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 PhaseIdealLoop;
42 class CountedLoopReserveKit;
43 class VectorSet;
44 class Invariance;
45 struct small_cache;
46
47 //
48 // I D E A L I Z E D L O O P S
49 //
50 // Idealized loops are the set of loops I perform more interesting
51 // transformations on, beyond simple hoisting.
52
53 //------------------------------LoopNode---------------------------------------
54 // Simple loop header. Fall in path on left, loop-back path on right.
55 class LoopNode : public RegionNode {
56 // Size is bigger to hold the flags. However, the flags do not change
57 // the semantics so it does not appear in the hash & cmp functions.
58 virtual uint size_of() const { return sizeof(*this); }
59 protected:
60 short _loop_flags;
61 // Names for flag bitfields
62 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
63 MainHasNoPreLoop=4,
64 HasExactTripCount=8,
65 InnerLoop=16,
66 PartialPeelLoop=32,
67 PartialPeelFailed=64,
68 HasReductions=128,
69 WasSlpAnalyzed=256,
70 PassedSlpAnalysis=512,
71 DoUnrollOnly=1024,
72 VectorizedLoop=2048,
73 HasAtomicPostLoop=4096,
74 HasRangeChecks=8192,
75 IsMultiversioned=16384,
76 StripMined=32768};
77 char _unswitch_count;
78 enum { _unswitch_max=3 };
79 char _postloop_flags;
80 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
81
82 public:
83 // Names for edge indices
84 enum { Self=0, EntryControl, LoopBackControl };
85
86 int is_inner_loop() const { return _loop_flags & InnerLoop; }
87 void set_inner_loop() { _loop_flags |= InnerLoop; }
88
89 int range_checks_present() const { return _loop_flags & HasRangeChecks; }
90 int is_multiversioned() const { return _loop_flags & IsMultiversioned; }
91 int is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
92 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
93 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
94 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
95 int is_strip_mined() const { return _loop_flags & StripMined; }
96
97 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
98 void mark_has_reductions() { _loop_flags |= HasReductions; }
99 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
100 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
101 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
102 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
103 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
104 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
105 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
106 void mark_strip_mined() { _loop_flags |= StripMined; }
107 void clear_strip_mined() { _loop_flags &= ~StripMined; }
108
109 int unswitch_max() { return _unswitch_max; }
110 int unswitch_count() { return _unswitch_count; }
111
112 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
113 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
114 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
115 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
116
117 void set_unswitch_count(int val) {
118 assert (val <= unswitch_max(), "too many unswitches");
119 _unswitch_count = val;
120 }
121
122 LoopNode(Node *entry, Node *backedge) : RegionNode(3), _loop_flags(0), _unswitch_count(0), _postloop_flags(0) {
123 init_class_id(Class_Loop);
124 init_req(EntryControl, entry);
125 init_req(LoopBackControl, backedge);
126 }
127
128 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
129 virtual int Opcode() const;
130 bool can_be_counted_loop(PhaseTransform* phase) const {
131 return req() == 3 && in(0) != NULL &&
132 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
133 in(2) != NULL && phase->type(in(2)) != Type::TOP;
134 }
135 bool is_valid_counted_loop() const;
136 #ifndef PRODUCT
137 virtual void dump_spec(outputStream *st) const;
138 #endif
139
140 void verify_strip_mined(int expect_skeleton) const;
141 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
142 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
143 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
144 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
145 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; }
146 };
147
148 //------------------------------Counted Loops----------------------------------
149 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
150 // path (and maybe some other exit paths). The trip-counter exit is always
151 // last in the loop. The trip-counter have to stride by a constant;
152 // the exit value is also loop invariant.
153
154 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
155 // CountedLoopNode has the incoming loop control and the loop-back-control
156 // which is always the IfTrue before the matching CountedLoopEndNode. The
157 // CountedLoopEndNode has an incoming control (possibly not the
158 // CountedLoopNode if there is control flow in the loop), the post-increment
159 // trip-counter value, and the limit. The trip-counter value is always of
160 // the form (Op old-trip-counter stride). The old-trip-counter is produced
161 // by a Phi connected to the CountedLoopNode. The stride is constant.
162 // The Op is any commutable opcode, including Add, Mul, Xor. The
163 // CountedLoopEndNode also takes in the loop-invariant limit value.
164
165 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
166 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
167 // via the old-trip-counter from the Op node.
168
169 //------------------------------CountedLoopNode--------------------------------
170 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
171 // inputs the incoming loop-start control and the loop-back control, so they
172 // act like RegionNodes. They also take in the initial trip counter, the
173 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
174 // produce a loop-body control and the trip counter value. Since
175 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
176
177 class CountedLoopNode : public LoopNode {
178 // Size is bigger to hold _main_idx. However, _main_idx does not change
179 // the semantics so it does not appear in the hash & cmp functions.
180 virtual uint size_of() const { return sizeof(*this); }
181
182 // For Pre- and Post-loops during debugging ONLY, this holds the index of
183 // the Main CountedLoop. Used to assert that we understand the graph shape.
184 node_idx_t _main_idx;
185
186 // Known trip count calculated by compute_exact_trip_count()
187 uint _trip_count;
188
189 // Expected trip count from profile data
190 float _profile_trip_cnt;
191
192 // Log2 of original loop bodies in unrolled loop
193 int _unrolled_count_log2;
194
195 // Node count prior to last unrolling - used to decide if
196 // unroll,optimize,unroll,optimize,... is making progress
197 int _node_count_before_unroll;
198
199 // If slp analysis is performed we record the maximum
200 // vector mapped unroll factor here
201 int _slp_maximum_unroll_factor;
202
203 public:
204 CountedLoopNode( Node *entry, Node *backedge )
205 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
206 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
207 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
208 init_class_id(Class_CountedLoop);
209 // Initialize _trip_count to the largest possible value.
210 // Will be reset (lower) if the loop's trip count is known.
211 }
212
213 virtual int Opcode() const;
214 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
215
216 Node *init_control() const { return in(EntryControl); }
217 Node *back_control() const { return in(LoopBackControl); }
218 CountedLoopEndNode *loopexit_or_null() const;
219 CountedLoopEndNode *loopexit() const;
220 Node *init_trip() const;
221 Node *stride() const;
222 int stride_con() const;
223 bool stride_is_con() const;
224 Node *limit() const;
225 Node *incr() const;
226 Node *phi() const;
227
228 // Match increment with optional truncation
229 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
230
231 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
232 // can run short a few iterations and may start a few iterations in.
233 // It will be RCE'd and unrolled and aligned.
234
235 // A following 'post' loop will run any remaining iterations. Used
236 // during Range Check Elimination, the 'post' loop will do any final
237 // iterations with full checks. Also used by Loop Unrolling, where
238 // the 'post' loop will do any epilog iterations needed. Basically,
239 // a 'post' loop can not profitably be further unrolled or RCE'd.
240
241 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
242 // it may do under-flow checks for RCE and may do alignment iterations
243 // so the following main loop 'knows' that it is striding down cache
244 // lines.
245
246 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
247 // Aligned, may be missing it's pre-loop.
248 int is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
249 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
250 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
251 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
252 int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
253 int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
254 int has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
255 int do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
256 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
257 int has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
258 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
259
260 int main_idx() const { return _main_idx; }
261
262
263 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
264 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
265 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
266 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
267
268 void set_trip_count(uint tc) { _trip_count = tc; }
269 uint trip_count() { return _trip_count; }
270
271 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
272 void set_exact_trip_count(uint tc) {
273 _trip_count = tc;
274 _loop_flags |= HasExactTripCount;
275 }
276 void set_nonexact_trip_count() {
277 _loop_flags &= ~HasExactTripCount;
278 }
279 void set_notpassed_slp() {
280 _loop_flags &= ~PassedSlpAnalysis;
281 }
282
283 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
284 float profile_trip_cnt() { return _profile_trip_cnt; }
285
286 void double_unrolled_count() { _unrolled_count_log2++; }
287 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
288
289 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
290 int node_count_before_unroll() { return _node_count_before_unroll; }
291 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
292 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
293
294 virtual LoopNode* skip_strip_mined(int expect_opaq = 1);
295 OuterStripMinedLoopNode* outer_loop() const;
296 virtual IfTrueNode* outer_loop_tail() const;
297 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
298 virtual IfFalseNode* outer_loop_exit() const;
299 virtual SafePointNode* outer_safepoint() const;
300
301 // If this is a main loop in a pre/main/post loop nest, walk over
302 // the predicates that were inserted by
303 // duplicate_predicates()/add_range_check_predicate()
304 Node* skip_predicates();
305
306 #ifndef PRODUCT
307 virtual void dump_spec(outputStream *st) const;
308 #endif
309 };
310
311 //------------------------------CountedLoopEndNode-----------------------------
312 // CountedLoopEndNodes end simple trip counted loops. They act much like
313 // IfNodes.
314 class CountedLoopEndNode : public IfNode {
315 public:
316 enum { TestControl, TestValue };
317
318 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
319 : IfNode( control, test, prob, cnt) {
320 init_class_id(Class_CountedLoopEnd);
321 }
322 virtual int Opcode() const;
323
324 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
325 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
326 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
327 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
328 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
329 int stride_con() const;
330 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
331 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
332 PhiNode *phi() const {
333 Node *tmp = incr();
334 if (tmp && tmp->req() == 3) {
335 Node* phi = tmp->in(1);
336 if (phi->is_Phi()) {
337 return phi->as_Phi();
338 }
339 }
340 return NULL;
341 }
342 CountedLoopNode *loopnode() const {
343 // The CountedLoopNode that goes with this CountedLoopEndNode may
344 // have been optimized out by the IGVN so be cautious with the
345 // pattern matching on the graph
346 PhiNode* iv_phi = phi();
347 if (iv_phi == NULL) {
348 return NULL;
349 }
350 Node *ln = iv_phi->in(0);
351 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) {
352 return (CountedLoopNode*)ln;
353 }
354 return NULL;
355 }
356
357 #ifndef PRODUCT
358 virtual void dump_spec(outputStream *st) const;
359 #endif
360 };
361
362
363 inline CountedLoopEndNode *CountedLoopNode::loopexit_or_null() const {
364 Node *bc = back_control();
365 if( bc == NULL ) return NULL;
366 Node *le = bc->in(0);
367 if( le->Opcode() != Op_CountedLoopEnd )
368 return NULL;
369 return (CountedLoopEndNode*)le;
370 }
371 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
372 CountedLoopEndNode* cle = loopexit_or_null();
373 assert(cle != NULL, "loopexit is NULL");
374 return cle;
375 }
376 inline Node *CountedLoopNode::init_trip() const { return loopexit_or_null() ? loopexit()->init_trip() : NULL; }
377 inline Node *CountedLoopNode::stride() const { return loopexit_or_null() ? loopexit()->stride() : NULL; }
378 inline int CountedLoopNode::stride_con() const { return loopexit_or_null() ? loopexit()->stride_con() : 0; }
379 inline bool CountedLoopNode::stride_is_con() const { return loopexit_or_null() && loopexit()->stride_is_con(); }
380 inline Node *CountedLoopNode::limit() const { return loopexit_or_null() ? loopexit()->limit() : NULL; }
381 inline Node *CountedLoopNode::incr() const { return loopexit_or_null() ? loopexit()->incr() : NULL; }
382 inline Node *CountedLoopNode::phi() const { return loopexit_or_null() ? loopexit()->phi() : NULL; }
383
384 //------------------------------LoopLimitNode-----------------------------
385 // Counted Loop limit node which represents exact final iterator value:
386 // trip_count = (limit - init_trip + stride - 1)/stride
387 // final_value= trip_count * stride + init_trip.
388 // Use HW instructions to calculate it when it can overflow in integer.
389 // Note, final_value should fit into integer since counted loop has
390 // limit check: limit <= max_int-stride.
391 class LoopLimitNode : public Node {
392 enum { Init=1, Limit=2, Stride=3 };
393 public:
394 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
395 // Put it on the Macro nodes list to optimize during macro nodes expansion.
396 init_flags(Flag_is_macro);
397 C->add_macro_node(this);
398 }
399 virtual int Opcode() const;
400 virtual const Type *bottom_type() const { return TypeInt::INT; }
401 virtual uint ideal_reg() const { return Op_RegI; }
402 virtual const Type* Value(PhaseGVN* phase) const;
403 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
404 virtual Node* Identity(PhaseGVN* phase);
405 };
406
407 // Support for strip mining
408 class OuterStripMinedLoopNode : public LoopNode {
409 private:
410 CountedLoopNode* inner_loop() const;
411 public:
412 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge)
413 : LoopNode(entry, backedge) {
414 init_class_id(Class_OuterStripMinedLoop);
415 init_flags(Flag_is_macro);
416 C->add_macro_node(this);
417 }
418
419 virtual int Opcode() const;
420
421 virtual IfTrueNode* outer_loop_tail() const;
422 virtual OuterStripMinedLoopEndNode* outer_loop_end() const;
423 virtual IfFalseNode* outer_loop_exit() const;
424 virtual SafePointNode* outer_safepoint() const;
425 void adjust_strip_mined_loop(PhaseIterGVN* igvn);
426 };
427
428 class OuterStripMinedLoopEndNode : public IfNode {
429 public:
430 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt)
431 : IfNode(control, test, prob, cnt) {
432 init_class_id(Class_OuterStripMinedLoopEnd);
433 }
434
435 virtual int Opcode() const;
436
437 virtual const Type* Value(PhaseGVN* phase) const;
438 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
439 };
440
441 // -----------------------------IdealLoopTree----------------------------------
442 class IdealLoopTree : public ResourceObj {
443 public:
444 IdealLoopTree *_parent; // Parent in loop tree
445 IdealLoopTree *_next; // Next sibling in loop tree
446 IdealLoopTree *_child; // First child in loop tree
447
448 // The head-tail backedge defines the loop.
449 // If tail is NULL then this loop has multiple backedges as part of the
450 // same loop. During cleanup I'll peel off the multiple backedges; merge
451 // them at the loop bottom and flow 1 real backedge into the loop.
452 Node *_head; // Head of loop
453 Node *_tail; // Tail of loop
454 inline Node *tail(); // Handle lazy update of _tail field
455 PhaseIdealLoop* _phase;
456 int _local_loop_unroll_limit;
457 int _local_loop_unroll_factor;
458
459 Node_List _body; // Loop body for inner loops
460
461 uint8_t _nest; // Nesting depth
462 uint8_t _irreducible:1, // True if irreducible
463 _has_call:1, // True if has call safepoint
464 _has_sfpt:1, // True if has non-call safepoint
465 _rce_candidate:1; // True if candidate for range check elimination
466
467 Node_List* _safepts; // List of safepoints in this loop
468 Node_List* _required_safept; // A inner loop cannot delete these safepts;
469 bool _allow_optimizations; // Allow loop optimizations
470
471 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
472 : _parent(0), _next(0), _child(0),
473 _head(head), _tail(tail),
474 _phase(phase),
475 _safepts(NULL),
476 _required_safept(NULL),
477 _allow_optimizations(true),
478 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
479 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
480 { }
481
482 // Is 'l' a member of 'this'?
483 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
484
485 // Set loop nesting depth. Accumulate has_call bits.
486 int set_nest( uint depth );
487
488 // Split out multiple fall-in edges from the loop header. Move them to a
489 // private RegionNode before the loop. This becomes the loop landing pad.
490 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
491
492 // Split out the outermost loop from this shared header.
493 void split_outer_loop( PhaseIdealLoop *phase );
494
495 // Merge all the backedges from the shared header into a private Region.
496 // Feed that region as the one backedge to this loop.
497 void merge_many_backedges( PhaseIdealLoop *phase );
498
499 // Split shared headers and insert loop landing pads.
500 // Insert a LoopNode to replace the RegionNode.
501 // Returns TRUE if loop tree is structurally changed.
502 bool beautify_loops( PhaseIdealLoop *phase );
503
504 // Perform optimization to use the loop predicates for null checks and range checks.
505 // Applies to any loop level (not just the innermost one)
506 bool loop_predication( PhaseIdealLoop *phase);
507
508 // Perform iteration-splitting on inner loops. Split iterations to
509 // avoid range checks or one-shot null checks. Returns false if the
510 // current round of loop opts should stop.
511 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
512
513 // Driver for various flavors of iteration splitting. Returns false
514 // if the current round of loop opts should stop.
515 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
516
517 // Given dominators, try to find loops with calls that must always be
518 // executed (call dominates loop tail). These loops do not need non-call
519 // safepoints (ncsfpt).
520 void check_safepts(VectorSet &visited, Node_List &stack);
521
522 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
523 // encountered.
524 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
525
526 // Remove safepoints from loop. Optionally keeping one.
527 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
528
529 // Convert to counted loops where possible
530 void counted_loop( PhaseIdealLoop *phase );
531
532 // Check for Node being a loop-breaking test
533 Node *is_loop_exit(Node *iff) const;
534
535 // Remove simplistic dead code from loop body
536 void DCE_loop_body();
537
538 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
539 // Replace with a 1-in-10 exit guess.
540 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
541
542 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
543 // Useful for unrolling loops with NO array accesses.
544 bool policy_peel_only( PhaseIdealLoop *phase ) const;
545
546 // Return TRUE or FALSE if the loop should be unswitched -- clone
547 // loop with an invariant test
548 bool policy_unswitching( PhaseIdealLoop *phase ) const;
549
550 // Micro-benchmark spamming. Remove empty loops.
551 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
552
553 // Convert one iteration loop into normal code.
554 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
555
556 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
557 // make some loop-invariant test (usually a null-check) happen before the
558 // loop.
559 bool policy_peeling( PhaseIdealLoop *phase ) const;
560
561 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
562 // known trip count in the counted loop node.
563 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
564
565 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
566 // the loop is a CountedLoop and the body is small enough.
567 bool policy_unroll(PhaseIdealLoop *phase);
568
569 // Loop analyses to map to a maximal superword unrolling for vectorization.
570 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
571
572 // Return TRUE or FALSE if the loop should be range-check-eliminated.
573 // Gather a list of IF tests that are dominated by iteration splitting;
574 // also gather the end of the first split and the start of the 2nd split.
575 bool policy_range_check( PhaseIdealLoop *phase ) const;
576
577 // Return TRUE or FALSE if the loop should be cache-line aligned.
578 // Gather the expression that does the alignment. Note that only
579 // one array base can be aligned in a loop (unless the VM guarantees
580 // mutual alignment). Note that if we vectorize short memory ops
581 // into longer memory ops, we may want to increase alignment.
582 bool policy_align( PhaseIdealLoop *phase ) const;
583
584 // Return TRUE if "iff" is a range check.
585 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
586
587 // Compute loop trip count if possible
588 void compute_trip_count(PhaseIdealLoop* phase);
589
590 // Compute loop trip count from profile data
591 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
592
593 // Reassociate invariant expressions.
594 void reassociate_invariants(PhaseIdealLoop *phase);
595 // Reassociate invariant add and subtract expressions.
596 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
597 // Return nonzero index of invariant operand if invariant and variant
598 // are combined with an Add or Sub. Helper for reassociate_invariants.
599 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
600
601 // Return true if n is invariant
602 bool is_invariant(Node* n) const;
603
604 // Put loop body on igvn work list
605 void record_for_igvn();
606
607 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
608 bool is_inner() { return is_loop() && _child == NULL; }
609 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
610
611 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
612
613 #ifndef PRODUCT
614 void dump_head( ) const; // Dump loop head only
615 void dump() const; // Dump this loop recursively
616 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
617 #endif
618
619 };
620
621 // -----------------------------PhaseIdealLoop---------------------------------
622 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
623 // loop tree. Drives the loop-based transformations on the ideal graph.
624 class PhaseIdealLoop : public PhaseTransform {
625 friend class IdealLoopTree;
626 friend class SuperWord;
627 friend class CountedLoopReserveKit;
628
629 // Pre-computed def-use info
630 PhaseIterGVN &_igvn;
631
632 // Head of loop tree
633 IdealLoopTree *_ltree_root;
634
635 // Array of pre-order numbers, plus post-visited bit.
636 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
637 // ODD for post-visited. Other bits are the pre-order number.
638 uint *_preorders;
639 uint _max_preorder;
640
641 const PhaseIdealLoop* _verify_me;
642 bool _verify_only;
643
644 // Allocate _preorders[] array
645 void allocate_preorders() {
646 _max_preorder = C->unique()+8;
647 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
648 memset(_preorders, 0, sizeof(uint) * _max_preorder);
649 }
650
651 // Allocate _preorders[] array
652 void reallocate_preorders() {
653 if ( _max_preorder < C->unique() ) {
654 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
655 _max_preorder = C->unique();
656 }
657 memset(_preorders, 0, sizeof(uint) * _max_preorder);
658 }
659
660 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
661 // adds new nodes.
662 void check_grow_preorders( ) {
663 if ( _max_preorder < C->unique() ) {
664 uint newsize = _max_preorder<<1; // double size of array
665 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
666 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
667 _max_preorder = newsize;
668 }
669 }
670 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
671 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
672 // Pre-order numbers are written to the Nodes array as low-bit-set values.
673 void set_preorder_visited( Node *n, int pre_order ) {
674 assert( !is_visited( n ), "already set" );
675 _preorders[n->_idx] = (pre_order<<1);
676 };
677 // Return pre-order number.
678 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
679
680 // Check for being post-visited.
681 // Should be previsited already (checked with assert(is_visited(n))).
682 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
683
684 // Mark as post visited
685 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
686
687 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
688 // Returns true if "n" is a data node, false if it's a control node.
689 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
690
691 // clear out dead code after build_loop_late
692 Node_List _deadlist;
693
694 // Support for faster execution of get_late_ctrl()/dom_lca()
695 // when a node has many uses and dominator depth is deep.
696 Node_Array _dom_lca_tags;
697 void init_dom_lca_tags();
698 void clear_dom_lca_tags();
699
700 // Helper for debugging bad dominance relationships
701 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
702
703 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
704
705 // Inline wrapper for frequent cases:
706 // 1) only one use
707 // 2) a use is the same as the current LCA passed as 'n1'
708 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
709 assert( n->is_CFG(), "" );
710 // Fast-path NULL lca
711 if( lca != NULL && lca != n ) {
712 assert( lca->is_CFG(), "" );
713 // find LCA of all uses
714 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
715 }
716 return find_non_split_ctrl(n);
717 }
718 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
719
720 // Helper function for directing control inputs away from CFG split
721 // points.
722 Node *find_non_split_ctrl( Node *ctrl ) const {
723 if (ctrl != NULL) {
724 if (ctrl->is_MultiBranch()) {
725 ctrl = ctrl->in(0);
726 }
727 assert(ctrl->is_CFG(), "CFG");
728 }
729 return ctrl;
730 }
731
732 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
733 void duplicate_predicates(CountedLoopNode* pre_head, Node *min_taken, Node* castii,
734 IdealLoopTree* outer_loop, LoopNode* outer_main_head,
735 uint dd_main_head);
736
737 public:
738
739 static bool is_canonical_loop_entry(CountedLoopNode* cl);
740
741 bool has_node( Node* n ) const {
742 guarantee(n != NULL, "No Node.");
743 return _nodes[n->_idx] != NULL;
744 }
745 // check if transform created new nodes that need _ctrl recorded
746 Node *get_late_ctrl( Node *n, Node *early );
747 Node *get_early_ctrl( Node *n );
748 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
749 void set_early_ctrl( Node *n );
750 void set_subtree_ctrl( Node *root );
751 void set_ctrl( Node *n, Node *ctrl ) {
752 assert( !has_node(n) || has_ctrl(n), "" );
753 assert( ctrl->in(0), "cannot set dead control node" );
754 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
755 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
756 }
757 // Set control and update loop membership
758 void set_ctrl_and_loop(Node* n, Node* ctrl) {
759 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
760 IdealLoopTree* new_loop = get_loop(ctrl);
761 if (old_loop != new_loop) {
762 if (old_loop->_child == NULL) old_loop->_body.yank(n);
763 if (new_loop->_child == NULL) new_loop->_body.push(n);
764 }
765 set_ctrl(n, ctrl);
766 }
767 // Control nodes can be replaced or subsumed. During this pass they
768 // get their replacement Node in slot 1. Instead of updating the block
769 // location of all Nodes in the subsumed block, we lazily do it. As we
770 // pull such a subsumed block out of the array, we write back the final
771 // correct block.
772 Node *get_ctrl( Node *i ) {
773 assert(has_node(i), "");
774 Node *n = get_ctrl_no_update(i);
775 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
776 assert(has_node(i) && has_ctrl(i), "");
777 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
778 return n;
779 }
780 // true if CFG node d dominates CFG node n
781 bool is_dominator(Node *d, Node *n);
782 // return get_ctrl for a data node and self(n) for a CFG node
783 Node* ctrl_or_self(Node* n) {
784 if (has_ctrl(n))
785 return get_ctrl(n);
786 else {
787 assert (n->is_CFG(), "must be a CFG node");
788 return n;
789 }
790 }
791
792 private:
793 Node *get_ctrl_no_update_helper(Node *i) const {
794 assert(has_ctrl(i), "should be control, not loop");
795 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
796 }
797
798 Node *get_ctrl_no_update(Node *i) const {
799 assert( has_ctrl(i), "" );
800 Node *n = get_ctrl_no_update_helper(i);
801 if (!n->in(0)) {
802 // Skip dead CFG nodes
803 do {
804 n = get_ctrl_no_update_helper(n);
805 } while (!n->in(0));
806 n = find_non_split_ctrl(n);
807 }
808 return n;
809 }
810
811 // Check for loop being set
812 // "n" must be a control node. Returns true if "n" is known to be in a loop.
813 bool has_loop( Node *n ) const {
814 assert(!has_node(n) || !has_ctrl(n), "");
815 return has_node(n);
816 }
817 // Set loop
818 void set_loop( Node *n, IdealLoopTree *loop ) {
819 _nodes.map(n->_idx, (Node*)loop);
820 }
821 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
822 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
823 // from old_node to new_node to support the lazy update. Reference
824 // replaces loop reference, since that is not needed for dead node.
825 public:
826 void lazy_update(Node *old_node, Node *new_node) {
827 assert(old_node != new_node, "no cycles please");
828 // Re-use the side array slot for this node to provide the
829 // forwarding pointer.
830 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
831 }
832 void lazy_replace(Node *old_node, Node *new_node) {
833 _igvn.replace_node(old_node, new_node);
834 lazy_update(old_node, new_node);
835 }
836
837 private:
838
839 // Place 'n' in some loop nest, where 'n' is a CFG node
840 void build_loop_tree();
841 int build_loop_tree_impl( Node *n, int pre_order );
842 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
843 // loop tree, not the root.
844 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
845
846 // Place Data nodes in some loop nest
847 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
848 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
849 void build_loop_late_post ( Node* n );
850 void verify_strip_mined_scheduling(Node *n, Node* least);
851
852 // Array of immediate dominance info for each CFG node indexed by node idx
853 private:
854 uint _idom_size;
855 Node **_idom; // Array of immediate dominators
856 uint *_dom_depth; // Used for fast LCA test
857 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
858
859 Node* idom_no_update(Node* d) const {
860 assert(d->_idx < _idom_size, "oob");
861 Node* n = _idom[d->_idx];
862 assert(n != NULL,"Bad immediate dominator info.");
863 while (n->in(0) == NULL) { // Skip dead CFG nodes
864 //n = n->in(1);
865 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
866 assert(n != NULL,"Bad immediate dominator info.");
867 }
868 return n;
869 }
870 Node *idom(Node* d) const {
871 uint didx = d->_idx;
872 Node *n = idom_no_update(d);
873 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
874 return n;
875 }
876 uint dom_depth(Node* d) const {
877 guarantee(d != NULL, "Null dominator info.");
878 guarantee(d->_idx < _idom_size, "");
879 return _dom_depth[d->_idx];
880 }
881 void set_idom(Node* d, Node* n, uint dom_depth);
882 // Locally compute IDOM using dom_lca call
883 Node *compute_idom( Node *region ) const;
884 // Recompute dom_depth
885 void recompute_dom_depth();
886
887 // Is safept not required by an outer loop?
888 bool is_deleteable_safept(Node* sfpt);
889
890 // Replace parallel induction variable (parallel to trip counter)
891 void replace_parallel_iv(IdealLoopTree *loop);
892
893 // Perform verification that the graph is valid.
894 PhaseIdealLoop( PhaseIterGVN &igvn) :
895 PhaseTransform(Ideal_Loop),
896 _igvn(igvn),
897 _dom_lca_tags(arena()), // Thread::resource_area
898 _verify_me(NULL),
899 _verify_only(true) {
900 build_and_optimize(false, false);
901 }
902
903 // build the loop tree and perform any requested optimizations
904 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
905
906 public:
907 // Dominators for the sea of nodes
908 void Dominators();
909 Node *dom_lca( Node *n1, Node *n2 ) const {
910 return find_non_split_ctrl(dom_lca_internal(n1, n2));
911 }
912 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
913
914 // Compute the Ideal Node to Loop mapping
915 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
916 PhaseTransform(Ideal_Loop),
917 _igvn(igvn),
918 _dom_lca_tags(arena()), // Thread::resource_area
919 _verify_me(NULL),
920 _verify_only(false) {
921 build_and_optimize(do_split_ifs, skip_loop_opts);
922 }
923
924 // Verify that verify_me made the same decisions as a fresh run.
925 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
926 PhaseTransform(Ideal_Loop),
927 _igvn(igvn),
928 _dom_lca_tags(arena()), // Thread::resource_area
929 _verify_me(verify_me),
930 _verify_only(false) {
931 build_and_optimize(false, false);
932 }
933
934 // Build and verify the loop tree without modifying the graph. This
935 // is useful to verify that all inputs properly dominate their uses.
936 static void verify(PhaseIterGVN& igvn) {
937 #ifdef ASSERT
938 PhaseIdealLoop v(igvn);
939 #endif
940 }
941
942 // True if the method has at least 1 irreducible loop
943 bool _has_irreducible_loops;
944
945 // Per-Node transform
946 virtual Node *transform( Node *a_node ) { return 0; }
947
948 bool is_counted_loop(Node* x, IdealLoopTree*& loop);
949 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
950 IdealLoopTree* loop, float cl_prob, float le_fcnt,
951 Node*& entry_control, Node*& iffalse);
952
953 Node* exact_limit( IdealLoopTree *loop );
954
955 // Return a post-walked LoopNode
956 IdealLoopTree *get_loop( Node *n ) const {
957 // Dead nodes have no loop, so return the top level loop instead
958 if (!has_node(n)) return _ltree_root;
959 assert(!has_ctrl(n), "");
960 return (IdealLoopTree*)_nodes[n->_idx];
961 }
962
963 // Is 'n' a (nested) member of 'loop'?
964 int is_member( const IdealLoopTree *loop, Node *n ) const {
965 return loop->is_member(get_loop(n)); }
966
967 // This is the basic building block of the loop optimizations. It clones an
968 // entire loop body. It makes an old_new loop body mapping; with this
969 // mapping you can find the new-loop equivalent to an old-loop node. All
970 // new-loop nodes are exactly equal to their old-loop counterparts, all
971 // edges are the same. All exits from the old-loop now have a RegionNode
972 // that merges the equivalent new-loop path. This is true even for the
973 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
974 // now come from (one or more) Phis that merge their new-loop equivalents.
975 // Parameter side_by_side_idom:
976 // When side_by_size_idom is NULL, the dominator tree is constructed for
977 // the clone loop to dominate the original. Used in construction of
978 // pre-main-post loop sequence.
979 // When nonnull, the clone and original are side-by-side, both are
980 // dominated by the passed in side_by_side_idom node. Used in
981 // construction of unswitched loops.
982 enum CloneLoopMode {
983 IgnoreStripMined = 0, // Only clone inner strip mined loop
984 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
985 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
986 // result control flow branches
987 // either to inner clone or outer
988 // strip mined loop.
989 };
990 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
991 CloneLoopMode mode, Node* side_by_side_idom = NULL);
992 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
993 IdealLoopTree* loop, IdealLoopTree* companion_loop,
994 Node_List*& split_if_set, Node_List*& split_bool_set,
995 Node_List*& split_cex_set, Node_List& worklist,
996 uint new_counter, CloneLoopMode mode);
997 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
998 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
999 Node_List& extra_data_nodes);
1000
1001 // If we got the effect of peeling, either by actually peeling or by
1002 // making a pre-loop which must execute at least once, we can remove
1003 // all loop-invariant dominated tests in the main body.
1004 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1005
1006 // Generate code to do a loop peel for the given loop (and body).
1007 // old_new is a temp array.
1008 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1009
1010 // Add pre and post loops around the given loop. These loops are used
1011 // during RCE, unrolling and aligning loops.
1012 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1013
1014 // Add post loop after the given loop.
1015 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1016 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1017 Node *incr, Node *limit, CountedLoopNode *&post_head);
1018
1019 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1020 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1021
1022 // Add a vector post loop between a vector main loop and the current post loop
1023 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1024 // If Node n lives in the back_ctrl block, we clone a private version of n
1025 // in preheader_ctrl block and return that, otherwise return n.
1026 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1027
1028 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1029 // unroll to do double iterations. The next round of major loop transforms
1030 // will repeat till the doubled loop body does all remaining iterations in 1
1031 // pass.
1032 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1033
1034 // Unroll the loop body one step - make each trip do 2 iterations.
1035 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1036
1037 // Mark vector reduction candidates before loop unrolling
1038 void mark_reductions( IdealLoopTree *loop );
1039
1040 // Return true if exp is a constant times an induction var
1041 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1042
1043 // Return true if exp is a scaled induction var plus (or minus) constant
1044 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1045
1046 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1047 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
1048 Deoptimization::DeoptReason reason,
1049 int opcode);
1050 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
1051
1052 // Clone loop predicates to cloned loops (peeled, unswitched)
1053 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
1054 Deoptimization::DeoptReason reason,
1055 PhaseIdealLoop* loop_phase,
1056 PhaseIterGVN* igvn);
1057
1058 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
1059 bool clone_limit_check,
1060 PhaseIdealLoop* loop_phase,
1061 PhaseIterGVN* igvn);
1062 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
1063
1064 static Node* skip_loop_predicates(Node* entry);
1065
1066 // Find a good location to insert a predicate
1067 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1068 // Find a predicate
1069 static Node* find_predicate(Node* entry);
1070 // Construct a range check for a predicate if
1071 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1072 int scale, Node* offset,
1073 Node* init, Node* limit, jint stride,
1074 Node* range, bool upper, bool &overflow);
1075
1076 // Implementation of the loop predication to promote checks outside the loop
1077 bool loop_predication_impl(IdealLoopTree *loop);
1078 ProjNode* insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1079 ProjNode* proj, ProjNode *predicate_proj,
1080 ProjNode* upper_bound_proj,
1081 int scale, Node* offset,
1082 Node* init, Node* limit, jint stride,
1083 Node* rng, bool& overflow);
1084 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1085 Node* predicate_proj, int scale_con, Node* offset,
1086 Node* limit, jint stride_con);
1087
1088 // Helper function to collect predicate for eliminating the useless ones
1089 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1090 void eliminate_useless_predicates();
1091
1092 // Change the control input of expensive nodes to allow commoning by
1093 // IGVN when it is guaranteed to not result in a more frequent
1094 // execution of the expensive node. Return true if progress.
1095 bool process_expensive_nodes();
1096
1097 // Check whether node has become unreachable
1098 bool is_node_unreachable(Node *n) const {
1099 return !has_node(n) || n->is_unreachable(_igvn);
1100 }
1101
1102 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1103 int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1104
1105 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1106 void has_range_checks(IdealLoopTree *loop);
1107
1108 // Process post loops which have range checks and try to build a multi-version
1109 // guard to safely determine if we can execute the post loop which was RCE'd.
1110 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1111
1112 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1113 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1114
1115 // Create a slow version of the loop by cloning the loop
1116 // and inserting an if to select fast-slow versions.
1117 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1118 Node_List &old_new,
1119 int opcode,
1120 CloneLoopMode mode);
1121
1122 // Clone a loop and return the clone head (clone_loop_head).
1123 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1124 // This routine was created for usage in CountedLoopReserveKit.
1125 //
1126 // int(1) -> If -> IfTrue -> original_loop_head
1127 // |
1128 // V
1129 // IfFalse -> clone_loop_head (returned by function pointer)
1130 //
1131 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1132 // Clone loop with an invariant test (that does not exit) and
1133 // insert a clone of the test that selects which version to
1134 // execute.
1135 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1136
1137 // Find candidate "if" for unswitching
1138 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1139
1140 // Range Check Elimination uses this function!
1141 // Constrain the main loop iterations so the affine function:
1142 // low_limit <= scale_con * I + offset < upper_limit
1143 // always holds true. That is, either increase the number of iterations in
1144 // the pre-loop or the post-loop until the condition holds true in the main
1145 // loop. Scale_con, offset and limit are all loop invariant.
1146 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 );
1147 // Helper function for add_constraint().
1148 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1149
1150 // Partially peel loop up through last_peel node.
1151 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1152
1153 // Create a scheduled list of nodes control dependent on ctrl set.
1154 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1155 // Has a use in the vector set
1156 bool has_use_in_set( Node* n, VectorSet& vset );
1157 // Has use internal to the vector set (ie. not in a phi at the loop head)
1158 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1159 // clone "n" for uses that are outside of loop
1160 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1161 // clone "n" for special uses that are in the not_peeled region
1162 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1163 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1164 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1165 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1166 #ifdef ASSERT
1167 // Validate the loop partition sets: peel and not_peel
1168 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1169 // Ensure that uses outside of loop are of the right form
1170 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1171 uint orig_exit_idx, uint clone_exit_idx);
1172 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1173 #endif
1174
1175 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1176 int stride_of_possible_iv( Node* iff );
1177 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1178 // Return the (unique) control output node that's in the loop (if it exists.)
1179 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1180 // Insert a signed compare loop exit cloned from an unsigned compare.
1181 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1182 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1183 // Utility to register node "n" with PhaseIdealLoop
1184 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1185 // Utility to create an if-projection
1186 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1187 // Force the iff control output to be the live_proj
1188 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1189 // Insert a region before an if projection
1190 RegionNode* insert_region_before_proj(ProjNode* proj);
1191 // Insert a new if before an if projection
1192 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1193
1194 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1195 // "Nearly" because all Nodes have been cloned from the original in the loop,
1196 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1197 // through the Phi recursively, and return a Bool.
1198 Node *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1199 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1200
1201
1202 // Rework addressing expressions to get the most loop-invariant stuff
1203 // moved out. We'd like to do all associative operators, but it's especially
1204 // important (common) to do address expressions.
1205 Node *remix_address_expressions( Node *n );
1206
1207 // Attempt to use a conditional move instead of a phi/branch
1208 Node *conditional_move( Node *n );
1209
1210 // Reorganize offset computations to lower register pressure.
1211 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1212 // (which are then alive with the post-incremented trip counter
1213 // forcing an extra register move)
1214 void reorg_offsets( IdealLoopTree *loop );
1215
1216 // Check for aggressive application of 'split-if' optimization,
1217 // using basic block level info.
1218 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1219 Node *split_if_with_blocks_pre ( Node *n );
1220 void split_if_with_blocks_post( Node *n );
1221 Node *has_local_phi_input( Node *n );
1222 // Mark an IfNode as being dominated by a prior test,
1223 // without actually altering the CFG (and hence IDOM info).
1224 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1225
1226 // Split Node 'n' through merge point
1227 Node *split_thru_region( Node *n, Node *region );
1228 // Split Node 'n' through merge point if there is enough win.
1229 Node *split_thru_phi( Node *n, Node *region, int policy );
1230 // Found an If getting its condition-code input from a Phi in the
1231 // same block. Split thru the Region.
1232 void do_split_if( Node *iff );
1233
1234 // Conversion of fill/copy patterns into intrisic versions
1235 bool do_intrinsify_fill();
1236 bool intrinsify_fill(IdealLoopTree* lpt);
1237 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1238 Node*& shift, Node*& offset);
1239
1240 private:
1241 // Return a type based on condition control flow
1242 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1243 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1244 // Helpers for filtered type
1245 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1246
1247 // Helper functions
1248 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1249 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1250 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 );
1251 bool split_up( Node *n, Node *blk1, Node *blk2 );
1252 void sink_use( Node *use, Node *post_loop );
1253 Node *place_near_use( Node *useblock ) const;
1254 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1255 void try_move_store_after_loop(Node* n);
1256 bool identical_backtoback_ifs(Node *n);
1257 bool can_split_if(Node *n_ctrl);
1258
1259 bool _created_loop_node;
1260 public:
1261 void set_created_loop_node() { _created_loop_node = true; }
1262 bool created_loop_node() { return _created_loop_node; }
1263 void register_new_node( Node *n, Node *blk );
1264
1265 #ifdef ASSERT
1266 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1267 #endif
1268
1269 #ifndef PRODUCT
1270 void dump( ) const;
1271 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1272 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1273 void verify() const; // Major slow :-)
1274 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1275 IdealLoopTree *get_loop_idx(Node* n) const {
1276 // Dead nodes have no loop, so return the top level loop instead
1277 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1278 }
1279 // Print some stats
1280 static void print_statistics();
1281 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1282 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1283 #endif
1284 };
1285
1286 // This kit may be used for making of a reserved copy of a loop before this loop
1287 // goes under non-reversible changes.
1288 //
1289 // Function create_reserve() creates a reserved copy (clone) of the loop.
1290 // The reserved copy is created by calling
1291 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1292 // the original and reserved loops are connected in the outer graph.
1293 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1294 //
1295 // By default the reserved copy (clone) of the loop is created as dead code - it is
1296 // dominated in the outer loop by this node chain:
1297 // intcon(1)->If->IfFalse->reserved_copy.
1298 // The original loop is dominated by the the same node chain but IfTrue projection:
1299 // intcon(0)->If->IfTrue->original_loop.
1300 //
1301 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1302 // and the dtor, checks _use_new value.
1303 // If _use_new == false, it "switches" control to reserved copy of the loop
1304 // by simple replacing of node intcon(1) with node intcon(0).
1305 //
1306 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1307 //
1308 // void CountedLoopReserveKit_example()
1309 // {
1310 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1311 // if (DoReserveCopy && !lrk.has_reserved()) {
1312 // return; //failed to create reserved loop copy
1313 // }
1314 // ...
1315 // //something is wrong, switch to original loop
1316 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1317 // ...
1318 // //everything worked ok, return with the newly modified loop
1319 // lrk.use_new();
1320 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1321 // }
1322 //
1323 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1324 // the dtor will "switch to the original" loop.
1325 // NOTE. You you modify outside of the original loop this class is no help.
1326 //
1327 class CountedLoopReserveKit {
1328 private:
1329 PhaseIdealLoop* _phase;
1330 IdealLoopTree* _lpt;
1331 LoopNode* _lp;
1332 IfNode* _iff;
1333 LoopNode* _lp_reserved;
1334 bool _has_reserved;
1335 bool _use_new;
1336 const bool _active; //may be set to false in ctor, then the object is dummy
1337
1338 public:
1339 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1340 ~CountedLoopReserveKit();
1341 void use_new() {_use_new = true;}
1342 void set_iff(IfNode* x) {_iff = x;}
1343 bool has_reserved() const { return _active && _has_reserved;}
1344 private:
1345 bool create_reserve();
1346 };// class CountedLoopReserveKit
1347
1348 inline Node* IdealLoopTree::tail() {
1349 // Handle lazy update of _tail field
1350 Node *n = _tail;
1351 //while( !n->in(0) ) // Skip dead CFG nodes
1352 //n = n->in(1);
1353 if (n->in(0) == NULL)
1354 n = _phase->get_ctrl(n);
1355 _tail = n;
1356 return n;
1357 }
1358
1359
1360 // Iterate over the loop tree using a preorder, left-to-right traversal.
1361 //
1362 // Example that visits all counted loops from within PhaseIdealLoop
1363 //
1364 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1365 // IdealLoopTree* lpt = iter.current();
1366 // if (!lpt->is_counted()) continue;
1367 // ...
1368 class LoopTreeIterator : public StackObj {
1369 private:
1370 IdealLoopTree* _root;
1371 IdealLoopTree* _curnt;
1372
1373 public:
1374 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1375
1376 bool done() { return _curnt == NULL; } // Finished iterating?
1377
1378 void next(); // Advance to next loop tree
1379
1380 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1381 };
1382
1383 #endif // SHARE_VM_OPTO_LOOPNODE_HPP