1 /*
2 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP
26 #define SHARE_VM_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/subnode.hpp"
32 #include "opto/type.hpp"
33
34 class CmpNode;
35 class CountedLoopEndNode;
36 class CountedLoopNode;
37 class IdealLoopTree;
38 class LoopNode;
39 class Node;
40 class PhaseIdealLoop;
41 class VectorSet;
42 class Invariance;
43 struct small_cache;
44
45 //
46 // I D E A L I Z E D L O O P S
47 //
48 // Idealized loops are the set of loops I perform more interesting
49 // transformations on, beyond simple hoisting.
50
51 //------------------------------LoopNode---------------------------------------
52 // Simple loop header. Fall in path on left, loop-back path on right.
53 class LoopNode : public RegionNode {
54 // Size is bigger to hold the flags. However, the flags do not change
55 // the semantics so it does not appear in the hash & cmp functions.
56 virtual uint size_of() const { return sizeof(*this); }
57 protected:
58 short _loop_flags;
59 // Names for flag bitfields
60 enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 };
61 char _unswitch_count;
62 enum { _unswitch_max=3 };
63
64 public:
65 // Names for edge indices
66 enum { Self=0, EntryControl, LoopBackControl };
67
68 int is_inner_loop() const { return _loop_flags & inner_loop; }
69 void set_inner_loop() { _loop_flags |= inner_loop; }
70
71 int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
72 void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
73 int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
74 void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
75
76 int unswitch_max() { return _unswitch_max; }
77 int unswitch_count() { return _unswitch_count; }
78 void set_unswitch_count(int val) {
79 assert (val <= unswitch_max(), "too many unswitches");
80 _unswitch_count = val;
81 }
82
83 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
84 init_class_id(Class_Loop);
85 init_req(EntryControl, entry);
86 init_req(LoopBackControl, backedge);
87 }
88
89 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
90 virtual int Opcode() const;
91 bool can_be_counted_loop(PhaseTransform* phase) const {
92 return req() == 3 && in(0) != NULL &&
93 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
94 in(2) != NULL && phase->type(in(2)) != Type::TOP;
95 }
96 #ifndef PRODUCT
97 virtual void dump_spec(outputStream *st) const;
98 #endif
99 };
100
101 //------------------------------Counted Loops----------------------------------
102 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
103 // path (and maybe some other exit paths). The trip-counter exit is always
104 // last in the loop. The trip-counter does not have to stride by a constant,
105 // but it does have to stride by a loop-invariant amount; the exit value is
106 // also loop invariant.
107
108 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
109 // CountedLoopNode has the incoming loop control and the loop-back-control
110 // which is always the IfTrue before the matching CountedLoopEndNode. The
111 // CountedLoopEndNode has an incoming control (possibly not the
112 // CountedLoopNode if there is control flow in the loop), the post-increment
113 // trip-counter value, and the limit. The trip-counter value is always of
114 // the form (Op old-trip-counter stride). The old-trip-counter is produced
115 // by a Phi connected to the CountedLoopNode. The stride is loop invariant.
116 // The Op is any commutable opcode, including Add, Mul, Xor. The
117 // CountedLoopEndNode also takes in the loop-invariant limit value.
118
119 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
120 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
121 // via the old-trip-counter from the Op node.
122
123 //------------------------------CountedLoopNode--------------------------------
124 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
125 // inputs the incoming loop-start control and the loop-back control, so they
126 // act like RegionNodes. They also take in the initial trip counter, the
127 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
128 // produce a loop-body control and the trip counter value. Since
129 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
130
131 class CountedLoopNode : public LoopNode {
132 // Size is bigger to hold _main_idx. However, _main_idx does not change
133 // the semantics so it does not appear in the hash & cmp functions.
134 virtual uint size_of() const { return sizeof(*this); }
135
136 // For Pre- and Post-loops during debugging ONLY, this holds the index of
137 // the Main CountedLoop. Used to assert that we understand the graph shape.
138 node_idx_t _main_idx;
139
140 // Known trip count calculated by policy_maximally_unroll
141 int _trip_count;
142
143 // Expected trip count from profile data
144 float _profile_trip_cnt;
145
146 // Log2 of original loop bodies in unrolled loop
147 int _unrolled_count_log2;
148
149 // Node count prior to last unrolling - used to decide if
150 // unroll,optimize,unroll,optimize,... is making progress
151 int _node_count_before_unroll;
152
153 public:
154 CountedLoopNode( Node *entry, Node *backedge )
155 : LoopNode(entry, backedge), _trip_count(max_jint),
156 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
157 _node_count_before_unroll(0) {
158 init_class_id(Class_CountedLoop);
159 // Initialize _trip_count to the largest possible value.
160 // Will be reset (lower) if the loop's trip count is known.
161 }
162
163 virtual int Opcode() const;
164 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
165
166 Node *init_control() const { return in(EntryControl); }
167 Node *back_control() const { return in(LoopBackControl); }
168 CountedLoopEndNode *loopexit() const;
169 Node *init_trip() const;
170 Node *stride() const;
171 int stride_con() const;
172 bool stride_is_con() const;
173 Node *limit() const;
174 Node *incr() const;
175 Node *phi() const;
176
177 // Match increment with optional truncation
178 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
179
180 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
181 // can run short a few iterations and may start a few iterations in.
182 // It will be RCE'd and unrolled and aligned.
183
184 // A following 'post' loop will run any remaining iterations. Used
185 // during Range Check Elimination, the 'post' loop will do any final
186 // iterations with full checks. Also used by Loop Unrolling, where
187 // the 'post' loop will do any epilog iterations needed. Basically,
188 // a 'post' loop can not profitably be further unrolled or RCE'd.
189
190 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
191 // it may do under-flow checks for RCE and may do alignment iterations
192 // so the following main loop 'knows' that it is striding down cache
193 // lines.
194
195 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
196 // Aligned, may be missing it's pre-loop.
197 enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
198 int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
199 int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; }
200 int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; }
201 int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; }
202 int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
203 void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
204
205 int main_idx() const { return _main_idx; }
206
207
208 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
209 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
210 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
211 void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; }
212
213 void set_trip_count(int tc) { _trip_count = tc; }
214 int trip_count() { return _trip_count; }
215
216 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
217 float profile_trip_cnt() { return _profile_trip_cnt; }
218
219 void double_unrolled_count() { _unrolled_count_log2++; }
220 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
221
222 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
223 int node_count_before_unroll() { return _node_count_before_unroll; }
224
225 #ifndef PRODUCT
226 virtual void dump_spec(outputStream *st) const;
227 #endif
228 };
229
230 //------------------------------CountedLoopEndNode-----------------------------
231 // CountedLoopEndNodes end simple trip counted loops. They act much like
232 // IfNodes.
233 class CountedLoopEndNode : public IfNode {
234 public:
235 enum { TestControl, TestValue };
236
237 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
238 : IfNode( control, test, prob, cnt) {
239 init_class_id(Class_CountedLoopEnd);
240 }
241 virtual int Opcode() const;
242
243 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
244 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
245 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
246 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
247 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
248 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
249 int stride_con() const;
250 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
251 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
252 CountedLoopNode *loopnode() const {
253 Node *ln = phi()->in(0);
254 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
255 return (CountedLoopNode*)ln; }
256
257 #ifndef PRODUCT
258 virtual void dump_spec(outputStream *st) const;
259 #endif
260 };
261
262
263 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
264 Node *bc = back_control();
265 if( bc == NULL ) return NULL;
266 Node *le = bc->in(0);
267 if( le->Opcode() != Op_CountedLoopEnd )
268 return NULL;
269 return (CountedLoopEndNode*)le;
270 }
271 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
272 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
273 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
274 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
275 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
276 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
277 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
278
279
280 // -----------------------------IdealLoopTree----------------------------------
281 class IdealLoopTree : public ResourceObj {
282 public:
283 IdealLoopTree *_parent; // Parent in loop tree
284 IdealLoopTree *_next; // Next sibling in loop tree
285 IdealLoopTree *_child; // First child in loop tree
286
287 // The head-tail backedge defines the loop.
288 // If tail is NULL then this loop has multiple backedges as part of the
289 // same loop. During cleanup I'll peel off the multiple backedges; merge
290 // them at the loop bottom and flow 1 real backedge into the loop.
291 Node *_head; // Head of loop
292 Node *_tail; // Tail of loop
293 inline Node *tail(); // Handle lazy update of _tail field
294 PhaseIdealLoop* _phase;
295
296 Node_List _body; // Loop body for inner loops
297
298 uint8 _nest; // Nesting depth
299 uint8 _irreducible:1, // True if irreducible
300 _has_call:1, // True if has call safepoint
301 _has_sfpt:1, // True if has non-call safepoint
302 _rce_candidate:1; // True if candidate for range check elimination
303
304 Node_List* _required_safept; // A inner loop cannot delete these safepts;
305 bool _allow_optimizations; // Allow loop optimizations
306
307 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
308 : _parent(0), _next(0), _child(0),
309 _head(head), _tail(tail),
310 _phase(phase),
311 _required_safept(NULL),
312 _allow_optimizations(true),
313 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
314 { }
315
316 // Is 'l' a member of 'this'?
317 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
318
319 // Set loop nesting depth. Accumulate has_call bits.
320 int set_nest( uint depth );
321
322 // Split out multiple fall-in edges from the loop header. Move them to a
323 // private RegionNode before the loop. This becomes the loop landing pad.
324 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
325
326 // Split out the outermost loop from this shared header.
327 void split_outer_loop( PhaseIdealLoop *phase );
328
329 // Merge all the backedges from the shared header into a private Region.
330 // Feed that region as the one backedge to this loop.
331 void merge_many_backedges( PhaseIdealLoop *phase );
332
333 // Split shared headers and insert loop landing pads.
334 // Insert a LoopNode to replace the RegionNode.
335 // Returns TRUE if loop tree is structurally changed.
336 bool beautify_loops( PhaseIdealLoop *phase );
337
338 // Perform optimization to use the loop predicates for null checks and range checks.
339 // Applies to any loop level (not just the innermost one)
340 bool loop_predication( PhaseIdealLoop *phase);
341
342 // Perform iteration-splitting on inner loops. Split iterations to
343 // avoid range checks or one-shot null checks. Returns false if the
344 // current round of loop opts should stop.
345 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
346
347 // Driver for various flavors of iteration splitting. Returns false
348 // if the current round of loop opts should stop.
349 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
350
351 // Given dominators, try to find loops with calls that must always be
352 // executed (call dominates loop tail). These loops do not need non-call
353 // safepoints (ncsfpt).
354 void check_safepts(VectorSet &visited, Node_List &stack);
355
356 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
357 // encountered.
358 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
359
360 // Convert to counted loops where possible
361 void counted_loop( PhaseIdealLoop *phase );
362
363 // Check for Node being a loop-breaking test
364 Node *is_loop_exit(Node *iff) const;
365
366 // Returns true if ctrl is executed on every complete iteration
367 bool dominates_backedge(Node* ctrl);
368
369 // Remove simplistic dead code from loop body
370 void DCE_loop_body();
371
372 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
373 // Replace with a 1-in-10 exit guess.
374 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
375
376 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
377 // Useful for unrolling loops with NO array accesses.
378 bool policy_peel_only( PhaseIdealLoop *phase ) const;
379
380 // Return TRUE or FALSE if the loop should be unswitched -- clone
381 // loop with an invariant test
382 bool policy_unswitching( PhaseIdealLoop *phase ) const;
383
384 // Micro-benchmark spamming. Remove empty loops.
385 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
386
387 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
388 // make some loop-invariant test (usually a null-check) happen before the
389 // loop.
390 bool policy_peeling( PhaseIdealLoop *phase ) const;
391
392 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
393 // known trip count in the counted loop node.
394 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
395
396 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
397 // the loop is a CountedLoop and the body is small enough.
398 bool policy_unroll( PhaseIdealLoop *phase ) const;
399
400 // Return TRUE or FALSE if the loop should be range-check-eliminated.
401 // Gather a list of IF tests that are dominated by iteration splitting;
402 // also gather the end of the first split and the start of the 2nd split.
403 bool policy_range_check( PhaseIdealLoop *phase ) const;
404
405 // Return TRUE or FALSE if the loop should be cache-line aligned.
406 // Gather the expression that does the alignment. Note that only
407 // one array base can be aligned in a loop (unless the VM guarantees
408 // mutual alignment). Note that if we vectorize short memory ops
409 // into longer memory ops, we may want to increase alignment.
410 bool policy_align( PhaseIdealLoop *phase ) const;
411
412 // Return TRUE if "iff" is a range check.
413 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
414
415 // Compute loop trip count from profile data
416 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
417
418 // Reassociate invariant expressions.
419 void reassociate_invariants(PhaseIdealLoop *phase);
420 // Reassociate invariant add and subtract expressions.
421 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
422 // Return nonzero index of invariant operand if invariant and variant
423 // are combined with an Add or Sub. Helper for reassociate_invariants.
424 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
425
426 // Return true if n is invariant
427 bool is_invariant(Node* n) const;
428
429 // Put loop body on igvn work list
430 void record_for_igvn();
431
432 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
433 bool is_inner() { return is_loop() && _child == NULL; }
434 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
435
436 #ifndef PRODUCT
437 void dump_head( ) const; // Dump loop head only
438 void dump() const; // Dump this loop recursively
439 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
440 #endif
441
442 };
443
444 // -----------------------------PhaseIdealLoop---------------------------------
445 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
446 // loop tree. Drives the loop-based transformations on the ideal graph.
447 class PhaseIdealLoop : public PhaseTransform {
448 friend class IdealLoopTree;
449 friend class SuperWord;
450 // Pre-computed def-use info
451 PhaseIterGVN &_igvn;
452
453 // Head of loop tree
454 IdealLoopTree *_ltree_root;
455
456 // Array of pre-order numbers, plus post-visited bit.
457 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
458 // ODD for post-visited. Other bits are the pre-order number.
459 uint *_preorders;
460 uint _max_preorder;
461
462 const PhaseIdealLoop* _verify_me;
463 bool _verify_only;
464
465 // Allocate _preorders[] array
466 void allocate_preorders() {
467 _max_preorder = C->unique()+8;
468 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
469 memset(_preorders, 0, sizeof(uint) * _max_preorder);
470 }
471
472 // Allocate _preorders[] array
473 void reallocate_preorders() {
474 if ( _max_preorder < C->unique() ) {
475 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
476 _max_preorder = C->unique();
477 }
478 memset(_preorders, 0, sizeof(uint) * _max_preorder);
479 }
480
481 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
482 // adds new nodes.
483 void check_grow_preorders( ) {
484 if ( _max_preorder < C->unique() ) {
485 uint newsize = _max_preorder<<1; // double size of array
486 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
487 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
488 _max_preorder = newsize;
489 }
490 }
491 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
492 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
493 // Pre-order numbers are written to the Nodes array as low-bit-set values.
494 void set_preorder_visited( Node *n, int pre_order ) {
495 assert( !is_visited( n ), "already set" );
496 _preorders[n->_idx] = (pre_order<<1);
497 };
498 // Return pre-order number.
499 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
500
501 // Check for being post-visited.
502 // Should be previsited already (checked with assert(is_visited(n))).
503 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
504
505 // Mark as post visited
506 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
507
508 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
509 // Returns true if "n" is a data node, false if it's a control node.
510 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
511
512 // clear out dead code after build_loop_late
513 Node_List _deadlist;
514
515 // Support for faster execution of get_late_ctrl()/dom_lca()
516 // when a node has many uses and dominator depth is deep.
517 Node_Array _dom_lca_tags;
518 void init_dom_lca_tags();
519 void clear_dom_lca_tags();
520
521 // Helper for debugging bad dominance relationships
522 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
523
524 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
525
526 // Inline wrapper for frequent cases:
527 // 1) only one use
528 // 2) a use is the same as the current LCA passed as 'n1'
529 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
530 assert( n->is_CFG(), "" );
531 // Fast-path NULL lca
532 if( lca != NULL && lca != n ) {
533 assert( lca->is_CFG(), "" );
534 // find LCA of all uses
535 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
536 }
537 return find_non_split_ctrl(n);
538 }
539 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
540
541 // Helper function for directing control inputs away from CFG split
542 // points.
543 Node *find_non_split_ctrl( Node *ctrl ) const {
544 if (ctrl != NULL) {
545 if (ctrl->is_MultiBranch()) {
546 ctrl = ctrl->in(0);
547 }
548 assert(ctrl->is_CFG(), "CFG");
549 }
550 return ctrl;
551 }
552
553 public:
554 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
555 // check if transform created new nodes that need _ctrl recorded
556 Node *get_late_ctrl( Node *n, Node *early );
557 Node *get_early_ctrl( Node *n );
558 void set_early_ctrl( Node *n );
559 void set_subtree_ctrl( Node *root );
560 void set_ctrl( Node *n, Node *ctrl ) {
561 assert( !has_node(n) || has_ctrl(n), "" );
562 assert( ctrl->in(0), "cannot set dead control node" );
563 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
564 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
565 }
566 // Set control and update loop membership
567 void set_ctrl_and_loop(Node* n, Node* ctrl) {
568 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
569 IdealLoopTree* new_loop = get_loop(ctrl);
570 if (old_loop != new_loop) {
571 if (old_loop->_child == NULL) old_loop->_body.yank(n);
572 if (new_loop->_child == NULL) new_loop->_body.push(n);
573 }
574 set_ctrl(n, ctrl);
575 }
576 // Control nodes can be replaced or subsumed. During this pass they
577 // get their replacement Node in slot 1. Instead of updating the block
578 // location of all Nodes in the subsumed block, we lazily do it. As we
579 // pull such a subsumed block out of the array, we write back the final
580 // correct block.
581 Node *get_ctrl( Node *i ) {
582 assert(has_node(i), "");
583 Node *n = get_ctrl_no_update(i);
584 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
585 assert(has_node(i) && has_ctrl(i), "");
586 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
587 return n;
588 }
589 // true if CFG node d dominates CFG node n
590 bool is_dominator(Node *d, Node *n);
591 // return get_ctrl for a data node and self(n) for a CFG node
592 Node* ctrl_or_self(Node* n) {
593 if (has_ctrl(n))
594 return get_ctrl(n);
595 else {
596 assert (n->is_CFG(), "must be a CFG node");
597 return n;
598 }
599 }
600
601 private:
602 Node *get_ctrl_no_update( Node *i ) const {
603 assert( has_ctrl(i), "" );
604 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
605 if (!n->in(0)) {
606 // Skip dead CFG nodes
607 do {
608 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
609 } while (!n->in(0));
610 n = find_non_split_ctrl(n);
611 }
612 return n;
613 }
614
615 // Check for loop being set
616 // "n" must be a control node. Returns true if "n" is known to be in a loop.
617 bool has_loop( Node *n ) const {
618 assert(!has_node(n) || !has_ctrl(n), "");
619 return has_node(n);
620 }
621 // Set loop
622 void set_loop( Node *n, IdealLoopTree *loop ) {
623 _nodes.map(n->_idx, (Node*)loop);
624 }
625 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
626 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
627 // from old_node to new_node to support the lazy update. Reference
628 // replaces loop reference, since that is not needed for dead node.
629 public:
630 void lazy_update( Node *old_node, Node *new_node ) {
631 assert( old_node != new_node, "no cycles please" );
632 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
633 // Nodes always have DU info now, so re-use the side array slot
634 // for this node to provide the forwarding pointer.
635 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
636 }
637 void lazy_replace( Node *old_node, Node *new_node ) {
638 _igvn.replace_node( old_node, new_node );
639 lazy_update( old_node, new_node );
640 }
641 void lazy_replace_proj( Node *old_node, Node *new_node ) {
642 assert( old_node->req() == 1, "use this for Projs" );
643 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
644 old_node->add_req( NULL );
645 lazy_replace( old_node, new_node );
646 }
647
648 private:
649
650 // Place 'n' in some loop nest, where 'n' is a CFG node
651 void build_loop_tree();
652 int build_loop_tree_impl( Node *n, int pre_order );
653 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
654 // loop tree, not the root.
655 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
656
657 // Place Data nodes in some loop nest
658 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
659 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
660 void build_loop_late_post ( Node* n );
661
662 // Array of immediate dominance info for each CFG node indexed by node idx
663 private:
664 uint _idom_size;
665 Node **_idom; // Array of immediate dominators
666 uint *_dom_depth; // Used for fast LCA test
667 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
668
669 Node* idom_no_update(Node* d) const {
670 assert(d->_idx < _idom_size, "oob");
671 Node* n = _idom[d->_idx];
672 assert(n != NULL,"Bad immediate dominator info.");
673 while (n->in(0) == NULL) { // Skip dead CFG nodes
674 //n = n->in(1);
675 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
676 assert(n != NULL,"Bad immediate dominator info.");
677 }
678 return n;
679 }
680 Node *idom(Node* d) const {
681 uint didx = d->_idx;
682 Node *n = idom_no_update(d);
683 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
684 return n;
685 }
686 uint dom_depth(Node* d) const {
687 assert(d->_idx < _idom_size, "");
688 return _dom_depth[d->_idx];
689 }
690 void set_idom(Node* d, Node* n, uint dom_depth);
691 // Locally compute IDOM using dom_lca call
692 Node *compute_idom( Node *region ) const;
693 // Recompute dom_depth
694 void recompute_dom_depth();
695
696 // Is safept not required by an outer loop?
697 bool is_deleteable_safept(Node* sfpt);
698
699 // Perform verification that the graph is valid.
700 PhaseIdealLoop( PhaseIterGVN &igvn) :
701 PhaseTransform(Ideal_Loop),
702 _igvn(igvn),
703 _dom_lca_tags(C->comp_arena()),
704 _verify_me(NULL),
705 _verify_only(true) {
706 build_and_optimize(false, false);
707 }
708
709 // build the loop tree and perform any requested optimizations
710 void build_and_optimize(bool do_split_if, bool do_loop_pred);
711
712 public:
713 // Dominators for the sea of nodes
714 void Dominators();
715 Node *dom_lca( Node *n1, Node *n2 ) const {
716 return find_non_split_ctrl(dom_lca_internal(n1, n2));
717 }
718 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
719
720 // Compute the Ideal Node to Loop mapping
721 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) :
722 PhaseTransform(Ideal_Loop),
723 _igvn(igvn),
724 _dom_lca_tags(C->comp_arena()),
725 _verify_me(NULL),
726 _verify_only(false) {
727 build_and_optimize(do_split_ifs, do_loop_pred);
728 }
729
730 // Verify that verify_me made the same decisions as a fresh run.
731 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
732 PhaseTransform(Ideal_Loop),
733 _igvn(igvn),
734 _dom_lca_tags(C->comp_arena()),
735 _verify_me(verify_me),
736 _verify_only(false) {
737 build_and_optimize(false, false);
738 }
739
740 // Build and verify the loop tree without modifying the graph. This
741 // is useful to verify that all inputs properly dominate their uses.
742 static void verify(PhaseIterGVN& igvn) {
743 #ifdef ASSERT
744 PhaseIdealLoop v(igvn);
745 #endif
746 }
747
748 // True if the method has at least 1 irreducible loop
749 bool _has_irreducible_loops;
750
751 // Per-Node transform
752 virtual Node *transform( Node *a_node ) { return 0; }
753
754 Node *is_counted_loop( Node *x, IdealLoopTree *loop );
755
756 // Return a post-walked LoopNode
757 IdealLoopTree *get_loop( Node *n ) const {
758 // Dead nodes have no loop, so return the top level loop instead
759 if (!has_node(n)) return _ltree_root;
760 assert(!has_ctrl(n), "");
761 return (IdealLoopTree*)_nodes[n->_idx];
762 }
763
764 // Is 'n' a (nested) member of 'loop'?
765 int is_member( const IdealLoopTree *loop, Node *n ) const {
766 return loop->is_member(get_loop(n)); }
767
768 // This is the basic building block of the loop optimizations. It clones an
769 // entire loop body. It makes an old_new loop body mapping; with this
770 // mapping you can find the new-loop equivalent to an old-loop node. All
771 // new-loop nodes are exactly equal to their old-loop counterparts, all
772 // edges are the same. All exits from the old-loop now have a RegionNode
773 // that merges the equivalent new-loop path. This is true even for the
774 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
775 // now come from (one or more) Phis that merge their new-loop equivalents.
776 // Parameter side_by_side_idom:
777 // When side_by_size_idom is NULL, the dominator tree is constructed for
778 // the clone loop to dominate the original. Used in construction of
779 // pre-main-post loop sequence.
780 // When nonnull, the clone and original are side-by-side, both are
781 // dominated by the passed in side_by_side_idom node. Used in
782 // construction of unswitched loops.
783 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
784 Node* side_by_side_idom = NULL);
785
786 // If we got the effect of peeling, either by actually peeling or by
787 // making a pre-loop which must execute at least once, we can remove
788 // all loop-invariant dominated tests in the main body.
789 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
790
791 // Generate code to do a loop peel for the given loop (and body).
792 // old_new is a temp array.
793 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
794
795 // Add pre and post loops around the given loop. These loops are used
796 // during RCE, unrolling and aligning loops.
797 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
798 // If Node n lives in the back_ctrl block, we clone a private version of n
799 // in preheader_ctrl block and return that, otherwise return n.
800 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );
801
802 // Take steps to maximally unroll the loop. Peel any odd iterations, then
803 // unroll to do double iterations. The next round of major loop transforms
804 // will repeat till the doubled loop body does all remaining iterations in 1
805 // pass.
806 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
807
808 // Unroll the loop body one step - make each trip do 2 iterations.
809 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
810
811 // Return true if exp is a constant times an induction var
812 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
813
814 // Return true if exp is a scaled induction var plus (or minus) constant
815 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
816
817 // Return true if proj is for "proj->[region->..]call_uct"
818 bool is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate = false);
819 // Return true for "if(test)-> proj -> ...
820 // |
821 // V
822 // other_proj->[region->..]call_uct"
823 bool is_uncommon_trap_if_pattern(ProjNode* proj, bool must_reason_predicate = false);
824 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
825 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj);
826 // Find a good location to insert a predicate
827 ProjNode* find_predicate_insertion_point(Node* start_c);
828 // Construct a range check for a predicate if
829 BoolNode* rc_predicate(Node* ctrl,
830 int scale, Node* offset,
831 Node* init, Node* limit, Node* stride,
832 Node* range, bool upper);
833
834 // Implementation of the loop predication to promote checks outside the loop
835 bool loop_predication_impl(IdealLoopTree *loop);
836
837 // Helper function to collect predicate for eliminating the useless ones
838 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
839 void eliminate_useless_predicates();
840
841 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
842 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
843
844 // Create a slow version of the loop by cloning the loop
845 // and inserting an if to select fast-slow versions.
846 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
847 Node_List &old_new);
848
849 // Clone loop with an invariant test (that does not exit) and
850 // insert a clone of the test that selects which version to
851 // execute.
852 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
853
854 // Find candidate "if" for unswitching
855 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
856
857 // Range Check Elimination uses this function!
858 // Constrain the main loop iterations so the affine function:
859 // scale_con * I + offset < limit
860 // always holds true. That is, either increase the number of iterations in
861 // the pre-loop or the post-loop until the condition holds true in the main
862 // loop. Scale_con, offset and limit are all loop invariant.
863 void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
864
865 // Partially peel loop up through last_peel node.
866 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
867
868 // Create a scheduled list of nodes control dependent on ctrl set.
869 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
870 // Has a use in the vector set
871 bool has_use_in_set( Node* n, VectorSet& vset );
872 // Has use internal to the vector set (ie. not in a phi at the loop head)
873 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
874 // clone "n" for uses that are outside of loop
875 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
876 // clone "n" for special uses that are in the not_peeled region
877 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
878 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
879 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
880 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
881 #ifdef ASSERT
882 // Validate the loop partition sets: peel and not_peel
883 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
884 // Ensure that uses outside of loop are of the right form
885 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
886 uint orig_exit_idx, uint clone_exit_idx);
887 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
888 #endif
889
890 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
891 int stride_of_possible_iv( Node* iff );
892 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
893 // Return the (unique) control output node that's in the loop (if it exists.)
894 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
895 // Insert a signed compare loop exit cloned from an unsigned compare.
896 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
897 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
898 // Utility to register node "n" with PhaseIdealLoop
899 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
900 // Utility to create an if-projection
901 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
902 // Force the iff control output to be the live_proj
903 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
904 // Insert a region before an if projection
905 RegionNode* insert_region_before_proj(ProjNode* proj);
906 // Insert a new if before an if projection
907 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
908
909 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
910 // "Nearly" because all Nodes have been cloned from the original in the loop,
911 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
912 // through the Phi recursively, and return a Bool.
913 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
914 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
915
916
917 // Rework addressing expressions to get the most loop-invariant stuff
918 // moved out. We'd like to do all associative operators, but it's especially
919 // important (common) to do address expressions.
920 Node *remix_address_expressions( Node *n );
921
922 // Attempt to use a conditional move instead of a phi/branch
923 Node *conditional_move( Node *n );
924
925 // Reorganize offset computations to lower register pressure.
926 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
927 // (which are then alive with the post-incremented trip counter
928 // forcing an extra register move)
929 void reorg_offsets( IdealLoopTree *loop );
930
931 // Check for aggressive application of 'split-if' optimization,
932 // using basic block level info.
933 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
934 Node *split_if_with_blocks_pre ( Node *n );
935 void split_if_with_blocks_post( Node *n );
936 Node *has_local_phi_input( Node *n );
937 // Mark an IfNode as being dominated by a prior test,
938 // without actually altering the CFG (and hence IDOM info).
939 void dominated_by( Node *prevdom, Node *iff );
940
941 // Split Node 'n' through merge point
942 Node *split_thru_region( Node *n, Node *region );
943 // Split Node 'n' through merge point if there is enough win.
944 Node *split_thru_phi( Node *n, Node *region, int policy );
945 // Found an If getting its condition-code input from a Phi in the
946 // same block. Split thru the Region.
947 void do_split_if( Node *iff );
948
949 // Conversion of fill/copy patterns into intrisic versions
950 bool do_intrinsify_fill();
951 bool intrinsify_fill(IdealLoopTree* lpt);
952 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
953 Node*& shift, Node*& offset);
954
955 private:
956 // Return a type based on condition control flow
957 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
958 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
959 // Helpers for filtered type
960 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
961
962 // Helper functions
963 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
964 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
965 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 );
966 bool split_up( Node *n, Node *blk1, Node *blk2 );
967 void sink_use( Node *use, Node *post_loop );
968 Node *place_near_use( Node *useblock ) const;
969
970 bool _created_loop_node;
971 public:
972 void set_created_loop_node() { _created_loop_node = true; }
973 bool created_loop_node() { return _created_loop_node; }
974 void register_new_node( Node *n, Node *blk );
975
976 #ifndef PRODUCT
977 void dump( ) const;
978 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
979 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
980 void verify() const; // Major slow :-)
981 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
982 IdealLoopTree *get_loop_idx(Node* n) const {
983 // Dead nodes have no loop, so return the top level loop instead
984 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
985 }
986 // Print some stats
987 static void print_statistics();
988 static int _loop_invokes; // Count of PhaseIdealLoop invokes
989 static int _loop_work; // Sum of PhaseIdealLoop x _unique
990 #endif
991 };
992
993 inline Node* IdealLoopTree::tail() {
994 // Handle lazy update of _tail field
995 Node *n = _tail;
996 //while( !n->in(0) ) // Skip dead CFG nodes
997 //n = n->in(1);
998 if (n->in(0) == NULL)
999 n = _phase->get_ctrl(n);
1000 _tail = n;
1001 return n;
1002 }
1003
1004
1005 // Iterate over the loop tree using a preorder, left-to-right traversal.
1006 //
1007 // Example that visits all counted loops from within PhaseIdealLoop
1008 //
1009 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1010 // IdealLoopTree* lpt = iter.current();
1011 // if (!lpt->is_counted()) continue;
1012 // ...
1013 class LoopTreeIterator : public StackObj {
1014 private:
1015 IdealLoopTree* _root;
1016 IdealLoopTree* _curnt;
1017
1018 public:
1019 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1020
1021 bool done() { return _curnt == NULL; } // Finished iterating?
1022
1023 void next(); // Advance to next loop tree
1024
1025 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1026 };
1027
1028 #endif // SHARE_VM_OPTO_LOOPNODE_HPP