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