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