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