1 /*
2 * Copyright (c) 1998, 2018, 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_VM_OPTO_LOOPNODE_HPP
26 #define SHARE_VM_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 CountedLoopEndNode;
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=4,
65 HasExactTripCount=8,
66 InnerLoop=16,
67 PartialPeelLoop=32,
68 PartialPeelFailed=64,
69 HasReductions=128,
70 WasSlpAnalyzed=256,
71 PassedSlpAnalysis=512,
72 DoUnrollOnly=1024,
73 VectorizedLoop=2048,
74 HasAtomicPostLoop=4096,
75 HasRangeChecks=8192,
76 IsMultiversioned=16384,
77 StripMined=32768,
78 SubwordLoop=65536,
79 ProfileTripFailed=131072};
80 char _unswitch_count;
81 enum { _unswitch_max=3 };
82 char _postloop_flags;
83 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
84
85 // Expected trip count from profile data
86 float _profile_trip_cnt;
87
88 public:
89 // Names for edge indices
90 enum { Self=0, EntryControl, LoopBackControl };
91
92 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
93 void set_inner_loop() { _loop_flags |= InnerLoop; }
94
95 bool range_checks_present() const { return _loop_flags & HasRangeChecks; }
96 bool is_multiversioned() const { return _loop_flags & IsMultiversioned; }
97 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
98 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
99 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
100 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
101 bool is_strip_mined() const { return _loop_flags & StripMined; }
102 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
103 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
104
105 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
106 void mark_has_reductions() { _loop_flags |= HasReductions; }
107 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
108 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
109 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
110 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
111 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
112 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
113 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
114 void mark_strip_mined() { _loop_flags |= StripMined; }
115 void clear_strip_mined() { _loop_flags &= ~StripMined; }
116 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
117 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
118
119 int unswitch_max() { return _unswitch_max; }
120 int unswitch_count() { return _unswitch_count; }
121
122 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
123 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
124 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
125 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
126
127 void set_unswitch_count(int val) {
128 assert (val <= unswitch_max(), "too many unswitches");
129 _unswitch_count = val;
130 }
131
132 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
133 float profile_trip_cnt() { return _profile_trip_cnt; }
134
135 LoopNode(Node *entry, Node *backedge)
136 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
137 _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) {
138 init_class_id(Class_Loop);
139 init_req(EntryControl, entry);
140 init_req(LoopBackControl, backedge);
141 }
142
143 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
144 virtual int Opcode() const;
145 bool can_be_counted_loop(PhaseTransform* phase) const {
146 return req() == 3 && in(0) != NULL &&
147 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
148 in(2) != NULL && phase->type(in(2)) != Type::TOP;
149 }
150 bool is_valid_counted_loop() const;
151 #ifndef PRODUCT
152 virtual void dump_spec(outputStream *st) const;
153 #endif
154
155 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN;
156 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
157 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
158 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
159 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
160 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; }
161 };
162
163 //------------------------------Counted Loops----------------------------------
164 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
165 // path (and maybe some other exit paths). The trip-counter exit is always
166 // last in the loop. The trip-counter have to stride by a constant;
167 // the exit value is also loop invariant.
168
169 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
170 // CountedLoopNode has the incoming loop control and the loop-back-control
171 // which is always the IfTrue before the matching CountedLoopEndNode. The
172 // CountedLoopEndNode has an incoming control (possibly not the
173 // CountedLoopNode if there is control flow in the loop), the post-increment
174 // trip-counter value, and the limit. The trip-counter value is always of
175 // the form (Op old-trip-counter stride). The old-trip-counter is produced
176 // by a Phi connected to the CountedLoopNode. The stride is constant.
177 // The Op is any commutable opcode, including Add, Mul, Xor. The
178 // CountedLoopEndNode also takes in the loop-invariant limit value.
179
180 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
181 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
182 // via the old-trip-counter from the Op node.
183
184 //------------------------------CountedLoopNode--------------------------------
185 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
186 // inputs the incoming loop-start control and the loop-back control, so they
187 // act like RegionNodes. They also take in the initial trip counter, the
188 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
189 // produce a loop-body control and the trip counter value. Since
190 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
191
192 class CountedLoopNode : public LoopNode {
193 // Size is bigger to hold _main_idx. However, _main_idx does not change
194 // the semantics so it does not appear in the hash & cmp functions.
195 virtual uint size_of() const { return sizeof(*this); }
196
197 // For Pre- and Post-loops during debugging ONLY, this holds the index of
198 // the Main CountedLoop. Used to assert that we understand the graph shape.
199 node_idx_t _main_idx;
200
201 // Known trip count calculated by compute_exact_trip_count()
202 uint _trip_count;
203
204 // Log2 of original loop bodies in unrolled loop
205 int _unrolled_count_log2;
206
207 // Node count prior to last unrolling - used to decide if
208 // unroll,optimize,unroll,optimize,... is making progress
209 int _node_count_before_unroll;
210
211 // If slp analysis is performed we record the maximum
212 // vector mapped unroll factor here
213 int _slp_maximum_unroll_factor;
214
215 public:
216 CountedLoopNode( Node *entry, Node *backedge )
217 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
218 _unrolled_count_log2(0), _node_count_before_unroll(0),
219 _slp_maximum_unroll_factor(0) {
220 init_class_id(Class_CountedLoop);
221 // Initialize _trip_count to the largest possible value.
222 // Will be reset (lower) if the loop's trip count is known.
223 }
224
225 virtual int Opcode() const;
226 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
227
228 Node *init_control() const { return in(EntryControl); }
229 Node *back_control() const { return in(LoopBackControl); }
230 CountedLoopEndNode *loopexit_or_null() const;
231 CountedLoopEndNode *loopexit() const;
232 Node *init_trip() const;
233 Node *stride() const;
234 int stride_con() const;
235 bool stride_is_con() const;
236 Node *limit() const;
237 Node *incr() const;
238 Node *phi() const;
239
240 // Match increment with optional truncation
241 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
242
243 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
244 // can run short a few iterations and may start a few iterations in.
245 // It will be RCE'd and unrolled and aligned.
246
247 // A following 'post' loop will run any remaining iterations. Used
248 // during Range Check Elimination, the 'post' loop will do any final
249 // iterations with full checks. Also used by Loop Unrolling, where
250 // the 'post' loop will do any epilog iterations needed. Basically,
251 // a 'post' loop can not profitably be further unrolled or RCE'd.
252
253 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
254 // it may do under-flow checks for RCE and may do alignment iterations
255 // so the following main loop 'knows' that it is striding down cache
256 // lines.
257
258 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
259 // Aligned, may be missing it's pre-loop.
260 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
261 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
262 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
263 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
264 bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
265 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
266 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
267 bool do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
268 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
269 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
270 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
271
272 int main_idx() const { return _main_idx; }
273
274
275 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
276 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
277 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
278 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
279
280 void set_trip_count(uint tc) { _trip_count = tc; }
281 uint trip_count() { return _trip_count; }
282
283 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
284 void set_exact_trip_count(uint tc) {
285 _trip_count = tc;
286 _loop_flags |= HasExactTripCount;
287 }
288 void set_nonexact_trip_count() {
289 _loop_flags &= ~HasExactTripCount;
290 }
291 void set_notpassed_slp() {
292 _loop_flags &= ~PassedSlpAnalysis;
293 }
294
295 void double_unrolled_count() { _unrolled_count_log2++; }
296 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
297
298 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
299 int node_count_before_unroll() { return _node_count_before_unroll; }
300 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
301 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
302
303 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1);
304 OuterStripMinedLoopNode* outer_loop() const;
305 virtual IfTrueNode* outer_loop_tail() const;
306 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
307 virtual IfFalseNode* outer_loop_exit() const;
308 virtual SafePointNode* outer_safepoint() const;
309
310 // If this is a main loop in a pre/main/post loop nest, walk over
311 // the predicates that were inserted by
312 // duplicate_predicates()/add_range_check_predicate()
313 static Node* skip_predicates_from_entry(Node* ctrl);
314 Node* skip_predicates();
315
316 #ifndef PRODUCT
317 virtual void dump_spec(outputStream *st) const;
318 #endif
319 };
320
321 //------------------------------CountedLoopEndNode-----------------------------
322 // CountedLoopEndNodes end simple trip counted loops. They act much like
323 // IfNodes.
324 class CountedLoopEndNode : public IfNode {
325 public:
326 enum { TestControl, TestValue };
327
328 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
329 : IfNode( control, test, prob, cnt) {
330 init_class_id(Class_CountedLoopEnd);
331 }
332 virtual int Opcode() const;
333
334 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
335 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
336 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
337 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
338 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
339 int stride_con() const;
340 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
341 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
342 PhiNode *phi() const {
343 Node *tmp = incr();
344 if (tmp && tmp->req() == 3) {
345 Node* phi = tmp->in(1);
346 if (phi->is_Phi()) {
347 return phi->as_Phi();
348 }
349 }
350 return NULL;
351 }
352 CountedLoopNode *loopnode() const {
353 // The CountedLoopNode that goes with this CountedLoopEndNode may
354 // have been optimized out by the IGVN so be cautious with the
355 // pattern matching on the graph
356 PhiNode* iv_phi = phi();
357 if (iv_phi == NULL) {
358 return NULL;
359 }
360 Node *ln = iv_phi->in(0);
361 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) {
362 return (CountedLoopNode*)ln;
363 }
364 return NULL;
365 }
366
367 #ifndef PRODUCT
368 virtual void dump_spec(outputStream *st) const;
369 #endif
370 };
371
372
373 inline CountedLoopEndNode *CountedLoopNode::loopexit_or_null() const {
374 Node *bc = back_control();
375 if( bc == NULL ) return NULL;
376 Node *le = bc->in(0);
377 if( le->Opcode() != Op_CountedLoopEnd )
378 return NULL;
379 return (CountedLoopEndNode*)le;
380 }
381 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
382 CountedLoopEndNode* cle = loopexit_or_null();
383 assert(cle != NULL, "loopexit is NULL");
384 return cle;
385 }
386 inline Node *CountedLoopNode::init_trip() const { return loopexit_or_null() ? loopexit()->init_trip() : NULL; }
387 inline Node *CountedLoopNode::stride() const { return loopexit_or_null() ? loopexit()->stride() : NULL; }
388 inline int CountedLoopNode::stride_con() const { return loopexit_or_null() ? loopexit()->stride_con() : 0; }
389 inline bool CountedLoopNode::stride_is_con() const { return loopexit_or_null() && loopexit()->stride_is_con(); }
390 inline Node *CountedLoopNode::limit() const { return loopexit_or_null() ? loopexit()->limit() : NULL; }
391 inline Node *CountedLoopNode::incr() const { return loopexit_or_null() ? loopexit()->incr() : NULL; }
392 inline Node *CountedLoopNode::phi() const { return loopexit_or_null() ? loopexit()->phi() : NULL; }
393
394 //------------------------------LoopLimitNode-----------------------------
395 // Counted Loop limit node which represents exact final iterator value:
396 // trip_count = (limit - init_trip + stride - 1)/stride
397 // final_value= trip_count * stride + init_trip.
398 // Use HW instructions to calculate it when it can overflow in integer.
399 // Note, final_value should fit into integer since counted loop has
400 // limit check: limit <= max_int-stride.
401 class LoopLimitNode : public Node {
402 enum { Init=1, Limit=2, Stride=3 };
403 public:
404 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
405 // Put it on the Macro nodes list to optimize during macro nodes expansion.
406 init_flags(Flag_is_macro);
407 C->add_macro_node(this);
408 }
409 virtual int Opcode() const;
410 virtual const Type *bottom_type() const { return TypeInt::INT; }
411 virtual uint ideal_reg() const { return Op_RegI; }
412 virtual const Type* Value(PhaseGVN* phase) const;
413 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
414 virtual Node* Identity(PhaseGVN* phase);
415 };
416
417 // Support for strip mining
418 class OuterStripMinedLoopNode : public LoopNode {
419 private:
420 CountedLoopNode* inner_loop() const;
421 public:
422 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
423 : LoopNode(entry, backedge) {
424 init_class_id(Class_OuterStripMinedLoop);
425 init_flags(Flag_is_macro);
426 C->add_macro_node(this);
427 }
428
429 virtual int Opcode() const;
430
431 virtual IfTrueNode* outer_loop_tail() const;
432 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
433 virtual IfFalseNode* outer_loop_exit() const;
434 virtual SafePointNode* outer_safepoint() const;
435 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
436 };
437
438 class OuterStripMinedLoopEndNode : public IfNode {
439 public:
440 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
441 : IfNode(control, test, prob, cnt) {
442 init_class_id(Class_OuterStripMinedLoopEnd);
443 }
444
445 virtual int Opcode() const;
446
447 virtual const Type* Value(PhaseGVN* phase) const;
448 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
449 };
450
451 // -----------------------------IdealLoopTree----------------------------------
452 class IdealLoopTree : public ResourceObj {
453 public:
454 IdealLoopTree *_parent; // Parent in loop tree
455 IdealLoopTree *_next; // Next sibling in loop tree
456 IdealLoopTree *_child; // First child in loop tree
457
458 // The head-tail backedge defines the loop.
459 // If tail is NULL then this loop has multiple backedges as part of the
460 // same loop. During cleanup I'll peel off the multiple backedges; merge
461 // them at the loop bottom and flow 1 real backedge into the loop.
462 Node *_head; // Head of loop
463 Node *_tail; // Tail of loop
464 inline Node *tail(); // Handle lazy update of _tail field
465 PhaseIdealLoop* _phase;
466 int _local_loop_unroll_limit;
467 int _local_loop_unroll_factor;
468
469 Node_List _body; // Loop body for inner loops
470
471 uint8_t _nest; // Nesting depth
472 uint8_t _irreducible:1, // True if irreducible
473 _has_call:1, // True if has call safepoint
474 _has_sfpt:1, // True if has non-call safepoint
475 _rce_candidate:1; // True if candidate for range check elimination
476
477 Node_List* _safepts; // List of safepoints in this loop
478 Node_List* _required_safept; // A inner loop cannot delete these safepts;
479 bool _allow_optimizations; // Allow loop optimizations
480
481 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
482 : _parent(0), _next(0), _child(0),
483 _head(head), _tail(tail),
484 _phase(phase),
485 _safepts(NULL),
486 _required_safept(NULL),
487 _allow_optimizations(true),
488 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
489 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
490 { }
491
492 // Is 'l' a member of 'this'?
493 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
494
495 // Set loop nesting depth. Accumulate has_call bits.
496 int set_nest( uint depth );
497
498 // Split out multiple fall-in edges from the loop header. Move them to a
499 // private RegionNode before the loop. This becomes the loop landing pad.
500 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
501
502 // Split out the outermost loop from this shared header.
503 void split_outer_loop( PhaseIdealLoop *phase );
504
505 // Merge all the backedges from the shared header into a private Region.
506 // Feed that region as the one backedge to this loop.
507 void merge_many_backedges( PhaseIdealLoop *phase );
508
509 // Split shared headers and insert loop landing pads.
510 // Insert a LoopNode to replace the RegionNode.
511 // Returns TRUE if loop tree is structurally changed.
512 bool beautify_loops( PhaseIdealLoop *phase );
513
514 // Perform optimization to use the loop predicates for null checks and range checks.
515 // Applies to any loop level (not just the innermost one)
516 bool loop_predication( PhaseIdealLoop *phase);
517
518 // Perform iteration-splitting on inner loops. Split iterations to
519 // avoid range checks or one-shot null checks. Returns false if the
520 // current round of loop opts should stop.
521 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
522
523 // Driver for various flavors of iteration splitting. Returns false
524 // if the current round of loop opts should stop.
525 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
526
527 // Given dominators, try to find loops with calls that must always be
528 // executed (call dominates loop tail). These loops do not need non-call
529 // safepoints (ncsfpt).
530 void check_safepts(VectorSet &visited, Node_List &stack);
531
532 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
533 // encountered.
534 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
535
536 // Remove safepoints from loop. Optionally keeping one.
537 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
538
539 // Convert to counted loops where possible
540 void counted_loop( PhaseIdealLoop *phase );
541
542 // Check for Node being a loop-breaking test
543 Node *is_loop_exit(Node *iff) const;
544
545 // Remove simplistic dead code from loop body
546 void DCE_loop_body();
547
548 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
549 // Replace with a 1-in-10 exit guess.
550 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
551
552 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
553 // Useful for unrolling loops with NO array accesses.
554 bool policy_peel_only( PhaseIdealLoop *phase ) const;
555
556 // Return TRUE or FALSE if the loop should be unswitched -- clone
557 // loop with an invariant test
558 bool policy_unswitching( PhaseIdealLoop *phase ) const;
559
560 // Micro-benchmark spamming. Remove empty loops.
561 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
562
563 // Convert one iteration loop into normal code.
564 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
565
566 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
567 // make some loop-invariant test (usually a null-check) happen before the
568 // loop.
569 bool policy_peeling( PhaseIdealLoop *phase ) const;
570
571 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
572 // known trip count in the counted loop node.
573 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
574
575 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
576 // the loop is a CountedLoop and the body is small enough.
577 bool policy_unroll(PhaseIdealLoop *phase);
578
579 // Loop analyses to map to a maximal superword unrolling for vectorization.
580 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
581
582 // Return TRUE or FALSE if the loop should be range-check-eliminated.
583 // Gather a list of IF tests that are dominated by iteration splitting;
584 // also gather the end of the first split and the start of the 2nd split.
585 bool policy_range_check( PhaseIdealLoop *phase ) const;
586
587 // Return TRUE or FALSE if the loop should be cache-line aligned.
588 // Gather the expression that does the alignment. Note that only
589 // one array base can be aligned in a loop (unless the VM guarantees
590 // mutual alignment). Note that if we vectorize short memory ops
591 // into longer memory ops, we may want to increase alignment.
592 bool policy_align( PhaseIdealLoop *phase ) const;
593
594 // Return TRUE if "iff" is a range check.
595 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
596
597 // Compute loop trip count if possible
598 void compute_trip_count(PhaseIdealLoop* phase);
599
600 // Compute loop trip count from profile data
601 float compute_profile_trip_cnt_helper(Node* n);
602 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
603
604 // Reassociate invariant expressions.
605 void reassociate_invariants(PhaseIdealLoop *phase);
606 // Reassociate invariant add and subtract expressions.
607 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
608 // Return nonzero index of invariant operand if invariant and variant
609 // are combined with an Add or Sub. Helper for reassociate_invariants.
610 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
611
612 // Return true if n is invariant
613 bool is_invariant(Node* n) const;
614
615 // Put loop body on igvn work list
616 void record_for_igvn();
617
618 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
619 bool is_inner() { return is_loop() && _child == NULL; }
620 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
621
622 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
623
624 #ifndef PRODUCT
625 void dump_head( ) const; // Dump loop head only
626 void dump() const; // Dump this loop recursively
627 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
628 #endif
629
630 };
631
632 // -----------------------------PhaseIdealLoop---------------------------------
633 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
634 // loop tree. Drives the loop-based transformations on the ideal graph.
635 class PhaseIdealLoop : public PhaseTransform {
636 friend class IdealLoopTree;
637 friend class SuperWord;
638 friend class CountedLoopReserveKit;
639
640 // Pre-computed def-use info
641 PhaseIterGVN &_igvn;
642
643 // Head of loop tree
644 IdealLoopTree *_ltree_root;
645
646 // Array of pre-order numbers, plus post-visited bit.
647 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
648 // ODD for post-visited. Other bits are the pre-order number.
649 uint *_preorders;
650 uint _max_preorder;
651
652 const PhaseIdealLoop* _verify_me;
653 bool _verify_only;
654
655 // Allocate _preorders[] array
656 void allocate_preorders() {
657 _max_preorder = C->unique()+8;
658 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
659 memset(_preorders, 0, sizeof(uint) * _max_preorder);
660 }
661
662 // Allocate _preorders[] array
663 void reallocate_preorders() {
664 if ( _max_preorder < C->unique() ) {
665 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
666 _max_preorder = C->unique();
667 }
668 memset(_preorders, 0, sizeof(uint) * _max_preorder);
669 }
670
671 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
672 // adds new nodes.
673 void check_grow_preorders( ) {
674 if ( _max_preorder < C->unique() ) {
675 uint newsize = _max_preorder<<1; // double size of array
676 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
677 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
678 _max_preorder = newsize;
679 }
680 }
681 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
682 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
683 // Pre-order numbers are written to the Nodes array as low-bit-set values.
684 void set_preorder_visited( Node *n, int pre_order ) {
685 assert( !is_visited( n ), "already set" );
686 _preorders[n->_idx] = (pre_order<<1);
687 };
688 // Return pre-order number.
689 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
690
691 // Check for being post-visited.
692 // Should be previsited already (checked with assert(is_visited(n))).
693 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
694
695 // Mark as post visited
696 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
697
698 public:
699 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
700 // Returns true if "n" is a data node, false if it's a control node.
701 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
702
703 private:
704 // clear out dead code after build_loop_late
705 Node_List _deadlist;
706
707 // Support for faster execution of get_late_ctrl()/dom_lca()
708 // when a node has many uses and dominator depth is deep.
709 Node_Array _dom_lca_tags;
710 void init_dom_lca_tags();
711 void clear_dom_lca_tags();
712
713 // Helper for debugging bad dominance relationships
714 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
715
716 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
717
718 // Inline wrapper for frequent cases:
719 // 1) only one use
720 // 2) a use is the same as the current LCA passed as 'n1'
721 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
722 assert( n->is_CFG(), "" );
723 // Fast-path NULL lca
724 if( lca != NULL && lca != n ) {
725 assert( lca->is_CFG(), "" );
726 // find LCA of all uses
727 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
728 }
729 return find_non_split_ctrl(n);
730 }
731 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
732
733 // Helper function for directing control inputs away from CFG split
734 // points.
735 Node *find_non_split_ctrl( Node *ctrl ) const {
736 if (ctrl != NULL) {
737 if (ctrl->is_MultiBranch()) {
738 ctrl = ctrl->in(0);
739 }
740 assert(ctrl->is_CFG(), "CFG");
741 }
742 return ctrl;
743 }
744
745 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
746
747 #ifdef ASSERT
748 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
749 #endif
750 void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* start, Node* end, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
751 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
752 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
753 void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* start, Node* end, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
754 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
755 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
756 Node* clone_skeleton_predicate(Node* iff, Node* value, Node* predicate, Node* uncommon_proj,
757 Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj);
758 bool skeleton_predicate_has_opaque(IfNode* iff);
759 void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
760 void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol);
761
762 public:
763
764 PhaseIterGVN &igvn() const { return _igvn; }
765
766 static bool is_canonical_loop_entry(CountedLoopNode* cl);
767
768 bool has_node( Node* n ) const {
769 guarantee(n != NULL, "No Node.");
770 return _nodes[n->_idx] != NULL;
771 }
772 // check if transform created new nodes that need _ctrl recorded
773 Node *get_late_ctrl( Node *n, Node *early );
774 Node *get_early_ctrl( Node *n );
775 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
776 void set_early_ctrl( Node *n );
777 void set_subtree_ctrl( Node *root );
778 void set_ctrl( Node *n, Node *ctrl ) {
779 assert( !has_node(n) || has_ctrl(n), "" );
780 assert( ctrl->in(0), "cannot set dead control node" );
781 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
782 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
783 }
784 // Set control and update loop membership
785 void set_ctrl_and_loop(Node* n, Node* ctrl) {
786 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
787 IdealLoopTree* new_loop = get_loop(ctrl);
788 if (old_loop != new_loop) {
789 if (old_loop->_child == NULL) old_loop->_body.yank(n);
790 if (new_loop->_child == NULL) new_loop->_body.push(n);
791 }
792 set_ctrl(n, ctrl);
793 }
794 // Control nodes can be replaced or subsumed. During this pass they
795 // get their replacement Node in slot 1. Instead of updating the block
796 // location of all Nodes in the subsumed block, we lazily do it. As we
797 // pull such a subsumed block out of the array, we write back the final
798 // correct block.
799 Node *get_ctrl( Node *i ) {
800 assert(has_node(i), "");
801 Node *n = get_ctrl_no_update(i);
802 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
803 assert(has_node(i) && has_ctrl(i), "");
804 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
805 return n;
806 }
807 // true if CFG node d dominates CFG node n
808 bool is_dominator(Node *d, Node *n);
809 // return get_ctrl for a data node and self(n) for a CFG node
810 Node* ctrl_or_self(Node* n) {
811 if (has_ctrl(n))
812 return get_ctrl(n);
813 else {
814 assert (n->is_CFG(), "must be a CFG node");
815 return n;
816 }
817 }
818
819 Node *get_ctrl_no_update_helper(Node *i) const {
820 assert(has_ctrl(i), "should be control, not loop");
821 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
822 }
823
824 Node *get_ctrl_no_update(Node *i) const {
825 assert( has_ctrl(i), "" );
826 Node *n = get_ctrl_no_update_helper(i);
827 if (!n->in(0)) {
828 // Skip dead CFG nodes
829 do {
830 n = get_ctrl_no_update_helper(n);
831 } while (!n->in(0));
832 n = find_non_split_ctrl(n);
833 }
834 return n;
835 }
836
837 // Check for loop being set
838 // "n" must be a control node. Returns true if "n" is known to be in a loop.
839 bool has_loop( Node *n ) const {
840 assert(!has_node(n) || !has_ctrl(n), "");
841 return has_node(n);
842 }
843 // Set loop
844 void set_loop( Node *n, IdealLoopTree *loop ) {
845 _nodes.map(n->_idx, (Node*)loop);
846 }
847 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
848 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
849 // from old_node to new_node to support the lazy update. Reference
850 // replaces loop reference, since that is not needed for dead node.
851 void lazy_update(Node *old_node, Node *new_node) {
852 assert(old_node != new_node, "no cycles please");
853 // Re-use the side array slot for this node to provide the
854 // forwarding pointer.
855 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
856 }
857 void lazy_replace(Node *old_node, Node *new_node) {
858 _igvn.replace_node(old_node, new_node);
859 lazy_update(old_node, new_node);
860 }
861
862 private:
863
864 // Place 'n' in some loop nest, where 'n' is a CFG node
865 void build_loop_tree();
866 int build_loop_tree_impl( Node *n, int pre_order );
867 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
868 // loop tree, not the root.
869 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
870
871 // Place Data nodes in some loop nest
872 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
873 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
874 void build_loop_late_post ( Node* n );
875 void verify_strip_mined_scheduling(Node *n, Node* least);
876
877 // Array of immediate dominance info for each CFG node indexed by node idx
878 private:
879 uint _idom_size;
880 Node **_idom; // Array of immediate dominators
881 uint *_dom_depth; // Used for fast LCA test
882 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
883
884 public:
885 Node* idom_no_update(Node* d) const {
886 return idom_no_update(d->_idx);
887 }
888
889 Node* idom_no_update(uint didx) const {
890 assert(didx < _idom_size, "oob");
891 Node* n = _idom[didx];
892 assert(n != NULL,"Bad immediate dominator info.");
893 while (n->in(0) == NULL) { // Skip dead CFG nodes
894 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
895 assert(n != NULL,"Bad immediate dominator info.");
896 }
897 return n;
898 }
899
900 Node *idom(Node* d) const {
901 return idom(d->_idx);
902 }
903
904 Node *idom(uint didx) const {
905 Node *n = idom_no_update(didx);
906 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
907 return n;
908 }
909
910 uint dom_depth(Node* d) const {
911 guarantee(d != NULL, "Null dominator info.");
912 guarantee(d->_idx < _idom_size, "");
913 return _dom_depth[d->_idx];
914 }
915 void set_idom(Node* d, Node* n, uint dom_depth);
916 // Locally compute IDOM using dom_lca call
917 Node *compute_idom( Node *region ) const;
918 // Recompute dom_depth
919 void recompute_dom_depth();
920
921 // Is safept not required by an outer loop?
922 bool is_deleteable_safept(Node* sfpt);
923
924 // Replace parallel induction variable (parallel to trip counter)
925 void replace_parallel_iv(IdealLoopTree *loop);
926
927 // Perform verification that the graph is valid.
928 PhaseIdealLoop( PhaseIterGVN &igvn) :
929 PhaseTransform(Ideal_Loop),
930 _igvn(igvn),
931 _dom_lca_tags(arena()), // Thread::resource_area
932 _verify_me(NULL),
933 _verify_only(true) {
934 build_and_optimize(LoopOptsVerify);
935 }
936
937 // build the loop tree and perform any requested optimizations
938 void build_and_optimize(LoopOptsMode mode);
939
940 // Dominators for the sea of nodes
941 void Dominators();
942 Node *dom_lca( Node *n1, Node *n2 ) const {
943 return find_non_split_ctrl(dom_lca_internal(n1, n2));
944 }
945 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
946
947 // Compute the Ideal Node to Loop mapping
948 PhaseIdealLoop(PhaseIterGVN &igvn, LoopOptsMode mode) :
949 PhaseTransform(Ideal_Loop),
950 _igvn(igvn),
951 _dom_lca_tags(arena()), // Thread::resource_area
952 _verify_me(NULL),
953 _verify_only(false) {
954 build_and_optimize(mode);
955 }
956
957 // Verify that verify_me made the same decisions as a fresh run.
958 PhaseIdealLoop(PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
959 PhaseTransform(Ideal_Loop),
960 _igvn(igvn),
961 _dom_lca_tags(arena()), // Thread::resource_area
962 _verify_me(verify_me),
963 _verify_only(false) {
964 build_and_optimize(LoopOptsVerify);
965 }
966
967 // Build and verify the loop tree without modifying the graph. This
968 // is useful to verify that all inputs properly dominate their uses.
969 static void verify(PhaseIterGVN& igvn) {
970 #ifdef ASSERT
971 PhaseIdealLoop v(igvn);
972 #endif
973 }
974
975 // True if the method has at least 1 irreducible loop
976 bool _has_irreducible_loops;
977
978 // Per-Node transform
979 virtual Node *transform( Node *a_node ) { return 0; }
980
981 bool is_counted_loop(Node* x, IdealLoopTree*& loop);
982 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
983 IdealLoopTree* loop, float cl_prob, float le_fcnt,
984 Node*& entry_control, Node*& iffalse);
985
986 Node* exact_limit( IdealLoopTree *loop );
987
988 // Return a post-walked LoopNode
989 IdealLoopTree *get_loop( Node *n ) const {
990 // Dead nodes have no loop, so return the top level loop instead
991 if (!has_node(n)) return _ltree_root;
992 assert(!has_ctrl(n), "");
993 return (IdealLoopTree*)_nodes[n->_idx];
994 }
995
996 IdealLoopTree *ltree_root() const { return _ltree_root; }
997
998 // Is 'n' a (nested) member of 'loop'?
999 int is_member( const IdealLoopTree *loop, Node *n ) const {
1000 return loop->is_member(get_loop(n)); }
1001
1002 // This is the basic building block of the loop optimizations. It clones an
1003 // entire loop body. It makes an old_new loop body mapping; with this
1004 // mapping you can find the new-loop equivalent to an old-loop node. All
1005 // new-loop nodes are exactly equal to their old-loop counterparts, all
1006 // edges are the same. All exits from the old-loop now have a RegionNode
1007 // that merges the equivalent new-loop path. This is true even for the
1008 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1009 // now come from (one or more) Phis that merge their new-loop equivalents.
1010 // Parameter side_by_side_idom:
1011 // When side_by_size_idom is NULL, the dominator tree is constructed for
1012 // the clone loop to dominate the original. Used in construction of
1013 // pre-main-post loop sequence.
1014 // When nonnull, the clone and original are side-by-side, both are
1015 // dominated by the passed in side_by_side_idom node. Used in
1016 // construction of unswitched loops.
1017 enum CloneLoopMode {
1018 IgnoreStripMined = 0, // Only clone inner strip mined loop
1019 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1020 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1021 // result control flow branches
1022 // either to inner clone or outer
1023 // strip mined loop.
1024 };
1025 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1026 CloneLoopMode mode, Node* side_by_side_idom = NULL);
1027 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1028 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1029 Node_List*& split_if_set, Node_List*& split_bool_set,
1030 Node_List*& split_cex_set, Node_List& worklist,
1031 uint new_counter, CloneLoopMode mode);
1032 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1033 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1034 Node_List& extra_data_nodes);
1035
1036 // If we got the effect of peeling, either by actually peeling or by
1037 // making a pre-loop which must execute at least once, we can remove
1038 // all loop-invariant dominated tests in the main body.
1039 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1040
1041 // Generate code to do a loop peel for the given loop (and body).
1042 // old_new is a temp array.
1043 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1044
1045 // Add pre and post loops around the given loop. These loops are used
1046 // during RCE, unrolling and aligning loops.
1047 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1048
1049 // Add post loop after the given loop.
1050 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1051 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1052 Node *incr, Node *limit, CountedLoopNode *&post_head);
1053
1054 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1055 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1056
1057 // Add a vector post loop between a vector main loop and the current post loop
1058 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1059 // If Node n lives in the back_ctrl block, we clone a private version of n
1060 // in preheader_ctrl block and return that, otherwise return n.
1061 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1062
1063 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1064 // unroll to do double iterations. The next round of major loop transforms
1065 // will repeat till the doubled loop body does all remaining iterations in 1
1066 // pass.
1067 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1068
1069 // Unroll the loop body one step - make each trip do 2 iterations.
1070 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1071
1072 // Mark vector reduction candidates before loop unrolling
1073 void mark_reductions( IdealLoopTree *loop );
1074
1075 // Return true if exp is a constant times an induction var
1076 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1077
1078 // Return true if exp is a scaled induction var plus (or minus) constant
1079 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1080
1081 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1082 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1083 int opcode, bool if_cont_is_true_proj = true);
1084
1085 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
1086
1087 static Node* skip_all_loop_predicates(Node* entry);
1088 static Node* skip_loop_predicates(Node* entry);
1089
1090 // Find a good location to insert a predicate
1091 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1092 // Find a predicate
1093 static Node* find_predicate(Node* entry);
1094 // Construct a range check for a predicate if
1095 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1096 int scale, Node* offset,
1097 Node* init, Node* limit, jint stride,
1098 Node* range, bool upper, bool &overflow);
1099
1100 // Implementation of the loop predication to promote checks outside the loop
1101 bool loop_predication_impl(IdealLoopTree *loop);
1102 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1103 CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1104 Deoptimization::DeoptReason reason);
1105 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt);
1106 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1107 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1108 Node_List& if_proj_list);
1109 ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1110 ProjNode* proj, ProjNode *predicate_proj,
1111 ProjNode* upper_bound_proj,
1112 int scale, Node* offset,
1113 Node* init, Node* limit, jint stride,
1114 Node* rng, bool& overflow,
1115 Deoptimization::DeoptReason reason);
1116 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1117 Node* predicate_proj, int scale_con, Node* offset,
1118 Node* limit, jint stride_con, Node* value);
1119
1120 // Helper function to collect predicate for eliminating the useless ones
1121 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1122 void eliminate_useless_predicates();
1123
1124 // Change the control input of expensive nodes to allow commoning by
1125 // IGVN when it is guaranteed to not result in a more frequent
1126 // execution of the expensive node. Return true if progress.
1127 bool process_expensive_nodes();
1128
1129 // Check whether node has become unreachable
1130 bool is_node_unreachable(Node *n) const {
1131 return !has_node(n) || n->is_unreachable(_igvn);
1132 }
1133
1134 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1135 int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1136
1137 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1138 void has_range_checks(IdealLoopTree *loop);
1139
1140 // Process post loops which have range checks and try to build a multi-version
1141 // guard to safely determine if we can execute the post loop which was RCE'd.
1142 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1143
1144 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1145 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1146
1147 // Create a slow version of the loop by cloning the loop
1148 // and inserting an if to select fast-slow versions.
1149 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1150 Node_List &old_new,
1151 int opcode,
1152 CloneLoopMode mode);
1153
1154 // Clone a loop and return the clone head (clone_loop_head).
1155 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1156 // This routine was created for usage in CountedLoopReserveKit.
1157 //
1158 // int(1) -> If -> IfTrue -> original_loop_head
1159 // |
1160 // V
1161 // IfFalse -> clone_loop_head (returned by function pointer)
1162 //
1163 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1164 // Clone loop with an invariant test (that does not exit) and
1165 // insert a clone of the test that selects which version to
1166 // execute.
1167 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1168
1169 // Find candidate "if" for unswitching
1170 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1171
1172 // Range Check Elimination uses this function!
1173 // Constrain the main loop iterations so the affine function:
1174 // low_limit <= scale_con * I + offset < upper_limit
1175 // always holds true. That is, either increase the number of iterations in
1176 // the pre-loop or the post-loop until the condition holds true in the main
1177 // loop. Scale_con, offset and limit are all loop invariant.
1178 void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
1179 // Helper function for add_constraint().
1180 Node* adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up);
1181
1182 // Partially peel loop up through last_peel node.
1183 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1184
1185 // Create a scheduled list of nodes control dependent on ctrl set.
1186 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1187 // Has a use in the vector set
1188 bool has_use_in_set( Node* n, VectorSet& vset );
1189 // Has use internal to the vector set (ie. not in a phi at the loop head)
1190 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1191 // clone "n" for uses that are outside of loop
1192 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1193 // clone "n" for special uses that are in the not_peeled region
1194 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1195 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1196 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1197 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1198 #ifdef ASSERT
1199 // Validate the loop partition sets: peel and not_peel
1200 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1201 // Ensure that uses outside of loop are of the right form
1202 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1203 uint orig_exit_idx, uint clone_exit_idx);
1204 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1205 #endif
1206
1207 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1208 int stride_of_possible_iv( Node* iff );
1209 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1210 // Return the (unique) control output node that's in the loop (if it exists.)
1211 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1212 // Insert a signed compare loop exit cloned from an unsigned compare.
1213 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1214 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1215 // Utility to register node "n" with PhaseIdealLoop
1216 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1217 // Utility to create an if-projection
1218 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1219 // Force the iff control output to be the live_proj
1220 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1221 // Insert a region before an if projection
1222 RegionNode* insert_region_before_proj(ProjNode* proj);
1223 // Insert a new if before an if projection
1224 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1225
1226 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1227 // "Nearly" because all Nodes have been cloned from the original in the loop,
1228 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1229 // through the Phi recursively, and return a Bool.
1230 Node *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1231 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1232
1233
1234 // Rework addressing expressions to get the most loop-invariant stuff
1235 // moved out. We'd like to do all associative operators, but it's especially
1236 // important (common) to do address expressions.
1237 Node *remix_address_expressions( Node *n );
1238
1239 // Attempt to use a conditional move instead of a phi/branch
1240 Node *conditional_move( Node *n );
1241
1242 // Reorganize offset computations to lower register pressure.
1243 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1244 // (which are then alive with the post-incremented trip counter
1245 // forcing an extra register move)
1246 void reorg_offsets( IdealLoopTree *loop );
1247
1248 // Check for aggressive application of 'split-if' optimization,
1249 // using basic block level info.
1250 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack, bool last_round );
1251 Node *split_if_with_blocks_pre ( Node *n );
1252 void split_if_with_blocks_post( Node *n, bool last_round );
1253 Node *has_local_phi_input( Node *n );
1254 // Mark an IfNode as being dominated by a prior test,
1255 // without actually altering the CFG (and hence IDOM info).
1256 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1257
1258 // Split Node 'n' through merge point
1259 Node *split_thru_region( Node *n, Node *region );
1260 // Split Node 'n' through merge point if there is enough win.
1261 Node *split_thru_phi( Node *n, Node *region, int policy );
1262 // Found an If getting its condition-code input from a Phi in the
1263 // same block. Split thru the Region.
1264 void do_split_if( Node *iff );
1265
1266 // Conversion of fill/copy patterns into intrisic versions
1267 bool do_intrinsify_fill();
1268 bool intrinsify_fill(IdealLoopTree* lpt);
1269 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1270 Node*& shift, Node*& offset);
1271
1272 private:
1273 // Return a type based on condition control flow
1274 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1275 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1276 // Helpers for filtered type
1277 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1278
1279 // Helper functions
1280 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1281 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1282 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 );
1283 bool split_up( Node *n, Node *blk1, Node *blk2 );
1284 void sink_use( Node *use, Node *post_loop );
1285 Node *place_near_use( Node *useblock ) const;
1286 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1287 void try_move_store_after_loop(Node* n);
1288 bool identical_backtoback_ifs(Node *n);
1289 bool can_split_if(Node *n_ctrl);
1290
1291 // Clone loop predicates to slow and fast loop when unswitching a loop
1292 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check, bool is_slow_loop,
1293 uint idx_before_clone, Node_List &old_new);
1294 ProjNode* clone_loop_predicate(ProjNode* predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1295 bool is_slow_loop, uint idx_before_clone, Node_List &old_new);
1296 void clone_concrete_loop_predicates(Deoptimization::DeoptReason reason, ProjNode* old_predicate_proj,
1297 ProjNode* new_predicate_proj, bool is_slow_loop,
1298 uint idx_before_clone, Node_List &old_new);
1299
1300 bool _created_loop_node;
1301 public:
1302 void set_created_loop_node() { _created_loop_node = true; }
1303 bool created_loop_node() { return _created_loop_node; }
1304 void register_new_node( Node *n, Node *blk );
1305
1306 #ifdef ASSERT
1307 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1308 #endif
1309
1310 #ifndef PRODUCT
1311 void dump( ) const;
1312 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1313 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1314 void verify() const; // Major slow :-)
1315 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1316 IdealLoopTree *get_loop_idx(Node* n) const {
1317 // Dead nodes have no loop, so return the top level loop instead
1318 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1319 }
1320 // Print some stats
1321 static void print_statistics();
1322 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1323 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1324 #endif
1325 };
1326
1327 // This kit may be used for making of a reserved copy of a loop before this loop
1328 // goes under non-reversible changes.
1329 //
1330 // Function create_reserve() creates a reserved copy (clone) of the loop.
1331 // The reserved copy is created by calling
1332 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1333 // the original and reserved loops are connected in the outer graph.
1334 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1335 //
1336 // By default the reserved copy (clone) of the loop is created as dead code - it is
1337 // dominated in the outer loop by this node chain:
1338 // intcon(1)->If->IfFalse->reserved_copy.
1339 // The original loop is dominated by the the same node chain but IfTrue projection:
1340 // intcon(0)->If->IfTrue->original_loop.
1341 //
1342 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1343 // and the dtor, checks _use_new value.
1344 // If _use_new == false, it "switches" control to reserved copy of the loop
1345 // by simple replacing of node intcon(1) with node intcon(0).
1346 //
1347 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1348 //
1349 // void CountedLoopReserveKit_example()
1350 // {
1351 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1352 // if (DoReserveCopy && !lrk.has_reserved()) {
1353 // return; //failed to create reserved loop copy
1354 // }
1355 // ...
1356 // //something is wrong, switch to original loop
1357 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1358 // ...
1359 // //everything worked ok, return with the newly modified loop
1360 // lrk.use_new();
1361 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1362 // }
1363 //
1364 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1365 // the dtor will "switch to the original" loop.
1366 // NOTE. You you modify outside of the original loop this class is no help.
1367 //
1368 class CountedLoopReserveKit {
1369 private:
1370 PhaseIdealLoop* _phase;
1371 IdealLoopTree* _lpt;
1372 LoopNode* _lp;
1373 IfNode* _iff;
1374 LoopNode* _lp_reserved;
1375 bool _has_reserved;
1376 bool _use_new;
1377 const bool _active; //may be set to false in ctor, then the object is dummy
1378
1379 public:
1380 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1381 ~CountedLoopReserveKit();
1382 void use_new() {_use_new = true;}
1383 void set_iff(IfNode* x) {_iff = x;}
1384 bool has_reserved() const { return _active && _has_reserved;}
1385 private:
1386 bool create_reserve();
1387 };// class CountedLoopReserveKit
1388
1389 inline Node* IdealLoopTree::tail() {
1390 // Handle lazy update of _tail field
1391 Node *n = _tail;
1392 //while( !n->in(0) ) // Skip dead CFG nodes
1393 //n = n->in(1);
1394 if (n->in(0) == NULL)
1395 n = _phase->get_ctrl(n);
1396 _tail = n;
1397 return n;
1398 }
1399
1400
1401 // Iterate over the loop tree using a preorder, left-to-right traversal.
1402 //
1403 // Example that visits all counted loops from within PhaseIdealLoop
1404 //
1405 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1406 // IdealLoopTree* lpt = iter.current();
1407 // if (!lpt->is_counted()) continue;
1408 // ...
1409 class LoopTreeIterator : public StackObj {
1410 private:
1411 IdealLoopTree* _root;
1412 IdealLoopTree* _curnt;
1413
1414 public:
1415 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1416
1417 bool done() { return _curnt == NULL; } // Finished iterating?
1418
1419 void next(); // Advance to next loop tree
1420
1421 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1422 };
1423
1424 #endif // SHARE_VM_OPTO_LOOPNODE_HPP