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 UnswitchCandidate;
47 class LoopSelector;
48 class UnswitchedLoopSelector;
49 class VectorSet;
50 class VSharedData;
51 class Invariance;
52 struct small_cache;
53
54 //
55 // I D E A L I Z E D L O O P S
56 //
57 // Idealized loops are the set of loops I perform more interesting
58 // transformations on, beyond simple hoisting.
59
60 //------------------------------LoopNode---------------------------------------
61 // Simple loop header. Fall in path on left, loop-back path on right.
62 class LoopNode : public RegionNode {
63 // Size is bigger to hold the flags. However, the flags do not change
64 // the semantics so it does not appear in the hash & cmp functions.
65 virtual uint size_of() const { return sizeof(*this); }
66 protected:
67 uint _loop_flags;
68 // Names for flag bitfields
69 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
70 MainHasNoPreLoop = 1<<2,
71 HasExactTripCount = 1<<3,
72 InnerLoop = 1<<4,
73 PartialPeelLoop = 1<<5,
74 PartialPeelFailed = 1<<6,
75 WasSlpAnalyzed = 1<<7,
76 PassedSlpAnalysis = 1<<8,
77 DoUnrollOnly = 1<<9,
78 VectorizedLoop = 1<<10,
79 HasAtomicPostLoop = 1<<11,
80 StripMined = 1<<12,
81 SubwordLoop = 1<<13,
82 ProfileTripFailed = 1<<14,
83 LoopNestInnerLoop = 1<<15,
84 LoopNestLongOuterLoop = 1<<16,
85 MultiversionFastLoop = 1<<17,
86 MultiversionSlowLoop = 2<<17,
87 MultiversionDelayedSlowLoop = 3<<17,
88 MultiversionFlagsMask = 3<<17,
89 FlatArrays = 1<<18};
90 char _unswitch_count;
91 enum { _unswitch_max=3 };
92
93 // Expected trip count from profile data
94 float _profile_trip_cnt;
95
96 public:
97 // Names for edge indices
98 enum { Self=0, EntryControl, LoopBackControl };
99
100 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
101 void set_inner_loop() { _loop_flags |= InnerLoop; }
102
103 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
104 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
105 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
106 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
107 bool is_strip_mined() const { return _loop_flags & StripMined; }
108 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
109 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
110 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
111 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
112 bool is_flat_arrays() const { return _loop_flags & FlatArrays; }
113
114 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
115 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
116 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
117 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
118 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
119 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
120 void mark_strip_mined() { _loop_flags |= StripMined; }
121 void clear_strip_mined() { _loop_flags &= ~StripMined; }
122 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
123 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
124 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
125 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
126 void mark_flat_arrays() { _loop_flags |= FlatArrays; }
127
128 int unswitch_max() { return _unswitch_max; }
129 int unswitch_count() { return _unswitch_count; }
130
131 void set_unswitch_count(int val) {
132 assert (val <= unswitch_max(), "too many unswitches");
133 _unswitch_count = val;
134 }
135
136 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
137 float profile_trip_cnt() { return _profile_trip_cnt; }
138
139 LoopNode(Node *entry, Node *backedge)
140 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
141 _profile_trip_cnt(COUNT_UNKNOWN) {
142 init_class_id(Class_Loop);
143 init_req(EntryControl, entry);
144 init_req(LoopBackControl, backedge);
145 }
146
147 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
148 virtual int Opcode() const;
149 bool can_be_counted_loop(PhaseValues* phase) const {
150 return req() == 3 && in(0) != nullptr &&
151 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
152 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
153 }
154 bool is_valid_counted_loop(BasicType bt) const;
155 #ifndef PRODUCT
156 virtual void dump_spec(outputStream *st) const;
157 #endif
158
159 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
160 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
161 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
162 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
163 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
164 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
165 };
166
167 //------------------------------Counted Loops----------------------------------
168 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
169 // path (and maybe some other exit paths). The trip-counter exit is always
170 // last in the loop. The trip-counter have to stride by a constant;
171 // the exit value is also loop invariant.
172
173 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
174 // CountedLoopNode has the incoming loop control and the loop-back-control
175 // which is always the IfTrue before the matching CountedLoopEndNode. The
176 // CountedLoopEndNode has an incoming control (possibly not the
177 // CountedLoopNode if there is control flow in the loop), the post-increment
178 // trip-counter value, and the limit. The trip-counter value is always of
179 // the form (Op old-trip-counter stride). The old-trip-counter is produced
180 // by a Phi connected to the CountedLoopNode. The stride is constant.
181 // The Op is any commutable opcode, including Add, Mul, Xor. The
182 // CountedLoopEndNode also takes in the loop-invariant limit value.
183
184 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
185 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
186 // via the old-trip-counter from the Op node.
187
188 //------------------------------CountedLoopNode--------------------------------
189 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
190 // inputs the incoming loop-start control and the loop-back control, so they
191 // act like RegionNodes. They also take in the initial trip counter, the
192 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
193 // produce a loop-body control and the trip counter value. Since
194 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
195
196 class BaseCountedLoopNode : public LoopNode {
197 public:
198 BaseCountedLoopNode(Node *entry, Node *backedge)
199 : LoopNode(entry, backedge) {
200 }
201
202 Node *init_control() const { return in(EntryControl); }
203 Node *back_control() const { return in(LoopBackControl); }
204
205 Node* init_trip() const;
206 Node* stride() const;
207 bool stride_is_con() const;
208 Node* limit() const;
209 Node* incr() const;
210 Node* phi() const;
211
212 BaseCountedLoopEndNode* loopexit_or_null() const;
213 BaseCountedLoopEndNode* loopexit() const;
214
215 virtual BasicType bt() const = 0;
216
217 jlong stride_con() const;
218
219 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
220 };
221
222
223 class CountedLoopNode : public BaseCountedLoopNode {
224 // Size is bigger to hold _main_idx. However, _main_idx does not change
225 // the semantics so it does not appear in the hash & cmp functions.
226 virtual uint size_of() const { return sizeof(*this); }
227
228 // For Pre- and Post-loops during debugging ONLY, this holds the index of
229 // the Main CountedLoop. Used to assert that we understand the graph shape.
230 node_idx_t _main_idx;
231
232 // Known trip count calculated by compute_exact_trip_count()
233 uint _trip_count;
234
235 // Log2 of original loop bodies in unrolled loop
236 int _unrolled_count_log2;
237
238 // Node count prior to last unrolling - used to decide if
239 // unroll,optimize,unroll,optimize,... is making progress
240 int _node_count_before_unroll;
241
242 // If slp analysis is performed we record the maximum
243 // vector mapped unroll factor here
244 int _slp_maximum_unroll_factor;
245
246 public:
247 CountedLoopNode(Node *entry, Node *backedge)
248 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
249 _unrolled_count_log2(0), _node_count_before_unroll(0),
250 _slp_maximum_unroll_factor(0) {
251 init_class_id(Class_CountedLoop);
252 // Initialize _trip_count to the largest possible value.
253 // Will be reset (lower) if the loop's trip count is known.
254 }
255
256 virtual int Opcode() const;
257 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
258
259 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
260 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
261 int stride_con() const;
262
263 // Match increment with optional truncation
264 static Node*
265 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
266 BasicType bt);
267
268 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
269 // can run short a few iterations and may start a few iterations in.
270 // It will be RCE'd and unrolled and aligned.
271
272 // A following 'post' loop will run any remaining iterations. Used
273 // during Range Check Elimination, the 'post' loop will do any final
274 // iterations with full checks. Also used by Loop Unrolling, where
275 // the 'post' loop will do any epilog iterations needed. Basically,
276 // a 'post' loop can not profitably be further unrolled or RCE'd.
277
278 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
279 // it may do under-flow checks for RCE and may do alignment iterations
280 // so the following main loop 'knows' that it is striding down cache
281 // lines.
282
283 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
284 // Aligned, may be missing it's pre-loop.
285 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
286 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
287 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
288 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
289 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
290 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
291 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
292 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
293 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
294 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
295
296 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
297
298 int main_idx() const { return _main_idx; }
299
300
301 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
302 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
303 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
304 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
305
306 void set_trip_count(uint tc) { _trip_count = tc; }
307 uint trip_count() { return _trip_count; }
308
309 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
310 void set_exact_trip_count(uint tc) {
311 _trip_count = tc;
312 _loop_flags |= HasExactTripCount;
313 }
314 void set_nonexact_trip_count() {
315 _loop_flags &= ~HasExactTripCount;
316 }
317 void set_notpassed_slp() {
318 _loop_flags &= ~PassedSlpAnalysis;
319 }
320
321 void double_unrolled_count() { _unrolled_count_log2++; }
322 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
323
324 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
325 int node_count_before_unroll() { return _node_count_before_unroll; }
326 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
327 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
328
329 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
330 // decides which one is taken:
331 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
332 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
333 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
334 // assumption fails we take the slow_loop instead, where we do not make the same speculative
335 // assumption.
336 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
337 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
338 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
339 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
340 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
341 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
342 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
343 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
344 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
345 // Hence, we expect the slow_loop to execute slower.
346 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
347 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
348 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
349 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
350 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
351 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
352 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
353 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
354
355 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
356 OuterStripMinedLoopNode* outer_loop() const;
357 virtual IfTrueNode* outer_loop_tail() const;
358 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
359 virtual IfFalseNode* outer_loop_exit() const;
360 virtual SafePointNode* outer_safepoint() const;
361
362 Node* skip_assertion_predicates_with_halt();
363
364 virtual BasicType bt() const {
365 return T_INT;
366 }
367
368 Node* is_canonical_loop_entry();
369 CountedLoopEndNode* find_pre_loop_end();
370
371 Node* uncasted_init_trip(bool uncasted);
372
373 #ifndef PRODUCT
374 virtual void dump_spec(outputStream *st) const;
375 #endif
376 };
377
378 class LongCountedLoopNode : public BaseCountedLoopNode {
379 public:
380 LongCountedLoopNode(Node *entry, Node *backedge)
381 : BaseCountedLoopNode(entry, backedge) {
382 init_class_id(Class_LongCountedLoop);
383 }
384
385 virtual int Opcode() const;
386
387 virtual BasicType bt() const {
388 return T_LONG;
389 }
390
391 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
392 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
393 };
394
395
396 //------------------------------CountedLoopEndNode-----------------------------
397 // CountedLoopEndNodes end simple trip counted loops. They act much like
398 // IfNodes.
399
400 class BaseCountedLoopEndNode : public IfNode {
401 public:
402 enum { TestControl, TestValue };
403 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
404 : IfNode(control, test, prob, cnt) {
405 init_class_id(Class_BaseCountedLoopEnd);
406 }
407
408 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
409 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
410 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
411 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
412 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
413 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
414
415 PhiNode* phi() const {
416 Node* tmp = incr();
417 if (tmp && tmp->req() == 3) {
418 Node* phi = tmp->in(1);
419 if (phi->is_Phi()) {
420 return phi->as_Phi();
421 }
422 }
423 return nullptr;
424 }
425
426 BaseCountedLoopNode* loopnode() const {
427 // The CountedLoopNode that goes with this CountedLoopEndNode may
428 // have been optimized out by the IGVN so be cautious with the
429 // pattern matching on the graph
430 PhiNode* iv_phi = phi();
431 if (iv_phi == nullptr) {
432 return nullptr;
433 }
434 Node* ln = iv_phi->in(0);
435 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
436 return nullptr;
437 }
438 if (ln->as_BaseCountedLoop()->bt() != bt()) {
439 return nullptr;
440 }
441 return ln->as_BaseCountedLoop();
442 }
443
444 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
445
446 jlong stride_con() const;
447 virtual BasicType bt() const = 0;
448
449 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
450 };
451
452 class CountedLoopEndNode : public BaseCountedLoopEndNode {
453 public:
454
455 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
456 : BaseCountedLoopEndNode(control, test, prob, cnt) {
457 init_class_id(Class_CountedLoopEnd);
458 }
459 virtual int Opcode() const;
460
461 CountedLoopNode* loopnode() const {
462 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
463 }
464
465 virtual BasicType bt() const {
466 return T_INT;
467 }
468
469 #ifndef PRODUCT
470 virtual void dump_spec(outputStream *st) const;
471 #endif
472 };
473
474 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
475 public:
476 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
477 : BaseCountedLoopEndNode(control, test, prob, cnt) {
478 init_class_id(Class_LongCountedLoopEnd);
479 }
480
481 LongCountedLoopNode* loopnode() const {
482 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
483 }
484
485 virtual int Opcode() const;
486
487 virtual BasicType bt() const {
488 return T_LONG;
489 }
490 };
491
492
493 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
494 Node* bctrl = back_control();
495 if (bctrl == nullptr) return nullptr;
496
497 Node* lexit = bctrl->in(0);
498 if (!lexit->is_BaseCountedLoopEnd()) {
499 return nullptr;
500 }
501 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
502 if (result->bt() != bt()) {
503 return nullptr;
504 }
505 return result;
506 }
507
508 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
509 BaseCountedLoopEndNode* cle = loopexit_or_null();
510 assert(cle != nullptr, "loopexit is null");
511 return cle;
512 }
513
514 inline Node* BaseCountedLoopNode::init_trip() const {
515 BaseCountedLoopEndNode* cle = loopexit_or_null();
516 return cle != nullptr ? cle->init_trip() : nullptr;
517 }
518 inline Node* BaseCountedLoopNode::stride() const {
519 BaseCountedLoopEndNode* cle = loopexit_or_null();
520 return cle != nullptr ? cle->stride() : nullptr;
521 }
522
523 inline bool BaseCountedLoopNode::stride_is_con() const {
524 BaseCountedLoopEndNode* cle = loopexit_or_null();
525 return cle != nullptr && cle->stride_is_con();
526 }
527 inline Node* BaseCountedLoopNode::limit() const {
528 BaseCountedLoopEndNode* cle = loopexit_or_null();
529 return cle != nullptr ? cle->limit() : nullptr;
530 }
531 inline Node* BaseCountedLoopNode::incr() const {
532 BaseCountedLoopEndNode* cle = loopexit_or_null();
533 return cle != nullptr ? cle->incr() : nullptr;
534 }
535 inline Node* BaseCountedLoopNode::phi() const {
536 BaseCountedLoopEndNode* cle = loopexit_or_null();
537 return cle != nullptr ? cle->phi() : nullptr;
538 }
539
540 inline jlong BaseCountedLoopNode::stride_con() const {
541 BaseCountedLoopEndNode* cle = loopexit_or_null();
542 return cle != nullptr ? cle->stride_con() : 0;
543 }
544
545
546 //------------------------------LoopLimitNode-----------------------------
547 // Counted Loop limit node which represents exact final iterator value:
548 // trip_count = (limit - init_trip + stride - 1)/stride
549 // final_value= trip_count * stride + init_trip.
550 // Use HW instructions to calculate it when it can overflow in integer.
551 // Note, final_value should fit into integer since counted loop has
552 // limit check: limit <= max_int-stride.
553 class LoopLimitNode : public Node {
554 enum { Init=1, Limit=2, Stride=3 };
555 public:
556 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
557 // Put it on the Macro nodes list to optimize during macro nodes expansion.
558 init_flags(Flag_is_macro);
559 C->add_macro_node(this);
560 }
561 virtual int Opcode() const;
562 virtual const Type *bottom_type() const { return TypeInt::INT; }
563 virtual uint ideal_reg() const { return Op_RegI; }
564 virtual const Type* Value(PhaseGVN* phase) const;
565 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
566 virtual Node* Identity(PhaseGVN* phase);
567 };
568
569 // Support for strip mining
570 class OuterStripMinedLoopNode : public LoopNode {
571 private:
572 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
573
574 public:
575 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
576 : LoopNode(entry, backedge) {
577 init_class_id(Class_OuterStripMinedLoop);
578 init_flags(Flag_is_macro);
579 C->add_macro_node(this);
580 }
581
582 virtual int Opcode() const;
583
584 virtual IfTrueNode* outer_loop_tail() const;
585 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
586 virtual IfFalseNode* outer_loop_exit() const;
587 virtual SafePointNode* outer_safepoint() const;
588 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
589
590 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
591
592 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
593
594 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
595
596 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
597 PhaseIdealLoop* iloop);
598 };
599
600 class OuterStripMinedLoopEndNode : public IfNode {
601 public:
602 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
603 : IfNode(control, test, prob, cnt) {
604 init_class_id(Class_OuterStripMinedLoopEnd);
605 }
606
607 virtual int Opcode() const;
608
609 virtual const Type* Value(PhaseGVN* phase) const;
610 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
611
612 bool is_expanded(PhaseGVN *phase) const;
613 };
614
615 // -----------------------------IdealLoopTree----------------------------------
616 class IdealLoopTree : public ResourceObj {
617 public:
618 IdealLoopTree *_parent; // Parent in loop tree
619 IdealLoopTree *_next; // Next sibling in loop tree
620 IdealLoopTree *_child; // First child in loop tree
621
622 // The head-tail backedge defines the loop.
623 // If a loop has multiple backedges, this is addressed during cleanup where
624 // we peel off the multiple backedges, merging all edges at the bottom and
625 // ensuring that one proper backedge flow into the loop.
626 Node *_head; // Head of loop
627 Node *_tail; // Tail of loop
628 inline Node *tail(); // Handle lazy update of _tail field
629 inline Node *head(); // Handle lazy update of _head field
630 PhaseIdealLoop* _phase;
631 int _local_loop_unroll_limit;
632 int _local_loop_unroll_factor;
633
634 Node_List _body; // Loop body for inner loops
635
636 uint16_t _nest; // Nesting depth
637 uint8_t _irreducible:1, // True if irreducible
638 _has_call:1, // True if has call safepoint
639 _has_sfpt:1, // True if has non-call safepoint
640 _rce_candidate:1, // True if candidate for range check elimination
641 _has_range_checks:1,
642 _has_range_checks_computed:1;
643
644 Node_List* _safepts; // List of safepoints in this loop
645 Node_List* _required_safept; // A inner loop cannot delete these safepts;
646 bool _allow_optimizations; // Allow loop optimizations
647
648 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
649 : _parent(nullptr), _next(nullptr), _child(nullptr),
650 _head(head), _tail(tail),
651 _phase(phase),
652 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
653 _body(Compile::current()->comp_arena()),
654 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
655 _has_range_checks(0), _has_range_checks_computed(0),
656 _safepts(nullptr),
657 _required_safept(nullptr),
658 _allow_optimizations(true)
659 {
660 precond(_head != nullptr);
661 precond(_tail != nullptr);
662 }
663
664 // Is 'l' a member of 'this'?
665 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
666
667 // Set loop nesting depth. Accumulate has_call bits.
668 int set_nest( uint depth );
669
670 // Split out multiple fall-in edges from the loop header. Move them to a
671 // private RegionNode before the loop. This becomes the loop landing pad.
672 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
673
674 // Split out the outermost loop from this shared header.
675 void split_outer_loop( PhaseIdealLoop *phase );
676
677 // Merge all the backedges from the shared header into a private Region.
678 // Feed that region as the one backedge to this loop.
679 void merge_many_backedges( PhaseIdealLoop *phase );
680
681 // Split shared headers and insert loop landing pads.
682 // Insert a LoopNode to replace the RegionNode.
683 // Returns TRUE if loop tree is structurally changed.
684 bool beautify_loops( PhaseIdealLoop *phase );
685
686 // Perform optimization to use the loop predicates for null checks and range checks.
687 // Applies to any loop level (not just the innermost one)
688 bool loop_predication( PhaseIdealLoop *phase);
689 bool can_apply_loop_predication();
690
691 // Perform iteration-splitting on inner loops. Split iterations to
692 // avoid range checks or one-shot null checks. Returns false if the
693 // current round of loop opts should stop.
694 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
695
696 // Driver for various flavors of iteration splitting. Returns false
697 // if the current round of loop opts should stop.
698 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
699
700 // Given dominators, try to find loops with calls that must always be
701 // executed (call dominates loop tail). These loops do not need non-call
702 // safepoints (ncsfpt).
703 void check_safepts(VectorSet &visited, Node_List &stack);
704
705 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
706 // encountered.
707 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
708
709 // Remove safepoints from loop. Optionally keeping one.
710 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
711
712 // Convert to counted loops where possible
713 void counted_loop( PhaseIdealLoop *phase );
714
715 // Check for Node being a loop-breaking test
716 Node *is_loop_exit(Node *iff) const;
717
718 // Remove simplistic dead code from loop body
719 void DCE_loop_body();
720
721 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
722 // Replace with a 1-in-10 exit guess.
723 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
724
725 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
726 // Useful for unrolling loops with NO array accesses.
727 bool policy_peel_only( PhaseIdealLoop *phase ) const;
728
729 // Return TRUE or FALSE if the loop should be unswitched -- clone
730 // loop with an invariant test
731 bool policy_unswitching( PhaseIdealLoop *phase ) const;
732 bool no_unswitch_candidate() const;
733
734 // Micro-benchmark spamming. Remove empty loops.
735 bool do_remove_empty_loop( PhaseIdealLoop *phase );
736
737 // Convert one iteration loop into normal code.
738 bool do_one_iteration_loop( PhaseIdealLoop *phase );
739
740 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
741 // move some loop-invariant test (usually a null-check) before the loop.
742 bool policy_peeling(PhaseIdealLoop *phase);
743
744 uint estimate_peeling(PhaseIdealLoop *phase);
745
746 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
747 // known trip count in the counted loop node.
748 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
749
750 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
751 // if the loop is a counted loop and the loop body is small enough.
752 bool policy_unroll(PhaseIdealLoop *phase);
753
754 // Loop analyses to map to a maximal superword unrolling for vectorization.
755 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
756
757 // Return TRUE or FALSE if the loop should be range-check-eliminated.
758 // Gather a list of IF tests that are dominated by iteration splitting;
759 // also gather the end of the first split and the start of the 2nd split.
760 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
761
762 // Return TRUE if "iff" is a range check.
763 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
764 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
765 jlong& scale) const;
766
767 // Estimate the number of nodes required when cloning a loop (body).
768 uint est_loop_clone_sz(uint factor) const;
769 // Estimate the number of nodes required when unrolling a loop (body).
770 uint est_loop_unroll_sz(uint factor) const;
771
772 // Compute loop trip count if possible
773 void compute_trip_count(PhaseIdealLoop* phase);
774
775 // Compute loop trip count from profile data
776 float compute_profile_trip_cnt_helper(Node* n);
777 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
778
779 // Reassociate invariant expressions.
780 void reassociate_invariants(PhaseIdealLoop *phase);
781 // Reassociate invariant binary expressions.
782 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
783 // Reassociate invariant add, subtract, and compare expressions.
784 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
785 // Return nonzero index of invariant operand if invariant and variant
786 // are combined with an associative binary. Helper for reassociate_invariants.
787 int find_invariant(Node* n, PhaseIdealLoop *phase);
788 // Return TRUE if "n" is associative.
789 bool is_associative(Node* n, Node* base=nullptr);
790 // Return TRUE if "n" is an associative cmp node.
791 bool is_associative_cmp(Node* n);
792
793 // Return true if n is invariant
794 bool is_invariant(Node* n) const;
795
796 // Put loop body on igvn work list
797 void record_for_igvn();
798
799 bool is_root() { return _parent == nullptr; }
800 // A proper/reducible loop w/o any (occasional) dead back-edge.
801 bool is_loop() { return !_irreducible && !tail()->is_top(); }
802 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
803 bool is_innermost() { return is_loop() && _child == nullptr; }
804
805 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
806
807 bool compute_has_range_checks() const;
808 bool range_checks_present() {
809 if (!_has_range_checks_computed) {
810 if (compute_has_range_checks()) {
811 _has_range_checks = 1;
812 }
813 _has_range_checks_computed = 1;
814 }
815 return _has_range_checks;
816 }
817
818 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
819 // In all other cases, return this.
820 IdealLoopTree* skip_strip_mined() {
821 return _head->as_Loop()->is_strip_mined() ? _parent : this;
822 }
823
824 #ifndef PRODUCT
825 void dump_head(); // Dump loop head only
826 void dump(); // Dump this loop recursively
827 #endif
828
829 #ifdef ASSERT
830 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
831 bool verify_tree(IdealLoopTree* loop_verify) const;
832 #endif
833
834 private:
835 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
836
837 // Estimate the number of nodes resulting from control and data flow merge.
838 uint est_loop_flow_merge_sz() const;
839
840 // Check if the number of residual iterations is large with unroll_cnt.
841 // Return true if the residual iterations are more than 10% of the trip count.
842 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
843 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
844 }
845
846 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
847
848 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
849
850 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
851
852 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
853 Node* sfpt) const;
854
855 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
856
857 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
858 };
859
860 // -----------------------------PhaseIdealLoop---------------------------------
861 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
862 // into a loop tree. Drives the loop-based transformations on the ideal graph.
863 class PhaseIdealLoop : public PhaseTransform {
864 friend class IdealLoopTree;
865 friend class SuperWord;
866 friend class ShenandoahBarrierC2Support;
867 friend class AutoNodeBudget;
868
869 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
870 Node_List _loop_or_ctrl;
871
872 // Pre-computed def-use info
873 PhaseIterGVN &_igvn;
874
875 // Head of loop tree
876 IdealLoopTree* _ltree_root;
877
878 // Array of pre-order numbers, plus post-visited bit.
879 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
880 // ODD for post-visited. Other bits are the pre-order number.
881 uint *_preorders;
882 uint _max_preorder;
883
884 ReallocMark _nesting; // Safety checks for arena reallocation
885
886 const PhaseIdealLoop* _verify_me;
887 bool _verify_only;
888
889 // Allocate _preorders[] array
890 void allocate_preorders() {
891 _max_preorder = C->unique()+8;
892 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
893 memset(_preorders, 0, sizeof(uint) * _max_preorder);
894 }
895
896 // Allocate _preorders[] array
897 void reallocate_preorders() {
898 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
899 if ( _max_preorder < C->unique() ) {
900 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
901 _max_preorder = C->unique();
902 }
903 memset(_preorders, 0, sizeof(uint) * _max_preorder);
904 }
905
906 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
907 // adds new nodes.
908 void check_grow_preorders( ) {
909 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
910 if ( _max_preorder < C->unique() ) {
911 uint newsize = _max_preorder<<1; // double size of array
912 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
913 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
914 _max_preorder = newsize;
915 }
916 }
917 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
918 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
919 // Pre-order numbers are written to the Nodes array as low-bit-set values.
920 void set_preorder_visited( Node *n, int pre_order ) {
921 assert( !is_visited( n ), "already set" );
922 _preorders[n->_idx] = (pre_order<<1);
923 };
924 // Return pre-order number.
925 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
926
927 // Check for being post-visited.
928 // Should be previsited already (checked with assert(is_visited(n))).
929 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
930
931 // Mark as post visited
932 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
933
934 public:
935 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
936 // Returns true if "n" is a data node, false if it's a control node.
937 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
938
939 private:
940 // clear out dead code after build_loop_late
941 Node_List _deadlist;
942 Node_List _zero_trip_guard_opaque_nodes;
943 Node_List _multiversion_opaque_nodes;
944
945 // Support for faster execution of get_late_ctrl()/dom_lca()
946 // when a node has many uses and dominator depth is deep.
947 GrowableArray<jlong> _dom_lca_tags;
948 uint _dom_lca_tags_round;
949 void init_dom_lca_tags();
950
951 // Helper for debugging bad dominance relationships
952 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
953
954 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
955
956 // Inline wrapper for frequent cases:
957 // 1) only one use
958 // 2) a use is the same as the current LCA passed as 'n1'
959 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
960 assert( n->is_CFG(), "" );
961 // Fast-path null lca
962 if( lca != nullptr && lca != n ) {
963 assert( lca->is_CFG(), "" );
964 // find LCA of all uses
965 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
966 }
967 return find_non_split_ctrl(n);
968 }
969 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
970
971 // Helper function for directing control inputs away from CFG split points.
972 Node *find_non_split_ctrl( Node *ctrl ) const {
973 if (ctrl != nullptr) {
974 if (ctrl->is_MultiBranch()) {
975 ctrl = ctrl->in(0);
976 }
977 assert(ctrl->is_CFG(), "CFG");
978 }
979 return ctrl;
980 }
981
982 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
983
984 #ifdef ASSERT
985 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
986 #endif
987 private:
988 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
989 Unique_Node_List& list);
990 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
991 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
992 CountedLoopNode* remaining_loop_head,
993 uint first_node_index_in_cloned_loop_body,
994 const Node_List& old_new);
995 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
996 CountedLoopNode* main_loop_head,
997 uint first_node_index_in_pre_loop_body,
998 uint last_node_index_in_pre_loop_body,
999 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1000 const Node_List& old_new);
1001 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1002 uint first_node_index_in_cloned_loop_body);
1003 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1004 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1005 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1006 CountedLoopNode* target_loop_head,
1007 const NodeInLoopBody& _node_in_loop_body,
1008 bool kill_old_template);
1009 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1010 const Node* old_target_loop_entry,
1011 uint node_index_before_new_assertion_predicate_nodes);
1012 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
1013 Node* bol);
1014 void log_loop_tree();
1015
1016 public:
1017
1018 PhaseIterGVN &igvn() const { return _igvn; }
1019
1020 bool has_node(const Node* n) const {
1021 guarantee(n != nullptr, "No Node.");
1022 return _loop_or_ctrl[n->_idx] != nullptr;
1023 }
1024 // check if transform created new nodes that need _ctrl recorded
1025 Node *get_late_ctrl( Node *n, Node *early );
1026 Node *get_early_ctrl( Node *n );
1027 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1028 void set_early_ctrl(Node* n, bool update_body);
1029 void set_subtree_ctrl(Node* n, bool update_body);
1030 void set_ctrl( Node *n, Node *ctrl ) {
1031 assert( !has_node(n) || has_ctrl(n), "" );
1032 assert( ctrl->in(0), "cannot set dead control node" );
1033 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1034 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1035 }
1036 void set_root_as_ctrl(Node* n) {
1037 assert(!has_node(n) || has_ctrl(n), "");
1038 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1039 }
1040 // Set control and update loop membership
1041 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1042 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1043 IdealLoopTree* new_loop = get_loop(ctrl);
1044 if (old_loop != new_loop) {
1045 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1046 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1047 }
1048 set_ctrl(n, ctrl);
1049 }
1050 // Control nodes can be replaced or subsumed. During this pass they
1051 // get their replacement Node in slot 1. Instead of updating the block
1052 // location of all Nodes in the subsumed block, we lazily do it. As we
1053 // pull such a subsumed block out of the array, we write back the final
1054 // correct block.
1055 Node* get_ctrl(const Node* i) {
1056 assert(has_node(i), "");
1057 Node *n = get_ctrl_no_update(i);
1058 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1059 assert(has_node(i) && has_ctrl(i), "");
1060 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1061 return n;
1062 }
1063
1064 bool is_dominator(Node* dominator, Node* n);
1065 bool is_strict_dominator(Node* dominator, Node* n);
1066
1067 // return get_ctrl for a data node and self(n) for a CFG node
1068 Node* ctrl_or_self(Node* n) {
1069 if (has_ctrl(n))
1070 return get_ctrl(n);
1071 else {
1072 assert (n->is_CFG(), "must be a CFG node");
1073 return n;
1074 }
1075 }
1076
1077 Node* get_ctrl_no_update_helper(const Node* i) const {
1078 assert(has_ctrl(i), "should be control, not loop");
1079 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1080 }
1081
1082 Node* get_ctrl_no_update(const Node* i) const {
1083 assert( has_ctrl(i), "" );
1084 Node *n = get_ctrl_no_update_helper(i);
1085 if (!n->in(0)) {
1086 // Skip dead CFG nodes
1087 do {
1088 n = get_ctrl_no_update_helper(n);
1089 } while (!n->in(0));
1090 n = find_non_split_ctrl(n);
1091 }
1092 return n;
1093 }
1094
1095 // Check for loop being set
1096 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1097 bool has_loop( Node *n ) const {
1098 assert(!has_node(n) || !has_ctrl(n), "");
1099 return has_node(n);
1100 }
1101 // Set loop
1102 void set_loop( Node *n, IdealLoopTree *loop ) {
1103 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1104 }
1105 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1106 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1107 // from old_node to new_node to support the lazy update. Reference
1108 // replaces loop reference, since that is not needed for dead node.
1109 void lazy_update(Node *old_node, Node *new_node) {
1110 assert(old_node != new_node, "no cycles please");
1111 // Re-use the side array slot for this node to provide the
1112 // forwarding pointer.
1113 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1114 }
1115 void lazy_replace(Node *old_node, Node *new_node) {
1116 _igvn.replace_node(old_node, new_node);
1117 lazy_update(old_node, new_node);
1118 }
1119
1120 private:
1121
1122 // Place 'n' in some loop nest, where 'n' is a CFG node
1123 void build_loop_tree();
1124 int build_loop_tree_impl(Node* n, int pre_order);
1125 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1126 // loop tree, not the root.
1127 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1128
1129 #ifdef ASSERT
1130 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1131 void verify_regions_in_irreducible_loops();
1132 bool is_in_irreducible_loop(RegionNode* region);
1133 #endif
1134
1135 // Place Data nodes in some loop nest
1136 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1137 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1138 void build_loop_late_post_work(Node* n, bool pinned);
1139 void build_loop_late_post(Node* n);
1140 void verify_strip_mined_scheduling(Node *n, Node* least);
1141
1142 // Array of immediate dominance info for each CFG node indexed by node idx
1143 private:
1144 uint _idom_size;
1145 Node **_idom; // Array of immediate dominators
1146 uint *_dom_depth; // Used for fast LCA test
1147 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1148 LoopOptsMode _mode;
1149
1150 // build the loop tree and perform any requested optimizations
1151 void build_and_optimize();
1152
1153 // Dominators for the sea of nodes
1154 void Dominators();
1155
1156 // Compute the Ideal Node to Loop mapping
1157 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1158 PhaseTransform(Ideal_Loop),
1159 _loop_or_ctrl(igvn.C->comp_arena()),
1160 _igvn(igvn),
1161 _verify_me(nullptr),
1162 _verify_only(false),
1163 _mode(mode),
1164 _nodes_required(UINT_MAX) {
1165 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1166 build_and_optimize();
1167 }
1168
1169 #ifndef PRODUCT
1170 // Verify that verify_me made the same decisions as a fresh run
1171 // or only verify that the graph is valid if verify_me is null.
1172 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1173 PhaseTransform(Ideal_Loop),
1174 _loop_or_ctrl(igvn.C->comp_arena()),
1175 _igvn(igvn),
1176 _verify_me(verify_me),
1177 _verify_only(verify_me == nullptr),
1178 _mode(LoopOptsVerify),
1179 _nodes_required(UINT_MAX) {
1180 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1181 build_and_optimize();
1182 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1183 }
1184 #endif
1185
1186 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1187
1188 public:
1189 Node* idom_no_update(Node* d) const {
1190 return idom_no_update(d->_idx);
1191 }
1192
1193 Node* idom_no_update(uint didx) const {
1194 assert(didx < _idom_size, "oob");
1195 Node* n = _idom[didx];
1196 assert(n != nullptr,"Bad immediate dominator info.");
1197 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1198 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1199 assert(n != nullptr,"Bad immediate dominator info.");
1200 }
1201 return n;
1202 }
1203
1204 Node *idom(Node* d) const {
1205 return idom(d->_idx);
1206 }
1207
1208 Node *idom(uint didx) const {
1209 Node *n = idom_no_update(didx);
1210 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1211 return n;
1212 }
1213
1214 uint dom_depth(Node* d) const {
1215 guarantee(d != nullptr, "Null dominator info.");
1216 guarantee(d->_idx < _idom_size, "");
1217 return _dom_depth[d->_idx];
1218 }
1219 void set_idom(Node* d, Node* n, uint dom_depth);
1220 // Locally compute IDOM using dom_lca call
1221 Node *compute_idom( Node *region ) const;
1222 // Recompute dom_depth
1223 void recompute_dom_depth();
1224
1225 // Is safept not required by an outer loop?
1226 bool is_deleteable_safept(Node* sfpt);
1227
1228 // Replace parallel induction variable (parallel to trip counter)
1229 void replace_parallel_iv(IdealLoopTree *loop);
1230
1231 Node *dom_lca( Node *n1, Node *n2 ) const {
1232 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1233 }
1234 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1235
1236 Node* dominated_node(Node* c1, Node* c2) {
1237 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1238 return is_dominator(c1, c2) ? c2 : c1;
1239 }
1240
1241 // Return control node that's dominated by the 2 others
1242 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1243 return dominated_node(c1, dominated_node(c2, c3));
1244 }
1245
1246 // Build and verify the loop tree without modifying the graph. This
1247 // is useful to verify that all inputs properly dominate their uses.
1248 static void verify(PhaseIterGVN& igvn) {
1249 #ifdef ASSERT
1250 ResourceMark rm;
1251 Compile::TracePhase tp(_t_idealLoopVerify);
1252 PhaseIdealLoop v(igvn);
1253 #endif
1254 }
1255
1256 // Recommended way to use PhaseIdealLoop.
1257 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1258 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1259 ResourceMark rm;
1260 PhaseIdealLoop v(igvn, mode);
1261
1262 Compile* C = Compile::current();
1263 if (!C->failing()) {
1264 // Cleanup any modified bits
1265 igvn.optimize();
1266 if (C->failing()) { return; }
1267 v.log_loop_tree();
1268 }
1269 }
1270
1271 // True if the method has at least 1 irreducible loop
1272 bool _has_irreducible_loops;
1273
1274 // Per-Node transform
1275 virtual Node* transform(Node* n) { return nullptr; }
1276
1277 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1278 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1279 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1280 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1281 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1282
1283 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1284
1285 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1286 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1287 #ifdef ASSERT
1288 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1289 #endif
1290 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1291 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1292 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1293 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1294 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1295 Node*& entry_control, Node*& iffalse);
1296
1297 Node* exact_limit( IdealLoopTree *loop );
1298
1299 // Return a post-walked LoopNode
1300 IdealLoopTree *get_loop( Node *n ) const {
1301 // Dead nodes have no loop, so return the top level loop instead
1302 if (!has_node(n)) return _ltree_root;
1303 assert(!has_ctrl(n), "");
1304 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1305 }
1306
1307 IdealLoopTree* ltree_root() const { return _ltree_root; }
1308
1309 // Is 'n' a (nested) member of 'loop'?
1310 int is_member( const IdealLoopTree *loop, Node *n ) const {
1311 return loop->is_member(get_loop(n)); }
1312
1313 // This is the basic building block of the loop optimizations. It clones an
1314 // entire loop body. It makes an old_new loop body mapping; with this
1315 // mapping you can find the new-loop equivalent to an old-loop node. All
1316 // new-loop nodes are exactly equal to their old-loop counterparts, all
1317 // edges are the same. All exits from the old-loop now have a RegionNode
1318 // that merges the equivalent new-loop path. This is true even for the
1319 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1320 // now come from (one or more) Phis that merge their new-loop equivalents.
1321 // Parameter side_by_side_idom:
1322 // When side_by_size_idom is null, the dominator tree is constructed for
1323 // the clone loop to dominate the original. Used in construction of
1324 // pre-main-post loop sequence.
1325 // When nonnull, the clone and original are side-by-side, both are
1326 // dominated by the passed in side_by_side_idom node. Used in
1327 // construction of unswitched loops.
1328 enum CloneLoopMode {
1329 IgnoreStripMined = 0, // Only clone inner strip mined loop
1330 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1331 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1332 // result control flow branches
1333 // either to inner clone or outer
1334 // strip mined loop.
1335 };
1336 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1337 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1338 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1339 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1340 Node_List*& split_if_set, Node_List*& split_bool_set,
1341 Node_List*& split_cex_set, Node_List& worklist,
1342 uint new_counter, CloneLoopMode mode);
1343 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1344 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1345 Node_List& extra_data_nodes);
1346
1347 // If we got the effect of peeling, either by actually peeling or by
1348 // making a pre-loop which must execute at least once, we can remove
1349 // all loop-invariant dominated tests in the main body.
1350 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1351
1352 // Generate code to do a loop peel for the given loop (and body).
1353 // old_new is a temp array.
1354 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1355
1356 // Add pre and post loops around the given loop. These loops are used
1357 // during RCE, unrolling and aligning loops.
1358 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1359
1360 // Add post loop after the given loop.
1361 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1362 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1363 Node* incr, Node* limit, CountedLoopNode*& post_head);
1364
1365 // Add a vector post loop between a vector main loop and the current post loop
1366 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1367 // If Node n lives in the back_ctrl block, we clone a private version of n
1368 // in preheader_ctrl block and return that, otherwise return n.
1369 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1370
1371 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1372 // unroll to do double iterations. The next round of major loop transforms
1373 // will repeat till the doubled loop body does all remaining iterations in 1
1374 // pass.
1375 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1376
1377 // Unroll the loop body one step - make each trip do 2 iterations.
1378 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1379
1380 // Return true if exp is a constant times an induction var
1381 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1382
1383 bool is_iv(Node* exp, Node* iv, BasicType bt);
1384
1385 // Return true if exp is a scaled induction var plus (or minus) constant
1386 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);
1387 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1388 jlong long_scale;
1389 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1390 int int_scale = checked_cast<int>(long_scale);
1391 if (p_scale != nullptr) {
1392 *p_scale = int_scale;
1393 }
1394 return true;
1395 }
1396 return false;
1397 }
1398 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1399 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1400 BasicType bt,
1401 jlong* p_scale, Node** p_offset,
1402 bool* p_short_scale, int depth);
1403
1404 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1405 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1406 Deoptimization::DeoptReason reason, int opcode,
1407 bool rewire_uncommon_proj_phi_inputs = false);
1408
1409 private:
1410 // Helper functions for create_new_if_for_predicate()
1411 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1412 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1413 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1414 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1415 const OrigToNewHashtable& orig_to_clone);
1416 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
1417 Node* loop_entry);
1418
1419 public:
1420 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1421
1422 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1423 void replace_control(Node* node, Node* new_control) {
1424 _igvn.replace_input_of(node, 0, new_control);
1425 set_idom(node, new_control, dom_depth(new_control));
1426 }
1427
1428 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1429 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1430 set_idom(loop_head, new_entry, dom_depth(new_entry));
1431 }
1432
1433 // Construct a range check for a predicate if
1434 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1435 jint stride, Node* range, bool upper, bool& overflow);
1436
1437 // Implementation of the loop predication to promote checks outside the loop
1438 bool loop_predication_impl(IdealLoopTree *loop);
1439
1440 private:
1441 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1442 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1443 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1444 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1445 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1446 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1447 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1448 Node_List& if_proj_list);
1449 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1450 IfProjNode* new_control, int scale, Node* offset, Node* range);
1451 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1452
1453 // Helper function to collect predicate for eliminating the useless ones
1454 void eliminate_useless_predicates() const;
1455
1456 void eliminate_useless_zero_trip_guard();
1457 void eliminate_useless_multiversion_if();
1458
1459 public:
1460 // Change the control input of expensive nodes to allow commoning by
1461 // IGVN when it is guaranteed to not result in a more frequent
1462 // execution of the expensive node. Return true if progress.
1463 bool process_expensive_nodes();
1464
1465 // Check whether node has become unreachable
1466 bool is_node_unreachable(Node *n) const {
1467 return !has_node(n) || n->is_unreachable(_igvn);
1468 }
1469
1470 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1471 void do_range_check(IdealLoopTree* loop);
1472
1473 // Clone loop with an invariant test (that does not exit) and
1474 // insert a clone of the test that selects which version to
1475 // execute.
1476 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1477
1478 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1479 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1480
1481 private:
1482 static bool has_control_dependencies_from_predicates(LoopNode* head);
1483 static void revert_to_normal_loop(const LoopNode* loop_head);
1484
1485 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1486 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1487 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1488 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1489 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1490 #ifndef PRODUCT
1491 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1492 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1493 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1494 const UnswitchCandidate& unswitch_candidate,
1495 const LoopNode* original_head, const LoopNode* new_head);
1496 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1497 const LoopNode* original_head, const LoopNode* new_head);
1498 #endif
1499
1500 public:
1501
1502 // Range Check Elimination uses this function!
1503 // Constrain the main loop iterations so the affine function:
1504 // low_limit <= scale_con * I + offset < upper_limit
1505 // always holds true. That is, either increase the number of iterations in
1506 // the pre-loop or the post-loop until the condition holds true in the main
1507 // loop. Scale_con, offset and limit are all loop invariant.
1508 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);
1509 // Helper function for add_constraint().
1510 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1511
1512 // Partially peel loop up through last_peel node.
1513 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1514 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1515
1516 // AutoVectorize the loop: replace scalar ops with vector ops.
1517 enum AutoVectorizeStatus {
1518 Impossible, // This loop has the wrong shape to even try vectorization.
1519 Success, // We just successfully vectorized the loop.
1520 TriedAndFailed, // We tried to vectorize, but failed.
1521 };
1522 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1523
1524 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1525 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1526 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1527 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1528
1529 // Move an unordered Reduction out of loop if possible
1530 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop);
1531
1532 // Create a scheduled list of nodes control dependent on ctrl set.
1533 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1534 // Has a use in the vector set
1535 bool has_use_in_set( Node* n, VectorSet& vset );
1536 // Has use internal to the vector set (ie. not in a phi at the loop head)
1537 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1538 // clone "n" for uses that are outside of loop
1539 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1540 // clone "n" for special uses that are in the not_peeled region
1541 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1542 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1543 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1544 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1545 #ifdef ASSERT
1546 // Validate the loop partition sets: peel and not_peel
1547 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1548 // Ensure that uses outside of loop are of the right form
1549 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1550 uint orig_exit_idx, uint clone_exit_idx);
1551 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1552 #endif
1553
1554 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1555 int stride_of_possible_iv( Node* iff );
1556 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1557 // Return the (unique) control output node that's in the loop (if it exists.)
1558 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1559 // Insert a signed compare loop exit cloned from an unsigned compare.
1560 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1561 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1562 // Utility to register node "n" with PhaseIdealLoop
1563 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1564 // Utility to create an if-projection
1565 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1566 // Force the iff control output to be the live_proj
1567 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1568 // Insert a region before an if projection
1569 RegionNode* insert_region_before_proj(ProjNode* proj);
1570 // Insert a new if before an if projection
1571 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1572
1573 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1574 // "Nearly" because all Nodes have been cloned from the original in the loop,
1575 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1576 // through the Phi recursively, and return a Bool.
1577 Node* clone_iff(PhiNode* phi);
1578 CmpNode* clone_bool(PhiNode* phi);
1579
1580
1581 // Rework addressing expressions to get the most loop-invariant stuff
1582 // moved out. We'd like to do all associative operators, but it's especially
1583 // important (common) to do address expressions.
1584 Node* remix_address_expressions(Node* n);
1585 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1586
1587 // Convert add to muladd to generate MuladdS2I under certain criteria
1588 Node * convert_add_to_muladd(Node * n);
1589
1590 // Attempt to use a conditional move instead of a phi/branch
1591 Node *conditional_move( Node *n );
1592
1593 // Check for aggressive application of 'split-if' optimization,
1594 // using basic block level info.
1595 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1596 Node *split_if_with_blocks_pre ( Node *n );
1597 void split_if_with_blocks_post( Node *n );
1598 Node *has_local_phi_input( Node *n );
1599 // Mark an IfNode as being dominated by a prior test,
1600 // without actually altering the CFG (and hence IDOM info).
1601 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool pin_array_access_nodes = false);
1602 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool pin_array_access_nodes);
1603
1604 // Split Node 'n' through merge point
1605 RegionNode* split_thru_region(Node* n, RegionNode* region);
1606 // Split Node 'n' through merge point if there is enough win.
1607 Node *split_thru_phi( Node *n, Node *region, int policy );
1608 // Found an If getting its condition-code input from a Phi in the
1609 // same block. Split thru the Region.
1610 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1611
1612 // Conversion of fill/copy patterns into intrinsic versions
1613 bool do_intrinsify_fill();
1614 bool intrinsify_fill(IdealLoopTree* lpt);
1615 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1616 Node*& shift, Node*& offset);
1617
1618 private:
1619 // Return a type based on condition control flow
1620 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1621 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1622 // Helpers for filtered type
1623 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1624
1625 // Helper functions
1626 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1627 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1628 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 );
1629 bool split_up( Node *n, Node *blk1, Node *blk2 );
1630
1631 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1632 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1633 void try_move_store_after_loop(Node* n);
1634 void move_flat_array_check_out_of_loop(Node* n);
1635 bool identical_backtoback_ifs(Node *n);
1636 bool flat_array_element_type_check(Node *n);
1637 bool can_split_if(Node *n_ctrl);
1638 bool cannot_split_division(const Node* n, const Node* region) const;
1639 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1640 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1641
1642 // Determine if a method is too big for a/another round of split-if, based on
1643 // a magic (approximate) ratio derived from the equally magic constant 35000,
1644 // previously used for this purpose (but without relating to the node limit).
1645 bool must_throttle_split_if() {
1646 uint threshold = C->max_node_limit() * 2 / 5;
1647 return C->live_nodes() > threshold;
1648 }
1649
1650 // A simplistic node request tracking mechanism, where
1651 // = UINT_MAX Request not valid or made final.
1652 // < UINT_MAX Nodes currently requested (estimate).
1653 uint _nodes_required;
1654
1655 enum { REQUIRE_MIN = 70 };
1656
1657 uint nodes_required() const { return _nodes_required; }
1658
1659 // Given the _currently_ available number of nodes, check whether there is
1660 // "room" for an additional request or not, considering the already required
1661 // number of nodes. Return TRUE if the new request is exceeding the node
1662 // budget limit, otherwise return FALSE. Note that this interpretation will
1663 // act pessimistic on additional requests when new nodes have already been
1664 // generated since the 'begin'. This behaviour fits with the intention that
1665 // node estimates/requests should be made upfront.
1666 bool exceeding_node_budget(uint required = 0) {
1667 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1668 uint available = C->max_node_limit() - C->live_nodes();
1669 return available < required + _nodes_required + REQUIRE_MIN;
1670 }
1671
1672 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1673 precond(require > 0);
1674 _nodes_required += MAX2(require, minreq);
1675 return _nodes_required;
1676 }
1677
1678 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1679 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1680 }
1681
1682 uint require_nodes_begin() {
1683 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1684 _nodes_required = 0;
1685 return C->live_nodes();
1686 }
1687
1688 // When a node request is final, optionally check that the requested number
1689 // of nodes was reasonably correct with respect to the number of new nodes
1690 // introduced since the last 'begin'. Always check that we have not exceeded
1691 // the maximum node limit.
1692 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1693 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1694
1695 #ifdef ASSERT
1696 if (check_estimate) {
1697 // Check that the node budget request was not off by too much (x2).
1698 // Should this be the case we _surely_ need to improve the estimates
1699 // used in our budget calculations.
1700 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1701 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1702 C->live_nodes() - live_at_begin, _nodes_required);
1703 }
1704 }
1705 #endif
1706 // Assert that we have stayed within the node budget limit.
1707 assert(C->live_nodes() < C->max_node_limit(),
1708 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1709 C->live_nodes() - live_at_begin, live_at_begin,
1710 C->max_node_limit(), _nodes_required);
1711
1712 _nodes_required = UINT_MAX;
1713 }
1714
1715 private:
1716
1717 bool _created_loop_node;
1718 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1719 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1720
1721 public:
1722 void set_created_loop_node() { _created_loop_node = true; }
1723 bool created_loop_node() { return _created_loop_node; }
1724 void register_new_node(Node* n, Node* blk);
1725 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1726 register_new_node(new_node, get_ctrl(ctrl_of));
1727 }
1728
1729 Node* clone_and_register(Node* n, Node* ctrl) {
1730 n = n->clone();
1731 register_new_node(n, ctrl);
1732 return n;
1733 }
1734
1735 #ifdef ASSERT
1736 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1737 #endif
1738
1739 #ifndef PRODUCT
1740 void dump() const;
1741 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1742 void dump_idom(Node* n, uint count) const;
1743 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1744 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1745 IdealLoopTree* get_loop_idx(Node* n) const {
1746 // Dead nodes have no loop, so return the top level loop instead
1747 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1748 }
1749 // Print some stats
1750 static void print_statistics();
1751 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1752 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1753 static volatile int _long_loop_candidates;
1754 static volatile int _long_loop_nests;
1755 static volatile int _long_loop_counted_loops;
1756 #endif
1757
1758 #ifdef ASSERT
1759 void verify() const;
1760 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1761 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1762 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1763 #endif
1764
1765 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1766
1767 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1768
1769 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1770
1771
1772 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1773 Node_List &range_checks);
1774
1775 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1776 Node* inner_iters_actual_int, Node* inner_phi,
1777 Node* iv_add, LoopNode* inner_head);
1778
1779 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1780
1781 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1782
1783 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1784
1785 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1786
1787 void try_sink_out_of_loop(Node* n);
1788
1789 Node* clamp(Node* R, Node* L, Node* H);
1790
1791 bool safe_for_if_replacement(const Node* dom) const;
1792
1793 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1794
1795 bool try_merge_identical_ifs(Node* n);
1796
1797 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1798
1799 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1800 IdealLoopTree* parent, bool partial);
1801
1802 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1803 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1804
1805 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1806 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1807 Node_List*& split_bool_set, Node_List*& split_cex_set);
1808
1809 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1810
1811 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1812
1813 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1814 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1815 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1816 void clone_template_assertion_expression_down(Node* node);
1817
1818 Node* similar_subtype_check(const Node* x, Node* r_in);
1819
1820 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1821
1822 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1823
1824 void pin_array_access_nodes_dependent_on(Node* ctrl);
1825
1826 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1827
1828 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1829
1830 ConINode* intcon(jint i);
1831
1832 ConLNode* longcon(jlong i);
1833
1834 ConNode* makecon(const Type* t);
1835
1836 ConNode* integercon(jlong l, BasicType bt);
1837
1838 ConNode* zerocon(BasicType bt);
1839 };
1840
1841
1842 class AutoNodeBudget : public StackObj
1843 {
1844 public:
1845 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1846
1847 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1848 : _phase(phase),
1849 _check_at_final(chk == BUDGET_CHECK),
1850 _nodes_at_begin(0)
1851 {
1852 precond(_phase != nullptr);
1853
1854 _nodes_at_begin = _phase->require_nodes_begin();
1855 }
1856
1857 ~AutoNodeBudget() {
1858 #ifndef PRODUCT
1859 if (TraceLoopOpts) {
1860 uint request = _phase->nodes_required();
1861 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1862
1863 if (request < delta) {
1864 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1865 } else {
1866 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1867 // Identify the worst estimates as "poor" ones.
1868 if (request > REQUIRE_MIN && delta > 0) {
1869 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1870 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1871 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1872 }
1873 }
1874 }
1875 }
1876 #endif // PRODUCT
1877 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1878 }
1879
1880 private:
1881 PhaseIdealLoop* _phase;
1882 bool _check_at_final;
1883 uint _nodes_at_begin;
1884 };
1885
1886 inline Node* IdealLoopTree::tail() {
1887 // Handle lazy update of _tail field.
1888 if (_tail->in(0) == nullptr) {
1889 _tail = _phase->get_ctrl(_tail);
1890 }
1891 return _tail;
1892 }
1893
1894 inline Node* IdealLoopTree::head() {
1895 // Handle lazy update of _head field.
1896 if (_head->in(0) == nullptr) {
1897 _head = _phase->get_ctrl(_head);
1898 }
1899 return _head;
1900 }
1901
1902 // Iterate over the loop tree using a preorder, left-to-right traversal.
1903 //
1904 // Example that visits all counted loops from within PhaseIdealLoop
1905 //
1906 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1907 // IdealLoopTree* lpt = iter.current();
1908 // if (!lpt->is_counted()) continue;
1909 // ...
1910 class LoopTreeIterator : public StackObj {
1911 private:
1912 IdealLoopTree* _root;
1913 IdealLoopTree* _curnt;
1914
1915 public:
1916 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1917
1918 bool done() { return _curnt == nullptr; } // Finished iterating?
1919
1920 void next(); // Advance to next loop tree
1921
1922 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1923 };
1924
1925 // Compute probability of reaching some CFG node from a fixed
1926 // dominating CFG node
1927 class PathFrequency {
1928 private:
1929 Node* _dom; // frequencies are computed relative to this node
1930 Node_Stack _stack;
1931 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1932 GrowableArray<float> _freqs; // cache frequencies
1933 PhaseIdealLoop* _phase;
1934
1935 float check_and_truncate_frequency(float f) {
1936 assert(f >= 0, "Incorrect frequency");
1937 // We do not perform an exact (f <= 1) check
1938 // this would be error prone with rounding of floats.
1939 // Performing a check like (f <= 1+eps) would be of benefit,
1940 // however, it is not evident how to determine such an eps,
1941 // given that an arbitrary number of add/mul operations
1942 // are performed on these frequencies.
1943 return (f > 1) ? 1 : f;
1944 }
1945
1946 public:
1947 PathFrequency(Node* dom, PhaseIdealLoop* phase)
1948 : _dom(dom), _stack(0), _phase(phase) {
1949 }
1950
1951 float to(Node* n);
1952 };
1953
1954 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
1955 // 1. Clone the data nodes
1956 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
1957 class DataNodeGraph : public StackObj {
1958 PhaseIdealLoop* const _phase;
1959 const Unique_Node_List& _data_nodes;
1960 OrigToNewHashtable _orig_to_new;
1961
1962 public:
1963 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
1964 : _phase(phase),
1965 _data_nodes(data_nodes),
1966 // Use 107 as best guess which is the first resize value in ResizeableResourceHashtable::large_table_sizes.
1967 _orig_to_new(107, MaxNodeLimit)
1968 {
1969 #ifdef ASSERT
1970 for (uint i = 0; i < data_nodes.size(); i++) {
1971 assert(!data_nodes[i]->is_CFG(), "only data nodes");
1972 }
1973 #endif
1974 }
1975 NONCOPYABLE(DataNodeGraph);
1976
1977 private:
1978 void clone(Node* node, Node* new_ctrl);
1979 void clone_data_nodes(Node* new_ctrl);
1980 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
1981 Node* new_ctrl);
1982 void rewire_clones_to_cloned_inputs();
1983 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
1984
1985 public:
1986 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
1987 // Set 'new_ctrl' as ctrl for the cloned nodes.
1988 const OrigToNewHashtable& clone(Node* new_ctrl) {
1989 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
1990 clone_data_nodes(new_ctrl);
1991 rewire_clones_to_cloned_inputs();
1992 return _orig_to_new;
1993 }
1994
1995 // Create a copy of the data nodes provided to the constructor by doing the following:
1996 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
1997 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
1998 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
1999 // `new_ctrl` as ctrl.
2000 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2001 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2002 Node* new_ctrl) {
2003 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2004 rewire_clones_to_cloned_inputs();
2005 return _orig_to_new;
2006 }
2007 };
2008 #endif // SHARE_OPTO_LOOPNODE_HPP