1 /*
2 * Copyright (c) 1998, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_OPTO_LOOPNODE_HPP
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class LoopSelector;
47 class UnswitchedLoopSelector;
48 class VectorSet;
49 class VSharedData;
50 class Invariance;
51 struct small_cache;
52
53 //
54 // I D E A L I Z E D L O O P S
55 //
56 // Idealized loops are the set of loops I perform more interesting
57 // transformations on, beyond simple hoisting.
58
59 //------------------------------LoopNode---------------------------------------
60 // Simple loop header. Fall in path on left, loop-back path on right.
61 class LoopNode : public RegionNode {
62 // Size is bigger to hold the flags. However, the flags do not change
63 // the semantics so it does not appear in the hash & cmp functions.
64 virtual uint size_of() const { return sizeof(*this); }
65 protected:
66 uint _loop_flags;
67 // Names for flag bitfields
68 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
69 MainHasNoPreLoop = 1<<2,
70 HasExactTripCount = 1<<3,
71 InnerLoop = 1<<4,
72 PartialPeelLoop = 1<<5,
73 PartialPeelFailed = 1<<6,
74 WasSlpAnalyzed = 1<<7,
75 PassedSlpAnalysis = 1<<8,
76 DoUnrollOnly = 1<<9,
77 VectorizedLoop = 1<<10,
78 HasAtomicPostLoop = 1<<11,
79 StripMined = 1<<12,
80 SubwordLoop = 1<<13,
81 ProfileTripFailed = 1<<14,
82 LoopNestInnerLoop = 1<<15,
83 LoopNestLongOuterLoop = 1<<16,
84 MultiversionFastLoop = 1<<17,
85 MultiversionSlowLoop = 2<<17,
86 MultiversionDelayedSlowLoop = 3<<17,
87 MultiversionFlagsMask = 3<<17,
88 };
89 char _unswitch_count;
90 enum { _unswitch_max=3 };
91
92 // Expected trip count from profile data
93 float _profile_trip_cnt;
94
95 public:
96 // Names for edge indices
97 enum { Self=0, EntryControl, LoopBackControl };
98
99 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
100 void set_inner_loop() { _loop_flags |= InnerLoop; }
101
102 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
103 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
104 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
105 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
106 bool is_strip_mined() const { return _loop_flags & StripMined; }
107 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
108 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
109 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
110 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
111
112 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
113 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
114 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
115 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
116 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
117 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
118 void mark_strip_mined() { _loop_flags |= StripMined; }
119 void clear_strip_mined() { _loop_flags &= ~StripMined; }
120 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
121 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
122 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
123 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
124
125 int unswitch_max() { return _unswitch_max; }
126 int unswitch_count() { return _unswitch_count; }
127
128 void set_unswitch_count(int val) {
129 assert (val <= unswitch_max(), "too many unswitches");
130 _unswitch_count = val;
131 }
132
133 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
134 float profile_trip_cnt() { return _profile_trip_cnt; }
135
136 #ifndef PRODUCT
137 uint _stress_peeling_attempts = 0;
138 #endif
139
140 LoopNode(Node *entry, Node *backedge)
141 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
142 _profile_trip_cnt(COUNT_UNKNOWN) {
143 init_class_id(Class_Loop);
144 init_req(EntryControl, entry);
145 init_req(LoopBackControl, backedge);
146 }
147
148 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
149 virtual int Opcode() const;
150 bool can_be_counted_loop(PhaseValues* phase) const {
151 return req() == 3 && in(0) != nullptr &&
152 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
153 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
154 }
155 bool is_valid_counted_loop(BasicType bt) const;
156 #ifndef PRODUCT
157 virtual void dump_spec(outputStream *st) const;
158 #endif
159
160 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
161 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
162 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
163 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
164 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
165 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
166 };
167
168 //------------------------------Counted Loops----------------------------------
169 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
170 // path (and maybe some other exit paths). The trip-counter exit is always
171 // last in the loop. The trip-counter have to stride by a constant;
172 // the exit value is also loop invariant.
173
174 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
175 // CountedLoopNode has the incoming loop control and the loop-back-control
176 // which is always the IfTrue before the matching CountedLoopEndNode. The
177 // CountedLoopEndNode has an incoming control (possibly not the
178 // CountedLoopNode if there is control flow in the loop), the post-increment
179 // trip-counter value, and the limit. The trip-counter value is always of
180 // the form (Op old-trip-counter stride). The old-trip-counter is produced
181 // by a Phi connected to the CountedLoopNode. The stride is constant.
182 // The Op is any commutable opcode, including Add, Mul, Xor. The
183 // CountedLoopEndNode also takes in the loop-invariant limit value.
184
185 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
186 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
187 // via the old-trip-counter from the Op node.
188
189 //------------------------------CountedLoopNode--------------------------------
190 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
191 // inputs the incoming loop-start control and the loop-back control, so they
192 // act like RegionNodes. They also take in the initial trip counter, the
193 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
194 // produce a loop-body control and the trip counter value. Since
195 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
196
197 class BaseCountedLoopNode : public LoopNode {
198 public:
199 BaseCountedLoopNode(Node *entry, Node *backedge)
200 : LoopNode(entry, backedge) {
201 }
202
203 Node *init_control() const { return in(EntryControl); }
204 Node *back_control() const { return in(LoopBackControl); }
205
206 Node* init_trip() const;
207 Node* stride() const;
208 bool stride_is_con() const;
209 Node* limit() const;
210 Node* incr() const;
211 Node* phi() const;
212
213 BaseCountedLoopEndNode* loopexit_or_null() const;
214 BaseCountedLoopEndNode* loopexit() const;
215
216 virtual BasicType bt() const = 0;
217
218 jlong stride_con() const;
219
220 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
221
222 virtual void set_trip_count(julong tc) = 0;
223 virtual julong trip_count() const = 0;
224
225 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
226 void set_exact_trip_count(julong tc) {
227 set_trip_count(tc);
228 _loop_flags |= HasExactTripCount;
229 }
230 void set_nonexact_trip_count() {
231 _loop_flags &= ~HasExactTripCount;
232 }
233 };
234
235
236 class CountedLoopNode : public BaseCountedLoopNode {
237 // Size is bigger to hold _main_idx. However, _main_idx does not change
238 // the semantics so it does not appear in the hash & cmp functions.
239 virtual uint size_of() const { return sizeof(*this); }
240
241 // For Pre- and Post-loops during debugging ONLY, this holds the index of
242 // the Main CountedLoop. Used to assert that we understand the graph shape.
243 node_idx_t _main_idx;
244
245 // Known trip count calculated by compute_exact_trip_count()
246 uint _trip_count;
247
248 // Log2 of original loop bodies in unrolled loop
249 int _unrolled_count_log2;
250
251 // Node count prior to last unrolling - used to decide if
252 // unroll,optimize,unroll,optimize,... is making progress
253 int _node_count_before_unroll;
254
255 // If slp analysis is performed we record the maximum
256 // vector mapped unroll factor here
257 int _slp_maximum_unroll_factor;
258
259 public:
260 CountedLoopNode(Node *entry, Node *backedge)
261 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
262 _unrolled_count_log2(0), _node_count_before_unroll(0),
263 _slp_maximum_unroll_factor(0) {
264 init_class_id(Class_CountedLoop);
265 // Initialize _trip_count to the largest possible value.
266 // Will be reset (lower) if the loop's trip count is known.
267 }
268
269 virtual int Opcode() const;
270 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
271
272 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
273 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
274 int stride_con() const;
275
276 // Match increment with optional truncation
277 static Node*
278 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
279 BasicType bt);
280
281 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
282 // can run short a few iterations and may start a few iterations in.
283 // It will be RCE'd and unrolled and aligned.
284
285 // A following 'post' loop will run any remaining iterations. Used
286 // during Range Check Elimination, the 'post' loop will do any final
287 // iterations with full checks. Also used by Loop Unrolling, where
288 // the 'post' loop will do any epilog iterations needed. Basically,
289 // a 'post' loop can not profitably be further unrolled or RCE'd.
290
291 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
292 // it may do under-flow checks for RCE and may do alignment iterations
293 // so the following main loop 'knows' that it is striding down cache
294 // lines.
295
296 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
297 // Aligned, may be missing it's pre-loop.
298 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
299 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
300 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
301 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
302 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
303 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
304 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
305 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
306 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
307 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
308
309 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
310
311 int main_idx() const { return _main_idx; }
312
313 void set_trip_count(julong tc) {
314 assert(tc < max_juint, "Cannot set trip count to max_juint");
315 _trip_count = checked_cast<uint>(tc);
316 }
317 julong trip_count() const { return _trip_count; }
318
319 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
320 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
321 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
322 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
323
324 void set_notpassed_slp() {
325 _loop_flags &= ~PassedSlpAnalysis;
326 }
327
328 void double_unrolled_count() { _unrolled_count_log2++; }
329 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
330
331 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
332 int node_count_before_unroll() { return _node_count_before_unroll; }
333 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
334 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
335
336 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
337 // decides which one is taken:
338 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
339 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
340 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
341 // assumption fails we take the slow_loop instead, where we do not make the same speculative
342 // assumption.
343 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
344 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
345 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
346 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
347 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
348 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
349 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
350 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
351 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
352 // Hence, we expect the slow_loop to execute slower.
353 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
354 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
355 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
356 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
357 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
358 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
359 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
360 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
361
362 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
363 OuterStripMinedLoopNode* outer_loop() const;
364 virtual IfTrueNode* outer_loop_tail() const;
365 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
366 virtual IfFalseNode* outer_loop_exit() const;
367 virtual SafePointNode* outer_safepoint() const;
368
369 Node* skip_assertion_predicates_with_halt();
370
371 virtual BasicType bt() const {
372 return T_INT;
373 }
374
375 Node* is_canonical_loop_entry();
376 CountedLoopEndNode* find_pre_loop_end();
377
378 Node* uncasted_init_trip(bool uncasted);
379
380 #ifndef PRODUCT
381 virtual void dump_spec(outputStream *st) const;
382 #endif
383 };
384
385 class LongCountedLoopNode : public BaseCountedLoopNode {
386 private:
387 virtual uint size_of() const { return sizeof(*this); }
388
389 // Known trip count calculated by compute_exact_trip_count()
390 julong _trip_count;
391
392 public:
393 LongCountedLoopNode(Node *entry, Node *backedge)
394 : BaseCountedLoopNode(entry, backedge), _trip_count(max_julong) {
395 init_class_id(Class_LongCountedLoop);
396 }
397
398 virtual int Opcode() const;
399
400 virtual BasicType bt() const {
401 return T_LONG;
402 }
403
404 void set_trip_count(julong tc) {
405 assert(tc < max_julong, "Cannot set trip count to max_julong");
406 _trip_count = tc;
407 }
408 julong trip_count() const { return _trip_count; }
409
410 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
411 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
412 };
413
414
415 //------------------------------CountedLoopEndNode-----------------------------
416 // CountedLoopEndNodes end simple trip counted loops. They act much like
417 // IfNodes.
418
419 class BaseCountedLoopEndNode : public IfNode {
420 public:
421 enum { TestControl, TestValue };
422 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
423 : IfNode(control, test, prob, cnt) {
424 init_class_id(Class_BaseCountedLoopEnd);
425 }
426
427 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
428 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
429 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
430 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
431 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
432 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
433
434 PhiNode* phi() const {
435 Node* tmp = incr();
436 if (tmp && tmp->req() == 3) {
437 Node* phi = tmp->in(1);
438 if (phi->is_Phi()) {
439 return phi->as_Phi();
440 }
441 }
442 return nullptr;
443 }
444
445 BaseCountedLoopNode* loopnode() const {
446 // The CountedLoopNode that goes with this CountedLoopEndNode may
447 // have been optimized out by the IGVN so be cautious with the
448 // pattern matching on the graph
449 PhiNode* iv_phi = phi();
450 if (iv_phi == nullptr) {
451 return nullptr;
452 }
453 Node* ln = iv_phi->in(0);
454 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
455 return nullptr;
456 }
457 if (ln->as_BaseCountedLoop()->bt() != bt()) {
458 return nullptr;
459 }
460 return ln->as_BaseCountedLoop();
461 }
462
463 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
464
465 jlong stride_con() const;
466 virtual BasicType bt() const = 0;
467
468 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
469 };
470
471 class CountedLoopEndNode : public BaseCountedLoopEndNode {
472 public:
473
474 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
475 : BaseCountedLoopEndNode(control, test, prob, cnt) {
476 init_class_id(Class_CountedLoopEnd);
477 }
478 virtual int Opcode() const;
479
480 CountedLoopNode* loopnode() const {
481 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
482 }
483
484 virtual BasicType bt() const {
485 return T_INT;
486 }
487
488 #ifndef PRODUCT
489 virtual void dump_spec(outputStream *st) const;
490 #endif
491 };
492
493 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
494 public:
495 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
496 : BaseCountedLoopEndNode(control, test, prob, cnt) {
497 init_class_id(Class_LongCountedLoopEnd);
498 }
499
500 LongCountedLoopNode* loopnode() const {
501 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
502 }
503
504 virtual int Opcode() const;
505
506 virtual BasicType bt() const {
507 return T_LONG;
508 }
509 };
510
511
512 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
513 Node* bctrl = back_control();
514 if (bctrl == nullptr) return nullptr;
515
516 Node* lexit = bctrl->in(0);
517 if (!lexit->is_BaseCountedLoopEnd()) {
518 return nullptr;
519 }
520 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
521 if (result->bt() != bt()) {
522 return nullptr;
523 }
524 return result;
525 }
526
527 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
528 BaseCountedLoopEndNode* cle = loopexit_or_null();
529 assert(cle != nullptr, "loopexit is null");
530 return cle;
531 }
532
533 inline Node* BaseCountedLoopNode::init_trip() const {
534 BaseCountedLoopEndNode* cle = loopexit_or_null();
535 return cle != nullptr ? cle->init_trip() : nullptr;
536 }
537 inline Node* BaseCountedLoopNode::stride() const {
538 BaseCountedLoopEndNode* cle = loopexit_or_null();
539 return cle != nullptr ? cle->stride() : nullptr;
540 }
541
542 inline bool BaseCountedLoopNode::stride_is_con() const {
543 BaseCountedLoopEndNode* cle = loopexit_or_null();
544 return cle != nullptr && cle->stride_is_con();
545 }
546 inline Node* BaseCountedLoopNode::limit() const {
547 BaseCountedLoopEndNode* cle = loopexit_or_null();
548 return cle != nullptr ? cle->limit() : nullptr;
549 }
550 inline Node* BaseCountedLoopNode::incr() const {
551 BaseCountedLoopEndNode* cle = loopexit_or_null();
552 return cle != nullptr ? cle->incr() : nullptr;
553 }
554 inline Node* BaseCountedLoopNode::phi() const {
555 BaseCountedLoopEndNode* cle = loopexit_or_null();
556 return cle != nullptr ? cle->phi() : nullptr;
557 }
558
559 inline jlong BaseCountedLoopNode::stride_con() const {
560 BaseCountedLoopEndNode* cle = loopexit_or_null();
561 return cle != nullptr ? cle->stride_con() : 0;
562 }
563
564
565 //------------------------------LoopLimitNode-----------------------------
566 // Counted Loop limit node which represents exact final iterator value:
567 // trip_count = (limit - init_trip + stride - 1)/stride
568 // final_value= trip_count * stride + init_trip.
569 // Use HW instructions to calculate it when it can overflow in integer.
570 // Note, final_value should fit into integer since counted loop has
571 // limit check: limit <= max_int-stride.
572 class LoopLimitNode : public Node {
573 enum { Init=1, Limit=2, Stride=3 };
574 public:
575 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
576 // Put it on the Macro nodes list to optimize during macro nodes expansion.
577 init_flags(Flag_is_macro);
578 C->add_macro_node(this);
579 }
580 virtual int Opcode() const;
581 virtual const Type *bottom_type() const { return TypeInt::INT; }
582 virtual uint ideal_reg() const { return Op_RegI; }
583 virtual const Type* Value(PhaseGVN* phase) const;
584 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
585 virtual Node* Identity(PhaseGVN* phase);
586 };
587
588 // Support for strip mining
589 class OuterStripMinedLoopNode : public LoopNode {
590 private:
591 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const;
592 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn);
593
594 public:
595 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
596 : LoopNode(entry, backedge) {
597 init_class_id(Class_OuterStripMinedLoop);
598 init_flags(Flag_is_macro);
599 C->add_macro_node(this);
600 }
601
602 virtual int Opcode() const;
603
604 virtual IfTrueNode* outer_loop_tail() const;
605 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
606 virtual IfFalseNode* outer_loop_exit() const;
607 virtual SafePointNode* outer_safepoint() const;
608 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); }
609 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); }
610 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->false_proj(); }
611
612 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
613
614 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
615
616 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
617
618 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
619
620 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
621 PhaseIdealLoop* iloop);
622 };
623
624 class OuterStripMinedLoopEndNode : public IfNode {
625 public:
626 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
627 : IfNode(control, test, prob, cnt) {
628 init_class_id(Class_OuterStripMinedLoopEnd);
629 }
630
631 virtual int Opcode() const;
632
633 virtual const Type* Value(PhaseGVN* phase) const;
634 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
635
636 bool is_expanded(PhaseGVN *phase) const;
637 };
638
639 // -----------------------------IdealLoopTree----------------------------------
640 class IdealLoopTree : public ResourceObj {
641 public:
642 IdealLoopTree *_parent; // Parent in loop tree
643 IdealLoopTree *_next; // Next sibling in loop tree
644 IdealLoopTree *_child; // First child in loop tree
645
646 // The head-tail backedge defines the loop.
647 // If a loop has multiple backedges, this is addressed during cleanup where
648 // we peel off the multiple backedges, merging all edges at the bottom and
649 // ensuring that one proper backedge flow into the loop.
650 Node *_head; // Head of loop
651 Node *_tail; // Tail of loop
652 inline Node *tail(); // Handle lazy update of _tail field
653 inline Node *head(); // Handle lazy update of _head field
654 PhaseIdealLoop* _phase;
655 int _local_loop_unroll_limit;
656 int _local_loop_unroll_factor;
657
658 Node_List _body; // Loop body for inner loops
659
660 uint16_t _nest; // Nesting depth
661 uint8_t _irreducible:1, // True if irreducible
662 _has_call:1, // True if has call safepoint
663 _has_sfpt:1, // True if has non-call safepoint
664 _rce_candidate:1, // True if candidate for range check elimination
665 _has_range_checks:1,
666 _has_range_checks_computed:1;
667
668 Node_List* _safepts; // List of safepoints in this loop
669 Node_List* _required_safept; // A inner loop cannot delete these safepts;
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 // Remove simplistic dead code from loop body
729 void DCE_loop_body();
730
731 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
732 // Replace with a 1-in-10 exit guess.
733 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
734
735 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
736 // Useful for unrolling loops with NO array accesses.
737 bool policy_peel_only( PhaseIdealLoop *phase ) const;
738
739 // Return TRUE or FALSE if the loop should be unswitched -- clone
740 // loop with an invariant test
741 bool policy_unswitching( PhaseIdealLoop *phase ) const;
742
743 // Micro-benchmark spamming. Remove empty loops.
744 bool do_remove_empty_loop( PhaseIdealLoop *phase );
745
746 // Convert one-iteration loop into normal code.
747 bool do_one_iteration_loop( PhaseIdealLoop *phase );
748
749 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
750 // move some loop-invariant test (usually a null-check) before the loop.
751 bool policy_peeling(PhaseIdealLoop *phase);
752
753 uint estimate_peeling(PhaseIdealLoop *phase);
754
755 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
756 // known trip count in the counted loop node.
757 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
758
759 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
760 // if the loop is a counted loop and the loop body is small enough.
761 bool policy_unroll(PhaseIdealLoop *phase);
762
763 // Loop analyses to map to a maximal superword unrolling for vectorization.
764 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
765
766 // Return TRUE or FALSE if the loop should be range-check-eliminated.
767 // Gather a list of IF tests that are dominated by iteration splitting;
768 // also gather the end of the first split and the start of the 2nd split.
769 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
770
771 // Return TRUE if "iff" is a range check.
772 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
773 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
774 jlong& scale) const;
775
776 // Estimate the number of nodes required when cloning a loop (body).
777 uint est_loop_clone_sz(uint factor) const;
778 // Estimate the number of nodes required when unrolling a loop (body).
779 uint est_loop_unroll_sz(uint factor) const;
780
781 // Compute loop trip count if possible
782 void compute_trip_count(PhaseIdealLoop* phase, BasicType bt);
783
784 // Compute loop trip count from profile data
785 float compute_profile_trip_cnt_helper(Node* n);
786 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
787
788 // Reassociate invariant expressions.
789 void reassociate_invariants(PhaseIdealLoop *phase);
790 // Reassociate invariant binary expressions.
791 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
792 // Reassociate invariant add, subtract, and compare expressions.
793 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
794 // Return nonzero index of invariant operand if invariant and variant
795 // are combined with an associative binary. Helper for reassociate_invariants.
796 int find_invariant(Node* n, PhaseIdealLoop *phase);
797 // Return TRUE if "n" is associative.
798 bool is_associative(Node* n, Node* base=nullptr);
799 // Return TRUE if "n" is an associative cmp node.
800 bool is_associative_cmp(Node* n);
801
802 // Return true if n is invariant
803 bool is_invariant(Node* n) const;
804
805 // Put loop body on igvn work list
806 void record_for_igvn();
807
808 bool is_root() { return _parent == nullptr; }
809 // A proper/reducible loop w/o any (occasional) dead back-edge.
810 bool is_loop() { return !_irreducible && !tail()->is_top(); }
811 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
812 bool is_innermost() { return is_loop() && _child == nullptr; }
813
814 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
815
816 bool compute_has_range_checks() const;
817 bool range_checks_present() {
818 if (!_has_range_checks_computed) {
819 if (compute_has_range_checks()) {
820 _has_range_checks = 1;
821 }
822 _has_range_checks_computed = 1;
823 }
824 return _has_range_checks;
825 }
826
827 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
828 // In all other cases, return this.
829 IdealLoopTree* skip_strip_mined() {
830 return _head->as_Loop()->is_strip_mined() ? _parent : this;
831 }
832
833 #ifndef PRODUCT
834 void dump_head(); // Dump loop head only
835 void dump(); // Dump this loop recursively
836 #endif
837
838 #ifdef ASSERT
839 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
840 bool verify_tree(IdealLoopTree* loop_verify) const;
841 #endif
842
843 private:
844 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
845
846 // Estimate the number of nodes resulting from control and data flow merge.
847 uint est_loop_flow_merge_sz() const;
848
849 // Check if the number of residual iterations is large with unroll_cnt.
850 // Return true if the residual iterations are more than 10% of the trip count.
851 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
852 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
853 }
854
855 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
856
857 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
858
859 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
860
861 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
862 Node* sfpt) const;
863
864 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
865
866 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
867 };
868
869 // -----------------------------PhaseIdealLoop---------------------------------
870 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
871 // into a loop tree. Drives the loop-based transformations on the ideal graph.
872 class PhaseIdealLoop : public PhaseTransform {
873 friend class IdealLoopTree;
874 friend class SuperWord;
875 friend class ShenandoahBarrierC2Support;
876 friend class AutoNodeBudget;
877
878 Arena _arena; // For data whose lifetime is a single pass of loop optimizations
879
880 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
881 //
882 // Exception: dead CFG nodes may instead have a ctrl/idom forwarding
883 // installed. See: forward_ctrl
884 Node_List _loop_or_ctrl;
885
886 // Pre-computed def-use info
887 PhaseIterGVN &_igvn;
888
889 // Head of loop tree
890 IdealLoopTree* _ltree_root;
891
892 // Array of pre-order numbers, plus post-visited bit.
893 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
894 // ODD for post-visited. Other bits are the pre-order number.
895 uint *_preorders;
896 uint _max_preorder;
897
898 ReallocMark _nesting; // Safety checks for arena reallocation
899
900 const PhaseIdealLoop* _verify_me;
901 bool _verify_only;
902
903 // Allocate _preorders[] array
904 void allocate_preorders() {
905 _max_preorder = C->unique()+8;
906 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
907 memset(_preorders, 0, sizeof(uint) * _max_preorder);
908 }
909
910 // Allocate _preorders[] array
911 void reallocate_preorders() {
912 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
913 if ( _max_preorder < C->unique() ) {
914 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
915 _max_preorder = C->unique();
916 }
917 memset(_preorders, 0, sizeof(uint) * _max_preorder);
918 }
919
920 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
921 // adds new nodes.
922 void check_grow_preorders( ) {
923 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
924 if ( _max_preorder < C->unique() ) {
925 uint newsize = _max_preorder<<1; // double size of array
926 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
927 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
928 _max_preorder = newsize;
929 }
930 }
931 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
932 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
933 // Pre-order numbers are written to the Nodes array as low-bit-set values.
934 void set_preorder_visited( Node *n, int pre_order ) {
935 assert( !is_visited( n ), "already set" );
936 _preorders[n->_idx] = (pre_order<<1);
937 };
938 // Return pre-order number.
939 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
940
941 // Check for being post-visited.
942 // Should be previsited already (checked with assert(is_visited(n))).
943 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
944
945 // Mark as post visited
946 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
947
948 public:
949 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
950 // Returns true if "n" is a data node, false if it's a CFG node.
951 //
952 // Exception:
953 // control nodes that are dead because of "replace_node_and_forward_ctrl"
954 // or have otherwise modified their ctrl state by "forward_ctrl".
955 // They return "true", because they have a ctrl "forwarding" to the other ctrl node they
956 // were replaced with.
957 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
958
959 private:
960 // clear out dead code after build_loop_late
961 Node_List _deadlist;
962 Node_List _zero_trip_guard_opaque_nodes;
963 Node_List _multiversion_opaque_nodes;
964
965 // Support for faster execution of get_late_ctrl()/dom_lca()
966 // when a node has many uses and dominator depth is deep.
967 GrowableArray<jlong> _dom_lca_tags;
968 uint _dom_lca_tags_round;
969 void init_dom_lca_tags();
970
971 // Helper for debugging bad dominance relationships
972 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
973
974 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
975
976 // Inline wrapper for frequent cases:
977 // 1) only one use
978 // 2) a use is the same as the current LCA passed as 'n1'
979 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
980 assert( n->is_CFG(), "" );
981 // Fast-path null lca
982 if( lca != nullptr && lca != n ) {
983 assert( lca->is_CFG(), "" );
984 // find LCA of all uses
985 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
986 }
987 return find_non_split_ctrl(n);
988 }
989 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
990
991 // Helper function for directing control inputs away from CFG split points.
992 Node *find_non_split_ctrl( Node *ctrl ) const {
993 if (ctrl != nullptr) {
994 if (ctrl->is_MultiBranch()) {
995 ctrl = ctrl->in(0);
996 }
997 assert(ctrl->is_CFG(), "CFG");
998 }
999 return ctrl;
1000 }
1001
1002 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
1003
1004 #ifdef ASSERT
1005 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
1006 #endif
1007 private:
1008 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
1009 Unique_Node_List& list);
1010 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
1011 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1012 CountedLoopNode* remaining_loop_head,
1013 uint first_node_index_in_cloned_loop_body,
1014 const Node_List& old_new);
1015 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1016 CountedLoopNode* main_loop_head,
1017 uint first_node_index_in_pre_loop_body,
1018 uint last_node_index_in_pre_loop_body,
1019 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1020 const Node_List& old_new);
1021 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1022 uint first_node_index_in_cloned_loop_body);
1023 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1024 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1025 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1026 CountedLoopNode* target_loop_head,
1027 const NodeInLoopBody& _node_in_loop_body,
1028 bool kill_old_template);
1029 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1030 const Node* old_target_loop_entry,
1031 uint node_index_before_new_assertion_predicate_nodes);
1032 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
1033 Node* bol);
1034 void log_loop_tree();
1035
1036 public:
1037
1038 PhaseIterGVN &igvn() const { return _igvn; }
1039
1040 Arena* arena() { return &_arena; };
1041
1042 bool has_node(const Node* n) const {
1043 guarantee(n != nullptr, "No Node.");
1044 return _loop_or_ctrl[n->_idx] != nullptr;
1045 }
1046 // check if transform created new nodes that need _ctrl recorded
1047 Node *get_late_ctrl( Node *n, Node *early );
1048 Node *get_early_ctrl( Node *n );
1049 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1050 void set_early_ctrl(Node* n, bool update_body);
1051 void set_subtree_ctrl(Node* n, bool update_body);
1052 void set_ctrl( Node *n, Node *ctrl ) {
1053 assert( !has_node(n) || has_ctrl(n), "" );
1054 assert( ctrl->in(0), "cannot set dead control node" );
1055 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1056 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1057 }
1058 void set_root_as_ctrl(Node* n) {
1059 assert(!has_node(n) || has_ctrl(n), "");
1060 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1061 }
1062 // Set control and update loop membership
1063 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1064 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1065 IdealLoopTree* new_loop = get_loop(ctrl);
1066 if (old_loop != new_loop) {
1067 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1068 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1069 }
1070 set_ctrl(n, ctrl);
1071 }
1072
1073 // Retrieves the ctrl for a data node i.
1074 Node* get_ctrl(const Node* i) {
1075 assert(has_node(i) && has_ctrl(i), "must be data node with ctrl");
1076 Node* n = get_ctrl_no_update(i);
1077 // We store the found ctrl in the side-table again. In most cases,
1078 // this is a no-op, since we just read from _loop_or_ctrl. But in cases
1079 // where there was a ctrl forwarding via dead ctrl nodes, this shortens the path.
1080 // See: forward_ctrl
1081 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1082 assert(has_node(i) && has_ctrl(i), "must still be data node with ctrl");
1083 assert(n == find_non_split_ctrl(n), "must return legal ctrl");
1084 return n;
1085 }
1086
1087 bool is_dominator(Node* dominator, Node* n);
1088 bool is_strict_dominator(Node* dominator, Node* n);
1089
1090 // return get_ctrl for a data node and self(n) for a CFG node
1091 Node* ctrl_or_self(Node* n) {
1092 if (has_ctrl(n))
1093 return get_ctrl(n);
1094 else {
1095 assert (n->is_CFG(), "must be a CFG node");
1096 return n;
1097 }
1098 }
1099
1100 private:
1101 Node* get_ctrl_no_update_helper(const Node* i) const {
1102 // We expect only data nodes (which must have a ctrl set), or
1103 // dead ctrl nodes that have a ctrl "forwarding".
1104 // See: forward_ctrl.
1105 assert(has_ctrl(i), "only data nodes or ctrl nodes with ctrl forwarding expected");
1106 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1107 }
1108
1109 // Compute the ctrl of node i, jumping over ctrl forwardings.
1110 Node* get_ctrl_no_update(const Node* i) const {
1111 assert(has_ctrl(i), "only data nodes expected");
1112 Node* n = get_ctrl_no_update_helper(i);
1113 if (n->in(0) == nullptr) {
1114 // We encountered a dead CFG node.
1115 // If everything went right, this dead CFG node should have had a ctrl
1116 // forwarding installed, using "forward_ctrl". We now have to jump from
1117 // the old (dead) ctrl node to the new (live) ctrl node, in possibly
1118 // multiple ctrl forwarding steps.
1119 do {
1120 n = get_ctrl_no_update_helper(n);
1121 } while (n->in(0) == nullptr);
1122 n = find_non_split_ctrl(n);
1123 }
1124 return n;
1125 }
1126
1127 public:
1128 // Check for loop being set
1129 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1130 bool has_loop( Node *n ) const {
1131 assert(!has_node(n) || !has_ctrl(n), "");
1132 return has_node(n);
1133 }
1134 // Set loop
1135 void set_loop( Node *n, IdealLoopTree *loop ) {
1136 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1137 }
1138
1139 // Install a ctrl "forwarding" from an old (dead) control node.
1140 // This is a "lazy" update of the "get_ctrl" and "idom" mechanism:
1141 // - Install a forwarding from old_node (dead ctrl) to new_node.
1142 // - When querying "get_ctrl": jump from data node over possibly
1143 // multiple dead ctrl nodes with ctrl forwarding to eventually
1144 // reach a live ctrl node. Shorten the path to avoid chasing the
1145 // forwarding in the future.
1146 // - When querying "idom": from some node get its old idom, which
1147 // may be dead but has an idom forwarding to the new and live
1148 // idom. Shorten the path to avoid chasing the forwarding in the
1149 // future.
1150 // Note: while the "idom" information is stored in the "_idom"
1151 // side-table, the idom forwarding piggybacks on the ctrl
1152 // forwarding on "_loop_or_ctrl".
1153 // Using "forward_ctrl" allows us to only edit the entry for the old
1154 // dead node now, and we do not have to update all the nodes that had
1155 // the old_node as their "get_ctrl" or "idom". We clean up the forwarding
1156 // links when we query "get_ctrl" or "idom" for these nodes the next time.
1157 void forward_ctrl(Node* old_node, Node* new_node) {
1158 assert(!has_ctrl(old_node) && old_node->is_CFG() && old_node->in(0) == nullptr,
1159 "must be dead ctrl (CFG) node");
1160 assert(!has_ctrl(new_node) && new_node->is_CFG() && new_node->in(0) != nullptr,
1161 "must be live ctrl (CFG) node");
1162 assert(old_node != new_node, "no cycles please");
1163 // Re-use the side array slot for this node to provide the
1164 // forwarding pointer.
1165 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1166 assert(has_ctrl(old_node), "must have installed ctrl forwarding");
1167 }
1168
1169 // Replace the old ctrl node with a new ctrl node.
1170 // - Update the node inputs of all uses.
1171 // - Lazily update the ctrl and idom info of all uses, via a ctrl/idom forwarding.
1172 void replace_node_and_forward_ctrl(Node* old_node, Node* new_node) {
1173 _igvn.replace_node(old_node, new_node);
1174 forward_ctrl(old_node, new_node);
1175 }
1176
1177 private:
1178
1179 // Place 'n' in some loop nest, where 'n' is a CFG node
1180 void build_loop_tree();
1181 int build_loop_tree_impl(Node* n, int pre_order);
1182 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1183 // loop tree, not the root.
1184 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1185
1186 #ifdef ASSERT
1187 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1188 void verify_regions_in_irreducible_loops();
1189 bool is_in_irreducible_loop(RegionNode* region);
1190 #endif
1191
1192 // Place Data nodes in some loop nest
1193 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1194 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1195 void build_loop_late_post_work(Node* n, bool pinned);
1196 void build_loop_late_post(Node* n);
1197 void verify_strip_mined_scheduling(Node *n, Node* least);
1198
1199 // Array of immediate dominance info for each CFG node indexed by node idx
1200 private:
1201 uint _idom_size;
1202 Node **_idom; // Array of immediate dominators
1203 uint *_dom_depth; // Used for fast LCA test
1204 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1205 LoopOptsMode _mode;
1206
1207 // build the loop tree and perform any requested optimizations
1208 void build_and_optimize();
1209
1210 // Dominators for the sea of nodes
1211 void Dominators();
1212
1213 // Compute the Ideal Node to Loop mapping
1214 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1215 PhaseTransform(Ideal_Loop),
1216 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1217 _loop_or_ctrl(&_arena),
1218 _igvn(igvn),
1219 _verify_me(nullptr),
1220 _verify_only(false),
1221 _mode(mode),
1222 _nodes_required(UINT_MAX) {
1223 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1224 build_and_optimize();
1225 }
1226
1227 #ifndef PRODUCT
1228 // Verify that verify_me made the same decisions as a fresh run
1229 // or only verify that the graph is valid if verify_me is null.
1230 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1231 PhaseTransform(Ideal_Loop),
1232 _arena(mtCompiler, Arena::Tag::tag_idealloop),
1233 _loop_or_ctrl(&_arena),
1234 _igvn(igvn),
1235 _verify_me(verify_me),
1236 _verify_only(verify_me == nullptr),
1237 _mode(LoopOptsVerify),
1238 _nodes_required(UINT_MAX) {
1239 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1240 build_and_optimize();
1241 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1242 }
1243 #endif
1244
1245 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1246
1247 Node* idom_no_update(Node* d) const {
1248 return idom_no_update(d->_idx);
1249 }
1250
1251 Node* idom_no_update(uint node_idx) const {
1252 assert(node_idx < _idom_size, "oob");
1253 Node* n = _idom[node_idx];
1254 assert(n != nullptr,"Bad immediate dominator info.");
1255 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1256 // We encountered a dead CFG node.
1257 // If everything went right, this dead CFG node should have had an idom
1258 // forwarding installed, using "forward_ctrl". We now have to jump from
1259 // the old (dead) idom node to the new (live) idom node, in possibly
1260 // multiple idom forwarding steps.
1261 // Note that we piggyback on "_loop_or_ctrl" to do the forwarding,
1262 // since we forward both "get_ctrl" and "idom" from the dead to the
1263 // new live ctrl/idom nodes.
1264 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1265 assert(n != nullptr,"Bad immediate dominator info.");
1266 }
1267 return n;
1268 }
1269
1270 public:
1271 Node* idom(Node* n) const {
1272 return idom(n->_idx);
1273 }
1274
1275 Node* idom(uint node_idx) const {
1276 Node* n = idom_no_update(node_idx);
1277 // We store the found idom in the side-table again. In most cases,
1278 // this is a no-op, since we just read from _idom. But in cases where
1279 // there was an idom forwarding via dead idom nodes, this shortens the path.
1280 // See: forward_ctrl
1281 _idom[node_idx] = n;
1282 return n;
1283 }
1284
1285 uint dom_depth(Node* d) const {
1286 guarantee(d != nullptr, "Null dominator info.");
1287 guarantee(d->_idx < _idom_size, "");
1288 return _dom_depth[d->_idx];
1289 }
1290 void set_idom(Node* d, Node* n, uint dom_depth);
1291 // Locally compute IDOM using dom_lca call
1292 Node *compute_idom( Node *region ) const;
1293 // Recompute dom_depth
1294 void recompute_dom_depth();
1295
1296 // Is safept not required by an outer loop?
1297 bool is_deleteable_safept(Node* sfpt);
1298
1299 // Replace parallel induction variable (parallel to trip counter)
1300 void replace_parallel_iv(IdealLoopTree *loop);
1301
1302 Node *dom_lca( Node *n1, Node *n2 ) const {
1303 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1304 }
1305 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1306
1307 Node* dominated_node(Node* c1, Node* c2) {
1308 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1309 return is_dominator(c1, c2) ? c2 : c1;
1310 }
1311
1312 // Return control node that's dominated by the 2 others
1313 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1314 return dominated_node(c1, dominated_node(c2, c3));
1315 }
1316
1317 // Build and verify the loop tree without modifying the graph. This
1318 // is useful to verify that all inputs properly dominate their uses.
1319 static void verify(PhaseIterGVN& igvn) {
1320 #ifdef ASSERT
1321 ResourceMark rm;
1322 Compile::TracePhase tp(_t_idealLoopVerify);
1323 PhaseIdealLoop v(igvn);
1324 #endif
1325 }
1326
1327 // Recommended way to use PhaseIdealLoop.
1328 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1329 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1330 ResourceMark rm;
1331 PhaseIdealLoop v(igvn, mode);
1332
1333 Compile* C = Compile::current();
1334 if (!C->failing()) {
1335 // Cleanup any modified bits
1336 igvn.optimize();
1337 if (C->failing()) { return; }
1338 v.log_loop_tree();
1339 }
1340 }
1341
1342 // True if the method has at least 1 irreducible loop
1343 bool _has_irreducible_loops;
1344
1345 // Per-Node transform
1346 virtual Node* transform(Node* n) { return nullptr; }
1347
1348 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1349 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1350 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1351 Node* loop_iv_stride(Node* incr, Node*& xphi);
1352 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x);
1353
1354 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1355
1356 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1357 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1358 #ifdef ASSERT
1359 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1360 #endif
1361 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1362 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1363
1364 void add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt);
1365
1366 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1367 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1368 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1369 Node*& entry_control, Node*& iffalse);
1370
1371 Node* exact_limit( IdealLoopTree *loop );
1372
1373 // Return a post-walked LoopNode
1374 IdealLoopTree* get_loop(const Node* n) const {
1375 // Dead nodes have no loop, so return the top level loop instead
1376 if (!has_node(n)) return _ltree_root;
1377 assert(!has_ctrl(n), "");
1378 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1379 }
1380
1381 IdealLoopTree* ltree_root() const { return _ltree_root; }
1382
1383 // Is 'n' a (nested) member of 'loop'?
1384 bool is_member(const IdealLoopTree* loop, const Node* n) const {
1385 return loop->is_member(get_loop(n));
1386 }
1387
1388 // is the control for 'n' a (nested) member of 'loop'?
1389 bool ctrl_is_member(const IdealLoopTree* loop, const Node* n) {
1390 return is_member(loop, get_ctrl(n));
1391 }
1392
1393 // This is the basic building block of the loop optimizations. It clones an
1394 // entire loop body. It makes an old_new loop body mapping; with this
1395 // mapping you can find the new-loop equivalent to an old-loop node. All
1396 // new-loop nodes are exactly equal to their old-loop counterparts, all
1397 // edges are the same. All exits from the old-loop now have a RegionNode
1398 // that merges the equivalent new-loop path. This is true even for the
1399 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1400 // now come from (one or more) Phis that merge their new-loop equivalents.
1401 // Parameter side_by_side_idom:
1402 // When side_by_size_idom is null, the dominator tree is constructed for
1403 // the clone loop to dominate the original. Used in construction of
1404 // pre-main-post loop sequence.
1405 // When nonnull, the clone and original are side-by-side, both are
1406 // dominated by the passed in side_by_side_idom node. Used in
1407 // construction of unswitched loops.
1408 enum CloneLoopMode {
1409 IgnoreStripMined = 0, // Only clone inner strip mined loop
1410 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1411 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1412 // result control flow branches
1413 // either to inner clone or outer
1414 // strip mined loop.
1415 };
1416 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1417 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1418 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1419 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1420 Node_List*& split_if_set, Node_List*& split_bool_set,
1421 Node_List*& split_cex_set, Node_List& worklist,
1422 uint new_counter, CloneLoopMode mode);
1423 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1424 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1425 Node_List& extra_data_nodes);
1426
1427 // If we got the effect of peeling, either by actually peeling or by
1428 // making a pre-loop which must execute at least once, we can remove
1429 // all loop-invariant dominated tests in the main body.
1430 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1431
1432 // Generate code to do a loop peel for the given loop (and body).
1433 // old_new is a temp array.
1434 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1435
1436 // Add pre and post loops around the given loop. These loops are used
1437 // during RCE, unrolling and aligning loops.
1438 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1439
1440 // Find the last store in the body of an OuterStripMinedLoop when following memory uses
1441 Node *find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop);
1442
1443 // Add post loop after the given loop.
1444 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1445 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1446 Node* incr, Node* limit, CountedLoopNode*& post_head);
1447
1448 // Add a vector post loop between a vector main loop and the current post loop
1449 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1450 // If Node n lives in the back_ctrl block, we clone a private version of n
1451 // in preheader_ctrl block and return that, otherwise return n.
1452 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1453
1454 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1455 // unroll to do double iterations. The next round of major loop transforms
1456 // will repeat till the doubled loop body does all remaining iterations in 1
1457 // pass.
1458 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1459
1460 // Unroll the loop body one step - make each trip do 2 iterations.
1461 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1462
1463 // Return true if exp is a constant times an induction var
1464 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1465
1466 bool is_iv(Node* exp, Node* iv, BasicType bt);
1467
1468 // Return true if exp is a scaled induction var plus (or minus) constant
1469 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);
1470 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1471 jlong long_scale;
1472 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1473 int int_scale = checked_cast<int>(long_scale);
1474 if (p_scale != nullptr) {
1475 *p_scale = int_scale;
1476 }
1477 return true;
1478 }
1479 return false;
1480 }
1481 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1482 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1483 BasicType bt,
1484 jlong* p_scale, Node** p_offset,
1485 bool* p_short_scale, int depth);
1486
1487 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1488 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1489 Deoptimization::DeoptReason reason, int opcode,
1490 bool rewire_uncommon_proj_phi_inputs = false);
1491
1492 private:
1493 // Helper functions for create_new_if_for_predicate()
1494 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1495 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1496 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1497 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1498 const OrigToNewHashtable& orig_to_clone);
1499 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
1500 Node* loop_entry);
1501
1502 public:
1503 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1504
1505 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1506 void replace_control(Node* node, Node* new_control) {
1507 _igvn.replace_input_of(node, 0, new_control);
1508 set_idom(node, new_control, dom_depth(new_control));
1509 }
1510
1511 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1512 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1513 set_idom(loop_head, new_entry, dom_depth(new_entry));
1514 }
1515
1516 // Construct a range check for a predicate if
1517 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1518 jint stride, Node* range, bool upper, bool& overflow);
1519
1520 // Implementation of the loop predication to promote checks outside the loop
1521 bool loop_predication_impl(IdealLoopTree *loop);
1522
1523 private:
1524 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1525 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1526 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1527 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1528 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1529 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1530 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1531 Node_List& if_proj_list);
1532 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1533 IfProjNode* new_control, int scale, Node* offset, Node* range);
1534 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1535
1536 // Helper function to collect predicate for eliminating the useless ones
1537 void eliminate_useless_predicates() const;
1538
1539 void eliminate_useless_zero_trip_guard();
1540 void eliminate_useless_multiversion_if();
1541
1542 public:
1543 // Change the control input of expensive nodes to allow commoning by
1544 // IGVN when it is guaranteed to not result in a more frequent
1545 // execution of the expensive node. Return true if progress.
1546 bool process_expensive_nodes();
1547
1548 // Check whether node has become unreachable
1549 bool is_node_unreachable(Node *n) const {
1550 return !has_node(n) || n->is_unreachable(_igvn);
1551 }
1552
1553 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1554 void do_range_check(IdealLoopTree* loop);
1555
1556 // Clone loop with an invariant test (that does not exit) and
1557 // insert a clone of the test that selects which version to
1558 // execute.
1559 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1560
1561 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1562
1563 private:
1564 static bool has_control_dependencies_from_predicates(LoopNode* head);
1565 static void revert_to_normal_loop(const LoopNode* loop_head);
1566
1567 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1568 const UnswitchedLoopSelector& unswitched_loop_selector);
1569 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1570 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1571 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1572 #ifndef PRODUCT
1573 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1574 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1575 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1576 const LoopNode* original_head, const LoopNode* new_head);
1577 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1578 const LoopNode* original_head, const LoopNode* new_head);
1579 #endif
1580
1581 public:
1582
1583 // Range Check Elimination uses this function!
1584 // Constrain the main loop iterations so the affine function:
1585 // low_limit <= scale_con * I + offset < upper_limit
1586 // always holds true. That is, either increase the number of iterations in
1587 // the pre-loop or the post-loop until the condition holds true in the main
1588 // loop. Scale_con, offset and limit are all loop invariant.
1589 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);
1590 // Helper function for add_constraint().
1591 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1592
1593 // Partially peel loop up through last_peel node.
1594 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1595 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1596
1597 // AutoVectorize the loop: replace scalar ops with vector ops.
1598 enum AutoVectorizeStatus {
1599 Impossible, // This loop has the wrong shape to even try vectorization.
1600 Success, // We just successfully vectorized the loop.
1601 TriedAndFailed, // We tried to vectorize, but failed.
1602 };
1603 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1604
1605 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1606 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1607 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1608 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1609
1610 // Create a scheduled list of nodes control dependent on ctrl set.
1611 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1612 // Has a use in the vector set
1613 bool has_use_in_set( Node* n, VectorSet& vset );
1614 // Has use internal to the vector set (ie. not in a phi at the loop head)
1615 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1616 // clone "n" for uses that are outside of loop
1617 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1618 // clone "n" for special uses that are in the not_peeled region
1619 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1620 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1621 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1622 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1623 #ifdef ASSERT
1624 // Validate the loop partition sets: peel and not_peel
1625 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1626 // Ensure that uses outside of loop are of the right form
1627 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1628 uint orig_exit_idx, uint clone_exit_idx);
1629 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1630 #endif
1631
1632 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1633 int stride_of_possible_iv( Node* iff );
1634 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1635 // Return the (unique) control output node that's in the loop (if it exists.)
1636 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1637 // Insert a signed compare loop exit cloned from an unsigned compare.
1638 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1639 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1640 // Utility to register node "n" with PhaseIdealLoop
1641 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1642 // Utility to create an if-projection
1643 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1644 // Force the iff control output to be the live_proj
1645 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1646 // Insert a region before an if projection
1647 RegionNode* insert_region_before_proj(ProjNode* proj);
1648 // Insert a new if before an if projection
1649 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1650
1651 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1652 // "Nearly" because all Nodes have been cloned from the original in the loop,
1653 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1654 // through the Phi recursively, and return a Bool.
1655 Node* clone_iff(PhiNode* phi);
1656 CmpNode* clone_bool(PhiNode* phi);
1657
1658
1659 // Rework addressing expressions to get the most loop-invariant stuff
1660 // moved out. We'd like to do all associative operators, but it's especially
1661 // important (common) to do address expressions.
1662 Node* remix_address_expressions(Node* n);
1663 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1664
1665 // Convert add to muladd to generate MuladdS2I under certain criteria
1666 Node * convert_add_to_muladd(Node * n);
1667
1668 // Attempt to use a conditional move instead of a phi/branch
1669 Node *conditional_move( Node *n );
1670
1671 // Check for aggressive application of 'split-if' optimization,
1672 // using basic block level info.
1673 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1674 Node *split_if_with_blocks_pre ( Node *n );
1675 void split_if_with_blocks_post( Node *n );
1676 Node *has_local_phi_input( Node *n );
1677 // Mark an IfNode as being dominated by a prior test,
1678 // without actually altering the CFG (and hence IDOM info).
1679 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool prev_dom_not_imply_this = false);
1680 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool dominator_not_imply_source);
1681
1682 // Split Node 'n' through merge point
1683 RegionNode* split_thru_region(Node* n, RegionNode* region);
1684 // Split Node 'n' through merge point if there is enough win.
1685 Node *split_thru_phi( Node *n, Node *region, int policy );
1686 // Found an If getting its condition-code input from a Phi in the
1687 // same block. Split thru the Region.
1688 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1689
1690 private:
1691 // Class to keep track of wins in split_thru_phi.
1692 class SplitThruPhiWins {
1693 private:
1694 // Region containing the phi we are splitting through.
1695 const Node* _region;
1696
1697 // Sum of all wins regardless of where they happen. This applies to Loops phis as well as non-loop phis.
1698 int _total_wins;
1699
1700 // For Loops, wins have different impact depending on if they happen on loop entry or on the backedge.
1701 // Number of wins on a loop entry edge if the split is through a loop head,
1702 // otherwise 0. Entry edge wins only pay dividends once on loop entry.
1703 int _loop_entry_wins;
1704 // Number of wins on a loop back-edge, which pay dividends on every iteration.
1705 int _loop_back_wins;
1706
1707 public:
1708 SplitThruPhiWins(const Node* region) :
1709 _region(region),
1710 _total_wins(0),
1711 _loop_entry_wins(0),
1712 _loop_back_wins(0) {};
1713
1714 void reset() {_total_wins = 0; _loop_entry_wins = 0; _loop_back_wins = 0;}
1715 void add_win(int ctrl_index) {
1716 if (_region->is_Loop() && ctrl_index == LoopNode::EntryControl) {
1717 _loop_entry_wins++;
1718 } else if (_region->is_Loop() && ctrl_index == LoopNode::LoopBackControl) {
1719 _loop_back_wins++;
1720 }
1721 _total_wins++;
1722 }
1723 // Is this split profitable with respect to the policy?
1724 bool profitable(int policy) const {
1725 assert(_region->is_Loop() || (_loop_entry_wins == 0 && _loop_back_wins == 0), "wins on loop edges without a loop");
1726 assert(!_region->is_Loop() || _total_wins == _loop_entry_wins + _loop_back_wins, "missed some win");
1727 // In general this means that the split has to have more wins than specified
1728 // in the policy. However, for loops we need to take into account where the
1729 // wins happen. We need to be careful when splitting, because splitting nodes
1730 // related to the iv through the phi can sufficiently rearrange the loop
1731 // structure to prevent RCE and thus vectorization. Thus, we only deem splitting
1732 // profitable if the win of a split is not on the entry edge, as such wins
1733 // only pay off once and have a high chance of messing up the loop structure.
1734 return (_loop_entry_wins == 0 && _total_wins > policy) ||
1735 // If there are wins on the entry edge but the backadge also has sufficient wins,
1736 // there is sufficient profitability to spilt regardless of the risk of messing
1737 // up the loop structure.
1738 _loop_back_wins > policy ||
1739 // If the policy is less than 0, a split is always profitable, i.e. we always
1740 // split. This is needed when we split a node and then must also split a
1741 // dependant node, i.e. spliting a Bool node after splitting a Cmp node.
1742 policy < 0;
1743 }
1744 };
1745
1746
1747 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1748
1749 public:
1750
1751 // Conversion of fill/copy patterns into intrinsic versions
1752 bool do_intrinsify_fill();
1753 bool intrinsify_fill(IdealLoopTree* lpt);
1754 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1755 Node*& shift, Node*& offset);
1756
1757 private:
1758 // Return a type based on condition control flow
1759 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1760 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1761 // Helpers for filtered type
1762 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1763
1764 // Helper functions
1765 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1766 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1767 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 );
1768 bool split_up( Node *n, Node *blk1, Node *blk2 );
1769
1770 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1771 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1772 void try_move_store_after_loop(Node* n);
1773 bool identical_backtoback_ifs(Node *n);
1774 bool can_split_if(Node *n_ctrl);
1775 bool cannot_split_division(const Node* n, const Node* region) const;
1776 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1777 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1778
1779 // Determine if a method is too big for a/another round of split-if, based on
1780 // a magic (approximate) ratio derived from the equally magic constant 35000,
1781 // previously used for this purpose (but without relating to the node limit).
1782 bool must_throttle_split_if() {
1783 uint threshold = C->max_node_limit() * 2 / 5;
1784 return C->live_nodes() > threshold;
1785 }
1786
1787 // A simplistic node request tracking mechanism, where
1788 // = UINT_MAX Request not valid or made final.
1789 // < UINT_MAX Nodes currently requested (estimate).
1790 uint _nodes_required;
1791
1792 enum { REQUIRE_MIN = 70 };
1793
1794 uint nodes_required() const { return _nodes_required; }
1795
1796 // Given the _currently_ available number of nodes, check whether there is
1797 // "room" for an additional request or not, considering the already required
1798 // number of nodes. Return TRUE if the new request is exceeding the node
1799 // budget limit, otherwise return FALSE. Note that this interpretation will
1800 // act pessimistic on additional requests when new nodes have already been
1801 // generated since the 'begin'. This behaviour fits with the intention that
1802 // node estimates/requests should be made upfront.
1803 bool exceeding_node_budget(uint required = 0) {
1804 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1805 uint available = C->max_node_limit() - C->live_nodes();
1806 return available < required + _nodes_required + REQUIRE_MIN;
1807 }
1808
1809 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1810 precond(require > 0);
1811 _nodes_required += MAX2(require, minreq);
1812 return _nodes_required;
1813 }
1814
1815 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1816 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1817 }
1818
1819 uint require_nodes_begin() {
1820 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1821 _nodes_required = 0;
1822 return C->live_nodes();
1823 }
1824
1825 // When a node request is final, optionally check that the requested number
1826 // of nodes was reasonably correct with respect to the number of new nodes
1827 // introduced since the last 'begin'. Always check that we have not exceeded
1828 // the maximum node limit.
1829 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1830 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1831
1832 #ifdef ASSERT
1833 if (check_estimate) {
1834 // Check that the node budget request was not off by too much (x2).
1835 // Should this be the case we _surely_ need to improve the estimates
1836 // used in our budget calculations.
1837 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1838 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1839 C->live_nodes() - live_at_begin, _nodes_required);
1840 }
1841 }
1842 #endif
1843 // Assert that we have stayed within the node budget limit.
1844 assert(C->live_nodes() < C->max_node_limit(),
1845 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1846 C->live_nodes() - live_at_begin, live_at_begin,
1847 C->max_node_limit(), _nodes_required);
1848
1849 _nodes_required = UINT_MAX;
1850 }
1851
1852 private:
1853
1854 bool _created_loop_node;
1855 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1856 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1857
1858 public:
1859 void set_created_loop_node() { _created_loop_node = true; }
1860 bool created_loop_node() { return _created_loop_node; }
1861 void register_new_node(Node* n, Node* blk);
1862 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1863 register_new_node(new_node, get_ctrl(ctrl_of));
1864 }
1865
1866 Node* clone_and_register(Node* n, Node* ctrl) {
1867 n = n->clone();
1868 register_new_node(n, ctrl);
1869 return n;
1870 }
1871
1872 #ifdef ASSERT
1873 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1874 #endif
1875
1876 #ifndef PRODUCT
1877 void dump() const;
1878 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1879 void dump_idom(Node* n, uint count) const;
1880 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1881 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1882 IdealLoopTree* get_loop_idx(Node* n) const {
1883 // Dead nodes have no loop, so return the top level loop instead
1884 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1885 }
1886 // Print some stats
1887 static void print_statistics();
1888 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1889 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1890 static volatile int _long_loop_candidates;
1891 static volatile int _long_loop_nests;
1892 static volatile int _long_loop_counted_loops;
1893 #endif
1894
1895 #ifdef ASSERT
1896 void verify() const;
1897 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1898 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1899 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1900 #endif
1901
1902 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1903
1904 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1905
1906 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1907
1908
1909 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1910 Node_List &range_checks);
1911
1912 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1913 Node* inner_iters_actual_int, Node* inner_phi,
1914 Node* iv_add, LoopNode* inner_head);
1915
1916 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1917
1918 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1919
1920 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1921
1922 bool would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl);
1923
1924 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1925
1926 void try_sink_out_of_loop(Node* n);
1927
1928 Node* clamp(Node* R, Node* L, Node* H);
1929
1930 bool safe_for_if_replacement(const Node* dom) const;
1931
1932 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1933
1934 bool try_merge_identical_ifs(Node* n);
1935
1936 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1937
1938 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1939 IdealLoopTree* parent, bool partial);
1940
1941 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1942 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1943
1944 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1945 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1946 Node_List*& split_bool_set, Node_List*& split_cex_set);
1947
1948 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1949
1950 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1951
1952 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1953 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1954 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1955 void clone_template_assertion_expression_down(Node* node);
1956
1957 Node* similar_subtype_check(const Node* x, Node* r_in);
1958
1959 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1960
1961 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1962
1963 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
1964
1965 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1966
1967 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
1968
1969 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
1970
1971 ConINode* intcon(jint i);
1972
1973 ConLNode* longcon(jlong i);
1974
1975 ConNode* makecon(const Type* t);
1976
1977 ConNode* integercon(jlong l, BasicType bt);
1978
1979 ConNode* zerocon(BasicType bt);
1980 };
1981
1982
1983 class AutoNodeBudget : public StackObj
1984 {
1985 public:
1986 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1987
1988 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1989 : _phase(phase),
1990 _check_at_final(chk == BUDGET_CHECK),
1991 _nodes_at_begin(0)
1992 {
1993 precond(_phase != nullptr);
1994
1995 _nodes_at_begin = _phase->require_nodes_begin();
1996 }
1997
1998 ~AutoNodeBudget() {
1999 #ifndef PRODUCT
2000 if (TraceLoopOpts) {
2001 uint request = _phase->nodes_required();
2002 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
2003
2004 if (request < delta) {
2005 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
2006 } else {
2007 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
2008 // Identify the worst estimates as "poor" ones.
2009 if (request > REQUIRE_MIN && delta > 0) {
2010 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
2011 (delta <= REQUIRE_MIN && request > 10 * delta)) {
2012 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
2013 }
2014 }
2015 }
2016 }
2017 #endif // PRODUCT
2018 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
2019 }
2020
2021 private:
2022 PhaseIdealLoop* _phase;
2023 bool _check_at_final;
2024 uint _nodes_at_begin;
2025 };
2026
2027 inline Node* IdealLoopTree::tail() {
2028 // Handle lazy update of _tail field.
2029 if (_tail->in(0) == nullptr) {
2030 _tail = _phase->get_ctrl(_tail);
2031 }
2032 return _tail;
2033 }
2034
2035 inline Node* IdealLoopTree::head() {
2036 // Handle lazy update of _head field.
2037 if (_head->in(0) == nullptr) {
2038 _head = _phase->get_ctrl(_head);
2039 }
2040 return _head;
2041 }
2042
2043 // Iterate over the loop tree using a preorder, left-to-right traversal.
2044 //
2045 // Example that visits all counted loops from within PhaseIdealLoop
2046 //
2047 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2048 // IdealLoopTree* lpt = iter.current();
2049 // if (!lpt->is_counted()) continue;
2050 // ...
2051 class LoopTreeIterator : public StackObj {
2052 private:
2053 IdealLoopTree* _root;
2054 IdealLoopTree* _curnt;
2055
2056 public:
2057 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
2058
2059 bool done() { return _curnt == nullptr; } // Finished iterating?
2060
2061 void next(); // Advance to next loop tree
2062
2063 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
2064 };
2065
2066 // Compute probability of reaching some CFG node from a fixed
2067 // dominating CFG node
2068 class PathFrequency {
2069 private:
2070 Node* _dom; // frequencies are computed relative to this node
2071 Node_Stack _stack;
2072 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
2073 GrowableArray<float> _freqs; // cache frequencies
2074 PhaseIdealLoop* _phase;
2075
2076 float check_and_truncate_frequency(float f) {
2077 assert(f >= 0, "Incorrect frequency");
2078 // We do not perform an exact (f <= 1) check
2079 // this would be error prone with rounding of floats.
2080 // Performing a check like (f <= 1+eps) would be of benefit,
2081 // however, it is not evident how to determine such an eps,
2082 // given that an arbitrary number of add/mul operations
2083 // are performed on these frequencies.
2084 return (f > 1) ? 1 : f;
2085 }
2086
2087 public:
2088 PathFrequency(Node* dom, PhaseIdealLoop* phase)
2089 : _dom(dom), _stack(0), _phase(phase) {
2090 }
2091
2092 float to(Node* n);
2093 };
2094
2095 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
2096 // 1. Clone the data nodes
2097 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
2098 class DataNodeGraph : public StackObj {
2099 PhaseIdealLoop* const _phase;
2100 const Unique_Node_List& _data_nodes;
2101 OrigToNewHashtable _orig_to_new;
2102
2103 public:
2104 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
2105 : _phase(phase),
2106 _data_nodes(data_nodes),
2107 // Use 107 as best guess which is the first resize value in ResizeableHashTable::large_table_sizes.
2108 _orig_to_new(107, MaxNodeLimit)
2109 {
2110 #ifdef ASSERT
2111 for (uint i = 0; i < data_nodes.size(); i++) {
2112 assert(!data_nodes[i]->is_CFG(), "only data nodes");
2113 }
2114 #endif
2115 }
2116 NONCOPYABLE(DataNodeGraph);
2117
2118 private:
2119 void clone(Node* node, Node* new_ctrl);
2120 void clone_data_nodes(Node* new_ctrl);
2121 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2122 Node* new_ctrl);
2123 void rewire_clones_to_cloned_inputs();
2124 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
2125
2126 public:
2127 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
2128 // Set 'new_ctrl' as ctrl for the cloned nodes.
2129 const OrigToNewHashtable& clone(Node* new_ctrl) {
2130 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
2131 clone_data_nodes(new_ctrl);
2132 rewire_clones_to_cloned_inputs();
2133 return _orig_to_new;
2134 }
2135
2136 // Create a copy of the data nodes provided to the constructor by doing the following:
2137 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
2138 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2139 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2140 // `new_ctrl` as ctrl.
2141 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2142 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2143 Node* new_ctrl) {
2144 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2145 rewire_clones_to_cloned_inputs();
2146 return _orig_to_new;
2147 }
2148 };
2149 #endif // SHARE_OPTO_LOOPNODE_HPP