1 /*
2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP
26 #define SHARE_VM_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/subnode.hpp"
32 #include "opto/type.hpp"
33
34 class CmpNode;
35 class CountedLoopEndNode;
36 class CountedLoopNode;
37 class IdealLoopTree;
38 class LoopNode;
39 class Node;
40 class PhaseIdealLoop;
41 class VectorSet;
42 class Invariance;
43 struct small_cache;
44
45 //
46 // I D E A L I Z E D L O O P S
47 //
48 // Idealized loops are the set of loops I perform more interesting
49 // transformations on, beyond simple hoisting.
50
51 //------------------------------LoopNode---------------------------------------
52 // Simple loop header. Fall in path on left, loop-back path on right.
53 class LoopNode : public RegionNode {
54 // Size is bigger to hold the flags. However, the flags do not change
55 // the semantics so it does not appear in the hash & cmp functions.
56 virtual uint size_of() const { return sizeof(*this); }
57 protected:
58 short _loop_flags;
59 // Names for flag bitfields
60 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
61 MainHasNoPreLoop=4,
62 HasExactTripCount=8,
63 InnerLoop=16,
64 PartialPeelLoop=32,
65 PartialPeelFailed=64,
66 HasReductions=128,
67 PassedSlpAnalysis=256 };
68 char _unswitch_count;
69 enum { _unswitch_max=3 };
70
71 public:
72 // Names for edge indices
73 enum { Self=0, EntryControl, LoopBackControl };
74
75 int is_inner_loop() const { return _loop_flags & InnerLoop; }
76 void set_inner_loop() { _loop_flags |= InnerLoop; }
77
78 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
79 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
80 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
81 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
82 void mark_has_reductions() { _loop_flags |= HasReductions; }
83 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
84
85 int unswitch_max() { return _unswitch_max; }
86 int unswitch_count() { return _unswitch_count; }
87 void set_unswitch_count(int val) {
88 assert (val <= unswitch_max(), "too many unswitches");
89 _unswitch_count = val;
90 }
91
92 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
93 init_class_id(Class_Loop);
94 init_req(EntryControl, entry);
95 init_req(LoopBackControl, backedge);
96 }
97
98 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
99 virtual int Opcode() const;
100 bool can_be_counted_loop(PhaseTransform* phase) const {
101 return req() == 3 && in(0) != NULL &&
102 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
103 in(2) != NULL && phase->type(in(2)) != Type::TOP;
104 }
105 bool is_valid_counted_loop() const;
106 #ifndef PRODUCT
107 virtual void dump_spec(outputStream *st) const;
108 #endif
109 };
110
111 //------------------------------Counted Loops----------------------------------
112 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
113 // path (and maybe some other exit paths). The trip-counter exit is always
114 // last in the loop. The trip-counter have to stride by a constant;
115 // the exit value is also loop invariant.
116
117 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
118 // CountedLoopNode has the incoming loop control and the loop-back-control
119 // which is always the IfTrue before the matching CountedLoopEndNode. The
120 // CountedLoopEndNode has an incoming control (possibly not the
121 // CountedLoopNode if there is control flow in the loop), the post-increment
122 // trip-counter value, and the limit. The trip-counter value is always of
123 // the form (Op old-trip-counter stride). The old-trip-counter is produced
124 // by a Phi connected to the CountedLoopNode. The stride is constant.
125 // The Op is any commutable opcode, including Add, Mul, Xor. The
126 // CountedLoopEndNode also takes in the loop-invariant limit value.
127
128 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
129 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
130 // via the old-trip-counter from the Op node.
131
132 //------------------------------CountedLoopNode--------------------------------
133 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
134 // inputs the incoming loop-start control and the loop-back control, so they
135 // act like RegionNodes. They also take in the initial trip counter, the
136 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
137 // produce a loop-body control and the trip counter value. Since
138 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
139
140 class CountedLoopNode : public LoopNode {
141 // Size is bigger to hold _main_idx. However, _main_idx does not change
142 // the semantics so it does not appear in the hash & cmp functions.
143 virtual uint size_of() const { return sizeof(*this); }
144
145 // For Pre- and Post-loops during debugging ONLY, this holds the index of
146 // the Main CountedLoop. Used to assert that we understand the graph shape.
147 node_idx_t _main_idx;
148
149 // Known trip count calculated by compute_exact_trip_count()
150 uint _trip_count;
151
152 // Expected trip count from profile data
153 float _profile_trip_cnt;
154
155 // Log2 of original loop bodies in unrolled loop
156 int _unrolled_count_log2;
157
158 // Node count prior to last unrolling - used to decide if
159 // unroll,optimize,unroll,optimize,... is making progress
160 int _node_count_before_unroll;
161
162 // If slp analysis is performed we record the maximum
163 // vector mapped unroll factor here
164 int _slp_maximum_unroll_factor;
165
166 public:
167 CountedLoopNode( Node *entry, Node *backedge )
168 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
169 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
170 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
171 init_class_id(Class_CountedLoop);
172 // Initialize _trip_count to the largest possible value.
173 // Will be reset (lower) if the loop's trip count is known.
174 }
175
176 virtual int Opcode() const;
177 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
178
179 Node *init_control() const { return in(EntryControl); }
180 Node *back_control() const { return in(LoopBackControl); }
181 CountedLoopEndNode *loopexit() const;
182 Node *init_trip() const;
183 Node *stride() const;
184 int stride_con() const;
185 bool stride_is_con() const;
186 Node *limit() const;
187 Node *incr() const;
188 Node *phi() const;
189
190 // Match increment with optional truncation
191 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
192
193 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
194 // can run short a few iterations and may start a few iterations in.
195 // It will be RCE'd and unrolled and aligned.
196
197 // A following 'post' loop will run any remaining iterations. Used
198 // during Range Check Elimination, the 'post' loop will do any final
199 // iterations with full checks. Also used by Loop Unrolling, where
200 // the 'post' loop will do any epilog iterations needed. Basically,
201 // a 'post' loop can not profitably be further unrolled or RCE'd.
202
203 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
204 // it may do under-flow checks for RCE and may do alignment iterations
205 // so the following main loop 'knows' that it is striding down cache
206 // lines.
207
208 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
209 // Aligned, may be missing it's pre-loop.
210 int is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
211 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
212 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
213 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
214 int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
215 int has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
216 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
217 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
218
219 int main_idx() const { return _main_idx; }
220
221
222 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
223 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
224 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
225 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
226
227 void set_trip_count(uint tc) { _trip_count = tc; }
228 uint trip_count() { return _trip_count; }
229
230 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
231 void set_exact_trip_count(uint tc) {
232 _trip_count = tc;
233 _loop_flags |= HasExactTripCount;
234 }
235 void set_nonexact_trip_count() {
236 _loop_flags &= ~HasExactTripCount;
237 }
238
239 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
240 float profile_trip_cnt() { return _profile_trip_cnt; }
241
242 void double_unrolled_count() { _unrolled_count_log2++; }
243 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
244
245 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
246 int node_count_before_unroll() { return _node_count_before_unroll; }
247 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
248 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
249
250 #ifndef PRODUCT
251 virtual void dump_spec(outputStream *st) const;
252 #endif
253 };
254
255 //------------------------------CountedLoopEndNode-----------------------------
256 // CountedLoopEndNodes end simple trip counted loops. They act much like
257 // IfNodes.
258 class CountedLoopEndNode : public IfNode {
259 public:
260 enum { TestControl, TestValue };
261
262 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
263 : IfNode( control, test, prob, cnt) {
264 init_class_id(Class_CountedLoopEnd);
265 }
266 virtual int Opcode() const;
267
268 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
269 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
270 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
271 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
272 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
273 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
274 int stride_con() const;
275 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
276 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
277 CountedLoopNode *loopnode() const {
278 // The CountedLoopNode that goes with this CountedLoopEndNode may
279 // have been optimized out by the IGVN so be cautious with the
280 // pattern matching on the graph
281 if (phi() == NULL) {
282 return NULL;
283 }
284 Node *ln = phi()->in(0);
285 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
286 return (CountedLoopNode*)ln;
287 }
288 return NULL;
289 }
290
291 #ifndef PRODUCT
292 virtual void dump_spec(outputStream *st) const;
293 #endif
294 };
295
296
297 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
298 Node *bc = back_control();
299 if( bc == NULL ) return NULL;
300 Node *le = bc->in(0);
301 if( le->Opcode() != Op_CountedLoopEnd )
302 return NULL;
303 return (CountedLoopEndNode*)le;
304 }
305 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
306 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
307 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
308 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
309 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
310 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
311 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
312
313 //------------------------------LoopLimitNode-----------------------------
314 // Counted Loop limit node which represents exact final iterator value:
315 // trip_count = (limit - init_trip + stride - 1)/stride
316 // final_value= trip_count * stride + init_trip.
317 // Use HW instructions to calculate it when it can overflow in integer.
318 // Note, final_value should fit into integer since counted loop has
319 // limit check: limit <= max_int-stride.
320 class LoopLimitNode : public Node {
321 enum { Init=1, Limit=2, Stride=3 };
322 public:
323 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
324 // Put it on the Macro nodes list to optimize during macro nodes expansion.
325 init_flags(Flag_is_macro);
326 C->add_macro_node(this);
327 }
328 virtual int Opcode() const;
329 virtual const Type *bottom_type() const { return TypeInt::INT; }
330 virtual uint ideal_reg() const { return Op_RegI; }
331 virtual const Type *Value( PhaseTransform *phase ) const;
332 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
333 virtual Node *Identity( PhaseTransform *phase );
334 };
335
336 // -----------------------------IdealLoopTree----------------------------------
337 class IdealLoopTree : public ResourceObj {
338 public:
339 IdealLoopTree *_parent; // Parent in loop tree
340 IdealLoopTree *_next; // Next sibling in loop tree
341 IdealLoopTree *_child; // First child in loop tree
342
343 // The head-tail backedge defines the loop.
344 // If tail is NULL then this loop has multiple backedges as part of the
345 // same loop. During cleanup I'll peel off the multiple backedges; merge
346 // them at the loop bottom and flow 1 real backedge into the loop.
347 Node *_head; // Head of loop
348 Node *_tail; // Tail of loop
349 inline Node *tail(); // Handle lazy update of _tail field
350 PhaseIdealLoop* _phase;
351 int _local_loop_unroll_limit;
352 int _local_loop_unroll_factor;
353
354 Node_List _body; // Loop body for inner loops
355
356 uint8_t _nest; // Nesting depth
357 uint8_t _irreducible:1, // True if irreducible
358 _has_call:1, // True if has call safepoint
359 _has_sfpt:1, // True if has non-call safepoint
360 _rce_candidate:1; // True if candidate for range check elimination
361
362 Node_List* _safepts; // List of safepoints in this loop
363 Node_List* _required_safept; // A inner loop cannot delete these safepts;
364 bool _allow_optimizations; // Allow loop optimizations
365
366 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
367 : _parent(0), _next(0), _child(0),
368 _head(head), _tail(tail),
369 _phase(phase),
370 _safepts(NULL),
371 _required_safept(NULL),
372 _allow_optimizations(true),
373 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
374 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
375 { }
376
377 // Is 'l' a member of 'this'?
378 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
379
380 // Set loop nesting depth. Accumulate has_call bits.
381 int set_nest( uint depth );
382
383 // Split out multiple fall-in edges from the loop header. Move them to a
384 // private RegionNode before the loop. This becomes the loop landing pad.
385 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
386
387 // Split out the outermost loop from this shared header.
388 void split_outer_loop( PhaseIdealLoop *phase );
389
390 // Merge all the backedges from the shared header into a private Region.
391 // Feed that region as the one backedge to this loop.
392 void merge_many_backedges( PhaseIdealLoop *phase );
393
394 // Split shared headers and insert loop landing pads.
395 // Insert a LoopNode to replace the RegionNode.
396 // Returns TRUE if loop tree is structurally changed.
397 bool beautify_loops( PhaseIdealLoop *phase );
398
399 // Perform optimization to use the loop predicates for null checks and range checks.
400 // Applies to any loop level (not just the innermost one)
401 bool loop_predication( PhaseIdealLoop *phase);
402
403 // Perform iteration-splitting on inner loops. Split iterations to
404 // avoid range checks or one-shot null checks. Returns false if the
405 // current round of loop opts should stop.
406 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
407
408 // Driver for various flavors of iteration splitting. Returns false
409 // if the current round of loop opts should stop.
410 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
411
412 // Given dominators, try to find loops with calls that must always be
413 // executed (call dominates loop tail). These loops do not need non-call
414 // safepoints (ncsfpt).
415 void check_safepts(VectorSet &visited, Node_List &stack);
416
417 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
418 // encountered.
419 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
420
421 // Convert to counted loops where possible
422 void counted_loop( PhaseIdealLoop *phase );
423
424 // Check for Node being a loop-breaking test
425 Node *is_loop_exit(Node *iff) const;
426
427 // Returns true if ctrl is executed on every complete iteration
428 bool dominates_backedge(Node* ctrl);
429
430 // Remove simplistic dead code from loop body
431 void DCE_loop_body();
432
433 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
434 // Replace with a 1-in-10 exit guess.
435 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
436
437 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
438 // Useful for unrolling loops with NO array accesses.
439 bool policy_peel_only( PhaseIdealLoop *phase ) const;
440
441 // Return TRUE or FALSE if the loop should be unswitched -- clone
442 // loop with an invariant test
443 bool policy_unswitching( PhaseIdealLoop *phase ) const;
444
445 // Micro-benchmark spamming. Remove empty loops.
446 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
447
448 // Convert one iteration loop into normal code.
449 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
450
451 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
452 // make some loop-invariant test (usually a null-check) happen before the
453 // loop.
454 bool policy_peeling( PhaseIdealLoop *phase ) const;
455
456 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
457 // known trip count in the counted loop node.
458 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
459
460 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
461 // the loop is a CountedLoop and the body is small enough.
462 bool policy_unroll(PhaseIdealLoop *phase);
463
464 // Loop analyses to map to a maximal superword unrolling for vectorization.
465 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
466
467 // Return TRUE or FALSE if the loop should be range-check-eliminated.
468 // Gather a list of IF tests that are dominated by iteration splitting;
469 // also gather the end of the first split and the start of the 2nd split.
470 bool policy_range_check( PhaseIdealLoop *phase ) const;
471
472 // Return TRUE or FALSE if the loop should be cache-line aligned.
473 // Gather the expression that does the alignment. Note that only
474 // one array base can be aligned in a loop (unless the VM guarantees
475 // mutual alignment). Note that if we vectorize short memory ops
476 // into longer memory ops, we may want to increase alignment.
477 bool policy_align( PhaseIdealLoop *phase ) const;
478
479 // Return TRUE if "iff" is a range check.
480 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
481
482 // Compute loop exact trip count if possible
483 void compute_exact_trip_count( PhaseIdealLoop *phase );
484
485 // Compute loop trip count from profile data
486 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
487
488 // Reassociate invariant expressions.
489 void reassociate_invariants(PhaseIdealLoop *phase);
490 // Reassociate invariant add and subtract expressions.
491 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
492 // Return nonzero index of invariant operand if invariant and variant
493 // are combined with an Add or Sub. Helper for reassociate_invariants.
494 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
495
496 // Return true if n is invariant
497 bool is_invariant(Node* n) const;
498
499 // Put loop body on igvn work list
500 void record_for_igvn();
501
502 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
503 bool is_inner() { return is_loop() && _child == NULL; }
504 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
505
506 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
507
508 #ifndef PRODUCT
509 void dump_head( ) const; // Dump loop head only
510 void dump() const; // Dump this loop recursively
511 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
512 #endif
513
514 };
515
516 // -----------------------------PhaseIdealLoop---------------------------------
517 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
518 // loop tree. Drives the loop-based transformations on the ideal graph.
519 class PhaseIdealLoop : public PhaseTransform {
520 friend class IdealLoopTree;
521 friend class SuperWord;
522 // Pre-computed def-use info
523 PhaseIterGVN &_igvn;
524
525 // Head of loop tree
526 IdealLoopTree *_ltree_root;
527
528 // Array of pre-order numbers, plus post-visited bit.
529 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
530 // ODD for post-visited. Other bits are the pre-order number.
531 uint *_preorders;
532 uint _max_preorder;
533
534 const PhaseIdealLoop* _verify_me;
535 bool _verify_only;
536
537 // Allocate _preorders[] array
538 void allocate_preorders() {
539 _max_preorder = C->unique()+8;
540 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
541 memset(_preorders, 0, sizeof(uint) * _max_preorder);
542 }
543
544 // Allocate _preorders[] array
545 void reallocate_preorders() {
546 if ( _max_preorder < C->unique() ) {
547 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
548 _max_preorder = C->unique();
549 }
550 memset(_preorders, 0, sizeof(uint) * _max_preorder);
551 }
552
553 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
554 // adds new nodes.
555 void check_grow_preorders( ) {
556 if ( _max_preorder < C->unique() ) {
557 uint newsize = _max_preorder<<1; // double size of array
558 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
559 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
560 _max_preorder = newsize;
561 }
562 }
563 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
564 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
565 // Pre-order numbers are written to the Nodes array as low-bit-set values.
566 void set_preorder_visited( Node *n, int pre_order ) {
567 assert( !is_visited( n ), "already set" );
568 _preorders[n->_idx] = (pre_order<<1);
569 };
570 // Return pre-order number.
571 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
572
573 // Check for being post-visited.
574 // Should be previsited already (checked with assert(is_visited(n))).
575 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
576
577 // Mark as post visited
578 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
579
580 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
581 // Returns true if "n" is a data node, false if it's a control node.
582 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
583
584 // clear out dead code after build_loop_late
585 Node_List _deadlist;
586
587 // Support for faster execution of get_late_ctrl()/dom_lca()
588 // when a node has many uses and dominator depth is deep.
589 Node_Array _dom_lca_tags;
590 void init_dom_lca_tags();
591 void clear_dom_lca_tags();
592
593 // Helper for debugging bad dominance relationships
594 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
595
596 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
597
598 // Inline wrapper for frequent cases:
599 // 1) only one use
600 // 2) a use is the same as the current LCA passed as 'n1'
601 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
602 assert( n->is_CFG(), "" );
603 // Fast-path NULL lca
604 if( lca != NULL && lca != n ) {
605 assert( lca->is_CFG(), "" );
606 // find LCA of all uses
607 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
608 }
609 return find_non_split_ctrl(n);
610 }
611 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
612
613 // Helper function for directing control inputs away from CFG split
614 // points.
615 Node *find_non_split_ctrl( Node *ctrl ) const {
616 if (ctrl != NULL) {
617 if (ctrl->is_MultiBranch()) {
618 ctrl = ctrl->in(0);
619 }
620 assert(ctrl->is_CFG(), "CFG");
621 }
622 return ctrl;
623 }
624
625 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
626
627 public:
628 bool has_node( Node* n ) const {
629 guarantee(n != NULL, "No Node.");
630 return _nodes[n->_idx] != NULL;
631 }
632 // check if transform created new nodes that need _ctrl recorded
633 Node *get_late_ctrl( Node *n, Node *early );
634 Node *get_early_ctrl( Node *n );
635 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
636 void set_early_ctrl( Node *n );
637 void set_subtree_ctrl( Node *root );
638 void set_ctrl( Node *n, Node *ctrl ) {
639 assert( !has_node(n) || has_ctrl(n), "" );
640 assert( ctrl->in(0), "cannot set dead control node" );
641 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
642 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
643 }
644 // Set control and update loop membership
645 void set_ctrl_and_loop(Node* n, Node* ctrl) {
646 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
647 IdealLoopTree* new_loop = get_loop(ctrl);
648 if (old_loop != new_loop) {
649 if (old_loop->_child == NULL) old_loop->_body.yank(n);
650 if (new_loop->_child == NULL) new_loop->_body.push(n);
651 }
652 set_ctrl(n, ctrl);
653 }
654 // Control nodes can be replaced or subsumed. During this pass they
655 // get their replacement Node in slot 1. Instead of updating the block
656 // location of all Nodes in the subsumed block, we lazily do it. As we
657 // pull such a subsumed block out of the array, we write back the final
658 // correct block.
659 Node *get_ctrl( Node *i ) {
660 assert(has_node(i), "");
661 Node *n = get_ctrl_no_update(i);
662 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
663 assert(has_node(i) && has_ctrl(i), "");
664 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
665 return n;
666 }
667 // true if CFG node d dominates CFG node n
668 bool is_dominator(Node *d, Node *n);
669 // return get_ctrl for a data node and self(n) for a CFG node
670 Node* ctrl_or_self(Node* n) {
671 if (has_ctrl(n))
672 return get_ctrl(n);
673 else {
674 assert (n->is_CFG(), "must be a CFG node");
675 return n;
676 }
677 }
678
679 private:
680 Node *get_ctrl_no_update( Node *i ) const {
681 assert( has_ctrl(i), "" );
682 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
683 if (!n->in(0)) {
684 // Skip dead CFG nodes
685 do {
686 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
687 } while (!n->in(0));
688 n = find_non_split_ctrl(n);
689 }
690 return n;
691 }
692
693 // Check for loop being set
694 // "n" must be a control node. Returns true if "n" is known to be in a loop.
695 bool has_loop( Node *n ) const {
696 assert(!has_node(n) || !has_ctrl(n), "");
697 return has_node(n);
698 }
699 // Set loop
700 void set_loop( Node *n, IdealLoopTree *loop ) {
701 _nodes.map(n->_idx, (Node*)loop);
702 }
703 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
704 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
705 // from old_node to new_node to support the lazy update. Reference
706 // replaces loop reference, since that is not needed for dead node.
707 public:
708 void lazy_update( Node *old_node, Node *new_node ) {
709 assert( old_node != new_node, "no cycles please" );
710 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
711 // Nodes always have DU info now, so re-use the side array slot
712 // for this node to provide the forwarding pointer.
713 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
714 }
715 void lazy_replace( Node *old_node, Node *new_node ) {
716 _igvn.replace_node( old_node, new_node );
717 lazy_update( old_node, new_node );
718 }
719 void lazy_replace_proj( Node *old_node, Node *new_node ) {
720 assert( old_node->req() == 1, "use this for Projs" );
721 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
722 old_node->add_req( NULL );
723 lazy_replace( old_node, new_node );
724 }
725
726 private:
727
728 // Place 'n' in some loop nest, where 'n' is a CFG node
729 void build_loop_tree();
730 int build_loop_tree_impl( Node *n, int pre_order );
731 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
732 // loop tree, not the root.
733 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
734
735 // Place Data nodes in some loop nest
736 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
737 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
738 void build_loop_late_post ( Node* n );
739
740 // Array of immediate dominance info for each CFG node indexed by node idx
741 private:
742 uint _idom_size;
743 Node **_idom; // Array of immediate dominators
744 uint *_dom_depth; // Used for fast LCA test
745 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
746
747 Node* idom_no_update(Node* d) const {
748 assert(d->_idx < _idom_size, "oob");
749 Node* n = _idom[d->_idx];
750 assert(n != NULL,"Bad immediate dominator info.");
751 while (n->in(0) == NULL) { // Skip dead CFG nodes
752 //n = n->in(1);
753 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
754 assert(n != NULL,"Bad immediate dominator info.");
755 }
756 return n;
757 }
758 Node *idom(Node* d) const {
759 uint didx = d->_idx;
760 Node *n = idom_no_update(d);
761 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
762 return n;
763 }
764 uint dom_depth(Node* d) const {
765 guarantee(d != NULL, "Null dominator info.");
766 guarantee(d->_idx < _idom_size, "");
767 return _dom_depth[d->_idx];
768 }
769 void set_idom(Node* d, Node* n, uint dom_depth);
770 // Locally compute IDOM using dom_lca call
771 Node *compute_idom( Node *region ) const;
772 // Recompute dom_depth
773 void recompute_dom_depth();
774
775 // Is safept not required by an outer loop?
776 bool is_deleteable_safept(Node* sfpt);
777
778 // Replace parallel induction variable (parallel to trip counter)
779 void replace_parallel_iv(IdealLoopTree *loop);
780
781 // Perform verification that the graph is valid.
782 PhaseIdealLoop( PhaseIterGVN &igvn) :
783 PhaseTransform(Ideal_Loop),
784 _igvn(igvn),
785 _dom_lca_tags(arena()), // Thread::resource_area
786 _verify_me(NULL),
787 _verify_only(true) {
788 build_and_optimize(false, false);
789 }
790
791 // build the loop tree and perform any requested optimizations
792 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
793
794 public:
795 // Dominators for the sea of nodes
796 void Dominators();
797 Node *dom_lca( Node *n1, Node *n2 ) const {
798 return find_non_split_ctrl(dom_lca_internal(n1, n2));
799 }
800 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
801
802 // Compute the Ideal Node to Loop mapping
803 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
804 PhaseTransform(Ideal_Loop),
805 _igvn(igvn),
806 _dom_lca_tags(arena()), // Thread::resource_area
807 _verify_me(NULL),
808 _verify_only(false) {
809 build_and_optimize(do_split_ifs, skip_loop_opts);
810 }
811
812 // Verify that verify_me made the same decisions as a fresh run.
813 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
814 PhaseTransform(Ideal_Loop),
815 _igvn(igvn),
816 _dom_lca_tags(arena()), // Thread::resource_area
817 _verify_me(verify_me),
818 _verify_only(false) {
819 build_and_optimize(false, false);
820 }
821
822 // Build and verify the loop tree without modifying the graph. This
823 // is useful to verify that all inputs properly dominate their uses.
824 static void verify(PhaseIterGVN& igvn) {
825 #ifdef ASSERT
826 PhaseIdealLoop v(igvn);
827 #endif
828 }
829
830 // True if the method has at least 1 irreducible loop
831 bool _has_irreducible_loops;
832
833 // Per-Node transform
834 virtual Node *transform( Node *a_node ) { return 0; }
835
836 bool is_counted_loop( Node *x, IdealLoopTree *loop );
837
838 Node* exact_limit( IdealLoopTree *loop );
839
840 // Return a post-walked LoopNode
841 IdealLoopTree *get_loop( Node *n ) const {
842 // Dead nodes have no loop, so return the top level loop instead
843 if (!has_node(n)) return _ltree_root;
844 assert(!has_ctrl(n), "");
845 return (IdealLoopTree*)_nodes[n->_idx];
846 }
847
848 // Is 'n' a (nested) member of 'loop'?
849 int is_member( const IdealLoopTree *loop, Node *n ) const {
850 return loop->is_member(get_loop(n)); }
851
852 // This is the basic building block of the loop optimizations. It clones an
853 // entire loop body. It makes an old_new loop body mapping; with this
854 // mapping you can find the new-loop equivalent to an old-loop node. All
855 // new-loop nodes are exactly equal to their old-loop counterparts, all
856 // edges are the same. All exits from the old-loop now have a RegionNode
857 // that merges the equivalent new-loop path. This is true even for the
858 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
859 // now come from (one or more) Phis that merge their new-loop equivalents.
860 // Parameter side_by_side_idom:
861 // When side_by_size_idom is NULL, the dominator tree is constructed for
862 // the clone loop to dominate the original. Used in construction of
863 // pre-main-post loop sequence.
864 // When nonnull, the clone and original are side-by-side, both are
865 // dominated by the passed in side_by_side_idom node. Used in
866 // construction of unswitched loops.
867 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
868 Node* side_by_side_idom = NULL);
869
870 // If we got the effect of peeling, either by actually peeling or by
871 // making a pre-loop which must execute at least once, we can remove
872 // all loop-invariant dominated tests in the main body.
873 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
874
875 // Generate code to do a loop peel for the given loop (and body).
876 // old_new is a temp array.
877 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
878
879 // Add pre and post loops around the given loop. These loops are used
880 // during RCE, unrolling and aligning loops.
881 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
882 // If Node n lives in the back_ctrl block, we clone a private version of n
883 // in preheader_ctrl block and return that, otherwise return n.
884 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
885
886 // Take steps to maximally unroll the loop. Peel any odd iterations, then
887 // unroll to do double iterations. The next round of major loop transforms
888 // will repeat till the doubled loop body does all remaining iterations in 1
889 // pass.
890 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
891
892 // Unroll the loop body one step - make each trip do 2 iterations.
893 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
894
895 // Mark vector reduction candidates before loop unrolling
896 void mark_reductions( IdealLoopTree *loop );
897
898 // Return true if exp is a constant times an induction var
899 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
900
901 // Return true if exp is a scaled induction var plus (or minus) constant
902 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
903
904 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
905 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
906 Deoptimization::DeoptReason reason);
907 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
908
909 // Clone loop predicates to cloned loops (peeled, unswitched)
910 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
911 Deoptimization::DeoptReason reason,
912 PhaseIdealLoop* loop_phase,
913 PhaseIterGVN* igvn);
914
915 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
916 bool clone_limit_check,
917 PhaseIdealLoop* loop_phase,
918 PhaseIterGVN* igvn);
919 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
920
921 static Node* skip_loop_predicates(Node* entry);
922
923 // Find a good location to insert a predicate
924 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
925 // Find a predicate
926 static Node* find_predicate(Node* entry);
927 // Construct a range check for a predicate if
928 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
929 int scale, Node* offset,
930 Node* init, Node* limit, Node* stride,
931 Node* range, bool upper);
932
933 // Implementation of the loop predication to promote checks outside the loop
934 bool loop_predication_impl(IdealLoopTree *loop);
935
936 // Helper function to collect predicate for eliminating the useless ones
937 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
938 void eliminate_useless_predicates();
939
940 // Change the control input of expensive nodes to allow commoning by
941 // IGVN when it is guaranteed to not result in a more frequent
942 // execution of the expensive node. Return true if progress.
943 bool process_expensive_nodes();
944
945 // Check whether node has become unreachable
946 bool is_node_unreachable(Node *n) const {
947 return !has_node(n) || n->is_unreachable(_igvn);
948 }
949
950 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
951 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
952
953 // Create a slow version of the loop by cloning the loop
954 // and inserting an if to select fast-slow versions.
955 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
956 Node_List &old_new);
957
958 // Clone loop with an invariant test (that does not exit) and
959 // insert a clone of the test that selects which version to
960 // execute.
961 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
962
963 // Find candidate "if" for unswitching
964 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
965
966 // Range Check Elimination uses this function!
967 // Constrain the main loop iterations so the affine function:
968 // low_limit <= scale_con * I + offset < upper_limit
969 // always holds true. That is, either increase the number of iterations in
970 // the pre-loop or the post-loop until the condition holds true in the main
971 // loop. Scale_con, offset and limit are all loop invariant.
972 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 );
973 // Helper function for add_constraint().
974 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
975
976 // Partially peel loop up through last_peel node.
977 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
978
979 // Create a scheduled list of nodes control dependent on ctrl set.
980 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
981 // Has a use in the vector set
982 bool has_use_in_set( Node* n, VectorSet& vset );
983 // Has use internal to the vector set (ie. not in a phi at the loop head)
984 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
985 // clone "n" for uses that are outside of loop
986 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
987 // clone "n" for special uses that are in the not_peeled region
988 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
989 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
990 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
991 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
992 #ifdef ASSERT
993 // Validate the loop partition sets: peel and not_peel
994 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
995 // Ensure that uses outside of loop are of the right form
996 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
997 uint orig_exit_idx, uint clone_exit_idx);
998 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
999 #endif
1000
1001 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1002 int stride_of_possible_iv( Node* iff );
1003 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1004 // Return the (unique) control output node that's in the loop (if it exists.)
1005 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1006 // Insert a signed compare loop exit cloned from an unsigned compare.
1007 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1008 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1009 // Utility to register node "n" with PhaseIdealLoop
1010 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1011 // Utility to create an if-projection
1012 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1013 // Force the iff control output to be the live_proj
1014 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1015 // Insert a region before an if projection
1016 RegionNode* insert_region_before_proj(ProjNode* proj);
1017 // Insert a new if before an if projection
1018 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1019
1020 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1021 // "Nearly" because all Nodes have been cloned from the original in the loop,
1022 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1023 // through the Phi recursively, and return a Bool.
1024 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1025 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1026
1027
1028 // Rework addressing expressions to get the most loop-invariant stuff
1029 // moved out. We'd like to do all associative operators, but it's especially
1030 // important (common) to do address expressions.
1031 Node *remix_address_expressions( Node *n );
1032
1033 // Attempt to use a conditional move instead of a phi/branch
1034 Node *conditional_move( Node *n );
1035
1036 // Reorganize offset computations to lower register pressure.
1037 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1038 // (which are then alive with the post-incremented trip counter
1039 // forcing an extra register move)
1040 void reorg_offsets( IdealLoopTree *loop );
1041
1042 // Check for aggressive application of 'split-if' optimization,
1043 // using basic block level info.
1044 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1045 Node *split_if_with_blocks_pre ( Node *n );
1046 void split_if_with_blocks_post( Node *n );
1047 Node *has_local_phi_input( Node *n );
1048 // Mark an IfNode as being dominated by a prior test,
1049 // without actually altering the CFG (and hence IDOM info).
1050 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1051
1052 // Split Node 'n' through merge point
1053 Node *split_thru_region( Node *n, Node *region );
1054 // Split Node 'n' through merge point if there is enough win.
1055 Node *split_thru_phi( Node *n, Node *region, int policy );
1056 // Found an If getting its condition-code input from a Phi in the
1057 // same block. Split thru the Region.
1058 void do_split_if( Node *iff );
1059
1060 // Conversion of fill/copy patterns into intrisic versions
1061 bool do_intrinsify_fill();
1062 bool intrinsify_fill(IdealLoopTree* lpt);
1063 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1064 Node*& shift, Node*& offset);
1065
1066 private:
1067 // Return a type based on condition control flow
1068 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1069 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1070 // Helpers for filtered type
1071 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1072
1073 // Helper functions
1074 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1075 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1076 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 );
1077 bool split_up( Node *n, Node *blk1, Node *blk2 );
1078 void sink_use( Node *use, Node *post_loop );
1079 Node *place_near_use( Node *useblock ) const;
1080
1081 bool _created_loop_node;
1082 public:
1083 void set_created_loop_node() { _created_loop_node = true; }
1084 bool created_loop_node() { return _created_loop_node; }
1085 void register_new_node( Node *n, Node *blk );
1086
1087 #ifdef ASSERT
1088 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1089 #endif
1090
1091 #ifndef PRODUCT
1092 void dump( ) const;
1093 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1094 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1095 void verify() const; // Major slow :-)
1096 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1097 IdealLoopTree *get_loop_idx(Node* n) const {
1098 // Dead nodes have no loop, so return the top level loop instead
1099 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1100 }
1101 // Print some stats
1102 static void print_statistics();
1103 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1104 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1105 #endif
1106 };
1107
1108 inline Node* IdealLoopTree::tail() {
1109 // Handle lazy update of _tail field
1110 Node *n = _tail;
1111 //while( !n->in(0) ) // Skip dead CFG nodes
1112 //n = n->in(1);
1113 if (n->in(0) == NULL)
1114 n = _phase->get_ctrl(n);
1115 _tail = n;
1116 return n;
1117 }
1118
1119
1120 // Iterate over the loop tree using a preorder, left-to-right traversal.
1121 //
1122 // Example that visits all counted loops from within PhaseIdealLoop
1123 //
1124 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1125 // IdealLoopTree* lpt = iter.current();
1126 // if (!lpt->is_counted()) continue;
1127 // ...
1128 class LoopTreeIterator : public StackObj {
1129 private:
1130 IdealLoopTree* _root;
1131 IdealLoopTree* _curnt;
1132
1133 public:
1134 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1135
1136 bool done() { return _curnt == NULL; } // Finished iterating?
1137
1138 void next(); // Advance to next loop tree
1139
1140 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1141 };
1142
1143 #endif // SHARE_VM_OPTO_LOOPNODE_HPP