1 /*
2 * Copyright (c) 1998, 2016, 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 PhaseIdealLoop;
41 class CountedLoopReserveKit;
42 class VectorSet;
43 class Invariance;
44 struct small_cache;
45
46 //
47 // I D E A L I Z E D L O O P S
48 //
49 // Idealized loops are the set of loops I perform more interesting
50 // transformations on, beyond simple hoisting.
51
52 //------------------------------LoopNode---------------------------------------
53 // Simple loop header. Fall in path on left, loop-back path on right.
54 class LoopNode : public RegionNode {
55 // Size is bigger to hold the flags. However, the flags do not change
56 // the semantics so it does not appear in the hash & cmp functions.
57 virtual uint size_of() const { return sizeof(*this); }
58 protected:
59 uint _loop_flags;
60 // Names for flag bitfields
61 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
62 MainHasNoPreLoop=4,
63 HasExactTripCount=8,
64 InnerLoop=16,
65 PartialPeelLoop=32,
66 PartialPeelFailed=64,
67 HasReductions=128,
68 WasSlpAnalyzed=256,
69 PassedSlpAnalysis=512,
70 DoUnrollOnly=1024,
71 VectorizedLoop=2048,
72 HasAtomicPostLoop=4096,
73 HasRangeChecks=8192,
74 IsMultiversioned=16384,
75 StripMined=32768};
76 char _unswitch_count;
77 enum { _unswitch_max=3 };
78 char _postloop_flags;
79 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 };
80
81 public:
82 // Names for edge indices
83 enum { Self=0, EntryControl, LoopBackControl };
84
85 uint is_inner_loop() const { return _loop_flags & InnerLoop; }
86 void set_inner_loop() { _loop_flags |= InnerLoop; }
87
88 uint range_checks_present() const { return _loop_flags & HasRangeChecks; }
89 uint is_multiversioned() const { return _loop_flags & IsMultiversioned; }
90 uint is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
91 uint is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
92 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
93 uint partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
94 uint is_strip_mined() const { return _loop_flags & StripMined; }
95
96 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
97 void mark_has_reductions() { _loop_flags |= HasReductions; }
98 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
99 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
100 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
101 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
102 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
103 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
104 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
105 void mark_strip_mined() { _loop_flags |= StripMined; }
106 void clear_strip_mined() { _loop_flags &= ~StripMined; }
107
108 int unswitch_max() { return _unswitch_max; }
109 int unswitch_count() { return _unswitch_count; }
110
111 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
112 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
113 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
114 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
115
116 void set_unswitch_count(int val) {
117 assert (val <= unswitch_max(), "too many unswitches");
118 _unswitch_count = val;
119 }
120
121 LoopNode(Node *entry, Node *backedge) : RegionNode(3), _loop_flags(0), _unswitch_count(0), _postloop_flags(0) {
122 init_class_id(Class_Loop);
123 init_req(EntryControl, entry);
124 init_req(LoopBackControl, backedge);
125 }
126
127 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
128 virtual int Opcode() const;
129 bool can_be_counted_loop(PhaseTransform* phase) const {
130 return req() == 3 && in(0) != NULL &&
131 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
132 in(2) != NULL && phase->type(in(2)) != Type::TOP;
133 }
134 bool is_valid_counted_loop() const;
135 #ifndef PRODUCT
136 virtual void dump_spec(outputStream *st) const;
137 #endif
138
139 void verify_strip_mined(int expect_opaq) const;
140 virtual LoopNode* skip_strip_mined(int expect_opaq = 1) { return this; }
141 };
142
143 //------------------------------Counted Loops----------------------------------
144 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
145 // path (and maybe some other exit paths). The trip-counter exit is always
146 // last in the loop. The trip-counter have to stride by a constant;
147 // the exit value is also loop invariant.
148
149 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
150 // CountedLoopNode has the incoming loop control and the loop-back-control
151 // which is always the IfTrue before the matching CountedLoopEndNode. The
152 // CountedLoopEndNode has an incoming control (possibly not the
153 // CountedLoopNode if there is control flow in the loop), the post-increment
154 // trip-counter value, and the limit. The trip-counter value is always of
155 // the form (Op old-trip-counter stride). The old-trip-counter is produced
156 // by a Phi connected to the CountedLoopNode. The stride is constant.
157 // The Op is any commutable opcode, including Add, Mul, Xor. The
158 // CountedLoopEndNode also takes in the loop-invariant limit value.
159
160 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
161 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
162 // via the old-trip-counter from the Op node.
163
164 //------------------------------CountedLoopNode--------------------------------
165 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
166 // inputs the incoming loop-start control and the loop-back control, so they
167 // act like RegionNodes. They also take in the initial trip counter, the
168 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
169 // produce a loop-body control and the trip counter value. Since
170 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
171
172 class CountedLoopNode : public LoopNode {
173 // Size is bigger to hold _main_idx. However, _main_idx does not change
174 // the semantics so it does not appear in the hash & cmp functions.
175 virtual uint size_of() const { return sizeof(*this); }
176
177 // For Pre- and Post-loops during debugging ONLY, this holds the index of
178 // the Main CountedLoop. Used to assert that we understand the graph shape.
179 node_idx_t _main_idx;
180
181 // Known trip count calculated by compute_exact_trip_count()
182 uint _trip_count;
183
184 // Expected trip count from profile data
185 float _profile_trip_cnt;
186
187 // Log2 of original loop bodies in unrolled loop
188 int _unrolled_count_log2;
189
190 // Node count prior to last unrolling - used to decide if
191 // unroll,optimize,unroll,optimize,... is making progress
192 int _node_count_before_unroll;
193
194 // If slp analysis is performed we record the maximum
195 // vector mapped unroll factor here
196 int _slp_maximum_unroll_factor;
197
198 public:
199 CountedLoopNode( Node *entry, Node *backedge )
200 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
201 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
202 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
203 init_class_id(Class_CountedLoop);
204 // Initialize _trip_count to the largest possible value.
205 // Will be reset (lower) if the loop's trip count is known.
206 }
207
208 virtual int Opcode() const;
209 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
210
211 Node *init_control() const { return in(EntryControl); }
212 Node *back_control() const { return in(LoopBackControl); }
213 CountedLoopEndNode *loopexit() const;
214 Node *init_trip() const;
215 Node *stride() const;
216 int stride_con() const;
217 bool stride_is_con() const;
218 Node *limit() const;
219 Node *incr() const;
220 Node *phi() const;
221
222 // Match increment with optional truncation
223 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
224
225 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
226 // can run short a few iterations and may start a few iterations in.
227 // It will be RCE'd and unrolled and aligned.
228
229 // A following 'post' loop will run any remaining iterations. Used
230 // during Range Check Elimination, the 'post' loop will do any final
231 // iterations with full checks. Also used by Loop Unrolling, where
232 // the 'post' loop will do any epilog iterations needed. Basically,
233 // a 'post' loop can not profitably be further unrolled or RCE'd.
234
235 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
236 // it may do under-flow checks for RCE and may do alignment iterations
237 // so the following main loop 'knows' that it is striding down cache
238 // lines.
239
240 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
241 // Aligned, may be missing it's pre-loop.
242 uint is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
243 uint is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
244 uint is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
245 uint is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
246 uint is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
247 uint was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
248 uint has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
249 uint do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
250 uint is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
251 uint has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
252 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
253
254 int main_idx() const { return _main_idx; }
255
256
257 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
258 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
259 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
260 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
261
262 void set_trip_count(uint tc) { _trip_count = tc; }
263 uint trip_count() { return _trip_count; }
264
265 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
266 void set_exact_trip_count(uint tc) {
267 _trip_count = tc;
268 _loop_flags |= HasExactTripCount;
269 }
270 void set_nonexact_trip_count() {
271 _loop_flags &= ~HasExactTripCount;
272 }
273 void set_notpassed_slp() {
274 _loop_flags &= ~PassedSlpAnalysis;
275 }
276
277 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
278 float profile_trip_cnt() { return _profile_trip_cnt; }
279
280 void double_unrolled_count() { _unrolled_count_log2++; }
281 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
282
283 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
284 int node_count_before_unroll() { return _node_count_before_unroll; }
285 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
286 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
287
288 virtual LoopNode* skip_strip_mined(int expect_opaq = 1);
289
290 #ifndef PRODUCT
291 virtual void dump_spec(outputStream *st) const;
292 #endif
293 };
294
295 //------------------------------CountedLoopEndNode-----------------------------
296 // CountedLoopEndNodes end simple trip counted loops. They act much like
297 // IfNodes.
298 class CountedLoopEndNode : public IfNode {
299 public:
300 enum { TestControl, TestValue };
301
302 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
303 : IfNode( control, test, prob, cnt) {
304 init_class_id(Class_CountedLoopEnd);
305 }
306 virtual int Opcode() const;
307
308 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
309 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
310 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
311 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
312 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
313 int stride_con() const;
314 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
315 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
316 PhiNode *phi() const {
317 Node *tmp = incr();
318 if (tmp && tmp->req() == 3) {
319 Node* phi = tmp->in(1);
320 if (phi->is_Phi()) {
321 return phi->as_Phi();
322 }
323 }
324 return NULL;
325 }
326 CountedLoopNode *loopnode() const {
327 // The CountedLoopNode that goes with this CountedLoopEndNode may
328 // have been optimized out by the IGVN so be cautious with the
329 // pattern matching on the graph
330 PhiNode* iv_phi = phi();
331 if (iv_phi == NULL) {
332 return NULL;
333 }
334 Node *ln = iv_phi->in(0);
335 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
336 return (CountedLoopNode*)ln;
337 }
338 return NULL;
339 }
340
341 #ifndef PRODUCT
342 virtual void dump_spec(outputStream *st) const;
343 #endif
344 };
345
346
347 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
348 Node *bc = back_control();
349 if( bc == NULL ) return NULL;
350 Node *le = bc->in(0);
351 if( le->Opcode() != Op_CountedLoopEnd )
352 return NULL;
353 return (CountedLoopEndNode*)le;
354 }
355 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
356 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
357 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
358 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
359 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
360 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
361 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
362
363 //------------------------------LoopLimitNode-----------------------------
364 // Counted Loop limit node which represents exact final iterator value:
365 // trip_count = (limit - init_trip + stride - 1)/stride
366 // final_value= trip_count * stride + init_trip.
367 // Use HW instructions to calculate it when it can overflow in integer.
368 // Note, final_value should fit into integer since counted loop has
369 // limit check: limit <= max_int-stride.
370 class LoopLimitNode : public Node {
371 enum { Init=1, Limit=2, Stride=3 };
372 public:
373 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
374 // Put it on the Macro nodes list to optimize during macro nodes expansion.
375 init_flags(Flag_is_macro);
376 C->add_macro_node(this);
377 }
378 virtual int Opcode() const;
379 virtual const Type *bottom_type() const { return TypeInt::INT; }
380 virtual uint ideal_reg() const { return Op_RegI; }
381 virtual const Type* Value(PhaseGVN* phase) const;
382 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
383 virtual Node* Identity(PhaseGVN* phase);
384 };
385
386 // -----------------------------IdealLoopTree----------------------------------
387 class IdealLoopTree : public ResourceObj {
388 public:
389 IdealLoopTree *_parent; // Parent in loop tree
390 IdealLoopTree *_next; // Next sibling in loop tree
391 IdealLoopTree *_child; // First child in loop tree
392
393 // The head-tail backedge defines the loop.
394 // If tail is NULL then this loop has multiple backedges as part of the
395 // same loop. During cleanup I'll peel off the multiple backedges; merge
396 // them at the loop bottom and flow 1 real backedge into the loop.
397 Node *_head; // Head of loop
398 Node *_tail; // Tail of loop
399 inline Node *tail(); // Handle lazy update of _tail field
400 PhaseIdealLoop* _phase;
401 int _local_loop_unroll_limit;
402 int _local_loop_unroll_factor;
403
404 Node_List _body; // Loop body for inner loops
405
406 uint8_t _nest; // Nesting depth
407 uint8_t _irreducible:1, // True if irreducible
408 _has_call:1, // True if has call safepoint
409 _has_sfpt:1, // True if has non-call safepoint
410 _rce_candidate:1; // True if candidate for range check elimination
411
412 Node_List* _safepts; // List of safepoints in this loop
413 Node_List* _required_safept; // A inner loop cannot delete these safepts;
414 bool _allow_optimizations; // Allow loop optimizations
415
416 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
417 : _parent(0), _next(0), _child(0),
418 _head(head), _tail(tail),
419 _phase(phase),
420 _safepts(NULL),
421 _required_safept(NULL),
422 _allow_optimizations(true),
423 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
424 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
425 { }
426
427 // Is 'l' a member of 'this'?
428 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
429
430 // Set loop nesting depth. Accumulate has_call bits.
431 int set_nest( uint depth );
432
433 // Split out multiple fall-in edges from the loop header. Move them to a
434 // private RegionNode before the loop. This becomes the loop landing pad.
435 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
436
437 // Split out the outermost loop from this shared header.
438 void split_outer_loop( PhaseIdealLoop *phase );
439
440 // Merge all the backedges from the shared header into a private Region.
441 // Feed that region as the one backedge to this loop.
442 void merge_many_backedges( PhaseIdealLoop *phase );
443
444 // Split shared headers and insert loop landing pads.
445 // Insert a LoopNode to replace the RegionNode.
446 // Returns TRUE if loop tree is structurally changed.
447 bool beautify_loops( PhaseIdealLoop *phase );
448
449 // Perform optimization to use the loop predicates for null checks and range checks.
450 // Applies to any loop level (not just the innermost one)
451 bool loop_predication( PhaseIdealLoop *phase);
452
453 // Perform iteration-splitting on inner loops. Split iterations to
454 // avoid range checks or one-shot null checks. Returns false if the
455 // current round of loop opts should stop.
456 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
457
458 // Driver for various flavors of iteration splitting. Returns false
459 // if the current round of loop opts should stop.
460 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
461
462 // Given dominators, try to find loops with calls that must always be
463 // executed (call dominates loop tail). These loops do not need non-call
464 // safepoints (ncsfpt).
465 void check_safepts(VectorSet &visited, Node_List &stack);
466
467 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
468 // encountered.
469 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
470
471 // Remove safepoints from loop. Optionally keeping one.
472 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
473
474 // Convert to counted loops where possible
475 void counted_loop( PhaseIdealLoop *phase );
476
477 // Check for Node being a loop-breaking test
478 Node *is_loop_exit(Node *iff) const;
479
480 // Remove simplistic dead code from loop body
481 void DCE_loop_body();
482
483 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
484 // Replace with a 1-in-10 exit guess.
485 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
486
487 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
488 // Useful for unrolling loops with NO array accesses.
489 bool policy_peel_only( PhaseIdealLoop *phase ) const;
490
491 // Return TRUE or FALSE if the loop should be unswitched -- clone
492 // loop with an invariant test
493 bool policy_unswitching( PhaseIdealLoop *phase ) const;
494
495 // Micro-benchmark spamming. Remove empty loops.
496 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
497
498 // Convert one iteration loop into normal code.
499 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
500
501 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
502 // make some loop-invariant test (usually a null-check) happen before the
503 // loop.
504 bool policy_peeling( PhaseIdealLoop *phase ) const;
505
506 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
507 // known trip count in the counted loop node.
508 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
509
510 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
511 // the loop is a CountedLoop and the body is small enough.
512 bool policy_unroll(PhaseIdealLoop *phase);
513
514 // Loop analyses to map to a maximal superword unrolling for vectorization.
515 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
516
517 // Return TRUE or FALSE if the loop should be range-check-eliminated.
518 // Gather a list of IF tests that are dominated by iteration splitting;
519 // also gather the end of the first split and the start of the 2nd split.
520 bool policy_range_check( PhaseIdealLoop *phase ) const;
521
522 // Return TRUE or FALSE if the loop should be cache-line aligned.
523 // Gather the expression that does the alignment. Note that only
524 // one array base can be aligned in a loop (unless the VM guarantees
525 // mutual alignment). Note that if we vectorize short memory ops
526 // into longer memory ops, we may want to increase alignment.
527 bool policy_align( PhaseIdealLoop *phase ) const;
528
529 // Return TRUE if "iff" is a range check.
530 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
531
532 // Compute loop trip count if possible
533 void compute_trip_count(PhaseIdealLoop* phase);
534
535 // Compute loop trip count from profile data
536 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
537
538 // Reassociate invariant expressions.
539 void reassociate_invariants(PhaseIdealLoop *phase);
540 // Reassociate invariant add and subtract expressions.
541 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
542 // Return nonzero index of invariant operand if invariant and variant
543 // are combined with an Add or Sub. Helper for reassociate_invariants.
544 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
545
546 // Return true if n is invariant
547 bool is_invariant(Node* n) const;
548
549 // Put loop body on igvn work list
550 void record_for_igvn();
551
552 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
553 bool is_inner() { return is_loop() && _child == NULL; }
554 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
555
556 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
557
558 #ifndef PRODUCT
559 void dump_head( ) const; // Dump loop head only
560 void dump() const; // Dump this loop recursively
561 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
562 #endif
563
564 };
565
566 // -----------------------------PhaseIdealLoop---------------------------------
567 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
568 // loop tree. Drives the loop-based transformations on the ideal graph.
569 class PhaseIdealLoop : public PhaseTransform {
570 friend class IdealLoopTree;
571 friend class SuperWord;
572 friend class CountedLoopReserveKit;
573
574 // Pre-computed def-use info
575 PhaseIterGVN &_igvn;
576
577 // Head of loop tree
578 IdealLoopTree *_ltree_root;
579
580 // Array of pre-order numbers, plus post-visited bit.
581 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
582 // ODD for post-visited. Other bits are the pre-order number.
583 uint *_preorders;
584 uint _max_preorder;
585
586 const PhaseIdealLoop* _verify_me;
587 bool _verify_only;
588
589 // Allocate _preorders[] array
590 void allocate_preorders() {
591 _max_preorder = C->unique()+8;
592 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
593 memset(_preorders, 0, sizeof(uint) * _max_preorder);
594 }
595
596 // Allocate _preorders[] array
597 void reallocate_preorders() {
598 if ( _max_preorder < C->unique() ) {
599 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
600 _max_preorder = C->unique();
601 }
602 memset(_preorders, 0, sizeof(uint) * _max_preorder);
603 }
604
605 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
606 // adds new nodes.
607 void check_grow_preorders( ) {
608 if ( _max_preorder < C->unique() ) {
609 uint newsize = _max_preorder<<1; // double size of array
610 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
611 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
612 _max_preorder = newsize;
613 }
614 }
615 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
616 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
617 // Pre-order numbers are written to the Nodes array as low-bit-set values.
618 void set_preorder_visited( Node *n, int pre_order ) {
619 assert( !is_visited( n ), "already set" );
620 _preorders[n->_idx] = (pre_order<<1);
621 };
622 // Return pre-order number.
623 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
624
625 // Check for being post-visited.
626 // Should be previsited already (checked with assert(is_visited(n))).
627 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
628
629 // Mark as post visited
630 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
631
632 public:
633 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
634 // Returns true if "n" is a data node, false if it's a control node.
635 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
636
637 private:
638 // clear out dead code after build_loop_late
639 Node_List _deadlist;
640
641 // Support for faster execution of get_late_ctrl()/dom_lca()
642 // when a node has many uses and dominator depth is deep.
643 Node_Array _dom_lca_tags;
644 void init_dom_lca_tags();
645 void clear_dom_lca_tags();
646
647 // Helper for debugging bad dominance relationships
648 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
649
650 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
651
652 // Inline wrapper for frequent cases:
653 // 1) only one use
654 // 2) a use is the same as the current LCA passed as 'n1'
655 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
656 assert( n->is_CFG(), "" );
657 // Fast-path NULL lca
658 if( lca != NULL && lca != n ) {
659 assert( lca->is_CFG(), "" );
660 // find LCA of all uses
661 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
662 }
663 return find_non_split_ctrl(n);
664 }
665 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
666
667 // Helper function for directing control inputs away from CFG split
668 // points.
669 Node *find_non_split_ctrl( Node *ctrl ) const {
670 if (ctrl != NULL) {
671 if (ctrl->is_MultiBranch()) {
672 ctrl = ctrl->in(0);
673 }
674 assert(ctrl->is_CFG(), "CFG");
675 }
676 return ctrl;
677 }
678
679 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
680
681 public:
682
683 static bool is_canonical_loop_entry(CountedLoopNode* cl);
684
685 bool has_node( Node* n ) const {
686 guarantee(n != NULL, "No Node.");
687 return _nodes[n->_idx] != NULL;
688 }
689 // check if transform created new nodes that need _ctrl recorded
690 Node *get_late_ctrl( Node *n, Node *early );
691 Node *get_early_ctrl( Node *n );
692 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
693 void set_early_ctrl( Node *n );
694 void set_subtree_ctrl( Node *root );
695 void set_ctrl( Node *n, Node *ctrl ) {
696 assert( !has_node(n) || has_ctrl(n), "" );
697 assert( ctrl->in(0), "cannot set dead control node" );
698 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
699 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
700 }
701 // Set control and update loop membership
702 void set_ctrl_and_loop(Node* n, Node* ctrl) {
703 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
704 IdealLoopTree* new_loop = get_loop(ctrl);
705 if (old_loop != new_loop) {
706 if (old_loop->_child == NULL) old_loop->_body.yank(n);
707 if (new_loop->_child == NULL) new_loop->_body.push(n);
708 }
709 set_ctrl(n, ctrl);
710 }
711 // Control nodes can be replaced or subsumed. During this pass they
712 // get their replacement Node in slot 1. Instead of updating the block
713 // location of all Nodes in the subsumed block, we lazily do it. As we
714 // pull such a subsumed block out of the array, we write back the final
715 // correct block.
716 Node *get_ctrl( Node *i ) {
717 assert(has_node(i), "");
718 Node *n = get_ctrl_no_update(i);
719 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
720 assert(has_node(i) && has_ctrl(i), "");
721 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
722 return n;
723 }
724 // true if CFG node d dominates CFG node n
725 bool is_dominator(Node *d, Node *n);
726 // return get_ctrl for a data node and self(n) for a CFG node
727 Node* ctrl_or_self(Node* n) {
728 if (has_ctrl(n))
729 return get_ctrl(n);
730 else {
731 assert (n->is_CFG(), "must be a CFG node");
732 return n;
733 }
734 }
735
736 private:
737 Node *get_ctrl_no_update_helper(Node *i) const {
738 assert(has_ctrl(i), "should be control, not loop");
739 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
740 }
741
742 Node *get_ctrl_no_update(Node *i) const {
743 assert( has_ctrl(i), "" );
744 Node *n = get_ctrl_no_update_helper(i);
745 if (!n->in(0)) {
746 // Skip dead CFG nodes
747 do {
748 n = get_ctrl_no_update_helper(n);
749 } while (!n->in(0));
750 n = find_non_split_ctrl(n);
751 }
752 return n;
753 }
754
755 // Check for loop being set
756 // "n" must be a control node. Returns true if "n" is known to be in a loop.
757 bool has_loop( Node *n ) const {
758 assert(!has_node(n) || !has_ctrl(n), "");
759 return has_node(n);
760 }
761 // Set loop
762 void set_loop( Node *n, IdealLoopTree *loop ) {
763 _nodes.map(n->_idx, (Node*)loop);
764 }
765 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
766 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
767 // from old_node to new_node to support the lazy update. Reference
768 // replaces loop reference, since that is not needed for dead node.
769 public:
770 void lazy_update(Node *old_node, Node *new_node) {
771 assert(old_node != new_node, "no cycles please");
772 // Re-use the side array slot for this node to provide the
773 // forwarding pointer.
774 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
775 }
776 void lazy_replace(Node *old_node, Node *new_node) {
777 _igvn.replace_node(old_node, new_node);
778 lazy_update(old_node, new_node);
779 }
780
781 private:
782
783 // Place 'n' in some loop nest, where 'n' is a CFG node
784 void build_loop_tree();
785 int build_loop_tree_impl( Node *n, int pre_order );
786 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
787 // loop tree, not the root.
788 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
789
790 // Place Data nodes in some loop nest
791 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
792 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
793 void build_loop_late_post ( Node* n );
794
795 // Array of immediate dominance info for each CFG node indexed by node idx
796 private:
797 uint _idom_size;
798 Node **_idom; // Array of immediate dominators
799 uint *_dom_depth; // Used for fast LCA test
800 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
801
802 Node* idom_no_update(Node* d) const {
803 assert(d->_idx < _idom_size, "oob");
804 Node* n = _idom[d->_idx];
805 assert(n != NULL,"Bad immediate dominator info.");
806 while (n->in(0) == NULL) { // Skip dead CFG nodes
807 //n = n->in(1);
808 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
809 assert(n != NULL,"Bad immediate dominator info.");
810 }
811 return n;
812 }
813 public:
814 Node *idom(Node* d) const {
815 uint didx = d->_idx;
816 Node *n = idom_no_update(d);
817 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
818 return n;
819 }
820 uint dom_depth(Node* d) const {
821 guarantee(d != NULL, "Null dominator info.");
822 guarantee(d->_idx < _idom_size, "");
823 return _dom_depth[d->_idx];
824 }
825 void set_idom(Node* d, Node* n, uint dom_depth);
826
827 private:
828 // Locally compute IDOM using dom_lca call
829 Node *compute_idom( Node *region ) const;
830 // Recompute dom_depth
831 void recompute_dom_depth();
832
833 // Is safept not required by an outer loop?
834 bool is_deleteable_safept(Node* sfpt);
835
836 // Replace parallel induction variable (parallel to trip counter)
837 void replace_parallel_iv(IdealLoopTree *loop);
838
839 // Perform verification that the graph is valid.
840 PhaseIdealLoop( PhaseIterGVN &igvn) :
841 PhaseTransform(Ideal_Loop),
842 _igvn(igvn),
843 _dom_lca_tags(arena()), // Thread::resource_area
844 _verify_me(NULL),
845 _verify_only(true) {
846 build_and_optimize(LoopOptsVerify);
847 }
848
849 // build the loop tree and perform any requested optimizations
850 void build_and_optimize(LoopOptsMode mode);
851
852 public:
853 // Dominators for the sea of nodes
854 void Dominators();
855 Node *dom_lca( Node *n1, Node *n2 ) const {
856 return find_non_split_ctrl(dom_lca_internal(n1, n2));
857 }
858 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
859
860 // Compute the Ideal Node to Loop mapping
861 PhaseIdealLoop( PhaseIterGVN &igvn, LoopOptsMode mode) :
862 PhaseTransform(Ideal_Loop),
863 _igvn(igvn),
864 _dom_lca_tags(arena()), // Thread::resource_area
865 _verify_me(NULL),
866 _verify_only(false) {
867 build_and_optimize(mode);
868 }
869
870 // Verify that verify_me made the same decisions as a fresh run.
871 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
872 PhaseTransform(Ideal_Loop),
873 _igvn(igvn),
874 _dom_lca_tags(arena()), // Thread::resource_area
875 _verify_me(verify_me),
876 _verify_only(false) {
877 build_and_optimize(LoopOptsVerify);
878 }
879
880 // Build and verify the loop tree without modifying the graph. This
881 // is useful to verify that all inputs properly dominate their uses.
882 static void verify(PhaseIterGVN& igvn) {
883 #ifdef ASSERT
884 PhaseIdealLoop v(igvn);
885 #endif
886 }
887
888 // True if the method has at least 1 irreducible loop
889 bool _has_irreducible_loops;
890
891 // Per-Node transform
892 virtual Node *transform( Node *a_node ) { return 0; }
893
894 bool is_counted_loop(Node* x, IdealLoopTree*& loop);
895 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
896 IdealLoopTree* loop, float cl_prob, float le_fcnt,
897 Node*& entry_control, Node*& iffalse);
898
899 Node* exact_limit( IdealLoopTree *loop );
900
901 // Return a post-walked LoopNode
902 IdealLoopTree *get_loop( Node *n ) const {
903 // Dead nodes have no loop, so return the top level loop instead
904 if (!has_node(n)) return _ltree_root;
905 assert(!has_ctrl(n), "");
906 return (IdealLoopTree*)_nodes[n->_idx];
907 }
908
909 // Is 'n' a (nested) member of 'loop'?
910 int is_member( const IdealLoopTree *loop, Node *n ) const {
911 return loop->is_member(get_loop(n)); }
912
913 // This is the basic building block of the loop optimizations. It clones an
914 // entire loop body. It makes an old_new loop body mapping; with this
915 // mapping you can find the new-loop equivalent to an old-loop node. All
916 // new-loop nodes are exactly equal to their old-loop counterparts, all
917 // edges are the same. All exits from the old-loop now have a RegionNode
918 // that merges the equivalent new-loop path. This is true even for the
919 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
920 // now come from (one or more) Phis that merge their new-loop equivalents.
921 // Parameter side_by_side_idom:
922 // When side_by_size_idom is NULL, the dominator tree is constructed for
923 // the clone loop to dominate the original. Used in construction of
924 // pre-main-post loop sequence.
925 // When nonnull, the clone and original are side-by-side, both are
926 // dominated by the passed in side_by_side_idom node. Used in
927 // construction of unswitched loops.
928 enum CloneLoopMode {
929 IgnoreStripMined = 0, // Only clone inner strip mined loop
930 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
931 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
932 // result control flow branches
933 // either to inner clone or outer
934 // strip mined loop.
935 };
936 void clone_loop(IdealLoopTree *loop, Node_List &old_new, int dom_depth,
937 CloneLoopMode mode, Node* side_by_side_idom = NULL);
938 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
939 IdealLoopTree* loop, IdealLoopTree* companion_loop,
940 Node_List*& split_if_set, Node_List*& split_bool_set,
941 Node_List*& split_cex_set, Node_List& worklist,
942 uint new_counter, CloneLoopMode mode);
943 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
944 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
945 Node_List& extra_data_nodes);
946
947 // If we got the effect of peeling, either by actually peeling or by
948 // making a pre-loop which must execute at least once, we can remove
949 // all loop-invariant dominated tests in the main body.
950 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
951
952 // Generate code to do a loop peel for the given loop (and body).
953 // old_new is a temp array.
954 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
955
956 // Add pre and post loops around the given loop. These loops are used
957 // during RCE, unrolling and aligning loops.
958 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
959
960 // Add post loop after the given loop.
961 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
962 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
963 Node *incr, Node *limit, CountedLoopNode *&post_head);
964
965 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
966 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
967
968 // Add a vector post loop between a vector main loop and the current post loop
969 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
970 // If Node n lives in the back_ctrl block, we clone a private version of n
971 // in preheader_ctrl block and return that, otherwise return n.
972 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
973
974 // Take steps to maximally unroll the loop. Peel any odd iterations, then
975 // unroll to do double iterations. The next round of major loop transforms
976 // will repeat till the doubled loop body does all remaining iterations in 1
977 // pass.
978 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
979
980 // Unroll the loop body one step - make each trip do 2 iterations.
981 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
982
983 // Mark vector reduction candidates before loop unrolling
984 void mark_reductions( IdealLoopTree *loop );
985
986 // Return true if exp is a constant times an induction var
987 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
988
989 // Return true if exp is a scaled induction var plus (or minus) constant
990 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
991
992 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
993 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
994 Deoptimization::DeoptReason reason,
995 int opcode);
996 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
997
998 // Clone loop predicates to cloned loops (peeled, unswitched)
999 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
1000 Deoptimization::DeoptReason reason,
1001 PhaseIdealLoop* loop_phase,
1002 PhaseIterGVN* igvn);
1003
1004 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
1005 bool clone_limit_check,
1006 PhaseIdealLoop* loop_phase,
1007 PhaseIterGVN* igvn);
1008 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
1009
1010 static Node* skip_loop_predicates(Node* entry);
1011
1012 // Find a good location to insert a predicate
1013 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1014 // Find a predicate
1015 static Node* find_predicate(Node* entry);
1016 // Construct a range check for a predicate if
1017 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1018 int scale, Node* offset,
1019 Node* init, Node* limit, Node* stride,
1020 Node* range, bool upper);
1021
1022 // Implementation of the loop predication to promote checks outside the loop
1023 bool loop_predication_impl(IdealLoopTree *loop);
1024
1025 // Helper function to collect predicate for eliminating the useless ones
1026 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1027 void eliminate_useless_predicates();
1028
1029 // Change the control input of expensive nodes to allow commoning by
1030 // IGVN when it is guaranteed to not result in a more frequent
1031 // execution of the expensive node. Return true if progress.
1032 bool process_expensive_nodes();
1033
1034 // Check whether node has become unreachable
1035 bool is_node_unreachable(Node *n) const {
1036 return !has_node(n) || n->is_unreachable(_igvn);
1037 }
1038
1039 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1040 int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1041
1042 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1043 void has_range_checks(IdealLoopTree *loop);
1044
1045 // Process post loops which have range checks and try to build a multi-version
1046 // guard to safely determine if we can execute the post loop which was RCE'd.
1047 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1048
1049 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1050 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1051
1052 // Create a slow version of the loop by cloning the loop
1053 // and inserting an if to select fast-slow versions.
1054 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1055 Node_List &old_new,
1056 int opcode,
1057 CloneLoopMode mode);
1058
1059 // Clone a loop and return the clone head (clone_loop_head).
1060 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1061 // This routine was created for usage in CountedLoopReserveKit.
1062 //
1063 // int(1) -> If -> IfTrue -> original_loop_head
1064 // |
1065 // V
1066 // IfFalse -> clone_loop_head (returned by function pointer)
1067 //
1068 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1069 // Clone loop with an invariant test (that does not exit) and
1070 // insert a clone of the test that selects which version to
1071 // execute.
1072 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1073
1074 // Find candidate "if" for unswitching
1075 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1076
1077 // Range Check Elimination uses this function!
1078 // Constrain the main loop iterations so the affine function:
1079 // low_limit <= scale_con * I + offset < upper_limit
1080 // always holds true. That is, either increase the number of iterations in
1081 // the pre-loop or the post-loop until the condition holds true in the main
1082 // loop. Scale_con, offset and limit are all loop invariant.
1083 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 );
1084 // Helper function for add_constraint().
1085 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1086
1087 // Partially peel loop up through last_peel node.
1088 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1089
1090 // Create a scheduled list of nodes control dependent on ctrl set.
1091 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1092 // Has a use in the vector set
1093 bool has_use_in_set( Node* n, VectorSet& vset );
1094 // Has use internal to the vector set (ie. not in a phi at the loop head)
1095 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1096 // clone "n" for uses that are outside of loop
1097 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1098 // clone "n" for special uses that are in the not_peeled region
1099 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1100 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1101 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1102 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1103 #ifdef ASSERT
1104 // Validate the loop partition sets: peel and not_peel
1105 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1106 // Ensure that uses outside of loop are of the right form
1107 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1108 uint orig_exit_idx, uint clone_exit_idx);
1109 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1110 #endif
1111
1112 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1113 int stride_of_possible_iv( Node* iff );
1114 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1115 // Return the (unique) control output node that's in the loop (if it exists.)
1116 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1117 // Insert a signed compare loop exit cloned from an unsigned compare.
1118 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1119 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1120 // Utility to register node "n" with PhaseIdealLoop
1121 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1122 // Utility to create an if-projection
1123 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1124 // Force the iff control output to be the live_proj
1125 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1126 // Insert a region before an if projection
1127 RegionNode* insert_region_before_proj(ProjNode* proj);
1128 // Insert a new if before an if projection
1129 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1130
1131 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1132 // "Nearly" because all Nodes have been cloned from the original in the loop,
1133 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1134 // through the Phi recursively, and return a Bool.
1135 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1136 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1137
1138
1139 // Rework addressing expressions to get the most loop-invariant stuff
1140 // moved out. We'd like to do all associative operators, but it's especially
1141 // important (common) to do address expressions.
1142 Node *remix_address_expressions( Node *n );
1143
1144 // Attempt to use a conditional move instead of a phi/branch
1145 Node *conditional_move( Node *n );
1146
1147 // Reorganize offset computations to lower register pressure.
1148 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1149 // (which are then alive with the post-incremented trip counter
1150 // forcing an extra register move)
1151 void reorg_offsets( IdealLoopTree *loop );
1152
1153 // Check for aggressive application of 'split-if' optimization,
1154 // using basic block level info.
1155 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1156 Node *split_if_with_blocks_pre ( Node *n );
1157 void split_if_with_blocks_post( Node *n );
1158 Node *has_local_phi_input( Node *n );
1159 // Mark an IfNode as being dominated by a prior test,
1160 // without actually altering the CFG (and hence IDOM info).
1161 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1162
1163 // Split Node 'n' through merge point
1164 Node *split_thru_region( Node *n, Node *region );
1165 // Split Node 'n' through merge point if there is enough win.
1166 Node *split_thru_phi( Node *n, Node *region, int policy );
1167 void split_mem_thru_phi(Node*, Node* r, Node* phi);
1168
1169 // Found an If getting its condition-code input from a Phi in the
1170 // same block. Split thru the Region.
1171 void do_split_if( Node *iff );
1172
1173 // Conversion of fill/copy patterns into intrisic versions
1174 bool do_intrinsify_fill();
1175 bool intrinsify_fill(IdealLoopTree* lpt);
1176 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1177 Node*& shift, Node*& offset);
1178
1179 private:
1180 // Return a type based on condition control flow
1181 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1182 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1183 // Helpers for filtered type
1184 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1185
1186 // Helper functions
1187 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1188 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1189 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 );
1190 bool split_up( Node *n, Node *blk1, Node *blk2 );
1191 void sink_use( Node *use, Node *post_loop );
1192 Node *place_near_use( Node *useblock ) const;
1193 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1194 void try_move_store_after_loop(Node* n);
1195 public:
1196 bool identical_backtoback_ifs(Node *n);
1197 bool can_split_if(Node *n_ctrl);
1198 private:
1199
1200 Unique_Node_List _shenandoah_evacuation_tests;
1201
1202 bool _created_loop_node;
1203 public:
1204 void set_created_loop_node() { _created_loop_node = true; }
1205 bool created_loop_node() { return _created_loop_node; }
1206 void register_new_node( Node *n, Node *blk );
1207
1208 #ifdef ASSERT
1209 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1210 #endif
1211
1212 #ifndef PRODUCT
1213 void dump( ) const;
1214 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1215 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1216 void verify() const; // Major slow :-)
1217 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1218 IdealLoopTree *get_loop_idx(Node* n) const {
1219 // Dead nodes have no loop, so return the top level loop instead
1220 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1221 }
1222 // Print some stats
1223 static void print_statistics();
1224 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1225 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1226 #endif
1227
1228 PhaseIterGVN& igvn() { return _igvn; }
1229 IdealLoopTree* ltree_root() const { return _ltree_root; }
1230 };
1231
1232 // This kit may be used for making of a reserved copy of a loop before this loop
1233 // goes under non-reversible changes.
1234 //
1235 // Function create_reserve() creates a reserved copy (clone) of the loop.
1236 // The reserved copy is created by calling
1237 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1238 // the original and reserved loops are connected in the outer graph.
1239 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1240 //
1241 // By default the reserved copy (clone) of the loop is created as dead code - it is
1242 // dominated in the outer loop by this node chain:
1243 // intcon(1)->If->IfFalse->reserved_copy.
1244 // The original loop is dominated by the the same node chain but IfTrue projection:
1245 // intcon(0)->If->IfTrue->original_loop.
1246 //
1247 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1248 // and the dtor, checks _use_new value.
1249 // If _use_new == false, it "switches" control to reserved copy of the loop
1250 // by simple replacing of node intcon(1) with node intcon(0).
1251 //
1252 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1253 //
1254 // void CountedLoopReserveKit_example()
1255 // {
1256 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1257 // if (DoReserveCopy && !lrk.has_reserved()) {
1258 // return; //failed to create reserved loop copy
1259 // }
1260 // ...
1261 // //something is wrong, switch to original loop
1262 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1263 // ...
1264 // //everything worked ok, return with the newly modified loop
1265 // lrk.use_new();
1266 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1267 // }
1268 //
1269 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1270 // the dtor will "switch to the original" loop.
1271 // NOTE. You you modify outside of the original loop this class is no help.
1272 //
1273 class CountedLoopReserveKit {
1274 private:
1275 PhaseIdealLoop* _phase;
1276 IdealLoopTree* _lpt;
1277 LoopNode* _lp;
1278 IfNode* _iff;
1279 LoopNode* _lp_reserved;
1280 bool _has_reserved;
1281 bool _use_new;
1282 const bool _active; //may be set to false in ctor, then the object is dummy
1283
1284 public:
1285 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1286 ~CountedLoopReserveKit();
1287 void use_new() {_use_new = true;}
1288 void set_iff(IfNode* x) {_iff = x;}
1289 bool has_reserved() const { return _active && _has_reserved;}
1290 private:
1291 bool create_reserve();
1292 };// class CountedLoopReserveKit
1293
1294 inline Node* IdealLoopTree::tail() {
1295 // Handle lazy update of _tail field
1296 Node *n = _tail;
1297 //while( !n->in(0) ) // Skip dead CFG nodes
1298 //n = n->in(1);
1299 if (n->in(0) == NULL)
1300 n = _phase->get_ctrl(n);
1301 _tail = n;
1302 return n;
1303 }
1304
1305
1306 // Iterate over the loop tree using a preorder, left-to-right traversal.
1307 //
1308 // Example that visits all counted loops from within PhaseIdealLoop
1309 //
1310 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1311 // IdealLoopTree* lpt = iter.current();
1312 // if (!lpt->is_counted()) continue;
1313 // ...
1314 class LoopTreeIterator : public StackObj {
1315 private:
1316 IdealLoopTree* _root;
1317 IdealLoopTree* _curnt;
1318
1319 public:
1320 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1321
1322 bool done() { return _curnt == NULL; } // Finished iterating?
1323
1324 void next(); // Advance to next loop tree
1325
1326 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1327 };
1328
1329 #endif // SHARE_VM_OPTO_LOOPNODE_HPP