1 /*
   2  * Copyright (c) 1998, 2017, 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   char _unswitch_count;
  67   enum { _unswitch_max=3 };
  68 
  69 public:
  70   // Names for edge indices
  71   enum { Self=0, EntryControl, LoopBackControl };
  72 
  73   int is_inner_loop() const { return _loop_flags & InnerLoop; }
  74   void set_inner_loop() { _loop_flags |= InnerLoop; }
  75 
  76   int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
  77   void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
  78   int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
  79   void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
  80 
  81   int unswitch_max() { return _unswitch_max; }
  82   int unswitch_count() { return _unswitch_count; }
  83   void set_unswitch_count(int val) {
  84     assert (val <= unswitch_max(), "too many unswitches");
  85     _unswitch_count = val;
  86   }
  87 
  88   LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
  89     init_class_id(Class_Loop);
  90     init_req(EntryControl, entry);
  91     init_req(LoopBackControl, backedge);
  92   }
  93 
  94   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
  95   virtual int Opcode() const;
  96   bool can_be_counted_loop(PhaseTransform* phase) const {
  97     return req() == 3 && in(0) != NULL &&
  98       in(1) != NULL && phase->type(in(1)) != Type::TOP &&
  99       in(2) != NULL && phase->type(in(2)) != Type::TOP;
 100   }
 101   bool is_valid_counted_loop() const;
 102 #ifndef PRODUCT
 103   virtual void dump_spec(outputStream *st) const;
 104 #endif
 105 };
 106 
 107 //------------------------------Counted Loops----------------------------------
 108 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
 109 // path (and maybe some other exit paths).  The trip-counter exit is always
 110 // last in the loop.  The trip-counter have to stride by a constant;
 111 // the exit value is also loop invariant.
 112 
 113 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
 114 // CountedLoopNode has the incoming loop control and the loop-back-control
 115 // which is always the IfTrue before the matching CountedLoopEndNode.  The
 116 // CountedLoopEndNode has an incoming control (possibly not the
 117 // CountedLoopNode if there is control flow in the loop), the post-increment
 118 // trip-counter value, and the limit.  The trip-counter value is always of
 119 // the form (Op old-trip-counter stride).  The old-trip-counter is produced
 120 // by a Phi connected to the CountedLoopNode.  The stride is constant.
 121 // The Op is any commutable opcode, including Add, Mul, Xor.  The
 122 // CountedLoopEndNode also takes in the loop-invariant limit value.
 123 
 124 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
 125 // loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
 126 // via the old-trip-counter from the Op node.
 127 
 128 //------------------------------CountedLoopNode--------------------------------
 129 // CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
 130 // inputs the incoming loop-start control and the loop-back control, so they
 131 // act like RegionNodes.  They also take in the initial trip counter, the
 132 // loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
 133 // produce a loop-body control and the trip counter value.  Since
 134 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
 135 
 136 class CountedLoopNode : public LoopNode {
 137   // Size is bigger to hold _main_idx.  However, _main_idx does not change
 138   // the semantics so it does not appear in the hash & cmp functions.
 139   virtual uint size_of() const { return sizeof(*this); }
 140 
 141   // For Pre- and Post-loops during debugging ONLY, this holds the index of
 142   // the Main CountedLoop.  Used to assert that we understand the graph shape.
 143   node_idx_t _main_idx;
 144 
 145   // Known trip count calculated by compute_exact_trip_count()
 146   uint  _trip_count;
 147 
 148   // Expected trip count from profile data
 149   float _profile_trip_cnt;
 150 
 151   // Log2 of original loop bodies in unrolled loop
 152   int _unrolled_count_log2;
 153 
 154   // Node count prior to last unrolling - used to decide if
 155   // unroll,optimize,unroll,optimize,... is making progress
 156   int _node_count_before_unroll;
 157 
 158 public:
 159   CountedLoopNode( Node *entry, Node *backedge )
 160     : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
 161       _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
 162       _node_count_before_unroll(0) {
 163     init_class_id(Class_CountedLoop);
 164     // Initialize _trip_count to the largest possible value.
 165     // Will be reset (lower) if the loop's trip count is known.
 166   }
 167 
 168   virtual int Opcode() const;
 169   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 170 
 171   Node *init_control() const { return in(EntryControl); }
 172   Node *back_control() const { return in(LoopBackControl); }
 173   CountedLoopEndNode *loopexit() const;
 174   Node *init_trip() const;
 175   Node *stride() const;
 176   int   stride_con() const;
 177   bool  stride_is_con() const;
 178   Node *limit() const;
 179   Node *incr() const;
 180   Node *phi() const;
 181 
 182   // Match increment with optional truncation
 183   static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
 184 
 185   // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
 186   // can run short a few iterations and may start a few iterations in.
 187   // It will be RCE'd and unrolled and aligned.
 188 
 189   // A following 'post' loop will run any remaining iterations.  Used
 190   // during Range Check Elimination, the 'post' loop will do any final
 191   // iterations with full checks.  Also used by Loop Unrolling, where
 192   // the 'post' loop will do any epilog iterations needed.  Basically,
 193   // a 'post' loop can not profitably be further unrolled or RCE'd.
 194 
 195   // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
 196   // it may do under-flow checks for RCE and may do alignment iterations
 197   // so the following main loop 'knows' that it is striding down cache
 198   // lines.
 199 
 200   // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
 201   // Aligned, may be missing it's pre-loop.
 202   int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
 203   int is_pre_loop   () const { return (_loop_flags&PreMainPostFlagsMask) == Pre;    }
 204   int is_main_loop  () const { return (_loop_flags&PreMainPostFlagsMask) == Main;   }
 205   int is_post_loop  () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
 206   int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
 207   void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
 208 
 209   int main_idx() const { return _main_idx; }
 210 
 211 
 212   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
 213   void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
 214   void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
 215   void set_normal_loop(                    ) { _loop_flags &= ~PreMainPostFlagsMask; }
 216 
 217   void set_trip_count(uint tc) { _trip_count = tc; }
 218   uint trip_count()            { return _trip_count; }
 219 
 220   bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
 221   void set_exact_trip_count(uint tc) {
 222     _trip_count = tc;
 223     _loop_flags |= HasExactTripCount;
 224   }
 225   void set_nonexact_trip_count() {
 226     _loop_flags &= ~HasExactTripCount;
 227   }
 228 
 229   void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
 230   float profile_trip_cnt()             { return _profile_trip_cnt; }
 231 
 232   void double_unrolled_count() { _unrolled_count_log2++; }
 233   int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
 234 
 235   void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
 236   int  node_count_before_unroll()           { return _node_count_before_unroll; }
 237 
 238 #ifndef PRODUCT
 239   virtual void dump_spec(outputStream *st) const;
 240 #endif
 241 };
 242 
 243 //------------------------------CountedLoopEndNode-----------------------------
 244 // CountedLoopEndNodes end simple trip counted loops.  They act much like
 245 // IfNodes.
 246 class CountedLoopEndNode : public IfNode {
 247 public:
 248   enum { TestControl, TestValue };
 249 
 250   CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
 251     : IfNode( control, test, prob, cnt) {
 252     init_class_id(Class_CountedLoopEnd);
 253   }
 254   virtual int Opcode() const;
 255 
 256   Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
 257   Node *incr() const                { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 258   Node *limit() const               { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 259   Node *stride() const              { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
 260   Node *init_trip() const           { Node *tmp = phi     (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
 261   int stride_con() const;
 262   bool stride_is_con() const        { Node *tmp = stride  (); return (tmp != NULL && tmp->is_Con()); }
 263   BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
 264   PhiNode *phi() const {
 265     Node *tmp = incr();
 266     if (tmp && tmp->req() == 3) {
 267       Node* phi = tmp->in(1);
 268       if (phi->is_Phi()) {
 269         return phi->as_Phi();
 270       }
 271     }
 272     return NULL;
 273   }
 274   CountedLoopNode *loopnode() const {
 275     // The CountedLoopNode that goes with this CountedLoopEndNode may
 276     // have been optimized out by the IGVN so be cautious with the
 277     // pattern matching on the graph
 278     PhiNode* iv_phi = phi();
 279     if (iv_phi == NULL) {
 280       return NULL;
 281     }
 282     assert(iv_phi->is_Phi(), "should be PhiNode");
 283     Node *ln = iv_phi->in(0);
 284     if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
 285       return (CountedLoopNode*)ln;
 286     }
 287     return NULL;
 288   }
 289 
 290 #ifndef PRODUCT
 291   virtual void dump_spec(outputStream *st) const;
 292 #endif
 293 };
 294 
 295 
 296 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
 297   Node *bc = back_control();
 298   if( bc == NULL ) return NULL;
 299   Node *le = bc->in(0);
 300   if( le->Opcode() != Op_CountedLoopEnd )
 301     return NULL;
 302   return (CountedLoopEndNode*)le;
 303 }
 304 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
 305 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
 306 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
 307 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
 308 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
 309 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
 310 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
 311 
 312 //------------------------------LoopLimitNode-----------------------------
 313 // Counted Loop limit node which represents exact final iterator value:
 314 // trip_count = (limit - init_trip + stride - 1)/stride
 315 // final_value= trip_count * stride + init_trip.
 316 // Use HW instructions to calculate it when it can overflow in integer.
 317 // Note, final_value should fit into integer since counted loop has
 318 // limit check: limit <= max_int-stride.
 319 class LoopLimitNode : public Node {
 320   enum { Init=1, Limit=2, Stride=3 };
 321  public:
 322   LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
 323     // Put it on the Macro nodes list to optimize during macro nodes expansion.
 324     init_flags(Flag_is_macro);
 325     C->add_macro_node(this);
 326   }
 327   virtual int Opcode() const;
 328   virtual const Type *bottom_type() const { return TypeInt::INT; }
 329   virtual uint ideal_reg() const { return Op_RegI; }
 330   virtual const Type *Value( PhaseTransform *phase ) const;
 331   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
 332   virtual Node *Identity( PhaseTransform *phase );
 333 };
 334 
 335 // -----------------------------IdealLoopTree----------------------------------
 336 class IdealLoopTree : public ResourceObj {
 337 public:
 338   IdealLoopTree *_parent;       // Parent in loop tree
 339   IdealLoopTree *_next;         // Next sibling in loop tree
 340   IdealLoopTree *_child;        // First child in loop tree
 341 
 342   // The head-tail backedge defines the loop.
 343   // If tail is NULL then this loop has multiple backedges as part of the
 344   // same loop.  During cleanup I'll peel off the multiple backedges; merge
 345   // them at the loop bottom and flow 1 real backedge into the loop.
 346   Node *_head;                  // Head of loop
 347   Node *_tail;                  // Tail of loop
 348   inline Node *tail();          // Handle lazy update of _tail field
 349   PhaseIdealLoop* _phase;
 350 
 351   Node_List _body;              // Loop body for inner loops
 352 
 353   uint8 _nest;                  // Nesting depth
 354   uint8 _irreducible:1,         // True if irreducible
 355         _has_call:1,            // True if has call safepoint
 356         _has_sfpt:1,            // True if has non-call safepoint
 357         _rce_candidate:1;       // True if candidate for range check elimination
 358 
 359   Node_List* _safepts;          // List of safepoints in this loop
 360   Node_List* _required_safept;  // A inner loop cannot delete these safepts;
 361   bool  _allow_optimizations;   // Allow loop optimizations
 362 
 363   IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
 364     : _parent(0), _next(0), _child(0),
 365       _head(head), _tail(tail),
 366       _phase(phase),
 367       _safepts(NULL),
 368       _required_safept(NULL),
 369       _allow_optimizations(true),
 370       _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
 371   { }
 372 
 373   // Is 'l' a member of 'this'?
 374   int is_member( const IdealLoopTree *l ) const; // Test for nested membership
 375 
 376   // Set loop nesting depth.  Accumulate has_call bits.
 377   int set_nest( uint depth );
 378 
 379   // Split out multiple fall-in edges from the loop header.  Move them to a
 380   // private RegionNode before the loop.  This becomes the loop landing pad.
 381   void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
 382 
 383   // Split out the outermost loop from this shared header.
 384   void split_outer_loop( PhaseIdealLoop *phase );
 385 
 386   // Merge all the backedges from the shared header into a private Region.
 387   // Feed that region as the one backedge to this loop.
 388   void merge_many_backedges( PhaseIdealLoop *phase );
 389 
 390   // Split shared headers and insert loop landing pads.
 391   // Insert a LoopNode to replace the RegionNode.
 392   // Returns TRUE if loop tree is structurally changed.
 393   bool beautify_loops( PhaseIdealLoop *phase );
 394 
 395   // Perform optimization to use the loop predicates for null checks and range checks.
 396   // Applies to any loop level (not just the innermost one)
 397   bool loop_predication( PhaseIdealLoop *phase);
 398 
 399   // Perform iteration-splitting on inner loops.  Split iterations to
 400   // avoid range checks or one-shot null checks.  Returns false if the
 401   // current round of loop opts should stop.
 402   bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
 403 
 404   // Driver for various flavors of iteration splitting.  Returns false
 405   // if the current round of loop opts should stop.
 406   bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
 407 
 408   // Given dominators, try to find loops with calls that must always be
 409   // executed (call dominates loop tail).  These loops do not need non-call
 410   // safepoints (ncsfpt).
 411   void check_safepts(VectorSet &visited, Node_List &stack);
 412 
 413   // Allpaths backwards scan from loop tail, terminating each path at first safepoint
 414   // encountered.
 415   void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
 416 
 417   // Remove safepoints from loop. Optionally keeping one.
 418   void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
 419 
 420   // Convert to counted loops where possible
 421   void counted_loop( PhaseIdealLoop *phase );
 422 
 423   // Check for Node being a loop-breaking test
 424   Node *is_loop_exit(Node *iff) const;
 425 
 426   // Returns true if ctrl is executed on every complete iteration
 427   bool dominates_backedge(Node* ctrl);
 428 
 429   // Remove simplistic dead code from loop body
 430   void DCE_loop_body();
 431 
 432   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
 433   // Replace with a 1-in-10 exit guess.
 434   void adjust_loop_exit_prob( PhaseIdealLoop *phase );
 435 
 436   // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
 437   // Useful for unrolling loops with NO array accesses.
 438   bool policy_peel_only( PhaseIdealLoop *phase ) const;
 439 
 440   // Return TRUE or FALSE if the loop should be unswitched -- clone
 441   // loop with an invariant test
 442   bool policy_unswitching( PhaseIdealLoop *phase ) const;
 443 
 444   // Micro-benchmark spamming.  Remove empty loops.
 445   bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
 446 
 447   // Convert one iteration loop into normal code.
 448   bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
 449 
 450   // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
 451   // make some loop-invariant test (usually a null-check) happen before the
 452   // loop.
 453   bool policy_peeling( PhaseIdealLoop *phase ) const;
 454 
 455   // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
 456   // known trip count in the counted loop node.
 457   bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
 458 
 459   // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
 460   // the loop is a CountedLoop and the body is small enough.
 461   bool policy_unroll( PhaseIdealLoop *phase ) const;
 462 
 463   // Return TRUE or FALSE if the loop should be range-check-eliminated.
 464   // Gather a list of IF tests that are dominated by iteration splitting;
 465   // also gather the end of the first split and the start of the 2nd split.
 466   bool policy_range_check( PhaseIdealLoop *phase ) const;
 467 
 468   // Return TRUE or FALSE if the loop should be cache-line aligned.
 469   // Gather the expression that does the alignment.  Note that only
 470   // one array base can be aligned in a loop (unless the VM guarantees
 471   // mutual alignment).  Note that if we vectorize short memory ops
 472   // into longer memory ops, we may want to increase alignment.
 473   bool policy_align( PhaseIdealLoop *phase ) const;
 474 
 475   // Return TRUE if "iff" is a range check.
 476   bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
 477 
 478   // Compute loop exact trip count if possible
 479   void compute_exact_trip_count( PhaseIdealLoop *phase );
 480 
 481   // Compute loop trip count from profile data
 482   void compute_profile_trip_cnt( PhaseIdealLoop *phase );
 483 
 484   // Reassociate invariant expressions.
 485   void reassociate_invariants(PhaseIdealLoop *phase);
 486   // Reassociate invariant add and subtract expressions.
 487   Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
 488   // Return nonzero index of invariant operand if invariant and variant
 489   // are combined with an Add or Sub. Helper for reassociate_invariants.
 490   int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
 491 
 492   // Return true if n is invariant
 493   bool is_invariant(Node* n) const;
 494 
 495   // Put loop body on igvn work list
 496   void record_for_igvn();
 497 
 498   bool is_loop()    { return !_irreducible && _tail && !_tail->is_top(); }
 499   bool is_inner()   { return is_loop() && _child == NULL; }
 500   bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
 501 
 502 #ifndef PRODUCT
 503   void dump_head( ) const;      // Dump loop head only
 504   void dump() const;            // Dump this loop recursively
 505   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 506 #endif
 507 
 508 };
 509 
 510 // -----------------------------PhaseIdealLoop---------------------------------
 511 // Computes the mapping from Nodes to IdealLoopTrees.  Organizes IdealLoopTrees into a
 512 // loop tree.  Drives the loop-based transformations on the ideal graph.
 513 class PhaseIdealLoop : public PhaseTransform {
 514   friend class IdealLoopTree;
 515   friend class SuperWord;
 516   // Pre-computed def-use info
 517   PhaseIterGVN &_igvn;
 518 
 519   // Head of loop tree
 520   IdealLoopTree *_ltree_root;
 521 
 522   // Array of pre-order numbers, plus post-visited bit.
 523   // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
 524   // ODD for post-visited.  Other bits are the pre-order number.
 525   uint *_preorders;
 526   uint _max_preorder;
 527 
 528   const PhaseIdealLoop* _verify_me;
 529   bool _verify_only;
 530 
 531   // Allocate _preorders[] array
 532   void allocate_preorders() {
 533     _max_preorder = C->unique()+8;
 534     _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
 535     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 536   }
 537 
 538   // Allocate _preorders[] array
 539   void reallocate_preorders() {
 540     if ( _max_preorder < C->unique() ) {
 541       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
 542       _max_preorder = C->unique();
 543     }
 544     memset(_preorders, 0, sizeof(uint) * _max_preorder);
 545   }
 546 
 547   // Check to grow _preorders[] array for the case when build_loop_tree_impl()
 548   // adds new nodes.
 549   void check_grow_preorders( ) {
 550     if ( _max_preorder < C->unique() ) {
 551       uint newsize = _max_preorder<<1;  // double size of array
 552       _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
 553       memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
 554       _max_preorder = newsize;
 555     }
 556   }
 557   // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
 558   int is_visited( Node *n ) const { return _preorders[n->_idx]; }
 559   // Pre-order numbers are written to the Nodes array as low-bit-set values.
 560   void set_preorder_visited( Node *n, int pre_order ) {
 561     assert( !is_visited( n ), "already set" );
 562     _preorders[n->_idx] = (pre_order<<1);
 563   };
 564   // Return pre-order number.
 565   int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
 566 
 567   // Check for being post-visited.
 568   // Should be previsited already (checked with assert(is_visited(n))).
 569   int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
 570 
 571   // Mark as post visited
 572   void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
 573 
 574 public:
 575   // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
 576   // Returns true if "n" is a data node, false if it's a control node.
 577   bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
 578 
 579 private:
 580   // clear out dead code after build_loop_late
 581   Node_List _deadlist;
 582 
 583   // Support for faster execution of get_late_ctrl()/dom_lca()
 584   // when a node has many uses and dominator depth is deep.
 585   Node_Array _dom_lca_tags;
 586   void   init_dom_lca_tags();
 587   void   clear_dom_lca_tags();
 588 
 589   // Helper for debugging bad dominance relationships
 590   bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
 591 
 592   Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
 593 
 594   // Inline wrapper for frequent cases:
 595   // 1) only one use
 596   // 2) a use is the same as the current LCA passed as 'n1'
 597   Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
 598     assert( n->is_CFG(), "" );
 599     // Fast-path NULL lca
 600     if( lca != NULL && lca != n ) {
 601       assert( lca->is_CFG(), "" );
 602       // find LCA of all uses
 603       n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
 604     }
 605     return find_non_split_ctrl(n);
 606   }
 607   Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
 608 
 609   // Helper function for directing control inputs away from CFG split
 610   // points.
 611   Node *find_non_split_ctrl( Node *ctrl ) const {
 612     if (ctrl != NULL) {
 613       if (ctrl->is_MultiBranch()) {
 614         ctrl = ctrl->in(0);
 615       }
 616       assert(ctrl->is_CFG(), "CFG");
 617     }
 618     return ctrl;
 619   }
 620 
 621   bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
 622 
 623 public:
 624 
 625   static bool is_canonical_main_loop_entry(CountedLoopNode* cl);
 626 
 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_helper(Node *i) const {
 680     assert(has_ctrl(i), "should be control, not loop");
 681     return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
 682   }
 683 
 684   Node *get_ctrl_no_update(Node *i) const {
 685     assert( has_ctrl(i), "" );
 686     Node *n = get_ctrl_no_update_helper(i);
 687     if (!n->in(0)) {
 688       // Skip dead CFG nodes
 689       do {
 690         n = get_ctrl_no_update_helper(n);
 691       } while (!n->in(0));
 692       n = find_non_split_ctrl(n);
 693     }
 694     return n;
 695   }
 696 
 697   // Check for loop being set
 698   // "n" must be a control node. Returns true if "n" is known to be in a loop.
 699   bool has_loop( Node *n ) const {
 700     assert(!has_node(n) || !has_ctrl(n), "");
 701     return has_node(n);
 702   }
 703   // Set loop
 704   void set_loop( Node *n, IdealLoopTree *loop ) {
 705     _nodes.map(n->_idx, (Node*)loop);
 706   }
 707   // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
 708   // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
 709   // from old_node to new_node to support the lazy update.  Reference
 710   // replaces loop reference, since that is not needed for dead node.
 711 public:
 712   void lazy_update(Node *old_node, Node *new_node) {
 713     assert(old_node != new_node, "no cycles please");
 714     // Re-use the side array slot for this node to provide the
 715     // forwarding pointer.
 716     _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
 717   }
 718   void lazy_replace(Node *old_node, Node *new_node) {
 719     _igvn.replace_node(old_node, new_node);
 720     lazy_update(old_node, new_node);
 721   }
 722 
 723 private:
 724 
 725   // Place 'n' in some loop nest, where 'n' is a CFG node
 726   void build_loop_tree();
 727   int build_loop_tree_impl( Node *n, int pre_order );
 728   // Insert loop into the existing loop tree.  'innermost' is a leaf of the
 729   // loop tree, not the root.
 730   IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
 731 
 732   // Place Data nodes in some loop nest
 733   void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 734   void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
 735   void build_loop_late_post ( Node* n );
 736 
 737   // Array of immediate dominance info for each CFG node indexed by node idx
 738 private:
 739   uint _idom_size;
 740   Node **_idom;                 // Array of immediate dominators
 741   uint *_dom_depth;           // Used for fast LCA test
 742   GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
 743 
 744   Node* idom_no_update(Node* d) const {
 745     assert(d->_idx < _idom_size, "oob");
 746     Node* n = _idom[d->_idx];
 747     assert(n != NULL,"Bad immediate dominator info.");
 748     while (n->in(0) == NULL) {  // Skip dead CFG nodes
 749       //n = n->in(1);
 750       n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
 751       assert(n != NULL,"Bad immediate dominator info.");
 752     }
 753     return n;
 754   }
 755 public:
 756   Node *idom(Node* d) const {
 757     uint didx = d->_idx;
 758     Node *n = idom_no_update(d);
 759     _idom[didx] = n;            // Lazily remove dead CFG nodes from table.
 760     return n;
 761   }
 762   uint dom_depth(Node* d) const {
 763     guarantee(d != NULL, "Null dominator info.");
 764     guarantee(d->_idx < _idom_size, "");
 765     return _dom_depth[d->_idx];
 766   }
 767   void set_idom(Node* d, Node* n, uint dom_depth);
 768 
 769 private:
 770   // Locally compute IDOM using dom_lca call
 771   Node *compute_idom( Node *region ) const;
 772   // Recompute dom_depth
 773   void recompute_dom_depth();
 774 
 775   // Is safept not required by an outer loop?
 776   bool is_deleteable_safept(Node* sfpt);
 777 
 778   // Replace parallel induction variable (parallel to trip counter)
 779   void replace_parallel_iv(IdealLoopTree *loop);
 780 
 781   // Perform verification that the graph is valid.
 782   PhaseIdealLoop( PhaseIterGVN &igvn) :
 783     PhaseTransform(Ideal_Loop),
 784     _igvn(igvn),
 785     _dom_lca_tags(arena()), // Thread::resource_area
 786     _verify_me(NULL),
 787     _verify_only(true) {
 788     build_and_optimize(false, false);
 789   }
 790 
 791   // build the loop tree and perform any requested optimizations
 792   void build_and_optimize(bool do_split_if, bool skip_loop_opts);
 793 
 794 public:
 795   // Dominators for the sea of nodes
 796   void Dominators();
 797   Node *dom_lca( Node *n1, Node *n2 ) const {
 798     return find_non_split_ctrl(dom_lca_internal(n1, n2));
 799   }
 800   Node *dom_lca_internal( Node *n1, Node *n2 ) const;
 801 
 802   // Compute the Ideal Node to Loop mapping
 803   PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
 804     PhaseTransform(Ideal_Loop),
 805     _igvn(igvn),
 806     _dom_lca_tags(arena()), // Thread::resource_area
 807     _verify_me(NULL),
 808     _verify_only(false) {
 809     build_and_optimize(do_split_ifs, skip_loop_opts);
 810   }
 811 
 812   // Verify that verify_me made the same decisions as a fresh run.
 813   PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
 814     PhaseTransform(Ideal_Loop),
 815     _igvn(igvn),
 816     _dom_lca_tags(arena()), // Thread::resource_area
 817     _verify_me(verify_me),
 818     _verify_only(false) {
 819     build_and_optimize(false, false);
 820   }
 821 
 822   // Build and verify the loop tree without modifying the graph.  This
 823   // is useful to verify that all inputs properly dominate their uses.
 824   static void verify(PhaseIterGVN& igvn) {
 825 #ifdef ASSERT
 826     PhaseIdealLoop v(igvn);
 827 #endif
 828   }
 829 
 830   // True if the method has at least 1 irreducible loop
 831   bool _has_irreducible_loops;
 832 
 833   // Per-Node transform
 834   virtual Node *transform( Node *a_node ) { return 0; }
 835 
 836   bool is_counted_loop( Node *x, IdealLoopTree *loop );
 837 
 838   Node* exact_limit( IdealLoopTree *loop );
 839 
 840   // Return a post-walked LoopNode
 841   IdealLoopTree *get_loop( Node *n ) const {
 842     // Dead nodes have no loop, so return the top level loop instead
 843     if (!has_node(n))  return _ltree_root;
 844     assert(!has_ctrl(n), "");
 845     return (IdealLoopTree*)_nodes[n->_idx];
 846   }
 847 
 848   // Is 'n' a (nested) member of 'loop'?
 849   int is_member( const IdealLoopTree *loop, Node *n ) const {
 850     return loop->is_member(get_loop(n)); }
 851 
 852   // This is the basic building block of the loop optimizations.  It clones an
 853   // entire loop body.  It makes an old_new loop body mapping; with this
 854   // mapping you can find the new-loop equivalent to an old-loop node.  All
 855   // new-loop nodes are exactly equal to their old-loop counterparts, all
 856   // edges are the same.  All exits from the old-loop now have a RegionNode
 857   // that merges the equivalent new-loop path.  This is true even for the
 858   // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
 859   // now come from (one or more) Phis that merge their new-loop equivalents.
 860   // Parameter side_by_side_idom:
 861   //   When side_by_size_idom is NULL, the dominator tree is constructed for
 862   //      the clone loop to dominate the original.  Used in construction of
 863   //      pre-main-post loop sequence.
 864   //   When nonnull, the clone and original are side-by-side, both are
 865   //      dominated by the passed in side_by_side_idom node.  Used in
 866   //      construction of unswitched loops.
 867   void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
 868                    Node* side_by_side_idom = NULL);
 869 
 870   // If we got the effect of peeling, either by actually peeling or by
 871   // making a pre-loop which must execute at least once, we can remove
 872   // all loop-invariant dominated tests in the main body.
 873   void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
 874 
 875   // Generate code to do a loop peel for the given loop (and body).
 876   // old_new is a temp array.
 877   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
 878 
 879   // Add pre and post loops around the given loop.  These loops are used
 880   // during RCE, unrolling and aligning loops.
 881   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
 882   // If Node n lives in the back_ctrl block, we clone a private version of n
 883   // in preheader_ctrl block and return that, otherwise return n.
 884   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
 885 
 886   // Take steps to maximally unroll the loop.  Peel any odd iterations, then
 887   // unroll to do double iterations.  The next round of major loop transforms
 888   // will repeat till the doubled loop body does all remaining iterations in 1
 889   // pass.
 890   void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
 891 
 892   // Unroll the loop body one step - make each trip do 2 iterations.
 893   void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
 894 
 895   // Return true if exp is a constant times an induction var
 896   bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
 897 
 898   // Return true if exp is a scaled induction var plus (or minus) constant
 899   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
 900 
 901   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
 902   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 903                                         Deoptimization::DeoptReason reason);
 904   void register_control(Node* n, IdealLoopTree *loop, Node* pred);
 905 
 906   // Clone loop predicates to cloned loops (peeled, unswitched)
 907   static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 908                                    Deoptimization::DeoptReason reason,
 909                                    PhaseIdealLoop* loop_phase,
 910                                    PhaseIterGVN* igvn);
 911 
 912   static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
 913                                          bool clone_limit_check,
 914                                          PhaseIdealLoop* loop_phase,
 915                                          PhaseIterGVN* igvn);
 916   Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
 917 
 918   static Node* skip_loop_predicates(Node* entry);
 919 
 920   // Find a good location to insert a predicate
 921   static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
 922   // Find a predicate
 923   static Node* find_predicate(Node* entry);
 924   // Construct a range check for a predicate if
 925   BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
 926                          int scale, Node* offset,
 927                          Node* init, Node* limit, jint stride,
 928                          Node* range, bool upper, bool &overflow);
 929 
 930   // Implementation of the loop predication to promote checks outside the loop
 931   bool loop_predication_impl(IdealLoopTree *loop);
 932 
 933   // Helper function to collect predicate for eliminating the useless ones
 934   void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
 935   void eliminate_useless_predicates();
 936 
 937   // Change the control input of expensive nodes to allow commoning by
 938   // IGVN when it is guaranteed to not result in a more frequent
 939   // execution of the expensive node. Return true if progress.
 940   bool process_expensive_nodes();
 941 
 942   // Check whether node has become unreachable
 943   bool is_node_unreachable(Node *n) const {
 944     return !has_node(n) || n->is_unreachable(_igvn);
 945   }
 946 
 947   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
 948   void do_range_check( IdealLoopTree *loop, Node_List &old_new );
 949 
 950   // Create a slow version of the loop by cloning the loop
 951   // and inserting an if to select fast-slow versions.
 952   ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
 953                                         Node_List &old_new);
 954 
 955   // Clone loop with an invariant test (that does not exit) and
 956   // insert a clone of the test that selects which version to
 957   // execute.
 958   void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
 959 
 960   // Find candidate "if" for unswitching
 961   IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
 962 
 963   // Range Check Elimination uses this function!
 964   // Constrain the main loop iterations so the affine function:
 965   //    low_limit <= scale_con * I + offset  <  upper_limit
 966   // always holds true.  That is, either increase the number of iterations in
 967   // the pre-loop or the post-loop until the condition holds true in the main
 968   // loop.  Scale_con, offset and limit are all loop invariant.
 969   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 );
 970   // Helper function for add_constraint().
 971   Node* adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up);
 972 
 973   // Partially peel loop up through last_peel node.
 974   bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
 975 
 976   // Create a scheduled list of nodes control dependent on ctrl set.
 977   void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
 978   // Has a use in the vector set
 979   bool has_use_in_set( Node* n, VectorSet& vset );
 980   // Has use internal to the vector set (ie. not in a phi at the loop head)
 981   bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
 982   // clone "n" for uses that are outside of loop
 983   int  clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
 984   // clone "n" for special uses that are in the not_peeled region
 985   void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
 986                                           VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
 987   // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
 988   void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
 989 #ifdef ASSERT
 990   // Validate the loop partition sets: peel and not_peel
 991   bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
 992   // Ensure that uses outside of loop are of the right form
 993   bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
 994                                  uint orig_exit_idx, uint clone_exit_idx);
 995   bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
 996 #endif
 997 
 998   // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
 999   int stride_of_possible_iv( Node* iff );
1000   bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1001   // Return the (unique) control output node that's in the loop (if it exists.)
1002   Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1003   // Insert a signed compare loop exit cloned from an unsigned compare.
1004   IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1005   void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1006   // Utility to register node "n" with PhaseIdealLoop
1007   void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1008   // Utility to create an if-projection
1009   ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1010   // Force the iff control output to be the live_proj
1011   Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1012   // Insert a region before an if projection
1013   RegionNode* insert_region_before_proj(ProjNode* proj);
1014   // Insert a new if before an if projection
1015   ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1016 
1017   // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1018   // "Nearly" because all Nodes have been cloned from the original in the loop,
1019   // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1020   // through the Phi recursively, and return a Bool.
1021   BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1022   CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1023 
1024 
1025   // Rework addressing expressions to get the most loop-invariant stuff
1026   // moved out.  We'd like to do all associative operators, but it's especially
1027   // important (common) to do address expressions.
1028   Node *remix_address_expressions( Node *n );
1029 
1030   // Attempt to use a conditional move instead of a phi/branch
1031   Node *conditional_move( Node *n );
1032 
1033   // Reorganize offset computations to lower register pressure.
1034   // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1035   // (which are then alive with the post-incremented trip counter
1036   // forcing an extra register move)
1037   void reorg_offsets( IdealLoopTree *loop );
1038 
1039   // Check for aggressive application of 'split-if' optimization,
1040   // using basic block level info.
1041   void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack );
1042   Node *split_if_with_blocks_pre ( Node *n );
1043   void  split_if_with_blocks_post( Node *n );
1044   Node *has_local_phi_input( Node *n );
1045   // Mark an IfNode as being dominated by a prior test,
1046   // without actually altering the CFG (and hence IDOM info).
1047   void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1048 
1049   // Split Node 'n' through merge point
1050   Node *split_thru_region( Node *n, Node *region );
1051   // Split Node 'n' through merge point if there is enough win.
1052   Node *split_thru_phi( Node *n, Node *region, int policy );
1053 
1054   // Found an If getting its condition-code input from a Phi in the
1055   // same block.  Split thru the Region.
1056   void do_split_if( Node *iff );
1057 
1058   // Conversion of fill/copy patterns into intrisic versions
1059   bool do_intrinsify_fill();
1060   bool intrinsify_fill(IdealLoopTree* lpt);
1061   bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1062                        Node*& shift, Node*& offset);
1063 
1064 private:
1065   // Return a type based on condition control flow
1066   const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1067   const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1068  // Helpers for filtered type
1069   const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1070 
1071   // Helper functions
1072   Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1073   Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1074   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 );
1075   bool split_up( Node *n, Node *blk1, Node *blk2 );
1076   void sink_use( Node *use, Node *post_loop );
1077   Node *place_near_use( Node *useblock ) const;
1078 
1079   bool _created_loop_node;
1080 public:
1081   void set_created_loop_node() { _created_loop_node = true; }
1082   bool created_loop_node()     { return _created_loop_node; }
1083   void register_new_node( Node *n, Node *blk );
1084 
1085 #ifdef ASSERT
1086   void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1087 #endif
1088   void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1089 
1090 #ifndef PRODUCT
1091   void dump( ) const;
1092   void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1093   void verify() const;          // Major slow  :-)
1094   void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1095   IdealLoopTree *get_loop_idx(Node* n) const {
1096     // Dead nodes have no loop, so return the top level loop instead
1097     return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1098   }
1099   // Print some stats
1100   static void print_statistics();
1101   static int _loop_invokes;     // Count of PhaseIdealLoop invokes
1102   static int _loop_work;        // Sum of PhaseIdealLoop x _unique
1103 #endif
1104 
1105   PhaseIterGVN& igvn() { return _igvn; }
1106   IdealLoopTree* ltree_root() const { return _ltree_root; }
1107 };
1108 
1109 inline Node* IdealLoopTree::tail() {
1110 // Handle lazy update of _tail field
1111   Node *n = _tail;
1112   //while( !n->in(0) )  // Skip dead CFG nodes
1113     //n = n->in(1);
1114   if (n->in(0) == NULL)
1115     n = _phase->get_ctrl(n);
1116   _tail = n;
1117   return n;
1118 }
1119 
1120 
1121 // Iterate over the loop tree using a preorder, left-to-right traversal.
1122 //
1123 // Example that visits all counted loops from within PhaseIdealLoop
1124 //
1125 //  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1126 //   IdealLoopTree* lpt = iter.current();
1127 //   if (!lpt->is_counted()) continue;
1128 //   ...
1129 class LoopTreeIterator : public StackObj {
1130 private:
1131   IdealLoopTree* _root;
1132   IdealLoopTree* _curnt;
1133 
1134 public:
1135   LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1136 
1137   bool done() { return _curnt == NULL; }       // Finished iterating?
1138 
1139   void next();                                 // Advance to next loop tree
1140 
1141   IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
1142 };
1143 
1144 #endif // SHARE_VM_OPTO_LOOPNODE_HPP