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