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