1 /*
2 * Copyright (c) 1998, 2015, 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 CountedLoopReserveKit;
42 class VectorSet;
43 class Invariance;
44 struct small_cache;
45
46 //
47 // I D E A L I Z E D L O O P S
48 //
49 // Idealized loops are the set of loops I perform more interesting
50 // transformations on, beyond simple hoisting.
51
52 //------------------------------LoopNode---------------------------------------
53 // Simple loop header. Fall in path on left, loop-back path on right.
54 class LoopNode : public RegionNode {
55 // Size is bigger to hold the flags. However, the flags do not change
56 // the semantics so it does not appear in the hash & cmp functions.
57 virtual uint size_of() const { return sizeof(*this); }
58 protected:
59 short _loop_flags;
60 // Names for flag bitfields
61 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
62 MainHasNoPreLoop=4,
63 HasExactTripCount=8,
64 InnerLoop=16,
65 PartialPeelLoop=32,
66 PartialPeelFailed=64,
67 HasReductions=128,
68 WasSlpAnalyzed=256,
69 PassedSlpAnalysis=512,
70 DoUnrollOnly=1024,
71 VectorizedLoop=2048,
72 HasAtomicPostLoop=4096,
73 HasRangeChecks=8192,
74 IsMultiversioned=16384};
75 char _unswitch_count;
76 enum { _unswitch_max=3 };
77
78 public:
79 // Names for edge indices
80 enum { Self=0, EntryControl, LoopBackControl };
81
82 int is_inner_loop() const { return _loop_flags & InnerLoop; }
83 void set_inner_loop() { _loop_flags |= InnerLoop; }
84
85 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
86 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
87 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
88 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
89 void mark_has_reductions() { _loop_flags |= HasReductions; }
90 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
91 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
92 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
93 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
94 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
95 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
96 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
97
98 int unswitch_max() { return _unswitch_max; }
99 int unswitch_count() { return _unswitch_count; }
100 void set_unswitch_count(int val) {
101 assert (val <= unswitch_max(), "too many unswitches");
102 _unswitch_count = val;
103 }
104
105 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
106 init_class_id(Class_Loop);
107 init_req(EntryControl, entry);
108 init_req(LoopBackControl, backedge);
109 }
110
111 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
112 virtual int Opcode() const;
113 bool can_be_counted_loop(PhaseTransform* phase) const {
114 return req() == 3 && in(0) != NULL &&
115 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
116 in(2) != NULL && phase->type(in(2)) != Type::TOP;
117 }
118 bool is_valid_counted_loop() const;
119 #ifndef PRODUCT
120 virtual void dump_spec(outputStream *st) const;
121 #endif
122 };
123
124 //------------------------------Counted Loops----------------------------------
125 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
126 // path (and maybe some other exit paths). The trip-counter exit is always
127 // last in the loop. The trip-counter have to stride by a constant;
128 // the exit value is also loop invariant.
129
130 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
131 // CountedLoopNode has the incoming loop control and the loop-back-control
132 // which is always the IfTrue before the matching CountedLoopEndNode. The
133 // CountedLoopEndNode has an incoming control (possibly not the
134 // CountedLoopNode if there is control flow in the loop), the post-increment
135 // trip-counter value, and the limit. The trip-counter value is always of
136 // the form (Op old-trip-counter stride). The old-trip-counter is produced
137 // by a Phi connected to the CountedLoopNode. The stride is constant.
138 // The Op is any commutable opcode, including Add, Mul, Xor. The
139 // CountedLoopEndNode also takes in the loop-invariant limit value.
140
141 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
142 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
143 // via the old-trip-counter from the Op node.
144
145 //------------------------------CountedLoopNode--------------------------------
146 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
147 // inputs the incoming loop-start control and the loop-back control, so they
148 // act like RegionNodes. They also take in the initial trip counter, the
149 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
150 // produce a loop-body control and the trip counter value. Since
151 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
152
153 class CountedLoopNode : public LoopNode {
154 // Size is bigger to hold _main_idx. However, _main_idx does not change
155 // the semantics so it does not appear in the hash & cmp functions.
156 virtual uint size_of() const { return sizeof(*this); }
157
158 // For Pre- and Post-loops during debugging ONLY, this holds the index of
159 // the Main CountedLoop. Used to assert that we understand the graph shape.
160 node_idx_t _main_idx;
161
162 // Known trip count calculated by compute_exact_trip_count()
163 uint _trip_count;
164
165 // Expected trip count from profile data
166 float _profile_trip_cnt;
167
168 // Log2 of original loop bodies in unrolled loop
169 int _unrolled_count_log2;
170
171 // Node count prior to last unrolling - used to decide if
172 // unroll,optimize,unroll,optimize,... is making progress
173 int _node_count_before_unroll;
174
175 // If slp analysis is performed we record the maximum
176 // vector mapped unroll factor here
177 int _slp_maximum_unroll_factor;
178
179 public:
180 CountedLoopNode( Node *entry, Node *backedge )
181 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
182 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
183 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
184 init_class_id(Class_CountedLoop);
185 // Initialize _trip_count to the largest possible value.
186 // Will be reset (lower) if the loop's trip count is known.
187 }
188
189 virtual int Opcode() const;
190 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
191
192 Node *init_control() const { return in(EntryControl); }
193 Node *back_control() const { return in(LoopBackControl); }
194 CountedLoopEndNode *loopexit() const;
195 Node *init_trip() const;
196 Node *stride() const;
197 int stride_con() const;
198 bool stride_is_con() const;
199 Node *limit() const;
200 Node *incr() const;
201 Node *phi() const;
202
203 // Match increment with optional truncation
204 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
205
206 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
207 // can run short a few iterations and may start a few iterations in.
208 // It will be RCE'd and unrolled and aligned.
209
210 // A following 'post' loop will run any remaining iterations. Used
211 // during Range Check Elimination, the 'post' loop will do any final
212 // iterations with full checks. Also used by Loop Unrolling, where
213 // the 'post' loop will do any epilog iterations needed. Basically,
214 // a 'post' loop can not profitably be further unrolled or RCE'd.
215
216 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
217 // it may do under-flow checks for RCE and may do alignment iterations
218 // so the following main loop 'knows' that it is striding down cache
219 // lines.
220
221 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
222 // Aligned, may be missing it's pre-loop.
223 int is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
224 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
225 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
226 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
227 int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
228 int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
229 int has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
230 int do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
231 int is_main_no_pre_loop () const { return _loop_flags & MainHasNoPreLoop; }
232 int is_vectorized_loop () const { return (_loop_flags & VectorizedLoop) == VectorizedLoop; }
233 int has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
234 int loop_has_range_checks () const { return (_loop_flags & HasRangeChecks) == HasRangeChecks; }
235 int loop_is_multiversioned() const { return (_loop_flags & IsMultiversioned) == IsMultiversioned; }
236 void set_main_no_pre_loop () { _loop_flags |= MainHasNoPreLoop; }
237
238 int main_idx() const { return _main_idx; }
239
240
241 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
242 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
243 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
244 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
245
246 void set_trip_count(uint tc) { _trip_count = tc; }
247 uint trip_count() { return _trip_count; }
248
249 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
250 void set_exact_trip_count(uint tc) {
251 _trip_count = tc;
252 _loop_flags |= HasExactTripCount;
253 }
254 void set_nonexact_trip_count() {
255 _loop_flags &= ~HasExactTripCount;
256 }
257 void set_notpassed_slp() {
258 _loop_flags &= ~PassedSlpAnalysis;
259 }
260
261 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
262 float profile_trip_cnt() { return _profile_trip_cnt; }
263
264 void double_unrolled_count() { _unrolled_count_log2++; }
265 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
266
267 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
268 int node_count_before_unroll() { return _node_count_before_unroll; }
269 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
270 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
271
272 #ifndef PRODUCT
273 virtual void dump_spec(outputStream *st) const;
274 #endif
275 };
276
277 //------------------------------CountedLoopEndNode-----------------------------
278 // CountedLoopEndNodes end simple trip counted loops. They act much like
279 // IfNodes.
280 class CountedLoopEndNode : public IfNode {
281 public:
282 enum { TestControl, TestValue };
283
284 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
285 : IfNode( control, test, prob, cnt) {
286 init_class_id(Class_CountedLoopEnd);
287 }
288 virtual int Opcode() const;
289
290 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
291 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
292 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
293 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
294 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
295 int stride_con() const;
296 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
297 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
298 PhiNode *phi() const {
299 Node *tmp = incr();
300 if (tmp && tmp->req() == 3) {
301 Node* phi = tmp->in(1);
302 if (phi->is_Phi()) {
303 return phi->as_Phi();
304 }
305 }
306 return NULL;
307 }
308 CountedLoopNode *loopnode() const {
309 // The CountedLoopNode that goes with this CountedLoopEndNode may
310 // have been optimized out by the IGVN so be cautious with the
311 // pattern matching on the graph
312 PhiNode* iv_phi = phi();
313 if (iv_phi == NULL) {
314 return NULL;
315 }
316 Node *ln = iv_phi->in(0);
317 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
318 return (CountedLoopNode*)ln;
319 }
320 return NULL;
321 }
322
323 #ifndef PRODUCT
324 virtual void dump_spec(outputStream *st) const;
325 #endif
326 };
327
328
329 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
330 Node *bc = back_control();
331 if( bc == NULL ) return NULL;
332 Node *le = bc->in(0);
333 if( le->Opcode() != Op_CountedLoopEnd )
334 return NULL;
335 return (CountedLoopEndNode*)le;
336 }
337 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
338 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
339 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
340 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
341 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
342 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
343 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
344
345 //------------------------------LoopLimitNode-----------------------------
346 // Counted Loop limit node which represents exact final iterator value:
347 // trip_count = (limit - init_trip + stride - 1)/stride
348 // final_value= trip_count * stride + init_trip.
349 // Use HW instructions to calculate it when it can overflow in integer.
350 // Note, final_value should fit into integer since counted loop has
351 // limit check: limit <= max_int-stride.
352 class LoopLimitNode : public Node {
353 enum { Init=1, Limit=2, Stride=3 };
354 public:
355 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
356 // Put it on the Macro nodes list to optimize during macro nodes expansion.
357 init_flags(Flag_is_macro);
358 C->add_macro_node(this);
359 }
360 virtual int Opcode() const;
361 virtual const Type *bottom_type() const { return TypeInt::INT; }
362 virtual uint ideal_reg() const { return Op_RegI; }
363 virtual const Type* Value(PhaseGVN* phase) const;
364 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
365 virtual Node* Identity(PhaseGVN* phase);
366 };
367
368 // -----------------------------IdealLoopTree----------------------------------
369 class IdealLoopTree : public ResourceObj {
370 public:
371 IdealLoopTree *_parent; // Parent in loop tree
372 IdealLoopTree *_next; // Next sibling in loop tree
373 IdealLoopTree *_child; // First child in loop tree
374
375 // The head-tail backedge defines the loop.
376 // If tail is NULL then this loop has multiple backedges as part of the
377 // same loop. During cleanup I'll peel off the multiple backedges; merge
378 // them at the loop bottom and flow 1 real backedge into the loop.
379 Node *_head; // Head of loop
380 Node *_tail; // Tail of loop
381 inline Node *tail(); // Handle lazy update of _tail field
382 PhaseIdealLoop* _phase;
383 int _local_loop_unroll_limit;
384 int _local_loop_unroll_factor;
385
386 Node_List _body; // Loop body for inner loops
387
388 uint8_t _nest; // Nesting depth
389 uint8_t _irreducible:1, // True if irreducible
390 _has_call:1, // True if has call safepoint
391 _has_sfpt:1, // True if has non-call safepoint
392 _rce_candidate:1; // True if candidate for range check elimination
393
394 Node_List* _safepts; // List of safepoints in this loop
395 Node_List* _required_safept; // A inner loop cannot delete these safepts;
396 bool _allow_optimizations; // Allow loop optimizations
397
398 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
399 : _parent(0), _next(0), _child(0),
400 _head(head), _tail(tail),
401 _phase(phase),
402 _safepts(NULL),
403 _required_safept(NULL),
404 _allow_optimizations(true),
405 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
406 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
407 { }
408
409 // Is 'l' a member of 'this'?
410 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
411
412 // Set loop nesting depth. Accumulate has_call bits.
413 int set_nest( uint depth );
414
415 // Split out multiple fall-in edges from the loop header. Move them to a
416 // private RegionNode before the loop. This becomes the loop landing pad.
417 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
418
419 // Split out the outermost loop from this shared header.
420 void split_outer_loop( PhaseIdealLoop *phase );
421
422 // Merge all the backedges from the shared header into a private Region.
423 // Feed that region as the one backedge to this loop.
424 void merge_many_backedges( PhaseIdealLoop *phase );
425
426 // Split shared headers and insert loop landing pads.
427 // Insert a LoopNode to replace the RegionNode.
428 // Returns TRUE if loop tree is structurally changed.
429 bool beautify_loops( PhaseIdealLoop *phase );
430
431 // Perform optimization to use the loop predicates for null checks and range checks.
432 // Applies to any loop level (not just the innermost one)
433 bool loop_predication( PhaseIdealLoop *phase);
434
435 // Perform iteration-splitting on inner loops. Split iterations to
436 // avoid range checks or one-shot null checks. Returns false if the
437 // current round of loop opts should stop.
438 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
439
440 // Driver for various flavors of iteration splitting. Returns false
441 // if the current round of loop opts should stop.
442 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
443
444 // Given dominators, try to find loops with calls that must always be
445 // executed (call dominates loop tail). These loops do not need non-call
446 // safepoints (ncsfpt).
447 void check_safepts(VectorSet &visited, Node_List &stack);
448
449 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
450 // encountered.
451 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
452
453 // Remove safepoints from loop. Optionally keeping one.
454 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
455
456 // Convert to counted loops where possible
457 void counted_loop( PhaseIdealLoop *phase );
458
459 // Check for Node being a loop-breaking test
460 Node *is_loop_exit(Node *iff) const;
461
462 // Returns true if ctrl is executed on every complete iteration
463 bool dominates_backedge(Node* ctrl);
464
465 // Remove simplistic dead code from loop body
466 void DCE_loop_body();
467
468 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
469 // Replace with a 1-in-10 exit guess.
470 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
471
472 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
473 // Useful for unrolling loops with NO array accesses.
474 bool policy_peel_only( PhaseIdealLoop *phase ) const;
475
476 // Return TRUE or FALSE if the loop should be unswitched -- clone
477 // loop with an invariant test
478 bool policy_unswitching( PhaseIdealLoop *phase ) const;
479
480 // Micro-benchmark spamming. Remove empty loops.
481 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
482
483 // Convert one iteration loop into normal code.
484 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
485
486 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
487 // make some loop-invariant test (usually a null-check) happen before the
488 // loop.
489 bool policy_peeling( PhaseIdealLoop *phase ) const;
490
491 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
492 // known trip count in the counted loop node.
493 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
494
495 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
496 // the loop is a CountedLoop and the body is small enough.
497 bool policy_unroll(PhaseIdealLoop *phase);
498
499 // Loop analyses to map to a maximal superword unrolling for vectorization.
500 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
501
502 // Return TRUE or FALSE if the loop should be range-check-eliminated.
503 // Gather a list of IF tests that are dominated by iteration splitting;
504 // also gather the end of the first split and the start of the 2nd split.
505 bool policy_range_check( PhaseIdealLoop *phase ) const;
506
507 // Return TRUE or FALSE if the loop should be cache-line aligned.
508 // Gather the expression that does the alignment. Note that only
509 // one array base can be aligned in a loop (unless the VM guarantees
510 // mutual alignment). Note that if we vectorize short memory ops
511 // into longer memory ops, we may want to increase alignment.
512 bool policy_align( PhaseIdealLoop *phase ) const;
513
514 // Return TRUE if "iff" is a range check.
515 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
516
517 // Compute loop exact trip count if possible
518 void compute_exact_trip_count( PhaseIdealLoop *phase );
519
520 // Compute loop trip count from profile data
521 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
522
523 // Reassociate invariant expressions.
524 void reassociate_invariants(PhaseIdealLoop *phase);
525 // Reassociate invariant add and subtract expressions.
526 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
527 // Return nonzero index of invariant operand if invariant and variant
528 // are combined with an Add or Sub. Helper for reassociate_invariants.
529 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
530
531 // Return true if n is invariant
532 bool is_invariant(Node* n) const;
533
534 // Put loop body on igvn work list
535 void record_for_igvn();
536
537 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
538 bool is_inner() { return is_loop() && _child == NULL; }
539 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
540
541 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
542
543 #ifndef PRODUCT
544 void dump_head( ) const; // Dump loop head only
545 void dump() const; // Dump this loop recursively
546 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
547 #endif
548
549 };
550
551 class PostLoopInfo {
552 public:
553 Node *new_main_exit;
554 CountedLoopNode *post_head;
555
556 PostLoopInfo() { init(); }
557
558 void init() { new_main_exit = NULL; post_head = NULL; }
559 };
560
561 // -----------------------------PhaseIdealLoop---------------------------------
562 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
563 // loop tree. Drives the loop-based transformations on the ideal graph.
564 class PhaseIdealLoop : public PhaseTransform {
565 friend class IdealLoopTree;
566 friend class SuperWord;
567 friend class CountedLoopReserveKit;
568
569 // Pre-computed def-use info
570 PhaseIterGVN &_igvn;
571
572 // Head of loop tree
573 IdealLoopTree *_ltree_root;
574
575 // Array of pre-order numbers, plus post-visited bit.
576 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
577 // ODD for post-visited. Other bits are the pre-order number.
578 uint *_preorders;
579 uint _max_preorder;
580
581 const PhaseIdealLoop* _verify_me;
582 bool _verify_only;
583
584 // Allocate _preorders[] array
585 void allocate_preorders() {
586 _max_preorder = C->unique()+8;
587 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
588 memset(_preorders, 0, sizeof(uint) * _max_preorder);
589 }
590
591 // Allocate _preorders[] array
592 void reallocate_preorders() {
593 if ( _max_preorder < C->unique() ) {
594 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
595 _max_preorder = C->unique();
596 }
597 memset(_preorders, 0, sizeof(uint) * _max_preorder);
598 }
599
600 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
601 // adds new nodes.
602 void check_grow_preorders( ) {
603 if ( _max_preorder < C->unique() ) {
604 uint newsize = _max_preorder<<1; // double size of array
605 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
606 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
607 _max_preorder = newsize;
608 }
609 }
610 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
611 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
612 // Pre-order numbers are written to the Nodes array as low-bit-set values.
613 void set_preorder_visited( Node *n, int pre_order ) {
614 assert( !is_visited( n ), "already set" );
615 _preorders[n->_idx] = (pre_order<<1);
616 };
617 // Return pre-order number.
618 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
619
620 // Check for being post-visited.
621 // Should be previsited already (checked with assert(is_visited(n))).
622 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
623
624 // Mark as post visited
625 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
626
627 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
628 // Returns true if "n" is a data node, false if it's a control node.
629 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
630
631 // clear out dead code after build_loop_late
632 Node_List _deadlist;
633
634 // Support for faster execution of get_late_ctrl()/dom_lca()
635 // when a node has many uses and dominator depth is deep.
636 Node_Array _dom_lca_tags;
637 void init_dom_lca_tags();
638 void clear_dom_lca_tags();
639
640 // Helper for debugging bad dominance relationships
641 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
642
643 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
644
645 // Inline wrapper for frequent cases:
646 // 1) only one use
647 // 2) a use is the same as the current LCA passed as 'n1'
648 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
649 assert( n->is_CFG(), "" );
650 // Fast-path NULL lca
651 if( lca != NULL && lca != n ) {
652 assert( lca->is_CFG(), "" );
653 // find LCA of all uses
654 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
655 }
656 return find_non_split_ctrl(n);
657 }
658 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
659
660 // Helper function for directing control inputs away from CFG split
661 // points.
662 Node *find_non_split_ctrl( Node *ctrl ) const {
663 if (ctrl != NULL) {
664 if (ctrl->is_MultiBranch()) {
665 ctrl = ctrl->in(0);
666 }
667 assert(ctrl->is_CFG(), "CFG");
668 }
669 return ctrl;
670 }
671
672 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
673
674 public:
675 bool has_node( Node* n ) const {
676 guarantee(n != NULL, "No Node.");
677 return _nodes[n->_idx] != NULL;
678 }
679 // check if transform created new nodes that need _ctrl recorded
680 Node *get_late_ctrl( Node *n, Node *early );
681 Node *get_early_ctrl( Node *n );
682 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
683 void set_early_ctrl( Node *n );
684 void set_subtree_ctrl( Node *root );
685 void set_ctrl( Node *n, Node *ctrl ) {
686 assert( !has_node(n) || has_ctrl(n), "" );
687 assert( ctrl->in(0), "cannot set dead control node" );
688 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
689 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
690 }
691 // Set control and update loop membership
692 void set_ctrl_and_loop(Node* n, Node* ctrl) {
693 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
694 IdealLoopTree* new_loop = get_loop(ctrl);
695 if (old_loop != new_loop) {
696 if (old_loop->_child == NULL) old_loop->_body.yank(n);
697 if (new_loop->_child == NULL) new_loop->_body.push(n);
698 }
699 set_ctrl(n, ctrl);
700 }
701 // Control nodes can be replaced or subsumed. During this pass they
702 // get their replacement Node in slot 1. Instead of updating the block
703 // location of all Nodes in the subsumed block, we lazily do it. As we
704 // pull such a subsumed block out of the array, we write back the final
705 // correct block.
706 Node *get_ctrl( Node *i ) {
707 assert(has_node(i), "");
708 Node *n = get_ctrl_no_update(i);
709 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
710 assert(has_node(i) && has_ctrl(i), "");
711 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
712 return n;
713 }
714 // true if CFG node d dominates CFG node n
715 bool is_dominator(Node *d, Node *n);
716 // return get_ctrl for a data node and self(n) for a CFG node
717 Node* ctrl_or_self(Node* n) {
718 if (has_ctrl(n))
719 return get_ctrl(n);
720 else {
721 assert (n->is_CFG(), "must be a CFG node");
722 return n;
723 }
724 }
725
726 private:
727 Node *get_ctrl_no_update_helper(Node *i) const {
728 assert(has_ctrl(i), "should be control, not loop");
729 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
730 }
731
732 Node *get_ctrl_no_update(Node *i) const {
733 assert( has_ctrl(i), "" );
734 Node *n = get_ctrl_no_update_helper(i);
735 if (!n->in(0)) {
736 // Skip dead CFG nodes
737 do {
738 n = get_ctrl_no_update_helper(n);
739 } while (!n->in(0));
740 n = find_non_split_ctrl(n);
741 }
742 return n;
743 }
744
745 // Check for loop being set
746 // "n" must be a control node. Returns true if "n" is known to be in a loop.
747 bool has_loop( Node *n ) const {
748 assert(!has_node(n) || !has_ctrl(n), "");
749 return has_node(n);
750 }
751 // Set loop
752 void set_loop( Node *n, IdealLoopTree *loop ) {
753 _nodes.map(n->_idx, (Node*)loop);
754 }
755 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
756 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
757 // from old_node to new_node to support the lazy update. Reference
758 // replaces loop reference, since that is not needed for dead node.
759 public:
760 void lazy_update(Node *old_node, Node *new_node) {
761 assert(old_node != new_node, "no cycles please");
762 // Re-use the side array slot for this node to provide the
763 // forwarding pointer.
764 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
765 }
766 void lazy_replace(Node *old_node, Node *new_node) {
767 _igvn.replace_node(old_node, new_node);
768 lazy_update(old_node, new_node);
769 }
770
771 private:
772
773 // Place 'n' in some loop nest, where 'n' is a CFG node
774 void build_loop_tree();
775 int build_loop_tree_impl( Node *n, int pre_order );
776 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
777 // loop tree, not the root.
778 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
779
780 // Place Data nodes in some loop nest
781 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
782 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
783 void build_loop_late_post ( Node* n );
784
785 // Array of immediate dominance info for each CFG node indexed by node idx
786 private:
787 uint _idom_size;
788 Node **_idom; // Array of immediate dominators
789 uint *_dom_depth; // Used for fast LCA test
790 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
791
792 Node* idom_no_update(Node* d) const {
793 assert(d->_idx < _idom_size, "oob");
794 Node* n = _idom[d->_idx];
795 assert(n != NULL,"Bad immediate dominator info.");
796 while (n->in(0) == NULL) { // Skip dead CFG nodes
797 //n = n->in(1);
798 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
799 assert(n != NULL,"Bad immediate dominator info.");
800 }
801 return n;
802 }
803 Node *idom(Node* d) const {
804 uint didx = d->_idx;
805 Node *n = idom_no_update(d);
806 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
807 return n;
808 }
809 uint dom_depth(Node* d) const {
810 guarantee(d != NULL, "Null dominator info.");
811 guarantee(d->_idx < _idom_size, "");
812 return _dom_depth[d->_idx];
813 }
814 void set_idom(Node* d, Node* n, uint dom_depth);
815 // Locally compute IDOM using dom_lca call
816 Node *compute_idom( Node *region ) const;
817 // Recompute dom_depth
818 void recompute_dom_depth();
819
820 // Is safept not required by an outer loop?
821 bool is_deleteable_safept(Node* sfpt);
822
823 // Replace parallel induction variable (parallel to trip counter)
824 void replace_parallel_iv(IdealLoopTree *loop);
825
826 // Perform verification that the graph is valid.
827 PhaseIdealLoop( PhaseIterGVN &igvn) :
828 PhaseTransform(Ideal_Loop),
829 _igvn(igvn),
830 _dom_lca_tags(arena()), // Thread::resource_area
831 _verify_me(NULL),
832 _verify_only(true) {
833 build_and_optimize(false, false);
834 }
835
836 // build the loop tree and perform any requested optimizations
837 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
838
839 public:
840 // Dominators for the sea of nodes
841 void Dominators();
842 Node *dom_lca( Node *n1, Node *n2 ) const {
843 return find_non_split_ctrl(dom_lca_internal(n1, n2));
844 }
845 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
846
847 // Compute the Ideal Node to Loop mapping
848 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
849 PhaseTransform(Ideal_Loop),
850 _igvn(igvn),
851 _dom_lca_tags(arena()), // Thread::resource_area
852 _verify_me(NULL),
853 _verify_only(false) {
854 build_and_optimize(do_split_ifs, skip_loop_opts);
855 }
856
857 // Verify that verify_me made the same decisions as a fresh run.
858 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
859 PhaseTransform(Ideal_Loop),
860 _igvn(igvn),
861 _dom_lca_tags(arena()), // Thread::resource_area
862 _verify_me(verify_me),
863 _verify_only(false) {
864 build_and_optimize(false, false);
865 }
866
867 // Build and verify the loop tree without modifying the graph. This
868 // is useful to verify that all inputs properly dominate their uses.
869 static void verify(PhaseIterGVN& igvn) {
870 #ifdef ASSERT
871 PhaseIdealLoop v(igvn);
872 #endif
873 }
874
875 // True if the method has at least 1 irreducible loop
876 bool _has_irreducible_loops;
877
878 // Per-Node transform
879 virtual Node *transform( Node *a_node ) { return 0; }
880
881 bool is_counted_loop( Node *x, IdealLoopTree *loop );
882
883 Node* exact_limit( IdealLoopTree *loop );
884
885 // Return a post-walked LoopNode
886 IdealLoopTree *get_loop( Node *n ) const {
887 // Dead nodes have no loop, so return the top level loop instead
888 if (!has_node(n)) return _ltree_root;
889 assert(!has_ctrl(n), "");
890 return (IdealLoopTree*)_nodes[n->_idx];
891 }
892
893 // Is 'n' a (nested) member of 'loop'?
894 int is_member( const IdealLoopTree *loop, Node *n ) const {
895 return loop->is_member(get_loop(n)); }
896
897 // This is the basic building block of the loop optimizations. It clones an
898 // entire loop body. It makes an old_new loop body mapping; with this
899 // mapping you can find the new-loop equivalent to an old-loop node. All
900 // new-loop nodes are exactly equal to their old-loop counterparts, all
901 // edges are the same. All exits from the old-loop now have a RegionNode
902 // that merges the equivalent new-loop path. This is true even for the
903 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
904 // now come from (one or more) Phis that merge their new-loop equivalents.
905 // Parameter side_by_side_idom:
906 // When side_by_size_idom is NULL, the dominator tree is constructed for
907 // the clone loop to dominate the original. Used in construction of
908 // pre-main-post loop sequence.
909 // When nonnull, the clone and original are side-by-side, both are
910 // dominated by the passed in side_by_side_idom node. Used in
911 // construction of unswitched loops.
912 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
913 Node* side_by_side_idom = NULL);
914
915 // If we got the effect of peeling, either by actually peeling or by
916 // making a pre-loop which must execute at least once, we can remove
917 // all loop-invariant dominated tests in the main body.
918 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
919
920 // Generate code to do a loop peel for the given loop (and body).
921 // old_new is a temp array.
922 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
923
924 // Add pre and post loops around the given loop. These loops are used
925 // during RCE, unrolling and aligning loops.
926 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
927
928 // Add post loop after the given loop.
929 void insert_post_loop( IdealLoopTree *loop, Node_List &old_new,
930 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
931 Node *incr, Node *limit, PostLoopInfo &post_loop_info );
932
933 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
934 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
935
936 // Add a vector post loop between a vector main loop and the current post loop
937 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
938
939 // If Node n lives in the back_ctrl block, we clone a private version of n
940 // in preheader_ctrl block and return that, otherwise return n.
941 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
942
943 // Take steps to maximally unroll the loop. Peel any odd iterations, then
944 // unroll to do double iterations. The next round of major loop transforms
945 // will repeat till the doubled loop body does all remaining iterations in 1
946 // pass.
947 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
948
949 // Unroll the loop body one step - make each trip do 2 iterations.
950 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
951
952 // Mark vector reduction candidates before loop unrolling
953 void mark_reductions( IdealLoopTree *loop );
954
955 // Return true if exp is a constant times an induction var
956 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
957
958 // Return true if exp is a scaled induction var plus (or minus) constant
959 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
960
961 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
962 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
963 Deoptimization::DeoptReason reason,
964 int opcode);
965 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
966
967 // Clone loop predicates to cloned loops (peeled, unswitched)
968 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
969 Deoptimization::DeoptReason reason,
970 PhaseIdealLoop* loop_phase,
971 PhaseIterGVN* igvn);
972
973 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
974 bool clone_limit_check,
975 PhaseIdealLoop* loop_phase,
976 PhaseIterGVN* igvn);
977 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
978
979 static Node* skip_loop_predicates(Node* entry);
980
981 // Find a good location to insert a predicate
982 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
983 // Find a predicate
984 static Node* find_predicate(Node* entry);
985 // Construct a range check for a predicate if
986 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
987 int scale, Node* offset,
988 Node* init, Node* limit, Node* stride,
989 Node* range, bool upper);
990
991 // Implementation of the loop predication to promote checks outside the loop
992 bool loop_predication_impl(IdealLoopTree *loop);
993
994 // Helper function to collect predicate for eliminating the useless ones
995 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
996 void eliminate_useless_predicates();
997
998 // Change the control input of expensive nodes to allow commoning by
999 // IGVN when it is guaranteed to not result in a more frequent
1000 // execution of the expensive node. Return true if progress.
1001 bool process_expensive_nodes();
1002
1003 // Check whether node has become unreachable
1004 bool is_node_unreachable(Node *n) const {
1005 return !has_node(n) || n->is_unreachable(_igvn);
1006 }
1007
1008 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1009 void do_range_check(IdealLoopTree *loop, Node_List &range_check_list);
1010
1011 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1012 void has_range_checks(IdealLoopTree *loop);
1013
1014 // Process post loops which have range checks and try to build a multi-version
1015 // guard to safely determine if we can execute the post loop which was RCE'd.
1016 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1017
1018 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1019 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1020
1021 // Create a slow version of the loop by cloning the loop
1022 // and inserting an if to select fast-slow versions.
1023 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1024 Node_List &old_new,
1025 int opcode);
1026
1027 // Clone a loop and return the clone head (clone_loop_head).
1028 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1029 // This routine was created for usage in CountedLoopReserveKit.
1030 //
1031 // int(1) -> If -> IfTrue -> original_loop_head
1032 // |
1033 // V
1034 // IfFalse -> clone_loop_head (returned by function pointer)
1035 //
1036 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1037 // Clone loop with an invariant test (that does not exit) and
1038 // insert a clone of the test that selects which version to
1039 // execute.
1040 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1041
1042 // Find candidate "if" for unswitching
1043 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1044
1045 // Range Check Elimination uses this function!
1046 // Constrain the main loop iterations so the affine function:
1047 // low_limit <= scale_con * I + offset < upper_limit
1048 // always holds true. That is, either increase the number of iterations in
1049 // the pre-loop or the post-loop until the condition holds true in the main
1050 // loop. Scale_con, offset and limit are all loop invariant.
1051 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 );
1052 // Helper function for add_constraint().
1053 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1054
1055 // Partially peel loop up through last_peel node.
1056 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1057
1058 // Create a scheduled list of nodes control dependent on ctrl set.
1059 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1060 // Has a use in the vector set
1061 bool has_use_in_set( Node* n, VectorSet& vset );
1062 // Has use internal to the vector set (ie. not in a phi at the loop head)
1063 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1064 // clone "n" for uses that are outside of loop
1065 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1066 // clone "n" for special uses that are in the not_peeled region
1067 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1068 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1069 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1070 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1071 #ifdef ASSERT
1072 // Validate the loop partition sets: peel and not_peel
1073 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1074 // Ensure that uses outside of loop are of the right form
1075 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1076 uint orig_exit_idx, uint clone_exit_idx);
1077 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1078 #endif
1079
1080 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1081 int stride_of_possible_iv( Node* iff );
1082 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1083 // Return the (unique) control output node that's in the loop (if it exists.)
1084 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1085 // Insert a signed compare loop exit cloned from an unsigned compare.
1086 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1087 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1088 // Utility to register node "n" with PhaseIdealLoop
1089 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1090 // Utility to create an if-projection
1091 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1092 // Force the iff control output to be the live_proj
1093 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1094 // Insert a region before an if projection
1095 RegionNode* insert_region_before_proj(ProjNode* proj);
1096 // Insert a new if before an if projection
1097 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1098
1099 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1100 // "Nearly" because all Nodes have been cloned from the original in the loop,
1101 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1102 // through the Phi recursively, and return a Bool.
1103 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1104 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1105
1106
1107 // Rework addressing expressions to get the most loop-invariant stuff
1108 // moved out. We'd like to do all associative operators, but it's especially
1109 // important (common) to do address expressions.
1110 Node *remix_address_expressions( Node *n );
1111
1112 // Attempt to use a conditional move instead of a phi/branch
1113 Node *conditional_move( Node *n );
1114
1115 // Reorganize offset computations to lower register pressure.
1116 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1117 // (which are then alive with the post-incremented trip counter
1118 // forcing an extra register move)
1119 void reorg_offsets( IdealLoopTree *loop );
1120
1121 // Check for aggressive application of 'split-if' optimization,
1122 // using basic block level info.
1123 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1124 Node *split_if_with_blocks_pre ( Node *n );
1125 void split_if_with_blocks_post( Node *n );
1126 Node *has_local_phi_input( Node *n );
1127 // Mark an IfNode as being dominated by a prior test,
1128 // without actually altering the CFG (and hence IDOM info).
1129 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1130
1131 // Split Node 'n' through merge point
1132 Node *split_thru_region( Node *n, Node *region );
1133 // Split Node 'n' through merge point if there is enough win.
1134 Node *split_thru_phi( Node *n, Node *region, int policy );
1135 // Found an If getting its condition-code input from a Phi in the
1136 // same block. Split thru the Region.
1137 void do_split_if( Node *iff );
1138
1139 // Conversion of fill/copy patterns into intrisic versions
1140 bool do_intrinsify_fill();
1141 bool intrinsify_fill(IdealLoopTree* lpt);
1142 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1143 Node*& shift, Node*& offset);
1144
1145 private:
1146 // Return a type based on condition control flow
1147 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1148 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1149 // Helpers for filtered type
1150 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1151
1152 // Helper functions
1153 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1154 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1155 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 );
1156 bool split_up( Node *n, Node *blk1, Node *blk2 );
1157 void sink_use( Node *use, Node *post_loop );
1158 Node *place_near_use( Node *useblock ) const;
1159 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1160 void try_move_store_after_loop(Node* n);
1161 bool identical_backtoback_ifs(Node *n);
1162 bool can_split_if(Node *n_ctrl);
1163
1164 bool _created_loop_node;
1165 public:
1166 void set_created_loop_node() { _created_loop_node = true; }
1167 bool created_loop_node() { return _created_loop_node; }
1168 void register_new_node( Node *n, Node *blk );
1169
1170 #ifdef ASSERT
1171 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1172 #endif
1173
1174 #ifndef PRODUCT
1175 void dump( ) const;
1176 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1177 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1178 void verify() const; // Major slow :-)
1179 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1180 IdealLoopTree *get_loop_idx(Node* n) const {
1181 // Dead nodes have no loop, so return the top level loop instead
1182 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1183 }
1184 // Print some stats
1185 static void print_statistics();
1186 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1187 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1188 #endif
1189 };
1190
1191 // This kit may be used for making of a reserved copy of a loop before this loop
1192 // goes under non-reversible changes.
1193 //
1194 // Function create_reserve() creates a reserved copy (clone) of the loop.
1195 // The reserved copy is created by calling
1196 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1197 // the original and reserved loops are connected in the outer graph.
1198 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1199 //
1200 // By default the reserved copy (clone) of the loop is created as dead code - it is
1201 // dominated in the outer loop by this node chain:
1202 // intcon(1)->If->IfFalse->reserved_copy.
1203 // The original loop is dominated by the the same node chain but IfTrue projection:
1204 // intcon(0)->If->IfTrue->original_loop.
1205 //
1206 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1207 // and the dtor, checks _use_new value.
1208 // If _use_new == false, it "switches" control to reserved copy of the loop
1209 // by simple replacing of node intcon(1) with node intcon(0).
1210 //
1211 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1212 //
1213 // void CountedLoopReserveKit_example()
1214 // {
1215 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1216 // if (DoReserveCopy && !lrk.has_reserved()) {
1217 // return; //failed to create reserved loop copy
1218 // }
1219 // ...
1220 // //something is wrong, switch to original loop
1221 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1222 // ...
1223 // //everything worked ok, return with the newly modified loop
1224 // lrk.use_new();
1225 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1226 // }
1227 //
1228 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1229 // the dtor will "switch to the original" loop.
1230 // NOTE. You you modify outside of the original loop this class is no help.
1231 //
1232 class CountedLoopReserveKit {
1233 private:
1234 PhaseIdealLoop* _phase;
1235 IdealLoopTree* _lpt;
1236 LoopNode* _lp;
1237 IfNode* _iff;
1238 LoopNode* _lp_reserved;
1239 bool _has_reserved;
1240 bool _use_new;
1241 const bool _active; //may be set to false in ctor, then the object is dummy
1242
1243 public:
1244 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1245 ~CountedLoopReserveKit();
1246 void use_new() {_use_new = true;}
1247 void set_iff(IfNode* x) {_iff = x;}
1248 bool has_reserved() const { return _active && _has_reserved;}
1249 private:
1250 bool create_reserve();
1251 };// class CountedLoopReserveKit
1252
1253 inline Node* IdealLoopTree::tail() {
1254 // Handle lazy update of _tail field
1255 Node *n = _tail;
1256 //while( !n->in(0) ) // Skip dead CFG nodes
1257 //n = n->in(1);
1258 if (n->in(0) == NULL)
1259 n = _phase->get_ctrl(n);
1260 _tail = n;
1261 return n;
1262 }
1263
1264
1265 // Iterate over the loop tree using a preorder, left-to-right traversal.
1266 //
1267 // Example that visits all counted loops from within PhaseIdealLoop
1268 //
1269 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1270 // IdealLoopTree* lpt = iter.current();
1271 // if (!lpt->is_counted()) continue;
1272 // ...
1273 class LoopTreeIterator : public StackObj {
1274 private:
1275 IdealLoopTree* _root;
1276 IdealLoopTree* _curnt;
1277
1278 public:
1279 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1280
1281 bool done() { return _curnt == NULL; } // Finished iterating?
1282
1283 void next(); // Advance to next loop tree
1284
1285 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1286 };
1287
1288 #endif // SHARE_VM_OPTO_LOOPNODE_HPP