loopnode.hpp 44.4 KB
Newer Older
D
duke 已提交
1
/*
2
 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
D
duke 已提交
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This code is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 only, as
 * published by the Free Software Foundation.
 *
 * This code is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 * version 2 for more details (a copy is included in the LICENSE file that
 * accompanied this code).
 *
 * You should have received a copy of the GNU General Public License version
 * 2 along with this work; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 *
19 20 21
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 * or visit www.oracle.com if you need additional information or have any
 * questions.
D
duke 已提交
22 23 24
 *
 */

25 26 27 28 29 30 31 32 33
#ifndef SHARE_VM_OPTO_LOOPNODE_HPP
#define SHARE_VM_OPTO_LOOPNODE_HPP

#include "opto/cfgnode.hpp"
#include "opto/multnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
#include "opto/type.hpp"

D
duke 已提交
34 35 36 37 38 39 40 41
class CmpNode;
class CountedLoopEndNode;
class CountedLoopNode;
class IdealLoopTree;
class LoopNode;
class Node;
class PhaseIdealLoop;
class VectorSet;
42
class Invariance;
D
duke 已提交
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
struct small_cache;

//
//                  I D E A L I Z E D   L O O P S
//
// Idealized loops are the set of loops I perform more interesting
// transformations on, beyond simple hoisting.

//------------------------------LoopNode---------------------------------------
// Simple loop header.  Fall in path on left, loop-back path on right.
class LoopNode : public RegionNode {
  // Size is bigger to hold the flags.  However, the flags do not change
  // the semantics so it does not appear in the hash & cmp functions.
  virtual uint size_of() const { return sizeof(*this); }
protected:
  short _loop_flags;
  // Names for flag bitfields
  enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32  };
  char _unswitch_count;
  enum { _unswitch_max=3 };

public:
  // Names for edge indices
  enum { Self=0, EntryControl, LoopBackControl };

  int is_inner_loop() const { return _loop_flags & inner_loop; }
  void set_inner_loop() { _loop_flags |= inner_loop; }

  int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
  void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
  int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
  void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }

  int unswitch_max() { return _unswitch_max; }
  int unswitch_count() { return _unswitch_count; }
  void set_unswitch_count(int val) {
    assert (val <= unswitch_max(), "too many unswitches");
    _unswitch_count = val;
  }

  LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
    init_class_id(Class_Loop);
    init_req(EntryControl, entry);
    init_req(LoopBackControl, backedge);
  }

  virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
  virtual int Opcode() const;
  bool can_be_counted_loop(PhaseTransform* phase) const {
    return req() == 3 && in(0) != NULL &&
      in(1) != NULL && phase->type(in(1)) != Type::TOP &&
      in(2) != NULL && phase->type(in(2)) != Type::TOP;
  }
96
  bool is_valid_counted_loop() const;
D
duke 已提交
97 98 99 100 101 102 103 104
#ifndef PRODUCT
  virtual void dump_spec(outputStream *st) const;
#endif
};

//------------------------------Counted Loops----------------------------------
// Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
// path (and maybe some other exit paths).  The trip-counter exit is always
105 106
// last in the loop.  The trip-counter have to stride by a constant;
// the exit value is also loop invariant.
D
duke 已提交
107 108 109 110 111 112 113 114

// CountedLoopNodes and CountedLoopEndNodes come in matched pairs.  The
// CountedLoopNode has the incoming loop control and the loop-back-control
// which is always the IfTrue before the matching CountedLoopEndNode.  The
// CountedLoopEndNode has an incoming control (possibly not the
// CountedLoopNode if there is control flow in the loop), the post-increment
// trip-counter value, and the limit.  The trip-counter value is always of
// the form (Op old-trip-counter stride).  The old-trip-counter is produced
115
// by a Phi connected to the CountedLoopNode.  The stride is constant.
D
duke 已提交
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
// The Op is any commutable opcode, including Add, Mul, Xor.  The
// CountedLoopEndNode also takes in the loop-invariant limit value.

// From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
// loop-back control.  From CountedLoopEndNodes I can reach CountedLoopNodes
// via the old-trip-counter from the Op node.

//------------------------------CountedLoopNode--------------------------------
// CountedLoopNodes head simple counted loops.  CountedLoopNodes have as
// inputs the incoming loop-start control and the loop-back control, so they
// act like RegionNodes.  They also take in the initial trip counter, the
// loop-invariant stride and the loop-invariant limit value.  CountedLoopNodes
// produce a loop-body control and the trip counter value.  Since
// CountedLoopNodes behave like RegionNodes I still have a standard CFG model.

class CountedLoopNode : public LoopNode {
  // Size is bigger to hold _main_idx.  However, _main_idx does not change
  // the semantics so it does not appear in the hash & cmp functions.
  virtual uint size_of() const { return sizeof(*this); }

  // For Pre- and Post-loops during debugging ONLY, this holds the index of
  // the Main CountedLoop.  Used to assert that we understand the graph shape.
  node_idx_t _main_idx;

  // Known trip count calculated by policy_maximally_unroll
  int   _trip_count;

  // Expected trip count from profile data
  float _profile_trip_cnt;

  // Log2 of original loop bodies in unrolled loop
  int _unrolled_count_log2;

  // Node count prior to last unrolling - used to decide if
  // unroll,optimize,unroll,optimize,... is making progress
  int _node_count_before_unroll;

public:
  CountedLoopNode( Node *entry, Node *backedge )
    : LoopNode(entry, backedge), _trip_count(max_jint),
      _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
      _node_count_before_unroll(0) {
    init_class_id(Class_CountedLoop);
    // Initialize _trip_count to the largest possible value.
    // Will be reset (lower) if the loop's trip count is known.
  }

  virtual int Opcode() const;
  virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);

  Node *init_control() const { return in(EntryControl); }
  Node *back_control() const { return in(LoopBackControl); }
  CountedLoopEndNode *loopexit() const;
  Node *init_trip() const;
  Node *stride() const;
  int   stride_con() const;
  bool  stride_is_con() const;
  Node *limit() const;
  Node *incr() const;
  Node *phi() const;

  // Match increment with optional truncation
  static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);

  // A 'main' loop has a pre-loop and a post-loop.  The 'main' loop
  // can run short a few iterations and may start a few iterations in.
  // It will be RCE'd and unrolled and aligned.

  // A following 'post' loop will run any remaining iterations.  Used
  // during Range Check Elimination, the 'post' loop will do any final
  // iterations with full checks.  Also used by Loop Unrolling, where
  // the 'post' loop will do any epilog iterations needed.  Basically,
  // a 'post' loop can not profitably be further unrolled or RCE'd.

  // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
  // it may do under-flow checks for RCE and may do alignment iterations
  // so the following main loop 'knows' that it is striding down cache
  // lines.

  // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
  // Aligned, may be missing it's pre-loop.
  enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
  int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
  int is_pre_loop   () const { return (_loop_flags&PrePostFlagsMask) == Pre;    }
  int is_main_loop  () const { return (_loop_flags&PrePostFlagsMask) == Main;   }
  int is_post_loop  () const { return (_loop_flags&PrePostFlagsMask) == Post;   }
  int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
  void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }

205 206
  int main_idx() const { return _main_idx; }

D
duke 已提交
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303

  void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
  void set_main_loop (                     ) { assert(is_normal_loop(),""); _loop_flags |= Main;                         }
  void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
  void set_normal_loop(                    ) { _loop_flags &= ~PrePostFlagsMask; }

  void set_trip_count(int tc) { _trip_count = tc; }
  int trip_count()            { return _trip_count; }

  void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
  float profile_trip_cnt()             { return _profile_trip_cnt; }

  void double_unrolled_count() { _unrolled_count_log2++; }
  int  unrolled_count()        { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }

  void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
  int  node_count_before_unroll()           { return _node_count_before_unroll; }

#ifndef PRODUCT
  virtual void dump_spec(outputStream *st) const;
#endif
};

//------------------------------CountedLoopEndNode-----------------------------
// CountedLoopEndNodes end simple trip counted loops.  They act much like
// IfNodes.
class CountedLoopEndNode : public IfNode {
public:
  enum { TestControl, TestValue };

  CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
    : IfNode( control, test, prob, cnt) {
    init_class_id(Class_CountedLoopEnd);
  }
  virtual int Opcode() const;

  Node *cmp_node() const            { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
  Node *incr() const                { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
  Node *limit() const               { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
  Node *stride() const              { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
  Node *phi() const                 { Node *tmp = incr    (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
  Node *init_trip() const           { Node *tmp = phi     (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
  int stride_con() const;
  bool stride_is_con() const        { Node *tmp = stride  (); return (tmp != NULL && tmp->is_Con()); }
  BoolTest::mask test_trip() const  { return in(TestValue)->as_Bool()->_test._test; }
  CountedLoopNode *loopnode() const {
    Node *ln = phi()->in(0);
    assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
    return (CountedLoopNode*)ln; }

#ifndef PRODUCT
  virtual void dump_spec(outputStream *st) const;
#endif
};


inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
  Node *bc = back_control();
  if( bc == NULL ) return NULL;
  Node *le = bc->in(0);
  if( le->Opcode() != Op_CountedLoopEnd )
    return NULL;
  return (CountedLoopEndNode*)le;
}
inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }


// -----------------------------IdealLoopTree----------------------------------
class IdealLoopTree : public ResourceObj {
public:
  IdealLoopTree *_parent;       // Parent in loop tree
  IdealLoopTree *_next;         // Next sibling in loop tree
  IdealLoopTree *_child;        // First child in loop tree

  // The head-tail backedge defines the loop.
  // If tail is NULL then this loop has multiple backedges as part of the
  // same loop.  During cleanup I'll peel off the multiple backedges; merge
  // them at the loop bottom and flow 1 real backedge into the loop.
  Node *_head;                  // Head of loop
  Node *_tail;                  // Tail of loop
  inline Node *tail();          // Handle lazy update of _tail field
  PhaseIdealLoop* _phase;

  Node_List _body;              // Loop body for inner loops

  uint8 _nest;                  // Nesting depth
  uint8 _irreducible:1,         // True if irreducible
        _has_call:1,            // True if has call safepoint
        _has_sfpt:1,            // True if has non-call safepoint
        _rce_candidate:1;       // True if candidate for range check elimination

304 305
  Node_List* _required_safept;  // A inner loop cannot delete these safepts;
  bool  _allow_optimizations;   // Allow loop optimizations
D
duke 已提交
306 307 308 309 310 311

  IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
    : _parent(0), _next(0), _child(0),
      _head(head), _tail(tail),
      _phase(phase),
      _required_safept(NULL),
312
      _allow_optimizations(true),
D
duke 已提交
313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337
      _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
  { }

  // Is 'l' a member of 'this'?
  int is_member( const IdealLoopTree *l ) const; // Test for nested membership

  // Set loop nesting depth.  Accumulate has_call bits.
  int set_nest( uint depth );

  // Split out multiple fall-in edges from the loop header.  Move them to a
  // private RegionNode before the loop.  This becomes the loop landing pad.
  void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );

  // Split out the outermost loop from this shared header.
  void split_outer_loop( PhaseIdealLoop *phase );

  // Merge all the backedges from the shared header into a private Region.
  // Feed that region as the one backedge to this loop.
  void merge_many_backedges( PhaseIdealLoop *phase );

  // Split shared headers and insert loop landing pads.
  // Insert a LoopNode to replace the RegionNode.
  // Returns TRUE if loop tree is structurally changed.
  bool beautify_loops( PhaseIdealLoop *phase );

338 339 340 341
  // Perform optimization to use the loop predicates for null checks and range checks.
  // Applies to any loop level (not just the innermost one)
  bool loop_predication( PhaseIdealLoop *phase);

342 343 344 345 346 347 348 349
  // Perform iteration-splitting on inner loops.  Split iterations to
  // avoid range checks or one-shot null checks.  Returns false if the
  // current round of loop opts should stop.
  bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );

  // Driver for various flavors of iteration splitting.  Returns false
  // if the current round of loop opts should stop.
  bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
D
duke 已提交
350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406

  // Given dominators, try to find loops with calls that must always be
  // executed (call dominates loop tail).  These loops do not need non-call
  // safepoints (ncsfpt).
  void check_safepts(VectorSet &visited, Node_List &stack);

  // Allpaths backwards scan from loop tail, terminating each path at first safepoint
  // encountered.
  void allpaths_check_safepts(VectorSet &visited, Node_List &stack);

  // Convert to counted loops where possible
  void counted_loop( PhaseIdealLoop *phase );

  // Check for Node being a loop-breaking test
  Node *is_loop_exit(Node *iff) const;

  // Returns true if ctrl is executed on every complete iteration
  bool dominates_backedge(Node* ctrl);

  // Remove simplistic dead code from loop body
  void DCE_loop_body();

  // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
  // Replace with a 1-in-10 exit guess.
  void adjust_loop_exit_prob( PhaseIdealLoop *phase );

  // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
  // Useful for unrolling loops with NO array accesses.
  bool policy_peel_only( PhaseIdealLoop *phase ) const;

  // Return TRUE or FALSE if the loop should be unswitched -- clone
  // loop with an invariant test
  bool policy_unswitching( PhaseIdealLoop *phase ) const;

  // Micro-benchmark spamming.  Remove empty loops.
  bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );

  // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
  // make some loop-invariant test (usually a null-check) happen before the
  // loop.
  bool policy_peeling( PhaseIdealLoop *phase ) const;

  // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
  // known trip count in the counted loop node.
  bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;

  // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
  // the loop is a CountedLoop and the body is small enough.
  bool policy_unroll( PhaseIdealLoop *phase ) const;

  // Return TRUE or FALSE if the loop should be range-check-eliminated.
  // Gather a list of IF tests that are dominated by iteration splitting;
  // also gather the end of the first split and the start of the 2nd split.
  bool policy_range_check( PhaseIdealLoop *phase ) const;

  // Return TRUE or FALSE if the loop should be cache-line aligned.
  // Gather the expression that does the alignment.  Note that only
T
twisti 已提交
407
  // one array base can be aligned in a loop (unless the VM guarantees
D
duke 已提交
408 409 410 411
  // mutual alignment).  Note that if we vectorize short memory ops
  // into longer memory ops, we may want to increase alignment.
  bool policy_align( PhaseIdealLoop *phase ) const;

412 413 414
  // Return TRUE if "iff" is a range check.
  bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;

D
duke 已提交
415 416 417 418 419 420 421 422
  // Compute loop trip count from profile data
  void compute_profile_trip_cnt( PhaseIdealLoop *phase );

  // Reassociate invariant expressions.
  void reassociate_invariants(PhaseIdealLoop *phase);
  // Reassociate invariant add and subtract expressions.
  Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
  // Return nonzero index of invariant operand if invariant and variant
T
twisti 已提交
423
  // are combined with an Add or Sub. Helper for reassociate_invariants.
D
duke 已提交
424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
  int is_invariant_addition(Node* n, PhaseIdealLoop *phase);

  // Return true if n is invariant
  bool is_invariant(Node* n) const;

  // Put loop body on igvn work list
  void record_for_igvn();

  bool is_loop()    { return !_irreducible && _tail && !_tail->is_top(); }
  bool is_inner()   { return is_loop() && _child == NULL; }
  bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }

#ifndef PRODUCT
  void dump_head( ) const;      // Dump loop head only
  void dump() const;            // Dump this loop recursively
  void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
#endif

};

// -----------------------------PhaseIdealLoop---------------------------------
// Computes the mapping from Nodes to IdealLoopTrees.  Organizes IdealLoopTrees into a
// loop tree.  Drives the loop-based transformations on the ideal graph.
class PhaseIdealLoop : public PhaseTransform {
  friend class IdealLoopTree;
  friend class SuperWord;
  // Pre-computed def-use info
  PhaseIterGVN &_igvn;

  // Head of loop tree
  IdealLoopTree *_ltree_root;

  // Array of pre-order numbers, plus post-visited bit.
  // ZERO for not pre-visited.  EVEN for pre-visited but not post-visited.
  // ODD for post-visited.  Other bits are the pre-order number.
  uint *_preorders;
  uint _max_preorder;

462 463 464
  const PhaseIdealLoop* _verify_me;
  bool _verify_only;

D
duke 已提交
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
  // Allocate _preorders[] array
  void allocate_preorders() {
    _max_preorder = C->unique()+8;
    _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
    memset(_preorders, 0, sizeof(uint) * _max_preorder);
  }

  // Allocate _preorders[] array
  void reallocate_preorders() {
    if ( _max_preorder < C->unique() ) {
      _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
      _max_preorder = C->unique();
    }
    memset(_preorders, 0, sizeof(uint) * _max_preorder);
  }

  // Check to grow _preorders[] array for the case when build_loop_tree_impl()
  // adds new nodes.
  void check_grow_preorders( ) {
    if ( _max_preorder < C->unique() ) {
      uint newsize = _max_preorder<<1;  // double size of array
      _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
      memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
      _max_preorder = newsize;
    }
  }
  // Check for pre-visited.  Zero for NOT visited; non-zero for visited.
  int is_visited( Node *n ) const { return _preorders[n->_idx]; }
  // Pre-order numbers are written to the Nodes array as low-bit-set values.
  void set_preorder_visited( Node *n, int pre_order ) {
    assert( !is_visited( n ), "already set" );
    _preorders[n->_idx] = (pre_order<<1);
  };
  // Return pre-order number.
  int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }

  // Check for being post-visited.
  // Should be previsited already (checked with assert(is_visited(n))).
  int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }

  // Mark as post visited
  void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }

  // Set/get control node out.  Set lower bit to distinguish from IdealLoopTree
  // Returns true if "n" is a data node, false if it's a control node.
  bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }

  // clear out dead code after build_loop_late
  Node_List _deadlist;

  // Support for faster execution of get_late_ctrl()/dom_lca()
  // when a node has many uses and dominator depth is deep.
  Node_Array _dom_lca_tags;
  void   init_dom_lca_tags();
  void   clear_dom_lca_tags();
520 521 522 523 524 525

  // Helper for debugging bad dominance relationships
  bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);

  Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);

D
duke 已提交
526 527 528 529 530 531 532 533 534 535 536 537 538 539
  // Inline wrapper for frequent cases:
  // 1) only one use
  // 2) a use is the same as the current LCA passed as 'n1'
  Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
    assert( n->is_CFG(), "" );
    // Fast-path NULL lca
    if( lca != NULL && lca != n ) {
      assert( lca->is_CFG(), "" );
      // find LCA of all uses
      n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
    }
    return find_non_split_ctrl(n);
  }
  Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
540

D
duke 已提交
541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588
  // Helper function for directing control inputs away from CFG split
  // points.
  Node *find_non_split_ctrl( Node *ctrl ) const {
    if (ctrl != NULL) {
      if (ctrl->is_MultiBranch()) {
        ctrl = ctrl->in(0);
      }
      assert(ctrl->is_CFG(), "CFG");
    }
    return ctrl;
  }

public:
  bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
  // check if transform created new nodes that need _ctrl recorded
  Node *get_late_ctrl( Node *n, Node *early );
  Node *get_early_ctrl( Node *n );
  void set_early_ctrl( Node *n );
  void set_subtree_ctrl( Node *root );
  void set_ctrl( Node *n, Node *ctrl ) {
    assert( !has_node(n) || has_ctrl(n), "" );
    assert( ctrl->in(0), "cannot set dead control node" );
    assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
    _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
  }
  // Set control and update loop membership
  void set_ctrl_and_loop(Node* n, Node* ctrl) {
    IdealLoopTree* old_loop = get_loop(get_ctrl(n));
    IdealLoopTree* new_loop = get_loop(ctrl);
    if (old_loop != new_loop) {
      if (old_loop->_child == NULL) old_loop->_body.yank(n);
      if (new_loop->_child == NULL) new_loop->_body.push(n);
    }
    set_ctrl(n, ctrl);
  }
  // Control nodes can be replaced or subsumed.  During this pass they
  // get their replacement Node in slot 1.  Instead of updating the block
  // location of all Nodes in the subsumed block, we lazily do it.  As we
  // pull such a subsumed block out of the array, we write back the final
  // correct block.
  Node *get_ctrl( Node *i ) {
    assert(has_node(i), "");
    Node *n = get_ctrl_no_update(i);
    _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
    assert(has_node(i) && has_ctrl(i), "");
    assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
    return n;
  }
589 590 591 592 593 594 595 596 597 598 599
  // true if CFG node d dominates CFG node n
  bool is_dominator(Node *d, Node *n);
  // return get_ctrl for a data node and self(n) for a CFG node
  Node* ctrl_or_self(Node* n) {
    if (has_ctrl(n))
      return get_ctrl(n);
    else {
      assert (n->is_CFG(), "must be a CFG node");
      return n;
    }
  }
D
duke 已提交
600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627

private:
  Node *get_ctrl_no_update( Node *i ) const {
    assert( has_ctrl(i), "" );
    Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
    if (!n->in(0)) {
      // Skip dead CFG nodes
      do {
        n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
      } while (!n->in(0));
      n = find_non_split_ctrl(n);
    }
    return n;
  }

  // Check for loop being set
  // "n" must be a control node. Returns true if "n" is known to be in a loop.
  bool has_loop( Node *n ) const {
    assert(!has_node(n) || !has_ctrl(n), "");
    return has_node(n);
  }
  // Set loop
  void set_loop( Node *n, IdealLoopTree *loop ) {
    _nodes.map(n->_idx, (Node*)loop);
  }
  // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms.  Replace
  // the 'old_node' with 'new_node'.  Kill old-node.  Add a reference
  // from old_node to new_node to support the lazy update.  Reference
628
  // replaces loop reference, since that is not needed for dead node.
D
duke 已提交
629 630 631 632 633 634 635 636 637
public:
  void lazy_update( Node *old_node, Node *new_node ) {
    assert( old_node != new_node, "no cycles please" );
    //old_node->set_req( 1, new_node /*NO DU INFO*/ );
    // Nodes always have DU info now, so re-use the side array slot
    // for this node to provide the forwarding pointer.
    _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
  }
  void lazy_replace( Node *old_node, Node *new_node ) {
638
    _igvn.replace_node( old_node, new_node );
D
duke 已提交
639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657
    lazy_update( old_node, new_node );
  }
  void lazy_replace_proj( Node *old_node, Node *new_node ) {
    assert( old_node->req() == 1, "use this for Projs" );
    _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
    old_node->add_req( NULL );
    lazy_replace( old_node, new_node );
  }

private:

  // Place 'n' in some loop nest, where 'n' is a CFG node
  void build_loop_tree();
  int build_loop_tree_impl( Node *n, int pre_order );
  // Insert loop into the existing loop tree.  'innermost' is a leaf of the
  // loop tree, not the root.
  IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );

  // Place Data nodes in some loop nest
658 659 660
  void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
  void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
  void build_loop_late_post ( Node* n );
D
duke 已提交
661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698

  // Array of immediate dominance info for each CFG node indexed by node idx
private:
  uint _idom_size;
  Node **_idom;                 // Array of immediate dominators
  uint *_dom_depth;           // Used for fast LCA test
  GrowableArray<uint>* _dom_stk; // For recomputation of dom depth

  Node* idom_no_update(Node* d) const {
    assert(d->_idx < _idom_size, "oob");
    Node* n = _idom[d->_idx];
    assert(n != NULL,"Bad immediate dominator info.");
    while (n->in(0) == NULL) {  // Skip dead CFG nodes
      //n = n->in(1);
      n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
      assert(n != NULL,"Bad immediate dominator info.");
    }
    return n;
  }
  Node *idom(Node* d) const {
    uint didx = d->_idx;
    Node *n = idom_no_update(d);
    _idom[didx] = n;            // Lazily remove dead CFG nodes from table.
    return n;
  }
  uint dom_depth(Node* d) const {
    assert(d->_idx < _idom_size, "");
    return _dom_depth[d->_idx];
  }
  void set_idom(Node* d, Node* n, uint dom_depth);
  // Locally compute IDOM using dom_lca call
  Node *compute_idom( Node *region ) const;
  // Recompute dom_depth
  void recompute_dom_depth();

  // Is safept not required by an outer loop?
  bool is_deleteable_safept(Node* sfpt);

699 700 701
  // Replace parallel induction variable (parallel to trip counter)
  void replace_parallel_iv(IdealLoopTree *loop);

702 703 704 705
  // Perform verification that the graph is valid.
  PhaseIdealLoop( PhaseIterGVN &igvn) :
    PhaseTransform(Ideal_Loop),
    _igvn(igvn),
706
    _dom_lca_tags(arena()), // Thread::resource_area
707 708
    _verify_me(NULL),
    _verify_only(true) {
709
    build_and_optimize(false);
710 711 712
  }

  // build the loop tree and perform any requested optimizations
713
  void build_and_optimize(bool do_split_if);
714

D
duke 已提交
715 716 717 718 719 720 721 722 723
public:
  // Dominators for the sea of nodes
  void Dominators();
  Node *dom_lca( Node *n1, Node *n2 ) const {
    return find_non_split_ctrl(dom_lca_internal(n1, n2));
  }
  Node *dom_lca_internal( Node *n1, Node *n2 ) const;

  // Compute the Ideal Node to Loop mapping
724
  PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs) :
725 726
    PhaseTransform(Ideal_Loop),
    _igvn(igvn),
727
    _dom_lca_tags(arena()), // Thread::resource_area
728 729
    _verify_me(NULL),
    _verify_only(false) {
730
    build_and_optimize(do_split_ifs);
731 732 733 734 735 736
  }

  // Verify that verify_me made the same decisions as a fresh run.
  PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
    PhaseTransform(Ideal_Loop),
    _igvn(igvn),
737
    _dom_lca_tags(arena()), // Thread::resource_area
738 739
    _verify_me(verify_me),
    _verify_only(false) {
740
    build_and_optimize(false);
741 742 743 744 745 746 747 748 749
  }

  // Build and verify the loop tree without modifying the graph.  This
  // is useful to verify that all inputs properly dominate their uses.
  static void verify(PhaseIterGVN& igvn) {
#ifdef ASSERT
    PhaseIdealLoop v(igvn);
#endif
  }
D
duke 已提交
750 751 752 753 754 755 756

  // True if the method has at least 1 irreducible loop
  bool _has_irreducible_loops;

  // Per-Node transform
  virtual Node *transform( Node *a_node ) { return 0; }

757
  bool is_counted_loop( Node *x, IdealLoopTree *loop );
D
duke 已提交
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819

  // Return a post-walked LoopNode
  IdealLoopTree *get_loop( Node *n ) const {
    // Dead nodes have no loop, so return the top level loop instead
    if (!has_node(n))  return _ltree_root;
    assert(!has_ctrl(n), "");
    return (IdealLoopTree*)_nodes[n->_idx];
  }

  // Is 'n' a (nested) member of 'loop'?
  int is_member( const IdealLoopTree *loop, Node *n ) const {
    return loop->is_member(get_loop(n)); }

  // This is the basic building block of the loop optimizations.  It clones an
  // entire loop body.  It makes an old_new loop body mapping; with this
  // mapping you can find the new-loop equivalent to an old-loop node.  All
  // new-loop nodes are exactly equal to their old-loop counterparts, all
  // edges are the same.  All exits from the old-loop now have a RegionNode
  // that merges the equivalent new-loop path.  This is true even for the
  // normal "loop-exit" condition.  All uses of loop-invariant old-loop values
  // now come from (one or more) Phis that merge their new-loop equivalents.
  // Parameter side_by_side_idom:
  //   When side_by_size_idom is NULL, the dominator tree is constructed for
  //      the clone loop to dominate the original.  Used in construction of
  //      pre-main-post loop sequence.
  //   When nonnull, the clone and original are side-by-side, both are
  //      dominated by the passed in side_by_side_idom node.  Used in
  //      construction of unswitched loops.
  void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
                   Node* side_by_side_idom = NULL);

  // If we got the effect of peeling, either by actually peeling or by
  // making a pre-loop which must execute at least once, we can remove
  // all loop-invariant dominated tests in the main body.
  void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );

  // Generate code to do a loop peel for the given loop (and body).
  // old_new is a temp array.
  void do_peeling( IdealLoopTree *loop, Node_List &old_new );

  // Add pre and post loops around the given loop.  These loops are used
  // during RCE, unrolling and aligning loops.
  void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
  // If Node n lives in the back_ctrl block, we clone a private version of n
  // in preheader_ctrl block and return that, otherwise return n.
  Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );

  // Take steps to maximally unroll the loop.  Peel any odd iterations, then
  // unroll to do double iterations.  The next round of major loop transforms
  // will repeat till the doubled loop body does all remaining iterations in 1
  // pass.
  void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );

  // Unroll the loop body one step - make each trip do 2 iterations.
  void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );

  // Return true if exp is a constant times an induction var
  bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);

  // Return true if exp is a scaled induction var plus (or minus) constant
  bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);

820
  // Return true if proj is for "proj->[region->..]call_uct"
821 822
  // Return true if proj is for "proj->[region->..]call_uct"
  static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason);
823 824 825 826
  // Return true for    "if(test)-> proj -> ...
  //                          |
  //                          V
  //                      other_proj->[region->..]call_uct"
827
  static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason);
828
  // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
829 830 831 832
  ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
                                        Deoptimization::DeoptReason reason);
  void register_control(Node* n, IdealLoopTree *loop, Node* pred);

833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852
  // Clone loop predicates to cloned loops (peeled, unswitched)
  static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
                                   Deoptimization::DeoptReason reason,
                                   PhaseIdealLoop* loop_phase,
                                   PhaseIterGVN* igvn);
  static ProjNode*  move_predicate(ProjNode* predicate_proj, Node* new_entry,
                                   Deoptimization::DeoptReason reason,
                                   PhaseIdealLoop* loop_phase,
                                   PhaseIterGVN* igvn);
  static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
                                         bool move_predicates,
                                         PhaseIdealLoop* loop_phase,
                                         PhaseIterGVN* igvn);
  Node* clone_loop_predicates(Node* old_entry, Node* new_entry);
  Node*  move_loop_predicates(Node* old_entry, Node* new_entry);

  void eliminate_loop_predicates(Node* entry);
  static Node* skip_loop_predicates(Node* entry);

  // Find a good location to insert a predicate
853 854 855
  static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
  // Find a predicate
  static Node* find_predicate(Node* entry);
856 857 858 859
  // Construct a range check for a predicate if
  BoolNode* rc_predicate(Node* ctrl,
                         int scale, Node* offset,
                         Node* init, Node* limit, Node* stride,
860
                         Node* range, bool upper);
861 862 863 864 865 866 867 868

  // Implementation of the loop predication to promote checks outside the loop
  bool loop_predication_impl(IdealLoopTree *loop);

  // Helper function to collect predicate for eliminating the useless ones
  void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
  void eliminate_useless_predicates();

D
duke 已提交
869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966
  // Eliminate range-checks and other trip-counter vs loop-invariant tests.
  void do_range_check( IdealLoopTree *loop, Node_List &old_new );

  // Create a slow version of the loop by cloning the loop
  // and inserting an if to select fast-slow versions.
  ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
                                        Node_List &old_new);

  // Clone loop with an invariant test (that does not exit) and
  // insert a clone of the test that selects which version to
  // execute.
  void do_unswitching (IdealLoopTree *loop, Node_List &old_new);

  // Find candidate "if" for unswitching
  IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;

  // Range Check Elimination uses this function!
  // Constrain the main loop iterations so the affine function:
  //    scale_con * I + offset  <  limit
  // always holds true.  That is, either increase the number of iterations in
  // the pre-loop or the post-loop until the condition holds true in the main
  // loop.  Scale_con, offset and limit are all loop invariant.
  void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );

  // Partially peel loop up through last_peel node.
  bool partial_peel( IdealLoopTree *loop, Node_List &old_new );

  // Create a scheduled list of nodes control dependent on ctrl set.
  void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
  // Has a use in the vector set
  bool has_use_in_set( Node* n, VectorSet& vset );
  // Has use internal to the vector set (ie. not in a phi at the loop head)
  bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
  // clone "n" for uses that are outside of loop
  void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
  // clone "n" for special uses that are in the not_peeled region
  void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
                                          VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
  // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
  void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
#ifdef ASSERT
  // Validate the loop partition sets: peel and not_peel
  bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
  // Ensure that uses outside of loop are of the right form
  bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
                                 uint orig_exit_idx, uint clone_exit_idx);
  bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
#endif

  // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
  int stride_of_possible_iv( Node* iff );
  bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
  // Return the (unique) control output node that's in the loop (if it exists.)
  Node* stay_in_loop( Node* n, IdealLoopTree *loop);
  // Insert a signed compare loop exit cloned from an unsigned compare.
  IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
  void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
  // Utility to register node "n" with PhaseIdealLoop
  void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
  // Utility to create an if-projection
  ProjNode* proj_clone(ProjNode* p, IfNode* iff);
  // Force the iff control output to be the live_proj
  Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
  // Insert a region before an if projection
  RegionNode* insert_region_before_proj(ProjNode* proj);
  // Insert a new if before an if projection
  ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);

  // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
  // "Nearly" because all Nodes have been cloned from the original in the loop,
  // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
  // through the Phi recursively, and return a Bool.
  BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
  CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );


  // Rework addressing expressions to get the most loop-invariant stuff
  // moved out.  We'd like to do all associative operators, but it's especially
  // important (common) to do address expressions.
  Node *remix_address_expressions( Node *n );

  // Attempt to use a conditional move instead of a phi/branch
  Node *conditional_move( Node *n );

  // Reorganize offset computations to lower register pressure.
  // Mostly prevent loop-fallout uses of the pre-incremented trip counter
  // (which are then alive with the post-incremented trip counter
  // forcing an extra register move)
  void reorg_offsets( IdealLoopTree *loop );

  // Check for aggressive application of 'split-if' optimization,
  // using basic block level info.
  void  split_if_with_blocks     ( VectorSet &visited, Node_Stack &nstack );
  Node *split_if_with_blocks_pre ( Node *n );
  void  split_if_with_blocks_post( Node *n );
  Node *has_local_phi_input( Node *n );
  // Mark an IfNode as being dominated by a prior test,
  // without actually altering the CFG (and hence IDOM info).
967
  void dominated_by( Node *prevdom, Node *iff, bool flip = false );
D
duke 已提交
968 969 970 971 972 973 974 975 976

  // Split Node 'n' through merge point
  Node *split_thru_region( Node *n, Node *region );
  // Split Node 'n' through merge point if there is enough win.
  Node *split_thru_phi( Node *n, Node *region, int policy );
  // Found an If getting its condition-code input from a Phi in the
  // same block.  Split thru the Region.
  void do_split_if( Node *iff );

N
never 已提交
977 978 979 980 981 982
  // Conversion of fill/copy patterns into intrisic versions
  bool do_intrinsify_fill();
  bool intrinsify_fill(IdealLoopTree* lpt);
  bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
                       Node*& shift, Node*& offset);

D
duke 已提交
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001
private:
  // Return a type based on condition control flow
  const TypeInt* filtered_type( Node *n, Node* n_ctrl);
  const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
 // Helpers for filtered type
  const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);

  // Helper functions
  Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
  Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
  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 );
  bool split_up( Node *n, Node *blk1, Node *blk2 );
  void sink_use( Node *use, Node *post_loop );
  Node *place_near_use( Node *useblock ) const;

  bool _created_loop_node;
public:
  void set_created_loop_node() { _created_loop_node = true; }
  bool created_loop_node()     { return _created_loop_node; }
1002
  void register_new_node( Node *n, Node *blk );
D
duke 已提交
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054

#ifndef PRODUCT
  void dump( ) const;
  void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
  void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
  void verify() const;          // Major slow  :-)
  void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
  IdealLoopTree *get_loop_idx(Node* n) const {
    // Dead nodes have no loop, so return the top level loop instead
    return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
  }
  // Print some stats
  static void print_statistics();
  static int _loop_invokes;     // Count of PhaseIdealLoop invokes
  static int _loop_work;        // Sum of PhaseIdealLoop x _unique
#endif
};

inline Node* IdealLoopTree::tail() {
// Handle lazy update of _tail field
  Node *n = _tail;
  //while( !n->in(0) )  // Skip dead CFG nodes
    //n = n->in(1);
  if (n->in(0) == NULL)
    n = _phase->get_ctrl(n);
  _tail = n;
  return n;
}


// Iterate over the loop tree using a preorder, left-to-right traversal.
//
// Example that visits all counted loops from within PhaseIdealLoop
//
//  for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
//   IdealLoopTree* lpt = iter.current();
//   if (!lpt->is_counted()) continue;
//   ...
class LoopTreeIterator : public StackObj {
private:
  IdealLoopTree* _root;
  IdealLoopTree* _curnt;

public:
  LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}

  bool done() { return _curnt == NULL; }       // Finished iterating?

  void next();                                 // Advance to next loop tree

  IdealLoopTree* current() { return _curnt; }  // Return current value of iterator.
};
1055 1056

#endif // SHARE_VM_OPTO_LOOPNODE_HPP