1 /*
2 * Copyright (c) 1998, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_OPTO_LOOPNODE_HPP
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/subnode.hpp"
32 #include "opto/type.hpp"
33
34 class CmpNode;
35 class BaseCountedLoopEndNode;
36 class CountedLoopNode;
37 class IdealLoopTree;
38 class LoopNode;
39 class Node;
40 class OuterStripMinedLoopEndNode;
41 class PathFrequency;
42 class PhaseIdealLoop;
43 class CountedLoopReserveKit;
44 class VectorSet;
45 class Invariance;
46 struct small_cache;
47
48 //
49 // I D E A L I Z E D L O O P S
50 //
51 // Idealized loops are the set of loops I perform more interesting
52 // transformations on, beyond simple hoisting.
53
54 //------------------------------LoopNode---------------------------------------
55 // Simple loop header. Fall in path on left, loop-back path on right.
56 class LoopNode : public RegionNode {
57 // Size is bigger to hold the flags. However, the flags do not change
58 // the semantics so it does not appear in the hash & cmp functions.
59 virtual uint size_of() const { return sizeof(*this); }
60 protected:
61 uint _loop_flags;
62 // Names for flag bitfields
63 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
64 MainHasNoPreLoop = 1<<2,
65 HasExactTripCount = 1<<3,
66 InnerLoop = 1<<4,
67 PartialPeelLoop = 1<<5,
68 PartialPeelFailed = 1<<6,
69 HasReductions = 1<<7,
70 WasSlpAnalyzed = 1<<8,
71 PassedSlpAnalysis = 1<<9,
72 DoUnrollOnly = 1<<10,
73 VectorizedLoop = 1<<11,
74 HasAtomicPostLoop = 1<<12,
75 HasRangeChecks = 1<<13,
76 IsMultiversioned = 1<<14,
77 StripMined = 1<<15,
78 SubwordLoop = 1<<16,
79 ProfileTripFailed = 1<<17,
80 LoopNestInnerLoop = 1 << 18,
81 LoopNestLongOuterLoop = 1 << 19};
82 char _unswitch_count;
83 enum { _unswitch_max=3 };
84 char _postloop_flags;
85 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
86
87 // Expected trip count from profile data
88 float _profile_trip_cnt;
89
90 public:
91 // Names for edge indices
92 enum { Self=0, EntryControl, LoopBackControl };
93
94 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
95 void set_inner_loop() { _loop_flags |= InnerLoop; }
96
97 bool range_checks_present() const { return _loop_flags & HasRangeChecks; }
98 bool is_multiversioned() const { return _loop_flags & IsMultiversioned; }
99 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
100 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
101 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
102 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
103 bool is_strip_mined() const { return _loop_flags & StripMined; }
104 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
105 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
106 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
107 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
108
109 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
110 void mark_has_reductions() { _loop_flags |= HasReductions; }
111 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
112 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
113 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
114 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
115 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
116 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
117 void clear_has_range_checks() { _loop_flags &= ~HasRangeChecks; }
118 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
119 void mark_strip_mined() { _loop_flags |= StripMined; }
120 void clear_strip_mined() { _loop_flags &= ~StripMined; }
121 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
122 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
123 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
124 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
125
126 int unswitch_max() { return _unswitch_max; }
127 int unswitch_count() { return _unswitch_count; }
128
129 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
130 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
131 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
132 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
133
134 void set_unswitch_count(int val) {
135 assert (val <= unswitch_max(), "too many unswitches");
136 _unswitch_count = val;
137 }
138
139 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
140 float profile_trip_cnt() { return _profile_trip_cnt; }
141
142 LoopNode(Node *entry, Node *backedge)
143 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
144 _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) {
145 init_class_id(Class_Loop);
146 init_req(EntryControl, entry);
147 init_req(LoopBackControl, backedge);
148 }
149
150 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
151 virtual int Opcode() const;
152 bool can_be_counted_loop(PhaseTransform* phase) const {
153 return req() == 3 && in(0) != NULL &&
154 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
155 in(2) != NULL && phase->type(in(2)) != Type::TOP;
156 }
157 bool is_valid_counted_loop(BasicType bt) const;
158 #ifndef PRODUCT
159 virtual void dump_spec(outputStream *st) const;
160 #endif
161
162 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
163 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
164 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
165 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
166 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
167 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; }
168 };
169
170 //------------------------------Counted Loops----------------------------------
171 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
172 // path (and maybe some other exit paths). The trip-counter exit is always
173 // last in the loop. The trip-counter have to stride by a constant;
174 // the exit value is also loop invariant.
175
176 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
177 // CountedLoopNode has the incoming loop control and the loop-back-control
178 // which is always the IfTrue before the matching CountedLoopEndNode. The
179 // CountedLoopEndNode has an incoming control (possibly not the
180 // CountedLoopNode if there is control flow in the loop), the post-increment
181 // trip-counter value, and the limit. The trip-counter value is always of
182 // the form (Op old-trip-counter stride). The old-trip-counter is produced
183 // by a Phi connected to the CountedLoopNode. The stride is constant.
184 // The Op is any commutable opcode, including Add, Mul, Xor. The
185 // CountedLoopEndNode also takes in the loop-invariant limit value.
186
187 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
188 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
189 // via the old-trip-counter from the Op node.
190
191 //------------------------------CountedLoopNode--------------------------------
192 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
193 // inputs the incoming loop-start control and the loop-back control, so they
194 // act like RegionNodes. They also take in the initial trip counter, the
195 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
196 // produce a loop-body control and the trip counter value. Since
197 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
198
199 class BaseCountedLoopNode : public LoopNode {
200 public:
201 BaseCountedLoopNode(Node *entry, Node *backedge)
202 : LoopNode(entry, backedge) {
203 }
204
205 Node *init_control() const { return in(EntryControl); }
206 Node *back_control() const { return in(LoopBackControl); }
207
208 Node* init_trip() const;
209 Node* stride() const;
210 bool stride_is_con() const;
211 Node* limit() const;
212 Node* incr() const;
213 Node* phi() const;
214
215 BaseCountedLoopEndNode* loopexit_or_null() const;
216 BaseCountedLoopEndNode* loopexit() const;
217
218 virtual BasicType bt() const = 0;
219
220 jlong stride_con() const;
221
222 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt);
223 };
224
225
226 class CountedLoopNode : public BaseCountedLoopNode {
227 // Size is bigger to hold _main_idx. However, _main_idx does not change
228 // the semantics so it does not appear in the hash & cmp functions.
229 virtual uint size_of() const { return sizeof(*this); }
230
231 // For Pre- and Post-loops during debugging ONLY, this holds the index of
232 // the Main CountedLoop. Used to assert that we understand the graph shape.
233 node_idx_t _main_idx;
234
235 // Known trip count calculated by compute_exact_trip_count()
236 uint _trip_count;
237
238 // Log2 of original loop bodies in unrolled loop
239 int _unrolled_count_log2;
240
241 // Node count prior to last unrolling - used to decide if
242 // unroll,optimize,unroll,optimize,... is making progress
243 int _node_count_before_unroll;
244
245 // If slp analysis is performed we record the maximum
246 // vector mapped unroll factor here
247 int _slp_maximum_unroll_factor;
248
249 // The eventual count of vectorizable packs in slp
250 int _slp_vector_pack_count;
251
252 public:
253 CountedLoopNode(Node *entry, Node *backedge)
254 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
255 _unrolled_count_log2(0), _node_count_before_unroll(0),
256 _slp_maximum_unroll_factor(0), _slp_vector_pack_count(0) {
257 init_class_id(Class_CountedLoop);
258 // Initialize _trip_count to the largest possible value.
259 // Will be reset (lower) if the loop's trip count is known.
260 }
261
262 virtual int Opcode() const;
263 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
264
265 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
266 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
267 int stride_con() const;
268
269 // Match increment with optional truncation
270 static Node*
271 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
272 BasicType bt);
273
274 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
275 // can run short a few iterations and may start a few iterations in.
276 // It will be RCE'd and unrolled and aligned.
277
278 // A following 'post' loop will run any remaining iterations. Used
279 // during Range Check Elimination, the 'post' loop will do any final
280 // iterations with full checks. Also used by Loop Unrolling, where
281 // the 'post' loop will do any epilog iterations needed. Basically,
282 // a 'post' loop can not profitably be further unrolled or RCE'd.
283
284 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
285 // it may do under-flow checks for RCE and may do alignment iterations
286 // so the following main loop 'knows' that it is striding down cache
287 // lines.
288
289 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
290 // Aligned, may be missing it's pre-loop.
291 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
292 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
293 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
294 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
295 bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
296 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
297 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
298 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
299 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
300 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
301 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
302
303 int main_idx() const { return _main_idx; }
304
305
306 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
307 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
308 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
309 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
310
311 void set_trip_count(uint tc) { _trip_count = tc; }
312 uint trip_count() { return _trip_count; }
313
314 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
315 void set_exact_trip_count(uint tc) {
316 _trip_count = tc;
317 _loop_flags |= HasExactTripCount;
318 }
319 void set_nonexact_trip_count() {
320 _loop_flags &= ~HasExactTripCount;
321 }
322 void set_notpassed_slp() {
323 _loop_flags &= ~PassedSlpAnalysis;
324 }
325
326 void double_unrolled_count() { _unrolled_count_log2++; }
327 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
328
329 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
330 int node_count_before_unroll() { return _node_count_before_unroll; }
331 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
332 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
333 void set_slp_pack_count(int pack_count) { _slp_vector_pack_count = pack_count; }
334 int slp_pack_count() const { return _slp_vector_pack_count; }
335
336 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
337 OuterStripMinedLoopNode* outer_loop() const;
338 virtual IfTrueNode* outer_loop_tail() const;
339 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
340 virtual IfFalseNode* outer_loop_exit() const;
341 virtual SafePointNode* outer_safepoint() const;
342
343 // If this is a main loop in a pre/main/post loop nest, walk over
344 // the predicates that were inserted by
345 // duplicate_predicates()/add_range_check_predicate()
346 static Node* skip_predicates_from_entry(Node* ctrl);
347 Node* skip_predicates();
348
349 virtual BasicType bt() const {
350 return T_INT;
351 }
352
353 Node* is_canonical_loop_entry();
354
355 #ifndef PRODUCT
356 virtual void dump_spec(outputStream *st) const;
357 #endif
358 };
359
360 class LongCountedLoopNode : public BaseCountedLoopNode {
361 public:
362 LongCountedLoopNode(Node *entry, Node *backedge)
363 : BaseCountedLoopNode(entry, backedge) {
364 init_class_id(Class_LongCountedLoop);
365 }
366
367 virtual int Opcode() const;
368
369 virtual BasicType bt() const {
370 return T_LONG;
371 }
372
373 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); }
374 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); }
375 };
376
377
378 //------------------------------CountedLoopEndNode-----------------------------
379 // CountedLoopEndNodes end simple trip counted loops. They act much like
380 // IfNodes.
381
382 class BaseCountedLoopEndNode : public IfNode {
383 public:
384 enum { TestControl, TestValue };
385 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
386 : IfNode(control, test, prob, cnt) {
387 init_class_id(Class_BaseCountedLoopEnd);
388 }
389
390 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
391 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; }
392 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; }
393 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; }
394 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; }
395 bool stride_is_con() const { Node *tmp = stride(); return (tmp != NULL && tmp->is_Con()); }
396
397 PhiNode* phi() const {
398 Node* tmp = incr();
399 if (tmp && tmp->req() == 3) {
400 Node* phi = tmp->in(1);
401 if (phi->is_Phi()) {
402 return phi->as_Phi();
403 }
404 }
405 return NULL;
406 }
407
408 BaseCountedLoopNode* loopnode() const {
409 // The CountedLoopNode that goes with this CountedLoopEndNode may
410 // have been optimized out by the IGVN so be cautious with the
411 // pattern matching on the graph
412 PhiNode* iv_phi = phi();
413 if (iv_phi == NULL) {
414 return NULL;
415 }
416 Node* ln = iv_phi->in(0);
417 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) {
418 return NULL;
419 }
420 if (ln->as_BaseCountedLoop()->bt() != bt()) {
421 return NULL;
422 }
423 return ln->as_BaseCountedLoop();
424 }
425
426 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
427
428 jlong stride_con() const;
429 virtual BasicType bt() const = 0;
430
431 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt);
432 };
433
434 class CountedLoopEndNode : public BaseCountedLoopEndNode {
435 public:
436
437 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
438 : BaseCountedLoopEndNode(control, test, prob, cnt) {
439 init_class_id(Class_CountedLoopEnd);
440 }
441 virtual int Opcode() const;
442
443 CountedLoopNode* loopnode() const {
444 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode();
445 }
446
447 virtual BasicType bt() const {
448 return T_INT;
449 }
450
451 #ifndef PRODUCT
452 virtual void dump_spec(outputStream *st) const;
453 #endif
454 };
455
456 class LongCountedLoopEndNode : public BaseCountedLoopEndNode {
457 public:
458 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt)
459 : BaseCountedLoopEndNode(control, test, prob, cnt) {
460 init_class_id(Class_LongCountedLoopEnd);
461 }
462
463 LongCountedLoopNode* loopnode() const {
464 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode();
465 }
466
467 virtual int Opcode() const;
468
469 virtual BasicType bt() const {
470 return T_LONG;
471 }
472 };
473
474
475 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const {
476 Node* bctrl = back_control();
477 if (bctrl == NULL) return NULL;
478
479 Node* lexit = bctrl->in(0);
480 if (!lexit->is_BaseCountedLoopEnd()) {
481 return NULL;
482 }
483 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd();
484 if (result->bt() != bt()) {
485 return NULL;
486 }
487 return result;
488 }
489
490 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const {
491 BaseCountedLoopEndNode* cle = loopexit_or_null();
492 assert(cle != NULL, "loopexit is NULL");
493 return cle;
494 }
495
496 inline Node* BaseCountedLoopNode::init_trip() const {
497 BaseCountedLoopEndNode* cle = loopexit_or_null();
498 return cle != NULL ? cle->init_trip() : NULL;
499 }
500 inline Node* BaseCountedLoopNode::stride() const {
501 BaseCountedLoopEndNode* cle = loopexit_or_null();
502 return cle != NULL ? cle->stride() : NULL;
503 }
504
505 inline bool BaseCountedLoopNode::stride_is_con() const {
506 BaseCountedLoopEndNode* cle = loopexit_or_null();
507 return cle != NULL && cle->stride_is_con();
508 }
509 inline Node* BaseCountedLoopNode::limit() const {
510 BaseCountedLoopEndNode* cle = loopexit_or_null();
511 return cle != NULL ? cle->limit() : NULL;
512 }
513 inline Node* BaseCountedLoopNode::incr() const {
514 BaseCountedLoopEndNode* cle = loopexit_or_null();
515 return cle != NULL ? cle->incr() : NULL;
516 }
517 inline Node* BaseCountedLoopNode::phi() const {
518 BaseCountedLoopEndNode* cle = loopexit_or_null();
519 return cle != NULL ? cle->phi() : NULL;
520 }
521
522 inline jlong BaseCountedLoopNode::stride_con() const {
523 BaseCountedLoopEndNode* cle = loopexit_or_null();
524 return cle != NULL ? cle->stride_con() : 0;
525 }
526
527
528 //------------------------------LoopLimitNode-----------------------------
529 // Counted Loop limit node which represents exact final iterator value:
530 // trip_count = (limit - init_trip + stride - 1)/stride
531 // final_value= trip_count * stride + init_trip.
532 // Use HW instructions to calculate it when it can overflow in integer.
533 // Note, final_value should fit into integer since counted loop has
534 // limit check: limit <= max_int-stride.
535 class LoopLimitNode : public Node {
536 enum { Init=1, Limit=2, Stride=3 };
537 public:
538 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
539 // Put it on the Macro nodes list to optimize during macro nodes expansion.
540 init_flags(Flag_is_macro);
541 C->add_macro_node(this);
542 }
543 virtual int Opcode() const;
544 virtual const Type *bottom_type() const { return TypeInt::INT; }
545 virtual uint ideal_reg() const { return Op_RegI; }
546 virtual const Type* Value(PhaseGVN* phase) const;
547 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
548 virtual Node* Identity(PhaseGVN* phase);
549 };
550
551 // Support for strip mining
552 class OuterStripMinedLoopNode : public LoopNode {
553 private:
554 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
555
556 public:
557 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
558 : LoopNode(entry, backedge) {
559 init_class_id(Class_OuterStripMinedLoop);
560 init_flags(Flag_is_macro);
561 C->add_macro_node(this);
562 }
563
564 virtual int Opcode() const;
565
566 virtual IfTrueNode* outer_loop_tail() const;
567 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
568 virtual IfFalseNode* outer_loop_exit() const;
569 virtual SafePointNode* outer_safepoint() const;
570 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
571
572 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const;
573
574 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
575
576 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop);
577
578 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
579 PhaseIdealLoop* iloop);
580 };
581
582 class OuterStripMinedLoopEndNode : public IfNode {
583 public:
584 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
585 : IfNode(control, test, prob, cnt) {
586 init_class_id(Class_OuterStripMinedLoopEnd);
587 }
588
589 virtual int Opcode() const;
590
591 virtual const Type* Value(PhaseGVN* phase) const;
592 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
593
594 bool is_expanded(PhaseGVN *phase) const;
595 };
596
597 // -----------------------------IdealLoopTree----------------------------------
598 class IdealLoopTree : public ResourceObj {
599 public:
600 IdealLoopTree *_parent; // Parent in loop tree
601 IdealLoopTree *_next; // Next sibling in loop tree
602 IdealLoopTree *_child; // First child in loop tree
603
604 // The head-tail backedge defines the loop.
605 // If a loop has multiple backedges, this is addressed during cleanup where
606 // we peel off the multiple backedges, merging all edges at the bottom and
607 // ensuring that one proper backedge flow into the loop.
608 Node *_head; // Head of loop
609 Node *_tail; // Tail of loop
610 inline Node *tail(); // Handle lazy update of _tail field
611 inline Node *head(); // Handle lazy update of _head field
612 PhaseIdealLoop* _phase;
613 int _local_loop_unroll_limit;
614 int _local_loop_unroll_factor;
615
616 Node_List _body; // Loop body for inner loops
617
618 uint16_t _nest; // Nesting depth
619 uint8_t _irreducible:1, // True if irreducible
620 _has_call:1, // True if has call safepoint
621 _has_sfpt:1, // True if has non-call safepoint
622 _rce_candidate:1; // True if candidate for range check elimination
623
624 Node_List* _safepts; // List of safepoints in this loop
625 Node_List* _required_safept; // A inner loop cannot delete these safepts;
626 bool _allow_optimizations; // Allow loop optimizations
627
628 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
629 : _parent(0), _next(0), _child(0),
630 _head(head), _tail(tail),
631 _phase(phase),
632 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
633 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
634 _safepts(NULL),
635 _required_safept(NULL),
636 _allow_optimizations(true)
637 {
638 precond(_head != NULL);
639 precond(_tail != NULL);
640 }
641
642 // Is 'l' a member of 'this'?
643 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
644
645 // Set loop nesting depth. Accumulate has_call bits.
646 int set_nest( uint depth );
647
648 // Split out multiple fall-in edges from the loop header. Move them to a
649 // private RegionNode before the loop. This becomes the loop landing pad.
650 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
651
652 // Split out the outermost loop from this shared header.
653 void split_outer_loop( PhaseIdealLoop *phase );
654
655 // Merge all the backedges from the shared header into a private Region.
656 // Feed that region as the one backedge to this loop.
657 void merge_many_backedges( PhaseIdealLoop *phase );
658
659 // Split shared headers and insert loop landing pads.
660 // Insert a LoopNode to replace the RegionNode.
661 // Returns TRUE if loop tree is structurally changed.
662 bool beautify_loops( PhaseIdealLoop *phase );
663
664 // Perform optimization to use the loop predicates for null checks and range checks.
665 // Applies to any loop level (not just the innermost one)
666 bool loop_predication( PhaseIdealLoop *phase);
667
668 // Perform iteration-splitting on inner loops. Split iterations to
669 // avoid range checks or one-shot null checks. Returns false if the
670 // current round of loop opts should stop.
671 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
672
673 // Driver for various flavors of iteration splitting. Returns false
674 // if the current round of loop opts should stop.
675 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
676
677 // Given dominators, try to find loops with calls that must always be
678 // executed (call dominates loop tail). These loops do not need non-call
679 // safepoints (ncsfpt).
680 void check_safepts(VectorSet &visited, Node_List &stack);
681
682 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
683 // encountered.
684 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
685
686 // Remove safepoints from loop. Optionally keeping one.
687 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
688
689 // Convert to counted loops where possible
690 void counted_loop( PhaseIdealLoop *phase );
691
692 // Check for Node being a loop-breaking test
693 Node *is_loop_exit(Node *iff) const;
694
695 // Remove simplistic dead code from loop body
696 void DCE_loop_body();
697
698 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
699 // Replace with a 1-in-10 exit guess.
700 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
701
702 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
703 // Useful for unrolling loops with NO array accesses.
704 bool policy_peel_only( PhaseIdealLoop *phase ) const;
705
706 // Return TRUE or FALSE if the loop should be unswitched -- clone
707 // loop with an invariant test
708 bool policy_unswitching( PhaseIdealLoop *phase ) const;
709
710 // Micro-benchmark spamming. Remove empty loops.
711 bool do_remove_empty_loop( PhaseIdealLoop *phase );
712
713 // Convert one iteration loop into normal code.
714 bool do_one_iteration_loop( PhaseIdealLoop *phase );
715
716 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
717 // move some loop-invariant test (usually a null-check) before the loop.
718 bool policy_peeling(PhaseIdealLoop *phase);
719
720 uint estimate_peeling(PhaseIdealLoop *phase);
721
722 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
723 // known trip count in the counted loop node.
724 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
725
726 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
727 // if the loop is a counted loop and the loop body is small enough.
728 bool policy_unroll(PhaseIdealLoop *phase);
729
730 // Loop analyses to map to a maximal superword unrolling for vectorization.
731 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
732
733 // Return TRUE or FALSE if the loop should be range-check-eliminated.
734 // Gather a list of IF tests that are dominated by iteration splitting;
735 // also gather the end of the first split and the start of the 2nd split.
736 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const;
737
738 // Return TRUE if "iff" is a range check.
739 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode *predicate_proj)) const;
740 bool is_range_check_if(IfNode* iff, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset,
741 jlong& scale) const;
742
743 // Estimate the number of nodes required when cloning a loop (body).
744 uint est_loop_clone_sz(uint factor) const;
745 // Estimate the number of nodes required when unrolling a loop (body).
746 uint est_loop_unroll_sz(uint factor) const;
747
748 // Compute loop trip count if possible
749 void compute_trip_count(PhaseIdealLoop* phase);
750
751 // Compute loop trip count from profile data
752 float compute_profile_trip_cnt_helper(Node* n);
753 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
754
755 // Reassociate invariant expressions.
756 void reassociate_invariants(PhaseIdealLoop *phase);
757 // Reassociate invariant binary expressions.
758 Node* reassociate(Node* n1, PhaseIdealLoop *phase);
759 // Reassociate invariant add and subtract expressions.
760 Node* reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase);
761 // Return nonzero index of invariant operand if invariant and variant
762 // are combined with an associative binary. Helper for reassociate_invariants.
763 int find_invariant(Node* n, PhaseIdealLoop *phase);
764 // Return TRUE if "n" is associative.
765 bool is_associative(Node* n, Node* base=NULL);
766
767 // Return true if n is invariant
768 bool is_invariant(Node* n) const;
769
770 // Put loop body on igvn work list
771 void record_for_igvn();
772
773 bool is_root() { return _parent == NULL; }
774 // A proper/reducible loop w/o any (occasional) dead back-edge.
775 bool is_loop() { return !_irreducible && !tail()->is_top(); }
776 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
777 bool is_innermost() { return is_loop() && _child == NULL; }
778
779 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
780
781 #ifdef ASSERT
782 // Tell whether the body contains nodes marked as reductions.
783 bool has_reduction_nodes() const;
784 #endif // ASSERT
785
786 #ifndef PRODUCT
787 void dump_head() const; // Dump loop head only
788 void dump() const; // Dump this loop recursively
789 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
790 #endif
791
792 private:
793 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
794
795 // Estimate the number of nodes resulting from control and data flow merge.
796 uint est_loop_flow_merge_sz() const;
797
798 // Check if the number of residual iterations is large with unroll_cnt.
799 // Return true if the residual iterations are more than 10% of the trip count.
800 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const {
801 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt();
802 }
803 };
804
805 // -----------------------------PhaseIdealLoop---------------------------------
806 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
807 // into a loop tree. Drives the loop-based transformations on the ideal graph.
808 class PhaseIdealLoop : public PhaseTransform {
809 friend class IdealLoopTree;
810 friend class SuperWord;
811 friend class CountedLoopReserveKit;
812 friend class ShenandoahBarrierC2Support;
813 friend class AutoNodeBudget;
814
815 // Pre-computed def-use info
816 PhaseIterGVN &_igvn;
817
818 // Head of loop tree
819 IdealLoopTree* _ltree_root;
820
821 // Array of pre-order numbers, plus post-visited bit.
822 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
823 // ODD for post-visited. Other bits are the pre-order number.
824 uint *_preorders;
825 uint _max_preorder;
826
827 const PhaseIdealLoop* _verify_me;
828 bool _verify_only;
829
830 // Allocate _preorders[] array
831 void allocate_preorders() {
832 _max_preorder = C->unique()+8;
833 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
834 memset(_preorders, 0, sizeof(uint) * _max_preorder);
835 }
836
837 // Allocate _preorders[] array
838 void reallocate_preorders() {
839 if ( _max_preorder < C->unique() ) {
840 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
841 _max_preorder = C->unique();
842 }
843 memset(_preorders, 0, sizeof(uint) * _max_preorder);
844 }
845
846 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
847 // adds new nodes.
848 void check_grow_preorders( ) {
849 if ( _max_preorder < C->unique() ) {
850 uint newsize = _max_preorder<<1; // double size of array
851 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
852 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
853 _max_preorder = newsize;
854 }
855 }
856 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
857 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
858 // Pre-order numbers are written to the Nodes array as low-bit-set values.
859 void set_preorder_visited( Node *n, int pre_order ) {
860 assert( !is_visited( n ), "already set" );
861 _preorders[n->_idx] = (pre_order<<1);
862 };
863 // Return pre-order number.
864 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
865
866 // Check for being post-visited.
867 // Should be previsited already (checked with assert(is_visited(n))).
868 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
869
870 // Mark as post visited
871 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
872
873 public:
874 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
875 // Returns true if "n" is a data node, false if it's a control node.
876 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
877
878 private:
879 // clear out dead code after build_loop_late
880 Node_List _deadlist;
881
882 // Support for faster execution of get_late_ctrl()/dom_lca()
883 // when a node has many uses and dominator depth is deep.
884 GrowableArray<jlong> _dom_lca_tags;
885 uint _dom_lca_tags_round;
886 void init_dom_lca_tags();
887
888 // Helper for debugging bad dominance relationships
889 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
890
891 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
892
893 // Inline wrapper for frequent cases:
894 // 1) only one use
895 // 2) a use is the same as the current LCA passed as 'n1'
896 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
897 assert( n->is_CFG(), "" );
898 // Fast-path NULL lca
899 if( lca != NULL && lca != n ) {
900 assert( lca->is_CFG(), "" );
901 // find LCA of all uses
902 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
903 }
904 return find_non_split_ctrl(n);
905 }
906 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
907
908 // Helper function for directing control inputs away from CFG split points.
909 Node *find_non_split_ctrl( Node *ctrl ) const {
910 if (ctrl != NULL) {
911 if (ctrl->is_MultiBranch()) {
912 ctrl = ctrl->in(0);
913 }
914 assert(ctrl->is_CFG(), "CFG");
915 }
916 return ctrl;
917 }
918
919 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
920
921 #ifdef ASSERT
922 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
923 #endif
924 void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
925 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
926 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
927 void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
928 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
929 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
930 Node* clone_skeleton_predicate_and_initialize(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj, Node* control,
931 IdealLoopTree* outer_loop, Node* input_proj);
932 Node* clone_skeleton_predicate_bool(Node* iff, Node* new_init, Node* new_stride, Node* control);
933 static bool skeleton_predicate_has_opaque(IfNode* iff);
934 static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride);
935 static bool subgraph_has_opaque(Node* n);
936 static void get_skeleton_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false);
937 void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
938 void copy_skeleton_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head, Node* init, Node* stride);
939 void initialize_skeleton_predicates_for_peeled_loop(ProjNode* predicate, LoopNode* outer_loop_head, int dd_outer_loop_head,
940 Node* init, Node* stride, IdealLoopTree* outer_loop,
941 const uint idx_before_clone, const Node_List& old_new);
942 void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol);
943 #ifdef ASSERT
944 bool only_has_infinite_loops();
945 #endif
946
947 void log_loop_tree();
948
949 public:
950
951 PhaseIterGVN &igvn() const { return _igvn; }
952
953 bool has_node( Node* n ) const {
954 guarantee(n != NULL, "No Node.");
955 return _nodes[n->_idx] != NULL;
956 }
957 // check if transform created new nodes that need _ctrl recorded
958 Node *get_late_ctrl( Node *n, Node *early );
959 Node *get_early_ctrl( Node *n );
960 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
961 void set_early_ctrl(Node* n, bool update_body);
962 void set_subtree_ctrl(Node* n, bool update_body);
963 void set_ctrl( Node *n, Node *ctrl ) {
964 assert( !has_node(n) || has_ctrl(n), "" );
965 assert( ctrl->in(0), "cannot set dead control node" );
966 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
967 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
968 }
969 // Set control and update loop membership
970 void set_ctrl_and_loop(Node* n, Node* ctrl) {
971 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
972 IdealLoopTree* new_loop = get_loop(ctrl);
973 if (old_loop != new_loop) {
974 if (old_loop->_child == NULL) old_loop->_body.yank(n);
975 if (new_loop->_child == NULL) new_loop->_body.push(n);
976 }
977 set_ctrl(n, ctrl);
978 }
979 // Control nodes can be replaced or subsumed. During this pass they
980 // get their replacement Node in slot 1. Instead of updating the block
981 // location of all Nodes in the subsumed block, we lazily do it. As we
982 // pull such a subsumed block out of the array, we write back the final
983 // correct block.
984 Node *get_ctrl( Node *i ) {
985
986 assert(has_node(i), "");
987 Node *n = get_ctrl_no_update(i);
988 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
989 assert(has_node(i) && has_ctrl(i), "");
990 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
991 return n;
992 }
993 // true if CFG node d dominates CFG node n
994 bool is_dominator(Node *d, Node *n);
995 // return get_ctrl for a data node and self(n) for a CFG node
996 Node* ctrl_or_self(Node* n) {
997 if (has_ctrl(n))
998 return get_ctrl(n);
999 else {
1000 assert (n->is_CFG(), "must be a CFG node");
1001 return n;
1002 }
1003 }
1004
1005 Node *get_ctrl_no_update_helper(Node *i) const {
1006 assert(has_ctrl(i), "should be control, not loop");
1007 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
1008 }
1009
1010 Node *get_ctrl_no_update(Node *i) const {
1011 assert( has_ctrl(i), "" );
1012 Node *n = get_ctrl_no_update_helper(i);
1013 if (!n->in(0)) {
1014 // Skip dead CFG nodes
1015 do {
1016 n = get_ctrl_no_update_helper(n);
1017 } while (!n->in(0));
1018 n = find_non_split_ctrl(n);
1019 }
1020 return n;
1021 }
1022
1023 // Check for loop being set
1024 // "n" must be a control node. Returns true if "n" is known to be in a loop.
1025 bool has_loop( Node *n ) const {
1026 assert(!has_node(n) || !has_ctrl(n), "");
1027 return has_node(n);
1028 }
1029 // Set loop
1030 void set_loop( Node *n, IdealLoopTree *loop ) {
1031 _nodes.map(n->_idx, (Node*)loop);
1032 }
1033 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1034 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1035 // from old_node to new_node to support the lazy update. Reference
1036 // replaces loop reference, since that is not needed for dead node.
1037 void lazy_update(Node *old_node, Node *new_node) {
1038 assert(old_node != new_node, "no cycles please");
1039 // Re-use the side array slot for this node to provide the
1040 // forwarding pointer.
1041 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
1042 }
1043 void lazy_replace(Node *old_node, Node *new_node) {
1044 _igvn.replace_node(old_node, new_node);
1045 lazy_update(old_node, new_node);
1046 }
1047
1048 private:
1049
1050 // Place 'n' in some loop nest, where 'n' is a CFG node
1051 void build_loop_tree();
1052 int build_loop_tree_impl( Node *n, int pre_order );
1053 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1054 // loop tree, not the root.
1055 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1056
1057 // Place Data nodes in some loop nest
1058 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1059 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1060 void build_loop_late_post_work(Node* n, bool pinned);
1061 void build_loop_late_post(Node* n);
1062 void verify_strip_mined_scheduling(Node *n, Node* least);
1063
1064 // Array of immediate dominance info for each CFG node indexed by node idx
1065 private:
1066 uint _idom_size;
1067 Node **_idom; // Array of immediate dominators
1068 uint *_dom_depth; // Used for fast LCA test
1069 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1070 LoopOptsMode _mode;
1071
1072 // build the loop tree and perform any requested optimizations
1073 void build_and_optimize();
1074
1075 // Dominators for the sea of nodes
1076 void Dominators();
1077
1078 // Compute the Ideal Node to Loop mapping
1079 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) :
1080 PhaseTransform(Ideal_Loop),
1081 _igvn(igvn),
1082 _verify_me(nullptr),
1083 _verify_only(false),
1084 _mode(mode),
1085 _nodes_required(UINT_MAX) {
1086 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop");
1087 build_and_optimize();
1088 }
1089
1090 #ifndef PRODUCT
1091 // Verify that verify_me made the same decisions as a fresh run
1092 // or only verify that the graph is valid if verify_me is null.
1093 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) :
1094 PhaseTransform(Ideal_Loop),
1095 _igvn(igvn),
1096 _verify_me(verify_me),
1097 _verify_only(verify_me == nullptr),
1098 _mode(LoopOptsVerify),
1099 _nodes_required(UINT_MAX) {
1100 build_and_optimize();
1101 }
1102 #endif
1103
1104 public:
1105 Node* idom_no_update(Node* d) const {
1106 return idom_no_update(d->_idx);
1107 }
1108
1109 Node* idom_no_update(uint didx) const {
1110 assert(didx < _idom_size, "oob");
1111 Node* n = _idom[didx];
1112 assert(n != NULL,"Bad immediate dominator info.");
1113 while (n->in(0) == NULL) { // Skip dead CFG nodes
1114 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1115 assert(n != NULL,"Bad immediate dominator info.");
1116 }
1117 return n;
1118 }
1119
1120 Node *idom(Node* d) const {
1121 return idom(d->_idx);
1122 }
1123
1124 Node *idom(uint didx) const {
1125 Node *n = idom_no_update(didx);
1126 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1127 return n;
1128 }
1129
1130 uint dom_depth(Node* d) const {
1131 guarantee(d != NULL, "Null dominator info.");
1132 guarantee(d->_idx < _idom_size, "");
1133 return _dom_depth[d->_idx];
1134 }
1135 void set_idom(Node* d, Node* n, uint dom_depth);
1136 // Locally compute IDOM using dom_lca call
1137 Node *compute_idom( Node *region ) const;
1138 // Recompute dom_depth
1139 void recompute_dom_depth();
1140
1141 // Is safept not required by an outer loop?
1142 bool is_deleteable_safept(Node* sfpt);
1143
1144 // Replace parallel induction variable (parallel to trip counter)
1145 void replace_parallel_iv(IdealLoopTree *loop);
1146
1147 Node *dom_lca( Node *n1, Node *n2 ) const {
1148 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1149 }
1150 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1151
1152 // Build and verify the loop tree without modifying the graph. This
1153 // is useful to verify that all inputs properly dominate their uses.
1154 static void verify(PhaseIterGVN& igvn) {
1155 #ifdef ASSERT
1156 ResourceMark rm;
1157 Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
1158 PhaseIdealLoop v(igvn);
1159 #endif
1160 }
1161
1162 // Recommended way to use PhaseIdealLoop.
1163 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
1164 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1165 ResourceMark rm;
1166 PhaseIdealLoop v(igvn, mode);
1167
1168 Compile* C = Compile::current();
1169 if (!C->failing()) {
1170 // Cleanup any modified bits
1171 igvn.optimize();
1172
1173 v.log_loop_tree();
1174 }
1175 }
1176
1177 // True if the method has at least 1 irreducible loop
1178 bool _has_irreducible_loops;
1179
1180 // Per-Node transform
1181 virtual Node* transform(Node* n) { return NULL; }
1182
1183 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1184 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1185 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1186 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1187 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1188
1189 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt);
1190
1191 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt);
1192 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new);
1193 #ifdef ASSERT
1194 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop);
1195 #endif
1196 void add_empty_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt);
1197 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop);
1198 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1199 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1200 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1201 Node*& entry_control, Node*& iffalse);
1202
1203 Node* exact_limit( IdealLoopTree *loop );
1204
1205 // Return a post-walked LoopNode
1206 IdealLoopTree *get_loop( Node *n ) const {
1207 // Dead nodes have no loop, so return the top level loop instead
1208 if (!has_node(n)) return _ltree_root;
1209 assert(!has_ctrl(n), "");
1210 return (IdealLoopTree*)_nodes[n->_idx];
1211 }
1212
1213 IdealLoopTree* ltree_root() const { return _ltree_root; }
1214
1215 // Is 'n' a (nested) member of 'loop'?
1216 int is_member( const IdealLoopTree *loop, Node *n ) const {
1217 return loop->is_member(get_loop(n)); }
1218
1219 // This is the basic building block of the loop optimizations. It clones an
1220 // entire loop body. It makes an old_new loop body mapping; with this
1221 // mapping you can find the new-loop equivalent to an old-loop node. All
1222 // new-loop nodes are exactly equal to their old-loop counterparts, all
1223 // edges are the same. All exits from the old-loop now have a RegionNode
1224 // that merges the equivalent new-loop path. This is true even for the
1225 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1226 // now come from (one or more) Phis that merge their new-loop equivalents.
1227 // Parameter side_by_side_idom:
1228 // When side_by_size_idom is NULL, the dominator tree is constructed for
1229 // the clone loop to dominate the original. Used in construction of
1230 // pre-main-post loop sequence.
1231 // When nonnull, the clone and original are side-by-side, both are
1232 // dominated by the passed in side_by_side_idom node. Used in
1233 // construction of unswitched loops.
1234 enum CloneLoopMode {
1235 IgnoreStripMined = 0, // Only clone inner strip mined loop
1236 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1237 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1238 // result control flow branches
1239 // either to inner clone or outer
1240 // strip mined loop.
1241 };
1242 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1243 CloneLoopMode mode, Node* side_by_side_idom = NULL);
1244 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1245 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1246 Node_List*& split_if_set, Node_List*& split_bool_set,
1247 Node_List*& split_cex_set, Node_List& worklist,
1248 uint new_counter, CloneLoopMode mode);
1249 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1250 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1251 Node_List& extra_data_nodes);
1252
1253 // If we got the effect of peeling, either by actually peeling or by
1254 // making a pre-loop which must execute at least once, we can remove
1255 // all loop-invariant dominated tests in the main body.
1256 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1257
1258 // Generate code to do a loop peel for the given loop (and body).
1259 // old_new is a temp array.
1260 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1261
1262 // Add pre and post loops around the given loop. These loops are used
1263 // during RCE, unrolling and aligning loops.
1264 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1265
1266 // Add post loop after the given loop.
1267 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1268 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1269 Node*& incr, Node* limit, CountedLoopNode*& post_head);
1270
1271 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1272 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1273
1274 // Add a vector post loop between a vector main loop and the current post loop
1275 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1276 // If Node n lives in the back_ctrl block, we clone a private version of n
1277 // in preheader_ctrl block and return that, otherwise return n.
1278 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1279
1280 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1281 // unroll to do double iterations. The next round of major loop transforms
1282 // will repeat till the doubled loop body does all remaining iterations in 1
1283 // pass.
1284 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1285
1286 // Unroll the loop body one step - make each trip do 2 iterations.
1287 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1288
1289 // Mark vector reduction candidates before loop unrolling
1290 void mark_reductions( IdealLoopTree *loop );
1291
1292 // Return true if exp is a constant times an induction var
1293 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0);
1294
1295 bool is_iv(Node* exp, Node* iv, BasicType bt);
1296
1297 // Return true if exp is a scaled induction var plus (or minus) constant
1298 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = NULL, int depth = 0);
1299 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) {
1300 jlong long_scale;
1301 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) {
1302 int int_scale = checked_cast<int>(long_scale);
1303 if (p_scale != NULL) {
1304 *p_scale = int_scale;
1305 }
1306 return true;
1307 }
1308 return false;
1309 }
1310 // Helper for finding more complex matches to is_scaled_iv_plus_offset.
1311 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv,
1312 BasicType bt,
1313 jlong* p_scale, Node** p_offset,
1314 bool* p_short_scale, int depth);
1315
1316
1317 // Enum to determine the action to be performed in create_new_if_for_predicate() when processing phis of UCT regions.
1318 enum class UnswitchingAction {
1319 None, // No special action.
1320 FastLoopCloning, // Need to clone nodes for the fast loop.
1321 SlowLoopRewiring // Need to rewire nodes for the slow loop.
1322 };
1323
1324 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1325 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1326 int opcode, bool if_cont_is_true_proj = true, Node_List* old_new = NULL,
1327 UnswitchingAction unswitching_action = UnswitchingAction::None);
1328
1329 // Clone data nodes for the fast loop while creating a new If with create_new_if_for_predicate.
1330 Node* clone_data_nodes_for_fast_loop(Node* phi_input, ProjNode* uncommon_proj, Node* if_uct, Node_List* old_new);
1331
1332 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1333
1334 static Node* skip_all_loop_predicates(Node* entry);
1335 static Node* skip_loop_predicates(Node* entry);
1336 static ProjNode* next_predicate(ProjNode* predicate);
1337
1338 // Find a good location to insert a predicate
1339 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1340
1341 class Predicates {
1342 public:
1343 // given loop entry, find all predicates above loop
1344 Predicates(Node* entry);
1345
1346 // Proj of empty loop limit check predicate
1347 ProjNode* loop_limit_check() {
1348 return _loop_limit_check;
1349 }
1350
1351 // Proj of empty profile predicate
1352 ProjNode* profile_predicate() {
1353 return _profile_predicate;
1354 }
1355
1356 // Proj of empty predicate
1357 ProjNode* predicate() {
1358 return _predicate;
1359 }
1360
1361 // First control node above all predicates
1362 Node* skip_all() {
1363 return _entry_to_all_predicates;
1364 }
1365
1366 private:
1367 ProjNode*_loop_limit_check = nullptr;
1368 ProjNode* _profile_predicate = nullptr;
1369 ProjNode* _predicate = nullptr;
1370 Node* _entry_to_all_predicates = nullptr;
1371 };
1372
1373 // Find a predicate
1374 static Node* find_predicate(Node* entry);
1375 // Construct a range check for a predicate if
1376 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1377 int scale, Node* offset,
1378 Node* init, Node* limit, jint stride,
1379 Node* range, bool upper, bool &overflow,
1380 bool negate);
1381
1382 // Implementation of the loop predication to promote checks outside the loop
1383 bool loop_predication_impl(IdealLoopTree *loop);
1384 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1385 CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1386 Deoptimization::DeoptReason reason);
1387 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt);
1388 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1389 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1390 Node_List& if_proj_list);
1391 ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1392 ProjNode* proj, ProjNode *predicate_proj,
1393 ProjNode* upper_bound_proj,
1394 int scale, Node* offset,
1395 Node* init, Node* limit, jint stride,
1396 Node* rng, bool& overflow,
1397 Deoptimization::DeoptReason reason);
1398 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1399 Node* predicate_proj, int scale_con, Node* offset,
1400 Node* limit, jint stride_con, Node* value);
1401
1402 // Helper function to collect predicate for eliminating the useless ones
1403 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1404 void eliminate_useless_predicates();
1405
1406 // Change the control input of expensive nodes to allow commoning by
1407 // IGVN when it is guaranteed to not result in a more frequent
1408 // execution of the expensive node. Return true if progress.
1409 bool process_expensive_nodes();
1410
1411 // Check whether node has become unreachable
1412 bool is_node_unreachable(Node *n) const {
1413 return !has_node(n) || n->is_unreachable(_igvn);
1414 }
1415
1416 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1417 int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1418
1419 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1420 void has_range_checks(IdealLoopTree *loop);
1421
1422 // Process post loops which have range checks and try to build a multi-version
1423 // guard to safely determine if we can execute the post loop which was RCE'd.
1424 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1425
1426 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1427 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1428
1429 // Create a slow version of the loop by cloning the loop
1430 // and inserting an if to select fast-slow versions.
1431 // Return the inserted if.
1432 IfNode* create_slow_version_of_loop(IdealLoopTree *loop,
1433 Node_List &old_new,
1434 IfNode* unswitch_iff,
1435 CloneLoopMode mode);
1436
1437 // Clone a loop and return the clone head (clone_loop_head).
1438 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1439 // This routine was created for usage in CountedLoopReserveKit.
1440 //
1441 // int(1) -> If -> IfTrue -> original_loop_head
1442 // |
1443 // V
1444 // IfFalse -> clone_loop_head (returned by function pointer)
1445 //
1446 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1447 // Clone loop with an invariant test (that does not exit) and
1448 // insert a clone of the test that selects which version to
1449 // execute.
1450 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1451
1452 // Find candidate "if" for unswitching
1453 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1454
1455 // Range Check Elimination uses this function!
1456 // Constrain the main loop iterations so the affine function:
1457 // low_limit <= scale_con * I + offset < upper_limit
1458 // always holds true. That is, either increase the number of iterations in
1459 // the pre-loop or the post-loop until the condition holds true in the main
1460 // loop. Scale_con, offset and limit are all loop invariant.
1461 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);
1462 // Helper function for add_constraint().
1463 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1464
1465 // Partially peel loop up through last_peel node.
1466 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1467 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1468
1469 // Create a scheduled list of nodes control dependent on ctrl set.
1470 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1471 // Has a use in the vector set
1472 bool has_use_in_set( Node* n, VectorSet& vset );
1473 // Has use internal to the vector set (ie. not in a phi at the loop head)
1474 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1475 // clone "n" for uses that are outside of loop
1476 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1477 // clone "n" for special uses that are in the not_peeled region
1478 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1479 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1480 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1481 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1482 #ifdef ASSERT
1483 // Validate the loop partition sets: peel and not_peel
1484 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1485 // Ensure that uses outside of loop are of the right form
1486 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1487 uint orig_exit_idx, uint clone_exit_idx);
1488 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1489 #endif
1490
1491 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1492 int stride_of_possible_iv( Node* iff );
1493 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1494 // Return the (unique) control output node that's in the loop (if it exists.)
1495 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1496 // Insert a signed compare loop exit cloned from an unsigned compare.
1497 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1498 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1499 // Utility to register node "n" with PhaseIdealLoop
1500 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1501 // Utility to create an if-projection
1502 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1503 // Force the iff control output to be the live_proj
1504 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1505 // Insert a region before an if projection
1506 RegionNode* insert_region_before_proj(ProjNode* proj);
1507 // Insert a new if before an if projection
1508 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1509
1510 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1511 // "Nearly" because all Nodes have been cloned from the original in the loop,
1512 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1513 // through the Phi recursively, and return a Bool.
1514 Node* clone_iff(PhiNode* phi);
1515 CmpNode* clone_bool(PhiNode* phi);
1516
1517
1518 // Rework addressing expressions to get the most loop-invariant stuff
1519 // moved out. We'd like to do all associative operators, but it's especially
1520 // important (common) to do address expressions.
1521 Node* remix_address_expressions(Node* n);
1522 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt);
1523
1524 // Convert add to muladd to generate MuladdS2I under certain criteria
1525 Node * convert_add_to_muladd(Node * n);
1526
1527 // Attempt to use a conditional move instead of a phi/branch
1528 Node *conditional_move( Node *n );
1529
1530 // Reorganize offset computations to lower register pressure.
1531 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1532 // (which are then alive with the post-incremented trip counter
1533 // forcing an extra register move)
1534 void reorg_offsets( IdealLoopTree *loop );
1535
1536 // Check for aggressive application of 'split-if' optimization,
1537 // using basic block level info.
1538 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1539 Node *split_if_with_blocks_pre ( Node *n );
1540 void split_if_with_blocks_post( Node *n );
1541 Node *has_local_phi_input( Node *n );
1542 // Mark an IfNode as being dominated by a prior test,
1543 // without actually altering the CFG (and hence IDOM info).
1544 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool exclude_loop_predicate = false);
1545
1546 // Split Node 'n' through merge point
1547 RegionNode* split_thru_region(Node* n, RegionNode* region);
1548 // Split Node 'n' through merge point if there is enough win.
1549 Node *split_thru_phi( Node *n, Node *region, int policy );
1550 // Found an If getting its condition-code input from a Phi in the
1551 // same block. Split thru the Region.
1552 void do_split_if(Node *iff, RegionNode** new_false_region = NULL, RegionNode** new_true_region = NULL);
1553
1554 // Conversion of fill/copy patterns into intrinsic versions
1555 bool do_intrinsify_fill();
1556 bool intrinsify_fill(IdealLoopTree* lpt);
1557 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1558 Node*& shift, Node*& offset);
1559
1560 private:
1561 // Return a type based on condition control flow
1562 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1563 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1564 // Helpers for filtered type
1565 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1566
1567 // Helper functions
1568 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1569 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1570 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 );
1571 bool split_up( Node *n, Node *blk1, Node *blk2 );
1572 void sink_use( Node *use, Node *post_loop );
1573 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1574 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1575 void try_move_store_after_loop(Node* n);
1576 bool identical_backtoback_ifs(Node *n);
1577 bool can_split_if(Node *n_ctrl);
1578
1579 // Determine if a method is too big for a/another round of split-if, based on
1580 // a magic (approximate) ratio derived from the equally magic constant 35000,
1581 // previously used for this purpose (but without relating to the node limit).
1582 bool must_throttle_split_if() {
1583 uint threshold = C->max_node_limit() * 2 / 5;
1584 return C->live_nodes() > threshold;
1585 }
1586
1587 // A simplistic node request tracking mechanism, where
1588 // = UINT_MAX Request not valid or made final.
1589 // < UINT_MAX Nodes currently requested (estimate).
1590 uint _nodes_required;
1591
1592 enum { REQUIRE_MIN = 70 };
1593
1594 uint nodes_required() const { return _nodes_required; }
1595
1596 // Given the _currently_ available number of nodes, check whether there is
1597 // "room" for an additional request or not, considering the already required
1598 // number of nodes. Return TRUE if the new request is exceeding the node
1599 // budget limit, otherwise return FALSE. Note that this interpretation will
1600 // act pessimistic on additional requests when new nodes have already been
1601 // generated since the 'begin'. This behaviour fits with the intention that
1602 // node estimates/requests should be made upfront.
1603 bool exceeding_node_budget(uint required = 0) {
1604 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1605 uint available = C->max_node_limit() - C->live_nodes();
1606 return available < required + _nodes_required + REQUIRE_MIN;
1607 }
1608
1609 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1610 precond(require > 0);
1611 _nodes_required += MAX2(require, minreq);
1612 return _nodes_required;
1613 }
1614
1615 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1616 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1617 }
1618
1619 uint require_nodes_begin() {
1620 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1621 _nodes_required = 0;
1622 return C->live_nodes();
1623 }
1624
1625 // When a node request is final, optionally check that the requested number
1626 // of nodes was reasonably correct with respect to the number of new nodes
1627 // introduced since the last 'begin'. Always check that we have not exceeded
1628 // the maximum node limit.
1629 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1630 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1631
1632 #ifdef ASSERT
1633 if (check_estimate) {
1634 // Check that the node budget request was not off by too much (x2).
1635 // Should this be the case we _surely_ need to improve the estimates
1636 // used in our budget calculations.
1637 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1638 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1639 C->live_nodes() - live_at_begin, _nodes_required);
1640 }
1641 }
1642 #endif
1643 // Assert that we have stayed within the node budget limit.
1644 assert(C->live_nodes() < C->max_node_limit(),
1645 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1646 C->live_nodes() - live_at_begin, live_at_begin,
1647 C->max_node_limit(), _nodes_required);
1648
1649 _nodes_required = UINT_MAX;
1650 }
1651
1652 // Clone loop predicates to slow and fast loop when unswitching a loop
1653 void clone_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new, ProjNode*& iffast_pred, ProjNode*& ifslow_pred);
1654 ProjNode* clone_predicate_to_unswitched_loop(ProjNode* predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1655 Node_List* old_new = NULL);
1656 void clone_skeleton_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new, Deoptimization::DeoptReason reason,
1657 ProjNode* old_predicate_proj, ProjNode* iffast_pred, ProjNode* ifslow_pred);
1658 ProjNode* clone_skeleton_predicate_for_unswitched_loops(Node* iff, ProjNode* predicate,
1659 Deoptimization::DeoptReason reason,
1660 ProjNode* output_proj);
1661 static void check_created_predicate_for_unswitching(const Node* new_entry) PRODUCT_RETURN;
1662
1663 bool _created_loop_node;
1664 #ifdef ASSERT
1665 void dump_real_LCA(Node* early, Node* wrong_lca);
1666 bool check_idom_chains_intersection(const Node* n, uint& idom_idx_new, uint& idom_idx_other, const Node_List* nodes_seen) const;
1667 #endif
1668
1669 public:
1670 void set_created_loop_node() { _created_loop_node = true; }
1671 bool created_loop_node() { return _created_loop_node; }
1672 void register_new_node(Node* n, Node* blk);
1673
1674 #ifdef ASSERT
1675 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1676 #endif
1677
1678 #ifndef PRODUCT
1679 void dump() const;
1680 void dump_idom(Node* n) const;
1681 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1682 void verify() const; // Major slow :-)
1683 void verify_compare(Node* n, const PhaseIdealLoop* loop_verify, VectorSet &visited) const;
1684 IdealLoopTree* get_loop_idx(Node* n) const {
1685 // Dead nodes have no loop, so return the top level loop instead
1686 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1687 }
1688 // Print some stats
1689 static void print_statistics();
1690 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1691 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1692 static volatile int _long_loop_candidates;
1693 static volatile int _long_loop_nests;
1694 static volatile int _long_loop_counted_loops;
1695 #endif
1696
1697 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1698
1699 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN;
1700
1701 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test);
1702
1703
1704 int extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi,
1705 Node_List &range_checks);
1706
1707 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1708 Node* inner_iters_actual_int, Node* inner_phi,
1709 Node* iv_add, LoopNode* inner_head);
1710
1711 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA);
1712
1713 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl);
1714
1715 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop);
1716
1717 Node* compute_early_ctrl(Node* n, Node* n_ctrl);
1718
1719 void try_sink_out_of_loop(Node* n);
1720
1721 Node* clamp(Node* R, Node* L, Node* H);
1722
1723 bool safe_for_if_replacement(const Node* dom) const;
1724
1725 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region);
1726
1727 bool try_merge_identical_ifs(Node* n);
1728
1729 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm);
1730
1731 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
1732 IdealLoopTree* parent, bool partial);
1733
1734 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
1735 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist);
1736
1737 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
1738 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1739 Node_List*& split_bool_set, Node_List*& split_cex_set);
1740
1741 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1742 };
1743
1744
1745 class AutoNodeBudget : public StackObj
1746 {
1747 public:
1748 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1749
1750 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1751 : _phase(phase),
1752 _check_at_final(chk == BUDGET_CHECK),
1753 _nodes_at_begin(0)
1754 {
1755 precond(_phase != NULL);
1756
1757 _nodes_at_begin = _phase->require_nodes_begin();
1758 }
1759
1760 ~AutoNodeBudget() {
1761 #ifndef PRODUCT
1762 if (TraceLoopOpts) {
1763 uint request = _phase->nodes_required();
1764 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1765
1766 if (request < delta) {
1767 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1768 } else {
1769 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1770 // Identify the worst estimates as "poor" ones.
1771 if (request > REQUIRE_MIN && delta > 0) {
1772 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1773 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1774 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1775 }
1776 }
1777 }
1778 }
1779 #endif // PRODUCT
1780 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1781 }
1782
1783 private:
1784 PhaseIdealLoop* _phase;
1785 bool _check_at_final;
1786 uint _nodes_at_begin;
1787 };
1788
1789
1790 // This kit may be used for making of a reserved copy of a loop before this loop
1791 // goes under non-reversible changes.
1792 //
1793 // Function create_reserve() creates a reserved copy (clone) of the loop.
1794 // The reserved copy is created by calling
1795 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1796 // the original and reserved loops are connected in the outer graph.
1797 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1798 //
1799 // By default the reserved copy (clone) of the loop is created as dead code - it is
1800 // dominated in the outer loop by this node chain:
1801 // intcon(1)->If->IfFalse->reserved_copy.
1802 // The original loop is dominated by the same node chain but IfTrue projection:
1803 // intcon(0)->If->IfTrue->original_loop.
1804 //
1805 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1806 // and the dtor, checks _use_new value.
1807 // If _use_new == false, it "switches" control to reserved copy of the loop
1808 // by simple replacing of node intcon(1) with node intcon(0).
1809 //
1810 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1811 //
1812 // void CountedLoopReserveKit_example()
1813 // {
1814 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1815 // if (DoReserveCopy && !lrk.has_reserved()) {
1816 // return; //failed to create reserved loop copy
1817 // }
1818 // ...
1819 // //something is wrong, switch to original loop
1820 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1821 // ...
1822 // //everything worked ok, return with the newly modified loop
1823 // lrk.use_new();
1824 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1825 // }
1826 //
1827 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1828 // the dtor will "switch to the original" loop.
1829 // NOTE. You you modify outside of the original loop this class is no help.
1830 //
1831 class CountedLoopReserveKit {
1832 private:
1833 PhaseIdealLoop* _phase;
1834 IdealLoopTree* _lpt;
1835 LoopNode* _lp;
1836 IfNode* _iff;
1837 LoopNode* _lp_reserved;
1838 bool _has_reserved;
1839 bool _use_new;
1840 const bool _active; //may be set to false in ctor, then the object is dummy
1841
1842 public:
1843 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1844 ~CountedLoopReserveKit();
1845 void use_new() {_use_new = true;}
1846 void set_iff(IfNode* x) {_iff = x;}
1847 bool has_reserved() const { return _active && _has_reserved;}
1848 private:
1849 bool create_reserve();
1850 };// class CountedLoopReserveKit
1851
1852 inline Node* IdealLoopTree::tail() {
1853 // Handle lazy update of _tail field.
1854 if (_tail->in(0) == NULL) {
1855 _tail = _phase->get_ctrl(_tail);
1856 }
1857 return _tail;
1858 }
1859
1860 inline Node* IdealLoopTree::head() {
1861 // Handle lazy update of _head field.
1862 if (_head->in(0) == NULL) {
1863 _head = _phase->get_ctrl(_head);
1864 }
1865 return _head;
1866 }
1867
1868 // Iterate over the loop tree using a preorder, left-to-right traversal.
1869 //
1870 // Example that visits all counted loops from within PhaseIdealLoop
1871 //
1872 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1873 // IdealLoopTree* lpt = iter.current();
1874 // if (!lpt->is_counted()) continue;
1875 // ...
1876 class LoopTreeIterator : public StackObj {
1877 private:
1878 IdealLoopTree* _root;
1879 IdealLoopTree* _curnt;
1880
1881 public:
1882 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1883
1884 bool done() { return _curnt == NULL; } // Finished iterating?
1885
1886 void next(); // Advance to next loop tree
1887
1888 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1889 };
1890
1891 // Compute probability of reaching some CFG node from a fixed
1892 // dominating CFG node
1893 class PathFrequency {
1894 private:
1895 Node* _dom; // frequencies are computed relative to this node
1896 Node_Stack _stack;
1897 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
1898 GrowableArray<float> _freqs; // cache frequencies
1899 PhaseIdealLoop* _phase;
1900
1901 float check_and_truncate_frequency(float f) {
1902 assert(f >= 0, "Incorrect frequency");
1903 // We do not perform an exact (f <= 1) check
1904 // this would be error prone with rounding of floats.
1905 // Performing a check like (f <= 1+eps) would be of benefit,
1906 // however, it is not evident how to determine such an eps,
1907 // given that an arbitrary number of add/mul operations
1908 // are performed on these frequencies.
1909 return (f > 1) ? 1 : f;
1910 }
1911
1912 public:
1913 PathFrequency(Node* dom, PhaseIdealLoop* phase)
1914 : _dom(dom), _stack(0), _phase(phase) {
1915 }
1916
1917 float to(Node* n);
1918 };
1919
1920 #endif // SHARE_OPTO_LOOPNODE_HPP