1 /*
   2  * Copyright (c) 1998, 2015, 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_VM_OPTO_LOOPNODE_HPP
  26 #define SHARE_VM_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 PhaseIdealLoop;
  41 class CountedLoopReserveKit;
  42 class VectorSet;
  43 class Invariance;
  44 struct small_cache;
  45 
  46 //
  47 //                  I D E A L I Z E D   L O O P S
  48 //
  49 // Idealized loops are the set of loops I perform more interesting
  50 // transformations on, beyond simple hoisting.
  51 
  52 //------------------------------LoopNode---------------------------------------
  53 // Simple loop header.  Fall in path on left, loop-back path on right.
  54 class LoopNode : public RegionNode {
  55   // Size is bigger to hold the flags.  However, the flags do not change
  56   // the semantics so it does not appear in the hash & cmp functions.
  57   virtual uint size_of() const { return sizeof(*this); }
  58 protected:
  59   short _loop_flags;
  60   // Names for flag bitfields
  61   enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
  62          MainHasNoPreLoop=4,
  63          HasExactTripCount=8,
  64          InnerLoop=16,
  65          PartialPeelLoop=32,
  66          PartialPeelFailed=64,
  67          HasReductions=128,
  68          WasSlpAnalyzed=256,
  69          PassedSlpAnalysis=512,
  70          DoUnrollOnly=1024,
  71          VectorizedLoop=2048,
  72          HasAtomicPostLoop=4096,
  73          HasRangeChecks=8192,
  74          IsMultiversioned=16384};
  75   char _unswitch_count;
  76   enum { _unswitch_max=3 };
  77   char _postloop_flags;
  78   enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
  79 
  80 public:
  81   // Names for edge indices
  82   enum { Self=0, EntryControl, LoopBackControl };
  83 
  84   int is_inner_loop() const { return _loop_flags & InnerLoop; }
  85   void set_inner_loop() { _loop_flags |= InnerLoop; }
  86 
  87   int range_checks_present() const { return _loop_flags & HasRangeChecks; }
  88   int is_multiversioned() const { return _loop_flags & IsMultiversioned; }
  89   int is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
  90   int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
  91   void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
  92   int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
  93 
  94   void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
  95   void mark_has_reductions() { _loop_flags |= HasReductions; }
  96   void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
  97   void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
  98   void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
  99   void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
 100   void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
 101   void mark_has_range_checks() { _loop_flags |=  HasRangeChecks; }
 102   void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
 103 
 104   int unswitch_max() { return _unswitch_max; }
 105   int unswitch_count() { return _unswitch_count; }
 106 
 107   int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
 108   void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
 109   int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
 110   void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
 111 
 112   void set_unswitch_count(int val) {
 113     assert (val <= unswitch_max(), "too many unswitches");
 114     _unswitch_count = val;
 115   }
 116 
 117   LoopNode(Node *entry, Node *backedge) : RegionNode(3), _loop_flags(0), _unswitch_count(0), _postloop_flags(0) {
 118     init_class_id(Class_Loop);
 119     init_req(EntryControl, entry);
 120     init_req(LoopBackControl, backedge);
 121   }
 122 
 123   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 124   virtual uint Opcode() const;
 125   bool can_be_counted_loop(PhaseTransform* phase) const {
 126     return req() == 3 && in(0) != NULL &&
 127       in(1) != NULL && phase->type(in(1)) != Type::TOP &&
 128       in(2) != NULL && phase->type(in(2)) != Type::TOP;
 129   }
 130   bool is_valid_counted_loop() const;
 131 #ifndef PRODUCT
 132   virtual void dump_spec(outputStream *st) const;
 133 #endif
 134 };
 135 
 136 //------------------------------Counted Loops----------------------------------
 137 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
 138 // path (and maybe some other exit paths).  The trip-counter exit is always
 139 // last in the loop.  The trip-counter have to stride by a constant;
 140 // the exit value is also loop invariant.
 141 
 142 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
 143 // CountedLoopNode has the incoming loop control and the loop-back-control
 144 // which is always the IfTrue before the matching CountedLoopEndNode.  The
 145 // CountedLoopEndNode has an incoming control (possibly not the
 146 // CountedLoopNode if there is control flow in the loop), the post-increment
 147 // trip-counter value, and the limit.  The trip-counter value is always of
 148 // the form (Op old-trip-counter stride).  The old-trip-counter is produced
 149 // by a Phi connected to the CountedLoopNode.  The stride is constant.
 150 // The Op is any commutable opcode, including Add, Mul, Xor.  The
 151 // CountedLoopEndNode also takes in the loop-invariant limit value.
 152 
 153 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
 154 // loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
 155 // via the old-trip-counter from the Op node.
 156 
 157 //------------------------------CountedLoopNode--------------------------------
 158 // CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
 159 // inputs the incoming loop-start control and the loop-back control, so they
 160 // act like RegionNodes.  They also take in the initial trip counter, the
 161 // loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
 162 // produce a loop-body control and the trip counter value.  Since
 163 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
 164 
 165 class CountedLoopNode : public LoopNode {
 166   // Size is bigger to hold _main_idx.  However, _main_idx does not change
 167   // the semantics so it does not appear in the hash & cmp functions.
 168   virtual uint size_of() const { return sizeof(*this); }
 169 
 170   // For Pre- and Post-loops during debugging ONLY, this holds the index of
 171   // the Main CountedLoop.  Used to assert that we understand the graph shape.
 172   node_idx_t _main_idx;
 173 
 174   // Known trip count calculated by compute_exact_trip_count()
 175   uint  _trip_count;
 176 
 177   // Expected trip count from profile data
 178   float _profile_trip_cnt;
 179 
 180   // Log2 of original loop bodies in unrolled loop
 181   int _unrolled_count_log2;
 182 
 183   // Node count prior to last unrolling - used to decide if
 184   // unroll,optimize,unroll,optimize,... is making progress
 185   int _node_count_before_unroll;
 186 
 187   // If slp analysis is performed we record the maximum
 188   // vector mapped unroll factor here
 189   int _slp_maximum_unroll_factor;
 190 
 191 public:
 192   CountedLoopNode( Node *entry, Node *backedge )
 193     : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
 194       _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
 195       _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
 196     init_class_id(Class_CountedLoop);
 197     // Initialize _trip_count to the largest possible value.
 198     // Will be reset (lower) if the loop's trip count is known.
 199   }
 200 
 201   virtual uint Opcode() const;
 202   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 203 
 204   Node *init_control() const { return in(EntryControl); }
 205   Node *back_control() const { return in(LoopBackControl); }
 206   CountedLoopEndNode *loopexit() const;
 207   Node *init_trip() const;
 208   Node *stride() const;
 209   int   stride_con() const;
 210   bool  stride_is_con() const;
 211   Node *limit() const;
 212   Node *incr() const;
 213   Node *phi() const;
 214 
 215   // Match increment with optional truncation
 216   static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
 217 
 218   // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
 219   // can run short a few iterations and may start a few iterations in.
 220   // It will be RCE'd and unrolled and aligned.
 221 
 222   // A following 'post' loop will run any remaining iterations.  Used
 223   // during Range Check Elimination, the 'post' loop will do any final
 224   // iterations with full checks.  Also used by Loop Unrolling, where
 225   // the 'post' loop will do any epilog iterations needed.  Basically,
 226   // a 'post' loop can not profitably be further unrolled or RCE'd.
 227 
 228   // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
 229   // it may do under-flow checks for RCE and may do alignment iterations
 230   // so the following main loop 'knows' that it is striding down cache
 231   // lines.
 232 
 233   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
 234   // Aligned, may be missing it's pre-loop.
 235   int is_normal_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
 236   int is_pre_loop      () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
 237   int is_main_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
 238   int is_post_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
 239   int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
 240   int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
 241   int has_passed_slp   () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
 242   int do_unroll_only      () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
 243   int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
 244   int has_atomic_post_loop  () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
 245   void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 246 
 247   int main_idx() const { return _main_idx; }
 248 
 249 
 250   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
 251   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
 252   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
 253   void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 254 
 255   void set_trip_count(uint tc) { _trip_count = tc; }
 256   uint trip_count()            { return _trip_count; }
 257 
 258   bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
 259   void set_exact_trip_count(uint tc) {
 260     _trip_count = tc;
 261     _loop_flags |= HasExactTripCount;
 262   }
 263   void set_nonexact_trip_count() {
 264     _loop_flags &= ~HasExactTripCount;
 265   }
 266   void set_notpassed_slp() {
 267     _loop_flags &= ~PassedSlpAnalysis;
 268   }
 269 
 270   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
 271   float profile_trip_cnt()             { return _profile_trip_cnt; }
 272 
 273   void double_unrolled_count() { _unrolled_count_log2++; }
 274   int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
 275 
 276   void set_node_count_before_unroll(int ct)  { _node_count_before_unroll = ct; }
 277   int  node_count_before_unroll()            { return _node_count_before_unroll; }
 278   void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
 279   int  slp_max_unroll() const                { return _slp_maximum_unroll_factor; }
 280 
 281 #ifndef PRODUCT
 282   virtual void dump_spec(outputStream *st) const;
 283 #endif
 284 };
 285 
 286 //------------------------------CountedLoopEndNode-----------------------------
 287 // CountedLoopEndNodes end simple trip counted loops.  They act much like
 288 // IfNodes.
 289 class CountedLoopEndNode : public IfNode {
 290 public:
 291   enum { TestControl, TestValue };
 292 
 293   CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
 294     : IfNode( control, test, prob, cnt) {
 295     init_class_id(Class_CountedLoopEnd);
 296   }
 297   virtual uint Opcode() const;
 298 
 299   Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
 300   Node *incr() const                { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 301   Node *limit() const               { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 302   Node *stride() const              { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 303   Node *init_trip() const           { Node *tmp = phi     (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 304   int stride_con() const;
 305   bool stride_is_con() const        { Node *tmp = stride  (); return (tmp != NULL && tmp->is_Con()); }
 306   BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
 307   PhiNode *phi() const {
 308     Node *tmp = incr();
 309     if (tmp && tmp->req() == 3) {
 310       Node* phi = tmp->in(1);
 311       if (phi->is_Phi()) {
 312         return phi->as_Phi();
 313       }
 314     }
 315     return NULL;
 316   }
 317   CountedLoopNode *loopnode() const {
 318     // The CountedLoopNode that goes with this CountedLoopEndNode may
 319     // have been optimized out by the IGVN so be cautious with the
 320     // pattern matching on the graph
 321     PhiNode* iv_phi = phi();
 322     if (iv_phi == NULL) {
 323       return NULL;
 324     }
 325     Node *ln = iv_phi->in(0);
 326     if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
 327       return (CountedLoopNode*)ln;
 328     }
 329     return NULL;
 330   }
 331 
 332 #ifndef PRODUCT
 333   virtual void dump_spec(outputStream *st) const;
 334 #endif
 335 };
 336 
 337 
 338 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
 339   Node *bc = back_control();
 340   if( bc == NULL ) return NULL;
 341   Node *le = bc->in(0);
 342   if( le->Opcode() != Op_CountedLoopEnd )
 343     return NULL;
 344   return (CountedLoopEndNode*)le;
 345 }
 346 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
 347 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
 348 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
 349 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
 350 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
 351 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
 352 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
 353 
 354 //------------------------------LoopLimitNode-----------------------------
 355 // Counted Loop limit node which represents exact final iterator value:
 356 // trip_count = (limit - init_trip + stride - 1)/stride
 357 // final_value= trip_count * stride + init_trip.
 358 // Use HW instructions to calculate it when it can overflow in integer.
 359 // Note, final_value should fit into integer since counted loop has
 360 // limit check: limit <= max_int-stride.
 361 class LoopLimitNode : public Node {
 362   enum { Init=1, Limit=2, Stride=3 };
 363  public:
 364   LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
 365     // Put it on the Macro nodes list to optimize during macro nodes expansion.
 366     init_flags(Flag_is_macro);
 367     C->add_macro_node(this);
 368   }
 369   virtual uint Opcode() const;
 370   virtual const Type *bottom_type() const { return TypeInt::INT; }
 371   virtual uint ideal_reg() const { return Op_RegI; }
 372   virtual const Type* Value(PhaseGVN* phase) const;
 373   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 374   virtual Node* Identity(PhaseGVN* phase);
 375 };
 376 
 377 // -----------------------------IdealLoopTree----------------------------------
 378 class IdealLoopTree : public ResourceObj {
 379 public:
 380   IdealLoopTree *_parent;       // Parent in loop tree
 381   IdealLoopTree *_next;         // Next sibling in loop tree
 382   IdealLoopTree *_child;        // First child in loop tree
 383 
 384   // The head-tail backedge defines the loop.
 385   // If tail is NULL then this loop has multiple backedges as part of the
 386   // same loop.  During cleanup I'll peel off the multiple backedges; merge
 387   // them at the loop bottom and flow 1 real backedge into the loop.
 388   Node *_head;                  // Head of loop
 389   Node *_tail;                  // Tail of loop
 390   inline Node *tail();          // Handle lazy update of _tail field
 391   PhaseIdealLoop* _phase;
 392   int _local_loop_unroll_limit;
 393   int _local_loop_unroll_factor;
 394 
 395   Node_List _body;              // Loop body for inner loops
 396 
 397   uint8_t _nest;                // Nesting depth
 398   uint8_t _irreducible:1,       // True if irreducible
 399           _has_call:1,          // True if has call safepoint
 400           _has_sfpt:1,          // True if has non-call safepoint
 401           _rce_candidate:1;     // True if candidate for range check elimination
 402 
 403   Node_List* _safepts;          // List of safepoints in this loop
 404   Node_List* _required_safept;  // A inner loop cannot delete these safepts;
 405   bool  _allow_optimizations;   // Allow loop optimizations
 406 
 407   IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
 408     : _parent(0), _next(0), _child(0),
 409       _head(head), _tail(tail),
 410       _phase(phase),
 411       _safepts(NULL),
 412       _required_safept(NULL),
 413       _allow_optimizations(true),
 414       _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
 415       _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
 416   { }
 417 
 418   // Is 'l' a member of 'this'?
 419   bool is_member(const IdealLoopTree *l) const; // Test for nested membership
 420 
 421   // Set loop nesting depth.  Accumulate has_call bits.
 422   int set_nest( uint depth );
 423 
 424   // Split out multiple fall-in edges from the loop header.  Move them to a
 425   // private RegionNode before the loop.  This becomes the loop landing pad.
 426   void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
 427 
 428   // Split out the outermost loop from this shared header.
 429   void split_outer_loop( PhaseIdealLoop *phase );
 430 
 431   // Merge all the backedges from the shared header into a private Region.
 432   // Feed that region as the one backedge to this loop.
 433   void merge_many_backedges( PhaseIdealLoop *phase );
 434 
 435   // Split shared headers and insert loop landing pads.
 436   // Insert a LoopNode to replace the RegionNode.
 437   // Returns TRUE if loop tree is structurally changed.
 438   bool beautify_loops( PhaseIdealLoop *phase );
 439 
 440   // Perform optimization to use the loop predicates for null checks and range checks.
 441   // Applies to any loop level (not just the innermost one)
 442   bool loop_predication( PhaseIdealLoop *phase);
 443 
 444   // Perform iteration-splitting on inner loops.  Split iterations to
 445   // avoid range checks or one-shot null checks.  Returns false if the
 446   // current round of loop opts should stop.
 447   bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
 448 
 449   // Driver for various flavors of iteration splitting.  Returns false
 450   // if the current round of loop opts should stop.
 451   bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
 452 
 453   // Given dominators, try to find loops with calls that must always be
 454   // executed (call dominates loop tail).  These loops do not need non-call
 455   // safepoints (ncsfpt).
 456   void check_safepts(VectorSet &visited, Node_List &stack);
 457 
 458   // Allpaths backwards scan from loop tail, terminating each path at first safepoint
 459   // encountered.
 460   void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
 461 
 462   // Remove safepoints from loop. Optionally keeping one.
 463   void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
 464 
 465   // Convert to counted loops where possible
 466   void counted_loop( PhaseIdealLoop *phase );
 467 
 468   // Check for Node being a loop-breaking test
 469   Node *is_loop_exit(Node *iff) const;
 470 
 471   // Remove simplistic dead code from loop body
 472   void DCE_loop_body();
 473 
 474   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
 475   // Replace with a 1-in-10 exit guess.
 476   void adjust_loop_exit_prob( PhaseIdealLoop *phase );
 477 
 478   // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
 479   // Useful for unrolling loops with NO array accesses.
 480   bool policy_peel_only( PhaseIdealLoop *phase ) const;
 481 
 482   // Return TRUE or FALSE if the loop should be unswitched -- clone
 483   // loop with an invariant test
 484   bool policy_unswitching( PhaseIdealLoop *phase ) const;
 485 
 486   // Micro-benchmark spamming.  Remove empty loops.
 487   bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
 488 
 489   // Convert one iteration loop into normal code.
 490   bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
 491 
 492   // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
 493   // make some loop-invariant test (usually a null-check) happen before the
 494   // loop.
 495   bool policy_peeling( PhaseIdealLoop *phase ) const;
 496 
 497   // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
 498   // known trip count in the counted loop node.
 499   bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
 500 
 501   // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
 502   // the loop is a CountedLoop and the body is small enough.
 503   bool policy_unroll(PhaseIdealLoop *phase);
 504 
 505   // Loop analyses to map to a maximal superword unrolling for vectorization.
 506   void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
 507 
 508   // Return TRUE or FALSE if the loop should be range-check-eliminated.
 509   // Gather a list of IF tests that are dominated by iteration splitting;
 510   // also gather the end of the first split and the start of the 2nd split.
 511   bool policy_range_check( PhaseIdealLoop *phase ) const;
 512 
 513   // Return TRUE or FALSE if the loop should be cache-line aligned.
 514   // Gather the expression that does the alignment.  Note that only
 515   // one array base can be aligned in a loop (unless the VM guarantees
 516   // mutual alignment).  Note that if we vectorize short memory ops
 517   // into longer memory ops, we may want to increase alignment.
 518   bool policy_align( PhaseIdealLoop *phase ) const;
 519 
 520   // Return TRUE if "iff" is a range check.
 521   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
 522 
 523   // Compute loop exact trip count if possible
 524   void compute_exact_trip_count( PhaseIdealLoop *phase );
 525 
 526   // Compute loop trip count from profile data
 527   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 528 
 529   // Reassociate invariant expressions.
 530   void reassociate_invariants(PhaseIdealLoop *phase);
 531   // Reassociate invariant add and subtract expressions.
 532   Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
 533   // Return nonzero index of invariant operand if invariant and variant
 534   // are combined with an Add or Sub. Helper for reassociate_invariants.
 535   int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
 536 
 537   // Return true if n is invariant
 538   bool is_invariant(Node* n) const;
 539 
 540   // Put loop body on igvn work list
 541   void record_for_igvn();
 542 
 543   bool is_loop()    { return !_irreducible && _tail && !_tail->is_top(); }
 544   bool is_inner()   { return is_loop() && _child == NULL; }
 545   bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
 546 
 547   void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
 548 
 549 #ifndef PRODUCT
 550   void dump_head( ) const;      // Dump loop head only
 551   void dump() const;            // Dump this loop recursively
 552   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 553 #endif
 554 
 555 };
 556 
 557 // -----------------------------PhaseIdealLoop---------------------------------
 558 // Computes the mapping from Nodes to IdealLoopTrees.  Organizes IdealLoopTrees into a
 559 // loop tree.  Drives the loop-based transformations on the ideal graph.
 560 class PhaseIdealLoop : public PhaseTransform {
 561   friend class IdealLoopTree;
 562   friend class SuperWord;
 563   friend class CountedLoopReserveKit;
 564 
 565   // Pre-computed def-use info
 566   PhaseIterGVN &_igvn;
 567 
 568   // Head of loop tree
 569   IdealLoopTree *_ltree_root;
 570 
 571   // Array of pre-order numbers, plus post-visited bit.
 572   // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
 573   // ODD for post-visited.  Other bits are the pre-order number.
 574   uint *_preorders;
 575   uint _max_preorder;
 576 
 577   const PhaseIdealLoop* _verify_me;
 578   bool _verify_only;
 579 
 580   // Allocate _preorders[] array
 581   void allocate_preorders() {
 582     _max_preorder = C->unique()+8;
 583     _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
 584     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 585   }
 586 
 587   // Allocate _preorders[] array
 588   void reallocate_preorders() {
 589     if ( _max_preorder < C->unique() ) {
 590       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
 591       _max_preorder = C->unique();
 592     }
 593     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 594   }
 595 
 596   // Check to grow _preorders[] array for the case when build_loop_tree_impl()
 597   // adds new nodes.
 598   void check_grow_preorders( ) {
 599     if ( _max_preorder < C->unique() ) {
 600       uint newsize = _max_preorder<<1;  // double size of array
 601       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
 602       memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
 603       _max_preorder = newsize;
 604     }
 605   }
 606   // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
 607   int is_visited( Node *n ) const { return _preorders[n->_idx]; }
 608   // Pre-order numbers are written to the Nodes array as low-bit-set values.
 609   void set_preorder_visited( Node *n, int pre_order ) {
 610     assert( !is_visited( n ), "already set" );
 611     _preorders[n->_idx] = (pre_order<<1);
 612   };
 613   // Return pre-order number.
 614   int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
 615 
 616   // Check for being post-visited.
 617   // Should be previsited already (checked with assert(is_visited(n))).
 618   int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
 619 
 620   // Mark as post visited
 621   void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
 622 
 623   // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
 624   // Returns true if "n" is a data node, false if it's a control node.
 625   bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
 626 
 627   // clear out dead code after build_loop_late
 628   Node_List _deadlist;
 629 
 630   // Support for faster execution of get_late_ctrl()/dom_lca()
 631   // when a node has many uses and dominator depth is deep.
 632   Node_Array _dom_lca_tags;
 633   void   init_dom_lca_tags();
 634   void   clear_dom_lca_tags();
 635 
 636   // Helper for debugging bad dominance relationships
 637   bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
 638 
 639   Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
 640 
 641   // Inline wrapper for frequent cases:
 642   // 1) only one use
 643   // 2) a use is the same as the current LCA passed as 'n1'
 644   Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
 645     assert( n->is_CFG(), "" );
 646     // Fast-path NULL lca
 647     if( lca != NULL && lca != n ) {
 648       assert( lca->is_CFG(), "" );
 649       // find LCA of all uses
 650       n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
 651     }
 652     return find_non_split_ctrl(n);
 653   }
 654   Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
 655 
 656   // Helper function for directing control inputs away from CFG split
 657   // points.
 658   Node *find_non_split_ctrl( Node *ctrl ) const {
 659     if (ctrl != NULL) {
 660       if (ctrl->is_MultiBranch()) {
 661         ctrl = ctrl->in(0);
 662       }
 663       assert(ctrl->is_CFG(), "CFG");
 664     }
 665     return ctrl;
 666   }
 667 
 668   bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
 669 
 670 public:
 671 
 672   static bool is_canonical_loop_entry(CountedLoopNode* cl);
 673 
 674   bool has_node( Node* n ) const {
 675     guarantee(n != NULL, "No Node.");
 676     return _nodes[n->_idx] != NULL;
 677   }
 678   // check if transform created new nodes that need _ctrl recorded
 679   Node *get_late_ctrl( Node *n, Node *early );
 680   Node *get_early_ctrl( Node *n );
 681   Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
 682   void set_early_ctrl( Node *n );
 683   void set_subtree_ctrl( Node *root );
 684   void set_ctrl( Node *n, Node *ctrl ) {
 685     assert( !has_node(n) || has_ctrl(n), "" );
 686     assert( ctrl->in(0), "cannot set dead control node" );
 687     assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
 688     _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
 689   }
 690   // Set control and update loop membership
 691   void set_ctrl_and_loop(Node* n, Node* ctrl) {
 692     IdealLoopTree* old_loop = get_loop(get_ctrl(n));
 693     IdealLoopTree* new_loop = get_loop(ctrl);
 694     if (old_loop != new_loop) {
 695       if (old_loop->_child == NULL) old_loop->_body.yank(n);
 696       if (new_loop->_child == NULL) new_loop->_body.push(n);
 697     }
 698     set_ctrl(n, ctrl);
 699   }
 700   // Control nodes can be replaced or subsumed.  During this pass they
 701   // get their replacement Node in slot 1.  Instead of updating the block
 702   // location of all Nodes in the subsumed block, we lazily do it.  As we
 703   // pull such a subsumed block out of the array, we write back the final
 704   // correct block.
 705   Node *get_ctrl( Node *i ) {
 706     assert(has_node(i), "");
 707     Node *n = get_ctrl_no_update(i);
 708     _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
 709     assert(has_node(i) && has_ctrl(i), "");
 710     assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
 711     return n;
 712   }
 713   // true if CFG node d dominates CFG node n
 714   bool is_dominator(Node *d, Node *n);
 715   // return get_ctrl for a data node and self(n) for a CFG node
 716   Node* ctrl_or_self(Node* n) {
 717     if (has_ctrl(n))
 718       return get_ctrl(n);
 719     else {
 720       assert (n->is_CFG(), "must be a CFG node");
 721       return n;
 722     }
 723   }
 724 
 725 private:
 726   Node *get_ctrl_no_update_helper(Node *i) const {
 727     assert(has_ctrl(i), "should be control, not loop");
 728     return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
 729   }
 730 
 731   Node *get_ctrl_no_update(Node *i) const {
 732     assert( has_ctrl(i), "" );
 733     Node *n = get_ctrl_no_update_helper(i);
 734     if (!n->in(0)) {
 735       // Skip dead CFG nodes
 736       do {
 737         n = get_ctrl_no_update_helper(n);
 738       } while (!n->in(0));
 739       n = find_non_split_ctrl(n);
 740     }
 741     return n;
 742   }
 743 
 744   // Check for loop being set
 745   // "n" must be a control node. Returns true if "n" is known to be in a loop.
 746   bool has_loop( Node *n ) const {
 747     assert(!has_node(n) || !has_ctrl(n), "");
 748     return has_node(n);
 749   }
 750   // Set loop
 751   void set_loop( Node *n, IdealLoopTree *loop ) {
 752     _nodes.map(n->_idx, (Node*)loop);
 753   }
 754   // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
 755   // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
 756   // from old_node to new_node to support the lazy update.  Reference
 757   // replaces loop reference, since that is not needed for dead node.
 758 public:
 759   void lazy_update(Node *old_node, Node *new_node) {
 760     assert(old_node != new_node, "no cycles please");
 761     // Re-use the side array slot for this node to provide the
 762     // forwarding pointer.
 763     _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
 764   }
 765   void lazy_replace(Node *old_node, Node *new_node) {
 766     _igvn.replace_node(old_node, new_node);
 767     lazy_update(old_node, new_node);
 768   }
 769 
 770 private:
 771 
 772   // Place 'n' in some loop nest, where 'n' is a CFG node
 773   void build_loop_tree();
 774   int build_loop_tree_impl( Node *n, int pre_order );
 775   // Insert loop into the existing loop tree.  'innermost' is a leaf of the
 776   // loop tree, not the root.
 777   IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
 778 
 779   // Place Data nodes in some loop nest
 780   void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 781   void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 782   void build_loop_late_post ( Node* n );
 783 
 784   // Array of immediate dominance info for each CFG node indexed by node idx
 785 private:
 786   uint _idom_size;
 787   Node **_idom;                 // Array of immediate dominators
 788   uint *_dom_depth;           // Used for fast LCA test
 789   GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
 790 
 791   Node* idom_no_update(Node* d) const {
 792     assert(d->_idx < _idom_size, "oob");
 793     Node* n = _idom[d->_idx];
 794     assert(n != NULL,"Bad immediate dominator info.");
 795     while (n->in(0) == NULL) {  // Skip dead CFG nodes
 796       //n = n->in(1);
 797       n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
 798       assert(n != NULL,"Bad immediate dominator info.");
 799     }
 800     return n;
 801   }
 802   Node *idom(Node* d) const {
 803     uint didx = d->_idx;
 804     Node *n = idom_no_update(d);
 805     _idom[didx] = n;            // Lazily remove dead CFG nodes from table.
 806     return n;
 807   }
 808   uint dom_depth(Node* d) const {
 809     guarantee(d != NULL, "Null dominator info.");
 810     guarantee(d->_idx < _idom_size, "");
 811     return _dom_depth[d->_idx];
 812   }
 813   void set_idom(Node* d, Node* n, uint dom_depth);
 814   // Locally compute IDOM using dom_lca call
 815   Node *compute_idom( Node *region ) const;
 816   // Recompute dom_depth
 817   void recompute_dom_depth();
 818 
 819   // Is safept not required by an outer loop?
 820   bool is_deleteable_safept(Node* sfpt);
 821 
 822   // Replace parallel induction variable (parallel to trip counter)
 823   void replace_parallel_iv(IdealLoopTree *loop);
 824 
 825   // Perform verification that the graph is valid.
 826   PhaseIdealLoop( PhaseIterGVN &igvn) :
 827     PhaseTransform(Ideal_Loop),
 828     _igvn(igvn),
 829     _dom_lca_tags(arena()), // Thread::resource_area
 830     _verify_me(NULL),
 831     _verify_only(true) {
 832     build_and_optimize(false, false);
 833   }
 834 
 835   // build the loop tree and perform any requested optimizations
 836   void build_and_optimize(bool do_split_if, bool skip_loop_opts);
 837 
 838 public:
 839   // Dominators for the sea of nodes
 840   void Dominators();
 841   Node *dom_lca( Node *n1, Node *n2 ) const {
 842     return find_non_split_ctrl(dom_lca_internal(n1, n2));
 843   }
 844   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
 845 
 846   // Compute the Ideal Node to Loop mapping
 847   PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
 848     PhaseTransform(Ideal_Loop),
 849     _igvn(igvn),
 850     _dom_lca_tags(arena()), // Thread::resource_area
 851     _verify_me(NULL),
 852     _verify_only(false) {
 853     build_and_optimize(do_split_ifs, skip_loop_opts);
 854   }
 855 
 856   // Verify that verify_me made the same decisions as a fresh run.
 857   PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
 858     PhaseTransform(Ideal_Loop),
 859     _igvn(igvn),
 860     _dom_lca_tags(arena()), // Thread::resource_area
 861     _verify_me(verify_me),
 862     _verify_only(false) {
 863     build_and_optimize(false, false);
 864   }
 865 
 866   // Build and verify the loop tree without modifying the graph.  This
 867   // is useful to verify that all inputs properly dominate their uses.
 868   static void verify(PhaseIterGVN& igvn) {
 869 #ifdef ASSERT
 870     PhaseIdealLoop v(igvn);
 871 #endif
 872   }
 873 
 874   // True if the method has at least 1 irreducible loop
 875   bool _has_irreducible_loops;
 876 
 877   // Per-Node transform
 878   virtual Node *transform( Node *a_node ) { return 0; }
 879 
 880   bool is_counted_loop( Node *x, IdealLoopTree *loop );
 881 
 882   Node* exact_limit( IdealLoopTree *loop );
 883 
 884   // Return a post-walked LoopNode
 885   IdealLoopTree *get_loop( Node *n ) const {
 886     // Dead nodes have no loop, so return the top level loop instead
 887     if (!has_node(n))  return _ltree_root;
 888     assert(!has_ctrl(n), "");
 889     return (IdealLoopTree*)_nodes[n->_idx];
 890   }
 891 
 892   // Is 'n' a (nested) member of 'loop'?
 893   int is_member( const IdealLoopTree *loop, Node *n ) const {
 894     return loop->is_member(get_loop(n)); }
 895 
 896   // This is the basic building block of the loop optimizations.  It clones an
 897   // entire loop body.  It makes an old_new loop body mapping; with this
 898   // mapping you can find the new-loop equivalent to an old-loop node.  All
 899   // new-loop nodes are exactly equal to their old-loop counterparts, all
 900   // edges are the same.  All exits from the old-loop now have a RegionNode
 901   // that merges the equivalent new-loop path.  This is true even for the
 902   // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
 903   // now come from (one or more) Phis that merge their new-loop equivalents.
 904   // Parameter side_by_side_idom:
 905   //   When side_by_size_idom is NULL, the dominator tree is constructed for
 906   //      the clone loop to dominate the original.  Used in construction of
 907   //      pre-main-post loop sequence.
 908   //   When nonnull, the clone and original are side-by-side, both are
 909   //      dominated by the passed in side_by_side_idom node.  Used in
 910   //      construction of unswitched loops.
 911   void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
 912                    Node* side_by_side_idom = NULL);
 913 
 914   // If we got the effect of peeling, either by actually peeling or by
 915   // making a pre-loop which must execute at least once, we can remove
 916   // all loop-invariant dominated tests in the main body.
 917   void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
 918 
 919   // Generate code to do a loop peel for the given loop (and body).
 920   // old_new is a temp array.
 921   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
 922 
 923   // Add pre and post loops around the given loop.  These loops are used
 924   // during RCE, unrolling and aligning loops.
 925   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
 926 
 927   // Add post loop after the given loop.
 928   Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
 929                          CountedLoopNode *main_head, CountedLoopEndNode *main_end,
 930                          Node *incr, Node *limit, CountedLoopNode *&post_head);
 931 
 932   // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
 933   void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
 934 
 935   // Add a vector post loop between a vector main loop and the current post loop
 936   void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
 937   // If Node n lives in the back_ctrl block, we clone a private version of n
 938   // in preheader_ctrl block and return that, otherwise return n.
 939   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
 940 
 941   // Take steps to maximally unroll the loop.  Peel any odd iterations, then
 942   // unroll to do double iterations.  The next round of major loop transforms
 943   // will repeat till the doubled loop body does all remaining iterations in 1
 944   // pass.
 945   void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
 946 
 947   // Unroll the loop body one step - make each trip do 2 iterations.
 948   void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
 949 
 950   // Mark vector reduction candidates before loop unrolling
 951   void mark_reductions( IdealLoopTree *loop );
 952 
 953   // Return true if exp is a constant times an induction var
 954   bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
 955 
 956   // Return true if exp is a scaled induction var plus (or minus) constant
 957   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
 958 
 959   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
 960   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 961                                         Deoptimization::DeoptReason reason,
 962                                         int opcode);
 963   void register_control(Node* n, IdealLoopTree *loop, Node* pred);
 964 
 965   // Clone loop predicates to cloned loops (peeled, unswitched)
 966   static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 967                                    Deoptimization::DeoptReason reason,
 968                                    PhaseIdealLoop* loop_phase,
 969                                    PhaseIterGVN* igvn);
 970 
 971   static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
 972                                          bool clone_limit_check,
 973                                          PhaseIdealLoop* loop_phase,
 974                                          PhaseIterGVN* igvn);
 975   Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
 976 
 977   static Node* skip_loop_predicates(Node* entry);
 978 
 979   // Find a good location to insert a predicate
 980   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
 981   // Find a predicate
 982   static Node* find_predicate(Node* entry);
 983   // Construct a range check for a predicate if
 984   BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
 985                          int scale, Node* offset,
 986                          Node* init, Node* limit, Node* stride,
 987                          Node* range, bool upper);
 988 
 989   // Implementation of the loop predication to promote checks outside the loop
 990   bool loop_predication_impl(IdealLoopTree *loop);
 991 
 992   // Helper function to collect predicate for eliminating the useless ones
 993   void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
 994   void eliminate_useless_predicates();
 995 
 996   // Change the control input of expensive nodes to allow commoning by
 997   // IGVN when it is guaranteed to not result in a more frequent
 998   // execution of the expensive node. Return true if progress.
 999   bool process_expensive_nodes();
1000 
1001   // Check whether node has become unreachable
1002   bool is_node_unreachable(Node *n) const {
1003     return !has_node(n) || n->is_unreachable(_igvn);
1004   }
1005 
1006   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1007   int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1008 
1009   // Check to see if do_range_check(...) cleaned the main loop of range-checks
1010   void has_range_checks(IdealLoopTree *loop);
1011 
1012   // Process post loops which have range checks and try to build a multi-version
1013   // guard to safely determine if we can execute the post loop which was RCE'd.
1014   bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1015 
1016   // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1017   void poison_rce_post_loop(IdealLoopTree *rce_loop);
1018 
1019   // Create a slow version of the loop by cloning the loop
1020   // and inserting an if to select fast-slow versions.
1021   ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1022                                         Node_List &old_new,
1023                                         int opcode);
1024 
1025   // Clone a loop and return the clone head (clone_loop_head).
1026   // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1027   // This routine was created for usage in CountedLoopReserveKit.
1028   //
1029   //    int(1) -> If -> IfTrue -> original_loop_head
1030   //              |
1031   //              V
1032   //           IfFalse -> clone_loop_head (returned by function pointer)
1033   //
1034   LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1035   // Clone loop with an invariant test (that does not exit) and
1036   // insert a clone of the test that selects which version to
1037   // execute.
1038   void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1039 
1040   // Find candidate "if" for unswitching
1041   IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1042 
1043   // Range Check Elimination uses this function!
1044   // Constrain the main loop iterations so the affine function:
1045   //    low_limit <= scale_con * I + offset  <  upper_limit
1046   // always holds true.  That is, either increase the number of iterations in
1047   // the pre-loop or the post-loop until the condition holds true in the main
1048   // loop.  Scale_con, offset and limit are all loop invariant.
1049   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 );
1050   // Helper function for add_constraint().
1051   Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1052 
1053   // Partially peel loop up through last_peel node.
1054   bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1055 
1056   // Create a scheduled list of nodes control dependent on ctrl set.
1057   void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1058   // Has a use in the vector set
1059   bool has_use_in_set( Node* n, VectorSet& vset );
1060   // Has use internal to the vector set (ie. not in a phi at the loop head)
1061   bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1062   // clone "n" for uses that are outside of loop
1063   int  clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1064   // clone "n" for special uses that are in the not_peeled region
1065   void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1066                                           VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1067   // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1068   void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1069 #ifdef ASSERT
1070   // Validate the loop partition sets: peel and not_peel
1071   bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1072   // Ensure that uses outside of loop are of the right form
1073   bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1074                                  uint orig_exit_idx, uint clone_exit_idx);
1075   bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1076 #endif
1077 
1078   // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1079   int stride_of_possible_iv( Node* iff );
1080   bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1081   // Return the (unique) control output node that's in the loop (if it exists.)
1082   Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1083   // Insert a signed compare loop exit cloned from an unsigned compare.
1084   IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1085   void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1086   // Utility to register node "n" with PhaseIdealLoop
1087   void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1088   // Utility to create an if-projection
1089   ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1090   // Force the iff control output to be the live_proj
1091   Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1092   // Insert a region before an if projection
1093   RegionNode* insert_region_before_proj(ProjNode* proj);
1094   // Insert a new if before an if projection
1095   ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1096 
1097   // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1098   // "Nearly" because all Nodes have been cloned from the original in the loop,
1099   // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1100   // through the Phi recursively, and return a Bool.
1101   BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1102   CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1103 
1104 
1105   // Rework addressing expressions to get the most loop-invariant stuff
1106   // moved out.  We'd like to do all associative operators, but it's especially
1107   // important (common) to do address expressions.
1108   Node *remix_address_expressions( Node *n );
1109 
1110   // Attempt to use a conditional move instead of a phi/branch
1111   Node *conditional_move( Node *n );
1112 
1113   // Reorganize offset computations to lower register pressure.
1114   // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1115   // (which are then alive with the post-incremented trip counter
1116   // forcing an extra register move)
1117   void reorg_offsets( IdealLoopTree *loop );
1118 
1119   // Check for aggressive application of 'split-if' optimization,
1120   // using basic block level info.
1121   void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack );
1122   Node *split_if_with_blocks_pre ( Node *n );
1123   void  split_if_with_blocks_post( Node *n );
1124   Node *has_local_phi_input( Node *n );
1125   // Mark an IfNode as being dominated by a prior test,
1126   // without actually altering the CFG (and hence IDOM info).
1127   void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1128 
1129   // Split Node 'n' through merge point
1130   Node *split_thru_region( Node *n, Node *region );
1131   // Split Node 'n' through merge point if there is enough win.
1132   Node *split_thru_phi( Node *n, Node *region, int policy );
1133   // Found an If getting its condition-code input from a Phi in the
1134   // same block.  Split thru the Region.
1135   void do_split_if( Node *iff );
1136 
1137   // Conversion of fill/copy patterns into intrisic versions
1138   bool do_intrinsify_fill();
1139   bool intrinsify_fill(IdealLoopTree* lpt);
1140   bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1141                        Node*& shift, Node*& offset);
1142 
1143 private:
1144   // Return a type based on condition control flow
1145   const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1146   const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1147  // Helpers for filtered type
1148   const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1149 
1150   // Helper functions
1151   Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1152   Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1153   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 );
1154   bool split_up( Node *n, Node *blk1, Node *blk2 );
1155   void sink_use( Node *use, Node *post_loop );
1156   Node *place_near_use( Node *useblock ) const;
1157   Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1158   void try_move_store_after_loop(Node* n);
1159   bool identical_backtoback_ifs(Node *n);
1160   bool can_split_if(Node *n_ctrl);
1161 
1162   bool _created_loop_node;
1163 public:
1164   void set_created_loop_node() { _created_loop_node = true; }
1165   bool created_loop_node()     { return _created_loop_node; }
1166   void register_new_node( Node *n, Node *blk );
1167 
1168 #ifdef ASSERT
1169   void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1170 #endif
1171 
1172 #ifndef PRODUCT
1173   void dump( ) const;
1174   void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1175   void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1176   void verify() const;          // Major slow  :-)
1177   void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1178   IdealLoopTree *get_loop_idx(Node* n) const {
1179     // Dead nodes have no loop, so return the top level loop instead
1180     return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1181   }
1182   // Print some stats
1183   static void print_statistics();
1184   static int _loop_invokes;     // Count of PhaseIdealLoop invokes
1185   static int _loop_work;        // Sum of PhaseIdealLoop x _unique
1186 #endif
1187 };
1188 
1189 // This kit may be used for making of a reserved copy of a loop before this loop
1190 //  goes under non-reversible changes.
1191 //
1192 // Function create_reserve() creates a reserved copy (clone) of the loop.
1193 // The reserved copy is created by calling
1194 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1195 // the original and reserved loops are connected in the outer graph.
1196 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1197 //
1198 // By default the reserved copy (clone) of the loop is created as dead code - it is
1199 // dominated in the outer loop by this node chain:
1200 //   intcon(1)->If->IfFalse->reserved_copy.
1201 // The original loop is dominated by the the same node chain but IfTrue projection:
1202 //   intcon(0)->If->IfTrue->original_loop.
1203 //
1204 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1205 // and the dtor, checks _use_new value.
1206 // If _use_new == false, it "switches" control to reserved copy of the loop
1207 // by simple replacing of node intcon(1) with node intcon(0).
1208 //
1209 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1210 //
1211 // void CountedLoopReserveKit_example()
1212 // {
1213 //    CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1214 //    if (DoReserveCopy && !lrk.has_reserved()) {
1215 //      return; //failed to create reserved loop copy
1216 //    }
1217 //    ...
1218 //    //something is wrong, switch to original loop
1219 ///   if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1220 //    ...
1221 //    //everything worked ok, return with the newly modified loop
1222 //    lrk.use_new();
1223 //    return; // ~CountedLoopReserveKit does nothing once use_new() was called
1224 //  }
1225 //
1226 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1227 // the dtor will "switch to the original" loop.
1228 // NOTE. You you modify outside of the original loop this class is no help.
1229 //
1230 class CountedLoopReserveKit {
1231   private:
1232     PhaseIdealLoop* _phase;
1233     IdealLoopTree*  _lpt;
1234     LoopNode*       _lp;
1235     IfNode*         _iff;
1236     LoopNode*       _lp_reserved;
1237     bool            _has_reserved;
1238     bool            _use_new;
1239     const bool      _active; //may be set to false in ctor, then the object is dummy
1240 
1241   public:
1242     CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1243     ~CountedLoopReserveKit();
1244     void use_new()                {_use_new = true;}
1245     void set_iff(IfNode* x)       {_iff = x;}
1246     bool has_reserved()     const { return _active && _has_reserved;}
1247   private:
1248     bool create_reserve();
1249 };// class CountedLoopReserveKit
1250 
1251 inline Node* IdealLoopTree::tail() {
1252 // Handle lazy update of _tail field
1253   Node *n = _tail;
1254   //while( !n->in(0) )  // Skip dead CFG nodes
1255     //n = n->in(1);
1256   if (n->in(0) == NULL)
1257     n = _phase->get_ctrl(n);
1258   _tail = n;
1259   return n;
1260 }
1261 
1262 
1263 // Iterate over the loop tree using a preorder, left-to-right traversal.
1264 //
1265 // Example that visits all counted loops from within PhaseIdealLoop
1266 //
1267 //  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1268 //   IdealLoopTree* lpt = iter.current();
1269 //   if (!lpt->is_counted()) continue;
1270 //   ...
1271 class LoopTreeIterator : public StackObj {
1272 private:
1273   IdealLoopTree* _root;
1274   IdealLoopTree* _curnt;
1275 
1276 public:
1277   LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1278 
1279   bool done() { return _curnt == NULL; }       // Finished iterating?
1280 
1281   void next();                                 // Advance to next loop tree
1282 
1283   IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
1284 };
1285 
1286 #endif // SHARE_VM_OPTO_LOOPNODE_HPP