1 /*
2 * Copyright (c) 1998, 2026, 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_OPTO_LOOPNODE_HPP
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class UnswitchCandidate;
47 class LoopSelector;
48 class ReachabilityFenceNode;
49 class UnswitchedLoopSelector;
50 class VectorSet;
51 class VSharedData;
52 class Invariance;
53 struct small_cache;
54
55 //
56 // I D E A L I Z E D L O O P S
57 //
58 // Idealized loops are the set of loops I perform more interesting
59 // transformations on, beyond simple hoisting.
60
61 //------------------------------LoopNode---------------------------------------
62 // Simple loop header. Fall in path on left, loop-back path on right.
63 class LoopNode : public RegionNode {
64 // Size is bigger to hold the flags. However, the flags do not change
65 // the semantics so it does not appear in the hash & cmp functions.
66 virtual uint size_of() const { return sizeof(*this); }
67 protected:
68 uint _loop_flags;
69 // Names for flag bitfields
70 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
71 MainHasNoPreLoop = 1<<2,
72 HasExactTripCount = 1<<3,
73 InnerLoop = 1<<4,
74 PartialPeelLoop = 1<<5,
75 PartialPeelFailed = 1<<6,
76 WasSlpAnalyzed = 1<<7,
77 PassedSlpAnalysis = 1<<8,
78 DoUnrollOnly = 1<<9,
79 VectorizedLoop = 1<<10,
80 HasAtomicPostLoop = 1<<11,
81 StripMined = 1<<12,
82 SubwordLoop = 1<<13,
83 ProfileTripFailed = 1<<14,
84 LoopNestInnerLoop = 1<<15,
85 LoopNestLongOuterLoop = 1<<16,
86 MultiversionFastLoop = 1<<17,
87 MultiversionSlowLoop = 2<<17,
88 MultiversionDelayedSlowLoop = 3<<17,
89 MultiversionFlagsMask = 3<<17,
90 FlatArrays = 1<<18};
91 char _unswitch_count;
92 enum { _unswitch_max=3 };
93
94 // Expected trip count from profile data
95 float _profile_trip_cnt;
96
97 public:
98 // Names for edge indices
99 enum { Self=0, EntryControl, LoopBackControl };
100
101 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
102 void set_inner_loop() { _loop_flags |= InnerLoop; }
103
104 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
105 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
106 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
107 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
108 bool is_strip_mined() const { return _loop_flags & StripMined; }
109 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
110 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
111 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
112 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
113 bool is_flat_arrays() const { return _loop_flags & FlatArrays; }
114
115 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
116 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
117 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
118 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
119 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
120 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
121 void mark_strip_mined() { _loop_flags |= StripMined; }
122 void clear_strip_mined() { _loop_flags &= ~StripMined; }
123 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
124 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
125 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
126 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
127 void mark_flat_arrays() { _loop_flags |= FlatArrays; }
128
129 int unswitch_max() { return _unswitch_max; }
130 int unswitch_count() { return _unswitch_count; }
131
132 void set_unswitch_count(int val) {
133 assert (val <= unswitch_max(), "too many unswitches");
134 _unswitch_count = val;
135 }
136
137 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
138 float profile_trip_cnt() { return _profile_trip_cnt; }
139
140 #ifndef PRODUCT
141 uint _stress_peeling_attempts = 0;
142 #endif
143
144 LoopNode(Node *entry, Node *backedge)
145 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
146 _profile_trip_cnt(COUNT_UNKNOWN) {
147 init_class_id(Class_Loop);
148 init_req(EntryControl, entry);
149 init_req(LoopBackControl, backedge);
150 }
151
152 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
153 virtual int Opcode() const;
154 bool can_be_counted_loop(PhaseValues* phase) const {
155 return req() == 3 && in(0) != nullptr &&
156 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
157 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
158 }
159 bool is_valid_counted_loop(BasicType bt) const;
160 #ifndef PRODUCT
161 virtual void dump_spec(outputStream *st) const;
162 #endif
163
164 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
165 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
166 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
167 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
168 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
169 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
170 };
171
172 //------------------------------Counted Loops----------------------------------
173 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
174 // path (and maybe some other exit paths). The trip-counter exit is always
175 // last in the loop. The trip-counter have to stride by a constant;
176 // the exit value is also loop invariant.
177
178 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
179 // CountedLoopNode has the incoming loop control and the loop-back-control
180 // which is always the IfTrue before the matching CountedLoopEndNode. The
181 // CountedLoopEndNode has an incoming control (possibly not the
182 // CountedLoopNode if there is control flow in the loop), the post-increment
183 // trip-counter value, and the limit. The trip-counter value is always of
184 // the form (Op old-trip-counter stride). The old-trip-counter is produced
185 // by a Phi connected to the CountedLoopNode. The stride is constant.
186 // The Op is any commutable opcode, including Add, Mul, Xor. The
187 // CountedLoopEndNode also takes in the loop-invariant limit value.
188
189 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
190 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
191 // via the old-trip-counter from the Op node.
192
193 //------------------------------CountedLoopNode--------------------------------
194 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
195 // inputs the incoming loop-start control and the loop-back control, so they
196 // act like RegionNodes. They also take in the initial trip counter, the
197 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
198 // produce a loop-body control and the trip counter value. Since
199 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
200
201 class BaseCountedLoopNode : public LoopNode {
202 public:
203 BaseCountedLoopNode(Node *entry, Node *backedge)
204 : LoopNode(entry, backedge) {
205 }
206
207 Node *init_control() const { return in(EntryControl); }
208 Node *back_control() const { return in(LoopBackControl); }
209
210 Node* init_trip() const;
211 Node* stride() const;
212 bool stride_is_con() const;
213 Node* limit() const;
214 Node* incr() const;
215 Node* phi() const;
216
217 BaseCountedLoopEndNode* loopexit_or_null() const;
218 BaseCountedLoopEndNode* loopexit() const;
219
220 virtual BasicType bt() const = 0;
221
222 jlong stride_con() const;
223
224 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
225
226 virtual void set_trip_count(julong tc) = 0;
227 virtual julong trip_count() const = 0;
228
229 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
230 void set_exact_trip_count(julong tc) {
231 set_trip_count(tc);
232 _loop_flags |= HasExactTripCount;
233 }
234 void set_nonexact_trip_count() {
235 _loop_flags &= ~HasExactTripCount;
236 }
237 };
238
239
240 class CountedLoopNode : public BaseCountedLoopNode {
241 // Size is bigger to hold _main_idx. However, _main_idx does not change
242 // the semantics so it does not appear in the hash & cmp functions.
243 virtual uint size_of() const { return sizeof(*this); }
244
245 // For Pre- and Post-loops during debugging ONLY, this holds the index of
246 // the Main CountedLoop. Used to assert that we understand the graph shape.
247 node_idx_t _main_idx;
248
249 // Known trip count calculated by compute_exact_trip_count()
250 uint _trip_count;
251
252 // Log2 of original loop bodies in unrolled loop
253 int _unrolled_count_log2;
254
255 // Node count prior to last unrolling - used to decide if
256 // unroll,optimize,unroll,optimize,... is making progress
257 int _node_count_before_unroll;
258
259 // If slp analysis is performed we record the maximum
260 // vector mapped unroll factor here
261 int _slp_maximum_unroll_factor;
262
263 public:
264 CountedLoopNode(Node *entry, Node *backedge)
265 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
266 _unrolled_count_log2(0), _node_count_before_unroll(0),
267 _slp_maximum_unroll_factor(0) {
268 init_class_id(Class_CountedLoop);
269 // Initialize _trip_count to the largest possible value.
270 // Will be reset (lower) if the loop's trip count is known.
271 }
272
273 virtual int Opcode() const;
274 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
275
276 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
277 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
278 int stride_con() const;
279
280 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
281 // can run short a few iterations and may start a few iterations in.
282 // It will be RCE'd and unrolled and aligned.
283
284 // A following 'post' loop will run any remaining iterations. Used
285 // during Range Check Elimination, the 'post' loop will do any final
286 // iterations with full checks. Also used by Loop Unrolling, where
287 // the 'post' loop will do any epilog iterations needed. Basically,
288 // a 'post' loop can not profitably be further unrolled or RCE'd.
289
290 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
291 // it may do under-flow checks for RCE and may do alignment iterations
292 // so the following main loop 'knows' that it is striding down cache
293 // lines.
294
295 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
296 // Aligned, may be missing it's pre-loop.
297 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
298 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
299 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
300 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
301 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
302 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
303 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
304 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
305 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
306 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
307
308 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
309
310 int main_idx() const { return _main_idx; }
311
312 void set_trip_count(julong tc) {
313 assert(tc < max_juint, "Cannot set trip count to max_juint");
314 _trip_count = checked_cast<uint>(tc);
315 }
316 julong trip_count() const { return _trip_count; }
317
318 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
319 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
320 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
321 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
322
323 void set_notpassed_slp() {
324 _loop_flags &= ~PassedSlpAnalysis;
325 }
326
327 void double_unrolled_count() { _unrolled_count_log2++; }
328 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
329
330 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
331 int node_count_before_unroll() { return _node_count_before_unroll; }
332 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
333 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
334
335 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
336 // decides which one is taken:
337 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
338 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
339 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
340 // assumption fails we take the slow_loop instead, where we do not make the same speculative
341 // assumption.
342 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
343 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
344 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
345 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
346 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
347 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
348 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
349 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
350 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
351 // Hence, we expect the slow_loop to execute slower.
352 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
353 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
354 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
355 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
356 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
357 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
358 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
359 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
360
361 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
362 OuterStripMinedLoopNode* outer_loop() const;
363 virtual IfTrueNode* outer_loop_tail() const;
364 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
365 virtual IfFalseNode* outer_loop_exit() const;
366 virtual SafePointNode* outer_safepoint() const;
367
368 Node* skip_assertion_predicates_with_halt();
369
370 virtual BasicType bt() const {
371 return T_INT;
372 }
373
374 Node* is_canonical_loop_entry();
375 CountedLoopEndNode* find_pre_loop_end();
376
377 Node* uncasted_init_trip(bool uncasted);
378
379 #ifndef PRODUCT
380 virtual void dump_spec(outputStream *st) const;
381 #endif
382 };
383
384 class LongCountedLoopNode : public BaseCountedLoopNode {
385 private:
386 virtual uint size_of() const { return sizeof(*this); }
387
388 // Known trip count calculated by compute_exact_trip_count()
389 julong _trip_count;
390
391 public:
392 LongCountedLoopNode(Node *entry, Node *backedge)
393 : BaseCountedLoopNode(entry, backedge), _trip_count(max_julong) {
394 init_class_id(Class_LongCountedLoop);
395 }
396
397 virtual int Opcode() const;
398
399 virtual BasicType bt() const {
400 return T_LONG;
401 }
402
403 void set_trip_count(julong tc) {
404 assert(tc < max_julong, "Cannot set trip count to max_julong");
405 _trip_count = tc;
406 }
407 julong trip_count() const { return _trip_count; }
408
409 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
410 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
411 };
412
413
414 //------------------------------CountedLoopEndNode-----------------------------
415 // CountedLoopEndNodes end simple trip counted loops. They act much like
416 // IfNodes.
417
418 class BaseCountedLoopEndNode : public IfNode {
419 public:
420 enum { TestControl, TestValue };
421 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
422 : IfNode(control, test, prob, cnt) {
423 init_class_id(Class_BaseCountedLoopEnd);
424 }
425
426 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
427 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
428 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
429 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
430 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
431 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
432
433 PhiNode* phi() const {
434 Node* tmp = incr();
435 if (tmp && tmp->req() == 3) {
436 Node* phi = tmp->in(1);
437 if (phi->is_Phi()) {
438 return phi->as_Phi();
439 }
440 }
441 return nullptr;
442 }
443
444 BaseCountedLoopNode* loopnode() const {
445 // The CountedLoopNode that goes with this CountedLoopEndNode may
446 // have been optimized out by the IGVN so be cautious with the
447 // pattern matching on the graph
448 PhiNode* iv_phi = phi();
449 if (iv_phi == nullptr) {
450 return nullptr;
451 }
452 Node* ln = iv_phi->in(0);
453 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
454 return nullptr;
455 }
456 if (ln->as_BaseCountedLoop()->bt() != bt()) {
457 return nullptr;
458 }
459 return ln->as_BaseCountedLoop();
460 }
461
462 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
463
464 jlong stride_con() const;
465 virtual BasicType bt() const = 0;
466
467 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
468 };
469
470 class CountedLoopEndNode : public BaseCountedLoopEndNode {
471 public:
472
473 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
474 : BaseCountedLoopEndNode(control, test, prob, cnt) {
475 init_class_id(Class_CountedLoopEnd);
476 }
477 virtual int Opcode() const;
478
479 CountedLoopNode* loopnode() const {
480 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
481 }
482
483 virtual BasicType bt() const {
484 return T_INT;
485 }
486
487 #ifndef PRODUCT
488 virtual void dump_spec(outputStream *st) const;
489 #endif
490 };
491
492 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
493 public:
494 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
495 : BaseCountedLoopEndNode(control, test, prob, cnt) {
496 init_class_id(Class_LongCountedLoopEnd);
497 }
498
499 LongCountedLoopNode* loopnode() const {
500 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
501 }
502
503 virtual int Opcode() const;
504
505 virtual BasicType bt() const {
506 return T_LONG;
507 }
508 };
509
510
511 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
512 Node* bctrl = back_control();
513 if (bctrl == nullptr) return nullptr;
514
515 Node* lexit = bctrl->in(0);
516 if (!lexit->is_BaseCountedLoopEnd()) {
517 return nullptr;
518 }
519 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
520 if (result->bt() != bt()) {
521 return nullptr;
522 }
523 return result;
524 }
525
526 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
527 BaseCountedLoopEndNode* cle = loopexit_or_null();
528 assert(cle != nullptr, "loopexit is null");
529 return cle;
530 }
531
532 inline Node* BaseCountedLoopNode::init_trip() const {
533 BaseCountedLoopEndNode* cle = loopexit_or_null();
534 return cle != nullptr ? cle->init_trip() : nullptr;
535 }
536 inline Node* BaseCountedLoopNode::stride() const {
537 BaseCountedLoopEndNode* cle = loopexit_or_null();
538 return cle != nullptr ? cle->stride() : nullptr;
539 }
540
541 inline bool BaseCountedLoopNode::stride_is_con() const {
542 BaseCountedLoopEndNode* cle = loopexit_or_null();
543 return cle != nullptr && cle->stride_is_con();
544 }
545 inline Node* BaseCountedLoopNode::limit() const {
546 BaseCountedLoopEndNode* cle = loopexit_or_null();
547 return cle != nullptr ? cle->limit() : nullptr;
548 }
549 inline Node* BaseCountedLoopNode::incr() const {
550 BaseCountedLoopEndNode* cle = loopexit_or_null();
551 return cle != nullptr ? cle->incr() : nullptr;
552 }
553 inline Node* BaseCountedLoopNode::phi() const {
554 BaseCountedLoopEndNode* cle = loopexit_or_null();
555 return cle != nullptr ? cle->phi() : nullptr;
556 }
557
558 inline jlong BaseCountedLoopNode::stride_con() const {
559 BaseCountedLoopEndNode* cle = loopexit_or_null();
560 return cle != nullptr ? cle->stride_con() : 0;
561 }
562
563
564 //------------------------------LoopLimitNode-----------------------------
565 // Counted Loop limit node which represents exact final iterator value:
566 // trip_count = (limit - init_trip + stride - 1)/stride
567 // final_value= trip_count * stride + init_trip.
568 // Use HW instructions to calculate it when it can overflow in integer.
569 // Note, final_value should fit into integer since counted loop has
570 // limit check: limit <= max_int-stride.
571 class LoopLimitNode : public Node {
572 enum { Init=1, Limit=2, Stride=3 };
573 public:
574 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
575 // Put it on the Macro nodes list to optimize during macro nodes expansion.
576 init_flags(Flag_is_macro);
577 C->add_macro_node(this);
578 }
579 virtual int Opcode() const;
580 virtual const Type *bottom_type() const { return TypeInt::INT; }
581 virtual uint ideal_reg() const { return Op_RegI; }
582 virtual const Type* Value(PhaseGVN* phase) const;
583 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
584 virtual Node* Identity(PhaseGVN* phase);
585 };
586
587 // Support for strip mining
588 class OuterStripMinedLoopNode : public LoopNode {
589 private:
590 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const;
591 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn);
592
593 public:
594 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
595 : LoopNode(entry, backedge) {
596 init_class_id(Class_OuterStripMinedLoop);
597 init_flags(Flag_is_macro);
598 C->add_macro_node(this);
599 }
600
601 virtual int Opcode() const;
602
603 virtual IfTrueNode* outer_loop_tail() const;
604 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
605 virtual IfFalseNode* outer_loop_exit() const;
606 virtual SafePointNode* outer_safepoint() const;
607 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); }
608 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); }
609 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->false_proj(); }
610
611 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
612
613 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
614
615 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
616
617 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
618
619 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
620 PhaseIdealLoop* iloop);
621 };
622
623 class OuterStripMinedLoopEndNode : public IfNode {
624 public:
625 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
626 : IfNode(control, test, prob, cnt) {
627 init_class_id(Class_OuterStripMinedLoopEnd);
628 }
629
630 virtual int Opcode() const;
631
632 virtual const Type* Value(PhaseGVN* phase) const;
633 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
634
635 bool is_expanded(PhaseGVN *phase) const;
636 };
637
638 // -----------------------------IdealLoopTree----------------------------------
639 class IdealLoopTree : public ResourceObj {
640 public:
641 IdealLoopTree *_parent; // Parent in loop tree
642 IdealLoopTree *_next; // Next sibling in loop tree
643 IdealLoopTree *_child; // First child in loop tree
644
645 // The head-tail backedge defines the loop.
646 // If a loop has multiple backedges, this is addressed during cleanup where
647 // we peel off the multiple backedges, merging all edges at the bottom and
648 // ensuring that one proper backedge flow into the loop.
649 Node *_head; // Head of loop
650 Node *_tail; // Tail of loop
651 inline Node *tail(); // Handle lazy update of _tail field
652 inline Node *head(); // Handle lazy update of _head field
653 PhaseIdealLoop* _phase;
654 int _local_loop_unroll_limit;
655 int _local_loop_unroll_factor;
656
657 Node_List _body; // Loop body for inner loops
658
659 uint16_t _nest; // Nesting depth
660 uint8_t _irreducible:1, // True if irreducible
661 _has_call:1, // True if has call safepoint
662 _has_sfpt:1, // True if has non-call safepoint
663 _rce_candidate:1, // True if candidate for range check elimination
664 _has_range_checks:1,
665 _has_range_checks_computed:1;
666
667 Node_List* _safepts; // List of safepoints in this loop
668 Node_List* _required_safept; // A inner loop cannot delete these safepts;
669 Node_List* _reachability_fences; // List of reachability fences in this loop
670 bool _allow_optimizations; // Allow loop optimizations
671
672 IdealLoopTree(PhaseIdealLoop* phase, Node* head, Node* tail);
673
674 // Is 'l' a member of 'this'?
675 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
676
677 // Set loop nesting depth. Accumulate has_call bits.
678 int set_nest( uint depth );
679
680 // Split out multiple fall-in edges from the loop header. Move them to a
681 // private RegionNode before the loop. This becomes the loop landing pad.
682 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
683
684 // Split out the outermost loop from this shared header.
685 void split_outer_loop( PhaseIdealLoop *phase );
686
687 // Merge all the backedges from the shared header into a private Region.
688 // Feed that region as the one backedge to this loop.
689 void merge_many_backedges( PhaseIdealLoop *phase );
690
691 // Split shared headers and insert loop landing pads.
692 // Insert a LoopNode to replace the RegionNode.
693 // Returns TRUE if loop tree is structurally changed.
694 bool beautify_loops( PhaseIdealLoop *phase );
695
696 // Perform optimization to use the loop predicates for null checks and range checks.
697 // Applies to any loop level (not just the innermost one)
698 bool loop_predication( PhaseIdealLoop *phase);
699 bool can_apply_loop_predication();
700
701 // Perform iteration-splitting on inner loops. Split iterations to
702 // avoid range checks or one-shot null checks. Returns false if the
703 // current round of loop opts should stop.
704 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
705
706 // Driver for various flavors of iteration splitting. Returns false
707 // if the current round of loop opts should stop.
708 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
709
710 // Given dominators, try to find loops with calls that must always be
711 // executed (call dominates loop tail). These loops do not need non-call
712 // safepoints (ncsfpt).
713 void check_safepts(VectorSet &visited, Node_List &stack);
714
715 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
716 // encountered.
717 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
718
719 // Remove safepoints from loop. Optionally keeping one.
720 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
721
722 // Convert to counted loops where possible
723 void counted_loop( PhaseIdealLoop *phase );
724
725 // Check for Node being a loop-breaking test
726 Node *is_loop_exit(Node *iff) const;
727
728 // Return unique loop-exit projection or null if the loop has multiple exits.
729 IfFalseNode* unique_loop_exit_proj_or_null();
730
731 // Remove simplistic dead code from loop body
732 void DCE_loop_body();
733
734 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
735 // Replace with a 1-in-10 exit guess.
736 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
737
738 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
739 // Useful for unrolling loops with NO array accesses.
740 bool policy_peel_only( PhaseIdealLoop *phase ) const;
741
742 // Return TRUE or FALSE if the loop should be unswitched -- clone
743 // loop with an invariant test
744 bool policy_unswitching( PhaseIdealLoop *phase ) const;
745 bool no_unswitch_candidate() const;
746
747 // Micro-benchmark spamming. Remove empty loops.
748 bool do_remove_empty_loop( PhaseIdealLoop *phase );
749
750 // Convert one-iteration loop into normal code.
751 bool do_one_iteration_loop( PhaseIdealLoop *phase );
752
753 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
754 // move some loop-invariant test (usually a null-check) before the loop.
755 bool policy_peeling(PhaseIdealLoop *phase);
756
757 uint estimate_peeling(PhaseIdealLoop *phase);
758
759 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
760 // known trip count in the counted loop node.
761 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
762
763 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
764 // if the loop is a counted loop and the loop body is small enough.
765 bool policy_unroll(PhaseIdealLoop *phase);
766
767 // Loop analyses to map to a maximal superword unrolling for vectorization.
768 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
769
770 // Return TRUE or FALSE if the loop should be range-check-eliminated.
771 // Gather a list of IF tests that are dominated by iteration splitting;
772 // also gather the end of the first split and the start of the 2nd split.
773 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
774
775 // Return TRUE if "iff" is a range check.
776 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
777 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
778 jlong& scale) const;
779
780 // Estimate the number of nodes required when cloning a loop (body).
781 uint est_loop_clone_sz(uint factor) const;
782 // Estimate the number of nodes required when unrolling a loop (body).
783 uint est_loop_unroll_sz(uint factor) const;
784
785 // Compute loop trip count if possible
786 void compute_trip_count(PhaseIdealLoop* phase, BasicType bt);
787
788 // Compute loop trip count from profile data
789 float compute_profile_trip_cnt_helper(Node* n);
790 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
791
792 // Reassociate invariant expressions.
793 void reassociate_invariants(PhaseIdealLoop *phase);
794 // Reassociate invariant binary expressions.
795 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
796 // Reassociate invariant add, subtract, and compare expressions.
797 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
798 // Return nonzero index of invariant operand if invariant and variant
799 // are combined with an associative binary. Helper for reassociate_invariants.
800 int find_invariant(Node* n, PhaseIdealLoop *phase);
801 // Return TRUE if "n" is associative.
802 bool is_associative(Node* n, Node* base=nullptr);
803 // Return TRUE if "n" is an associative cmp node.
804 bool is_associative_cmp(Node* n);
805
806 // Return true if n is invariant
807 bool is_invariant(Node* n) const;
808
809 // Put loop body on igvn work list
810 void record_for_igvn();
811
812 bool is_root() { return _parent == nullptr; }
813 // A proper/reducible loop w/o any (occasional) dead back-edge.
814 bool is_loop() { return !_irreducible && !tail()->is_top(); }
815 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
816 bool is_innermost() { return is_loop() && _child == nullptr; }
817
818 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
819
820 bool compute_has_range_checks() const;
821 bool range_checks_present() {
822 if (!_has_range_checks_computed) {
823 if (compute_has_range_checks()) {
824 _has_range_checks = 1;
825 }
826 _has_range_checks_computed = 1;
827 }
828 return _has_range_checks;
829 }
830
831 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
832 // In all other cases, return this.
833 IdealLoopTree* skip_strip_mined() {
834 return _head->as_Loop()->is_strip_mined() ? _parent : this;
835 }
836
837 // Registers a reachability fence node in the loop.
838 void register_reachability_fence(ReachabilityFenceNode* rf);
839
840 #ifndef PRODUCT
841 void dump_head(); // Dump loop head only
842 void dump(); // Dump this loop recursively
843 #endif
844
845 #ifdef ASSERT
846 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
847 bool verify_tree(IdealLoopTree* loop_verify) const;
848 #endif
849
850 private:
851 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
852
853 // Estimate the number of nodes resulting from control and data flow merge.
854 uint est_loop_flow_merge_sz() const;
855
856 // Check if the number of residual iterations is large with unroll_cnt.
857 // Return true if the residual iterations are more than 10% of the trip count.
858 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
859 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
860 }
861
862 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
863
864 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
865
866 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
867
868 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
869 Node* sfpt) const;
870
871 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
872
873 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
874 };
875
876 // -----------------------------PhaseIdealLoop---------------------------------
877 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
878 // into a loop tree. Drives the loop-based transformations on the ideal graph.
879 class PhaseIdealLoop : public PhaseTransform {
880 friend class IdealLoopTree;
881 friend class SuperWord;
882 friend class AutoNodeBudget;
883
884 Arena _arena; // For data whose lifetime is a single pass of loop optimizations
885
886 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
887 //
888 // Exception: dead CFG nodes may instead have a ctrl/idom forwarding
889 // installed. See: forward_ctrl
890 Node_List _loop_or_ctrl;
891
892 // Pre-computed def-use info
893 PhaseIterGVN &_igvn;
894
895 // Head of loop tree
896 IdealLoopTree* _ltree_root;
897
898 // Array of pre-order numbers, plus post-visited bit.
899 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
900 // ODD for post-visited. Other bits are the pre-order number.
901 uint *_preorders;
902 uint _max_preorder;
903
904 ReallocMark _nesting; // Safety checks for arena reallocation
905
906 const PhaseIdealLoop* _verify_me;
907 bool _verify_only;
908
909 // Allocate _preorders[] array
910 void allocate_preorders() {
911 _max_preorder = C->unique()+8;
912 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
913 memset(_preorders, 0, sizeof(uint) * _max_preorder);
914 }
915
916 // Allocate _preorders[] array
917 void reallocate_preorders() {
918 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
919 if ( _max_preorder < C->unique() ) {
920 _preorders = REALLOC_RESOURCE_ARRAY(_preorders, _max_preorder, C->unique());
921 _max_preorder = C->unique();
922 }
923 memset(_preorders, 0, sizeof(uint) * _max_preorder);
924 }
925
926 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
927 // adds new nodes.
928 void check_grow_preorders( ) {
929 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
930 if ( _max_preorder < C->unique() ) {
931 uint newsize = _max_preorder<<1; // double size of array
932 _preorders = REALLOC_RESOURCE_ARRAY(_preorders, _max_preorder, newsize);
933 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
934 _max_preorder = newsize;
935 }
936 }
937 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
938 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
939 // Pre-order numbers are written to the Nodes array as low-bit-set values.
940 void set_preorder_visited( Node *n, int pre_order ) {
941 assert( !is_visited( n ), "already set" );
942 _preorders[n->_idx] = (pre_order<<1);
943 };
944 // Return pre-order number.
945 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
946
947 // Check for being post-visited.
948 // Should be previsited already (checked with assert(is_visited(n))).
949 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
950
951 // Mark as post visited
952 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
953
954 public:
955 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
956 // Returns true if "n" is a data node, false if it's a CFG node.
957 //
958 // Exception:
959 // control nodes that are dead because of "replace_node_and_forward_ctrl"
960 // or have otherwise modified their ctrl state by "forward_ctrl".
961 // They return "true", because they have a ctrl "forwarding" to the other ctrl node they
962 // were replaced with.
963 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
964
965 private:
966 // clear out dead code after build_loop_late
967 Node_List _deadlist;
968 Node_List _zero_trip_guard_opaque_nodes;
969 Node_List _multiversion_opaque_nodes;
970
971 // Support for faster execution of get_late_ctrl()/dom_lca()
972 // when a node has many uses and dominator depth is deep.
973 GrowableArray<jlong> _dom_lca_tags;
974 uint _dom_lca_tags_round;
975 void init_dom_lca_tags();
976
977 // Helper for debugging bad dominance relationships
978 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
979
980 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
981
982 // Inline wrapper for frequent cases:
983 // 1) only one use
984 // 2) a use is the same as the current LCA passed as 'n1'
985 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
986 assert( n->is_CFG(), "" );
987 // Fast-path null lca
988 if( lca != nullptr && lca != n ) {
989 assert( lca->is_CFG(), "" );
990 // find LCA of all uses
991 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
992 }
993 return find_non_split_ctrl(n);
994 }
995 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
996
997 // Helper function for directing control inputs away from CFG split points.
998 Node *find_non_split_ctrl( Node *ctrl ) const {
999 if (ctrl != nullptr) {
1000 if (ctrl->is_MultiBranch()) {
1001 ctrl = ctrl->in(0);
1002 }
1003 assert(ctrl->is_CFG(), "CFG");
1004 }
1005 return ctrl;
1006 }
1007
1008 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
1009
1010 #ifdef ASSERT
1011 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
1012 #endif
1013 private:
1014 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
1015 Unique_Node_List& list);
1016 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
1017 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1018 CountedLoopNode* remaining_loop_head,
1019 uint first_node_index_in_cloned_loop_body,
1020 const Node_List& old_new);
1021 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1022 CountedLoopNode* main_loop_head,
1023 uint first_node_index_in_pre_loop_body,
1024 uint last_node_index_in_pre_loop_body,
1025 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1026 const Node_List& old_new);
1027 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1028 uint first_node_index_in_cloned_loop_body);
1029 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1030 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1031 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1032 CountedLoopNode* target_loop_head,
1033 const NodeInLoopBody& _node_in_loop_body,
1034 bool kill_old_template);
1035 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1036 const Node* old_target_loop_entry,
1037 uint node_index_before_new_assertion_predicate_nodes);
1038 void log_loop_tree();
1039
1040 public:
1041
1042 PhaseIterGVN &igvn() const { return _igvn; }
1043
1044 Arena* arena() { return &_arena; };
1045
1046 bool has_node(const Node* n) const {
1047 guarantee(n != nullptr, "No Node.");
1048 return _loop_or_ctrl[n->_idx] != nullptr;
1049 }
1050 // check if transform created new nodes that need _ctrl recorded
1051 Node *get_late_ctrl( Node *n, Node *early );
1052 Node *get_early_ctrl( Node *n );
1053 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1054 void set_early_ctrl(Node* n, bool update_body);
1055 void set_subtree_ctrl(Node* n, bool update_body);
1056 void set_ctrl( Node *n, Node *ctrl ) {
1057 assert( !has_node(n) || has_ctrl(n), "" );
1058 assert( ctrl->in(0), "cannot set dead control node" );
1059 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1060 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1061 }
1062 void set_root_as_ctrl(Node* n) {
1063 assert(!has_node(n) || has_ctrl(n), "");
1064 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1065 }
1066 // Set control and update loop membership
1067 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1068 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1069 IdealLoopTree* new_loop = get_loop(ctrl);
1070 if (old_loop != new_loop) {
1071 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1072 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1073 }
1074 set_ctrl(n, ctrl);
1075 }
1076
1077 // Retrieves the ctrl for a data node i.
1078 Node* get_ctrl(const Node* i) {
1079 assert(has_node(i) && has_ctrl(i), "must be data node with ctrl");
1080 Node* n = get_ctrl_no_update(i);
1081 // We store the found ctrl in the side-table again. In most cases,
1082 // this is a no-op, since we just read from _loop_or_ctrl. But in cases
1083 // where there was a ctrl forwarding via dead ctrl nodes, this shortens the path.
1084 // See: forward_ctrl
1085 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1086 assert(has_node(i) && has_ctrl(i), "must still be data node with ctrl");
1087 assert(n == find_non_split_ctrl(n), "must return legal ctrl");
1088 return n;
1089 }
1090
1091 bool is_dominator(Node* dominator, Node* n);
1092 bool is_strict_dominator(Node* dominator, Node* n);
1093
1094 // return get_ctrl for a data node and self(n) for a CFG node
1095 Node* ctrl_or_self(Node* n) {
1096 if (has_ctrl(n))
1097 return get_ctrl(n);
1098 else {
1099 assert (n->is_CFG(), "must be a CFG node");
1100 return n;
1101 }
1102 }
1103
1104 private:
1105 Node* get_ctrl_no_update_helper(const Node* i) const {
1106 // We expect only data nodes (which must have a ctrl set), or
1107 // dead ctrl nodes that have a ctrl "forwarding".
1108 // See: forward_ctrl.
1109 assert(has_ctrl(i), "only data nodes or ctrl nodes with ctrl forwarding expected");
1110 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1111 }
1112
1113 // Compute the ctrl of node i, jumping over ctrl forwardings.
1114 Node* get_ctrl_no_update(const Node* i) const {
1115 assert(has_ctrl(i), "only data nodes expected");
1116 Node* n = get_ctrl_no_update_helper(i);
1117 if (n->in(0) == nullptr) {
1118 // We encountered a dead CFG node.
1119 // If everything went right, this dead CFG node should have had a ctrl
1120 // forwarding installed, using "forward_ctrl". We now have to jump from
1121 // the old (dead) ctrl node to the new (live) ctrl node, in possibly
1122 // multiple ctrl forwarding steps.
1123 do {
1124 n = get_ctrl_no_update_helper(n);
1125 } while (n->in(0) == nullptr);
1126 n = find_non_split_ctrl(n);
1127 }
1128 return n;
1129 }
1130
1131 public:
1132 // Check for loop being set
1133 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1134 bool has_loop( Node *n ) const {
1135 assert(!has_node(n) || !has_ctrl(n), "");
1136 return has_node(n);
1137 }
1138 // Set loop
1139 void set_loop( Node *n, IdealLoopTree *loop ) {
1140 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1141 }
1142
1143 // Install a ctrl "forwarding" from an old (dead) control node.
1144 // This is a "lazy" update of the "get_ctrl" and "idom" mechanism:
1145 // - Install a forwarding from old_node (dead ctrl) to new_node.
1146 // - When querying "get_ctrl": jump from data node over possibly
1147 // multiple dead ctrl nodes with ctrl forwarding to eventually
1148 // reach a live ctrl node. Shorten the path to avoid chasing the
1149 // forwarding in the future.
1150 // - When querying "idom": from some node get its old idom, which
1151 // may be dead but has an idom forwarding to the new and live
1152 // idom. Shorten the path to avoid chasing the forwarding in the
1153 // future.
1154 // Note: while the "idom" information is stored in the "_idom"
1155 // side-table, the idom forwarding piggybacks on the ctrl
1156 // forwarding on "_loop_or_ctrl".
1157 // Using "forward_ctrl" allows us to only edit the entry for the old
1158 // dead node now, and we do not have to update all the nodes that had
1159 // the old_node as their "get_ctrl" or "idom". We clean up the forwarding
1160 // links when we query "get_ctrl" or "idom" for these nodes the next time.
1161 void forward_ctrl(Node* old_node, Node* new_node) {
1162 assert(!has_ctrl(old_node) && old_node->is_CFG() && old_node->in(0) == nullptr,
1163 "must be dead ctrl (CFG) node");
1164 assert(!has_ctrl(new_node) && new_node->is_CFG() && new_node->in(0) != nullptr,
1165 "must be live ctrl (CFG) node");
1166 assert(old_node != new_node, "no cycles please");
1167 // Re-use the side array slot for this node to provide the
1168 // forwarding pointer.
1169 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1170 assert(has_ctrl(old_node), "must have installed ctrl forwarding");
1171 }
1172
1173 // Replace the old ctrl node with a new ctrl node.
1174 // - Update the node inputs of all uses.
1175 // - Lazily update the ctrl and idom info of all uses, via a ctrl/idom forwarding.
1176 void replace_node_and_forward_ctrl(Node* old_node, Node* new_node) {
1177 _igvn.replace_node(old_node, new_node);
1178 forward_ctrl(old_node, new_node);
1179 }
1180
1181 void remove_dead_data_node(Node* dead) {
1182 assert(dead->outcnt() == 0 && !dead->is_top(), "must be dead");
1183 assert(!dead->is_CFG(), "not a data node");
1184 Node* c = get_ctrl(dead);
1185 IdealLoopTree* lpt = get_loop(c);
1186 _loop_or_ctrl.map(dead->_idx, nullptr); // This node is useless
1187 lpt->_body.yank(dead);
1188 igvn().remove_dead_node(dead, PhaseIterGVN::NodeOrigin::Graph);
1189 }
1190
1191 private:
1192
1193 // Place 'n' in some loop nest, where 'n' is a CFG node
1194 void build_loop_tree();
1195 int build_loop_tree_impl(Node* n, int pre_order);
1196 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1197 // loop tree, not the root.
1198 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1199
1200 #ifdef ASSERT
1201 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1202 void verify_regions_in_irreducible_loops();
1203 bool is_in_irreducible_loop(RegionNode* region);
1204 #endif
1205
1206 // Place Data nodes in some loop nest
1207 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1208 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1209 void build_loop_late_post_work(Node* n, bool pinned);
1210 void build_loop_late_post(Node* n);
1211 void verify_strip_mined_scheduling(Node *n, Node* least);
1212
1213 // Array of immediate dominance info for each CFG node indexed by node idx
1214 private:
1215 uint _idom_size;
1216 Node **_idom; // Array of immediate dominators
1217 uint *_dom_depth; // Used for fast LCA test
1218 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1219 LoopOptsMode _mode;
1220
1221 // build the loop tree and perform any requested optimizations
1222 void build_and_optimize();
1223
1224 // Dominators for the sea of nodes
1225 void Dominators();
1226
1227 // Compute the Ideal Node to Loop mapping
1228 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1229 PhaseTransform(Ideal_Loop),
1230 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1231 _loop_or_ctrl(&_arena),
1232 _igvn(igvn),
1233 _verify_me(nullptr),
1234 _verify_only(false),
1235 _mode(mode),
1236 _nodes_required(UINT_MAX) {
1237 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1238 build_and_optimize();
1239 }
1240
1241 #ifndef PRODUCT
1242 // Verify that verify_me made the same decisions as a fresh run
1243 // or only verify that the graph is valid if verify_me is null.
1244 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1245 PhaseTransform(Ideal_Loop),
1246 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1247 _loop_or_ctrl(&_arena),
1248 _igvn(igvn),
1249 _verify_me(verify_me),
1250 _verify_only(verify_me == nullptr),
1251 _mode(LoopOptsVerify),
1252 _nodes_required(UINT_MAX) {
1253 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1254 build_and_optimize();
1255 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1256 }
1257 #endif
1258
1259 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1260
1261 Node* idom_no_update(Node* d) const {
1262 return idom_no_update(d->_idx);
1263 }
1264
1265 Node* idom_no_update(uint node_idx) const {
1266 assert(node_idx < _idom_size, "oob");
1267 Node* n = _idom[node_idx];
1268 assert(n != nullptr,"Bad immediate dominator info.");
1269 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1270 // We encountered a dead CFG node.
1271 // If everything went right, this dead CFG node should have had an idom
1272 // forwarding installed, using "forward_ctrl". We now have to jump from
1273 // the old (dead) idom node to the new (live) idom node, in possibly
1274 // multiple idom forwarding steps.
1275 // Note that we piggyback on "_loop_or_ctrl" to do the forwarding,
1276 // since we forward both "get_ctrl" and "idom" from the dead to the
1277 // new live ctrl/idom nodes.
1278 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1279 assert(n != nullptr,"Bad immediate dominator info.");
1280 }
1281 return n;
1282 }
1283
1284 public:
1285 Node* idom(Node* n) const {
1286 return idom(n->_idx);
1287 }
1288
1289 Node* idom(uint node_idx) const {
1290 Node* n = idom_no_update(node_idx);
1291 // We store the found idom in the side-table again. In most cases,
1292 // this is a no-op, since we just read from _idom. But in cases where
1293 // there was an idom forwarding via dead idom nodes, this shortens the path.
1294 // See: forward_ctrl
1295 _idom[node_idx] = n;
1296 return n;
1297 }
1298
1299 uint dom_depth(Node* d) const {
1300 guarantee(d != nullptr, "Null dominator info.");
1301 guarantee(d->_idx < _idom_size, "");
1302 return _dom_depth[d->_idx];
1303 }
1304 void set_idom(Node* d, Node* n, uint dom_depth);
1305 // Locally compute IDOM using dom_lca call
1306 Node *compute_idom( Node *region ) const;
1307 // Recompute dom_depth
1308 void recompute_dom_depth();
1309
1310 // Is safept not required by an outer loop?
1311 bool is_deleteable_safept(Node* sfpt) const;
1312
1313 // Replace parallel induction variable (parallel to trip counter)
1314 void replace_parallel_iv(IdealLoopTree *loop);
1315
1316 Node *dom_lca( Node *n1, Node *n2 ) const {
1317 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1318 }
1319 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1320
1321 Node* dominated_node(Node* c1, Node* c2) {
1322 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1323 return is_dominator(c1, c2) ? c2 : c1;
1324 }
1325
1326 // Return control node that's dominated by the 2 others
1327 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1328 return dominated_node(c1, dominated_node(c2, c3));
1329 }
1330
1331 // Build and verify the loop tree without modifying the graph. This
1332 // is useful to verify that all inputs properly dominate their uses.
1333 static void verify(PhaseIterGVN& igvn) {
1334 #ifdef ASSERT
1335 ResourceMark rm;
1336 Compile::TracePhase tp(_t_idealLoopVerify);
1337 PhaseIdealLoop v(igvn);
1338 #endif
1339 }
1340
1341 // Recommended way to use PhaseIdealLoop.
1342 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1343 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1344 ResourceMark rm;
1345 PhaseIdealLoop v(igvn, mode);
1346
1347 Compile* C = Compile::current();
1348 if (!C->failing()) {
1349 // Cleanup any modified bits
1350 igvn.optimize();
1351 if (C->failing()) { return; }
1352 v.log_loop_tree();
1353 }
1354 }
1355
1356 // True if the method has at least 1 irreducible loop
1357 bool _has_irreducible_loops;
1358
1359 // Per-Node transform
1360 virtual Node* transform(Node* n) { return nullptr; }
1361
1362 Node* loop_exit_control(const IdealLoopTree* loop) const;
1363
1364 class LoopExitTest {
1365 bool _is_valid;
1366
1367 const Node* _back_control;
1368 const IdealLoopTree* _loop;
1369 PhaseIdealLoop* _phase;
1370
1371 Node* _cmp;
1372 Node* _incr;
1373 Node* _limit;
1374 BoolTest::mask _mask;
1375 float _cl_prob;
1376
1377 public:
1378 LoopExitTest(const Node* back_control, const IdealLoopTree* loop, PhaseIdealLoop* phase) :
1379 _is_valid(false),
1380 _back_control(back_control),
1381 _loop(loop),
1382 _phase(phase),
1383 _cmp(nullptr),
1384 _incr(nullptr),
1385 _limit(nullptr),
1386 _mask(BoolTest::illegal),
1387 _cl_prob(0.0f) {}
1388
1389 void build();
1390 void canonicalize_mask(jlong stride_con);
1391
1392 bool is_valid_with_bt(BasicType bt) const {
1393 return _is_valid && _cmp != nullptr && _cmp->Opcode() == Op_Cmp(bt);
1394 }
1395
1396 bool should_include_limit() const { return _mask == BoolTest::le || _mask == BoolTest::ge; }
1397
1398 CmpNode* cmp() const { return _cmp->as_Cmp(); }
1399 Node* incr() const { return _incr; }
1400 Node* limit() const { return _limit; }
1401 BoolTest::mask mask() const { return _mask; }
1402 float cl_prob() const { return _cl_prob; }
1403 };
1404
1405 class LoopIVIncr {
1406 bool _is_valid;
1407
1408 const Node* _head;
1409 const IdealLoopTree* _loop;
1410
1411 Node* _incr;
1412 Node* _phi_incr;
1413
1414 public:
1415 LoopIVIncr(const Node* head, const IdealLoopTree* loop) :
1416 _is_valid(false),
1417 _head(head),
1418 _loop(loop),
1419 _incr(nullptr),
1420 _phi_incr(nullptr) {}
1421
1422 void build(Node* old_incr);
1423
1424 bool is_valid() const { return _is_valid; }
1425 bool is_valid_with_bt(const BasicType bt) const {
1426 return _is_valid && _incr->Opcode() == Op_Add(bt);
1427 }
1428
1429 Node* incr() const { return _incr; }
1430 Node* phi_incr() const { return _phi_incr; }
1431 };
1432
1433 class LoopIVStride {
1434 bool _is_valid;
1435
1436 BasicType _iv_bt;
1437 Node* _stride_node;
1438 Node* _xphi;
1439
1440 public:
1441 LoopIVStride(BasicType iv_bt) :
1442 _is_valid(false),
1443 _iv_bt(iv_bt),
1444 _stride_node(nullptr),
1445 _xphi(nullptr) {}
1446
1447 void build(const Node* incr);
1448
1449 bool is_valid() const { return _is_valid && _stride_node != nullptr; }
1450 Node* stride_node() const { return _stride_node; }
1451 Node* xphi() const { return _xphi; }
1452
1453 jlong compute_non_zero_stride_con(BoolTest::mask mask, BasicType iv_bt) const;
1454 };
1455
1456 static PhiNode* loop_iv_phi(const Node* xphi, const Node* phi_incr, const Node* head);
1457
1458 bool try_convert_to_counted_loop(Node* head, IdealLoopTree*& loop, BasicType iv_bt);
1459
1460 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1461 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1462
1463 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1464 SafePointNode* find_safepoint(Node* back_control, const Node* head, const IdealLoopTree* loop);
1465
1466 void add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt);
1467
1468 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1469 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1470 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1471 Node*& entry_control, Node*& iffalse);
1472
1473 Node* exact_limit( IdealLoopTree *loop );
1474
1475 // Return a post-walked LoopNode
1476 IdealLoopTree* get_loop(const Node* n) const {
1477 // Dead nodes have no loop, so return the top level loop instead
1478 if (!has_node(n)) return _ltree_root;
1479 assert(!has_ctrl(n), "");
1480 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1481 }
1482
1483 IdealLoopTree* ltree_root() const { return _ltree_root; }
1484
1485 // Is 'n' a (nested) member of 'loop'?
1486 bool is_member(const IdealLoopTree* loop, const Node* n) const {
1487 return loop->is_member(get_loop(n));
1488 }
1489
1490 // is the control for 'n' a (nested) member of 'loop'?
1491 bool ctrl_is_member(const IdealLoopTree* loop, const Node* n) {
1492 return is_member(loop, get_ctrl(n));
1493 }
1494
1495 // This is the basic building block of the loop optimizations. It clones an
1496 // entire loop body. It makes an old_new loop body mapping; with this
1497 // mapping you can find the new-loop equivalent to an old-loop node. All
1498 // new-loop nodes are exactly equal to their old-loop counterparts, all
1499 // edges are the same. All exits from the old-loop now have a RegionNode
1500 // that merges the equivalent new-loop path. This is true even for the
1501 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1502 // now come from (one or more) Phis that merge their new-loop equivalents.
1503 // Parameter side_by_side_idom:
1504 // When side_by_size_idom is null, the dominator tree is constructed for
1505 // the clone loop to dominate the original. Used in construction of
1506 // pre-main-post loop sequence.
1507 // When nonnull, the clone and original are side-by-side, both are
1508 // dominated by the passed in side_by_side_idom node. Used in
1509 // construction of unswitched loops.
1510 enum CloneLoopMode {
1511 IgnoreStripMined = 0, // Only clone inner strip mined loop
1512 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1513 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1514 // result control flow branches
1515 // either to inner clone or outer
1516 // strip mined loop.
1517 };
1518 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1519 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1520 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1521 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1522 Node_List*& split_if_set, Node_List*& split_bool_set,
1523 Node_List*& split_cex_set, Node_List& worklist,
1524 uint new_counter, CloneLoopMode mode);
1525 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1526 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1527 Node_List& extra_data_nodes);
1528
1529 // If we got the effect of peeling, either by actually peeling or by
1530 // making a pre-loop which must execute at least once, we can remove
1531 // all loop-invariant dominated tests in the main body.
1532 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1533
1534 // Generate code to do a loop peel for the given loop (and body).
1535 // old_new is a temp array.
1536 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1537
1538 // Add pre and post loops around the given loop. These loops are used
1539 // during RCE, unrolling and aligning loops.
1540 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1541
1542 // Find the last store in the body of an OuterStripMinedLoop when following memory uses
1543 Node *find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop);
1544
1545 // Add post loop after the given loop.
1546 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1547 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1548 Node* incr, Node* limit, CountedLoopNode*& post_head);
1549
1550 // Add a vector post loop between a vector main loop and the current post loop
1551 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1552 // If Node n lives in the back_ctrl block, we clone a private version of n
1553 // in preheader_ctrl block and return that, otherwise return n.
1554 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1555
1556 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1557 // unroll to do double iterations. The next round of major loop transforms
1558 // will repeat till the doubled loop body does all remaining iterations in 1
1559 // pass.
1560 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1561
1562 // Unroll the loop body one step - make each trip do 2 iterations.
1563 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1564
1565 // Return true if exp is a constant times an induction var
1566 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1567
1568 bool is_iv(Node* exp, Node* iv, BasicType bt);
1569
1570 // Return true if exp is a scaled induction var plus (or minus) constant
1571 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = nullptr, int depth = 0);
1572 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1573 jlong long_scale;
1574 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1575 int int_scale = checked_cast<int>(long_scale);
1576 if (p_scale != nullptr) {
1577 *p_scale = int_scale;
1578 }
1579 return true;
1580 }
1581 return false;
1582 }
1583 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1584 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1585 BasicType bt,
1586 jlong* p_scale, Node** p_offset,
1587 bool* p_short_scale, int depth);
1588
1589 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1590 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1591 Deoptimization::DeoptReason reason, int opcode,
1592 bool rewire_uncommon_proj_phi_inputs = false);
1593
1594 private:
1595 // Helper functions for create_new_if_for_predicate()
1596 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1597 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1598 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1599 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1600 const OrigToNewHashtable& orig_to_clone);
1601
1602 public:
1603 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1604
1605 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1606 void replace_control(Node* node, Node* new_control) {
1607 _igvn.replace_input_of(node, 0, new_control);
1608 set_idom(node, new_control, dom_depth(new_control));
1609 }
1610
1611 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1612 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1613 set_idom(loop_head, new_entry, dom_depth(new_entry));
1614 }
1615
1616 // Construct a range check for a predicate if
1617 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1618 jint stride, Node* range, bool upper, bool& overflow);
1619
1620 // Implementation of the loop predication to promote checks outside the loop
1621 bool loop_predication_impl(IdealLoopTree *loop);
1622
1623 // Reachability Fence (RF) support.
1624 private:
1625 void insert_rf(Node* ctrl, Node* referent);
1626 void replace_rf(Node* old_node, Node* new_node);
1627 void remove_rf(ReachabilityFenceNode* rf);
1628 public:
1629 bool optimize_reachability_fences();
1630 bool expand_reachability_fences();
1631
1632 private:
1633 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1634 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1635 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1636 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1637 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1638 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1639 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1640 Node_List& if_proj_list);
1641 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1642 IfProjNode* new_control, int scale, Node* offset, Node* range);
1643 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1644
1645 // Helper function to collect predicate for eliminating the useless ones
1646 void eliminate_useless_predicates() const;
1647
1648 void eliminate_useless_zero_trip_guard();
1649 void eliminate_useless_multiversion_if();
1650
1651 public:
1652 // Change the control input of expensive nodes to allow commoning by
1653 // IGVN when it is guaranteed to not result in a more frequent
1654 // execution of the expensive node. Return true if progress.
1655 bool process_expensive_nodes();
1656
1657 // Check whether node has become unreachable
1658 bool is_node_unreachable(Node *n) const {
1659 return !has_node(n) || n->is_unreachable(_igvn);
1660 }
1661
1662 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1663 void do_range_check(IdealLoopTree* loop);
1664
1665 // Clone loop with an invariant test (that does not exit) and
1666 // insert a clone of the test that selects which version to
1667 // execute.
1668 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1669
1670 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1671 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1672
1673 private:
1674 static bool has_control_dependencies_from_predicates(LoopNode* head);
1675 static void revert_to_normal_loop(const LoopNode* loop_head);
1676
1677 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1678 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1679 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1680 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1681 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1682 #ifndef PRODUCT
1683 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1684 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1685 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1686 const UnswitchCandidate& unswitch_candidate,
1687 const LoopNode* original_head, const LoopNode* new_head);
1688 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1689 const LoopNode* original_head, const LoopNode* new_head);
1690 #endif
1691
1692 public:
1693
1694 // Range Check Elimination uses this function!
1695 // Constrain the main loop iterations so the affine function:
1696 // low_limit <= scale_con * I + offset < upper_limit
1697 // always holds true. That is, either increase the number of iterations in
1698 // the pre-loop or the post-loop until the condition holds true in the main
1699 // loop. Scale_con, offset and limit are all loop invariant.
1700 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit);
1701 // Helper function for add_constraint().
1702 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1703
1704 // Partially peel loop up through last_peel node.
1705 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1706 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1707
1708 // AutoVectorize the loop: replace scalar ops with vector ops.
1709 enum AutoVectorizeStatus {
1710 Impossible, // This loop has the wrong shape to even try vectorization.
1711 Success, // We just successfully vectorized the loop.
1712 TriedAndFailed, // We tried to vectorize, but failed.
1713 };
1714 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1715
1716 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1717 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1718 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1719 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1720
1721 // Create a scheduled list of nodes control dependent on ctrl set.
1722 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1723 // Has a use in the vector set
1724 bool has_use_in_set( Node* n, VectorSet& vset );
1725 // Has use internal to the vector set (ie. not in a phi at the loop head)
1726 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1727 // clone "n" for uses that are outside of loop
1728 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1729 // clone "n" for special uses that are in the not_peeled region
1730 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1731 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1732 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1733 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1734 #ifdef ASSERT
1735 // Validate the loop partition sets: peel and not_peel
1736 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1737 // Ensure that uses outside of loop are of the right form
1738 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1739 uint orig_exit_idx, uint clone_exit_idx);
1740 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1741 #endif
1742
1743 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1744 int stride_of_possible_iv( Node* iff );
1745 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1746 // Return the (unique) control output node that's in the loop (if it exists.)
1747 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1748 // Insert a signed compare loop exit cloned from an unsigned compare.
1749 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1750 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1751 // Utility to register node "n" with PhaseIdealLoop
1752 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1753 // Utility to create an if-projection
1754 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1755 // Force the iff control output to be the live_proj
1756 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1757 // Insert a region before an if projection
1758 RegionNode* insert_region_before_proj(ProjNode* proj);
1759 // Insert a new if before an if projection
1760 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1761
1762 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1763 // "Nearly" because all Nodes have been cloned from the original in the loop,
1764 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1765 // through the Phi recursively, and return a Bool.
1766 Node* clone_iff(PhiNode* phi);
1767 CmpNode* clone_bool(PhiNode* phi);
1768
1769
1770 // Rework addressing expressions to get the most loop-invariant stuff
1771 // moved out. We'd like to do all associative operators, but it's especially
1772 // important (common) to do address expressions.
1773 Node* remix_address_expressions(Node* n);
1774 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1775
1776 // Convert add to muladd to generate MuladdS2I under certain criteria
1777 Node * convert_add_to_muladd(Node * n);
1778
1779 // Attempt to use a conditional move instead of a phi/branch
1780 Node *conditional_move( Node *n );
1781
1782 // Check for aggressive application of 'split-if' optimization,
1783 // using basic block level info.
1784 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1785 Node *split_if_with_blocks_pre ( Node *n );
1786 void split_if_with_blocks_post( Node *n );
1787 Node *has_local_phi_input( Node *n );
1788 // Mark an IfNode as being dominated by a prior test,
1789 // without actually altering the CFG (and hence IDOM info).
1790 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool prev_dom_not_imply_this = false);
1791 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool dominator_not_imply_source);
1792
1793 // Split Node 'n' through merge point
1794 RegionNode* split_thru_region(Node* n, RegionNode* region);
1795 // Split Node 'n' through merge point if there is enough win.
1796 Node *split_thru_phi( Node *n, Node *region, int policy );
1797 // Found an If getting its condition-code input from a Phi in the
1798 // same block. Split thru the Region.
1799 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1800
1801 private:
1802 // Class to keep track of wins in split_thru_phi.
1803 class SplitThruPhiWins {
1804 private:
1805 // Region containing the phi we are splitting through.
1806 const Node* _region;
1807
1808 // Sum of all wins regardless of where they happen. This applies to Loops phis as well as non-loop phis.
1809 int _total_wins;
1810
1811 // For Loops, wins have different impact depending on if they happen on loop entry or on the backedge.
1812 // Number of wins on a loop entry edge if the split is through a loop head,
1813 // otherwise 0. Entry edge wins only pay dividends once on loop entry.
1814 int _loop_entry_wins;
1815 // Number of wins on a loop back-edge, which pay dividends on every iteration.
1816 int _loop_back_wins;
1817
1818 public:
1819 SplitThruPhiWins(const Node* region) :
1820 _region(region),
1821 _total_wins(0),
1822 _loop_entry_wins(0),
1823 _loop_back_wins(0) {};
1824
1825 void reset() {_total_wins = 0; _loop_entry_wins = 0; _loop_back_wins = 0;}
1826 void add_win(int ctrl_index) {
1827 if (_region->is_Loop() && ctrl_index == LoopNode::EntryControl) {
1828 _loop_entry_wins++;
1829 } else if (_region->is_Loop() && ctrl_index == LoopNode::LoopBackControl) {
1830 _loop_back_wins++;
1831 }
1832 _total_wins++;
1833 }
1834 // Is this split profitable with respect to the policy?
1835 bool profitable(int policy) const {
1836 assert(_region->is_Loop() || (_loop_entry_wins == 0 && _loop_back_wins == 0), "wins on loop edges without a loop");
1837 assert(!_region->is_Loop() || _total_wins == _loop_entry_wins + _loop_back_wins, "missed some win");
1838 // In general this means that the split has to have more wins than specified
1839 // in the policy. However, for loops we need to take into account where the
1840 // wins happen. We need to be careful when splitting, because splitting nodes
1841 // related to the iv through the phi can sufficiently rearrange the loop
1842 // structure to prevent RCE and thus vectorization. Thus, we only deem splitting
1843 // profitable if the win of a split is not on the entry edge, as such wins
1844 // only pay off once and have a high chance of messing up the loop structure.
1845 return (_loop_entry_wins == 0 && _total_wins > policy) ||
1846 // If there are wins on the entry edge but the backadge also has sufficient wins,
1847 // there is sufficient profitability to spilt regardless of the risk of messing
1848 // up the loop structure.
1849 _loop_back_wins > policy ||
1850 // If the policy is less than 0, a split is always profitable, i.e. we always
1851 // split. This is needed when we split a node and then must also split a
1852 // dependant node, i.e. spliting a Bool node after splitting a Cmp node.
1853 policy < 0;
1854 }
1855 };
1856
1857
1858 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1859
1860 public:
1861
1862 // Conversion of fill/copy patterns into intrinsic versions
1863 bool do_intrinsify_fill();
1864 bool intrinsify_fill(IdealLoopTree* lpt);
1865 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1866 Node*& shift, Node*& offset);
1867
1868 private:
1869 // Helper functions
1870 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1871 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1872 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 );
1873 bool split_up( Node *n, Node *blk1, Node *blk2 );
1874
1875 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1876 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1877 void try_move_store_after_loop(Node* n);
1878 void move_flat_array_check_out_of_loop(Node* n);
1879 bool identical_backtoback_ifs(Node *n);
1880 bool flat_array_element_type_check(Node *n);
1881 bool can_split_if(Node *n_ctrl);
1882 bool cannot_split_division(const Node* n, const Node* region) const;
1883 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1884 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1885
1886 // Determine if a method is too big for a/another round of split-if, based on
1887 // a magic (approximate) ratio derived from the equally magic constant 35000,
1888 // previously used for this purpose (but without relating to the node limit).
1889 bool must_throttle_split_if() {
1890 uint threshold = C->max_node_limit() * 2 / 5;
1891 return C->live_nodes() > threshold;
1892 }
1893
1894 // A simplistic node request tracking mechanism, where
1895 // = UINT_MAX Request not valid or made final.
1896 // < UINT_MAX Nodes currently requested (estimate).
1897 uint _nodes_required;
1898
1899 enum { REQUIRE_MIN = 70 };
1900
1901 uint nodes_required() const { return _nodes_required; }
1902
1903 // Given the _currently_ available number of nodes, check whether there is
1904 // "room" for an additional request or not, considering the already required
1905 // number of nodes. Return TRUE if the new request is exceeding the node
1906 // budget limit, otherwise return FALSE. Note that this interpretation will
1907 // act pessimistic on additional requests when new nodes have already been
1908 // generated since the 'begin'. This behaviour fits with the intention that
1909 // node estimates/requests should be made upfront.
1910 bool exceeding_node_budget(uint required = 0) {
1911 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1912 uint available = C->max_node_limit() - C->live_nodes();
1913 return available < required + _nodes_required + REQUIRE_MIN;
1914 }
1915
1916 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1917 precond(require > 0);
1918 _nodes_required += MAX2(require, minreq);
1919 return _nodes_required;
1920 }
1921
1922 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1923 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1924 }
1925
1926 uint require_nodes_begin() {
1927 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1928 _nodes_required = 0;
1929 return C->live_nodes();
1930 }
1931
1932 // When a node request is final, optionally check that the requested number
1933 // of nodes was reasonably correct with respect to the number of new nodes
1934 // introduced since the last 'begin'. Always check that we have not exceeded
1935 // the maximum node limit.
1936 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1937 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1938
1939 #ifdef ASSERT
1940 if (check_estimate) {
1941 // Check that the node budget request was not off by too much (x2).
1942 // Should this be the case we _surely_ need to improve the estimates
1943 // used in our budget calculations.
1944 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1945 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1946 C->live_nodes() - live_at_begin, _nodes_required);
1947 }
1948 }
1949 #endif
1950 // Assert that we have stayed within the node budget limit.
1951 assert(C->live_nodes() < C->max_node_limit(),
1952 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1953 C->live_nodes() - live_at_begin, live_at_begin,
1954 C->max_node_limit(), _nodes_required);
1955
1956 _nodes_required = UINT_MAX;
1957 }
1958
1959 private:
1960
1961 bool _created_loop_node;
1962 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1963 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1964
1965 public:
1966 void set_created_loop_node() { _created_loop_node = true; }
1967 bool created_loop_node() { return _created_loop_node; }
1968 void register_new_node(Node* n, Node* blk);
1969 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1970 register_new_node(new_node, get_ctrl(ctrl_of));
1971 }
1972
1973 Node* clone_and_register(Node* n, Node* ctrl) {
1974 n = n->clone();
1975 register_new_node(n, ctrl);
1976 return n;
1977 }
1978
1979 #ifdef ASSERT
1980 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1981 #endif
1982
1983 #ifndef PRODUCT
1984 void dump() const;
1985 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1986 void dump_idom(Node* n, uint count) const;
1987 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1988 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1989 IdealLoopTree* get_loop_idx(Node* n) const {
1990 // Dead nodes have no loop, so return the top level loop instead
1991 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1992 }
1993 // Print some stats
1994 static void print_statistics();
1995 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1996 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1997 static volatile int _long_loop_candidates;
1998 static volatile int _long_loop_nests;
1999 #endif
2000
2001 #ifdef ASSERT
2002 void verify() const;
2003 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
2004 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
2005 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
2006 #endif
2007
2008 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
2009
2010 void check_counted_loop_shape(IdealLoopTree* loop, Node* head, BasicType bt) NOT_DEBUG_RETURN;
2011
2012 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
2013
2014
2015 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
2016 Node_List &range_checks);
2017
2018 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
2019 Node* inner_iters_actual_int, Node* inner_phi,
2020 Node* iv_add, LoopNode* inner_head);
2021
2022 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
2023
2024 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
2025
2026 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
2027
2028 bool would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl);
2029
2030 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
2031
2032 void try_sink_out_of_loop(Node* n);
2033
2034 Node* clamp(Node* R, Node* L, Node* H);
2035
2036 bool safe_for_if_replacement(const Node* dom) const;
2037
2038 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
2039
2040 bool try_merge_identical_ifs(Node* n);
2041
2042 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
2043
2044 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2045 IdealLoopTree* parent, bool partial);
2046
2047 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2048 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
2049
2050 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2051 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
2052 Node_List*& split_bool_set, Node_List*& split_cex_set);
2053
2054 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
2055
2056 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
2057
2058 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
2059 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
2060 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
2061 void clone_template_assertion_expression_down(Node* node);
2062
2063 Node* similar_subtype_check(const Node* x, Node* r_in);
2064
2065 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
2066
2067 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
2068
2069 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
2070
2071 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
2072
2073 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
2074
2075 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
2076
2077 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
2078
2079 ConINode* intcon(jint i);
2080
2081 ConLNode* longcon(jlong i);
2082
2083 ConNode* makecon(const Type* t);
2084
2085 ConNode* integercon(jlong l, BasicType bt);
2086
2087 ConNode* zerocon(BasicType bt);
2088 };
2089
2090 class CountedLoopConverter {
2091 friend class PhaseIdealLoop;
2092
2093 // Match increment with optional truncation
2094 class TruncatedIncrement {
2095 bool _is_valid;
2096
2097 BasicType _bt;
2098
2099 Node* _incr;
2100 Node* _outer_trunc;
2101 Node* _inner_trunc;
2102 const TypeInteger* _trunc_type;
2103
2104 public:
2105 TruncatedIncrement(BasicType bt) :
2106 _is_valid(false),
2107 _bt(bt),
2108 _incr(nullptr),
2109 _outer_trunc(nullptr),
2110 _inner_trunc(nullptr),
2111 _trunc_type(nullptr) {}
2112
2113 void build(Node* expr);
2114
2115 bool is_valid() const { return _is_valid; }
2116 Node* incr() const { return _incr; }
2117
2118 // Optional truncation for: CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2119 Node* outer_trunc() const { return _outer_trunc; } // the outermost truncating node (either the & or the final >>)
2120 Node* inner_trunc() const { return _inner_trunc; } // the inner truncating node, if applicable (the << in a <</>> pair)
2121 const TypeInteger* trunc_type() const { return _trunc_type; }
2122 };
2123
2124 class LoopStructure {
2125 bool _is_valid;
2126
2127 const Node* _head;
2128 const IdealLoopTree* _loop;
2129 PhaseIdealLoop* _phase;
2130 BasicType _iv_bt;
2131
2132 Node* _back_control;
2133 PhaseIdealLoop::LoopExitTest _exit_test;
2134 PhaseIdealLoop::LoopIVIncr _iv_incr;
2135 TruncatedIncrement _truncated_increment;
2136 PhaseIdealLoop::LoopIVStride _stride;
2137 PhiNode* _phi;
2138 SafePointNode* _safepoint;
2139
2140 public:
2141 LoopStructure(const Node* head, const IdealLoopTree* loop, PhaseIdealLoop* phase, const BasicType iv_bt) :
2142 _is_valid(false),
2143 _head(head),
2144 _loop(loop),
2145 _phase(phase),
2146 _iv_bt(iv_bt),
2147 _back_control(_phase->loop_exit_control(_loop)),
2148 _exit_test(_back_control, _loop, _phase),
2149 _iv_incr(_head, _loop),
2150 _truncated_increment(_iv_bt),
2151 _stride(PhaseIdealLoop::LoopIVStride(_iv_bt)),
2152 _phi(nullptr),
2153 _safepoint(nullptr) {}
2154
2155 void build();
2156
2157 jlong final_limit_correction() const; // compute adjusted loop limit correction
2158 bool is_infinite_loop() const;
2159
2160 bool is_valid() const { return _is_valid; }
2161
2162 Node* back_control() const { return _back_control; }
2163 PhaseIdealLoop::LoopExitTest& exit_test() { return _exit_test; }
2164 PhaseIdealLoop::LoopIVIncr& iv_incr() { return _iv_incr; }
2165 TruncatedIncrement& truncated_increment() { return _truncated_increment; }
2166 PhaseIdealLoop::LoopIVStride& stride() { return _stride; }
2167 PhiNode* phi() const { return _phi; }
2168 SafePointNode* sfpt() const { return _safepoint; }
2169 jlong stride_con() const { return _stride.compute_non_zero_stride_con(_exit_test.mask(), _iv_bt); }
2170 Node* limit() const { return _exit_test.limit(); }
2171 };
2172
2173 PhaseIdealLoop* const _phase;
2174 Node* const _head;
2175 IdealLoopTree* const _loop;
2176 const BasicType _iv_bt;
2177
2178 LoopStructure _structure;
2179 bool _should_insert_stride_overflow_limit_check = false;
2180 bool _should_insert_init_trip_limit_check = false;
2181
2182 DEBUG_ONLY(bool _checked_for_counted_loop = false;)
2183
2184 // stats for PhaseIdealLoop::print_statistics()
2185 static volatile int _long_loop_counted_loops;
2186
2187 // Return a type based on condition control flow
2188 const TypeInt* filtered_type(Node* n, Node* n_ctrl);
2189 const TypeInt* filtered_type(Node* n) { return filtered_type(n, nullptr); }
2190 // Helpers for filtered type
2191 const TypeInt* filtered_type_from_dominators(Node* val, Node* val_ctrl);
2192
2193 void insert_loop_limit_check_predicate(const ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* bol) const;
2194 void insert_stride_overflow_limit_check() const;
2195 void insert_init_trip_limit_check() const;
2196 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
2197 Node* loop_entry) const;
2198
2199 bool is_iv_overflowing(const TypeInteger* init_t, jlong stride_con, Node* phi_increment, BoolTest::mask mask) const;
2200 bool has_truncation_wrap(const TruncatedIncrement& truncation, Node* phi, jlong stride_con);
2201 SafePointNode* find_safepoint(Node* iftrue);
2202 bool is_safepoint_invalid(SafePointNode* sfpt) const;
2203
2204 public:
2205 CountedLoopConverter(PhaseIdealLoop* phase, Node* head, IdealLoopTree* loop, const BasicType iv_bt)
2206 : _phase(phase),
2207 _head(head),
2208 _loop(loop),
2209 _iv_bt(iv_bt),
2210 _structure(LoopStructure(_head, _loop, _phase, _iv_bt)) {
2211 assert(phase != nullptr, "must be"); // Fail early if mandatory parameters are null.
2212 assert(head != nullptr, "must be");
2213 assert(loop != nullptr, "must be");
2214 assert(iv_bt == T_INT || iv_bt == T_LONG, "either int or long loops");
2215 }
2216
2217 bool is_counted_loop();
2218 IdealLoopTree* convert();
2219
2220 DEBUG_ONLY(bool should_stress_long_counted_loop();)
2221 DEBUG_ONLY(bool stress_long_counted_loop();)
2222
2223 enum StrideOverflowState {
2224 Overflow = -1,
2225 NoOverflow = 0,
2226 RequireLimitCheck = 1
2227 };
2228 static StrideOverflowState check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt);
2229 };
2230
2231 class AutoNodeBudget : public StackObj
2232 {
2233 public:
2234 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
2235
2236 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
2237 : _phase(phase),
2238 _check_at_final(chk == BUDGET_CHECK),
2239 _nodes_at_begin(0)
2240 {
2241 precond(_phase != nullptr);
2242
2243 _nodes_at_begin = _phase->require_nodes_begin();
2244 }
2245
2246 ~AutoNodeBudget() {
2247 #ifndef PRODUCT
2248 if (TraceLoopOpts) {
2249 uint request = _phase->nodes_required();
2250 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
2251
2252 if (request < delta) {
2253 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
2254 } else {
2255 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
2256 // Identify the worst estimates as "poor" ones.
2257 if (request > REQUIRE_MIN && delta > 0) {
2258 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
2259 (delta <= REQUIRE_MIN && request > 10 * delta)) {
2260 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
2261 }
2262 }
2263 }
2264 }
2265 #endif // PRODUCT
2266 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
2267 }
2268
2269 private:
2270 PhaseIdealLoop* _phase;
2271 bool _check_at_final;
2272 uint _nodes_at_begin;
2273 };
2274
2275 inline Node* IdealLoopTree::tail() {
2276 // Handle lazy update of _tail field.
2277 if (_tail->in(0) == nullptr) {
2278 _tail = _phase->get_ctrl(_tail);
2279 }
2280 return _tail;
2281 }
2282
2283 inline Node* IdealLoopTree::head() {
2284 // Handle lazy update of _head field.
2285 if (_head->in(0) == nullptr) {
2286 _head = _phase->get_ctrl(_head);
2287 }
2288 return _head;
2289 }
2290
2291 // Iterate over the loop tree using a preorder, left-to-right traversal.
2292 //
2293 // Example that visits all counted loops from within PhaseIdealLoop
2294 //
2295 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2296 // IdealLoopTree* lpt = iter.current();
2297 // if (!lpt->is_counted()) continue;
2298 // ...
2299 class LoopTreeIterator : public StackObj {
2300 private:
2301 IdealLoopTree* _root;
2302 IdealLoopTree* _curnt;
2303
2304 public:
2305 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
2306
2307 bool done() { return _curnt == nullptr; } // Finished iterating?
2308
2309 void next(); // Advance to next loop tree
2310
2311 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
2312 };
2313
2314 // Compute probability of reaching some CFG node from a fixed
2315 // dominating CFG node
2316 class PathFrequency {
2317 private:
2318 Node* _dom; // frequencies are computed relative to this node
2319 Node_Stack _stack;
2320 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
2321 GrowableArray<float> _freqs; // cache frequencies
2322 PhaseIdealLoop* _phase;
2323
2324 float check_and_truncate_frequency(float f) {
2325 assert(f >= 0, "Incorrect frequency");
2326 // We do not perform an exact (f <= 1) check
2327 // this would be error prone with rounding of floats.
2328 // Performing a check like (f <= 1+eps) would be of benefit,
2329 // however, it is not evident how to determine such an eps,
2330 // given that an arbitrary number of add/mul operations
2331 // are performed on these frequencies.
2332 return (f > 1) ? 1 : f;
2333 }
2334
2335 public:
2336 PathFrequency(Node* dom, PhaseIdealLoop* phase)
2337 : _dom(dom), _stack(0), _phase(phase) {
2338 }
2339
2340 float to(Node* n);
2341 };
2342
2343 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
2344 // 1. Clone the data nodes
2345 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
2346 class DataNodeGraph : public StackObj {
2347 PhaseIdealLoop* const _phase;
2348 const Unique_Node_List& _data_nodes;
2349 OrigToNewHashtable _orig_to_new;
2350
2351 public:
2352 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
2353 : _phase(phase),
2354 _data_nodes(data_nodes),
2355 // Use 107 as best guess which is the first resize value in ResizeableHashTable::large_table_sizes.
2356 _orig_to_new(107, MaxNodeLimit)
2357 {
2358 #ifdef ASSERT
2359 for (uint i = 0; i < data_nodes.size(); i++) {
2360 assert(!data_nodes[i]->is_CFG(), "only data nodes");
2361 }
2362 #endif
2363 }
2364 NONCOPYABLE(DataNodeGraph);
2365
2366 private:
2367 void clone(Node* node, Node* new_ctrl);
2368 void clone_data_nodes(Node* new_ctrl);
2369 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2370 Node* new_ctrl);
2371 void rewire_clones_to_cloned_inputs();
2372 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
2373
2374 public:
2375 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
2376 // Set 'new_ctrl' as ctrl for the cloned nodes.
2377 const OrigToNewHashtable& clone(Node* new_ctrl) {
2378 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
2379 clone_data_nodes(new_ctrl);
2380 rewire_clones_to_cloned_inputs();
2381 return _orig_to_new;
2382 }
2383
2384 // Create a copy of the data nodes provided to the constructor by doing the following:
2385 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
2386 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2387 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2388 // `new_ctrl` as ctrl.
2389 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2390 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2391 Node* new_ctrl) {
2392 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2393 rewire_clones_to_cloned_inputs();
2394 return _orig_to_new;
2395 }
2396 };
2397 #endif // SHARE_OPTO_LOOPNODE_HPP