1 /*
2 * Copyright (c) 1998, 2017, 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.
size_of() const56 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 char _unswitch_count;
67 enum { _unswitch_max=3 };
68
69 public:
70 // Names for edge indices
71 enum { Self=0, EntryControl, LoopBackControl };
72
is_inner_loop() const73 int is_inner_loop() const { return _loop_flags & InnerLoop; }
set_inner_loop()74 void set_inner_loop() { _loop_flags |= InnerLoop; }
75
is_partial_peel_loop() const76 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
set_partial_peel_loop()77 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
partial_peel_has_failed() const78 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
mark_partial_peel_failed()79 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
80
unswitch_max()81 int unswitch_max() { return _unswitch_max; }
unswitch_count()82 int unswitch_count() { return _unswitch_count; }
set_unswitch_count(int val)83 void set_unswitch_count(int val) {
84 assert (val <= unswitch_max(), "too many unswitches");
85 _unswitch_count = val;
86 }
87
LoopNode(Node * entry,Node * backedge)88 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
89 init_class_id(Class_Loop);
90 init_req(EntryControl, entry);
91 init_req(LoopBackControl, backedge);
92 }
93
94 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
95 virtual int Opcode() const;
can_be_counted_loop(PhaseTransform * phase) const96 bool can_be_counted_loop(PhaseTransform* phase) const {
97 return req() == 3 && in(0) != NULL &&
98 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
99 in(2) != NULL && phase->type(in(2)) != Type::TOP;
100 }
101 bool is_valid_counted_loop() const;
102 #ifndef PRODUCT
103 virtual void dump_spec(outputStream *st) const;
104 #endif
105 };
106
107 //------------------------------Counted Loops----------------------------------
108 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
109 // path (and maybe some other exit paths). The trip-counter exit is always
110 // last in the loop. The trip-counter have to stride by a constant;
111 // the exit value is also loop invariant.
112
113 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
114 // CountedLoopNode has the incoming loop control and the loop-back-control
115 // which is always the IfTrue before the matching CountedLoopEndNode. The
116 // CountedLoopEndNode has an incoming control (possibly not the
117 // CountedLoopNode if there is control flow in the loop), the post-increment
118 // trip-counter value, and the limit. The trip-counter value is always of
119 // the form (Op old-trip-counter stride). The old-trip-counter is produced
120 // by a Phi connected to the CountedLoopNode. The stride is constant.
121 // The Op is any commutable opcode, including Add, Mul, Xor. The
122 // CountedLoopEndNode also takes in the loop-invariant limit value.
123
124 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
125 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
126 // via the old-trip-counter from the Op node.
127
128 //------------------------------CountedLoopNode--------------------------------
129 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
130 // inputs the incoming loop-start control and the loop-back control, so they
131 // act like RegionNodes. They also take in the initial trip counter, the
132 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
133 // produce a loop-body control and the trip counter value. Since
134 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
135
136 class CountedLoopNode : public LoopNode {
137 // Size is bigger to hold _main_idx. However, _main_idx does not change
138 // the semantics so it does not appear in the hash & cmp functions.
size_of() const139 virtual uint size_of() const { return sizeof(*this); }
140
141 // For Pre- and Post-loops during debugging ONLY, this holds the index of
142 // the Main CountedLoop. Used to assert that we understand the graph shape.
143 node_idx_t _main_idx;
144
145 // Known trip count calculated by compute_exact_trip_count()
146 uint _trip_count;
147
148 // Expected trip count from profile data
149 float _profile_trip_cnt;
150
151 // Log2 of original loop bodies in unrolled loop
152 int _unrolled_count_log2;
153
154 // Node count prior to last unrolling - used to decide if
155 // unroll,optimize,unroll,optimize,... is making progress
156 int _node_count_before_unroll;
157
158 public:
CountedLoopNode(Node * entry,Node * backedge)159 CountedLoopNode( Node *entry, Node *backedge )
160 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
161 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
162 _node_count_before_unroll(0) {
163 init_class_id(Class_CountedLoop);
164 // Initialize _trip_count to the largest possible value.
165 // Will be reset (lower) if the loop's trip count is known.
166 }
167
168 virtual int Opcode() const;
169 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
170
init_control() const171 Node *init_control() const { return in(EntryControl); }
back_control() const172 Node *back_control() const { return in(LoopBackControl); }
173 CountedLoopEndNode *loopexit() const;
174 Node *init_trip() const;
175 Node *stride() const;
176 int stride_con() const;
177 bool stride_is_con() const;
178 Node *limit() const;
179 Node *incr() const;
180 Node *phi() const;
181
182 // Match increment with optional truncation
183 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
184
185 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
186 // can run short a few iterations and may start a few iterations in.
187 // It will be RCE'd and unrolled and aligned.
188
189 // A following 'post' loop will run any remaining iterations. Used
190 // during Range Check Elimination, the 'post' loop will do any final
191 // iterations with full checks. Also used by Loop Unrolling, where
192 // the 'post' loop will do any epilog iterations needed. Basically,
193 // a 'post' loop can not profitably be further unrolled or RCE'd.
194
195 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
196 // it may do under-flow checks for RCE and may do alignment iterations
197 // so the following main loop 'knows' that it is striding down cache
198 // lines.
199
200 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
201 // Aligned, may be missing it's pre-loop.
is_normal_loop() const202 int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
is_pre_loop() const203 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
is_main_loop() const204 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
is_post_loop() const205 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
is_main_no_pre_loop() const206 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
set_main_no_pre_loop()207 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
208
main_idx() const209 int main_idx() const { return _main_idx; }
210
211
set_pre_loop(CountedLoopNode * main)212 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
set_main_loop()213 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
set_post_loop(CountedLoopNode * main)214 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
set_normal_loop()215 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
216
set_trip_count(uint tc)217 void set_trip_count(uint tc) { _trip_count = tc; }
trip_count()218 uint trip_count() { return _trip_count; }
219
has_exact_trip_count() const220 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
set_exact_trip_count(uint tc)221 void set_exact_trip_count(uint tc) {
222 _trip_count = tc;
223 _loop_flags |= HasExactTripCount;
224 }
set_nonexact_trip_count()225 void set_nonexact_trip_count() {
226 _loop_flags &= ~HasExactTripCount;
227 }
228
set_profile_trip_cnt(float ptc)229 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
profile_trip_cnt()230 float profile_trip_cnt() { return _profile_trip_cnt; }
231
double_unrolled_count()232 void double_unrolled_count() { _unrolled_count_log2++; }
unrolled_count()233 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
234
set_node_count_before_unroll(int ct)235 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
node_count_before_unroll()236 int node_count_before_unroll() { return _node_count_before_unroll; }
237
238 #ifndef PRODUCT
239 virtual void dump_spec(outputStream *st) const;
240 #endif
241 };
242
243 //------------------------------CountedLoopEndNode-----------------------------
244 // CountedLoopEndNodes end simple trip counted loops. They act much like
245 // IfNodes.
246 class CountedLoopEndNode : public IfNode {
247 public:
248 enum { TestControl, TestValue };
249
CountedLoopEndNode(Node * control,Node * test,float prob,float cnt)250 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
251 : IfNode( control, test, prob, cnt) {
252 init_class_id(Class_CountedLoopEnd);
253 }
254 virtual int Opcode() const;
255
cmp_node() const256 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
incr() const257 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
limit() const258 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
stride() const259 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
init_trip() const260 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
261 int stride_con() const;
stride_is_con() const262 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
test_trip() const263 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
phi() const264 PhiNode *phi() const {
265 Node *tmp = incr();
266 if (tmp && tmp->req() == 3) {
267 Node* phi = tmp->in(1);
268 if (phi->is_Phi()) {
269 return phi->as_Phi();
270 }
271 }
272 return NULL;
273 }
loopnode() const274 CountedLoopNode *loopnode() const {
275 // The CountedLoopNode that goes with this CountedLoopEndNode may
276 // have been optimized out by the IGVN so be cautious with the
277 // pattern matching on the graph
278 PhiNode* iv_phi = phi();
279 if (iv_phi == NULL) {
280 return NULL;
281 }
282 assert(iv_phi->is_Phi(), "should be PhiNode");
283 Node *ln = iv_phi->in(0);
284 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
285 return (CountedLoopNode*)ln;
286 }
287 return NULL;
288 }
289
290 #ifndef PRODUCT
291 virtual void dump_spec(outputStream *st) const;
292 #endif
293 };
294
295
loopexit() const296 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
297 Node *bc = back_control();
298 if( bc == NULL ) return NULL;
299 Node *le = bc->in(0);
300 if( le->Opcode() != Op_CountedLoopEnd )
301 return NULL;
302 return (CountedLoopEndNode*)le;
303 }
init_trip() const304 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
stride() const305 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
stride_con() const306 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
stride_is_con() const307 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
limit() const308 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
incr() const309 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
phi() const310 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
311
312 //------------------------------LoopLimitNode-----------------------------
313 // Counted Loop limit node which represents exact final iterator value:
314 // trip_count = (limit - init_trip + stride - 1)/stride
315 // final_value= trip_count * stride + init_trip.
316 // Use HW instructions to calculate it when it can overflow in integer.
317 // Note, final_value should fit into integer since counted loop has
318 // limit check: limit <= max_int-stride.
319 class LoopLimitNode : public Node {
320 enum { Init=1, Limit=2, Stride=3 };
321 public:
LoopLimitNode(Compile * C,Node * init,Node * limit,Node * stride)322 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
323 // Put it on the Macro nodes list to optimize during macro nodes expansion.
324 init_flags(Flag_is_macro);
325 C->add_macro_node(this);
326 }
327 virtual int Opcode() const;
bottom_type() const328 virtual const Type *bottom_type() const { return TypeInt::INT; }
ideal_reg() const329 virtual uint ideal_reg() const { return Op_RegI; }
330 virtual const Type *Value( PhaseTransform *phase ) const;
331 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
332 virtual Node *Identity( PhaseTransform *phase );
333 };
334
335 // -----------------------------IdealLoopTree----------------------------------
336 class IdealLoopTree : public ResourceObj {
337 public:
338 IdealLoopTree *_parent; // Parent in loop tree
339 IdealLoopTree *_next; // Next sibling in loop tree
340 IdealLoopTree *_child; // First child in loop tree
341
342 // The head-tail backedge defines the loop.
343 // If tail is NULL then this loop has multiple backedges as part of the
344 // same loop. During cleanup I'll peel off the multiple backedges; merge
345 // them at the loop bottom and flow 1 real backedge into the loop.
346 Node *_head; // Head of loop
347 Node *_tail; // Tail of loop
348 inline Node *tail(); // Handle lazy update of _tail field
349 PhaseIdealLoop* _phase;
350
351 Node_List _body; // Loop body for inner loops
352
353 uint8 _nest; // Nesting depth
354 uint8 _irreducible:1, // True if irreducible
355 _has_call:1, // True if has call safepoint
356 _has_sfpt:1, // True if has non-call safepoint
357 _rce_candidate:1; // True if candidate for range check elimination
358
359 Node_List* _safepts; // List of safepoints in this loop
360 Node_List* _required_safept; // A inner loop cannot delete these safepts;
361 bool _allow_optimizations; // Allow loop optimizations
362
IdealLoopTree(PhaseIdealLoop * phase,Node * head,Node * tail)363 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
364 : _parent(0), _next(0), _child(0),
365 _head(head), _tail(tail),
366 _phase(phase),
367 _safepts(NULL),
368 _required_safept(NULL),
369 _allow_optimizations(true),
370 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
371 { }
372
373 // Is 'l' a member of 'this'?
374 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
375
376 // Set loop nesting depth. Accumulate has_call bits.
377 int set_nest( uint depth );
378
379 // Split out multiple fall-in edges from the loop header. Move them to a
380 // private RegionNode before the loop. This becomes the loop landing pad.
381 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
382
383 // Split out the outermost loop from this shared header.
384 void split_outer_loop( PhaseIdealLoop *phase );
385
386 // Merge all the backedges from the shared header into a private Region.
387 // Feed that region as the one backedge to this loop.
388 void merge_many_backedges( PhaseIdealLoop *phase );
389
390 // Split shared headers and insert loop landing pads.
391 // Insert a LoopNode to replace the RegionNode.
392 // Returns TRUE if loop tree is structurally changed.
393 bool beautify_loops( PhaseIdealLoop *phase );
394
395 // Perform optimization to use the loop predicates for null checks and range checks.
396 // Applies to any loop level (not just the innermost one)
397 bool loop_predication( PhaseIdealLoop *phase);
398
399 // Perform iteration-splitting on inner loops. Split iterations to
400 // avoid range checks or one-shot null checks. Returns false if the
401 // current round of loop opts should stop.
402 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
403
404 // Driver for various flavors of iteration splitting. Returns false
405 // if the current round of loop opts should stop.
406 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
407
408 // Given dominators, try to find loops with calls that must always be
409 // executed (call dominates loop tail). These loops do not need non-call
410 // safepoints (ncsfpt).
411 void check_safepts(VectorSet &visited, Node_List &stack);
412
413 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
414 // encountered.
415 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
416
417 // Remove safepoints from loop. Optionally keeping one.
418 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
419
420 // Convert to counted loops where possible
421 void counted_loop( PhaseIdealLoop *phase );
422
423 // Check for Node being a loop-breaking test
424 Node *is_loop_exit(Node *iff) const;
425
426 // Returns true if ctrl is executed on every complete iteration
427 bool dominates_backedge(Node* ctrl);
428
429 // Remove simplistic dead code from loop body
430 void DCE_loop_body();
431
432 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
433 // Replace with a 1-in-10 exit guess.
434 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
435
436 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
437 // Useful for unrolling loops with NO array accesses.
438 bool policy_peel_only( PhaseIdealLoop *phase ) const;
439
440 // Return TRUE or FALSE if the loop should be unswitched -- clone
441 // loop with an invariant test
442 bool policy_unswitching( PhaseIdealLoop *phase ) const;
443
444 // Micro-benchmark spamming. Remove empty loops.
445 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
446
447 // Convert one iteration loop into normal code.
448 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
449
450 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
451 // make some loop-invariant test (usually a null-check) happen before the
452 // loop.
453 bool policy_peeling( PhaseIdealLoop *phase ) const;
454
455 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
456 // known trip count in the counted loop node.
457 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
458
459 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
460 // the loop is a CountedLoop and the body is small enough.
461 bool policy_unroll( PhaseIdealLoop *phase ) const;
462
463 // Return TRUE or FALSE if the loop should be range-check-eliminated.
464 // Gather a list of IF tests that are dominated by iteration splitting;
465 // also gather the end of the first split and the start of the 2nd split.
466 bool policy_range_check( PhaseIdealLoop *phase ) const;
467
468 // Return TRUE or FALSE if the loop should be cache-line aligned.
469 // Gather the expression that does the alignment. Note that only
470 // one array base can be aligned in a loop (unless the VM guarantees
471 // mutual alignment). Note that if we vectorize short memory ops
472 // into longer memory ops, we may want to increase alignment.
473 bool policy_align( PhaseIdealLoop *phase ) const;
474
475 // Return TRUE if "iff" is a range check.
476 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
477
478 // Compute loop exact trip count if possible
479 void compute_exact_trip_count( PhaseIdealLoop *phase );
480
481 // Compute loop trip count from profile data
482 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
483
484 // Reassociate invariant expressions.
485 void reassociate_invariants(PhaseIdealLoop *phase);
486 // Reassociate invariant add and subtract expressions.
487 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
488 // Return nonzero index of invariant operand if invariant and variant
489 // are combined with an Add or Sub. Helper for reassociate_invariants.
490 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
491
492 // Return true if n is invariant
493 bool is_invariant(Node* n) const;
494
495 // Put loop body on igvn work list
496 void record_for_igvn();
497
is_loop()498 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
is_inner()499 bool is_inner() { return is_loop() && _child == NULL; }
is_counted()500 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
501
502 #ifndef PRODUCT
503 void dump_head( ) const; // Dump loop head only
504 void dump() const; // Dump this loop recursively
505 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
506 #endif
507
508 };
509
510 // -----------------------------PhaseIdealLoop---------------------------------
511 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
512 // loop tree. Drives the loop-based transformations on the ideal graph.
513 class PhaseIdealLoop : public PhaseTransform {
514 friend class IdealLoopTree;
515 friend class SuperWord;
516 // Pre-computed def-use info
517 PhaseIterGVN &_igvn;
518
519 // Head of loop tree
520 IdealLoopTree *_ltree_root;
521
522 // Array of pre-order numbers, plus post-visited bit.
523 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
524 // ODD for post-visited. Other bits are the pre-order number.
525 uint *_preorders;
526 uint _max_preorder;
527
528 const PhaseIdealLoop* _verify_me;
529 bool _verify_only;
530
531 // Allocate _preorders[] array
allocate_preorders()532 void allocate_preorders() {
533 _max_preorder = C->unique()+8;
534 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
535 memset(_preorders, 0, sizeof(uint) * _max_preorder);
536 }
537
538 // Allocate _preorders[] array
reallocate_preorders()539 void reallocate_preorders() {
540 if ( _max_preorder < C->unique() ) {
541 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
542 _max_preorder = C->unique();
543 }
544 memset(_preorders, 0, sizeof(uint) * _max_preorder);
545 }
546
547 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
548 // adds new nodes.
check_grow_preorders()549 void check_grow_preorders( ) {
550 if ( _max_preorder < C->unique() ) {
551 uint newsize = _max_preorder<<1; // double size of array
552 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
553 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
554 _max_preorder = newsize;
555 }
556 }
557 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
is_visited(Node * n) const558 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
559 // Pre-order numbers are written to the Nodes array as low-bit-set values.
set_preorder_visited(Node * n,int pre_order)560 void set_preorder_visited( Node *n, int pre_order ) {
561 assert( !is_visited( n ), "already set" );
562 _preorders[n->_idx] = (pre_order<<1);
563 };
564 // Return pre-order number.
get_preorder(Node * n) const565 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
566
567 // Check for being post-visited.
568 // Should be previsited already (checked with assert(is_visited(n))).
is_postvisited(Node * n) const569 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
570
571 // Mark as post visited
set_postvisited(Node * n)572 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
573
574 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
575 // Returns true if "n" is a data node, false if it's a control node.
has_ctrl(Node * n) const576 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
577
578 // clear out dead code after build_loop_late
579 Node_List _deadlist;
580
581 // Support for faster execution of get_late_ctrl()/dom_lca()
582 // when a node has many uses and dominator depth is deep.
583 Node_Array _dom_lca_tags;
584 void init_dom_lca_tags();
585 void clear_dom_lca_tags();
586
587 // Helper for debugging bad dominance relationships
588 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
589
590 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
591
592 // Inline wrapper for frequent cases:
593 // 1) only one use
594 // 2) a use is the same as the current LCA passed as 'n1'
dom_lca_for_get_late_ctrl(Node * lca,Node * n,Node * tag)595 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
596 assert( n->is_CFG(), "" );
597 // Fast-path NULL lca
598 if( lca != NULL && lca != n ) {
599 assert( lca->is_CFG(), "" );
600 // find LCA of all uses
601 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
602 }
603 return find_non_split_ctrl(n);
604 }
605 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
606
607 // Helper function for directing control inputs away from CFG split
608 // points.
find_non_split_ctrl(Node * ctrl) const609 Node *find_non_split_ctrl( Node *ctrl ) const {
610 if (ctrl != NULL) {
611 if (ctrl->is_MultiBranch()) {
612 ctrl = ctrl->in(0);
613 }
614 assert(ctrl->is_CFG(), "CFG");
615 }
616 return ctrl;
617 }
618
619 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
620
621 public:
622
623 static bool is_canonical_main_loop_entry(CountedLoopNode* cl);
624
has_node(Node * n) const625 bool has_node( Node* n ) const {
626 guarantee(n != NULL, "No Node.");
627 return _nodes[n->_idx] != NULL;
628 }
629 // check if transform created new nodes that need _ctrl recorded
630 Node *get_late_ctrl( Node *n, Node *early );
631 Node *get_early_ctrl( Node *n );
632 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
633 void set_early_ctrl( Node *n );
634 void set_subtree_ctrl( Node *root );
set_ctrl(Node * n,Node * ctrl)635 void set_ctrl( Node *n, Node *ctrl ) {
636 assert( !has_node(n) || has_ctrl(n), "" );
637 assert( ctrl->in(0), "cannot set dead control node" );
638 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
639 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
640 }
641 // Set control and update loop membership
set_ctrl_and_loop(Node * n,Node * ctrl)642 void set_ctrl_and_loop(Node* n, Node* ctrl) {
643 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
644 IdealLoopTree* new_loop = get_loop(ctrl);
645 if (old_loop != new_loop) {
646 if (old_loop->_child == NULL) old_loop->_body.yank(n);
647 if (new_loop->_child == NULL) new_loop->_body.push(n);
648 }
649 set_ctrl(n, ctrl);
650 }
651 // Control nodes can be replaced or subsumed. During this pass they
652 // get their replacement Node in slot 1. Instead of updating the block
653 // location of all Nodes in the subsumed block, we lazily do it. As we
654 // pull such a subsumed block out of the array, we write back the final
655 // correct block.
get_ctrl(Node * i)656 Node *get_ctrl( Node *i ) {
657 assert(has_node(i), "");
658 Node *n = get_ctrl_no_update(i);
659 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
660 assert(has_node(i) && has_ctrl(i), "");
661 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
662 return n;
663 }
664 // true if CFG node d dominates CFG node n
665 bool is_dominator(Node *d, Node *n);
666 // return get_ctrl for a data node and self(n) for a CFG node
ctrl_or_self(Node * n)667 Node* ctrl_or_self(Node* n) {
668 if (has_ctrl(n))
669 return get_ctrl(n);
670 else {
671 assert (n->is_CFG(), "must be a CFG node");
672 return n;
673 }
674 }
675
676 private:
get_ctrl_no_update_helper(Node * i) const677 Node *get_ctrl_no_update_helper(Node *i) const {
678 assert(has_ctrl(i), "should be control, not loop");
679 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
680 }
681
get_ctrl_no_update(Node * i) const682 Node *get_ctrl_no_update(Node *i) const {
683 assert( has_ctrl(i), "" );
684 Node *n = get_ctrl_no_update_helper(i);
685 if (!n->in(0)) {
686 // Skip dead CFG nodes
687 do {
688 n = get_ctrl_no_update_helper(n);
689 } while (!n->in(0));
690 n = find_non_split_ctrl(n);
691 }
692 return n;
693 }
694
695 // Check for loop being set
696 // "n" must be a control node. Returns true if "n" is known to be in a loop.
has_loop(Node * n) const697 bool has_loop( Node *n ) const {
698 assert(!has_node(n) || !has_ctrl(n), "");
699 return has_node(n);
700 }
701 // Set loop
set_loop(Node * n,IdealLoopTree * loop)702 void set_loop( Node *n, IdealLoopTree *loop ) {
703 _nodes.map(n->_idx, (Node*)loop);
704 }
705 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
706 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
707 // from old_node to new_node to support the lazy update. Reference
708 // replaces loop reference, since that is not needed for dead node.
709 public:
lazy_update(Node * old_node,Node * new_node)710 void lazy_update(Node *old_node, Node *new_node) {
711 assert(old_node != new_node, "no cycles please");
712 // Re-use the side array slot for this node to provide the
713 // forwarding pointer.
714 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
715 }
lazy_replace(Node * old_node,Node * new_node)716 void lazy_replace(Node *old_node, Node *new_node) {
717 _igvn.replace_node(old_node, new_node);
718 lazy_update(old_node, new_node);
719 }
720
721 private:
722
723 // Place 'n' in some loop nest, where 'n' is a CFG node
724 void build_loop_tree();
725 int build_loop_tree_impl( Node *n, int pre_order );
726 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
727 // loop tree, not the root.
728 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
729
730 // Place Data nodes in some loop nest
731 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
732 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
733 void build_loop_late_post ( Node* n );
734
735 // Array of immediate dominance info for each CFG node indexed by node idx
736 private:
737 uint _idom_size;
738 Node **_idom; // Array of immediate dominators
739 uint *_dom_depth; // Used for fast LCA test
740 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
741
idom_no_update(Node * d) const742 Node* idom_no_update(Node* d) const {
743 assert(d->_idx < _idom_size, "oob");
744 Node* n = _idom[d->_idx];
745 assert(n != NULL,"Bad immediate dominator info.");
746 while (n->in(0) == NULL) { // Skip dead CFG nodes
747 //n = n->in(1);
748 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
749 assert(n != NULL,"Bad immediate dominator info.");
750 }
751 return n;
752 }
idom(Node * d) const753 Node *idom(Node* d) const {
754 uint didx = d->_idx;
755 Node *n = idom_no_update(d);
756 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
757 return n;
758 }
dom_depth(Node * d) const759 uint dom_depth(Node* d) const {
760 guarantee(d != NULL, "Null dominator info.");
761 guarantee(d->_idx < _idom_size, "");
762 return _dom_depth[d->_idx];
763 }
764 void set_idom(Node* d, Node* n, uint dom_depth);
765 // Locally compute IDOM using dom_lca call
766 Node *compute_idom( Node *region ) const;
767 // Recompute dom_depth
768 void recompute_dom_depth();
769
770 // Is safept not required by an outer loop?
771 bool is_deleteable_safept(Node* sfpt);
772
773 // Replace parallel induction variable (parallel to trip counter)
774 void replace_parallel_iv(IdealLoopTree *loop);
775
776 // Perform verification that the graph is valid.
PhaseIdealLoop(PhaseIterGVN & igvn)777 PhaseIdealLoop( PhaseIterGVN &igvn) :
778 PhaseTransform(Ideal_Loop),
779 _igvn(igvn),
780 _dom_lca_tags(arena()), // Thread::resource_area
781 _verify_me(NULL),
782 _verify_only(true) {
783 build_and_optimize(false, false);
784 }
785
786 // build the loop tree and perform any requested optimizations
787 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
788
789 public:
790 // Dominators for the sea of nodes
791 void Dominators();
dom_lca(Node * n1,Node * n2) const792 Node *dom_lca( Node *n1, Node *n2 ) const {
793 return find_non_split_ctrl(dom_lca_internal(n1, n2));
794 }
795 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
796
797 // Compute the Ideal Node to Loop mapping
PhaseIdealLoop(PhaseIterGVN & igvn,bool do_split_ifs,bool skip_loop_opts=false)798 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
799 PhaseTransform(Ideal_Loop),
800 _igvn(igvn),
801 _dom_lca_tags(arena()), // Thread::resource_area
802 _verify_me(NULL),
803 _verify_only(false) {
804 build_and_optimize(do_split_ifs, skip_loop_opts);
805 }
806
807 // Verify that verify_me made the same decisions as a fresh run.
PhaseIdealLoop(PhaseIterGVN & igvn,const PhaseIdealLoop * verify_me)808 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
809 PhaseTransform(Ideal_Loop),
810 _igvn(igvn),
811 _dom_lca_tags(arena()), // Thread::resource_area
812 _verify_me(verify_me),
813 _verify_only(false) {
814 build_and_optimize(false, false);
815 }
816
817 // Build and verify the loop tree without modifying the graph. This
818 // is useful to verify that all inputs properly dominate their uses.
verify(PhaseIterGVN & igvn)819 static void verify(PhaseIterGVN& igvn) {
820 #ifdef ASSERT
821 PhaseIdealLoop v(igvn);
822 #endif
823 }
824
825 // True if the method has at least 1 irreducible loop
826 bool _has_irreducible_loops;
827
828 // Per-Node transform
transform(Node * a_node)829 virtual Node *transform( Node *a_node ) { return 0; }
830
831 bool is_counted_loop( Node *x, IdealLoopTree *loop );
832
833 Node* exact_limit( IdealLoopTree *loop );
834
835 // Return a post-walked LoopNode
get_loop(Node * n) const836 IdealLoopTree *get_loop( Node *n ) const {
837 // Dead nodes have no loop, so return the top level loop instead
838 if (!has_node(n)) return _ltree_root;
839 assert(!has_ctrl(n), "");
840 return (IdealLoopTree*)_nodes[n->_idx];
841 }
842
843 // Is 'n' a (nested) member of 'loop'?
is_member(const IdealLoopTree * loop,Node * n) const844 int is_member( const IdealLoopTree *loop, Node *n ) const {
845 return loop->is_member(get_loop(n)); }
846
847 // This is the basic building block of the loop optimizations. It clones an
848 // entire loop body. It makes an old_new loop body mapping; with this
849 // mapping you can find the new-loop equivalent to an old-loop node. All
850 // new-loop nodes are exactly equal to their old-loop counterparts, all
851 // edges are the same. All exits from the old-loop now have a RegionNode
852 // that merges the equivalent new-loop path. This is true even for the
853 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
854 // now come from (one or more) Phis that merge their new-loop equivalents.
855 // Parameter side_by_side_idom:
856 // When side_by_size_idom is NULL, the dominator tree is constructed for
857 // the clone loop to dominate the original. Used in construction of
858 // pre-main-post loop sequence.
859 // When nonnull, the clone and original are side-by-side, both are
860 // dominated by the passed in side_by_side_idom node. Used in
861 // construction of unswitched loops.
862 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
863 Node* side_by_side_idom = NULL);
864
865 // If we got the effect of peeling, either by actually peeling or by
866 // making a pre-loop which must execute at least once, we can remove
867 // all loop-invariant dominated tests in the main body.
868 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
869
870 // Generate code to do a loop peel for the given loop (and body).
871 // old_new is a temp array.
872 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
873
874 // Add pre and post loops around the given loop. These loops are used
875 // during RCE, unrolling and aligning loops.
876 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
877 // If Node n lives in the back_ctrl block, we clone a private version of n
878 // in preheader_ctrl block and return that, otherwise return n.
879 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
880
881 // Take steps to maximally unroll the loop. Peel any odd iterations, then
882 // unroll to do double iterations. The next round of major loop transforms
883 // will repeat till the doubled loop body does all remaining iterations in 1
884 // pass.
885 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
886
887 // Unroll the loop body one step - make each trip do 2 iterations.
888 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
889
890 // Return true if exp is a constant times an induction var
891 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
892
893 // Return true if exp is a scaled induction var plus (or minus) constant
894 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
895
896 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
897 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
898 Deoptimization::DeoptReason reason);
899 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
900
901 // Clone loop predicates to cloned loops (peeled, unswitched)
902 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
903 Deoptimization::DeoptReason reason,
904 PhaseIdealLoop* loop_phase,
905 PhaseIterGVN* igvn);
906
907 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
908 bool clone_limit_check,
909 PhaseIdealLoop* loop_phase,
910 PhaseIterGVN* igvn);
911 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
912
913 static Node* skip_loop_predicates(Node* entry);
914
915 // Find a good location to insert a predicate
916 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
917 // Find a predicate
918 static Node* find_predicate(Node* entry);
919 // Construct a range check for a predicate if
920 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
921 int scale, Node* offset,
922 Node* init, Node* limit, jint stride,
923 Node* range, bool upper, bool &overflow);
924
925 // Implementation of the loop predication to promote checks outside the loop
926 bool loop_predication_impl(IdealLoopTree *loop);
927
928 // Helper function to collect predicate for eliminating the useless ones
929 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
930 void eliminate_useless_predicates();
931
932 // Change the control input of expensive nodes to allow commoning by
933 // IGVN when it is guaranteed to not result in a more frequent
934 // execution of the expensive node. Return true if progress.
935 bool process_expensive_nodes();
936
937 // Check whether node has become unreachable
is_node_unreachable(Node * n) const938 bool is_node_unreachable(Node *n) const {
939 return !has_node(n) || n->is_unreachable(_igvn);
940 }
941
942 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
943 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
944
945 // Create a slow version of the loop by cloning the loop
946 // and inserting an if to select fast-slow versions.
947 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
948 Node_List &old_new);
949
950 // Clone loop with an invariant test (that does not exit) and
951 // insert a clone of the test that selects which version to
952 // execute.
953 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
954
955 // Find candidate "if" for unswitching
956 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
957
958 // Range Check Elimination uses this function!
959 // Constrain the main loop iterations so the affine function:
960 // low_limit <= scale_con * I + offset < upper_limit
961 // always holds true. That is, either increase the number of iterations in
962 // the pre-loop or the post-loop until the condition holds true in the main
963 // loop. Scale_con, offset and limit are all loop invariant.
964 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit);
965 // Helper function for add_constraint().
966 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
967
968 // Partially peel loop up through last_peel node.
969 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
970
971 // Create a scheduled list of nodes control dependent on ctrl set.
972 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
973 // Has a use in the vector set
974 bool has_use_in_set( Node* n, VectorSet& vset );
975 // Has use internal to the vector set (ie. not in a phi at the loop head)
976 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
977 // clone "n" for uses that are outside of loop
978 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
979 // clone "n" for special uses that are in the not_peeled region
980 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
981 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
982 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
983 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
984 #ifdef ASSERT
985 // Validate the loop partition sets: peel and not_peel
986 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
987 // Ensure that uses outside of loop are of the right form
988 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
989 uint orig_exit_idx, uint clone_exit_idx);
990 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
991 #endif
992
993 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
994 int stride_of_possible_iv( Node* iff );
is_possible_iv_test(Node * iff)995 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
996 // Return the (unique) control output node that's in the loop (if it exists.)
997 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
998 // Insert a signed compare loop exit cloned from an unsigned compare.
999 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1000 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1001 // Utility to register node "n" with PhaseIdealLoop
1002 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1003 // Utility to create an if-projection
1004 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1005 // Force the iff control output to be the live_proj
1006 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1007 // Insert a region before an if projection
1008 RegionNode* insert_region_before_proj(ProjNode* proj);
1009 // Insert a new if before an if projection
1010 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1011
1012 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1013 // "Nearly" because all Nodes have been cloned from the original in the loop,
1014 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1015 // through the Phi recursively, and return a Bool.
1016 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1017 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1018
1019
1020 // Rework addressing expressions to get the most loop-invariant stuff
1021 // moved out. We'd like to do all associative operators, but it's especially
1022 // important (common) to do address expressions.
1023 Node *remix_address_expressions( Node *n );
1024
1025 // Attempt to use a conditional move instead of a phi/branch
1026 Node *conditional_move( Node *n );
1027
1028 // Reorganize offset computations to lower register pressure.
1029 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1030 // (which are then alive with the post-incremented trip counter
1031 // forcing an extra register move)
1032 void reorg_offsets( IdealLoopTree *loop );
1033
1034 // Check for aggressive application of 'split-if' optimization,
1035 // using basic block level info.
1036 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1037 Node *split_if_with_blocks_pre ( Node *n );
1038 void split_if_with_blocks_post( Node *n );
1039 Node *has_local_phi_input( Node *n );
1040 // Mark an IfNode as being dominated by a prior test,
1041 // without actually altering the CFG (and hence IDOM info).
1042 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1043
1044 // Split Node 'n' through merge point
1045 Node *split_thru_region( Node *n, Node *region );
1046 // Split Node 'n' through merge point if there is enough win.
1047 Node *split_thru_phi( Node *n, Node *region, int policy );
1048 // Found an If getting its condition-code input from a Phi in the
1049 // same block. Split thru the Region.
1050 void do_split_if( Node *iff );
1051
1052 // Conversion of fill/copy patterns into intrisic versions
1053 bool do_intrinsify_fill();
1054 bool intrinsify_fill(IdealLoopTree* lpt);
1055 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1056 Node*& shift, Node*& offset);
1057
1058 private:
1059 // Return a type based on condition control flow
1060 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
filtered_type(Node * n)1061 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1062 // Helpers for filtered type
1063 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1064
1065 // Helper functions
1066 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1067 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1068 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 );
1069 bool split_up( Node *n, Node *blk1, Node *blk2 );
1070 void sink_use( Node *use, Node *post_loop );
1071 Node *place_near_use( Node *useblock ) const;
1072
1073 bool _created_loop_node;
1074 public:
set_created_loop_node()1075 void set_created_loop_node() { _created_loop_node = true; }
created_loop_node()1076 bool created_loop_node() { return _created_loop_node; }
1077 void register_new_node( Node *n, Node *blk );
1078
1079 #ifdef ASSERT
1080 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1081 #endif
1082
1083 #ifndef PRODUCT
1084 void dump( ) const;
1085 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1086 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1087 void verify() const; // Major slow :-)
1088 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
get_loop_idx(Node * n) const1089 IdealLoopTree *get_loop_idx(Node* n) const {
1090 // Dead nodes have no loop, so return the top level loop instead
1091 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1092 }
1093 // Print some stats
1094 static void print_statistics();
1095 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1096 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1097 #endif
1098 };
1099
tail()1100 inline Node* IdealLoopTree::tail() {
1101 // Handle lazy update of _tail field
1102 Node *n = _tail;
1103 //while( !n->in(0) ) // Skip dead CFG nodes
1104 //n = n->in(1);
1105 if (n->in(0) == NULL)
1106 n = _phase->get_ctrl(n);
1107 _tail = n;
1108 return n;
1109 }
1110
1111
1112 // Iterate over the loop tree using a preorder, left-to-right traversal.
1113 //
1114 // Example that visits all counted loops from within PhaseIdealLoop
1115 //
1116 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1117 // IdealLoopTree* lpt = iter.current();
1118 // if (!lpt->is_counted()) continue;
1119 // ...
1120 class LoopTreeIterator : public StackObj {
1121 private:
1122 IdealLoopTree* _root;
1123 IdealLoopTree* _curnt;
1124
1125 public:
LoopTreeIterator(IdealLoopTree * root)1126 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1127
done()1128 bool done() { return _curnt == NULL; } // Finished iterating?
1129
1130 void next(); // Advance to next loop tree
1131
current()1132 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1133 };
1134
1135 #endif // SHARE_VM_OPTO_LOOPNODE_HPP
1136