1 /*
2 * Copyright (c) 1998, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_OPTO_LOOPNODE_HPP
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class UnswitchCandidate;
47 class LoopSelector;
48 class ReachabilityFenceNode;
49 class UnswitchedLoopSelector;
50 class VectorSet;
51 class VSharedData;
52 class Invariance;
53 struct small_cache;
54
55 //
56 // I D E A L I Z E D L O O P S
57 //
58 // Idealized loops are the set of loops I perform more interesting
59 // transformations on, beyond simple hoisting.
60
61 //------------------------------LoopNode---------------------------------------
62 // Simple loop header. Fall in path on left, loop-back path on right.
63 class LoopNode : public RegionNode {
64 // Size is bigger to hold the flags. However, the flags do not change
65 // the semantics so it does not appear in the hash & cmp functions.
66 virtual uint size_of() const { return sizeof(*this); }
67 protected:
68 uint _loop_flags;
69 // Names for flag bitfields
70 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
71 MainHasNoPreLoop = 1<<2,
72 HasExactTripCount = 1<<3,
73 InnerLoop = 1<<4,
74 PartialPeelLoop = 1<<5,
75 PartialPeelFailed = 1<<6,
76 WasSlpAnalyzed = 1<<7,
77 PassedSlpAnalysis = 1<<8,
78 DoUnrollOnly = 1<<9,
79 VectorizedLoop = 1<<10,
80 HasAtomicPostLoop = 1<<11,
81 StripMined = 1<<12,
82 SubwordLoop = 1<<13,
83 ProfileTripFailed = 1<<14,
84 LoopNestInnerLoop = 1<<15,
85 LoopNestLongOuterLoop = 1<<16,
86 MultiversionFastLoop = 1<<17,
87 MultiversionSlowLoop = 2<<17,
88 MultiversionDelayedSlowLoop = 3<<17,
89 MultiversionFlagsMask = 3<<17,
90 FlatArrays = 1<<18};
91 char _unswitch_count;
92 enum { _unswitch_max=3 };
93
94 // Expected trip count from profile data
95 float _profile_trip_cnt;
96
97 public:
98 // Names for edge indices
99 enum { Self=0, EntryControl, LoopBackControl };
100
101 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
102 void set_inner_loop() { _loop_flags |= InnerLoop; }
103
104 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
105 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
106 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
107 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
108 bool is_strip_mined() const { return _loop_flags & StripMined; }
109 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
110 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
111 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
112 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
113 bool is_flat_arrays() const { return _loop_flags & FlatArrays; }
114
115 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
116 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
117 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
118 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
119 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
120 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
121 void mark_strip_mined() { _loop_flags |= StripMined; }
122 void clear_strip_mined() { _loop_flags &= ~StripMined; }
123 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
124 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
125 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
126 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
127 void mark_flat_arrays() { _loop_flags |= FlatArrays; }
128
129 int unswitch_max() { return _unswitch_max; }
130 int unswitch_count() { return _unswitch_count; }
131
132 void set_unswitch_count(int val) {
133 assert (val <= unswitch_max(), "too many unswitches");
134 _unswitch_count = val;
135 }
136
137 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
138 float profile_trip_cnt() { return _profile_trip_cnt; }
139
140 #ifndef PRODUCT
141 uint _stress_peeling_attempts = 0;
142 #endif
143
144 LoopNode(Node *entry, Node *backedge)
145 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
146 _profile_trip_cnt(COUNT_UNKNOWN) {
147 init_class_id(Class_Loop);
148 init_req(EntryControl, entry);
149 init_req(LoopBackControl, backedge);
150 }
151
152 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
153 virtual int Opcode() const;
154 bool can_be_counted_loop(PhaseValues* phase) const {
155 return req() == 3 && in(0) != nullptr &&
156 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
157 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
158 }
159 bool is_valid_counted_loop(BasicType bt) const;
160 #ifndef PRODUCT
161 virtual void dump_spec(outputStream *st) const;
162 #endif
163
164 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
165 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
166 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
167 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
168 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
169 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
170 };
171
172 //------------------------------Counted Loops----------------------------------
173 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
174 // path (and maybe some other exit paths). The trip-counter exit is always
175 // last in the loop. The trip-counter have to stride by a constant;
176 // the exit value is also loop invariant.
177
178 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
179 // CountedLoopNode has the incoming loop control and the loop-back-control
180 // which is always the IfTrue before the matching CountedLoopEndNode. The
181 // CountedLoopEndNode has an incoming control (possibly not the
182 // CountedLoopNode if there is control flow in the loop), the post-increment
183 // trip-counter value, and the limit. The trip-counter value is always of
184 // the form (Op old-trip-counter stride). The old-trip-counter is produced
185 // by a Phi connected to the CountedLoopNode. The stride is constant.
186 // The Op is any commutable opcode, including Add, Mul, Xor. The
187 // CountedLoopEndNode also takes in the loop-invariant limit value.
188
189 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
190 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
191 // via the old-trip-counter from the Op node.
192
193 //------------------------------CountedLoopNode--------------------------------
194 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
195 // inputs the incoming loop-start control and the loop-back control, so they
196 // act like RegionNodes. They also take in the initial trip counter, the
197 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
198 // produce a loop-body control and the trip counter value. Since
199 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
200
201 class BaseCountedLoopNode : public LoopNode {
202 public:
203 BaseCountedLoopNode(Node *entry, Node *backedge)
204 : LoopNode(entry, backedge) {
205 }
206
207 Node *init_control() const { return in(EntryControl); }
208 Node *back_control() const { return in(LoopBackControl); }
209
210 Node* init_trip() const;
211 Node* stride() const;
212 bool stride_is_con() const;
213 Node* limit() const;
214 Node* incr() const;
215 Node* phi() const;
216
217 BaseCountedLoopEndNode* loopexit_or_null() const;
218 BaseCountedLoopEndNode* loopexit() const;
219
220 virtual BasicType bt() const = 0;
221
222 jlong stride_con() const;
223
224 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
225
226 virtual void set_trip_count(julong tc) = 0;
227 virtual julong trip_count() const = 0;
228
229 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
230 void set_exact_trip_count(julong tc) {
231 set_trip_count(tc);
232 _loop_flags |= HasExactTripCount;
233 }
234 void set_nonexact_trip_count() {
235 _loop_flags &= ~HasExactTripCount;
236 }
237 };
238
239
240 class CountedLoopNode : public BaseCountedLoopNode {
241 // Size is bigger to hold _main_idx. However, _main_idx does not change
242 // the semantics so it does not appear in the hash & cmp functions.
243 virtual uint size_of() const { return sizeof(*this); }
244
245 // For Pre- and Post-loops during debugging ONLY, this holds the index of
246 // the Main CountedLoop. Used to assert that we understand the graph shape.
247 node_idx_t _main_idx;
248
249 // Known trip count calculated by compute_exact_trip_count()
250 uint _trip_count;
251
252 // Log2 of original loop bodies in unrolled loop
253 int _unrolled_count_log2;
254
255 // Node count prior to last unrolling - used to decide if
256 // unroll,optimize,unroll,optimize,... is making progress
257 int _node_count_before_unroll;
258
259 // If slp analysis is performed we record the maximum
260 // vector mapped unroll factor here
261 int _slp_maximum_unroll_factor;
262
263 public:
264 CountedLoopNode(Node *entry, Node *backedge)
265 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
266 _unrolled_count_log2(0), _node_count_before_unroll(0),
267 _slp_maximum_unroll_factor(0) {
268 init_class_id(Class_CountedLoop);
269 // Initialize _trip_count to the largest possible value.
270 // Will be reset (lower) if the loop's trip count is known.
271 }
272
273 virtual int Opcode() const;
274 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
275
276 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
277 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
278 int stride_con() const;
279
280 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
281 // can run short a few iterations and may start a few iterations in.
282 // It will be RCE'd and unrolled and aligned.
283
284 // A following 'post' loop will run any remaining iterations. Used
285 // during Range Check Elimination, the 'post' loop will do any final
286 // iterations with full checks. Also used by Loop Unrolling, where
287 // the 'post' loop will do any epilog iterations needed. Basically,
288 // a 'post' loop can not profitably be further unrolled or RCE'd.
289
290 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
291 // it may do under-flow checks for RCE and may do alignment iterations
292 // so the following main loop 'knows' that it is striding down cache
293 // lines.
294
295 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
296 // Aligned, may be missing it's pre-loop.
297 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
298 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
299 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
300 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
301 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
302 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
303 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
304 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
305 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
306 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
307
308 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
309
310 int main_idx() const { return _main_idx; }
311
312 void set_trip_count(julong tc) {
313 assert(tc < max_juint, "Cannot set trip count to max_juint");
314 _trip_count = checked_cast<uint>(tc);
315 }
316 julong trip_count() const { return _trip_count; }
317
318 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
319 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
320 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
321 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
322
323 void set_notpassed_slp() {
324 _loop_flags &= ~PassedSlpAnalysis;
325 }
326
327 void double_unrolled_count() { _unrolled_count_log2++; }
328 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
329
330 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
331 int node_count_before_unroll() { return _node_count_before_unroll; }
332 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
333 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
334
335 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
336 // decides which one is taken:
337 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
338 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
339 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
340 // assumption fails we take the slow_loop instead, where we do not make the same speculative
341 // assumption.
342 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
343 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
344 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
345 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
346 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
347 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
348 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
349 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
350 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
351 // Hence, we expect the slow_loop to execute slower.
352 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
353 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
354 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
355 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
356 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
357 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
358 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
359 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
360
361 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
362 OuterStripMinedLoopNode* outer_loop() const;
363 virtual IfTrueNode* outer_loop_tail() const;
364 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
365 virtual IfFalseNode* outer_loop_exit() const;
366 virtual SafePointNode* outer_safepoint() const;
367
368 Node* skip_assertion_predicates_with_halt();
369
370 virtual BasicType bt() const {
371 return T_INT;
372 }
373
374 Node* is_canonical_loop_entry();
375 CountedLoopEndNode* find_pre_loop_end();
376
377 Node* uncasted_init_trip(bool uncasted);
378
379 #ifndef PRODUCT
380 virtual void dump_spec(outputStream *st) const;
381 #endif
382 };
383
384 class LongCountedLoopNode : public BaseCountedLoopNode {
385 private:
386 virtual uint size_of() const { return sizeof(*this); }
387
388 // Known trip count calculated by compute_exact_trip_count()
389 julong _trip_count;
390
391 public:
392 LongCountedLoopNode(Node *entry, Node *backedge)
393 : BaseCountedLoopNode(entry, backedge), _trip_count(max_julong) {
394 init_class_id(Class_LongCountedLoop);
395 }
396
397 virtual int Opcode() const;
398
399 virtual BasicType bt() const {
400 return T_LONG;
401 }
402
403 void set_trip_count(julong tc) {
404 assert(tc < max_julong, "Cannot set trip count to max_julong");
405 _trip_count = tc;
406 }
407 julong trip_count() const { return _trip_count; }
408
409 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
410 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
411 };
412
413
414 //------------------------------CountedLoopEndNode-----------------------------
415 // CountedLoopEndNodes end simple trip counted loops. They act much like
416 // IfNodes.
417
418 class BaseCountedLoopEndNode : public IfNode {
419 public:
420 enum { TestControl, TestValue };
421 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
422 : IfNode(control, test, prob, cnt) {
423 init_class_id(Class_BaseCountedLoopEnd);
424 }
425
426 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
427 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
428 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
429 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
430 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
431 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
432
433 PhiNode* phi() const {
434 Node* tmp = incr();
435 if (tmp && tmp->req() == 3) {
436 Node* phi = tmp->in(1);
437 if (phi->is_Phi()) {
438 return phi->as_Phi();
439 }
440 }
441 return nullptr;
442 }
443
444 BaseCountedLoopNode* loopnode() const {
445 // The CountedLoopNode that goes with this CountedLoopEndNode may
446 // have been optimized out by the IGVN so be cautious with the
447 // pattern matching on the graph
448 PhiNode* iv_phi = phi();
449 if (iv_phi == nullptr) {
450 return nullptr;
451 }
452 Node* ln = iv_phi->in(0);
453 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
454 return nullptr;
455 }
456 if (ln->as_BaseCountedLoop()->bt() != bt()) {
457 return nullptr;
458 }
459 return ln->as_BaseCountedLoop();
460 }
461
462 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
463
464 jlong stride_con() const;
465 virtual BasicType bt() const = 0;
466
467 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
468 };
469
470 class CountedLoopEndNode : public BaseCountedLoopEndNode {
471 public:
472
473 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
474 : BaseCountedLoopEndNode(control, test, prob, cnt) {
475 init_class_id(Class_CountedLoopEnd);
476 }
477 virtual int Opcode() const;
478
479 CountedLoopNode* loopnode() const {
480 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
481 }
482
483 virtual BasicType bt() const {
484 return T_INT;
485 }
486
487 #ifndef PRODUCT
488 virtual void dump_spec(outputStream *st) const;
489 #endif
490 };
491
492 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
493 public:
494 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
495 : BaseCountedLoopEndNode(control, test, prob, cnt) {
496 init_class_id(Class_LongCountedLoopEnd);
497 }
498
499 LongCountedLoopNode* loopnode() const {
500 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
501 }
502
503 virtual int Opcode() const;
504
505 virtual BasicType bt() const {
506 return T_LONG;
507 }
508 };
509
510
511 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
512 Node* bctrl = back_control();
513 if (bctrl == nullptr) return nullptr;
514
515 Node* lexit = bctrl->in(0);
516 if (!lexit->is_BaseCountedLoopEnd()) {
517 return nullptr;
518 }
519 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
520 if (result->bt() != bt()) {
521 return nullptr;
522 }
523 return result;
524 }
525
526 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
527 BaseCountedLoopEndNode* cle = loopexit_or_null();
528 assert(cle != nullptr, "loopexit is null");
529 return cle;
530 }
531
532 inline Node* BaseCountedLoopNode::init_trip() const {
533 BaseCountedLoopEndNode* cle = loopexit_or_null();
534 return cle != nullptr ? cle->init_trip() : nullptr;
535 }
536 inline Node* BaseCountedLoopNode::stride() const {
537 BaseCountedLoopEndNode* cle = loopexit_or_null();
538 return cle != nullptr ? cle->stride() : nullptr;
539 }
540
541 inline bool BaseCountedLoopNode::stride_is_con() const {
542 BaseCountedLoopEndNode* cle = loopexit_or_null();
543 return cle != nullptr && cle->stride_is_con();
544 }
545 inline Node* BaseCountedLoopNode::limit() const {
546 BaseCountedLoopEndNode* cle = loopexit_or_null();
547 return cle != nullptr ? cle->limit() : nullptr;
548 }
549 inline Node* BaseCountedLoopNode::incr() const {
550 BaseCountedLoopEndNode* cle = loopexit_or_null();
551 return cle != nullptr ? cle->incr() : nullptr;
552 }
553 inline Node* BaseCountedLoopNode::phi() const {
554 BaseCountedLoopEndNode* cle = loopexit_or_null();
555 return cle != nullptr ? cle->phi() : nullptr;
556 }
557
558 inline jlong BaseCountedLoopNode::stride_con() const {
559 BaseCountedLoopEndNode* cle = loopexit_or_null();
560 return cle != nullptr ? cle->stride_con() : 0;
561 }
562
563
564 //------------------------------LoopLimitNode-----------------------------
565 // Counted Loop limit node which represents exact final iterator value:
566 // trip_count = (limit - init_trip + stride - 1)/stride
567 // final_value= trip_count * stride + init_trip.
568 // Use HW instructions to calculate it when it can overflow in integer.
569 // Note, final_value should fit into integer since counted loop has
570 // limit check: limit <= max_int-stride.
571 class LoopLimitNode : public Node {
572 enum { Init=1, Limit=2, Stride=3 };
573 public:
574 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
575 // Put it on the Macro nodes list to optimize during macro nodes expansion.
576 init_flags(Flag_is_macro);
577 C->add_macro_node(this);
578 }
579 virtual int Opcode() const;
580 virtual const Type *bottom_type() const { return TypeInt::INT; }
581 virtual uint ideal_reg() const { return Op_RegI; }
582 virtual const Type* Value(PhaseGVN* phase) const;
583 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
584 virtual Node* Identity(PhaseGVN* phase);
585 };
586
587 // Support for strip mining
588 class OuterStripMinedLoopNode : public LoopNode {
589 private:
590 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const;
591 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn);
592
593 public:
594 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
595 : LoopNode(entry, backedge) {
596 init_class_id(Class_OuterStripMinedLoop);
597 init_flags(Flag_is_macro);
598 C->add_macro_node(this);
599 }
600
601 virtual int Opcode() const;
602
603 virtual IfTrueNode* outer_loop_tail() const;
604 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
605 virtual IfFalseNode* outer_loop_exit() const;
606 virtual SafePointNode* outer_safepoint() const;
607 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); }
608 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); }
609 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->false_proj(); }
610
611 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
612
613 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
614
615 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
616
617 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
618
619 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
620 PhaseIdealLoop* iloop);
621 };
622
623 class OuterStripMinedLoopEndNode : public IfNode {
624 public:
625 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
626 : IfNode(control, test, prob, cnt) {
627 init_class_id(Class_OuterStripMinedLoopEnd);
628 }
629
630 virtual int Opcode() const;
631
632 virtual const Type* Value(PhaseGVN* phase) const;
633 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
634
635 bool is_expanded(PhaseGVN *phase) const;
636 };
637
638 // -----------------------------IdealLoopTree----------------------------------
639 class IdealLoopTree : public ResourceObj {
640 public:
641 IdealLoopTree *_parent; // Parent in loop tree
642 IdealLoopTree *_next; // Next sibling in loop tree
643 IdealLoopTree *_child; // First child in loop tree
644
645 // The head-tail backedge defines the loop.
646 // If a loop has multiple backedges, this is addressed during cleanup where
647 // we peel off the multiple backedges, merging all edges at the bottom and
648 // ensuring that one proper backedge flow into the loop.
649 Node *_head; // Head of loop
650 Node *_tail; // Tail of loop
651 inline Node *tail(); // Handle lazy update of _tail field
652 inline Node *head(); // Handle lazy update of _head field
653 PhaseIdealLoop* _phase;
654 int _local_loop_unroll_limit;
655 int _local_loop_unroll_factor;
656
657 Node_List _body; // Loop body for inner loops
658
659 uint16_t _nest; // Nesting depth
660 uint8_t _irreducible:1, // True if irreducible
661 _has_call:1, // True if has call safepoint
662 _has_sfpt:1, // True if has non-call safepoint
663 _rce_candidate:1, // True if candidate for range check elimination
664 _has_range_checks:1,
665 _has_range_checks_computed:1;
666
667 Node_List* _safepts; // List of safepoints in this loop
668 Node_List* _required_safept; // A inner loop cannot delete these safepts;
669 Node_List* _reachability_fences; // List of reachability fences in this loop
670 bool _allow_optimizations; // Allow loop optimizations
671
672 IdealLoopTree(PhaseIdealLoop* phase, Node* head, Node* tail);
673
674 // Is 'l' a member of 'this'?
675 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
676
677 // Set loop nesting depth. Accumulate has_call bits.
678 int set_nest( uint depth );
679
680 // Split out multiple fall-in edges from the loop header. Move them to a
681 // private RegionNode before the loop. This becomes the loop landing pad.
682 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
683
684 // Split out the outermost loop from this shared header.
685 void split_outer_loop( PhaseIdealLoop *phase );
686
687 // Merge all the backedges from the shared header into a private Region.
688 // Feed that region as the one backedge to this loop.
689 void merge_many_backedges( PhaseIdealLoop *phase );
690
691 // Split shared headers and insert loop landing pads.
692 // Insert a LoopNode to replace the RegionNode.
693 // Returns TRUE if loop tree is structurally changed.
694 bool beautify_loops( PhaseIdealLoop *phase );
695
696 // Perform optimization to use the loop predicates for null checks and range checks.
697 // Applies to any loop level (not just the innermost one)
698 bool loop_predication( PhaseIdealLoop *phase);
699 bool can_apply_loop_predication();
700
701 // Perform iteration-splitting on inner loops. Split iterations to
702 // avoid range checks or one-shot null checks. Returns false if the
703 // current round of loop opts should stop.
704 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
705
706 // Driver for various flavors of iteration splitting. Returns false
707 // if the current round of loop opts should stop.
708 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
709
710 // Given dominators, try to find loops with calls that must always be
711 // executed (call dominates loop tail). These loops do not need non-call
712 // safepoints (ncsfpt).
713 void check_safepts(VectorSet &visited, Node_List &stack);
714
715 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
716 // encountered.
717 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
718
719 // Remove safepoints from loop. Optionally keeping one.
720 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
721
722 // Convert to counted loops where possible
723 void counted_loop( PhaseIdealLoop *phase );
724
725 // Check for Node being a loop-breaking test
726 Node *is_loop_exit(Node *iff) const;
727
728 // Return unique loop-exit projection or null if the loop has multiple exits.
729 IfFalseNode* unique_loop_exit_proj_or_null();
730
731 // Remove simplistic dead code from loop body
732 void DCE_loop_body();
733
734 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
735 // Replace with a 1-in-10 exit guess.
736 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
737
738 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
739 // Useful for unrolling loops with NO array accesses.
740 bool policy_peel_only( PhaseIdealLoop *phase ) const;
741
742 // Return TRUE or FALSE if the loop should be unswitched -- clone
743 // loop with an invariant test
744 bool policy_unswitching( PhaseIdealLoop *phase ) const;
745 bool no_unswitch_candidate() const;
746
747 // Micro-benchmark spamming. Remove empty loops.
748 bool do_remove_empty_loop( PhaseIdealLoop *phase );
749
750 // Convert one-iteration loop into normal code.
751 bool do_one_iteration_loop( PhaseIdealLoop *phase );
752
753 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
754 // move some loop-invariant test (usually a null-check) before the loop.
755 bool policy_peeling(PhaseIdealLoop *phase);
756
757 uint estimate_peeling(PhaseIdealLoop *phase);
758
759 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
760 // known trip count in the counted loop node.
761 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
762
763 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
764 // if the loop is a counted loop and the loop body is small enough.
765 bool policy_unroll(PhaseIdealLoop *phase);
766
767 // Loop analyses to map to a maximal superword unrolling for vectorization.
768 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
769
770 // Return TRUE or FALSE if the loop should be range-check-eliminated.
771 // Gather a list of IF tests that are dominated by iteration splitting;
772 // also gather the end of the first split and the start of the 2nd split.
773 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
774
775 // Return TRUE if "iff" is a range check.
776 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
777 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
778 jlong& scale) const;
779
780 // Estimate the number of nodes required when cloning a loop (body).
781 uint est_loop_clone_sz(uint factor) const;
782 // Estimate the number of nodes required when unrolling a loop (body).
783 uint est_loop_unroll_sz(uint factor) const;
784
785 // Compute loop trip count if possible
786 void compute_trip_count(PhaseIdealLoop* phase, BasicType bt);
787
788 // Compute loop trip count from profile data
789 float compute_profile_trip_cnt_helper(Node* n);
790 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
791
792 // Reassociate invariant expressions.
793 void reassociate_invariants(PhaseIdealLoop *phase);
794 // Reassociate invariant binary expressions.
795 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
796 // Reassociate invariant add, subtract, and compare expressions.
797 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
798 // Return nonzero index of invariant operand if invariant and variant
799 // are combined with an associative binary. Helper for reassociate_invariants.
800 int find_invariant(Node* n, PhaseIdealLoop *phase);
801 // Return TRUE if "n" is associative.
802 bool is_associative(Node* n, Node* base=nullptr);
803 // Return TRUE if "n" is an associative cmp node.
804 bool is_associative_cmp(Node* n);
805
806 // Return true if n is invariant
807 bool is_invariant(Node* n) const;
808
809 // Put loop body on igvn work list
810 void record_for_igvn();
811
812 bool is_root() { return _parent == nullptr; }
813 // A proper/reducible loop w/o any (occasional) dead back-edge.
814 bool is_loop() { return !_irreducible && !tail()->is_top(); }
815 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
816 bool is_innermost() { return is_loop() && _child == nullptr; }
817
818 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
819
820 bool compute_has_range_checks() const;
821 bool range_checks_present() {
822 if (!_has_range_checks_computed) {
823 if (compute_has_range_checks()) {
824 _has_range_checks = 1;
825 }
826 _has_range_checks_computed = 1;
827 }
828 return _has_range_checks;
829 }
830
831 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
832 // In all other cases, return this.
833 IdealLoopTree* skip_strip_mined() {
834 return _head->as_Loop()->is_strip_mined() ? _parent : this;
835 }
836
837 // Registers a reachability fence node in the loop.
838 void register_reachability_fence(ReachabilityFenceNode* rf);
839
840 #ifndef PRODUCT
841 void dump_head(); // Dump loop head only
842 void dump(); // Dump this loop recursively
843 #endif
844
845 #ifdef ASSERT
846 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
847 bool verify_tree(IdealLoopTree* loop_verify) const;
848 #endif
849
850 private:
851 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
852
853 // Estimate the number of nodes resulting from control and data flow merge.
854 uint est_loop_flow_merge_sz() const;
855
856 // Check if the number of residual iterations is large with unroll_cnt.
857 // Return true if the residual iterations are more than 10% of the trip count.
858 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
859 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
860 }
861
862 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
863
864 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
865
866 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
867
868 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
869 Node* sfpt) const;
870
871 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
872
873 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
874 };
875
876 // -----------------------------PhaseIdealLoop---------------------------------
877 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
878 // into a loop tree. Drives the loop-based transformations on the ideal graph.
879 class PhaseIdealLoop : public PhaseTransform {
880 friend class IdealLoopTree;
881 friend class SuperWord;
882 friend class ShenandoahBarrierC2Support;
883 friend class AutoNodeBudget;
884
885 Arena _arena; // For data whose lifetime is a single pass of loop optimizations
886
887 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
888 //
889 // Exception: dead CFG nodes may instead have a ctrl/idom forwarding
890 // installed. See: forward_ctrl
891 Node_List _loop_or_ctrl;
892
893 // Pre-computed def-use info
894 PhaseIterGVN &_igvn;
895
896 // Head of loop tree
897 IdealLoopTree* _ltree_root;
898
899 // Array of pre-order numbers, plus post-visited bit.
900 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
901 // ODD for post-visited. Other bits are the pre-order number.
902 uint *_preorders;
903 uint _max_preorder;
904
905 ReallocMark _nesting; // Safety checks for arena reallocation
906
907 const PhaseIdealLoop* _verify_me;
908 bool _verify_only;
909
910 // Allocate _preorders[] array
911 void allocate_preorders() {
912 _max_preorder = C->unique()+8;
913 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
914 memset(_preorders, 0, sizeof(uint) * _max_preorder);
915 }
916
917 // Allocate _preorders[] array
918 void reallocate_preorders() {
919 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
920 if ( _max_preorder < C->unique() ) {
921 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
922 _max_preorder = C->unique();
923 }
924 memset(_preorders, 0, sizeof(uint) * _max_preorder);
925 }
926
927 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
928 // adds new nodes.
929 void check_grow_preorders( ) {
930 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
931 if ( _max_preorder < C->unique() ) {
932 uint newsize = _max_preorder<<1; // double size of array
933 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
934 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
935 _max_preorder = newsize;
936 }
937 }
938 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
939 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
940 // Pre-order numbers are written to the Nodes array as low-bit-set values.
941 void set_preorder_visited( Node *n, int pre_order ) {
942 assert( !is_visited( n ), "already set" );
943 _preorders[n->_idx] = (pre_order<<1);
944 };
945 // Return pre-order number.
946 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
947
948 // Check for being post-visited.
949 // Should be previsited already (checked with assert(is_visited(n))).
950 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
951
952 // Mark as post visited
953 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
954
955 public:
956 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
957 // Returns true if "n" is a data node, false if it's a CFG node.
958 //
959 // Exception:
960 // control nodes that are dead because of "replace_node_and_forward_ctrl"
961 // or have otherwise modified their ctrl state by "forward_ctrl".
962 // They return "true", because they have a ctrl "forwarding" to the other ctrl node they
963 // were replaced with.
964 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
965
966 private:
967 // clear out dead code after build_loop_late
968 Node_List _deadlist;
969 Node_List _zero_trip_guard_opaque_nodes;
970 Node_List _multiversion_opaque_nodes;
971
972 // Support for faster execution of get_late_ctrl()/dom_lca()
973 // when a node has many uses and dominator depth is deep.
974 GrowableArray<jlong> _dom_lca_tags;
975 uint _dom_lca_tags_round;
976 void init_dom_lca_tags();
977
978 // Helper for debugging bad dominance relationships
979 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
980
981 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
982
983 // Inline wrapper for frequent cases:
984 // 1) only one use
985 // 2) a use is the same as the current LCA passed as 'n1'
986 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
987 assert( n->is_CFG(), "" );
988 // Fast-path null lca
989 if( lca != nullptr && lca != n ) {
990 assert( lca->is_CFG(), "" );
991 // find LCA of all uses
992 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
993 }
994 return find_non_split_ctrl(n);
995 }
996 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
997
998 // Helper function for directing control inputs away from CFG split points.
999 Node *find_non_split_ctrl( Node *ctrl ) const {
1000 if (ctrl != nullptr) {
1001 if (ctrl->is_MultiBranch()) {
1002 ctrl = ctrl->in(0);
1003 }
1004 assert(ctrl->is_CFG(), "CFG");
1005 }
1006 return ctrl;
1007 }
1008
1009 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
1010
1011 #ifdef ASSERT
1012 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
1013 #endif
1014 private:
1015 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
1016 Unique_Node_List& list);
1017 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
1018 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1019 CountedLoopNode* remaining_loop_head,
1020 uint first_node_index_in_cloned_loop_body,
1021 const Node_List& old_new);
1022 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1023 CountedLoopNode* main_loop_head,
1024 uint first_node_index_in_pre_loop_body,
1025 uint last_node_index_in_pre_loop_body,
1026 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1027 const Node_List& old_new);
1028 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1029 uint first_node_index_in_cloned_loop_body);
1030 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1031 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1032 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1033 CountedLoopNode* target_loop_head,
1034 const NodeInLoopBody& _node_in_loop_body,
1035 bool kill_old_template);
1036 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1037 const Node* old_target_loop_entry,
1038 uint node_index_before_new_assertion_predicate_nodes);
1039 void log_loop_tree();
1040
1041 public:
1042
1043 PhaseIterGVN &igvn() const { return _igvn; }
1044
1045 Arena* arena() { return &_arena; };
1046
1047 bool has_node(const Node* n) const {
1048 guarantee(n != nullptr, "No Node.");
1049 return _loop_or_ctrl[n->_idx] != nullptr;
1050 }
1051 // check if transform created new nodes that need _ctrl recorded
1052 Node *get_late_ctrl( Node *n, Node *early );
1053 Node *get_early_ctrl( Node *n );
1054 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1055 void set_early_ctrl(Node* n, bool update_body);
1056 void set_subtree_ctrl(Node* n, bool update_body);
1057 void set_ctrl( Node *n, Node *ctrl ) {
1058 assert( !has_node(n) || has_ctrl(n), "" );
1059 assert( ctrl->in(0), "cannot set dead control node" );
1060 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1061 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1062 }
1063 void set_root_as_ctrl(Node* n) {
1064 assert(!has_node(n) || has_ctrl(n), "");
1065 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1066 }
1067 // Set control and update loop membership
1068 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1069 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1070 IdealLoopTree* new_loop = get_loop(ctrl);
1071 if (old_loop != new_loop) {
1072 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1073 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1074 }
1075 set_ctrl(n, ctrl);
1076 }
1077
1078 // Retrieves the ctrl for a data node i.
1079 Node* get_ctrl(const Node* i) {
1080 assert(has_node(i) && has_ctrl(i), "must be data node with ctrl");
1081 Node* n = get_ctrl_no_update(i);
1082 // We store the found ctrl in the side-table again. In most cases,
1083 // this is a no-op, since we just read from _loop_or_ctrl. But in cases
1084 // where there was a ctrl forwarding via dead ctrl nodes, this shortens the path.
1085 // See: forward_ctrl
1086 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1087 assert(has_node(i) && has_ctrl(i), "must still be data node with ctrl");
1088 assert(n == find_non_split_ctrl(n), "must return legal ctrl");
1089 return n;
1090 }
1091
1092 bool is_dominator(Node* dominator, Node* n);
1093 bool is_strict_dominator(Node* dominator, Node* n);
1094
1095 // return get_ctrl for a data node and self(n) for a CFG node
1096 Node* ctrl_or_self(Node* n) {
1097 if (has_ctrl(n))
1098 return get_ctrl(n);
1099 else {
1100 assert (n->is_CFG(), "must be a CFG node");
1101 return n;
1102 }
1103 }
1104
1105 private:
1106 Node* get_ctrl_no_update_helper(const Node* i) const {
1107 // We expect only data nodes (which must have a ctrl set), or
1108 // dead ctrl nodes that have a ctrl "forwarding".
1109 // See: forward_ctrl.
1110 assert(has_ctrl(i), "only data nodes or ctrl nodes with ctrl forwarding expected");
1111 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1112 }
1113
1114 // Compute the ctrl of node i, jumping over ctrl forwardings.
1115 Node* get_ctrl_no_update(const Node* i) const {
1116 assert(has_ctrl(i), "only data nodes expected");
1117 Node* n = get_ctrl_no_update_helper(i);
1118 if (n->in(0) == nullptr) {
1119 // We encountered a dead CFG node.
1120 // If everything went right, this dead CFG node should have had a ctrl
1121 // forwarding installed, using "forward_ctrl". We now have to jump from
1122 // the old (dead) ctrl node to the new (live) ctrl node, in possibly
1123 // multiple ctrl forwarding steps.
1124 do {
1125 n = get_ctrl_no_update_helper(n);
1126 } while (n->in(0) == nullptr);
1127 n = find_non_split_ctrl(n);
1128 }
1129 return n;
1130 }
1131
1132 public:
1133 // Check for loop being set
1134 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1135 bool has_loop( Node *n ) const {
1136 assert(!has_node(n) || !has_ctrl(n), "");
1137 return has_node(n);
1138 }
1139 // Set loop
1140 void set_loop( Node *n, IdealLoopTree *loop ) {
1141 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1142 }
1143
1144 // Install a ctrl "forwarding" from an old (dead) control node.
1145 // This is a "lazy" update of the "get_ctrl" and "idom" mechanism:
1146 // - Install a forwarding from old_node (dead ctrl) to new_node.
1147 // - When querying "get_ctrl": jump from data node over possibly
1148 // multiple dead ctrl nodes with ctrl forwarding to eventually
1149 // reach a live ctrl node. Shorten the path to avoid chasing the
1150 // forwarding in the future.
1151 // - When querying "idom": from some node get its old idom, which
1152 // may be dead but has an idom forwarding to the new and live
1153 // idom. Shorten the path to avoid chasing the forwarding in the
1154 // future.
1155 // Note: while the "idom" information is stored in the "_idom"
1156 // side-table, the idom forwarding piggybacks on the ctrl
1157 // forwarding on "_loop_or_ctrl".
1158 // Using "forward_ctrl" allows us to only edit the entry for the old
1159 // dead node now, and we do not have to update all the nodes that had
1160 // the old_node as their "get_ctrl" or "idom". We clean up the forwarding
1161 // links when we query "get_ctrl" or "idom" for these nodes the next time.
1162 void forward_ctrl(Node* old_node, Node* new_node) {
1163 assert(!has_ctrl(old_node) && old_node->is_CFG() && old_node->in(0) == nullptr,
1164 "must be dead ctrl (CFG) node");
1165 assert(!has_ctrl(new_node) && new_node->is_CFG() && new_node->in(0) != nullptr,
1166 "must be live ctrl (CFG) node");
1167 assert(old_node != new_node, "no cycles please");
1168 // Re-use the side array slot for this node to provide the
1169 // forwarding pointer.
1170 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1171 assert(has_ctrl(old_node), "must have installed ctrl forwarding");
1172 }
1173
1174 // Replace the old ctrl node with a new ctrl node.
1175 // - Update the node inputs of all uses.
1176 // - Lazily update the ctrl and idom info of all uses, via a ctrl/idom forwarding.
1177 void replace_node_and_forward_ctrl(Node* old_node, Node* new_node) {
1178 _igvn.replace_node(old_node, new_node);
1179 forward_ctrl(old_node, new_node);
1180 }
1181
1182 void remove_dead_data_node(Node* dead) {
1183 assert(dead->outcnt() == 0 && !dead->is_top(), "must be dead");
1184 assert(!dead->is_CFG(), "not a data node");
1185 Node* c = get_ctrl(dead);
1186 IdealLoopTree* lpt = get_loop(c);
1187 _loop_or_ctrl.map(dead->_idx, nullptr); // This node is useless
1188 lpt->_body.yank(dead);
1189 igvn().remove_dead_node(dead, PhaseIterGVN::NodeOrigin::Graph);
1190 }
1191
1192 private:
1193
1194 // Place 'n' in some loop nest, where 'n' is a CFG node
1195 void build_loop_tree();
1196 int build_loop_tree_impl(Node* n, int pre_order);
1197 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1198 // loop tree, not the root.
1199 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1200
1201 #ifdef ASSERT
1202 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1203 void verify_regions_in_irreducible_loops();
1204 bool is_in_irreducible_loop(RegionNode* region);
1205 #endif
1206
1207 // Place Data nodes in some loop nest
1208 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1209 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1210 void build_loop_late_post_work(Node* n, bool pinned);
1211 void build_loop_late_post(Node* n);
1212 void verify_strip_mined_scheduling(Node *n, Node* least);
1213
1214 // Array of immediate dominance info for each CFG node indexed by node idx
1215 private:
1216 uint _idom_size;
1217 Node **_idom; // Array of immediate dominators
1218 uint *_dom_depth; // Used for fast LCA test
1219 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1220 LoopOptsMode _mode;
1221
1222 // build the loop tree and perform any requested optimizations
1223 void build_and_optimize();
1224
1225 // Dominators for the sea of nodes
1226 void Dominators();
1227
1228 // Compute the Ideal Node to Loop mapping
1229 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1230 PhaseTransform(Ideal_Loop),
1231 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1232 _loop_or_ctrl(&_arena),
1233 _igvn(igvn),
1234 _verify_me(nullptr),
1235 _verify_only(false),
1236 _mode(mode),
1237 _nodes_required(UINT_MAX) {
1238 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1239 build_and_optimize();
1240 }
1241
1242 #ifndef PRODUCT
1243 // Verify that verify_me made the same decisions as a fresh run
1244 // or only verify that the graph is valid if verify_me is null.
1245 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1246 PhaseTransform(Ideal_Loop),
1247 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1248 _loop_or_ctrl(&_arena),
1249 _igvn(igvn),
1250 _verify_me(verify_me),
1251 _verify_only(verify_me == nullptr),
1252 _mode(LoopOptsVerify),
1253 _nodes_required(UINT_MAX) {
1254 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1255 build_and_optimize();
1256 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1257 }
1258 #endif
1259
1260 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1261
1262 Node* idom_no_update(Node* d) const {
1263 return idom_no_update(d->_idx);
1264 }
1265
1266 Node* idom_no_update(uint node_idx) const {
1267 assert(node_idx < _idom_size, "oob");
1268 Node* n = _idom[node_idx];
1269 assert(n != nullptr,"Bad immediate dominator info.");
1270 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1271 // We encountered a dead CFG node.
1272 // If everything went right, this dead CFG node should have had an idom
1273 // forwarding installed, using "forward_ctrl". We now have to jump from
1274 // the old (dead) idom node to the new (live) idom node, in possibly
1275 // multiple idom forwarding steps.
1276 // Note that we piggyback on "_loop_or_ctrl" to do the forwarding,
1277 // since we forward both "get_ctrl" and "idom" from the dead to the
1278 // new live ctrl/idom nodes.
1279 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1280 assert(n != nullptr,"Bad immediate dominator info.");
1281 }
1282 return n;
1283 }
1284
1285 public:
1286 Node* idom(Node* n) const {
1287 return idom(n->_idx);
1288 }
1289
1290 Node* idom(uint node_idx) const {
1291 Node* n = idom_no_update(node_idx);
1292 // We store the found idom in the side-table again. In most cases,
1293 // this is a no-op, since we just read from _idom. But in cases where
1294 // there was an idom forwarding via dead idom nodes, this shortens the path.
1295 // See: forward_ctrl
1296 _idom[node_idx] = n;
1297 return n;
1298 }
1299
1300 uint dom_depth(Node* d) const {
1301 guarantee(d != nullptr, "Null dominator info.");
1302 guarantee(d->_idx < _idom_size, "");
1303 return _dom_depth[d->_idx];
1304 }
1305 void set_idom(Node* d, Node* n, uint dom_depth);
1306 // Locally compute IDOM using dom_lca call
1307 Node *compute_idom( Node *region ) const;
1308 // Recompute dom_depth
1309 void recompute_dom_depth();
1310
1311 // Is safept not required by an outer loop?
1312 bool is_deleteable_safept(Node* sfpt) const;
1313
1314 // Replace parallel induction variable (parallel to trip counter)
1315 void replace_parallel_iv(IdealLoopTree *loop);
1316
1317 Node *dom_lca( Node *n1, Node *n2 ) const {
1318 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1319 }
1320 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1321
1322 Node* dominated_node(Node* c1, Node* c2) {
1323 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1324 return is_dominator(c1, c2) ? c2 : c1;
1325 }
1326
1327 // Return control node that's dominated by the 2 others
1328 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1329 return dominated_node(c1, dominated_node(c2, c3));
1330 }
1331
1332 // Build and verify the loop tree without modifying the graph. This
1333 // is useful to verify that all inputs properly dominate their uses.
1334 static void verify(PhaseIterGVN& igvn) {
1335 #ifdef ASSERT
1336 ResourceMark rm;
1337 Compile::TracePhase tp(_t_idealLoopVerify);
1338 PhaseIdealLoop v(igvn);
1339 #endif
1340 }
1341
1342 // Recommended way to use PhaseIdealLoop.
1343 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1344 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1345 ResourceMark rm;
1346 PhaseIdealLoop v(igvn, mode);
1347
1348 Compile* C = Compile::current();
1349 if (!C->failing()) {
1350 // Cleanup any modified bits
1351 igvn.optimize();
1352 if (C->failing()) { return; }
1353 v.log_loop_tree();
1354 }
1355 }
1356
1357 // True if the method has at least 1 irreducible loop
1358 bool _has_irreducible_loops;
1359
1360 // Per-Node transform
1361 virtual Node* transform(Node* n) { return nullptr; }
1362
1363 Node* loop_exit_control(const IdealLoopTree* loop) const;
1364
1365 class LoopExitTest {
1366 bool _is_valid;
1367
1368 const Node* _back_control;
1369 const IdealLoopTree* _loop;
1370 PhaseIdealLoop* _phase;
1371
1372 Node* _cmp;
1373 Node* _incr;
1374 Node* _limit;
1375 BoolTest::mask _mask;
1376 float _cl_prob;
1377
1378 public:
1379 LoopExitTest(const Node* back_control, const IdealLoopTree* loop, PhaseIdealLoop* phase) :
1380 _is_valid(false),
1381 _back_control(back_control),
1382 _loop(loop),
1383 _phase(phase),
1384 _cmp(nullptr),
1385 _incr(nullptr),
1386 _limit(nullptr),
1387 _mask(BoolTest::illegal),
1388 _cl_prob(0.0f) {}
1389
1390 void build();
1391 void canonicalize_mask(jlong stride_con);
1392
1393 bool is_valid_with_bt(BasicType bt) const {
1394 return _is_valid && _cmp != nullptr && _cmp->Opcode() == Op_Cmp(bt);
1395 }
1396
1397 bool should_include_limit() const { return _mask == BoolTest::le || _mask == BoolTest::ge; }
1398
1399 CmpNode* cmp() const { return _cmp->as_Cmp(); }
1400 Node* incr() const { return _incr; }
1401 Node* limit() const { return _limit; }
1402 BoolTest::mask mask() const { return _mask; }
1403 float cl_prob() const { return _cl_prob; }
1404 };
1405
1406 class LoopIVIncr {
1407 bool _is_valid;
1408
1409 const Node* _head;
1410 const IdealLoopTree* _loop;
1411
1412 Node* _incr;
1413 Node* _phi_incr;
1414
1415 public:
1416 LoopIVIncr(const Node* head, const IdealLoopTree* loop) :
1417 _is_valid(false),
1418 _head(head),
1419 _loop(loop),
1420 _incr(nullptr),
1421 _phi_incr(nullptr) {}
1422
1423 void build(Node* old_incr);
1424
1425 bool is_valid() const { return _is_valid; }
1426 bool is_valid_with_bt(const BasicType bt) const {
1427 return _is_valid && _incr->Opcode() == Op_Add(bt);
1428 }
1429
1430 Node* incr() const { return _incr; }
1431 Node* phi_incr() const { return _phi_incr; }
1432 };
1433
1434 class LoopIVStride {
1435 bool _is_valid;
1436
1437 BasicType _iv_bt;
1438 Node* _stride_node;
1439 Node* _xphi;
1440
1441 public:
1442 LoopIVStride(BasicType iv_bt) :
1443 _is_valid(false),
1444 _iv_bt(iv_bt),
1445 _stride_node(nullptr),
1446 _xphi(nullptr) {}
1447
1448 void build(const Node* incr);
1449
1450 bool is_valid() const { return _is_valid && _stride_node != nullptr; }
1451 Node* stride_node() const { return _stride_node; }
1452 Node* xphi() const { return _xphi; }
1453
1454 jlong compute_non_zero_stride_con(BoolTest::mask mask, BasicType iv_bt) const;
1455 };
1456
1457 static PhiNode* loop_iv_phi(const Node* xphi, const Node* phi_incr, const Node* head);
1458
1459 bool try_convert_to_counted_loop(Node* head, IdealLoopTree*& loop, BasicType iv_bt);
1460
1461 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1462 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1463
1464 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1465 SafePointNode* find_safepoint(Node* back_control, const Node* head, const IdealLoopTree* loop);
1466
1467 void add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt);
1468
1469 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1470 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1471 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1472 Node*& entry_control, Node*& iffalse);
1473
1474 Node* exact_limit( IdealLoopTree *loop );
1475
1476 // Return a post-walked LoopNode
1477 IdealLoopTree* get_loop(const Node* n) const {
1478 // Dead nodes have no loop, so return the top level loop instead
1479 if (!has_node(n)) return _ltree_root;
1480 assert(!has_ctrl(n), "");
1481 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1482 }
1483
1484 IdealLoopTree* ltree_root() const { return _ltree_root; }
1485
1486 // Is 'n' a (nested) member of 'loop'?
1487 bool is_member(const IdealLoopTree* loop, const Node* n) const {
1488 return loop->is_member(get_loop(n));
1489 }
1490
1491 // is the control for 'n' a (nested) member of 'loop'?
1492 bool ctrl_is_member(const IdealLoopTree* loop, const Node* n) {
1493 return is_member(loop, get_ctrl(n));
1494 }
1495
1496 // This is the basic building block of the loop optimizations. It clones an
1497 // entire loop body. It makes an old_new loop body mapping; with this
1498 // mapping you can find the new-loop equivalent to an old-loop node. All
1499 // new-loop nodes are exactly equal to their old-loop counterparts, all
1500 // edges are the same. All exits from the old-loop now have a RegionNode
1501 // that merges the equivalent new-loop path. This is true even for the
1502 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1503 // now come from (one or more) Phis that merge their new-loop equivalents.
1504 // Parameter side_by_side_idom:
1505 // When side_by_size_idom is null, the dominator tree is constructed for
1506 // the clone loop to dominate the original. Used in construction of
1507 // pre-main-post loop sequence.
1508 // When nonnull, the clone and original are side-by-side, both are
1509 // dominated by the passed in side_by_side_idom node. Used in
1510 // construction of unswitched loops.
1511 enum CloneLoopMode {
1512 IgnoreStripMined = 0, // Only clone inner strip mined loop
1513 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1514 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1515 // result control flow branches
1516 // either to inner clone or outer
1517 // strip mined loop.
1518 };
1519 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1520 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1521 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1522 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1523 Node_List*& split_if_set, Node_List*& split_bool_set,
1524 Node_List*& split_cex_set, Node_List& worklist,
1525 uint new_counter, CloneLoopMode mode);
1526 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1527 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1528 Node_List& extra_data_nodes);
1529
1530 // If we got the effect of peeling, either by actually peeling or by
1531 // making a pre-loop which must execute at least once, we can remove
1532 // all loop-invariant dominated tests in the main body.
1533 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1534
1535 // Generate code to do a loop peel for the given loop (and body).
1536 // old_new is a temp array.
1537 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1538
1539 // Add pre and post loops around the given loop. These loops are used
1540 // during RCE, unrolling and aligning loops.
1541 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1542
1543 // Find the last store in the body of an OuterStripMinedLoop when following memory uses
1544 Node *find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop);
1545
1546 // Add post loop after the given loop.
1547 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1548 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1549 Node* incr, Node* limit, CountedLoopNode*& post_head);
1550
1551 // Add a vector post loop between a vector main loop and the current post loop
1552 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1553 // If Node n lives in the back_ctrl block, we clone a private version of n
1554 // in preheader_ctrl block and return that, otherwise return n.
1555 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1556
1557 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1558 // unroll to do double iterations. The next round of major loop transforms
1559 // will repeat till the doubled loop body does all remaining iterations in 1
1560 // pass.
1561 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1562
1563 // Unroll the loop body one step - make each trip do 2 iterations.
1564 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1565
1566 // Return true if exp is a constant times an induction var
1567 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1568
1569 bool is_iv(Node* exp, Node* iv, BasicType bt);
1570
1571 // Return true if exp is a scaled induction var plus (or minus) constant
1572 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);
1573 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1574 jlong long_scale;
1575 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1576 int int_scale = checked_cast<int>(long_scale);
1577 if (p_scale != nullptr) {
1578 *p_scale = int_scale;
1579 }
1580 return true;
1581 }
1582 return false;
1583 }
1584 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1585 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1586 BasicType bt,
1587 jlong* p_scale, Node** p_offset,
1588 bool* p_short_scale, int depth);
1589
1590 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1591 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1592 Deoptimization::DeoptReason reason, int opcode,
1593 bool rewire_uncommon_proj_phi_inputs = false);
1594
1595 private:
1596 // Helper functions for create_new_if_for_predicate()
1597 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1598 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1599 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1600 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1601 const OrigToNewHashtable& orig_to_clone);
1602
1603 public:
1604 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1605
1606 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1607 void replace_control(Node* node, Node* new_control) {
1608 _igvn.replace_input_of(node, 0, new_control);
1609 set_idom(node, new_control, dom_depth(new_control));
1610 }
1611
1612 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1613 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1614 set_idom(loop_head, new_entry, dom_depth(new_entry));
1615 }
1616
1617 // Construct a range check for a predicate if
1618 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1619 jint stride, Node* range, bool upper, bool& overflow);
1620
1621 // Implementation of the loop predication to promote checks outside the loop
1622 bool loop_predication_impl(IdealLoopTree *loop);
1623
1624 // Reachability Fence (RF) support.
1625 private:
1626 void insert_rf(Node* ctrl, Node* referent);
1627 void replace_rf(Node* old_node, Node* new_node);
1628 void remove_rf(ReachabilityFenceNode* rf);
1629 public:
1630 bool optimize_reachability_fences();
1631 bool expand_reachability_fences();
1632
1633 private:
1634 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1635 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1636 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1637 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1638 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1639 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1640 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1641 Node_List& if_proj_list);
1642 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1643 IfProjNode* new_control, int scale, Node* offset, Node* range);
1644 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1645
1646 // Helper function to collect predicate for eliminating the useless ones
1647 void eliminate_useless_predicates() const;
1648
1649 void eliminate_useless_zero_trip_guard();
1650 void eliminate_useless_multiversion_if();
1651
1652 public:
1653 // Change the control input of expensive nodes to allow commoning by
1654 // IGVN when it is guaranteed to not result in a more frequent
1655 // execution of the expensive node. Return true if progress.
1656 bool process_expensive_nodes();
1657
1658 // Check whether node has become unreachable
1659 bool is_node_unreachable(Node *n) const {
1660 return !has_node(n) || n->is_unreachable(_igvn);
1661 }
1662
1663 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1664 void do_range_check(IdealLoopTree* loop);
1665
1666 // Clone loop with an invariant test (that does not exit) and
1667 // insert a clone of the test that selects which version to
1668 // execute.
1669 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1670
1671 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1672 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1673
1674 private:
1675 static bool has_control_dependencies_from_predicates(LoopNode* head);
1676 static void revert_to_normal_loop(const LoopNode* loop_head);
1677
1678 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1679 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1680 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1681 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1682 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1683 #ifndef PRODUCT
1684 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1685 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1686 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1687 const UnswitchCandidate& unswitch_candidate,
1688 const LoopNode* original_head, const LoopNode* new_head);
1689 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1690 const LoopNode* original_head, const LoopNode* new_head);
1691 #endif
1692
1693 public:
1694
1695 // Range Check Elimination uses this function!
1696 // Constrain the main loop iterations so the affine function:
1697 // low_limit <= scale_con * I + offset < upper_limit
1698 // always holds true. That is, either increase the number of iterations in
1699 // the pre-loop or the post-loop until the condition holds true in the main
1700 // loop. Scale_con, offset and limit are all loop invariant.
1701 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);
1702 // Helper function for add_constraint().
1703 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1704
1705 // Partially peel loop up through last_peel node.
1706 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1707 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1708
1709 // AutoVectorize the loop: replace scalar ops with vector ops.
1710 enum AutoVectorizeStatus {
1711 Impossible, // This loop has the wrong shape to even try vectorization.
1712 Success, // We just successfully vectorized the loop.
1713 TriedAndFailed, // We tried to vectorize, but failed.
1714 };
1715 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1716
1717 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1718 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1719 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1720 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1721
1722 // Create a scheduled list of nodes control dependent on ctrl set.
1723 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1724 // Has a use in the vector set
1725 bool has_use_in_set( Node* n, VectorSet& vset );
1726 // Has use internal to the vector set (ie. not in a phi at the loop head)
1727 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1728 // clone "n" for uses that are outside of loop
1729 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1730 // clone "n" for special uses that are in the not_peeled region
1731 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1732 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1733 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1734 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1735 #ifdef ASSERT
1736 // Validate the loop partition sets: peel and not_peel
1737 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1738 // Ensure that uses outside of loop are of the right form
1739 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1740 uint orig_exit_idx, uint clone_exit_idx);
1741 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1742 #endif
1743
1744 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1745 int stride_of_possible_iv( Node* iff );
1746 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1747 // Return the (unique) control output node that's in the loop (if it exists.)
1748 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1749 // Insert a signed compare loop exit cloned from an unsigned compare.
1750 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1751 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1752 // Utility to register node "n" with PhaseIdealLoop
1753 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1754 // Utility to create an if-projection
1755 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1756 // Force the iff control output to be the live_proj
1757 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1758 // Insert a region before an if projection
1759 RegionNode* insert_region_before_proj(ProjNode* proj);
1760 // Insert a new if before an if projection
1761 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1762
1763 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1764 // "Nearly" because all Nodes have been cloned from the original in the loop,
1765 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1766 // through the Phi recursively, and return a Bool.
1767 Node* clone_iff(PhiNode* phi);
1768 CmpNode* clone_bool(PhiNode* phi);
1769
1770
1771 // Rework addressing expressions to get the most loop-invariant stuff
1772 // moved out. We'd like to do all associative operators, but it's especially
1773 // important (common) to do address expressions.
1774 Node* remix_address_expressions(Node* n);
1775 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1776
1777 // Convert add to muladd to generate MuladdS2I under certain criteria
1778 Node * convert_add_to_muladd(Node * n);
1779
1780 // Attempt to use a conditional move instead of a phi/branch
1781 Node *conditional_move( Node *n );
1782
1783 // Check for aggressive application of 'split-if' optimization,
1784 // using basic block level info.
1785 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1786 Node *split_if_with_blocks_pre ( Node *n );
1787 void split_if_with_blocks_post( Node *n );
1788 Node *has_local_phi_input( Node *n );
1789 // Mark an IfNode as being dominated by a prior test,
1790 // without actually altering the CFG (and hence IDOM info).
1791 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool prev_dom_not_imply_this = false);
1792 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool dominator_not_imply_source);
1793
1794 // Split Node 'n' through merge point
1795 RegionNode* split_thru_region(Node* n, RegionNode* region);
1796 // Split Node 'n' through merge point if there is enough win.
1797 Node *split_thru_phi( Node *n, Node *region, int policy );
1798 // Found an If getting its condition-code input from a Phi in the
1799 // same block. Split thru the Region.
1800 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1801
1802 private:
1803 // Class to keep track of wins in split_thru_phi.
1804 class SplitThruPhiWins {
1805 private:
1806 // Region containing the phi we are splitting through.
1807 const Node* _region;
1808
1809 // Sum of all wins regardless of where they happen. This applies to Loops phis as well as non-loop phis.
1810 int _total_wins;
1811
1812 // For Loops, wins have different impact depending on if they happen on loop entry or on the backedge.
1813 // Number of wins on a loop entry edge if the split is through a loop head,
1814 // otherwise 0. Entry edge wins only pay dividends once on loop entry.
1815 int _loop_entry_wins;
1816 // Number of wins on a loop back-edge, which pay dividends on every iteration.
1817 int _loop_back_wins;
1818
1819 public:
1820 SplitThruPhiWins(const Node* region) :
1821 _region(region),
1822 _total_wins(0),
1823 _loop_entry_wins(0),
1824 _loop_back_wins(0) {};
1825
1826 void reset() {_total_wins = 0; _loop_entry_wins = 0; _loop_back_wins = 0;}
1827 void add_win(int ctrl_index) {
1828 if (_region->is_Loop() && ctrl_index == LoopNode::EntryControl) {
1829 _loop_entry_wins++;
1830 } else if (_region->is_Loop() && ctrl_index == LoopNode::LoopBackControl) {
1831 _loop_back_wins++;
1832 }
1833 _total_wins++;
1834 }
1835 // Is this split profitable with respect to the policy?
1836 bool profitable(int policy) const {
1837 assert(_region->is_Loop() || (_loop_entry_wins == 0 && _loop_back_wins == 0), "wins on loop edges without a loop");
1838 assert(!_region->is_Loop() || _total_wins == _loop_entry_wins + _loop_back_wins, "missed some win");
1839 // In general this means that the split has to have more wins than specified
1840 // in the policy. However, for loops we need to take into account where the
1841 // wins happen. We need to be careful when splitting, because splitting nodes
1842 // related to the iv through the phi can sufficiently rearrange the loop
1843 // structure to prevent RCE and thus vectorization. Thus, we only deem splitting
1844 // profitable if the win of a split is not on the entry edge, as such wins
1845 // only pay off once and have a high chance of messing up the loop structure.
1846 return (_loop_entry_wins == 0 && _total_wins > policy) ||
1847 // If there are wins on the entry edge but the backadge also has sufficient wins,
1848 // there is sufficient profitability to spilt regardless of the risk of messing
1849 // up the loop structure.
1850 _loop_back_wins > policy ||
1851 // If the policy is less than 0, a split is always profitable, i.e. we always
1852 // split. This is needed when we split a node and then must also split a
1853 // dependant node, i.e. spliting a Bool node after splitting a Cmp node.
1854 policy < 0;
1855 }
1856 };
1857
1858
1859 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1860
1861 public:
1862
1863 // Conversion of fill/copy patterns into intrinsic versions
1864 bool do_intrinsify_fill();
1865 bool intrinsify_fill(IdealLoopTree* lpt);
1866 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1867 Node*& shift, Node*& offset);
1868
1869 private:
1870 // Helper functions
1871 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1872 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1873 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 );
1874 bool split_up( Node *n, Node *blk1, Node *blk2 );
1875
1876 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1877 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1878 void try_move_store_after_loop(Node* n);
1879 void move_flat_array_check_out_of_loop(Node* n);
1880 bool identical_backtoback_ifs(Node *n);
1881 bool flat_array_element_type_check(Node *n);
1882 bool can_split_if(Node *n_ctrl);
1883 bool cannot_split_division(const Node* n, const Node* region) const;
1884 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1885 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1886
1887 // Determine if a method is too big for a/another round of split-if, based on
1888 // a magic (approximate) ratio derived from the equally magic constant 35000,
1889 // previously used for this purpose (but without relating to the node limit).
1890 bool must_throttle_split_if() {
1891 uint threshold = C->max_node_limit() * 2 / 5;
1892 return C->live_nodes() > threshold;
1893 }
1894
1895 // A simplistic node request tracking mechanism, where
1896 // = UINT_MAX Request not valid or made final.
1897 // < UINT_MAX Nodes currently requested (estimate).
1898 uint _nodes_required;
1899
1900 enum { REQUIRE_MIN = 70 };
1901
1902 uint nodes_required() const { return _nodes_required; }
1903
1904 // Given the _currently_ available number of nodes, check whether there is
1905 // "room" for an additional request or not, considering the already required
1906 // number of nodes. Return TRUE if the new request is exceeding the node
1907 // budget limit, otherwise return FALSE. Note that this interpretation will
1908 // act pessimistic on additional requests when new nodes have already been
1909 // generated since the 'begin'. This behaviour fits with the intention that
1910 // node estimates/requests should be made upfront.
1911 bool exceeding_node_budget(uint required = 0) {
1912 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1913 uint available = C->max_node_limit() - C->live_nodes();
1914 return available < required + _nodes_required + REQUIRE_MIN;
1915 }
1916
1917 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1918 precond(require > 0);
1919 _nodes_required += MAX2(require, minreq);
1920 return _nodes_required;
1921 }
1922
1923 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1924 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1925 }
1926
1927 uint require_nodes_begin() {
1928 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1929 _nodes_required = 0;
1930 return C->live_nodes();
1931 }
1932
1933 // When a node request is final, optionally check that the requested number
1934 // of nodes was reasonably correct with respect to the number of new nodes
1935 // introduced since the last 'begin'. Always check that we have not exceeded
1936 // the maximum node limit.
1937 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1938 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1939
1940 #ifdef ASSERT
1941 if (check_estimate) {
1942 // Check that the node budget request was not off by too much (x2).
1943 // Should this be the case we _surely_ need to improve the estimates
1944 // used in our budget calculations.
1945 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1946 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1947 C->live_nodes() - live_at_begin, _nodes_required);
1948 }
1949 }
1950 #endif
1951 // Assert that we have stayed within the node budget limit.
1952 assert(C->live_nodes() < C->max_node_limit(),
1953 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1954 C->live_nodes() - live_at_begin, live_at_begin,
1955 C->max_node_limit(), _nodes_required);
1956
1957 _nodes_required = UINT_MAX;
1958 }
1959
1960 private:
1961
1962 bool _created_loop_node;
1963 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1964 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1965
1966 public:
1967 void set_created_loop_node() { _created_loop_node = true; }
1968 bool created_loop_node() { return _created_loop_node; }
1969 void register_new_node(Node* n, Node* blk);
1970 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1971 register_new_node(new_node, get_ctrl(ctrl_of));
1972 }
1973
1974 Node* clone_and_register(Node* n, Node* ctrl) {
1975 n = n->clone();
1976 register_new_node(n, ctrl);
1977 return n;
1978 }
1979
1980 #ifdef ASSERT
1981 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1982 #endif
1983
1984 #ifndef PRODUCT
1985 void dump() const;
1986 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1987 void dump_idom(Node* n, uint count) const;
1988 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1989 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1990 IdealLoopTree* get_loop_idx(Node* n) const {
1991 // Dead nodes have no loop, so return the top level loop instead
1992 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1993 }
1994 // Print some stats
1995 static void print_statistics();
1996 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1997 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1998 static volatile int _long_loop_candidates;
1999 static volatile int _long_loop_nests;
2000 #endif
2001
2002 #ifdef ASSERT
2003 void verify() const;
2004 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
2005 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
2006 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
2007 #endif
2008
2009 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
2010
2011 void check_counted_loop_shape(IdealLoopTree* loop, Node* head, BasicType bt) NOT_DEBUG_RETURN;
2012
2013 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
2014
2015
2016 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
2017 Node_List &range_checks);
2018
2019 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
2020 Node* inner_iters_actual_int, Node* inner_phi,
2021 Node* iv_add, LoopNode* inner_head);
2022
2023 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
2024
2025 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
2026
2027 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
2028
2029 bool would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl);
2030
2031 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
2032
2033 void try_sink_out_of_loop(Node* n);
2034
2035 Node* clamp(Node* R, Node* L, Node* H);
2036
2037 bool safe_for_if_replacement(const Node* dom) const;
2038
2039 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
2040
2041 bool try_merge_identical_ifs(Node* n);
2042
2043 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
2044
2045 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2046 IdealLoopTree* parent, bool partial);
2047
2048 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2049 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
2050
2051 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2052 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
2053 Node_List*& split_bool_set, Node_List*& split_cex_set);
2054
2055 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
2056
2057 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
2058
2059 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
2060 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
2061 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
2062 void clone_template_assertion_expression_down(Node* node);
2063
2064 Node* similar_subtype_check(const Node* x, Node* r_in);
2065
2066 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
2067
2068 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
2069
2070 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
2071
2072 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
2073
2074 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
2075
2076 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
2077
2078 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
2079
2080 ConINode* intcon(jint i);
2081
2082 ConLNode* longcon(jlong i);
2083
2084 ConNode* makecon(const Type* t);
2085
2086 ConNode* integercon(jlong l, BasicType bt);
2087
2088 ConNode* zerocon(BasicType bt);
2089 };
2090
2091 class CountedLoopConverter {
2092 friend class PhaseIdealLoop;
2093
2094 // Match increment with optional truncation
2095 class TruncatedIncrement {
2096 bool _is_valid;
2097
2098 BasicType _bt;
2099
2100 Node* _incr;
2101 Node* _outer_trunc;
2102 Node* _inner_trunc;
2103 const TypeInteger* _trunc_type;
2104
2105 public:
2106 TruncatedIncrement(BasicType bt) :
2107 _is_valid(false),
2108 _bt(bt),
2109 _incr(nullptr),
2110 _outer_trunc(nullptr),
2111 _inner_trunc(nullptr),
2112 _trunc_type(nullptr) {}
2113
2114 void build(Node* expr);
2115
2116 bool is_valid() const { return _is_valid; }
2117 Node* incr() const { return _incr; }
2118
2119 // Optional truncation for: CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2120 Node* outer_trunc() const { return _outer_trunc; } // the outermost truncating node (either the & or the final >>)
2121 Node* inner_trunc() const { return _inner_trunc; } // the inner truncating node, if applicable (the << in a <</>> pair)
2122 const TypeInteger* trunc_type() const { return _trunc_type; }
2123 };
2124
2125 class LoopStructure {
2126 bool _is_valid;
2127
2128 const Node* _head;
2129 const IdealLoopTree* _loop;
2130 PhaseIdealLoop* _phase;
2131 BasicType _iv_bt;
2132
2133 Node* _back_control;
2134 PhaseIdealLoop::LoopExitTest _exit_test;
2135 PhaseIdealLoop::LoopIVIncr _iv_incr;
2136 TruncatedIncrement _truncated_increment;
2137 PhaseIdealLoop::LoopIVStride _stride;
2138 PhiNode* _phi;
2139 SafePointNode* _safepoint;
2140
2141 public:
2142 LoopStructure(const Node* head, const IdealLoopTree* loop, PhaseIdealLoop* phase, const BasicType iv_bt) :
2143 _is_valid(false),
2144 _head(head),
2145 _loop(loop),
2146 _phase(phase),
2147 _iv_bt(iv_bt),
2148 _back_control(_phase->loop_exit_control(_loop)),
2149 _exit_test(_back_control, _loop, _phase),
2150 _iv_incr(_head, _loop),
2151 _truncated_increment(_iv_bt),
2152 _stride(PhaseIdealLoop::LoopIVStride(_iv_bt)),
2153 _phi(nullptr),
2154 _safepoint(nullptr) {}
2155
2156 void build();
2157
2158 jlong final_limit_correction() const; // compute adjusted loop limit correction
2159 bool is_infinite_loop() const;
2160
2161 bool is_valid() const { return _is_valid; }
2162
2163 Node* back_control() const { return _back_control; }
2164 PhaseIdealLoop::LoopExitTest& exit_test() { return _exit_test; }
2165 PhaseIdealLoop::LoopIVIncr& iv_incr() { return _iv_incr; }
2166 TruncatedIncrement& truncated_increment() { return _truncated_increment; }
2167 PhaseIdealLoop::LoopIVStride& stride() { return _stride; }
2168 PhiNode* phi() const { return _phi; }
2169 SafePointNode* sfpt() const { return _safepoint; }
2170 jlong stride_con() const { return _stride.compute_non_zero_stride_con(_exit_test.mask(), _iv_bt); }
2171 Node* limit() const { return _exit_test.limit(); }
2172 };
2173
2174 PhaseIdealLoop* const _phase;
2175 Node* const _head;
2176 IdealLoopTree* const _loop;
2177 const BasicType _iv_bt;
2178
2179 LoopStructure _structure;
2180 bool _should_insert_stride_overflow_limit_check = false;
2181 bool _should_insert_init_trip_limit_check = false;
2182
2183 DEBUG_ONLY(bool _checked_for_counted_loop = false;)
2184
2185 // stats for PhaseIdealLoop::print_statistics()
2186 static volatile int _long_loop_counted_loops;
2187
2188 // Return a type based on condition control flow
2189 const TypeInt* filtered_type(Node* n, Node* n_ctrl);
2190 const TypeInt* filtered_type(Node* n) { return filtered_type(n, nullptr); }
2191 // Helpers for filtered type
2192 const TypeInt* filtered_type_from_dominators(Node* val, Node* val_ctrl);
2193
2194 void insert_loop_limit_check_predicate(const ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* bol) const;
2195 void insert_stride_overflow_limit_check() const;
2196 void insert_init_trip_limit_check() const;
2197 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
2198 Node* loop_entry) const;
2199
2200 bool is_iv_overflowing(const TypeInteger* init_t, jlong stride_con, Node* phi_increment, BoolTest::mask mask) const;
2201 bool has_truncation_wrap(const TruncatedIncrement& truncation, Node* phi, jlong stride_con);
2202 SafePointNode* find_safepoint(Node* iftrue);
2203 bool is_safepoint_invalid(SafePointNode* sfpt) const;
2204
2205 public:
2206 CountedLoopConverter(PhaseIdealLoop* phase, Node* head, IdealLoopTree* loop, const BasicType iv_bt)
2207 : _phase(phase),
2208 _head(head),
2209 _loop(loop),
2210 _iv_bt(iv_bt),
2211 _structure(LoopStructure(_head, _loop, _phase, _iv_bt)) {
2212 assert(phase != nullptr, "must be"); // Fail early if mandatory parameters are null.
2213 assert(head != nullptr, "must be");
2214 assert(loop != nullptr, "must be");
2215 assert(iv_bt == T_INT || iv_bt == T_LONG, "either int or long loops");
2216 }
2217
2218 bool is_counted_loop();
2219 IdealLoopTree* convert();
2220
2221 DEBUG_ONLY(bool should_stress_long_counted_loop();)
2222 DEBUG_ONLY(bool stress_long_counted_loop();)
2223
2224 enum StrideOverflowState {
2225 Overflow = -1,
2226 NoOverflow = 0,
2227 RequireLimitCheck = 1
2228 };
2229 static StrideOverflowState check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt);
2230 };
2231
2232 class AutoNodeBudget : public StackObj
2233 {
2234 public:
2235 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
2236
2237 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
2238 : _phase(phase),
2239 _check_at_final(chk == BUDGET_CHECK),
2240 _nodes_at_begin(0)
2241 {
2242 precond(_phase != nullptr);
2243
2244 _nodes_at_begin = _phase->require_nodes_begin();
2245 }
2246
2247 ~AutoNodeBudget() {
2248 #ifndef PRODUCT
2249 if (TraceLoopOpts) {
2250 uint request = _phase->nodes_required();
2251 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
2252
2253 if (request < delta) {
2254 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
2255 } else {
2256 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
2257 // Identify the worst estimates as "poor" ones.
2258 if (request > REQUIRE_MIN && delta > 0) {
2259 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
2260 (delta <= REQUIRE_MIN && request > 10 * delta)) {
2261 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
2262 }
2263 }
2264 }
2265 }
2266 #endif // PRODUCT
2267 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
2268 }
2269
2270 private:
2271 PhaseIdealLoop* _phase;
2272 bool _check_at_final;
2273 uint _nodes_at_begin;
2274 };
2275
2276 inline Node* IdealLoopTree::tail() {
2277 // Handle lazy update of _tail field.
2278 if (_tail->in(0) == nullptr) {
2279 _tail = _phase->get_ctrl(_tail);
2280 }
2281 return _tail;
2282 }
2283
2284 inline Node* IdealLoopTree::head() {
2285 // Handle lazy update of _head field.
2286 if (_head->in(0) == nullptr) {
2287 _head = _phase->get_ctrl(_head);
2288 }
2289 return _head;
2290 }
2291
2292 // Iterate over the loop tree using a preorder, left-to-right traversal.
2293 //
2294 // Example that visits all counted loops from within PhaseIdealLoop
2295 //
2296 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2297 // IdealLoopTree* lpt = iter.current();
2298 // if (!lpt->is_counted()) continue;
2299 // ...
2300 class LoopTreeIterator : public StackObj {
2301 private:
2302 IdealLoopTree* _root;
2303 IdealLoopTree* _curnt;
2304
2305 public:
2306 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
2307
2308 bool done() { return _curnt == nullptr; } // Finished iterating?
2309
2310 void next(); // Advance to next loop tree
2311
2312 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
2313 };
2314
2315 // Compute probability of reaching some CFG node from a fixed
2316 // dominating CFG node
2317 class PathFrequency {
2318 private:
2319 Node* _dom; // frequencies are computed relative to this node
2320 Node_Stack _stack;
2321 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
2322 GrowableArray<float> _freqs; // cache frequencies
2323 PhaseIdealLoop* _phase;
2324
2325 float check_and_truncate_frequency(float f) {
2326 assert(f >= 0, "Incorrect frequency");
2327 // We do not perform an exact (f <= 1) check
2328 // this would be error prone with rounding of floats.
2329 // Performing a check like (f <= 1+eps) would be of benefit,
2330 // however, it is not evident how to determine such an eps,
2331 // given that an arbitrary number of add/mul operations
2332 // are performed on these frequencies.
2333 return (f > 1) ? 1 : f;
2334 }
2335
2336 public:
2337 PathFrequency(Node* dom, PhaseIdealLoop* phase)
2338 : _dom(dom), _stack(0), _phase(phase) {
2339 }
2340
2341 float to(Node* n);
2342 };
2343
2344 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
2345 // 1. Clone the data nodes
2346 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
2347 class DataNodeGraph : public StackObj {
2348 PhaseIdealLoop* const _phase;
2349 const Unique_Node_List& _data_nodes;
2350 OrigToNewHashtable _orig_to_new;
2351
2352 public:
2353 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
2354 : _phase(phase),
2355 _data_nodes(data_nodes),
2356 // Use 107 as best guess which is the first resize value in ResizeableHashTable::large_table_sizes.
2357 _orig_to_new(107, MaxNodeLimit)
2358 {
2359 #ifdef ASSERT
2360 for (uint i = 0; i < data_nodes.size(); i++) {
2361 assert(!data_nodes[i]->is_CFG(), "only data nodes");
2362 }
2363 #endif
2364 }
2365 NONCOPYABLE(DataNodeGraph);
2366
2367 private:
2368 void clone(Node* node, Node* new_ctrl);
2369 void clone_data_nodes(Node* new_ctrl);
2370 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2371 Node* new_ctrl);
2372 void rewire_clones_to_cloned_inputs();
2373 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
2374
2375 public:
2376 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
2377 // Set 'new_ctrl' as ctrl for the cloned nodes.
2378 const OrigToNewHashtable& clone(Node* new_ctrl) {
2379 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
2380 clone_data_nodes(new_ctrl);
2381 rewire_clones_to_cloned_inputs();
2382 return _orig_to_new;
2383 }
2384
2385 // Create a copy of the data nodes provided to the constructor by doing the following:
2386 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
2387 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2388 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2389 // `new_ctrl` as ctrl.
2390 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2391 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2392 Node* new_ctrl) {
2393 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2394 rewire_clones_to_cloned_inputs();
2395 return _orig_to_new;
2396 }
2397 };
2398 #endif // SHARE_OPTO_LOOPNODE_HPP