1 /* 2 * Copyright (c) 1998, 2025, 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/predicates.hpp" 32 #include "opto/subnode.hpp" 33 #include "opto/type.hpp" 34 #include "utilities/checkedCast.hpp" 35 36 class CmpNode; 37 class BaseCountedLoopEndNode; 38 class CountedLoopNode; 39 class IdealLoopTree; 40 class LoopNode; 41 class Node; 42 class OuterStripMinedLoopEndNode; 43 class PredicateBlock; 44 class PathFrequency; 45 class PhaseIdealLoop; 46 class UnswitchCandidate; 47 class LoopSelector; 48 class UnswitchedLoopSelector; 49 class VectorSet; 50 class VSharedData; 51 class Invariance; 52 struct small_cache; 53 54 // 55 // I D E A L I Z E D L O O P S 56 // 57 // Idealized loops are the set of loops I perform more interesting 58 // transformations on, beyond simple hoisting. 59 60 //------------------------------LoopNode--------------------------------------- 61 // Simple loop header. Fall in path on left, loop-back path on right. 62 class LoopNode : public RegionNode { 63 // Size is bigger to hold the flags. However, the flags do not change 64 // the semantics so it does not appear in the hash & cmp functions. 65 virtual uint size_of() const { return sizeof(*this); } 66 protected: 67 uint _loop_flags; 68 // Names for flag bitfields 69 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 70 MainHasNoPreLoop = 1<<2, 71 HasExactTripCount = 1<<3, 72 InnerLoop = 1<<4, 73 PartialPeelLoop = 1<<5, 74 PartialPeelFailed = 1<<6, 75 WasSlpAnalyzed = 1<<7, 76 PassedSlpAnalysis = 1<<8, 77 DoUnrollOnly = 1<<9, 78 VectorizedLoop = 1<<10, 79 HasAtomicPostLoop = 1<<11, 80 StripMined = 1<<12, 81 SubwordLoop = 1<<13, 82 ProfileTripFailed = 1<<14, 83 LoopNestInnerLoop = 1<<15, 84 LoopNestLongOuterLoop = 1<<16, 85 MultiversionFastLoop = 1<<17, 86 MultiversionSlowLoop = 2<<17, 87 MultiversionDelayedSlowLoop = 3<<17, 88 MultiversionFlagsMask = 3<<17, 89 FlatArrays = 1<<18}; 90 char _unswitch_count; 91 enum { _unswitch_max=3 }; 92 93 // Expected trip count from profile data 94 float _profile_trip_cnt; 95 96 public: 97 // Names for edge indices 98 enum { Self=0, EntryControl, LoopBackControl }; 99 100 bool is_inner_loop() const { return _loop_flags & InnerLoop; } 101 void set_inner_loop() { _loop_flags |= InnerLoop; } 102 103 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } 104 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 105 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 106 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 107 bool is_strip_mined() const { return _loop_flags & StripMined; } 108 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } 109 bool is_subword_loop() const { return _loop_flags & SubwordLoop; } 110 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; } 111 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; } 112 bool is_flat_arrays() const { return _loop_flags & FlatArrays; } 113 114 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 115 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 116 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 117 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 118 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 119 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 120 void mark_strip_mined() { _loop_flags |= StripMined; } 121 void clear_strip_mined() { _loop_flags &= ~StripMined; } 122 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } 123 void mark_subword_loop() { _loop_flags |= SubwordLoop; } 124 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; } 125 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; } 126 void mark_flat_arrays() { _loop_flags |= FlatArrays; } 127 128 int unswitch_max() { return _unswitch_max; } 129 int unswitch_count() { return _unswitch_count; } 130 131 void set_unswitch_count(int val) { 132 assert (val <= unswitch_max(), "too many unswitches"); 133 _unswitch_count = val; 134 } 135 136 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 137 float profile_trip_cnt() { return _profile_trip_cnt; } 138 139 #ifndef PRODUCT 140 uint _stress_peeling_attempts = 0; 141 #endif 142 143 LoopNode(Node *entry, Node *backedge) 144 : RegionNode(3), _loop_flags(0), _unswitch_count(0), 145 _profile_trip_cnt(COUNT_UNKNOWN) { 146 init_class_id(Class_Loop); 147 init_req(EntryControl, entry); 148 init_req(LoopBackControl, backedge); 149 } 150 151 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 152 virtual int Opcode() const; 153 bool can_be_counted_loop(PhaseValues* phase) const { 154 return req() == 3 && in(0) != nullptr && 155 in(1) != nullptr && phase->type(in(1)) != Type::TOP && 156 in(2) != nullptr && phase->type(in(2)) != Type::TOP; 157 } 158 bool is_valid_counted_loop(BasicType bt) const; 159 #ifndef PRODUCT 160 virtual void dump_spec(outputStream *st) const; 161 #endif 162 163 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN; 164 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } 165 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; } 166 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; } 167 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; } 168 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; } 169 }; 170 171 //------------------------------Counted Loops---------------------------------- 172 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 173 // path (and maybe some other exit paths). The trip-counter exit is always 174 // last in the loop. The trip-counter have to stride by a constant; 175 // the exit value is also loop invariant. 176 177 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 178 // CountedLoopNode has the incoming loop control and the loop-back-control 179 // which is always the IfTrue before the matching CountedLoopEndNode. The 180 // CountedLoopEndNode has an incoming control (possibly not the 181 // CountedLoopNode if there is control flow in the loop), the post-increment 182 // trip-counter value, and the limit. The trip-counter value is always of 183 // the form (Op old-trip-counter stride). The old-trip-counter is produced 184 // by a Phi connected to the CountedLoopNode. The stride is constant. 185 // The Op is any commutable opcode, including Add, Mul, Xor. The 186 // CountedLoopEndNode also takes in the loop-invariant limit value. 187 188 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 189 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 190 // via the old-trip-counter from the Op node. 191 192 //------------------------------CountedLoopNode-------------------------------- 193 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 194 // inputs the incoming loop-start control and the loop-back control, so they 195 // act like RegionNodes. They also take in the initial trip counter, the 196 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 197 // produce a loop-body control and the trip counter value. Since 198 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 199 200 class BaseCountedLoopNode : public LoopNode { 201 public: 202 BaseCountedLoopNode(Node *entry, Node *backedge) 203 : LoopNode(entry, backedge) { 204 } 205 206 Node *init_control() const { return in(EntryControl); } 207 Node *back_control() const { return in(LoopBackControl); } 208 209 Node* init_trip() const; 210 Node* stride() const; 211 bool stride_is_con() const; 212 Node* limit() const; 213 Node* incr() const; 214 Node* phi() const; 215 216 BaseCountedLoopEndNode* loopexit_or_null() const; 217 BaseCountedLoopEndNode* loopexit() const; 218 219 virtual BasicType bt() const = 0; 220 221 jlong stride_con() const; 222 223 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt); 224 }; 225 226 227 class CountedLoopNode : public BaseCountedLoopNode { 228 // Size is bigger to hold _main_idx. However, _main_idx does not change 229 // the semantics so it does not appear in the hash & cmp functions. 230 virtual uint size_of() const { return sizeof(*this); } 231 232 // For Pre- and Post-loops during debugging ONLY, this holds the index of 233 // the Main CountedLoop. Used to assert that we understand the graph shape. 234 node_idx_t _main_idx; 235 236 // Known trip count calculated by compute_exact_trip_count() 237 uint _trip_count; 238 239 // Log2 of original loop bodies in unrolled loop 240 int _unrolled_count_log2; 241 242 // Node count prior to last unrolling - used to decide if 243 // unroll,optimize,unroll,optimize,... is making progress 244 int _node_count_before_unroll; 245 246 // If slp analysis is performed we record the maximum 247 // vector mapped unroll factor here 248 int _slp_maximum_unroll_factor; 249 250 public: 251 CountedLoopNode(Node *entry, Node *backedge) 252 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 253 _unrolled_count_log2(0), _node_count_before_unroll(0), 254 _slp_maximum_unroll_factor(0) { 255 init_class_id(Class_CountedLoop); 256 // Initialize _trip_count to the largest possible value. 257 // Will be reset (lower) if the loop's trip count is known. 258 } 259 260 virtual int Opcode() const; 261 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 262 263 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 264 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 265 int stride_con() const; 266 267 // Match increment with optional truncation 268 static Node* 269 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type, 270 BasicType bt); 271 272 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 273 // can run short a few iterations and may start a few iterations in. 274 // It will be RCE'd and unrolled and aligned. 275 276 // A following 'post' loop will run any remaining iterations. Used 277 // during Range Check Elimination, the 'post' loop will do any final 278 // iterations with full checks. Also used by Loop Unrolling, where 279 // the 'post' loop will do any epilog iterations needed. Basically, 280 // a 'post' loop can not profitably be further unrolled or RCE'd. 281 282 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 283 // it may do under-flow checks for RCE and may do alignment iterations 284 // so the following main loop 'knows' that it is striding down cache 285 // lines. 286 287 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 288 // Aligned, may be missing it's pre-loop. 289 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 290 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 291 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 292 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 293 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 294 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 295 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 296 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 297 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 298 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 299 300 IfNode* find_multiversion_if_from_multiversion_fast_main_loop(); 301 302 int main_idx() const { return _main_idx; } 303 304 305 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 306 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 307 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 308 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 309 310 // We use max_juint for the default value of _trip_count to signal it wasn't set. 311 // We shouldn't set _trip_count to max_juint explicitly. 312 void set_trip_count(uint tc) { assert(tc < max_juint, "Cannot set trip count to max_juint"); _trip_count = tc; } 313 uint trip_count() { return _trip_count; } 314 315 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 316 void set_exact_trip_count(uint tc) { 317 assert(tc < max_juint, "Cannot set trip count to max_juint"); 318 _trip_count = tc; 319 _loop_flags |= HasExactTripCount; 320 } 321 void set_nonexact_trip_count() { 322 _loop_flags &= ~HasExactTripCount; 323 } 324 void set_notpassed_slp() { 325 _loop_flags &= ~PassedSlpAnalysis; 326 } 327 328 void double_unrolled_count() { _unrolled_count_log2++; } 329 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 330 331 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 332 int node_count_before_unroll() { return _node_count_before_unroll; } 333 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 334 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 335 336 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if 337 // decides which one is taken: 338 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to 339 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption 340 // for an optimization, we can add the runtime-check to the multiversion_if, and if the 341 // assumption fails we take the slow_loop instead, where we do not make the same speculative 342 // assumption. 343 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative 344 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster. 345 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while 346 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts 347 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save 348 // compile time, we delay the optimization of the slow_loop until a runtime-check is added 349 // to the multiversion_if, at which point we resume optimizations for the slow_loop. 350 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back 351 // loop where we do not make any of the speculative assumptions we make for the fast_loop. 352 // Hence, we expect the slow_loop to execute slower. 353 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; } 354 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; } 355 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; } 356 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; } 357 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; } 358 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; } 359 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; } 360 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; } 361 362 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1); 363 OuterStripMinedLoopNode* outer_loop() const; 364 virtual IfTrueNode* outer_loop_tail() const; 365 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 366 virtual IfFalseNode* outer_loop_exit() const; 367 virtual SafePointNode* outer_safepoint() const; 368 369 Node* skip_assertion_predicates_with_halt(); 370 371 virtual BasicType bt() const { 372 return T_INT; 373 } 374 375 Node* is_canonical_loop_entry(); 376 CountedLoopEndNode* find_pre_loop_end(); 377 378 Node* uncasted_init_trip(bool uncasted); 379 380 #ifndef PRODUCT 381 virtual void dump_spec(outputStream *st) const; 382 #endif 383 }; 384 385 class LongCountedLoopNode : public BaseCountedLoopNode { 386 public: 387 LongCountedLoopNode(Node *entry, Node *backedge) 388 : BaseCountedLoopNode(entry, backedge) { 389 init_class_id(Class_LongCountedLoop); 390 } 391 392 virtual int Opcode() const; 393 394 virtual BasicType bt() const { 395 return T_LONG; 396 } 397 398 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 399 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 400 }; 401 402 403 //------------------------------CountedLoopEndNode----------------------------- 404 // CountedLoopEndNodes end simple trip counted loops. They act much like 405 // IfNodes. 406 407 class BaseCountedLoopEndNode : public IfNode { 408 public: 409 enum { TestControl, TestValue }; 410 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 411 : IfNode(control, test, prob, cnt) { 412 init_class_id(Class_BaseCountedLoopEnd); 413 } 414 415 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; } 416 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; } 417 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; } 418 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; } 419 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; } 420 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); } 421 422 PhiNode* phi() const { 423 Node* tmp = incr(); 424 if (tmp && tmp->req() == 3) { 425 Node* phi = tmp->in(1); 426 if (phi->is_Phi()) { 427 return phi->as_Phi(); 428 } 429 } 430 return nullptr; 431 } 432 433 BaseCountedLoopNode* loopnode() const { 434 // The CountedLoopNode that goes with this CountedLoopEndNode may 435 // have been optimized out by the IGVN so be cautious with the 436 // pattern matching on the graph 437 PhiNode* iv_phi = phi(); 438 if (iv_phi == nullptr) { 439 return nullptr; 440 } 441 Node* ln = iv_phi->in(0); 442 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) { 443 return nullptr; 444 } 445 if (ln->as_BaseCountedLoop()->bt() != bt()) { 446 return nullptr; 447 } 448 return ln->as_BaseCountedLoop(); 449 } 450 451 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 452 453 jlong stride_con() const; 454 virtual BasicType bt() const = 0; 455 456 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt); 457 }; 458 459 class CountedLoopEndNode : public BaseCountedLoopEndNode { 460 public: 461 462 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 463 : BaseCountedLoopEndNode(control, test, prob, cnt) { 464 init_class_id(Class_CountedLoopEnd); 465 } 466 virtual int Opcode() const; 467 468 CountedLoopNode* loopnode() const { 469 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 470 } 471 472 virtual BasicType bt() const { 473 return T_INT; 474 } 475 476 #ifndef PRODUCT 477 virtual void dump_spec(outputStream *st) const; 478 #endif 479 }; 480 481 class LongCountedLoopEndNode : public BaseCountedLoopEndNode { 482 public: 483 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 484 : BaseCountedLoopEndNode(control, test, prob, cnt) { 485 init_class_id(Class_LongCountedLoopEnd); 486 } 487 488 LongCountedLoopNode* loopnode() const { 489 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 490 } 491 492 virtual int Opcode() const; 493 494 virtual BasicType bt() const { 495 return T_LONG; 496 } 497 }; 498 499 500 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const { 501 Node* bctrl = back_control(); 502 if (bctrl == nullptr) return nullptr; 503 504 Node* lexit = bctrl->in(0); 505 if (!lexit->is_BaseCountedLoopEnd()) { 506 return nullptr; 507 } 508 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd(); 509 if (result->bt() != bt()) { 510 return nullptr; 511 } 512 return result; 513 } 514 515 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const { 516 BaseCountedLoopEndNode* cle = loopexit_or_null(); 517 assert(cle != nullptr, "loopexit is null"); 518 return cle; 519 } 520 521 inline Node* BaseCountedLoopNode::init_trip() const { 522 BaseCountedLoopEndNode* cle = loopexit_or_null(); 523 return cle != nullptr ? cle->init_trip() : nullptr; 524 } 525 inline Node* BaseCountedLoopNode::stride() const { 526 BaseCountedLoopEndNode* cle = loopexit_or_null(); 527 return cle != nullptr ? cle->stride() : nullptr; 528 } 529 530 inline bool BaseCountedLoopNode::stride_is_con() const { 531 BaseCountedLoopEndNode* cle = loopexit_or_null(); 532 return cle != nullptr && cle->stride_is_con(); 533 } 534 inline Node* BaseCountedLoopNode::limit() const { 535 BaseCountedLoopEndNode* cle = loopexit_or_null(); 536 return cle != nullptr ? cle->limit() : nullptr; 537 } 538 inline Node* BaseCountedLoopNode::incr() const { 539 BaseCountedLoopEndNode* cle = loopexit_or_null(); 540 return cle != nullptr ? cle->incr() : nullptr; 541 } 542 inline Node* BaseCountedLoopNode::phi() const { 543 BaseCountedLoopEndNode* cle = loopexit_or_null(); 544 return cle != nullptr ? cle->phi() : nullptr; 545 } 546 547 inline jlong BaseCountedLoopNode::stride_con() const { 548 BaseCountedLoopEndNode* cle = loopexit_or_null(); 549 return cle != nullptr ? cle->stride_con() : 0; 550 } 551 552 553 //------------------------------LoopLimitNode----------------------------- 554 // Counted Loop limit node which represents exact final iterator value: 555 // trip_count = (limit - init_trip + stride - 1)/stride 556 // final_value= trip_count * stride + init_trip. 557 // Use HW instructions to calculate it when it can overflow in integer. 558 // Note, final_value should fit into integer since counted loop has 559 // limit check: limit <= max_int-stride. 560 class LoopLimitNode : public Node { 561 enum { Init=1, Limit=2, Stride=3 }; 562 public: 563 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) { 564 // Put it on the Macro nodes list to optimize during macro nodes expansion. 565 init_flags(Flag_is_macro); 566 C->add_macro_node(this); 567 } 568 virtual int Opcode() const; 569 virtual const Type *bottom_type() const { return TypeInt::INT; } 570 virtual uint ideal_reg() const { return Op_RegI; } 571 virtual const Type* Value(PhaseGVN* phase) const; 572 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 573 virtual Node* Identity(PhaseGVN* phase); 574 }; 575 576 // Support for strip mining 577 class OuterStripMinedLoopNode : public LoopNode { 578 private: 579 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const; 580 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn); 581 582 public: 583 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) 584 : LoopNode(entry, backedge) { 585 init_class_id(Class_OuterStripMinedLoop); 586 init_flags(Flag_is_macro); 587 C->add_macro_node(this); 588 } 589 590 virtual int Opcode() const; 591 592 virtual IfTrueNode* outer_loop_tail() const; 593 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 594 virtual IfFalseNode* outer_loop_exit() const; 595 virtual SafePointNode* outer_safepoint() const; 596 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); } 597 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); } 598 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->proj_out(false)->as_IfFalse(); } 599 600 void adjust_strip_mined_loop(PhaseIterGVN* igvn); 601 602 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const; 603 604 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 605 606 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 607 608 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn, 609 PhaseIdealLoop* iloop); 610 }; 611 612 class OuterStripMinedLoopEndNode : public IfNode { 613 public: 614 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) 615 : IfNode(control, test, prob, cnt) { 616 init_class_id(Class_OuterStripMinedLoopEnd); 617 } 618 619 virtual int Opcode() const; 620 621 virtual const Type* Value(PhaseGVN* phase) const; 622 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 623 624 bool is_expanded(PhaseGVN *phase) const; 625 }; 626 627 // -----------------------------IdealLoopTree---------------------------------- 628 class IdealLoopTree : public ResourceObj { 629 public: 630 IdealLoopTree *_parent; // Parent in loop tree 631 IdealLoopTree *_next; // Next sibling in loop tree 632 IdealLoopTree *_child; // First child in loop tree 633 634 // The head-tail backedge defines the loop. 635 // If a loop has multiple backedges, this is addressed during cleanup where 636 // we peel off the multiple backedges, merging all edges at the bottom and 637 // ensuring that one proper backedge flow into the loop. 638 Node *_head; // Head of loop 639 Node *_tail; // Tail of loop 640 inline Node *tail(); // Handle lazy update of _tail field 641 inline Node *head(); // Handle lazy update of _head field 642 PhaseIdealLoop* _phase; 643 int _local_loop_unroll_limit; 644 int _local_loop_unroll_factor; 645 646 Node_List _body; // Loop body for inner loops 647 648 uint16_t _nest; // Nesting depth 649 uint8_t _irreducible:1, // True if irreducible 650 _has_call:1, // True if has call safepoint 651 _has_sfpt:1, // True if has non-call safepoint 652 _rce_candidate:1, // True if candidate for range check elimination 653 _has_range_checks:1, 654 _has_range_checks_computed:1; 655 656 Node_List* _safepts; // List of safepoints in this loop 657 Node_List* _required_safept; // A inner loop cannot delete these safepts; 658 bool _allow_optimizations; // Allow loop optimizations 659 660 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 661 : _parent(nullptr), _next(nullptr), _child(nullptr), 662 _head(head), _tail(tail), 663 _phase(phase), 664 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0), 665 _body(Compile::current()->comp_arena()), 666 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 667 _has_range_checks(0), _has_range_checks_computed(0), 668 _safepts(nullptr), 669 _required_safept(nullptr), 670 _allow_optimizations(true) 671 { 672 precond(_head != nullptr); 673 precond(_tail != nullptr); 674 } 675 676 // Is 'l' a member of 'this'? 677 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 678 679 // Set loop nesting depth. Accumulate has_call bits. 680 int set_nest( uint depth ); 681 682 // Split out multiple fall-in edges from the loop header. Move them to a 683 // private RegionNode before the loop. This becomes the loop landing pad. 684 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 685 686 // Split out the outermost loop from this shared header. 687 void split_outer_loop( PhaseIdealLoop *phase ); 688 689 // Merge all the backedges from the shared header into a private Region. 690 // Feed that region as the one backedge to this loop. 691 void merge_many_backedges( PhaseIdealLoop *phase ); 692 693 // Split shared headers and insert loop landing pads. 694 // Insert a LoopNode to replace the RegionNode. 695 // Returns TRUE if loop tree is structurally changed. 696 bool beautify_loops( PhaseIdealLoop *phase ); 697 698 // Perform optimization to use the loop predicates for null checks and range checks. 699 // Applies to any loop level (not just the innermost one) 700 bool loop_predication( PhaseIdealLoop *phase); 701 bool can_apply_loop_predication(); 702 703 // Perform iteration-splitting on inner loops. Split iterations to 704 // avoid range checks or one-shot null checks. Returns false if the 705 // current round of loop opts should stop. 706 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 707 708 // Driver for various flavors of iteration splitting. Returns false 709 // if the current round of loop opts should stop. 710 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 711 712 // Given dominators, try to find loops with calls that must always be 713 // executed (call dominates loop tail). These loops do not need non-call 714 // safepoints (ncsfpt). 715 void check_safepts(VectorSet &visited, Node_List &stack); 716 717 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 718 // encountered. 719 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 720 721 // Remove safepoints from loop. Optionally keeping one. 722 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 723 724 // Convert to counted loops where possible 725 void counted_loop( PhaseIdealLoop *phase ); 726 727 // Check for Node being a loop-breaking test 728 Node *is_loop_exit(Node *iff) const; 729 730 // Remove simplistic dead code from loop body 731 void DCE_loop_body(); 732 733 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 734 // Replace with a 1-in-10 exit guess. 735 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 736 737 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 738 // Useful for unrolling loops with NO array accesses. 739 bool policy_peel_only( PhaseIdealLoop *phase ) const; 740 741 // Return TRUE or FALSE if the loop should be unswitched -- clone 742 // loop with an invariant test 743 bool policy_unswitching( PhaseIdealLoop *phase ) const; 744 bool no_unswitch_candidate() const; 745 746 // Micro-benchmark spamming. Remove empty loops. 747 bool do_remove_empty_loop( PhaseIdealLoop *phase ); 748 749 // Convert one iteration loop into normal code. 750 bool do_one_iteration_loop( PhaseIdealLoop *phase ); 751 752 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 753 // move some loop-invariant test (usually a null-check) before the loop. 754 bool policy_peeling(PhaseIdealLoop *phase); 755 756 uint estimate_peeling(PhaseIdealLoop *phase); 757 758 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 759 // known trip count in the counted loop node. 760 bool policy_maximally_unroll(PhaseIdealLoop *phase) const; 761 762 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll 763 // if the loop is a counted loop and the loop body is small enough. 764 bool policy_unroll(PhaseIdealLoop *phase); 765 766 // Loop analyses to map to a maximal superword unrolling for vectorization. 767 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 768 769 // Return TRUE or FALSE if the loop should be range-check-eliminated. 770 // Gather a list of IF tests that are dominated by iteration splitting; 771 // also gather the end of the first split and the start of the 2nd split. 772 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const; 773 774 // Return TRUE if "iff" is a range check. 775 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const; 776 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset, 777 jlong& scale) const; 778 779 // Estimate the number of nodes required when cloning a loop (body). 780 uint est_loop_clone_sz(uint factor) const; 781 // Estimate the number of nodes required when unrolling a loop (body). 782 uint est_loop_unroll_sz(uint factor) const; 783 784 // Compute loop trip count if possible 785 void compute_trip_count(PhaseIdealLoop* phase); 786 787 // Compute loop trip count from profile data 788 float compute_profile_trip_cnt_helper(Node* n); 789 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 790 791 // Reassociate invariant expressions. 792 void reassociate_invariants(PhaseIdealLoop *phase); 793 // Reassociate invariant binary expressions. 794 Node* reassociate(Node* n1, PhaseIdealLoop *phase); 795 // Reassociate invariant add, subtract, and compare expressions. 796 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase); 797 // Return nonzero index of invariant operand if invariant and variant 798 // are combined with an associative binary. Helper for reassociate_invariants. 799 int find_invariant(Node* n, PhaseIdealLoop *phase); 800 // Return TRUE if "n" is associative. 801 bool is_associative(Node* n, Node* base=nullptr); 802 // Return TRUE if "n" is an associative cmp node. 803 bool is_associative_cmp(Node* n); 804 805 // Return true if n is invariant 806 bool is_invariant(Node* n) const; 807 808 // Put loop body on igvn work list 809 void record_for_igvn(); 810 811 bool is_root() { return _parent == nullptr; } 812 // A proper/reducible loop w/o any (occasional) dead back-edge. 813 bool is_loop() { return !_irreducible && !tail()->is_top(); } 814 bool is_counted() { return is_loop() && _head->is_CountedLoop(); } 815 bool is_innermost() { return is_loop() && _child == nullptr; } 816 817 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 818 819 bool compute_has_range_checks() const; 820 bool range_checks_present() { 821 if (!_has_range_checks_computed) { 822 if (compute_has_range_checks()) { 823 _has_range_checks = 1; 824 } 825 _has_range_checks_computed = 1; 826 } 827 return _has_range_checks; 828 } 829 830 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop. 831 // In all other cases, return this. 832 IdealLoopTree* skip_strip_mined() { 833 return _head->as_Loop()->is_strip_mined() ? _parent : this; 834 } 835 836 #ifndef PRODUCT 837 void dump_head(); // Dump loop head only 838 void dump(); // Dump this loop recursively 839 #endif 840 841 #ifdef ASSERT 842 GrowableArray<IdealLoopTree*> collect_sorted_children() const; 843 bool verify_tree(IdealLoopTree* loop_verify) const; 844 #endif 845 846 private: 847 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop. 848 849 // Estimate the number of nodes resulting from control and data flow merge. 850 uint est_loop_flow_merge_sz() const; 851 852 // Check if the number of residual iterations is large with unroll_cnt. 853 // Return true if the residual iterations are more than 10% of the trip count. 854 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const { 855 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt(); 856 } 857 858 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const; 859 860 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const; 861 862 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const; 863 864 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq, 865 Node* sfpt) const; 866 867 bool empty_loop_candidate(PhaseIdealLoop* phase) const; 868 869 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const; 870 }; 871 872 // -----------------------------PhaseIdealLoop--------------------------------- 873 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees 874 // into a loop tree. Drives the loop-based transformations on the ideal graph. 875 class PhaseIdealLoop : public PhaseTransform { 876 friend class IdealLoopTree; 877 friend class SuperWord; 878 friend class ShenandoahBarrierC2Support; 879 friend class AutoNodeBudget; 880 881 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes. 882 Node_List _loop_or_ctrl; 883 884 // Pre-computed def-use info 885 PhaseIterGVN &_igvn; 886 887 // Head of loop tree 888 IdealLoopTree* _ltree_root; 889 890 // Array of pre-order numbers, plus post-visited bit. 891 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 892 // ODD for post-visited. Other bits are the pre-order number. 893 uint *_preorders; 894 uint _max_preorder; 895 896 ReallocMark _nesting; // Safety checks for arena reallocation 897 898 const PhaseIdealLoop* _verify_me; 899 bool _verify_only; 900 901 // Allocate _preorders[] array 902 void allocate_preorders() { 903 _max_preorder = C->unique()+8; 904 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 905 memset(_preorders, 0, sizeof(uint) * _max_preorder); 906 } 907 908 // Allocate _preorders[] array 909 void reallocate_preorders() { 910 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe 911 if ( _max_preorder < C->unique() ) { 912 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 913 _max_preorder = C->unique(); 914 } 915 memset(_preorders, 0, sizeof(uint) * _max_preorder); 916 } 917 918 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 919 // adds new nodes. 920 void check_grow_preorders( ) { 921 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe 922 if ( _max_preorder < C->unique() ) { 923 uint newsize = _max_preorder<<1; // double size of array 924 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 925 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 926 _max_preorder = newsize; 927 } 928 } 929 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 930 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 931 // Pre-order numbers are written to the Nodes array as low-bit-set values. 932 void set_preorder_visited( Node *n, int pre_order ) { 933 assert( !is_visited( n ), "already set" ); 934 _preorders[n->_idx] = (pre_order<<1); 935 }; 936 // Return pre-order number. 937 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 938 939 // Check for being post-visited. 940 // Should be previsited already (checked with assert(is_visited(n))). 941 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 942 943 // Mark as post visited 944 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 945 946 public: 947 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 948 // Returns true if "n" is a data node, false if it's a control node. 949 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; } 950 951 private: 952 // clear out dead code after build_loop_late 953 Node_List _deadlist; 954 Node_List _zero_trip_guard_opaque_nodes; 955 Node_List _multiversion_opaque_nodes; 956 957 // Support for faster execution of get_late_ctrl()/dom_lca() 958 // when a node has many uses and dominator depth is deep. 959 GrowableArray<jlong> _dom_lca_tags; 960 uint _dom_lca_tags_round; 961 void init_dom_lca_tags(); 962 963 // Helper for debugging bad dominance relationships 964 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 965 966 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 967 968 // Inline wrapper for frequent cases: 969 // 1) only one use 970 // 2) a use is the same as the current LCA passed as 'n1' 971 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 972 assert( n->is_CFG(), "" ); 973 // Fast-path null lca 974 if( lca != nullptr && lca != n ) { 975 assert( lca->is_CFG(), "" ); 976 // find LCA of all uses 977 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 978 } 979 return find_non_split_ctrl(n); 980 } 981 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 982 983 // Helper function for directing control inputs away from CFG split points. 984 Node *find_non_split_ctrl( Node *ctrl ) const { 985 if (ctrl != nullptr) { 986 if (ctrl->is_MultiBranch()) { 987 ctrl = ctrl->in(0); 988 } 989 assert(ctrl->is_CFG(), "CFG"); 990 } 991 return ctrl; 992 } 993 994 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop); 995 996 #ifdef ASSERT 997 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop); 998 #endif 999 private: 1000 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj, 1001 Unique_Node_List& list); 1002 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll); 1003 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head, 1004 CountedLoopNode* remaining_loop_head, 1005 uint first_node_index_in_cloned_loop_body, 1006 const Node_List& old_new); 1007 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head, 1008 CountedLoopNode* main_loop_head, 1009 uint first_node_index_in_pre_loop_body, 1010 uint last_node_index_in_pre_loop_body, 1011 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA) 1012 const Node_List& old_new); 1013 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head, 1014 uint first_node_index_in_cloned_loop_body); 1015 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head, 1016 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template); 1017 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head, 1018 CountedLoopNode* target_loop_head, 1019 const NodeInLoopBody& _node_in_loop_body, 1020 bool kill_old_template); 1021 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head, 1022 const Node* old_target_loop_entry, 1023 uint node_index_before_new_assertion_predicate_nodes); 1024 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit, 1025 Node* bol); 1026 void log_loop_tree(); 1027 1028 public: 1029 1030 PhaseIterGVN &igvn() const { return _igvn; } 1031 1032 bool has_node(const Node* n) const { 1033 guarantee(n != nullptr, "No Node."); 1034 return _loop_or_ctrl[n->_idx] != nullptr; 1035 } 1036 // check if transform created new nodes that need _ctrl recorded 1037 Node *get_late_ctrl( Node *n, Node *early ); 1038 Node *get_early_ctrl( Node *n ); 1039 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 1040 void set_early_ctrl(Node* n, bool update_body); 1041 void set_subtree_ctrl(Node* n, bool update_body); 1042 void set_ctrl( Node *n, Node *ctrl ) { 1043 assert( !has_node(n) || has_ctrl(n), "" ); 1044 assert( ctrl->in(0), "cannot set dead control node" ); 1045 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 1046 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1)); 1047 } 1048 void set_root_as_ctrl(Node* n) { 1049 assert(!has_node(n) || has_ctrl(n), ""); 1050 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1)); 1051 } 1052 // Set control and update loop membership 1053 void set_ctrl_and_loop(Node* n, Node* ctrl) { 1054 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 1055 IdealLoopTree* new_loop = get_loop(ctrl); 1056 if (old_loop != new_loop) { 1057 if (old_loop->_child == nullptr) old_loop->_body.yank(n); 1058 if (new_loop->_child == nullptr) new_loop->_body.push(n); 1059 } 1060 set_ctrl(n, ctrl); 1061 } 1062 // Control nodes can be replaced or subsumed. During this pass they 1063 // get their replacement Node in slot 1. Instead of updating the block 1064 // location of all Nodes in the subsumed block, we lazily do it. As we 1065 // pull such a subsumed block out of the array, we write back the final 1066 // correct block. 1067 Node* get_ctrl(const Node* i) { 1068 assert(has_node(i), ""); 1069 Node *n = get_ctrl_no_update(i); 1070 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1)); 1071 assert(has_node(i) && has_ctrl(i), ""); 1072 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 1073 return n; 1074 } 1075 1076 bool is_dominator(Node* dominator, Node* n); 1077 bool is_strict_dominator(Node* dominator, Node* n); 1078 1079 // return get_ctrl for a data node and self(n) for a CFG node 1080 Node* ctrl_or_self(Node* n) { 1081 if (has_ctrl(n)) 1082 return get_ctrl(n); 1083 else { 1084 assert (n->is_CFG(), "must be a CFG node"); 1085 return n; 1086 } 1087 } 1088 1089 Node* get_ctrl_no_update_helper(const Node* i) const { 1090 assert(has_ctrl(i), "should be control, not loop"); 1091 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1); 1092 } 1093 1094 Node* get_ctrl_no_update(const Node* i) const { 1095 assert( has_ctrl(i), "" ); 1096 Node *n = get_ctrl_no_update_helper(i); 1097 if (!n->in(0)) { 1098 // Skip dead CFG nodes 1099 do { 1100 n = get_ctrl_no_update_helper(n); 1101 } while (!n->in(0)); 1102 n = find_non_split_ctrl(n); 1103 } 1104 return n; 1105 } 1106 1107 // Check for loop being set 1108 // "n" must be a control node. Returns true if "n" is known to be in a loop. 1109 bool has_loop( Node *n ) const { 1110 assert(!has_node(n) || !has_ctrl(n), ""); 1111 return has_node(n); 1112 } 1113 // Set loop 1114 void set_loop( Node *n, IdealLoopTree *loop ) { 1115 _loop_or_ctrl.map(n->_idx, (Node*)loop); 1116 } 1117 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 1118 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 1119 // from old_node to new_node to support the lazy update. Reference 1120 // replaces loop reference, since that is not needed for dead node. 1121 void lazy_update(Node *old_node, Node *new_node) { 1122 assert(old_node != new_node, "no cycles please"); 1123 // Re-use the side array slot for this node to provide the 1124 // forwarding pointer. 1125 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 1126 } 1127 void lazy_replace(Node *old_node, Node *new_node) { 1128 _igvn.replace_node(old_node, new_node); 1129 lazy_update(old_node, new_node); 1130 } 1131 1132 private: 1133 1134 // Place 'n' in some loop nest, where 'n' is a CFG node 1135 void build_loop_tree(); 1136 int build_loop_tree_impl(Node* n, int pre_order); 1137 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 1138 // loop tree, not the root. 1139 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 1140 1141 #ifdef ASSERT 1142 // verify that regions in irreducible loops are marked is_in_irreducible_loop 1143 void verify_regions_in_irreducible_loops(); 1144 bool is_in_irreducible_loop(RegionNode* region); 1145 #endif 1146 1147 // Place Data nodes in some loop nest 1148 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1149 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1150 void build_loop_late_post_work(Node* n, bool pinned); 1151 void build_loop_late_post(Node* n); 1152 void verify_strip_mined_scheduling(Node *n, Node* least); 1153 1154 // Array of immediate dominance info for each CFG node indexed by node idx 1155 private: 1156 uint _idom_size; 1157 Node **_idom; // Array of immediate dominators 1158 uint *_dom_depth; // Used for fast LCA test 1159 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 1160 LoopOptsMode _mode; 1161 1162 // build the loop tree and perform any requested optimizations 1163 void build_and_optimize(); 1164 1165 // Dominators for the sea of nodes 1166 void Dominators(); 1167 1168 // Compute the Ideal Node to Loop mapping 1169 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) : 1170 PhaseTransform(Ideal_Loop), 1171 _loop_or_ctrl(igvn.C->comp_arena()), 1172 _igvn(igvn), 1173 _verify_me(nullptr), 1174 _verify_only(false), 1175 _mode(mode), 1176 _nodes_required(UINT_MAX) { 1177 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop"); 1178 build_and_optimize(); 1179 } 1180 1181 #ifndef PRODUCT 1182 // Verify that verify_me made the same decisions as a fresh run 1183 // or only verify that the graph is valid if verify_me is null. 1184 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) : 1185 PhaseTransform(Ideal_Loop), 1186 _loop_or_ctrl(igvn.C->comp_arena()), 1187 _igvn(igvn), 1188 _verify_me(verify_me), 1189 _verify_only(verify_me == nullptr), 1190 _mode(LoopOptsVerify), 1191 _nodes_required(UINT_MAX) { 1192 DEBUG_ONLY(C->set_phase_verify_ideal_loop();) 1193 build_and_optimize(); 1194 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();) 1195 } 1196 #endif 1197 1198 Node* insert_convert_node_if_needed(BasicType target, Node* input); 1199 1200 public: 1201 Node* idom_no_update(Node* d) const { 1202 return idom_no_update(d->_idx); 1203 } 1204 1205 Node* idom_no_update(uint didx) const { 1206 assert(didx < _idom_size, "oob"); 1207 Node* n = _idom[didx]; 1208 assert(n != nullptr,"Bad immediate dominator info."); 1209 while (n->in(0) == nullptr) { // Skip dead CFG nodes 1210 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1); 1211 assert(n != nullptr,"Bad immediate dominator info."); 1212 } 1213 return n; 1214 } 1215 1216 Node *idom(Node* d) const { 1217 return idom(d->_idx); 1218 } 1219 1220 Node *idom(uint didx) const { 1221 Node *n = idom_no_update(didx); 1222 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 1223 return n; 1224 } 1225 1226 uint dom_depth(Node* d) const { 1227 guarantee(d != nullptr, "Null dominator info."); 1228 guarantee(d->_idx < _idom_size, ""); 1229 return _dom_depth[d->_idx]; 1230 } 1231 void set_idom(Node* d, Node* n, uint dom_depth); 1232 // Locally compute IDOM using dom_lca call 1233 Node *compute_idom( Node *region ) const; 1234 // Recompute dom_depth 1235 void recompute_dom_depth(); 1236 1237 // Is safept not required by an outer loop? 1238 bool is_deleteable_safept(Node* sfpt); 1239 1240 // Replace parallel induction variable (parallel to trip counter) 1241 void replace_parallel_iv(IdealLoopTree *loop); 1242 1243 Node *dom_lca( Node *n1, Node *n2 ) const { 1244 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 1245 } 1246 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 1247 1248 Node* dominated_node(Node* c1, Node* c2) { 1249 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related"); 1250 return is_dominator(c1, c2) ? c2 : c1; 1251 } 1252 1253 // Return control node that's dominated by the 2 others 1254 Node* dominated_node(Node* c1, Node* c2, Node* c3) { 1255 return dominated_node(c1, dominated_node(c2, c3)); 1256 } 1257 1258 // Build and verify the loop tree without modifying the graph. This 1259 // is useful to verify that all inputs properly dominate their uses. 1260 static void verify(PhaseIterGVN& igvn) { 1261 #ifdef ASSERT 1262 ResourceMark rm; 1263 Compile::TracePhase tp(_t_idealLoopVerify); 1264 PhaseIdealLoop v(igvn); 1265 #endif 1266 } 1267 1268 // Recommended way to use PhaseIdealLoop. 1269 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations. 1270 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) { 1271 ResourceMark rm; 1272 PhaseIdealLoop v(igvn, mode); 1273 1274 Compile* C = Compile::current(); 1275 if (!C->failing()) { 1276 // Cleanup any modified bits 1277 igvn.optimize(); 1278 if (C->failing()) { return; } 1279 v.log_loop_tree(); 1280 } 1281 } 1282 1283 // True if the method has at least 1 irreducible loop 1284 bool _has_irreducible_loops; 1285 1286 // Per-Node transform 1287 virtual Node* transform(Node* n) { return nullptr; } 1288 1289 Node* loop_exit_control(Node* x, IdealLoopTree* loop); 1290 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob); 1291 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr); 1292 Node* loop_iv_stride(Node* incr, Node*& xphi); 1293 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x); 1294 1295 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt); 1296 1297 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt); 1298 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new); 1299 #ifdef ASSERT 1300 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop); 1301 #endif 1302 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt); 1303 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop); 1304 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift); 1305 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control, 1306 IdealLoopTree* loop, float cl_prob, float le_fcnt, 1307 Node*& entry_control, Node*& iffalse); 1308 1309 Node* exact_limit( IdealLoopTree *loop ); 1310 1311 // Return a post-walked LoopNode 1312 IdealLoopTree *get_loop( Node *n ) const { 1313 // Dead nodes have no loop, so return the top level loop instead 1314 if (!has_node(n)) return _ltree_root; 1315 assert(!has_ctrl(n), ""); 1316 return (IdealLoopTree*)_loop_or_ctrl[n->_idx]; 1317 } 1318 1319 IdealLoopTree* ltree_root() const { return _ltree_root; } 1320 1321 // Is 'n' a (nested) member of 'loop'? 1322 int is_member( const IdealLoopTree *loop, Node *n ) const { 1323 return loop->is_member(get_loop(n)); } 1324 1325 // This is the basic building block of the loop optimizations. It clones an 1326 // entire loop body. It makes an old_new loop body mapping; with this 1327 // mapping you can find the new-loop equivalent to an old-loop node. All 1328 // new-loop nodes are exactly equal to their old-loop counterparts, all 1329 // edges are the same. All exits from the old-loop now have a RegionNode 1330 // that merges the equivalent new-loop path. This is true even for the 1331 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 1332 // now come from (one or more) Phis that merge their new-loop equivalents. 1333 // Parameter side_by_side_idom: 1334 // When side_by_size_idom is null, the dominator tree is constructed for 1335 // the clone loop to dominate the original. Used in construction of 1336 // pre-main-post loop sequence. 1337 // When nonnull, the clone and original are side-by-side, both are 1338 // dominated by the passed in side_by_side_idom node. Used in 1339 // construction of unswitched loops. 1340 enum CloneLoopMode { 1341 IgnoreStripMined = 0, // Only clone inner strip mined loop 1342 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops 1343 ControlAroundStripMined = 2 // Only clone inner strip mined loop, 1344 // result control flow branches 1345 // either to inner clone or outer 1346 // strip mined loop. 1347 }; 1348 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 1349 CloneLoopMode mode, Node* side_by_side_idom = nullptr); 1350 void clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1351 IdealLoopTree* loop, IdealLoopTree* companion_loop, 1352 Node_List*& split_if_set, Node_List*& split_bool_set, 1353 Node_List*& split_cex_set, Node_List& worklist, 1354 uint new_counter, CloneLoopMode mode); 1355 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1356 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1357 Node_List& extra_data_nodes); 1358 1359 // If we got the effect of peeling, either by actually peeling or by 1360 // making a pre-loop which must execute at least once, we can remove 1361 // all loop-invariant dominated tests in the main body. 1362 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 1363 1364 // Generate code to do a loop peel for the given loop (and body). 1365 // old_new is a temp array. 1366 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 1367 1368 // Add pre and post loops around the given loop. These loops are used 1369 // during RCE, unrolling and aligning loops. 1370 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 1371 1372 // Add post loop after the given loop. 1373 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new, 1374 CountedLoopNode* main_head, CountedLoopEndNode* main_end, 1375 Node* incr, Node* limit, CountedLoopNode*& post_head); 1376 1377 // Add a vector post loop between a vector main loop and the current post loop 1378 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 1379 // If Node n lives in the back_ctrl block, we clone a private version of n 1380 // in preheader_ctrl block and return that, otherwise return n. 1381 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 1382 1383 // Take steps to maximally unroll the loop. Peel any odd iterations, then 1384 // unroll to do double iterations. The next round of major loop transforms 1385 // will repeat till the doubled loop body does all remaining iterations in 1 1386 // pass. 1387 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 1388 1389 // Unroll the loop body one step - make each trip do 2 iterations. 1390 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 1391 1392 // Return true if exp is a constant times an induction var 1393 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0); 1394 1395 bool is_iv(Node* exp, Node* iv, BasicType bt); 1396 1397 // Return true if exp is a scaled induction var plus (or minus) constant 1398 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = nullptr, int depth = 0); 1399 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) { 1400 jlong long_scale; 1401 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) { 1402 int int_scale = checked_cast<int>(long_scale); 1403 if (p_scale != nullptr) { 1404 *p_scale = int_scale; 1405 } 1406 return true; 1407 } 1408 return false; 1409 } 1410 // Helper for finding more complex matches to is_scaled_iv_plus_offset. 1411 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv, 1412 BasicType bt, 1413 jlong* p_scale, Node** p_offset, 1414 bool* p_short_scale, int depth); 1415 1416 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 1417 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry, 1418 Deoptimization::DeoptReason reason, int opcode, 1419 bool rewire_uncommon_proj_phi_inputs = false); 1420 1421 private: 1422 // Helper functions for create_new_if_for_predicate() 1423 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj); 1424 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl); 1425 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj); 1426 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj, 1427 const OrigToNewHashtable& orig_to_clone); 1428 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt, 1429 Node* loop_entry); 1430 1431 public: 1432 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true); 1433 1434 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'. 1435 void replace_control(Node* node, Node* new_control) { 1436 _igvn.replace_input_of(node, 0, new_control); 1437 set_idom(node, new_control, dom_depth(new_control)); 1438 } 1439 1440 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) { 1441 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry); 1442 set_idom(loop_head, new_entry, dom_depth(new_entry)); 1443 } 1444 1445 // Construct a range check for a predicate if 1446 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit, 1447 jint stride, Node* range, bool upper, bool& overflow); 1448 1449 // Implementation of the loop predication to promote checks outside the loop 1450 bool loop_predication_impl(IdealLoopTree *loop); 1451 1452 private: 1453 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj, 1454 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero, 1455 Invariance& invar, Deoptimization::DeoptReason deopt_reason); 1456 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const; 1457 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt); 1458 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, 1459 PathFrequency& pf, Node_Stack& stack, VectorSet& seen, 1460 Node_List& if_proj_list); 1461 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate, 1462 IfProjNode* new_control, int scale, Node* offset, Node* range); 1463 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj); 1464 1465 // Helper function to collect predicate for eliminating the useless ones 1466 void eliminate_useless_predicates() const; 1467 1468 void eliminate_useless_zero_trip_guard(); 1469 void eliminate_useless_multiversion_if(); 1470 1471 public: 1472 // Change the control input of expensive nodes to allow commoning by 1473 // IGVN when it is guaranteed to not result in a more frequent 1474 // execution of the expensive node. Return true if progress. 1475 bool process_expensive_nodes(); 1476 1477 // Check whether node has become unreachable 1478 bool is_node_unreachable(Node *n) const { 1479 return !has_node(n) || n->is_unreachable(_igvn); 1480 } 1481 1482 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1483 void do_range_check(IdealLoopTree* loop); 1484 1485 // Clone loop with an invariant test (that does not exit) and 1486 // insert a clone of the test that selects which version to 1487 // execute. 1488 void do_unswitching(IdealLoopTree* loop, Node_List& old_new); 1489 1490 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const; 1491 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const; 1492 1493 private: 1494 static bool has_control_dependencies_from_predicates(LoopNode* head); 1495 static void revert_to_normal_loop(const LoopNode* loop_head); 1496 1497 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new, 1498 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector); 1499 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new); 1500 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head); 1501 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector); 1502 #ifndef PRODUCT 1503 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head); 1504 static void trace_loop_unswitching_impossible(const LoopNode* original_head); 1505 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector, 1506 const UnswitchCandidate& unswitch_candidate, 1507 const LoopNode* original_head, const LoopNode* new_head); 1508 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector, 1509 const LoopNode* original_head, const LoopNode* new_head); 1510 #endif 1511 1512 public: 1513 1514 // Range Check Elimination uses this function! 1515 // Constrain the main loop iterations so the affine function: 1516 // low_limit <= scale_con * I + offset < upper_limit 1517 // always holds true. That is, either increase the number of iterations in 1518 // the pre-loop or the post-loop until the condition holds true in the main 1519 // loop. Scale_con, offset and limit are all loop invariant. 1520 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit); 1521 // Helper function for add_constraint(). 1522 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round); 1523 1524 // Partially peel loop up through last_peel node. 1525 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1526 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new); 1527 1528 // AutoVectorize the loop: replace scalar ops with vector ops. 1529 enum AutoVectorizeStatus { 1530 Impossible, // This loop has the wrong shape to even try vectorization. 1531 Success, // We just successfully vectorized the loop. 1532 TriedAndFailed, // We tried to vectorize, but failed. 1533 }; 1534 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared); 1535 1536 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new); 1537 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new); 1538 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj); 1539 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt); 1540 1541 // Move an unordered Reduction out of loop if possible 1542 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop); 1543 1544 // Create a scheduled list of nodes control dependent on ctrl set. 1545 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1546 // Has a use in the vector set 1547 bool has_use_in_set( Node* n, VectorSet& vset ); 1548 // Has use internal to the vector set (ie. not in a phi at the loop head) 1549 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1550 // clone "n" for uses that are outside of loop 1551 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1552 // clone "n" for special uses that are in the not_peeled region 1553 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1554 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1555 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1556 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1557 #ifdef ASSERT 1558 // Validate the loop partition sets: peel and not_peel 1559 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1560 // Ensure that uses outside of loop are of the right form 1561 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1562 uint orig_exit_idx, uint clone_exit_idx); 1563 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1564 #endif 1565 1566 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1567 int stride_of_possible_iv( Node* iff ); 1568 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1569 // Return the (unique) control output node that's in the loop (if it exists.) 1570 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1571 // Insert a signed compare loop exit cloned from an unsigned compare. 1572 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1573 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1574 // Utility to register node "n" with PhaseIdealLoop 1575 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth); 1576 // Utility to create an if-projection 1577 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1578 // Force the iff control output to be the live_proj 1579 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1580 // Insert a region before an if projection 1581 RegionNode* insert_region_before_proj(ProjNode* proj); 1582 // Insert a new if before an if projection 1583 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1584 1585 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1586 // "Nearly" because all Nodes have been cloned from the original in the loop, 1587 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1588 // through the Phi recursively, and return a Bool. 1589 Node* clone_iff(PhiNode* phi); 1590 CmpNode* clone_bool(PhiNode* phi); 1591 1592 1593 // Rework addressing expressions to get the most loop-invariant stuff 1594 // moved out. We'd like to do all associative operators, but it's especially 1595 // important (common) to do address expressions. 1596 Node* remix_address_expressions(Node* n); 1597 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt); 1598 1599 // Convert add to muladd to generate MuladdS2I under certain criteria 1600 Node * convert_add_to_muladd(Node * n); 1601 1602 // Attempt to use a conditional move instead of a phi/branch 1603 Node *conditional_move( Node *n ); 1604 1605 // Check for aggressive application of 'split-if' optimization, 1606 // using basic block level info. 1607 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack); 1608 Node *split_if_with_blocks_pre ( Node *n ); 1609 void split_if_with_blocks_post( Node *n ); 1610 Node *has_local_phi_input( Node *n ); 1611 // Mark an IfNode as being dominated by a prior test, 1612 // without actually altering the CFG (and hence IDOM info). 1613 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool pin_array_access_nodes = false); 1614 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool pin_array_access_nodes); 1615 1616 // Split Node 'n' through merge point 1617 RegionNode* split_thru_region(Node* n, RegionNode* region); 1618 // Split Node 'n' through merge point if there is enough win. 1619 Node *split_thru_phi( Node *n, Node *region, int policy ); 1620 // Found an If getting its condition-code input from a Phi in the 1621 // same block. Split thru the Region. 1622 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr); 1623 1624 // Conversion of fill/copy patterns into intrinsic versions 1625 bool do_intrinsify_fill(); 1626 bool intrinsify_fill(IdealLoopTree* lpt); 1627 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1628 Node*& shift, Node*& offset); 1629 1630 private: 1631 // Return a type based on condition control flow 1632 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1633 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); } 1634 // Helpers for filtered type 1635 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1636 1637 // Helper functions 1638 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1639 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1640 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 ); 1641 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1642 1643 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const; 1644 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1645 void try_move_store_after_loop(Node* n); 1646 void move_flat_array_check_out_of_loop(Node* n); 1647 bool identical_backtoback_ifs(Node *n); 1648 bool flat_array_element_type_check(Node *n); 1649 bool can_split_if(Node *n_ctrl); 1650 bool cannot_split_division(const Node* n, const Node* region) const; 1651 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop); 1652 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const; 1653 1654 // Determine if a method is too big for a/another round of split-if, based on 1655 // a magic (approximate) ratio derived from the equally magic constant 35000, 1656 // previously used for this purpose (but without relating to the node limit). 1657 bool must_throttle_split_if() { 1658 uint threshold = C->max_node_limit() * 2 / 5; 1659 return C->live_nodes() > threshold; 1660 } 1661 1662 // A simplistic node request tracking mechanism, where 1663 // = UINT_MAX Request not valid or made final. 1664 // < UINT_MAX Nodes currently requested (estimate). 1665 uint _nodes_required; 1666 1667 enum { REQUIRE_MIN = 70 }; 1668 1669 uint nodes_required() const { return _nodes_required; } 1670 1671 // Given the _currently_ available number of nodes, check whether there is 1672 // "room" for an additional request or not, considering the already required 1673 // number of nodes. Return TRUE if the new request is exceeding the node 1674 // budget limit, otherwise return FALSE. Note that this interpretation will 1675 // act pessimistic on additional requests when new nodes have already been 1676 // generated since the 'begin'. This behaviour fits with the intention that 1677 // node estimates/requests should be made upfront. 1678 bool exceeding_node_budget(uint required = 0) { 1679 assert(C->live_nodes() < C->max_node_limit(), "sanity"); 1680 uint available = C->max_node_limit() - C->live_nodes(); 1681 return available < required + _nodes_required + REQUIRE_MIN; 1682 } 1683 1684 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1685 precond(require > 0); 1686 _nodes_required += MAX2(require, minreq); 1687 return _nodes_required; 1688 } 1689 1690 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1691 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0; 1692 } 1693 1694 uint require_nodes_begin() { 1695 assert(_nodes_required == UINT_MAX, "Bad state (begin)."); 1696 _nodes_required = 0; 1697 return C->live_nodes(); 1698 } 1699 1700 // When a node request is final, optionally check that the requested number 1701 // of nodes was reasonably correct with respect to the number of new nodes 1702 // introduced since the last 'begin'. Always check that we have not exceeded 1703 // the maximum node limit. 1704 void require_nodes_final(uint live_at_begin, bool check_estimate) { 1705 assert(_nodes_required < UINT_MAX, "Bad state (final)."); 1706 1707 #ifdef ASSERT 1708 if (check_estimate) { 1709 // Check that the node budget request was not off by too much (x2). 1710 // Should this be the case we _surely_ need to improve the estimates 1711 // used in our budget calculations. 1712 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) { 1713 log_info(compilation)("Bad node estimate: actual = %d >> request = %d", 1714 C->live_nodes() - live_at_begin, _nodes_required); 1715 } 1716 } 1717 #endif 1718 // Assert that we have stayed within the node budget limit. 1719 assert(C->live_nodes() < C->max_node_limit(), 1720 "Exceeding node budget limit: %d + %d > %d (request = %d)", 1721 C->live_nodes() - live_at_begin, live_at_begin, 1722 C->max_node_limit(), _nodes_required); 1723 1724 _nodes_required = UINT_MAX; 1725 } 1726 1727 private: 1728 1729 bool _created_loop_node; 1730 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);) 1731 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;) 1732 1733 public: 1734 void set_created_loop_node() { _created_loop_node = true; } 1735 bool created_loop_node() { return _created_loop_node; } 1736 void register_new_node(Node* n, Node* blk); 1737 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) { 1738 register_new_node(new_node, get_ctrl(ctrl_of)); 1739 } 1740 1741 Node* clone_and_register(Node* n, Node* ctrl) { 1742 n = n->clone(); 1743 register_new_node(n, ctrl); 1744 return n; 1745 } 1746 1747 #ifdef ASSERT 1748 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1749 #endif 1750 1751 #ifndef PRODUCT 1752 void dump() const; 1753 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging 1754 void dump_idom(Node* n, uint count) const; 1755 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const; 1756 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const; 1757 IdealLoopTree* get_loop_idx(Node* n) const { 1758 // Dead nodes have no loop, so return the top level loop instead 1759 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root; 1760 } 1761 // Print some stats 1762 static void print_statistics(); 1763 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1764 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1765 static volatile int _long_loop_candidates; 1766 static volatile int _long_loop_nests; 1767 static volatile int _long_loop_counted_loops; 1768 #endif 1769 1770 #ifdef ASSERT 1771 void verify() const; 1772 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const; 1773 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const; 1774 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const; 1775 #endif 1776 1777 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const; 1778 1779 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN; 1780 1781 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test); 1782 1783 1784 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi, 1785 Node_List &range_checks); 1786 1787 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi, 1788 Node* inner_iters_actual_int, Node* inner_phi, 1789 Node* iv_add, LoopNode* inner_head); 1790 1791 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA); 1792 1793 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl); 1794 1795 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop); 1796 1797 Node* compute_early_ctrl(Node* n, Node* n_ctrl); 1798 1799 void try_sink_out_of_loop(Node* n); 1800 1801 Node* clamp(Node* R, Node* L, Node* H); 1802 1803 bool safe_for_if_replacement(const Node* dom) const; 1804 1805 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region); 1806 1807 bool try_merge_identical_ifs(Node* n); 1808 1809 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm); 1810 1811 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 1812 IdealLoopTree* parent, bool partial); 1813 1814 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 1815 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist); 1816 1817 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 1818 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set, 1819 Node_List*& split_bool_set, Node_List*& split_cex_set); 1820 1821 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set); 1822 1823 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2); 1824 1825 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2); 1826 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i); 1827 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2); 1828 void clone_template_assertion_expression_down(Node* node); 1829 1830 Node* similar_subtype_check(const Node* x, Node* r_in); 1831 1832 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base); 1833 1834 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x); 1835 1836 void pin_array_access_nodes_dependent_on(Node* ctrl); 1837 1838 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const; 1839 1840 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node); 1841 1842 ConINode* intcon(jint i); 1843 1844 ConLNode* longcon(jlong i); 1845 1846 ConNode* makecon(const Type* t); 1847 1848 ConNode* integercon(jlong l, BasicType bt); 1849 1850 ConNode* zerocon(BasicType bt); 1851 }; 1852 1853 1854 class AutoNodeBudget : public StackObj 1855 { 1856 public: 1857 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK }; 1858 1859 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK) 1860 : _phase(phase), 1861 _check_at_final(chk == BUDGET_CHECK), 1862 _nodes_at_begin(0) 1863 { 1864 precond(_phase != nullptr); 1865 1866 _nodes_at_begin = _phase->require_nodes_begin(); 1867 } 1868 1869 ~AutoNodeBudget() { 1870 #ifndef PRODUCT 1871 if (TraceLoopOpts) { 1872 uint request = _phase->nodes_required(); 1873 uint delta = _phase->C->live_nodes() - _nodes_at_begin; 1874 1875 if (request < delta) { 1876 tty->print_cr("Exceeding node budget: %d < %d", request, delta); 1877 } else { 1878 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN; 1879 // Identify the worst estimates as "poor" ones. 1880 if (request > REQUIRE_MIN && delta > 0) { 1881 if ((delta > REQUIRE_MIN && request > 3 * delta) || 1882 (delta <= REQUIRE_MIN && request > 10 * delta)) { 1883 tty->print_cr("Poor node estimate: %d >> %d", request, delta); 1884 } 1885 } 1886 } 1887 } 1888 #endif // PRODUCT 1889 _phase->require_nodes_final(_nodes_at_begin, _check_at_final); 1890 } 1891 1892 private: 1893 PhaseIdealLoop* _phase; 1894 bool _check_at_final; 1895 uint _nodes_at_begin; 1896 }; 1897 1898 inline Node* IdealLoopTree::tail() { 1899 // Handle lazy update of _tail field. 1900 if (_tail->in(0) == nullptr) { 1901 _tail = _phase->get_ctrl(_tail); 1902 } 1903 return _tail; 1904 } 1905 1906 inline Node* IdealLoopTree::head() { 1907 // Handle lazy update of _head field. 1908 if (_head->in(0) == nullptr) { 1909 _head = _phase->get_ctrl(_head); 1910 } 1911 return _head; 1912 } 1913 1914 // Iterate over the loop tree using a preorder, left-to-right traversal. 1915 // 1916 // Example that visits all counted loops from within PhaseIdealLoop 1917 // 1918 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1919 // IdealLoopTree* lpt = iter.current(); 1920 // if (!lpt->is_counted()) continue; 1921 // ... 1922 class LoopTreeIterator : public StackObj { 1923 private: 1924 IdealLoopTree* _root; 1925 IdealLoopTree* _curnt; 1926 1927 public: 1928 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1929 1930 bool done() { return _curnt == nullptr; } // Finished iterating? 1931 1932 void next(); // Advance to next loop tree 1933 1934 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1935 }; 1936 1937 // Compute probability of reaching some CFG node from a fixed 1938 // dominating CFG node 1939 class PathFrequency { 1940 private: 1941 Node* _dom; // frequencies are computed relative to this node 1942 Node_Stack _stack; 1943 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions 1944 GrowableArray<float> _freqs; // cache frequencies 1945 PhaseIdealLoop* _phase; 1946 1947 float check_and_truncate_frequency(float f) { 1948 assert(f >= 0, "Incorrect frequency"); 1949 // We do not perform an exact (f <= 1) check 1950 // this would be error prone with rounding of floats. 1951 // Performing a check like (f <= 1+eps) would be of benefit, 1952 // however, it is not evident how to determine such an eps, 1953 // given that an arbitrary number of add/mul operations 1954 // are performed on these frequencies. 1955 return (f > 1) ? 1 : f; 1956 } 1957 1958 public: 1959 PathFrequency(Node* dom, PhaseIdealLoop* phase) 1960 : _dom(dom), _stack(0), _phase(phase) { 1961 } 1962 1963 float to(Node* n); 1964 }; 1965 1966 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps: 1967 // 1. Clone the data nodes 1968 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping. 1969 class DataNodeGraph : public StackObj { 1970 PhaseIdealLoop* const _phase; 1971 const Unique_Node_List& _data_nodes; 1972 OrigToNewHashtable _orig_to_new; 1973 1974 public: 1975 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase) 1976 : _phase(phase), 1977 _data_nodes(data_nodes), 1978 // Use 107 as best guess which is the first resize value in ResizeableResourceHashtable::large_table_sizes. 1979 _orig_to_new(107, MaxNodeLimit) 1980 { 1981 #ifdef ASSERT 1982 for (uint i = 0; i < data_nodes.size(); i++) { 1983 assert(!data_nodes[i]->is_CFG(), "only data nodes"); 1984 } 1985 #endif 1986 } 1987 NONCOPYABLE(DataNodeGraph); 1988 1989 private: 1990 void clone(Node* node, Node* new_ctrl); 1991 void clone_data_nodes(Node* new_ctrl); 1992 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy, 1993 Node* new_ctrl); 1994 void rewire_clones_to_cloned_inputs(); 1995 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node); 1996 1997 public: 1998 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy. 1999 // Set 'new_ctrl' as ctrl for the cloned nodes. 2000 const OrigToNewHashtable& clone(Node* new_ctrl) { 2001 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row"); 2002 clone_data_nodes(new_ctrl); 2003 rewire_clones_to_cloned_inputs(); 2004 return _orig_to_new; 2005 } 2006 2007 // Create a copy of the data nodes provided to the constructor by doing the following: 2008 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes, 2009 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete 2010 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set 2011 // `new_ctrl` as ctrl. 2012 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy( 2013 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 2014 Node* new_ctrl) { 2015 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl); 2016 rewire_clones_to_cloned_inputs(); 2017 return _orig_to_new; 2018 } 2019 }; 2020 #endif // SHARE_OPTO_LOOPNODE_HPP