1 /*
2 * Copyright (c) 1998, 2013, 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 short _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 char _unswitch_count;
72 enum { _unswitch_max=3 };
73
74 public:
75 // Names for edge indices
76 enum { Self=0, EntryControl, LoopBackControl };
77
78 int is_inner_loop() const { return _loop_flags & InnerLoop; }
79 void set_inner_loop() { _loop_flags |= InnerLoop; }
80
81 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
82 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
83 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
84 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
85 void mark_has_reductions() { _loop_flags |= HasReductions; }
86 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
87 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
88 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
89
90 int unswitch_max() { return _unswitch_max; }
91 int unswitch_count() { return _unswitch_count; }
92 void set_unswitch_count(int val) {
93 assert (val <= unswitch_max(), "too many unswitches");
94 _unswitch_count = val;
95 }
96
97 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
98 init_class_id(Class_Loop);
99 init_req(EntryControl, entry);
100 init_req(LoopBackControl, backedge);
101 }
102
103 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
104 virtual int Opcode() const;
105 bool can_be_counted_loop(PhaseTransform* phase) const {
106 return req() == 3 && in(0) != NULL &&
107 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
108 in(2) != NULL && phase->type(in(2)) != Type::TOP;
109 }
110 bool is_valid_counted_loop() const;
111 #ifndef PRODUCT
112 virtual void dump_spec(outputStream *st) const;
113 #endif
114 };
115
116 //------------------------------Counted Loops----------------------------------
117 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
118 // path (and maybe some other exit paths). The trip-counter exit is always
119 // last in the loop. The trip-counter have to stride by a constant;
120 // the exit value is also loop invariant.
121
122 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
123 // CountedLoopNode has the incoming loop control and the loop-back-control
124 // which is always the IfTrue before the matching CountedLoopEndNode. The
125 // CountedLoopEndNode has an incoming control (possibly not the
126 // CountedLoopNode if there is control flow in the loop), the post-increment
127 // trip-counter value, and the limit. The trip-counter value is always of
128 // the form (Op old-trip-counter stride). The old-trip-counter is produced
129 // by a Phi connected to the CountedLoopNode. The stride is constant.
130 // The Op is any commutable opcode, including Add, Mul, Xor. The
131 // CountedLoopEndNode also takes in the loop-invariant limit value.
132
133 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
134 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
135 // via the old-trip-counter from the Op node.
136
137 //------------------------------CountedLoopNode--------------------------------
138 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
139 // inputs the incoming loop-start control and the loop-back control, so they
140 // act like RegionNodes. They also take in the initial trip counter, the
141 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
142 // produce a loop-body control and the trip counter value. Since
143 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
144
145 class CountedLoopNode : public LoopNode {
146 // Size is bigger to hold _main_idx. However, _main_idx does not change
147 // the semantics so it does not appear in the hash & cmp functions.
148 virtual uint size_of() const { return sizeof(*this); }
149
150 // For Pre- and Post-loops during debugging ONLY, this holds the index of
151 // the Main CountedLoop. Used to assert that we understand the graph shape.
152 node_idx_t _main_idx;
153
154 // Known trip count calculated by compute_exact_trip_count()
155 uint _trip_count;
156
157 // Expected trip count from profile data
158 float _profile_trip_cnt;
159
160 // Log2 of original loop bodies in unrolled loop
161 int _unrolled_count_log2;
162
163 // Node count prior to last unrolling - used to decide if
164 // unroll,optimize,unroll,optimize,... is making progress
165 int _node_count_before_unroll;
166
167 // If slp analysis is performed we record the maximum
168 // vector mapped unroll factor here
169 int _slp_maximum_unroll_factor;
170
171 public:
172 CountedLoopNode( Node *entry, Node *backedge )
173 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
174 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
175 _node_count_before_unroll(0), _slp_maximum_unroll_factor(0) {
176 init_class_id(Class_CountedLoop);
177 // Initialize _trip_count to the largest possible value.
178 // Will be reset (lower) if the loop's trip count is known.
179 }
180
181 virtual int Opcode() const;
182 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
183
184 Node *init_control() const { return in(EntryControl); }
185 Node *back_control() const { return in(LoopBackControl); }
186 CountedLoopEndNode *loopexit() const;
187 Node *init_trip() const;
188 Node *stride() const;
189 int stride_con() const;
190 bool stride_is_con() const;
191 Node *limit() const;
192 Node *incr() const;
193 Node *phi() const;
194
195 // Match increment with optional truncation
196 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
197
198 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
199 // can run short a few iterations and may start a few iterations in.
200 // It will be RCE'd and unrolled and aligned.
201
202 // A following 'post' loop will run any remaining iterations. Used
203 // during Range Check Elimination, the 'post' loop will do any final
204 // iterations with full checks. Also used by Loop Unrolling, where
205 // the 'post' loop will do any epilog iterations needed. Basically,
206 // a 'post' loop can not profitably be further unrolled or RCE'd.
207
208 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
209 // it may do under-flow checks for RCE and may do alignment iterations
210 // so the following main loop 'knows' that it is striding down cache
211 // lines.
212
213 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
214 // Aligned, may be missing it's pre-loop.
215 int is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
216 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
217 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
218 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
219 int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
220 int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
221 int has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
222 int do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
223 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
224 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
225
226 int main_idx() const { return _main_idx; }
227
228
229 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
230 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
231 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
232 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
233
234 void set_trip_count(uint tc) { _trip_count = tc; }
235 uint trip_count() { return _trip_count; }
236
237 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
238 void set_exact_trip_count(uint tc) {
239 _trip_count = tc;
240 _loop_flags |= HasExactTripCount;
241 }
242 void set_nonexact_trip_count() {
243 _loop_flags &= ~HasExactTripCount;
244 }
245 void set_notpassed_slp() {
246 _loop_flags &= ~PassedSlpAnalysis;
247 }
248
249 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
250 float profile_trip_cnt() { return _profile_trip_cnt; }
251
252 void double_unrolled_count() { _unrolled_count_log2++; }
253 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
254
255 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
256 int node_count_before_unroll() { return _node_count_before_unroll; }
257 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
258 int slp_max_unroll() const { return _slp_maximum_unroll_factor; }
259
260 #ifndef PRODUCT
261 virtual void dump_spec(outputStream *st) const;
262 #endif
263 };
264
265 //------------------------------CountedLoopEndNode-----------------------------
266 // CountedLoopEndNodes end simple trip counted loops. They act much like
267 // IfNodes.
268 class CountedLoopEndNode : public IfNode {
269 public:
270 enum { TestControl, TestValue };
271
272 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
273 : IfNode( control, test, prob, cnt) {
274 init_class_id(Class_CountedLoopEnd);
275 }
276 virtual int Opcode() const;
277
278 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
279 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
280 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
281 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
282 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
283 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
284 int stride_con() const;
285 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
286 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
287 CountedLoopNode *loopnode() const {
288 // The CountedLoopNode that goes with this CountedLoopEndNode may
289 // have been optimized out by the IGVN so be cautious with the
290 // pattern matching on the graph
291 if (phi() == NULL) {
292 return NULL;
293 }
294 assert(phi()->is_Phi(), "should be PhiNode");
295 Node *ln = phi()->in(0);
296 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
297 return (CountedLoopNode*)ln;
298 }
299 return NULL;
300 }
301
302 #ifndef PRODUCT
303 virtual void dump_spec(outputStream *st) const;
304 #endif
305 };
306
307
308 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
309 Node *bc = back_control();
310 if( bc == NULL ) return NULL;
311 Node *le = bc->in(0);
312 if( le->Opcode() != Op_CountedLoopEnd )
313 return NULL;
314 return (CountedLoopEndNode*)le;
315 }
316 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
317 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
318 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
319 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
320 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
321 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
322 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
323
324 //------------------------------LoopLimitNode-----------------------------
325 // Counted Loop limit node which represents exact final iterator value:
326 // trip_count = (limit - init_trip + stride - 1)/stride
327 // final_value= trip_count * stride + init_trip.
328 // Use HW instructions to calculate it when it can overflow in integer.
329 // Note, final_value should fit into integer since counted loop has
330 // limit check: limit <= max_int-stride.
331 class LoopLimitNode : public Node {
332 enum { Init=1, Limit=2, Stride=3 };
333 public:
334 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
335 // Put it on the Macro nodes list to optimize during macro nodes expansion.
336 init_flags(Flag_is_macro);
337 C->add_macro_node(this);
338 }
339 virtual int Opcode() const;
340 virtual const Type *bottom_type() const { return TypeInt::INT; }
341 virtual uint ideal_reg() const { return Op_RegI; }
342 virtual const Type *Value( PhaseTransform *phase ) const;
343 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
344 virtual Node *Identity( PhaseTransform *phase );
345 };
346
347 // -----------------------------IdealLoopTree----------------------------------
348 class IdealLoopTree : public ResourceObj {
349 public:
350 IdealLoopTree *_parent; // Parent in loop tree
351 IdealLoopTree *_next; // Next sibling in loop tree
352 IdealLoopTree *_child; // First child in loop tree
353
354 // The head-tail backedge defines the loop.
355 // If tail is NULL then this loop has multiple backedges as part of the
356 // same loop. During cleanup I'll peel off the multiple backedges; merge
357 // them at the loop bottom and flow 1 real backedge into the loop.
358 Node *_head; // Head of loop
359 Node *_tail; // Tail of loop
360 inline Node *tail(); // Handle lazy update of _tail field
361 PhaseIdealLoop* _phase;
362 int _local_loop_unroll_limit;
363 int _local_loop_unroll_factor;
364
365 Node_List _body; // Loop body for inner loops
366
367 uint8_t _nest; // Nesting depth
368 uint8_t _irreducible:1, // True if irreducible
369 _has_call:1, // True if has call safepoint
370 _has_sfpt:1, // True if has non-call safepoint
371 _rce_candidate:1; // True if candidate for range check elimination
372
373 Node_List* _safepts; // List of safepoints in this loop
374 Node_List* _required_safept; // A inner loop cannot delete these safepts;
375 bool _allow_optimizations; // Allow loop optimizations
376
377 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
378 : _parent(0), _next(0), _child(0),
379 _head(head), _tail(tail),
380 _phase(phase),
381 _safepts(NULL),
382 _required_safept(NULL),
383 _allow_optimizations(true),
384 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
385 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0)
386 { }
387
388 // Is 'l' a member of 'this'?
389 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
390
391 // Set loop nesting depth. Accumulate has_call bits.
392 int set_nest( uint depth );
393
394 // Split out multiple fall-in edges from the loop header. Move them to a
395 // private RegionNode before the loop. This becomes the loop landing pad.
396 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
397
398 // Split out the outermost loop from this shared header.
399 void split_outer_loop( PhaseIdealLoop *phase );
400
401 // Merge all the backedges from the shared header into a private Region.
402 // Feed that region as the one backedge to this loop.
403 void merge_many_backedges( PhaseIdealLoop *phase );
404
405 // Split shared headers and insert loop landing pads.
406 // Insert a LoopNode to replace the RegionNode.
407 // Returns TRUE if loop tree is structurally changed.
408 bool beautify_loops( PhaseIdealLoop *phase );
409
410 // Perform optimization to use the loop predicates for null checks and range checks.
411 // Applies to any loop level (not just the innermost one)
412 bool loop_predication( PhaseIdealLoop *phase);
413
414 // Perform iteration-splitting on inner loops. Split iterations to
415 // avoid range checks or one-shot null checks. Returns false if the
416 // current round of loop opts should stop.
417 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
418
419 // Driver for various flavors of iteration splitting. Returns false
420 // if the current round of loop opts should stop.
421 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
422
423 // Given dominators, try to find loops with calls that must always be
424 // executed (call dominates loop tail). These loops do not need non-call
425 // safepoints (ncsfpt).
426 void check_safepts(VectorSet &visited, Node_List &stack);
427
428 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
429 // encountered.
430 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
431
432 // Convert to counted loops where possible
433 void counted_loop( PhaseIdealLoop *phase );
434
435 // Check for Node being a loop-breaking test
436 Node *is_loop_exit(Node *iff) const;
437
438 // Returns true if ctrl is executed on every complete iteration
439 bool dominates_backedge(Node* ctrl);
440
441 // Remove simplistic dead code from loop body
442 void DCE_loop_body();
443
444 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
445 // Replace with a 1-in-10 exit guess.
446 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
447
448 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
449 // Useful for unrolling loops with NO array accesses.
450 bool policy_peel_only( PhaseIdealLoop *phase ) const;
451
452 // Return TRUE or FALSE if the loop should be unswitched -- clone
453 // loop with an invariant test
454 bool policy_unswitching( PhaseIdealLoop *phase ) const;
455
456 // Micro-benchmark spamming. Remove empty loops.
457 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
458
459 // Convert one iteration loop into normal code.
460 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
461
462 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
463 // make some loop-invariant test (usually a null-check) happen before the
464 // loop.
465 bool policy_peeling( PhaseIdealLoop *phase ) const;
466
467 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
468 // known trip count in the counted loop node.
469 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
470
471 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
472 // the loop is a CountedLoop and the body is small enough.
473 bool policy_unroll(PhaseIdealLoop *phase);
474
475 // Loop analyses to map to a maximal superword unrolling for vectorization.
476 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
477
478 // Return TRUE or FALSE if the loop should be range-check-eliminated.
479 // Gather a list of IF tests that are dominated by iteration splitting;
480 // also gather the end of the first split and the start of the 2nd split.
481 bool policy_range_check( PhaseIdealLoop *phase ) const;
482
483 // Return TRUE or FALSE if the loop should be cache-line aligned.
484 // Gather the expression that does the alignment. Note that only
485 // one array base can be aligned in a loop (unless the VM guarantees
486 // mutual alignment). Note that if we vectorize short memory ops
487 // into longer memory ops, we may want to increase alignment.
488 bool policy_align( PhaseIdealLoop *phase ) const;
489
490 // Return TRUE if "iff" is a range check.
491 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
492
493 // Compute loop exact trip count if possible
494 void compute_exact_trip_count( PhaseIdealLoop *phase );
495
496 // Compute loop trip count from profile data
497 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
498
499 // Reassociate invariant expressions.
500 void reassociate_invariants(PhaseIdealLoop *phase);
501 // Reassociate invariant add and subtract expressions.
502 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
503 // Return nonzero index of invariant operand if invariant and variant
504 // are combined with an Add or Sub. Helper for reassociate_invariants.
505 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
506
507 // Return true if n is invariant
508 bool is_invariant(Node* n) const;
509
510 // Put loop body on igvn work list
511 void record_for_igvn();
512
513 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
514 bool is_inner() { return is_loop() && _child == NULL; }
515 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
516
517 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
518
519 #ifndef PRODUCT
520 void dump_head( ) const; // Dump loop head only
521 void dump() const; // Dump this loop recursively
522 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
523 #endif
524
525 };
526
527 // -----------------------------PhaseIdealLoop---------------------------------
528 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
529 // loop tree. Drives the loop-based transformations on the ideal graph.
530 class PhaseIdealLoop : public PhaseTransform {
531 friend class IdealLoopTree;
532 friend class SuperWord;
533 friend class CountedLoopReserveKit;
534
535 // Pre-computed def-use info
536 PhaseIterGVN &_igvn;
537
538 // Head of loop tree
539 IdealLoopTree *_ltree_root;
540
541 // Array of pre-order numbers, plus post-visited bit.
542 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
543 // ODD for post-visited. Other bits are the pre-order number.
544 uint *_preorders;
545 uint _max_preorder;
546
547 const PhaseIdealLoop* _verify_me;
548 bool _verify_only;
549
550 // Allocate _preorders[] array
551 void allocate_preorders() {
552 _max_preorder = C->unique()+8;
553 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
554 memset(_preorders, 0, sizeof(uint) * _max_preorder);
555 }
556
557 // Allocate _preorders[] array
558 void reallocate_preorders() {
559 if ( _max_preorder < C->unique() ) {
560 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
561 _max_preorder = C->unique();
562 }
563 memset(_preorders, 0, sizeof(uint) * _max_preorder);
564 }
565
566 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
567 // adds new nodes.
568 void check_grow_preorders( ) {
569 if ( _max_preorder < C->unique() ) {
570 uint newsize = _max_preorder<<1; // double size of array
571 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
572 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
573 _max_preorder = newsize;
574 }
575 }
576 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
577 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
578 // Pre-order numbers are written to the Nodes array as low-bit-set values.
579 void set_preorder_visited( Node *n, int pre_order ) {
580 assert( !is_visited( n ), "already set" );
581 _preorders[n->_idx] = (pre_order<<1);
582 };
583 // Return pre-order number.
584 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
585
586 // Check for being post-visited.
587 // Should be previsited already (checked with assert(is_visited(n))).
588 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
589
590 // Mark as post visited
591 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
592
593 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
594 // Returns true if "n" is a data node, false if it's a control node.
595 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
596
597 // clear out dead code after build_loop_late
598 Node_List _deadlist;
599
600 // Support for faster execution of get_late_ctrl()/dom_lca()
601 // when a node has many uses and dominator depth is deep.
602 Node_Array _dom_lca_tags;
603 void init_dom_lca_tags();
604 void clear_dom_lca_tags();
605
606 // Helper for debugging bad dominance relationships
607 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
608
609 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
610
611 // Inline wrapper for frequent cases:
612 // 1) only one use
613 // 2) a use is the same as the current LCA passed as 'n1'
614 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
615 assert( n->is_CFG(), "" );
616 // Fast-path NULL lca
617 if( lca != NULL && lca != n ) {
618 assert( lca->is_CFG(), "" );
619 // find LCA of all uses
620 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
621 }
622 return find_non_split_ctrl(n);
623 }
624 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
625
626 // Helper function for directing control inputs away from CFG split
627 // points.
628 Node *find_non_split_ctrl( Node *ctrl ) const {
629 if (ctrl != NULL) {
630 if (ctrl->is_MultiBranch()) {
631 ctrl = ctrl->in(0);
632 }
633 assert(ctrl->is_CFG(), "CFG");
634 }
635 return ctrl;
636 }
637
638 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
639
640 public:
641 bool has_node( Node* n ) const {
642 guarantee(n != NULL, "No Node.");
643 return _nodes[n->_idx] != NULL;
644 }
645 // check if transform created new nodes that need _ctrl recorded
646 Node *get_late_ctrl( Node *n, Node *early );
647 Node *get_early_ctrl( Node *n );
648 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
649 void set_early_ctrl( Node *n );
650 void set_subtree_ctrl( Node *root );
651 void set_ctrl( Node *n, Node *ctrl ) {
652 assert( !has_node(n) || has_ctrl(n), "" );
653 assert( ctrl->in(0), "cannot set dead control node" );
654 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
655 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
656 }
657 // Set control and update loop membership
658 void set_ctrl_and_loop(Node* n, Node* ctrl) {
659 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
660 IdealLoopTree* new_loop = get_loop(ctrl);
661 if (old_loop != new_loop) {
662 if (old_loop->_child == NULL) old_loop->_body.yank(n);
663 if (new_loop->_child == NULL) new_loop->_body.push(n);
664 }
665 set_ctrl(n, ctrl);
666 }
667 // Control nodes can be replaced or subsumed. During this pass they
668 // get their replacement Node in slot 1. Instead of updating the block
669 // location of all Nodes in the subsumed block, we lazily do it. As we
670 // pull such a subsumed block out of the array, we write back the final
671 // correct block.
672 Node *get_ctrl( Node *i ) {
673 assert(has_node(i), "");
674 Node *n = get_ctrl_no_update(i);
675 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
676 assert(has_node(i) && has_ctrl(i), "");
677 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
678 return n;
679 }
680 // true if CFG node d dominates CFG node n
681 bool is_dominator(Node *d, Node *n);
682 // return get_ctrl for a data node and self(n) for a CFG node
683 Node* ctrl_or_self(Node* n) {
684 if (has_ctrl(n))
685 return get_ctrl(n);
686 else {
687 assert (n->is_CFG(), "must be a CFG node");
688 return n;
689 }
690 }
691
692 private:
693 Node *get_ctrl_no_update( Node *i ) const {
694 assert( has_ctrl(i), "" );
695 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
696 if (!n->in(0)) {
697 // Skip dead CFG nodes
698 do {
699 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
700 } while (!n->in(0));
701 n = find_non_split_ctrl(n);
702 }
703 return n;
704 }
705
706 // Check for loop being set
707 // "n" must be a control node. Returns true if "n" is known to be in a loop.
708 bool has_loop( Node *n ) const {
709 assert(!has_node(n) || !has_ctrl(n), "");
710 return has_node(n);
711 }
712 // Set loop
713 void set_loop( Node *n, IdealLoopTree *loop ) {
714 _nodes.map(n->_idx, (Node*)loop);
715 }
716 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
717 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
718 // from old_node to new_node to support the lazy update. Reference
719 // replaces loop reference, since that is not needed for dead node.
720 public:
721 void lazy_update( Node *old_node, Node *new_node ) {
722 assert( old_node != new_node, "no cycles please" );
723 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
724 // Nodes always have DU info now, so re-use the side array slot
725 // for this node to provide the forwarding pointer.
726 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
727 }
728 void lazy_replace( Node *old_node, Node *new_node ) {
729 _igvn.replace_node( old_node, new_node );
730 lazy_update( old_node, new_node );
731 }
732 void lazy_replace_proj( Node *old_node, Node *new_node ) {
733 assert( old_node->req() == 1, "use this for Projs" );
734 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
735 old_node->add_req( NULL );
736 lazy_replace( old_node, new_node );
737 }
738
739 private:
740
741 // Place 'n' in some loop nest, where 'n' is a CFG node
742 void build_loop_tree();
743 int build_loop_tree_impl( Node *n, int pre_order );
744 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
745 // loop tree, not the root.
746 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
747
748 // Place Data nodes in some loop nest
749 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
750 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
751 void build_loop_late_post ( Node* n );
752
753 // Array of immediate dominance info for each CFG node indexed by node idx
754 private:
755 uint _idom_size;
756 Node **_idom; // Array of immediate dominators
757 uint *_dom_depth; // Used for fast LCA test
758 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
759
760 Node* idom_no_update(Node* d) const {
761 assert(d->_idx < _idom_size, "oob");
762 Node* n = _idom[d->_idx];
763 assert(n != NULL,"Bad immediate dominator info.");
764 while (n->in(0) == NULL) { // Skip dead CFG nodes
765 //n = n->in(1);
766 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
767 assert(n != NULL,"Bad immediate dominator info.");
768 }
769 return n;
770 }
771 Node *idom(Node* d) const {
772 uint didx = d->_idx;
773 Node *n = idom_no_update(d);
774 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
775 return n;
776 }
777 uint dom_depth(Node* d) const {
778 guarantee(d != NULL, "Null dominator info.");
779 guarantee(d->_idx < _idom_size, "");
780 return _dom_depth[d->_idx];
781 }
782 void set_idom(Node* d, Node* n, uint dom_depth);
783 // Locally compute IDOM using dom_lca call
784 Node *compute_idom( Node *region ) const;
785 // Recompute dom_depth
786 void recompute_dom_depth();
787
788 // Is safept not required by an outer loop?
789 bool is_deleteable_safept(Node* sfpt);
790
791 // Replace parallel induction variable (parallel to trip counter)
792 void replace_parallel_iv(IdealLoopTree *loop);
793
794 // Perform verification that the graph is valid.
795 PhaseIdealLoop( PhaseIterGVN &igvn) :
796 PhaseTransform(Ideal_Loop),
797 _igvn(igvn),
798 _dom_lca_tags(arena()), // Thread::resource_area
799 _verify_me(NULL),
800 _verify_only(true) {
801 build_and_optimize(false, false);
802 }
803
804 // build the loop tree and perform any requested optimizations
805 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
806
807 public:
808 // Dominators for the sea of nodes
809 void Dominators();
810 Node *dom_lca( Node *n1, Node *n2 ) const {
811 return find_non_split_ctrl(dom_lca_internal(n1, n2));
812 }
813 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
814
815 // Compute the Ideal Node to Loop mapping
816 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
817 PhaseTransform(Ideal_Loop),
818 _igvn(igvn),
819 _dom_lca_tags(arena()), // Thread::resource_area
820 _verify_me(NULL),
821 _verify_only(false) {
822 build_and_optimize(do_split_ifs, skip_loop_opts);
823 }
824
825 // Verify that verify_me made the same decisions as a fresh run.
826 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
827 PhaseTransform(Ideal_Loop),
828 _igvn(igvn),
829 _dom_lca_tags(arena()), // Thread::resource_area
830 _verify_me(verify_me),
831 _verify_only(false) {
832 build_and_optimize(false, false);
833 }
834
835 // Build and verify the loop tree without modifying the graph. This
836 // is useful to verify that all inputs properly dominate their uses.
837 static void verify(PhaseIterGVN& igvn) {
838 #ifdef ASSERT
839 PhaseIdealLoop v(igvn);
840 #endif
841 }
842
843 // True if the method has at least 1 irreducible loop
844 bool _has_irreducible_loops;
845
846 // Per-Node transform
847 virtual Node *transform( Node *a_node ) { return 0; }
848
849 bool is_counted_loop( Node *x, IdealLoopTree *loop );
850
851 Node* exact_limit( IdealLoopTree *loop );
852
853 // Return a post-walked LoopNode
854 IdealLoopTree *get_loop( Node *n ) const {
855 // Dead nodes have no loop, so return the top level loop instead
856 if (!has_node(n)) return _ltree_root;
857 assert(!has_ctrl(n), "");
858 return (IdealLoopTree*)_nodes[n->_idx];
859 }
860
861 // Is 'n' a (nested) member of 'loop'?
862 int is_member( const IdealLoopTree *loop, Node *n ) const {
863 return loop->is_member(get_loop(n)); }
864
865 // This is the basic building block of the loop optimizations. It clones an
866 // entire loop body. It makes an old_new loop body mapping; with this
867 // mapping you can find the new-loop equivalent to an old-loop node. All
868 // new-loop nodes are exactly equal to their old-loop counterparts, all
869 // edges are the same. All exits from the old-loop now have a RegionNode
870 // that merges the equivalent new-loop path. This is true even for the
871 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
872 // now come from (one or more) Phis that merge their new-loop equivalents.
873 // Parameter side_by_side_idom:
874 // When side_by_size_idom is NULL, the dominator tree is constructed for
875 // the clone loop to dominate the original. Used in construction of
876 // pre-main-post loop sequence.
877 // When nonnull, the clone and original are side-by-side, both are
878 // dominated by the passed in side_by_side_idom node. Used in
879 // construction of unswitched loops.
880 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
881 Node* side_by_side_idom = NULL);
882
883 // If we got the effect of peeling, either by actually peeling or by
884 // making a pre-loop which must execute at least once, we can remove
885 // all loop-invariant dominated tests in the main body.
886 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
887
888 // Generate code to do a loop peel for the given loop (and body).
889 // old_new is a temp array.
890 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
891
892 // Add pre and post loops around the given loop. These loops are used
893 // during RCE, unrolling and aligning loops.
894 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
895 // If Node n lives in the back_ctrl block, we clone a private version of n
896 // in preheader_ctrl block and return that, otherwise return n.
897 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
898
899 // Take steps to maximally unroll the loop. Peel any odd iterations, then
900 // unroll to do double iterations. The next round of major loop transforms
901 // will repeat till the doubled loop body does all remaining iterations in 1
902 // pass.
903 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
904
905 // Unroll the loop body one step - make each trip do 2 iterations.
906 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
907
908 // Mark vector reduction candidates before loop unrolling
909 void mark_reductions( IdealLoopTree *loop );
910
911 // Return true if exp is a constant times an induction var
912 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
913
914 // Return true if exp is a scaled induction var plus (or minus) constant
915 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
916
917 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
918 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
919 Deoptimization::DeoptReason reason,
920 int opcode);
921 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
922
923 // Clone loop predicates to cloned loops (peeled, unswitched)
924 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
925 Deoptimization::DeoptReason reason,
926 PhaseIdealLoop* loop_phase,
927 PhaseIterGVN* igvn);
928
929 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
930 bool clone_limit_check,
931 PhaseIdealLoop* loop_phase,
932 PhaseIterGVN* igvn);
933 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
934
935 static Node* skip_loop_predicates(Node* entry);
936
937 // Find a good location to insert a predicate
938 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
939 // Find a predicate
940 static Node* find_predicate(Node* entry);
941 // Construct a range check for a predicate if
942 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
943 int scale, Node* offset,
944 Node* init, Node* limit, Node* stride,
945 Node* range, bool upper);
946
947 // Implementation of the loop predication to promote checks outside the loop
948 bool loop_predication_impl(IdealLoopTree *loop);
949
950 // Helper function to collect predicate for eliminating the useless ones
951 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
952 void eliminate_useless_predicates();
953
954 // Change the control input of expensive nodes to allow commoning by
955 // IGVN when it is guaranteed to not result in a more frequent
956 // execution of the expensive node. Return true if progress.
957 bool process_expensive_nodes();
958
959 // Check whether node has become unreachable
960 bool is_node_unreachable(Node *n) const {
961 return !has_node(n) || n->is_unreachable(_igvn);
962 }
963
964 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
965 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
966
967 // Create a slow version of the loop by cloning the loop
968 // and inserting an if to select fast-slow versions.
969 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
970 Node_List &old_new,
971 int opcode);
972
973 // Clone a loop and return the clone head (clone_loop_head).
974 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
975 // This routine was created for usage in CountedLoopReserveKit.
976 //
977 // int(1) -> If -> IfTrue -> original_loop_head
978 // |
979 // V
980 // IfFalse -> clone_loop_head (returned by function pointer)
981 //
982 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
983 // Clone loop with an invariant test (that does not exit) and
984 // insert a clone of the test that selects which version to
985 // execute.
986 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
987
988 // Find candidate "if" for unswitching
989 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
990
991 // Range Check Elimination uses this function!
992 // Constrain the main loop iterations so the affine function:
993 // low_limit <= scale_con * I + offset < upper_limit
994 // always holds true. That is, either increase the number of iterations in
995 // the pre-loop or the post-loop until the condition holds true in the main
996 // loop. Scale_con, offset and limit are all loop invariant.
997 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 );
998 // Helper function for add_constraint().
999 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1000
1001 // Partially peel loop up through last_peel node.
1002 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1003
1004 // Create a scheduled list of nodes control dependent on ctrl set.
1005 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1006 // Has a use in the vector set
1007 bool has_use_in_set( Node* n, VectorSet& vset );
1008 // Has use internal to the vector set (ie. not in a phi at the loop head)
1009 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1010 // clone "n" for uses that are outside of loop
1011 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1012 // clone "n" for special uses that are in the not_peeled region
1013 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1014 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1015 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1016 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1017 #ifdef ASSERT
1018 // Validate the loop partition sets: peel and not_peel
1019 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1020 // Ensure that uses outside of loop are of the right form
1021 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1022 uint orig_exit_idx, uint clone_exit_idx);
1023 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1024 #endif
1025
1026 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1027 int stride_of_possible_iv( Node* iff );
1028 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1029 // Return the (unique) control output node that's in the loop (if it exists.)
1030 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1031 // Insert a signed compare loop exit cloned from an unsigned compare.
1032 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1033 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1034 // Utility to register node "n" with PhaseIdealLoop
1035 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1036 // Utility to create an if-projection
1037 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1038 // Force the iff control output to be the live_proj
1039 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1040 // Insert a region before an if projection
1041 RegionNode* insert_region_before_proj(ProjNode* proj);
1042 // Insert a new if before an if projection
1043 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1044
1045 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1046 // "Nearly" because all Nodes have been cloned from the original in the loop,
1047 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1048 // through the Phi recursively, and return a Bool.
1049 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1050 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1051
1052
1053 // Rework addressing expressions to get the most loop-invariant stuff
1054 // moved out. We'd like to do all associative operators, but it's especially
1055 // important (common) to do address expressions.
1056 Node *remix_address_expressions( Node *n );
1057
1058 // Attempt to use a conditional move instead of a phi/branch
1059 Node *conditional_move( Node *n );
1060
1061 // Reorganize offset computations to lower register pressure.
1062 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1063 // (which are then alive with the post-incremented trip counter
1064 // forcing an extra register move)
1065 void reorg_offsets( IdealLoopTree *loop );
1066
1067 // Check for aggressive application of 'split-if' optimization,
1068 // using basic block level info.
1069 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1070 Node *split_if_with_blocks_pre ( Node *n );
1071 void split_if_with_blocks_post( Node *n );
1072 Node *has_local_phi_input( Node *n );
1073 // Mark an IfNode as being dominated by a prior test,
1074 // without actually altering the CFG (and hence IDOM info).
1075 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1076
1077 // Split Node 'n' through merge point
1078 Node *split_thru_region( Node *n, Node *region );
1079 // Split Node 'n' through merge point if there is enough win.
1080 Node *split_thru_phi( Node *n, Node *region, int policy );
1081 // Found an If getting its condition-code input from a Phi in the
1082 // same block. Split thru the Region.
1083 void do_split_if( Node *iff );
1084
1085 // Conversion of fill/copy patterns into intrisic versions
1086 bool do_intrinsify_fill();
1087 bool intrinsify_fill(IdealLoopTree* lpt);
1088 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1089 Node*& shift, Node*& offset);
1090
1091 private:
1092 // Return a type based on condition control flow
1093 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1094 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1095 // Helpers for filtered type
1096 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1097
1098 // Helper functions
1099 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1100 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1101 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 );
1102 bool split_up( Node *n, Node *blk1, Node *blk2 );
1103 void sink_use( Node *use, Node *post_loop );
1104 Node *place_near_use( Node *useblock ) const;
1105 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1106 void try_move_store_after_loop(Node* n);
1107
1108 bool _created_loop_node;
1109 public:
1110 void set_created_loop_node() { _created_loop_node = true; }
1111 bool created_loop_node() { return _created_loop_node; }
1112 void register_new_node( Node *n, Node *blk );
1113
1114 #ifdef ASSERT
1115 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1116 #endif
1117
1118 #ifndef PRODUCT
1119 void dump( ) const;
1120 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1121 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1122 void verify() const; // Major slow :-)
1123 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1124 IdealLoopTree *get_loop_idx(Node* n) const {
1125 // Dead nodes have no loop, so return the top level loop instead
1126 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1127 }
1128 // Print some stats
1129 static void print_statistics();
1130 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1131 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1132 #endif
1133 };
1134
1135 // This kit may be used for making of a reserved copy of a loop before this loop
1136 // goes under non-reversible changes.
1137 //
1138 // Function create_reserve() creates a reserved copy (clone) of the loop.
1139 // The reserved copy is created by calling
1140 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1141 // the original and reserved loops are connected in the outer graph.
1142 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1143 //
1144 // By default the reserved copy (clone) of the loop is created as dead code - it is
1145 // dominated in the outer loop by this node chain:
1146 // intcon(1)->If->IfFalse->reserved_copy.
1147 // The original loop is dominated by the the same node chain but IfTrue projection:
1148 // intcon(0)->If->IfTrue->original_loop.
1149 //
1150 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1151 // and the dtor, checks _use_new value.
1152 // If _use_new == false, it "switches" control to reserved copy of the loop
1153 // by simple replacing of node intcon(1) with node intcon(0).
1154 //
1155 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1156 //
1157 // void CountedLoopReserveKit_example()
1158 // {
1159 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1160 // if (DoReserveCopy && !lrk.has_reserved()) {
1161 // return; //failed to create reserved loop copy
1162 // }
1163 // ...
1164 // //something is wrong, switch to original loop
1165 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1166 // ...
1167 // //everything worked ok, return with the newly modified loop
1168 // lrk.use_new();
1169 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1170 // }
1171 //
1172 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1173 // the dtor will "switch to the original" loop.
1174 // NOTE. You you modify outside of the original loop this class is no help.
1175 //
1176 class CountedLoopReserveKit {
1177 private:
1178 PhaseIdealLoop* _phase;
1179 IdealLoopTree* _lpt;
1180 LoopNode* _lp;
1181 IfNode* _iff;
1182 LoopNode* _lp_reserved;
1183 bool _has_reserved;
1184 bool _use_new;
1185 const bool _active; //may be set to false in ctor, then the object is dummy
1186
1187 public:
1188 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1189 ~CountedLoopReserveKit();
1190 void use_new() {_use_new = true;}
1191 void set_iff(IfNode* x) {_iff = x;}
1192 bool has_reserved() const { return _active && _has_reserved;}
1193 private:
1194 bool create_reserve();
1195 };// class CountedLoopReserveKit
1196
1197 inline Node* IdealLoopTree::tail() {
1198 // Handle lazy update of _tail field
1199 Node *n = _tail;
1200 //while( !n->in(0) ) // Skip dead CFG nodes
1201 //n = n->in(1);
1202 if (n->in(0) == NULL)
1203 n = _phase->get_ctrl(n);
1204 _tail = n;
1205 return n;
1206 }
1207
1208
1209 // Iterate over the loop tree using a preorder, left-to-right traversal.
1210 //
1211 // Example that visits all counted loops from within PhaseIdealLoop
1212 //
1213 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1214 // IdealLoopTree* lpt = iter.current();
1215 // if (!lpt->is_counted()) continue;
1216 // ...
1217 class LoopTreeIterator : public StackObj {
1218 private:
1219 IdealLoopTree* _root;
1220 IdealLoopTree* _curnt;
1221
1222 public:
1223 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1224
1225 bool done() { return _curnt == NULL; } // Finished iterating?
1226
1227 void next(); // Advance to next loop tree
1228
1229 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1230 };
1231
1232 #endif // SHARE_VM_OPTO_LOOPNODE_HPP