1 /*
   2  * Copyright (c) 1998, 2013, 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 VectorSet;
  42 class Invariance;
  43 struct small_cache;
  44 
  45 //
  46 //                  I D E A L I Z E D   L O O P S
  47 //
  48 // Idealized loops are the set of loops I perform more interesting
  49 // transformations on, beyond simple hoisting.
  50 
  51 //------------------------------LoopNode---------------------------------------
  52 // Simple loop header.  Fall in path on left, loop-back path on right.
  53 class LoopNode : public RegionNode {
  54   // Size is bigger to hold the flags.  However, the flags do not change
  55   // the semantics so it does not appear in the hash & cmp functions.
  56   virtual uint size_of() const { return sizeof(*this); }
  57 protected:
  58   short _loop_flags;
  59   // Names for flag bitfields
  60   enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
  61          MainHasNoPreLoop=4,
  62          HasExactTripCount=8,
  63          InnerLoop=16,
  64          PartialPeelLoop=32,
  65          PartialPeelFailed=64,
  66          HasReductions=128,
  67          WasSlpAnalyzed=256,
  68          PassedSlpAnalysis=512,
  69          DoUnrollOnly=1024 };
  70   char _unswitch_count;
  71   enum { _unswitch_max=3 };
  72 
  73 public:
  74   // Names for edge indices
  75   enum { Self=0, EntryControl, LoopBackControl };
  76 
  77   int is_inner_loop() const { return _loop_flags & InnerLoop; }
  78   void set_inner_loop() { _loop_flags |= InnerLoop; }
  79 
  80   int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
  81   void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
  82   int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
  83   void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
  84   void mark_has_reductions() { _loop_flags |= HasReductions; }
  85   void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
  86   void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
  87   void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
  88 
  89   int unswitch_max() { return _unswitch_max; }
  90   int unswitch_count() { return _unswitch_count; }
  91   void set_unswitch_count(int val) {
  92     assert (val <= unswitch_max(), "too many unswitches");
  93     _unswitch_count = val;
  94   }
  95 
  96   LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
  97     init_class_id(Class_Loop);
  98     init_req(EntryControl, entry);
  99     init_req(LoopBackControl, backedge);
 100   }
 101 
 102   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 103   virtual int Opcode() const;
 104   bool can_be_counted_loop(PhaseTransform* phase) const {
 105     return req() == 3 && in(0) != NULL &&
 106       in(1) != NULL && phase->type(in(1)) != Type::TOP &&
 107       in(2) != NULL && phase->type(in(2)) != Type::TOP;
 108   }
 109   bool is_valid_counted_loop() const;
 110 #ifndef PRODUCT
 111   virtual void dump_spec(outputStream *st) const;
 112 #endif
 113 };
 114 
 115 //------------------------------Counted Loops----------------------------------
 116 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
 117 // path (and maybe some other exit paths).  The trip-counter exit is always
 118 // last in the loop.  The trip-counter have to stride by a constant;
 119 // the exit value is also loop invariant.
 120 
 121 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
 122 // CountedLoopNode has the incoming loop control and the loop-back-control
 123 // which is always the IfTrue before the matching CountedLoopEndNode.  The
 124 // CountedLoopEndNode has an incoming control (possibly not the
 125 // CountedLoopNode if there is control flow in the loop), the post-increment
 126 // trip-counter value, and the limit.  The trip-counter value is always of
 127 // the form (Op old-trip-counter stride).  The old-trip-counter is produced
 128 // by a Phi connected to the CountedLoopNode.  The stride is constant.
 129 // The Op is any commutable opcode, including Add, Mul, Xor.  The
 130 // CountedLoopEndNode also takes in the loop-invariant limit value.
 131 
 132 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
 133 // loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
 134 // via the old-trip-counter from the Op node.
 135 
 136 //------------------------------CountedLoopNode--------------------------------
 137 // CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
 138 // inputs the incoming loop-start control and the loop-back control, so they
 139 // act like RegionNodes.  They also take in the initial trip counter, the
 140 // loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
 141 // produce a loop-body control and the trip counter value.  Since
 142 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
 143 
 144 class CountedLoopNode : public LoopNode {
 145   // Size is bigger to hold _main_idx.  However, _main_idx does not change
 146   // the semantics so it does not appear in the hash & cmp functions.
 147   virtual uint size_of() const { return sizeof(*this); }
 148 
 149   // For Pre- and Post-loops during debugging ONLY, this holds the index of
 150   // the Main CountedLoop.  Used to assert that we understand the graph shape.
 151   node_idx_t _main_idx;
 152 
 153   // Known trip count calculated by compute_exact_trip_count()
 154   uint  _trip_count;
 155 
 156   // Expected trip count from profile data
 157   float _profile_trip_cnt;
 158 
 159   // Log2 of original loop bodies in unrolled loop
 160   int _unrolled_count_log2;
 161 
 162   // Node count prior to last unrolling - used to decide if
 163   // unroll,optimize,unroll,optimize,... is making progress
 164   int _node_count_before_unroll;
 165 
 166   // If slp analysis is performed we record the maximum
 167   // vector mapped unroll factor here
 168   int _slp_maximum_unroll_factor;
 169 
 170 public:
 171   CountedLoopNode( Node *entry, Node *backedge )
 172     : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
 173       _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
 174       _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
 175     init_class_id(Class_CountedLoop);
 176     // Initialize _trip_count to the largest possible value.
 177     // Will be reset (lower) if the loop's trip count is known.
 178   }
 179 
 180   virtual int Opcode() const;
 181   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 182 
 183   Node *init_control() const { return in(EntryControl); }
 184   Node *back_control() const { return in(LoopBackControl); }
 185   CountedLoopEndNode *loopexit() const;
 186   Node *init_trip() const;
 187   Node *stride() const;
 188   int   stride_con() const;
 189   bool  stride_is_con() const;
 190   Node *limit() const;
 191   Node *incr() const;
 192   Node *phi() const;
 193 
 194   // Match increment with optional truncation
 195   static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
 196 
 197   // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
 198   // can run short a few iterations and may start a few iterations in.
 199   // It will be RCE'd and unrolled and aligned.
 200 
 201   // A following 'post' loop will run any remaining iterations.  Used
 202   // during Range Check Elimination, the 'post' loop will do any final
 203   // iterations with full checks.  Also used by Loop Unrolling, where
 204   // the 'post' loop will do any epilog iterations needed.  Basically,
 205   // a 'post' loop can not profitably be further unrolled or RCE'd.
 206 
 207   // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
 208   // it may do under-flow checks for RCE and may do alignment iterations
 209   // so the following main loop 'knows' that it is striding down cache
 210   // lines.
 211 
 212   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
 213   // Aligned, may be missing it's pre-loop.
 214   int is_normal_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
 215   int is_pre_loop      () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
 216   int is_main_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
 217   int is_post_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
 218   int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
 219   int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
 220   int has_passed_slp   () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
 221   int do_unroll_only      () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
 222   int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
 223   void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 224 
 225   int main_idx() const { return _main_idx; }
 226 
 227 
 228   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
 229   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
 230   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
 231   void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 232 
 233   void set_trip_count(uint tc) { _trip_count = tc; }
 234   uint trip_count()            { return _trip_count; }
 235 
 236   bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
 237   void set_exact_trip_count(uint tc) {
 238     _trip_count = tc;
 239     _loop_flags |= HasExactTripCount;
 240   }
 241   void set_nonexact_trip_count() {
 242     _loop_flags &= ~HasExactTripCount;
 243   }
 244   void set_notpassed_slp() {
 245     _loop_flags &= ~PassedSlpAnalysis;
 246   }
 247 
 248   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
 249   float profile_trip_cnt()             { return _profile_trip_cnt; }
 250 
 251   void double_unrolled_count() { _unrolled_count_log2++; }
 252   int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
 253 
 254   void set_node_count_before_unroll(int ct)  { _node_count_before_unroll = ct; }
 255   int  node_count_before_unroll()            { return _node_count_before_unroll; }
 256   void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
 257   int  slp_max_unroll() const                { return _slp_maximum_unroll_factor; }
 258 
 259 #ifndef PRODUCT
 260   virtual void dump_spec(outputStream *st) const;
 261 #endif
 262 };
 263 
 264 //------------------------------CountedLoopEndNode-----------------------------
 265 // CountedLoopEndNodes end simple trip counted loops.  They act much like
 266 // IfNodes.
 267 class CountedLoopEndNode : public IfNode {
 268 public:
 269   enum { TestControl, TestValue };
 270 
 271   CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
 272     : IfNode( control, test, prob, cnt) {
 273     init_class_id(Class_CountedLoopEnd);
 274   }
 275   virtual int Opcode() const;
 276 
 277   Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
 278   Node *incr() const                { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 279   Node *limit() const               { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 280   Node *stride() const              { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 281   Node *phi() const                 { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 282   Node *init_trip() const           { Node *tmp = phi     (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 283   int stride_con() const;
 284   bool stride_is_con() const        { Node *tmp = stride  (); return (tmp != NULL && tmp->is_Con()); }
 285   BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
 286   CountedLoopNode *loopnode() const {
 287     // The CountedLoopNode that goes with this CountedLoopEndNode may
 288     // have been optimized out by the IGVN so be cautious with the
 289     // pattern matching on the graph
 290     if (phi() == NULL) {
 291       return NULL;
 292     }
 293     assert(phi()->is_Phi(), "should be PhiNode");
 294     Node *ln = phi()->in(0);
 295     if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
 296       return (CountedLoopNode*)ln;
 297     }
 298     return NULL;
 299   }
 300 
 301 #ifndef PRODUCT
 302   virtual void dump_spec(outputStream *st) const;
 303 #endif
 304 };
 305 
 306 
 307 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
 308   Node *bc = back_control();
 309   if( bc == NULL ) return NULL;
 310   Node *le = bc->in(0);
 311   if( le->Opcode() != Op_CountedLoopEnd )
 312     return NULL;
 313   return (CountedLoopEndNode*)le;
 314 }
 315 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
 316 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
 317 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
 318 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
 319 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
 320 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
 321 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
 322 
 323 //------------------------------LoopLimitNode-----------------------------
 324 // Counted Loop limit node which represents exact final iterator value:
 325 // trip_count = (limit - init_trip + stride - 1)/stride
 326 // final_value= trip_count * stride + init_trip.
 327 // Use HW instructions to calculate it when it can overflow in integer.
 328 // Note, final_value should fit into integer since counted loop has
 329 // limit check: limit <= max_int-stride.
 330 class LoopLimitNode : public Node {
 331   enum { Init=1, Limit=2, Stride=3 };
 332  public:
 333   LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
 334     // Put it on the Macro nodes list to optimize during macro nodes expansion.
 335     init_flags(Flag_is_macro);
 336     C->add_macro_node(this);
 337   }
 338   virtual int Opcode() const;
 339   virtual const Type *bottom_type() const { return TypeInt::INT; }
 340   virtual uint ideal_reg() const { return Op_RegI; }
 341   virtual const Type *Value( PhaseTransform *phase ) const;
 342   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 343   virtual Node *Identity( PhaseTransform *phase );
 344 };
 345 
 346 // -----------------------------IdealLoopTree----------------------------------
 347 class IdealLoopTree : public ResourceObj {
 348 public:
 349   IdealLoopTree *_parent;       // Parent in loop tree
 350   IdealLoopTree *_next;         // Next sibling in loop tree
 351   IdealLoopTree *_child;        // First child in loop tree
 352 
 353   // The head-tail backedge defines the loop.
 354   // If tail is NULL then this loop has multiple backedges as part of the
 355   // same loop.  During cleanup I'll peel off the multiple backedges; merge
 356   // them at the loop bottom and flow 1 real backedge into the loop.
 357   Node *_head;                  // Head of loop
 358   Node *_tail;                  // Tail of loop
 359   inline Node *tail();          // Handle lazy update of _tail field
 360   PhaseIdealLoop* _phase;
 361   int _local_loop_unroll_limit;
 362   int _local_loop_unroll_factor;
 363 
 364   Node_List _body;              // Loop body for inner loops
 365 
 366   uint8_t _nest;                // Nesting depth
 367   uint8_t _irreducible:1,       // True if irreducible
 368           _has_call:1,          // True if has call safepoint
 369           _has_sfpt:1,          // True if has non-call safepoint
 370           _rce_candidate:1;     // True if candidate for range check elimination
 371 
 372   Node_List* _safepts;          // List of safepoints in this loop
 373   Node_List* _required_safept;  // A inner loop cannot delete these safepts;
 374   bool  _allow_optimizations;   // Allow loop optimizations
 375 
 376   IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
 377     : _parent(0), _next(0), _child(0),
 378       _head(head), _tail(tail),
 379       _phase(phase),
 380       _safepts(NULL),
 381       _required_safept(NULL),
 382       _allow_optimizations(true),
 383       _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
 384       _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
 385   { }
 386 
 387   // Is 'l' a member of 'this'?
 388   bool is_member(const IdealLoopTree *l) const; // Test for nested membership
 389 
 390   // Set loop nesting depth.  Accumulate has_call bits.
 391   int set_nest( uint depth );
 392 
 393   // Split out multiple fall-in edges from the loop header.  Move them to a
 394   // private RegionNode before the loop.  This becomes the loop landing pad.
 395   void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
 396 
 397   // Split out the outermost loop from this shared header.
 398   void split_outer_loop( PhaseIdealLoop *phase );
 399 
 400   // Merge all the backedges from the shared header into a private Region.
 401   // Feed that region as the one backedge to this loop.
 402   void merge_many_backedges( PhaseIdealLoop *phase );
 403 
 404   // Split shared headers and insert loop landing pads.
 405   // Insert a LoopNode to replace the RegionNode.
 406   // Returns TRUE if loop tree is structurally changed.
 407   bool beautify_loops( PhaseIdealLoop *phase );
 408 
 409   // Perform optimization to use the loop predicates for null checks and range checks.
 410   // Applies to any loop level (not just the innermost one)
 411   bool loop_predication( PhaseIdealLoop *phase);
 412 
 413   // Perform iteration-splitting on inner loops.  Split iterations to
 414   // avoid range checks or one-shot null checks.  Returns false if the
 415   // current round of loop opts should stop.
 416   bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
 417 
 418   // Driver for various flavors of iteration splitting.  Returns false
 419   // if the current round of loop opts should stop.
 420   bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
 421 
 422   // Given dominators, try to find loops with calls that must always be
 423   // executed (call dominates loop tail).  These loops do not need non-call
 424   // safepoints (ncsfpt).
 425   void check_safepts(VectorSet &visited, Node_List &stack);
 426 
 427   // Allpaths backwards scan from loop tail, terminating each path at first safepoint
 428   // encountered.
 429   void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
 430 
 431   // Convert to counted loops where possible
 432   void counted_loop( PhaseIdealLoop *phase );
 433 
 434   // Check for Node being a loop-breaking test
 435   Node *is_loop_exit(Node *iff) const;
 436 
 437   // Returns true if ctrl is executed on every complete iteration
 438   bool dominates_backedge(Node* ctrl);
 439 
 440   // Remove simplistic dead code from loop body
 441   void DCE_loop_body();
 442 
 443   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
 444   // Replace with a 1-in-10 exit guess.
 445   void adjust_loop_exit_prob( PhaseIdealLoop *phase );
 446 
 447   // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
 448   // Useful for unrolling loops with NO array accesses.
 449   bool policy_peel_only( PhaseIdealLoop *phase ) const;
 450 
 451   // Return TRUE or FALSE if the loop should be unswitched -- clone
 452   // loop with an invariant test
 453   bool policy_unswitching( PhaseIdealLoop *phase ) const;
 454 
 455   // Micro-benchmark spamming.  Remove empty loops.
 456   bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
 457 
 458   // Convert one iteration loop into normal code.
 459   bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
 460 
 461   // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
 462   // make some loop-invariant test (usually a null-check) happen before the
 463   // loop.
 464   bool policy_peeling( PhaseIdealLoop *phase ) const;
 465 
 466   // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
 467   // known trip count in the counted loop node.
 468   bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
 469 
 470   // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
 471   // the loop is a CountedLoop and the body is small enough.
 472   bool policy_unroll(PhaseIdealLoop *phase);
 473 
 474   // Loop analyses to map to a maximal superword unrolling for vectorization.
 475   void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
 476 
 477   // Return TRUE or FALSE if the loop should be range-check-eliminated.
 478   // Gather a list of IF tests that are dominated by iteration splitting;
 479   // also gather the end of the first split and the start of the 2nd split.
 480   bool policy_range_check( PhaseIdealLoop *phase ) const;
 481 
 482   // Return TRUE or FALSE if the loop should be cache-line aligned.
 483   // Gather the expression that does the alignment.  Note that only
 484   // one array base can be aligned in a loop (unless the VM guarantees
 485   // mutual alignment).  Note that if we vectorize short memory ops
 486   // into longer memory ops, we may want to increase alignment.
 487   bool policy_align( PhaseIdealLoop *phase ) const;
 488 
 489   // Return TRUE if "iff" is a range check.
 490   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
 491 
 492   // Compute loop exact trip count if possible
 493   void compute_exact_trip_count( PhaseIdealLoop *phase );
 494 
 495   // Compute loop trip count from profile data
 496   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 497 
 498   // Reassociate invariant expressions.
 499   void reassociate_invariants(PhaseIdealLoop *phase);
 500   // Reassociate invariant add and subtract expressions.
 501   Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
 502   // Return nonzero index of invariant operand if invariant and variant
 503   // are combined with an Add or Sub. Helper for reassociate_invariants.
 504   int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
 505 
 506   // Return true if n is invariant
 507   bool is_invariant(Node* n) const;
 508 
 509   // Put loop body on igvn work list
 510   void record_for_igvn();
 511 
 512   bool is_loop()    { return !_irreducible && _tail && !_tail->is_top(); }
 513   bool is_inner()   { return is_loop() && _child == NULL; }
 514   bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
 515 
 516   void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
 517 
 518 #ifndef PRODUCT
 519   void dump_head( ) const;      // Dump loop head only
 520   void dump() const;            // Dump this loop recursively
 521   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 522 #endif
 523 
 524 };
 525 
 526 // -----------------------------PhaseIdealLoop---------------------------------
 527 // Computes the mapping from Nodes to IdealLoopTrees.  Organizes IdealLoopTrees into a
 528 // loop tree.  Drives the loop-based transformations on the ideal graph.
 529 class PhaseIdealLoop : public PhaseTransform {
 530   friend class IdealLoopTree;
 531   friend class SuperWord;
 532   // Pre-computed def-use info
 533   PhaseIterGVN &_igvn;
 534 
 535   // Head of loop tree
 536   IdealLoopTree *_ltree_root;
 537 
 538   // Array of pre-order numbers, plus post-visited bit.
 539   // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
 540   // ODD for post-visited.  Other bits are the pre-order number.
 541   uint *_preorders;
 542   uint _max_preorder;
 543 
 544   const PhaseIdealLoop* _verify_me;
 545   bool _verify_only;
 546 
 547   // Allocate _preorders[] array
 548   void allocate_preorders() {
 549     _max_preorder = C->unique()+8;
 550     _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
 551     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 552   }
 553 
 554   // Allocate _preorders[] array
 555   void reallocate_preorders() {
 556     if ( _max_preorder < C->unique() ) {
 557       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
 558       _max_preorder = C->unique();
 559     }
 560     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 561   }
 562 
 563   // Check to grow _preorders[] array for the case when build_loop_tree_impl()
 564   // adds new nodes.
 565   void check_grow_preorders( ) {
 566     if ( _max_preorder < C->unique() ) {
 567       uint newsize = _max_preorder<<1;  // double size of array
 568       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
 569       memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
 570       _max_preorder = newsize;
 571     }
 572   }
 573   // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
 574   int is_visited( Node *n ) const { return _preorders[n->_idx]; }
 575   // Pre-order numbers are written to the Nodes array as low-bit-set values.
 576   void set_preorder_visited( Node *n, int pre_order ) {
 577     assert( !is_visited( n ), "already set" );
 578     _preorders[n->_idx] = (pre_order<<1);
 579   };
 580   // Return pre-order number.
 581   int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
 582 
 583   // Check for being post-visited.
 584   // Should be previsited already (checked with assert(is_visited(n))).
 585   int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
 586 
 587   // Mark as post visited
 588   void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
 589 
 590   // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
 591   // Returns true if "n" is a data node, false if it's a control node.
 592   bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
 593 
 594   // clear out dead code after build_loop_late
 595   Node_List _deadlist;
 596 
 597   // Support for faster execution of get_late_ctrl()/dom_lca()
 598   // when a node has many uses and dominator depth is deep.
 599   Node_Array _dom_lca_tags;
 600   void   init_dom_lca_tags();
 601   void   clear_dom_lca_tags();
 602 
 603   // Helper for debugging bad dominance relationships
 604   bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
 605 
 606   Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
 607 
 608   // Inline wrapper for frequent cases:
 609   // 1) only one use
 610   // 2) a use is the same as the current LCA passed as 'n1'
 611   Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
 612     assert( n->is_CFG(), "" );
 613     // Fast-path NULL lca
 614     if( lca != NULL && lca != n ) {
 615       assert( lca->is_CFG(), "" );
 616       // find LCA of all uses
 617       n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
 618     }
 619     return find_non_split_ctrl(n);
 620   }
 621   Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
 622 
 623   // Helper function for directing control inputs away from CFG split
 624   // points.
 625   Node *find_non_split_ctrl( Node *ctrl ) const {
 626     if (ctrl != NULL) {
 627       if (ctrl->is_MultiBranch()) {
 628         ctrl = ctrl->in(0);
 629       }
 630       assert(ctrl->is_CFG(), "CFG");
 631     }
 632     return ctrl;
 633   }
 634 
 635   bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
 636 
 637 public:
 638   bool has_node( Node* n ) const {
 639     guarantee(n != NULL, "No Node.");
 640     return _nodes[n->_idx] != NULL;
 641   }
 642   // check if transform created new nodes that need _ctrl recorded
 643   Node *get_late_ctrl( Node *n, Node *early );
 644   Node *get_early_ctrl( Node *n );
 645   Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
 646   void set_early_ctrl( Node *n );
 647   void set_subtree_ctrl( Node *root );
 648   void set_ctrl( Node *n, Node *ctrl ) {
 649     assert( !has_node(n) || has_ctrl(n), "" );
 650     assert( ctrl->in(0), "cannot set dead control node" );
 651     assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
 652     _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
 653   }
 654   // Set control and update loop membership
 655   void set_ctrl_and_loop(Node* n, Node* ctrl) {
 656     IdealLoopTree* old_loop = get_loop(get_ctrl(n));
 657     IdealLoopTree* new_loop = get_loop(ctrl);
 658     if (old_loop != new_loop) {
 659       if (old_loop->_child == NULL) old_loop->_body.yank(n);
 660       if (new_loop->_child == NULL) new_loop->_body.push(n);
 661     }
 662     set_ctrl(n, ctrl);
 663   }
 664   // Control nodes can be replaced or subsumed.  During this pass they
 665   // get their replacement Node in slot 1.  Instead of updating the block
 666   // location of all Nodes in the subsumed block, we lazily do it.  As we
 667   // pull such a subsumed block out of the array, we write back the final
 668   // correct block.
 669   Node *get_ctrl( Node *i ) {
 670     assert(has_node(i), "");
 671     Node *n = get_ctrl_no_update(i);
 672     _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
 673     assert(has_node(i) && has_ctrl(i), "");
 674     assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
 675     return n;
 676   }
 677   // true if CFG node d dominates CFG node n
 678   bool is_dominator(Node *d, Node *n);
 679   // return get_ctrl for a data node and self(n) for a CFG node
 680   Node* ctrl_or_self(Node* n) {
 681     if (has_ctrl(n))
 682       return get_ctrl(n);
 683     else {
 684       assert (n->is_CFG(), "must be a CFG node");
 685       return n;
 686     }
 687   }
 688 
 689 private:
 690   Node *get_ctrl_no_update( Node *i ) const {
 691     assert( has_ctrl(i), "" );
 692     Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
 693     if (!n->in(0)) {
 694       // Skip dead CFG nodes
 695       do {
 696         n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
 697       } while (!n->in(0));
 698       n = find_non_split_ctrl(n);
 699     }
 700     return n;
 701   }
 702 
 703   // Check for loop being set
 704   // "n" must be a control node. Returns true if "n" is known to be in a loop.
 705   bool has_loop( Node *n ) const {
 706     assert(!has_node(n) || !has_ctrl(n), "");
 707     return has_node(n);
 708   }
 709   // Set loop
 710   void set_loop( Node *n, IdealLoopTree *loop ) {
 711     _nodes.map(n->_idx, (Node*)loop);
 712   }
 713   // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
 714   // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
 715   // from old_node to new_node to support the lazy update.  Reference
 716   // replaces loop reference, since that is not needed for dead node.
 717 public:
 718   void lazy_update( Node *old_node, Node *new_node ) {
 719     assert( old_node != new_node, "no cycles please" );
 720     //old_node->set_req( 1, new_node /*NO DU INFO*/ );
 721     // Nodes always have DU info now, so re-use the side array slot
 722     // for this node to provide the forwarding pointer.
 723     _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
 724   }
 725   void lazy_replace( Node *old_node, Node *new_node ) {
 726     _igvn.replace_node( old_node, new_node );
 727     lazy_update( old_node, new_node );
 728   }
 729   void lazy_replace_proj( Node *old_node, Node *new_node ) {
 730     assert( old_node->req() == 1, "use this for Projs" );
 731     _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
 732     old_node->add_req( NULL );
 733     lazy_replace( old_node, new_node );
 734   }
 735 
 736 private:
 737 
 738   // Place 'n' in some loop nest, where 'n' is a CFG node
 739   void build_loop_tree();
 740   int build_loop_tree_impl( Node *n, int pre_order );
 741   // Insert loop into the existing loop tree.  'innermost' is a leaf of the
 742   // loop tree, not the root.
 743   IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
 744 
 745   // Place Data nodes in some loop nest
 746   void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 747   void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 748   void build_loop_late_post ( Node* n );
 749 
 750   // Array of immediate dominance info for each CFG node indexed by node idx
 751 private:
 752   uint _idom_size;
 753   Node **_idom;                 // Array of immediate dominators
 754   uint *_dom_depth;           // Used for fast LCA test
 755   GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
 756 
 757   Node* idom_no_update(Node* d) const {
 758     assert(d->_idx < _idom_size, "oob");
 759     Node* n = _idom[d->_idx];
 760     assert(n != NULL,"Bad immediate dominator info.");
 761     while (n->in(0) == NULL) {  // Skip dead CFG nodes
 762       //n = n->in(1);
 763       n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
 764       assert(n != NULL,"Bad immediate dominator info.");
 765     }
 766     return n;
 767   }
 768   Node *idom(Node* d) const {
 769     uint didx = d->_idx;
 770     Node *n = idom_no_update(d);
 771     _idom[didx] = n;            // Lazily remove dead CFG nodes from table.
 772     return n;
 773   }
 774   uint dom_depth(Node* d) const {
 775     guarantee(d != NULL, "Null dominator info.");
 776     guarantee(d->_idx < _idom_size, "");
 777     return _dom_depth[d->_idx];
 778   }
 779   void set_idom(Node* d, Node* n, uint dom_depth);
 780   // Locally compute IDOM using dom_lca call
 781   Node *compute_idom( Node *region ) const;
 782   // Recompute dom_depth
 783   void recompute_dom_depth();
 784 
 785   // Is safept not required by an outer loop?
 786   bool is_deleteable_safept(Node* sfpt);
 787 
 788   // Replace parallel induction variable (parallel to trip counter)
 789   void replace_parallel_iv(IdealLoopTree *loop);
 790 
 791   // Perform verification that the graph is valid.
 792   PhaseIdealLoop( PhaseIterGVN &igvn) :
 793     PhaseTransform(Ideal_Loop),
 794     _igvn(igvn),
 795     _dom_lca_tags(arena()), // Thread::resource_area
 796     _verify_me(NULL),
 797     _verify_only(true) {
 798     build_and_optimize(false, false);
 799   }
 800 
 801   // build the loop tree and perform any requested optimizations
 802   void build_and_optimize(bool do_split_if, bool skip_loop_opts);
 803 
 804 public:
 805   // Dominators for the sea of nodes
 806   void Dominators();
 807   Node *dom_lca( Node *n1, Node *n2 ) const {
 808     return find_non_split_ctrl(dom_lca_internal(n1, n2));
 809   }
 810   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
 811 
 812   // Compute the Ideal Node to Loop mapping
 813   PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
 814     PhaseTransform(Ideal_Loop),
 815     _igvn(igvn),
 816     _dom_lca_tags(arena()), // Thread::resource_area
 817     _verify_me(NULL),
 818     _verify_only(false) {
 819     build_and_optimize(do_split_ifs, skip_loop_opts);
 820   }
 821 
 822   // Verify that verify_me made the same decisions as a fresh run.
 823   PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
 824     PhaseTransform(Ideal_Loop),
 825     _igvn(igvn),
 826     _dom_lca_tags(arena()), // Thread::resource_area
 827     _verify_me(verify_me),
 828     _verify_only(false) {
 829     build_and_optimize(false, false);
 830   }
 831 
 832   // Build and verify the loop tree without modifying the graph.  This
 833   // is useful to verify that all inputs properly dominate their uses.
 834   static void verify(PhaseIterGVN& igvn) {
 835 #ifdef ASSERT
 836     PhaseIdealLoop v(igvn);
 837 #endif
 838   }
 839 
 840   // True if the method has at least 1 irreducible loop
 841   bool _has_irreducible_loops;
 842 
 843   // Per-Node transform
 844   virtual Node *transform( Node *a_node ) { return 0; }
 845 
 846   bool is_counted_loop( Node *x, IdealLoopTree *loop );
 847 
 848   Node* exact_limit( IdealLoopTree *loop );
 849 
 850   // Return a post-walked LoopNode
 851   IdealLoopTree *get_loop( Node *n ) const {
 852     // Dead nodes have no loop, so return the top level loop instead
 853     if (!has_node(n))  return _ltree_root;
 854     assert(!has_ctrl(n), "");
 855     return (IdealLoopTree*)_nodes[n->_idx];
 856   }
 857 
 858   // Is 'n' a (nested) member of 'loop'?
 859   int is_member( const IdealLoopTree *loop, Node *n ) const {
 860     return loop->is_member(get_loop(n)); }
 861 
 862   // This is the basic building block of the loop optimizations.  It clones an
 863   // entire loop body.  It makes an old_new loop body mapping; with this
 864   // mapping you can find the new-loop equivalent to an old-loop node.  All
 865   // new-loop nodes are exactly equal to their old-loop counterparts, all
 866   // edges are the same.  All exits from the old-loop now have a RegionNode
 867   // that merges the equivalent new-loop path.  This is true even for the
 868   // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
 869   // now come from (one or more) Phis that merge their new-loop equivalents.
 870   // Parameter side_by_side_idom:
 871   //   When side_by_size_idom is NULL, the dominator tree is constructed for
 872   //      the clone loop to dominate the original.  Used in construction of
 873   //      pre-main-post loop sequence.
 874   //   When nonnull, the clone and original are side-by-side, both are
 875   //      dominated by the passed in side_by_side_idom node.  Used in
 876   //      construction of unswitched loops.
 877   void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
 878                    Node* side_by_side_idom = NULL);
 879 
 880   // If we got the effect of peeling, either by actually peeling or by
 881   // making a pre-loop which must execute at least once, we can remove
 882   // all loop-invariant dominated tests in the main body.
 883   void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
 884 
 885   // Generate code to do a loop peel for the given loop (and body).
 886   // old_new is a temp array.
 887   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
 888 
 889   // Add pre and post loops around the given loop.  These loops are used
 890   // during RCE, unrolling and aligning loops.
 891   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
 892   // If Node n lives in the back_ctrl block, we clone a private version of n
 893   // in preheader_ctrl block and return that, otherwise return n.
 894   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
 895 
 896   // Take steps to maximally unroll the loop.  Peel any odd iterations, then
 897   // unroll to do double iterations.  The next round of major loop transforms
 898   // will repeat till the doubled loop body does all remaining iterations in 1
 899   // pass.
 900   void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
 901 
 902   // Unroll the loop body one step - make each trip do 2 iterations.
 903   void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
 904 
 905   // Mark vector reduction candidates before loop unrolling
 906   void mark_reductions( IdealLoopTree *loop );
 907 
 908   // Return true if exp is a constant times an induction var
 909   bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
 910 
 911   // Return true if exp is a scaled induction var plus (or minus) constant
 912   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
 913 
 914   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
 915   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 916                                         Deoptimization::DeoptReason reason,
 917                                         int opcode);
 918   void register_control(Node* n, IdealLoopTree *loop, Node* pred);
 919 
 920   // Clone loop predicates to cloned loops (peeled, unswitched)
 921   static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 922                                    Deoptimization::DeoptReason reason,
 923                                    PhaseIdealLoop* loop_phase,
 924                                    PhaseIterGVN* igvn);
 925 
 926   static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
 927                                          bool clone_limit_check,
 928                                          PhaseIdealLoop* loop_phase,
 929                                          PhaseIterGVN* igvn);
 930   Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
 931 
 932   static Node* skip_loop_predicates(Node* entry);
 933 
 934   // Find a good location to insert a predicate
 935   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
 936   // Find a predicate
 937   static Node* find_predicate(Node* entry);
 938   // Construct a range check for a predicate if
 939   BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
 940                          int scale, Node* offset,
 941                          Node* init, Node* limit, Node* stride,
 942                          Node* range, bool upper);
 943 
 944   // Implementation of the loop predication to promote checks outside the loop
 945   bool loop_predication_impl(IdealLoopTree *loop);
 946 
 947   // Helper function to collect predicate for eliminating the useless ones
 948   void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
 949   void eliminate_useless_predicates();
 950 
 951   // Change the control input of expensive nodes to allow commoning by
 952   // IGVN when it is guaranteed to not result in a more frequent
 953   // execution of the expensive node. Return true if progress.
 954   bool process_expensive_nodes();
 955 
 956   // Check whether node has become unreachable
 957   bool is_node_unreachable(Node *n) const {
 958     return !has_node(n) || n->is_unreachable(_igvn);
 959   }
 960 
 961   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
 962   void do_range_check( IdealLoopTree *loop, Node_List &old_new );
 963 
 964   // Create a slow version of the loop by cloning the loop
 965   // and inserting an if to select fast-slow versions.
 966   ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
 967                                         Node_List &old_new,
 968                                         int opcode);
 969 
 970   // Clone loop with an invariant test (that does not exit) and
 971   // insert a clone of the test that selects which version to
 972   // execute.
 973   void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
 974 
 975   // Find candidate "if" for unswitching
 976   IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
 977 
 978   // Range Check Elimination uses this function!
 979   // Constrain the main loop iterations so the affine function:
 980   //    low_limit <= scale_con * I + offset  <  upper_limit
 981   // always holds true.  That is, either increase the number of iterations in
 982   // the pre-loop or the post-loop until the condition holds true in the main
 983   // loop.  Scale_con, offset and limit are all loop invariant.
 984   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 );
 985   // Helper function for add_constraint().
 986   Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
 987 
 988   // Partially peel loop up through last_peel node.
 989   bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
 990 
 991   // Create a scheduled list of nodes control dependent on ctrl set.
 992   void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
 993   // Has a use in the vector set
 994   bool has_use_in_set( Node* n, VectorSet& vset );
 995   // Has use internal to the vector set (ie. not in a phi at the loop head)
 996   bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
 997   // clone "n" for uses that are outside of loop
 998   int  clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
 999   // clone "n" for special uses that are in the not_peeled region
1000   void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1001                                           VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1002   // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1003   void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1004 #ifdef ASSERT
1005   // Validate the loop partition sets: peel and not_peel
1006   bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1007   // Ensure that uses outside of loop are of the right form
1008   bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1009                                  uint orig_exit_idx, uint clone_exit_idx);
1010   bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1011 #endif
1012 
1013   // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1014   int stride_of_possible_iv( Node* iff );
1015   bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1016   // Return the (unique) control output node that's in the loop (if it exists.)
1017   Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1018   // Insert a signed compare loop exit cloned from an unsigned compare.
1019   IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1020   void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1021   // Utility to register node "n" with PhaseIdealLoop
1022   void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1023   // Utility to create an if-projection
1024   ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1025   // Force the iff control output to be the live_proj
1026   Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1027   // Insert a region before an if projection
1028   RegionNode* insert_region_before_proj(ProjNode* proj);
1029   // Insert a new if before an if projection
1030   ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1031 
1032   // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1033   // "Nearly" because all Nodes have been cloned from the original in the loop,
1034   // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1035   // through the Phi recursively, and return a Bool.
1036   BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1037   CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1038 
1039 
1040   // Rework addressing expressions to get the most loop-invariant stuff
1041   // moved out.  We'd like to do all associative operators, but it's especially
1042   // important (common) to do address expressions.
1043   Node *remix_address_expressions( Node *n );
1044 
1045   // Attempt to use a conditional move instead of a phi/branch
1046   Node *conditional_move( Node *n );
1047 
1048   // Reorganize offset computations to lower register pressure.
1049   // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1050   // (which are then alive with the post-incremented trip counter
1051   // forcing an extra register move)
1052   void reorg_offsets( IdealLoopTree *loop );
1053 
1054   // Check for aggressive application of 'split-if' optimization,
1055   // using basic block level info.
1056   void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack );
1057   Node *split_if_with_blocks_pre ( Node *n );
1058   void  split_if_with_blocks_post( Node *n );
1059   Node *has_local_phi_input( Node *n );
1060   // Mark an IfNode as being dominated by a prior test,
1061   // without actually altering the CFG (and hence IDOM info).
1062   void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1063 
1064   // Split Node 'n' through merge point
1065   Node *split_thru_region( Node *n, Node *region );
1066   // Split Node 'n' through merge point if there is enough win.
1067   Node *split_thru_phi( Node *n, Node *region, int policy );
1068   // Found an If getting its condition-code input from a Phi in the
1069   // same block.  Split thru the Region.
1070   void do_split_if( Node *iff );
1071 
1072   // Conversion of fill/copy patterns into intrisic versions
1073   bool do_intrinsify_fill();
1074   bool intrinsify_fill(IdealLoopTree* lpt);
1075   bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1076                        Node*& shift, Node*& offset);
1077 
1078 private:
1079   // Return a type based on condition control flow
1080   const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1081   const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1082  // Helpers for filtered type
1083   const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1084 
1085   // Helper functions
1086   Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1087   Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1088   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 );
1089   bool split_up( Node *n, Node *blk1, Node *blk2 );
1090   void sink_use( Node *use, Node *post_loop );
1091   Node *place_near_use( Node *useblock ) const;
1092   Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1093   void try_move_store_after_loop(Node* n);
1094 
1095   bool _created_loop_node;
1096 public:
1097   void set_created_loop_node() { _created_loop_node = true; }
1098   bool created_loop_node()     { return _created_loop_node; }
1099   void register_new_node( Node *n, Node *blk );
1100 
1101 #ifdef ASSERT
1102   void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1103 #endif
1104 
1105 #ifndef PRODUCT
1106   void dump( ) const;
1107   void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1108   void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1109   void verify() const;          // Major slow  :-)
1110   void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1111   IdealLoopTree *get_loop_idx(Node* n) const {
1112     // Dead nodes have no loop, so return the top level loop instead
1113     return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1114   }
1115   // Print some stats
1116   static void print_statistics();
1117   static int _loop_invokes;     // Count of PhaseIdealLoop invokes
1118   static int _loop_work;        // Sum of PhaseIdealLoop x _unique
1119 #endif
1120 };
1121 
1122 inline Node* IdealLoopTree::tail() {
1123 // Handle lazy update of _tail field
1124   Node *n = _tail;
1125   //while( !n->in(0) )  // Skip dead CFG nodes
1126     //n = n->in(1);
1127   if (n->in(0) == NULL)
1128     n = _phase->get_ctrl(n);
1129   _tail = n;
1130   return n;
1131 }
1132 
1133 
1134 // Iterate over the loop tree using a preorder, left-to-right traversal.
1135 //
1136 // Example that visits all counted loops from within PhaseIdealLoop
1137 //
1138 //  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1139 //   IdealLoopTree* lpt = iter.current();
1140 //   if (!lpt->is_counted()) continue;
1141 //   ...
1142 class LoopTreeIterator : public StackObj {
1143 private:
1144   IdealLoopTree* _root;
1145   IdealLoopTree* _curnt;
1146 
1147 public:
1148   LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1149 
1150   bool done() { return _curnt == NULL; }       // Finished iterating?
1151 
1152   void next();                                 // Advance to next loop tree
1153 
1154   IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
1155 };
1156 
1157 #endif // SHARE_VM_OPTO_LOOPNODE_HPP