1 /*
   2  * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_OPTO_LOOPNODE_HPP
  26 #define SHARE_OPTO_LOOPNODE_HPP
  27 
  28 #include "opto/cfgnode.hpp"
  29 #include "opto/multnode.hpp"
  30 #include "opto/phaseX.hpp"
  31 #include "opto/subnode.hpp"
  32 #include "opto/type.hpp"
  33 
  34 class CmpNode;
  35 class CountedLoopEndNode;
  36 class CountedLoopNode;
  37 class IdealLoopTree;
  38 class LoopNode;
  39 class Node;
  40 class OuterStripMinedLoopEndNode;
  41 class PathFrequency;
  42 class PhaseIdealLoop;
  43 class CountedLoopReserveKit;
  44 class VectorSet;
  45 class Invariance;
  46 struct small_cache;
  47 
  48 //
  49 //                  I D E A L I Z E D   L O O P S
  50 //
  51 // Idealized loops are the set of loops I perform more interesting
  52 // transformations on, beyond simple hoisting.
  53 
  54 //------------------------------LoopNode---------------------------------------
  55 // Simple loop header.  Fall in path on left, loop-back path on right.
  56 class LoopNode : public RegionNode {
  57   // Size is bigger to hold the flags.  However, the flags do not change
  58   // the semantics so it does not appear in the hash & cmp functions.
  59   virtual uint size_of() const { return sizeof(*this); }
  60 protected:
  61   uint _loop_flags;
  62   // Names for flag bitfields
  63   enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
  64          MainHasNoPreLoop=4,
  65          HasExactTripCount=8,
  66          InnerLoop=16,
  67          PartialPeelLoop=32,
  68          PartialPeelFailed=64,
  69          HasReductions=128,
  70          WasSlpAnalyzed=256,
  71          PassedSlpAnalysis=512,
  72          DoUnrollOnly=1024,
  73          VectorizedLoop=2048,
  74          HasAtomicPostLoop=4096,
  75          HasRangeChecks=8192,
  76          IsMultiversioned=16384,
  77          StripMined=32768,
  78          SubwordLoop=65536,
  79          ProfileTripFailed=131072};
  80   char _unswitch_count;
  81   enum { _unswitch_max=3 };
  82   char _postloop_flags;
  83   enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
  84 
  85   // Expected trip count from profile data
  86   float _profile_trip_cnt;
  87 
  88 public:
  89   // Names for edge indices
  90   enum { Self=0, EntryControl, LoopBackControl };
  91 
  92   bool is_inner_loop() const { return _loop_flags & InnerLoop; }
  93   void set_inner_loop() { _loop_flags |= InnerLoop; }
  94 
  95   bool range_checks_present() const { return _loop_flags & HasRangeChecks; }
  96   bool is_multiversioned() const { return _loop_flags & IsMultiversioned; }
  97   bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
  98   bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
  99   void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
 100   bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
 101   bool is_strip_mined() const { return _loop_flags & StripMined; }
 102   bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
 103   bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
 104 
 105   void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
 106   void mark_has_reductions() { _loop_flags |= HasReductions; }
 107   void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
 108   void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
 109   void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
 110   void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
 111   void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
 112   void mark_has_range_checks() { _loop_flags |=  HasRangeChecks; }
 113   void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
 114   void mark_strip_mined() { _loop_flags |= StripMined; }
 115   void clear_strip_mined() { _loop_flags &= ~StripMined; }
 116   void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
 117   void mark_subword_loop() { _loop_flags |= SubwordLoop; }
 118 
 119   int unswitch_max() { return _unswitch_max; }
 120   int unswitch_count() { return _unswitch_count; }
 121 
 122   int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
 123   void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
 124   int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
 125   void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
 126 
 127   void set_unswitch_count(int val) {
 128     assert (val <= unswitch_max(), "too many unswitches");
 129     _unswitch_count = val;
 130   }
 131 
 132   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
 133   float profile_trip_cnt()             { return _profile_trip_cnt; }
 134 
 135   LoopNode(Node *entry, Node *backedge)
 136     : RegionNode(3), _loop_flags(0), _unswitch_count(0),
 137       _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN)  {
 138     init_class_id(Class_Loop);
 139     init_req(EntryControl, entry);
 140     init_req(LoopBackControl, backedge);
 141   }
 142 
 143   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 144   virtual int Opcode() const;
 145   bool can_be_counted_loop(PhaseTransform* phase) const {
 146     return req() == 3 && in(0) != NULL &&
 147       in(1) != NULL && phase->type(in(1)) != Type::TOP &&
 148       in(2) != NULL && phase->type(in(2)) != Type::TOP;
 149   }
 150   bool is_valid_counted_loop() const;
 151 #ifndef PRODUCT
 152   virtual void dump_spec(outputStream *st) const;
 153 #endif
 154 
 155   void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
 156   virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
 157   virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
 158   virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
 159   virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
 160   virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; }
 161 };
 162 
 163 //------------------------------Counted Loops----------------------------------
 164 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
 165 // path (and maybe some other exit paths).  The trip-counter exit is always
 166 // last in the loop.  The trip-counter have to stride by a constant;
 167 // the exit value is also loop invariant.
 168 
 169 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
 170 // CountedLoopNode has the incoming loop control and the loop-back-control
 171 // which is always the IfTrue before the matching CountedLoopEndNode.  The
 172 // CountedLoopEndNode has an incoming control (possibly not the
 173 // CountedLoopNode if there is control flow in the loop), the post-increment
 174 // trip-counter value, and the limit.  The trip-counter value is always of
 175 // the form (Op old-trip-counter stride).  The old-trip-counter is produced
 176 // by a Phi connected to the CountedLoopNode.  The stride is constant.
 177 // The Op is any commutable opcode, including Add, Mul, Xor.  The
 178 // CountedLoopEndNode also takes in the loop-invariant limit value.
 179 
 180 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
 181 // loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
 182 // via the old-trip-counter from the Op node.
 183 
 184 //------------------------------CountedLoopNode--------------------------------
 185 // CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
 186 // inputs the incoming loop-start control and the loop-back control, so they
 187 // act like RegionNodes.  They also take in the initial trip counter, the
 188 // loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
 189 // produce a loop-body control and the trip counter value.  Since
 190 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
 191 
 192 class CountedLoopNode : public LoopNode {
 193   // Size is bigger to hold _main_idx.  However, _main_idx does not change
 194   // the semantics so it does not appear in the hash & cmp functions.
 195   virtual uint size_of() const { return sizeof(*this); }
 196 
 197   // For Pre- and Post-loops during debugging ONLY, this holds the index of
 198   // the Main CountedLoop.  Used to assert that we understand the graph shape.
 199   node_idx_t _main_idx;
 200 
 201   // Known trip count calculated by compute_exact_trip_count()
 202   uint  _trip_count;
 203 
 204   // Log2 of original loop bodies in unrolled loop
 205   int _unrolled_count_log2;
 206 
 207   // Node count prior to last unrolling - used to decide if
 208   // unroll,optimize,unroll,optimize,... is making progress
 209   int _node_count_before_unroll;
 210 
 211   // If slp analysis is performed we record the maximum
 212   // vector mapped unroll factor here
 213   int _slp_maximum_unroll_factor;
 214 
 215 public:
 216   CountedLoopNode( Node *entry, Node *backedge )
 217     : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
 218       _unrolled_count_log2(0), _node_count_before_unroll(0),
 219       _slp_maximum_unroll_factor(0) {
 220     init_class_id(Class_CountedLoop);
 221     // Initialize _trip_count to the largest possible value.
 222     // Will be reset (lower) if the loop's trip count is known.
 223   }
 224 
 225   virtual int Opcode() const;
 226   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 227 
 228   Node *init_control() const { return in(EntryControl); }
 229   Node *back_control() const { return in(LoopBackControl); }
 230   CountedLoopEndNode *loopexit_or_null() const;
 231   CountedLoopEndNode *loopexit() const;
 232   Node *init_trip() const;
 233   Node *stride() const;
 234   int   stride_con() const;
 235   bool  stride_is_con() const;
 236   Node *limit() const;
 237   Node *incr() const;
 238   Node *phi() const;
 239 
 240   // Match increment with optional truncation
 241   static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
 242 
 243   // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
 244   // can run short a few iterations and may start a few iterations in.
 245   // It will be RCE'd and unrolled and aligned.
 246 
 247   // A following 'post' loop will run any remaining iterations.  Used
 248   // during Range Check Elimination, the 'post' loop will do any final
 249   // iterations with full checks.  Also used by Loop Unrolling, where
 250   // the 'post' loop will do any epilog iterations needed.  Basically,
 251   // a 'post' loop can not profitably be further unrolled or RCE'd.
 252 
 253   // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
 254   // it may do under-flow checks for RCE and may do alignment iterations
 255   // so the following main loop 'knows' that it is striding down cache
 256   // lines.
 257 
 258   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
 259   // Aligned, may be missing it's pre-loop.
 260   bool is_normal_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
 261   bool is_pre_loop      () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
 262   bool is_main_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
 263   bool is_post_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
 264   bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
 265   bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
 266   bool has_passed_slp   () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
 267   bool is_unroll_only   () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
 268   bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
 269   bool has_atomic_post_loop  () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
 270   void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 271 
 272   int main_idx() const { return _main_idx; }
 273 
 274 
 275   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
 276   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
 277   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
 278   void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 279 
 280   void set_trip_count(uint tc) { _trip_count = tc; }
 281   uint trip_count()            { return _trip_count; }
 282 
 283   bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
 284   void set_exact_trip_count(uint tc) {
 285     _trip_count = tc;
 286     _loop_flags |= HasExactTripCount;
 287   }
 288   void set_nonexact_trip_count() {
 289     _loop_flags &= ~HasExactTripCount;
 290   }
 291   void set_notpassed_slp() {
 292     _loop_flags &= ~PassedSlpAnalysis;
 293   }
 294 
 295   void double_unrolled_count() { _unrolled_count_log2++; }
 296   int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
 297 
 298   void set_node_count_before_unroll(int ct)  { _node_count_before_unroll = ct; }
 299   int  node_count_before_unroll()            { return _node_count_before_unroll; }
 300   void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
 301   int  slp_max_unroll() const                { return _slp_maximum_unroll_factor; }
 302 
 303   virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
 304   OuterStripMinedLoopNode* outer_loop() const;
 305   virtual IfTrueNode* outer_loop_tail() const;
 306   virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
 307   virtual IfFalseNode* outer_loop_exit() const;
 308   virtual SafePointNode* outer_safepoint() const;
 309 
 310   // If this is a main loop in a pre/main/post loop nest, walk over
 311   // the predicates that were inserted by
 312   // duplicate_predicates()/add_range_check_predicate()
 313   static Node* skip_predicates_from_entry(Node* ctrl);
 314   Node* skip_predicates();
 315 
 316 #ifndef PRODUCT
 317   virtual void dump_spec(outputStream *st) const;
 318 #endif
 319 };
 320 
 321 //------------------------------CountedLoopEndNode-----------------------------
 322 // CountedLoopEndNodes end simple trip counted loops.  They act much like
 323 // IfNodes.
 324 class CountedLoopEndNode : public IfNode {
 325 public:
 326   enum { TestControl, TestValue };
 327 
 328   CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
 329     : IfNode( control, test, prob, cnt) {
 330     init_class_id(Class_CountedLoopEnd);
 331   }
 332   virtual int Opcode() const;
 333 
 334   Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
 335   Node *incr() const                { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 336   Node *limit() const               { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 337   Node *stride() const              { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 338   Node *init_trip() const           { Node *tmp = phi     (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 339   int stride_con() const;
 340   bool stride_is_con() const        { Node *tmp = stride  (); return (tmp != NULL && tmp->is_Con()); }
 341   BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
 342   PhiNode *phi() const {
 343     Node *tmp = incr();
 344     if (tmp && tmp->req() == 3) {
 345       Node* phi = tmp->in(1);
 346       if (phi->is_Phi()) {
 347         return phi->as_Phi();
 348       }
 349     }
 350     return NULL;
 351   }
 352   CountedLoopNode *loopnode() const {
 353     // The CountedLoopNode that goes with this CountedLoopEndNode may
 354     // have been optimized out by the IGVN so be cautious with the
 355     // pattern matching on the graph
 356     PhiNode* iv_phi = phi();
 357     if (iv_phi == NULL) {
 358       return NULL;
 359     }
 360     Node *ln = iv_phi->in(0);
 361     if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) {
 362       return (CountedLoopNode*)ln;
 363     }
 364     return NULL;
 365   }
 366 
 367 #ifndef PRODUCT
 368   virtual void dump_spec(outputStream *st) const;
 369 #endif
 370 };
 371 
 372 
 373 inline CountedLoopEndNode* CountedLoopNode::loopexit_or_null() const {
 374   Node* bctrl = back_control();
 375   if (bctrl == NULL) return NULL;
 376 
 377   Node* lexit = bctrl->in(0);
 378   return (CountedLoopEndNode*)
 379       (lexit->Opcode() == Op_CountedLoopEnd ? lexit : NULL);
 380 }
 381 
 382 inline CountedLoopEndNode* CountedLoopNode::loopexit() const {
 383   CountedLoopEndNode* cle = loopexit_or_null();
 384   assert(cle != NULL, "loopexit is NULL");
 385   return cle;
 386 }
 387 
 388 inline Node* CountedLoopNode::init_trip() const {
 389   CountedLoopEndNode* cle = loopexit_or_null();
 390   return cle != NULL ? cle->init_trip() : NULL;
 391 }
 392 inline Node* CountedLoopNode::stride() const {
 393   CountedLoopEndNode* cle = loopexit_or_null();
 394   return cle != NULL ? cle->stride() : NULL;
 395 }
 396 inline int CountedLoopNode::stride_con() const {
 397   CountedLoopEndNode* cle = loopexit_or_null();
 398   return cle != NULL ? cle->stride_con() : 0;
 399 }
 400 inline bool CountedLoopNode::stride_is_con() const {
 401   CountedLoopEndNode* cle = loopexit_or_null();
 402   return cle != NULL && cle->stride_is_con();
 403 }
 404 inline Node* CountedLoopNode::limit() const {
 405   CountedLoopEndNode* cle = loopexit_or_null();
 406   return cle != NULL ? cle->limit() : NULL;
 407 }
 408 inline Node* CountedLoopNode::incr() const {
 409   CountedLoopEndNode* cle = loopexit_or_null();
 410   return cle != NULL ? cle->incr() : NULL;
 411 }
 412 inline Node* CountedLoopNode::phi() const {
 413   CountedLoopEndNode* cle = loopexit_or_null();
 414   return cle != NULL ? cle->phi() : NULL;
 415 }
 416 
 417 //------------------------------LoopLimitNode-----------------------------
 418 // Counted Loop limit node which represents exact final iterator value:
 419 // trip_count = (limit - init_trip + stride - 1)/stride
 420 // final_value= trip_count * stride + init_trip.
 421 // Use HW instructions to calculate it when it can overflow in integer.
 422 // Note, final_value should fit into integer since counted loop has
 423 // limit check: limit <= max_int-stride.
 424 class LoopLimitNode : public Node {
 425   enum { Init=1, Limit=2, Stride=3 };
 426  public:
 427   LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
 428     // Put it on the Macro nodes list to optimize during macro nodes expansion.
 429     init_flags(Flag_is_macro);
 430     C->add_macro_node(this);
 431   }
 432   virtual int Opcode() const;
 433   virtual const Type *bottom_type() const { return TypeInt::INT; }
 434   virtual uint ideal_reg() const { return Op_RegI; }
 435   virtual const Type* Value(PhaseGVN* phase) const;
 436   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 437   virtual Node* Identity(PhaseGVN* phase);
 438 };
 439 
 440 // Support for strip mining
 441 class OuterStripMinedLoopNode : public LoopNode {
 442 private:
 443   CountedLoopNode* inner_loop() const;
 444 public:
 445   OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
 446     : LoopNode(entry, backedge) {
 447     init_class_id(Class_OuterStripMinedLoop);
 448     init_flags(Flag_is_macro);
 449     C->add_macro_node(this);
 450   }
 451 
 452   virtual int Opcode() const;
 453 
 454   virtual IfTrueNode* outer_loop_tail() const;
 455   virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
 456   virtual IfFalseNode* outer_loop_exit() const;
 457   virtual SafePointNode* outer_safepoint() const;
 458   void adjust_strip_mined_loop(PhaseIterGVN* igvn);
 459 };
 460 
 461 class OuterStripMinedLoopEndNode : public IfNode {
 462 public:
 463   OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
 464     : IfNode(control, test, prob, cnt) {
 465     init_class_id(Class_OuterStripMinedLoopEnd);
 466   }
 467 
 468   virtual int Opcode() const;
 469 
 470   virtual const Type* Value(PhaseGVN* phase) const;
 471   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 472 };
 473 
 474 // -----------------------------IdealLoopTree----------------------------------
 475 class IdealLoopTree : public ResourceObj {
 476 public:
 477   IdealLoopTree *_parent;       // Parent in loop tree
 478   IdealLoopTree *_next;         // Next sibling in loop tree
 479   IdealLoopTree *_child;        // First child in loop tree
 480 
 481   // The head-tail backedge defines the loop.
 482   // If a loop has multiple backedges, this is addressed during cleanup where
 483   // we peel off the multiple backedges,  merging all edges at the bottom and
 484   // ensuring that one proper backedge flow into the loop.
 485   Node *_head;                  // Head of loop
 486   Node *_tail;                  // Tail of loop
 487   inline Node *tail();          // Handle lazy update of _tail field
 488   inline Node *head();          // Handle lazy update of _head field
 489   PhaseIdealLoop* _phase;
 490   int _local_loop_unroll_limit;
 491   int _local_loop_unroll_factor;
 492 
 493   Node_List _body;              // Loop body for inner loops
 494 
 495   uint8_t _nest;                // Nesting depth
 496   uint8_t _irreducible:1,       // True if irreducible
 497           _has_call:1,          // True if has call safepoint
 498           _has_sfpt:1,          // True if has non-call safepoint
 499           _rce_candidate:1;     // True if candidate for range check elimination
 500 
 501   Node_List* _safepts;          // List of safepoints in this loop
 502   Node_List* _required_safept;  // A inner loop cannot delete these safepts;
 503   bool  _allow_optimizations;   // Allow loop optimizations
 504 
 505   IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
 506     : _parent(0), _next(0), _child(0),
 507       _head(head), _tail(tail),
 508       _phase(phase),
 509       _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
 510       _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
 511       _safepts(NULL),
 512       _required_safept(NULL),
 513       _allow_optimizations(true)
 514   {
 515     precond(_head != NULL);
 516     precond(_tail != NULL);
 517   }
 518 
 519   // Is 'l' a member of 'this'?
 520   bool is_member(const IdealLoopTree *l) const; // Test for nested membership
 521 
 522   // Set loop nesting depth.  Accumulate has_call bits.
 523   int set_nest( uint depth );
 524 
 525   // Split out multiple fall-in edges from the loop header.  Move them to a
 526   // private RegionNode before the loop.  This becomes the loop landing pad.
 527   void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
 528 
 529   // Split out the outermost loop from this shared header.
 530   void split_outer_loop( PhaseIdealLoop *phase );
 531 
 532   // Merge all the backedges from the shared header into a private Region.
 533   // Feed that region as the one backedge to this loop.
 534   void merge_many_backedges( PhaseIdealLoop *phase );
 535 
 536   // Split shared headers and insert loop landing pads.
 537   // Insert a LoopNode to replace the RegionNode.
 538   // Returns TRUE if loop tree is structurally changed.
 539   bool beautify_loops( PhaseIdealLoop *phase );
 540 
 541   // Perform optimization to use the loop predicates for null checks and range checks.
 542   // Applies to any loop level (not just the innermost one)
 543   bool loop_predication( PhaseIdealLoop *phase);
 544 
 545   // Perform iteration-splitting on inner loops.  Split iterations to
 546   // avoid range checks or one-shot null checks.  Returns false if the
 547   // current round of loop opts should stop.
 548   bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
 549 
 550   // Driver for various flavors of iteration splitting.  Returns false
 551   // if the current round of loop opts should stop.
 552   bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
 553 
 554   // Given dominators, try to find loops with calls that must always be
 555   // executed (call dominates loop tail).  These loops do not need non-call
 556   // safepoints (ncsfpt).
 557   void check_safepts(VectorSet &visited, Node_List &stack);
 558 
 559   // Allpaths backwards scan from loop tail, terminating each path at first safepoint
 560   // encountered.
 561   void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
 562 
 563   // Remove safepoints from loop. Optionally keeping one.
 564   void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
 565 
 566   // Convert to counted loops where possible
 567   void counted_loop( PhaseIdealLoop *phase );
 568 
 569   // Check for Node being a loop-breaking test
 570   Node *is_loop_exit(Node *iff) const;
 571 
 572   // Remove simplistic dead code from loop body
 573   void DCE_loop_body();
 574 
 575   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
 576   // Replace with a 1-in-10 exit guess.
 577   void adjust_loop_exit_prob( PhaseIdealLoop *phase );
 578 
 579   // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
 580   // Useful for unrolling loops with NO array accesses.
 581   bool policy_peel_only( PhaseIdealLoop *phase ) const;
 582 
 583   // Return TRUE or FALSE if the loop should be unswitched -- clone
 584   // loop with an invariant test
 585   bool policy_unswitching( PhaseIdealLoop *phase ) const;
 586 
 587   // Micro-benchmark spamming.  Remove empty loops.
 588   bool do_remove_empty_loop( PhaseIdealLoop *phase );
 589 
 590   // Convert one iteration loop into normal code.
 591   bool do_one_iteration_loop( PhaseIdealLoop *phase );
 592 
 593   // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
 594   // move some loop-invariant test (usually a null-check) before the loop.
 595   bool policy_peeling(PhaseIdealLoop *phase);
 596 
 597   uint estimate_peeling(PhaseIdealLoop *phase);
 598 
 599   // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
 600   // known trip count in the counted loop node.
 601   bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
 602 
 603   // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
 604   // if the loop is a counted loop and the loop body is small enough.
 605   bool policy_unroll(PhaseIdealLoop *phase);
 606 
 607   // Loop analyses to map to a maximal superword unrolling for vectorization.
 608   void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
 609 
 610   // Return TRUE or FALSE if the loop should be range-check-eliminated.
 611   // Gather a list of IF tests that are dominated by iteration splitting;
 612   // also gather the end of the first split and the start of the 2nd split.
 613   bool policy_range_check( PhaseIdealLoop *phase ) const;
 614 
 615   // Return TRUE or FALSE if the loop should be cache-line aligned.
 616   // Gather the expression that does the alignment.  Note that only
 617   // one array base can be aligned in a loop (unless the VM guarantees
 618   // mutual alignment).  Note that if we vectorize short memory ops
 619   // into longer memory ops, we may want to increase alignment.
 620   bool policy_align( PhaseIdealLoop *phase ) const;
 621 
 622   // Return TRUE if "iff" is a range check.
 623   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
 624 
 625   // Estimate the number of nodes required when cloning a loop (body).
 626   uint est_loop_clone_sz(uint factor) const;
 627   // Estimate the number of nodes required when unrolling a loop (body).
 628   uint est_loop_unroll_sz(uint factor) const;
 629 
 630   // Compute loop trip count if possible
 631   void compute_trip_count(PhaseIdealLoop* phase);
 632 
 633   // Compute loop trip count from profile data
 634   float compute_profile_trip_cnt_helper(Node* n);
 635   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 636 
 637   // Reassociate invariant expressions.
 638   void reassociate_invariants(PhaseIdealLoop *phase);
 639   // Reassociate invariant add and subtract expressions.
 640   Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
 641   // Return nonzero index of invariant operand if invariant and variant
 642   // are combined with an Add or Sub. Helper for reassociate_invariants.
 643   int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
 644 
 645   // Return true if n is invariant
 646   bool is_invariant(Node* n) const;
 647 
 648   // Put loop body on igvn work list
 649   void record_for_igvn();
 650 
 651   bool is_root() { return _parent == NULL; }
 652   // A proper/reducible loop w/o any (occasional) dead back-edge.
 653   bool is_loop() { return !_irreducible && !tail()->is_top(); }
 654   bool is_counted()   { return is_loop() && _head->is_CountedLoop(); }
 655   bool is_innermost() { return is_loop() && _child == NULL; }
 656 
 657   void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
 658 
 659 #ifndef PRODUCT
 660   void dump_head() const;       // Dump loop head only
 661   void dump() const;            // Dump this loop recursively
 662   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 663 #endif
 664 
 665  private:
 666   enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
 667 
 668   // Estimate the number of nodes resulting from control and data flow merge.
 669   uint est_loop_flow_merge_sz() const;
 670 };
 671 
 672 // -----------------------------PhaseIdealLoop---------------------------------
 673 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
 674 // into a loop tree. Drives the loop-based transformations on the ideal graph.
 675 class PhaseIdealLoop : public PhaseTransform {
 676   friend class IdealLoopTree;
 677   friend class SuperWord;
 678   friend class CountedLoopReserveKit;
 679   friend class ShenandoahBarrierC2Support;
 680   friend class AutoNodeBudget;
 681 
 682   // Pre-computed def-use info
 683   PhaseIterGVN &_igvn;
 684 
 685   // Head of loop tree
 686   IdealLoopTree* _ltree_root;
 687 
 688   // Array of pre-order numbers, plus post-visited bit.
 689   // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
 690   // ODD for post-visited.  Other bits are the pre-order number.
 691   uint *_preorders;
 692   uint _max_preorder;
 693 
 694   const PhaseIdealLoop* _verify_me;
 695   bool _verify_only;
 696 
 697   // Allocate _preorders[] array
 698   void allocate_preorders() {
 699     _max_preorder = C->unique()+8;
 700     _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
 701     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 702   }
 703 
 704   // Allocate _preorders[] array
 705   void reallocate_preorders() {
 706     if ( _max_preorder < C->unique() ) {
 707       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
 708       _max_preorder = C->unique();
 709     }
 710     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 711   }
 712 
 713   // Check to grow _preorders[] array for the case when build_loop_tree_impl()
 714   // adds new nodes.
 715   void check_grow_preorders( ) {
 716     if ( _max_preorder < C->unique() ) {
 717       uint newsize = _max_preorder<<1;  // double size of array
 718       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
 719       memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
 720       _max_preorder = newsize;
 721     }
 722   }
 723   // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
 724   int is_visited( Node *n ) const { return _preorders[n->_idx]; }
 725   // Pre-order numbers are written to the Nodes array as low-bit-set values.
 726   void set_preorder_visited( Node *n, int pre_order ) {
 727     assert( !is_visited( n ), "already set" );
 728     _preorders[n->_idx] = (pre_order<<1);
 729   };
 730   // Return pre-order number.
 731   int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
 732 
 733   // Check for being post-visited.
 734   // Should be previsited already (checked with assert(is_visited(n))).
 735   int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
 736 
 737   // Mark as post visited
 738   void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
 739 
 740 public:
 741   // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
 742   // Returns true if "n" is a data node, false if it's a control node.
 743   bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
 744 
 745 private:
 746   // clear out dead code after build_loop_late
 747   Node_List _deadlist;
 748 
 749   // Support for faster execution of get_late_ctrl()/dom_lca()
 750   // when a node has many uses and dominator depth is deep.
 751   Node_Array _dom_lca_tags;
 752   void   init_dom_lca_tags();
 753   void   clear_dom_lca_tags();
 754 
 755   // Helper for debugging bad dominance relationships
 756   bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
 757 
 758   Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
 759 
 760   // Inline wrapper for frequent cases:
 761   // 1) only one use
 762   // 2) a use is the same as the current LCA passed as 'n1'
 763   Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
 764     assert( n->is_CFG(), "" );
 765     // Fast-path NULL lca
 766     if( lca != NULL && lca != n ) {
 767       assert( lca->is_CFG(), "" );
 768       // find LCA of all uses
 769       n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
 770     }
 771     return find_non_split_ctrl(n);
 772   }
 773   Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
 774 
 775   // Helper function for directing control inputs away from CFG split points.
 776   Node *find_non_split_ctrl( Node *ctrl ) const {
 777     if (ctrl != NULL) {
 778       if (ctrl->is_MultiBranch()) {
 779         ctrl = ctrl->in(0);
 780       }
 781       assert(ctrl->is_CFG(), "CFG");
 782     }
 783     return ctrl;
 784   }
 785 
 786   Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
 787 
 788 #ifdef ASSERT
 789   void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
 790 #endif
 791   void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* start, Node* end, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
 792                                                     uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
 793                                                     Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
 794   void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* start, Node* end, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
 795                                              uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
 796                                              Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
 797   Node* clone_skeleton_predicate(Node* iff, Node* value, Node* predicate, Node* uncommon_proj,
 798                                  Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj);
 799   bool skeleton_predicate_has_opaque(IfNode* iff);
 800   void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
 801   void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol);
 802 
 803 public:
 804 
 805   PhaseIterGVN &igvn() const { return _igvn; }
 806 
 807   static bool is_canonical_loop_entry(CountedLoopNode* cl);
 808 
 809   bool has_node( Node* n ) const {
 810     guarantee(n != NULL, "No Node.");
 811     return _nodes[n->_idx] != NULL;
 812   }
 813   // check if transform created new nodes that need _ctrl recorded
 814   Node *get_late_ctrl( Node *n, Node *early );
 815   Node *get_early_ctrl( Node *n );
 816   Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
 817   void set_early_ctrl( Node *n );
 818   void set_subtree_ctrl( Node *root );
 819   void set_ctrl( Node *n, Node *ctrl ) {
 820     assert( !has_node(n) || has_ctrl(n), "" );
 821     assert( ctrl->in(0), "cannot set dead control node" );
 822     assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
 823     _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
 824   }
 825   // Set control and update loop membership
 826   void set_ctrl_and_loop(Node* n, Node* ctrl) {
 827     IdealLoopTree* old_loop = get_loop(get_ctrl(n));
 828     IdealLoopTree* new_loop = get_loop(ctrl);
 829     if (old_loop != new_loop) {
 830       if (old_loop->_child == NULL) old_loop->_body.yank(n);
 831       if (new_loop->_child == NULL) new_loop->_body.push(n);
 832     }
 833     set_ctrl(n, ctrl);
 834   }
 835   // Control nodes can be replaced or subsumed.  During this pass they
 836   // get their replacement Node in slot 1.  Instead of updating the block
 837   // location of all Nodes in the subsumed block, we lazily do it.  As we
 838   // pull such a subsumed block out of the array, we write back the final
 839   // correct block.
 840   Node *get_ctrl( Node *i ) {
 841 
 842     assert(has_node(i), "");
 843     Node *n = get_ctrl_no_update(i);
 844     _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
 845     assert(has_node(i) && has_ctrl(i), "");
 846     assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
 847     return n;
 848   }
 849   // true if CFG node d dominates CFG node n
 850   bool is_dominator(Node *d, Node *n);
 851   // return get_ctrl for a data node and self(n) for a CFG node
 852   Node* ctrl_or_self(Node* n) {
 853     if (has_ctrl(n))
 854       return get_ctrl(n);
 855     else {
 856       assert (n->is_CFG(), "must be a CFG node");
 857       return n;
 858     }
 859   }
 860 
 861   Node *get_ctrl_no_update_helper(Node *i) const {
 862     assert(has_ctrl(i), "should be control, not loop");
 863     return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
 864   }
 865 
 866   Node *get_ctrl_no_update(Node *i) const {
 867     assert( has_ctrl(i), "" );
 868     Node *n = get_ctrl_no_update_helper(i);
 869     if (!n->in(0)) {
 870       // Skip dead CFG nodes
 871       do {
 872         n = get_ctrl_no_update_helper(n);
 873       } while (!n->in(0));
 874       n = find_non_split_ctrl(n);
 875     }
 876     return n;
 877   }
 878 
 879   // Check for loop being set
 880   // "n" must be a control node. Returns true if "n" is known to be in a loop.
 881   bool has_loop( Node *n ) const {
 882     assert(!has_node(n) || !has_ctrl(n), "");
 883     return has_node(n);
 884   }
 885   // Set loop
 886   void set_loop( Node *n, IdealLoopTree *loop ) {
 887     _nodes.map(n->_idx, (Node*)loop);
 888   }
 889   // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
 890   // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
 891   // from old_node to new_node to support the lazy update.  Reference
 892   // replaces loop reference, since that is not needed for dead node.
 893   void lazy_update(Node *old_node, Node *new_node) {
 894     assert(old_node != new_node, "no cycles please");
 895     // Re-use the side array slot for this node to provide the
 896     // forwarding pointer.
 897     _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
 898   }
 899   void lazy_replace(Node *old_node, Node *new_node) {
 900     _igvn.replace_node(old_node, new_node);
 901     lazy_update(old_node, new_node);
 902   }
 903 
 904 private:
 905 
 906   // Place 'n' in some loop nest, where 'n' is a CFG node
 907   void build_loop_tree();
 908   int build_loop_tree_impl( Node *n, int pre_order );
 909   // Insert loop into the existing loop tree.  'innermost' is a leaf of the
 910   // loop tree, not the root.
 911   IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
 912 
 913   // Place Data nodes in some loop nest
 914   void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 915   void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 916   void build_loop_late_post_work(Node* n, bool pinned);
 917   void build_loop_late_post(Node* n);
 918   void verify_strip_mined_scheduling(Node *n, Node* least);
 919 
 920   // Array of immediate dominance info for each CFG node indexed by node idx
 921 private:
 922   uint _idom_size;
 923   Node **_idom;                  // Array of immediate dominators
 924   uint *_dom_depth;              // Used for fast LCA test
 925   GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
 926 
 927   // Perform verification that the graph is valid.
 928   PhaseIdealLoop( PhaseIterGVN &igvn) :
 929     PhaseTransform(Ideal_Loop),
 930     _igvn(igvn),
 931     _verify_me(NULL),
 932     _verify_only(true),
 933     _dom_lca_tags(arena()),  // Thread::resource_area
 934     _nodes_required(UINT_MAX) {
 935     build_and_optimize(LoopOptsVerify);
 936   }
 937 
 938   // build the loop tree and perform any requested optimizations
 939   void build_and_optimize(LoopOptsMode mode);
 940 
 941   // Dominators for the sea of nodes
 942   void Dominators();
 943 
 944   // Compute the Ideal Node to Loop mapping
 945   PhaseIdealLoop(PhaseIterGVN &igvn, LoopOptsMode mode) :
 946     PhaseTransform(Ideal_Loop),
 947     _igvn(igvn),
 948     _verify_me(NULL),
 949     _verify_only(false),
 950     _dom_lca_tags(arena()),  // Thread::resource_area
 951     _nodes_required(UINT_MAX) {
 952     build_and_optimize(mode);
 953   }
 954 
 955   // Verify that verify_me made the same decisions as a fresh run.
 956   PhaseIdealLoop(PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
 957     PhaseTransform(Ideal_Loop),
 958     _igvn(igvn),
 959     _verify_me(verify_me),
 960     _verify_only(false),
 961     _dom_lca_tags(arena()),  // Thread::resource_area
 962     _nodes_required(UINT_MAX) {
 963     build_and_optimize(LoopOptsVerify);
 964   }
 965 
 966 public:
 967   Node* idom_no_update(Node* d) const {
 968     return idom_no_update(d->_idx);
 969   }
 970 
 971   Node* idom_no_update(uint didx) const {
 972     assert(didx < _idom_size, "oob");
 973     Node* n = _idom[didx];
 974     assert(n != NULL,"Bad immediate dominator info.");
 975     while (n->in(0) == NULL) { // Skip dead CFG nodes
 976       n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
 977       assert(n != NULL,"Bad immediate dominator info.");
 978     }
 979     return n;
 980   }
 981 
 982   Node *idom(Node* d) const {
 983     return idom(d->_idx);
 984   }
 985 
 986   Node *idom(uint didx) const {
 987     Node *n = idom_no_update(didx);
 988     _idom[didx] = n; // Lazily remove dead CFG nodes from table.
 989     return n;
 990   }
 991 
 992   uint dom_depth(Node* d) const {
 993     guarantee(d != NULL, "Null dominator info.");
 994     guarantee(d->_idx < _idom_size, "");
 995     return _dom_depth[d->_idx];
 996   }
 997   void set_idom(Node* d, Node* n, uint dom_depth);
 998   // Locally compute IDOM using dom_lca call
 999   Node *compute_idom( Node *region ) const;
1000   // Recompute dom_depth
1001   void recompute_dom_depth();
1002 
1003   // Is safept not required by an outer loop?
1004   bool is_deleteable_safept(Node* sfpt);
1005 
1006   // Replace parallel induction variable (parallel to trip counter)
1007   void replace_parallel_iv(IdealLoopTree *loop);
1008 
1009   Node *dom_lca( Node *n1, Node *n2 ) const {
1010     return find_non_split_ctrl(dom_lca_internal(n1, n2));
1011   }
1012   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1013 
1014   // Build and verify the loop tree without modifying the graph.  This
1015   // is useful to verify that all inputs properly dominate their uses.
1016   static void verify(PhaseIterGVN& igvn) {
1017 #ifdef ASSERT
1018     ResourceMark rm;
1019     PhaseIdealLoop v(igvn);
1020 #endif
1021   }
1022 
1023   // Recommended way to use PhaseIdealLoop.
1024   // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1025   static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1026     ResourceMark rm;
1027     PhaseIdealLoop v(igvn, mode);
1028   }
1029 
1030   // True if the method has at least 1 irreducible loop
1031   bool _has_irreducible_loops;
1032 
1033   // Per-Node transform
1034   virtual Node* transform(Node* n) { return NULL; }
1035 
1036   Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1037   Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1038   Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1039   Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1040   PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1041 
1042   bool is_counted_loop(Node* n, IdealLoopTree* &loop);
1043   IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1044   IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1045                                                IdealLoopTree* loop, float cl_prob, float le_fcnt,
1046                                                Node*& entry_control, Node*& iffalse);
1047 
1048   Node* exact_limit( IdealLoopTree *loop );
1049 
1050   // Return a post-walked LoopNode
1051   IdealLoopTree *get_loop( Node *n ) const {
1052     // Dead nodes have no loop, so return the top level loop instead
1053     if (!has_node(n))  return _ltree_root;
1054     assert(!has_ctrl(n), "");
1055     return (IdealLoopTree*)_nodes[n->_idx];
1056   }
1057 
1058   IdealLoopTree* ltree_root() const { return _ltree_root; }
1059 
1060   // Is 'n' a (nested) member of 'loop'?
1061   int is_member( const IdealLoopTree *loop, Node *n ) const {
1062     return loop->is_member(get_loop(n)); }
1063 
1064   // This is the basic building block of the loop optimizations.  It clones an
1065   // entire loop body.  It makes an old_new loop body mapping; with this
1066   // mapping you can find the new-loop equivalent to an old-loop node.  All
1067   // new-loop nodes are exactly equal to their old-loop counterparts, all
1068   // edges are the same.  All exits from the old-loop now have a RegionNode
1069   // that merges the equivalent new-loop path.  This is true even for the
1070   // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
1071   // now come from (one or more) Phis that merge their new-loop equivalents.
1072   // Parameter side_by_side_idom:
1073   //   When side_by_size_idom is NULL, the dominator tree is constructed for
1074   //      the clone loop to dominate the original.  Used in construction of
1075   //      pre-main-post loop sequence.
1076   //   When nonnull, the clone and original are side-by-side, both are
1077   //      dominated by the passed in side_by_side_idom node.  Used in
1078   //      construction of unswitched loops.
1079   enum CloneLoopMode {
1080     IgnoreStripMined = 0,        // Only clone inner strip mined loop
1081     CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1082     ControlAroundStripMined = 2  // Only clone inner strip mined loop,
1083                                  // result control flow branches
1084                                  // either to inner clone or outer
1085                                  // strip mined loop.
1086   };
1087   void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1088                   CloneLoopMode mode, Node* side_by_side_idom = NULL);
1089   void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1090                                    IdealLoopTree* loop, IdealLoopTree* companion_loop,
1091                                    Node_List*& split_if_set, Node_List*& split_bool_set,
1092                                    Node_List*& split_cex_set, Node_List& worklist,
1093                                    uint new_counter, CloneLoopMode mode);
1094   void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1095                         IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1096                         Node_List& extra_data_nodes);
1097 
1098   // If we got the effect of peeling, either by actually peeling or by
1099   // making a pre-loop which must execute at least once, we can remove
1100   // all loop-invariant dominated tests in the main body.
1101   void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1102 
1103   // Generate code to do a loop peel for the given loop (and body).
1104   // old_new is a temp array.
1105   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1106 
1107   // Add pre and post loops around the given loop.  These loops are used
1108   // during RCE, unrolling and aligning loops.
1109   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1110 
1111   // Add post loop after the given loop.
1112   Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1113                          CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1114                          Node *incr, Node *limit, CountedLoopNode *&post_head);
1115 
1116   // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1117   void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1118 
1119   // Add a vector post loop between a vector main loop and the current post loop
1120   void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1121   // If Node n lives in the back_ctrl block, we clone a private version of n
1122   // in preheader_ctrl block and return that, otherwise return n.
1123   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1124 
1125   // Take steps to maximally unroll the loop.  Peel any odd iterations, then
1126   // unroll to do double iterations.  The next round of major loop transforms
1127   // will repeat till the doubled loop body does all remaining iterations in 1
1128   // pass.
1129   void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1130 
1131   // Unroll the loop body one step - make each trip do 2 iterations.
1132   void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1133 
1134   // Mark vector reduction candidates before loop unrolling
1135   void mark_reductions( IdealLoopTree *loop );
1136 
1137   // Return true if exp is a constant times an induction var
1138   bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1139 
1140   // Return true if exp is a scaled induction var plus (or minus) constant
1141   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1142 
1143   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1144   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1145                                         int opcode, bool if_cont_is_true_proj = true);
1146 
1147   void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1148 
1149   static Node* skip_all_loop_predicates(Node* entry);
1150   static Node* skip_loop_predicates(Node* entry);
1151 
1152   // Find a good location to insert a predicate
1153   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1154   // Find a predicate
1155   static Node* find_predicate(Node* entry);
1156   // Construct a range check for a predicate if
1157   BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1158                          int scale, Node* offset,
1159                          Node* init, Node* limit, jint stride,
1160                          Node* range, bool upper, bool &overflow);
1161 
1162   // Implementation of the loop predication to promote checks outside the loop
1163   bool loop_predication_impl(IdealLoopTree *loop);
1164   bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1165                                     CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1166                                     Deoptimization::DeoptReason reason);
1167   bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt);
1168   void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1169                                         PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1170                                         Node_List& if_proj_list);
1171   ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1172                                               ProjNode* proj, ProjNode *predicate_proj,
1173                                               ProjNode* upper_bound_proj,
1174                                               int scale, Node* offset,
1175                                               Node* init, Node* limit, jint stride,
1176                                               Node* rng, bool& overflow,
1177                                               Deoptimization::DeoptReason reason);
1178   Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1179                                   Node* predicate_proj, int scale_con, Node* offset,
1180                                   Node* limit, jint stride_con, Node* value);
1181 
1182   // Helper function to collect predicate for eliminating the useless ones
1183   void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1184   void eliminate_useless_predicates();
1185 
1186   // Change the control input of expensive nodes to allow commoning by
1187   // IGVN when it is guaranteed to not result in a more frequent
1188   // execution of the expensive node. Return true if progress.
1189   bool process_expensive_nodes();
1190 
1191   // Check whether node has become unreachable
1192   bool is_node_unreachable(Node *n) const {
1193     return !has_node(n) || n->is_unreachable(_igvn);
1194   }
1195 
1196   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1197   int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1198 
1199   // Check to see if do_range_check(...) cleaned the main loop of range-checks
1200   void has_range_checks(IdealLoopTree *loop);
1201 
1202   // Process post loops which have range checks and try to build a multi-version
1203   // guard to safely determine if we can execute the post loop which was RCE'd.
1204   bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1205 
1206   // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1207   void poison_rce_post_loop(IdealLoopTree *rce_loop);
1208 
1209   // Create a slow version of the loop by cloning the loop
1210   // and inserting an if to select fast-slow versions.
1211   ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1212                                         Node_List &old_new,
1213                                         int opcode,
1214                                         CloneLoopMode mode);
1215 
1216   // Clone a loop and return the clone head (clone_loop_head).
1217   // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1218   // This routine was created for usage in CountedLoopReserveKit.
1219   //
1220   //    int(1) -> If -> IfTrue -> original_loop_head
1221   //              |
1222   //              V
1223   //           IfFalse -> clone_loop_head (returned by function pointer)
1224   //
1225   LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1226   // Clone loop with an invariant test (that does not exit) and
1227   // insert a clone of the test that selects which version to
1228   // execute.
1229   void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1230 
1231   // Find candidate "if" for unswitching
1232   IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1233 
1234   // Range Check Elimination uses this function!
1235   // Constrain the main loop iterations so the affine function:
1236   //    low_limit <= scale_con * I + offset  <  upper_limit
1237   // always holds true.  That is, either increase the number of iterations in
1238   // the pre-loop or the post-loop until the condition holds true in the main
1239   // loop.  Scale_con, offset and limit are all loop invariant.
1240   void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
1241   // Helper function for add_constraint().
1242   Node* adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up);
1243 
1244   // Partially peel loop up through last_peel node.
1245   bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1246 
1247   // Create a scheduled list of nodes control dependent on ctrl set.
1248   void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1249   // Has a use in the vector set
1250   bool has_use_in_set( Node* n, VectorSet& vset );
1251   // Has use internal to the vector set (ie. not in a phi at the loop head)
1252   bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1253   // clone "n" for uses that are outside of loop
1254   int  clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1255   // clone "n" for special uses that are in the not_peeled region
1256   void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1257                                           VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1258   // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1259   void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1260 #ifdef ASSERT
1261   // Validate the loop partition sets: peel and not_peel
1262   bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1263   // Ensure that uses outside of loop are of the right form
1264   bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1265                                  uint orig_exit_idx, uint clone_exit_idx);
1266   bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1267 #endif
1268 
1269   // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1270   int stride_of_possible_iv( Node* iff );
1271   bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1272   // Return the (unique) control output node that's in the loop (if it exists.)
1273   Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1274   // Insert a signed compare loop exit cloned from an unsigned compare.
1275   IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1276   void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1277   // Utility to register node "n" with PhaseIdealLoop
1278   void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1279   // Utility to create an if-projection
1280   ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1281   // Force the iff control output to be the live_proj
1282   Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1283   // Insert a region before an if projection
1284   RegionNode* insert_region_before_proj(ProjNode* proj);
1285   // Insert a new if before an if projection
1286   ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1287 
1288   // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1289   // "Nearly" because all Nodes have been cloned from the original in the loop,
1290   // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1291   // through the Phi recursively, and return a Bool.
1292   Node *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1293   CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1294 
1295 
1296   // Rework addressing expressions to get the most loop-invariant stuff
1297   // moved out.  We'd like to do all associative operators, but it's especially
1298   // important (common) to do address expressions.
1299   Node *remix_address_expressions( Node *n );
1300 
1301   // Convert add to muladd to generate MuladdS2I under certain criteria
1302   Node * convert_add_to_muladd(Node * n);
1303 
1304   // Attempt to use a conditional move instead of a phi/branch
1305   Node *conditional_move( Node *n );
1306 
1307   // Reorganize offset computations to lower register pressure.
1308   // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1309   // (which are then alive with the post-incremented trip counter
1310   // forcing an extra register move)
1311   void reorg_offsets( IdealLoopTree *loop );
1312 
1313   // Check for aggressive application of 'split-if' optimization,
1314   // using basic block level info.
1315   void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack);
1316   Node *split_if_with_blocks_pre ( Node *n );
1317   void  split_if_with_blocks_post( Node *n );
1318   Node *has_local_phi_input( Node *n );
1319   // Mark an IfNode as being dominated by a prior test,
1320   // without actually altering the CFG (and hence IDOM info).
1321   void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1322 
1323   // Split Node 'n' through merge point
1324   Node *split_thru_region( Node *n, Node *region );
1325   // Split Node 'n' through merge point if there is enough win.
1326   Node *split_thru_phi( Node *n, Node *region, int policy );
1327   // Found an If getting its condition-code input from a Phi in the
1328   // same block.  Split thru the Region.
1329   void do_split_if( Node *iff );
1330 
1331   // Conversion of fill/copy patterns into intrinsic versions
1332   bool do_intrinsify_fill();
1333   bool intrinsify_fill(IdealLoopTree* lpt);
1334   bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1335                        Node*& shift, Node*& offset);
1336 
1337 private:
1338   // Return a type based on condition control flow
1339   const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1340   const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1341  // Helpers for filtered type
1342   const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1343 
1344   // Helper functions
1345   Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1346   Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1347   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 );
1348   bool split_up( Node *n, Node *blk1, Node *blk2 );
1349   void sink_use( Node *use, Node *post_loop );
1350   Node *place_near_use( Node *useblock ) const;
1351   Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1352   void try_move_store_after_loop(Node* n);
1353   bool identical_backtoback_ifs(Node *n);
1354   bool can_split_if(Node *n_ctrl);
1355   DEBUG_ONLY(void check_divide_by_zero(const Node* n, const Type* zero) const;)
1356 
1357   // Determine if a method is too big for a/another round of split-if, based on
1358   // a magic (approximate) ratio derived from the equally magic constant 35000,
1359   // previously used for this purpose (but without relating to the node limit).
1360   bool must_throttle_split_if() {
1361     uint threshold = C->max_node_limit() * 2 / 5;
1362     return C->live_nodes() > threshold;
1363   }
1364 
1365   // A simplistic node request tracking mechanism, where
1366   //   = UINT_MAX   Request not valid or made final.
1367   //   < UINT_MAX   Nodes currently requested (estimate).
1368   uint _nodes_required;
1369 
1370   enum { REQUIRE_MIN = 70 };
1371 
1372   uint nodes_required() const { return _nodes_required; }
1373 
1374   // Given the _currently_  available number of nodes, check  whether there is
1375   // "room" for an additional request or not, considering the already required
1376   // number of  nodes.  Return TRUE if  the new request is  exceeding the node
1377   // budget limit, otherwise return FALSE.  Note that this interpretation will
1378   // act pessimistic on  additional requests when new nodes  have already been
1379   // generated since the 'begin'.  This behaviour fits with the intention that
1380   // node estimates/requests should be made upfront.
1381   bool exceeding_node_budget(uint required = 0) {
1382     assert(C->live_nodes() < C->max_node_limit(), "sanity");
1383     uint available = C->max_node_limit() - C->live_nodes();
1384     return available < required + _nodes_required + REQUIRE_MIN;
1385   }
1386 
1387   uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1388     precond(require > 0);
1389     _nodes_required += MAX2(require, minreq);
1390     return _nodes_required;
1391   }
1392 
1393   bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1394     return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1395   }
1396 
1397   uint require_nodes_begin() {
1398     assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1399     _nodes_required = 0;
1400     return C->live_nodes();
1401   }
1402 
1403   // When a node request is final,  optionally check that the requested number
1404   // of nodes was  reasonably correct with respect to the  number of new nodes
1405   // introduced since the last 'begin'. Always check that we have not exceeded
1406   // the maximum node limit.
1407   void require_nodes_final(uint live_at_begin, bool check_estimate) {
1408     assert(_nodes_required < UINT_MAX, "Bad state (final).");
1409 
1410 #ifdef ASSERT
1411     if (check_estimate) {
1412       // Check that the node budget request was not off by too much (x2).
1413       // Should this be the case we _surely_ need to improve the estimates
1414       // used in our budget calculations.
1415       if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1416         log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1417                               C->live_nodes() - live_at_begin, _nodes_required);
1418       }
1419     }
1420 #endif
1421     // Assert that we have stayed within the node budget limit.
1422     assert(C->live_nodes() < C->max_node_limit(),
1423            "Exceeding node budget limit: %d + %d > %d (request = %d)",
1424            C->live_nodes() - live_at_begin, live_at_begin,
1425            C->max_node_limit(), _nodes_required);
1426 
1427     _nodes_required = UINT_MAX;
1428   }
1429 
1430   // Clone loop predicates to slow and fast loop when unswitching a loop
1431   Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check, bool is_slow_loop,
1432                               uint idx_before_clone, Node_List &old_new);
1433   ProjNode* clone_loop_predicate(ProjNode* predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1434                                  bool is_slow_loop, uint idx_before_clone, Node_List &old_new);
1435   void clone_concrete_loop_predicates(Deoptimization::DeoptReason reason, ProjNode* old_predicate_proj,
1436                                       ProjNode* new_predicate_proj, bool is_slow_loop,
1437                                       uint idx_before_clone, Node_List &old_new);
1438 
1439   bool _created_loop_node;
1440 
1441 public:
1442   void set_created_loop_node() { _created_loop_node = true; }
1443   bool created_loop_node()     { return _created_loop_node; }
1444   void register_new_node(Node* n, Node* blk);
1445 
1446 #ifdef ASSERT
1447   void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1448 #endif
1449 
1450 #ifndef PRODUCT
1451   void dump() const;
1452   void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1453   void verify() const;          // Major slow  :-)
1454   void verify_compare(Node* n, const PhaseIdealLoop* loop_verify, VectorSet &visited) const;
1455   IdealLoopTree* get_loop_idx(Node* n) const {
1456     // Dead nodes have no loop, so return the top level loop instead
1457     return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1458   }
1459   // Print some stats
1460   static void print_statistics();
1461   static int _loop_invokes;     // Count of PhaseIdealLoop invokes
1462   static int _loop_work;        // Sum of PhaseIdealLoop x _unique
1463 #endif
1464 
1465   void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1466 };
1467 
1468 
1469 class AutoNodeBudget : public StackObj
1470 {
1471 public:
1472   enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1473 
1474   AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1475     : _phase(phase),
1476       _check_at_final(chk == BUDGET_CHECK),
1477       _nodes_at_begin(0)
1478   {
1479     precond(_phase != NULL);
1480 
1481     _nodes_at_begin = _phase->require_nodes_begin();
1482   }
1483 
1484   ~AutoNodeBudget() {
1485 #ifndef PRODUCT
1486     if (TraceLoopOpts) {
1487       uint request = _phase->nodes_required();
1488       uint delta   = _phase->C->live_nodes() - _nodes_at_begin;
1489 
1490       if (request < delta) {
1491         tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1492       } else {
1493         uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1494         // Identify the worst estimates as "poor" ones.
1495         if (request > REQUIRE_MIN && delta > 0) {
1496           if ((delta >  REQUIRE_MIN && request >  3 * delta) ||
1497               (delta <= REQUIRE_MIN && request > 10 * delta)) {
1498             tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1499           }
1500         }
1501       }
1502     }
1503 #endif // PRODUCT
1504     _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1505   }
1506 
1507 private:
1508   PhaseIdealLoop* _phase;
1509   bool _check_at_final;
1510   uint _nodes_at_begin;
1511 };
1512 
1513 
1514 // This kit may be used for making of a reserved copy of a loop before this loop
1515 //  goes under non-reversible changes.
1516 //
1517 // Function create_reserve() creates a reserved copy (clone) of the loop.
1518 // The reserved copy is created by calling
1519 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1520 // the original and reserved loops are connected in the outer graph.
1521 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1522 //
1523 // By default the reserved copy (clone) of the loop is created as dead code - it is
1524 // dominated in the outer loop by this node chain:
1525 //   intcon(1)->If->IfFalse->reserved_copy.
1526 // The original loop is dominated by the the same node chain but IfTrue projection:
1527 //   intcon(0)->If->IfTrue->original_loop.
1528 //
1529 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1530 // and the dtor, checks _use_new value.
1531 // If _use_new == false, it "switches" control to reserved copy of the loop
1532 // by simple replacing of node intcon(1) with node intcon(0).
1533 //
1534 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1535 //
1536 // void CountedLoopReserveKit_example()
1537 // {
1538 //    CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1539 //    if (DoReserveCopy && !lrk.has_reserved()) {
1540 //      return; //failed to create reserved loop copy
1541 //    }
1542 //    ...
1543 //    //something is wrong, switch to original loop
1544 ///   if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1545 //    ...
1546 //    //everything worked ok, return with the newly modified loop
1547 //    lrk.use_new();
1548 //    return; // ~CountedLoopReserveKit does nothing once use_new() was called
1549 //  }
1550 //
1551 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1552 // the dtor will "switch to the original" loop.
1553 // NOTE. You you modify outside of the original loop this class is no help.
1554 //
1555 class CountedLoopReserveKit {
1556   private:
1557     PhaseIdealLoop* _phase;
1558     IdealLoopTree*  _lpt;
1559     LoopNode*       _lp;
1560     IfNode*         _iff;
1561     LoopNode*       _lp_reserved;
1562     bool            _has_reserved;
1563     bool            _use_new;
1564     const bool      _active; //may be set to false in ctor, then the object is dummy
1565 
1566   public:
1567     CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1568     ~CountedLoopReserveKit();
1569     void use_new()                {_use_new = true;}
1570     void set_iff(IfNode* x)       {_iff = x;}
1571     bool has_reserved()     const { return _active && _has_reserved;}
1572   private:
1573     bool create_reserve();
1574 };// class CountedLoopReserveKit
1575 
1576 inline Node* IdealLoopTree::tail() {
1577   // Handle lazy update of _tail field.
1578   if (_tail->in(0) == NULL) {
1579     _tail = _phase->get_ctrl(_tail);
1580   }
1581   return _tail;
1582 }
1583 
1584 inline Node* IdealLoopTree::head() {
1585   // Handle lazy update of _head field.
1586   if (_head->in(0) == NULL) {
1587     _head = _phase->get_ctrl(_head);
1588   }
1589   return _head;
1590 }
1591 
1592 // Iterate over the loop tree using a preorder, left-to-right traversal.
1593 //
1594 // Example that visits all counted loops from within PhaseIdealLoop
1595 //
1596 //  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1597 //   IdealLoopTree* lpt = iter.current();
1598 //   if (!lpt->is_counted()) continue;
1599 //   ...
1600 class LoopTreeIterator : public StackObj {
1601 private:
1602   IdealLoopTree* _root;
1603   IdealLoopTree* _curnt;
1604 
1605 public:
1606   LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1607 
1608   bool done() { return _curnt == NULL; }       // Finished iterating?
1609 
1610   void next();                                 // Advance to next loop tree
1611 
1612   IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
1613 };
1614 
1615 #endif // SHARE_OPTO_LOOPNODE_HPP