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
223
224 class CountedLoopNode : public BaseCountedLoopNode {
225 // Size is bigger to hold _main_idx. However, _main_idx does not change
226 // the semantics so it does not appear in the hash & cmp functions.
227 virtual uint size_of() const { return sizeof(*this); }
228
229 // For Pre- and Post-loops during debugging ONLY, this holds the index of
230 // the Main CountedLoop. Used to assert that we understand the graph shape.
231 node_idx_t _main_idx;
232
233 // Known trip count calculated by compute_exact_trip_count()
234 uint _trip_count;
235
236 // Log2 of original loop bodies in unrolled loop
237 int _unrolled_count_log2;
238
239 // Node count prior to last unrolling - used to decide if
240 // unroll,optimize,unroll,optimize,... is making progress
241 int _node_count_before_unroll;
242
243 // If slp analysis is performed we record the maximum
244 // vector mapped unroll factor here
245 int _slp_maximum_unroll_factor;
246
247 public:
248 CountedLoopNode(Node *entry, Node *backedge)
249 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
250 _unrolled_count_log2(0), _node_count_before_unroll(0),
251 _slp_maximum_unroll_factor(0) {
252 init_class_id(Class_CountedLoop);
253 // Initialize _trip_count to the largest possible value.
254 // Will be reset (lower) if the loop's trip count is known.
255 }
256
257 virtual int Opcode() const;
258 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
259
260 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
261 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
262 int stride_con() const;
263
264 // Match increment with optional truncation
265 static Node*
266 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
267 BasicType bt);
268
269 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
270 // can run short a few iterations and may start a few iterations in.
271 // It will be RCE'd and unrolled and aligned.
272
273 // A following 'post' loop will run any remaining iterations. Used
274 // during Range Check Elimination, the 'post' loop will do any final
275 // iterations with full checks. Also used by Loop Unrolling, where
276 // the 'post' loop will do any epilog iterations needed. Basically,
277 // a 'post' loop can not profitably be further unrolled or RCE'd.
278
279 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
280 // it may do under-flow checks for RCE and may do alignment iterations
281 // so the following main loop 'knows' that it is striding down cache
282 // lines.
283
284 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
285 // Aligned, may be missing it's pre-loop.
286 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
287 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
288 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
289 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
290 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
291 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
292 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
293 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
294 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
295 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
296
297 IfNode* find_multiversion_if_from_multiversion_fast_main_loop();
298
299 int main_idx() const { return _main_idx; }
300
301
302 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
303 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
304 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
305 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
306
307 // We use max_juint for the default value of _trip_count to signal it wasn't set.
308 // We shouldn't set _trip_count to max_juint explicitly.
309 void set_trip_count(uint tc) { assert(tc < max_juint, "Cannot set trip count to max_juint"); _trip_count = tc; }
310 uint trip_count() { return _trip_count; }
311
312 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
313 void set_exact_trip_count(uint tc) {
314 assert(tc < max_juint, "Cannot set trip count to max_juint");
315 _trip_count = tc;
316 _loop_flags |= HasExactTripCount;
317 }
318 void set_nonexact_trip_count() {
319 _loop_flags &= ~HasExactTripCount;
320 }
321 void set_notpassed_slp() {
322 _loop_flags &= ~PassedSlpAnalysis;
323 }
324
325 void double_unrolled_count() { _unrolled_count_log2++; }
326 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
327
328 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
329 int node_count_before_unroll() { return _node_count_before_unroll; }
330 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
331 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
332
333 // Multiversioning allows us to duplicate a CountedLoop, and have two versions, and the multiversion_if
334 // decides which one is taken:
335 // (1) fast_loop: We enter this loop by default, by default the multiversion_if has its condition set to
336 // "true", guarded by a OpaqueMultiversioning. If we want to make a speculative assumption
337 // for an optimization, we can add the runtime-check to the multiversion_if, and if the
338 // assumption fails we take the slow_loop instead, where we do not make the same speculative
339 // assumption.
340 // We call it the "fast_loop" because it has more optimizations, enabled by the speculative
341 // runtime-checks at the multiversion_if, and we expect the fast_loop to execute faster.
342 // (2) slow_loop: By default, it is not taken, until a runtime-check is added to the multiversion_if while
343 // optimizing the fast_looop. If such a runtime-check is never added, then after loop-opts
344 // the multiversion_if constant folds to true, and the slow_loop is folded away. To save
345 // compile time, we delay the optimization of the slow_loop until a runtime-check is added
346 // to the multiversion_if, at which point we resume optimizations for the slow_loop.
347 // We call it the "slow_loop" because it has fewer optimizations, since this is the fall-back
348 // loop where we do not make any of the speculative assumptions we make for the fast_loop.
349 // Hence, we expect the slow_loop to execute slower.
350 bool is_multiversion() const { return (_loop_flags & MultiversionFlagsMask) != Normal; }
351 bool is_multiversion_fast_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionFastLoop; }
352 bool is_multiversion_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionSlowLoop; }
353 bool is_multiversion_delayed_slow_loop() const { return (_loop_flags & MultiversionFlagsMask) == MultiversionDelayedSlowLoop; }
354 void set_multiversion_fast_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionFastLoop; }
355 void set_multiversion_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionSlowLoop; }
356 void set_multiversion_delayed_slow_loop() { assert(!is_multiversion(), ""); _loop_flags |= MultiversionDelayedSlowLoop; }
357 void set_no_multiversion() { assert( is_multiversion(), ""); _loop_flags &= ~MultiversionFlagsMask; }
358
359 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
360 OuterStripMinedLoopNode* outer_loop() const;
361 virtual IfTrueNode* outer_loop_tail() const;
362 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
363 virtual IfFalseNode* outer_loop_exit() const;
364 virtual SafePointNode* outer_safepoint() const;
365
366 Node* skip_assertion_predicates_with_halt();
367
368 virtual BasicType bt() const {
369 return T_INT;
370 }
371
372 Node* is_canonical_loop_entry();
373 CountedLoopEndNode* find_pre_loop_end();
374
375 Node* uncasted_init_trip(bool uncasted);
376
377 #ifndef PRODUCT
378 virtual void dump_spec(outputStream *st) const;
379 #endif
380 };
381
382 class LongCountedLoopNode : public BaseCountedLoopNode {
383 public:
384 LongCountedLoopNode(Node *entry, Node *backedge)
385 : BaseCountedLoopNode(entry, backedge) {
386 init_class_id(Class_LongCountedLoop);
387 }
388
389 virtual int Opcode() const;
390
391 virtual BasicType bt() const {
392 return T_LONG;
393 }
394
395 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
396 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
397 };
398
399
400 //------------------------------CountedLoopEndNode-----------------------------
401 // CountedLoopEndNodes end simple trip counted loops. They act much like
402 // IfNodes.
403
404 class BaseCountedLoopEndNode : public IfNode {
405 public:
406 enum { TestControl, TestValue };
407 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
408 : IfNode(control, test, prob, cnt) {
409 init_class_id(Class_BaseCountedLoopEnd);
410 }
411
412 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
413 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
414 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
415 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
416 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
417 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
418
419 PhiNode* phi() const {
420 Node* tmp = incr();
421 if (tmp && tmp->req() == 3) {
422 Node* phi = tmp->in(1);
423 if (phi->is_Phi()) {
424 return phi->as_Phi();
425 }
426 }
427 return nullptr;
428 }
429
430 BaseCountedLoopNode* loopnode() const {
431 // The CountedLoopNode that goes with this CountedLoopEndNode may
432 // have been optimized out by the IGVN so be cautious with the
433 // pattern matching on the graph
434 PhiNode* iv_phi = phi();
435 if (iv_phi == nullptr) {
436 return nullptr;
437 }
438 Node* ln = iv_phi->in(0);
439 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
440 return nullptr;
441 }
442 if (ln->as_BaseCountedLoop()->bt() != bt()) {
443 return nullptr;
444 }
445 return ln->as_BaseCountedLoop();
446 }
447
448 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
449
450 jlong stride_con() const;
451 virtual BasicType bt() const = 0;
452
453 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
454 };
455
456 class CountedLoopEndNode : public BaseCountedLoopEndNode {
457 public:
458
459 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
460 : BaseCountedLoopEndNode(control, test, prob, cnt) {
461 init_class_id(Class_CountedLoopEnd);
462 }
463 virtual int Opcode() const;
464
465 CountedLoopNode* loopnode() const {
466 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
467 }
468
469 virtual BasicType bt() const {
470 return T_INT;
471 }
472
473 #ifndef PRODUCT
474 virtual void dump_spec(outputStream *st) const;
475 #endif
476 };
477
478 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
479 public:
480 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
481 : BaseCountedLoopEndNode(control, test, prob, cnt) {
482 init_class_id(Class_LongCountedLoopEnd);
483 }
484
485 LongCountedLoopNode* loopnode() const {
486 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
487 }
488
489 virtual int Opcode() const;
490
491 virtual BasicType bt() const {
492 return T_LONG;
493 }
494 };
495
496
497 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
498 Node* bctrl = back_control();
499 if (bctrl == nullptr) return nullptr;
500
501 Node* lexit = bctrl->in(0);
502 if (!lexit->is_BaseCountedLoopEnd()) {
503 return nullptr;
504 }
505 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
506 if (result->bt() != bt()) {
507 return nullptr;
508 }
509 return result;
510 }
511
512 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
513 BaseCountedLoopEndNode* cle = loopexit_or_null();
514 assert(cle != nullptr, "loopexit is null");
515 return cle;
516 }
517
518 inline Node* BaseCountedLoopNode::init_trip() const {
519 BaseCountedLoopEndNode* cle = loopexit_or_null();
520 return cle != nullptr ? cle->init_trip() : nullptr;
521 }
522 inline Node* BaseCountedLoopNode::stride() const {
523 BaseCountedLoopEndNode* cle = loopexit_or_null();
524 return cle != nullptr ? cle->stride() : nullptr;
525 }
526
527 inline bool BaseCountedLoopNode::stride_is_con() const {
528 BaseCountedLoopEndNode* cle = loopexit_or_null();
529 return cle != nullptr && cle->stride_is_con();
530 }
531 inline Node* BaseCountedLoopNode::limit() const {
532 BaseCountedLoopEndNode* cle = loopexit_or_null();
533 return cle != nullptr ? cle->limit() : nullptr;
534 }
535 inline Node* BaseCountedLoopNode::incr() const {
536 BaseCountedLoopEndNode* cle = loopexit_or_null();
537 return cle != nullptr ? cle->incr() : nullptr;
538 }
539 inline Node* BaseCountedLoopNode::phi() const {
540 BaseCountedLoopEndNode* cle = loopexit_or_null();
541 return cle != nullptr ? cle->phi() : nullptr;
542 }
543
544 inline jlong BaseCountedLoopNode::stride_con() const {
545 BaseCountedLoopEndNode* cle = loopexit_or_null();
546 return cle != nullptr ? cle->stride_con() : 0;
547 }
548
549
550 //------------------------------LoopLimitNode-----------------------------
551 // Counted Loop limit node which represents exact final iterator value:
552 // trip_count = (limit - init_trip + stride - 1)/stride
553 // final_value= trip_count * stride + init_trip.
554 // Use HW instructions to calculate it when it can overflow in integer.
555 // Note, final_value should fit into integer since counted loop has
556 // limit check: limit <= max_int-stride.
557 class LoopLimitNode : public Node {
558 enum { Init=1, Limit=2, Stride=3 };
559 public:
560 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
561 // Put it on the Macro nodes list to optimize during macro nodes expansion.
562 init_flags(Flag_is_macro);
563 C->add_macro_node(this);
564 }
565 virtual int Opcode() const;
566 virtual const Type *bottom_type() const { return TypeInt::INT; }
567 virtual uint ideal_reg() const { return Op_RegI; }
568 virtual const Type* Value(PhaseGVN* phase) const;
569 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
570 virtual Node* Identity(PhaseGVN* phase);
571 };
572
573 // Support for strip mining
574 class OuterStripMinedLoopNode : public LoopNode {
575 private:
576 void fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) const;
577 void handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn);
578
579 public:
580 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
581 : LoopNode(entry, backedge) {
582 init_class_id(Class_OuterStripMinedLoop);
583 init_flags(Flag_is_macro);
584 C->add_macro_node(this);
585 }
586
587 virtual int Opcode() const;
588
589 virtual IfTrueNode* outer_loop_tail() const;
590 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
591 virtual IfFalseNode* outer_loop_exit() const;
592 virtual SafePointNode* outer_safepoint() const;
593 CountedLoopNode* inner_counted_loop() const { return unique_ctrl_out()->as_CountedLoop(); }
594 CountedLoopEndNode* inner_counted_loop_end() const { return inner_counted_loop()->loopexit(); }
595 IfFalseNode* inner_loop_exit() const { return inner_counted_loop_end()->proj_out(false)->as_IfFalse(); }
596
597 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
598
599 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
600
601 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
602
603 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
604
605 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
606 PhaseIdealLoop* iloop);
607 };
608
609 class OuterStripMinedLoopEndNode : public IfNode {
610 public:
611 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
612 : IfNode(control, test, prob, cnt) {
613 init_class_id(Class_OuterStripMinedLoopEnd);
614 }
615
616 virtual int Opcode() const;
617
618 virtual const Type* Value(PhaseGVN* phase) const;
619 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
620
621 bool is_expanded(PhaseGVN *phase) const;
622 };
623
624 // -----------------------------IdealLoopTree----------------------------------
625 class IdealLoopTree : public ResourceObj {
626 public:
627 IdealLoopTree *_parent; // Parent in loop tree
628 IdealLoopTree *_next; // Next sibling in loop tree
629 IdealLoopTree *_child; // First child in loop tree
630
631 // The head-tail backedge defines the loop.
632 // If a loop has multiple backedges, this is addressed during cleanup where
633 // we peel off the multiple backedges, merging all edges at the bottom and
634 // ensuring that one proper backedge flow into the loop.
635 Node *_head; // Head of loop
636 Node *_tail; // Tail of loop
637 inline Node *tail(); // Handle lazy update of _tail field
638 inline Node *head(); // Handle lazy update of _head field
639 PhaseIdealLoop* _phase;
640 int _local_loop_unroll_limit;
641 int _local_loop_unroll_factor;
642
643 Node_List _body; // Loop body for inner loops
644
645 uint16_t _nest; // Nesting depth
646 uint8_t _irreducible:1, // True if irreducible
647 _has_call:1, // True if has call safepoint
648 _has_sfpt:1, // True if has non-call safepoint
649 _rce_candidate:1, // True if candidate for range check elimination
650 _has_range_checks:1,
651 _has_range_checks_computed:1;
652
653 Node_List* _safepts; // List of safepoints in this loop
654 Node_List* _required_safept; // A inner loop cannot delete these safepts;
655 bool _allow_optimizations; // Allow loop optimizations
656
657 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
658 : _parent(nullptr), _next(nullptr), _child(nullptr),
659 _head(head), _tail(tail),
660 _phase(phase),
661 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
662 _body(Compile::current()->comp_arena()),
663 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
664 _has_range_checks(0), _has_range_checks_computed(0),
665 _safepts(nullptr),
666 _required_safept(nullptr),
667 _allow_optimizations(true)
668 {
669 precond(_head != nullptr);
670 precond(_tail != nullptr);
671 }
672
673 // Is 'l' a member of 'this'?
674 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
675
676 // Set loop nesting depth. Accumulate has_call bits.
677 int set_nest( uint depth );
678
679 // Split out multiple fall-in edges from the loop header. Move them to a
680 // private RegionNode before the loop. This becomes the loop landing pad.
681 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
682
683 // Split out the outermost loop from this shared header.
684 void split_outer_loop( PhaseIdealLoop *phase );
685
686 // Merge all the backedges from the shared header into a private Region.
687 // Feed that region as the one backedge to this loop.
688 void merge_many_backedges( PhaseIdealLoop *phase );
689
690 // Split shared headers and insert loop landing pads.
691 // Insert a LoopNode to replace the RegionNode.
692 // Returns TRUE if loop tree is structurally changed.
693 bool beautify_loops( PhaseIdealLoop *phase );
694
695 // Perform optimization to use the loop predicates for null checks and range checks.
696 // Applies to any loop level (not just the innermost one)
697 bool loop_predication( PhaseIdealLoop *phase);
698 bool can_apply_loop_predication();
699
700 // Perform iteration-splitting on inner loops. Split iterations to
701 // avoid range checks or one-shot null checks. Returns false if the
702 // current round of loop opts should stop.
703 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
704
705 // Driver for various flavors of iteration splitting. Returns false
706 // if the current round of loop opts should stop.
707 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
708
709 // Given dominators, try to find loops with calls that must always be
710 // executed (call dominates loop tail). These loops do not need non-call
711 // safepoints (ncsfpt).
712 void check_safepts(VectorSet &visited, Node_List &stack);
713
714 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
715 // encountered.
716 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
717
718 // Remove safepoints from loop. Optionally keeping one.
719 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
720
721 // Convert to counted loops where possible
722 void counted_loop( PhaseIdealLoop *phase );
723
724 // Check for Node being a loop-breaking test
725 Node *is_loop_exit(Node *iff) const;
726
727 // Remove simplistic dead code from loop body
728 void DCE_loop_body();
729
730 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
731 // Replace with a 1-in-10 exit guess.
732 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
733
734 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
735 // Useful for unrolling loops with NO array accesses.
736 bool policy_peel_only( PhaseIdealLoop *phase ) const;
737
738 // Return TRUE or FALSE if the loop should be unswitched -- clone
739 // loop with an invariant test
740 bool policy_unswitching( PhaseIdealLoop *phase ) const;
741
742 // Micro-benchmark spamming. Remove empty loops.
743 bool do_remove_empty_loop( PhaseIdealLoop *phase );
744
745 // Convert one iteration loop into normal code.
746 bool do_one_iteration_loop( PhaseIdealLoop *phase );
747
748 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
749 // move some loop-invariant test (usually a null-check) before the loop.
750 bool policy_peeling(PhaseIdealLoop *phase);
751
752 uint estimate_peeling(PhaseIdealLoop *phase);
753
754 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
755 // known trip count in the counted loop node.
756 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
757
758 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
759 // if the loop is a counted loop and the loop body is small enough.
760 bool policy_unroll(PhaseIdealLoop *phase);
761
762 // Loop analyses to map to a maximal superword unrolling for vectorization.
763 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
764
765 // Return TRUE or FALSE if the loop should be range-check-eliminated.
766 // Gather a list of IF tests that are dominated by iteration splitting;
767 // also gather the end of the first split and the start of the 2nd split.
768 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
769
770 // Return TRUE if "iff" is a range check.
771 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
772 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
773 jlong& scale) const;
774
775 // Estimate the number of nodes required when cloning a loop (body).
776 uint est_loop_clone_sz(uint factor) const;
777 // Estimate the number of nodes required when unrolling a loop (body).
778 uint est_loop_unroll_sz(uint factor) const;
779
780 // Compute loop trip count if possible
781 void compute_trip_count(PhaseIdealLoop* phase);
782
783 // Compute loop trip count from profile data
784 float compute_profile_trip_cnt_helper(Node* n);
785 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
786
787 // Reassociate invariant expressions.
788 void reassociate_invariants(PhaseIdealLoop *phase);
789 // Reassociate invariant binary expressions.
790 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
791 // Reassociate invariant add, subtract, and compare expressions.
792 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
793 // Return nonzero index of invariant operand if invariant and variant
794 // are combined with an associative binary. Helper for reassociate_invariants.
795 int find_invariant(Node* n, PhaseIdealLoop *phase);
796 // Return TRUE if "n" is associative.
797 bool is_associative(Node* n, Node* base=nullptr);
798 // Return TRUE if "n" is an associative cmp node.
799 bool is_associative_cmp(Node* n);
800
801 // Return true if n is invariant
802 bool is_invariant(Node* n) const;
803
804 // Put loop body on igvn work list
805 void record_for_igvn();
806
807 bool is_root() { return _parent == nullptr; }
808 // A proper/reducible loop w/o any (occasional) dead back-edge.
809 bool is_loop() { return !_irreducible && !tail()->is_top(); }
810 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
811 bool is_innermost() { return is_loop() && _child == nullptr; }
812
813 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
814
815 bool compute_has_range_checks() const;
816 bool range_checks_present() {
817 if (!_has_range_checks_computed) {
818 if (compute_has_range_checks()) {
819 _has_range_checks = 1;
820 }
821 _has_range_checks_computed = 1;
822 }
823 return _has_range_checks;
824 }
825
826 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
827 // In all other cases, return this.
828 IdealLoopTree* skip_strip_mined() {
829 return _head->as_Loop()->is_strip_mined() ? _parent : this;
830 }
831
832 #ifndef PRODUCT
833 void dump_head(); // Dump loop head only
834 void dump(); // Dump this loop recursively
835 #endif
836
837 #ifdef ASSERT
838 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
839 bool verify_tree(IdealLoopTree* loop_verify) const;
840 #endif
841
842 private:
843 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
844
845 // Estimate the number of nodes resulting from control and data flow merge.
846 uint est_loop_flow_merge_sz() const;
847
848 // Check if the number of residual iterations is large with unroll_cnt.
849 // Return true if the residual iterations are more than 10% of the trip count.
850 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
851 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
852 }
853
854 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
855
856 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
857
858 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
859
860 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
861 Node* sfpt) const;
862
863 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
864
865 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
866 };
867
868 // -----------------------------PhaseIdealLoop---------------------------------
869 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
870 // into a loop tree. Drives the loop-based transformations on the ideal graph.
871 class PhaseIdealLoop : public PhaseTransform {
872 friend class IdealLoopTree;
873 friend class SuperWord;
874 friend class ShenandoahBarrierC2Support;
875 friend class AutoNodeBudget;
876
877 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
878 Node_List _loop_or_ctrl;
879
880 // Pre-computed def-use info
881 PhaseIterGVN &_igvn;
882
883 // Head of loop tree
884 IdealLoopTree* _ltree_root;
885
886 // Array of pre-order numbers, plus post-visited bit.
887 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
888 // ODD for post-visited. Other bits are the pre-order number.
889 uint *_preorders;
890 uint _max_preorder;
891
892 ReallocMark _nesting; // Safety checks for arena reallocation
893
894 const PhaseIdealLoop* _verify_me;
895 bool _verify_only;
896
897 // Allocate _preorders[] array
898 void allocate_preorders() {
899 _max_preorder = C->unique()+8;
900 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
901 memset(_preorders, 0, sizeof(uint) * _max_preorder);
902 }
903
904 // Allocate _preorders[] array
905 void reallocate_preorders() {
906 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
907 if ( _max_preorder < C->unique() ) {
908 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
909 _max_preorder = C->unique();
910 }
911 memset(_preorders, 0, sizeof(uint) * _max_preorder);
912 }
913
914 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
915 // adds new nodes.
916 void check_grow_preorders( ) {
917 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
918 if ( _max_preorder < C->unique() ) {
919 uint newsize = _max_preorder<<1; // double size of array
920 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
921 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
922 _max_preorder = newsize;
923 }
924 }
925 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
926 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
927 // Pre-order numbers are written to the Nodes array as low-bit-set values.
928 void set_preorder_visited( Node *n, int pre_order ) {
929 assert( !is_visited( n ), "already set" );
930 _preorders[n->_idx] = (pre_order<<1);
931 };
932 // Return pre-order number.
933 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
934
935 // Check for being post-visited.
936 // Should be previsited already (checked with assert(is_visited(n))).
937 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
938
939 // Mark as post visited
940 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
941
942 public:
943 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
944 // Returns true if "n" is a data node, false if it's a control node.
945 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
946
947 private:
948 // clear out dead code after build_loop_late
949 Node_List _deadlist;
950 Node_List _zero_trip_guard_opaque_nodes;
951 Node_List _multiversion_opaque_nodes;
952
953 // Support for faster execution of get_late_ctrl()/dom_lca()
954 // when a node has many uses and dominator depth is deep.
955 GrowableArray<jlong> _dom_lca_tags;
956 uint _dom_lca_tags_round;
957 void init_dom_lca_tags();
958
959 // Helper for debugging bad dominance relationships
960 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
961
962 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
963
964 // Inline wrapper for frequent cases:
965 // 1) only one use
966 // 2) a use is the same as the current LCA passed as 'n1'
967 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
968 assert( n->is_CFG(), "" );
969 // Fast-path null lca
970 if( lca != nullptr && lca != n ) {
971 assert( lca->is_CFG(), "" );
972 // find LCA of all uses
973 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
974 }
975 return find_non_split_ctrl(n);
976 }
977 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
978
979 // Helper function for directing control inputs away from CFG split points.
980 Node *find_non_split_ctrl( Node *ctrl ) const {
981 if (ctrl != nullptr) {
982 if (ctrl->is_MultiBranch()) {
983 ctrl = ctrl->in(0);
984 }
985 assert(ctrl->is_CFG(), "CFG");
986 }
987 return ctrl;
988 }
989
990 void cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop);
991
992 #ifdef ASSERT
993 static void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
994 #endif
995 private:
996 static void get_opaque_template_assertion_predicate_nodes(ParsePredicateSuccessProj* parse_predicate_proj,
997 Unique_Node_List& list);
998 void update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head, int stride_con_before_unroll);
999 void initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1000 CountedLoopNode* remaining_loop_head,
1001 uint first_node_index_in_cloned_loop_body,
1002 const Node_List& old_new);
1003 void initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1004 CountedLoopNode* main_loop_head,
1005 uint first_node_index_in_pre_loop_body,
1006 uint last_node_index_in_pre_loop_body,
1007 DEBUG_ONLY(uint last_node_index_from_backedge_goo COMMA)
1008 const Node_List& old_new);
1009 void initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head, CountedLoopNode* post_loop_head,
1010 uint first_node_index_in_cloned_loop_body);
1011 void create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head, CountedLoopNode* target_loop_head,
1012 const NodeInLoopBody& _node_in_loop_body, bool kill_old_template);
1013 void create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1014 CountedLoopNode* target_loop_head,
1015 const NodeInLoopBody& _node_in_loop_body,
1016 bool kill_old_template);
1017 void rewire_old_target_loop_entry_dependency_to_new_entry(CountedLoopNode* target_loop_head,
1018 const Node* old_target_loop_entry,
1019 uint node_index_before_new_assertion_predicate_nodes);
1020 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
1021 Node* bol);
1022 void log_loop_tree();
1023
1024 public:
1025
1026 PhaseIterGVN &igvn() const { return _igvn; }
1027
1028 bool has_node(const Node* n) const {
1029 guarantee(n != nullptr, "No Node.");
1030 return _loop_or_ctrl[n->_idx] != nullptr;
1031 }
1032 // check if transform created new nodes that need _ctrl recorded
1033 Node *get_late_ctrl( Node *n, Node *early );
1034 Node *get_early_ctrl( Node *n );
1035 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1036 void set_early_ctrl(Node* n, bool update_body);
1037 void set_subtree_ctrl(Node* n, bool update_body);
1038 void set_ctrl( Node *n, Node *ctrl ) {
1039 assert( !has_node(n) || has_ctrl(n), "" );
1040 assert( ctrl->in(0), "cannot set dead control node" );
1041 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1042 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
1043 }
1044 void set_root_as_ctrl(Node* n) {
1045 assert(!has_node(n) || has_ctrl(n), "");
1046 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)C->root() + 1));
1047 }
1048 // Set control and update loop membership
1049 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1050 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1051 IdealLoopTree* new_loop = get_loop(ctrl);
1052 if (old_loop != new_loop) {
1053 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1054 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1055 }
1056 set_ctrl(n, ctrl);
1057 }
1058 // Control nodes can be replaced or subsumed. During this pass they
1059 // get their replacement Node in slot 1. Instead of updating the block
1060 // location of all Nodes in the subsumed block, we lazily do it. As we
1061 // pull such a subsumed block out of the array, we write back the final
1062 // correct block.
1063 Node* get_ctrl(const Node* i) {
1064 assert(has_node(i), "");
1065 Node *n = get_ctrl_no_update(i);
1066 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1067 assert(has_node(i) && has_ctrl(i), "");
1068 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1069 return n;
1070 }
1071
1072 bool is_dominator(Node* dominator, Node* n);
1073 bool is_strict_dominator(Node* dominator, Node* n);
1074
1075 // return get_ctrl for a data node and self(n) for a CFG node
1076 Node* ctrl_or_self(Node* n) {
1077 if (has_ctrl(n))
1078 return get_ctrl(n);
1079 else {
1080 assert (n->is_CFG(), "must be a CFG node");
1081 return n;
1082 }
1083 }
1084
1085 Node* get_ctrl_no_update_helper(const Node* i) const {
1086 assert(has_ctrl(i), "should be control, not loop");
1087 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1088 }
1089
1090 Node* get_ctrl_no_update(const Node* i) const {
1091 assert( has_ctrl(i), "" );
1092 Node *n = get_ctrl_no_update_helper(i);
1093 if (!n->in(0)) {
1094 // Skip dead CFG nodes
1095 do {
1096 n = get_ctrl_no_update_helper(n);
1097 } while (!n->in(0));
1098 n = find_non_split_ctrl(n);
1099 }
1100 return n;
1101 }
1102
1103 // Check for loop being set
1104 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1105 bool has_loop( Node *n ) const {
1106 assert(!has_node(n) || !has_ctrl(n), "");
1107 return has_node(n);
1108 }
1109 // Set loop
1110 void set_loop( Node *n, IdealLoopTree *loop ) {
1111 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1112 }
1113 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1114 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1115 // from old_node to new_node to support the lazy update. Reference
1116 // replaces loop reference, since that is not needed for dead node.
1117 void lazy_update(Node *old_node, Node *new_node) {
1118 assert(old_node != new_node, "no cycles please");
1119 // Re-use the side array slot for this node to provide the
1120 // forwarding pointer.
1121 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1122 }
1123 void lazy_replace(Node *old_node, Node *new_node) {
1124 _igvn.replace_node(old_node, new_node);
1125 lazy_update(old_node, new_node);
1126 }
1127
1128 private:
1129
1130 // Place 'n' in some loop nest, where 'n' is a CFG node
1131 void build_loop_tree();
1132 int build_loop_tree_impl(Node* n, int pre_order);
1133 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1134 // loop tree, not the root.
1135 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1136
1137 #ifdef ASSERT
1138 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1139 void verify_regions_in_irreducible_loops();
1140 bool is_in_irreducible_loop(RegionNode* region);
1141 #endif
1142
1143 // Place Data nodes in some loop nest
1144 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1145 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1146 void build_loop_late_post_work(Node* n, bool pinned);
1147 void build_loop_late_post(Node* n);
1148 void verify_strip_mined_scheduling(Node *n, Node* least);
1149
1150 // Array of immediate dominance info for each CFG node indexed by node idx
1151 private:
1152 uint _idom_size;
1153 Node **_idom; // Array of immediate dominators
1154 uint *_dom_depth; // Used for fast LCA test
1155 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1156 LoopOptsMode _mode;
1157
1158 // build the loop tree and perform any requested optimizations
1159 void build_and_optimize();
1160
1161 // Dominators for the sea of nodes
1162 void Dominators();
1163
1164 // Compute the Ideal Node to Loop mapping
1165 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1166 PhaseTransform(Ideal_Loop),
1167 _loop_or_ctrl(igvn.C->comp_arena()),
1168 _igvn(igvn),
1169 _verify_me(nullptr),
1170 _verify_only(false),
1171 _mode(mode),
1172 _nodes_required(UINT_MAX) {
1173 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1174 build_and_optimize();
1175 }
1176
1177 #ifndef PRODUCT
1178 // Verify that verify_me made the same decisions as a fresh run
1179 // or only verify that the graph is valid if verify_me is null.
1180 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1181 PhaseTransform(Ideal_Loop),
1182 _loop_or_ctrl(igvn.C->comp_arena()),
1183 _igvn(igvn),
1184 _verify_me(verify_me),
1185 _verify_only(verify_me == nullptr),
1186 _mode(LoopOptsVerify),
1187 _nodes_required(UINT_MAX) {
1188 DEBUG_ONLY(C->set_phase_verify_ideal_loop();)
1189 build_and_optimize();
1190 DEBUG_ONLY(C->reset_phase_verify_ideal_loop();)
1191 }
1192 #endif
1193
1194 Node* insert_convert_node_if_needed(BasicType target, Node* input);
1195
1196 public:
1197 Node* idom_no_update(Node* d) const {
1198 return idom_no_update(d->_idx);
1199 }
1200
1201 Node* idom_no_update(uint didx) const {
1202 assert(didx < _idom_size, "oob");
1203 Node* n = _idom[didx];
1204 assert(n != nullptr,"Bad immediate dominator info.");
1205 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1206 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1207 assert(n != nullptr,"Bad immediate dominator info.");
1208 }
1209 return n;
1210 }
1211
1212 Node *idom(Node* d) const {
1213 return idom(d->_idx);
1214 }
1215
1216 Node *idom(uint didx) const {
1217 Node *n = idom_no_update(didx);
1218 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1219 return n;
1220 }
1221
1222 uint dom_depth(Node* d) const {
1223 guarantee(d != nullptr, "Null dominator info.");
1224 guarantee(d->_idx < _idom_size, "");
1225 return _dom_depth[d->_idx];
1226 }
1227 void set_idom(Node* d, Node* n, uint dom_depth);
1228 // Locally compute IDOM using dom_lca call
1229 Node *compute_idom( Node *region ) const;
1230 // Recompute dom_depth
1231 void recompute_dom_depth();
1232
1233 // Is safept not required by an outer loop?
1234 bool is_deleteable_safept(Node* sfpt);
1235
1236 // Replace parallel induction variable (parallel to trip counter)
1237 void replace_parallel_iv(IdealLoopTree *loop);
1238
1239 Node *dom_lca( Node *n1, Node *n2 ) const {
1240 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1241 }
1242 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1243
1244 Node* dominated_node(Node* c1, Node* c2) {
1245 assert(is_dominator(c1, c2) || is_dominator(c2, c1), "nodes must be related");
1246 return is_dominator(c1, c2) ? c2 : c1;
1247 }
1248
1249 // Return control node that's dominated by the 2 others
1250 Node* dominated_node(Node* c1, Node* c2, Node* c3) {
1251 return dominated_node(c1, dominated_node(c2, c3));
1252 }
1253
1254 // Build and verify the loop tree without modifying the graph. This
1255 // is useful to verify that all inputs properly dominate their uses.
1256 static void verify(PhaseIterGVN& igvn) {
1257 #ifdef ASSERT
1258 ResourceMark rm;
1259 Compile::TracePhase tp(_t_idealLoopVerify);
1260 PhaseIdealLoop v(igvn);
1261 #endif
1262 }
1263
1264 // Recommended way to use PhaseIdealLoop.
1265 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1266 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1267 ResourceMark rm;
1268 PhaseIdealLoop v(igvn, mode);
1269
1270 Compile* C = Compile::current();
1271 if (!C->failing()) {
1272 // Cleanup any modified bits
1273 igvn.optimize();
1274 if (C->failing()) { return; }
1275 v.log_loop_tree();
1276 }
1277 }
1278
1279 // True if the method has at least 1 irreducible loop
1280 bool _has_irreducible_loops;
1281
1282 // Per-Node transform
1283 virtual Node* transform(Node* n) { return nullptr; }
1284
1285 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1286 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1287 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1288 Node* loop_iv_stride(Node* incr, Node*& xphi);
1289 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x);
1290
1291 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1292
1293 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1294 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1295 #ifdef ASSERT
1296 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1297 #endif
1298 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1299 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1300 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1301 IdealLoopTree* create_outer_strip_mined_loop(Node* init_control,
1302 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1303 Node*& entry_control, Node*& iffalse);
1304
1305 Node* exact_limit( IdealLoopTree *loop );
1306
1307 // Return a post-walked LoopNode
1308 IdealLoopTree *get_loop( Node *n ) const {
1309 // Dead nodes have no loop, so return the top level loop instead
1310 if (!has_node(n)) return _ltree_root;
1311 assert(!has_ctrl(n), "");
1312 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1313 }
1314
1315 IdealLoopTree* ltree_root() const { return _ltree_root; }
1316
1317 // Is 'n' a (nested) member of 'loop'?
1318 int is_member( const IdealLoopTree *loop, Node *n ) const {
1319 return loop->is_member(get_loop(n)); }
1320
1321 // This is the basic building block of the loop optimizations. It clones an
1322 // entire loop body. It makes an old_new loop body mapping; with this
1323 // mapping you can find the new-loop equivalent to an old-loop node. All
1324 // new-loop nodes are exactly equal to their old-loop counterparts, all
1325 // edges are the same. All exits from the old-loop now have a RegionNode
1326 // that merges the equivalent new-loop path. This is true even for the
1327 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1328 // now come from (one or more) Phis that merge their new-loop equivalents.
1329 // Parameter side_by_side_idom:
1330 // When side_by_size_idom is null, the dominator tree is constructed for
1331 // the clone loop to dominate the original. Used in construction of
1332 // pre-main-post loop sequence.
1333 // When nonnull, the clone and original are side-by-side, both are
1334 // dominated by the passed in side_by_side_idom node. Used in
1335 // construction of unswitched loops.
1336 enum CloneLoopMode {
1337 IgnoreStripMined = 0, // Only clone inner strip mined loop
1338 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1339 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1340 // result control flow branches
1341 // either to inner clone or outer
1342 // strip mined loop.
1343 };
1344 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1345 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1346 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1347 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1348 Node_List*& split_if_set, Node_List*& split_bool_set,
1349 Node_List*& split_cex_set, Node_List& worklist,
1350 uint new_counter, CloneLoopMode mode);
1351 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1352 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1353 Node_List& extra_data_nodes);
1354
1355 // If we got the effect of peeling, either by actually peeling or by
1356 // making a pre-loop which must execute at least once, we can remove
1357 // all loop-invariant dominated tests in the main body.
1358 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1359
1360 // Generate code to do a loop peel for the given loop (and body).
1361 // old_new is a temp array.
1362 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1363
1364 // Add pre and post loops around the given loop. These loops are used
1365 // during RCE, unrolling and aligning loops.
1366 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1367
1368 // Add post loop after the given loop.
1369 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1370 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1371 Node* incr, Node* limit, CountedLoopNode*& post_head);
1372
1373 // Add a vector post loop between a vector main loop and the current post loop
1374 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1375 // If Node n lives in the back_ctrl block, we clone a private version of n
1376 // in preheader_ctrl block and return that, otherwise return n.
1377 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1378
1379 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1380 // unroll to do double iterations. The next round of major loop transforms
1381 // will repeat till the doubled loop body does all remaining iterations in 1
1382 // pass.
1383 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1384
1385 // Unroll the loop body one step - make each trip do 2 iterations.
1386 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1387
1388 // Return true if exp is a constant times an induction var
1389 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1390
1391 bool is_iv(Node* exp, Node* iv, BasicType bt);
1392
1393 // Return true if exp is a scaled induction var plus (or minus) constant
1394 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);
1395 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1396 jlong long_scale;
1397 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1398 int int_scale = checked_cast<int>(long_scale);
1399 if (p_scale != nullptr) {
1400 *p_scale = int_scale;
1401 }
1402 return true;
1403 }
1404 return false;
1405 }
1406 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1407 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1408 BasicType bt,
1409 jlong* p_scale, Node** p_offset,
1410 bool* p_short_scale, int depth);
1411
1412 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1413 IfTrueNode* create_new_if_for_predicate(const ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1414 Deoptimization::DeoptReason reason, int opcode,
1415 bool rewire_uncommon_proj_phi_inputs = false);
1416
1417 private:
1418 // Helper functions for create_new_if_for_predicate()
1419 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1420 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1421 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1422 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1423 const OrigToNewHashtable& orig_to_clone);
1424 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
1425 Node* loop_entry);
1426
1427 public:
1428 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1429
1430 // Replace the control input of 'node' with 'new_control' and set the dom depth to the one of 'new_control'.
1431 void replace_control(Node* node, Node* new_control) {
1432 _igvn.replace_input_of(node, 0, new_control);
1433 set_idom(node, new_control, dom_depth(new_control));
1434 }
1435
1436 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1437 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1438 set_idom(loop_head, new_entry, dom_depth(new_entry));
1439 }
1440
1441 // Construct a range check for a predicate if
1442 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1443 jint stride, Node* range, bool upper, bool& overflow);
1444
1445 // Implementation of the loop predication to promote checks outside the loop
1446 bool loop_predication_impl(IdealLoopTree *loop);
1447
1448 private:
1449 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1450 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1451 Invariance& invar, Deoptimization::DeoptReason deopt_reason);
1452 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1453 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1454 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1455 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1456 Node_List& if_proj_list);
1457 IfTrueNode* create_template_assertion_predicate(CountedLoopNode* loop_head, ParsePredicateNode* parse_predicate,
1458 IfProjNode* new_control, int scale, Node* offset, Node* range);
1459 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1460
1461 // Helper function to collect predicate for eliminating the useless ones
1462 void eliminate_useless_predicates() const;
1463
1464 void eliminate_useless_zero_trip_guard();
1465 void eliminate_useless_multiversion_if();
1466
1467 public:
1468 // Change the control input of expensive nodes to allow commoning by
1469 // IGVN when it is guaranteed to not result in a more frequent
1470 // execution of the expensive node. Return true if progress.
1471 bool process_expensive_nodes();
1472
1473 // Check whether node has become unreachable
1474 bool is_node_unreachable(Node *n) const {
1475 return !has_node(n) || n->is_unreachable(_igvn);
1476 }
1477
1478 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1479 void do_range_check(IdealLoopTree* loop);
1480
1481 // Clone loop with an invariant test (that does not exit) and
1482 // insert a clone of the test that selects which version to
1483 // execute.
1484 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1485
1486 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1487
1488 private:
1489 static bool has_control_dependencies_from_predicates(LoopNode* head);
1490 static void revert_to_normal_loop(const LoopNode* loop_head);
1491
1492 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1493 const UnswitchedLoopSelector& unswitched_loop_selector);
1494 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1495 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1496 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1497 #ifndef PRODUCT
1498 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1499 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1500 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1501 const LoopNode* original_head, const LoopNode* new_head);
1502 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1503 const LoopNode* original_head, const LoopNode* new_head);
1504 #endif
1505
1506 public:
1507
1508 // Range Check Elimination uses this function!
1509 // Constrain the main loop iterations so the affine function:
1510 // low_limit <= scale_con * I + offset < upper_limit
1511 // always holds true. That is, either increase the number of iterations in
1512 // the pre-loop or the post-loop until the condition holds true in the main
1513 // loop. Scale_con, offset and limit are all loop invariant.
1514 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);
1515 // Helper function for add_constraint().
1516 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1517
1518 // Partially peel loop up through last_peel node.
1519 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1520 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1521
1522 // AutoVectorize the loop: replace scalar ops with vector ops.
1523 enum AutoVectorizeStatus {
1524 Impossible, // This loop has the wrong shape to even try vectorization.
1525 Success, // We just successfully vectorized the loop.
1526 TriedAndFailed, // We tried to vectorize, but failed.
1527 };
1528 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1529
1530 void maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new);
1531 void do_multiversioning(IdealLoopTree* lpt, Node_List& old_new);
1532 IfTrueNode* create_new_if_for_multiversion(IfTrueNode* multiversioning_fast_proj);
1533 bool try_resume_optimizations_for_delayed_slow_loop(IdealLoopTree* lpt);
1534
1535 // Move an unordered Reduction out of loop if possible
1536 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop);
1537
1538 // Create a scheduled list of nodes control dependent on ctrl set.
1539 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1540 // Has a use in the vector set
1541 bool has_use_in_set( Node* n, VectorSet& vset );
1542 // Has use internal to the vector set (ie. not in a phi at the loop head)
1543 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1544 // clone "n" for uses that are outside of loop
1545 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1546 // clone "n" for special uses that are in the not_peeled region
1547 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1548 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1549 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1550 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1551 #ifdef ASSERT
1552 // Validate the loop partition sets: peel and not_peel
1553 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1554 // Ensure that uses outside of loop are of the right form
1555 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1556 uint orig_exit_idx, uint clone_exit_idx);
1557 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1558 #endif
1559
1560 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1561 int stride_of_possible_iv( Node* iff );
1562 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1563 // Return the (unique) control output node that's in the loop (if it exists.)
1564 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1565 // Insert a signed compare loop exit cloned from an unsigned compare.
1566 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1567 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1568 // Utility to register node "n" with PhaseIdealLoop
1569 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1570 // Utility to create an if-projection
1571 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1572 // Force the iff control output to be the live_proj
1573 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1574 // Insert a region before an if projection
1575 RegionNode* insert_region_before_proj(ProjNode* proj);
1576 // Insert a new if before an if projection
1577 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1578
1579 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1580 // "Nearly" because all Nodes have been cloned from the original in the loop,
1581 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1582 // through the Phi recursively, and return a Bool.
1583 Node* clone_iff(PhiNode* phi);
1584 CmpNode* clone_bool(PhiNode* phi);
1585
1586
1587 // Rework addressing expressions to get the most loop-invariant stuff
1588 // moved out. We'd like to do all associative operators, but it's especially
1589 // important (common) to do address expressions.
1590 Node* remix_address_expressions(Node* n);
1591 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1592
1593 // Convert add to muladd to generate MuladdS2I under certain criteria
1594 Node * convert_add_to_muladd(Node * n);
1595
1596 // Attempt to use a conditional move instead of a phi/branch
1597 Node *conditional_move( Node *n );
1598
1599 // Check for aggressive application of 'split-if' optimization,
1600 // using basic block level info.
1601 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1602 Node *split_if_with_blocks_pre ( Node *n );
1603 void split_if_with_blocks_post( Node *n );
1604 Node *has_local_phi_input( Node *n );
1605 // Mark an IfNode as being dominated by a prior test,
1606 // without actually altering the CFG (and hence IDOM info).
1607 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool pin_array_access_nodes = false);
1608 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool pin_array_access_nodes);
1609
1610 // Split Node 'n' through merge point
1611 RegionNode* split_thru_region(Node* n, RegionNode* region);
1612 // Split Node 'n' through merge point if there is enough win.
1613 Node *split_thru_phi( Node *n, Node *region, int policy );
1614 // Found an If getting its condition-code input from a Phi in the
1615 // same block. Split thru the Region.
1616 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1617
1618 // Conversion of fill/copy patterns into intrinsic versions
1619 bool do_intrinsify_fill();
1620 bool intrinsify_fill(IdealLoopTree* lpt);
1621 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1622 Node*& shift, Node*& offset);
1623
1624 private:
1625 // Return a type based on condition control flow
1626 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1627 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1628 // Helpers for filtered type
1629 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1630
1631 // Helper functions
1632 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1633 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1634 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 );
1635 bool split_up( Node *n, Node *blk1, Node *blk2 );
1636
1637 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1638 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1639 void try_move_store_after_loop(Node* n);
1640 bool identical_backtoback_ifs(Node *n);
1641 bool can_split_if(Node *n_ctrl);
1642 bool cannot_split_division(const Node* n, const Node* region) const;
1643 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1644 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1645
1646 // Determine if a method is too big for a/another round of split-if, based on
1647 // a magic (approximate) ratio derived from the equally magic constant 35000,
1648 // previously used for this purpose (but without relating to the node limit).
1649 bool must_throttle_split_if() {
1650 uint threshold = C->max_node_limit() * 2 / 5;
1651 return C->live_nodes() > threshold;
1652 }
1653
1654 // A simplistic node request tracking mechanism, where
1655 // = UINT_MAX Request not valid or made final.
1656 // < UINT_MAX Nodes currently requested (estimate).
1657 uint _nodes_required;
1658
1659 enum { REQUIRE_MIN = 70 };
1660
1661 uint nodes_required() const { return _nodes_required; }
1662
1663 // Given the _currently_ available number of nodes, check whether there is
1664 // "room" for an additional request or not, considering the already required
1665 // number of nodes. Return TRUE if the new request is exceeding the node
1666 // budget limit, otherwise return FALSE. Note that this interpretation will
1667 // act pessimistic on additional requests when new nodes have already been
1668 // generated since the 'begin'. This behaviour fits with the intention that
1669 // node estimates/requests should be made upfront.
1670 bool exceeding_node_budget(uint required = 0) {
1671 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1672 uint available = C->max_node_limit() - C->live_nodes();
1673 return available < required + _nodes_required + REQUIRE_MIN;
1674 }
1675
1676 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1677 precond(require > 0);
1678 _nodes_required += MAX2(require, minreq);
1679 return _nodes_required;
1680 }
1681
1682 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1683 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1684 }
1685
1686 uint require_nodes_begin() {
1687 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1688 _nodes_required = 0;
1689 return C->live_nodes();
1690 }
1691
1692 // When a node request is final, optionally check that the requested number
1693 // of nodes was reasonably correct with respect to the number of new nodes
1694 // introduced since the last 'begin'. Always check that we have not exceeded
1695 // the maximum node limit.
1696 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1697 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1698
1699 #ifdef ASSERT
1700 if (check_estimate) {
1701 // Check that the node budget request was not off by too much (x2).
1702 // Should this be the case we _surely_ need to improve the estimates
1703 // used in our budget calculations.
1704 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1705 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1706 C->live_nodes() - live_at_begin, _nodes_required);
1707 }
1708 }
1709 #endif
1710 // Assert that we have stayed within the node budget limit.
1711 assert(C->live_nodes() < C->max_node_limit(),
1712 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1713 C->live_nodes() - live_at_begin, live_at_begin,
1714 C->max_node_limit(), _nodes_required);
1715
1716 _nodes_required = UINT_MAX;
1717 }
1718
1719 private:
1720
1721 bool _created_loop_node;
1722 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1723 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1724
1725 public:
1726 void set_created_loop_node() { _created_loop_node = true; }
1727 bool created_loop_node() { return _created_loop_node; }
1728 void register_new_node(Node* n, Node* blk);
1729 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1730 register_new_node(new_node, get_ctrl(ctrl_of));
1731 }
1732
1733 Node* clone_and_register(Node* n, Node* ctrl) {
1734 n = n->clone();
1735 register_new_node(n, ctrl);
1736 return n;
1737 }
1738
1739 #ifdef ASSERT
1740 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1741 #endif
1742
1743 #ifndef PRODUCT
1744 void dump() const;
1745 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1746 void dump_idom(Node* n, uint count) const;
1747 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1748 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1749 IdealLoopTree* get_loop_idx(Node* n) const {
1750 // Dead nodes have no loop, so return the top level loop instead
1751 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1752 }
1753 // Print some stats
1754 static void print_statistics();
1755 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1756 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1757 static volatile int _long_loop_candidates;
1758 static volatile int _long_loop_nests;
1759 static volatile int _long_loop_counted_loops;
1760 #endif
1761
1762 #ifdef ASSERT
1763 void verify() const;
1764 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1765 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1766 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1767 #endif
1768
1769 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1770
1771 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1772
1773 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1774
1775
1776 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1777 Node_List &range_checks);
1778
1779 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1780 Node* inner_iters_actual_int, Node* inner_phi,
1781 Node* iv_add, LoopNode* inner_head);
1782
1783 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1784
1785 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1786
1787 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1788
1789 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1790
1791 void try_sink_out_of_loop(Node* n);
1792
1793 Node* clamp(Node* R, Node* L, Node* H);
1794
1795 bool safe_for_if_replacement(const Node* dom) const;
1796
1797 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1798
1799 bool try_merge_identical_ifs(Node* n);
1800
1801 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1802
1803 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1804 IdealLoopTree* parent, bool partial);
1805
1806 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1807 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1808
1809 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1810 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1811 Node_List*& split_bool_set, Node_List*& split_cex_set);
1812
1813 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1814
1815 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1816
1817 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1818 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1819 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1820 void clone_template_assertion_expression_down(Node* node);
1821
1822 Node* similar_subtype_check(const Node* x, Node* r_in);
1823
1824 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1825
1826 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1827
1828 void pin_array_access_nodes_dependent_on(Node* ctrl);
1829
1830 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1831
1832 ConINode* intcon(jint i);
1833
1834 ConLNode* longcon(jlong i);
1835
1836 ConNode* makecon(const Type* t);
1837
1838 ConNode* integercon(jlong l, BasicType bt);
1839
1840 ConNode* zerocon(BasicType bt);
1841 };
1842
1843
1844 class AutoNodeBudget : public StackObj
1845 {
1846 public:
1847 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1848
1849 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1850 : _phase(phase),
1851 _check_at_final(chk == BUDGET_CHECK),
1852 _nodes_at_begin(0)
1853 {
1854 precond(_phase != nullptr);
1855
1856 _nodes_at_begin = _phase->require_nodes_begin();
1857 }
1858
1859 ~AutoNodeBudget() {
1860 #ifndef PRODUCT
1861 if (TraceLoopOpts) {
1862 uint request = _phase->nodes_required();
1863 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1864
1865 if (request < delta) {
1866 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1867 } else {
1868 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1869 // Identify the worst estimates as "poor" ones.
1870 if (request > REQUIRE_MIN && delta > 0) {
1871 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1872 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1873 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1874 }
1875 }
1876 }
1877 }
1878 #endif // PRODUCT
1879 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1880 }
1881
1882 private:
1883 PhaseIdealLoop* _phase;
1884 bool _check_at_final;
1885 uint _nodes_at_begin;
1886 };
1887
1888 inline Node* IdealLoopTree::tail() {
1889 // Handle lazy update of _tail field.
1890 if (_tail->in(0) == nullptr) {
1891 _tail = _phase->get_ctrl(_tail);
1892 }
1893 return _tail;
1894 }
1895
1896 inline Node* IdealLoopTree::head() {
1897 // Handle lazy update of _head field.
1898 if (_head->in(0) == nullptr) {
1899 _head = _phase->get_ctrl(_head);
1900 }
1901 return _head;
1902 }
1903
1904 // Iterate over the loop tree using a preorder, left-to-right traversal.
1905 //
1906 // Example that visits all counted loops from within PhaseIdealLoop
1907 //
1908 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1909 // IdealLoopTree* lpt = iter.current();
1910 // if (!lpt->is_counted()) continue;
1911 // ...
1912 class LoopTreeIterator : public StackObj {
1913 private:
1914 IdealLoopTree* _root;
1915 IdealLoopTree* _curnt;
1916
1917 public:
1918 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1919
1920 bool done() { return _curnt == nullptr; } // Finished iterating?
1921
1922 void next(); // Advance to next loop tree
1923
1924 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1925 };
1926
1927 // Compute probability of reaching some CFG node from a fixed
1928 // dominating CFG node
1929 class PathFrequency {
1930 private:
1931 Node* _dom; // frequencies are computed relative to this node
1932 Node_Stack _stack;
1933 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1934 GrowableArray<float> _freqs; // cache frequencies
1935 PhaseIdealLoop* _phase;
1936
1937 float check_and_truncate_frequency(float f) {
1938 assert(f >= 0, "Incorrect frequency");
1939 // We do not perform an exact (f <= 1) check
1940 // this would be error prone with rounding of floats.
1941 // Performing a check like (f <= 1+eps) would be of benefit,
1942 // however, it is not evident how to determine such an eps,
1943 // given that an arbitrary number of add/mul operations
1944 // are performed on these frequencies.
1945 return (f > 1) ? 1 : f;
1946 }
1947
1948 public:
1949 PathFrequency(Node* dom, PhaseIdealLoop* phase)
1950 : _dom(dom), _stack(0), _phase(phase) {
1951 }
1952
1953 float to(Node* n);
1954 };
1955
1956 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
1957 // 1. Clone the data nodes
1958 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
1959 class DataNodeGraph : public StackObj {
1960 PhaseIdealLoop* const _phase;
1961 const Unique_Node_List& _data_nodes;
1962 OrigToNewHashtable _orig_to_new;
1963
1964 public:
1965 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
1966 : _phase(phase),
1967 _data_nodes(data_nodes),
1968 // Use 107 as best guess which is the first resize value in ResizeableResourceHashtable::large_table_sizes.
1969 _orig_to_new(107, MaxNodeLimit)
1970 {
1971 #ifdef ASSERT
1972 for (uint i = 0; i < data_nodes.size(); i++) {
1973 assert(!data_nodes[i]->is_CFG(), "only data nodes");
1974 }
1975 #endif
1976 }
1977 NONCOPYABLE(DataNodeGraph);
1978
1979 private:
1980 void clone(Node* node, Node* new_ctrl);
1981 void clone_data_nodes(Node* new_ctrl);
1982 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
1983 Node* new_ctrl);
1984 void rewire_clones_to_cloned_inputs();
1985 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
1986
1987 public:
1988 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
1989 // Set 'new_ctrl' as ctrl for the cloned nodes.
1990 const OrigToNewHashtable& clone(Node* new_ctrl) {
1991 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
1992 clone_data_nodes(new_ctrl);
1993 rewire_clones_to_cloned_inputs();
1994 return _orig_to_new;
1995 }
1996
1997 // Create a copy of the data nodes provided to the constructor by doing the following:
1998 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
1999 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
2000 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
2001 // `new_ctrl` as ctrl.
2002 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
2003 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
2004 Node* new_ctrl) {
2005 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
2006 rewire_clones_to_cloned_inputs();
2007 return _orig_to_new;
2008 }
2009 };
2010 #endif // SHARE_OPTO_LOOPNODE_HPP