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