Print this page
rev 4533 : 8014811: loopTransform.cpp assert(cmp_end->in(2) == limit) failed
Summary: Stop current iteration of loop opts if partial_peel() failed and it created node clones outside processed loop.
Reviewed-by: roland
| Split |
Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/opto/loopnode.hpp
+++ new/src/share/vm/opto/loopnode.hpp
1 1 /*
2 2 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP
26 26 #define SHARE_VM_OPTO_LOOPNODE_HPP
27 27
28 28 #include "opto/cfgnode.hpp"
29 29 #include "opto/multnode.hpp"
30 30 #include "opto/phaseX.hpp"
31 31 #include "opto/subnode.hpp"
32 32 #include "opto/type.hpp"
33 33
34 34 class CmpNode;
35 35 class CountedLoopEndNode;
36 36 class CountedLoopNode;
37 37 class IdealLoopTree;
38 38 class LoopNode;
39 39 class Node;
40 40 class PhaseIdealLoop;
41 41 class VectorSet;
42 42 class Invariance;
43 43 struct small_cache;
44 44
45 45 //
46 46 // I D E A L I Z E D L O O P S
47 47 //
48 48 // Idealized loops are the set of loops I perform more interesting
49 49 // transformations on, beyond simple hoisting.
50 50
51 51 //------------------------------LoopNode---------------------------------------
52 52 // Simple loop header. Fall in path on left, loop-back path on right.
53 53 class LoopNode : public RegionNode {
54 54 // Size is bigger to hold the flags. However, the flags do not change
55 55 // the semantics so it does not appear in the hash & cmp functions.
56 56 virtual uint size_of() const { return sizeof(*this); }
57 57 protected:
58 58 short _loop_flags;
59 59 // Names for flag bitfields
60 60 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
61 61 MainHasNoPreLoop=4,
62 62 HasExactTripCount=8,
63 63 InnerLoop=16,
64 64 PartialPeelLoop=32,
65 65 PartialPeelFailed=64 };
66 66 char _unswitch_count;
67 67 enum { _unswitch_max=3 };
68 68
69 69 public:
70 70 // Names for edge indices
71 71 enum { Self=0, EntryControl, LoopBackControl };
72 72
73 73 int is_inner_loop() const { return _loop_flags & InnerLoop; }
74 74 void set_inner_loop() { _loop_flags |= InnerLoop; }
75 75
76 76 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
77 77 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
78 78 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
79 79 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
80 80
81 81 int unswitch_max() { return _unswitch_max; }
82 82 int unswitch_count() { return _unswitch_count; }
83 83 void set_unswitch_count(int val) {
84 84 assert (val <= unswitch_max(), "too many unswitches");
85 85 _unswitch_count = val;
86 86 }
87 87
88 88 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
89 89 init_class_id(Class_Loop);
90 90 init_req(EntryControl, entry);
91 91 init_req(LoopBackControl, backedge);
92 92 }
93 93
94 94 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
95 95 virtual int Opcode() const;
96 96 bool can_be_counted_loop(PhaseTransform* phase) const {
97 97 return req() == 3 && in(0) != NULL &&
98 98 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
99 99 in(2) != NULL && phase->type(in(2)) != Type::TOP;
100 100 }
101 101 bool is_valid_counted_loop() const;
102 102 #ifndef PRODUCT
103 103 virtual void dump_spec(outputStream *st) const;
104 104 #endif
105 105 };
106 106
107 107 //------------------------------Counted Loops----------------------------------
108 108 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
109 109 // path (and maybe some other exit paths). The trip-counter exit is always
110 110 // last in the loop. The trip-counter have to stride by a constant;
111 111 // the exit value is also loop invariant.
112 112
113 113 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
114 114 // CountedLoopNode has the incoming loop control and the loop-back-control
115 115 // which is always the IfTrue before the matching CountedLoopEndNode. The
116 116 // CountedLoopEndNode has an incoming control (possibly not the
117 117 // CountedLoopNode if there is control flow in the loop), the post-increment
118 118 // trip-counter value, and the limit. The trip-counter value is always of
119 119 // the form (Op old-trip-counter stride). The old-trip-counter is produced
120 120 // by a Phi connected to the CountedLoopNode. The stride is constant.
121 121 // The Op is any commutable opcode, including Add, Mul, Xor. The
122 122 // CountedLoopEndNode also takes in the loop-invariant limit value.
123 123
124 124 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
125 125 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
126 126 // via the old-trip-counter from the Op node.
127 127
128 128 //------------------------------CountedLoopNode--------------------------------
129 129 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
130 130 // inputs the incoming loop-start control and the loop-back control, so they
131 131 // act like RegionNodes. They also take in the initial trip counter, the
132 132 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
133 133 // produce a loop-body control and the trip counter value. Since
134 134 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
135 135
136 136 class CountedLoopNode : public LoopNode {
137 137 // Size is bigger to hold _main_idx. However, _main_idx does not change
138 138 // the semantics so it does not appear in the hash & cmp functions.
139 139 virtual uint size_of() const { return sizeof(*this); }
140 140
141 141 // For Pre- and Post-loops during debugging ONLY, this holds the index of
142 142 // the Main CountedLoop. Used to assert that we understand the graph shape.
143 143 node_idx_t _main_idx;
144 144
145 145 // Known trip count calculated by compute_exact_trip_count()
146 146 uint _trip_count;
147 147
148 148 // Expected trip count from profile data
149 149 float _profile_trip_cnt;
150 150
151 151 // Log2 of original loop bodies in unrolled loop
152 152 int _unrolled_count_log2;
153 153
154 154 // Node count prior to last unrolling - used to decide if
155 155 // unroll,optimize,unroll,optimize,... is making progress
156 156 int _node_count_before_unroll;
157 157
158 158 public:
159 159 CountedLoopNode( Node *entry, Node *backedge )
160 160 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
161 161 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
162 162 _node_count_before_unroll(0) {
163 163 init_class_id(Class_CountedLoop);
164 164 // Initialize _trip_count to the largest possible value.
165 165 // Will be reset (lower) if the loop's trip count is known.
166 166 }
167 167
168 168 virtual int Opcode() const;
169 169 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
170 170
171 171 Node *init_control() const { return in(EntryControl); }
172 172 Node *back_control() const { return in(LoopBackControl); }
173 173 CountedLoopEndNode *loopexit() const;
174 174 Node *init_trip() const;
175 175 Node *stride() const;
176 176 int stride_con() const;
177 177 bool stride_is_con() const;
178 178 Node *limit() const;
179 179 Node *incr() const;
180 180 Node *phi() const;
181 181
182 182 // Match increment with optional truncation
183 183 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
184 184
185 185 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
186 186 // can run short a few iterations and may start a few iterations in.
187 187 // It will be RCE'd and unrolled and aligned.
188 188
189 189 // A following 'post' loop will run any remaining iterations. Used
190 190 // during Range Check Elimination, the 'post' loop will do any final
191 191 // iterations with full checks. Also used by Loop Unrolling, where
192 192 // the 'post' loop will do any epilog iterations needed. Basically,
193 193 // a 'post' loop can not profitably be further unrolled or RCE'd.
194 194
195 195 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
196 196 // it may do under-flow checks for RCE and may do alignment iterations
197 197 // so the following main loop 'knows' that it is striding down cache
198 198 // lines.
199 199
200 200 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
201 201 // Aligned, may be missing it's pre-loop.
202 202 int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
203 203 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
204 204 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
205 205 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
206 206 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
207 207 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
208 208
209 209 int main_idx() const { return _main_idx; }
210 210
211 211
212 212 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
213 213 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
214 214 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
215 215 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
216 216
217 217 void set_trip_count(uint tc) { _trip_count = tc; }
218 218 uint trip_count() { return _trip_count; }
219 219
220 220 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
221 221 void set_exact_trip_count(uint tc) {
222 222 _trip_count = tc;
223 223 _loop_flags |= HasExactTripCount;
224 224 }
225 225 void set_nonexact_trip_count() {
226 226 _loop_flags &= ~HasExactTripCount;
227 227 }
228 228
229 229 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
230 230 float profile_trip_cnt() { return _profile_trip_cnt; }
231 231
232 232 void double_unrolled_count() { _unrolled_count_log2++; }
233 233 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
234 234
235 235 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
236 236 int node_count_before_unroll() { return _node_count_before_unroll; }
237 237
238 238 #ifndef PRODUCT
239 239 virtual void dump_spec(outputStream *st) const;
240 240 #endif
241 241 };
242 242
243 243 //------------------------------CountedLoopEndNode-----------------------------
244 244 // CountedLoopEndNodes end simple trip counted loops. They act much like
245 245 // IfNodes.
246 246 class CountedLoopEndNode : public IfNode {
247 247 public:
248 248 enum { TestControl, TestValue };
249 249
250 250 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
251 251 : IfNode( control, test, prob, cnt) {
252 252 init_class_id(Class_CountedLoopEnd);
253 253 }
254 254 virtual int Opcode() const;
255 255
256 256 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
257 257 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
258 258 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
259 259 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
260 260 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
261 261 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
262 262 int stride_con() const;
263 263 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
264 264 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
265 265 CountedLoopNode *loopnode() const {
266 266 // The CountedLoopNode that goes with this CountedLoopEndNode may
267 267 // have been optimized out by the IGVN so be cautious with the
268 268 // pattern matching on the graph
269 269 if (phi() == NULL) {
270 270 return NULL;
271 271 }
272 272 Node *ln = phi()->in(0);
273 273 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
274 274 return (CountedLoopNode*)ln;
275 275 }
276 276 return NULL;
277 277 }
278 278
279 279 #ifndef PRODUCT
280 280 virtual void dump_spec(outputStream *st) const;
281 281 #endif
282 282 };
283 283
284 284
285 285 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
286 286 Node *bc = back_control();
287 287 if( bc == NULL ) return NULL;
288 288 Node *le = bc->in(0);
289 289 if( le->Opcode() != Op_CountedLoopEnd )
290 290 return NULL;
291 291 return (CountedLoopEndNode*)le;
292 292 }
293 293 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
294 294 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
295 295 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
296 296 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
297 297 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
298 298 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
299 299 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
300 300
301 301 //------------------------------LoopLimitNode-----------------------------
302 302 // Counted Loop limit node which represents exact final iterator value:
303 303 // trip_count = (limit - init_trip + stride - 1)/stride
304 304 // final_value= trip_count * stride + init_trip.
305 305 // Use HW instructions to calculate it when it can overflow in integer.
306 306 // Note, final_value should fit into integer since counted loop has
307 307 // limit check: limit <= max_int-stride.
308 308 class LoopLimitNode : public Node {
309 309 enum { Init=1, Limit=2, Stride=3 };
310 310 public:
311 311 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
312 312 // Put it on the Macro nodes list to optimize during macro nodes expansion.
313 313 init_flags(Flag_is_macro);
314 314 C->add_macro_node(this);
315 315 }
316 316 virtual int Opcode() const;
317 317 virtual const Type *bottom_type() const { return TypeInt::INT; }
318 318 virtual uint ideal_reg() const { return Op_RegI; }
319 319 virtual const Type *Value( PhaseTransform *phase ) const;
320 320 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
321 321 virtual Node *Identity( PhaseTransform *phase );
322 322 };
323 323
324 324 // -----------------------------IdealLoopTree----------------------------------
325 325 class IdealLoopTree : public ResourceObj {
326 326 public:
327 327 IdealLoopTree *_parent; // Parent in loop tree
328 328 IdealLoopTree *_next; // Next sibling in loop tree
329 329 IdealLoopTree *_child; // First child in loop tree
330 330
331 331 // The head-tail backedge defines the loop.
332 332 // If tail is NULL then this loop has multiple backedges as part of the
333 333 // same loop. During cleanup I'll peel off the multiple backedges; merge
334 334 // them at the loop bottom and flow 1 real backedge into the loop.
335 335 Node *_head; // Head of loop
336 336 Node *_tail; // Tail of loop
337 337 inline Node *tail(); // Handle lazy update of _tail field
338 338 PhaseIdealLoop* _phase;
339 339
340 340 Node_List _body; // Loop body for inner loops
341 341
342 342 uint8 _nest; // Nesting depth
343 343 uint8 _irreducible:1, // True if irreducible
344 344 _has_call:1, // True if has call safepoint
345 345 _has_sfpt:1, // True if has non-call safepoint
346 346 _rce_candidate:1; // True if candidate for range check elimination
347 347
348 348 Node_List* _safepts; // List of safepoints in this loop
349 349 Node_List* _required_safept; // A inner loop cannot delete these safepts;
350 350 bool _allow_optimizations; // Allow loop optimizations
351 351
352 352 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
353 353 : _parent(0), _next(0), _child(0),
354 354 _head(head), _tail(tail),
355 355 _phase(phase),
356 356 _safepts(NULL),
357 357 _required_safept(NULL),
358 358 _allow_optimizations(true),
359 359 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
360 360 { }
361 361
362 362 // Is 'l' a member of 'this'?
363 363 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
364 364
365 365 // Set loop nesting depth. Accumulate has_call bits.
366 366 int set_nest( uint depth );
367 367
368 368 // Split out multiple fall-in edges from the loop header. Move them to a
369 369 // private RegionNode before the loop. This becomes the loop landing pad.
370 370 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
371 371
372 372 // Split out the outermost loop from this shared header.
373 373 void split_outer_loop( PhaseIdealLoop *phase );
374 374
375 375 // Merge all the backedges from the shared header into a private Region.
376 376 // Feed that region as the one backedge to this loop.
377 377 void merge_many_backedges( PhaseIdealLoop *phase );
378 378
379 379 // Split shared headers and insert loop landing pads.
380 380 // Insert a LoopNode to replace the RegionNode.
381 381 // Returns TRUE if loop tree is structurally changed.
382 382 bool beautify_loops( PhaseIdealLoop *phase );
383 383
384 384 // Perform optimization to use the loop predicates for null checks and range checks.
385 385 // Applies to any loop level (not just the innermost one)
386 386 bool loop_predication( PhaseIdealLoop *phase);
387 387
388 388 // Perform iteration-splitting on inner loops. Split iterations to
389 389 // avoid range checks or one-shot null checks. Returns false if the
390 390 // current round of loop opts should stop.
391 391 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
392 392
393 393 // Driver for various flavors of iteration splitting. Returns false
394 394 // if the current round of loop opts should stop.
395 395 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
396 396
397 397 // Given dominators, try to find loops with calls that must always be
398 398 // executed (call dominates loop tail). These loops do not need non-call
399 399 // safepoints (ncsfpt).
400 400 void check_safepts(VectorSet &visited, Node_List &stack);
401 401
402 402 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
403 403 // encountered.
404 404 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
405 405
406 406 // Convert to counted loops where possible
407 407 void counted_loop( PhaseIdealLoop *phase );
408 408
409 409 // Check for Node being a loop-breaking test
410 410 Node *is_loop_exit(Node *iff) const;
411 411
412 412 // Returns true if ctrl is executed on every complete iteration
413 413 bool dominates_backedge(Node* ctrl);
414 414
415 415 // Remove simplistic dead code from loop body
416 416 void DCE_loop_body();
417 417
418 418 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
419 419 // Replace with a 1-in-10 exit guess.
420 420 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
421 421
422 422 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
423 423 // Useful for unrolling loops with NO array accesses.
424 424 bool policy_peel_only( PhaseIdealLoop *phase ) const;
425 425
426 426 // Return TRUE or FALSE if the loop should be unswitched -- clone
427 427 // loop with an invariant test
428 428 bool policy_unswitching( PhaseIdealLoop *phase ) const;
429 429
430 430 // Micro-benchmark spamming. Remove empty loops.
431 431 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
432 432
433 433 // Convert one iteration loop into normal code.
434 434 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
435 435
436 436 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
437 437 // make some loop-invariant test (usually a null-check) happen before the
438 438 // loop.
439 439 bool policy_peeling( PhaseIdealLoop *phase ) const;
440 440
441 441 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
442 442 // known trip count in the counted loop node.
443 443 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
444 444
445 445 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
446 446 // the loop is a CountedLoop and the body is small enough.
447 447 bool policy_unroll( PhaseIdealLoop *phase ) const;
448 448
449 449 // Return TRUE or FALSE if the loop should be range-check-eliminated.
450 450 // Gather a list of IF tests that are dominated by iteration splitting;
451 451 // also gather the end of the first split and the start of the 2nd split.
452 452 bool policy_range_check( PhaseIdealLoop *phase ) const;
453 453
454 454 // Return TRUE or FALSE if the loop should be cache-line aligned.
455 455 // Gather the expression that does the alignment. Note that only
456 456 // one array base can be aligned in a loop (unless the VM guarantees
457 457 // mutual alignment). Note that if we vectorize short memory ops
458 458 // into longer memory ops, we may want to increase alignment.
459 459 bool policy_align( PhaseIdealLoop *phase ) const;
460 460
461 461 // Return TRUE if "iff" is a range check.
462 462 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
463 463
464 464 // Compute loop exact trip count if possible
465 465 void compute_exact_trip_count( PhaseIdealLoop *phase );
466 466
467 467 // Compute loop trip count from profile data
468 468 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
469 469
470 470 // Reassociate invariant expressions.
471 471 void reassociate_invariants(PhaseIdealLoop *phase);
472 472 // Reassociate invariant add and subtract expressions.
473 473 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
474 474 // Return nonzero index of invariant operand if invariant and variant
475 475 // are combined with an Add or Sub. Helper for reassociate_invariants.
476 476 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
477 477
478 478 // Return true if n is invariant
479 479 bool is_invariant(Node* n) const;
480 480
481 481 // Put loop body on igvn work list
482 482 void record_for_igvn();
483 483
484 484 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
485 485 bool is_inner() { return is_loop() && _child == NULL; }
486 486 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
487 487
488 488 #ifndef PRODUCT
489 489 void dump_head( ) const; // Dump loop head only
490 490 void dump() const; // Dump this loop recursively
491 491 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
492 492 #endif
493 493
494 494 };
495 495
496 496 // -----------------------------PhaseIdealLoop---------------------------------
497 497 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
498 498 // loop tree. Drives the loop-based transformations on the ideal graph.
499 499 class PhaseIdealLoop : public PhaseTransform {
500 500 friend class IdealLoopTree;
501 501 friend class SuperWord;
502 502 // Pre-computed def-use info
503 503 PhaseIterGVN &_igvn;
504 504
505 505 // Head of loop tree
506 506 IdealLoopTree *_ltree_root;
507 507
508 508 // Array of pre-order numbers, plus post-visited bit.
509 509 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
510 510 // ODD for post-visited. Other bits are the pre-order number.
511 511 uint *_preorders;
512 512 uint _max_preorder;
513 513
514 514 const PhaseIdealLoop* _verify_me;
515 515 bool _verify_only;
516 516
517 517 // Allocate _preorders[] array
518 518 void allocate_preorders() {
519 519 _max_preorder = C->unique()+8;
520 520 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
521 521 memset(_preorders, 0, sizeof(uint) * _max_preorder);
522 522 }
523 523
524 524 // Allocate _preorders[] array
525 525 void reallocate_preorders() {
526 526 if ( _max_preorder < C->unique() ) {
527 527 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
528 528 _max_preorder = C->unique();
529 529 }
530 530 memset(_preorders, 0, sizeof(uint) * _max_preorder);
531 531 }
532 532
533 533 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
534 534 // adds new nodes.
535 535 void check_grow_preorders( ) {
536 536 if ( _max_preorder < C->unique() ) {
537 537 uint newsize = _max_preorder<<1; // double size of array
538 538 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
539 539 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
540 540 _max_preorder = newsize;
541 541 }
542 542 }
543 543 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
544 544 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
545 545 // Pre-order numbers are written to the Nodes array as low-bit-set values.
546 546 void set_preorder_visited( Node *n, int pre_order ) {
547 547 assert( !is_visited( n ), "already set" );
548 548 _preorders[n->_idx] = (pre_order<<1);
549 549 };
550 550 // Return pre-order number.
551 551 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
552 552
553 553 // Check for being post-visited.
554 554 // Should be previsited already (checked with assert(is_visited(n))).
555 555 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
556 556
557 557 // Mark as post visited
558 558 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
559 559
560 560 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
561 561 // Returns true if "n" is a data node, false if it's a control node.
562 562 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
563 563
564 564 // clear out dead code after build_loop_late
565 565 Node_List _deadlist;
566 566
567 567 // Support for faster execution of get_late_ctrl()/dom_lca()
568 568 // when a node has many uses and dominator depth is deep.
569 569 Node_Array _dom_lca_tags;
570 570 void init_dom_lca_tags();
571 571 void clear_dom_lca_tags();
572 572
573 573 // Helper for debugging bad dominance relationships
574 574 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
575 575
576 576 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
577 577
578 578 // Inline wrapper for frequent cases:
579 579 // 1) only one use
580 580 // 2) a use is the same as the current LCA passed as 'n1'
581 581 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
582 582 assert( n->is_CFG(), "" );
583 583 // Fast-path NULL lca
584 584 if( lca != NULL && lca != n ) {
585 585 assert( lca->is_CFG(), "" );
586 586 // find LCA of all uses
587 587 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
588 588 }
589 589 return find_non_split_ctrl(n);
590 590 }
591 591 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
592 592
593 593 // Helper function for directing control inputs away from CFG split
594 594 // points.
595 595 Node *find_non_split_ctrl( Node *ctrl ) const {
596 596 if (ctrl != NULL) {
597 597 if (ctrl->is_MultiBranch()) {
598 598 ctrl = ctrl->in(0);
599 599 }
600 600 assert(ctrl->is_CFG(), "CFG");
601 601 }
602 602 return ctrl;
603 603 }
604 604
605 605 public:
606 606 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
607 607 // check if transform created new nodes that need _ctrl recorded
608 608 Node *get_late_ctrl( Node *n, Node *early );
609 609 Node *get_early_ctrl( Node *n );
610 610 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
611 611 void set_early_ctrl( Node *n );
612 612 void set_subtree_ctrl( Node *root );
613 613 void set_ctrl( Node *n, Node *ctrl ) {
614 614 assert( !has_node(n) || has_ctrl(n), "" );
615 615 assert( ctrl->in(0), "cannot set dead control node" );
616 616 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
617 617 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
618 618 }
619 619 // Set control and update loop membership
620 620 void set_ctrl_and_loop(Node* n, Node* ctrl) {
621 621 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
622 622 IdealLoopTree* new_loop = get_loop(ctrl);
623 623 if (old_loop != new_loop) {
624 624 if (old_loop->_child == NULL) old_loop->_body.yank(n);
625 625 if (new_loop->_child == NULL) new_loop->_body.push(n);
626 626 }
627 627 set_ctrl(n, ctrl);
628 628 }
629 629 // Control nodes can be replaced or subsumed. During this pass they
630 630 // get their replacement Node in slot 1. Instead of updating the block
631 631 // location of all Nodes in the subsumed block, we lazily do it. As we
632 632 // pull such a subsumed block out of the array, we write back the final
633 633 // correct block.
634 634 Node *get_ctrl( Node *i ) {
635 635 assert(has_node(i), "");
636 636 Node *n = get_ctrl_no_update(i);
637 637 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
638 638 assert(has_node(i) && has_ctrl(i), "");
639 639 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
640 640 return n;
641 641 }
642 642 // true if CFG node d dominates CFG node n
643 643 bool is_dominator(Node *d, Node *n);
644 644 // return get_ctrl for a data node and self(n) for a CFG node
645 645 Node* ctrl_or_self(Node* n) {
646 646 if (has_ctrl(n))
647 647 return get_ctrl(n);
648 648 else {
649 649 assert (n->is_CFG(), "must be a CFG node");
650 650 return n;
651 651 }
652 652 }
653 653
654 654 private:
655 655 Node *get_ctrl_no_update( Node *i ) const {
656 656 assert( has_ctrl(i), "" );
657 657 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
658 658 if (!n->in(0)) {
659 659 // Skip dead CFG nodes
660 660 do {
661 661 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
662 662 } while (!n->in(0));
663 663 n = find_non_split_ctrl(n);
664 664 }
665 665 return n;
666 666 }
667 667
668 668 // Check for loop being set
669 669 // "n" must be a control node. Returns true if "n" is known to be in a loop.
670 670 bool has_loop( Node *n ) const {
671 671 assert(!has_node(n) || !has_ctrl(n), "");
672 672 return has_node(n);
673 673 }
674 674 // Set loop
675 675 void set_loop( Node *n, IdealLoopTree *loop ) {
676 676 _nodes.map(n->_idx, (Node*)loop);
677 677 }
678 678 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
679 679 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
680 680 // from old_node to new_node to support the lazy update. Reference
681 681 // replaces loop reference, since that is not needed for dead node.
682 682 public:
683 683 void lazy_update( Node *old_node, Node *new_node ) {
684 684 assert( old_node != new_node, "no cycles please" );
685 685 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
686 686 // Nodes always have DU info now, so re-use the side array slot
687 687 // for this node to provide the forwarding pointer.
688 688 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
689 689 }
690 690 void lazy_replace( Node *old_node, Node *new_node ) {
691 691 _igvn.replace_node( old_node, new_node );
692 692 lazy_update( old_node, new_node );
693 693 }
694 694 void lazy_replace_proj( Node *old_node, Node *new_node ) {
695 695 assert( old_node->req() == 1, "use this for Projs" );
696 696 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
697 697 old_node->add_req( NULL );
698 698 lazy_replace( old_node, new_node );
699 699 }
700 700
701 701 private:
702 702
703 703 // Place 'n' in some loop nest, where 'n' is a CFG node
704 704 void build_loop_tree();
705 705 int build_loop_tree_impl( Node *n, int pre_order );
706 706 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
707 707 // loop tree, not the root.
708 708 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
709 709
710 710 // Place Data nodes in some loop nest
711 711 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
712 712 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
713 713 void build_loop_late_post ( Node* n );
714 714
715 715 // Array of immediate dominance info for each CFG node indexed by node idx
716 716 private:
717 717 uint _idom_size;
718 718 Node **_idom; // Array of immediate dominators
719 719 uint *_dom_depth; // Used for fast LCA test
720 720 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
721 721
722 722 Node* idom_no_update(Node* d) const {
723 723 assert(d->_idx < _idom_size, "oob");
724 724 Node* n = _idom[d->_idx];
725 725 assert(n != NULL,"Bad immediate dominator info.");
726 726 while (n->in(0) == NULL) { // Skip dead CFG nodes
727 727 //n = n->in(1);
728 728 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
729 729 assert(n != NULL,"Bad immediate dominator info.");
730 730 }
731 731 return n;
732 732 }
733 733 Node *idom(Node* d) const {
734 734 uint didx = d->_idx;
735 735 Node *n = idom_no_update(d);
736 736 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
737 737 return n;
738 738 }
739 739 uint dom_depth(Node* d) const {
740 740 assert(d->_idx < _idom_size, "");
741 741 return _dom_depth[d->_idx];
742 742 }
743 743 void set_idom(Node* d, Node* n, uint dom_depth);
744 744 // Locally compute IDOM using dom_lca call
745 745 Node *compute_idom( Node *region ) const;
746 746 // Recompute dom_depth
747 747 void recompute_dom_depth();
748 748
749 749 // Is safept not required by an outer loop?
750 750 bool is_deleteable_safept(Node* sfpt);
751 751
752 752 // Replace parallel induction variable (parallel to trip counter)
753 753 void replace_parallel_iv(IdealLoopTree *loop);
754 754
755 755 // Perform verification that the graph is valid.
756 756 PhaseIdealLoop( PhaseIterGVN &igvn) :
757 757 PhaseTransform(Ideal_Loop),
758 758 _igvn(igvn),
759 759 _dom_lca_tags(arena()), // Thread::resource_area
760 760 _verify_me(NULL),
761 761 _verify_only(true) {
762 762 build_and_optimize(false, false);
763 763 }
764 764
765 765 // build the loop tree and perform any requested optimizations
766 766 void build_and_optimize(bool do_split_if, bool skip_loop_opts);
767 767
768 768 public:
769 769 // Dominators for the sea of nodes
770 770 void Dominators();
771 771 Node *dom_lca( Node *n1, Node *n2 ) const {
772 772 return find_non_split_ctrl(dom_lca_internal(n1, n2));
773 773 }
774 774 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
775 775
776 776 // Compute the Ideal Node to Loop mapping
777 777 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
778 778 PhaseTransform(Ideal_Loop),
779 779 _igvn(igvn),
780 780 _dom_lca_tags(arena()), // Thread::resource_area
781 781 _verify_me(NULL),
782 782 _verify_only(false) {
783 783 build_and_optimize(do_split_ifs, skip_loop_opts);
784 784 }
785 785
786 786 // Verify that verify_me made the same decisions as a fresh run.
787 787 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
788 788 PhaseTransform(Ideal_Loop),
789 789 _igvn(igvn),
790 790 _dom_lca_tags(arena()), // Thread::resource_area
791 791 _verify_me(verify_me),
792 792 _verify_only(false) {
793 793 build_and_optimize(false, false);
794 794 }
795 795
796 796 // Build and verify the loop tree without modifying the graph. This
797 797 // is useful to verify that all inputs properly dominate their uses.
798 798 static void verify(PhaseIterGVN& igvn) {
799 799 #ifdef ASSERT
800 800 PhaseIdealLoop v(igvn);
801 801 #endif
802 802 }
803 803
804 804 // True if the method has at least 1 irreducible loop
805 805 bool _has_irreducible_loops;
806 806
807 807 // Per-Node transform
808 808 virtual Node *transform( Node *a_node ) { return 0; }
809 809
810 810 bool is_counted_loop( Node *x, IdealLoopTree *loop );
811 811
812 812 Node* exact_limit( IdealLoopTree *loop );
813 813
814 814 // Return a post-walked LoopNode
815 815 IdealLoopTree *get_loop( Node *n ) const {
816 816 // Dead nodes have no loop, so return the top level loop instead
817 817 if (!has_node(n)) return _ltree_root;
818 818 assert(!has_ctrl(n), "");
819 819 return (IdealLoopTree*)_nodes[n->_idx];
820 820 }
821 821
822 822 // Is 'n' a (nested) member of 'loop'?
823 823 int is_member( const IdealLoopTree *loop, Node *n ) const {
824 824 return loop->is_member(get_loop(n)); }
825 825
826 826 // This is the basic building block of the loop optimizations. It clones an
827 827 // entire loop body. It makes an old_new loop body mapping; with this
828 828 // mapping you can find the new-loop equivalent to an old-loop node. All
829 829 // new-loop nodes are exactly equal to their old-loop counterparts, all
830 830 // edges are the same. All exits from the old-loop now have a RegionNode
831 831 // that merges the equivalent new-loop path. This is true even for the
832 832 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
833 833 // now come from (one or more) Phis that merge their new-loop equivalents.
834 834 // Parameter side_by_side_idom:
835 835 // When side_by_size_idom is NULL, the dominator tree is constructed for
836 836 // the clone loop to dominate the original. Used in construction of
837 837 // pre-main-post loop sequence.
838 838 // When nonnull, the clone and original are side-by-side, both are
839 839 // dominated by the passed in side_by_side_idom node. Used in
840 840 // construction of unswitched loops.
841 841 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
842 842 Node* side_by_side_idom = NULL);
843 843
844 844 // If we got the effect of peeling, either by actually peeling or by
845 845 // making a pre-loop which must execute at least once, we can remove
846 846 // all loop-invariant dominated tests in the main body.
847 847 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
848 848
849 849 // Generate code to do a loop peel for the given loop (and body).
850 850 // old_new is a temp array.
851 851 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
852 852
853 853 // Add pre and post loops around the given loop. These loops are used
854 854 // during RCE, unrolling and aligning loops.
855 855 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
856 856 // If Node n lives in the back_ctrl block, we clone a private version of n
857 857 // in preheader_ctrl block and return that, otherwise return n.
858 858 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
859 859
860 860 // Take steps to maximally unroll the loop. Peel any odd iterations, then
861 861 // unroll to do double iterations. The next round of major loop transforms
862 862 // will repeat till the doubled loop body does all remaining iterations in 1
863 863 // pass.
864 864 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
865 865
866 866 // Unroll the loop body one step - make each trip do 2 iterations.
867 867 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
868 868
869 869 // Return true if exp is a constant times an induction var
870 870 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
871 871
872 872 // Return true if exp is a scaled induction var plus (or minus) constant
873 873 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
874 874
875 875 // Return true if proj is for "proj->[region->..]call_uct"
876 876 static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason);
877 877 // Return true for "if(test)-> proj -> ...
878 878 // |
879 879 // V
880 880 // other_proj->[region->..]call_uct"
881 881 static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason);
882 882 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
883 883 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
884 884 Deoptimization::DeoptReason reason);
885 885 void register_control(Node* n, IdealLoopTree *loop, Node* pred);
886 886
887 887 // Clone loop predicates to cloned loops (peeled, unswitched)
888 888 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
889 889 Deoptimization::DeoptReason reason,
890 890 PhaseIdealLoop* loop_phase,
891 891 PhaseIterGVN* igvn);
892 892
893 893 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
894 894 bool clone_limit_check,
895 895 PhaseIdealLoop* loop_phase,
896 896 PhaseIterGVN* igvn);
897 897 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
898 898
899 899 static Node* skip_loop_predicates(Node* entry);
900 900
901 901 // Find a good location to insert a predicate
902 902 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
903 903 // Find a predicate
904 904 static Node* find_predicate(Node* entry);
905 905 // Construct a range check for a predicate if
906 906 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
907 907 int scale, Node* offset,
908 908 Node* init, Node* limit, Node* stride,
909 909 Node* range, bool upper);
910 910
911 911 // Implementation of the loop predication to promote checks outside the loop
912 912 bool loop_predication_impl(IdealLoopTree *loop);
913 913
914 914 // Helper function to collect predicate for eliminating the useless ones
915 915 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
916 916 void eliminate_useless_predicates();
917 917
918 918 // Change the control input of expensive nodes to allow commoning by
919 919 // IGVN when it is guaranteed to not result in a more frequent
920 920 // execution of the expensive node. Return true if progress.
921 921 bool process_expensive_nodes();
922 922
923 923 // Check whether node has become unreachable
924 924 bool is_node_unreachable(Node *n) const {
925 925 return !has_node(n) || n->is_unreachable(_igvn);
926 926 }
927 927
928 928 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
929 929 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
930 930
931 931 // Create a slow version of the loop by cloning the loop
932 932 // and inserting an if to select fast-slow versions.
933 933 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
934 934 Node_List &old_new);
935 935
936 936 // Clone loop with an invariant test (that does not exit) and
937 937 // insert a clone of the test that selects which version to
938 938 // execute.
939 939 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
940 940
941 941 // Find candidate "if" for unswitching
942 942 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
943 943
944 944 // Range Check Elimination uses this function!
945 945 // Constrain the main loop iterations so the affine function:
946 946 // low_limit <= scale_con * I + offset < upper_limit
947 947 // always holds true. That is, either increase the number of iterations in
948 948 // the pre-loop or the post-loop until the condition holds true in the main
949 949 // loop. Scale_con, offset and limit are all loop invariant.
950 950 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 );
951 951 // Helper function for add_constraint().
952 952 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
953 953
|
↓ open down ↓ |
953 lines elided |
↑ open up ↑ |
954 954 // Partially peel loop up through last_peel node.
955 955 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
956 956
957 957 // Create a scheduled list of nodes control dependent on ctrl set.
958 958 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
959 959 // Has a use in the vector set
960 960 bool has_use_in_set( Node* n, VectorSet& vset );
961 961 // Has use internal to the vector set (ie. not in a phi at the loop head)
962 962 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
963 963 // clone "n" for uses that are outside of loop
964 - void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
964 + int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
965 965 // clone "n" for special uses that are in the not_peeled region
966 966 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
967 967 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
968 968 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
969 969 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
970 970 #ifdef ASSERT
971 971 // Validate the loop partition sets: peel and not_peel
972 972 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
973 973 // Ensure that uses outside of loop are of the right form
974 974 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
975 975 uint orig_exit_idx, uint clone_exit_idx);
976 976 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
977 977 #endif
978 978
979 979 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
980 980 int stride_of_possible_iv( Node* iff );
981 981 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
982 982 // Return the (unique) control output node that's in the loop (if it exists.)
983 983 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
984 984 // Insert a signed compare loop exit cloned from an unsigned compare.
985 985 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
986 986 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
987 987 // Utility to register node "n" with PhaseIdealLoop
988 988 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
989 989 // Utility to create an if-projection
990 990 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
991 991 // Force the iff control output to be the live_proj
992 992 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
993 993 // Insert a region before an if projection
994 994 RegionNode* insert_region_before_proj(ProjNode* proj);
995 995 // Insert a new if before an if projection
996 996 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
997 997
998 998 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
999 999 // "Nearly" because all Nodes have been cloned from the original in the loop,
1000 1000 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1001 1001 // through the Phi recursively, and return a Bool.
1002 1002 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1003 1003 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1004 1004
1005 1005
1006 1006 // Rework addressing expressions to get the most loop-invariant stuff
1007 1007 // moved out. We'd like to do all associative operators, but it's especially
1008 1008 // important (common) to do address expressions.
1009 1009 Node *remix_address_expressions( Node *n );
1010 1010
1011 1011 // Attempt to use a conditional move instead of a phi/branch
1012 1012 Node *conditional_move( Node *n );
1013 1013
1014 1014 // Reorganize offset computations to lower register pressure.
1015 1015 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1016 1016 // (which are then alive with the post-incremented trip counter
1017 1017 // forcing an extra register move)
1018 1018 void reorg_offsets( IdealLoopTree *loop );
1019 1019
1020 1020 // Check for aggressive application of 'split-if' optimization,
1021 1021 // using basic block level info.
1022 1022 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1023 1023 Node *split_if_with_blocks_pre ( Node *n );
1024 1024 void split_if_with_blocks_post( Node *n );
1025 1025 Node *has_local_phi_input( Node *n );
1026 1026 // Mark an IfNode as being dominated by a prior test,
1027 1027 // without actually altering the CFG (and hence IDOM info).
1028 1028 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1029 1029
1030 1030 // Split Node 'n' through merge point
1031 1031 Node *split_thru_region( Node *n, Node *region );
1032 1032 // Split Node 'n' through merge point if there is enough win.
1033 1033 Node *split_thru_phi( Node *n, Node *region, int policy );
1034 1034 // Found an If getting its condition-code input from a Phi in the
1035 1035 // same block. Split thru the Region.
1036 1036 void do_split_if( Node *iff );
1037 1037
1038 1038 // Conversion of fill/copy patterns into intrisic versions
1039 1039 bool do_intrinsify_fill();
1040 1040 bool intrinsify_fill(IdealLoopTree* lpt);
1041 1041 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1042 1042 Node*& shift, Node*& offset);
1043 1043
1044 1044 private:
1045 1045 // Return a type based on condition control flow
1046 1046 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1047 1047 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1048 1048 // Helpers for filtered type
1049 1049 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1050 1050
1051 1051 // Helper functions
1052 1052 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1053 1053 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1054 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 1055 bool split_up( Node *n, Node *blk1, Node *blk2 );
1056 1056 void sink_use( Node *use, Node *post_loop );
1057 1057 Node *place_near_use( Node *useblock ) const;
1058 1058
1059 1059 bool _created_loop_node;
1060 1060 public:
1061 1061 void set_created_loop_node() { _created_loop_node = true; }
1062 1062 bool created_loop_node() { return _created_loop_node; }
1063 1063 void register_new_node( Node *n, Node *blk );
1064 1064
1065 1065 #ifdef ASSERT
1066 1066 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1067 1067 #endif
1068 1068
1069 1069 #ifndef PRODUCT
1070 1070 void dump( ) const;
1071 1071 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1072 1072 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1073 1073 void verify() const; // Major slow :-)
1074 1074 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1075 1075 IdealLoopTree *get_loop_idx(Node* n) const {
1076 1076 // Dead nodes have no loop, so return the top level loop instead
1077 1077 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1078 1078 }
1079 1079 // Print some stats
1080 1080 static void print_statistics();
1081 1081 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1082 1082 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1083 1083 #endif
1084 1084 };
1085 1085
1086 1086 inline Node* IdealLoopTree::tail() {
1087 1087 // Handle lazy update of _tail field
1088 1088 Node *n = _tail;
1089 1089 //while( !n->in(0) ) // Skip dead CFG nodes
1090 1090 //n = n->in(1);
1091 1091 if (n->in(0) == NULL)
1092 1092 n = _phase->get_ctrl(n);
1093 1093 _tail = n;
1094 1094 return n;
1095 1095 }
1096 1096
1097 1097
1098 1098 // Iterate over the loop tree using a preorder, left-to-right traversal.
1099 1099 //
1100 1100 // Example that visits all counted loops from within PhaseIdealLoop
1101 1101 //
1102 1102 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1103 1103 // IdealLoopTree* lpt = iter.current();
1104 1104 // if (!lpt->is_counted()) continue;
1105 1105 // ...
1106 1106 class LoopTreeIterator : public StackObj {
1107 1107 private:
1108 1108 IdealLoopTree* _root;
1109 1109 IdealLoopTree* _curnt;
1110 1110
1111 1111 public:
1112 1112 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1113 1113
1114 1114 bool done() { return _curnt == NULL; } // Finished iterating?
1115 1115
1116 1116 void next(); // Advance to next loop tree
1117 1117
1118 1118 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1119 1119 };
1120 1120
1121 1121 #endif // SHARE_VM_OPTO_LOOPNODE_HPP
|
↓ open down ↓ |
147 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX