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