1 /*
   2  * Copyright (c) 1998, 2022, 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 BaseCountedLoopEndNode;
  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    = 1<<2,
  65          HasExactTripCount   = 1<<3,
  66          InnerLoop           = 1<<4,
  67          PartialPeelLoop     = 1<<5,
  68          PartialPeelFailed   = 1<<6,
  69          HasReductions       = 1<<7,
  70          WasSlpAnalyzed      = 1<<8,
  71          PassedSlpAnalysis   = 1<<9,
  72          DoUnrollOnly        = 1<<10,
  73          VectorizedLoop      = 1<<11,
  74          HasAtomicPostLoop   = 1<<12,
  75          HasRangeChecks      = 1<<13,
  76          IsMultiversioned    = 1<<14,
  77          StripMined          = 1<<15,
  78          SubwordLoop         = 1<<16,
  79          ProfileTripFailed   = 1<<17,
  80          LoopNestInnerLoop   = 1<< 18,
  81          LoopNestLongOuterLoop = 1<< 19,
  82          FlattenedArrays     = 1<<20};
  83 
  84   char _unswitch_count;
  85   enum { _unswitch_max=3 };
  86   char _postloop_flags;
  87   enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
  88 
  89   // Expected trip count from profile data
  90   float _profile_trip_cnt;
  91 
  92 public:
  93   // Names for edge indices
  94   enum { Self=0, EntryControl, LoopBackControl };
  95 
  96   bool is_inner_loop() const { return _loop_flags & InnerLoop; }
  97   void set_inner_loop() { _loop_flags |= InnerLoop; }
  98 
  99   bool range_checks_present() const { return _loop_flags & HasRangeChecks; }
 100   bool is_multiversioned() const { return _loop_flags & IsMultiversioned; }
 101   bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
 102   bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
 103   void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
 104   bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
 105   bool is_strip_mined() const { return _loop_flags & StripMined; }
 106   bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
 107   bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
 108   bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
 109   bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
 110   bool is_flattened_arrays() const { return _loop_flags & FlattenedArrays; }
 111 
 112   void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
 113   void mark_has_reductions() { _loop_flags |= HasReductions; }
 114   void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
 115   void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
 116   void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
 117   void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
 118   void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
 119   void mark_has_range_checks() { _loop_flags |=  HasRangeChecks; }
 120   void clear_has_range_checks() { _loop_flags &= ~HasRangeChecks; }
 121   void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
 122   void mark_strip_mined() { _loop_flags |= StripMined; }
 123   void clear_strip_mined() { _loop_flags &= ~StripMined; }
 124   void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
 125   void mark_subword_loop() { _loop_flags |= SubwordLoop; }
 126   void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
 127   void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
 128   void mark_flattened_arrays() { _loop_flags |= FlattenedArrays; }
 129 
 130   int unswitch_max() { return _unswitch_max; }
 131   int unswitch_count() { return _unswitch_count; }
 132 
 133   int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
 134   void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
 135   int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
 136   void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
 137 
 138   void set_unswitch_count(int val) {
 139     assert (val <= unswitch_max(), "too many unswitches");
 140     _unswitch_count = val;
 141   }
 142 
 143   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
 144   float profile_trip_cnt()             { return _profile_trip_cnt; }
 145 
 146   LoopNode(Node *entry, Node *backedge)
 147     : RegionNode(3), _loop_flags(0), _unswitch_count(0),
 148       _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN)  {
 149     init_class_id(Class_Loop);
 150     init_req(EntryControl, entry);
 151     init_req(LoopBackControl, backedge);
 152   }
 153 
 154   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 155   virtual int Opcode() const;
 156   bool can_be_counted_loop(PhaseTransform* phase) const {
 157     return req() == 3 && in(0) != NULL &&
 158       in(1) != NULL && phase->type(in(1)) != Type::TOP &&
 159       in(2) != NULL && phase->type(in(2)) != Type::TOP;
 160   }
 161   bool is_valid_counted_loop(BasicType bt) const;
 162 #ifndef PRODUCT
 163   virtual void dump_spec(outputStream *st) const;
 164 #endif
 165 
 166   void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
 167   virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
 168   virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
 169   virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
 170   virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
 171   virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; }
 172 };
 173 
 174 //------------------------------Counted Loops----------------------------------
 175 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
 176 // path (and maybe some other exit paths).  The trip-counter exit is always
 177 // last in the loop.  The trip-counter have to stride by a constant;
 178 // the exit value is also loop invariant.
 179 
 180 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
 181 // CountedLoopNode has the incoming loop control and the loop-back-control
 182 // which is always the IfTrue before the matching CountedLoopEndNode.  The
 183 // CountedLoopEndNode has an incoming control (possibly not the
 184 // CountedLoopNode if there is control flow in the loop), the post-increment
 185 // trip-counter value, and the limit.  The trip-counter value is always of
 186 // the form (Op old-trip-counter stride).  The old-trip-counter is produced
 187 // by a Phi connected to the CountedLoopNode.  The stride is constant.
 188 // The Op is any commutable opcode, including Add, Mul, Xor.  The
 189 // CountedLoopEndNode also takes in the loop-invariant limit value.
 190 
 191 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
 192 // loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
 193 // via the old-trip-counter from the Op node.
 194 
 195 //------------------------------CountedLoopNode--------------------------------
 196 // CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
 197 // inputs the incoming loop-start control and the loop-back control, so they
 198 // act like RegionNodes.  They also take in the initial trip counter, the
 199 // loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
 200 // produce a loop-body control and the trip counter value.  Since
 201 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
 202 
 203 class BaseCountedLoopNode : public LoopNode {
 204 public:
 205   BaseCountedLoopNode(Node *entry, Node *backedge)
 206     : LoopNode(entry, backedge) {
 207   }
 208 
 209   Node *init_control() const { return in(EntryControl); }
 210   Node *back_control() const { return in(LoopBackControl); }
 211 
 212   Node* init_trip() const;
 213   Node* stride() const;
 214   bool stride_is_con() const;
 215   Node* limit() const;
 216   Node* incr() const;
 217   Node* phi() const;
 218 
 219   BaseCountedLoopEndNode* loopexit_or_null() const;
 220   BaseCountedLoopEndNode* loopexit() const;
 221 
 222   virtual BasicType bt() const = 0;
 223 
 224   jlong stride_con() const;
 225 
 226   static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
 227 };
 228 
 229 
 230 class CountedLoopNode : public BaseCountedLoopNode {
 231   // Size is bigger to hold _main_idx.  However, _main_idx does not change
 232   // the semantics so it does not appear in the hash & cmp functions.
 233   virtual uint size_of() const { return sizeof(*this); }
 234 
 235   // For Pre- and Post-loops during debugging ONLY, this holds the index of
 236   // the Main CountedLoop.  Used to assert that we understand the graph shape.
 237   node_idx_t _main_idx;
 238 
 239   // Known trip count calculated by compute_exact_trip_count()
 240   uint  _trip_count;
 241 
 242   // Log2 of original loop bodies in unrolled loop
 243   int _unrolled_count_log2;
 244 
 245   // Node count prior to last unrolling - used to decide if
 246   // unroll,optimize,unroll,optimize,... is making progress
 247   int _node_count_before_unroll;
 248 
 249   // If slp analysis is performed we record the maximum
 250   // vector mapped unroll factor here
 251   int _slp_maximum_unroll_factor;
 252 
 253   // The eventual count of vectorizable packs in slp
 254   int _slp_vector_pack_count;
 255 
 256 public:
 257   CountedLoopNode(Node *entry, Node *backedge)
 258     : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
 259       _unrolled_count_log2(0), _node_count_before_unroll(0),
 260       _slp_maximum_unroll_factor(0), _slp_vector_pack_count(0) {
 261     init_class_id(Class_CountedLoop);
 262     // Initialize _trip_count to the largest possible value.
 263     // Will be reset (lower) if the loop's trip count is known.
 264   }
 265 
 266   virtual int Opcode() const;
 267   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 268 
 269   CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
 270   CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
 271   int   stride_con() const;
 272 
 273   // Match increment with optional truncation
 274   static Node*
 275   match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
 276                                       BasicType bt);
 277 
 278   // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
 279   // can run short a few iterations and may start a few iterations in.
 280   // It will be RCE'd and unrolled and aligned.
 281 
 282   // A following 'post' loop will run any remaining iterations.  Used
 283   // during Range Check Elimination, the 'post' loop will do any final
 284   // iterations with full checks.  Also used by Loop Unrolling, where
 285   // the 'post' loop will do any epilog iterations needed.  Basically,
 286   // a 'post' loop can not profitably be further unrolled or RCE'd.
 287 
 288   // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
 289   // it may do under-flow checks for RCE and may do alignment iterations
 290   // so the following main loop 'knows' that it is striding down cache
 291   // lines.
 292 
 293   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
 294   // Aligned, may be missing it's pre-loop.
 295   bool is_normal_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
 296   bool is_pre_loop      () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
 297   bool is_main_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
 298   bool is_post_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
 299   bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
 300   bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
 301   bool has_passed_slp   () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
 302   bool is_unroll_only   () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
 303   bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
 304   bool has_atomic_post_loop  () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
 305   void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 306 
 307   int main_idx() const { return _main_idx; }
 308 
 309 
 310   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
 311   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
 312   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
 313   void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 314 
 315   void set_trip_count(uint tc) { _trip_count = tc; }
 316   uint trip_count()            { return _trip_count; }
 317 
 318   bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
 319   void set_exact_trip_count(uint tc) {
 320     _trip_count = tc;
 321     _loop_flags |= HasExactTripCount;
 322   }
 323   void set_nonexact_trip_count() {
 324     _loop_flags &= ~HasExactTripCount;
 325   }
 326   void set_notpassed_slp() {
 327     _loop_flags &= ~PassedSlpAnalysis;
 328   }
 329 
 330   void double_unrolled_count() { _unrolled_count_log2++; }
 331   int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
 332 
 333   void set_node_count_before_unroll(int ct)  { _node_count_before_unroll = ct; }
 334   int  node_count_before_unroll()            { return _node_count_before_unroll; }
 335   void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
 336   int  slp_max_unroll() const                { return _slp_maximum_unroll_factor; }
 337   void set_slp_pack_count(int pack_count)    { _slp_vector_pack_count = pack_count; }
 338   int  slp_pack_count() const                { return _slp_vector_pack_count; }
 339 
 340   virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
 341   OuterStripMinedLoopNode* outer_loop() const;
 342   virtual IfTrueNode* outer_loop_tail() const;
 343   virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
 344   virtual IfFalseNode* outer_loop_exit() const;
 345   virtual SafePointNode* outer_safepoint() const;
 346 
 347   // If this is a main loop in a pre/main/post loop nest, walk over
 348   // the predicates that were inserted by
 349   // duplicate_predicates()/add_range_check_predicate()
 350   static Node* skip_predicates_from_entry(Node* ctrl);
 351   Node* skip_predicates();
 352 
 353   virtual BasicType bt() const {
 354     return T_INT;
 355   }
 356 
 357   Node* is_canonical_loop_entry();
 358 
 359 #ifndef PRODUCT
 360   virtual void dump_spec(outputStream *st) const;
 361 #endif
 362 };
 363 
 364 class LongCountedLoopNode : public BaseCountedLoopNode {
 365 public:
 366   LongCountedLoopNode(Node *entry, Node *backedge)
 367     : BaseCountedLoopNode(entry, backedge) {
 368     init_class_id(Class_LongCountedLoop);
 369   }
 370 
 371   virtual int Opcode() const;
 372 
 373   virtual BasicType bt() const {
 374     return T_LONG;
 375   }
 376 
 377   LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
 378   LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
 379 };
 380 
 381 
 382 //------------------------------CountedLoopEndNode-----------------------------
 383 // CountedLoopEndNodes end simple trip counted loops.  They act much like
 384 // IfNodes.
 385 
 386 class BaseCountedLoopEndNode : public IfNode {
 387 public:
 388   enum { TestControl, TestValue };
 389   BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
 390     : IfNode(control, test, prob, cnt) {
 391     init_class_id(Class_BaseCountedLoopEnd);
 392   }
 393 
 394   Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
 395   Node* incr() const                { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; }
 396   Node* limit() const               { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; }
 397   Node* stride() const              { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; }
 398   Node* init_trip() const           { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; }
 399   bool stride_is_con() const        { Node *tmp = stride(); return (tmp != NULL && tmp->is_Con()); }
 400 
 401   PhiNode* phi() const {
 402     Node* tmp = incr();
 403     if (tmp && tmp->req() == 3) {
 404       Node* phi = tmp->in(1);
 405       if (phi->is_Phi()) {
 406         return phi->as_Phi();
 407       }
 408     }
 409     return NULL;
 410   }
 411 
 412   BaseCountedLoopNode* loopnode() const {
 413     // The CountedLoopNode that goes with this CountedLoopEndNode may
 414     // have been optimized out by the IGVN so be cautious with the
 415     // pattern matching on the graph
 416     PhiNode* iv_phi = phi();
 417     if (iv_phi == NULL) {
 418       return NULL;
 419     }
 420     Node* ln = iv_phi->in(0);
 421     if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
 422       return NULL;
 423     }
 424     if (ln->as_BaseCountedLoop()->bt() != bt()) {
 425       return NULL;
 426     }
 427     return ln->as_BaseCountedLoop();
 428   }
 429 
 430   BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
 431 
 432   jlong stride_con() const;
 433   virtual BasicType bt() const = 0;
 434 
 435   static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
 436 };
 437 
 438 class CountedLoopEndNode : public BaseCountedLoopEndNode {
 439 public:
 440 
 441   CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
 442     : BaseCountedLoopEndNode(control, test, prob, cnt) {
 443     init_class_id(Class_CountedLoopEnd);
 444   }
 445   virtual int Opcode() const;
 446 
 447   CountedLoopNode* loopnode() const {
 448     return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
 449   }
 450 
 451   virtual BasicType bt() const {
 452     return T_INT;
 453   }
 454 
 455 #ifndef PRODUCT
 456   virtual void dump_spec(outputStream *st) const;
 457 #endif
 458 };
 459 
 460 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
 461 public:
 462   LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
 463     : BaseCountedLoopEndNode(control, test, prob, cnt) {
 464     init_class_id(Class_LongCountedLoopEnd);
 465   }
 466 
 467   LongCountedLoopNode* loopnode() const {
 468     return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
 469   }
 470 
 471   virtual int Opcode() const;
 472 
 473   virtual BasicType bt() const {
 474     return T_LONG;
 475   }
 476 };
 477 
 478 
 479 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
 480   Node* bctrl = back_control();
 481   if (bctrl == NULL) return NULL;
 482 
 483   Node* lexit = bctrl->in(0);
 484   if (!lexit->is_BaseCountedLoopEnd()) {
 485     return NULL;
 486   }
 487   BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
 488   if (result->bt() != bt()) {
 489     return NULL;
 490   }
 491   return result;
 492 }
 493 
 494 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
 495   BaseCountedLoopEndNode* cle = loopexit_or_null();
 496   assert(cle != NULL, "loopexit is NULL");
 497   return cle;
 498 }
 499 
 500 inline Node* BaseCountedLoopNode::init_trip() const {
 501   BaseCountedLoopEndNode* cle = loopexit_or_null();
 502   return cle != NULL ? cle->init_trip() : NULL;
 503 }
 504 inline Node* BaseCountedLoopNode::stride() const {
 505   BaseCountedLoopEndNode* cle = loopexit_or_null();
 506   return cle != NULL ? cle->stride() : NULL;
 507 }
 508 
 509 inline bool BaseCountedLoopNode::stride_is_con() const {
 510   BaseCountedLoopEndNode* cle = loopexit_or_null();
 511   return cle != NULL && cle->stride_is_con();
 512 }
 513 inline Node* BaseCountedLoopNode::limit() const {
 514   BaseCountedLoopEndNode* cle = loopexit_or_null();
 515   return cle != NULL ? cle->limit() : NULL;
 516 }
 517 inline Node* BaseCountedLoopNode::incr() const {
 518   BaseCountedLoopEndNode* cle = loopexit_or_null();
 519   return cle != NULL ? cle->incr() : NULL;
 520 }
 521 inline Node* BaseCountedLoopNode::phi() const {
 522   BaseCountedLoopEndNode* cle = loopexit_or_null();
 523   return cle != NULL ? cle->phi() : NULL;
 524 }
 525 
 526 inline jlong BaseCountedLoopNode::stride_con() const {
 527   BaseCountedLoopEndNode* cle = loopexit_or_null();
 528   return cle != NULL ? cle->stride_con() : 0;
 529 }
 530 
 531 
 532 //------------------------------LoopLimitNode-----------------------------
 533 // Counted Loop limit node which represents exact final iterator value:
 534 // trip_count = (limit - init_trip + stride - 1)/stride
 535 // final_value= trip_count * stride + init_trip.
 536 // Use HW instructions to calculate it when it can overflow in integer.
 537 // Note, final_value should fit into integer since counted loop has
 538 // limit check: limit <= max_int-stride.
 539 class LoopLimitNode : public Node {
 540   enum { Init=1, Limit=2, Stride=3 };
 541  public:
 542   LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
 543     // Put it on the Macro nodes list to optimize during macro nodes expansion.
 544     init_flags(Flag_is_macro);
 545     C->add_macro_node(this);
 546   }
 547   virtual int Opcode() const;
 548   virtual const Type *bottom_type() const { return TypeInt::INT; }
 549   virtual uint ideal_reg() const { return Op_RegI; }
 550   virtual const Type* Value(PhaseGVN* phase) const;
 551   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 552   virtual Node* Identity(PhaseGVN* phase);
 553 };
 554 
 555 // Support for strip mining
 556 class OuterStripMinedLoopNode : public LoopNode {
 557 private:
 558   static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
 559 
 560 public:
 561   OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
 562     : LoopNode(entry, backedge) {
 563     init_class_id(Class_OuterStripMinedLoop);
 564     init_flags(Flag_is_macro);
 565     C->add_macro_node(this);
 566   }
 567 
 568   virtual int Opcode() const;
 569 
 570   virtual IfTrueNode* outer_loop_tail() const;
 571   virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
 572   virtual IfFalseNode* outer_loop_exit() const;
 573   virtual SafePointNode* outer_safepoint() const;
 574   void adjust_strip_mined_loop(PhaseIterGVN* igvn);
 575 
 576   void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
 577 
 578   void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
 579 
 580   static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
 581 
 582   Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
 583                          PhaseIdealLoop* iloop);
 584 };
 585 
 586 class OuterStripMinedLoopEndNode : public IfNode {
 587 public:
 588   OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
 589     : IfNode(control, test, prob, cnt) {
 590     init_class_id(Class_OuterStripMinedLoopEnd);
 591   }
 592 
 593   virtual int Opcode() const;
 594 
 595   virtual const Type* Value(PhaseGVN* phase) const;
 596   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 597 
 598   bool is_expanded(PhaseGVN *phase) const;
 599 };
 600 
 601 // -----------------------------IdealLoopTree----------------------------------
 602 class IdealLoopTree : public ResourceObj {
 603 public:
 604   IdealLoopTree *_parent;       // Parent in loop tree
 605   IdealLoopTree *_next;         // Next sibling in loop tree
 606   IdealLoopTree *_child;        // First child in loop tree
 607 
 608   // The head-tail backedge defines the loop.
 609   // If a loop has multiple backedges, this is addressed during cleanup where
 610   // we peel off the multiple backedges,  merging all edges at the bottom and
 611   // ensuring that one proper backedge flow into the loop.
 612   Node *_head;                  // Head of loop
 613   Node *_tail;                  // Tail of loop
 614   inline Node *tail();          // Handle lazy update of _tail field
 615   inline Node *head();          // Handle lazy update of _head field
 616   PhaseIdealLoop* _phase;
 617   int _local_loop_unroll_limit;
 618   int _local_loop_unroll_factor;
 619 
 620   Node_List _body;              // Loop body for inner loops
 621 
 622   uint16_t _nest;               // Nesting depth
 623   uint8_t _irreducible:1,       // True if irreducible
 624           _has_call:1,          // True if has call safepoint
 625           _has_sfpt:1,          // True if has non-call safepoint
 626           _rce_candidate:1;     // True if candidate for range check elimination
 627 
 628   Node_List* _safepts;          // List of safepoints in this loop
 629   Node_List* _required_safept;  // A inner loop cannot delete these safepts;
 630   bool  _allow_optimizations;   // Allow loop optimizations
 631 
 632   IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
 633     : _parent(0), _next(0), _child(0),
 634       _head(head), _tail(tail),
 635       _phase(phase),
 636       _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
 637       _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
 638       _safepts(NULL),
 639       _required_safept(NULL),
 640       _allow_optimizations(true)
 641   {
 642     precond(_head != NULL);
 643     precond(_tail != NULL);
 644   }
 645 
 646   // Is 'l' a member of 'this'?
 647   bool is_member(const IdealLoopTree *l) const; // Test for nested membership
 648 
 649   // Set loop nesting depth.  Accumulate has_call bits.
 650   int set_nest( uint depth );
 651 
 652   // Split out multiple fall-in edges from the loop header.  Move them to a
 653   // private RegionNode before the loop.  This becomes the loop landing pad.
 654   void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
 655 
 656   // Split out the outermost loop from this shared header.
 657   void split_outer_loop( PhaseIdealLoop *phase );
 658 
 659   // Merge all the backedges from the shared header into a private Region.
 660   // Feed that region as the one backedge to this loop.
 661   void merge_many_backedges( PhaseIdealLoop *phase );
 662 
 663   // Split shared headers and insert loop landing pads.
 664   // Insert a LoopNode to replace the RegionNode.
 665   // Returns TRUE if loop tree is structurally changed.
 666   bool beautify_loops( PhaseIdealLoop *phase );
 667 
 668   // Perform optimization to use the loop predicates for null checks and range checks.
 669   // Applies to any loop level (not just the innermost one)
 670   bool loop_predication( PhaseIdealLoop *phase);
 671 
 672   // Perform iteration-splitting on inner loops.  Split iterations to
 673   // avoid range checks or one-shot null checks.  Returns false if the
 674   // current round of loop opts should stop.
 675   bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
 676 
 677   // Driver for various flavors of iteration splitting.  Returns false
 678   // if the current round of loop opts should stop.
 679   bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
 680 
 681   // Given dominators, try to find loops with calls that must always be
 682   // executed (call dominates loop tail).  These loops do not need non-call
 683   // safepoints (ncsfpt).
 684   void check_safepts(VectorSet &visited, Node_List &stack);
 685 
 686   // Allpaths backwards scan from loop tail, terminating each path at first safepoint
 687   // encountered.
 688   void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
 689 
 690   // Remove safepoints from loop. Optionally keeping one.
 691   void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
 692 
 693   // Convert to counted loops where possible
 694   void counted_loop( PhaseIdealLoop *phase );
 695 
 696   // Check for Node being a loop-breaking test
 697   Node *is_loop_exit(Node *iff) const;
 698 
 699   // Remove simplistic dead code from loop body
 700   void DCE_loop_body();
 701 
 702   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
 703   // Replace with a 1-in-10 exit guess.
 704   void adjust_loop_exit_prob( PhaseIdealLoop *phase );
 705 
 706   // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
 707   // Useful for unrolling loops with NO array accesses.
 708   bool policy_peel_only( PhaseIdealLoop *phase ) const;
 709 
 710   // Return TRUE or FALSE if the loop should be unswitched -- clone
 711   // loop with an invariant test
 712   bool policy_unswitching( PhaseIdealLoop *phase ) const;
 713 
 714   // Micro-benchmark spamming.  Remove empty loops.
 715   bool do_remove_empty_loop( PhaseIdealLoop *phase );
 716 
 717   // Convert one iteration loop into normal code.
 718   bool do_one_iteration_loop( PhaseIdealLoop *phase );
 719 
 720   // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
 721   // move some loop-invariant test (usually a null-check) before the loop.
 722   bool policy_peeling(PhaseIdealLoop *phase);
 723 
 724   uint estimate_peeling(PhaseIdealLoop *phase);
 725 
 726   // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
 727   // known trip count in the counted loop node.
 728   bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
 729 
 730   // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
 731   // if the loop is a counted loop and the loop body is small enough.
 732   bool policy_unroll(PhaseIdealLoop *phase);
 733 
 734   // Loop analyses to map to a maximal superword unrolling for vectorization.
 735   void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
 736 
 737   // Return TRUE or FALSE if the loop should be range-check-eliminated.
 738   // Gather a list of IF tests that are dominated by iteration splitting;
 739   // also gather the end of the first split and the start of the 2nd split.
 740   bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
 741 
 742   // Return TRUE if "iff" is a range check.
 743   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode *predicate_proj)) const;
 744   bool is_range_check_if(IfNode* iff, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
 745                          jlong& scale) const;
 746 
 747   // Estimate the number of nodes required when cloning a loop (body).
 748   uint est_loop_clone_sz(uint factor) const;
 749   // Estimate the number of nodes required when unrolling a loop (body).
 750   uint est_loop_unroll_sz(uint factor) const;
 751 
 752   // Compute loop trip count if possible
 753   void compute_trip_count(PhaseIdealLoop* phase);
 754 
 755   // Compute loop trip count from profile data
 756   float compute_profile_trip_cnt_helper(Node* n);
 757   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 758 
 759   // Reassociate invariant expressions.
 760   void reassociate_invariants(PhaseIdealLoop *phase);
 761   // Reassociate invariant binary expressions.
 762   Node* reassociate(Node* n1, PhaseIdealLoop *phase);
 763   // Reassociate invariant add and subtract expressions.
 764   Node* reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase);
 765   // Return nonzero index of invariant operand if invariant and variant
 766   // are combined with an associative binary. Helper for reassociate_invariants.
 767   int find_invariant(Node* n, PhaseIdealLoop *phase);
 768   // Return TRUE if "n" is associative.
 769   bool is_associative(Node* n, Node* base=NULL);
 770 
 771   // Return true if n is invariant
 772   bool is_invariant(Node* n) const;
 773 
 774   // Put loop body on igvn work list
 775   void record_for_igvn();
 776 
 777   bool is_root() { return _parent == NULL; }
 778   // A proper/reducible loop w/o any (occasional) dead back-edge.
 779   bool is_loop() { return !_irreducible && !tail()->is_top(); }
 780   bool is_counted()   { return is_loop() && _head->is_CountedLoop(); }
 781   bool is_innermost() { return is_loop() && _child == NULL; }
 782 
 783   void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
 784 
 785 #ifndef PRODUCT
 786   void dump_head() const;       // Dump loop head only
 787   void dump() const;            // Dump this loop recursively
 788   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 789 #endif
 790 
 791  private:
 792   enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
 793 
 794   // Estimate the number of nodes resulting from control and data flow merge.
 795   uint est_loop_flow_merge_sz() const;
 796 
 797   // Check if the number of residual iterations is large with unroll_cnt.
 798   // Return true if the residual iterations are more than 10% of the trip count.
 799   bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
 800     return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
 801   }
 802 };
 803 
 804 // -----------------------------PhaseIdealLoop---------------------------------
 805 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
 806 // into a loop tree. Drives the loop-based transformations on the ideal graph.
 807 class PhaseIdealLoop : public PhaseTransform {
 808   friend class IdealLoopTree;
 809   friend class SuperWord;
 810   friend class CountedLoopReserveKit;
 811   friend class ShenandoahBarrierC2Support;
 812   friend class AutoNodeBudget;
 813 
 814   // Pre-computed def-use info
 815   PhaseIterGVN &_igvn;
 816 
 817   // Head of loop tree
 818   IdealLoopTree* _ltree_root;
 819 
 820   // Array of pre-order numbers, plus post-visited bit.
 821   // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
 822   // ODD for post-visited.  Other bits are the pre-order number.
 823   uint *_preorders;
 824   uint _max_preorder;
 825 
 826   const PhaseIdealLoop* _verify_me;
 827   bool _verify_only;
 828 
 829   // Allocate _preorders[] array
 830   void allocate_preorders() {
 831     _max_preorder = C->unique()+8;
 832     _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
 833     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 834   }
 835 
 836   // Allocate _preorders[] array
 837   void reallocate_preorders() {
 838     if ( _max_preorder < C->unique() ) {
 839       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
 840       _max_preorder = C->unique();
 841     }
 842     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 843   }
 844 
 845   // Check to grow _preorders[] array for the case when build_loop_tree_impl()
 846   // adds new nodes.
 847   void check_grow_preorders( ) {
 848     if ( _max_preorder < C->unique() ) {
 849       uint newsize = _max_preorder<<1;  // double size of array
 850       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
 851       memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
 852       _max_preorder = newsize;
 853     }
 854   }
 855   // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
 856   int is_visited( Node *n ) const { return _preorders[n->_idx]; }
 857   // Pre-order numbers are written to the Nodes array as low-bit-set values.
 858   void set_preorder_visited( Node *n, int pre_order ) {
 859     assert( !is_visited( n ), "already set" );
 860     _preorders[n->_idx] = (pre_order<<1);
 861   };
 862   // Return pre-order number.
 863   int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
 864 
 865   // Check for being post-visited.
 866   // Should be previsited already (checked with assert(is_visited(n))).
 867   int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
 868 
 869   // Mark as post visited
 870   void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
 871 
 872 public:
 873   // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
 874   // Returns true if "n" is a data node, false if it's a control node.
 875   bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
 876 
 877 private:
 878   // clear out dead code after build_loop_late
 879   Node_List _deadlist;
 880 
 881   // Support for faster execution of get_late_ctrl()/dom_lca()
 882   // when a node has many uses and dominator depth is deep.
 883   GrowableArray<jlong> _dom_lca_tags;
 884   uint _dom_lca_tags_round;
 885   void   init_dom_lca_tags();
 886 
 887   // Helper for debugging bad dominance relationships
 888   bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
 889 
 890   Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
 891 
 892   // Inline wrapper for frequent cases:
 893   // 1) only one use
 894   // 2) a use is the same as the current LCA passed as 'n1'
 895   Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
 896     assert( n->is_CFG(), "" );
 897     // Fast-path NULL lca
 898     if( lca != NULL && lca != n ) {
 899       assert( lca->is_CFG(), "" );
 900       // find LCA of all uses
 901       n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
 902     }
 903     return find_non_split_ctrl(n);
 904   }
 905   Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
 906 
 907   // Helper function for directing control inputs away from CFG split points.
 908   Node *find_non_split_ctrl( Node *ctrl ) const {
 909     if (ctrl != NULL) {
 910       if (ctrl->is_MultiBranch()) {
 911         ctrl = ctrl->in(0);
 912       }
 913       assert(ctrl->is_CFG(), "CFG");
 914     }
 915     return ctrl;
 916   }
 917 
 918   Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
 919 
 920 #ifdef ASSERT
 921   void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
 922 #endif
 923   void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
 924                                                     uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
 925                                                     Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
 926   void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
 927                                              uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
 928                                              Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
 929   Node* clone_skeleton_predicate_and_initialize(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj, Node* control,
 930                                                 IdealLoopTree* outer_loop, Node* input_proj);
 931   Node* clone_skeleton_predicate_bool(Node* iff, Node* new_init, Node* new_stride, Node* control);
 932   static bool skeleton_predicate_has_opaque(IfNode* iff);
 933   static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride);
 934   static bool subgraph_has_opaque(Node* n);
 935   static void get_skeleton_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false);
 936   void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
 937   void copy_skeleton_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head, Node* init, Node* stride);
 938   void initialize_skeleton_predicates_for_peeled_loop(ProjNode* predicate, LoopNode* outer_loop_head, int dd_outer_loop_head,
 939                                                       Node* init, Node* stride, IdealLoopTree* outer_loop,
 940                                                       const uint idx_before_clone, const Node_List& old_new);
 941   void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol);
 942 #ifdef ASSERT
 943   bool only_has_infinite_loops();
 944 #endif
 945 
 946   void log_loop_tree();
 947 
 948 public:
 949 
 950   PhaseIterGVN &igvn() const { return _igvn; }
 951 
 952   bool has_node( Node* n ) const {
 953     guarantee(n != NULL, "No Node.");
 954     return _nodes[n->_idx] != NULL;
 955   }
 956   // check if transform created new nodes that need _ctrl recorded
 957   Node *get_late_ctrl( Node *n, Node *early );
 958   Node *get_early_ctrl( Node *n );
 959   Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
 960   void set_early_ctrl(Node* n, bool update_body);
 961   void set_subtree_ctrl(Node* n, bool update_body);
 962   void set_ctrl( Node *n, Node *ctrl ) {
 963     assert( !has_node(n) || has_ctrl(n), "" );
 964     assert( ctrl->in(0), "cannot set dead control node" );
 965     assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
 966     _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
 967   }
 968   // Set control and update loop membership
 969   void set_ctrl_and_loop(Node* n, Node* ctrl) {
 970     IdealLoopTree* old_loop = get_loop(get_ctrl(n));
 971     IdealLoopTree* new_loop = get_loop(ctrl);
 972     if (old_loop != new_loop) {
 973       if (old_loop->_child == NULL) old_loop->_body.yank(n);
 974       if (new_loop->_child == NULL) new_loop->_body.push(n);
 975     }
 976     set_ctrl(n, ctrl);
 977   }
 978   // Control nodes can be replaced or subsumed.  During this pass they
 979   // get their replacement Node in slot 1.  Instead of updating the block
 980   // location of all Nodes in the subsumed block, we lazily do it.  As we
 981   // pull such a subsumed block out of the array, we write back the final
 982   // correct block.
 983   Node *get_ctrl( Node *i ) {
 984 
 985     assert(has_node(i), "");
 986     Node *n = get_ctrl_no_update(i);
 987     _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
 988     assert(has_node(i) && has_ctrl(i), "");
 989     assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
 990     return n;
 991   }
 992   // true if CFG node d dominates CFG node n
 993   bool is_dominator(Node *d, Node *n);
 994   // return get_ctrl for a data node and self(n) for a CFG node
 995   Node* ctrl_or_self(Node* n) {
 996     if (has_ctrl(n))
 997       return get_ctrl(n);
 998     else {
 999       assert (n->is_CFG(), "must be a CFG node");
1000       return n;
1001     }
1002   }
1003 
1004   Node *get_ctrl_no_update_helper(Node *i) const {
1005     assert(has_ctrl(i), "should be control, not loop");
1006     return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
1007   }
1008 
1009   Node *get_ctrl_no_update(Node *i) const {
1010     assert( has_ctrl(i), "" );
1011     Node *n = get_ctrl_no_update_helper(i);
1012     if (!n->in(0)) {
1013       // Skip dead CFG nodes
1014       do {
1015         n = get_ctrl_no_update_helper(n);
1016       } while (!n->in(0));
1017       n = find_non_split_ctrl(n);
1018     }
1019     return n;
1020   }
1021 
1022   // Check for loop being set
1023   // "n" must be a control node. Returns true if "n" is known to be in a loop.
1024   bool has_loop( Node *n ) const {
1025     assert(!has_node(n) || !has_ctrl(n), "");
1026     return has_node(n);
1027   }
1028   // Set loop
1029   void set_loop( Node *n, IdealLoopTree *loop ) {
1030     _nodes.map(n->_idx, (Node*)loop);
1031   }
1032   // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
1033   // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
1034   // from old_node to new_node to support the lazy update.  Reference
1035   // replaces loop reference, since that is not needed for dead node.
1036   void lazy_update(Node *old_node, Node *new_node) {
1037     assert(old_node != new_node, "no cycles please");
1038     // Re-use the side array slot for this node to provide the
1039     // forwarding pointer.
1040     _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1041   }
1042   void lazy_replace(Node *old_node, Node *new_node) {
1043     _igvn.replace_node(old_node, new_node);
1044     lazy_update(old_node, new_node);
1045   }
1046 
1047 private:
1048 
1049   // Place 'n' in some loop nest, where 'n' is a CFG node
1050   void build_loop_tree();
1051   int build_loop_tree_impl( Node *n, int pre_order );
1052   // Insert loop into the existing loop tree.  'innermost' is a leaf of the
1053   // loop tree, not the root.
1054   IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1055 
1056   // Place Data nodes in some loop nest
1057   void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1058   void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1059   void build_loop_late_post_work(Node* n, bool pinned);
1060   void build_loop_late_post(Node* n);
1061   void verify_strip_mined_scheduling(Node *n, Node* least);
1062 
1063   // Array of immediate dominance info for each CFG node indexed by node idx
1064 private:
1065   uint _idom_size;
1066   Node **_idom;                  // Array of immediate dominators
1067   uint *_dom_depth;              // Used for fast LCA test
1068   GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1069   LoopOptsMode _mode;
1070 
1071   // build the loop tree and perform any requested optimizations
1072   void build_and_optimize();
1073 
1074   // Dominators for the sea of nodes
1075   void Dominators();
1076 
1077   // Compute the Ideal Node to Loop mapping
1078   PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1079     PhaseTransform(Ideal_Loop),
1080     _igvn(igvn),
1081     _verify_me(nullptr),
1082     _verify_only(false),
1083     _mode(mode),
1084     _nodes_required(UINT_MAX) {
1085     assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1086     build_and_optimize();
1087   }
1088 
1089 #ifndef PRODUCT
1090   // Verify that verify_me made the same decisions as a fresh run
1091   // or only verify that the graph is valid if verify_me is null.
1092   PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1093     PhaseTransform(Ideal_Loop),
1094     _igvn(igvn),
1095     _verify_me(verify_me),
1096     _verify_only(verify_me == nullptr),
1097     _mode(LoopOptsVerify),
1098     _nodes_required(UINT_MAX) {
1099     build_and_optimize();
1100   }
1101 #endif
1102 
1103 public:
1104   Node* idom_no_update(Node* d) const {
1105     return idom_no_update(d->_idx);
1106   }
1107 
1108   Node* idom_no_update(uint didx) const {
1109     assert(didx < _idom_size, "oob");
1110     Node* n = _idom[didx];
1111     assert(n != NULL,"Bad immediate dominator info.");
1112     while (n->in(0) == NULL) { // Skip dead CFG nodes
1113       n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1114       assert(n != NULL,"Bad immediate dominator info.");
1115     }
1116     return n;
1117   }
1118 
1119   Node *idom(Node* d) const {
1120     return idom(d->_idx);
1121   }
1122 
1123   Node *idom(uint didx) const {
1124     Node *n = idom_no_update(didx);
1125     _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1126     return n;
1127   }
1128 
1129   uint dom_depth(Node* d) const {
1130     guarantee(d != NULL, "Null dominator info.");
1131     guarantee(d->_idx < _idom_size, "");
1132     return _dom_depth[d->_idx];
1133   }
1134   void set_idom(Node* d, Node* n, uint dom_depth);
1135   // Locally compute IDOM using dom_lca call
1136   Node *compute_idom( Node *region ) const;
1137   // Recompute dom_depth
1138   void recompute_dom_depth();
1139 
1140   // Is safept not required by an outer loop?
1141   bool is_deleteable_safept(Node* sfpt);
1142 
1143   // Replace parallel induction variable (parallel to trip counter)
1144   void replace_parallel_iv(IdealLoopTree *loop);
1145 
1146   Node *dom_lca( Node *n1, Node *n2 ) const {
1147     return find_non_split_ctrl(dom_lca_internal(n1, n2));
1148   }
1149   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1150 
1151   // Build and verify the loop tree without modifying the graph.  This
1152   // is useful to verify that all inputs properly dominate their uses.
1153   static void verify(PhaseIterGVN& igvn) {
1154 #ifdef ASSERT
1155     ResourceMark rm;
1156     Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
1157     PhaseIdealLoop v(igvn);
1158 #endif
1159   }
1160 
1161   // Recommended way to use PhaseIdealLoop.
1162   // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1163   static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1164     ResourceMark rm;
1165     PhaseIdealLoop v(igvn, mode);
1166 
1167     Compile* C = Compile::current();
1168     if (!C->failing()) {
1169       // Cleanup any modified bits
1170       igvn.optimize();
1171 
1172       v.log_loop_tree();
1173     }
1174   }
1175 
1176   // True if the method has at least 1 irreducible loop
1177   bool _has_irreducible_loops;
1178 
1179   // Per-Node transform
1180   virtual Node* transform(Node* n) { return NULL; }
1181 
1182   Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1183   Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1184   Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1185   Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1186   PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1187 
1188   bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1189 
1190   Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1191   bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1192 #ifdef ASSERT
1193   bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1194 #endif
1195   void add_empty_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1196   SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1197   IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1198   IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1199                                                IdealLoopTree* loop, float cl_prob, float le_fcnt,
1200                                                Node*& entry_control, Node*& iffalse);
1201 
1202   Node* exact_limit( IdealLoopTree *loop );
1203 
1204   // Return a post-walked LoopNode
1205   IdealLoopTree *get_loop( Node *n ) const {
1206     // Dead nodes have no loop, so return the top level loop instead
1207     if (!has_node(n))  return _ltree_root;
1208     assert(!has_ctrl(n), "");
1209     return (IdealLoopTree*)_nodes[n->_idx];
1210   }
1211 
1212   IdealLoopTree* ltree_root() const { return _ltree_root; }
1213 
1214   // Is 'n' a (nested) member of 'loop'?
1215   int is_member( const IdealLoopTree *loop, Node *n ) const {
1216     return loop->is_member(get_loop(n)); }
1217 
1218   // This is the basic building block of the loop optimizations.  It clones an
1219   // entire loop body.  It makes an old_new loop body mapping; with this
1220   // mapping you can find the new-loop equivalent to an old-loop node.  All
1221   // new-loop nodes are exactly equal to their old-loop counterparts, all
1222   // edges are the same.  All exits from the old-loop now have a RegionNode
1223   // that merges the equivalent new-loop path.  This is true even for the
1224   // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
1225   // now come from (one or more) Phis that merge their new-loop equivalents.
1226   // Parameter side_by_side_idom:
1227   //   When side_by_size_idom is NULL, the dominator tree is constructed for
1228   //      the clone loop to dominate the original.  Used in construction of
1229   //      pre-main-post loop sequence.
1230   //   When nonnull, the clone and original are side-by-side, both are
1231   //      dominated by the passed in side_by_side_idom node.  Used in
1232   //      construction of unswitched loops.
1233   enum CloneLoopMode {
1234     IgnoreStripMined = 0,        // Only clone inner strip mined loop
1235     CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1236     ControlAroundStripMined = 2  // Only clone inner strip mined loop,
1237                                  // result control flow branches
1238                                  // either to inner clone or outer
1239                                  // strip mined loop.
1240   };
1241   void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1242                   CloneLoopMode mode, Node* side_by_side_idom = NULL);
1243   void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1244                                    IdealLoopTree* loop, IdealLoopTree* companion_loop,
1245                                    Node_List*& split_if_set, Node_List*& split_bool_set,
1246                                    Node_List*& split_cex_set, Node_List& worklist,
1247                                    uint new_counter, CloneLoopMode mode);
1248   void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1249                         IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1250                         Node_List& extra_data_nodes);
1251 
1252   // If we got the effect of peeling, either by actually peeling or by
1253   // making a pre-loop which must execute at least once, we can remove
1254   // all loop-invariant dominated tests in the main body.
1255   void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1256 
1257   // Generate code to do a loop peel for the given loop (and body).
1258   // old_new is a temp array.
1259   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1260 
1261   // Add pre and post loops around the given loop.  These loops are used
1262   // during RCE, unrolling and aligning loops.
1263   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1264 
1265   // Add post loop after the given loop.
1266   Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1267                          CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1268                          Node*& incr, Node* limit, CountedLoopNode*& post_head);
1269 
1270   // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1271   void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1272 
1273   // Add a vector post loop between a vector main loop and the current post loop
1274   void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1275   // If Node n lives in the back_ctrl block, we clone a private version of n
1276   // in preheader_ctrl block and return that, otherwise return n.
1277   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1278 
1279   // Take steps to maximally unroll the loop.  Peel any odd iterations, then
1280   // unroll to do double iterations.  The next round of major loop transforms
1281   // will repeat till the doubled loop body does all remaining iterations in 1
1282   // pass.
1283   void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1284 
1285   // Unroll the loop body one step - make each trip do 2 iterations.
1286   void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1287 
1288   // Mark vector reduction candidates before loop unrolling
1289   void mark_reductions( IdealLoopTree *loop );
1290 
1291   // Return true if exp is a constant times an induction var
1292   bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1293 
1294   bool is_iv(Node* exp, Node* iv, BasicType bt);
1295 
1296   // Return true if exp is a scaled induction var plus (or minus) constant
1297   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = NULL, int depth = 0);
1298   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1299     jlong long_scale;
1300     if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1301       int int_scale = checked_cast<int>(long_scale);
1302       if (p_scale != NULL) {
1303         *p_scale = int_scale;
1304       }
1305       return true;
1306     }
1307     return false;
1308   }
1309   // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1310   bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1311                                       BasicType bt,
1312                                       jlong* p_scale, Node** p_offset,
1313                                       bool* p_short_scale, int depth);
1314 
1315 
1316   // Enum to determine the action to be performed in create_new_if_for_predicate() when processing phis of UCT regions.
1317   enum class UnswitchingAction {
1318     None,            // No special action.
1319     FastLoopCloning, // Need to clone nodes for the fast loop.
1320     SlowLoopRewiring // Need to rewire nodes for the slow loop.
1321   };
1322 
1323   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1324   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1325                                         int opcode, bool if_cont_is_true_proj = true, Node_List* old_new = NULL,
1326                                         UnswitchingAction unswitching_action = UnswitchingAction::None);
1327 
1328   // Clone data nodes for the fast loop while creating a new If with create_new_if_for_predicate.
1329   Node* clone_data_nodes_for_fast_loop(Node* phi_input, ProjNode* uncommon_proj, Node* if_uct, Node_List* old_new);
1330 
1331   void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1332 
1333   static Node* skip_all_loop_predicates(Node* entry);
1334   static Node* skip_loop_predicates(Node* entry);
1335   static ProjNode* next_predicate(ProjNode* predicate);
1336 
1337   // Find a good location to insert a predicate
1338   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1339 
1340   class Predicates {
1341   public:
1342     // given loop entry, find all predicates above loop
1343     Predicates(Node* entry);
1344 
1345     // Proj of empty loop limit check predicate
1346     ProjNode* loop_limit_check() {
1347       return _loop_limit_check;
1348     }
1349 
1350     // Proj of empty profile predicate
1351     ProjNode* profile_predicate() {
1352       return _profile_predicate;
1353     }
1354 
1355     // Proj of empty predicate
1356     ProjNode* predicate() {
1357       return _predicate;
1358     }
1359 
1360     // First control node above all predicates
1361     Node* skip_all() {
1362       return _entry_to_all_predicates;
1363     }
1364 
1365   private:
1366     ProjNode*_loop_limit_check = nullptr;
1367     ProjNode* _profile_predicate = nullptr;
1368     ProjNode* _predicate = nullptr;
1369     Node* _entry_to_all_predicates = nullptr;
1370   };
1371 
1372   // Find a predicate
1373   static Node* find_predicate(Node* entry);
1374   // Construct a range check for a predicate if
1375   BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1376                          int scale, Node* offset,
1377                          Node* init, Node* limit, jint stride,
1378                          Node* range, bool upper, bool &overflow,
1379                          bool negate);
1380 
1381   // Implementation of the loop predication to promote checks outside the loop
1382   bool loop_predication_impl(IdealLoopTree *loop);
1383   bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1384                                     CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1385                                     Deoptimization::DeoptReason reason);
1386   bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt);
1387   void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1388                                         PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1389                                         Node_List& if_proj_list);
1390   ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1391                                               ProjNode* proj, ProjNode *predicate_proj,
1392                                               ProjNode* upper_bound_proj,
1393                                               int scale, Node* offset,
1394                                               Node* init, Node* limit, jint stride,
1395                                               Node* rng, bool& overflow,
1396                                               Deoptimization::DeoptReason reason);
1397   Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1398                                   Node* predicate_proj, int scale_con, Node* offset,
1399                                   Node* limit, jint stride_con, Node* value);
1400 
1401   // Helper function to collect predicate for eliminating the useless ones
1402   void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1403   void eliminate_useless_predicates();
1404 
1405   // Change the control input of expensive nodes to allow commoning by
1406   // IGVN when it is guaranteed to not result in a more frequent
1407   // execution of the expensive node. Return true if progress.
1408   bool process_expensive_nodes();
1409 
1410   // Check whether node has become unreachable
1411   bool is_node_unreachable(Node *n) const {
1412     return !has_node(n) || n->is_unreachable(_igvn);
1413   }
1414 
1415   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1416   int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1417 
1418   // Check to see if do_range_check(...) cleaned the main loop of range-checks
1419   void has_range_checks(IdealLoopTree *loop);
1420 
1421   // Process post loops which have range checks and try to build a multi-version
1422   // guard to safely determine if we can execute the post loop which was RCE'd.
1423   bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1424 
1425   // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1426   void poison_rce_post_loop(IdealLoopTree *rce_loop);
1427 
1428   // Create a slow version of the loop by cloning the loop
1429   // and inserting an if to select fast-slow versions.
1430   // Return the inserted if.
1431   IfNode* create_slow_version_of_loop(IdealLoopTree *loop,
1432                                       Node_List &old_new,
1433                                       Node_List &unswitch_iffs,
1434                                       CloneLoopMode mode);
1435 
1436   // Clone a loop and return the clone head (clone_loop_head).
1437   // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1438   // This routine was created for usage in CountedLoopReserveKit.
1439   //
1440   //    int(1) -> If -> IfTrue -> original_loop_head
1441   //              |
1442   //              V
1443   //           IfFalse -> clone_loop_head (returned by function pointer)
1444   //
1445   LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1446   // Clone loop with an invariant test (that does not exit) and
1447   // insert a clone of the test that selects which version to
1448   // execute.
1449   void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1450 
1451   // Find candidate "if" for unswitching
1452   IfNode* find_unswitching_candidate(const IdealLoopTree *loop, Node_List& unswitch_iffs) const;
1453 
1454   // Range Check Elimination uses this function!
1455   // Constrain the main loop iterations so the affine function:
1456   //    low_limit <= scale_con * I + offset  <  upper_limit
1457   // always holds true.  That is, either increase the number of iterations in
1458   // the pre-loop or the post-loop until the condition holds true in the main
1459   // loop.  Scale_con, offset and limit are all loop invariant.
1460   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);
1461   // Helper function for add_constraint().
1462   Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1463 
1464   // Partially peel loop up through last_peel node.
1465   bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1466   bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1467 
1468   // Create a scheduled list of nodes control dependent on ctrl set.
1469   void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1470   // Has a use in the vector set
1471   bool has_use_in_set( Node* n, VectorSet& vset );
1472   // Has use internal to the vector set (ie. not in a phi at the loop head)
1473   bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1474   // clone "n" for uses that are outside of loop
1475   int  clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1476   // clone "n" for special uses that are in the not_peeled region
1477   void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1478                                           VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1479   // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1480   void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1481 #ifdef ASSERT
1482   // Validate the loop partition sets: peel and not_peel
1483   bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1484   // Ensure that uses outside of loop are of the right form
1485   bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1486                                  uint orig_exit_idx, uint clone_exit_idx);
1487   bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1488 #endif
1489 
1490   // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1491   int stride_of_possible_iv( Node* iff );
1492   bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1493   // Return the (unique) control output node that's in the loop (if it exists.)
1494   Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1495   // Insert a signed compare loop exit cloned from an unsigned compare.
1496   IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1497   void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1498   // Utility to register node "n" with PhaseIdealLoop
1499   void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1500   // Utility to create an if-projection
1501   ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1502   // Force the iff control output to be the live_proj
1503   Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1504   // Insert a region before an if projection
1505   RegionNode* insert_region_before_proj(ProjNode* proj);
1506   // Insert a new if before an if projection
1507   ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1508 
1509   // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1510   // "Nearly" because all Nodes have been cloned from the original in the loop,
1511   // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1512   // through the Phi recursively, and return a Bool.
1513   Node* clone_iff(PhiNode* phi);
1514   CmpNode* clone_bool(PhiNode* phi);
1515 
1516 
1517   // Rework addressing expressions to get the most loop-invariant stuff
1518   // moved out.  We'd like to do all associative operators, but it's especially
1519   // important (common) to do address expressions.
1520   Node* remix_address_expressions(Node* n);
1521   Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1522 
1523   // Convert add to muladd to generate MuladdS2I under certain criteria
1524   Node * convert_add_to_muladd(Node * n);
1525 
1526   // Attempt to use a conditional move instead of a phi/branch
1527   Node *conditional_move( Node *n );
1528 
1529   // Reorganize offset computations to lower register pressure.
1530   // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1531   // (which are then alive with the post-incremented trip counter
1532   // forcing an extra register move)
1533   void reorg_offsets( IdealLoopTree *loop );
1534 
1535   // Check for aggressive application of 'split-if' optimization,
1536   // using basic block level info.
1537   void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack);
1538   Node *split_if_with_blocks_pre ( Node *n );
1539   void  split_if_with_blocks_post( Node *n );
1540   Node *has_local_phi_input( Node *n );
1541   // Mark an IfNode as being dominated by a prior test,
1542   // without actually altering the CFG (and hence IDOM info).
1543   void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool exclude_loop_predicate = false);
1544 
1545   // Split Node 'n' through merge point
1546   RegionNode* split_thru_region(Node* n, RegionNode* region);
1547   // Split Node 'n' through merge point if there is enough win.
1548   Node *split_thru_phi( Node *n, Node *region, int policy );
1549   // Found an If getting its condition-code input from a Phi in the
1550   // same block.  Split thru the Region.
1551   void do_split_if(Node *iff, RegionNode** new_false_region = NULL, RegionNode** new_true_region = NULL);
1552 
1553   // Conversion of fill/copy patterns into intrinsic versions
1554   bool do_intrinsify_fill();
1555   bool intrinsify_fill(IdealLoopTree* lpt);
1556   bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1557                        Node*& shift, Node*& offset);
1558 
1559 private:
1560   // Return a type based on condition control flow
1561   const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1562   const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1563  // Helpers for filtered type
1564   const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1565 
1566   // Helper functions
1567   Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1568   Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1569   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 );
1570   bool split_up( Node *n, Node *blk1, Node *blk2 );
1571   void sink_use( Node *use, Node *post_loop );
1572   Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1573   Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1574   void try_move_store_after_loop(Node* n);
1575   void move_flat_array_check_out_of_loop(Node* n);
1576   bool identical_backtoback_ifs(Node *n);
1577   bool flatten_array_element_type_check(Node *n);
1578   bool can_split_if(Node *n_ctrl);
1579 
1580   // Determine if a method is too big for a/another round of split-if, based on
1581   // a magic (approximate) ratio derived from the equally magic constant 35000,
1582   // previously used for this purpose (but without relating to the node limit).
1583   bool must_throttle_split_if() {
1584     uint threshold = C->max_node_limit() * 2 / 5;
1585     return C->live_nodes() > threshold;
1586   }
1587 
1588   // A simplistic node request tracking mechanism, where
1589   //   = UINT_MAX   Request not valid or made final.
1590   //   < UINT_MAX   Nodes currently requested (estimate).
1591   uint _nodes_required;
1592 
1593   enum { REQUIRE_MIN = 70 };
1594 
1595   uint nodes_required() const { return _nodes_required; }
1596 
1597   // Given the _currently_  available number of nodes, check  whether there is
1598   // "room" for an additional request or not, considering the already required
1599   // number of  nodes.  Return TRUE if  the new request is  exceeding the node
1600   // budget limit, otherwise return FALSE.  Note that this interpretation will
1601   // act pessimistic on  additional requests when new nodes  have already been
1602   // generated since the 'begin'.  This behaviour fits with the intention that
1603   // node estimates/requests should be made upfront.
1604   bool exceeding_node_budget(uint required = 0) {
1605     assert(C->live_nodes() < C->max_node_limit(), "sanity");
1606     uint available = C->max_node_limit() - C->live_nodes();
1607     return available < required + _nodes_required + REQUIRE_MIN;
1608   }
1609 
1610   uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1611     precond(require > 0);
1612     _nodes_required += MAX2(require, minreq);
1613     return _nodes_required;
1614   }
1615 
1616   bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1617     return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1618   }
1619 
1620   uint require_nodes_begin() {
1621     assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1622     _nodes_required = 0;
1623     return C->live_nodes();
1624   }
1625 
1626   // When a node request is final,  optionally check that the requested number
1627   // of nodes was  reasonably correct with respect to the  number of new nodes
1628   // introduced since the last 'begin'. Always check that we have not exceeded
1629   // the maximum node limit.
1630   void require_nodes_final(uint live_at_begin, bool check_estimate) {
1631     assert(_nodes_required < UINT_MAX, "Bad state (final).");
1632 
1633 #ifdef ASSERT
1634     if (check_estimate) {
1635       // Check that the node budget request was not off by too much (x2).
1636       // Should this be the case we _surely_ need to improve the estimates
1637       // used in our budget calculations.
1638       if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1639         log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1640                               C->live_nodes() - live_at_begin, _nodes_required);
1641       }
1642     }
1643 #endif
1644     // Assert that we have stayed within the node budget limit.
1645     assert(C->live_nodes() < C->max_node_limit(),
1646            "Exceeding node budget limit: %d + %d > %d (request = %d)",
1647            C->live_nodes() - live_at_begin, live_at_begin,
1648            C->max_node_limit(), _nodes_required);
1649 
1650     _nodes_required = UINT_MAX;
1651   }
1652 
1653   // Clone loop predicates to slow and fast loop when unswitching a loop
1654   void clone_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new, ProjNode*& iffast_pred, ProjNode*& ifslow_pred);
1655   ProjNode* clone_predicate_to_unswitched_loop(ProjNode* predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1656                                                Node_List* old_new = NULL);
1657   void clone_skeleton_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new, Deoptimization::DeoptReason reason,
1658                                                     ProjNode* old_predicate_proj, ProjNode* iffast_pred, ProjNode* ifslow_pred);
1659   ProjNode* clone_skeleton_predicate_for_unswitched_loops(Node* iff, ProjNode* predicate,
1660                                                           Deoptimization::DeoptReason reason,
1661                                                           ProjNode* output_proj);
1662   static void check_created_predicate_for_unswitching(const Node* new_entry) PRODUCT_RETURN;
1663 
1664   bool _created_loop_node;
1665 #ifdef ASSERT
1666   void dump_real_LCA(Node* early, Node* wrong_lca);
1667   bool check_idom_chains_intersection(const Node* n, uint& idom_idx_new, uint& idom_idx_other, const Node_List* nodes_seen) const;
1668 #endif
1669 
1670 public:
1671   void set_created_loop_node() { _created_loop_node = true; }
1672   bool created_loop_node()     { return _created_loop_node; }
1673   void register_new_node(Node* n, Node* blk);
1674 
1675 #ifdef ASSERT
1676   void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1677 #endif
1678 
1679 #ifndef PRODUCT
1680   void dump() const;
1681   void dump_idom(Node* n) const;
1682   void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1683   void verify() const;          // Major slow  :-)
1684   void verify_compare(Node* n, const PhaseIdealLoop* loop_verify, VectorSet &visited) const;
1685   IdealLoopTree* get_loop_idx(Node* n) const {
1686     // Dead nodes have no loop, so return the top level loop instead
1687     return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1688   }
1689   // Print some stats
1690   static void print_statistics();
1691   static int _loop_invokes;     // Count of PhaseIdealLoop invokes
1692   static int _loop_work;        // Sum of PhaseIdealLoop x _unique
1693   static volatile int _long_loop_candidates;
1694   static volatile int _long_loop_nests;
1695   static volatile int _long_loop_counted_loops;
1696 #endif
1697 
1698   void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1699 
1700   void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1701 
1702   LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1703 
1704 
1705   int extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi,
1706                                       Node_List &range_checks);
1707 
1708   void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1709                                    Node* inner_iters_actual_int, Node* inner_phi,
1710                                    Node* iv_add, LoopNode* inner_head);
1711 
1712   Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1713 
1714   bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1715 
1716   bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1717 
1718   Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1719 
1720   void try_sink_out_of_loop(Node* n);
1721 
1722   Node* clamp(Node* R, Node* L, Node* H);
1723 
1724   bool safe_for_if_replacement(const Node* dom) const;
1725 
1726   void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1727 
1728   bool try_merge_identical_ifs(Node* n);
1729 
1730   void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1731 
1732   void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1733                       IdealLoopTree* parent, bool partial);
1734 
1735   void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1736                 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1737 
1738   void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1739                      uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1740                      Node_List*& split_bool_set, Node_List*& split_cex_set);
1741 
1742   void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1743 };
1744 
1745 
1746 class AutoNodeBudget : public StackObj
1747 {
1748 public:
1749   enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1750 
1751   AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1752     : _phase(phase),
1753       _check_at_final(chk == BUDGET_CHECK),
1754       _nodes_at_begin(0)
1755   {
1756     precond(_phase != NULL);
1757 
1758     _nodes_at_begin = _phase->require_nodes_begin();
1759   }
1760 
1761   ~AutoNodeBudget() {
1762 #ifndef PRODUCT
1763     if (TraceLoopOpts) {
1764       uint request = _phase->nodes_required();
1765       uint delta   = _phase->C->live_nodes() - _nodes_at_begin;
1766 
1767       if (request < delta) {
1768         tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1769       } else {
1770         uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1771         // Identify the worst estimates as "poor" ones.
1772         if (request > REQUIRE_MIN && delta > 0) {
1773           if ((delta >  REQUIRE_MIN && request >  3 * delta) ||
1774               (delta <= REQUIRE_MIN && request > 10 * delta)) {
1775             tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1776           }
1777         }
1778       }
1779     }
1780 #endif // PRODUCT
1781     _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1782   }
1783 
1784 private:
1785   PhaseIdealLoop* _phase;
1786   bool _check_at_final;
1787   uint _nodes_at_begin;
1788 };
1789 
1790 
1791 // This kit may be used for making of a reserved copy of a loop before this loop
1792 //  goes under non-reversible changes.
1793 //
1794 // Function create_reserve() creates a reserved copy (clone) of the loop.
1795 // The reserved copy is created by calling
1796 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1797 // the original and reserved loops are connected in the outer graph.
1798 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1799 //
1800 // By default the reserved copy (clone) of the loop is created as dead code - it is
1801 // dominated in the outer loop by this node chain:
1802 //   intcon(1)->If->IfFalse->reserved_copy.
1803 // The original loop is dominated by the same node chain but IfTrue projection:
1804 //   intcon(0)->If->IfTrue->original_loop.
1805 //
1806 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1807 // and the dtor, checks _use_new value.
1808 // If _use_new == false, it "switches" control to reserved copy of the loop
1809 // by simple replacing of node intcon(1) with node intcon(0).
1810 //
1811 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1812 //
1813 // void CountedLoopReserveKit_example()
1814 // {
1815 //    CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1816 //    if (DoReserveCopy && !lrk.has_reserved()) {
1817 //      return; //failed to create reserved loop copy
1818 //    }
1819 //    ...
1820 //    //something is wrong, switch to original loop
1821 ///   if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1822 //    ...
1823 //    //everything worked ok, return with the newly modified loop
1824 //    lrk.use_new();
1825 //    return; // ~CountedLoopReserveKit does nothing once use_new() was called
1826 //  }
1827 //
1828 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1829 // the dtor will "switch to the original" loop.
1830 // NOTE. You you modify outside of the original loop this class is no help.
1831 //
1832 class CountedLoopReserveKit {
1833   private:
1834     PhaseIdealLoop* _phase;
1835     IdealLoopTree*  _lpt;
1836     LoopNode*       _lp;
1837     IfNode*         _iff;
1838     LoopNode*       _lp_reserved;
1839     bool            _has_reserved;
1840     bool            _use_new;
1841     const bool      _active; //may be set to false in ctor, then the object is dummy
1842 
1843   public:
1844     CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1845     ~CountedLoopReserveKit();
1846     void use_new()                {_use_new = true;}
1847     void set_iff(IfNode* x)       {_iff = x;}
1848     bool has_reserved()     const { return _active && _has_reserved;}
1849   private:
1850     bool create_reserve();
1851 };// class CountedLoopReserveKit
1852 
1853 inline Node* IdealLoopTree::tail() {
1854   // Handle lazy update of _tail field.
1855   if (_tail->in(0) == NULL) {
1856     _tail = _phase->get_ctrl(_tail);
1857   }
1858   return _tail;
1859 }
1860 
1861 inline Node* IdealLoopTree::head() {
1862   // Handle lazy update of _head field.
1863   if (_head->in(0) == NULL) {
1864     _head = _phase->get_ctrl(_head);
1865   }
1866   return _head;
1867 }
1868 
1869 // Iterate over the loop tree using a preorder, left-to-right traversal.
1870 //
1871 // Example that visits all counted loops from within PhaseIdealLoop
1872 //
1873 //  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1874 //   IdealLoopTree* lpt = iter.current();
1875 //   if (!lpt->is_counted()) continue;
1876 //   ...
1877 class LoopTreeIterator : public StackObj {
1878 private:
1879   IdealLoopTree* _root;
1880   IdealLoopTree* _curnt;
1881 
1882 public:
1883   LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1884 
1885   bool done() { return _curnt == NULL; }       // Finished iterating?
1886 
1887   void next();                                 // Advance to next loop tree
1888 
1889   IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
1890 };
1891 
1892 // Compute probability of reaching some CFG node from a fixed
1893 // dominating CFG node
1894 class PathFrequency {
1895 private:
1896   Node* _dom; // frequencies are computed relative to this node
1897   Node_Stack _stack;
1898   GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1899   GrowableArray<float> _freqs; // cache frequencies
1900   PhaseIdealLoop* _phase;
1901 
1902   float check_and_truncate_frequency(float f) {
1903     assert(f >= 0, "Incorrect frequency");
1904     // We do not perform an exact (f <= 1) check
1905     // this would be error prone with rounding of floats.
1906     // Performing a check like (f <= 1+eps) would be of benefit,
1907     // however, it is not evident how to determine such an eps,
1908     // given that an arbitrary number of add/mul operations
1909     // are performed on these frequencies.
1910     return (f > 1) ? 1 : f;
1911   }
1912 
1913 public:
1914   PathFrequency(Node* dom, PhaseIdealLoop* phase)
1915     : _dom(dom), _stack(0), _phase(phase) {
1916   }
1917 
1918   float to(Node* n);
1919 };
1920 
1921 #endif // SHARE_OPTO_LOOPNODE_HPP