Thu, 30 Mar 2017 15:28:33 +0200
8173770: Image conversion improvements
Reviewed-by: kvn, vlivanov, dlong, rhalade, mschoene, iignatyev
1 /*
2 * Copyright (c) 1998, 2015, 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 */
25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP
26 #define SHARE_VM_OPTO_LOOPNODE_HPP
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"
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;
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.
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 char _unswitch_count;
67 enum { _unswitch_max=3 };
69 public:
70 // Names for edge indices
71 enum { Self=0, EntryControl, LoopBackControl };
73 int is_inner_loop() const { return _loop_flags & InnerLoop; }
74 void set_inner_loop() { _loop_flags |= InnerLoop; }
76 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
77 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
78 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
79 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
81 int unswitch_max() { return _unswitch_max; }
82 int unswitch_count() { return _unswitch_count; }
83 void set_unswitch_count(int val) {
84 assert (val <= unswitch_max(), "too many unswitches");
85 _unswitch_count = val;
86 }
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 }
94 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
95 virtual int Opcode() const;
96 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 };
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.
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.
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.
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.
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.
139 virtual uint size_of() const { return sizeof(*this); }
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;
145 // Known trip count calculated by compute_exact_trip_count()
146 uint _trip_count;
148 // Expected trip count from profile data
149 float _profile_trip_cnt;
151 // Log2 of original loop bodies in unrolled loop
152 int _unrolled_count_log2;
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;
158 public:
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 }
168 virtual int Opcode() const;
169 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
171 Node *init_control() const { return in(EntryControl); }
172 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;
182 // Match increment with optional truncation
183 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
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.
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.
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.
200 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
201 // Aligned, may be missing it's pre-loop.
202 int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
203 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
204 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
205 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
206 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
207 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
209 int main_idx() const { return _main_idx; }
212 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
213 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
214 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
215 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
217 void set_trip_count(uint tc) { _trip_count = tc; }
218 uint trip_count() { return _trip_count; }
220 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
221 void set_exact_trip_count(uint tc) {
222 _trip_count = tc;
223 _loop_flags |= HasExactTripCount;
224 }
225 void set_nonexact_trip_count() {
226 _loop_flags &= ~HasExactTripCount;
227 }
229 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
230 float profile_trip_cnt() { return _profile_trip_cnt; }
232 void double_unrolled_count() { _unrolled_count_log2++; }
233 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
235 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
236 int node_count_before_unroll() { return _node_count_before_unroll; }
238 #ifndef PRODUCT
239 virtual void dump_spec(outputStream *st) const;
240 #endif
241 };
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 };
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;
256 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
257 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
258 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
259 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
260 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
261 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
262 int stride_con() const;
263 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
264 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
265 CountedLoopNode *loopnode() const {
266 // The CountedLoopNode that goes with this CountedLoopEndNode may
267 // have been optimized out by the IGVN so be cautious with the
268 // pattern matching on the graph
269 if (phi() == NULL) {
270 return NULL;
271 }
272 Node *ln = phi()->in(0);
273 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
274 return (CountedLoopNode*)ln;
275 }
276 return NULL;
277 }
279 #ifndef PRODUCT
280 virtual void dump_spec(outputStream *st) const;
281 #endif
282 };
285 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
286 Node *bc = back_control();
287 if( bc == NULL ) return NULL;
288 Node *le = bc->in(0);
289 if( le->Opcode() != Op_CountedLoopEnd )
290 return NULL;
291 return (CountedLoopEndNode*)le;
292 }
293 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
294 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
295 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
296 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
297 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
298 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
299 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
301 //------------------------------LoopLimitNode-----------------------------
302 // Counted Loop limit node which represents exact final iterator value:
303 // trip_count = (limit - init_trip + stride - 1)/stride
304 // final_value= trip_count * stride + init_trip.
305 // Use HW instructions to calculate it when it can overflow in integer.
306 // Note, final_value should fit into integer since counted loop has
307 // limit check: limit <= max_int-stride.
308 class LoopLimitNode : public Node {
309 enum { Init=1, Limit=2, Stride=3 };
310 public:
311 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
312 // Put it on the Macro nodes list to optimize during macro nodes expansion.
313 init_flags(Flag_is_macro);
314 C->add_macro_node(this);
315 }
316 virtual int Opcode() const;
317 virtual const Type *bottom_type() const { return TypeInt::INT; }
318 virtual uint ideal_reg() const { return Op_RegI; }
319 virtual const Type *Value( PhaseTransform *phase ) const;
320 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
321 virtual Node *Identity( PhaseTransform *phase );
322 };
324 // -----------------------------IdealLoopTree----------------------------------
325 class IdealLoopTree : public ResourceObj {
326 public:
327 IdealLoopTree *_parent; // Parent in loop tree
328 IdealLoopTree *_next; // Next sibling in loop tree
329 IdealLoopTree *_child; // First child in loop tree
331 // The head-tail backedge defines the loop.
332 // If tail is NULL then this loop has multiple backedges as part of the
333 // same loop. During cleanup I'll peel off the multiple backedges; merge
334 // them at the loop bottom and flow 1 real backedge into the loop.
335 Node *_head; // Head of loop
336 Node *_tail; // Tail of loop
337 inline Node *tail(); // Handle lazy update of _tail field
338 PhaseIdealLoop* _phase;
340 Node_List _body; // Loop body for inner loops
342 uint8 _nest; // Nesting depth
343 uint8 _irreducible:1, // True if irreducible
344 _has_call:1, // True if has call safepoint
345 _has_sfpt:1, // True if has non-call safepoint
346 _rce_candidate:1; // True if candidate for range check elimination
348 Node_List* _safepts; // List of safepoints in this loop
349 Node_List* _required_safept; // A inner loop cannot delete these safepts;
350 bool _allow_optimizations; // Allow loop optimizations
352 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
353 : _parent(0), _next(0), _child(0),
354 _head(head), _tail(tail),
355 _phase(phase),
356 _safepts(NULL),
357 _required_safept(NULL),
358 _allow_optimizations(true),
359 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
360 { }
362 // Is 'l' a member of 'this'?
363 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
365 // Set loop nesting depth. Accumulate has_call bits.
366 int set_nest( uint depth );
368 // Split out multiple fall-in edges from the loop header. Move them to a
369 // private RegionNode before the loop. This becomes the loop landing pad.
370 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
372 // Split out the outermost loop from this shared header.
373 void split_outer_loop( PhaseIdealLoop *phase );
375 // Merge all the backedges from the shared header into a private Region.
376 // Feed that region as the one backedge to this loop.
377 void merge_many_backedges( PhaseIdealLoop *phase );
379 // Split shared headers and insert loop landing pads.
380 // Insert a LoopNode to replace the RegionNode.
381 // Returns TRUE if loop tree is structurally changed.
382 bool beautify_loops( PhaseIdealLoop *phase );
384 // Perform optimization to use the loop predicates for null checks and range checks.
385 // Applies to any loop level (not just the innermost one)
386 bool loop_predication( PhaseIdealLoop *phase);
388 // Perform iteration-splitting on inner loops. Split iterations to
389 // avoid range checks or one-shot null checks. Returns false if the
390 // current round of loop opts should stop.
391 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
393 // Driver for various flavors of iteration splitting. Returns false
394 // if the current round of loop opts should stop.
395 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
397 // Given dominators, try to find loops with calls that must always be
398 // executed (call dominates loop tail). These loops do not need non-call
399 // safepoints (ncsfpt).
400 void check_safepts(VectorSet &visited, Node_List &stack);
402 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
403 // encountered.
404 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
406 // Remove safepoints from loop. Optionally keeping one.
407 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
409 // Convert to counted loops where possible
410 void counted_loop( PhaseIdealLoop *phase );
412 // Check for Node being a loop-breaking test
413 Node *is_loop_exit(Node *iff) const;
415 // Returns true if ctrl is executed on every complete iteration
416 bool dominates_backedge(Node* ctrl);
418 // Remove simplistic dead code from loop body
419 void DCE_loop_body();
421 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
422 // Replace with a 1-in-10 exit guess.
423 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
425 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
426 // Useful for unrolling loops with NO array accesses.
427 bool policy_peel_only( PhaseIdealLoop *phase ) const;
429 // Return TRUE or FALSE if the loop should be unswitched -- clone
430 // loop with an invariant test
431 bool policy_unswitching( PhaseIdealLoop *phase ) const;
433 // Micro-benchmark spamming. Remove empty loops.
434 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
436 // Convert one iteration loop into normal code.
437 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
439 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
440 // make some loop-invariant test (usually a null-check) happen before the
441 // loop.
442 bool policy_peeling( PhaseIdealLoop *phase ) const;
444 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
445 // known trip count in the counted loop node.
446 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
448 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
449 // the loop is a CountedLoop and the body is small enough.
450 bool policy_unroll( PhaseIdealLoop *phase ) const;
452 // Return TRUE or FALSE if the loop should be range-check-eliminated.
453 // Gather a list of IF tests that are dominated by iteration splitting;
454 // also gather the end of the first split and the start of the 2nd split.
455 bool policy_range_check( PhaseIdealLoop *phase ) const;
457 // Return TRUE or FALSE if the loop should be cache-line aligned.
458 // Gather the expression that does the alignment. Note that only
459 // one array base can be aligned in a loop (unless the VM guarantees
460 // mutual alignment). Note that if we vectorize short memory ops
461 // into longer memory ops, we may want to increase alignment.
462 bool policy_align( PhaseIdealLoop *phase ) const;
464 // Return TRUE if "iff" is a range check.
465 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
467 // Compute loop exact trip count if possible
468 void compute_exact_trip_count( PhaseIdealLoop *phase );
470 // Compute loop trip count from profile data
471 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
473 // Reassociate invariant expressions.
474 void reassociate_invariants(PhaseIdealLoop *phase);
475 // Reassociate invariant add and subtract expressions.
476 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
477 // Return nonzero index of invariant operand if invariant and variant
478 // are combined with an Add or Sub. Helper for reassociate_invariants.
479 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
481 // Return true if n is invariant
482 bool is_invariant(Node* n) const;
484 // Put loop body on igvn work list
485 void record_for_igvn();
487 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
488 bool is_inner() { return is_loop() && _child == NULL; }
489 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
491 #ifndef PRODUCT
492 void dump_head( ) const; // Dump loop head only
493 void dump() const; // Dump this loop recursively
494 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
495 #endif
497 };
499 // -----------------------------PhaseIdealLoop---------------------------------
500 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
501 // loop tree. Drives the loop-based transformations on the ideal graph.
502 class PhaseIdealLoop : public PhaseTransform {
503 friend class IdealLoopTree;
504 friend class SuperWord;
505 // Pre-computed def-use info
506 PhaseIterGVN &_igvn;
508 // Head of loop tree
509 IdealLoopTree *_ltree_root;
511 // Array of pre-order numbers, plus post-visited bit.
512 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
513 // ODD for post-visited. Other bits are the pre-order number.
514 uint *_preorders;
515 uint _max_preorder;
517 const PhaseIdealLoop* _verify_me;
518 bool _verify_only;
520 // Allocate _preorders[] array
521 void allocate_preorders() {
522 _max_preorder = C->unique()+8;
523 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
524 memset(_preorders, 0, sizeof(uint) * _max_preorder);
525 }
527 // Allocate _preorders[] array
528 void reallocate_preorders() {
529 if ( _max_preorder < C->unique() ) {
530 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
531 _max_preorder = C->unique();
532 }
533 memset(_preorders, 0, sizeof(uint) * _max_preorder);
534 }
536 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
537 // adds new nodes.
538 void check_grow_preorders( ) {
539 if ( _max_preorder < C->unique() ) {
540 uint newsize = _max_preorder<<1; // double size of array
541 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
542 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
543 _max_preorder = newsize;
544 }
545 }
546 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
547 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
548 // Pre-order numbers are written to the Nodes array as low-bit-set values.
549 void set_preorder_visited( Node *n, int pre_order ) {
550 assert( !is_visited( n ), "already set" );
551 _preorders[n->_idx] = (pre_order<<1);
552 };
553 // Return pre-order number.
554 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
556 // Check for being post-visited.
557 // Should be previsited already (checked with assert(is_visited(n))).
558 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
560 // Mark as post visited
561 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
563 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
564 // Returns true if "n" is a data node, false if it's a control node.
565 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
567 // clear out dead code after build_loop_late
568 Node_List _deadlist;
570 // Support for faster execution of get_late_ctrl()/dom_lca()
571 // when a node has many uses and dominator depth is deep.
572 Node_Array _dom_lca_tags;
573 void init_dom_lca_tags();
574 void clear_dom_lca_tags();
576 // Helper for debugging bad dominance relationships
577 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
579 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
581 // Inline wrapper for frequent cases:
582 // 1) only one use
583 // 2) a use is the same as the current LCA passed as 'n1'
584 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
585 assert( n->is_CFG(), "" );
586 // Fast-path NULL lca
587 if( lca != NULL && lca != n ) {
588 assert( lca->is_CFG(), "" );
589 // find LCA of all uses
590 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
591 }
592 return find_non_split_ctrl(n);
593 }
594 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
596 // Helper function for directing control inputs away from CFG split
597 // points.
598 Node *find_non_split_ctrl( Node *ctrl ) const {
599 if (ctrl != NULL) {
600 if (ctrl->is_MultiBranch()) {
601 ctrl = ctrl->in(0);
602 }
603 assert(ctrl->is_CFG(), "CFG");
604 }
605 return ctrl;
606 }
608 bool cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
610 public:
611 bool has_node( Node* n ) const {
612 guarantee(n != NULL, "No Node.");
613 return _nodes[n->_idx] != NULL;
614 }
615 // check if transform created new nodes that need _ctrl recorded
616 Node *get_late_ctrl( Node *n, Node *early );
617 Node *get_early_ctrl( Node *n );
618 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
619 void set_early_ctrl( Node *n );
620 void set_subtree_ctrl( Node *root );
621 void set_ctrl( Node *n, Node *ctrl ) {
622 assert( !has_node(n) || has_ctrl(n), "" );
623 assert( ctrl->in(0), "cannot set dead control node" );
624 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
625 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
626 }
627 // Set control and update loop membership
628 void set_ctrl_and_loop(Node* n, Node* ctrl) {
629 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
630 IdealLoopTree* new_loop = get_loop(ctrl);
631 if (old_loop != new_loop) {
632 if (old_loop->_child == NULL) old_loop->_body.yank(n);
633 if (new_loop->_child == NULL) new_loop->_body.push(n);
634 }
635 set_ctrl(n, ctrl);
636 }
637 // Control nodes can be replaced or subsumed. During this pass they
638 // get their replacement Node in slot 1. Instead of updating the block
639 // location of all Nodes in the subsumed block, we lazily do it. As we
640 // pull such a subsumed block out of the array, we write back the final
641 // correct block.
642 Node *get_ctrl( Node *i ) {
643 assert(has_node(i), "");
644 Node *n = get_ctrl_no_update(i);
645 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
646 assert(has_node(i) && has_ctrl(i), "");
647 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
648 return n;
649 }
650 // true if CFG node d dominates CFG node n
651 bool is_dominator(Node *d, Node *n);
652 // return get_ctrl for a data node and self(n) for a CFG node
653 Node* ctrl_or_self(Node* n) {
654 if (has_ctrl(n))
655 return get_ctrl(n);
656 else {
657 assert (n->is_CFG(), "must be a CFG node");
658 return n;
659 }
660 }
662 private:
663 Node *get_ctrl_no_update_helper(Node *i) const {
664 assert(has_ctrl(i), "should be control, not loop");
665 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
666 }
668 Node *get_ctrl_no_update(Node *i) const {
669 assert( has_ctrl(i), "" );
670 Node *n = get_ctrl_no_update_helper(i);
671 if (!n->in(0)) {
672 // Skip dead CFG nodes
673 do {
674 n = get_ctrl_no_update_helper(n);
675 } while (!n->in(0));
676 n = find_non_split_ctrl(n);
677 }
678 return n;
679 }
681 // Check for loop being set
682 // "n" must be a control node. Returns true if "n" is known to be in a loop.
683 bool has_loop( Node *n ) const {
684 assert(!has_node(n) || !has_ctrl(n), "");
685 return has_node(n);
686 }
687 // Set loop
688 void set_loop( Node *n, IdealLoopTree *loop ) {
689 _nodes.map(n->_idx, (Node*)loop);
690 }
691 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
692 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
693 // from old_node to new_node to support the lazy update. Reference
694 // replaces loop reference, since that is not needed for dead node.
695 public:
696 void lazy_update(Node *old_node, Node *new_node) {
697 assert(old_node != new_node, "no cycles please");
698 // Re-use the side array slot for this node to provide the
699 // forwarding pointer.
700 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
701 }
702 void lazy_replace(Node *old_node, Node *new_node) {
703 _igvn.replace_node(old_node, new_node);
704 lazy_update(old_node, new_node);
705 }
707 private:
709 // Place 'n' in some loop nest, where 'n' is a CFG node
710 void build_loop_tree();
711 int build_loop_tree_impl( Node *n, int pre_order );
712 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
713 // loop tree, not the root.
714 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
716 // Place Data nodes in some loop nest
717 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
718 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
719 void build_loop_late_post ( Node* n );
721 // Array of immediate dominance info for each CFG node indexed by node idx
722 private:
723 uint _idom_size;
724 Node **_idom; // Array of immediate dominators
725 uint *_dom_depth; // Used for fast LCA test
726 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
728 Node* idom_no_update(Node* d) const {
729 assert(d->_idx < _idom_size, "oob");
730 Node* n = _idom[d->_idx];
731 assert(n != NULL,"Bad immediate dominator info.");
732 while (n->in(0) == NULL) { // Skip dead CFG nodes
733 //n = n->in(1);
734 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
735 assert(n != NULL,"Bad immediate dominator info.");
736 }
737 return n;
738 }
739 Node *idom(Node* d) const {
740 uint didx = d->_idx;
741 Node *n = idom_no_update(d);
742 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
743 return n;
744 }
745 uint dom_depth(Node* d) const {
746 guarantee(d != NULL, "Null dominator info.");
747 guarantee(d->_idx < _idom_size, "");
748 return _dom_depth[d->_idx];
749 }
750 void set_idom(Node* d, Node* n, uint dom_depth);
751 // Locally compute IDOM using dom_lca call
752 Node *compute_idom( Node *region ) const;
753 // Recompute dom_depth
754 void recompute_dom_depth();
756 // Is safept not required by an outer loop?
757 bool is_deleteable_safept(Node* sfpt);
759 // Replace parallel induction variable (parallel to trip counter)
760 void replace_parallel_iv(IdealLoopTree *loop);
762 // Perform verification that the graph is valid.
763 PhaseIdealLoop( PhaseIterGVN &igvn) :
764 PhaseTransform(Ideal_Loop),
765 _igvn(igvn),
766 _dom_lca_tags(arena()), // Thread::resource_area
767 _verify_me(NULL),
768 _verify_only(true) {
769 build_and_optimize(false, false);
770 }
772 // build the loop tree and perform any requested optimizations
773 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
775 public:
776 // Dominators for the sea of nodes
777 void Dominators();
778 Node *dom_lca( Node *n1, Node *n2 ) const {
779 return find_non_split_ctrl(dom_lca_internal(n1, n2));
780 }
781 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
783 // Compute the Ideal Node to Loop mapping
784 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
785 PhaseTransform(Ideal_Loop),
786 _igvn(igvn),
787 _dom_lca_tags(arena()), // Thread::resource_area
788 _verify_me(NULL),
789 _verify_only(false) {
790 build_and_optimize(do_split_ifs, skip_loop_opts);
791 }
793 // Verify that verify_me made the same decisions as a fresh run.
794 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
795 PhaseTransform(Ideal_Loop),
796 _igvn(igvn),
797 _dom_lca_tags(arena()), // Thread::resource_area
798 _verify_me(verify_me),
799 _verify_only(false) {
800 build_and_optimize(false, false);
801 }
803 // Build and verify the loop tree without modifying the graph. This
804 // is useful to verify that all inputs properly dominate their uses.
805 static void verify(PhaseIterGVN& igvn) {
806 #ifdef ASSERT
807 PhaseIdealLoop v(igvn);
808 #endif
809 }
811 // True if the method has at least 1 irreducible loop
812 bool _has_irreducible_loops;
814 // Per-Node transform
815 virtual Node *transform( Node *a_node ) { return 0; }
817 bool is_counted_loop( Node *x, IdealLoopTree *loop );
819 Node* exact_limit( IdealLoopTree *loop );
821 // Return a post-walked LoopNode
822 IdealLoopTree *get_loop( Node *n ) const {
823 // Dead nodes have no loop, so return the top level loop instead
824 if (!has_node(n)) return _ltree_root;
825 assert(!has_ctrl(n), "");
826 return (IdealLoopTree*)_nodes[n->_idx];
827 }
829 // Is 'n' a (nested) member of 'loop'?
830 int is_member( const IdealLoopTree *loop, Node *n ) const {
831 return loop->is_member(get_loop(n)); }
833 // This is the basic building block of the loop optimizations. It clones an
834 // entire loop body. It makes an old_new loop body mapping; with this
835 // mapping you can find the new-loop equivalent to an old-loop node. All
836 // new-loop nodes are exactly equal to their old-loop counterparts, all
837 // edges are the same. All exits from the old-loop now have a RegionNode
838 // that merges the equivalent new-loop path. This is true even for the
839 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
840 // now come from (one or more) Phis that merge their new-loop equivalents.
841 // Parameter side_by_side_idom:
842 // When side_by_size_idom is NULL, the dominator tree is constructed for
843 // the clone loop to dominate the original. Used in construction of
844 // pre-main-post loop sequence.
845 // When nonnull, the clone and original are side-by-side, both are
846 // dominated by the passed in side_by_side_idom node. Used in
847 // construction of unswitched loops.
848 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
849 Node* side_by_side_idom = NULL);
851 // If we got the effect of peeling, either by actually peeling or by
852 // making a pre-loop which must execute at least once, we can remove
853 // all loop-invariant dominated tests in the main body.
854 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
856 // Generate code to do a loop peel for the given loop (and body).
857 // old_new is a temp array.
858 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
860 // Add pre and post loops around the given loop. These loops are used
861 // during RCE, unrolling and aligning loops.
862 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
863 // If Node n lives in the back_ctrl block, we clone a private version of n
864 // in preheader_ctrl block and return that, otherwise return n.
865 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
867 // Take steps to maximally unroll the loop. Peel any odd iterations, then
868 // unroll to do double iterations. The next round of major loop transforms
869 // will repeat till the doubled loop body does all remaining iterations in 1
870 // pass.
871 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
873 // Unroll the loop body one step - make each trip do 2 iterations.
874 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
876 // Return true if exp is a constant times an induction var
877 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
879 // Return true if exp is a scaled induction var plus (or minus) constant
880 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
882 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
883 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
884 Deoptimization::DeoptReason reason);
885 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
887 // Clone loop predicates to cloned loops (peeled, unswitched)
888 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
889 Deoptimization::DeoptReason reason,
890 PhaseIdealLoop* loop_phase,
891 PhaseIterGVN* igvn);
893 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
894 bool clone_limit_check,
895 PhaseIdealLoop* loop_phase,
896 PhaseIterGVN* igvn);
897 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
899 static Node* skip_loop_predicates(Node* entry);
901 // Find a good location to insert a predicate
902 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
903 // Find a predicate
904 static Node* find_predicate(Node* entry);
905 // Construct a range check for a predicate if
906 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
907 int scale, Node* offset,
908 Node* init, Node* limit, jint stride,
909 Node* range, bool upper, bool &overflow);
911 // Implementation of the loop predication to promote checks outside the loop
912 bool loop_predication_impl(IdealLoopTree *loop);
914 // Helper function to collect predicate for eliminating the useless ones
915 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
916 void eliminate_useless_predicates();
918 // Change the control input of expensive nodes to allow commoning by
919 // IGVN when it is guaranteed to not result in a more frequent
920 // execution of the expensive node. Return true if progress.
921 bool process_expensive_nodes();
923 // Check whether node has become unreachable
924 bool is_node_unreachable(Node *n) const {
925 return !has_node(n) || n->is_unreachable(_igvn);
926 }
928 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
929 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
931 // Create a slow version of the loop by cloning the loop
932 // and inserting an if to select fast-slow versions.
933 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
934 Node_List &old_new);
936 // Clone loop with an invariant test (that does not exit) and
937 // insert a clone of the test that selects which version to
938 // execute.
939 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
941 // Find candidate "if" for unswitching
942 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
944 // Range Check Elimination uses this function!
945 // Constrain the main loop iterations so the affine function:
946 // low_limit <= scale_con * I + offset < upper_limit
947 // always holds true. That is, either increase the number of iterations in
948 // the pre-loop or the post-loop until the condition holds true in the main
949 // loop. Scale_con, offset and limit are all loop invariant.
950 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 );
951 // Helper function for add_constraint().
952 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
954 // Partially peel loop up through last_peel node.
955 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
957 // Create a scheduled list of nodes control dependent on ctrl set.
958 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
959 // Has a use in the vector set
960 bool has_use_in_set( Node* n, VectorSet& vset );
961 // Has use internal to the vector set (ie. not in a phi at the loop head)
962 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
963 // clone "n" for uses that are outside of loop
964 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
965 // clone "n" for special uses that are in the not_peeled region
966 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
967 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
968 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
969 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
970 #ifdef ASSERT
971 // Validate the loop partition sets: peel and not_peel
972 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
973 // Ensure that uses outside of loop are of the right form
974 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
975 uint orig_exit_idx, uint clone_exit_idx);
976 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
977 #endif
979 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
980 int stride_of_possible_iv( Node* iff );
981 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
982 // Return the (unique) control output node that's in the loop (if it exists.)
983 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
984 // Insert a signed compare loop exit cloned from an unsigned compare.
985 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
986 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
987 // Utility to register node "n" with PhaseIdealLoop
988 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
989 // Utility to create an if-projection
990 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
991 // Force the iff control output to be the live_proj
992 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
993 // Insert a region before an if projection
994 RegionNode* insert_region_before_proj(ProjNode* proj);
995 // Insert a new if before an if projection
996 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
998 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
999 // "Nearly" because all Nodes have been cloned from the original in the loop,
1000 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1001 // through the Phi recursively, and return a Bool.
1002 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1003 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1006 // Rework addressing expressions to get the most loop-invariant stuff
1007 // moved out. We'd like to do all associative operators, but it's especially
1008 // important (common) to do address expressions.
1009 Node *remix_address_expressions( Node *n );
1011 // Attempt to use a conditional move instead of a phi/branch
1012 Node *conditional_move( Node *n );
1014 // Reorganize offset computations to lower register pressure.
1015 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1016 // (which are then alive with the post-incremented trip counter
1017 // forcing an extra register move)
1018 void reorg_offsets( IdealLoopTree *loop );
1020 // Check for aggressive application of 'split-if' optimization,
1021 // using basic block level info.
1022 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1023 Node *split_if_with_blocks_pre ( Node *n );
1024 void split_if_with_blocks_post( Node *n );
1025 Node *has_local_phi_input( Node *n );
1026 // Mark an IfNode as being dominated by a prior test,
1027 // without actually altering the CFG (and hence IDOM info).
1028 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1030 // Split Node 'n' through merge point
1031 Node *split_thru_region( Node *n, Node *region );
1032 // Split Node 'n' through merge point if there is enough win.
1033 Node *split_thru_phi( Node *n, Node *region, int policy );
1034 // Found an If getting its condition-code input from a Phi in the
1035 // same block. Split thru the Region.
1036 void do_split_if( Node *iff );
1038 // Conversion of fill/copy patterns into intrisic versions
1039 bool do_intrinsify_fill();
1040 bool intrinsify_fill(IdealLoopTree* lpt);
1041 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1042 Node*& shift, Node*& offset);
1044 private:
1045 // Return a type based on condition control flow
1046 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1047 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1048 // Helpers for filtered type
1049 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1051 // Helper functions
1052 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1053 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1054 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 );
1055 bool split_up( Node *n, Node *blk1, Node *blk2 );
1056 void sink_use( Node *use, Node *post_loop );
1057 Node *place_near_use( Node *useblock ) const;
1059 bool _created_loop_node;
1060 public:
1061 void set_created_loop_node() { _created_loop_node = true; }
1062 bool created_loop_node() { return _created_loop_node; }
1063 void register_new_node( Node *n, Node *blk );
1065 #ifdef ASSERT
1066 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1067 #endif
1069 #ifndef PRODUCT
1070 void dump( ) const;
1071 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1072 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1073 void verify() const; // Major slow :-)
1074 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1075 IdealLoopTree *get_loop_idx(Node* n) const {
1076 // Dead nodes have no loop, so return the top level loop instead
1077 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1078 }
1079 // Print some stats
1080 static void print_statistics();
1081 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1082 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1083 #endif
1084 };
1086 inline Node* IdealLoopTree::tail() {
1087 // Handle lazy update of _tail field
1088 Node *n = _tail;
1089 //while( !n->in(0) ) // Skip dead CFG nodes
1090 //n = n->in(1);
1091 if (n->in(0) == NULL)
1092 n = _phase->get_ctrl(n);
1093 _tail = n;
1094 return n;
1095 }
1098 // Iterate over the loop tree using a preorder, left-to-right traversal.
1099 //
1100 // Example that visits all counted loops from within PhaseIdealLoop
1101 //
1102 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1103 // IdealLoopTree* lpt = iter.current();
1104 // if (!lpt->is_counted()) continue;
1105 // ...
1106 class LoopTreeIterator : public StackObj {
1107 private:
1108 IdealLoopTree* _root;
1109 IdealLoopTree* _curnt;
1111 public:
1112 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1114 bool done() { return _curnt == NULL; } // Finished iterating?
1116 void next(); // Advance to next loop tree
1118 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1119 };
1121 #endif // SHARE_VM_OPTO_LOOPNODE_HPP