1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_OPTO_LOOPNODE_HPP
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class UnswitchCandidate;
47 class 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 FlatArrays = 1<<17};
85 char _unswitch_count;
86 enum { _unswitch_max=3 };
87
88 // Expected trip count from profile data
89 float _profile_trip_cnt;
90
91 public:
92 // Names for edge indices
93 enum { Self=0, EntryControl, LoopBackControl };
94
95 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
96 void set_inner_loop() { _loop_flags |= InnerLoop; }
97
98 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
99 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
100 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
101 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
102 bool is_strip_mined() const { return _loop_flags & StripMined; }
103 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
104 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
105 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
106 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
107 bool is_flat_arrays() const { return _loop_flags & FlatArrays; }
108
109 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
110 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
111 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
112 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
113 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
114 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
115 void mark_strip_mined() { _loop_flags |= StripMined; }
116 void clear_strip_mined() { _loop_flags &= ~StripMined; }
117 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
118 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
119 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
120 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
121 void mark_flat_arrays() { _loop_flags |= FlatArrays; }
122
123 int unswitch_max() { return _unswitch_max; }
124 int unswitch_count() { return _unswitch_count; }
125
126 void set_unswitch_count(int val) {
127 assert (val <= unswitch_max(), "too many unswitches");
128 _unswitch_count = val;
129 }
130
131 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
132 float profile_trip_cnt() { return _profile_trip_cnt; }
133
134 LoopNode(Node *entry, Node *backedge)
135 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
136 _profile_trip_cnt(COUNT_UNKNOWN) {
137 init_class_id(Class_Loop);
138 init_req(EntryControl, entry);
139 init_req(LoopBackControl, backedge);
140 }
141
142 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
143 virtual int Opcode() const;
144 bool can_be_counted_loop(PhaseValues* phase) const {
145 return req() == 3 && in(0) != nullptr &&
146 in(1) != nullptr && phase->type(in(1)) != Type::TOP &&
147 in(2) != nullptr && phase->type(in(2)) != Type::TOP;
148 }
149 bool is_valid_counted_loop(BasicType bt) const;
150 #ifndef PRODUCT
151 virtual void dump_spec(outputStream *st) const;
152 #endif
153
154 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
155 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
156 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; }
157 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; }
158 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; }
159 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; }
160 };
161
162 //------------------------------Counted Loops----------------------------------
163 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
164 // path (and maybe some other exit paths). The trip-counter exit is always
165 // last in the loop. The trip-counter have to stride by a constant;
166 // the exit value is also loop invariant.
167
168 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
169 // CountedLoopNode has the incoming loop control and the loop-back-control
170 // which is always the IfTrue before the matching CountedLoopEndNode. The
171 // CountedLoopEndNode has an incoming control (possibly not the
172 // CountedLoopNode if there is control flow in the loop), the post-increment
173 // trip-counter value, and the limit. The trip-counter value is always of
174 // the form (Op old-trip-counter stride). The old-trip-counter is produced
175 // by a Phi connected to the CountedLoopNode. The stride is constant.
176 // The Op is any commutable opcode, including Add, Mul, Xor. The
177 // CountedLoopEndNode also takes in the loop-invariant limit value.
178
179 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
180 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
181 // via the old-trip-counter from the Op node.
182
183 //------------------------------CountedLoopNode--------------------------------
184 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
185 // inputs the incoming loop-start control and the loop-back control, so they
186 // act like RegionNodes. They also take in the initial trip counter, the
187 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
188 // produce a loop-body control and the trip counter value. Since
189 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
190
191 class BaseCountedLoopNode : public LoopNode {
192 public:
193 BaseCountedLoopNode(Node *entry, Node *backedge)
194 : LoopNode(entry, backedge) {
195 }
196
197 Node *init_control() const { return in(EntryControl); }
198 Node *back_control() const { return in(LoopBackControl); }
199
200 Node* init_trip() const;
201 Node* stride() const;
202 bool stride_is_con() const;
203 Node* limit() const;
204 Node* incr() const;
205 Node* phi() const;
206
207 BaseCountedLoopEndNode* loopexit_or_null() const;
208 BaseCountedLoopEndNode* loopexit() const;
209
210 virtual BasicType bt() const = 0;
211
212 jlong stride_con() const;
213
214 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
215 };
216
217
218 class CountedLoopNode : public BaseCountedLoopNode {
219 // Size is bigger to hold _main_idx. However, _main_idx does not change
220 // the semantics so it does not appear in the hash & cmp functions.
221 virtual uint size_of() const { return sizeof(*this); }
222
223 // For Pre- and Post-loops during debugging ONLY, this holds the index of
224 // the Main CountedLoop. Used to assert that we understand the graph shape.
225 node_idx_t _main_idx;
226
227 // Known trip count calculated by compute_exact_trip_count()
228 uint _trip_count;
229
230 // Log2 of original loop bodies in unrolled loop
231 int _unrolled_count_log2;
232
233 // Node count prior to last unrolling - used to decide if
234 // unroll,optimize,unroll,optimize,... is making progress
235 int _node_count_before_unroll;
236
237 // If slp analysis is performed we record the maximum
238 // vector mapped unroll factor here
239 int _slp_maximum_unroll_factor;
240
241 public:
242 CountedLoopNode(Node *entry, Node *backedge)
243 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
244 _unrolled_count_log2(0), _node_count_before_unroll(0),
245 _slp_maximum_unroll_factor(0) {
246 init_class_id(Class_CountedLoop);
247 // Initialize _trip_count to the largest possible value.
248 // Will be reset (lower) if the loop's trip count is known.
249 }
250
251 virtual int Opcode() const;
252 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
253
254 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
255 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
256 int stride_con() const;
257
258 // Match increment with optional truncation
259 static Node*
260 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
261 BasicType bt);
262
263 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
264 // can run short a few iterations and may start a few iterations in.
265 // It will be RCE'd and unrolled and aligned.
266
267 // A following 'post' loop will run any remaining iterations. Used
268 // during Range Check Elimination, the 'post' loop will do any final
269 // iterations with full checks. Also used by Loop Unrolling, where
270 // the 'post' loop will do any epilog iterations needed. Basically,
271 // a 'post' loop can not profitably be further unrolled or RCE'd.
272
273 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
274 // it may do under-flow checks for RCE and may do alignment iterations
275 // so the following main loop 'knows' that it is striding down cache
276 // lines.
277
278 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
279 // Aligned, may be missing it's pre-loop.
280 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
281 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
282 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
283 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
284 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
285 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
286 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
287 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
288 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
289 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
290
291 int main_idx() const { return _main_idx; }
292
293
294 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
295 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
296 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
297 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
298
299 void set_trip_count(uint tc) { _trip_count = tc; }
300 uint trip_count() { return _trip_count; }
301
302 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
303 void set_exact_trip_count(uint tc) {
304 _trip_count = tc;
305 _loop_flags |= HasExactTripCount;
306 }
307 void set_nonexact_trip_count() {
308 _loop_flags &= ~HasExactTripCount;
309 }
310 void set_notpassed_slp() {
311 _loop_flags &= ~PassedSlpAnalysis;
312 }
313
314 void double_unrolled_count() { _unrolled_count_log2++; }
315 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
316
317 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
318 int node_count_before_unroll() { return _node_count_before_unroll; }
319 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
320 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
321
322 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
323 OuterStripMinedLoopNode* outer_loop() const;
324 virtual IfTrueNode* outer_loop_tail() const;
325 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
326 virtual IfFalseNode* outer_loop_exit() const;
327 virtual SafePointNode* outer_safepoint() const;
328
329 Node* skip_assertion_predicates_with_halt();
330
331 virtual BasicType bt() const {
332 return T_INT;
333 }
334
335 Node* is_canonical_loop_entry();
336 CountedLoopEndNode* find_pre_loop_end();
337
338 #ifndef PRODUCT
339 virtual void dump_spec(outputStream *st) const;
340 #endif
341 };
342
343 class LongCountedLoopNode : public BaseCountedLoopNode {
344 public:
345 LongCountedLoopNode(Node *entry, Node *backedge)
346 : BaseCountedLoopNode(entry, backedge) {
347 init_class_id(Class_LongCountedLoop);
348 }
349
350 virtual int Opcode() const;
351
352 virtual BasicType bt() const {
353 return T_LONG;
354 }
355
356 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
357 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
358 };
359
360
361 //------------------------------CountedLoopEndNode-----------------------------
362 // CountedLoopEndNodes end simple trip counted loops. They act much like
363 // IfNodes.
364
365 class BaseCountedLoopEndNode : public IfNode {
366 public:
367 enum { TestControl, TestValue };
368 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
369 : IfNode(control, test, prob, cnt) {
370 init_class_id(Class_BaseCountedLoopEnd);
371 }
372
373 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; }
374 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
375 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
376 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; }
377 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; }
378 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); }
379
380 PhiNode* phi() const {
381 Node* tmp = incr();
382 if (tmp && tmp->req() == 3) {
383 Node* phi = tmp->in(1);
384 if (phi->is_Phi()) {
385 return phi->as_Phi();
386 }
387 }
388 return nullptr;
389 }
390
391 BaseCountedLoopNode* loopnode() const {
392 // The CountedLoopNode that goes with this CountedLoopEndNode may
393 // have been optimized out by the IGVN so be cautious with the
394 // pattern matching on the graph
395 PhiNode* iv_phi = phi();
396 if (iv_phi == nullptr) {
397 return nullptr;
398 }
399 Node* ln = iv_phi->in(0);
400 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
401 return nullptr;
402 }
403 if (ln->as_BaseCountedLoop()->bt() != bt()) {
404 return nullptr;
405 }
406 return ln->as_BaseCountedLoop();
407 }
408
409 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
410
411 jlong stride_con() const;
412 virtual BasicType bt() const = 0;
413
414 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
415 };
416
417 class CountedLoopEndNode : public BaseCountedLoopEndNode {
418 public:
419
420 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
421 : BaseCountedLoopEndNode(control, test, prob, cnt) {
422 init_class_id(Class_CountedLoopEnd);
423 }
424 virtual int Opcode() const;
425
426 CountedLoopNode* loopnode() const {
427 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
428 }
429
430 virtual BasicType bt() const {
431 return T_INT;
432 }
433
434 #ifndef PRODUCT
435 virtual void dump_spec(outputStream *st) const;
436 #endif
437 };
438
439 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
440 public:
441 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
442 : BaseCountedLoopEndNode(control, test, prob, cnt) {
443 init_class_id(Class_LongCountedLoopEnd);
444 }
445
446 LongCountedLoopNode* loopnode() const {
447 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
448 }
449
450 virtual int Opcode() const;
451
452 virtual BasicType bt() const {
453 return T_LONG;
454 }
455 };
456
457
458 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
459 Node* bctrl = back_control();
460 if (bctrl == nullptr) return nullptr;
461
462 Node* lexit = bctrl->in(0);
463 if (!lexit->is_BaseCountedLoopEnd()) {
464 return nullptr;
465 }
466 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
467 if (result->bt() != bt()) {
468 return nullptr;
469 }
470 return result;
471 }
472
473 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
474 BaseCountedLoopEndNode* cle = loopexit_or_null();
475 assert(cle != nullptr, "loopexit is null");
476 return cle;
477 }
478
479 inline Node* BaseCountedLoopNode::init_trip() const {
480 BaseCountedLoopEndNode* cle = loopexit_or_null();
481 return cle != nullptr ? cle->init_trip() : nullptr;
482 }
483 inline Node* BaseCountedLoopNode::stride() const {
484 BaseCountedLoopEndNode* cle = loopexit_or_null();
485 return cle != nullptr ? cle->stride() : nullptr;
486 }
487
488 inline bool BaseCountedLoopNode::stride_is_con() const {
489 BaseCountedLoopEndNode* cle = loopexit_or_null();
490 return cle != nullptr && cle->stride_is_con();
491 }
492 inline Node* BaseCountedLoopNode::limit() const {
493 BaseCountedLoopEndNode* cle = loopexit_or_null();
494 return cle != nullptr ? cle->limit() : nullptr;
495 }
496 inline Node* BaseCountedLoopNode::incr() const {
497 BaseCountedLoopEndNode* cle = loopexit_or_null();
498 return cle != nullptr ? cle->incr() : nullptr;
499 }
500 inline Node* BaseCountedLoopNode::phi() const {
501 BaseCountedLoopEndNode* cle = loopexit_or_null();
502 return cle != nullptr ? cle->phi() : nullptr;
503 }
504
505 inline jlong BaseCountedLoopNode::stride_con() const {
506 BaseCountedLoopEndNode* cle = loopexit_or_null();
507 return cle != nullptr ? cle->stride_con() : 0;
508 }
509
510
511 //------------------------------LoopLimitNode-----------------------------
512 // Counted Loop limit node which represents exact final iterator value:
513 // trip_count = (limit - init_trip + stride - 1)/stride
514 // final_value= trip_count * stride + init_trip.
515 // Use HW instructions to calculate it when it can overflow in integer.
516 // Note, final_value should fit into integer since counted loop has
517 // limit check: limit <= max_int-stride.
518 class LoopLimitNode : public Node {
519 enum { Init=1, Limit=2, Stride=3 };
520 public:
521 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(nullptr,init,limit,stride) {
522 // Put it on the Macro nodes list to optimize during macro nodes expansion.
523 init_flags(Flag_is_macro);
524 C->add_macro_node(this);
525 }
526 virtual int Opcode() const;
527 virtual const Type *bottom_type() const { return TypeInt::INT; }
528 virtual uint ideal_reg() const { return Op_RegI; }
529 virtual const Type* Value(PhaseGVN* phase) const;
530 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
531 virtual Node* Identity(PhaseGVN* phase);
532 };
533
534 // Support for strip mining
535 class OuterStripMinedLoopNode : public LoopNode {
536 private:
537 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
538
539 public:
540 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
541 : LoopNode(entry, backedge) {
542 init_class_id(Class_OuterStripMinedLoop);
543 init_flags(Flag_is_macro);
544 C->add_macro_node(this);
545 }
546
547 virtual int Opcode() const;
548
549 virtual IfTrueNode* outer_loop_tail() const;
550 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
551 virtual IfFalseNode* outer_loop_exit() const;
552 virtual SafePointNode* outer_safepoint() const;
553 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
554
555 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
556
557 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
558
559 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
560
561 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
562 PhaseIdealLoop* iloop);
563 };
564
565 class OuterStripMinedLoopEndNode : public IfNode {
566 public:
567 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
568 : IfNode(control, test, prob, cnt) {
569 init_class_id(Class_OuterStripMinedLoopEnd);
570 }
571
572 virtual int Opcode() const;
573
574 virtual const Type* Value(PhaseGVN* phase) const;
575 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
576
577 bool is_expanded(PhaseGVN *phase) const;
578 };
579
580 // -----------------------------IdealLoopTree----------------------------------
581 class IdealLoopTree : public ResourceObj {
582 public:
583 IdealLoopTree *_parent; // Parent in loop tree
584 IdealLoopTree *_next; // Next sibling in loop tree
585 IdealLoopTree *_child; // First child in loop tree
586
587 // The head-tail backedge defines the loop.
588 // If a loop has multiple backedges, this is addressed during cleanup where
589 // we peel off the multiple backedges, merging all edges at the bottom and
590 // ensuring that one proper backedge flow into the loop.
591 Node *_head; // Head of loop
592 Node *_tail; // Tail of loop
593 inline Node *tail(); // Handle lazy update of _tail field
594 inline Node *head(); // Handle lazy update of _head field
595 PhaseIdealLoop* _phase;
596 int _local_loop_unroll_limit;
597 int _local_loop_unroll_factor;
598
599 Node_List _body; // Loop body for inner loops
600
601 uint16_t _nest; // Nesting depth
602 uint8_t _irreducible:1, // True if irreducible
603 _has_call:1, // True if has call safepoint
604 _has_sfpt:1, // True if has non-call safepoint
605 _rce_candidate:1, // True if candidate for range check elimination
606 _has_range_checks:1,
607 _has_range_checks_computed:1;
608
609 Node_List* _safepts; // List of safepoints in this loop
610 Node_List* _required_safept; // A inner loop cannot delete these safepts;
611 bool _allow_optimizations; // Allow loop optimizations
612
613 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
614 : _parent(nullptr), _next(nullptr), _child(nullptr),
615 _head(head), _tail(tail),
616 _phase(phase),
617 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
618 _body(Compile::current()->comp_arena()),
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 bool can_apply_loop_predication();
655
656 // Perform iteration-splitting on inner loops. Split iterations to
657 // avoid range checks or one-shot null checks. Returns false if the
658 // current round of loop opts should stop.
659 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
660
661 // Driver for various flavors of iteration splitting. Returns false
662 // if the current round of loop opts should stop.
663 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
664
665 // Given dominators, try to find loops with calls that must always be
666 // executed (call dominates loop tail). These loops do not need non-call
667 // safepoints (ncsfpt).
668 void check_safepts(VectorSet &visited, Node_List &stack);
669
670 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
671 // encountered.
672 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
673
674 // Remove safepoints from loop. Optionally keeping one.
675 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
676
677 // Convert to counted loops where possible
678 void counted_loop( PhaseIdealLoop *phase );
679
680 // Check for Node being a loop-breaking test
681 Node *is_loop_exit(Node *iff) const;
682
683 // Remove simplistic dead code from loop body
684 void DCE_loop_body();
685
686 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
687 // Replace with a 1-in-10 exit guess.
688 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
689
690 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
691 // Useful for unrolling loops with NO array accesses.
692 bool policy_peel_only( PhaseIdealLoop *phase ) const;
693
694 // Return TRUE or FALSE if the loop should be unswitched -- clone
695 // loop with an invariant test
696 bool policy_unswitching( PhaseIdealLoop *phase ) const;
697 bool no_unswitch_candidate() const;
698
699 // Micro-benchmark spamming. Remove empty loops.
700 bool do_remove_empty_loop( PhaseIdealLoop *phase );
701
702 // Convert one iteration loop into normal code.
703 bool do_one_iteration_loop( PhaseIdealLoop *phase );
704
705 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
706 // move some loop-invariant test (usually a null-check) before the loop.
707 bool policy_peeling(PhaseIdealLoop *phase);
708
709 uint estimate_peeling(PhaseIdealLoop *phase);
710
711 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
712 // known trip count in the counted loop node.
713 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
714
715 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
716 // if the loop is a counted loop and the loop body is small enough.
717 bool policy_unroll(PhaseIdealLoop *phase);
718
719 // Loop analyses to map to a maximal superword unrolling for vectorization.
720 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
721
722 // Return TRUE or FALSE if the loop should be range-check-eliminated.
723 // Gather a list of IF tests that are dominated by iteration splitting;
724 // also gather the end of the first split and the start of the 2nd split.
725 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
726
727 // Return TRUE if "iff" is a range check.
728 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const;
729 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
730 jlong& scale) const;
731
732 // Estimate the number of nodes required when cloning a loop (body).
733 uint est_loop_clone_sz(uint factor) const;
734 // Estimate the number of nodes required when unrolling a loop (body).
735 uint est_loop_unroll_sz(uint factor) const;
736
737 // Compute loop trip count if possible
738 void compute_trip_count(PhaseIdealLoop* phase);
739
740 // Compute loop trip count from profile data
741 float compute_profile_trip_cnt_helper(Node* n);
742 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
743
744 // Reassociate invariant expressions.
745 void reassociate_invariants(PhaseIdealLoop *phase);
746 // Reassociate invariant binary expressions.
747 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
748 // Reassociate invariant add, subtract, and compare expressions.
749 Node* reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase);
750 // Return nonzero index of invariant operand if invariant and variant
751 // are combined with an associative binary. Helper for reassociate_invariants.
752 int find_invariant(Node* n, PhaseIdealLoop *phase);
753 // Return TRUE if "n" is associative.
754 bool is_associative(Node* n, Node* base=nullptr);
755 // Return TRUE if "n" is an associative cmp node.
756 bool is_associative_cmp(Node* n);
757
758 // Return true if n is invariant
759 bool is_invariant(Node* n) const;
760
761 // Put loop body on igvn work list
762 void record_for_igvn();
763
764 bool is_root() { return _parent == nullptr; }
765 // A proper/reducible loop w/o any (occasional) dead back-edge.
766 bool is_loop() { return !_irreducible && !tail()->is_top(); }
767 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
768 bool is_innermost() { return is_loop() && _child == nullptr; }
769
770 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
771
772 bool compute_has_range_checks() const;
773 bool range_checks_present() {
774 if (!_has_range_checks_computed) {
775 if (compute_has_range_checks()) {
776 _has_range_checks = 1;
777 }
778 _has_range_checks_computed = 1;
779 }
780 return _has_range_checks;
781 }
782
783 // Return the parent's IdealLoopTree for a strip mined loop which is the outer strip mined loop.
784 // In all other cases, return this.
785 IdealLoopTree* skip_strip_mined() {
786 return _head->as_Loop()->is_strip_mined() ? _parent : this;
787 }
788
789 #ifndef PRODUCT
790 void dump_head(); // Dump loop head only
791 void dump(); // Dump this loop recursively
792 #endif
793
794 #ifdef ASSERT
795 GrowableArray<IdealLoopTree*> collect_sorted_children() const;
796 bool verify_tree(IdealLoopTree* loop_verify) const;
797 #endif
798
799 private:
800 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
801
802 // Estimate the number of nodes resulting from control and data flow merge.
803 uint est_loop_flow_merge_sz() const;
804
805 // Check if the number of residual iterations is large with unroll_cnt.
806 // Return true if the residual iterations are more than 10% of the trip count.
807 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
808 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
809 }
810
811 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const;
812
813 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const;
814
815 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const;
816
817 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
818 Node* sfpt) const;
819
820 bool empty_loop_candidate(PhaseIdealLoop* phase) const;
821
822 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const;
823 };
824
825 // -----------------------------PhaseIdealLoop---------------------------------
826 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
827 // into a loop tree. Drives the loop-based transformations on the ideal graph.
828 class PhaseIdealLoop : public PhaseTransform {
829 friend class IdealLoopTree;
830 friend class SuperWord;
831 friend class ShenandoahBarrierC2Support;
832 friend class AutoNodeBudget;
833
834 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes.
835 Node_List _loop_or_ctrl;
836
837 // Pre-computed def-use info
838 PhaseIterGVN &_igvn;
839
840 // Head of loop tree
841 IdealLoopTree* _ltree_root;
842
843 // Array of pre-order numbers, plus post-visited bit.
844 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
845 // ODD for post-visited. Other bits are the pre-order number.
846 uint *_preorders;
847 uint _max_preorder;
848
849 ReallocMark _nesting; // Safety checks for arena reallocation
850
851 const PhaseIdealLoop* _verify_me;
852 bool _verify_only;
853
854 // Allocate _preorders[] array
855 void allocate_preorders() {
856 _max_preorder = C->unique()+8;
857 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
858 memset(_preorders, 0, sizeof(uint) * _max_preorder);
859 }
860
861 // Allocate _preorders[] array
862 void reallocate_preorders() {
863 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
864 if ( _max_preorder < C->unique() ) {
865 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
866 _max_preorder = C->unique();
867 }
868 memset(_preorders, 0, sizeof(uint) * _max_preorder);
869 }
870
871 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
872 // adds new nodes.
873 void check_grow_preorders( ) {
874 _nesting.check(); // Check if a potential re-allocation in the resource arena is safe
875 if ( _max_preorder < C->unique() ) {
876 uint newsize = _max_preorder<<1; // double size of array
877 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
878 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
879 _max_preorder = newsize;
880 }
881 }
882 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
883 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
884 // Pre-order numbers are written to the Nodes array as low-bit-set values.
885 void set_preorder_visited( Node *n, int pre_order ) {
886 assert( !is_visited( n ), "already set" );
887 _preorders[n->_idx] = (pre_order<<1);
888 };
889 // Return pre-order number.
890 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
891
892 // Check for being post-visited.
893 // Should be previsited already (checked with assert(is_visited(n))).
894 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
895
896 // Mark as post visited
897 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
898
899 public:
900 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
901 // Returns true if "n" is a data node, false if it's a control node.
902 bool has_ctrl(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; }
903
904 private:
905 // clear out dead code after build_loop_late
906 Node_List _deadlist;
907 Node_List _zero_trip_guard_opaque_nodes;
908
909 // Support for faster execution of get_late_ctrl()/dom_lca()
910 // when a node has many uses and dominator depth is deep.
911 GrowableArray<jlong> _dom_lca_tags;
912 uint _dom_lca_tags_round;
913 void init_dom_lca_tags();
914
915 // Helper for debugging bad dominance relationships
916 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
917
918 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
919
920 // Inline wrapper for frequent cases:
921 // 1) only one use
922 // 2) a use is the same as the current LCA passed as 'n1'
923 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
924 assert( n->is_CFG(), "" );
925 // Fast-path null lca
926 if( lca != nullptr && lca != n ) {
927 assert( lca->is_CFG(), "" );
928 // find LCA of all uses
929 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
930 }
931 return find_non_split_ctrl(n);
932 }
933 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
934
935 // Helper function for directing control inputs away from CFG split points.
936 Node *find_non_split_ctrl( Node *ctrl ) const {
937 if (ctrl != nullptr) {
938 if (ctrl->is_MultiBranch()) {
939 ctrl = ctrl->in(0);
940 }
941 assert(ctrl->is_CFG(), "CFG");
942 }
943 return ctrl;
944 }
945
946 #ifdef ASSERT
947 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
948 #endif
949 void copy_assertion_predicates_to_main_loop_helper(const PredicateBlock* predicate_block, Node* init, Node* stride,
950 IdealLoopTree* outer_loop, LoopNode* outer_main_head,
951 uint dd_main_head, uint idx_before_pre_post,
952 uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main,
953 Node* zero_trip_guard_proj_post, const Node_List &old_new);
954 void copy_assertion_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop,
955 LoopNode* outer_main_head, uint dd_main_head, uint idx_before_pre_post,
956 uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main,
957 Node* zero_trip_guard_proj_post, const Node_List& old_new);
958 Node* clone_template_assertion_predicate(IfNode* iff, Node* new_init, Node* predicate, Node* uncommon_proj, Node* control,
959 IdealLoopTree* outer_loop, Node* input_proj);
960 IfTrueNode* create_initialized_assertion_predicate(IfNode* template_assertion_predicate, Node* new_init,
961 Node* new_stride, Node* control);
962 static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride);
963 static bool assertion_predicate_has_loop_opaque_node(IfNode* iff);
964 static void get_assertion_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false);
965 void update_main_loop_assertion_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
966 void copy_assertion_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head,
967 Node* stride);
968 void initialize_assertion_predicates_for_peeled_loop(const PredicateBlock* predicate_block, LoopNode* outer_loop_head,
969 int dd_outer_loop_head, Node* init, Node* stride,
970 IdealLoopTree* outer_loop, uint idx_before_clone,
971 const Node_List& old_new);
972 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit,
973 Node* bol);
974 #ifdef ASSERT
975 bool only_has_infinite_loops();
976 #endif
977
978 void log_loop_tree();
979
980 public:
981
982 PhaseIterGVN &igvn() const { return _igvn; }
983
984 bool has_node(const Node* n) const {
985 guarantee(n != nullptr, "No Node.");
986 return _loop_or_ctrl[n->_idx] != nullptr;
987 }
988 // check if transform created new nodes that need _ctrl recorded
989 Node *get_late_ctrl( Node *n, Node *early );
990 Node *get_early_ctrl( Node *n );
991 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
992 void set_early_ctrl(Node* n, bool update_body);
993 void set_subtree_ctrl(Node* n, bool update_body);
994 void set_ctrl( Node *n, Node *ctrl ) {
995 assert( !has_node(n) || has_ctrl(n), "" );
996 assert( ctrl->in(0), "cannot set dead control node" );
997 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
998 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1));
999 }
1000 // Set control and update loop membership
1001 void set_ctrl_and_loop(Node* n, Node* ctrl) {
1002 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1003 IdealLoopTree* new_loop = get_loop(ctrl);
1004 if (old_loop != new_loop) {
1005 if (old_loop->_child == nullptr) old_loop->_body.yank(n);
1006 if (new_loop->_child == nullptr) new_loop->_body.push(n);
1007 }
1008 set_ctrl(n, ctrl);
1009 }
1010 // Control nodes can be replaced or subsumed. During this pass they
1011 // get their replacement Node in slot 1. Instead of updating the block
1012 // location of all Nodes in the subsumed block, we lazily do it. As we
1013 // pull such a subsumed block out of the array, we write back the final
1014 // correct block.
1015 Node* get_ctrl(const Node* i) {
1016 assert(has_node(i), "");
1017 Node *n = get_ctrl_no_update(i);
1018 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1));
1019 assert(has_node(i) && has_ctrl(i), "");
1020 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1021 return n;
1022 }
1023
1024 bool is_dominator(Node* dominator, Node* n);
1025 bool is_strict_dominator(Node* dominator, Node* n);
1026
1027 // return get_ctrl for a data node and self(n) for a CFG node
1028 Node* ctrl_or_self(Node* n) {
1029 if (has_ctrl(n))
1030 return get_ctrl(n);
1031 else {
1032 assert (n->is_CFG(), "must be a CFG node");
1033 return n;
1034 }
1035 }
1036
1037 Node* get_ctrl_no_update_helper(const Node* i) const {
1038 assert(has_ctrl(i), "should be control, not loop");
1039 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1);
1040 }
1041
1042 Node* get_ctrl_no_update(const Node* i) const {
1043 assert( has_ctrl(i), "" );
1044 Node *n = get_ctrl_no_update_helper(i);
1045 if (!n->in(0)) {
1046 // Skip dead CFG nodes
1047 do {
1048 n = get_ctrl_no_update_helper(n);
1049 } while (!n->in(0));
1050 n = find_non_split_ctrl(n);
1051 }
1052 return n;
1053 }
1054
1055 // Check for loop being set
1056 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1057 bool has_loop( Node *n ) const {
1058 assert(!has_node(n) || !has_ctrl(n), "");
1059 return has_node(n);
1060 }
1061 // Set loop
1062 void set_loop( Node *n, IdealLoopTree *loop ) {
1063 _loop_or_ctrl.map(n->_idx, (Node*)loop);
1064 }
1065 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1066 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1067 // from old_node to new_node to support the lazy update. Reference
1068 // replaces loop reference, since that is not needed for dead node.
1069 void lazy_update(Node *old_node, Node *new_node) {
1070 assert(old_node != new_node, "no cycles please");
1071 // Re-use the side array slot for this node to provide the
1072 // forwarding pointer.
1073 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1074 }
1075 void lazy_replace(Node *old_node, Node *new_node) {
1076 _igvn.replace_node(old_node, new_node);
1077 lazy_update(old_node, new_node);
1078 }
1079
1080 private:
1081
1082 // Place 'n' in some loop nest, where 'n' is a CFG node
1083 void build_loop_tree();
1084 int build_loop_tree_impl( Node *n, int pre_order );
1085 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1086 // loop tree, not the root.
1087 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1088
1089 #ifdef ASSERT
1090 // verify that regions in irreducible loops are marked is_in_irreducible_loop
1091 void verify_regions_in_irreducible_loops();
1092 bool is_in_irreducible_loop(RegionNode* region);
1093 #endif
1094
1095 // Place Data nodes in some loop nest
1096 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1097 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1098 void build_loop_late_post_work(Node* n, bool pinned);
1099 void build_loop_late_post(Node* n);
1100 void verify_strip_mined_scheduling(Node *n, Node* least);
1101
1102 // Array of immediate dominance info for each CFG node indexed by node idx
1103 private:
1104 uint _idom_size;
1105 Node **_idom; // Array of immediate dominators
1106 uint *_dom_depth; // Used for fast LCA test
1107 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1108 LoopOptsMode _mode;
1109
1110 // build the loop tree and perform any requested optimizations
1111 void build_and_optimize();
1112
1113 // Dominators for the sea of nodes
1114 void Dominators();
1115
1116 // Compute the Ideal Node to Loop mapping
1117 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1118 PhaseTransform(Ideal_Loop),
1119 _loop_or_ctrl(igvn.C->comp_arena()),
1120 _igvn(igvn),
1121 _verify_me(nullptr),
1122 _verify_only(false),
1123 _mode(mode),
1124 _nodes_required(UINT_MAX) {
1125 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1126 build_and_optimize();
1127 }
1128
1129 #ifndef PRODUCT
1130 // Verify that verify_me made the same decisions as a fresh run
1131 // or only verify that the graph is valid if verify_me is null.
1132 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1133 PhaseTransform(Ideal_Loop),
1134 _loop_or_ctrl(igvn.C->comp_arena()),
1135 _igvn(igvn),
1136 _verify_me(verify_me),
1137 _verify_only(verify_me == nullptr),
1138 _mode(LoopOptsVerify),
1139 _nodes_required(UINT_MAX) {
1140 build_and_optimize();
1141 }
1142 #endif
1143
1144 public:
1145 Node* idom_no_update(Node* d) const {
1146 return idom_no_update(d->_idx);
1147 }
1148
1149 Node* idom_no_update(uint didx) const {
1150 assert(didx < _idom_size, "oob");
1151 Node* n = _idom[didx];
1152 assert(n != nullptr,"Bad immediate dominator info.");
1153 while (n->in(0) == nullptr) { // Skip dead CFG nodes
1154 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1);
1155 assert(n != nullptr,"Bad immediate dominator info.");
1156 }
1157 return n;
1158 }
1159
1160 Node *idom(Node* d) const {
1161 return idom(d->_idx);
1162 }
1163
1164 Node *idom(uint didx) const {
1165 Node *n = idom_no_update(didx);
1166 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1167 return n;
1168 }
1169
1170 uint dom_depth(Node* d) const {
1171 guarantee(d != nullptr, "Null dominator info.");
1172 guarantee(d->_idx < _idom_size, "");
1173 return _dom_depth[d->_idx];
1174 }
1175 void set_idom(Node* d, Node* n, uint dom_depth);
1176 // Locally compute IDOM using dom_lca call
1177 Node *compute_idom( Node *region ) const;
1178 // Recompute dom_depth
1179 void recompute_dom_depth();
1180
1181 // Is safept not required by an outer loop?
1182 bool is_deleteable_safept(Node* sfpt);
1183
1184 // Replace parallel induction variable (parallel to trip counter)
1185 void replace_parallel_iv(IdealLoopTree *loop);
1186
1187 Node *dom_lca( Node *n1, Node *n2 ) const {
1188 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1189 }
1190 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1191
1192 // Build and verify the loop tree without modifying the graph. This
1193 // is useful to verify that all inputs properly dominate their uses.
1194 static void verify(PhaseIterGVN& igvn) {
1195 #ifdef ASSERT
1196 ResourceMark rm;
1197 Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
1198 PhaseIdealLoop v(igvn);
1199 #endif
1200 }
1201
1202 // Recommended way to use PhaseIdealLoop.
1203 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1204 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1205 ResourceMark rm;
1206 PhaseIdealLoop v(igvn, mode);
1207
1208 Compile* C = Compile::current();
1209 if (!C->failing()) {
1210 // Cleanup any modified bits
1211 igvn.optimize();
1212 if (C->failing()) { return; }
1213 v.log_loop_tree();
1214 }
1215 }
1216
1217 // True if the method has at least 1 irreducible loop
1218 bool _has_irreducible_loops;
1219
1220 // Per-Node transform
1221 virtual Node* transform(Node* n) { return nullptr; }
1222
1223 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1224 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1225 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1226 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1227 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1228
1229 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1230
1231 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1232 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1233 #ifdef ASSERT
1234 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1235 #endif
1236 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1237 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1238 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1239 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1240 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1241 Node*& entry_control, Node*& iffalse);
1242
1243 Node* exact_limit( IdealLoopTree *loop );
1244
1245 // Return a post-walked LoopNode
1246 IdealLoopTree *get_loop( Node *n ) const {
1247 // Dead nodes have no loop, so return the top level loop instead
1248 if (!has_node(n)) return _ltree_root;
1249 assert(!has_ctrl(n), "");
1250 return (IdealLoopTree*)_loop_or_ctrl[n->_idx];
1251 }
1252
1253 IdealLoopTree* ltree_root() const { return _ltree_root; }
1254
1255 // Is 'n' a (nested) member of 'loop'?
1256 int is_member( const IdealLoopTree *loop, Node *n ) const {
1257 return loop->is_member(get_loop(n)); }
1258
1259 // This is the basic building block of the loop optimizations. It clones an
1260 // entire loop body. It makes an old_new loop body mapping; with this
1261 // mapping you can find the new-loop equivalent to an old-loop node. All
1262 // new-loop nodes are exactly equal to their old-loop counterparts, all
1263 // edges are the same. All exits from the old-loop now have a RegionNode
1264 // that merges the equivalent new-loop path. This is true even for the
1265 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1266 // now come from (one or more) Phis that merge their new-loop equivalents.
1267 // Parameter side_by_side_idom:
1268 // When side_by_size_idom is null, the dominator tree is constructed for
1269 // the clone loop to dominate the original. Used in construction of
1270 // pre-main-post loop sequence.
1271 // When nonnull, the clone and original are side-by-side, both are
1272 // dominated by the passed in side_by_side_idom node. Used in
1273 // construction of unswitched loops.
1274 enum CloneLoopMode {
1275 IgnoreStripMined = 0, // Only clone inner strip mined loop
1276 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1277 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1278 // result control flow branches
1279 // either to inner clone or outer
1280 // strip mined loop.
1281 };
1282 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1283 CloneLoopMode mode, Node* side_by_side_idom = nullptr);
1284 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1285 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1286 Node_List*& split_if_set, Node_List*& split_bool_set,
1287 Node_List*& split_cex_set, Node_List& worklist,
1288 uint new_counter, CloneLoopMode mode);
1289 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1290 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1291 Node_List& extra_data_nodes);
1292
1293 // If we got the effect of peeling, either by actually peeling or by
1294 // making a pre-loop which must execute at least once, we can remove
1295 // all loop-invariant dominated tests in the main body.
1296 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1297
1298 // Generate code to do a loop peel for the given loop (and body).
1299 // old_new is a temp array.
1300 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1301
1302 // Add pre and post loops around the given loop. These loops are used
1303 // during RCE, unrolling and aligning loops.
1304 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1305
1306 // Add post loop after the given loop.
1307 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1308 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1309 Node*& incr, Node* limit, CountedLoopNode*& post_head);
1310
1311 // Add a vector post loop between a vector main loop and the current post loop
1312 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1313 // If Node n lives in the back_ctrl block, we clone a private version of n
1314 // in preheader_ctrl block and return that, otherwise return n.
1315 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1316
1317 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1318 // unroll to do double iterations. The next round of major loop transforms
1319 // will repeat till the doubled loop body does all remaining iterations in 1
1320 // pass.
1321 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1322
1323 // Unroll the loop body one step - make each trip do 2 iterations.
1324 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1325
1326 // Return true if exp is a constant times an induction var
1327 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1328
1329 bool is_iv(Node* exp, Node* iv, BasicType bt);
1330
1331 // Return true if exp is a scaled induction var plus (or minus) constant
1332 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);
1333 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1334 jlong long_scale;
1335 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1336 int int_scale = checked_cast<int>(long_scale);
1337 if (p_scale != nullptr) {
1338 *p_scale = int_scale;
1339 }
1340 return true;
1341 }
1342 return false;
1343 }
1344 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1345 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1346 BasicType bt,
1347 jlong* p_scale, Node** p_offset,
1348 bool* p_short_scale, int depth);
1349
1350 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1351 IfTrueNode* create_new_if_for_predicate(
1352 ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason, int opcode,
1353 bool rewire_uncommon_proj_phi_inputs = false
1354 NOT_PRODUCT (COMMA AssertionPredicateType assertion_predicate_type = AssertionPredicateType::None));
1355
1356 private:
1357 // Helper functions for create_new_if_for_predicate()
1358 void set_ctrl_of_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1359 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl);
1360 Node* clone_nodes_with_same_ctrl(Node* start_node, ProjNode* old_uncommon_proj, Node* new_uncommon_proj);
1361 void fix_cloned_data_node_controls(const ProjNode* orig, Node* new_uncommon_proj,
1362 const OrigToNewHashtable& orig_to_clone);
1363 bool has_dominating_loop_limit_check(Node* init_trip, Node* limit, jlong stride_con, BasicType iv_bt,
1364 Node* loop_entry);
1365
1366 public:
1367 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1368
1369 void replace_loop_entry(LoopNode* loop_head, Node* new_entry) {
1370 _igvn.replace_input_of(loop_head, LoopNode::EntryControl, new_entry);
1371 set_idom(loop_head, new_entry, dom_depth(new_entry));
1372 }
1373
1374 // Construct a range check for a predicate if
1375 BoolNode* rc_predicate(Node* ctrl, int scale, Node* offset, Node* init, Node* limit,
1376 jint stride, Node* range, bool upper, bool& overflow);
1377
1378 // Implementation of the loop predication to promote checks outside the loop
1379 bool loop_predication_impl(IdealLoopTree *loop);
1380
1381 private:
1382 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj,
1383 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero,
1384 Invariance& invar, Deoptimization::DeoptReason reason);
1385 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const;
1386 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt);
1387 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1388 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1389 Node_List& if_proj_list);
1390 IfTrueNode* add_template_assertion_predicate(IfNode* iff, IdealLoopTree* loop, IfProjNode* if_proj,
1391 ParsePredicateSuccessProj* parse_predicate_proj,
1392 IfProjNode* upper_bound_proj, int scale, Node* offset, Node* init, Node* limit,
1393 jint stride, Node* rng, bool& overflow, Deoptimization::DeoptReason reason);
1394 void eliminate_hoisted_range_check(IfTrueNode* hoisted_check_proj, IfTrueNode* template_assertion_predicate_proj);
1395 Node* add_range_check_elimination_assertion_predicate(
1396 IdealLoopTree* loop, Node* predicate_proj, int scale_con, Node* offset, Node* limit, int stride_con, Node* value,
1397 bool is_template NOT_PRODUCT(COMMA AssertionPredicateType assertion_predicate_type = AssertionPredicateType::None));
1398
1399 // Helper function to collect predicate for eliminating the useless ones
1400 void eliminate_useless_predicates();
1401
1402 void eliminate_useless_parse_predicates();
1403 void mark_all_parse_predicates_useless() const;
1404 void mark_loop_associated_parse_predicates_useful();
1405 static void mark_useful_parse_predicates_for_loop(IdealLoopTree* loop);
1406 void add_useless_parse_predicates_to_igvn_worklist();
1407
1408 void eliminate_useless_template_assertion_predicates();
1409 void collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates);
1410 static void collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop, Unique_Node_List& useful_predicates);
1411 void eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates);
1412
1413 void eliminate_useless_zero_trip_guard();
1414
1415 public:
1416 // Change the control input of expensive nodes to allow commoning by
1417 // IGVN when it is guaranteed to not result in a more frequent
1418 // execution of the expensive node. Return true if progress.
1419 bool process_expensive_nodes();
1420
1421 // Check whether node has become unreachable
1422 bool is_node_unreachable(Node *n) const {
1423 return !has_node(n) || n->is_unreachable(_igvn);
1424 }
1425
1426 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1427 void do_range_check(IdealLoopTree *loop, Node_List &old_new);
1428
1429 // Clone loop with an invariant test (that does not exit) and
1430 // insert a clone of the test that selects which version to
1431 // execute.
1432 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1433
1434 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1435 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1436
1437 private:
1438 static bool has_control_dependencies_from_predicates(LoopNode* head);
1439 static void revert_to_normal_loop(const LoopNode* loop_head);
1440
1441 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1442 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1443 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1444 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1445 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1446 #ifndef PRODUCT
1447 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1448 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1449 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1450 const UnswitchCandidate& unswitch_candidate,
1451 const LoopNode* original_head, const LoopNode* new_head);
1452 #endif
1453
1454 public:
1455
1456 // Range Check Elimination uses this function!
1457 // Constrain the main loop iterations so the affine function:
1458 // low_limit <= scale_con * I + offset < upper_limit
1459 // always holds true. That is, either increase the number of iterations in
1460 // the pre-loop or the post-loop until the condition holds true in the main
1461 // loop. Scale_con, offset and limit are all loop invariant.
1462 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);
1463 // Helper function for add_constraint().
1464 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1465
1466 // Partially peel loop up through last_peel node.
1467 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1468 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1469
1470 // AutoVectorize the loop: replace scalar ops with vector ops.
1471 enum AutoVectorizeStatus {
1472 Impossible, // This loop has the wrong shape to even try vectorization.
1473 Success, // We just successfully vectorized the loop.
1474 TriedAndFailed, // We tried to vectorize, but failed.
1475 };
1476 AutoVectorizeStatus auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared);
1477
1478 // Move an unordered Reduction out of loop if possible
1479 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop);
1480
1481 // Create a scheduled list of nodes control dependent on ctrl set.
1482 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1483 // Has a use in the vector set
1484 bool has_use_in_set( Node* n, VectorSet& vset );
1485 // Has use internal to the vector set (ie. not in a phi at the loop head)
1486 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1487 // clone "n" for uses that are outside of loop
1488 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1489 // clone "n" for special uses that are in the not_peeled region
1490 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1491 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1492 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1493 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1494 #ifdef ASSERT
1495 // Validate the loop partition sets: peel and not_peel
1496 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1497 // Ensure that uses outside of loop are of the right form
1498 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1499 uint orig_exit_idx, uint clone_exit_idx);
1500 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1501 #endif
1502
1503 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1504 int stride_of_possible_iv( Node* iff );
1505 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1506 // Return the (unique) control output node that's in the loop (if it exists.)
1507 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1508 // Insert a signed compare loop exit cloned from an unsigned compare.
1509 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1510 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1511 // Utility to register node "n" with PhaseIdealLoop
1512 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth);
1513 // Utility to create an if-projection
1514 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1515 // Force the iff control output to be the live_proj
1516 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1517 // Insert a region before an if projection
1518 RegionNode* insert_region_before_proj(ProjNode* proj);
1519 // Insert a new if before an if projection
1520 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1521
1522 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1523 // "Nearly" because all Nodes have been cloned from the original in the loop,
1524 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1525 // through the Phi recursively, and return a Bool.
1526 Node* clone_iff(PhiNode* phi);
1527 CmpNode* clone_bool(PhiNode* phi);
1528
1529
1530 // Rework addressing expressions to get the most loop-invariant stuff
1531 // moved out. We'd like to do all associative operators, but it's especially
1532 // important (common) to do address expressions.
1533 Node* remix_address_expressions(Node* n);
1534 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1535
1536 // Convert add to muladd to generate MuladdS2I under certain criteria
1537 Node * convert_add_to_muladd(Node * n);
1538
1539 // Attempt to use a conditional move instead of a phi/branch
1540 Node *conditional_move( Node *n );
1541
1542 bool split_thru_phi_could_prevent_vectorization(Node* n, Node* n_blk);
1543
1544 // Check for aggressive application of 'split-if' optimization,
1545 // using basic block level info.
1546 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1547 Node *split_if_with_blocks_pre ( Node *n );
1548 void split_if_with_blocks_post( Node *n );
1549 Node *has_local_phi_input( Node *n );
1550 // Mark an IfNode as being dominated by a prior test,
1551 // without actually altering the CFG (and hence IDOM info).
1552 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool pin_array_access_nodes = false);
1553 void rewire_safe_outputs_to_dominator(Node* source, Node* dominator, bool pin_array_access_nodes);
1554
1555 // Split Node 'n' through merge point
1556 RegionNode* split_thru_region(Node* n, RegionNode* region);
1557 // Split Node 'n' through merge point if there is enough win.
1558 Node *split_thru_phi( Node *n, Node *region, int policy );
1559 // Found an If getting its condition-code input from a Phi in the
1560 // same block. Split thru the Region.
1561 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr);
1562
1563 // Conversion of fill/copy patterns into intrinsic versions
1564 bool do_intrinsify_fill();
1565 bool intrinsify_fill(IdealLoopTree* lpt);
1566 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1567 Node*& shift, Node*& offset);
1568
1569 private:
1570 // Return a type based on condition control flow
1571 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1572 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1573 // Helpers for filtered type
1574 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1575
1576 // Helper functions
1577 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1578 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1579 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 );
1580 bool split_up( Node *n, Node *blk1, Node *blk2 );
1581
1582 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1583 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1584 void try_move_store_after_loop(Node* n);
1585 void move_flat_array_check_out_of_loop(Node* n);
1586 bool identical_backtoback_ifs(Node *n);
1587 bool flat_array_element_type_check(Node *n);
1588 bool can_split_if(Node *n_ctrl);
1589 bool cannot_split_division(const Node* n, const Node* region) const;
1590 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1591 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1592
1593 // Determine if a method is too big for a/another round of split-if, based on
1594 // a magic (approximate) ratio derived from the equally magic constant 35000,
1595 // previously used for this purpose (but without relating to the node limit).
1596 bool must_throttle_split_if() {
1597 uint threshold = C->max_node_limit() * 2 / 5;
1598 return C->live_nodes() > threshold;
1599 }
1600
1601 // A simplistic node request tracking mechanism, where
1602 // = UINT_MAX Request not valid or made final.
1603 // < UINT_MAX Nodes currently requested (estimate).
1604 uint _nodes_required;
1605
1606 enum { REQUIRE_MIN = 70 };
1607
1608 uint nodes_required() const { return _nodes_required; }
1609
1610 // Given the _currently_ available number of nodes, check whether there is
1611 // "room" for an additional request or not, considering the already required
1612 // number of nodes. Return TRUE if the new request is exceeding the node
1613 // budget limit, otherwise return FALSE. Note that this interpretation will
1614 // act pessimistic on additional requests when new nodes have already been
1615 // generated since the 'begin'. This behaviour fits with the intention that
1616 // node estimates/requests should be made upfront.
1617 bool exceeding_node_budget(uint required = 0) {
1618 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1619 uint available = C->max_node_limit() - C->live_nodes();
1620 return available < required + _nodes_required + REQUIRE_MIN;
1621 }
1622
1623 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1624 precond(require > 0);
1625 _nodes_required += MAX2(require, minreq);
1626 return _nodes_required;
1627 }
1628
1629 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1630 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1631 }
1632
1633 uint require_nodes_begin() {
1634 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1635 _nodes_required = 0;
1636 return C->live_nodes();
1637 }
1638
1639 // When a node request is final, optionally check that the requested number
1640 // of nodes was reasonably correct with respect to the number of new nodes
1641 // introduced since the last 'begin'. Always check that we have not exceeded
1642 // the maximum node limit.
1643 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1644 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1645
1646 #ifdef ASSERT
1647 if (check_estimate) {
1648 // Check that the node budget request was not off by too much (x2).
1649 // Should this be the case we _surely_ need to improve the estimates
1650 // used in our budget calculations.
1651 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1652 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1653 C->live_nodes() - live_at_begin, _nodes_required);
1654 }
1655 }
1656 #endif
1657 // Assert that we have stayed within the node budget limit.
1658 assert(C->live_nodes() < C->max_node_limit(),
1659 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1660 C->live_nodes() - live_at_begin, live_at_begin,
1661 C->max_node_limit(), _nodes_required);
1662
1663 _nodes_required = UINT_MAX;
1664 }
1665
1666 public:
1667 // Clone Parse Predicates to slow and fast loop when unswitching a loop
1668 void clone_parse_and_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new,
1669 IfProjNode*& iffast_pred, IfProjNode*& ifslow_pred);
1670 private:
1671 void clone_loop_predication_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new,
1672 const PredicateBlock* predicate_block,
1673 Deoptimization::DeoptReason reason, IfProjNode*& iffast_pred,
1674 IfProjNode*& ifslow_pred);
1675 void clone_parse_predicate_to_unswitched_loops(const PredicateBlock* predicate_block, Deoptimization::DeoptReason reason,
1676 IfProjNode*& iffast_pred, IfProjNode*& ifslow_pred);
1677 IfProjNode* clone_parse_predicate_to_unswitched_loop(ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry,
1678 Deoptimization::DeoptReason reason, bool slow_loop);
1679 void clone_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new,
1680 Deoptimization::DeoptReason reason, IfProjNode* old_predicate_proj,
1681 ParsePredicateSuccessProj* fast_loop_parse_predicate_proj,
1682 ParsePredicateSuccessProj* slow_loop_parse_predicate_proj);
1683 IfProjNode* clone_assertion_predicate_for_unswitched_loops(IfNode* template_assertion_predicate, IfProjNode* predicate,
1684 Deoptimization::DeoptReason reason,
1685 ParsePredicateSuccessProj* parse_predicate_proj);
1686 static void check_cloned_parse_predicate_for_unswitching(const Node* new_entry, bool is_fast_loop) PRODUCT_RETURN;
1687
1688 bool _created_loop_node;
1689 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);)
1690 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;)
1691
1692 public:
1693 void set_created_loop_node() { _created_loop_node = true; }
1694 bool created_loop_node() { return _created_loop_node; }
1695 void register_new_node(Node* n, Node* blk);
1696 void register_new_node_with_ctrl_of(Node* new_node, Node* ctrl_of) {
1697 register_new_node(new_node, get_ctrl(ctrl_of));
1698 }
1699
1700 Node* clone_and_register(Node* n, Node* ctrl) {
1701 n = n->clone();
1702 register_new_node(n, ctrl);
1703 return n;
1704 }
1705
1706 #ifdef ASSERT
1707 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1708 #endif
1709
1710 #ifndef PRODUCT
1711 void dump() const;
1712 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging
1713 void dump_idom(Node* n, uint count) const;
1714 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const;
1715 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1716 IdealLoopTree* get_loop_idx(Node* n) const {
1717 // Dead nodes have no loop, so return the top level loop instead
1718 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root;
1719 }
1720 // Print some stats
1721 static void print_statistics();
1722 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1723 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1724 static volatile int _long_loop_candidates;
1725 static volatile int _long_loop_nests;
1726 static volatile int _long_loop_counted_loops;
1727 #endif
1728
1729 #ifdef ASSERT
1730 void verify() const;
1731 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const;
1732 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const;
1733 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const;
1734 #endif
1735
1736 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1737
1738 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1739
1740 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1741
1742
1743 int extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1744 Node_List &range_checks);
1745
1746 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1747 Node* inner_iters_actual_int, Node* inner_phi,
1748 Node* iv_add, LoopNode* inner_head);
1749
1750 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1751
1752 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1753
1754 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1755
1756 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1757
1758 void try_sink_out_of_loop(Node* n);
1759
1760 Node* clamp(Node* R, Node* L, Node* H);
1761
1762 bool safe_for_if_replacement(const Node* dom) const;
1763
1764 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1765
1766 bool try_merge_identical_ifs(Node* n);
1767
1768 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1769
1770 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1771 IdealLoopTree* parent, bool partial);
1772
1773 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1774 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1775
1776 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1777 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1778 Node_List*& split_bool_set, Node_List*& split_cex_set);
1779
1780 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1781
1782 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1783
1784 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1785 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1786 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1787 void clone_template_assertion_expression_down(Node* node);
1788
1789 Node* similar_subtype_check(const Node* x, Node* r_in);
1790
1791 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1792
1793 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1794
1795 void pin_array_access_nodes_dependent_on(Node* ctrl);
1796
1797 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1798 };
1799
1800
1801 class AutoNodeBudget : public StackObj
1802 {
1803 public:
1804 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1805
1806 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1807 : _phase(phase),
1808 _check_at_final(chk == BUDGET_CHECK),
1809 _nodes_at_begin(0)
1810 {
1811 precond(_phase != nullptr);
1812
1813 _nodes_at_begin = _phase->require_nodes_begin();
1814 }
1815
1816 ~AutoNodeBudget() {
1817 #ifndef PRODUCT
1818 if (TraceLoopOpts) {
1819 uint request = _phase->nodes_required();
1820 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1821
1822 if (request < delta) {
1823 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1824 } else {
1825 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1826 // Identify the worst estimates as "poor" ones.
1827 if (request > REQUIRE_MIN && delta > 0) {
1828 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1829 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1830 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1831 }
1832 }
1833 }
1834 }
1835 #endif // PRODUCT
1836 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1837 }
1838
1839 private:
1840 PhaseIdealLoop* _phase;
1841 bool _check_at_final;
1842 uint _nodes_at_begin;
1843 };
1844
1845 inline Node* IdealLoopTree::tail() {
1846 // Handle lazy update of _tail field.
1847 if (_tail->in(0) == nullptr) {
1848 _tail = _phase->get_ctrl(_tail);
1849 }
1850 return _tail;
1851 }
1852
1853 inline Node* IdealLoopTree::head() {
1854 // Handle lazy update of _head field.
1855 if (_head->in(0) == nullptr) {
1856 _head = _phase->get_ctrl(_head);
1857 }
1858 return _head;
1859 }
1860
1861 // Iterate over the loop tree using a preorder, left-to-right traversal.
1862 //
1863 // Example that visits all counted loops from within PhaseIdealLoop
1864 //
1865 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1866 // IdealLoopTree* lpt = iter.current();
1867 // if (!lpt->is_counted()) continue;
1868 // ...
1869 class LoopTreeIterator : public StackObj {
1870 private:
1871 IdealLoopTree* _root;
1872 IdealLoopTree* _curnt;
1873
1874 public:
1875 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1876
1877 bool done() { return _curnt == nullptr; } // Finished iterating?
1878
1879 void next(); // Advance to next loop tree
1880
1881 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1882 };
1883
1884 // Compute probability of reaching some CFG node from a fixed
1885 // dominating CFG node
1886 class PathFrequency {
1887 private:
1888 Node* _dom; // frequencies are computed relative to this node
1889 Node_Stack _stack;
1890 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1891 GrowableArray<float> _freqs; // cache frequencies
1892 PhaseIdealLoop* _phase;
1893
1894 float check_and_truncate_frequency(float f) {
1895 assert(f >= 0, "Incorrect frequency");
1896 // We do not perform an exact (f <= 1) check
1897 // this would be error prone with rounding of floats.
1898 // Performing a check like (f <= 1+eps) would be of benefit,
1899 // however, it is not evident how to determine such an eps,
1900 // given that an arbitrary number of add/mul operations
1901 // are performed on these frequencies.
1902 return (f > 1) ? 1 : f;
1903 }
1904
1905 public:
1906 PathFrequency(Node* dom, PhaseIdealLoop* phase)
1907 : _dom(dom), _stack(0), _phase(phase) {
1908 }
1909
1910 float to(Node* n);
1911 };
1912
1913 // Class to clone a data node graph by taking a list of data nodes. This is done in 2 steps:
1914 // 1. Clone the data nodes
1915 // 2. Fix the cloned data inputs pointing to the old nodes to the cloned inputs by using an old->new mapping.
1916 class DataNodeGraph : public StackObj {
1917 PhaseIdealLoop* const _phase;
1918 const Unique_Node_List& _data_nodes;
1919 OrigToNewHashtable _orig_to_new;
1920
1921 public:
1922 DataNodeGraph(const Unique_Node_List& data_nodes, PhaseIdealLoop* phase)
1923 : _phase(phase),
1924 _data_nodes(data_nodes),
1925 // Use 107 as best guess which is the first resize value in ResizeableResourceHashtable::large_table_sizes.
1926 _orig_to_new(107, MaxNodeLimit)
1927 {
1928 #ifdef ASSERT
1929 for (uint i = 0; i < data_nodes.size(); i++) {
1930 assert(!data_nodes[i]->is_CFG(), "only data nodes");
1931 }
1932 #endif
1933 }
1934 NONCOPYABLE(DataNodeGraph);
1935
1936 private:
1937 void clone(Node* node, Node* new_ctrl);
1938 void clone_data_nodes(Node* new_ctrl);
1939 void clone_data_nodes_and_transform_opaque_loop_nodes(const TransformStrategyForOpaqueLoopNodes& transform_strategy,
1940 Node* new_ctrl);
1941 void rewire_clones_to_cloned_inputs();
1942 void transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node);
1943
1944 public:
1945 // Clone the provided data node collection and rewire the clones in such a way to create an identical graph copy.
1946 // Set 'new_ctrl' as ctrl for the cloned nodes.
1947 const OrigToNewHashtable& clone(Node* new_ctrl) {
1948 assert(_orig_to_new.number_of_entries() == 0, "should not call this method twice in a row");
1949 clone_data_nodes(new_ctrl);
1950 rewire_clones_to_cloned_inputs();
1951 return _orig_to_new;
1952 }
1953
1954 // Create a copy of the data nodes provided to the constructor by doing the following:
1955 // Clone all non-OpaqueLoop* nodes and rewire them to create an identical subgraph copy. For the OpaqueLoop* nodes,
1956 // apply the provided transformation strategy and include the transformed node into the subgraph copy to get a complete
1957 // "cloned-and-transformed" graph copy. For all newly cloned nodes (which could also be new OpaqueLoop* nodes), set
1958 // `new_ctrl` as ctrl.
1959 const OrigToNewHashtable& clone_with_opaque_loop_transform_strategy(
1960 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
1961 Node* new_ctrl) {
1962 clone_data_nodes_and_transform_opaque_loop_nodes(transform_strategy, new_ctrl);
1963 rewire_clones_to_cloned_inputs();
1964 return _orig_to_new;
1965 }
1966 };
1967 #endif // SHARE_OPTO_LOOPNODE_HPP