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 LoopSelector;
47 class UnswitchedLoopSelector;
48 class VectorSet;
49 class VSharedData;
50 class Invariance;
51 struct small_cache;
52
53 //
54 // I D E A L I Z E D L O O P S
55 //
56 // Idealized loops are the set of loops I perform more interesting
57 // transformations on, beyond simple hoisting.
58
59 //------------------------------LoopNode---------------------------------------
60 // Simple loop header. Fall in path on left, loop-back path on right.
61 class LoopNode : public RegionNode {
62 // Size is bigger to hold the flags. However, the flags do not change
63 // the semantics so it does not appear in the hash & cmp functions.
64 virtual uint size_of() const { return sizeof(*this); }
65 protected:
66 uint _loop_flags;
67 // Names for flag bitfields
68 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
69 MainHasNoPreLoop = 1<<2,
70 HasExactTripCount = 1<<3,
71 InnerLoop = 1<<4,
72 PartialPeelLoop = 1<<5,
73 PartialPeelFailed = 1<<6,
74 WasSlpAnalyzed = 1<<7,
75 PassedSlpAnalysis = 1<<8,
76 DoUnrollOnly = 1<<9,
77 VectorizedLoop = 1<<10,
78 HasAtomicPostLoop = 1<<11,
79 StripMined = 1<<12,
80 SubwordLoop = 1<<13,
81 ProfileTripFailed = 1<<14,
82 LoopNestInnerLoop = 1<<15,
83 LoopNestLongOuterLoop = 1<<16,
84 MultiversionFastLoop = 1<<17,
85 MultiversionSlowLoop = 2<<17,
86 MultiversionDelayedSlowLoop = 3<<17,
87 MultiversionFlagsMask = 3<<17,
88 };
89 char _unswitch_count;
90 enum { _unswitch_max=3 };
91
92 // Expected trip count from profile data
93 float _profile_trip_cnt;
94
95 public:
96 // Names for edge indices
97 enum { Self=0, EntryControl, LoopBackControl };
98
99 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
100 void set_inner_loop() { _loop_flags |= InnerLoop; }
101
102 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
103 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
104 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
105 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
106 bool is_strip_mined() const { return _loop_flags & StripMined; }
107 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
108 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
109 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
110 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
111
112 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
113 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
114 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
115 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
116 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
117 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
118 void mark_strip_mined() { _loop_flags |= StripMined; }
119 void clear_strip_mined() { _loop_flags &= ~StripMined; }
120 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
121 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
122 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
123 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
124
125 int unswitch_max() { return _unswitch_max; }
126 int unswitch_count() { return _unswitch_count; }
127
128 void set_unswitch_count(int val) {
129 assert (val <= unswitch_max(), "too many unswitches");
130 _unswitch_count = val;
131 }
132
133 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
134 float profile_trip_cnt() { return _profile_trip_cnt; }
135
136 #ifndef PRODUCT
137 uint _stress_peeling_attempts = 0;
138 #endif
139
140 LoopNode(Node *entry, Node *backedge)
141 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
142 _profile_trip_cnt(COUNT_UNKNOWN) {
143 init_class_id(Class_Loop);
144 init_req(EntryControl, entry);
145 init_req(LoopBackControl, backedge);
146 }
147
148 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
149 virtual int Opcode() const;
150 bool can_be_counted_loop(PhaseValues* phase) const {
151 return req() == 3 && in(0) != nullptr &&
152 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
153 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
154 }
155 bool is_valid_counted_loop(BasicType bt) const;
156 #ifndef PRODUCT
157 virtual void dump_spec(outputStream *st) const;
158 #endif
159
160 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
161 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
162 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
163 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
164 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
165 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
166 };
167
168 //------------------------------Counted Loops----------------------------------
169 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
170 // path (and maybe some other exit paths). The trip-counter exit is always
171 // last in the loop. The trip-counter have to stride by a constant;
172 // the exit value is also loop invariant.
173
174 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
175 // CountedLoopNode has the incoming loop control and the loop-back-control
176 // which is always the IfTrue before the matching CountedLoopEndNode. The
177 // CountedLoopEndNode has an incoming control (possibly not the
178 // CountedLoopNode if there is control flow in the loop), the post-increment
179 // trip-counter value, and the limit. The trip-counter value is always of
180 // the form (Op old-trip-counter stride). The old-trip-counter is produced
181 // by a Phi connected to the CountedLoopNode. The stride is constant.
182 // The Op is any commutable opcode, including Add, Mul, Xor. The
183 // CountedLoopEndNode also takes in the loop-invariant limit value.
184
185 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
186 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
187 // via the old-trip-counter from the Op node.
188
189 //------------------------------CountedLoopNode--------------------------------
190 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
191 // inputs the incoming loop-start control and the loop-back control, so they
192 // act like RegionNodes. They also take in the initial trip counter, the
193 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
194 // produce a loop-body control and the trip counter value. Since
195 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
196
197 class BaseCountedLoopNode : public LoopNode {
198 public:
199 BaseCountedLoopNode(Node *entry, Node *backedge)
200 : LoopNode(entry, backedge) {
201 }
202
203 Node *init_control() const { return in(EntryControl); }
204 Node *back_control() const { return in(LoopBackControl); }
205
206 Node* init_trip() const;
207 Node* stride() const;
208 bool stride_is_con() const;
209 Node* limit() const;
210 Node* incr() const;
211 Node* phi() const;
212
213 BaseCountedLoopEndNode* loopexit_or_null() const;
214 BaseCountedLoopEndNode* loopexit() const;
215
216 virtual BasicType bt() const = 0;
217
218 jlong stride_con() const;
219
220 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
221
222 virtual void set_trip_count(julong tc) = 0;
223 virtual julong trip_count() const = 0;
224
225 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
226 void set_exact_trip_count(julong tc) {
227 set_trip_count(tc);
228 _loop_flags |= HasExactTripCount;
229 }
230 void set_nonexact_trip_count() {
231 _loop_flags &= ~HasExactTripCount;
232 }
233 };
234
235
236 class CountedLoopNode : public BaseCountedLoopNode {
237 // Size is bigger to hold _main_idx. However, _main_idx does not change
238 // the semantics so it does not appear in the hash & cmp functions.
239 virtual uint size_of() const { return sizeof(*this); }
240
241 // For Pre- and Post-loops during debugging ONLY, this holds the index of
242 // the Main CountedLoop. Used to assert that we understand the graph shape.
243 node_idx_t _main_idx;
244
245 // Known trip count calculated by compute_exact_trip_count()
246 uint _trip_count;
247
248 // Log2 of original loop bodies in unrolled loop
249 int _unrolled_count_log2;
250
251 // Node count prior to last unrolling - used to decide if
252 // unroll,optimize,unroll,optimize,... is making progress
253 int _node_count_before_unroll;
254
255 // If slp analysis is performed we record the maximum
256 // vector mapped unroll factor here
257 int _slp_maximum_unroll_factor;
258
259 public:
260 CountedLoopNode(Node *entry, Node *backedge)
261 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
262 _unrolled_count_log2(0), _node_count_before_unroll(0),
263 _slp_maximum_unroll_factor(0) {
264 init_class_id(Class_CountedLoop);
265 // Initialize _trip_count to the largest possible value.
266 // Will be reset (lower) if the loop's trip count is known.
267 }
268
269 virtual int Opcode() const;
270 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
271
272 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
273 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
274 int stride_con() const;
275
276 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
277 // can run short a few iterations and may start a few iterations in.
278 // It will be RCE'd and unrolled and aligned.
279
280 // A following 'post' loop will run any remaining iterations. Used
281 // during Range Check Elimination, the 'post' loop will do any final
282 // iterations with full checks. Also used by Loop Unrolling, where
283 // the 'post' loop will do any epilog iterations needed. Basically,
284 // a 'post' loop can not profitably be further unrolled or RCE'd.
285
286 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
287 // it may do under-flow checks for RCE and may do alignment iterations
288 // so the following main loop 'knows' that it is striding down cache
289 // lines.
290
291 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
292 // Aligned, may be missing it's pre-loop.
293 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
294 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
295 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
296 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
297 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
298 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
299 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
300 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
301 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
302 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
303
304 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
305
306 int main_idx() const { return _main_idx; }
307
308 void set_trip_count(julong tc) {
309 assert(tc < max_juint, "Cannot set trip count to max_juint");
310 _trip_count = checked_cast<uint>(tc);
311 }
312 julong trip_count() const { return _trip_count; }
313
314 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
315 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
316 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
317 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
318
319 void set_notpassed_slp() {
320 _loop_flags &= ~PassedSlpAnalysis;
321 }
322
323 void double_unrolled_count() { _unrolled_count_log2++; }
324 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
325
326 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
327 int node_count_before_unroll() { return _node_count_before_unroll; }
328 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
329 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
330
331 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
332 // decides which one is taken:
333 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
334 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
335 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
336 // assumption fails we take the slow_loop instead, where we do not make the same speculative
337 // assumption.
338 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
339 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
340 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
341 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
342 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
343 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
344 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
345 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
346 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
347 // Hence, we expect the slow_loop to execute slower.
348 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
349 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
350 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
351 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
352 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
353 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
354 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
355 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
356
357 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
358 OuterStripMinedLoopNode* outer_loop() const;
359 virtual IfTrueNode* outer_loop_tail() const;
360 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
361 virtual IfFalseNode* outer_loop_exit() const;
362 virtual SafePointNode* outer_safepoint() const;
363
364 Node* skip_assertion_predicates_with_halt();
365
366 virtual BasicType bt() const {
367 return T_INT;
368 }
369
370 Node* is_canonical_loop_entry();
371 CountedLoopEndNode* find_pre_loop_end();
372
373 Node* uncasted_init_trip(bool uncasted);
374
375 #ifndef PRODUCT
376 virtual void dump_spec(outputStream *st) const;
377 #endif
378 };
379
380 class LongCountedLoopNode : public BaseCountedLoopNode {
381 private:
382 virtual uint size_of() const { return sizeof(*this); }
383
384 // Known trip count calculated by compute_exact_trip_count()
385 julong _trip_count;
386
387 public:
388 LongCountedLoopNode(Node *entry, Node *backedge)
389 : BaseCountedLoopNode(entry, backedge), _trip_count(max_julong) {
390 init_class_id(Class_LongCountedLoop);
391 }
392
393 virtual int Opcode() const;
394
395 virtual BasicType bt() const {
396 return T_LONG;
397 }
398
399 void set_trip_count(julong tc) {
400 assert(tc < max_julong, "Cannot set trip count to max_julong");
401 _trip_count = tc;
402 }
403 julong trip_count() const { return _trip_count; }
404
405 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
406 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
407 };
408
409
410 //------------------------------CountedLoopEndNode-----------------------------
411 // CountedLoopEndNodes end simple trip counted loops. They act much like
412 // IfNodes.
413
414 class BaseCountedLoopEndNode : public IfNode {
415 public:
416 enum { TestControl, TestValue };
417 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
418 : IfNode(control, test, prob, cnt) {
419 init_class_id(Class_BaseCountedLoopEnd);
420 }
421
422 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
423 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
424 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
425 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
426 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
427 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
428
429 PhiNode* phi() const {
430 Node* tmp = incr();
431 if (tmp && tmp->req() == 3) {
432 Node* phi = tmp->in(1);
433 if (phi->is_Phi()) {
434 return phi->as_Phi();
435 }
436 }
437 return nullptr;
438 }
439
440 BaseCountedLoopNode* loopnode() const {
441 // The CountedLoopNode that goes with this CountedLoopEndNode may
442 // have been optimized out by the IGVN so be cautious with the
443 // pattern matching on the graph
444 PhiNode* iv_phi = phi();
445 if (iv_phi == nullptr) {
446 return nullptr;
447 }
448 Node* ln = iv_phi->in(0);
449 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
450 return nullptr;
451 }
452 if (ln->as_BaseCountedLoop()->bt() != bt()) {
453 return nullptr;
454 }
455 return ln->as_BaseCountedLoop();
456 }
457
458 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
459
460 jlong stride_con() const;
461 virtual BasicType bt() const = 0;
462
463 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
464 };
465
466 class CountedLoopEndNode : public BaseCountedLoopEndNode {
467 public:
468
469 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
470 : BaseCountedLoopEndNode(control, test, prob, cnt) {
471 init_class_id(Class_CountedLoopEnd);
472 }
473 virtual int Opcode() const;
474
475 CountedLoopNode* loopnode() const {
476 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
477 }
478
479 virtual BasicType bt() const {
480 return T_INT;
481 }
482
483 #ifndef PRODUCT
484 virtual void dump_spec(outputStream *st) const;
485 #endif
486 };
487
488 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
489 public:
490 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
491 : BaseCountedLoopEndNode(control, test, prob, cnt) {
492 init_class_id(Class_LongCountedLoopEnd);
493 }
494
495 LongCountedLoopNode* loopnode() const {
496 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
497 }
498
499 virtual int Opcode() const;
500
501 virtual BasicType bt() const {
502 return T_LONG;
503 }
504 };
505
506
507 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
508 Node* bctrl = back_control();
509 if (bctrl == nullptr) return nullptr;
510
511 Node* lexit = bctrl->in(0);
512 if (!lexit->is_BaseCountedLoopEnd()) {
513 return nullptr;
514 }
515 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
516 if (result->bt() != bt()) {
517 return nullptr;
518 }
519 return result;
520 }
521
522 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
523 BaseCountedLoopEndNode* cle = loopexit_or_null();
524 assert(cle != nullptr, "loopexit is null");
525 return cle;
526 }
527
528 inline Node* BaseCountedLoopNode::init_trip() const {
529 BaseCountedLoopEndNode* cle = loopexit_or_null();
530 return cle != nullptr ? cle->init_trip() : nullptr;
531 }
532 inline Node* BaseCountedLoopNode::stride() const {
533 BaseCountedLoopEndNode* cle = loopexit_or_null();
534 return cle != nullptr ? cle->stride() : nullptr;
535 }
536
537 inline bool BaseCountedLoopNode::stride_is_con() const {
538 BaseCountedLoopEndNode* cle = loopexit_or_null();
539 return cle != nullptr && cle->stride_is_con();
540 }
541 inline Node* BaseCountedLoopNode::limit() const {
542 BaseCountedLoopEndNode* cle = loopexit_or_null();
543 return cle != nullptr ? cle->limit() : nullptr;
544 }
545 inline Node* BaseCountedLoopNode::incr() const {
546 BaseCountedLoopEndNode* cle = loopexit_or_null();
547 return cle != nullptr ? cle->incr() : nullptr;
548 }
549 inline Node* BaseCountedLoopNode::phi() const {
550 BaseCountedLoopEndNode* cle = loopexit_or_null();
551 return cle != nullptr ? cle->phi() : nullptr;
552 }
553
554 inline jlong BaseCountedLoopNode::stride_con() const {
555 BaseCountedLoopEndNode* cle = loopexit_or_null();
556 return cle != nullptr ? cle->stride_con() : 0;
557 }
558
559
560 //------------------------------LoopLimitNode-----------------------------
561 // Counted Loop limit node which represents exact final iterator value:
562 // trip_count = (limit - init_trip + stride - 1)/stride
563 // final_value= trip_count * stride + init_trip.
564 // Use HW instructions to calculate it when it can overflow in integer.
565 // Note, final_value should fit into integer since counted loop has
566 // limit check: limit <= max_int-stride.
567 class LoopLimitNode : public Node {
568 enum { Init=1, Limit=2, Stride=3 };
569 public:
570 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
571 // Put it on the Macro nodes list to optimize during macro nodes expansion.
572 init_flags(Flag_is_macro);
573 C->add_macro_node(this);
574 }
575 virtual int Opcode() const;
576 virtual const Type *bottom_type() const { return TypeInt::INT; }
577 virtual uint ideal_reg() const { return Op_RegI; }
578 virtual const Type* Value(PhaseGVN* phase) const;
579 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
580 virtual Node* Identity(PhaseGVN* phase);
581 };
582
583 // Support for strip mining
584 class OuterStripMinedLoopNode : public LoopNode {
585 private:
586 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const;
587 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn);
588
589 public:
590 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
591 : LoopNode(entry, backedge) {
592 init_class_id(Class_OuterStripMinedLoop);
593 init_flags(Flag_is_macro);
594 C->add_macro_node(this);
595 }
596
597 virtual int Opcode() const;
598
599 virtual IfTrueNode* outer_loop_tail() const;
600 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
601 virtual IfFalseNode* outer_loop_exit() const;
602 virtual SafePointNode* outer_safepoint() const;
603 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); }
604 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); }
605 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->false_proj(); }
606
607 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
608
609 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
610
611 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
612
613 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
614
615 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
616 PhaseIdealLoop* iloop);
617 };
618
619 class OuterStripMinedLoopEndNode : public IfNode {
620 public:
621 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
622 : IfNode(control, test, prob, cnt) {
623 init_class_id(Class_OuterStripMinedLoopEnd);
624 }
625
626 virtual int Opcode() const;
627
628 virtual const Type* Value(PhaseGVN* phase) const;
629 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
630
631 bool is_expanded(PhaseGVN *phase) const;
632 };
633
634 // -----------------------------IdealLoopTree----------------------------------
635 class IdealLoopTree : public ResourceObj {
636 public:
637 IdealLoopTree *_parent; // Parent in loop tree
638 IdealLoopTree *_next; // Next sibling in loop tree
639 IdealLoopTree *_child; // First child in loop tree
640
641 // The head-tail backedge defines the loop.
642 // If a loop has multiple backedges, this is addressed during cleanup where
643 // we peel off the multiple backedges, merging all edges at the bottom and
644 // ensuring that one proper backedge flow into the loop.
645 Node *_head; // Head of loop
646 Node *_tail; // Tail of loop
647 inline Node *tail(); // Handle lazy update of _tail field
648 inline Node *head(); // Handle lazy update of _head field
649 PhaseIdealLoop* _phase;
650 int _local_loop_unroll_limit;
651 int _local_loop_unroll_factor;
652
653 Node_List _body; // Loop body for inner loops
654
655 uint16_t _nest; // Nesting depth
656 uint8_t _irreducible:1, // True if irreducible
657 _has_call:1, // True if has call safepoint
658 _has_sfpt:1, // True if has non-call safepoint
659 _rce_candidate:1, // True if candidate for range check elimination
660 _has_range_checks:1,
661 _has_range_checks_computed:1;
662
663 Node_List* _safepts; // List of safepoints in this loop
664 Node_List* _required_safept; // A inner loop cannot delete these safepts;
665 bool _allow_optimizations; // Allow loop optimizations
666
667 IdealLoopTree(PhaseIdealLoop* phase, Node* head, Node* tail);
668
669 // Is 'l' a member of 'this'?
670 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
671
672 // Set loop nesting depth. Accumulate has_call bits.
673 int set_nest( uint depth );
674
675 // Split out multiple fall-in edges from the loop header. Move them to a
676 // private RegionNode before the loop. This becomes the loop landing pad.
677 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
678
679 // Split out the outermost loop from this shared header.
680 void split_outer_loop( PhaseIdealLoop *phase );
681
682 // Merge all the backedges from the shared header into a private Region.
683 // Feed that region as the one backedge to this loop.
684 void merge_many_backedges( PhaseIdealLoop *phase );
685
686 // Split shared headers and insert loop landing pads.
687 // Insert a LoopNode to replace the RegionNode.
688 // Returns TRUE if loop tree is structurally changed.
689 bool beautify_loops( PhaseIdealLoop *phase );
690
691 // Perform optimization to use the loop predicates for null checks and range checks.
692 // Applies to any loop level (not just the innermost one)
693 bool loop_predication( PhaseIdealLoop *phase);
694 bool can_apply_loop_predication();
695
696 // Perform iteration-splitting on inner loops. Split iterations to
697 // avoid range checks or one-shot null checks. Returns false if the
698 // current round of loop opts should stop.
699 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
700
701 // Driver for various flavors of iteration splitting. Returns false
702 // if the current round of loop opts should stop.
703 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
704
705 // Given dominators, try to find loops with calls that must always be
706 // executed (call dominates loop tail). These loops do not need non-call
707 // safepoints (ncsfpt).
708 void check_safepts(VectorSet &visited, Node_List &stack);
709
710 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
711 // encountered.
712 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
713
714 // Remove safepoints from loop. Optionally keeping one.
715 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
716
717 // Convert to counted loops where possible
718 void counted_loop( PhaseIdealLoop *phase );
719
720 // Check for Node being a loop-breaking test
721 Node *is_loop_exit(Node *iff) const;
722
723 // Remove simplistic dead code from loop body
724 void DCE_loop_body();
725
726 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
727 // Replace with a 1-in-10 exit guess.
728 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
729
730 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
731 // Useful for unrolling loops with NO array accesses.
732 bool policy_peel_only( PhaseIdealLoop *phase ) const;
733
734 // Return TRUE or FALSE if the loop should be unswitched -- clone
735 // loop with an invariant test
736 bool policy_unswitching( PhaseIdealLoop *phase ) const;
737
738 // Micro-benchmark spamming. Remove empty loops.
739 bool do_remove_empty_loop( PhaseIdealLoop *phase );
740
741 // Convert one-iteration loop into normal code.
742 bool do_one_iteration_loop( PhaseIdealLoop *phase );
743
744 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
745 // move some loop-invariant test (usually a null-check) before the loop.
746 bool policy_peeling(PhaseIdealLoop *phase);
747
748 uint estimate_peeling(PhaseIdealLoop *phase);
749
750 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
751 // known trip count in the counted loop node.
752 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
753
754 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
755 // if the loop is a counted loop and the loop body is small enough.
756 bool policy_unroll(PhaseIdealLoop *phase);
757
758 // Loop analyses to map to a maximal superword unrolling for vectorization.
759 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
760
761 // Return TRUE or FALSE if the loop should be range-check-eliminated.
762 // Gather a list of IF tests that are dominated by iteration splitting;
763 // also gather the end of the first split and the start of the 2nd split.
764 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
765
766 // Return TRUE if "iff" is a range check.
767 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
768 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
769 jlong& scale) const;
770
771 // Estimate the number of nodes required when cloning a loop (body).
772 uint est_loop_clone_sz(uint factor) const;
773 // Estimate the number of nodes required when unrolling a loop (body).
774 uint est_loop_unroll_sz(uint factor) const;
775
776 // Compute loop trip count if possible
777 void compute_trip_count(PhaseIdealLoop* phase, BasicType bt);
778
779 // Compute loop trip count from profile data
780 float compute_profile_trip_cnt_helper(Node* n);
781 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
782
783 // Reassociate invariant expressions.
784 void reassociate_invariants(PhaseIdealLoop *phase);
785 // Reassociate invariant binary expressions.
786 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
787 // Reassociate invariant add, subtract, and compare expressions.
788 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
789 // Return nonzero index of invariant operand if invariant and variant
790 // are combined with an associative binary. Helper for reassociate_invariants.
791 int find_invariant(Node* n, PhaseIdealLoop *phase);
792 // Return TRUE if "n" is associative.
793 bool is_associative(Node* n, Node* base=nullptr);
794 // Return TRUE if "n" is an associative cmp node.
795 bool is_associative_cmp(Node* n);
796
797 // Return true if n is invariant
798 bool is_invariant(Node* n) const;
799
800 // Put loop body on igvn work list
801 void record_for_igvn();
802
803 bool is_root() { return _parent == nullptr; }
804 // A proper/reducible loop w/o any (occasional) dead back-edge.
805 bool is_loop() { return !_irreducible && !tail()->is_top(); }
806 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
807 bool is_innermost() { return is_loop() && _child == nullptr; }
808
809 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
810
811 bool compute_has_range_checks() const;
812 bool range_checks_present() {
813 if (!_has_range_checks_computed) {
814 if (compute_has_range_checks()) {
815 _has_range_checks = 1;
816 }
817 _has_range_checks_computed = 1;
818 }
819 return _has_range_checks;
820 }
821
822 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
823 // In all other cases, return this.
824 IdealLoopTree* skip_strip_mined() {
825 return _head->as_Loop()->is_strip_mined() ? _parent : this;
826 }
827
828 #ifndef PRODUCT
829 void dump_head(); // Dump loop head only
830 void dump(); // Dump this loop recursively
831 #endif
832
833 #ifdef ASSERT
834 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
835 bool verify_tree(IdealLoopTree* loop_verify) const;
836 #endif
837
838 private:
839 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
840
841 // Estimate the number of nodes resulting from control and data flow merge.
842 uint est_loop_flow_merge_sz() const;
843
844 // Check if the number of residual iterations is large with unroll_cnt.
845 // Return true if the residual iterations are more than 10% of the trip count.
846 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
847 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
848 }
849
850 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
851
852 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
853
854 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
855
856 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
857 Node* sfpt) const;
858
859 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
860
861 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
862 };
863
864 // -----------------------------PhaseIdealLoop---------------------------------
865 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
866 // into a loop tree. Drives the loop-based transformations on the ideal graph.
867 class PhaseIdealLoop : public PhaseTransform {
868 friend class IdealLoopTree;
869 friend class SuperWord;
870 friend class ShenandoahBarrierC2Support;
871 friend class AutoNodeBudget;
872
873 Arena _arena; // For data whose lifetime is a single pass of loop optimizations
874
875 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
876 //
877 // Exception: dead CFG nodes may instead have a ctrl/idom forwarding
878 // installed. See: forward_ctrl
879 Node_List _loop_or_ctrl;
880
881 // Pre-computed def-use info
882 PhaseIterGVN &_igvn;
883
884 // Head of loop tree
885 IdealLoopTree* _ltree_root;
886
887 // Array of pre-order numbers, plus post-visited bit.
888 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
889 // ODD for post-visited. Other bits are the pre-order number.
890 uint *_preorders;
891 uint _max_preorder;
892
893 ReallocMark _nesting; // Safety checks for arena reallocation
894
895 const PhaseIdealLoop* _verify_me;
896 bool _verify_only;
897
898 // Allocate _preorders[] array
899 void allocate_preorders() {
900 _max_preorder = C->unique()+8;
901 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
902 memset(_preorders, 0, sizeof(uint) * _max_preorder);
903 }
904
905 // Allocate _preorders[] array
906 void reallocate_preorders() {
907 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
908 if ( _max_preorder < C->unique() ) {
909 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
910 _max_preorder = C->unique();
911 }
912 memset(_preorders, 0, sizeof(uint) * _max_preorder);
913 }
914
915 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
916 // adds new nodes.
917 void check_grow_preorders( ) {
918 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
919 if ( _max_preorder < C->unique() ) {
920 uint newsize = _max_preorder<<1; // double size of array
921 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
922 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
923 _max_preorder = newsize;
924 }
925 }
926 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
927 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
928 // Pre-order numbers are written to the Nodes array as low-bit-set values.
929 void set_preorder_visited( Node *n, int pre_order ) {
930 assert( !is_visited( n ), "already set" );
931 _preorders[n->_idx] = (pre_order<<1);
932 };
933 // Return pre-order number.
934 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
935
936 // Check for being post-visited.
937 // Should be previsited already (checked with assert(is_visited(n))).
938 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
939
940 // Mark as post visited
941 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
942
943 public:
944 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
945 // Returns true if "n" is a data node, false if it's a CFG node.
946 //
947 // Exception:
948 // control nodes that are dead because of "replace_node_and_forward_ctrl"
949 // or have otherwise modified their ctrl state by "forward_ctrl".
950 // They return "true", because they have a ctrl "forwarding" to the other ctrl node they
951 // were replaced with.
952 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
953
954 private:
955 // clear out dead code after build_loop_late
956 Node_List _deadlist;
957 Node_List _zero_trip_guard_opaque_nodes;
958 Node_List _multiversion_opaque_nodes;
959
960 // Support for faster execution of get_late_ctrl()/dom_lca()
961 // when a node has many uses and dominator depth is deep.
962 GrowableArray<jlong> _dom_lca_tags;
963 uint _dom_lca_tags_round;
964 void init_dom_lca_tags();
965
966 // Helper for debugging bad dominance relationships
967 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
968
969 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
970
971 // Inline wrapper for frequent cases:
972 // 1) only one use
973 // 2) a use is the same as the current LCA passed as 'n1'
974 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
975 assert( n->is_CFG(), "" );
976 // Fast-path null lca
977 if( lca != nullptr && lca != n ) {
978 assert( lca->is_CFG(), "" );
979 // find LCA of all uses
980 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
981 }
982 return find_non_split_ctrl(n);
983 }
984 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
985
986 // Helper function for directing control inputs away from CFG split points.
987 Node *find_non_split_ctrl( Node *ctrl ) const {
988 if (ctrl != nullptr) {
989 if (ctrl->is_MultiBranch()) {
990 ctrl = ctrl->in(0);
991 }
992 assert(ctrl->is_CFG(), "CFG");
993 }
994 return ctrl;
995 }
996
997 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
998
999 #ifdef ASSERT
1000 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
1001 #endif
1002 private:
1003 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
1004 Unique_Node_List& list);
1005 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
1006 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1007 CountedLoopNode* remaining_loop_head,
1008 uint first_node_index_in_cloned_loop_body,
1009 const Node_List& old_new);
1010 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1011 CountedLoopNode* main_loop_head,
1012 uint first_node_index_in_pre_loop_body,
1013 uint last_node_index_in_pre_loop_body,
1014 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1015 const Node_List& old_new);
1016 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1017 uint first_node_index_in_cloned_loop_body);
1018 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1019 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1020 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1021 CountedLoopNode* target_loop_head,
1022 const NodeInLoopBody& _node_in_loop_body,
1023 bool kill_old_template);
1024 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1025 const Node* old_target_loop_entry,
1026 uint node_index_before_new_assertion_predicate_nodes);
1027 void log_loop_tree();
1028
1029 public:
1030
1031 PhaseIterGVN &igvn() const { return _igvn; }
1032
1033 Arena* arena() { return &_arena; };
1034
1035 bool has_node(const Node* n) const {
1036 guarantee(n != nullptr, "No Node.");
1037 return _loop_or_ctrl[n->_idx] != nullptr;
1038 }
1039 // check if transform created new nodes that need _ctrl recorded
1040 Node *get_late_ctrl( Node *n, Node *early );
1041 Node *get_early_ctrl( Node *n );
1042 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1043 void set_early_ctrl(Node* n, bool update_body);
1044 void set_subtree_ctrl(Node* n, bool update_body);
1045 void set_ctrl( Node *n, Node *ctrl ) {
1046 assert( !has_node(n) || has_ctrl(n), "" );
1047 assert( ctrl->in(0), "cannot set dead control node" );
1048 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1049 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1050 }
1051 void set_root_as_ctrl(Node* n) {
1052 assert(!has_node(n) || has_ctrl(n), "");
1053 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1054 }
1055 // Set control and update loop membership
1056 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1057 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1058 IdealLoopTree* new_loop = get_loop(ctrl);
1059 if (old_loop != new_loop) {
1060 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1061 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1062 }
1063 set_ctrl(n, ctrl);
1064 }
1065
1066 // Retrieves the ctrl for a data node i.
1067 Node* get_ctrl(const Node* i) {
1068 assert(has_node(i) && has_ctrl(i), "must be data node with ctrl");
1069 Node* n = get_ctrl_no_update(i);
1070 // We store the found ctrl in the side-table again. In most cases,
1071 // this is a no-op, since we just read from _loop_or_ctrl. But in cases
1072 // where there was a ctrl forwarding via dead ctrl nodes, this shortens the path.
1073 // See: forward_ctrl
1074 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1075 assert(has_node(i) && has_ctrl(i), "must still be data node with ctrl");
1076 assert(n == find_non_split_ctrl(n), "must return legal ctrl");
1077 return n;
1078 }
1079
1080 bool is_dominator(Node* dominator, Node* n);
1081 bool is_strict_dominator(Node* dominator, Node* n);
1082
1083 // return get_ctrl for a data node and self(n) for a CFG node
1084 Node* ctrl_or_self(Node* n) {
1085 if (has_ctrl(n))
1086 return get_ctrl(n);
1087 else {
1088 assert (n->is_CFG(), "must be a CFG node");
1089 return n;
1090 }
1091 }
1092
1093 private:
1094 Node* get_ctrl_no_update_helper(const Node* i) const {
1095 // We expect only data nodes (which must have a ctrl set), or
1096 // dead ctrl nodes that have a ctrl "forwarding".
1097 // See: forward_ctrl.
1098 assert(has_ctrl(i), "only data nodes or ctrl nodes with ctrl forwarding expected");
1099 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1100 }
1101
1102 // Compute the ctrl of node i, jumping over ctrl forwardings.
1103 Node* get_ctrl_no_update(const Node* i) const {
1104 assert(has_ctrl(i), "only data nodes expected");
1105 Node* n = get_ctrl_no_update_helper(i);
1106 if (n->in(0) == nullptr) {
1107 // We encountered a dead CFG node.
1108 // If everything went right, this dead CFG node should have had a ctrl
1109 // forwarding installed, using "forward_ctrl". We now have to jump from
1110 // the old (dead) ctrl node to the new (live) ctrl node, in possibly
1111 // multiple ctrl forwarding steps.
1112 do {
1113 n = get_ctrl_no_update_helper(n);
1114 } while (n->in(0) == nullptr);
1115 n = find_non_split_ctrl(n);
1116 }
1117 return n;
1118 }
1119
1120 public:
1121 // Check for loop being set
1122 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1123 bool has_loop( Node *n ) const {
1124 assert(!has_node(n) || !has_ctrl(n), "");
1125 return has_node(n);
1126 }
1127 // Set loop
1128 void set_loop( Node *n, IdealLoopTree *loop ) {
1129 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1130 }
1131
1132 // Install a ctrl "forwarding" from an old (dead) control node.
1133 // This is a "lazy" update of the "get_ctrl" and "idom" mechanism:
1134 // - Install a forwarding from old_node (dead ctrl) to new_node.
1135 // - When querying "get_ctrl": jump from data node over possibly
1136 // multiple dead ctrl nodes with ctrl forwarding to eventually
1137 // reach a live ctrl node. Shorten the path to avoid chasing the
1138 // forwarding in the future.
1139 // - When querying "idom": from some node get its old idom, which
1140 // may be dead but has an idom forwarding to the new and live
1141 // idom. Shorten the path to avoid chasing the forwarding in the
1142 // future.
1143 // Note: while the "idom" information is stored in the "_idom"
1144 // side-table, the idom forwarding piggybacks on the ctrl
1145 // forwarding on "_loop_or_ctrl".
1146 // Using "forward_ctrl" allows us to only edit the entry for the old
1147 // dead node now, and we do not have to update all the nodes that had
1148 // the old_node as their "get_ctrl" or "idom". We clean up the forwarding
1149 // links when we query "get_ctrl" or "idom" for these nodes the next time.
1150 void forward_ctrl(Node* old_node, Node* new_node) {
1151 assert(!has_ctrl(old_node) && old_node->is_CFG() && old_node->in(0) == nullptr,
1152 "must be dead ctrl (CFG) node");
1153 assert(!has_ctrl(new_node) && new_node->is_CFG() && new_node->in(0) != nullptr,
1154 "must be live ctrl (CFG) node");
1155 assert(old_node != new_node, "no cycles please");
1156 // Re-use the side array slot for this node to provide the
1157 // forwarding pointer.
1158 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1159 assert(has_ctrl(old_node), "must have installed ctrl forwarding");
1160 }
1161
1162 // Replace the old ctrl node with a new ctrl node.
1163 // - Update the node inputs of all uses.
1164 // - Lazily update the ctrl and idom info of all uses, via a ctrl/idom forwarding.
1165 void replace_node_and_forward_ctrl(Node* old_node, Node* new_node) {
1166 _igvn.replace_node(old_node, new_node);
1167 forward_ctrl(old_node, new_node);
1168 }
1169
1170 private:
1171
1172 // Place 'n' in some loop nest, where 'n' is a CFG node
1173 void build_loop_tree();
1174 int build_loop_tree_impl(Node* n, int pre_order);
1175 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1176 // loop tree, not the root.
1177 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1178
1179 #ifdef ASSERT
1180 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1181 void verify_regions_in_irreducible_loops();
1182 bool is_in_irreducible_loop(RegionNode* region);
1183 #endif
1184
1185 // Place Data nodes in some loop nest
1186 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1187 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1188 void build_loop_late_post_work(Node* n, bool pinned);
1189 void build_loop_late_post(Node* n);
1190 void verify_strip_mined_scheduling(Node *n, Node* least);
1191
1192 // Array of immediate dominance info for each CFG node indexed by node idx
1193 private:
1194 uint _idom_size;
1195 Node **_idom; // Array of immediate dominators
1196 uint *_dom_depth; // Used for fast LCA test
1197 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1198 LoopOptsMode _mode;
1199
1200 // build the loop tree and perform any requested optimizations
1201 void build_and_optimize();
1202
1203 // Dominators for the sea of nodes
1204 void Dominators();
1205
1206 // Compute the Ideal Node to Loop mapping
1207 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1208 PhaseTransform(Ideal_Loop),
1209 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1210 _loop_or_ctrl(&_arena),
1211 _igvn(igvn),
1212 _verify_me(nullptr),
1213 _verify_only(false),
1214 _mode(mode),
1215 _nodes_required(UINT_MAX) {
1216 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1217 build_and_optimize();
1218 }
1219
1220 #ifndef PRODUCT
1221 // Verify that verify_me made the same decisions as a fresh run
1222 // or only verify that the graph is valid if verify_me is null.
1223 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1224 PhaseTransform(Ideal_Loop),
1225 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1226 _loop_or_ctrl(&_arena),
1227 _igvn(igvn),
1228 _verify_me(verify_me),
1229 _verify_only(verify_me == nullptr),
1230 _mode(LoopOptsVerify),
1231 _nodes_required(UINT_MAX) {
1232 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1233 build_and_optimize();
1234 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1235 }
1236 #endif
1237
1238 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1239
1240 Node* idom_no_update(Node* d) const {
1241 return idom_no_update(d->_idx);
1242 }
1243
1244 Node* idom_no_update(uint node_idx) const {
1245 assert(node_idx < _idom_size, "oob");
1246 Node* n = _idom[node_idx];
1247 assert(n != nullptr,"Bad immediate dominator info.");
1248 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1249 // We encountered a dead CFG node.
1250 // If everything went right, this dead CFG node should have had an idom
1251 // forwarding installed, using "forward_ctrl". We now have to jump from
1252 // the old (dead) idom node to the new (live) idom node, in possibly
1253 // multiple idom forwarding steps.
1254 // Note that we piggyback on "_loop_or_ctrl" to do the forwarding,
1255 // since we forward both "get_ctrl" and "idom" from the dead to the
1256 // new live ctrl/idom nodes.
1257 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1258 assert(n != nullptr,"Bad immediate dominator info.");
1259 }
1260 return n;
1261 }
1262
1263 public:
1264 Node* idom(Node* n) const {
1265 return idom(n->_idx);
1266 }
1267
1268 Node* idom(uint node_idx) const {
1269 Node* n = idom_no_update(node_idx);
1270 // We store the found idom in the side-table again. In most cases,
1271 // this is a no-op, since we just read from _idom. But in cases where
1272 // there was an idom forwarding via dead idom nodes, this shortens the path.
1273 // See: forward_ctrl
1274 _idom[node_idx] = n;
1275 return n;
1276 }
1277
1278 uint dom_depth(Node* d) const {
1279 guarantee(d != nullptr, "Null dominator info.");
1280 guarantee(d->_idx < _idom_size, "");
1281 return _dom_depth[d->_idx];
1282 }
1283 void set_idom(Node* d, Node* n, uint dom_depth);
1284 // Locally compute IDOM using dom_lca call
1285 Node *compute_idom( Node *region ) const;
1286 // Recompute dom_depth
1287 void recompute_dom_depth();
1288
1289 // Is safept not required by an outer loop?
1290 bool is_deleteable_safept(Node* sfpt) const;
1291
1292 // Replace parallel induction variable (parallel to trip counter)
1293 void replace_parallel_iv(IdealLoopTree *loop);
1294
1295 Node *dom_lca( Node *n1, Node *n2 ) const {
1296 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1297 }
1298 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1299
1300 Node* dominated_node(Node* c1, Node* c2) {
1301 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1302 return is_dominator(c1, c2) ? c2 : c1;
1303 }
1304
1305 // Return control node that's dominated by the 2 others
1306 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1307 return dominated_node(c1, dominated_node(c2, c3));
1308 }
1309
1310 // Build and verify the loop tree without modifying the graph. This
1311 // is useful to verify that all inputs properly dominate their uses.
1312 static void verify(PhaseIterGVN& igvn) {
1313 #ifdef ASSERT
1314 ResourceMark rm;
1315 Compile::TracePhase tp(_t_idealLoopVerify);
1316 PhaseIdealLoop v(igvn);
1317 #endif
1318 }
1319
1320 // Recommended way to use PhaseIdealLoop.
1321 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1322 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1323 ResourceMark rm;
1324 PhaseIdealLoop v(igvn, mode);
1325
1326 Compile* C = Compile::current();
1327 if (!C->failing()) {
1328 // Cleanup any modified bits
1329 igvn.optimize();
1330 if (C->failing()) { return; }
1331 v.log_loop_tree();
1332 }
1333 }
1334
1335 // True if the method has at least 1 irreducible loop
1336 bool _has_irreducible_loops;
1337
1338 // Per-Node transform
1339 virtual Node* transform(Node* n) { return nullptr; }
1340
1341 Node* loop_exit_control(const IdealLoopTree* loop) const;
1342
1343 class LoopExitTest {
1344 bool _is_valid;
1345
1346 const Node* _back_control;
1347 const IdealLoopTree* _loop;
1348 PhaseIdealLoop* _phase;
1349
1350 Node* _cmp;
1351 Node* _incr;
1352 Node* _limit;
1353 BoolTest::mask _mask;
1354 float _cl_prob;
1355
1356 public:
1357 LoopExitTest(const Node* back_control, const IdealLoopTree* loop, PhaseIdealLoop* phase) :
1358 _is_valid(false),
1359 _back_control(back_control),
1360 _loop(loop),
1361 _phase(phase),
1362 _cmp(nullptr),
1363 _incr(nullptr),
1364 _limit(nullptr),
1365 _mask(BoolTest::illegal),
1366 _cl_prob(0.0f) {}
1367
1368 void build();
1369 void canonicalize_mask(jlong stride_con);
1370
1371 bool is_valid_with_bt(BasicType bt) const {
1372 return _is_valid && _cmp != nullptr && _cmp->Opcode() == Op_Cmp(bt);
1373 }
1374
1375 bool should_include_limit() const { return _mask == BoolTest::le || _mask == BoolTest::ge; }
1376
1377 CmpNode* cmp() const { return _cmp->as_Cmp(); }
1378 Node* incr() const { return _incr; }
1379 Node* limit() const { return _limit; }
1380 BoolTest::mask mask() const { return _mask; }
1381 float cl_prob() const { return _cl_prob; }
1382 };
1383
1384 class LoopIVIncr {
1385 bool _is_valid;
1386
1387 const Node* _head;
1388 const IdealLoopTree* _loop;
1389
1390 Node* _incr;
1391 Node* _phi_incr;
1392
1393 public:
1394 LoopIVIncr(const Node* head, const IdealLoopTree* loop) :
1395 _is_valid(false),
1396 _head(head),
1397 _loop(loop),
1398 _incr(nullptr),
1399 _phi_incr(nullptr) {}
1400
1401 void build(Node* old_incr);
1402
1403 bool is_valid() const { return _is_valid; }
1404 bool is_valid_with_bt(const BasicType bt) const {
1405 return _is_valid && _incr->Opcode() == Op_Add(bt);
1406 }
1407
1408 Node* incr() const { return _incr; }
1409 Node* phi_incr() const { return _phi_incr; }
1410 };
1411
1412 class LoopIVStride {
1413 bool _is_valid;
1414
1415 BasicType _iv_bt;
1416 Node* _stride_node;
1417 Node* _xphi;
1418
1419 public:
1420 LoopIVStride(BasicType iv_bt) :
1421 _is_valid(false),
1422 _iv_bt(iv_bt),
1423 _stride_node(nullptr),
1424 _xphi(nullptr) {}
1425
1426 void build(const Node* incr);
1427
1428 bool is_valid() const { return _is_valid && _stride_node != nullptr; }
1429 Node* stride_node() const { return _stride_node; }
1430 Node* xphi() const { return _xphi; }
1431
1432 jlong compute_non_zero_stride_con(BoolTest::mask mask, BasicType iv_bt) const;
1433 };
1434
1435 static PhiNode* loop_iv_phi(const Node* xphi, const Node* phi_incr, const Node* head);
1436
1437 bool try_convert_to_counted_loop(Node* head, IdealLoopTree*& loop, BasicType iv_bt);
1438
1439 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1440 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1441
1442 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1443 SafePointNode* find_safepoint(Node* back_control, const Node* head, const IdealLoopTree* loop);
1444
1445 void add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt);
1446
1447 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1448 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1449 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1450 Node*& entry_control, Node*& iffalse);
1451
1452 Node* exact_limit( IdealLoopTree *loop );
1453
1454 // Return a post-walked LoopNode
1455 IdealLoopTree* get_loop(const Node* n) const {
1456 // Dead nodes have no loop, so return the top level loop instead
1457 if (!has_node(n)) return _ltree_root;
1458 assert(!has_ctrl(n), "");
1459 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1460 }
1461
1462 IdealLoopTree* ltree_root() const { return _ltree_root; }
1463
1464 // Is 'n' a (nested) member of 'loop'?
1465 bool is_member(const IdealLoopTree* loop, const Node* n) const {
1466 return loop->is_member(get_loop(n));
1467 }
1468
1469 // is the control for 'n' a (nested) member of 'loop'?
1470 bool ctrl_is_member(const IdealLoopTree* loop, const Node* n) {
1471 return is_member(loop, get_ctrl(n));
1472 }
1473
1474 // This is the basic building block of the loop optimizations. It clones an
1475 // entire loop body. It makes an old_new loop body mapping; with this
1476 // mapping you can find the new-loop equivalent to an old-loop node. All
1477 // new-loop nodes are exactly equal to their old-loop counterparts, all
1478 // edges are the same. All exits from the old-loop now have a RegionNode
1479 // that merges the equivalent new-loop path. This is true even for the
1480 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1481 // now come from (one or more) Phis that merge their new-loop equivalents.
1482 // Parameter side_by_side_idom:
1483 // When side_by_size_idom is null, the dominator tree is constructed for
1484 // the clone loop to dominate the original. Used in construction of
1485 // pre-main-post loop sequence.
1486 // When nonnull, the clone and original are side-by-side, both are
1487 // dominated by the passed in side_by_side_idom node. Used in
1488 // construction of unswitched loops.
1489 enum CloneLoopMode {
1490 IgnoreStripMined = 0, // Only clone inner strip mined loop
1491 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1492 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1493 // result control flow branches
1494 // either to inner clone or outer
1495 // strip mined loop.
1496 };
1497 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1498 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1499 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1500 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1501 Node_List*& split_if_set, Node_List*& split_bool_set,
1502 Node_List*& split_cex_set, Node_List& worklist,
1503 uint new_counter, CloneLoopMode mode);
1504 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1505 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1506 Node_List& extra_data_nodes);
1507
1508 // If we got the effect of peeling, either by actually peeling or by
1509 // making a pre-loop which must execute at least once, we can remove
1510 // all loop-invariant dominated tests in the main body.
1511 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1512
1513 // Generate code to do a loop peel for the given loop (and body).
1514 // old_new is a temp array.
1515 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1516
1517 // Add pre and post loops around the given loop. These loops are used
1518 // during RCE, unrolling and aligning loops.
1519 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1520
1521 // Find the last store in the body of an OuterStripMinedLoop when following memory uses
1522 Node *find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop);
1523
1524 // Add post loop after the given loop.
1525 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1526 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1527 Node* incr, Node* limit, CountedLoopNode*& post_head);
1528
1529 // Add a vector post loop between a vector main loop and the current post loop
1530 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1531 // If Node n lives in the back_ctrl block, we clone a private version of n
1532 // in preheader_ctrl block and return that, otherwise return n.
1533 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1534
1535 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1536 // unroll to do double iterations. The next round of major loop transforms
1537 // will repeat till the doubled loop body does all remaining iterations in 1
1538 // pass.
1539 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1540
1541 // Unroll the loop body one step - make each trip do 2 iterations.
1542 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1543
1544 // Return true if exp is a constant times an induction var
1545 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1546
1547 bool is_iv(Node* exp, Node* iv, BasicType bt);
1548
1549 // Return true if exp is a scaled induction var plus (or minus) constant
1550 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);
1551 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1552 jlong long_scale;
1553 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1554 int int_scale = checked_cast<int>(long_scale);
1555 if (p_scale != nullptr) {
1556 *p_scale = int_scale;
1557 }
1558 return true;
1559 }
1560 return false;
1561 }
1562 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1563 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1564 BasicType bt,
1565 jlong* p_scale, Node** p_offset,
1566 bool* p_short_scale, int depth);
1567
1568 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1569 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1570 Deoptimization::DeoptReason reason, int opcode,
1571 bool rewire_uncommon_proj_phi_inputs = false);
1572
1573 private:
1574 // Helper functions for create_new_if_for_predicate()
1575 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1576 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1577 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1578 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1579 const OrigToNewHashtable& orig_to_clone);
1580
1581 public:
1582 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1583
1584 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1585 void replace_control(Node* node, Node* new_control) {
1586 _igvn.replace_input_of(node, 0, new_control);
1587 set_idom(node, new_control, dom_depth(new_control));
1588 }
1589
1590 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1591 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1592 set_idom(loop_head, new_entry, dom_depth(new_entry));
1593 }
1594
1595 // Construct a range check for a predicate if
1596 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1597 jint stride, Node* range, bool upper, bool& overflow);
1598
1599 // Implementation of the loop predication to promote checks outside the loop
1600 bool loop_predication_impl(IdealLoopTree *loop);
1601
1602 private:
1603 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1604 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1605 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1606 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1607 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1608 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1609 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1610 Node_List& if_proj_list);
1611 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1612 IfProjNode* new_control, int scale, Node* offset, Node* range);
1613 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1614
1615 // Helper function to collect predicate for eliminating the useless ones
1616 void eliminate_useless_predicates() const;
1617
1618 void eliminate_useless_zero_trip_guard();
1619 void eliminate_useless_multiversion_if();
1620
1621 public:
1622 // Change the control input of expensive nodes to allow commoning by
1623 // IGVN when it is guaranteed to not result in a more frequent
1624 // execution of the expensive node. Return true if progress.
1625 bool process_expensive_nodes();
1626
1627 // Check whether node has become unreachable
1628 bool is_node_unreachable(Node *n) const {
1629 return !has_node(n) || n->is_unreachable(_igvn);
1630 }
1631
1632 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1633 void do_range_check(IdealLoopTree* loop);
1634
1635 // Clone loop with an invariant test (that does not exit) and
1636 // insert a clone of the test that selects which version to
1637 // execute.
1638 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1639
1640 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1641
1642 private:
1643 static bool has_control_dependencies_from_predicates(LoopNode* head);
1644 static void revert_to_normal_loop(const LoopNode* loop_head);
1645
1646 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1647 const UnswitchedLoopSelector& unswitched_loop_selector);
1648 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1649 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1650 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1651 #ifndef PRODUCT
1652 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1653 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1654 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1655 const LoopNode* original_head, const LoopNode* new_head);
1656 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1657 const LoopNode* original_head, const LoopNode* new_head);
1658 #endif
1659
1660 public:
1661
1662 // Range Check Elimination uses this function!
1663 // Constrain the main loop iterations so the affine function:
1664 // low_limit <= scale_con * I + offset < upper_limit
1665 // always holds true. That is, either increase the number of iterations in
1666 // the pre-loop or the post-loop until the condition holds true in the main
1667 // loop. Scale_con, offset and limit are all loop invariant.
1668 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);
1669 // Helper function for add_constraint().
1670 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1671
1672 // Partially peel loop up through last_peel node.
1673 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1674 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1675
1676 // AutoVectorize the loop: replace scalar ops with vector ops.
1677 enum AutoVectorizeStatus {
1678 Impossible, // This loop has the wrong shape to even try vectorization.
1679 Success, // We just successfully vectorized the loop.
1680 TriedAndFailed, // We tried to vectorize, but failed.
1681 };
1682 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1683
1684 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1685 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1686 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1687 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1688
1689 // Create a scheduled list of nodes control dependent on ctrl set.
1690 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1691 // Has a use in the vector set
1692 bool has_use_in_set( Node* n, VectorSet& vset );
1693 // Has use internal to the vector set (ie. not in a phi at the loop head)
1694 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1695 // clone "n" for uses that are outside of loop
1696 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1697 // clone "n" for special uses that are in the not_peeled region
1698 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1699 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1700 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1701 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1702 #ifdef ASSERT
1703 // Validate the loop partition sets: peel and not_peel
1704 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1705 // Ensure that uses outside of loop are of the right form
1706 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1707 uint orig_exit_idx, uint clone_exit_idx);
1708 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1709 #endif
1710
1711 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1712 int stride_of_possible_iv( Node* iff );
1713 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1714 // Return the (unique) control output node that's in the loop (if it exists.)
1715 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1716 // Insert a signed compare loop exit cloned from an unsigned compare.
1717 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1718 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1719 // Utility to register node "n" with PhaseIdealLoop
1720 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1721 // Utility to create an if-projection
1722 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1723 // Force the iff control output to be the live_proj
1724 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1725 // Insert a region before an if projection
1726 RegionNode* insert_region_before_proj(ProjNode* proj);
1727 // Insert a new if before an if projection
1728 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1729
1730 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1731 // "Nearly" because all Nodes have been cloned from the original in the loop,
1732 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1733 // through the Phi recursively, and return a Bool.
1734 Node* clone_iff(PhiNode* phi);
1735 CmpNode* clone_bool(PhiNode* phi);
1736
1737
1738 // Rework addressing expressions to get the most loop-invariant stuff
1739 // moved out. We'd like to do all associative operators, but it's especially
1740 // important (common) to do address expressions.
1741 Node* remix_address_expressions(Node* n);
1742 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1743
1744 // Convert add to muladd to generate MuladdS2I under certain criteria
1745 Node * convert_add_to_muladd(Node * n);
1746
1747 // Attempt to use a conditional move instead of a phi/branch
1748 Node *conditional_move( Node *n );
1749
1750 // Check for aggressive application of 'split-if' optimization,
1751 // using basic block level info.
1752 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1753 Node *split_if_with_blocks_pre ( Node *n );
1754 void split_if_with_blocks_post( Node *n );
1755 Node *has_local_phi_input( Node *n );
1756 // Mark an IfNode as being dominated by a prior test,
1757 // without actually altering the CFG (and hence IDOM info).
1758 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool prev_dom_not_imply_this = false);
1759 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool dominator_not_imply_source);
1760
1761 // Split Node 'n' through merge point
1762 RegionNode* split_thru_region(Node* n, RegionNode* region);
1763 // Split Node 'n' through merge point if there is enough win.
1764 Node *split_thru_phi( Node *n, Node *region, int policy );
1765 // Found an If getting its condition-code input from a Phi in the
1766 // same block. Split thru the Region.
1767 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1768
1769 private:
1770 // Class to keep track of wins in split_thru_phi.
1771 class SplitThruPhiWins {
1772 private:
1773 // Region containing the phi we are splitting through.
1774 const Node* _region;
1775
1776 // Sum of all wins regardless of where they happen. This applies to Loops phis as well as non-loop phis.
1777 int _total_wins;
1778
1779 // For Loops, wins have different impact depending on if they happen on loop entry or on the backedge.
1780 // Number of wins on a loop entry edge if the split is through a loop head,
1781 // otherwise 0. Entry edge wins only pay dividends once on loop entry.
1782 int _loop_entry_wins;
1783 // Number of wins on a loop back-edge, which pay dividends on every iteration.
1784 int _loop_back_wins;
1785
1786 public:
1787 SplitThruPhiWins(const Node* region) :
1788 _region(region),
1789 _total_wins(0),
1790 _loop_entry_wins(0),
1791 _loop_back_wins(0) {};
1792
1793 void reset() {_total_wins = 0; _loop_entry_wins = 0; _loop_back_wins = 0;}
1794 void add_win(int ctrl_index) {
1795 if (_region->is_Loop() && ctrl_index == LoopNode::EntryControl) {
1796 _loop_entry_wins++;
1797 } else if (_region->is_Loop() && ctrl_index == LoopNode::LoopBackControl) {
1798 _loop_back_wins++;
1799 }
1800 _total_wins++;
1801 }
1802 // Is this split profitable with respect to the policy?
1803 bool profitable(int policy) const {
1804 assert(_region->is_Loop() || (_loop_entry_wins == 0 && _loop_back_wins == 0), "wins on loop edges without a loop");
1805 assert(!_region->is_Loop() || _total_wins == _loop_entry_wins + _loop_back_wins, "missed some win");
1806 // In general this means that the split has to have more wins than specified
1807 // in the policy. However, for loops we need to take into account where the
1808 // wins happen. We need to be careful when splitting, because splitting nodes
1809 // related to the iv through the phi can sufficiently rearrange the loop
1810 // structure to prevent RCE and thus vectorization. Thus, we only deem splitting
1811 // profitable if the win of a split is not on the entry edge, as such wins
1812 // only pay off once and have a high chance of messing up the loop structure.
1813 return (_loop_entry_wins == 0 && _total_wins > policy) ||
1814 // If there are wins on the entry edge but the backadge also has sufficient wins,
1815 // there is sufficient profitability to spilt regardless of the risk of messing
1816 // up the loop structure.
1817 _loop_back_wins > policy ||
1818 // If the policy is less than 0, a split is always profitable, i.e. we always
1819 // split. This is needed when we split a node and then must also split a
1820 // dependant node, i.e. spliting a Bool node after splitting a Cmp node.
1821 policy < 0;
1822 }
1823 };
1824
1825
1826 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1827
1828 public:
1829
1830 // Conversion of fill/copy patterns into intrinsic versions
1831 bool do_intrinsify_fill();
1832 bool intrinsify_fill(IdealLoopTree* lpt);
1833 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1834 Node*& shift, Node*& offset);
1835
1836 private:
1837 // Helper functions
1838 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1839 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1840 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 );
1841 bool split_up( Node *n, Node *blk1, Node *blk2 );
1842
1843 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1844 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1845 void try_move_store_after_loop(Node* n);
1846 bool identical_backtoback_ifs(Node *n);
1847 bool can_split_if(Node *n_ctrl);
1848 bool cannot_split_division(const Node* n, const Node* region) const;
1849 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1850 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1851
1852 // Determine if a method is too big for a/another round of split-if, based on
1853 // a magic (approximate) ratio derived from the equally magic constant 35000,
1854 // previously used for this purpose (but without relating to the node limit).
1855 bool must_throttle_split_if() {
1856 uint threshold = C->max_node_limit() * 2 / 5;
1857 return C->live_nodes() > threshold;
1858 }
1859
1860 // A simplistic node request tracking mechanism, where
1861 // = UINT_MAX Request not valid or made final.
1862 // < UINT_MAX Nodes currently requested (estimate).
1863 uint _nodes_required;
1864
1865 enum { REQUIRE_MIN = 70 };
1866
1867 uint nodes_required() const { return _nodes_required; }
1868
1869 // Given the _currently_ available number of nodes, check whether there is
1870 // "room" for an additional request or not, considering the already required
1871 // number of nodes. Return TRUE if the new request is exceeding the node
1872 // budget limit, otherwise return FALSE. Note that this interpretation will
1873 // act pessimistic on additional requests when new nodes have already been
1874 // generated since the 'begin'. This behaviour fits with the intention that
1875 // node estimates/requests should be made upfront.
1876 bool exceeding_node_budget(uint required = 0) {
1877 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1878 uint available = C->max_node_limit() - C->live_nodes();
1879 return available < required + _nodes_required + REQUIRE_MIN;
1880 }
1881
1882 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1883 precond(require > 0);
1884 _nodes_required += MAX2(require, minreq);
1885 return _nodes_required;
1886 }
1887
1888 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1889 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1890 }
1891
1892 uint require_nodes_begin() {
1893 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1894 _nodes_required = 0;
1895 return C->live_nodes();
1896 }
1897
1898 // When a node request is final, optionally check that the requested number
1899 // of nodes was reasonably correct with respect to the number of new nodes
1900 // introduced since the last 'begin'. Always check that we have not exceeded
1901 // the maximum node limit.
1902 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1903 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1904
1905 #ifdef ASSERT
1906 if (check_estimate) {
1907 // Check that the node budget request was not off by too much (x2).
1908 // Should this be the case we _surely_ need to improve the estimates
1909 // used in our budget calculations.
1910 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1911 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1912 C->live_nodes() - live_at_begin, _nodes_required);
1913 }
1914 }
1915 #endif
1916 // Assert that we have stayed within the node budget limit.
1917 assert(C->live_nodes() < C->max_node_limit(),
1918 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1919 C->live_nodes() - live_at_begin, live_at_begin,
1920 C->max_node_limit(), _nodes_required);
1921
1922 _nodes_required = UINT_MAX;
1923 }
1924
1925 private:
1926
1927 bool _created_loop_node;
1928 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1929 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1930
1931 public:
1932 void set_created_loop_node() { _created_loop_node = true; }
1933 bool created_loop_node() { return _created_loop_node; }
1934 void register_new_node(Node* n, Node* blk);
1935 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1936 register_new_node(new_node, get_ctrl(ctrl_of));
1937 }
1938
1939 Node* clone_and_register(Node* n, Node* ctrl) {
1940 n = n->clone();
1941 register_new_node(n, ctrl);
1942 return n;
1943 }
1944
1945 #ifdef ASSERT
1946 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1947 #endif
1948
1949 #ifndef PRODUCT
1950 void dump() const;
1951 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1952 void dump_idom(Node* n, uint count) const;
1953 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1954 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1955 IdealLoopTree* get_loop_idx(Node* n) const {
1956 // Dead nodes have no loop, so return the top level loop instead
1957 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1958 }
1959 // Print some stats
1960 static void print_statistics();
1961 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1962 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1963 static volatile int _long_loop_candidates;
1964 static volatile int _long_loop_nests;
1965 #endif
1966
1967 #ifdef ASSERT
1968 void verify() const;
1969 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1970 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1971 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1972 #endif
1973
1974 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1975
1976 void check_counted_loop_shape(IdealLoopTree* loop, Node* head, BasicType bt) NOT_DEBUG_RETURN;
1977
1978 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1979
1980
1981 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1982 Node_List &range_checks);
1983
1984 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1985 Node* inner_iters_actual_int, Node* inner_phi,
1986 Node* iv_add, LoopNode* inner_head);
1987
1988 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1989
1990 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1991
1992 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1993
1994 bool would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl);
1995
1996 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1997
1998 void try_sink_out_of_loop(Node* n);
1999
2000 Node* clamp(Node* R, Node* L, Node* H);
2001
2002 bool safe_for_if_replacement(const Node* dom) const;
2003
2004 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
2005
2006 bool try_merge_identical_ifs(Node* n);
2007
2008 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
2009
2010 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2011 IdealLoopTree* parent, bool partial);
2012
2013 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2014 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
2015
2016 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2017 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
2018 Node_List*& split_bool_set, Node_List*& split_cex_set);
2019
2020 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
2021
2022 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
2023
2024 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
2025 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
2026 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
2027 void clone_template_assertion_expression_down(Node* node);
2028
2029 Node* similar_subtype_check(const Node* x, Node* r_in);
2030
2031 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
2032
2033 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
2034
2035 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
2036
2037 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
2038
2039 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
2040
2041 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
2042
2043 ConINode* intcon(jint i);
2044
2045 ConLNode* longcon(jlong i);
2046
2047 ConNode* makecon(const Type* t);
2048
2049 ConNode* integercon(jlong l, BasicType bt);
2050
2051 ConNode* zerocon(BasicType bt);
2052 };
2053
2054 class CountedLoopConverter {
2055 friend class PhaseIdealLoop;
2056
2057 // Match increment with optional truncation
2058 class TruncatedIncrement {
2059 bool _is_valid;
2060
2061 BasicType _bt;
2062
2063 Node* _incr;
2064 Node* _outer_trunc;
2065 Node* _inner_trunc;
2066 const TypeInteger* _trunc_type;
2067
2068 public:
2069 TruncatedIncrement(BasicType bt) :
2070 _is_valid(false),
2071 _bt(bt),
2072 _incr(nullptr),
2073 _outer_trunc(nullptr),
2074 _inner_trunc(nullptr),
2075 _trunc_type(nullptr) {}
2076
2077 void build(Node* expr);
2078
2079 bool is_valid() const { return _is_valid; }
2080 Node* incr() const { return _incr; }
2081
2082 // Optional truncation for: CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2083 Node* outer_trunc() const { return _outer_trunc; } // the outermost truncating node (either the & or the final >>)
2084 Node* inner_trunc() const { return _inner_trunc; } // the inner truncating node, if applicable (the << in a <</>> pair)
2085 const TypeInteger* trunc_type() const { return _trunc_type; }
2086 };
2087
2088 class LoopStructure {
2089 bool _is_valid;
2090
2091 const Node* _head;
2092 const IdealLoopTree* _loop;
2093 PhaseIdealLoop* _phase;
2094 BasicType _iv_bt;
2095
2096 Node* _back_control;
2097 PhaseIdealLoop::LoopExitTest _exit_test;
2098 PhaseIdealLoop::LoopIVIncr _iv_incr;
2099 TruncatedIncrement _truncated_increment;
2100 PhaseIdealLoop::LoopIVStride _stride;
2101 PhiNode* _phi;
2102 SafePointNode* _safepoint;
2103
2104 public:
2105 LoopStructure(const Node* head, const IdealLoopTree* loop, PhaseIdealLoop* phase, const BasicType iv_bt) :
2106 _is_valid(false),
2107 _head(head),
2108 _loop(loop),
2109 _phase(phase),
2110 _iv_bt(iv_bt),
2111 _back_control(_phase->loop_exit_control(_loop)),
2112 _exit_test(_back_control, _loop, _phase),
2113 _iv_incr(_head, _loop),
2114 _truncated_increment(_iv_bt),
2115 _stride(PhaseIdealLoop::LoopIVStride(_iv_bt)),
2116 _phi(nullptr),
2117 _safepoint(nullptr) {}
2118
2119 void build();
2120
2121 jlong final_limit_correction() const; // compute adjusted loop limit correction
2122 bool is_infinite_loop() const;
2123
2124 bool is_valid() const { return _is_valid; }
2125
2126 Node* back_control() const { return _back_control; }
2127 PhaseIdealLoop::LoopExitTest& exit_test() { return _exit_test; }
2128 PhaseIdealLoop::LoopIVIncr& iv_incr() { return _iv_incr; }
2129 TruncatedIncrement& truncated_increment() { return _truncated_increment; }
2130 PhaseIdealLoop::LoopIVStride& stride() { return _stride; }
2131 PhiNode* phi() const { return _phi; }
2132 SafePointNode* sfpt() const { return _safepoint; }
2133 jlong stride_con() const { return _stride.compute_non_zero_stride_con(_exit_test.mask(), _iv_bt); }
2134 Node* limit() const { return _exit_test.limit(); }
2135 };
2136
2137 PhaseIdealLoop* const _phase;
2138 Node* const _head;
2139 IdealLoopTree* const _loop;
2140 const BasicType _iv_bt;
2141
2142 LoopStructure _structure;
2143 bool _should_insert_stride_overflow_limit_check = false;
2144 bool _should_insert_init_trip_limit_check = false;
2145
2146 DEBUG_ONLY(bool _checked_for_counted_loop = false;)
2147
2148 // stats for PhaseIdealLoop::print_statistics()
2149 static volatile int _long_loop_counted_loops;
2150
2151 // Return a type based on condition control flow
2152 const TypeInt* filtered_type(Node* n, Node* n_ctrl);
2153 const TypeInt* filtered_type(Node* n) { return filtered_type(n, nullptr); }
2154 // Helpers for filtered type
2155 const TypeInt* filtered_type_from_dominators(Node* val, Node* val_ctrl);
2156
2157 void insert_loop_limit_check_predicate(const ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* bol) const;
2158 void insert_stride_overflow_limit_check() const;
2159 void insert_init_trip_limit_check() const;
2160 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
2161 Node* loop_entry) const;
2162
2163 bool is_iv_overflowing(const TypeInteger* init_t, jlong stride_con, Node* phi_increment, BoolTest::mask mask) const;
2164 bool has_truncation_wrap(const TruncatedIncrement& truncation, Node* phi, jlong stride_con);
2165 SafePointNode* find_safepoint(Node* iftrue);
2166 bool is_safepoint_invalid(SafePointNode* sfpt) const;
2167
2168 public:
2169 CountedLoopConverter(PhaseIdealLoop* phase, Node* head, IdealLoopTree* loop, const BasicType iv_bt)
2170 : _phase(phase),
2171 _head(head),
2172 _loop(loop),
2173 _iv_bt(iv_bt),
2174 _structure(LoopStructure(_head, _loop, _phase, _iv_bt)) {
2175 assert(phase != nullptr, "must be"); // Fail early if mandatory parameters are null.
2176 assert(head != nullptr, "must be");
2177 assert(loop != nullptr, "must be");
2178 assert(iv_bt == T_INT || iv_bt == T_LONG, "either int or long loops");
2179 }
2180
2181 bool is_counted_loop();
2182 IdealLoopTree* convert();
2183
2184 DEBUG_ONLY(bool should_stress_long_counted_loop();)
2185 DEBUG_ONLY(bool stress_long_counted_loop();)
2186
2187 enum StrideOverflowState {
2188 Overflow = -1,
2189 NoOverflow = 0,
2190 RequireLimitCheck = 1
2191 };
2192 static StrideOverflowState check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt);
2193 };
2194
2195 class AutoNodeBudget : public StackObj
2196 {
2197 public:
2198 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
2199
2200 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
2201 : _phase(phase),
2202 _check_at_final(chk == BUDGET_CHECK),
2203 _nodes_at_begin(0)
2204 {
2205 precond(_phase != nullptr);
2206
2207 _nodes_at_begin = _phase->require_nodes_begin();
2208 }
2209
2210 ~AutoNodeBudget() {
2211 #ifndef PRODUCT
2212 if (TraceLoopOpts) {
2213 uint request = _phase->nodes_required();
2214 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
2215
2216 if (request < delta) {
2217 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
2218 } else {
2219 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
2220 // Identify the worst estimates as "poor" ones.
2221 if (request > REQUIRE_MIN && delta > 0) {
2222 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
2223 (delta <= REQUIRE_MIN && request > 10 * delta)) {
2224 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
2225 }
2226 }
2227 }
2228 }
2229 #endif // PRODUCT
2230 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
2231 }
2232
2233 private:
2234 PhaseIdealLoop* _phase;
2235 bool _check_at_final;
2236 uint _nodes_at_begin;
2237 };
2238
2239 inline Node* IdealLoopTree::tail() {
2240 // Handle lazy update of _tail field.
2241 if (_tail->in(0) == nullptr) {
2242 _tail = _phase->get_ctrl(_tail);
2243 }
2244 return _tail;
2245 }
2246
2247 inline Node* IdealLoopTree::head() {
2248 // Handle lazy update of _head field.
2249 if (_head->in(0) == nullptr) {
2250 _head = _phase->get_ctrl(_head);
2251 }
2252 return _head;
2253 }
2254
2255 // Iterate over the loop tree using a preorder, left-to-right traversal.
2256 //
2257 // Example that visits all counted loops from within PhaseIdealLoop
2258 //
2259 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2260 // IdealLoopTree* lpt = iter.current();
2261 // if (!lpt->is_counted()) continue;
2262 // ...
2263 class LoopTreeIterator : public StackObj {
2264 private:
2265 IdealLoopTree* _root;
2266 IdealLoopTree* _curnt;
2267
2268 public:
2269 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
2270
2271 bool done() { return _curnt == nullptr; } // Finished iterating?
2272
2273 void next(); // Advance to next loop tree
2274
2275 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
2276 };
2277
2278 // Compute probability of reaching some CFG node from a fixed
2279 // dominating CFG node
2280 class PathFrequency {
2281 private:
2282 Node* _dom; // frequencies are computed relative to this node
2283 Node_Stack _stack;
2284 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
2285 GrowableArray<float> _freqs; // cache frequencies
2286 PhaseIdealLoop* _phase;
2287
2288 float check_and_truncate_frequency(float f) {
2289 assert(f >= 0, "Incorrect frequency");
2290 // We do not perform an exact (f <= 1) check
2291 // this would be error prone with rounding of floats.
2292 // Performing a check like (f <= 1+eps) would be of benefit,
2293 // however, it is not evident how to determine such an eps,
2294 // given that an arbitrary number of add/mul operations
2295 // are performed on these frequencies.
2296 return (f > 1) ? 1 : f;
2297 }
2298
2299 public:
2300 PathFrequency(Node* dom, PhaseIdealLoop* phase)
2301 : _dom(dom), _stack(0), _phase(phase) {
2302 }
2303
2304 float to(Node* n);
2305 };
2306
2307 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
2308 // 1. Clone the data nodes
2309 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
2310 class DataNodeGraph : public StackObj {
2311 PhaseIdealLoop* const _phase;
2312 const Unique_Node_List& _data_nodes;
2313 OrigToNewHashtable _orig_to_new;
2314
2315 public:
2316 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
2317 : _phase(phase),
2318 _data_nodes(data_nodes),
2319 // Use 107 as best guess which is the first resize value in ResizeableHashTable::large_table_sizes.
2320 _orig_to_new(107, MaxNodeLimit)
2321 {
2322 #ifdef ASSERT
2323 for (uint i = 0; i < data_nodes.size(); i++) {
2324 assert(!data_nodes[i]->is_CFG(), "only data nodes");
2325 }
2326 #endif
2327 }
2328 NONCOPYABLE(DataNodeGraph);
2329
2330 private:
2331 void clone(Node* node, Node* new_ctrl);
2332 void clone_data_nodes(Node* new_ctrl);
2333 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2334 Node* new_ctrl);
2335 void rewire_clones_to_cloned_inputs();
2336 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
2337
2338 public:
2339 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
2340 // Set 'new_ctrl' as ctrl for the cloned nodes.
2341 const OrigToNewHashtable& clone(Node* new_ctrl) {
2342 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
2343 clone_data_nodes(new_ctrl);
2344 rewire_clones_to_cloned_inputs();
2345 return _orig_to_new;
2346 }
2347
2348 // Create a copy of the data nodes provided to the constructor by doing the following:
2349 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
2350 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2351 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2352 // `new_ctrl` as ctrl.
2353 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2354 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2355 Node* new_ctrl) {
2356 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2357 rewire_clones_to_cloned_inputs();
2358 return _orig_to_new;
2359 }
2360 };
2361 #endif // SHARE_OPTO_LOOPNODE_HPP