1 /*
2 * Copyright (c) 1998, 2023, 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/subnode.hpp"
32 #include "opto/type.hpp"
33
34 class CmpNode;
35 class BaseCountedLoopEndNode;
36 class CountedLoopNode;
37 class IdealLoopTree;
38 class LoopNode;
39 class Node;
40 class OuterStripMinedLoopEndNode;
41 class ParsePredicates;
42 class PathFrequency;
43 class PhaseIdealLoop;
44 class CountedLoopReserveKit;
45 class VectorSet;
46 class Invariance;
47 struct small_cache;
48
49 //
50 // I D E A L I Z E D L O O P S
51 //
52 // Idealized loops are the set of loops I perform more interesting
53 // transformations on, beyond simple hoisting.
54
55 //------------------------------LoopNode---------------------------------------
56 // Simple loop header. Fall in path on left, loop-back path on right.
57 class LoopNode : public RegionNode {
58 // Size is bigger to hold the flags. However, the flags do not change
59 // the semantics so it does not appear in the hash & cmp functions.
60 virtual uint size_of() const { return sizeof(*this); }
61 protected:
62 uint _loop_flags;
63 // Names for flag bitfields
64 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
65 MainHasNoPreLoop = 1<<2,
66 HasExactTripCount = 1<<3,
67 InnerLoop = 1<<4,
68 PartialPeelLoop = 1<<5,
69 PartialPeelFailed = 1<<6,
70 WasSlpAnalyzed = 1<<7,
71 PassedSlpAnalysis = 1<<8,
72 DoUnrollOnly = 1<<9,
73 VectorizedLoop = 1<<10,
74 HasAtomicPostLoop = 1<<11,
75 StripMined = 1<<12,
76 SubwordLoop = 1<<13,
77 ProfileTripFailed = 1<<14,
78 LoopNestInnerLoop = 1<<15,
79 LoopNestLongOuterLoop = 1<<16 };
80 char _unswitch_count;
81 enum { _unswitch_max=3 };
82
83 // Expected trip count from profile data
84 float _profile_trip_cnt;
85
86 public:
87 // Names for edge indices
88 enum { Self=0, EntryControl, LoopBackControl };
89
90 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
91 void set_inner_loop() { _loop_flags |= InnerLoop; }
92
93 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
94 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
95 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
96 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
97 bool is_strip_mined() const { return _loop_flags & StripMined; }
98 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
99 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
100 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
101 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
102
103 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
104 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
105 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
106 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
107 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
108 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
109 void mark_strip_mined() { _loop_flags |= StripMined; }
110 void clear_strip_mined() { _loop_flags &= ~StripMined; }
111 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
112 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
113 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
114 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
115
116 int unswitch_max() { return _unswitch_max; }
117 int unswitch_count() { return _unswitch_count; }
118
119 void set_unswitch_count(int val) {
120 assert (val <= unswitch_max(), "too many unswitches");
121 _unswitch_count = val;
122 }
123
124 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
125 float profile_trip_cnt() { return _profile_trip_cnt; }
126
127 LoopNode(Node *entry, Node *backedge)
128 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
129 _profile_trip_cnt(COUNT_UNKNOWN) {
130 init_class_id(Class_Loop);
131 init_req(EntryControl, entry);
132 init_req(LoopBackControl, backedge);
133 }
134
135 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
136 virtual int Opcode() const;
137 bool can_be_counted_loop(PhaseValues* phase) const {
138 return req() == 3 && in(0) != nullptr &&
139 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
140 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
141 }
142 bool is_valid_counted_loop(BasicType bt) const;
143 #ifndef PRODUCT
144 virtual void dump_spec(outputStream *st) const;
145 #endif
146
147 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
148 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
149 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
150 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
151 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
152 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
153 };
154
155 //------------------------------Counted Loops----------------------------------
156 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
157 // path (and maybe some other exit paths). The trip-counter exit is always
158 // last in the loop. The trip-counter have to stride by a constant;
159 // the exit value is also loop invariant.
160
161 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
162 // CountedLoopNode has the incoming loop control and the loop-back-control
163 // which is always the IfTrue before the matching CountedLoopEndNode. The
164 // CountedLoopEndNode has an incoming control (possibly not the
165 // CountedLoopNode if there is control flow in the loop), the post-increment
166 // trip-counter value, and the limit. The trip-counter value is always of
167 // the form (Op old-trip-counter stride). The old-trip-counter is produced
168 // by a Phi connected to the CountedLoopNode. The stride is constant.
169 // The Op is any commutable opcode, including Add, Mul, Xor. The
170 // CountedLoopEndNode also takes in the loop-invariant limit value.
171
172 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
173 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
174 // via the old-trip-counter from the Op node.
175
176 //------------------------------CountedLoopNode--------------------------------
177 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
178 // inputs the incoming loop-start control and the loop-back control, so they
179 // act like RegionNodes. They also take in the initial trip counter, the
180 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
181 // produce a loop-body control and the trip counter value. Since
182 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
183
184 class BaseCountedLoopNode : public LoopNode {
185 public:
186 BaseCountedLoopNode(Node *entry, Node *backedge)
187 : LoopNode(entry, backedge) {
188 }
189
190 Node *init_control() const { return in(EntryControl); }
191 Node *back_control() const { return in(LoopBackControl); }
192
193 Node* init_trip() const;
194 Node* stride() const;
195 bool stride_is_con() const;
196 Node* limit() const;
197 Node* incr() const;
198 Node* phi() const;
199
200 BaseCountedLoopEndNode* loopexit_or_null() const;
201 BaseCountedLoopEndNode* loopexit() const;
202
203 virtual BasicType bt() const = 0;
204
205 jlong stride_con() const;
206
207 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
208 };
209
210
211 class CountedLoopNode : public BaseCountedLoopNode {
212 // Size is bigger to hold _main_idx. However, _main_idx does not change
213 // the semantics so it does not appear in the hash & cmp functions.
214 virtual uint size_of() const { return sizeof(*this); }
215
216 // For Pre- and Post-loops during debugging ONLY, this holds the index of
217 // the Main CountedLoop. Used to assert that we understand the graph shape.
218 node_idx_t _main_idx;
219
220 // Known trip count calculated by compute_exact_trip_count()
221 uint _trip_count;
222
223 // Log2 of original loop bodies in unrolled loop
224 int _unrolled_count_log2;
225
226 // Node count prior to last unrolling - used to decide if
227 // unroll,optimize,unroll,optimize,... is making progress
228 int _node_count_before_unroll;
229
230 // If slp analysis is performed we record the maximum
231 // vector mapped unroll factor here
232 int _slp_maximum_unroll_factor;
233
234 // The eventual count of vectorizable packs in slp
235 int _slp_vector_pack_count;
236
237 public:
238 CountedLoopNode(Node *entry, Node *backedge)
239 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
240 _unrolled_count_log2(0), _node_count_before_unroll(0),
241 _slp_maximum_unroll_factor(0), _slp_vector_pack_count(0) {
242 init_class_id(Class_CountedLoop);
243 // Initialize _trip_count to the largest possible value.
244 // Will be reset (lower) if the loop's trip count is known.
245 }
246
247 virtual int Opcode() const;
248 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
249
250 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
251 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
252 int stride_con() const;
253
254 // Match increment with optional truncation
255 static Node*
256 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
257 BasicType bt);
258
259 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
260 // can run short a few iterations and may start a few iterations in.
261 // It will be RCE'd and unrolled and aligned.
262
263 // A following 'post' loop will run any remaining iterations. Used
264 // during Range Check Elimination, the 'post' loop will do any final
265 // iterations with full checks. Also used by Loop Unrolling, where
266 // the 'post' loop will do any epilog iterations needed. Basically,
267 // a 'post' loop can not profitably be further unrolled or RCE'd.
268
269 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
270 // it may do under-flow checks for RCE and may do alignment iterations
271 // so the following main loop 'knows' that it is striding down cache
272 // lines.
273
274 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
275 // Aligned, may be missing it's pre-loop.
276 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
277 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
278 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
279 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
280 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
281 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
282 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
283 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
284 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
285 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
286
287 int main_idx() const { return _main_idx; }
288
289
290 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
291 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
292 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
293 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
294
295 void set_trip_count(uint tc) { _trip_count = tc; }
296 uint trip_count() { return _trip_count; }
297
298 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
299 void set_exact_trip_count(uint tc) {
300 _trip_count = tc;
301 _loop_flags |= HasExactTripCount;
302 }
303 void set_nonexact_trip_count() {
304 _loop_flags &= ~HasExactTripCount;
305 }
306 void set_notpassed_slp() {
307 _loop_flags &= ~PassedSlpAnalysis;
308 }
309
310 void double_unrolled_count() { _unrolled_count_log2++; }
311 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
312
313 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
314 int node_count_before_unroll() { return _node_count_before_unroll; }
315 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
316 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
317
318 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
319 OuterStripMinedLoopNode* outer_loop() const;
320 virtual IfTrueNode* outer_loop_tail() const;
321 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
322 virtual IfFalseNode* outer_loop_exit() const;
323 virtual SafePointNode* outer_safepoint() const;
324
325 // If this is a main loop in a pre/main/post loop nest, walk over
326 // the predicates that were inserted by
327 // duplicate_predicates()/add_range_check_predicate()
328 static Node* skip_predicates_from_entry(Node* ctrl);
329 Node* skip_predicates();
330
331 virtual BasicType bt() const {
332 return T_INT;
333 }
334
335 Node* is_canonical_loop_entry();
336
337 #ifndef PRODUCT
338 virtual void dump_spec(outputStream *st) const;
339 #endif
340
341 static bool is_zero_trip_guard_if(const IfNode* iff);
342 };
343
344 class LongCountedLoopNode : public BaseCountedLoopNode {
345 public:
346 LongCountedLoopNode(Node *entry, Node *backedge)
347 : BaseCountedLoopNode(entry, backedge) {
348 init_class_id(Class_LongCountedLoop);
349 }
350
351 virtual int Opcode() const;
352
353 virtual BasicType bt() const {
354 return T_LONG;
355 }
356
357 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
358 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
359 };
360
361
362 //------------------------------CountedLoopEndNode-----------------------------
363 // CountedLoopEndNodes end simple trip counted loops. They act much like
364 // IfNodes.
365
366 class BaseCountedLoopEndNode : public IfNode {
367 public:
368 enum { TestControl, TestValue };
369 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
370 : IfNode(control, test, prob, cnt) {
371 init_class_id(Class_BaseCountedLoopEnd);
372 }
373
374 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
375 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
376 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
377 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
378 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
379 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
380
381 PhiNode* phi() const {
382 Node* tmp = incr();
383 if (tmp && tmp->req() == 3) {
384 Node* phi = tmp->in(1);
385 if (phi->is_Phi()) {
386 return phi->as_Phi();
387 }
388 }
389 return nullptr;
390 }
391
392 BaseCountedLoopNode* loopnode() const {
393 // The CountedLoopNode that goes with this CountedLoopEndNode may
394 // have been optimized out by the IGVN so be cautious with the
395 // pattern matching on the graph
396 PhiNode* iv_phi = phi();
397 if (iv_phi == nullptr) {
398 return nullptr;
399 }
400 Node* ln = iv_phi->in(0);
401 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
402 return nullptr;
403 }
404 if (ln->as_BaseCountedLoop()->bt() != bt()) {
405 return nullptr;
406 }
407 return ln->as_BaseCountedLoop();
408 }
409
410 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
411
412 jlong stride_con() const;
413 virtual BasicType bt() const = 0;
414
415 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
416 };
417
418 class CountedLoopEndNode : public BaseCountedLoopEndNode {
419 public:
420
421 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
422 : BaseCountedLoopEndNode(control, test, prob, cnt) {
423 init_class_id(Class_CountedLoopEnd);
424 }
425 virtual int Opcode() const;
426
427 CountedLoopNode* loopnode() const {
428 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
429 }
430
431 virtual BasicType bt() const {
432 return T_INT;
433 }
434
435 #ifndef PRODUCT
436 virtual void dump_spec(outputStream *st) const;
437 #endif
438 };
439
440 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
441 public:
442 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
443 : BaseCountedLoopEndNode(control, test, prob, cnt) {
444 init_class_id(Class_LongCountedLoopEnd);
445 }
446
447 LongCountedLoopNode* loopnode() const {
448 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
449 }
450
451 virtual int Opcode() const;
452
453 virtual BasicType bt() const {
454 return T_LONG;
455 }
456 };
457
458
459 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
460 Node* bctrl = back_control();
461 if (bctrl == nullptr) return nullptr;
462
463 Node* lexit = bctrl->in(0);
464 if (!lexit->is_BaseCountedLoopEnd()) {
465 return nullptr;
466 }
467 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
468 if (result->bt() != bt()) {
469 return nullptr;
470 }
471 return result;
472 }
473
474 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
475 BaseCountedLoopEndNode* cle = loopexit_or_null();
476 assert(cle != nullptr, "loopexit is null");
477 return cle;
478 }
479
480 inline Node* BaseCountedLoopNode::init_trip() const {
481 BaseCountedLoopEndNode* cle = loopexit_or_null();
482 return cle != nullptr ? cle->init_trip() : nullptr;
483 }
484 inline Node* BaseCountedLoopNode::stride() const {
485 BaseCountedLoopEndNode* cle = loopexit_or_null();
486 return cle != nullptr ? cle->stride() : nullptr;
487 }
488
489 inline bool BaseCountedLoopNode::stride_is_con() const {
490 BaseCountedLoopEndNode* cle = loopexit_or_null();
491 return cle != nullptr && cle->stride_is_con();
492 }
493 inline Node* BaseCountedLoopNode::limit() const {
494 BaseCountedLoopEndNode* cle = loopexit_or_null();
495 return cle != nullptr ? cle->limit() : nullptr;
496 }
497 inline Node* BaseCountedLoopNode::incr() const {
498 BaseCountedLoopEndNode* cle = loopexit_or_null();
499 return cle != nullptr ? cle->incr() : nullptr;
500 }
501 inline Node* BaseCountedLoopNode::phi() const {
502 BaseCountedLoopEndNode* cle = loopexit_or_null();
503 return cle != nullptr ? cle->phi() : nullptr;
504 }
505
506 inline jlong BaseCountedLoopNode::stride_con() const {
507 BaseCountedLoopEndNode* cle = loopexit_or_null();
508 return cle != nullptr ? cle->stride_con() : 0;
509 }
510
511
512 //------------------------------LoopLimitNode-----------------------------
513 // Counted Loop limit node which represents exact final iterator value:
514 // trip_count = (limit - init_trip + stride - 1)/stride
515 // final_value= trip_count * stride + init_trip.
516 // Use HW instructions to calculate it when it can overflow in integer.
517 // Note, final_value should fit into integer since counted loop has
518 // limit check: limit <= max_int-stride.
519 class LoopLimitNode : public Node {
520 enum { Init=1, Limit=2, Stride=3 };
521 public:
522 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
523 // Put it on the Macro nodes list to optimize during macro nodes expansion.
524 init_flags(Flag_is_macro);
525 C->add_macro_node(this);
526 }
527 virtual int Opcode() const;
528 virtual const Type *bottom_type() const { return TypeInt::INT; }
529 virtual uint ideal_reg() const { return Op_RegI; }
530 virtual const Type* Value(PhaseGVN* phase) const;
531 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
532 virtual Node* Identity(PhaseGVN* phase);
533 };
534
535 // Support for strip mining
536 class OuterStripMinedLoopNode : public LoopNode {
537 private:
538 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
539
540 public:
541 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
542 : LoopNode(entry, backedge) {
543 init_class_id(Class_OuterStripMinedLoop);
544 init_flags(Flag_is_macro);
545 C->add_macro_node(this);
546 }
547
548 virtual int Opcode() const;
549
550 virtual IfTrueNode* outer_loop_tail() const;
551 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
552 virtual IfFalseNode* outer_loop_exit() const;
553 virtual SafePointNode* outer_safepoint() const;
554 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
555
556 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
557
558 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
559
560 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
561
562 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
563 PhaseIdealLoop* iloop);
564 };
565
566 class OuterStripMinedLoopEndNode : public IfNode {
567 public:
568 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
569 : IfNode(control, test, prob, cnt) {
570 init_class_id(Class_OuterStripMinedLoopEnd);
571 }
572
573 virtual int Opcode() const;
574
575 virtual const Type* Value(PhaseGVN* phase) const;
576 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
577
578 bool is_expanded(PhaseGVN *phase) const;
579 };
580
581 // -----------------------------IdealLoopTree----------------------------------
582 class IdealLoopTree : public ResourceObj {
583 public:
584 IdealLoopTree *_parent; // Parent in loop tree
585 IdealLoopTree *_next; // Next sibling in loop tree
586 IdealLoopTree *_child; // First child in loop tree
587
588 // The head-tail backedge defines the loop.
589 // If a loop has multiple backedges, this is addressed during cleanup where
590 // we peel off the multiple backedges, merging all edges at the bottom and
591 // ensuring that one proper backedge flow into the loop.
592 Node *_head; // Head of loop
593 Node *_tail; // Tail of loop
594 inline Node *tail(); // Handle lazy update of _tail field
595 inline Node *head(); // Handle lazy update of _head field
596 PhaseIdealLoop* _phase;
597 int _local_loop_unroll_limit;
598 int _local_loop_unroll_factor;
599
600 Node_List _body; // Loop body for inner loops
601
602 uint16_t _nest; // Nesting depth
603 uint8_t _irreducible:1, // True if irreducible
604 _has_call:1, // True if has call safepoint
605 _has_sfpt:1, // True if has non-call safepoint
606 _rce_candidate:1, // True if candidate for range check elimination
607 _has_range_checks:1,
608 _has_range_checks_computed:1;
609
610 Node_List* _safepts; // List of safepoints in this loop
611 Node_List* _required_safept; // A inner loop cannot delete these safepts;
612 bool _allow_optimizations; // Allow loop optimizations
613
614 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
615 : _parent(0), _next(0), _child(0),
616 _head(head), _tail(tail),
617 _phase(phase),
618 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
619 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
620 _has_range_checks(0), _has_range_checks_computed(0),
621 _safepts(nullptr),
622 _required_safept(nullptr),
623 _allow_optimizations(true)
624 {
625 precond(_head != nullptr);
626 precond(_tail != nullptr);
627 }
628
629 // Is 'l' a member of 'this'?
630 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
631
632 // Set loop nesting depth. Accumulate has_call bits.
633 int set_nest( uint depth );
634
635 // Split out multiple fall-in edges from the loop header. Move them to a
636 // private RegionNode before the loop. This becomes the loop landing pad.
637 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
638
639 // Split out the outermost loop from this shared header.
640 void split_outer_loop( PhaseIdealLoop *phase );
641
642 // Merge all the backedges from the shared header into a private Region.
643 // Feed that region as the one backedge to this loop.
644 void merge_many_backedges( PhaseIdealLoop *phase );
645
646 // Split shared headers and insert loop landing pads.
647 // Insert a LoopNode to replace the RegionNode.
648 // Returns TRUE if loop tree is structurally changed.
649 bool beautify_loops( PhaseIdealLoop *phase );
650
651 // Perform optimization to use the loop predicates for null checks and range checks.
652 // Applies to any loop level (not just the innermost one)
653 bool loop_predication( PhaseIdealLoop *phase);
654
655 // Perform iteration-splitting on inner loops. Split iterations to
656 // avoid range checks or one-shot null checks. Returns false if the
657 // current round of loop opts should stop.
658 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
659
660 // Driver for various flavors of iteration splitting. Returns false
661 // if the current round of loop opts should stop.
662 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
663
664 // Given dominators, try to find loops with calls that must always be
665 // executed (call dominates loop tail). These loops do not need non-call
666 // safepoints (ncsfpt).
667 void check_safepts(VectorSet &visited, Node_List &stack);
668
669 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
670 // encountered.
671 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
672
673 // Remove safepoints from loop. Optionally keeping one.
674 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
675
676 // Convert to counted loops where possible
677 void counted_loop( PhaseIdealLoop *phase );
678
679 // Check for Node being a loop-breaking test
680 Node *is_loop_exit(Node *iff) const;
681
682 // Remove simplistic dead code from loop body
683 void DCE_loop_body();
684
685 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
686 // Replace with a 1-in-10 exit guess.
687 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
688
689 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
690 // Useful for unrolling loops with NO array accesses.
691 bool policy_peel_only( PhaseIdealLoop *phase ) const;
692
693 // Return TRUE or FALSE if the loop should be unswitched -- clone
694 // loop with an invariant test
695 bool policy_unswitching( PhaseIdealLoop *phase ) const;
696
697 // Micro-benchmark spamming. Remove empty loops.
698 bool do_remove_empty_loop( PhaseIdealLoop *phase );
699
700 // Convert one iteration loop into normal code.
701 bool do_one_iteration_loop( PhaseIdealLoop *phase );
702
703 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
704 // move some loop-invariant test (usually a null-check) before the loop.
705 bool policy_peeling(PhaseIdealLoop *phase);
706
707 uint estimate_peeling(PhaseIdealLoop *phase);
708
709 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
710 // known trip count in the counted loop node.
711 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
712
713 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
714 // if the loop is a counted loop and the loop body is small enough.
715 bool policy_unroll(PhaseIdealLoop *phase);
716
717 // Loop analyses to map to a maximal superword unrolling for vectorization.
718 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
719
720 // Return TRUE or FALSE if the loop should be range-check-eliminated.
721 // Gather a list of IF tests that are dominated by iteration splitting;
722 // also gather the end of the first split and the start of the 2nd split.
723 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
724
725 // Return TRUE if "iff" is a range check.
726 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
727 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
728 jlong& scale) const;
729
730 // Estimate the number of nodes required when cloning a loop (body).
731 uint est_loop_clone_sz(uint factor) const;
732 // Estimate the number of nodes required when unrolling a loop (body).
733 uint est_loop_unroll_sz(uint factor) const;
734
735 // Compute loop trip count if possible
736 void compute_trip_count(PhaseIdealLoop* phase);
737
738 // Compute loop trip count from profile data
739 float compute_profile_trip_cnt_helper(Node* n);
740 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
741
742 // Reassociate invariant expressions.
743 void reassociate_invariants(PhaseIdealLoop *phase);
744 // Reassociate invariant binary expressions.
745 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
746 // Reassociate invariant add and subtract expressions.
747 Node* reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase);
748 // Return nonzero index of invariant operand if invariant and variant
749 // are combined with an associative binary. Helper for reassociate_invariants.
750 int find_invariant(Node* n, PhaseIdealLoop *phase);
751 // Return TRUE if "n" is associative.
752 bool is_associative(Node* n, Node* base=nullptr);
753
754 // Return true if n is invariant
755 bool is_invariant(Node* n) const;
756
757 // Put loop body on igvn work list
758 void record_for_igvn();
759
760 bool is_root() { return _parent == nullptr; }
761 // A proper/reducible loop w/o any (occasional) dead back-edge.
762 bool is_loop() { return !_irreducible && !tail()->is_top(); }
763 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
764 bool is_innermost() { return is_loop() && _child == nullptr; }
765
766 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
767
768 bool compute_has_range_checks() const;
769 bool range_checks_present() {
770 if (!_has_range_checks_computed) {
771 if (compute_has_range_checks()) {
772 _has_range_checks = 1;
773 }
774 _has_range_checks_computed = 1;
775 }
776 return _has_range_checks;
777 }
778
779 #ifndef PRODUCT
780 void dump_head(); // Dump loop head only
781 void dump(); // Dump this loop recursively
782 #endif
783
784 #ifdef ASSERT
785 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
786 bool verify_tree(IdealLoopTree* loop_verify) const;
787 #endif
788
789 private:
790 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
791
792 // Estimate the number of nodes resulting from control and data flow merge.
793 uint est_loop_flow_merge_sz() const;
794
795 // Check if the number of residual iterations is large with unroll_cnt.
796 // Return true if the residual iterations are more than 10% of the trip count.
797 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
798 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
799 }
800
801 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
802
803 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
804
805 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
806
807 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
808 Node* sfpt) const;
809
810 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
811
812 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
813 };
814
815 // -----------------------------PhaseIdealLoop---------------------------------
816 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
817 // into a loop tree. Drives the loop-based transformations on the ideal graph.
818 class PhaseIdealLoop : public PhaseTransform {
819 friend class IdealLoopTree;
820 friend class SuperWord;
821 friend class CountedLoopReserveKit;
822 friend class ShenandoahBarrierC2Support;
823 friend class AutoNodeBudget;
824
825 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
826 Node_List _loop_or_ctrl;
827
828 // Pre-computed def-use info
829 PhaseIterGVN &_igvn;
830
831 // Head of loop tree
832 IdealLoopTree* _ltree_root;
833
834 // Array of pre-order numbers, plus post-visited bit.
835 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
836 // ODD for post-visited. Other bits are the pre-order number.
837 uint *_preorders;
838 uint _max_preorder;
839
840 const PhaseIdealLoop* _verify_me;
841 bool _verify_only;
842
843 // Allocate _preorders[] array
844 void allocate_preorders() {
845 _max_preorder = C->unique()+8;
846 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
847 memset(_preorders, 0, sizeof(uint) * _max_preorder);
848 }
849
850 // Allocate _preorders[] array
851 void reallocate_preorders() {
852 if ( _max_preorder < C->unique() ) {
853 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
854 _max_preorder = C->unique();
855 }
856 memset(_preorders, 0, sizeof(uint) * _max_preorder);
857 }
858
859 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
860 // adds new nodes.
861 void check_grow_preorders( ) {
862 if ( _max_preorder < C->unique() ) {
863 uint newsize = _max_preorder<<1; // double size of array
864 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
865 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
866 _max_preorder = newsize;
867 }
868 }
869 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
870 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
871 // Pre-order numbers are written to the Nodes array as low-bit-set values.
872 void set_preorder_visited( Node *n, int pre_order ) {
873 assert( !is_visited( n ), "already set" );
874 _preorders[n->_idx] = (pre_order<<1);
875 };
876 // Return pre-order number.
877 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
878
879 // Check for being post-visited.
880 // Should be previsited already (checked with assert(is_visited(n))).
881 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
882
883 // Mark as post visited
884 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
885
886 public:
887 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
888 // Returns true if "n" is a data node, false if it's a control node.
889 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
890
891 private:
892 // clear out dead code after build_loop_late
893 Node_List _deadlist;
894 Node_List _zero_trip_guard_opaque_nodes;
895
896 // Support for faster execution of get_late_ctrl()/dom_lca()
897 // when a node has many uses and dominator depth is deep.
898 GrowableArray<jlong> _dom_lca_tags;
899 uint _dom_lca_tags_round;
900 void init_dom_lca_tags();
901
902 // Helper for debugging bad dominance relationships
903 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
904
905 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
906
907 // Inline wrapper for frequent cases:
908 // 1) only one use
909 // 2) a use is the same as the current LCA passed as 'n1'
910 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
911 assert( n->is_CFG(), "" );
912 // Fast-path null lca
913 if( lca != nullptr && lca != n ) {
914 assert( lca->is_CFG(), "" );
915 // find LCA of all uses
916 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
917 }
918 return find_non_split_ctrl(n);
919 }
920 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
921
922 // Helper function for directing control inputs away from CFG split points.
923 Node *find_non_split_ctrl( Node *ctrl ) const {
924 if (ctrl != nullptr) {
925 if (ctrl->is_MultiBranch()) {
926 ctrl = ctrl->in(0);
927 }
928 assert(ctrl->is_CFG(), "CFG");
929 }
930 return ctrl;
931 }
932
933 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
934
935 #ifdef ASSERT
936 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
937 #endif
938 void copy_assertion_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride, IdealLoopTree* outer_loop,
939 LoopNode* outer_main_head, uint dd_main_head,
940 uint idx_before_pre_post, uint idx_after_post_before_pre,
941 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post,
942 const Node_List &old_new);
943 void copy_assertion_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop,
944 LoopNode* outer_main_head, uint dd_main_head, uint idx_before_pre_post,
945 uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main,
946 Node* zero_trip_guard_proj_post, const Node_List& old_new);
947 Node* clone_assertion_predicate_and_initialize(Node* iff, Node* new_init, Node* new_stride, Node* predicate,
948 Node* uncommon_proj, Node* control, IdealLoopTree* outer_loop,
949 Node* input_proj);
950 static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride);
951 static bool subgraph_has_opaque(Node* n);
952 Node* create_bool_from_template_assertion_predicate(Node* template_assertion_predicate, Node* new_init, Node* new_stride,
953 Node* control);
954 static bool assertion_predicate_has_loop_opaque_node(IfNode* iff);
955 static void get_assertion_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false);
956 void update_main_loop_assertion_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
957 void copy_assertion_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head, Node* init,
958 Node* stride);
959 void initialize_assertion_predicates_for_peeled_loop(IfProjNode* predicate_proj, LoopNode* outer_loop_head,
960 const int dd_outer_loop_head, Node* init, Node* stride,
961 IdealLoopTree* outer_loop, const uint idx_before_clone,
962 const Node_List& old_new);
963 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
964 Node* bol);
965 #ifdef ASSERT
966 bool only_has_infinite_loops();
967 #endif
968
969 void log_loop_tree();
970
971 public:
972
973 PhaseIterGVN &igvn() const { return _igvn; }
974
975 bool has_node(const Node* n) const {
976 guarantee(n != nullptr, "No Node.");
977 return _loop_or_ctrl[n->_idx] != nullptr;
978 }
979 // check if transform created new nodes that need _ctrl recorded
980 Node *get_late_ctrl( Node *n, Node *early );
981 Node *get_early_ctrl( Node *n );
982 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
983 void set_early_ctrl(Node* n, bool update_body);
984 void set_subtree_ctrl(Node* n, bool update_body);
985 void set_ctrl( Node *n, Node *ctrl ) {
986 assert( !has_node(n) || has_ctrl(n), "" );
987 assert( ctrl->in(0), "cannot set dead control node" );
988 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
989 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
990 }
991 // Set control and update loop membership
992 void set_ctrl_and_loop(Node* n, Node* ctrl) {
993 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
994 IdealLoopTree* new_loop = get_loop(ctrl);
995 if (old_loop != new_loop) {
996 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
997 if (new_loop->_child == nullptr) new_loop->_body.push(n);
998 }
999 set_ctrl(n, ctrl);
1000 }
1001 // Control nodes can be replaced or subsumed. During this pass they
1002 // get their replacement Node in slot 1. Instead of updating the block
1003 // location of all Nodes in the subsumed block, we lazily do it. As we
1004 // pull such a subsumed block out of the array, we write back the final
1005 // correct block.
1006 Node* get_ctrl(const Node* i) {
1007 assert(has_node(i), "");
1008 Node *n = get_ctrl_no_update(i);
1009 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1010 assert(has_node(i) && has_ctrl(i), "");
1011 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1012 return n;
1013 }
1014 // true if CFG node d dominates CFG node n
1015 bool is_dominator(Node *d, Node *n);
1016 // return get_ctrl for a data node and self(n) for a CFG node
1017 Node* ctrl_or_self(Node* n) {
1018 if (has_ctrl(n))
1019 return get_ctrl(n);
1020 else {
1021 assert (n->is_CFG(), "must be a CFG node");
1022 return n;
1023 }
1024 }
1025
1026 Node* get_ctrl_no_update_helper(const Node* i) const {
1027 assert(has_ctrl(i), "should be control, not loop");
1028 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1029 }
1030
1031 Node* get_ctrl_no_update(const Node* i) const {
1032 assert( has_ctrl(i), "" );
1033 Node *n = get_ctrl_no_update_helper(i);
1034 if (!n->in(0)) {
1035 // Skip dead CFG nodes
1036 do {
1037 n = get_ctrl_no_update_helper(n);
1038 } while (!n->in(0));
1039 n = find_non_split_ctrl(n);
1040 }
1041 return n;
1042 }
1043
1044 // Check for loop being set
1045 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1046 bool has_loop( Node *n ) const {
1047 assert(!has_node(n) || !has_ctrl(n), "");
1048 return has_node(n);
1049 }
1050 // Set loop
1051 void set_loop( Node *n, IdealLoopTree *loop ) {
1052 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1053 }
1054 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1055 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1056 // from old_node to new_node to support the lazy update. Reference
1057 // replaces loop reference, since that is not needed for dead node.
1058 void lazy_update(Node *old_node, Node *new_node) {
1059 assert(old_node != new_node, "no cycles please");
1060 // Re-use the side array slot for this node to provide the
1061 // forwarding pointer.
1062 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1063 }
1064 void lazy_replace(Node *old_node, Node *new_node) {
1065 _igvn.replace_node(old_node, new_node);
1066 lazy_update(old_node, new_node);
1067 }
1068
1069 private:
1070
1071 // Place 'n' in some loop nest, where 'n' is a CFG node
1072 void build_loop_tree();
1073 int build_loop_tree_impl( Node *n, int pre_order );
1074 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1075 // loop tree, not the root.
1076 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1077
1078 #ifdef ASSERT
1079 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1080 void verify_regions_in_irreducible_loops();
1081 bool is_in_irreducible_loop(RegionNode* region);
1082 #endif
1083
1084 // Place Data nodes in some loop nest
1085 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1086 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1087 void build_loop_late_post_work(Node* n, bool pinned);
1088 void build_loop_late_post(Node* n);
1089 void verify_strip_mined_scheduling(Node *n, Node* least);
1090
1091 // Array of immediate dominance info for each CFG node indexed by node idx
1092 private:
1093 uint _idom_size;
1094 Node **_idom; // Array of immediate dominators
1095 uint *_dom_depth; // Used for fast LCA test
1096 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1097 LoopOptsMode _mode;
1098
1099 // build the loop tree and perform any requested optimizations
1100 void build_and_optimize();
1101
1102 // Dominators for the sea of nodes
1103 void Dominators();
1104
1105 // Compute the Ideal Node to Loop mapping
1106 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1107 PhaseTransform(Ideal_Loop),
1108 _igvn(igvn),
1109 _verify_me(nullptr),
1110 _verify_only(false),
1111 _mode(mode),
1112 _nodes_required(UINT_MAX) {
1113 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1114 build_and_optimize();
1115 }
1116
1117 #ifndef PRODUCT
1118 // Verify that verify_me made the same decisions as a fresh run
1119 // or only verify that the graph is valid if verify_me is null.
1120 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1121 PhaseTransform(Ideal_Loop),
1122 _igvn(igvn),
1123 _verify_me(verify_me),
1124 _verify_only(verify_me == nullptr),
1125 _mode(LoopOptsVerify),
1126 _nodes_required(UINT_MAX) {
1127 build_and_optimize();
1128 }
1129 #endif
1130
1131 public:
1132 Node* idom_no_update(Node* d) const {
1133 return idom_no_update(d->_idx);
1134 }
1135
1136 Node* idom_no_update(uint didx) const {
1137 assert(didx < _idom_size, "oob");
1138 Node* n = _idom[didx];
1139 assert(n != nullptr,"Bad immediate dominator info.");
1140 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1141 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1142 assert(n != nullptr,"Bad immediate dominator info.");
1143 }
1144 return n;
1145 }
1146
1147 Node *idom(Node* d) const {
1148 return idom(d->_idx);
1149 }
1150
1151 Node *idom(uint didx) const {
1152 Node *n = idom_no_update(didx);
1153 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1154 return n;
1155 }
1156
1157 uint dom_depth(Node* d) const {
1158 guarantee(d != nullptr, "Null dominator info.");
1159 guarantee(d->_idx < _idom_size, "");
1160 return _dom_depth[d->_idx];
1161 }
1162 void set_idom(Node* d, Node* n, uint dom_depth);
1163 // Locally compute IDOM using dom_lca call
1164 Node *compute_idom( Node *region ) const;
1165 // Recompute dom_depth
1166 void recompute_dom_depth();
1167
1168 // Is safept not required by an outer loop?
1169 bool is_deleteable_safept(Node* sfpt);
1170
1171 // Replace parallel induction variable (parallel to trip counter)
1172 void replace_parallel_iv(IdealLoopTree *loop);
1173
1174 Node *dom_lca( Node *n1, Node *n2 ) const {
1175 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1176 }
1177 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1178
1179 // Build and verify the loop tree without modifying the graph. This
1180 // is useful to verify that all inputs properly dominate their uses.
1181 static void verify(PhaseIterGVN& igvn) {
1182 #ifdef ASSERT
1183 ResourceMark rm;
1184 Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
1185 PhaseIdealLoop v(igvn);
1186 #endif
1187 }
1188
1189 // Recommended way to use PhaseIdealLoop.
1190 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1191 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1192 ResourceMark rm;
1193 PhaseIdealLoop v(igvn, mode);
1194
1195 Compile* C = Compile::current();
1196 if (!C->failing()) {
1197 // Cleanup any modified bits
1198 igvn.optimize();
1199
1200 v.log_loop_tree();
1201 }
1202 }
1203
1204 // True if the method has at least 1 irreducible loop
1205 bool _has_irreducible_loops;
1206
1207 // Per-Node transform
1208 virtual Node* transform(Node* n) { return nullptr; }
1209
1210 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1211 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1212 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1213 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1214 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1215
1216 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1217
1218 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1219 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1220 #ifdef ASSERT
1221 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1222 #endif
1223 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1224 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1225 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1226 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1227 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1228 Node*& entry_control, Node*& iffalse);
1229
1230 Node* exact_limit( IdealLoopTree *loop );
1231
1232 // Return a post-walked LoopNode
1233 IdealLoopTree *get_loop( Node *n ) const {
1234 // Dead nodes have no loop, so return the top level loop instead
1235 if (!has_node(n)) return _ltree_root;
1236 assert(!has_ctrl(n), "");
1237 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1238 }
1239
1240 IdealLoopTree* ltree_root() const { return _ltree_root; }
1241
1242 // Is 'n' a (nested) member of 'loop'?
1243 int is_member( const IdealLoopTree *loop, Node *n ) const {
1244 return loop->is_member(get_loop(n)); }
1245
1246 // This is the basic building block of the loop optimizations. It clones an
1247 // entire loop body. It makes an old_new loop body mapping; with this
1248 // mapping you can find the new-loop equivalent to an old-loop node. All
1249 // new-loop nodes are exactly equal to their old-loop counterparts, all
1250 // edges are the same. All exits from the old-loop now have a RegionNode
1251 // that merges the equivalent new-loop path. This is true even for the
1252 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1253 // now come from (one or more) Phis that merge their new-loop equivalents.
1254 // Parameter side_by_side_idom:
1255 // When side_by_size_idom is null, the dominator tree is constructed for
1256 // the clone loop to dominate the original. Used in construction of
1257 // pre-main-post loop sequence.
1258 // When nonnull, the clone and original are side-by-side, both are
1259 // dominated by the passed in side_by_side_idom node. Used in
1260 // construction of unswitched loops.
1261 enum CloneLoopMode {
1262 IgnoreStripMined = 0, // Only clone inner strip mined loop
1263 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1264 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1265 // result control flow branches
1266 // either to inner clone or outer
1267 // strip mined loop.
1268 };
1269 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1270 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1271 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1272 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1273 Node_List*& split_if_set, Node_List*& split_bool_set,
1274 Node_List*& split_cex_set, Node_List& worklist,
1275 uint new_counter, CloneLoopMode mode);
1276 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1277 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1278 Node_List& extra_data_nodes);
1279
1280 // If we got the effect of peeling, either by actually peeling or by
1281 // making a pre-loop which must execute at least once, we can remove
1282 // all loop-invariant dominated tests in the main body.
1283 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1284
1285 // Generate code to do a loop peel for the given loop (and body).
1286 // old_new is a temp array.
1287 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1288
1289 // Add pre and post loops around the given loop. These loops are used
1290 // during RCE, unrolling and aligning loops.
1291 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1292
1293 // Add post loop after the given loop.
1294 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1295 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1296 Node*& incr, Node* limit, CountedLoopNode*& post_head);
1297
1298 // Add a vector post loop between a vector main loop and the current post loop
1299 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1300 // If Node n lives in the back_ctrl block, we clone a private version of n
1301 // in preheader_ctrl block and return that, otherwise return n.
1302 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1303
1304 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1305 // unroll to do double iterations. The next round of major loop transforms
1306 // will repeat till the doubled loop body does all remaining iterations in 1
1307 // pass.
1308 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1309
1310 // Unroll the loop body one step - make each trip do 2 iterations.
1311 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1312
1313 // Return true if exp is a constant times an induction var
1314 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1315
1316 bool is_iv(Node* exp, Node* iv, BasicType bt);
1317
1318 // Return true if exp is a scaled induction var plus (or minus) constant
1319 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);
1320 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1321 jlong long_scale;
1322 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1323 int int_scale = checked_cast<int>(long_scale);
1324 if (p_scale != nullptr) {
1325 *p_scale = int_scale;
1326 }
1327 return true;
1328 }
1329 return false;
1330 }
1331 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1332 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1333 BasicType bt,
1334 jlong* p_scale, Node** p_offset,
1335 bool* p_short_scale, int depth);
1336
1337 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1338 IfProjNode* create_new_if_for_predicate(IfProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1339 int opcode, bool rewire_uncommon_proj_phi_inputs = false,
1340 bool if_cont_is_true_proj = true);
1341
1342 private:
1343 // Helper functions for create_new_if_for_predicate()
1344 void set_ctrl_of_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl);
1345 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1346 Node* clone_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl);
1347 Dict clone_nodes(const Node_List& list_to_clone);
1348 void rewire_cloned_nodes_to_ctrl(const ProjNode* old_ctrl, Node* new_ctrl, const Node_List& nodes_with_same_ctrl,
1349 const Dict& old_new_mapping);
1350 void rewire_inputs_of_clones_to_clones(Node* new_ctrl, Node* clone, const Dict& old_new_mapping, const Node* next);
1351
1352 public:
1353 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1354
1355 // Construct a range check for a predicate if
1356 BoolNode* rc_predicate(IdealLoopTree* loop, Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1357 jint stride, Node* range, bool upper, bool& overflow);
1358
1359 // Implementation of the loop predication to promote checks outside the loop
1360 bool loop_predication_impl(IdealLoopTree *loop);
1361 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1362 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1363 Invariance& invar, Deoptimization::DeoptReason reason);
1364 bool loop_predication_should_follow_branches(IdealLoopTree* loop, IfProjNode* predicate_proj, float& loop_trip_cnt);
1365 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1366 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1367 Node_List& if_proj_list);
1368 IfProjNode* add_template_assertion_predicate(IfNode* iff, IdealLoopTree* loop, IfProjNode* if_proj, IfProjNode* predicate_proj,
1369 IfProjNode* upper_bound_proj, int scale, Node* offset, Node* init, Node* limit,
1370 jint stride, Node* rng, bool& overflow, Deoptimization::DeoptReason reason);
1371 Node* add_range_check_elimination_assertion_predicate(IdealLoopTree* loop, Node* predicate_proj, int scale_con,
1372 Node* offset, Node* limit, jint stride_con, Node* value);
1373
1374 // Helper function to collect predicate for eliminating the useless ones
1375 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1376 void eliminate_useless_predicates();
1377 void eliminate_useless_zero_trip_guard();
1378
1379 // Change the control input of expensive nodes to allow commoning by
1380 // IGVN when it is guaranteed to not result in a more frequent
1381 // execution of the expensive node. Return true if progress.
1382 bool process_expensive_nodes();
1383
1384 // Check whether node has become unreachable
1385 bool is_node_unreachable(Node *n) const {
1386 return !has_node(n) || n->is_unreachable(_igvn);
1387 }
1388
1389 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1390 void do_range_check(IdealLoopTree *loop, Node_List &old_new);
1391
1392 // Create a slow version of the loop by cloning the loop
1393 // and inserting an if to select fast-slow versions.
1394 // Return the inserted if.
1395 IfNode* create_slow_version_of_loop(IdealLoopTree *loop,
1396 Node_List &old_new,
1397 IfNode* unswitch_iff,
1398 CloneLoopMode mode);
1399
1400 // Clone a loop and return the clone head (clone_loop_head).
1401 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1402 // This routine was created for usage in CountedLoopReserveKit.
1403 //
1404 // int(1) -> If -> IfTrue -> original_loop_head
1405 // |
1406 // V
1407 // IfFalse -> clone_loop_head (returned by function pointer)
1408 //
1409 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1410 // Clone loop with an invariant test (that does not exit) and
1411 // insert a clone of the test that selects which version to
1412 // execute.
1413 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1414
1415 // Find candidate "if" for unswitching
1416 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1417
1418 // Range Check Elimination uses this function!
1419 // Constrain the main loop iterations so the affine function:
1420 // low_limit <= scale_con * I + offset < upper_limit
1421 // always holds true. That is, either increase the number of iterations in
1422 // the pre-loop or the post-loop until the condition holds true in the main
1423 // loop. Scale_con, offset and limit are all loop invariant.
1424 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);
1425 // Helper function for add_constraint().
1426 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1427
1428 // Partially peel loop up through last_peel node.
1429 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1430 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1431
1432 // Move UnorderedReduction out of loop if possible
1433 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop);
1434
1435 // Create a scheduled list of nodes control dependent on ctrl set.
1436 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1437 // Has a use in the vector set
1438 bool has_use_in_set( Node* n, VectorSet& vset );
1439 // Has use internal to the vector set (ie. not in a phi at the loop head)
1440 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1441 // clone "n" for uses that are outside of loop
1442 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1443 // clone "n" for special uses that are in the not_peeled region
1444 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1445 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1446 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1447 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1448 #ifdef ASSERT
1449 // Validate the loop partition sets: peel and not_peel
1450 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1451 // Ensure that uses outside of loop are of the right form
1452 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1453 uint orig_exit_idx, uint clone_exit_idx);
1454 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1455 #endif
1456
1457 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1458 int stride_of_possible_iv( Node* iff );
1459 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1460 // Return the (unique) control output node that's in the loop (if it exists.)
1461 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1462 // Insert a signed compare loop exit cloned from an unsigned compare.
1463 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1464 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1465 // Utility to register node "n" with PhaseIdealLoop
1466 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1467 // Utility to create an if-projection
1468 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1469 // Force the iff control output to be the live_proj
1470 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1471 // Insert a region before an if projection
1472 RegionNode* insert_region_before_proj(ProjNode* proj);
1473 // Insert a new if before an if projection
1474 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1475
1476 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1477 // "Nearly" because all Nodes have been cloned from the original in the loop,
1478 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1479 // through the Phi recursively, and return a Bool.
1480 Node* clone_iff(PhiNode* phi);
1481 CmpNode* clone_bool(PhiNode* phi);
1482
1483
1484 // Rework addressing expressions to get the most loop-invariant stuff
1485 // moved out. We'd like to do all associative operators, but it's especially
1486 // important (common) to do address expressions.
1487 Node* remix_address_expressions(Node* n);
1488 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1489
1490 // Convert add to muladd to generate MuladdS2I under certain criteria
1491 Node * convert_add_to_muladd(Node * n);
1492
1493 // Attempt to use a conditional move instead of a phi/branch
1494 Node *conditional_move( Node *n );
1495
1496 // Check for aggressive application of 'split-if' optimization,
1497 // using basic block level info.
1498 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1499 Node *split_if_with_blocks_pre ( Node *n );
1500 void split_if_with_blocks_post( Node *n );
1501 Node *has_local_phi_input( Node *n );
1502 // Mark an IfNode as being dominated by a prior test,
1503 // without actually altering the CFG (and hence IDOM info).
1504 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool exclude_loop_predicate = false);
1505
1506 // Split Node 'n' through merge point
1507 RegionNode* split_thru_region(Node* n, RegionNode* region);
1508 // Split Node 'n' through merge point if there is enough win.
1509 Node *split_thru_phi( Node *n, Node *region, int policy );
1510 // Found an If getting its condition-code input from a Phi in the
1511 // same block. Split thru the Region.
1512 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1513
1514 // Conversion of fill/copy patterns into intrinsic versions
1515 bool do_intrinsify_fill();
1516 bool intrinsify_fill(IdealLoopTree* lpt);
1517 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1518 Node*& shift, Node*& offset);
1519
1520 private:
1521 // Return a type based on condition control flow
1522 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1523 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1524 // Helpers for filtered type
1525 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1526
1527 // Helper functions
1528 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1529 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1530 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 );
1531 bool split_up( Node *n, Node *blk1, Node *blk2 );
1532 void sink_use( Node *use, Node *post_loop );
1533 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1534 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1535 void try_move_store_after_loop(Node* n);
1536 bool identical_backtoback_ifs(Node *n);
1537 bool can_split_if(Node *n_ctrl);
1538 bool cannot_split_division(const Node* n, const Node* region) const;
1539 static bool is_divisor_counted_loop_phi(const Node* divisor, const Node* loop);
1540 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1541
1542 // Determine if a method is too big for a/another round of split-if, based on
1543 // a magic (approximate) ratio derived from the equally magic constant 35000,
1544 // previously used for this purpose (but without relating to the node limit).
1545 bool must_throttle_split_if() {
1546 uint threshold = C->max_node_limit() * 2 / 5;
1547 return C->live_nodes() > threshold;
1548 }
1549
1550 // A simplistic node request tracking mechanism, where
1551 // = UINT_MAX Request not valid or made final.
1552 // < UINT_MAX Nodes currently requested (estimate).
1553 uint _nodes_required;
1554
1555 enum { REQUIRE_MIN = 70 };
1556
1557 uint nodes_required() const { return _nodes_required; }
1558
1559 // Given the _currently_ available number of nodes, check whether there is
1560 // "room" for an additional request or not, considering the already required
1561 // number of nodes. Return TRUE if the new request is exceeding the node
1562 // budget limit, otherwise return FALSE. Note that this interpretation will
1563 // act pessimistic on additional requests when new nodes have already been
1564 // generated since the 'begin'. This behaviour fits with the intention that
1565 // node estimates/requests should be made upfront.
1566 bool exceeding_node_budget(uint required = 0) {
1567 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1568 uint available = C->max_node_limit() - C->live_nodes();
1569 return available < required + _nodes_required + REQUIRE_MIN;
1570 }
1571
1572 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1573 precond(require > 0);
1574 _nodes_required += MAX2(require, minreq);
1575 return _nodes_required;
1576 }
1577
1578 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1579 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1580 }
1581
1582 uint require_nodes_begin() {
1583 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1584 _nodes_required = 0;
1585 return C->live_nodes();
1586 }
1587
1588 // When a node request is final, optionally check that the requested number
1589 // of nodes was reasonably correct with respect to the number of new nodes
1590 // introduced since the last 'begin'. Always check that we have not exceeded
1591 // the maximum node limit.
1592 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1593 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1594
1595 #ifdef ASSERT
1596 if (check_estimate) {
1597 // Check that the node budget request was not off by too much (x2).
1598 // Should this be the case we _surely_ need to improve the estimates
1599 // used in our budget calculations.
1600 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1601 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1602 C->live_nodes() - live_at_begin, _nodes_required);
1603 }
1604 }
1605 #endif
1606 // Assert that we have stayed within the node budget limit.
1607 assert(C->live_nodes() < C->max_node_limit(),
1608 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1609 C->live_nodes() - live_at_begin, live_at_begin,
1610 C->max_node_limit(), _nodes_required);
1611
1612 _nodes_required = UINT_MAX;
1613 }
1614
1615 // Clone Parse Predicates to slow and fast loop when unswitching a loop
1616 void clone_parse_and_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new,
1617 IfProjNode*& iffast_pred, IfProjNode*& ifslow_pred);
1618 IfProjNode* clone_parse_predicate_to_unswitched_loop(ParsePredicateSuccessProj* predicate_proj, Node* new_entry,
1619 Deoptimization::DeoptReason reason, bool slow_loop);
1620 void clone_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new,
1621 Deoptimization::DeoptReason reason, IfProjNode* old_predicate_proj,
1622 IfProjNode* iffast_pred, IfProjNode* ifslow_pred);
1623 IfProjNode* clone_assertion_predicate_for_unswitched_loops(Node* iff, IfProjNode* predicate,
1624 Deoptimization::DeoptReason reason,
1625 IfProjNode* output_proj);
1626 static void check_cloned_parse_predicate_for_unswitching(const Node* new_entry, bool is_fast_loop) PRODUCT_RETURN;
1627
1628 bool _created_loop_node;
1629 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1630 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1631
1632 public:
1633 void set_created_loop_node() { _created_loop_node = true; }
1634 bool created_loop_node() { return _created_loop_node; }
1635 void register_new_node(Node* n, Node* blk);
1636
1637 #ifdef ASSERT
1638 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1639 #endif
1640
1641 #ifndef PRODUCT
1642 void dump() const;
1643 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1644 void dump_idom(Node* n, uint count) const;
1645 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1646 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1647 IdealLoopTree* get_loop_idx(Node* n) const {
1648 // Dead nodes have no loop, so return the top level loop instead
1649 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1650 }
1651 // Print some stats
1652 static void print_statistics();
1653 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1654 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1655 static volatile int _long_loop_candidates;
1656 static volatile int _long_loop_nests;
1657 static volatile int _long_loop_counted_loops;
1658 #endif
1659
1660 #ifdef ASSERT
1661 void verify() const;
1662 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1663 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1664 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1665 #endif
1666
1667 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1668
1669 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1670
1671 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1672
1673
1674 int extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi,
1675 Node_List &range_checks);
1676
1677 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1678 Node* inner_iters_actual_int, Node* inner_phi,
1679 Node* iv_add, LoopNode* inner_head);
1680
1681 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1682
1683 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1684
1685 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1686
1687 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1688
1689 void try_sink_out_of_loop(Node* n);
1690
1691 Node* clamp(Node* R, Node* L, Node* H);
1692
1693 bool safe_for_if_replacement(const Node* dom) const;
1694
1695 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1696
1697 bool try_merge_identical_ifs(Node* n);
1698
1699 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1700
1701 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1702 IdealLoopTree* parent, bool partial);
1703
1704 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1705 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1706
1707 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1708 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1709 Node_List*& split_bool_set, Node_List*& split_cex_set);
1710
1711 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1712
1713 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1714
1715 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1716
1717 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1718
1719 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1720 };
1721
1722
1723 class AutoNodeBudget : public StackObj
1724 {
1725 public:
1726 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1727
1728 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1729 : _phase(phase),
1730 _check_at_final(chk == BUDGET_CHECK),
1731 _nodes_at_begin(0)
1732 {
1733 precond(_phase != nullptr);
1734
1735 _nodes_at_begin = _phase->require_nodes_begin();
1736 }
1737
1738 ~AutoNodeBudget() {
1739 #ifndef PRODUCT
1740 if (TraceLoopOpts) {
1741 uint request = _phase->nodes_required();
1742 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1743
1744 if (request < delta) {
1745 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1746 } else {
1747 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1748 // Identify the worst estimates as "poor" ones.
1749 if (request > REQUIRE_MIN && delta > 0) {
1750 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1751 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1752 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1753 }
1754 }
1755 }
1756 }
1757 #endif // PRODUCT
1758 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1759 }
1760
1761 private:
1762 PhaseIdealLoop* _phase;
1763 bool _check_at_final;
1764 uint _nodes_at_begin;
1765 };
1766
1767
1768 // This kit may be used for making of a reserved copy of a loop before this loop
1769 // goes under non-reversible changes.
1770 //
1771 // Function create_reserve() creates a reserved copy (clone) of the loop.
1772 // The reserved copy is created by calling
1773 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1774 // the original and reserved loops are connected in the outer graph.
1775 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1776 //
1777 // By default the reserved copy (clone) of the loop is created as dead code - it is
1778 // dominated in the outer loop by this node chain:
1779 // intcon(1)->If->IfFalse->reserved_copy.
1780 // The original loop is dominated by the same node chain but IfTrue projection:
1781 // intcon(0)->If->IfTrue->original_loop.
1782 //
1783 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1784 // and the dtor, checks _use_new value.
1785 // If _use_new == false, it "switches" control to reserved copy of the loop
1786 // by simple replacing of node intcon(1) with node intcon(0).
1787 //
1788 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1789 //
1790 // void CountedLoopReserveKit_example()
1791 // {
1792 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1793 // if (DoReserveCopy && !lrk.has_reserved()) {
1794 // return; //failed to create reserved loop copy
1795 // }
1796 // ...
1797 // //something is wrong, switch to original loop
1798 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1799 // ...
1800 // //everything worked ok, return with the newly modified loop
1801 // lrk.use_new();
1802 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1803 // }
1804 //
1805 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1806 // the dtor will "switch to the original" loop.
1807 // NOTE. You you modify outside of the original loop this class is no help.
1808 //
1809 class CountedLoopReserveKit {
1810 private:
1811 PhaseIdealLoop* _phase;
1812 IdealLoopTree* _lpt;
1813 LoopNode* _lp;
1814 IfNode* _iff;
1815 LoopNode* _lp_reserved;
1816 bool _has_reserved;
1817 bool _use_new;
1818 const bool _active; //may be set to false in ctor, then the object is dummy
1819
1820 public:
1821 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1822 ~CountedLoopReserveKit();
1823 void use_new() {_use_new = true;}
1824 void set_iff(IfNode* x) {_iff = x;}
1825 bool has_reserved() const { return _active && _has_reserved;}
1826 private:
1827 bool create_reserve();
1828 };// class CountedLoopReserveKit
1829
1830 inline Node* IdealLoopTree::tail() {
1831 // Handle lazy update of _tail field.
1832 if (_tail->in(0) == nullptr) {
1833 _tail = _phase->get_ctrl(_tail);
1834 }
1835 return _tail;
1836 }
1837
1838 inline Node* IdealLoopTree::head() {
1839 // Handle lazy update of _head field.
1840 if (_head->in(0) == nullptr) {
1841 _head = _phase->get_ctrl(_head);
1842 }
1843 return _head;
1844 }
1845
1846 // Iterate over the loop tree using a preorder, left-to-right traversal.
1847 //
1848 // Example that visits all counted loops from within PhaseIdealLoop
1849 //
1850 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1851 // IdealLoopTree* lpt = iter.current();
1852 // if (!lpt->is_counted()) continue;
1853 // ...
1854 class LoopTreeIterator : public StackObj {
1855 private:
1856 IdealLoopTree* _root;
1857 IdealLoopTree* _curnt;
1858
1859 public:
1860 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1861
1862 bool done() { return _curnt == nullptr; } // Finished iterating?
1863
1864 void next(); // Advance to next loop tree
1865
1866 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1867 };
1868
1869 // Compute probability of reaching some CFG node from a fixed
1870 // dominating CFG node
1871 class PathFrequency {
1872 private:
1873 Node* _dom; // frequencies are computed relative to this node
1874 Node_Stack _stack;
1875 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1876 GrowableArray<float> _freqs; // cache frequencies
1877 PhaseIdealLoop* _phase;
1878
1879 float check_and_truncate_frequency(float f) {
1880 assert(f >= 0, "Incorrect frequency");
1881 // We do not perform an exact (f <= 1) check
1882 // this would be error prone with rounding of floats.
1883 // Performing a check like (f <= 1+eps) would be of benefit,
1884 // however, it is not evident how to determine such an eps,
1885 // given that an arbitrary number of add/mul operations
1886 // are performed on these frequencies.
1887 return (f > 1) ? 1 : f;
1888 }
1889
1890 public:
1891 PathFrequency(Node* dom, PhaseIdealLoop* phase)
1892 : _dom(dom), _stack(0), _phase(phase) {
1893 }
1894
1895 float to(Node* n);
1896 };
1897
1898 // Utility class to work on predicates.
1899 class Predicates {
1900 public:
1901 static Node* skip_all_predicates(Node* node);
1902 static Node* skip_all_predicates(ParsePredicates& parse_predicates);
1903 static Node* skip_predicates_in_block(ParsePredicateSuccessProj* parse_predicate_success_proj);
1904 static IfProjNode* next_predicate_proj_in_block(IfProjNode* proj);
1905 static bool has_profiled_loop_predicates(ParsePredicates& parse_predicates);
1906 };
1907
1908 // Class representing the Parse Predicates that are added during parsing with ParsePredicateNodes.
1909 class ParsePredicates {
1910 private:
1911 ParsePredicateSuccessProj* _loop_predicate_proj = nullptr;
1912 ParsePredicateSuccessProj* _profiled_loop_predicate_proj = nullptr;
1913 ParsePredicateSuccessProj* _loop_limit_check_predicate_proj = nullptr;
1914 // The success projection of the Parse Predicate that comes first when starting from root.
1915 ParsePredicateSuccessProj* _top_predicate_proj;
1916 ParsePredicateSuccessProj* _starting_proj;
1917
1918 void find_parse_predicate_projections();
1919 static bool is_uct_proj(Node* node, Deoptimization::DeoptReason deopt_reason);
1920 static ParsePredicateNode* get_parse_predicate_or_null(Node* proj);
1921 bool assign_predicate_proj(ParsePredicateSuccessProj* parse_predicate_proj);
1922 public:
1923 ParsePredicates(Node* starting_proj);
1924
1925 // Success projection of Loop Parse Predicate.
1926 ParsePredicateSuccessProj* loop_predicate_proj() {
1927 return _loop_predicate_proj;
1928 }
1929
1930 // Success proj of Profiled Loop Parse Predicate.
1931 ParsePredicateSuccessProj* profiled_loop_predicate_proj() {
1932 return _profiled_loop_predicate_proj;
1933 }
1934
1935 // Success proj of Loop Limit Check Parse Predicate.
1936 ParsePredicateSuccessProj* loop_limit_check_predicate_proj() {
1937 return _loop_limit_check_predicate_proj;
1938 }
1939
1940 // Return the success projection of the Parse Predicate that comes first when starting from root.
1941 ParsePredicateSuccessProj* get_top_predicate_proj() {
1942 return _top_predicate_proj;
1943 }
1944
1945 static bool is_success_proj(Node* node);
1946
1947 // Are there any Parse Predicates?
1948 bool has_any() const {
1949 return _top_predicate_proj != nullptr;
1950 }
1951
1952 static bool is_loop_limit_check_predicate_proj(Node* node) {
1953 ParsePredicateNode* parse_predicate = get_parse_predicate_or_null(node);
1954 return parse_predicate != nullptr && parse_predicate->deopt_reason() == Deoptimization::DeoptReason::Reason_loop_limit_check;
1955 }
1956 };
1957 #endif // SHARE_OPTO_LOOPNODE_HPP