_limit; 493 int _local_loop_unroll_factor; 494 495 Node_List _body; // Loop body for inner loops 496 497 uint8_t _nest; // Nesting depth 498 uint8_t _irreducible:1, // True if irreducible 499 _has_call:1, // True if has call safepoint 500 _has_sfpt:1, // True if has non-call safepoint 501 _rce_candidate:1; // True if candidate for range check elimination 502 503 Node_List* _safepts; // List of safepoints in this loop 504 Node_List* _required_safept; // A inner loop cannot delete these safepts; 505 bool _allow_optimizations; // Allow loop optimizations 506 507 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 508 : _parent(0), _next(0), _child(0), 509 _head(head), _tail(tail), 510 _phase(phase), 511 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0), 512 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 513 _safepts(NULL), 514 _required_safept(NULL), 515 _allow_optimizations(true) 516 { 517 precond(_head != NULL); 518 precond(_tail != NULL); 519 } 520 521 // Is 'l' a member of 'this'? 522 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 523 524 // Set loop nesting depth. Accumulate has_call bits. 525 int set_nest( uint depth ); 526 527 // Split out multiple fall-in edges from the loop header. Move them to a 528 // private RegionNode before the loop. This becomes the loop landing pad. 529 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 530 531 // Split out the outermost loop from this shared header. 532 void split_outer_loop( PhaseIdealLoop *phase ); 533 534 // Merge all the backedges from the shared header into a private Region. 535 // Feed that region as the one backedge to this loop. 536 void merge_many_backedges( PhaseIdealLoop *phase ); 537 538 // Split shared headers and insert loop landing pads. 539 // Insert a LoopNode to replace the RegionNode. 540 // Returns TRUE if loop tree is structurally changed. 541 bool beautify_loops( PhaseIdealLoop *phase ); 542 543 // Perform optimization to use the loop predicates for null checks and range checks. 544 // Applies to any loop level (not just the innermost one) 545 bool loop_predication( PhaseIdealLoop *phase); 546 547 // Perform iteration-splitting on inner loops. Split iterations to 548 // avoid range checks or one-shot null checks. Returns false if the 549 // current round of loop opts should stop. 550 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 551 552 // Driver for various flavors of iteration splitting. Returns false 553 // if the current round of loop opts should stop. 554 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 555 556 // Given dominators, try to find loops with calls that must always be 557 // executed (call dominates loop tail). These loops do not need non-call 558 // safepoints (ncsfpt). 559 void check_safepts(VectorSet &visited, Node_List &stack); 560 561 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 562 // encountered. 563 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 564 565 // Remove safepoints from loop. Optionally keeping one. 566 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 567 568 // Convert to counted loops where possible 569 void counted_loop( PhaseIdealLoop *phase ); 570 571 // Check for Node being a loop-breaking test 572 Node *is_loop_exit(Node *iff) const; 573 574 // Remove simplistic dead code from loop body 575 void DCE_loop_body(); 576 577 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 578 // Replace with a 1-in-10 exit guess. 579 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 580 581 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 582 // Useful for unrolling loops with NO array accesses. 583 bool policy_peel_only( PhaseIdealLoop *phase ) const; 584 585 // Return TRUE or FALSE if the loop should be unswitched -- clone 586 // loop with an invariant test 587 bool policy_unswitching( PhaseIdealLoop *phase ) const; 588 589 // Micro-benchmark spamming. Remove empty loops. 590 bool do_remove_empty_loop( PhaseIdealLoop *phase ); 591 592 // Convert one iteration loop into normal code. 593 bool do_one_iteration_loop( PhaseIdealLoop *phase ); 594 595 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 596 // move some loop-invariant test (usually a null-check) before the loop. 597 bool policy_peeling(PhaseIdealLoop *phase); 598 599 uint estimate_peeling(PhaseIdealLoop *phase); 600 601 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 602 // known trip count in the counted loop node. 603 bool policy_maximally_unroll(PhaseIdealLoop *phase) const; 604 605 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll 606 // if the loop is a counted loop and the loop body is small enough. 607 bool policy_unroll(PhaseIdealLoop *phase); 608 609 // Loop analyses to map to a maximal superword unrolling for vectorization. 610 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 611 612 // Return TRUE or FALSE if the loop should be range-check-eliminated. 613 // Gather a list of IF tests that are dominated by iteration splitting; 614 // also gather the end of the first split and the start of the 2nd split. 615 bool policy_range_check( PhaseIdealLoop *phase ) const; 616 617 // Return TRUE or FALSE if the loop should be cache-line aligned. 618 // Gather the expression that does the alignment. Note that only 619 // one array base can be aligned in a loop (unless the VM guarantees 620 // mutual alignment). Note that if we vectorize short memory ops 621 // into longer memory ops, we may want to increase alignment. 622 bool policy_align( PhaseIdealLoop *phase ) const; 623 624 // Return TRUE if "iff" is a range check. 625 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; 626 627 // Estimate the number of nodes required when cloning a loop (body). 628 uint est_loop_clone_sz(uint factor) const; 629 630 // Compute loop trip count if possible 631 void compute_trip_count(PhaseIdealLoop* phase); 632 633 // Compute loop trip count from profile data 634 float compute_profile_trip_cnt_helper(Node* n); 635 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 636 637 // Reassociate invariant expressions. 638 void reassociate_invariants(PhaseIdealLoop *phase); 639 // Reassociate invariant add and subtract expressions. 640 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); 641 // Return nonzero index of invariant operand if invariant and variant 642 // are combined with an Add or Sub. Helper for reassociate_invariants. 643 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); 644 645 // Return true if n is invariant 646 bool is_invariant(Node* n) const; 647 648 // Put loop body on igvn work list 649 void record_for_igvn(); 650 651 bool is_root() { return _parent == NULL; } 652 // A proper/reducible loop w/o any (occasional) dead back-edge. 653 bool is_loop() { return !_irreducible && !tail()->is_top(); } 654 bool is_counted() { return is_loop() && _head->is_CountedLoop(); } 655 bool is_innermost() { return is_loop() && _child == NULL; } 656 657 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 658 659 #ifndef PRODUCT 660 void dump_head( ) const; // Dump loop head only 661 void dump() const; // Dump this loop recursively 662 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 663 #endif 664 665 }; 666 667 // -----------------------------PhaseIdealLoop--------------------------------- 668 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees 669 // into a loop tree. Drives the loop-based transformations on the ideal graph. 670 class PhaseIdealLoop : public PhaseTransform { 671 friend class IdealLoopTree; 672 friend class SuperWord; 673 friend class CountedLoopReserveKit; 674 friend class ShenandoahBarrierC2Support; 675 friend class AutoNodeBudget; 676 677 // Pre-computed def-use info 678 PhaseIterGVN &_igvn; 679 680 // Head of loop tree 681 IdealLoopTree *_ltree_root; 682 683 // Array of pre-order numbers, plus post-visited bit. 684 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 685 // ODD for post-visited. Other bits are the pre-order number. 686 uint *_preorders; 687 uint _max_preorder; 688 689 const PhaseIdealLoop* _verify_me; 690 bool _verify_only; 691 692 // Allocate _preorders[] array 693 void allocate_preorders() { 694 _max_preorder = C->unique()+8; 695 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 696 memset(_preorders, 0, sizeof(uint) * _max_preorder); 697 } 698 699 // Allocate _preorders[] array 700 void reallocate_preorders() { 701 if ( _max_preorder < C->unique() ) { 702 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 703 _max_preorder = C->unique(); 704 } 705 memset(_preorders, 0, sizeof(uint) * _max_preorder); 706 } 707 708 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 709 // adds new nodes. 710 void check_grow_preorders( ) { 711 if ( _max_preorder < C->unique() ) { 712 uint newsize = _max_preorder<<1; // double size of array 713 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 714 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 715 _max_preorder = newsize; 716 } 717 } 718 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 719 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 720 // Pre-order numbers are written to the Nodes array as low-bit-set values. 721 void set_preorder_visited( Node *n, int pre_order ) { 722 assert( !is_visited( n ), "already set" ); 723 _preorders[n->_idx] = (pre_order<<1); 724 }; 725 // Return pre-order number. 726 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 727 728 // Check for being post-visited. 729 // Should be previsited already (checked with assert(is_visited(n))). 730 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 731 732 // Mark as post visited 733 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 734 735 public: 736 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 737 // Returns true if "n" is a data node, false if it's a control node. 738 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 739 740 private: 741 // clear out dead code after build_loop_late 742 Node_List _deadlist; 743 744 // Support for faster execution of get_late_ctrl()/dom_lca() 745 // when a node has many uses and dominator depth is deep. 746 Node_Array _dom_lca_tags; 747 void init_dom_lca_tags(); 748 void clear_dom_lca_tags(); 749 750 // Helper for debugging bad dominance relationships 751 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 752 753 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 754 755 // Inline wrapper for frequent cases: 756 // 1) only one use 757 // 2) a use is the same as the current LCA passed as 'n1' 758 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 759 assert( n->is_CFG(), "" ); 760 // Fast-path NULL lca 761 if( lca != NULL && lca != n ) { 762 assert( lca->is_CFG(), "" ); 763 // find LCA of all uses 764 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 765 } 766 return find_non_split_ctrl(n); 767 } 768 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 769 770 // Helper function for directing control inputs away from CFG split points. 771 Node *find_non_split_ctrl( Node *ctrl ) const { 772 if (ctrl != NULL) { 773 if (ctrl->is_MultiBranch()) { 774 ctrl = ctrl->in(0); 775 } 776 assert(ctrl->is_CFG(), "CFG"); 777 } 778 return ctrl; 779 } 780 781 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 782 void duplicate_predicates_helper(Node* predicate, Node* start, Node* end, IdealLoopTree* outer_loop, 783 LoopNode* outer_main_head, uint dd_main_head); 784 void duplicate_predicates(CountedLoopNode* pre_head, Node* start, Node* end, IdealLoopTree* outer_loop, 785 LoopNode* outer_main_head, uint dd_main_head); 786 Node* clone_skeleton_predicate(Node* iff, Node* value, Node* predicate, Node* uncommon_proj, 787 Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj); 788 bool skeleton_predicate_has_opaque(IfNode* iff); 789 void update_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con); 790 void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol); 791 792 public: 793 794 PhaseIterGVN &igvn() const { return _igvn; } 795 796 static bool is_canonical_loop_entry(CountedLoopNode* cl); 797 798 bool has_node( Node* n ) const { 799 guarantee(n != NULL, "No Node."); 800 return _nodes[n->_idx] != NULL; 801 } 802 // check if transform created new nodes that need _ctrl recorded 803 Node *get_late_ctrl( Node *n, Node *early ); 804 Node *get_early_ctrl( Node *n ); 805 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 806 void set_early_ctrl( Node *n ); 807 void set_subtree_ctrl( Node *root ); 808 void set_ctrl( Node *n, Node *ctrl ) { 809 assert( !has_node(n) || has_ctrl(n), "" ); 810 assert( ctrl->in(0), "cannot set dead control node" ); 811 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 812 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 813 } 814 // Set control and update loop membership 815 void set_ctrl_and_loop(Node* n, Node* ctrl) { 816 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 817 IdealLoopTree* new_loop = get_loop(ctrl); 818 if (old_loop != new_loop) { 819 if (old_loop->_child == NULL) old_loop->_body.yank(n); 820 if (new_loop->_child == NULL) new_loop->_body.push(n); 821 } 822 set_ctrl(n, ctrl); 823 } 824 // Control nodes can be replaced or subsumed. During this pass they 825 // get their replacement Node in slot 1. Instead of updating the block 826 // location of all Nodes in the subsumed block, we lazily do it. As we 827 // pull such a subsumed block out of the array, we write back the final 828 // correct block. 829 Node *get_ctrl( Node *i ) { 830 831 assert(has_node(i), ""); 832 Node *n = get_ctrl_no_update(i); 833 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 834 assert(has_node(i) && has_ctrl(i), ""); 835 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 836 return n; 837 } 838 // true if CFG node d dominates CFG node n 839 bool is_dominator(Node *d, Node *n); 840 // return get_ctrl for a data node and self(n) for a CFG node 841 Node* ctrl_or_self(Node* n) { 842 if (has_ctrl(n)) 843 return get_ctrl(n); 844 else { 845 assert (n->is_CFG(), "must be a CFG node"); 846 return n; 847 } 848 } 849 850 Node *get_ctrl_no_update_helper(Node *i) const { 851 assert(has_ctrl(i), "should be control, not loop"); 852 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 853 } 854 855 Node *get_ctrl_no_update(Node *i) const { 856 assert( has_ctrl(i), "" ); 857 Node *n = get_ctrl_no_update_helper(i); 858 if (!n->in(0)) { 859 // Skip dead CFG nodes 860 do { 861 n = get_ctrl_no_update_helper(n); 862 } while (!n->in(0)); 863 n = find_non_split_ctrl(n); 864 } 865 return n; 866 } 867 868 // Check for loop being set 869 // "n" must be a control node. Returns true if "n" is known to be in a loop. 870 bool has_loop( Node *n ) const { 871 assert(!has_node(n) || !has_ctrl(n), ""); 872 return has_node(n); 873 } 874 // Set loop 875 void set_loop( Node *n, IdealLoopTree *loop ) { 876 _nodes.map(n->_idx, (Node*)loop); 877 } 878 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 879 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 880 // from old_node to new_node to support the lazy update. Reference 881 // replaces loop reference, since that is not needed for dead node. 882 void lazy_update(Node *old_node, Node *new_node) { 883 assert(old_node != new_node, "no cycles please"); 884 // Re-use the side array slot for this node to provide the 885 // forwarding pointer. 886 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 887 } 888 void lazy_replace(Node *old_node, Node *new_node) { 889 _igvn.replace_node(old_node, new_node); 890 lazy_update(old_node, new_node); 891 } 892 893 private: 894 895 // Place 'n' in some loop nest, where 'n' is a CFG node 896 void build_loop_tree(); 897 int build_loop_tree_impl( Node *n, int pre_order ); 898 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 899 // loop tree, not the root. 900 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 901 902 // Place Data nodes in some loop nest 903 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 904 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 905 void build_loop_late_post_work(Node* n, bool pinned); 906 void build_loop_late_post(Node* n); 907 void verify_strip_mined_scheduling(Node *n, Node* least); 908 909 // Array of immediate dominance info for each CFG node indexed by node idx 910 private: 911 uint _idom_size; 912 Node **_idom; // Array of immediate dominators 913 uint *_dom_depth; // Used for fast LCA test 914 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 915 916 // Perform verification that the graph is valid. 917 PhaseIdealLoop( PhaseIterGVN &igvn) : 918 PhaseTransform(Ideal_Loop), 919 _igvn(igvn), 920 _verify_me(NULL), 921 _verify_only(true), 922 _dom_lca_tags(arena()), // Thread::resource_area 923 _nodes_required(UINT_MAX) { 924 build_and_optimize(LoopOptsVerify); 925 } 926 927 // build the loop tree and perform any requested optimizations 928 void build_and_optimize(LoopOptsMode mode); 929 930 // Dominators for the sea of nodes 931 void Dominators(); 932 933 // Compute the Ideal Node to Loop mapping 934 PhaseIdealLoop(PhaseIterGVN &igvn, LoopOptsMode mode) : 935 PhaseTransform(Ideal_Loop), 936 _igvn(igvn), 937 _verify_me(NULL), 938 _verify_only(false), 939 _dom_lca_tags(arena()), // Thread::resource_area 940 _nodes_required(UINT_MAX) { 941 build_and_optimize(mode); 942 } 943 944 // Verify that verify_me made the same decisions as a fresh run. 945 PhaseIdealLoop(PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : 946 PhaseTransform(Ideal_Loop), 947 _igvn(igvn), 948 _verify_me(verify_me), 949 _verify_only(false), 950 _dom_lca_tags(arena()), // Thread::resource_area 951 _nodes_required(UINT_MAX) { 952 build_and_optimize(LoopOptsVerify); 953 } 954 955 public: 956 Node* idom_no_update(Node* d) const { 957 return idom_no_update(d->_idx); 958 } 959 960 Node* idom_no_update(uint didx) const { 961 assert(didx < _idom_size, "oob"); 962 Node* n = _idom[didx]; 963 assert(n != NULL,"Bad immediate dominator info."); 964 while (n->in(0) == NULL) { // Skip dead CFG nodes 965 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 966 assert(n != NULL,"Bad immediate dominator info."); 967 } 968 return n; 969 } 970 971 Node *idom(Node* d) const { 972 return idom(d->_idx); 973 } 974 975 Node *idom(uint didx) const { 976 Node *n = idom_no_update(didx); 977 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 978 return n; 979 } 980 981 uint dom_depth(Node* d) const { 982 guarantee(d != NULL, "Null dominator info."); 983 guarantee(d->_idx < _idom_size, ""); 984 return _dom_depth[d->_idx]; 985 } 986 void set_idom(Node* d, Node* n, uint dom_depth); 987 // Locally compute IDOM using dom_lca call 988 Node *compute_idom( Node *region ) const; 989 // Recompute dom_depth 990 void recompute_dom_depth(); 991 992 // Is safept not required by an outer loop? 993 bool is_deleteable_safept(Node* sfpt); 994 995 // Replace parallel induction variable (parallel to trip counter) 996 void replace_parallel_iv(IdealLoopTree *loop); 997 998 Node *dom_lca( Node *n1, Node *n2 ) const { 999 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 1000 } 1001 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 1002 1003 // Build and verify the loop tree without modifying the graph. This 1004 // is useful to verify that all inputs properly dominate their uses. 1005 static void verify(PhaseIterGVN& igvn) { 1006 #ifdef ASSERT 1007 ResourceMark rm; 1008 PhaseIdealLoop v(igvn); 1009 #endif 1010 } 1011 1012 // Recommended way to use PhaseIdealLoop. 1013 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations. 1014 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) { 1015 ResourceMark rm; 1016 PhaseIdealLoop v(igvn, mode); 1017 } 1018 1019 // True if the method has at least 1 irreducible loop 1020 bool _has_irreducible_loops; 1021 1022 // Per-Node transform 1023 virtual Node *transform( Node *a_node ) { return 0; } 1024 1025 bool is_counted_loop(Node* x, IdealLoopTree*& loop); 1026 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 1027 IdealLoopTree* loop, float cl_prob, float le_fcnt, 1028 Node*& entry_control, Node*& iffalse); 1029 1030 Node* exact_limit( IdealLoopTree *loop ); 1031 1032 // Return a post-walked LoopNode 1033 IdealLoopTree *get_loop( Node *n ) const { 1034 // Dead nodes have no loop, so return the top level loop instead 1035 if (!has_node(n)) return _ltree_root; 1036 assert(!has_ctrl(n), ""); 1037 return (IdealLoopTree*)_nodes[n->_idx]; 1038 } 1039 1040 IdealLoopTree *ltree_root() const { return _ltree_root; } 1041 1042 // Is 'n' a (nested) member of 'loop'? 1043 int is_member( const IdealLoopTree *loop, Node *n ) const { 1044 return loop->is_member(get_loop(n)); } 1045 1046 // This is the basic building block of the loop optimizations. It clones an 1047 // entire loop body. It makes an old_new loop body mapping; with this 1048 // mapping you can find the new-loop equivalent to an old-loop node. All 1049 // new-loop nodes are exactly equal to their old-loop counterparts, all 1050 // edges are the same. All exits from the old-loop now have a RegionNode 1051 // that merges the equivalent new-loop path. This is true even for the 1052 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 1053 // now come from (one or more) Phis that merge their new-loop equivalents. 1054 // Parameter side_by_side_idom: 1055 // When side_by_size_idom is NULL, the dominator tree is constructed for 1056 // the clone loop to dominate the original. Used in construction of 1057 // pre-main-post loop sequence. 1058 // When nonnull, the clone and original are side-by-side, both are 1059 // dominated by the passed in side_by_side_idom node. Used in 1060 // construction of unswitched loops. 1061 enum CloneLoopMode { 1062 IgnoreStripMined = 0, // Only clone inner strip mined loop 1063 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops 1064 ControlAroundStripMined = 2 // Only clone inner strip mined loop, 1065 // result control flow branches 1066 // either to inner clone or outer 1067 // strip mined loop. 1068 }; 1069 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 1070 CloneLoopMode mode, Node* side_by_side_idom = NULL); 1071 void clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1072 IdealLoopTree* loop, IdealLoopTree* companion_loop, 1073 Node_List*& split_if_set, Node_List*& split_bool_set, 1074 Node_List*& split_cex_set, Node_List& worklist, 1075 uint new_counter, CloneLoopMode mode); 1076 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1077 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1078 Node_List& extra_data_nodes); 1079 1080 // If we got the effect of peeling, either by actually peeling or by 1081 // making a pre-loop which must execute at least once, we can remove 1082 // all loop-invariant dominated tests in the main body. 1083 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 1084 1085 // Generate code to do a loop peel for the given loop (and body). 1086 // old_new is a temp array. 1087 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 1088 1089 // Add pre and post loops around the given loop. These loops are used 1090 // during RCE, unrolling and aligning loops. 1091 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 1092 1093 // Add post loop after the given loop. 1094 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new, 1095 CountedLoopNode *main_head, CountedLoopEndNode *main_end, 1096 Node *incr, Node *limit, CountedLoopNode *&post_head); 1097 1098 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage 1099 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new ); 1100 1101 // Add a vector post loop between a vector main loop and the current post loop 1102 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 1103 // If Node n lives in the back_ctrl block, we clone a private version of n 1104 // in preheader_ctrl block and return that, otherwise return n. 1105 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 1106 1107 // Take steps to maximally unroll the loop. Peel any odd iterations, then 1108 // unroll to do double iterations. The next round of major loop transforms 1109 // will repeat till the doubled loop body does all remaining iterations in 1 1110 // pass. 1111 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 1112 1113 // Unroll the loop body one step - make each trip do 2 iterations. 1114 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 1115 1116 // Mark vector reduction candidates before loop unrolling 1117 void mark_reductions( IdealLoopTree *loop ); 1118 1119 // Return true if exp is a constant times an induction var 1120 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); 1121 1122 // Return true if exp is a scaled induction var plus (or minus) constant 1123 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); 1124 1125 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 1126 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 1127 Deoptimization::DeoptReason reason, 1128 int opcode); 1129 void register_control(Node* n, IdealLoopTree *loop, Node* pred); 1130 1131 // Clone loop predicates to cloned loops (peeled, unswitched) 1132 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, 1133 Deoptimization::DeoptReason reason, 1134 PhaseIdealLoop* loop_phase, 1135 PhaseIterGVN* igvn); 1136 1137 static void clone_loop_predicates_fix_mem(ProjNode* dom_proj , ProjNode* proj, 1138 PhaseIdealLoop* loop_phase, 1139 PhaseIterGVN* igvn); 1140 1141 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, 1142 bool clone_limit_check, 1143 PhaseIdealLoop* loop_phase, 1144 PhaseIterGVN* igvn); 1145 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); 1146 1147 static Node* skip_all_loop_predicates(Node* entry); 1148 static Node* skip_loop_predicates(Node* entry); 1149 1150 // Find a good location to insert a predicate 1151 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); 1152 // Find a predicate 1153 static Node* find_predicate(Node* entry); 1154 // Construct a range check for a predicate if 1155 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, 1156 int scale, Node* offset, 1157 Node* init, Node* limit, jint stride, 1158 Node* range, bool upper, bool &overflow); 1159 1160 // Implementation of the loop predication to promote checks outside the loop 1161 bool loop_predication_impl(IdealLoopTree *loop); 1162 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj, 1163 CountedLoopNode *cl, ConNode* zero, Invariance& invar, 1164 Deoptimization::DeoptReason reason); 1165 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt); 1166 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, 1167 PathFrequency& pf, Node_Stack& stack, VectorSet& seen, 1168 Node_List& if_proj_list); 1169 ProjNode* insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop, 1170 ProjNode* proj, ProjNode *predicate_proj, 1171 ProjNode* upper_bound_proj, 1172 int scale, Node* offset, 1173 Node* init, Node* limit, jint stride, 1174 Node* rng, bool& overflow, 1175 Deoptimization::DeoptReason reason); 1176 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl, 1177 Node* predicate_proj, int scale_con, Node* offset, 1178 Node* limit, jint stride_con, Node* value); 1179 1180 // Helper function to collect predicate for eliminating the useless ones 1181 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); 1182 void eliminate_useless_predicates(); 1183 1184 // Change the control input of expensive nodes to allow commoning by 1185 // IGVN when it is guaranteed to not result in a more frequent 1186 // execution of the expensive node. Return true if progress. 1187 bool process_expensive_nodes(); 1188 1189 // Check whether node has become unreachable 1190 bool is_node_unreachable(Node *n) const { 1191 return !has_node(n) || n->is_unreachable(_igvn); 1192 } 1193 1194 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1195 int do_range_check( IdealLoopTree *loop, Node_List &old_new ); 1196 1197 // Check to see if do_range_check(...) cleaned the main loop of range-checks 1198 void has_range_checks(IdealLoopTree *loop); 1199 1200 // Process post loops which have range checks and try to build a multi-version 1201 // guard to safely determine if we can execute the post loop which was RCE'd. 1202 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop); 1203 1204 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning 1205 void poison_rce_post_loop(IdealLoopTree *rce_loop); 1206 1207 // Create a slow version of the loop by cloning the loop 1208 // and inserting an if to select fast-slow versions. 1209 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, 1210 Node_List &old_new, 1211 int opcode, 1212 CloneLoopMode mode); 1213 1214 // Clone a loop and return the clone head (clone_loop_head). 1215 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse, 1216 // This routine was created for usage in CountedLoopReserveKit. 1217 // 1218 // int(1) -> If -> IfTrue -> original_loop_head 1219 // | 1220 // V 1221 // IfFalse -> clone_loop_head (returned by function pointer) 1222 // 1223 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk); 1224 // Clone loop with an invariant test (that does not exit) and 1225 // insert a clone of the test that selects which version to 1226 // execute. 1227 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 1228 1229 // Find candidate "if" for unswitching 1230 IfNode* find_unswitching_candidate(const IdealLoopTree *loop, Node_List& flattened_checks) const; 1231 1232 // Range Check Elimination uses this function! 1233 // Constrain the main loop iterations so the affine function: 1234 // low_limit <= scale_con * I + offset < upper_limit 1235 // always holds true. That is, either increase the number of iterations in 1236 // the pre-loop or the post-loop until the condition holds true in the main 1237 // loop. Scale_con, offset and limit are all loop invariant. 1238 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 ); 1239 // Helper function for add_constraint(). 1240 Node* adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up); 1241 1242 // Partially peel loop up through last_peel node. 1243 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1244 1245 // Create a scheduled list of nodes control dependent on ctrl set. 1246 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1247 // Has a use in the vector set 1248 bool has_use_in_set( Node* n, VectorSet& vset ); 1249 // Has use internal to the vector set (ie. not in a phi at the loop head) 1250 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1251 // clone "n" for uses that are outside of loop 1252 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1253 // clone "n" for special uses that are in the not_peeled region 1254 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1255 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1256 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1257 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1258 #ifdef ASSERT 1259 // Validate the loop partition sets: peel and not_peel 1260 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1261 // Ensure that uses outside of loop are of the right form 1262 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1263 uint orig_exit_idx, uint clone_exit_idx); 1264 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1265 #endif 1266 1267 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1268 int stride_of_possible_iv( Node* iff ); 1269 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1270 // Return the (unique) control output node that's in the loop (if it exists.) 1271 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1272 // Insert a signed compare loop exit cloned from an unsigned compare. 1273 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1274 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1275 // Utility to register node "n" with PhaseIdealLoop 1276 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 1277 // Utility to create an if-projection 1278 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1279 // Force the iff control output to be the live_proj 1280 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1281 // Insert a region before an if projection 1282 RegionNode* insert_region_before_proj(ProjNode* proj); 1283 // Insert a new if before an if projection 1284 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1285 1286 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1287 // "Nearly" because all Nodes have been cloned from the original in the loop, 1288 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1289 // through the Phi recursively, and return a Bool. 1290 Node *clone_iff( PhiNode *phi, IdealLoopTree *loop ); 1291 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); 1292 1293 1294 // Rework addressing expressions to get the most loop-invariant stuff 1295 // moved out. We'd like to do all associative operators, but it's especially 1296 // important (common) to do address expressions. 1297 Node *remix_address_expressions( Node *n ); 1298 1299 // Convert add to muladd to generate MuladdS2I under certain criteria 1300 Node * convert_add_to_muladd(Node * n); 1301 1302 // Attempt to use a conditional move instead of a phi/branch 1303 Node *conditional_move( Node *n ); 1304 1305 // Reorganize offset computations to lower register pressure. 1306 // Mostly prevent loop-fallout uses of the pre-incremented trip counter 1307 // (which are then alive with the post-incremented trip counter 1308 // forcing an extra register move) 1309 void reorg_offsets( IdealLoopTree *loop ); 1310 1311 // Check for aggressive application of 'split-if' optimization, 1312 // using basic block level info. 1313 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack); 1314 Node *split_if_with_blocks_pre ( Node *n ); 1315 void split_if_with_blocks_post( Node *n ); 1316 Node *has_local_phi_input( Node *n ); 1317 // Mark an IfNode as being dominated by a prior test, 1318 // without actually altering the CFG (and hence IDOM info). 1319 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); 1320 1321 // Split Node 'n' through merge point 1322 Node *split_thru_region( Node *n, Node *region ); 1323 // Split Node 'n' through merge point if there is enough win. 1324 Node *split_thru_phi( Node *n, Node *region, int policy ); 1325 // Found an If getting its condition-code input from a Phi in the 1326 // same block. Split thru the Region. 1327 void do_split_if( Node *iff ); 1328 1329 // Conversion of fill/copy patterns into intrisic versions 1330 bool do_intrinsify_fill(); 1331 bool intrinsify_fill(IdealLoopTree* lpt); 1332 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1333 Node*& shift, Node*& offset); 1334 1335 private: 1336 // Return a type based on condition control flow 1337 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1338 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 1339 // Helpers for filtered type 1340 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1341 1342 // Helper functions 1343 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1344 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1345 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 ); 1346 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1347 void sink_use( Node *use, Node *post_loop ); 1348 Node *place_near_use( Node *useblock ) const; 1349 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1350 void try_move_store_after_loop(Node* n); 1351 bool identical_backtoback_ifs(Node *n); 1352 bool can_split_if(Node *n_ctrl); 1353 1354 // Determine if a method is too big for a/another round of split-if, based on 1355 // a magic (approximate) ratio derived from the equally magic constant 35000, 1356 // previously used for this purpose (but without relating to the node limit). 1357 bool must_throttle_split_if() { 1358 uint threshold = C->max_node_limit() * 2 / 5; 1359 return C->live_nodes() > threshold; 1360 } 1361 1362 // A simplistic node request tracking mechanism, where 1363 // = UINT_MAX Request not valid or made final. 1364 // < UINT_MAX Nodes currently requested (estimate). 1365 uint _nodes_required; 1366 1367 enum { REQUIRE_MIN = 70 }; 1368 1369 uint nodes_required() const { return _nodes_required; } 1370 1371 // Given the _currently_ available number of nodes, check whether there is 1372 // "room" for an additional request or not, considering the already required 1373 // number of nodes. Return TRUE if the new request is exceeding the node 1374 // budget limit, otherwise return FALSE. Note that this interpretation will 1375 // act pessimistic on additional requests when new nodes have already been 1376 // generated since the 'begin'. This behaviour fits with the intention that 1377 // node estimates/requests should be made upfront. 1378 bool exceeding_node_budget(uint required = 0) { 1379 assert(C->live_nodes() < C->max_node_limit(), "sanity"); 1380 uint available = C->max_node_limit() - C->live_nodes(); 1381 return available < required + _nodes_required; 1382 } 1383 1384 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1385 precond(require > 0); 1386 _nodes_required += MAX2(require, minreq); 1387 return _nodes_required; 1388 } 1389 1390 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1391 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0; 1392 } 1393 1394 uint require_nodes_begin() { 1395 assert(_nodes_required == UINT_MAX, "Bad state (begin)."); 1396 _nodes_required = 0; 1397 return C->live_nodes(); 1398 } 1399 1400 // When a node request is final, optionally check that the requested number 1401 // of nodes was reasonably correct with respect to the number of new nodes 1402 // introduced since the last 'begin'. Always check that we have not exceeded 1403 // the maximum node limit. 1404 void require_nodes_final(uint live_at_begin, bool check_estimate) { 1405 assert(_nodes_required < UINT_MAX, "Bad state (final)."); 1406 1407 if (check_estimate) { 1408 // Assert that the node budget request was not off by too much (x2). 1409 // Should this be the case we _surely_ need to improve the estimates 1410 // used in our budget calculations. 1411 assert(C->live_nodes() - live_at_begin <= 2 * _nodes_required, 1412 "Bad node estimate: actual = %d >> request = %d", 1413 C->live_nodes() - live_at_begin, _nodes_required); 1414 } 1415 // Assert that we have stayed within the node budget limit. 1416 assert(C->live_nodes() < C->max_node_limit(), 1417 "Exceeding node budget limit: %d + %d > %d (request = %d)", 1418 C->live_nodes() - live_at_begin, live_at_begin, 1419 C->max_node_limit(), _nodes_required); 1420 1421 _nodes_required = UINT_MAX; 1422 } 1423 1424 bool _created_loop_node; 1425 1426 public: 1427 void set_created_loop_node() { _created_loop_node = true; } 1428 bool created_loop_node() { return _created_loop_node; } 1429 void register_new_node( Node *n, Node *blk ); 1430 1431 #ifdef ASSERT 1432 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1433 #endif 1434 1435 #ifndef PRODUCT 1436 void dump( ) const; 1437 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; 1438 void verify() const; // Major slow :-) 1439 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; 1440 IdealLoopTree *get_loop_idx(Node* n) const { 1441 // Dead nodes have no loop, so return the top level loop instead 1442 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 1443 } 1444 // Print some stats 1445 static void print_statistics(); 1446 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1447 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1448 #endif 1449 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; 1450 }; 1451 1452 1453 class AutoNodeBudget : public StackObj 1454 { 1455 public: 1456 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK }; 1457 1458 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK) 1459 : _phase(phase), 1460 _check_at_final(chk == BUDGET_CHECK), 1461 _nodes_at_begin(0) 1462 { 1463 precond(_phase != NULL); 1464 1465 _nodes_at_begin = _phase->require_nodes_begin(); 1466 } 1467 1468 ~AutoNodeBudget() { 1469 #ifndef PRODUCT 1470 if (TraceLoopOpts) { 1471 uint request = _phase->nodes_required(); 1472 uint delta = _phase->C->live_nodes() - _nodes_at_begin; 1473 1474 if (request < delta) { 1475 tty->print_cr("Exceeding node budget: %d < %d", request, delta); 1476 } else { 1477 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN; 1478 // Identify the worst estimates as "poor" ones. 1479 if (request > REQUIRE_MIN && delta > 0) { 1480 if ((delta > REQUIRE_MIN && request > 3 * delta) || 1481 (delta <= REQUIRE_MIN && request > 10 * delta)) { 1482 tty->print_cr("Poor node estimate: %d >> %d", request, delta); 1483 } 1484 } 1485 } 1486 } 1487 #endif // PRODUCT 1488 _phase->require_nodes_final(_nodes_at_begin, _check_at_final); 1489 } 1490 1491 private: 1492 PhaseIdealLoop* _phase; 1493 bool _check_at_final; 1494 uint _nodes_at_begin; 1495 }; 1496 1497 1498 // This kit may be used for making of a reserved copy of a loop before this loop 1499 // goes under non-reversible changes. 1500 // 1501 // Function create_reserve() creates a reserved copy (clone) of the loop. 1502 // The reserved copy is created by calling 1503 // PhaseIdealLoop::create_reserve_version_of_loop - see there how 1504 // the original and reserved loops are connected in the outer graph. 1505 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'. 1506 // 1507 // By default the reserved copy (clone) of the loop is created as dead code - it is 1508 // dominated in the outer loop by this node chain: 1509 // intcon(1)->If->IfFalse->reserved_copy. 1510 // The original loop is dominated by the the same node chain but IfTrue projection: 1511 // intcon(0)->If->IfTrue->original_loop. 1512 // 1513 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve() 1514 // and the dtor, checks _use_new value. 1515 // If _use_new == false, it "switches" control to reserved copy of the loop 1516 // by simple replacing of node intcon(1) with node intcon(0). 1517 // 1518 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp). 1519 // 1520 // void CountedLoopReserveKit_example() 1521 // { 1522 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object 1523 // if (DoReserveCopy && !lrk.has_reserved()) { 1524 // return; //failed to create reserved loop copy 1525 // } 1526 // ... 1527 // //something is wrong, switch to original loop 1528 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch 1529 // ... 1530 // //everything worked ok, return with the newly modified loop 1531 // lrk.use_new(); 1532 // return; // ~CountedLoopReserveKit does nothing once use_new() was called 1533 // } 1534 // 1535 // Keep in mind, that by default if create_reserve() is not followed by use_new() 1536 // the dtor will "switch to the original" loop. 1537 // NOTE. You you modify outside of the original loop this class is no help. 1538 // 1539 class CountedLoopReserveKit { 1540 private: 1541 PhaseIdealLoop* _phase; 1542 IdealLoopTree* _lpt; 1543 LoopNode* _lp; 1544 IfNode* _iff; 1545 LoopNode* _lp_reserved; 1546 bool _has_reserved; 1547 bool _use_new; 1548 const bool _active; //may be set to false in ctor, then the object is dummy 1549 1550 public: 1551 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active); 1552 ~CountedLoopReserveKit(); 1553 void use_new() {_use_new = true;} 1554 void set_iff(IfNode* x) {_iff = x;} 1555 bool has_reserved() const { return _active && _has_reserved;} 1556 private: 1557 bool create_reserve(); 1558 };// class CountedLoopReserveKit 1559 1560 inline Node* IdealLoopTree::tail() { 1561 // Handle lazy update of _tail field. 1562 if (_tail->in(0) == NULL) { 1563 _tail = _phase->get_ctrl(_tail); 1564 } 1565 return _tail; 1566 } 1567 1568 1569 // Iterate over the loop tree using a preorder, left-to-right traversal. 1570 // 1571 // Example that visits all counted loops from within PhaseIdealLoop 1572 // 1573 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1574 // IdealLoopTree* lpt = iter.current(); 1575 // if (!lpt->is_counted()) continue; 1576 // ... 1577 class LoopTreeIterator : public StackObj { 1578 private: 1579 IdealLoopTree* _root; 1580 IdealLoopTree* _curnt; 1581 1582 public: 1583 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1584 1585 bool done() { return _curnt == NULL; } // Finished iterating? 1586 1587 void next(); // Advance to next loop tree 1588 1589 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1590 }; 1591 1592 #endif // SHARE_OPTO_LOOPNODE_HPP