1 /*
2 * Copyright (c) 1998, 2025, 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()->proj_out(false)->as_IfFalse(); }
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 : _parent(nullptr), _next(nullptr), _child(nullptr),
674 _head(head), _tail(tail),
675 _phase(phase),
676 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
677 _body(Compile::current()->comp_arena()),
678 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
679 _has_range_checks(0), _has_range_checks_computed(0),
680 _safepts(nullptr),
681 _required_safept(nullptr),
682 _allow_optimizations(true)
683 {
684 precond(_head != nullptr);
685 precond(_tail != nullptr);
686 }
687
688 // Is 'l' a member of 'this'?
689 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
690
691 // Set loop nesting depth. Accumulate has_call bits.
692 int set_nest( uint depth );
693
694 // Split out multiple fall-in edges from the loop header. Move them to a
695 // private RegionNode before the loop. This becomes the loop landing pad.
696 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
697
698 // Split out the outermost loop from this shared header.
699 void split_outer_loop( PhaseIdealLoop *phase );
700
701 // Merge all the backedges from the shared header into a private Region.
702 // Feed that region as the one backedge to this loop.
703 void merge_many_backedges( PhaseIdealLoop *phase );
704
705 // Split shared headers and insert loop landing pads.
706 // Insert a LoopNode to replace the RegionNode.
707 // Returns TRUE if loop tree is structurally changed.
708 bool beautify_loops( PhaseIdealLoop *phase );
709
710 // Perform optimization to use the loop predicates for null checks and range checks.
711 // Applies to any loop level (not just the innermost one)
712 bool loop_predication( PhaseIdealLoop *phase);
713 bool can_apply_loop_predication();
714
715 // Perform iteration-splitting on inner loops. Split iterations to
716 // avoid range checks or one-shot null checks. Returns false if the
717 // current round of loop opts should stop.
718 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
719
720 // Driver for various flavors of iteration splitting. Returns false
721 // if the current round of loop opts should stop.
722 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
723
724 // Given dominators, try to find loops with calls that must always be
725 // executed (call dominates loop tail). These loops do not need non-call
726 // safepoints (ncsfpt).
727 void check_safepts(VectorSet &visited, Node_List &stack);
728
729 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
730 // encountered.
731 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
732
733 // Remove safepoints from loop. Optionally keeping one.
734 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
735
736 // Convert to counted loops where possible
737 void counted_loop( PhaseIdealLoop *phase );
738
739 // Check for Node being a loop-breaking test
740 Node *is_loop_exit(Node *iff) const;
741
742 // Remove simplistic dead code from loop body
743 void DCE_loop_body();
744
745 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
746 // Replace with a 1-in-10 exit guess.
747 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
748
749 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
750 // Useful for unrolling loops with NO array accesses.
751 bool policy_peel_only( PhaseIdealLoop *phase ) const;
752
753 // Return TRUE or FALSE if the loop should be unswitched -- clone
754 // loop with an invariant test
755 bool policy_unswitching( PhaseIdealLoop *phase ) const;
756
757 // Micro-benchmark spamming. Remove empty loops.
758 bool do_remove_empty_loop( PhaseIdealLoop *phase );
759
760 // Convert one-iteration loop into normal code.
761 bool do_one_iteration_loop( PhaseIdealLoop *phase );
762
763 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
764 // move some loop-invariant test (usually a null-check) before the loop.
765 bool policy_peeling(PhaseIdealLoop *phase);
766
767 uint estimate_peeling(PhaseIdealLoop *phase);
768
769 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
770 // known trip count in the counted loop node.
771 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
772
773 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
774 // if the loop is a counted loop and the loop body is small enough.
775 bool policy_unroll(PhaseIdealLoop *phase);
776
777 // Loop analyses to map to a maximal superword unrolling for vectorization.
778 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
779
780 // Return TRUE or FALSE if the loop should be range-check-eliminated.
781 // Gather a list of IF tests that are dominated by iteration splitting;
782 // also gather the end of the first split and the start of the 2nd split.
783 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
784
785 // Return TRUE if "iff" is a range check.
786 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
787 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
788 jlong& scale) const;
789
790 // Estimate the number of nodes required when cloning a loop (body).
791 uint est_loop_clone_sz(uint factor) const;
792 // Estimate the number of nodes required when unrolling a loop (body).
793 uint est_loop_unroll_sz(uint factor) const;
794
795 // Compute loop trip count if possible
796 void compute_trip_count(PhaseIdealLoop* phase, BasicType bt);
797
798 // Compute loop trip count from profile data
799 float compute_profile_trip_cnt_helper(Node* n);
800 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
801
802 // Reassociate invariant expressions.
803 void reassociate_invariants(PhaseIdealLoop *phase);
804 // Reassociate invariant binary expressions.
805 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
806 // Reassociate invariant add, subtract, and compare expressions.
807 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
808 // Return nonzero index of invariant operand if invariant and variant
809 // are combined with an associative binary. Helper for reassociate_invariants.
810 int find_invariant(Node* n, PhaseIdealLoop *phase);
811 // Return TRUE if "n" is associative.
812 bool is_associative(Node* n, Node* base=nullptr);
813 // Return TRUE if "n" is an associative cmp node.
814 bool is_associative_cmp(Node* n);
815
816 // Return true if n is invariant
817 bool is_invariant(Node* n) const;
818
819 // Put loop body on igvn work list
820 void record_for_igvn();
821
822 bool is_root() { return _parent == nullptr; }
823 // A proper/reducible loop w/o any (occasional) dead back-edge.
824 bool is_loop() { return !_irreducible && !tail()->is_top(); }
825 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
826 bool is_innermost() { return is_loop() && _child == nullptr; }
827
828 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
829
830 bool compute_has_range_checks() const;
831 bool range_checks_present() {
832 if (!_has_range_checks_computed) {
833 if (compute_has_range_checks()) {
834 _has_range_checks = 1;
835 }
836 _has_range_checks_computed = 1;
837 }
838 return _has_range_checks;
839 }
840
841 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
842 // In all other cases, return this.
843 IdealLoopTree* skip_strip_mined() {
844 return _head->as_Loop()->is_strip_mined() ? _parent : this;
845 }
846
847 #ifndef PRODUCT
848 void dump_head(); // Dump loop head only
849 void dump(); // Dump this loop recursively
850 #endif
851
852 #ifdef ASSERT
853 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
854 bool verify_tree(IdealLoopTree* loop_verify) const;
855 #endif
856
857 private:
858 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
859
860 // Estimate the number of nodes resulting from control and data flow merge.
861 uint est_loop_flow_merge_sz() const;
862
863 // Check if the number of residual iterations is large with unroll_cnt.
864 // Return true if the residual iterations are more than 10% of the trip count.
865 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
866 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
867 }
868
869 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
870
871 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
872
873 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
874
875 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
876 Node* sfpt) const;
877
878 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
879
880 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
881 };
882
883 // -----------------------------PhaseIdealLoop---------------------------------
884 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
885 // into a loop tree. Drives the loop-based transformations on the ideal graph.
886 class PhaseIdealLoop : public PhaseTransform {
887 friend class IdealLoopTree;
888 friend class SuperWord;
889 friend class ShenandoahBarrierC2Support;
890 friend class AutoNodeBudget;
891
892 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
893 //
894 // Exception: dead CFG nodes may instead have a ctrl/idom forwarding
895 // installed. See: forward_ctrl
896 Node_List _loop_or_ctrl;
897
898 // Pre-computed def-use info
899 PhaseIterGVN &_igvn;
900
901 // Head of loop tree
902 IdealLoopTree* _ltree_root;
903
904 // Array of pre-order numbers, plus post-visited bit.
905 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
906 // ODD for post-visited. Other bits are the pre-order number.
907 uint *_preorders;
908 uint _max_preorder;
909
910 ReallocMark _nesting; // Safety checks for arena reallocation
911
912 const PhaseIdealLoop* _verify_me;
913 bool _verify_only;
914
915 // Allocate _preorders[] array
916 void allocate_preorders() {
917 _max_preorder = C->unique()+8;
918 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
919 memset(_preorders, 0, sizeof(uint) * _max_preorder);
920 }
921
922 // Allocate _preorders[] array
923 void reallocate_preorders() {
924 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
925 if ( _max_preorder < C->unique() ) {
926 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
927 _max_preorder = C->unique();
928 }
929 memset(_preorders, 0, sizeof(uint) * _max_preorder);
930 }
931
932 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
933 // adds new nodes.
934 void check_grow_preorders( ) {
935 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
936 if ( _max_preorder < C->unique() ) {
937 uint newsize = _max_preorder<<1; // double size of array
938 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
939 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
940 _max_preorder = newsize;
941 }
942 }
943 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
944 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
945 // Pre-order numbers are written to the Nodes array as low-bit-set values.
946 void set_preorder_visited( Node *n, int pre_order ) {
947 assert( !is_visited( n ), "already set" );
948 _preorders[n->_idx] = (pre_order<<1);
949 };
950 // Return pre-order number.
951 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
952
953 // Check for being post-visited.
954 // Should be previsited already (checked with assert(is_visited(n))).
955 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
956
957 // Mark as post visited
958 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
959
960 public:
961 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
962 // Returns true if "n" is a data node, false if it's a CFG node.
963 //
964 // Exception:
965 // control nodes that are dead because of "replace_node_and_forward_ctrl"
966 // or have otherwise modified their ctrl state by "forward_ctrl".
967 // They return "true", because they have a ctrl "forwarding" to the other ctrl node they
968 // were replaced with.
969 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
970
971 private:
972 // clear out dead code after build_loop_late
973 Node_List _deadlist;
974 Node_List _zero_trip_guard_opaque_nodes;
975 Node_List _multiversion_opaque_nodes;
976
977 // Support for faster execution of get_late_ctrl()/dom_lca()
978 // when a node has many uses and dominator depth is deep.
979 GrowableArray<jlong> _dom_lca_tags;
980 uint _dom_lca_tags_round;
981 void init_dom_lca_tags();
982
983 // Helper for debugging bad dominance relationships
984 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
985
986 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
987
988 // Inline wrapper for frequent cases:
989 // 1) only one use
990 // 2) a use is the same as the current LCA passed as 'n1'
991 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
992 assert( n->is_CFG(), "" );
993 // Fast-path null lca
994 if( lca != nullptr && lca != n ) {
995 assert( lca->is_CFG(), "" );
996 // find LCA of all uses
997 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
998 }
999 return find_non_split_ctrl(n);
1000 }
1001 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
1002
1003 // Helper function for directing control inputs away from CFG split points.
1004 Node *find_non_split_ctrl( Node *ctrl ) const {
1005 if (ctrl != nullptr) {
1006 if (ctrl->is_MultiBranch()) {
1007 ctrl = ctrl->in(0);
1008 }
1009 assert(ctrl->is_CFG(), "CFG");
1010 }
1011 return ctrl;
1012 }
1013
1014 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
1015
1016 #ifdef ASSERT
1017 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
1018 #endif
1019 private:
1020 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
1021 Unique_Node_List& list);
1022 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
1023 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1024 CountedLoopNode* remaining_loop_head,
1025 uint first_node_index_in_cloned_loop_body,
1026 const Node_List& old_new);
1027 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1028 CountedLoopNode* main_loop_head,
1029 uint first_node_index_in_pre_loop_body,
1030 uint last_node_index_in_pre_loop_body,
1031 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1032 const Node_List& old_new);
1033 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1034 uint first_node_index_in_cloned_loop_body);
1035 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1036 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1037 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1038 CountedLoopNode* target_loop_head,
1039 const NodeInLoopBody& _node_in_loop_body,
1040 bool kill_old_template);
1041 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1042 const Node* old_target_loop_entry,
1043 uint node_index_before_new_assertion_predicate_nodes);
1044 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
1045 Node* bol);
1046 void log_loop_tree();
1047
1048 public:
1049
1050 PhaseIterGVN &igvn() const { return _igvn; }
1051
1052 bool has_node(const Node* n) const {
1053 guarantee(n != nullptr, "No Node.");
1054 return _loop_or_ctrl[n->_idx] != nullptr;
1055 }
1056 // check if transform created new nodes that need _ctrl recorded
1057 Node *get_late_ctrl( Node *n, Node *early );
1058 Node *get_early_ctrl( Node *n );
1059 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1060 void set_early_ctrl(Node* n, bool update_body);
1061 void set_subtree_ctrl(Node* n, bool update_body);
1062 void set_ctrl( Node *n, Node *ctrl ) {
1063 assert( !has_node(n) || has_ctrl(n), "" );
1064 assert( ctrl->in(0), "cannot set dead control node" );
1065 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1066 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1067 }
1068 void set_root_as_ctrl(Node* n) {
1069 assert(!has_node(n) || has_ctrl(n), "");
1070 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1071 }
1072 // Set control and update loop membership
1073 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1074 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1075 IdealLoopTree* new_loop = get_loop(ctrl);
1076 if (old_loop != new_loop) {
1077 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1078 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1079 }
1080 set_ctrl(n, ctrl);
1081 }
1082
1083 // Retrieves the ctrl for a data node i.
1084 Node* get_ctrl(const Node* i) {
1085 assert(has_node(i) && has_ctrl(i), "must be data node with ctrl");
1086 Node* n = get_ctrl_no_update(i);
1087 // We store the found ctrl in the side-table again. In most cases,
1088 // this is a no-op, since we just read from _loop_or_ctrl. But in cases
1089 // where there was a ctrl forwarding via dead ctrl nodes, this shortens the path.
1090 // See: forward_ctrl
1091 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1092 assert(has_node(i) && has_ctrl(i), "must still be data node with ctrl");
1093 assert(n == find_non_split_ctrl(n), "must return legal ctrl");
1094 return n;
1095 }
1096
1097 bool is_dominator(Node* dominator, Node* n);
1098 bool is_strict_dominator(Node* dominator, Node* n);
1099
1100 // return get_ctrl for a data node and self(n) for a CFG node
1101 Node* ctrl_or_self(Node* n) {
1102 if (has_ctrl(n))
1103 return get_ctrl(n);
1104 else {
1105 assert (n->is_CFG(), "must be a CFG node");
1106 return n;
1107 }
1108 }
1109
1110 private:
1111 Node* get_ctrl_no_update_helper(const Node* i) const {
1112 // We expect only data nodes (which must have a ctrl set), or
1113 // dead ctrl nodes that have a ctrl "forwarding".
1114 // See: forward_ctrl.
1115 assert(has_ctrl(i), "only data nodes or ctrl nodes with ctrl forwarding expected");
1116 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1117 }
1118
1119 // Compute the ctrl of node i, jumping over ctrl forwardings.
1120 Node* get_ctrl_no_update(const Node* i) const {
1121 assert(has_ctrl(i), "only data nodes expected");
1122 Node* n = get_ctrl_no_update_helper(i);
1123 if (n->in(0) == nullptr) {
1124 // We encountered a dead CFG node.
1125 // If everything went right, this dead CFG node should have had a ctrl
1126 // forwarding installed, using "forward_ctrl". We now have to jump from
1127 // the old (dead) ctrl node to the new (live) ctrl node, in possibly
1128 // multiple ctrl forwarding steps.
1129 do {
1130 n = get_ctrl_no_update_helper(n);
1131 } while (n->in(0) == nullptr);
1132 n = find_non_split_ctrl(n);
1133 }
1134 return n;
1135 }
1136
1137 public:
1138 // Check for loop being set
1139 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1140 bool has_loop( Node *n ) const {
1141 assert(!has_node(n) || !has_ctrl(n), "");
1142 return has_node(n);
1143 }
1144 // Set loop
1145 void set_loop( Node *n, IdealLoopTree *loop ) {
1146 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1147 }
1148
1149 // Install a ctrl "forwarding" from an old (dead) control node.
1150 // This is a "lazy" update of the "get_ctrl" and "idom" mechanism:
1151 // - Install a forwarding from old_node (dead ctrl) to new_node.
1152 // - When querying "get_ctrl": jump from data node over possibly
1153 // multiple dead ctrl nodes with ctrl forwarding to eventually
1154 // reach a live ctrl node. Shorten the path to avoid chasing the
1155 // forwarding in the future.
1156 // - When querying "idom": from some node get its old idom, which
1157 // may be dead but has an idom forwarding to the new and live
1158 // idom. Shorten the path to avoid chasing the forwarding in the
1159 // future.
1160 // Note: while the "idom" information is stored in the "_idom"
1161 // side-table, the idom forwarding piggybacks on the ctrl
1162 // forwarding on "_loop_or_ctrl".
1163 // Using "forward_ctrl" allows us to only edit the entry for the old
1164 // dead node now, and we do not have to update all the nodes that had
1165 // the old_node as their "get_ctrl" or "idom". We clean up the forwarding
1166 // links when we query "get_ctrl" or "idom" for these nodes the next time.
1167 void forward_ctrl(Node* old_node, Node* new_node) {
1168 assert(!has_ctrl(old_node) && old_node->is_CFG() && old_node->in(0) == nullptr,
1169 "must be dead ctrl (CFG) node");
1170 assert(!has_ctrl(new_node) && new_node->is_CFG() && new_node->in(0) != nullptr,
1171 "must be live ctrl (CFG) node");
1172 assert(old_node != new_node, "no cycles please");
1173 // Re-use the side array slot for this node to provide the
1174 // forwarding pointer.
1175 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1176 assert(has_ctrl(old_node), "must have installed ctrl forwarding");
1177 }
1178
1179 // Replace the old ctrl node with a new ctrl node.
1180 // - Update the node inputs of all uses.
1181 // - Lazily update the ctrl and idom info of all uses, via a ctrl/idom forwarding.
1182 void replace_node_and_forward_ctrl(Node* old_node, Node* new_node) {
1183 _igvn.replace_node(old_node, new_node);
1184 forward_ctrl(old_node, new_node);
1185 }
1186
1187 private:
1188
1189 // Place 'n' in some loop nest, where 'n' is a CFG node
1190 void build_loop_tree();
1191 int build_loop_tree_impl(Node* n, int pre_order);
1192 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1193 // loop tree, not the root.
1194 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1195
1196 #ifdef ASSERT
1197 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1198 void verify_regions_in_irreducible_loops();
1199 bool is_in_irreducible_loop(RegionNode* region);
1200 #endif
1201
1202 // Place Data nodes in some loop nest
1203 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1204 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1205 void build_loop_late_post_work(Node* n, bool pinned);
1206 void build_loop_late_post(Node* n);
1207 void verify_strip_mined_scheduling(Node *n, Node* least);
1208
1209 // Array of immediate dominance info for each CFG node indexed by node idx
1210 private:
1211 uint _idom_size;
1212 Node **_idom; // Array of immediate dominators
1213 uint *_dom_depth; // Used for fast LCA test
1214 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1215 LoopOptsMode _mode;
1216
1217 // build the loop tree and perform any requested optimizations
1218 void build_and_optimize();
1219
1220 // Dominators for the sea of nodes
1221 void Dominators();
1222
1223 // Compute the Ideal Node to Loop mapping
1224 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1225 PhaseTransform(Ideal_Loop),
1226 _loop_or_ctrl(igvn.C->comp_arena()),
1227 _igvn(igvn),
1228 _verify_me(nullptr),
1229 _verify_only(false),
1230 _mode(mode),
1231 _nodes_required(UINT_MAX) {
1232 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1233 build_and_optimize();
1234 }
1235
1236 #ifndef PRODUCT
1237 // Verify that verify_me made the same decisions as a fresh run
1238 // or only verify that the graph is valid if verify_me is null.
1239 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1240 PhaseTransform(Ideal_Loop),
1241 _loop_or_ctrl(igvn.C->comp_arena()),
1242 _igvn(igvn),
1243 _verify_me(verify_me),
1244 _verify_only(verify_me == nullptr),
1245 _mode(LoopOptsVerify),
1246 _nodes_required(UINT_MAX) {
1247 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1248 build_and_optimize();
1249 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1250 }
1251 #endif
1252
1253 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1254
1255 Node* idom_no_update(Node* d) const {
1256 return idom_no_update(d->_idx);
1257 }
1258
1259 Node* idom_no_update(uint node_idx) const {
1260 assert(node_idx < _idom_size, "oob");
1261 Node* n = _idom[node_idx];
1262 assert(n != nullptr,"Bad immediate dominator info.");
1263 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1264 // We encountered a dead CFG node.
1265 // If everything went right, this dead CFG node should have had an idom
1266 // forwarding installed, using "forward_ctrl". We now have to jump from
1267 // the old (dead) idom node to the new (live) idom node, in possibly
1268 // multiple idom forwarding steps.
1269 // Note that we piggyback on "_loop_or_ctrl" to do the forwarding,
1270 // since we forward both "get_ctrl" and "idom" from the dead to the
1271 // new live ctrl/idom nodes.
1272 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1273 assert(n != nullptr,"Bad immediate dominator info.");
1274 }
1275 return n;
1276 }
1277
1278 public:
1279 Node* idom(Node* n) const {
1280 return idom(n->_idx);
1281 }
1282
1283 Node* idom(uint node_idx) const {
1284 Node* n = idom_no_update(node_idx);
1285 // We store the found idom in the side-table again. In most cases,
1286 // this is a no-op, since we just read from _idom. But in cases where
1287 // there was an idom forwarding via dead idom nodes, this shortens the path.
1288 // See: forward_ctrl
1289 _idom[node_idx] = n;
1290 return n;
1291 }
1292
1293 uint dom_depth(Node* d) const {
1294 guarantee(d != nullptr, "Null dominator info.");
1295 guarantee(d->_idx < _idom_size, "");
1296 return _dom_depth[d->_idx];
1297 }
1298 void set_idom(Node* d, Node* n, uint dom_depth);
1299 // Locally compute IDOM using dom_lca call
1300 Node *compute_idom( Node *region ) const;
1301 // Recompute dom_depth
1302 void recompute_dom_depth();
1303
1304 // Is safept not required by an outer loop?
1305 bool is_deleteable_safept(Node* sfpt);
1306
1307 // Replace parallel induction variable (parallel to trip counter)
1308 void replace_parallel_iv(IdealLoopTree *loop);
1309
1310 Node *dom_lca( Node *n1, Node *n2 ) const {
1311 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1312 }
1313 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1314
1315 Node* dominated_node(Node* c1, Node* c2) {
1316 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1317 return is_dominator(c1, c2) ? c2 : c1;
1318 }
1319
1320 // Return control node that's dominated by the 2 others
1321 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1322 return dominated_node(c1, dominated_node(c2, c3));
1323 }
1324
1325 // Build and verify the loop tree without modifying the graph. This
1326 // is useful to verify that all inputs properly dominate their uses.
1327 static void verify(PhaseIterGVN& igvn) {
1328 #ifdef ASSERT
1329 ResourceMark rm;
1330 Compile::TracePhase tp(_t_idealLoopVerify);
1331 PhaseIdealLoop v(igvn);
1332 #endif
1333 }
1334
1335 // Recommended way to use PhaseIdealLoop.
1336 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1337 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1338 ResourceMark rm;
1339 PhaseIdealLoop v(igvn, mode);
1340
1341 Compile* C = Compile::current();
1342 if (!C->failing()) {
1343 // Cleanup any modified bits
1344 igvn.optimize();
1345 if (C->failing()) { return; }
1346 v.log_loop_tree();
1347 }
1348 }
1349
1350 // True if the method has at least 1 irreducible loop
1351 bool _has_irreducible_loops;
1352
1353 // Per-Node transform
1354 virtual Node* transform(Node* n) { return nullptr; }
1355
1356 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1357 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1358 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1359 Node* loop_iv_stride(Node* incr, Node*& xphi);
1360 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x);
1361
1362 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1363
1364 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1365 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1366 #ifdef ASSERT
1367 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1368 #endif
1369 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1370 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1371 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1372 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1373 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1374 Node*& entry_control, Node*& iffalse);
1375
1376 Node* exact_limit( IdealLoopTree *loop );
1377
1378 // Return a post-walked LoopNode
1379 IdealLoopTree *get_loop( Node *n ) const {
1380 // Dead nodes have no loop, so return the top level loop instead
1381 if (!has_node(n)) return _ltree_root;
1382 assert(!has_ctrl(n), "");
1383 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1384 }
1385
1386 IdealLoopTree* ltree_root() const { return _ltree_root; }
1387
1388 // Is 'n' a (nested) member of 'loop'?
1389 int is_member( const IdealLoopTree *loop, Node *n ) const {
1390 return loop->is_member(get_loop(n)); }
1391
1392 // This is the basic building block of the loop optimizations. It clones an
1393 // entire loop body. It makes an old_new loop body mapping; with this
1394 // mapping you can find the new-loop equivalent to an old-loop node. All
1395 // new-loop nodes are exactly equal to their old-loop counterparts, all
1396 // edges are the same. All exits from the old-loop now have a RegionNode
1397 // that merges the equivalent new-loop path. This is true even for the
1398 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1399 // now come from (one or more) Phis that merge their new-loop equivalents.
1400 // Parameter side_by_side_idom:
1401 // When side_by_size_idom is null, the dominator tree is constructed for
1402 // the clone loop to dominate the original. Used in construction of
1403 // pre-main-post loop sequence.
1404 // When nonnull, the clone and original are side-by-side, both are
1405 // dominated by the passed in side_by_side_idom node. Used in
1406 // construction of unswitched loops.
1407 enum CloneLoopMode {
1408 IgnoreStripMined = 0, // Only clone inner strip mined loop
1409 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1410 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1411 // result control flow branches
1412 // either to inner clone or outer
1413 // strip mined loop.
1414 };
1415 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1416 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1417 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1418 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1419 Node_List*& split_if_set, Node_List*& split_bool_set,
1420 Node_List*& split_cex_set, Node_List& worklist,
1421 uint new_counter, CloneLoopMode mode);
1422 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1423 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1424 Node_List& extra_data_nodes);
1425
1426 // If we got the effect of peeling, either by actually peeling or by
1427 // making a pre-loop which must execute at least once, we can remove
1428 // all loop-invariant dominated tests in the main body.
1429 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1430
1431 // Generate code to do a loop peel for the given loop (and body).
1432 // old_new is a temp array.
1433 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1434
1435 // Add pre and post loops around the given loop. These loops are used
1436 // during RCE, unrolling and aligning loops.
1437 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1438
1439 // Find the last store in the body of an OuterStripMinedLoop when following memory uses
1440 Node *find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop);
1441
1442 // Add post loop after the given loop.
1443 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1444 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1445 Node* incr, Node* limit, CountedLoopNode*& post_head);
1446
1447 // Add a vector post loop between a vector main loop and the current post loop
1448 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1449 // If Node n lives in the back_ctrl block, we clone a private version of n
1450 // in preheader_ctrl block and return that, otherwise return n.
1451 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1452
1453 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1454 // unroll to do double iterations. The next round of major loop transforms
1455 // will repeat till the doubled loop body does all remaining iterations in 1
1456 // pass.
1457 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1458
1459 // Unroll the loop body one step - make each trip do 2 iterations.
1460 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1461
1462 // Return true if exp is a constant times an induction var
1463 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1464
1465 bool is_iv(Node* exp, Node* iv, BasicType bt);
1466
1467 // Return true if exp is a scaled induction var plus (or minus) constant
1468 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);
1469 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1470 jlong long_scale;
1471 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1472 int int_scale = checked_cast<int>(long_scale);
1473 if (p_scale != nullptr) {
1474 *p_scale = int_scale;
1475 }
1476 return true;
1477 }
1478 return false;
1479 }
1480 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1481 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1482 BasicType bt,
1483 jlong* p_scale, Node** p_offset,
1484 bool* p_short_scale, int depth);
1485
1486 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1487 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1488 Deoptimization::DeoptReason reason, int opcode,
1489 bool rewire_uncommon_proj_phi_inputs = false);
1490
1491 private:
1492 // Helper functions for create_new_if_for_predicate()
1493 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1494 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1495 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1496 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1497 const OrigToNewHashtable& orig_to_clone);
1498 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
1499 Node* loop_entry);
1500
1501 public:
1502 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1503
1504 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1505 void replace_control(Node* node, Node* new_control) {
1506 _igvn.replace_input_of(node, 0, new_control);
1507 set_idom(node, new_control, dom_depth(new_control));
1508 }
1509
1510 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1511 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1512 set_idom(loop_head, new_entry, dom_depth(new_entry));
1513 }
1514
1515 // Construct a range check for a predicate if
1516 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1517 jint stride, Node* range, bool upper, bool& overflow);
1518
1519 // Implementation of the loop predication to promote checks outside the loop
1520 bool loop_predication_impl(IdealLoopTree *loop);
1521
1522 private:
1523 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1524 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1525 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1526 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1527 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1528 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1529 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1530 Node_List& if_proj_list);
1531 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1532 IfProjNode* new_control, int scale, Node* offset, Node* range);
1533 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1534
1535 // Helper function to collect predicate for eliminating the useless ones
1536 void eliminate_useless_predicates() const;
1537
1538 void eliminate_useless_zero_trip_guard();
1539 void eliminate_useless_multiversion_if();
1540
1541 public:
1542 // Change the control input of expensive nodes to allow commoning by
1543 // IGVN when it is guaranteed to not result in a more frequent
1544 // execution of the expensive node. Return true if progress.
1545 bool process_expensive_nodes();
1546
1547 // Check whether node has become unreachable
1548 bool is_node_unreachable(Node *n) const {
1549 return !has_node(n) || n->is_unreachable(_igvn);
1550 }
1551
1552 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1553 void do_range_check(IdealLoopTree* loop);
1554
1555 // Clone loop with an invariant test (that does not exit) and
1556 // insert a clone of the test that selects which version to
1557 // execute.
1558 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1559
1560 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1561
1562 private:
1563 static bool has_control_dependencies_from_predicates(LoopNode* head);
1564 static void revert_to_normal_loop(const LoopNode* loop_head);
1565
1566 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1567 const UnswitchedLoopSelector& unswitched_loop_selector);
1568 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1569 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1570 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1571 #ifndef PRODUCT
1572 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1573 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1574 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1575 const LoopNode* original_head, const LoopNode* new_head);
1576 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1577 const LoopNode* original_head, const LoopNode* new_head);
1578 #endif
1579
1580 public:
1581
1582 // Range Check Elimination uses this function!
1583 // Constrain the main loop iterations so the affine function:
1584 // low_limit <= scale_con * I + offset < upper_limit
1585 // always holds true. That is, either increase the number of iterations in
1586 // the pre-loop or the post-loop until the condition holds true in the main
1587 // loop. Scale_con, offset and limit are all loop invariant.
1588 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);
1589 // Helper function for add_constraint().
1590 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1591
1592 // Partially peel loop up through last_peel node.
1593 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1594 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1595
1596 // AutoVectorize the loop: replace scalar ops with vector ops.
1597 enum AutoVectorizeStatus {
1598 Impossible, // This loop has the wrong shape to even try vectorization.
1599 Success, // We just successfully vectorized the loop.
1600 TriedAndFailed, // We tried to vectorize, but failed.
1601 };
1602 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1603
1604 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1605 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1606 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1607 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1608
1609 // Create a scheduled list of nodes control dependent on ctrl set.
1610 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1611 // Has a use in the vector set
1612 bool has_use_in_set( Node* n, VectorSet& vset );
1613 // Has use internal to the vector set (ie. not in a phi at the loop head)
1614 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1615 // clone "n" for uses that are outside of loop
1616 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1617 // clone "n" for special uses that are in the not_peeled region
1618 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1619 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1620 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1621 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1622 #ifdef ASSERT
1623 // Validate the loop partition sets: peel and not_peel
1624 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1625 // Ensure that uses outside of loop are of the right form
1626 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1627 uint orig_exit_idx, uint clone_exit_idx);
1628 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1629 #endif
1630
1631 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1632 int stride_of_possible_iv( Node* iff );
1633 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1634 // Return the (unique) control output node that's in the loop (if it exists.)
1635 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1636 // Insert a signed compare loop exit cloned from an unsigned compare.
1637 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1638 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1639 // Utility to register node "n" with PhaseIdealLoop
1640 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1641 // Utility to create an if-projection
1642 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1643 // Force the iff control output to be the live_proj
1644 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1645 // Insert a region before an if projection
1646 RegionNode* insert_region_before_proj(ProjNode* proj);
1647 // Insert a new if before an if projection
1648 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1649
1650 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1651 // "Nearly" because all Nodes have been cloned from the original in the loop,
1652 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1653 // through the Phi recursively, and return a Bool.
1654 Node* clone_iff(PhiNode* phi);
1655 CmpNode* clone_bool(PhiNode* phi);
1656
1657
1658 // Rework addressing expressions to get the most loop-invariant stuff
1659 // moved out. We'd like to do all associative operators, but it's especially
1660 // important (common) to do address expressions.
1661 Node* remix_address_expressions(Node* n);
1662 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1663
1664 // Convert add to muladd to generate MuladdS2I under certain criteria
1665 Node * convert_add_to_muladd(Node * n);
1666
1667 // Attempt to use a conditional move instead of a phi/branch
1668 Node *conditional_move( Node *n );
1669
1670 // Check for aggressive application of 'split-if' optimization,
1671 // using basic block level info.
1672 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1673 Node *split_if_with_blocks_pre ( Node *n );
1674 void split_if_with_blocks_post( Node *n );
1675 Node *has_local_phi_input( Node *n );
1676 // Mark an IfNode as being dominated by a prior test,
1677 // without actually altering the CFG (and hence IDOM info).
1678 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool pin_array_access_nodes = false);
1679 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool pin_array_access_nodes);
1680
1681 // Split Node 'n' through merge point
1682 RegionNode* split_thru_region(Node* n, RegionNode* region);
1683 // Split Node 'n' through merge point if there is enough win.
1684 Node *split_thru_phi( Node *n, Node *region, int policy );
1685 // Found an If getting its condition-code input from a Phi in the
1686 // same block. Split thru the Region.
1687 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1688
1689 private:
1690 // Class to keep track of wins in split_thru_phi.
1691 class SplitThruPhiWins {
1692 private:
1693 // Region containing the phi we are splitting through.
1694 const Node* _region;
1695
1696 // Sum of all wins regardless of where they happen. This applies to Loops phis as well as non-loop phis.
1697 int _total_wins;
1698
1699 // For Loops, wins have different impact depending on if they happen on loop entry or on the backedge.
1700 // Number of wins on a loop entry edge if the split is through a loop head,
1701 // otherwise 0. Entry edge wins only pay dividends once on loop entry.
1702 int _loop_entry_wins;
1703 // Number of wins on a loop back-edge, which pay dividends on every iteration.
1704 int _loop_back_wins;
1705
1706 public:
1707 SplitThruPhiWins(const Node* region) :
1708 _region(region),
1709 _total_wins(0),
1710 _loop_entry_wins(0),
1711 _loop_back_wins(0) {};
1712
1713 void reset() {_total_wins = 0; _loop_entry_wins = 0; _loop_back_wins = 0;}
1714 void add_win(int ctrl_index) {
1715 if (_region->is_Loop() && ctrl_index == LoopNode::EntryControl) {
1716 _loop_entry_wins++;
1717 } else if (_region->is_Loop() && ctrl_index == LoopNode::LoopBackControl) {
1718 _loop_back_wins++;
1719 }
1720 _total_wins++;
1721 }
1722 // Is this split profitable with respect to the policy?
1723 bool profitable(int policy) const {
1724 assert(_region->is_Loop() || (_loop_entry_wins == 0 && _loop_back_wins == 0), "wins on loop edges without a loop");
1725 assert(!_region->is_Loop() || _total_wins == _loop_entry_wins + _loop_back_wins, "missed some win");
1726 // In general this means that the split has to have more wins than specified
1727 // in the policy. However, for loops we need to take into account where the
1728 // wins happen. We need to be careful when splitting, because splitting nodes
1729 // related to the iv through the phi can sufficiently rearrange the loop
1730 // structure to prevent RCE and thus vectorization. Thus, we only deem splitting
1731 // profitable if the win of a split is not on the entry edge, as such wins
1732 // only pay off once and have a high chance of messing up the loop structure.
1733 return (_loop_entry_wins == 0 && _total_wins > policy) ||
1734 // If there are wins on the entry edge but the backadge also has sufficient wins,
1735 // there is sufficient profitability to spilt regardless of the risk of messing
1736 // up the loop structure.
1737 _loop_back_wins > policy ||
1738 // If the policy is less than 0, a split is always profitable, i.e. we always
1739 // split. This is needed when we split a node and then must also split a
1740 // dependant node, i.e. spliting a Bool node after splitting a Cmp node.
1741 policy < 0;
1742 }
1743 };
1744
1745 public:
1746
1747 // Conversion of fill/copy patterns into intrinsic versions
1748 bool do_intrinsify_fill();
1749 bool intrinsify_fill(IdealLoopTree* lpt);
1750 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1751 Node*& shift, Node*& offset);
1752
1753 private:
1754 // Return a type based on condition control flow
1755 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1756 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1757 // Helpers for filtered type
1758 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1759
1760 // Helper functions
1761 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1762 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1763 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 );
1764 bool split_up( Node *n, Node *blk1, Node *blk2 );
1765
1766 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1767 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1768 void try_move_store_after_loop(Node* n);
1769 bool identical_backtoback_ifs(Node *n);
1770 bool can_split_if(Node *n_ctrl);
1771 bool cannot_split_division(const Node* n, const Node* region) const;
1772 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1773 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1774
1775 // Determine if a method is too big for a/another round of split-if, based on
1776 // a magic (approximate) ratio derived from the equally magic constant 35000,
1777 // previously used for this purpose (but without relating to the node limit).
1778 bool must_throttle_split_if() {
1779 uint threshold = C->max_node_limit() * 2 / 5;
1780 return C->live_nodes() > threshold;
1781 }
1782
1783 // A simplistic node request tracking mechanism, where
1784 // = UINT_MAX Request not valid or made final.
1785 // < UINT_MAX Nodes currently requested (estimate).
1786 uint _nodes_required;
1787
1788 enum { REQUIRE_MIN = 70 };
1789
1790 uint nodes_required() const { return _nodes_required; }
1791
1792 // Given the _currently_ available number of nodes, check whether there is
1793 // "room" for an additional request or not, considering the already required
1794 // number of nodes. Return TRUE if the new request is exceeding the node
1795 // budget limit, otherwise return FALSE. Note that this interpretation will
1796 // act pessimistic on additional requests when new nodes have already been
1797 // generated since the 'begin'. This behaviour fits with the intention that
1798 // node estimates/requests should be made upfront.
1799 bool exceeding_node_budget(uint required = 0) {
1800 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1801 uint available = C->max_node_limit() - C->live_nodes();
1802 return available < required + _nodes_required + REQUIRE_MIN;
1803 }
1804
1805 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1806 precond(require > 0);
1807 _nodes_required += MAX2(require, minreq);
1808 return _nodes_required;
1809 }
1810
1811 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1812 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1813 }
1814
1815 uint require_nodes_begin() {
1816 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1817 _nodes_required = 0;
1818 return C->live_nodes();
1819 }
1820
1821 // When a node request is final, optionally check that the requested number
1822 // of nodes was reasonably correct with respect to the number of new nodes
1823 // introduced since the last 'begin'. Always check that we have not exceeded
1824 // the maximum node limit.
1825 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1826 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1827
1828 #ifdef ASSERT
1829 if (check_estimate) {
1830 // Check that the node budget request was not off by too much (x2).
1831 // Should this be the case we _surely_ need to improve the estimates
1832 // used in our budget calculations.
1833 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1834 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1835 C->live_nodes() - live_at_begin, _nodes_required);
1836 }
1837 }
1838 #endif
1839 // Assert that we have stayed within the node budget limit.
1840 assert(C->live_nodes() < C->max_node_limit(),
1841 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1842 C->live_nodes() - live_at_begin, live_at_begin,
1843 C->max_node_limit(), _nodes_required);
1844
1845 _nodes_required = UINT_MAX;
1846 }
1847
1848 private:
1849
1850 bool _created_loop_node;
1851 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1852 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1853
1854 public:
1855 void set_created_loop_node() { _created_loop_node = true; }
1856 bool created_loop_node() { return _created_loop_node; }
1857 void register_new_node(Node* n, Node* blk);
1858 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1859 register_new_node(new_node, get_ctrl(ctrl_of));
1860 }
1861
1862 Node* clone_and_register(Node* n, Node* ctrl) {
1863 n = n->clone();
1864 register_new_node(n, ctrl);
1865 return n;
1866 }
1867
1868 #ifdef ASSERT
1869 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1870 #endif
1871
1872 #ifndef PRODUCT
1873 void dump() const;
1874 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1875 void dump_idom(Node* n, uint count) const;
1876 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1877 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1878 IdealLoopTree* get_loop_idx(Node* n) const {
1879 // Dead nodes have no loop, so return the top level loop instead
1880 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1881 }
1882 // Print some stats
1883 static void print_statistics();
1884 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1885 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1886 static volatile int _long_loop_candidates;
1887 static volatile int _long_loop_nests;
1888 static volatile int _long_loop_counted_loops;
1889 #endif
1890
1891 #ifdef ASSERT
1892 void verify() const;
1893 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1894 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1895 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1896 #endif
1897
1898 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1899
1900 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1901
1902 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1903
1904
1905 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1906 Node_List &range_checks);
1907
1908 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1909 Node* inner_iters_actual_int, Node* inner_phi,
1910 Node* iv_add, LoopNode* inner_head);
1911
1912 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1913
1914 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1915
1916 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1917
1918 bool would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl);
1919
1920 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1921
1922 void try_sink_out_of_loop(Node* n);
1923
1924 Node* clamp(Node* R, Node* L, Node* H);
1925
1926 bool safe_for_if_replacement(const Node* dom) const;
1927
1928 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1929
1930 bool try_merge_identical_ifs(Node* n);
1931
1932 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1933
1934 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1935 IdealLoopTree* parent, bool partial);
1936
1937 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1938 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1939
1940 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1941 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1942 Node_List*& split_bool_set, Node_List*& split_cex_set);
1943
1944 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1945
1946 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1947
1948 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1949 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1950 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1951 void clone_template_assertion_expression_down(Node* node);
1952
1953 Node* similar_subtype_check(const Node* x, Node* r_in);
1954
1955 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1956
1957 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1958
1959 void pin_array_access_nodes_dependent_on(Node* ctrl);
1960
1961 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1962
1963 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
1964
1965 ConINode* intcon(jint i);
1966
1967 ConLNode* longcon(jlong i);
1968
1969 ConNode* makecon(const Type* t);
1970
1971 ConNode* integercon(jlong l, BasicType bt);
1972
1973 ConNode* zerocon(BasicType bt);
1974 };
1975
1976
1977 class AutoNodeBudget : public StackObj
1978 {
1979 public:
1980 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1981
1982 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1983 : _phase(phase),
1984 _check_at_final(chk == BUDGET_CHECK),
1985 _nodes_at_begin(0)
1986 {
1987 precond(_phase != nullptr);
1988
1989 _nodes_at_begin = _phase->require_nodes_begin();
1990 }
1991
1992 ~AutoNodeBudget() {
1993 #ifndef PRODUCT
1994 if (TraceLoopOpts) {
1995 uint request = _phase->nodes_required();
1996 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1997
1998 if (request < delta) {
1999 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
2000 } else {
2001 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
2002 // Identify the worst estimates as "poor" ones.
2003 if (request > REQUIRE_MIN && delta > 0) {
2004 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
2005 (delta <= REQUIRE_MIN && request > 10 * delta)) {
2006 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
2007 }
2008 }
2009 }
2010 }
2011 #endif // PRODUCT
2012 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
2013 }
2014
2015 private:
2016 PhaseIdealLoop* _phase;
2017 bool _check_at_final;
2018 uint _nodes_at_begin;
2019 };
2020
2021 inline Node* IdealLoopTree::tail() {
2022 // Handle lazy update of _tail field.
2023 if (_tail->in(0) == nullptr) {
2024 _tail = _phase->get_ctrl(_tail);
2025 }
2026 return _tail;
2027 }
2028
2029 inline Node* IdealLoopTree::head() {
2030 // Handle lazy update of _head field.
2031 if (_head->in(0) == nullptr) {
2032 _head = _phase->get_ctrl(_head);
2033 }
2034 return _head;
2035 }
2036
2037 // Iterate over the loop tree using a preorder, left-to-right traversal.
2038 //
2039 // Example that visits all counted loops from within PhaseIdealLoop
2040 //
2041 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2042 // IdealLoopTree* lpt = iter.current();
2043 // if (!lpt->is_counted()) continue;
2044 // ...
2045 class LoopTreeIterator : public StackObj {
2046 private:
2047 IdealLoopTree* _root;
2048 IdealLoopTree* _curnt;
2049
2050 public:
2051 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
2052
2053 bool done() { return _curnt == nullptr; } // Finished iterating?
2054
2055 void next(); // Advance to next loop tree
2056
2057 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
2058 };
2059
2060 // Compute probability of reaching some CFG node from a fixed
2061 // dominating CFG node
2062 class PathFrequency {
2063 private:
2064 Node* _dom; // frequencies are computed relative to this node
2065 Node_Stack _stack;
2066 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
2067 GrowableArray<float> _freqs; // cache frequencies
2068 PhaseIdealLoop* _phase;
2069
2070 float check_and_truncate_frequency(float f) {
2071 assert(f >= 0, "Incorrect frequency");
2072 // We do not perform an exact (f <= 1) check
2073 // this would be error prone with rounding of floats.
2074 // Performing a check like (f <= 1+eps) would be of benefit,
2075 // however, it is not evident how to determine such an eps,
2076 // given that an arbitrary number of add/mul operations
2077 // are performed on these frequencies.
2078 return (f > 1) ? 1 : f;
2079 }
2080
2081 public:
2082 PathFrequency(Node* dom, PhaseIdealLoop* phase)
2083 : _dom(dom), _stack(0), _phase(phase) {
2084 }
2085
2086 float to(Node* n);
2087 };
2088
2089 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
2090 // 1. Clone the data nodes
2091 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
2092 class DataNodeGraph : public StackObj {
2093 PhaseIdealLoop* const _phase;
2094 const Unique_Node_List& _data_nodes;
2095 OrigToNewHashtable _orig_to_new;
2096
2097 public:
2098 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
2099 : _phase(phase),
2100 _data_nodes(data_nodes),
2101 // Use 107 as best guess which is the first resize value in ResizeableHashTable::large_table_sizes.
2102 _orig_to_new(107, MaxNodeLimit)
2103 {
2104 #ifdef ASSERT
2105 for (uint i = 0; i < data_nodes.size(); i++) {
2106 assert(!data_nodes[i]->is_CFG(), "only data nodes");
2107 }
2108 #endif
2109 }
2110 NONCOPYABLE(DataNodeGraph);
2111
2112 private:
2113 void clone(Node* node, Node* new_ctrl);
2114 void clone_data_nodes(Node* new_ctrl);
2115 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2116 Node* new_ctrl);
2117 void rewire_clones_to_cloned_inputs();
2118 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
2119
2120 public:
2121 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
2122 // Set 'new_ctrl' as ctrl for the cloned nodes.
2123 const OrigToNewHashtable& clone(Node* new_ctrl) {
2124 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
2125 clone_data_nodes(new_ctrl);
2126 rewire_clones_to_cloned_inputs();
2127 return _orig_to_new;
2128 }
2129
2130 // Create a copy of the data nodes provided to the constructor by doing the following:
2131 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
2132 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2133 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2134 // `new_ctrl` as ctrl.
2135 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2136 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2137 Node* new_ctrl) {
2138 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2139 rewire_clones_to_cloned_inputs();
2140 return _orig_to_new;
2141 }
2142 };
2143 #endif // SHARE_OPTO_LOOPNODE_HPP