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