1 /*
2 * Copyright 1998-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 class CmpNode;
26 class CountedLoopEndNode;
27 class CountedLoopNode;
28 class IdealLoopTree;
29 class LoopNode;
30 class Node;
31 class PhaseIdealLoop;
32 class VectorSet;
33 struct small_cache;
34
35 //
36 // I D E A L I Z E D L O O P S
37 //
38 // Idealized loops are the set of loops I perform more interesting
39 // transformations on, beyond simple hoisting.
40
41 //------------------------------LoopNode---------------------------------------
42 // Simple loop header. Fall in path on left, loop-back path on right.
43 class LoopNode : public RegionNode {
44 // Size is bigger to hold the flags. However, the flags do not change
45 // the semantics so it does not appear in the hash & cmp functions.
46 virtual uint size_of() const { return sizeof(*this); }
47 protected:
48 short _loop_flags;
49 // Names for flag bitfields
50 enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 };
51 char _unswitch_count;
52 enum { _unswitch_max=3 };
53
54 public:
55 // Names for edge indices
56 enum { Self=0, EntryControl, LoopBackControl };
57
58 int is_inner_loop() const { return _loop_flags & inner_loop; }
59 void set_inner_loop() { _loop_flags |= inner_loop; }
60
61 int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
62 void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
63 int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
64 void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
65
66 int unswitch_max() { return _unswitch_max; }
67 int unswitch_count() { return _unswitch_count; }
68 void set_unswitch_count(int val) {
69 assert (val <= unswitch_max(), "too many unswitches");
70 _unswitch_count = val;
71 }
72
73 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
74 init_class_id(Class_Loop);
75 init_req(EntryControl, entry);
76 init_req(LoopBackControl, backedge);
77 }
78
79 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
80 virtual int Opcode() const;
81 bool can_be_counted_loop(PhaseTransform* phase) const {
82 return req() == 3 && in(0) != NULL &&
83 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
84 in(2) != NULL && phase->type(in(2)) != Type::TOP;
85 }
86 #ifndef PRODUCT
87 virtual void dump_spec(outputStream *st) const;
88 #endif
89 };
90
91 //------------------------------Counted Loops----------------------------------
92 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
93 // path (and maybe some other exit paths). The trip-counter exit is always
94 // last in the loop. The trip-counter does not have to stride by a constant,
95 // but it does have to stride by a loop-invariant amount; the exit value is
96 // also loop invariant.
97
98 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
99 // CountedLoopNode has the incoming loop control and the loop-back-control
100 // which is always the IfTrue before the matching CountedLoopEndNode. The
101 // CountedLoopEndNode has an incoming control (possibly not the
102 // CountedLoopNode if there is control flow in the loop), the post-increment
103 // trip-counter value, and the limit. The trip-counter value is always of
104 // the form (Op old-trip-counter stride). The old-trip-counter is produced
105 // by a Phi connected to the CountedLoopNode. The stride is loop invariant.
106 // The Op is any commutable opcode, including Add, Mul, Xor. The
107 // CountedLoopEndNode also takes in the loop-invariant limit value.
108
109 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
110 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
111 // via the old-trip-counter from the Op node.
112
113 //------------------------------CountedLoopNode--------------------------------
114 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
115 // inputs the incoming loop-start control and the loop-back control, so they
116 // act like RegionNodes. They also take in the initial trip counter, the
117 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
118 // produce a loop-body control and the trip counter value. Since
119 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
120
121 class CountedLoopNode : public LoopNode {
122 // Size is bigger to hold _main_idx. However, _main_idx does not change
123 // the semantics so it does not appear in the hash & cmp functions.
124 virtual uint size_of() const { return sizeof(*this); }
125
126 // For Pre- and Post-loops during debugging ONLY, this holds the index of
127 // the Main CountedLoop. Used to assert that we understand the graph shape.
128 node_idx_t _main_idx;
129
130 // Known trip count calculated by policy_maximally_unroll
131 int _trip_count;
132
133 // Expected trip count from profile data
134 float _profile_trip_cnt;
135
136 // Log2 of original loop bodies in unrolled loop
137 int _unrolled_count_log2;
138
139 // Node count prior to last unrolling - used to decide if
140 // unroll,optimize,unroll,optimize,... is making progress
141 int _node_count_before_unroll;
142
143 public:
144 CountedLoopNode( Node *entry, Node *backedge )
145 : LoopNode(entry, backedge), _trip_count(max_jint),
146 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
147 _node_count_before_unroll(0) {
148 init_class_id(Class_CountedLoop);
149 // Initialize _trip_count to the largest possible value.
150 // Will be reset (lower) if the loop's trip count is known.
151 }
152
153 virtual int Opcode() const;
154 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
155
156 Node *init_control() const { return in(EntryControl); }
157 Node *back_control() const { return in(LoopBackControl); }
158 CountedLoopEndNode *loopexit() const;
159 Node *init_trip() const;
160 Node *stride() const;
161 int stride_con() const;
162 bool stride_is_con() const;
163 Node *limit() const;
164 Node *incr() const;
165 Node *phi() const;
166
167 // Match increment with optional truncation
168 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
169
170 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
171 // can run short a few iterations and may start a few iterations in.
172 // It will be RCE'd and unrolled and aligned.
173
174 // A following 'post' loop will run any remaining iterations. Used
175 // during Range Check Elimination, the 'post' loop will do any final
176 // iterations with full checks. Also used by Loop Unrolling, where
177 // the 'post' loop will do any epilog iterations needed. Basically,
178 // a 'post' loop can not profitably be further unrolled or RCE'd.
179
180 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
181 // it may do under-flow checks for RCE and may do alignment iterations
182 // so the following main loop 'knows' that it is striding down cache
183 // lines.
184
185 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
186 // Aligned, may be missing it's pre-loop.
187 enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
188 int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
189 int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; }
190 int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; }
191 int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; }
192 int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
193 void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
194
195 int main_idx() const { return _main_idx; }
196
197
198 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
199 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
200 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
201 void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; }
202
203 void set_trip_count(int tc) { _trip_count = tc; }
204 int trip_count() { return _trip_count; }
205
206 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
207 float profile_trip_cnt() { return _profile_trip_cnt; }
208
209 void double_unrolled_count() { _unrolled_count_log2++; }
210 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
211
212 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
213 int node_count_before_unroll() { return _node_count_before_unroll; }
214
215 #ifndef PRODUCT
216 virtual void dump_spec(outputStream *st) const;
217 #endif
218 };
219
220 //------------------------------CountedLoopEndNode-----------------------------
221 // CountedLoopEndNodes end simple trip counted loops. They act much like
222 // IfNodes.
223 class CountedLoopEndNode : public IfNode {
224 public:
225 enum { TestControl, TestValue };
226
227 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
228 : IfNode( control, test, prob, cnt) {
229 init_class_id(Class_CountedLoopEnd);
230 }
231 virtual int Opcode() const;
232
233 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
234 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
235 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
236 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
237 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
238 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
239 int stride_con() const;
240 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
241 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
242 CountedLoopNode *loopnode() const {
243 Node *ln = phi()->in(0);
244 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
245 return (CountedLoopNode*)ln; }
246
247 #ifndef PRODUCT
248 virtual void dump_spec(outputStream *st) const;
249 #endif
250 };
251
252
253 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
254 Node *bc = back_control();
255 if( bc == NULL ) return NULL;
256 Node *le = bc->in(0);
257 if( le->Opcode() != Op_CountedLoopEnd )
258 return NULL;
259 return (CountedLoopEndNode*)le;
260 }
261 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
262 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
263 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
264 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
265 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
266 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
267 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
268
269
270 // -----------------------------IdealLoopTree----------------------------------
271 class IdealLoopTree : public ResourceObj {
272 public:
273 IdealLoopTree *_parent; // Parent in loop tree
274 IdealLoopTree *_next; // Next sibling in loop tree
275 IdealLoopTree *_child; // First child in loop tree
276
277 // The head-tail backedge defines the loop.
278 // If tail is NULL then this loop has multiple backedges as part of the
279 // same loop. During cleanup I'll peel off the multiple backedges; merge
280 // them at the loop bottom and flow 1 real backedge into the loop.
281 Node *_head; // Head of loop
282 Node *_tail; // Tail of loop
283 inline Node *tail(); // Handle lazy update of _tail field
284 PhaseIdealLoop* _phase;
285
286 Node_List _body; // Loop body for inner loops
287
288 uint8 _nest; // Nesting depth
289 uint8 _irreducible:1, // True if irreducible
290 _has_call:1, // True if has call safepoint
291 _has_sfpt:1, // True if has non-call safepoint
292 _rce_candidate:1; // True if candidate for range check elimination
293
294 Node_List* _required_safept; // A inner loop cannot delete these safepts;
295 bool _allow_optimizations; // Allow loop optimizations
296
297 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
298 : _parent(0), _next(0), _child(0),
299 _head(head), _tail(tail),
300 _phase(phase),
301 _required_safept(NULL),
302 _allow_optimizations(true),
303 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
304 { }
305
306 // Is 'l' a member of 'this'?
307 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
308
309 // Set loop nesting depth. Accumulate has_call bits.
310 int set_nest( uint depth );
311
312 // Split out multiple fall-in edges from the loop header. Move them to a
313 // private RegionNode before the loop. This becomes the loop landing pad.
314 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
315
316 // Split out the outermost loop from this shared header.
317 void split_outer_loop( PhaseIdealLoop *phase );
318
319 // Merge all the backedges from the shared header into a private Region.
320 // Feed that region as the one backedge to this loop.
321 void merge_many_backedges( PhaseIdealLoop *phase );
322
323 // Split shared headers and insert loop landing pads.
324 // Insert a LoopNode to replace the RegionNode.
325 // Returns TRUE if loop tree is structurally changed.
326 bool beautify_loops( PhaseIdealLoop *phase );
327
328 // Perform iteration-splitting on inner loops. Split iterations to
329 // avoid range checks or one-shot null checks. Returns false if the
330 // current round of loop opts should stop.
331 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
332
333 // Driver for various flavors of iteration splitting. Returns false
334 // if the current round of loop opts should stop.
335 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
336
337 // Given dominators, try to find loops with calls that must always be
338 // executed (call dominates loop tail). These loops do not need non-call
339 // safepoints (ncsfpt).
340 void check_safepts(VectorSet &visited, Node_List &stack);
341
342 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
343 // encountered.
344 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
345
346 // Convert to counted loops where possible
347 void counted_loop( PhaseIdealLoop *phase );
348
349 // Check for Node being a loop-breaking test
350 Node *is_loop_exit(Node *iff) const;
351
352 // Returns true if ctrl is executed on every complete iteration
353 bool dominates_backedge(Node* ctrl);
354
355 // Remove simplistic dead code from loop body
356 void DCE_loop_body();
357
358 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
359 // Replace with a 1-in-10 exit guess.
360 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
361
362 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
363 // Useful for unrolling loops with NO array accesses.
364 bool policy_peel_only( PhaseIdealLoop *phase ) const;
365
366 // Return TRUE or FALSE if the loop should be unswitched -- clone
367 // loop with an invariant test
368 bool policy_unswitching( PhaseIdealLoop *phase ) const;
369
370 // Micro-benchmark spamming. Remove empty loops.
371 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
372
373 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
374 // make some loop-invariant test (usually a null-check) happen before the
375 // loop.
376 bool policy_peeling( PhaseIdealLoop *phase ) const;
377
378 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
379 // known trip count in the counted loop node.
380 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
381
382 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
383 // the loop is a CountedLoop and the body is small enough.
384 bool policy_unroll( PhaseIdealLoop *phase ) const;
385
386 // Return TRUE or FALSE if the loop should be range-check-eliminated.
387 // Gather a list of IF tests that are dominated by iteration splitting;
388 // also gather the end of the first split and the start of the 2nd split.
389 bool policy_range_check( PhaseIdealLoop *phase ) const;
390
391 // Return TRUE or FALSE if the loop should be cache-line aligned.
392 // Gather the expression that does the alignment. Note that only
393 // one array base can be aligned in a loop (unless the VM guarentees
394 // mutual alignment). Note that if we vectorize short memory ops
395 // into longer memory ops, we may want to increase alignment.
396 bool policy_align( PhaseIdealLoop *phase ) const;
397
398 // Compute loop trip count from profile data
399 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
400
401 // Reassociate invariant expressions.
402 void reassociate_invariants(PhaseIdealLoop *phase);
403 // Reassociate invariant add and subtract expressions.
404 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
405 // Return nonzero index of invariant operand if invariant and variant
406 // are combined with an Add or Sub. Helper for reassoicate_invariants.
407 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
408
409 // Return true if n is invariant
410 bool is_invariant(Node* n) const;
411
412 // Put loop body on igvn work list
413 void record_for_igvn();
414
415 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
416 bool is_inner() { return is_loop() && _child == NULL; }
417 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
418
419 #ifndef PRODUCT
420 void dump_head( ) const; // Dump loop head only
421 void dump() const; // Dump this loop recursively
422 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
423 #endif
424
425 };
426
427 // -----------------------------PhaseIdealLoop---------------------------------
428 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
429 // loop tree. Drives the loop-based transformations on the ideal graph.
430 class PhaseIdealLoop : public PhaseTransform {
431 friend class IdealLoopTree;
432 friend class SuperWord;
433 // Pre-computed def-use info
434 PhaseIterGVN &_igvn;
435
436 // Head of loop tree
437 IdealLoopTree *_ltree_root;
438
439 // Array of pre-order numbers, plus post-visited bit.
440 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
441 // ODD for post-visited. Other bits are the pre-order number.
442 uint *_preorders;
443 uint _max_preorder;
444
445 // Allocate _preorders[] array
446 void allocate_preorders() {
447 _max_preorder = C->unique()+8;
448 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
449 memset(_preorders, 0, sizeof(uint) * _max_preorder);
450 }
451
452 // Allocate _preorders[] array
453 void reallocate_preorders() {
454 if ( _max_preorder < C->unique() ) {
455 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
456 _max_preorder = C->unique();
457 }
458 memset(_preorders, 0, sizeof(uint) * _max_preorder);
459 }
460
461 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
462 // adds new nodes.
463 void check_grow_preorders( ) {
464 if ( _max_preorder < C->unique() ) {
465 uint newsize = _max_preorder<<1; // double size of array
466 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
467 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
468 _max_preorder = newsize;
469 }
470 }
471 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
472 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
473 // Pre-order numbers are written to the Nodes array as low-bit-set values.
474 void set_preorder_visited( Node *n, int pre_order ) {
475 assert( !is_visited( n ), "already set" );
476 _preorders[n->_idx] = (pre_order<<1);
477 };
478 // Return pre-order number.
479 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
480
481 // Check for being post-visited.
482 // Should be previsited already (checked with assert(is_visited(n))).
483 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
484
485 // Mark as post visited
486 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
487
488 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
489 // Returns true if "n" is a data node, false if it's a control node.
490 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
491
492 // clear out dead code after build_loop_late
493 Node_List _deadlist;
494
495 // Support for faster execution of get_late_ctrl()/dom_lca()
496 // when a node has many uses and dominator depth is deep.
497 Node_Array _dom_lca_tags;
498 void init_dom_lca_tags();
499 void clear_dom_lca_tags();
500 // Inline wrapper for frequent cases:
501 // 1) only one use
502 // 2) a use is the same as the current LCA passed as 'n1'
503 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
504 assert( n->is_CFG(), "" );
505 // Fast-path NULL lca
506 if( lca != NULL && lca != n ) {
507 assert( lca->is_CFG(), "" );
508 // find LCA of all uses
509 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
510 }
511 return find_non_split_ctrl(n);
512 }
513 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
514 // true if CFG node d dominates CFG node n
515 bool is_dominator(Node *d, Node *n);
516
517 // Helper function for directing control inputs away from CFG split
518 // points.
519 Node *find_non_split_ctrl( Node *ctrl ) const {
520 if (ctrl != NULL) {
521 if (ctrl->is_MultiBranch()) {
522 ctrl = ctrl->in(0);
523 }
524 assert(ctrl->is_CFG(), "CFG");
525 }
526 return ctrl;
527 }
528
529 public:
530 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
531 // check if transform created new nodes that need _ctrl recorded
532 Node *get_late_ctrl( Node *n, Node *early );
533 Node *get_early_ctrl( Node *n );
534 void set_early_ctrl( Node *n );
535 void set_subtree_ctrl( Node *root );
536 void set_ctrl( Node *n, Node *ctrl ) {
537 assert( !has_node(n) || has_ctrl(n), "" );
538 assert( ctrl->in(0), "cannot set dead control node" );
539 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
540 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
541 }
542 // Set control and update loop membership
543 void set_ctrl_and_loop(Node* n, Node* ctrl) {
544 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
545 IdealLoopTree* new_loop = get_loop(ctrl);
546 if (old_loop != new_loop) {
547 if (old_loop->_child == NULL) old_loop->_body.yank(n);
548 if (new_loop->_child == NULL) new_loop->_body.push(n);
549 }
550 set_ctrl(n, ctrl);
551 }
552 // Control nodes can be replaced or subsumed. During this pass they
553 // get their replacement Node in slot 1. Instead of updating the block
554 // location of all Nodes in the subsumed block, we lazily do it. As we
555 // pull such a subsumed block out of the array, we write back the final
556 // correct block.
557 Node *get_ctrl( Node *i ) {
558 assert(has_node(i), "");
559 Node *n = get_ctrl_no_update(i);
560 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
561 assert(has_node(i) && has_ctrl(i), "");
562 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
563 return n;
564 }
565
566 private:
567 Node *get_ctrl_no_update( Node *i ) const {
568 assert( has_ctrl(i), "" );
569 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
570 if (!n->in(0)) {
571 // Skip dead CFG nodes
572 do {
573 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
574 } while (!n->in(0));
575 n = find_non_split_ctrl(n);
576 }
577 return n;
578 }
579
580 // Check for loop being set
581 // "n" must be a control node. Returns true if "n" is known to be in a loop.
582 bool has_loop( Node *n ) const {
583 assert(!has_node(n) || !has_ctrl(n), "");
584 return has_node(n);
585 }
586 // Set loop
587 void set_loop( Node *n, IdealLoopTree *loop ) {
588 _nodes.map(n->_idx, (Node*)loop);
589 }
590 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
591 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
592 // from old_node to new_node to support the lazy update. Reference
593 // replaces loop reference, since that is not neede for dead node.
594 public:
595 void lazy_update( Node *old_node, Node *new_node ) {
596 assert( old_node != new_node, "no cycles please" );
597 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
598 // Nodes always have DU info now, so re-use the side array slot
599 // for this node to provide the forwarding pointer.
600 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
601 }
602 void lazy_replace( Node *old_node, Node *new_node ) {
603 _igvn.hash_delete(old_node);
604 _igvn.subsume_node( old_node, new_node );
605 lazy_update( old_node, new_node );
606 }
607 void lazy_replace_proj( Node *old_node, Node *new_node ) {
608 assert( old_node->req() == 1, "use this for Projs" );
609 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
610 old_node->add_req( NULL );
611 lazy_replace( old_node, new_node );
612 }
613
614 private:
615
616 // Place 'n' in some loop nest, where 'n' is a CFG node
617 void build_loop_tree();
618 int build_loop_tree_impl( Node *n, int pre_order );
619 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
620 // loop tree, not the root.
621 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
622
623 // Place Data nodes in some loop nest
624 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me );
625 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me );
626 void build_loop_late_post ( Node* n, const PhaseIdealLoop *verify_me );
627
628 // Array of immediate dominance info for each CFG node indexed by node idx
629 private:
630 uint _idom_size;
631 Node **_idom; // Array of immediate dominators
632 uint *_dom_depth; // Used for fast LCA test
633 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
634
635 Node* idom_no_update(Node* d) const {
636 assert(d->_idx < _idom_size, "oob");
637 Node* n = _idom[d->_idx];
638 assert(n != NULL,"Bad immediate dominator info.");
639 while (n->in(0) == NULL) { // Skip dead CFG nodes
640 //n = n->in(1);
641 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
642 assert(n != NULL,"Bad immediate dominator info.");
643 }
644 return n;
645 }
646 Node *idom(Node* d) const {
647 uint didx = d->_idx;
648 Node *n = idom_no_update(d);
649 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
650 return n;
651 }
652 uint dom_depth(Node* d) const {
653 assert(d->_idx < _idom_size, "");
654 return _dom_depth[d->_idx];
655 }
656 void set_idom(Node* d, Node* n, uint dom_depth);
657 // Locally compute IDOM using dom_lca call
658 Node *compute_idom( Node *region ) const;
659 // Recompute dom_depth
660 void recompute_dom_depth();
661
662 // Is safept not required by an outer loop?
663 bool is_deleteable_safept(Node* sfpt);
664
665 public:
666 // Dominators for the sea of nodes
667 void Dominators();
668 Node *dom_lca( Node *n1, Node *n2 ) const {
669 return find_non_split_ctrl(dom_lca_internal(n1, n2));
670 }
671 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
672
673 // Compute the Ideal Node to Loop mapping
674 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs );
675
676 // True if the method has at least 1 irreducible loop
677 bool _has_irreducible_loops;
678
679 // Per-Node transform
680 virtual Node *transform( Node *a_node ) { return 0; }
681
682 Node *is_counted_loop( Node *x, IdealLoopTree *loop );
683
684 // Return a post-walked LoopNode
685 IdealLoopTree *get_loop( Node *n ) const {
686 // Dead nodes have no loop, so return the top level loop instead
687 if (!has_node(n)) return _ltree_root;
688 assert(!has_ctrl(n), "");
689 return (IdealLoopTree*)_nodes[n->_idx];
690 }
691
692 // Is 'n' a (nested) member of 'loop'?
693 int is_member( const IdealLoopTree *loop, Node *n ) const {
694 return loop->is_member(get_loop(n)); }
695
696 // This is the basic building block of the loop optimizations. It clones an
697 // entire loop body. It makes an old_new loop body mapping; with this
698 // mapping you can find the new-loop equivalent to an old-loop node. All
699 // new-loop nodes are exactly equal to their old-loop counterparts, all
700 // edges are the same. All exits from the old-loop now have a RegionNode
701 // that merges the equivalent new-loop path. This is true even for the
702 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
703 // now come from (one or more) Phis that merge their new-loop equivalents.
704 // Parameter side_by_side_idom:
705 // When side_by_size_idom is NULL, the dominator tree is constructed for
706 // the clone loop to dominate the original. Used in construction of
707 // pre-main-post loop sequence.
708 // When nonnull, the clone and original are side-by-side, both are
709 // dominated by the passed in side_by_side_idom node. Used in
710 // construction of unswitched loops.
711 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
712 Node* side_by_side_idom = NULL);
713
714 // If we got the effect of peeling, either by actually peeling or by
715 // making a pre-loop which must execute at least once, we can remove
716 // all loop-invariant dominated tests in the main body.
717 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
718
719 // Generate code to do a loop peel for the given loop (and body).
720 // old_new is a temp array.
721 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
722
723 // Add pre and post loops around the given loop. These loops are used
724 // during RCE, unrolling and aligning loops.
725 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
726 // If Node n lives in the back_ctrl block, we clone a private version of n
727 // in preheader_ctrl block and return that, otherwise return n.
728 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );
729
730 // Take steps to maximally unroll the loop. Peel any odd iterations, then
731 // unroll to do double iterations. The next round of major loop transforms
732 // will repeat till the doubled loop body does all remaining iterations in 1
733 // pass.
734 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
735
736 // Unroll the loop body one step - make each trip do 2 iterations.
737 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
738
739 // Return true if exp is a constant times an induction var
740 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
741
742 // Return true if exp is a scaled induction var plus (or minus) constant
743 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
744
745 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
746 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
747
748 // Create a slow version of the loop by cloning the loop
749 // and inserting an if to select fast-slow versions.
750 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
751 Node_List &old_new);
752
753 // Clone loop with an invariant test (that does not exit) and
754 // insert a clone of the test that selects which version to
755 // execute.
756 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
757
758 // Find candidate "if" for unswitching
759 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
760
761 // Range Check Elimination uses this function!
762 // Constrain the main loop iterations so the affine function:
763 // scale_con * I + offset < limit
764 // always holds true. That is, either increase the number of iterations in
765 // the pre-loop or the post-loop until the condition holds true in the main
766 // loop. Scale_con, offset and limit are all loop invariant.
767 void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
768
769 // Partially peel loop up through last_peel node.
770 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
771
772 // Create a scheduled list of nodes control dependent on ctrl set.
773 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
774 // Has a use in the vector set
775 bool has_use_in_set( Node* n, VectorSet& vset );
776 // Has use internal to the vector set (ie. not in a phi at the loop head)
777 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
778 // clone "n" for uses that are outside of loop
779 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
780 // clone "n" for special uses that are in the not_peeled region
781 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
782 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
783 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
784 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
785 #ifdef ASSERT
786 // Validate the loop partition sets: peel and not_peel
787 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
788 // Ensure that uses outside of loop are of the right form
789 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
790 uint orig_exit_idx, uint clone_exit_idx);
791 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
792 #endif
793
794 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
795 int stride_of_possible_iv( Node* iff );
796 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
797 // Return the (unique) control output node that's in the loop (if it exists.)
798 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
799 // Insert a signed compare loop exit cloned from an unsigned compare.
800 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
801 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
802 // Utility to register node "n" with PhaseIdealLoop
803 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
804 // Utility to create an if-projection
805 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
806 // Force the iff control output to be the live_proj
807 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
808 // Insert a region before an if projection
809 RegionNode* insert_region_before_proj(ProjNode* proj);
810 // Insert a new if before an if projection
811 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
812
813 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
814 // "Nearly" because all Nodes have been cloned from the original in the loop,
815 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
816 // through the Phi recursively, and return a Bool.
817 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
818 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
819
820
821 // Rework addressing expressions to get the most loop-invariant stuff
822 // moved out. We'd like to do all associative operators, but it's especially
823 // important (common) to do address expressions.
824 Node *remix_address_expressions( Node *n );
825
826 // Attempt to use a conditional move instead of a phi/branch
827 Node *conditional_move( Node *n );
828
829 // Reorganize offset computations to lower register pressure.
830 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
831 // (which are then alive with the post-incremented trip counter
832 // forcing an extra register move)
833 void reorg_offsets( IdealLoopTree *loop );
834
835 // Check for aggressive application of 'split-if' optimization,
836 // using basic block level info.
837 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
838 Node *split_if_with_blocks_pre ( Node *n );
839 void split_if_with_blocks_post( Node *n );
840 Node *has_local_phi_input( Node *n );
841 // Mark an IfNode as being dominated by a prior test,
842 // without actually altering the CFG (and hence IDOM info).
843 void dominated_by( Node *prevdom, Node *iff );
844
845 // Split Node 'n' through merge point
846 Node *split_thru_region( Node *n, Node *region );
847 // Split Node 'n' through merge point if there is enough win.
848 Node *split_thru_phi( Node *n, Node *region, int policy );
849 // Found an If getting its condition-code input from a Phi in the
850 // same block. Split thru the Region.
851 void do_split_if( Node *iff );
852
853 private:
854 // Return a type based on condition control flow
855 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
856 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
857 // Helpers for filtered type
858 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
859
860 // Helper functions
861 void register_new_node( Node *n, Node *blk );
862 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
863 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
864 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 );
865 bool split_up( Node *n, Node *blk1, Node *blk2 );
866 void sink_use( Node *use, Node *post_loop );
867 Node *place_near_use( Node *useblock ) const;
868
869 bool _created_loop_node;
870 public:
871 void set_created_loop_node() { _created_loop_node = true; }
872 bool created_loop_node() { return _created_loop_node; }
873
874 #ifndef PRODUCT
875 void dump( ) const;
876 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
877 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
878 void verify() const; // Major slow :-)
879 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
880 IdealLoopTree *get_loop_idx(Node* n) const {
881 // Dead nodes have no loop, so return the top level loop instead
882 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
883 }
884 // Print some stats
885 static void print_statistics();
886 static int _loop_invokes; // Count of PhaseIdealLoop invokes
887 static int _loop_work; // Sum of PhaseIdealLoop x _unique
888 #endif
889 };
890
891 inline Node* IdealLoopTree::tail() {
892 // Handle lazy update of _tail field
893 Node *n = _tail;
894 //while( !n->in(0) ) // Skip dead CFG nodes
895 //n = n->in(1);
896 if (n->in(0) == NULL)
897 n = _phase->get_ctrl(n);
898 _tail = n;
899 return n;
900 }
901
902
903 // Iterate over the loop tree using a preorder, left-to-right traversal.
904 //
905 // Example that visits all counted loops from within PhaseIdealLoop
906 //
907 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
908 // IdealLoopTree* lpt = iter.current();
909 // if (!lpt->is_counted()) continue;
910 // ...
911 class LoopTreeIterator : public StackObj {
912 private:
913 IdealLoopTree* _root;
914 IdealLoopTree* _curnt;
915
916 public:
917 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
918
919 bool done() { return _curnt == NULL; } // Finished iterating?
920
921 void next(); // Advance to next loop tree
922
923 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
924 };