re PR tree-optimization/45397 (Issues with integer narrowing conversions)
[gcc.git] / gcc / tree-ssa-pre.c
1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2017 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
12
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
21
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "backend.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "gimple.h"
29 #include "predict.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
32 #include "ssa.h"
33 #include "cgraph.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
36 #include "cfganal.h"
37 #include "gimple-fold.h"
38 #include "tree-eh.h"
39 #include "gimplify.h"
40 #include "gimple-iterator.h"
41 #include "tree-cfg.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
44 #include "tree-dfa.h"
45 #include "tree-ssa.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
51 #include "domwalk.h"
52 #include "tree-ssa-propagate.h"
53 #include "ipa-utils.h"
54 #include "tree-cfgcleanup.h"
55 #include "langhooks.h"
56 #include "alias.h"
57
58 /* Even though this file is called tree-ssa-pre.c, we actually
59 implement a bit more than just PRE here. All of them piggy-back
60 on GVN which is implemented in tree-ssa-sccvn.c.
61
62 1. Full Redundancy Elimination (FRE)
63 This is the elimination phase of GVN.
64
65 2. Partial Redundancy Elimination (PRE)
66 This is adds computation of AVAIL_OUT and ANTIC_IN and
67 doing expression insertion to form GVN-PRE.
68
69 3. Code hoisting
70 This optimization uses the ANTIC_IN sets computed for PRE
71 to move expressions further up than PRE would do, to make
72 multiple computations of the same value fully redundant.
73 This pass is explained below (after the explanation of the
74 basic algorithm for PRE).
75 */
76
77 /* TODO:
78
79 1. Avail sets can be shared by making an avail_find_leader that
80 walks up the dominator tree and looks in those avail sets.
81 This might affect code optimality, it's unclear right now.
82 Currently the AVAIL_OUT sets are the remaining quadraticness in
83 memory of GVN-PRE.
84 2. Strength reduction can be performed by anticipating expressions
85 we can repair later on.
86 3. We can do back-substitution or smarter value numbering to catch
87 commutative expressions split up over multiple statements.
88 */
89
90 /* For ease of terminology, "expression node" in the below refers to
91 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
92 represent the actual statement containing the expressions we care about,
93 and we cache the value number by putting it in the expression. */
94
95 /* Basic algorithm for Partial Redundancy Elimination:
96
97 First we walk the statements to generate the AVAIL sets, the
98 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
99 generation of values/expressions by a given block. We use them
100 when computing the ANTIC sets. The AVAIL sets consist of
101 SSA_NAME's that represent values, so we know what values are
102 available in what blocks. AVAIL is a forward dataflow problem. In
103 SSA, values are never killed, so we don't need a kill set, or a
104 fixpoint iteration, in order to calculate the AVAIL sets. In
105 traditional parlance, AVAIL sets tell us the downsafety of the
106 expressions/values.
107
108 Next, we generate the ANTIC sets. These sets represent the
109 anticipatable expressions. ANTIC is a backwards dataflow
110 problem. An expression is anticipatable in a given block if it could
111 be generated in that block. This means that if we had to perform
112 an insertion in that block, of the value of that expression, we
113 could. Calculating the ANTIC sets requires phi translation of
114 expressions, because the flow goes backwards through phis. We must
115 iterate to a fixpoint of the ANTIC sets, because we have a kill
116 set. Even in SSA form, values are not live over the entire
117 function, only from their definition point onwards. So we have to
118 remove values from the ANTIC set once we go past the definition
119 point of the leaders that make them up.
120 compute_antic/compute_antic_aux performs this computation.
121
122 Third, we perform insertions to make partially redundant
123 expressions fully redundant.
124
125 An expression is partially redundant (excluding partial
126 anticipation) if:
127
128 1. It is AVAIL in some, but not all, of the predecessors of a
129 given block.
130 2. It is ANTIC in all the predecessors.
131
132 In order to make it fully redundant, we insert the expression into
133 the predecessors where it is not available, but is ANTIC.
134
135 When optimizing for size, we only eliminate the partial redundancy
136 if we need to insert in only one predecessor. This avoids almost
137 completely the code size increase that PRE usually causes.
138
139 For the partial anticipation case, we only perform insertion if it
140 is partially anticipated in some block, and fully available in all
141 of the predecessors.
142
143 do_pre_regular_insertion/do_pre_partial_partial_insertion
144 performs these steps, driven by insert/insert_aux.
145
146 Fourth, we eliminate fully redundant expressions.
147 This is a simple statement walk that replaces redundant
148 calculations with the now available values. */
149
150 /* Basic algorithm for Code Hoisting:
151
152 Code hoisting is: Moving value computations up in the control flow
153 graph to make multiple copies redundant. Typically this is a size
154 optimization, but there are cases where it also is helpful for speed.
155
156 A simple code hoisting algorithm is implemented that piggy-backs on
157 the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
158 which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
159 computed for PRE, and we can use them to perform a limited version of
160 code hoisting, too.
161
162 For the purpose of this implementation, a value is hoistable to a basic
163 block B if the following properties are met:
164
165 1. The value is in ANTIC_IN(B) -- the value will be computed on all
166 paths from B to function exit and it can be computed in B);
167
168 2. The value is not in AVAIL_OUT(B) -- there would be no need to
169 compute the value again and make it available twice;
170
171 3. All successors of B are dominated by B -- makes sure that inserting
172 a computation of the value in B will make the remaining
173 computations fully redundant;
174
175 4. At least one successor has the value in AVAIL_OUT -- to avoid
176 hoisting values up too far;
177
178 5. There are at least two successors of B -- hoisting in straight
179 line code is pointless.
180
181 The third condition is not strictly necessary, but it would complicate
182 the hoisting pass a lot. In fact, I don't know of any code hoisting
183 algorithm that does not have this requirement. Fortunately, experiments
184 have show that most candidate hoistable values are in regions that meet
185 this condition (e.g. diamond-shape regions).
186
187 The forth condition is necessary to avoid hoisting things up too far
188 away from the uses of the value. Nothing else limits the algorithm
189 from hoisting everything up as far as ANTIC_IN allows. Experiments
190 with SPEC and CSiBE have shown that hoisting up too far results in more
191 spilling, less benefits for code size, and worse benchmark scores.
192 Fortunately, in practice most of the interesting hoisting opportunities
193 are caught despite this limitation.
194
195 For hoistable values that meet all conditions, expressions are inserted
196 to make the calculation of the hoistable value fully redundant. We
197 perform code hoisting insertions after each round of PRE insertions,
198 because code hoisting never exposes new PRE opportunities, but PRE can
199 create new code hoisting opportunities.
200
201 The code hoisting algorithm is implemented in do_hoist_insert, driven
202 by insert/insert_aux. */
203
204 /* Representations of value numbers:
205
206 Value numbers are represented by a representative SSA_NAME. We
207 will create fake SSA_NAME's in situations where we need a
208 representative but do not have one (because it is a complex
209 expression). In order to facilitate storing the value numbers in
210 bitmaps, and keep the number of wasted SSA_NAME's down, we also
211 associate a value_id with each value number, and create full blown
212 ssa_name's only where we actually need them (IE in operands of
213 existing expressions).
214
215 Theoretically you could replace all the value_id's with
216 SSA_NAME_VERSION, but this would allocate a large number of
217 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
218 It would also require an additional indirection at each point we
219 use the value id. */
220
221 /* Representation of expressions on value numbers:
222
223 Expressions consisting of value numbers are represented the same
224 way as our VN internally represents them, with an additional
225 "pre_expr" wrapping around them in order to facilitate storing all
226 of the expressions in the same sets. */
227
228 /* Representation of sets:
229
230 The dataflow sets do not need to be sorted in any particular order
231 for the majority of their lifetime, are simply represented as two
232 bitmaps, one that keeps track of values present in the set, and one
233 that keeps track of expressions present in the set.
234
235 When we need them in topological order, we produce it on demand by
236 transforming the bitmap into an array and sorting it into topo
237 order. */
238
239 /* Type of expression, used to know which member of the PRE_EXPR union
240 is valid. */
241
242 enum pre_expr_kind
243 {
244 NAME,
245 NARY,
246 REFERENCE,
247 CONSTANT
248 };
249
250 union pre_expr_union
251 {
252 tree name;
253 tree constant;
254 vn_nary_op_t nary;
255 vn_reference_t reference;
256 };
257
258 typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
259 {
260 enum pre_expr_kind kind;
261 unsigned int id;
262 pre_expr_union u;
263
264 /* hash_table support. */
265 static inline hashval_t hash (const pre_expr_d *);
266 static inline int equal (const pre_expr_d *, const pre_expr_d *);
267 } *pre_expr;
268
269 #define PRE_EXPR_NAME(e) (e)->u.name
270 #define PRE_EXPR_NARY(e) (e)->u.nary
271 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
272 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
273
274 /* Compare E1 and E1 for equality. */
275
276 inline int
277 pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
278 {
279 if (e1->kind != e2->kind)
280 return false;
281
282 switch (e1->kind)
283 {
284 case CONSTANT:
285 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
286 PRE_EXPR_CONSTANT (e2));
287 case NAME:
288 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
289 case NARY:
290 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
291 case REFERENCE:
292 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
293 PRE_EXPR_REFERENCE (e2));
294 default:
295 gcc_unreachable ();
296 }
297 }
298
299 /* Hash E. */
300
301 inline hashval_t
302 pre_expr_d::hash (const pre_expr_d *e)
303 {
304 switch (e->kind)
305 {
306 case CONSTANT:
307 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
308 case NAME:
309 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
310 case NARY:
311 return PRE_EXPR_NARY (e)->hashcode;
312 case REFERENCE:
313 return PRE_EXPR_REFERENCE (e)->hashcode;
314 default:
315 gcc_unreachable ();
316 }
317 }
318
319 /* Next global expression id number. */
320 static unsigned int next_expression_id;
321
322 /* Mapping from expression to id number we can use in bitmap sets. */
323 static vec<pre_expr> expressions;
324 static hash_table<pre_expr_d> *expression_to_id;
325 static vec<unsigned> name_to_id;
326
327 /* Allocate an expression id for EXPR. */
328
329 static inline unsigned int
330 alloc_expression_id (pre_expr expr)
331 {
332 struct pre_expr_d **slot;
333 /* Make sure we won't overflow. */
334 gcc_assert (next_expression_id + 1 > next_expression_id);
335 expr->id = next_expression_id++;
336 expressions.safe_push (expr);
337 if (expr->kind == NAME)
338 {
339 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
340 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
341 re-allocations by using vec::reserve upfront. */
342 unsigned old_len = name_to_id.length ();
343 name_to_id.reserve (num_ssa_names - old_len);
344 name_to_id.quick_grow_cleared (num_ssa_names);
345 gcc_assert (name_to_id[version] == 0);
346 name_to_id[version] = expr->id;
347 }
348 else
349 {
350 slot = expression_to_id->find_slot (expr, INSERT);
351 gcc_assert (!*slot);
352 *slot = expr;
353 }
354 return next_expression_id - 1;
355 }
356
357 /* Return the expression id for tree EXPR. */
358
359 static inline unsigned int
360 get_expression_id (const pre_expr expr)
361 {
362 return expr->id;
363 }
364
365 static inline unsigned int
366 lookup_expression_id (const pre_expr expr)
367 {
368 struct pre_expr_d **slot;
369
370 if (expr->kind == NAME)
371 {
372 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
373 if (name_to_id.length () <= version)
374 return 0;
375 return name_to_id[version];
376 }
377 else
378 {
379 slot = expression_to_id->find_slot (expr, NO_INSERT);
380 if (!slot)
381 return 0;
382 return ((pre_expr)*slot)->id;
383 }
384 }
385
386 /* Return the existing expression id for EXPR, or create one if one
387 does not exist yet. */
388
389 static inline unsigned int
390 get_or_alloc_expression_id (pre_expr expr)
391 {
392 unsigned int id = lookup_expression_id (expr);
393 if (id == 0)
394 return alloc_expression_id (expr);
395 return expr->id = id;
396 }
397
398 /* Return the expression that has expression id ID */
399
400 static inline pre_expr
401 expression_for_id (unsigned int id)
402 {
403 return expressions[id];
404 }
405
406 /* Free the expression id field in all of our expressions,
407 and then destroy the expressions array. */
408
409 static void
410 clear_expression_ids (void)
411 {
412 expressions.release ();
413 }
414
415 static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
416
417 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
418
419 static pre_expr
420 get_or_alloc_expr_for_name (tree name)
421 {
422 struct pre_expr_d expr;
423 pre_expr result;
424 unsigned int result_id;
425
426 expr.kind = NAME;
427 expr.id = 0;
428 PRE_EXPR_NAME (&expr) = name;
429 result_id = lookup_expression_id (&expr);
430 if (result_id != 0)
431 return expression_for_id (result_id);
432
433 result = pre_expr_pool.allocate ();
434 result->kind = NAME;
435 PRE_EXPR_NAME (result) = name;
436 alloc_expression_id (result);
437 return result;
438 }
439
440 /* An unordered bitmap set. One bitmap tracks values, the other,
441 expressions. */
442 typedef struct bitmap_set
443 {
444 bitmap_head expressions;
445 bitmap_head values;
446 } *bitmap_set_t;
447
448 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
449 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
450
451 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
452 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
453
454 /* Mapping from value id to expressions with that value_id. */
455 static vec<bitmap> value_expressions;
456
457 /* Sets that we need to keep track of. */
458 typedef struct bb_bitmap_sets
459 {
460 /* The EXP_GEN set, which represents expressions/values generated in
461 a basic block. */
462 bitmap_set_t exp_gen;
463
464 /* The PHI_GEN set, which represents PHI results generated in a
465 basic block. */
466 bitmap_set_t phi_gen;
467
468 /* The TMP_GEN set, which represents results/temporaries generated
469 in a basic block. IE the LHS of an expression. */
470 bitmap_set_t tmp_gen;
471
472 /* The AVAIL_OUT set, which represents which values are available in
473 a given basic block. */
474 bitmap_set_t avail_out;
475
476 /* The ANTIC_IN set, which represents which values are anticipatable
477 in a given basic block. */
478 bitmap_set_t antic_in;
479
480 /* The PA_IN set, which represents which values are
481 partially anticipatable in a given basic block. */
482 bitmap_set_t pa_in;
483
484 /* The NEW_SETS set, which is used during insertion to augment the
485 AVAIL_OUT set of blocks with the new insertions performed during
486 the current iteration. */
487 bitmap_set_t new_sets;
488
489 /* A cache for value_dies_in_block_x. */
490 bitmap expr_dies;
491
492 /* The live virtual operand on successor edges. */
493 tree vop_on_exit;
494
495 /* True if we have visited this block during ANTIC calculation. */
496 unsigned int visited : 1;
497
498 /* True when the block contains a call that might not return. */
499 unsigned int contains_may_not_return_call : 1;
500 } *bb_value_sets_t;
501
502 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
503 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
504 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
505 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
506 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
507 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
508 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
509 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
510 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
511 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
512 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
513
514
515 /* This structure is used to keep track of statistics on what
516 optimization PRE was able to perform. */
517 static struct
518 {
519 /* The number of RHS computations eliminated by PRE. */
520 int eliminations;
521
522 /* The number of new expressions/temporaries generated by PRE. */
523 int insertions;
524
525 /* The number of inserts found due to partial anticipation */
526 int pa_insert;
527
528 /* The number of inserts made for code hoisting. */
529 int hoist_insert;
530
531 /* The number of new PHI nodes added by PRE. */
532 int phis;
533 } pre_stats;
534
535 static bool do_partial_partial;
536 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
537 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
538 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
539 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
540 static void bitmap_set_and (bitmap_set_t, bitmap_set_t);
541 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
542 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
543 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
544 unsigned int, bool);
545 static bitmap_set_t bitmap_set_new (void);
546 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
547 tree);
548 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
549 static unsigned int get_expr_value_id (pre_expr);
550
551 /* We can add and remove elements and entries to and from sets
552 and hash tables, so we use alloc pools for them. */
553
554 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
555 static bitmap_obstack grand_bitmap_obstack;
556
557 /* Set of blocks with statements that have had their EH properties changed. */
558 static bitmap need_eh_cleanup;
559
560 /* Set of blocks with statements that have had their AB properties changed. */
561 static bitmap need_ab_cleanup;
562
563 /* A three tuple {e, pred, v} used to cache phi translations in the
564 phi_translate_table. */
565
566 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
567 {
568 /* The expression. */
569 pre_expr e;
570
571 /* The predecessor block along which we translated the expression. */
572 basic_block pred;
573
574 /* The value that resulted from the translation. */
575 pre_expr v;
576
577 /* The hashcode for the expression, pred pair. This is cached for
578 speed reasons. */
579 hashval_t hashcode;
580
581 /* hash_table support. */
582 static inline hashval_t hash (const expr_pred_trans_d *);
583 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
584 } *expr_pred_trans_t;
585 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
586
587 inline hashval_t
588 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
589 {
590 return e->hashcode;
591 }
592
593 inline int
594 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
595 const expr_pred_trans_d *ve2)
596 {
597 basic_block b1 = ve1->pred;
598 basic_block b2 = ve2->pred;
599
600 /* If they are not translations for the same basic block, they can't
601 be equal. */
602 if (b1 != b2)
603 return false;
604 return pre_expr_d::equal (ve1->e, ve2->e);
605 }
606
607 /* The phi_translate_table caches phi translations for a given
608 expression and predecessor. */
609 static hash_table<expr_pred_trans_d> *phi_translate_table;
610
611 /* Add the tuple mapping from {expression E, basic block PRED} to
612 the phi translation table and return whether it pre-existed. */
613
614 static inline bool
615 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
616 {
617 expr_pred_trans_t *slot;
618 expr_pred_trans_d tem;
619 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
620 pred->index);
621 tem.e = e;
622 tem.pred = pred;
623 tem.hashcode = hash;
624 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
625 if (*slot)
626 {
627 *entry = *slot;
628 return true;
629 }
630
631 *entry = *slot = XNEW (struct expr_pred_trans_d);
632 (*entry)->e = e;
633 (*entry)->pred = pred;
634 (*entry)->hashcode = hash;
635 return false;
636 }
637
638
639 /* Add expression E to the expression set of value id V. */
640
641 static void
642 add_to_value (unsigned int v, pre_expr e)
643 {
644 bitmap set;
645
646 gcc_checking_assert (get_expr_value_id (e) == v);
647
648 if (v >= value_expressions.length ())
649 {
650 value_expressions.safe_grow_cleared (v + 1);
651 }
652
653 set = value_expressions[v];
654 if (!set)
655 {
656 set = BITMAP_ALLOC (&grand_bitmap_obstack);
657 value_expressions[v] = set;
658 }
659
660 bitmap_set_bit (set, get_or_alloc_expression_id (e));
661 }
662
663 /* Create a new bitmap set and return it. */
664
665 static bitmap_set_t
666 bitmap_set_new (void)
667 {
668 bitmap_set_t ret = bitmap_set_pool.allocate ();
669 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
670 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
671 return ret;
672 }
673
674 /* Return the value id for a PRE expression EXPR. */
675
676 static unsigned int
677 get_expr_value_id (pre_expr expr)
678 {
679 unsigned int id;
680 switch (expr->kind)
681 {
682 case CONSTANT:
683 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
684 break;
685 case NAME:
686 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
687 break;
688 case NARY:
689 id = PRE_EXPR_NARY (expr)->value_id;
690 break;
691 case REFERENCE:
692 id = PRE_EXPR_REFERENCE (expr)->value_id;
693 break;
694 default:
695 gcc_unreachable ();
696 }
697 /* ??? We cannot assert that expr has a value-id (it can be 0), because
698 we assign value-ids only to expressions that have a result
699 in set_hashtable_value_ids. */
700 return id;
701 }
702
703 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
704
705 static tree
706 sccvn_valnum_from_value_id (unsigned int val)
707 {
708 bitmap_iterator bi;
709 unsigned int i;
710 bitmap exprset = value_expressions[val];
711 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
712 {
713 pre_expr vexpr = expression_for_id (i);
714 if (vexpr->kind == NAME)
715 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
716 else if (vexpr->kind == CONSTANT)
717 return PRE_EXPR_CONSTANT (vexpr);
718 }
719 return NULL_TREE;
720 }
721
722 /* Remove an expression EXPR from a bitmapped set. */
723
724 static void
725 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
726 {
727 unsigned int val = get_expr_value_id (expr);
728 if (!value_id_constant_p (val))
729 {
730 bitmap_clear_bit (&set->values, val);
731 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
732 }
733 }
734
735 static void
736 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
737 unsigned int val, bool allow_constants)
738 {
739 if (allow_constants || !value_id_constant_p (val))
740 {
741 /* We specifically expect this and only this function to be able to
742 insert constants into a set. */
743 bitmap_set_bit (&set->values, val);
744 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
745 }
746 }
747
748 /* Insert an expression EXPR into a bitmapped set. */
749
750 static void
751 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
752 {
753 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
754 }
755
756 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
757
758 static void
759 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
760 {
761 bitmap_copy (&dest->expressions, &orig->expressions);
762 bitmap_copy (&dest->values, &orig->values);
763 }
764
765
766 /* Free memory used up by SET. */
767 static void
768 bitmap_set_free (bitmap_set_t set)
769 {
770 bitmap_clear (&set->expressions);
771 bitmap_clear (&set->values);
772 }
773
774
775 /* Generate an topological-ordered array of bitmap set SET. */
776
777 static vec<pre_expr>
778 sorted_array_from_bitmap_set (bitmap_set_t set)
779 {
780 unsigned int i, j;
781 bitmap_iterator bi, bj;
782 vec<pre_expr> result;
783
784 /* Pre-allocate enough space for the array. */
785 result.create (bitmap_count_bits (&set->expressions));
786
787 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
788 {
789 /* The number of expressions having a given value is usually
790 relatively small. Thus, rather than making a vector of all
791 the expressions and sorting it by value-id, we walk the values
792 and check in the reverse mapping that tells us what expressions
793 have a given value, to filter those in our set. As a result,
794 the expressions are inserted in value-id order, which means
795 topological order.
796
797 If this is somehow a significant lose for some cases, we can
798 choose which set to walk based on the set size. */
799 bitmap exprset = value_expressions[i];
800 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
801 {
802 if (bitmap_bit_p (&set->expressions, j))
803 result.quick_push (expression_for_id (j));
804 }
805 }
806
807 return result;
808 }
809
810 /* Perform bitmapped set operation DEST &= ORIG. */
811
812 static void
813 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
814 {
815 bitmap_iterator bi;
816 unsigned int i;
817
818 if (dest != orig)
819 {
820 bitmap_head temp;
821 bitmap_initialize (&temp, &grand_bitmap_obstack);
822
823 bitmap_and_into (&dest->values, &orig->values);
824 bitmap_copy (&temp, &dest->expressions);
825 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
826 {
827 pre_expr expr = expression_for_id (i);
828 unsigned int value_id = get_expr_value_id (expr);
829 if (!bitmap_bit_p (&dest->values, value_id))
830 bitmap_clear_bit (&dest->expressions, i);
831 }
832 bitmap_clear (&temp);
833 }
834 }
835
836 /* Subtract all values and expressions contained in ORIG from DEST. */
837
838 static bitmap_set_t
839 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
840 {
841 bitmap_set_t result = bitmap_set_new ();
842 bitmap_iterator bi;
843 unsigned int i;
844
845 bitmap_and_compl (&result->expressions, &dest->expressions,
846 &orig->expressions);
847
848 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
849 {
850 pre_expr expr = expression_for_id (i);
851 unsigned int value_id = get_expr_value_id (expr);
852 bitmap_set_bit (&result->values, value_id);
853 }
854
855 return result;
856 }
857
858 /* Subtract all the values in bitmap set B from bitmap set A. */
859
860 static void
861 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
862 {
863 unsigned int i;
864 bitmap_iterator bi;
865 bitmap_head temp;
866
867 bitmap_initialize (&temp, &grand_bitmap_obstack);
868
869 bitmap_copy (&temp, &a->expressions);
870 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
871 {
872 pre_expr expr = expression_for_id (i);
873 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
874 bitmap_remove_from_set (a, expr);
875 }
876 bitmap_clear (&temp);
877 }
878
879
880 /* Return true if bitmapped set SET contains the value VALUE_ID. */
881
882 static bool
883 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
884 {
885 if (value_id_constant_p (value_id))
886 return true;
887
888 if (!set || bitmap_empty_p (&set->expressions))
889 return false;
890
891 return bitmap_bit_p (&set->values, value_id);
892 }
893
894 static inline bool
895 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
896 {
897 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
898 }
899
900 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
901
902 static void
903 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
904 const pre_expr expr)
905 {
906 bitmap exprset;
907 unsigned int i;
908 bitmap_iterator bi;
909
910 if (value_id_constant_p (lookfor))
911 return;
912
913 if (!bitmap_set_contains_value (set, lookfor))
914 return;
915
916 /* The number of expressions having a given value is usually
917 significantly less than the total number of expressions in SET.
918 Thus, rather than check, for each expression in SET, whether it
919 has the value LOOKFOR, we walk the reverse mapping that tells us
920 what expressions have a given value, and see if any of those
921 expressions are in our set. For large testcases, this is about
922 5-10x faster than walking the bitmap. If this is somehow a
923 significant lose for some cases, we can choose which set to walk
924 based on the set size. */
925 exprset = value_expressions[lookfor];
926 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
927 {
928 if (bitmap_clear_bit (&set->expressions, i))
929 {
930 bitmap_set_bit (&set->expressions, get_expression_id (expr));
931 return;
932 }
933 }
934
935 gcc_unreachable ();
936 }
937
938 /* Return true if two bitmap sets are equal. */
939
940 static bool
941 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
942 {
943 return bitmap_equal_p (&a->values, &b->values);
944 }
945
946 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
947 and add it otherwise. */
948
949 static void
950 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
951 {
952 unsigned int val = get_expr_value_id (expr);
953
954 if (bitmap_set_contains_value (set, val))
955 bitmap_set_replace_value (set, val, expr);
956 else
957 bitmap_insert_into_set (set, expr);
958 }
959
960 /* Insert EXPR into SET if EXPR's value is not already present in
961 SET. */
962
963 static void
964 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
965 {
966 unsigned int val = get_expr_value_id (expr);
967
968 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
969
970 /* Constant values are always considered to be part of the set. */
971 if (value_id_constant_p (val))
972 return;
973
974 /* If the value membership changed, add the expression. */
975 if (bitmap_set_bit (&set->values, val))
976 bitmap_set_bit (&set->expressions, expr->id);
977 }
978
979 /* Print out EXPR to outfile. */
980
981 static void
982 print_pre_expr (FILE *outfile, const pre_expr expr)
983 {
984 switch (expr->kind)
985 {
986 case CONSTANT:
987 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
988 break;
989 case NAME:
990 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
991 break;
992 case NARY:
993 {
994 unsigned int i;
995 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
996 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
997 for (i = 0; i < nary->length; i++)
998 {
999 print_generic_expr (outfile, nary->op[i], 0);
1000 if (i != (unsigned) nary->length - 1)
1001 fprintf (outfile, ",");
1002 }
1003 fprintf (outfile, "}");
1004 }
1005 break;
1006
1007 case REFERENCE:
1008 {
1009 vn_reference_op_t vro;
1010 unsigned int i;
1011 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1012 fprintf (outfile, "{");
1013 for (i = 0;
1014 ref->operands.iterate (i, &vro);
1015 i++)
1016 {
1017 bool closebrace = false;
1018 if (vro->opcode != SSA_NAME
1019 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
1020 {
1021 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
1022 if (vro->op0)
1023 {
1024 fprintf (outfile, "<");
1025 closebrace = true;
1026 }
1027 }
1028 if (vro->op0)
1029 {
1030 print_generic_expr (outfile, vro->op0, 0);
1031 if (vro->op1)
1032 {
1033 fprintf (outfile, ",");
1034 print_generic_expr (outfile, vro->op1, 0);
1035 }
1036 if (vro->op2)
1037 {
1038 fprintf (outfile, ",");
1039 print_generic_expr (outfile, vro->op2, 0);
1040 }
1041 }
1042 if (closebrace)
1043 fprintf (outfile, ">");
1044 if (i != ref->operands.length () - 1)
1045 fprintf (outfile, ",");
1046 }
1047 fprintf (outfile, "}");
1048 if (ref->vuse)
1049 {
1050 fprintf (outfile, "@");
1051 print_generic_expr (outfile, ref->vuse, 0);
1052 }
1053 }
1054 break;
1055 }
1056 }
1057 void debug_pre_expr (pre_expr);
1058
1059 /* Like print_pre_expr but always prints to stderr. */
1060 DEBUG_FUNCTION void
1061 debug_pre_expr (pre_expr e)
1062 {
1063 print_pre_expr (stderr, e);
1064 fprintf (stderr, "\n");
1065 }
1066
1067 /* Print out SET to OUTFILE. */
1068
1069 static void
1070 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1071 const char *setname, int blockindex)
1072 {
1073 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1074 if (set)
1075 {
1076 bool first = true;
1077 unsigned i;
1078 bitmap_iterator bi;
1079
1080 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1081 {
1082 const pre_expr expr = expression_for_id (i);
1083
1084 if (!first)
1085 fprintf (outfile, ", ");
1086 first = false;
1087 print_pre_expr (outfile, expr);
1088
1089 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1090 }
1091 }
1092 fprintf (outfile, " }\n");
1093 }
1094
1095 void debug_bitmap_set (bitmap_set_t);
1096
1097 DEBUG_FUNCTION void
1098 debug_bitmap_set (bitmap_set_t set)
1099 {
1100 print_bitmap_set (stderr, set, "debug", 0);
1101 }
1102
1103 void debug_bitmap_sets_for (basic_block);
1104
1105 DEBUG_FUNCTION void
1106 debug_bitmap_sets_for (basic_block bb)
1107 {
1108 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1109 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1110 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1111 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1112 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1113 if (do_partial_partial)
1114 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1115 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1116 }
1117
1118 /* Print out the expressions that have VAL to OUTFILE. */
1119
1120 static void
1121 print_value_expressions (FILE *outfile, unsigned int val)
1122 {
1123 bitmap set = value_expressions[val];
1124 if (set)
1125 {
1126 bitmap_set x;
1127 char s[10];
1128 sprintf (s, "%04d", val);
1129 x.expressions = *set;
1130 print_bitmap_set (outfile, &x, s, 0);
1131 }
1132 }
1133
1134
1135 DEBUG_FUNCTION void
1136 debug_value_expressions (unsigned int val)
1137 {
1138 print_value_expressions (stderr, val);
1139 }
1140
1141 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1142 represent it. */
1143
1144 static pre_expr
1145 get_or_alloc_expr_for_constant (tree constant)
1146 {
1147 unsigned int result_id;
1148 unsigned int value_id;
1149 struct pre_expr_d expr;
1150 pre_expr newexpr;
1151
1152 expr.kind = CONSTANT;
1153 PRE_EXPR_CONSTANT (&expr) = constant;
1154 result_id = lookup_expression_id (&expr);
1155 if (result_id != 0)
1156 return expression_for_id (result_id);
1157
1158 newexpr = pre_expr_pool.allocate ();
1159 newexpr->kind = CONSTANT;
1160 PRE_EXPR_CONSTANT (newexpr) = constant;
1161 alloc_expression_id (newexpr);
1162 value_id = get_or_alloc_constant_value_id (constant);
1163 add_to_value (value_id, newexpr);
1164 return newexpr;
1165 }
1166
1167 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1168 Currently only supports constants and SSA_NAMES. */
1169 static pre_expr
1170 get_or_alloc_expr_for (tree t)
1171 {
1172 if (TREE_CODE (t) == SSA_NAME)
1173 return get_or_alloc_expr_for_name (t);
1174 else if (is_gimple_min_invariant (t))
1175 return get_or_alloc_expr_for_constant (t);
1176 else
1177 {
1178 /* More complex expressions can result from SCCVN expression
1179 simplification that inserts values for them. As they all
1180 do not have VOPs the get handled by the nary ops struct. */
1181 vn_nary_op_t result;
1182 unsigned int result_id;
1183 vn_nary_op_lookup (t, &result);
1184 if (result != NULL)
1185 {
1186 pre_expr e = pre_expr_pool.allocate ();
1187 e->kind = NARY;
1188 PRE_EXPR_NARY (e) = result;
1189 result_id = lookup_expression_id (e);
1190 if (result_id != 0)
1191 {
1192 pre_expr_pool.remove (e);
1193 e = expression_for_id (result_id);
1194 return e;
1195 }
1196 alloc_expression_id (e);
1197 return e;
1198 }
1199 }
1200 return NULL;
1201 }
1202
1203 /* Return the folded version of T if T, when folded, is a gimple
1204 min_invariant or an SSA name. Otherwise, return T. */
1205
1206 static pre_expr
1207 fully_constant_expression (pre_expr e)
1208 {
1209 switch (e->kind)
1210 {
1211 case CONSTANT:
1212 return e;
1213 case NARY:
1214 {
1215 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1216 tree res = vn_nary_simplify (nary);
1217 if (!res)
1218 return e;
1219 if (is_gimple_min_invariant (res))
1220 return get_or_alloc_expr_for_constant (res);
1221 if (TREE_CODE (res) == SSA_NAME)
1222 return get_or_alloc_expr_for_name (res);
1223 return e;
1224 }
1225 case REFERENCE:
1226 {
1227 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1228 tree folded;
1229 if ((folded = fully_constant_vn_reference_p (ref)))
1230 return get_or_alloc_expr_for_constant (folded);
1231 return e;
1232 }
1233 default:
1234 return e;
1235 }
1236 return e;
1237 }
1238
1239 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1240 it has the value it would have in BLOCK. Set *SAME_VALID to true
1241 in case the new vuse doesn't change the value id of the OPERANDS. */
1242
1243 static tree
1244 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1245 alias_set_type set, tree type, tree vuse,
1246 basic_block phiblock,
1247 basic_block block, bool *same_valid)
1248 {
1249 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1250 ao_ref ref;
1251 edge e = NULL;
1252 bool use_oracle;
1253
1254 *same_valid = true;
1255
1256 if (gimple_bb (phi) != phiblock)
1257 return vuse;
1258
1259 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1260
1261 /* Use the alias-oracle to find either the PHI node in this block,
1262 the first VUSE used in this block that is equivalent to vuse or
1263 the first VUSE which definition in this block kills the value. */
1264 if (gimple_code (phi) == GIMPLE_PHI)
1265 e = find_edge (block, phiblock);
1266 else if (use_oracle)
1267 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1268 {
1269 vuse = gimple_vuse (phi);
1270 phi = SSA_NAME_DEF_STMT (vuse);
1271 if (gimple_bb (phi) != phiblock)
1272 return vuse;
1273 if (gimple_code (phi) == GIMPLE_PHI)
1274 {
1275 e = find_edge (block, phiblock);
1276 break;
1277 }
1278 }
1279 else
1280 return NULL_TREE;
1281
1282 if (e)
1283 {
1284 if (use_oracle)
1285 {
1286 bitmap visited = NULL;
1287 unsigned int cnt;
1288 /* Try to find a vuse that dominates this phi node by skipping
1289 non-clobbering statements. */
1290 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1291 NULL, NULL);
1292 if (visited)
1293 BITMAP_FREE (visited);
1294 }
1295 else
1296 vuse = NULL_TREE;
1297 if (!vuse)
1298 {
1299 /* If we didn't find any, the value ID can't stay the same,
1300 but return the translated vuse. */
1301 *same_valid = false;
1302 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1303 }
1304 /* ??? We would like to return vuse here as this is the canonical
1305 upmost vdef that this reference is associated with. But during
1306 insertion of the references into the hash tables we only ever
1307 directly insert with their direct gimple_vuse, hence returning
1308 something else would make us not find the other expression. */
1309 return PHI_ARG_DEF (phi, e->dest_idx);
1310 }
1311
1312 return NULL_TREE;
1313 }
1314
1315 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1316 SET2. This is used to avoid making a set consisting of the union
1317 of PA_IN and ANTIC_IN during insert. */
1318
1319 static inline pre_expr
1320 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1321 {
1322 pre_expr result;
1323
1324 result = bitmap_find_leader (set1, val);
1325 if (!result && set2)
1326 result = bitmap_find_leader (set2, val);
1327 return result;
1328 }
1329
1330 /* Get the tree type for our PRE expression e. */
1331
1332 static tree
1333 get_expr_type (const pre_expr e)
1334 {
1335 switch (e->kind)
1336 {
1337 case NAME:
1338 return TREE_TYPE (PRE_EXPR_NAME (e));
1339 case CONSTANT:
1340 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1341 case REFERENCE:
1342 return PRE_EXPR_REFERENCE (e)->type;
1343 case NARY:
1344 return PRE_EXPR_NARY (e)->type;
1345 }
1346 gcc_unreachable ();
1347 }
1348
1349 /* Get a representative SSA_NAME for a given expression.
1350 Since all of our sub-expressions are treated as values, we require
1351 them to be SSA_NAME's for simplicity.
1352 Prior versions of GVNPRE used to use "value handles" here, so that
1353 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1354 either case, the operands are really values (IE we do not expect
1355 them to be usable without finding leaders). */
1356
1357 static tree
1358 get_representative_for (const pre_expr e)
1359 {
1360 tree name;
1361 unsigned int value_id = get_expr_value_id (e);
1362
1363 switch (e->kind)
1364 {
1365 case NAME:
1366 return VN_INFO (PRE_EXPR_NAME (e))->valnum;
1367 case CONSTANT:
1368 return PRE_EXPR_CONSTANT (e);
1369 case NARY:
1370 case REFERENCE:
1371 {
1372 /* Go through all of the expressions representing this value
1373 and pick out an SSA_NAME. */
1374 unsigned int i;
1375 bitmap_iterator bi;
1376 bitmap exprs = value_expressions[value_id];
1377 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1378 {
1379 pre_expr rep = expression_for_id (i);
1380 if (rep->kind == NAME)
1381 return VN_INFO (PRE_EXPR_NAME (rep))->valnum;
1382 else if (rep->kind == CONSTANT)
1383 return PRE_EXPR_CONSTANT (rep);
1384 }
1385 }
1386 break;
1387 }
1388
1389 /* If we reached here we couldn't find an SSA_NAME. This can
1390 happen when we've discovered a value that has never appeared in
1391 the program as set to an SSA_NAME, as the result of phi translation.
1392 Create one here.
1393 ??? We should be able to re-use this when we insert the statement
1394 to compute it. */
1395 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1396 VN_INFO_GET (name)->value_id = value_id;
1397 VN_INFO (name)->valnum = name;
1398 /* ??? For now mark this SSA name for release by SCCVN. */
1399 VN_INFO (name)->needs_insertion = true;
1400 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1401 if (dump_file && (dump_flags & TDF_DETAILS))
1402 {
1403 fprintf (dump_file, "Created SSA_NAME representative ");
1404 print_generic_expr (dump_file, name, 0);
1405 fprintf (dump_file, " for expression:");
1406 print_pre_expr (dump_file, e);
1407 fprintf (dump_file, " (%04d)\n", value_id);
1408 }
1409
1410 return name;
1411 }
1412
1413
1414
1415 static pre_expr
1416 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1417 basic_block pred, basic_block phiblock);
1418
1419 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1420 the phis in PRED. Return NULL if we can't find a leader for each part
1421 of the translated expression. */
1422
1423 static pre_expr
1424 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1425 basic_block pred, basic_block phiblock)
1426 {
1427 switch (expr->kind)
1428 {
1429 case NARY:
1430 {
1431 unsigned int i;
1432 bool changed = false;
1433 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1434 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1435 sizeof_vn_nary_op (nary->length));
1436 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1437
1438 for (i = 0; i < newnary->length; i++)
1439 {
1440 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1441 continue;
1442 else
1443 {
1444 pre_expr leader, result;
1445 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1446 leader = find_leader_in_sets (op_val_id, set1, set2);
1447 result = phi_translate (leader, set1, set2, pred, phiblock);
1448 if (result && result != leader)
1449 newnary->op[i] = get_representative_for (result);
1450 else if (!result)
1451 return NULL;
1452
1453 changed |= newnary->op[i] != nary->op[i];
1454 }
1455 }
1456 if (changed)
1457 {
1458 pre_expr constant;
1459 unsigned int new_val_id;
1460
1461 PRE_EXPR_NARY (expr) = newnary;
1462 constant = fully_constant_expression (expr);
1463 PRE_EXPR_NARY (expr) = nary;
1464 if (constant != expr)
1465 {
1466 /* For non-CONSTANTs we have to make sure we can eventually
1467 insert the expression. Which means we need to have a
1468 leader for it. */
1469 if (constant->kind != CONSTANT)
1470 {
1471 unsigned value_id = get_expr_value_id (constant);
1472 constant = find_leader_in_sets (value_id, set1, set2);
1473 if (constant)
1474 return constant;
1475 }
1476 else
1477 return constant;
1478 }
1479
1480 tree result = vn_nary_op_lookup_pieces (newnary->length,
1481 newnary->opcode,
1482 newnary->type,
1483 &newnary->op[0],
1484 &nary);
1485 if (result && is_gimple_min_invariant (result))
1486 return get_or_alloc_expr_for_constant (result);
1487
1488 expr = pre_expr_pool.allocate ();
1489 expr->kind = NARY;
1490 expr->id = 0;
1491 if (nary)
1492 {
1493 PRE_EXPR_NARY (expr) = nary;
1494 new_val_id = nary->value_id;
1495 get_or_alloc_expression_id (expr);
1496 }
1497 else
1498 {
1499 new_val_id = get_next_value_id ();
1500 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1501 nary = vn_nary_op_insert_pieces (newnary->length,
1502 newnary->opcode,
1503 newnary->type,
1504 &newnary->op[0],
1505 result, new_val_id);
1506 PRE_EXPR_NARY (expr) = nary;
1507 get_or_alloc_expression_id (expr);
1508 }
1509 add_to_value (new_val_id, expr);
1510 }
1511 return expr;
1512 }
1513 break;
1514
1515 case REFERENCE:
1516 {
1517 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1518 vec<vn_reference_op_s> operands = ref->operands;
1519 tree vuse = ref->vuse;
1520 tree newvuse = vuse;
1521 vec<vn_reference_op_s> newoperands = vNULL;
1522 bool changed = false, same_valid = true;
1523 unsigned int i, n;
1524 vn_reference_op_t operand;
1525 vn_reference_t newref;
1526
1527 for (i = 0; operands.iterate (i, &operand); i++)
1528 {
1529 pre_expr opresult;
1530 pre_expr leader;
1531 tree op[3];
1532 tree type = operand->type;
1533 vn_reference_op_s newop = *operand;
1534 op[0] = operand->op0;
1535 op[1] = operand->op1;
1536 op[2] = operand->op2;
1537 for (n = 0; n < 3; ++n)
1538 {
1539 unsigned int op_val_id;
1540 if (!op[n])
1541 continue;
1542 if (TREE_CODE (op[n]) != SSA_NAME)
1543 {
1544 /* We can't possibly insert these. */
1545 if (n != 0
1546 && !is_gimple_min_invariant (op[n]))
1547 break;
1548 continue;
1549 }
1550 op_val_id = VN_INFO (op[n])->value_id;
1551 leader = find_leader_in_sets (op_val_id, set1, set2);
1552 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1553 if (opresult && opresult != leader)
1554 {
1555 tree name = get_representative_for (opresult);
1556 changed |= name != op[n];
1557 op[n] = name;
1558 }
1559 else if (!opresult)
1560 break;
1561 }
1562 if (n != 3)
1563 {
1564 newoperands.release ();
1565 return NULL;
1566 }
1567 if (!changed)
1568 continue;
1569 if (!newoperands.exists ())
1570 newoperands = operands.copy ();
1571 /* We may have changed from an SSA_NAME to a constant */
1572 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1573 newop.opcode = TREE_CODE (op[0]);
1574 newop.type = type;
1575 newop.op0 = op[0];
1576 newop.op1 = op[1];
1577 newop.op2 = op[2];
1578 newoperands[i] = newop;
1579 }
1580 gcc_checking_assert (i == operands.length ());
1581
1582 if (vuse)
1583 {
1584 newvuse = translate_vuse_through_block (newoperands.exists ()
1585 ? newoperands : operands,
1586 ref->set, ref->type,
1587 vuse, phiblock, pred,
1588 &same_valid);
1589 if (newvuse == NULL_TREE)
1590 {
1591 newoperands.release ();
1592 return NULL;
1593 }
1594 }
1595
1596 if (changed || newvuse != vuse)
1597 {
1598 unsigned int new_val_id;
1599 pre_expr constant;
1600
1601 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1602 ref->type,
1603 newoperands.exists ()
1604 ? newoperands : operands,
1605 &newref, VN_WALK);
1606 if (result)
1607 newoperands.release ();
1608
1609 /* We can always insert constants, so if we have a partial
1610 redundant constant load of another type try to translate it
1611 to a constant of appropriate type. */
1612 if (result && is_gimple_min_invariant (result))
1613 {
1614 tree tem = result;
1615 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1616 {
1617 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1618 if (tem && !is_gimple_min_invariant (tem))
1619 tem = NULL_TREE;
1620 }
1621 if (tem)
1622 return get_or_alloc_expr_for_constant (tem);
1623 }
1624
1625 /* If we'd have to convert things we would need to validate
1626 if we can insert the translated expression. So fail
1627 here for now - we cannot insert an alias with a different
1628 type in the VN tables either, as that would assert. */
1629 if (result
1630 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1631 return NULL;
1632 else if (!result && newref
1633 && !useless_type_conversion_p (ref->type, newref->type))
1634 {
1635 newoperands.release ();
1636 return NULL;
1637 }
1638
1639 expr = pre_expr_pool.allocate ();
1640 expr->kind = REFERENCE;
1641 expr->id = 0;
1642
1643 if (newref)
1644 {
1645 PRE_EXPR_REFERENCE (expr) = newref;
1646 constant = fully_constant_expression (expr);
1647 if (constant != expr)
1648 return constant;
1649
1650 new_val_id = newref->value_id;
1651 get_or_alloc_expression_id (expr);
1652 }
1653 else
1654 {
1655 if (changed || !same_valid)
1656 {
1657 new_val_id = get_next_value_id ();
1658 value_expressions.safe_grow_cleared
1659 (get_max_value_id () + 1);
1660 }
1661 else
1662 new_val_id = ref->value_id;
1663 if (!newoperands.exists ())
1664 newoperands = operands.copy ();
1665 newref = vn_reference_insert_pieces (newvuse, ref->set,
1666 ref->type,
1667 newoperands,
1668 result, new_val_id);
1669 newoperands = vNULL;
1670 PRE_EXPR_REFERENCE (expr) = newref;
1671 constant = fully_constant_expression (expr);
1672 if (constant != expr)
1673 return constant;
1674 get_or_alloc_expression_id (expr);
1675 }
1676 add_to_value (new_val_id, expr);
1677 }
1678 newoperands.release ();
1679 return expr;
1680 }
1681 break;
1682
1683 case NAME:
1684 {
1685 tree name = PRE_EXPR_NAME (expr);
1686 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1687 /* If the SSA name is defined by a PHI node in this block,
1688 translate it. */
1689 if (gimple_code (def_stmt) == GIMPLE_PHI
1690 && gimple_bb (def_stmt) == phiblock)
1691 {
1692 edge e = find_edge (pred, gimple_bb (def_stmt));
1693 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1694
1695 /* Handle constant. */
1696 if (is_gimple_min_invariant (def))
1697 return get_or_alloc_expr_for_constant (def);
1698
1699 return get_or_alloc_expr_for_name (def);
1700 }
1701 /* Otherwise return it unchanged - it will get removed if its
1702 value is not available in PREDs AVAIL_OUT set of expressions
1703 by the subtraction of TMP_GEN. */
1704 return expr;
1705 }
1706
1707 default:
1708 gcc_unreachable ();
1709 }
1710 }
1711
1712 /* Wrapper around phi_translate_1 providing caching functionality. */
1713
1714 static pre_expr
1715 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1716 basic_block pred, basic_block phiblock)
1717 {
1718 expr_pred_trans_t slot = NULL;
1719 pre_expr phitrans;
1720
1721 if (!expr)
1722 return NULL;
1723
1724 /* Constants contain no values that need translation. */
1725 if (expr->kind == CONSTANT)
1726 return expr;
1727
1728 if (value_id_constant_p (get_expr_value_id (expr)))
1729 return expr;
1730
1731 /* Don't add translations of NAMEs as those are cheap to translate. */
1732 if (expr->kind != NAME)
1733 {
1734 if (phi_trans_add (&slot, expr, pred))
1735 return slot->v;
1736 /* Store NULL for the value we want to return in the case of
1737 recursing. */
1738 slot->v = NULL;
1739 }
1740
1741 /* Translate. */
1742 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1743
1744 if (slot)
1745 {
1746 if (phitrans)
1747 slot->v = phitrans;
1748 else
1749 /* Remove failed translations again, they cause insert
1750 iteration to not pick up new opportunities reliably. */
1751 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1752 }
1753
1754 return phitrans;
1755 }
1756
1757
1758 /* For each expression in SET, translate the values through phi nodes
1759 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1760 expressions in DEST. */
1761
1762 static void
1763 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1764 basic_block phiblock)
1765 {
1766 vec<pre_expr> exprs;
1767 pre_expr expr;
1768 int i;
1769
1770 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1771 {
1772 bitmap_set_copy (dest, set);
1773 return;
1774 }
1775
1776 exprs = sorted_array_from_bitmap_set (set);
1777 FOR_EACH_VEC_ELT (exprs, i, expr)
1778 {
1779 pre_expr translated;
1780 translated = phi_translate (expr, set, NULL, pred, phiblock);
1781 if (!translated)
1782 continue;
1783
1784 /* We might end up with multiple expressions from SET being
1785 translated to the same value. In this case we do not want
1786 to retain the NARY or REFERENCE expression but prefer a NAME
1787 which would be the leader. */
1788 if (translated->kind == NAME)
1789 bitmap_value_replace_in_set (dest, translated);
1790 else
1791 bitmap_value_insert_into_set (dest, translated);
1792 }
1793 exprs.release ();
1794 }
1795
1796 /* Find the leader for a value (i.e., the name representing that
1797 value) in a given set, and return it. Return NULL if no leader
1798 is found. */
1799
1800 static pre_expr
1801 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1802 {
1803 if (value_id_constant_p (val))
1804 {
1805 unsigned int i;
1806 bitmap_iterator bi;
1807 bitmap exprset = value_expressions[val];
1808
1809 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1810 {
1811 pre_expr expr = expression_for_id (i);
1812 if (expr->kind == CONSTANT)
1813 return expr;
1814 }
1815 }
1816 if (bitmap_set_contains_value (set, val))
1817 {
1818 /* Rather than walk the entire bitmap of expressions, and see
1819 whether any of them has the value we are looking for, we look
1820 at the reverse mapping, which tells us the set of expressions
1821 that have a given value (IE value->expressions with that
1822 value) and see if any of those expressions are in our set.
1823 The number of expressions per value is usually significantly
1824 less than the number of expressions in the set. In fact, for
1825 large testcases, doing it this way is roughly 5-10x faster
1826 than walking the bitmap.
1827 If this is somehow a significant lose for some cases, we can
1828 choose which set to walk based on which set is smaller. */
1829 unsigned int i;
1830 bitmap_iterator bi;
1831 bitmap exprset = value_expressions[val];
1832
1833 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1834 return expression_for_id (i);
1835 }
1836 return NULL;
1837 }
1838
1839 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1840 BLOCK by seeing if it is not killed in the block. Note that we are
1841 only determining whether there is a store that kills it. Because
1842 of the order in which clean iterates over values, we are guaranteed
1843 that altered operands will have caused us to be eliminated from the
1844 ANTIC_IN set already. */
1845
1846 static bool
1847 value_dies_in_block_x (pre_expr expr, basic_block block)
1848 {
1849 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1850 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1851 gimple *def;
1852 gimple_stmt_iterator gsi;
1853 unsigned id = get_expression_id (expr);
1854 bool res = false;
1855 ao_ref ref;
1856
1857 if (!vuse)
1858 return false;
1859
1860 /* Lookup a previously calculated result. */
1861 if (EXPR_DIES (block)
1862 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1863 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1864
1865 /* A memory expression {e, VUSE} dies in the block if there is a
1866 statement that may clobber e. If, starting statement walk from the
1867 top of the basic block, a statement uses VUSE there can be no kill
1868 inbetween that use and the original statement that loaded {e, VUSE},
1869 so we can stop walking. */
1870 ref.base = NULL_TREE;
1871 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1872 {
1873 tree def_vuse, def_vdef;
1874 def = gsi_stmt (gsi);
1875 def_vuse = gimple_vuse (def);
1876 def_vdef = gimple_vdef (def);
1877
1878 /* Not a memory statement. */
1879 if (!def_vuse)
1880 continue;
1881
1882 /* Not a may-def. */
1883 if (!def_vdef)
1884 {
1885 /* A load with the same VUSE, we're done. */
1886 if (def_vuse == vuse)
1887 break;
1888
1889 continue;
1890 }
1891
1892 /* Init ref only if we really need it. */
1893 if (ref.base == NULL_TREE
1894 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1895 refx->operands))
1896 {
1897 res = true;
1898 break;
1899 }
1900 /* If the statement may clobber expr, it dies. */
1901 if (stmt_may_clobber_ref_p_1 (def, &ref))
1902 {
1903 res = true;
1904 break;
1905 }
1906 }
1907
1908 /* Remember the result. */
1909 if (!EXPR_DIES (block))
1910 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1911 bitmap_set_bit (EXPR_DIES (block), id * 2);
1912 if (res)
1913 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1914
1915 return res;
1916 }
1917
1918
1919 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1920 contains its value-id. */
1921
1922 static bool
1923 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1924 {
1925 if (op && TREE_CODE (op) == SSA_NAME)
1926 {
1927 unsigned int value_id = VN_INFO (op)->value_id;
1928 if (!(bitmap_set_contains_value (set1, value_id)
1929 || (set2 && bitmap_set_contains_value (set2, value_id))))
1930 return false;
1931 }
1932 return true;
1933 }
1934
1935 /* Determine if the expression EXPR is valid in SET1 U SET2.
1936 ONLY SET2 CAN BE NULL.
1937 This means that we have a leader for each part of the expression
1938 (if it consists of values), or the expression is an SSA_NAME.
1939 For loads/calls, we also see if the vuse is killed in this block. */
1940
1941 static bool
1942 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1943 {
1944 switch (expr->kind)
1945 {
1946 case NAME:
1947 /* By construction all NAMEs are available. Non-available
1948 NAMEs are removed by subtracting TMP_GEN from the sets. */
1949 return true;
1950 case NARY:
1951 {
1952 unsigned int i;
1953 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1954 for (i = 0; i < nary->length; i++)
1955 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1956 return false;
1957 return true;
1958 }
1959 break;
1960 case REFERENCE:
1961 {
1962 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1963 vn_reference_op_t vro;
1964 unsigned int i;
1965
1966 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1967 {
1968 if (!op_valid_in_sets (set1, set2, vro->op0)
1969 || !op_valid_in_sets (set1, set2, vro->op1)
1970 || !op_valid_in_sets (set1, set2, vro->op2))
1971 return false;
1972 }
1973 return true;
1974 }
1975 default:
1976 gcc_unreachable ();
1977 }
1978 }
1979
1980 /* Clean the set of expressions that are no longer valid in SET1 or
1981 SET2. This means expressions that are made up of values we have no
1982 leaders for in SET1 or SET2. This version is used for partial
1983 anticipation, which means it is not valid in either ANTIC_IN or
1984 PA_IN. */
1985
1986 static void
1987 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
1988 {
1989 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1990 pre_expr expr;
1991 int i;
1992
1993 FOR_EACH_VEC_ELT (exprs, i, expr)
1994 {
1995 if (!valid_in_sets (set1, set2, expr))
1996 bitmap_remove_from_set (set1, expr);
1997 }
1998 exprs.release ();
1999 }
2000
2001 /* Clean the set of expressions that are no longer valid in SET. This
2002 means expressions that are made up of values we have no leaders for
2003 in SET. */
2004
2005 static void
2006 clean (bitmap_set_t set)
2007 {
2008 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2009 pre_expr expr;
2010 int i;
2011
2012 FOR_EACH_VEC_ELT (exprs, i, expr)
2013 {
2014 if (!valid_in_sets (set, NULL, expr))
2015 bitmap_remove_from_set (set, expr);
2016 }
2017 exprs.release ();
2018 }
2019
2020 /* Clean the set of expressions that are no longer valid in SET because
2021 they are clobbered in BLOCK or because they trap and may not be executed. */
2022
2023 static void
2024 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2025 {
2026 bitmap_iterator bi;
2027 unsigned i;
2028 pre_expr to_remove = NULL;
2029
2030 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2031 {
2032 /* Remove queued expr. */
2033 if (to_remove)
2034 {
2035 bitmap_remove_from_set (set, to_remove);
2036 to_remove = NULL;
2037 }
2038
2039 pre_expr expr = expression_for_id (i);
2040 if (expr->kind == REFERENCE)
2041 {
2042 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2043 if (ref->vuse)
2044 {
2045 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2046 if (!gimple_nop_p (def_stmt)
2047 && ((gimple_bb (def_stmt) != block
2048 && !dominated_by_p (CDI_DOMINATORS,
2049 block, gimple_bb (def_stmt)))
2050 || (gimple_bb (def_stmt) == block
2051 && value_dies_in_block_x (expr, block))))
2052 to_remove = expr;
2053 }
2054 }
2055 else if (expr->kind == NARY)
2056 {
2057 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2058 /* If the NARY may trap make sure the block does not contain
2059 a possible exit point.
2060 ??? This is overly conservative if we translate AVAIL_OUT
2061 as the available expression might be after the exit point. */
2062 if (BB_MAY_NOTRETURN (block)
2063 && vn_nary_may_trap (nary))
2064 to_remove = expr;
2065 }
2066 }
2067
2068 /* Remove queued expr. */
2069 if (to_remove)
2070 bitmap_remove_from_set (set, to_remove);
2071 }
2072
2073 static sbitmap has_abnormal_preds;
2074
2075 /* Compute the ANTIC set for BLOCK.
2076
2077 If succs(BLOCK) > 1 then
2078 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2079 else if succs(BLOCK) == 1 then
2080 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2081
2082 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2083 */
2084
2085 static bool
2086 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2087 {
2088 bool changed = false;
2089 bitmap_set_t S, old, ANTIC_OUT;
2090 bitmap_iterator bi;
2091 unsigned int bii;
2092 edge e;
2093 edge_iterator ei;
2094 bool was_visited = BB_VISITED (block);
2095
2096 old = ANTIC_OUT = S = NULL;
2097 BB_VISITED (block) = 1;
2098
2099 /* If any edges from predecessors are abnormal, antic_in is empty,
2100 so do nothing. */
2101 if (block_has_abnormal_pred_edge)
2102 goto maybe_dump_sets;
2103
2104 old = ANTIC_IN (block);
2105 ANTIC_OUT = bitmap_set_new ();
2106
2107 /* If the block has no successors, ANTIC_OUT is empty. */
2108 if (EDGE_COUNT (block->succs) == 0)
2109 ;
2110 /* If we have one successor, we could have some phi nodes to
2111 translate through. */
2112 else if (single_succ_p (block))
2113 {
2114 basic_block succ_bb = single_succ (block);
2115 gcc_assert (BB_VISITED (succ_bb));
2116 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2117 }
2118 /* If we have multiple successors, we take the intersection of all of
2119 them. Note that in the case of loop exit phi nodes, we may have
2120 phis to translate through. */
2121 else
2122 {
2123 size_t i;
2124 basic_block bprime, first = NULL;
2125
2126 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2127 FOR_EACH_EDGE (e, ei, block->succs)
2128 {
2129 if (!first
2130 && BB_VISITED (e->dest))
2131 first = e->dest;
2132 else if (BB_VISITED (e->dest))
2133 worklist.quick_push (e->dest);
2134 else
2135 {
2136 /* Unvisited successors get their ANTIC_IN replaced by the
2137 maximal set to arrive at a maximum ANTIC_IN solution.
2138 We can ignore them in the intersection operation and thus
2139 need not explicitely represent that maximum solution. */
2140 if (dump_file && (dump_flags & TDF_DETAILS))
2141 fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2142 e->src->index, e->dest->index);
2143 }
2144 }
2145
2146 /* Of multiple successors we have to have visited one already
2147 which is guaranteed by iteration order. */
2148 gcc_assert (first != NULL);
2149
2150 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2151
2152 FOR_EACH_VEC_ELT (worklist, i, bprime)
2153 {
2154 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2155 {
2156 bitmap_set_t tmp = bitmap_set_new ();
2157 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2158 bitmap_set_and (ANTIC_OUT, tmp);
2159 bitmap_set_free (tmp);
2160 }
2161 else
2162 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2163 }
2164 }
2165
2166 /* Prune expressions that are clobbered in block and thus become
2167 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2168 prune_clobbered_mems (ANTIC_OUT, block);
2169
2170 /* Generate ANTIC_OUT - TMP_GEN. */
2171 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2172
2173 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2174 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2175 TMP_GEN (block));
2176
2177 /* Then union in the ANTIC_OUT - TMP_GEN values,
2178 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2179 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2180 bitmap_value_insert_into_set (ANTIC_IN (block),
2181 expression_for_id (bii));
2182
2183 clean (ANTIC_IN (block));
2184
2185 if (!was_visited || !bitmap_set_equal (old, ANTIC_IN (block)))
2186 changed = true;
2187
2188 maybe_dump_sets:
2189 if (dump_file && (dump_flags & TDF_DETAILS))
2190 {
2191 if (ANTIC_OUT)
2192 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2193
2194 if (changed)
2195 fprintf (dump_file, "[changed] ");
2196 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2197 block->index);
2198
2199 if (S)
2200 print_bitmap_set (dump_file, S, "S", block->index);
2201 }
2202 if (old)
2203 bitmap_set_free (old);
2204 if (S)
2205 bitmap_set_free (S);
2206 if (ANTIC_OUT)
2207 bitmap_set_free (ANTIC_OUT);
2208 return changed;
2209 }
2210
2211 /* Compute PARTIAL_ANTIC for BLOCK.
2212
2213 If succs(BLOCK) > 1 then
2214 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2215 in ANTIC_OUT for all succ(BLOCK)
2216 else if succs(BLOCK) == 1 then
2217 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2218
2219 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2220 - ANTIC_IN[BLOCK])
2221
2222 */
2223 static void
2224 compute_partial_antic_aux (basic_block block,
2225 bool block_has_abnormal_pred_edge)
2226 {
2227 bitmap_set_t old_PA_IN;
2228 bitmap_set_t PA_OUT;
2229 edge e;
2230 edge_iterator ei;
2231 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2232
2233 old_PA_IN = PA_OUT = NULL;
2234
2235 /* If any edges from predecessors are abnormal, antic_in is empty,
2236 so do nothing. */
2237 if (block_has_abnormal_pred_edge)
2238 goto maybe_dump_sets;
2239
2240 /* If there are too many partially anticipatable values in the
2241 block, phi_translate_set can take an exponential time: stop
2242 before the translation starts. */
2243 if (max_pa
2244 && single_succ_p (block)
2245 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2246 goto maybe_dump_sets;
2247
2248 old_PA_IN = PA_IN (block);
2249 PA_OUT = bitmap_set_new ();
2250
2251 /* If the block has no successors, ANTIC_OUT is empty. */
2252 if (EDGE_COUNT (block->succs) == 0)
2253 ;
2254 /* If we have one successor, we could have some phi nodes to
2255 translate through. Note that we can't phi translate across DFS
2256 back edges in partial antic, because it uses a union operation on
2257 the successors. For recurrences like IV's, we will end up
2258 generating a new value in the set on each go around (i + 3 (VH.1)
2259 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2260 else if (single_succ_p (block))
2261 {
2262 basic_block succ = single_succ (block);
2263 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2264 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2265 }
2266 /* If we have multiple successors, we take the union of all of
2267 them. */
2268 else
2269 {
2270 size_t i;
2271 basic_block bprime;
2272
2273 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2274 FOR_EACH_EDGE (e, ei, block->succs)
2275 {
2276 if (e->flags & EDGE_DFS_BACK)
2277 continue;
2278 worklist.quick_push (e->dest);
2279 }
2280 if (worklist.length () > 0)
2281 {
2282 FOR_EACH_VEC_ELT (worklist, i, bprime)
2283 {
2284 unsigned int i;
2285 bitmap_iterator bi;
2286
2287 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2288 bitmap_value_insert_into_set (PA_OUT,
2289 expression_for_id (i));
2290 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2291 {
2292 bitmap_set_t pa_in = bitmap_set_new ();
2293 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2294 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2295 bitmap_value_insert_into_set (PA_OUT,
2296 expression_for_id (i));
2297 bitmap_set_free (pa_in);
2298 }
2299 else
2300 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2301 bitmap_value_insert_into_set (PA_OUT,
2302 expression_for_id (i));
2303 }
2304 }
2305 }
2306
2307 /* Prune expressions that are clobbered in block and thus become
2308 invalid if translated from PA_OUT to PA_IN. */
2309 prune_clobbered_mems (PA_OUT, block);
2310
2311 /* PA_IN starts with PA_OUT - TMP_GEN.
2312 Then we subtract things from ANTIC_IN. */
2313 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2314
2315 /* For partial antic, we want to put back in the phi results, since
2316 we will properly avoid making them partially antic over backedges. */
2317 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2318 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2319
2320 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2321 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2322
2323 dependent_clean (PA_IN (block), ANTIC_IN (block));
2324
2325 maybe_dump_sets:
2326 if (dump_file && (dump_flags & TDF_DETAILS))
2327 {
2328 if (PA_OUT)
2329 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2330
2331 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2332 }
2333 if (old_PA_IN)
2334 bitmap_set_free (old_PA_IN);
2335 if (PA_OUT)
2336 bitmap_set_free (PA_OUT);
2337 }
2338
2339 /* Compute ANTIC and partial ANTIC sets. */
2340
2341 static void
2342 compute_antic (void)
2343 {
2344 bool changed = true;
2345 int num_iterations = 0;
2346 basic_block block;
2347 int i;
2348 edge_iterator ei;
2349 edge e;
2350
2351 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2352 We pre-build the map of blocks with incoming abnormal edges here. */
2353 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2354 bitmap_clear (has_abnormal_preds);
2355
2356 FOR_ALL_BB_FN (block, cfun)
2357 {
2358 BB_VISITED (block) = 0;
2359
2360 FOR_EACH_EDGE (e, ei, block->preds)
2361 if (e->flags & EDGE_ABNORMAL)
2362 {
2363 bitmap_set_bit (has_abnormal_preds, block->index);
2364
2365 /* We also anticipate nothing. */
2366 BB_VISITED (block) = 1;
2367 break;
2368 }
2369
2370 /* While we are here, give empty ANTIC_IN sets to each block. */
2371 ANTIC_IN (block) = bitmap_set_new ();
2372 if (do_partial_partial)
2373 PA_IN (block) = bitmap_set_new ();
2374 }
2375
2376 /* At the exit block we anticipate nothing. */
2377 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2378
2379 /* For ANTIC computation we need a postorder that also guarantees that
2380 a block with a single successor is visited after its successor.
2381 RPO on the inverted CFG has this property. */
2382 int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2383 int postorder_num = inverted_post_order_compute (postorder);
2384
2385 auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
2386 bitmap_ones (worklist);
2387 while (changed)
2388 {
2389 if (dump_file && (dump_flags & TDF_DETAILS))
2390 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2391 /* ??? We need to clear our PHI translation cache here as the
2392 ANTIC sets shrink and we restrict valid translations to
2393 those having operands with leaders in ANTIC. Same below
2394 for PA ANTIC computation. */
2395 num_iterations++;
2396 changed = false;
2397 for (i = postorder_num - 1; i >= 0; i--)
2398 {
2399 if (bitmap_bit_p (worklist, postorder[i]))
2400 {
2401 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2402 bitmap_clear_bit (worklist, block->index);
2403 if (compute_antic_aux (block,
2404 bitmap_bit_p (has_abnormal_preds,
2405 block->index)))
2406 {
2407 FOR_EACH_EDGE (e, ei, block->preds)
2408 bitmap_set_bit (worklist, e->src->index);
2409 changed = true;
2410 }
2411 }
2412 }
2413 /* Theoretically possible, but *highly* unlikely. */
2414 gcc_checking_assert (num_iterations < 500);
2415 }
2416
2417 statistics_histogram_event (cfun, "compute_antic iterations",
2418 num_iterations);
2419
2420 if (do_partial_partial)
2421 {
2422 /* For partial antic we ignore backedges and thus we do not need
2423 to perform any iteration when we process blocks in postorder. */
2424 postorder_num = pre_and_rev_post_order_compute (NULL, postorder, false);
2425 for (i = postorder_num - 1 ; i >= 0; i--)
2426 {
2427 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2428 compute_partial_antic_aux (block,
2429 bitmap_bit_p (has_abnormal_preds,
2430 block->index));
2431 }
2432 }
2433
2434 sbitmap_free (has_abnormal_preds);
2435 free (postorder);
2436 }
2437
2438
2439 /* Inserted expressions are placed onto this worklist, which is used
2440 for performing quick dead code elimination of insertions we made
2441 that didn't turn out to be necessary. */
2442 static bitmap inserted_exprs;
2443
2444 /* The actual worker for create_component_ref_by_pieces. */
2445
2446 static tree
2447 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2448 unsigned int *operand, gimple_seq *stmts)
2449 {
2450 vn_reference_op_t currop = &ref->operands[*operand];
2451 tree genop;
2452 ++*operand;
2453 switch (currop->opcode)
2454 {
2455 case CALL_EXPR:
2456 gcc_unreachable ();
2457
2458 case MEM_REF:
2459 {
2460 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2461 stmts);
2462 if (!baseop)
2463 return NULL_TREE;
2464 tree offset = currop->op0;
2465 if (TREE_CODE (baseop) == ADDR_EXPR
2466 && handled_component_p (TREE_OPERAND (baseop, 0)))
2467 {
2468 HOST_WIDE_INT off;
2469 tree base;
2470 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2471 &off);
2472 gcc_assert (base);
2473 offset = int_const_binop (PLUS_EXPR, offset,
2474 build_int_cst (TREE_TYPE (offset),
2475 off));
2476 baseop = build_fold_addr_expr (base);
2477 }
2478 genop = build2 (MEM_REF, currop->type, baseop, offset);
2479 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2480 MR_DEPENDENCE_BASE (genop) = currop->base;
2481 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2482 return genop;
2483 }
2484
2485 case TARGET_MEM_REF:
2486 {
2487 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2488 vn_reference_op_t nextop = &ref->operands[++*operand];
2489 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2490 stmts);
2491 if (!baseop)
2492 return NULL_TREE;
2493 if (currop->op0)
2494 {
2495 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2496 if (!genop0)
2497 return NULL_TREE;
2498 }
2499 if (nextop->op0)
2500 {
2501 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2502 if (!genop1)
2503 return NULL_TREE;
2504 }
2505 genop = build5 (TARGET_MEM_REF, currop->type,
2506 baseop, currop->op2, genop0, currop->op1, genop1);
2507
2508 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2509 MR_DEPENDENCE_BASE (genop) = currop->base;
2510 return genop;
2511 }
2512
2513 case ADDR_EXPR:
2514 if (currop->op0)
2515 {
2516 gcc_assert (is_gimple_min_invariant (currop->op0));
2517 return currop->op0;
2518 }
2519 /* Fallthrough. */
2520 case REALPART_EXPR:
2521 case IMAGPART_EXPR:
2522 case VIEW_CONVERT_EXPR:
2523 {
2524 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2525 stmts);
2526 if (!genop0)
2527 return NULL_TREE;
2528 return fold_build1 (currop->opcode, currop->type, genop0);
2529 }
2530
2531 case WITH_SIZE_EXPR:
2532 {
2533 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2534 stmts);
2535 if (!genop0)
2536 return NULL_TREE;
2537 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2538 if (!genop1)
2539 return NULL_TREE;
2540 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2541 }
2542
2543 case BIT_FIELD_REF:
2544 {
2545 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2546 stmts);
2547 if (!genop0)
2548 return NULL_TREE;
2549 tree op1 = currop->op0;
2550 tree op2 = currop->op1;
2551 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2552 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2553 return fold (t);
2554 }
2555
2556 /* For array ref vn_reference_op's, operand 1 of the array ref
2557 is op0 of the reference op and operand 3 of the array ref is
2558 op1. */
2559 case ARRAY_RANGE_REF:
2560 case ARRAY_REF:
2561 {
2562 tree genop0;
2563 tree genop1 = currop->op0;
2564 tree genop2 = currop->op1;
2565 tree genop3 = currop->op2;
2566 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2567 stmts);
2568 if (!genop0)
2569 return NULL_TREE;
2570 genop1 = find_or_generate_expression (block, genop1, stmts);
2571 if (!genop1)
2572 return NULL_TREE;
2573 if (genop2)
2574 {
2575 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2576 /* Drop zero minimum index if redundant. */
2577 if (integer_zerop (genop2)
2578 && (!domain_type
2579 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2580 genop2 = NULL_TREE;
2581 else
2582 {
2583 genop2 = find_or_generate_expression (block, genop2, stmts);
2584 if (!genop2)
2585 return NULL_TREE;
2586 }
2587 }
2588 if (genop3)
2589 {
2590 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2591 /* We can't always put a size in units of the element alignment
2592 here as the element alignment may be not visible. See
2593 PR43783. Simply drop the element size for constant
2594 sizes. */
2595 if (TREE_CODE (genop3) == INTEGER_CST
2596 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type)) == INTEGER_CST
2597 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type)),
2598 (wi::to_offset (genop3)
2599 * vn_ref_op_align_unit (currop))))
2600 genop3 = NULL_TREE;
2601 else
2602 {
2603 genop3 = find_or_generate_expression (block, genop3, stmts);
2604 if (!genop3)
2605 return NULL_TREE;
2606 }
2607 }
2608 return build4 (currop->opcode, currop->type, genop0, genop1,
2609 genop2, genop3);
2610 }
2611 case COMPONENT_REF:
2612 {
2613 tree op0;
2614 tree op1;
2615 tree genop2 = currop->op1;
2616 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2617 if (!op0)
2618 return NULL_TREE;
2619 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2620 op1 = currop->op0;
2621 if (genop2)
2622 {
2623 genop2 = find_or_generate_expression (block, genop2, stmts);
2624 if (!genop2)
2625 return NULL_TREE;
2626 }
2627 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2628 }
2629
2630 case SSA_NAME:
2631 {
2632 genop = find_or_generate_expression (block, currop->op0, stmts);
2633 return genop;
2634 }
2635 case STRING_CST:
2636 case INTEGER_CST:
2637 case COMPLEX_CST:
2638 case VECTOR_CST:
2639 case REAL_CST:
2640 case CONSTRUCTOR:
2641 case VAR_DECL:
2642 case PARM_DECL:
2643 case CONST_DECL:
2644 case RESULT_DECL:
2645 case FUNCTION_DECL:
2646 return currop->op0;
2647
2648 default:
2649 gcc_unreachable ();
2650 }
2651 }
2652
2653 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2654 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2655 trying to rename aggregates into ssa form directly, which is a no no.
2656
2657 Thus, this routine doesn't create temporaries, it just builds a
2658 single access expression for the array, calling
2659 find_or_generate_expression to build the innermost pieces.
2660
2661 This function is a subroutine of create_expression_by_pieces, and
2662 should not be called on it's own unless you really know what you
2663 are doing. */
2664
2665 static tree
2666 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2667 gimple_seq *stmts)
2668 {
2669 unsigned int op = 0;
2670 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2671 }
2672
2673 /* Find a simple leader for an expression, or generate one using
2674 create_expression_by_pieces from a NARY expression for the value.
2675 BLOCK is the basic_block we are looking for leaders in.
2676 OP is the tree expression to find a leader for or generate.
2677 Returns the leader or NULL_TREE on failure. */
2678
2679 static tree
2680 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2681 {
2682 pre_expr expr = get_or_alloc_expr_for (op);
2683 unsigned int lookfor = get_expr_value_id (expr);
2684 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2685 if (leader)
2686 {
2687 if (leader->kind == NAME)
2688 return PRE_EXPR_NAME (leader);
2689 else if (leader->kind == CONSTANT)
2690 return PRE_EXPR_CONSTANT (leader);
2691
2692 /* Defer. */
2693 return NULL_TREE;
2694 }
2695
2696 /* It must be a complex expression, so generate it recursively. Note
2697 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2698 where the insert algorithm fails to insert a required expression. */
2699 bitmap exprset = value_expressions[lookfor];
2700 bitmap_iterator bi;
2701 unsigned int i;
2702 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2703 {
2704 pre_expr temp = expression_for_id (i);
2705 /* We cannot insert random REFERENCE expressions at arbitrary
2706 places. We can insert NARYs which eventually re-materializes
2707 its operand values. */
2708 if (temp->kind == NARY)
2709 return create_expression_by_pieces (block, temp, stmts,
2710 get_expr_type (expr));
2711 }
2712
2713 /* Defer. */
2714 return NULL_TREE;
2715 }
2716
2717 #define NECESSARY GF_PLF_1
2718
2719 /* Create an expression in pieces, so that we can handle very complex
2720 expressions that may be ANTIC, but not necessary GIMPLE.
2721 BLOCK is the basic block the expression will be inserted into,
2722 EXPR is the expression to insert (in value form)
2723 STMTS is a statement list to append the necessary insertions into.
2724
2725 This function will die if we hit some value that shouldn't be
2726 ANTIC but is (IE there is no leader for it, or its components).
2727 The function returns NULL_TREE in case a different antic expression
2728 has to be inserted first.
2729 This function may also generate expressions that are themselves
2730 partially or fully redundant. Those that are will be either made
2731 fully redundant during the next iteration of insert (for partially
2732 redundant ones), or eliminated by eliminate (for fully redundant
2733 ones). */
2734
2735 static tree
2736 create_expression_by_pieces (basic_block block, pre_expr expr,
2737 gimple_seq *stmts, tree type)
2738 {
2739 tree name;
2740 tree folded;
2741 gimple_seq forced_stmts = NULL;
2742 unsigned int value_id;
2743 gimple_stmt_iterator gsi;
2744 tree exprtype = type ? type : get_expr_type (expr);
2745 pre_expr nameexpr;
2746 gassign *newstmt;
2747
2748 switch (expr->kind)
2749 {
2750 /* We may hit the NAME/CONSTANT case if we have to convert types
2751 that value numbering saw through. */
2752 case NAME:
2753 folded = PRE_EXPR_NAME (expr);
2754 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2755 return folded;
2756 break;
2757 case CONSTANT:
2758 {
2759 folded = PRE_EXPR_CONSTANT (expr);
2760 tree tem = fold_convert (exprtype, folded);
2761 if (is_gimple_min_invariant (tem))
2762 return tem;
2763 break;
2764 }
2765 case REFERENCE:
2766 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2767 {
2768 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2769 unsigned int operand = 1;
2770 vn_reference_op_t currop = &ref->operands[0];
2771 tree sc = NULL_TREE;
2772 tree fn;
2773 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2774 fn = currop->op0;
2775 else
2776 fn = find_or_generate_expression (block, currop->op0, stmts);
2777 if (!fn)
2778 return NULL_TREE;
2779 if (currop->op1)
2780 {
2781 sc = find_or_generate_expression (block, currop->op1, stmts);
2782 if (!sc)
2783 return NULL_TREE;
2784 }
2785 auto_vec<tree> args (ref->operands.length () - 1);
2786 while (operand < ref->operands.length ())
2787 {
2788 tree arg = create_component_ref_by_pieces_1 (block, ref,
2789 &operand, stmts);
2790 if (!arg)
2791 return NULL_TREE;
2792 args.quick_push (arg);
2793 }
2794 gcall *call
2795 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2796 ? build_fold_addr_expr (fn) : fn), args);
2797 gimple_call_set_with_bounds (call, currop->with_bounds);
2798 if (sc)
2799 gimple_call_set_chain (call, sc);
2800 tree forcedname = make_ssa_name (currop->type);
2801 gimple_call_set_lhs (call, forcedname);
2802 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2803 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2804 folded = forcedname;
2805 }
2806 else
2807 {
2808 folded = create_component_ref_by_pieces (block,
2809 PRE_EXPR_REFERENCE (expr),
2810 stmts);
2811 if (!folded)
2812 return NULL_TREE;
2813 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2814 newstmt = gimple_build_assign (name, folded);
2815 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2816 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2817 folded = name;
2818 }
2819 break;
2820 case NARY:
2821 {
2822 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2823 tree *genop = XALLOCAVEC (tree, nary->length);
2824 unsigned i;
2825 for (i = 0; i < nary->length; ++i)
2826 {
2827 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2828 if (!genop[i])
2829 return NULL_TREE;
2830 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2831 may have conversions stripped. */
2832 if (nary->opcode == POINTER_PLUS_EXPR)
2833 {
2834 if (i == 0)
2835 genop[i] = gimple_convert (&forced_stmts,
2836 nary->type, genop[i]);
2837 else if (i == 1)
2838 genop[i] = gimple_convert (&forced_stmts,
2839 sizetype, genop[i]);
2840 }
2841 else
2842 genop[i] = gimple_convert (&forced_stmts,
2843 TREE_TYPE (nary->op[i]), genop[i]);
2844 }
2845 if (nary->opcode == CONSTRUCTOR)
2846 {
2847 vec<constructor_elt, va_gc> *elts = NULL;
2848 for (i = 0; i < nary->length; ++i)
2849 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2850 folded = build_constructor (nary->type, elts);
2851 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2852 newstmt = gimple_build_assign (name, folded);
2853 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2854 folded = name;
2855 }
2856 else
2857 {
2858 switch (nary->length)
2859 {
2860 case 1:
2861 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2862 genop[0]);
2863 break;
2864 case 2:
2865 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2866 genop[0], genop[1]);
2867 break;
2868 case 3:
2869 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2870 genop[0], genop[1], genop[2]);
2871 break;
2872 default:
2873 gcc_unreachable ();
2874 }
2875 }
2876 }
2877 break;
2878 default:
2879 gcc_unreachable ();
2880 }
2881
2882 folded = gimple_convert (&forced_stmts, exprtype, folded);
2883
2884 /* If there is nothing to insert, return the simplified result. */
2885 if (gimple_seq_empty_p (forced_stmts))
2886 return folded;
2887 /* If we simplified to a constant return it and discard eventually
2888 built stmts. */
2889 if (is_gimple_min_invariant (folded))
2890 {
2891 gimple_seq_discard (forced_stmts);
2892 return folded;
2893 }
2894 /* Likewise if we simplified to sth not queued for insertion. */
2895 bool found = false;
2896 gsi = gsi_last (forced_stmts);
2897 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2898 {
2899 gimple *stmt = gsi_stmt (gsi);
2900 tree forcedname = gimple_get_lhs (stmt);
2901 if (forcedname == folded)
2902 {
2903 found = true;
2904 break;
2905 }
2906 }
2907 if (! found)
2908 {
2909 gimple_seq_discard (forced_stmts);
2910 return folded;
2911 }
2912 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2913
2914 /* If we have any intermediate expressions to the value sets, add them
2915 to the value sets and chain them in the instruction stream. */
2916 if (forced_stmts)
2917 {
2918 gsi = gsi_start (forced_stmts);
2919 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2920 {
2921 gimple *stmt = gsi_stmt (gsi);
2922 tree forcedname = gimple_get_lhs (stmt);
2923 pre_expr nameexpr;
2924
2925 if (forcedname != folded)
2926 {
2927 VN_INFO_GET (forcedname)->valnum = forcedname;
2928 VN_INFO (forcedname)->value_id = get_next_value_id ();
2929 nameexpr = get_or_alloc_expr_for_name (forcedname);
2930 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2931 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2932 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2933 }
2934
2935 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2936 gimple_set_plf (stmt, NECESSARY, false);
2937 }
2938 gimple_seq_add_seq (stmts, forced_stmts);
2939 }
2940
2941 name = folded;
2942
2943 /* Fold the last statement. */
2944 gsi = gsi_last (*stmts);
2945 if (fold_stmt_inplace (&gsi))
2946 update_stmt (gsi_stmt (gsi));
2947
2948 /* Add a value number to the temporary.
2949 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2950 we are creating the expression by pieces, and this particular piece of
2951 the expression may have been represented. There is no harm in replacing
2952 here. */
2953 value_id = get_expr_value_id (expr);
2954 VN_INFO_GET (name)->value_id = value_id;
2955 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2956 if (VN_INFO (name)->valnum == NULL_TREE)
2957 VN_INFO (name)->valnum = name;
2958 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2959 nameexpr = get_or_alloc_expr_for_name (name);
2960 add_to_value (value_id, nameexpr);
2961 if (NEW_SETS (block))
2962 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2963 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2964
2965 pre_stats.insertions++;
2966 if (dump_file && (dump_flags & TDF_DETAILS))
2967 {
2968 fprintf (dump_file, "Inserted ");
2969 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0, 0);
2970 fprintf (dump_file, " in predecessor %d (%04d)\n",
2971 block->index, value_id);
2972 }
2973
2974 return name;
2975 }
2976
2977
2978 /* Insert the to-be-made-available values of expression EXPRNUM for each
2979 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2980 merge the result with a phi node, given the same value number as
2981 NODE. Return true if we have inserted new stuff. */
2982
2983 static bool
2984 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2985 vec<pre_expr> avail)
2986 {
2987 pre_expr expr = expression_for_id (exprnum);
2988 pre_expr newphi;
2989 unsigned int val = get_expr_value_id (expr);
2990 edge pred;
2991 bool insertions = false;
2992 bool nophi = false;
2993 basic_block bprime;
2994 pre_expr eprime;
2995 edge_iterator ei;
2996 tree type = get_expr_type (expr);
2997 tree temp;
2998 gphi *phi;
2999
3000 /* Make sure we aren't creating an induction variable. */
3001 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3002 {
3003 bool firstinsideloop = false;
3004 bool secondinsideloop = false;
3005 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3006 EDGE_PRED (block, 0)->src);
3007 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3008 EDGE_PRED (block, 1)->src);
3009 /* Induction variables only have one edge inside the loop. */
3010 if ((firstinsideloop ^ secondinsideloop)
3011 && expr->kind != REFERENCE)
3012 {
3013 if (dump_file && (dump_flags & TDF_DETAILS))
3014 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3015 nophi = true;
3016 }
3017 }
3018
3019 /* Make the necessary insertions. */
3020 FOR_EACH_EDGE (pred, ei, block->preds)
3021 {
3022 gimple_seq stmts = NULL;
3023 tree builtexpr;
3024 bprime = pred->src;
3025 eprime = avail[pred->dest_idx];
3026 builtexpr = create_expression_by_pieces (bprime, eprime,
3027 &stmts, type);
3028 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3029 if (!gimple_seq_empty_p (stmts))
3030 {
3031 gsi_insert_seq_on_edge (pred, stmts);
3032 insertions = true;
3033 }
3034 if (!builtexpr)
3035 {
3036 /* We cannot insert a PHI node if we failed to insert
3037 on one edge. */
3038 nophi = true;
3039 continue;
3040 }
3041 if (is_gimple_min_invariant (builtexpr))
3042 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3043 else
3044 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3045 }
3046 /* If we didn't want a phi node, and we made insertions, we still have
3047 inserted new stuff, and thus return true. If we didn't want a phi node,
3048 and didn't make insertions, we haven't added anything new, so return
3049 false. */
3050 if (nophi && insertions)
3051 return true;
3052 else if (nophi && !insertions)
3053 return false;
3054
3055 /* Now build a phi for the new variable. */
3056 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3057 phi = create_phi_node (temp, block);
3058
3059 gimple_set_plf (phi, NECESSARY, false);
3060 VN_INFO_GET (temp)->value_id = val;
3061 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3062 if (VN_INFO (temp)->valnum == NULL_TREE)
3063 VN_INFO (temp)->valnum = temp;
3064 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3065 FOR_EACH_EDGE (pred, ei, block->preds)
3066 {
3067 pre_expr ae = avail[pred->dest_idx];
3068 gcc_assert (get_expr_type (ae) == type
3069 || useless_type_conversion_p (type, get_expr_type (ae)));
3070 if (ae->kind == CONSTANT)
3071 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3072 pred, UNKNOWN_LOCATION);
3073 else
3074 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3075 }
3076
3077 newphi = get_or_alloc_expr_for_name (temp);
3078 add_to_value (val, newphi);
3079
3080 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3081 this insertion, since we test for the existence of this value in PHI_GEN
3082 before proceeding with the partial redundancy checks in insert_aux.
3083
3084 The value may exist in AVAIL_OUT, in particular, it could be represented
3085 by the expression we are trying to eliminate, in which case we want the
3086 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3087 inserted there.
3088
3089 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3090 this block, because if it did, it would have existed in our dominator's
3091 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3092 */
3093
3094 bitmap_insert_into_set (PHI_GEN (block), newphi);
3095 bitmap_value_replace_in_set (AVAIL_OUT (block),
3096 newphi);
3097 bitmap_insert_into_set (NEW_SETS (block),
3098 newphi);
3099
3100 /* If we insert a PHI node for a conversion of another PHI node
3101 in the same basic-block try to preserve range information.
3102 This is important so that followup loop passes receive optimal
3103 number of iteration analysis results. See PR61743. */
3104 if (expr->kind == NARY
3105 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3106 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3107 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3108 && INTEGRAL_TYPE_P (type)
3109 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3110 && (TYPE_PRECISION (type)
3111 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3112 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3113 {
3114 wide_int min, max;
3115 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3116 && !wi::neg_p (min, SIGNED)
3117 && !wi::neg_p (max, SIGNED))
3118 /* Just handle extension and sign-changes of all-positive ranges. */
3119 set_range_info (temp,
3120 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3121 wide_int_storage::from (min, TYPE_PRECISION (type),
3122 TYPE_SIGN (type)),
3123 wide_int_storage::from (max, TYPE_PRECISION (type),
3124 TYPE_SIGN (type)));
3125 }
3126
3127 if (dump_file && (dump_flags & TDF_DETAILS))
3128 {
3129 fprintf (dump_file, "Created phi ");
3130 print_gimple_stmt (dump_file, phi, 0, 0);
3131 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3132 }
3133 pre_stats.phis++;
3134 return true;
3135 }
3136
3137
3138
3139 /* Perform insertion of partially redundant or hoistable values.
3140 For BLOCK, do the following:
3141 1. Propagate the NEW_SETS of the dominator into the current block.
3142 If the block has multiple predecessors,
3143 2a. Iterate over the ANTIC expressions for the block to see if
3144 any of them are partially redundant.
3145 2b. If so, insert them into the necessary predecessors to make
3146 the expression fully redundant.
3147 2c. Insert a new PHI merging the values of the predecessors.
3148 2d. Insert the new PHI, and the new expressions, into the
3149 NEW_SETS set.
3150 If the block has multiple successors,
3151 3a. Iterate over the ANTIC values for the block to see if
3152 any of them are good candidates for hoisting.
3153 3b. If so, insert expressions computing the values in BLOCK,
3154 and add the new expressions into the NEW_SETS set.
3155 4. Recursively call ourselves on the dominator children of BLOCK.
3156
3157 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3158 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3159 done in do_hoist_insertion.
3160 */
3161
3162 static bool
3163 do_pre_regular_insertion (basic_block block, basic_block dom)
3164 {
3165 bool new_stuff = false;
3166 vec<pre_expr> exprs;
3167 pre_expr expr;
3168 auto_vec<pre_expr> avail;
3169 int i;
3170
3171 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3172 avail.safe_grow (EDGE_COUNT (block->preds));
3173
3174 FOR_EACH_VEC_ELT (exprs, i, expr)
3175 {
3176 if (expr->kind == NARY
3177 || expr->kind == REFERENCE)
3178 {
3179 unsigned int val;
3180 bool by_some = false;
3181 bool cant_insert = false;
3182 bool all_same = true;
3183 pre_expr first_s = NULL;
3184 edge pred;
3185 basic_block bprime;
3186 pre_expr eprime = NULL;
3187 edge_iterator ei;
3188 pre_expr edoubleprime = NULL;
3189 bool do_insertion = false;
3190
3191 val = get_expr_value_id (expr);
3192 if (bitmap_set_contains_value (PHI_GEN (block), val))
3193 continue;
3194 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3195 {
3196 if (dump_file && (dump_flags & TDF_DETAILS))
3197 {
3198 fprintf (dump_file, "Found fully redundant value: ");
3199 print_pre_expr (dump_file, expr);
3200 fprintf (dump_file, "\n");
3201 }
3202 continue;
3203 }
3204
3205 FOR_EACH_EDGE (pred, ei, block->preds)
3206 {
3207 unsigned int vprime;
3208
3209 /* We should never run insertion for the exit block
3210 and so not come across fake pred edges. */
3211 gcc_assert (!(pred->flags & EDGE_FAKE));
3212 bprime = pred->src;
3213 /* We are looking at ANTIC_OUT of bprime. */
3214 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3215 bprime, block);
3216
3217 /* eprime will generally only be NULL if the
3218 value of the expression, translated
3219 through the PHI for this predecessor, is
3220 undefined. If that is the case, we can't
3221 make the expression fully redundant,
3222 because its value is undefined along a
3223 predecessor path. We can thus break out
3224 early because it doesn't matter what the
3225 rest of the results are. */
3226 if (eprime == NULL)
3227 {
3228 avail[pred->dest_idx] = NULL;
3229 cant_insert = true;
3230 break;
3231 }
3232
3233 vprime = get_expr_value_id (eprime);
3234 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3235 vprime);
3236 if (edoubleprime == NULL)
3237 {
3238 avail[pred->dest_idx] = eprime;
3239 all_same = false;
3240 }
3241 else
3242 {
3243 avail[pred->dest_idx] = edoubleprime;
3244 by_some = true;
3245 /* We want to perform insertions to remove a redundancy on
3246 a path in the CFG we want to optimize for speed. */
3247 if (optimize_edge_for_speed_p (pred))
3248 do_insertion = true;
3249 if (first_s == NULL)
3250 first_s = edoubleprime;
3251 else if (!pre_expr_d::equal (first_s, edoubleprime))
3252 all_same = false;
3253 }
3254 }
3255 /* If we can insert it, it's not the same value
3256 already existing along every predecessor, and
3257 it's defined by some predecessor, it is
3258 partially redundant. */
3259 if (!cant_insert && !all_same && by_some)
3260 {
3261 if (!do_insertion)
3262 {
3263 if (dump_file && (dump_flags & TDF_DETAILS))
3264 {
3265 fprintf (dump_file, "Skipping partial redundancy for "
3266 "expression ");
3267 print_pre_expr (dump_file, expr);
3268 fprintf (dump_file, " (%04d), no redundancy on to be "
3269 "optimized for speed edge\n", val);
3270 }
3271 }
3272 else if (dbg_cnt (treepre_insert))
3273 {
3274 if (dump_file && (dump_flags & TDF_DETAILS))
3275 {
3276 fprintf (dump_file, "Found partial redundancy for "
3277 "expression ");
3278 print_pre_expr (dump_file, expr);
3279 fprintf (dump_file, " (%04d)\n",
3280 get_expr_value_id (expr));
3281 }
3282 if (insert_into_preds_of_block (block,
3283 get_expression_id (expr),
3284 avail))
3285 new_stuff = true;
3286 }
3287 }
3288 /* If all edges produce the same value and that value is
3289 an invariant, then the PHI has the same value on all
3290 edges. Note this. */
3291 else if (!cant_insert && all_same)
3292 {
3293 gcc_assert (edoubleprime->kind == CONSTANT
3294 || edoubleprime->kind == NAME);
3295
3296 tree temp = make_temp_ssa_name (get_expr_type (expr),
3297 NULL, "pretmp");
3298 gassign *assign
3299 = gimple_build_assign (temp,
3300 edoubleprime->kind == CONSTANT ?
3301 PRE_EXPR_CONSTANT (edoubleprime) :
3302 PRE_EXPR_NAME (edoubleprime));
3303 gimple_stmt_iterator gsi = gsi_after_labels (block);
3304 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3305
3306 gimple_set_plf (assign, NECESSARY, false);
3307 VN_INFO_GET (temp)->value_id = val;
3308 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3309 if (VN_INFO (temp)->valnum == NULL_TREE)
3310 VN_INFO (temp)->valnum = temp;
3311 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3312 pre_expr newe = get_or_alloc_expr_for_name (temp);
3313 add_to_value (val, newe);
3314 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3315 bitmap_insert_into_set (NEW_SETS (block), newe);
3316 }
3317 }
3318 }
3319
3320 exprs.release ();
3321 return new_stuff;
3322 }
3323
3324
3325 /* Perform insertion for partially anticipatable expressions. There
3326 is only one case we will perform insertion for these. This case is
3327 if the expression is partially anticipatable, and fully available.
3328 In this case, we know that putting it earlier will enable us to
3329 remove the later computation. */
3330
3331 static bool
3332 do_pre_partial_partial_insertion (basic_block block, basic_block dom)
3333 {
3334 bool new_stuff = false;
3335 vec<pre_expr> exprs;
3336 pre_expr expr;
3337 auto_vec<pre_expr> avail;
3338 int i;
3339
3340 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3341 avail.safe_grow (EDGE_COUNT (block->preds));
3342
3343 FOR_EACH_VEC_ELT (exprs, i, expr)
3344 {
3345 if (expr->kind == NARY
3346 || expr->kind == REFERENCE)
3347 {
3348 unsigned int val;
3349 bool by_all = true;
3350 bool cant_insert = false;
3351 edge pred;
3352 basic_block bprime;
3353 pre_expr eprime = NULL;
3354 edge_iterator ei;
3355
3356 val = get_expr_value_id (expr);
3357 if (bitmap_set_contains_value (PHI_GEN (block), val))
3358 continue;
3359 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3360 continue;
3361
3362 FOR_EACH_EDGE (pred, ei, block->preds)
3363 {
3364 unsigned int vprime;
3365 pre_expr edoubleprime;
3366
3367 /* We should never run insertion for the exit block
3368 and so not come across fake pred edges. */
3369 gcc_assert (!(pred->flags & EDGE_FAKE));
3370 bprime = pred->src;
3371 eprime = phi_translate (expr, ANTIC_IN (block),
3372 PA_IN (block),
3373 bprime, block);
3374
3375 /* eprime will generally only be NULL if the
3376 value of the expression, translated
3377 through the PHI for this predecessor, is
3378 undefined. If that is the case, we can't
3379 make the expression fully redundant,
3380 because its value is undefined along a
3381 predecessor path. We can thus break out
3382 early because it doesn't matter what the
3383 rest of the results are. */
3384 if (eprime == NULL)
3385 {
3386 avail[pred->dest_idx] = NULL;
3387 cant_insert = true;
3388 break;
3389 }
3390
3391 vprime = get_expr_value_id (eprime);
3392 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3393 avail[pred->dest_idx] = edoubleprime;
3394 if (edoubleprime == NULL)
3395 {
3396 by_all = false;
3397 break;
3398 }
3399 }
3400
3401 /* If we can insert it, it's not the same value
3402 already existing along every predecessor, and
3403 it's defined by some predecessor, it is
3404 partially redundant. */
3405 if (!cant_insert && by_all)
3406 {
3407 edge succ;
3408 bool do_insertion = false;
3409
3410 /* Insert only if we can remove a later expression on a path
3411 that we want to optimize for speed.
3412 The phi node that we will be inserting in BLOCK is not free,
3413 and inserting it for the sake of !optimize_for_speed successor
3414 may cause regressions on the speed path. */
3415 FOR_EACH_EDGE (succ, ei, block->succs)
3416 {
3417 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3418 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3419 {
3420 if (optimize_edge_for_speed_p (succ))
3421 do_insertion = true;
3422 }
3423 }
3424
3425 if (!do_insertion)
3426 {
3427 if (dump_file && (dump_flags & TDF_DETAILS))
3428 {
3429 fprintf (dump_file, "Skipping partial partial redundancy "
3430 "for expression ");
3431 print_pre_expr (dump_file, expr);
3432 fprintf (dump_file, " (%04d), not (partially) anticipated "
3433 "on any to be optimized for speed edges\n", val);
3434 }
3435 }
3436 else if (dbg_cnt (treepre_insert))
3437 {
3438 pre_stats.pa_insert++;
3439 if (dump_file && (dump_flags & TDF_DETAILS))
3440 {
3441 fprintf (dump_file, "Found partial partial redundancy "
3442 "for expression ");
3443 print_pre_expr (dump_file, expr);
3444 fprintf (dump_file, " (%04d)\n",
3445 get_expr_value_id (expr));
3446 }
3447 if (insert_into_preds_of_block (block,
3448 get_expression_id (expr),
3449 avail))
3450 new_stuff = true;
3451 }
3452 }
3453 }
3454 }
3455
3456 exprs.release ();
3457 return new_stuff;
3458 }
3459
3460 /* Insert expressions in BLOCK to compute hoistable values up.
3461 Return TRUE if something was inserted, otherwise return FALSE.
3462 The caller has to make sure that BLOCK has at least two successors. */
3463
3464 static bool
3465 do_hoist_insertion (basic_block block)
3466 {
3467 edge e;
3468 edge_iterator ei;
3469 bool new_stuff = false;
3470 unsigned i;
3471 gimple_stmt_iterator last;
3472
3473 /* At least two successors, or else... */
3474 gcc_assert (EDGE_COUNT (block->succs) >= 2);
3475
3476 /* Check that all successors of BLOCK are dominated by block.
3477 We could use dominated_by_p() for this, but actually there is a much
3478 quicker check: any successor that is dominated by BLOCK can't have
3479 more than one predecessor edge. */
3480 FOR_EACH_EDGE (e, ei, block->succs)
3481 if (! single_pred_p (e->dest))
3482 return false;
3483
3484 /* Determine the insertion point. If we cannot safely insert before
3485 the last stmt if we'd have to, bail out. */
3486 last = gsi_last_bb (block);
3487 if (!gsi_end_p (last)
3488 && !is_ctrl_stmt (gsi_stmt (last))
3489 && stmt_ends_bb_p (gsi_stmt (last)))
3490 return false;
3491
3492 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3493 hoistable values. */
3494 bitmap_set hoistable_set;
3495
3496 /* A hoistable value must be in ANTIC_IN(block)
3497 but not in AVAIL_OUT(BLOCK). */
3498 bitmap_initialize (&hoistable_set.values, &grand_bitmap_obstack);
3499 bitmap_and_compl (&hoistable_set.values,
3500 &ANTIC_IN (block)->values, &AVAIL_OUT (block)->values);
3501
3502 /* Short-cut for a common case: hoistable_set is empty. */
3503 if (bitmap_empty_p (&hoistable_set.values))
3504 return false;
3505
3506 /* Compute which of the hoistable values is in AVAIL_OUT of
3507 at least one of the successors of BLOCK. */
3508 bitmap_head availout_in_some;
3509 bitmap_initialize (&availout_in_some, &grand_bitmap_obstack);
3510 FOR_EACH_EDGE (e, ei, block->succs)
3511 /* Do not consider expressions solely because their availability
3512 on loop exits. They'd be ANTIC-IN throughout the whole loop
3513 and thus effectively hoisted across loops by combination of
3514 PRE and hoisting. */
3515 if (! loop_exit_edge_p (block->loop_father, e))
3516 bitmap_ior_and_into (&availout_in_some, &hoistable_set.values,
3517 &AVAIL_OUT (e->dest)->values);
3518 bitmap_clear (&hoistable_set.values);
3519
3520 /* Short-cut for a common case: availout_in_some is empty. */
3521 if (bitmap_empty_p (&availout_in_some))
3522 return false;
3523
3524 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3525 hoistable_set.values = availout_in_some;
3526 hoistable_set.expressions = ANTIC_IN (block)->expressions;
3527
3528 /* Now finally construct the topological-ordered expression set. */
3529 vec<pre_expr> exprs = sorted_array_from_bitmap_set (&hoistable_set);
3530
3531 bitmap_clear (&hoistable_set.values);
3532
3533 /* If there are candidate values for hoisting, insert expressions
3534 strategically to make the hoistable expressions fully redundant. */
3535 pre_expr expr;
3536 FOR_EACH_VEC_ELT (exprs, i, expr)
3537 {
3538 /* While we try to sort expressions topologically above the
3539 sorting doesn't work out perfectly. Catch expressions we
3540 already inserted. */
3541 unsigned int value_id = get_expr_value_id (expr);
3542 if (bitmap_set_contains_value (AVAIL_OUT (block), value_id))
3543 {
3544 if (dump_file && (dump_flags & TDF_DETAILS))
3545 {
3546 fprintf (dump_file,
3547 "Already inserted expression for ");
3548 print_pre_expr (dump_file, expr);
3549 fprintf (dump_file, " (%04d)\n", value_id);
3550 }
3551 continue;
3552 }
3553
3554 /* OK, we should hoist this value. Perform the transformation. */
3555 pre_stats.hoist_insert++;
3556 if (dump_file && (dump_flags & TDF_DETAILS))
3557 {
3558 fprintf (dump_file,
3559 "Inserting expression in block %d for code hoisting: ",
3560 block->index);
3561 print_pre_expr (dump_file, expr);
3562 fprintf (dump_file, " (%04d)\n", value_id);
3563 }
3564
3565 gimple_seq stmts = NULL;
3566 tree res = create_expression_by_pieces (block, expr, &stmts,
3567 get_expr_type (expr));
3568
3569 /* Do not return true if expression creation ultimately
3570 did not insert any statements. */
3571 if (gimple_seq_empty_p (stmts))
3572 res = NULL_TREE;
3573 else
3574 {
3575 if (gsi_end_p (last) || is_ctrl_stmt (gsi_stmt (last)))
3576 gsi_insert_seq_before (&last, stmts, GSI_SAME_STMT);
3577 else
3578 gsi_insert_seq_after (&last, stmts, GSI_NEW_STMT);
3579 }
3580
3581 /* Make sure to not return true if expression creation ultimately
3582 failed but also make sure to insert any stmts produced as they
3583 are tracked in inserted_exprs. */
3584 if (! res)
3585 continue;
3586
3587 new_stuff = true;
3588 }
3589
3590 exprs.release ();
3591
3592 return new_stuff;
3593 }
3594
3595 /* Do a dominator walk on the control flow graph, and insert computations
3596 of values as necessary for PRE and hoisting. */
3597
3598 static bool
3599 insert_aux (basic_block block, bool do_pre, bool do_hoist)
3600 {
3601 basic_block son;
3602 bool new_stuff = false;
3603
3604 if (block)
3605 {
3606 basic_block dom;
3607 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3608 if (dom)
3609 {
3610 unsigned i;
3611 bitmap_iterator bi;
3612 bitmap_set_t newset;
3613
3614 /* First, update the AVAIL_OUT set with anything we may have
3615 inserted higher up in the dominator tree. */
3616 newset = NEW_SETS (dom);
3617 if (newset)
3618 {
3619 /* Note that we need to value_replace both NEW_SETS, and
3620 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3621 represented by some non-simple expression here that we want
3622 to replace it with. */
3623 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3624 {
3625 pre_expr expr = expression_for_id (i);
3626 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3627 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3628 }
3629 }
3630
3631 /* Insert expressions for partial redundancies. */
3632 if (do_pre && !single_pred_p (block))
3633 {
3634 new_stuff |= do_pre_regular_insertion (block, dom);
3635 if (do_partial_partial)
3636 new_stuff |= do_pre_partial_partial_insertion (block, dom);
3637 }
3638
3639 /* Insert expressions for hoisting. */
3640 if (do_hoist && EDGE_COUNT (block->succs) >= 2)
3641 new_stuff |= do_hoist_insertion (block);
3642 }
3643 }
3644 for (son = first_dom_son (CDI_DOMINATORS, block);
3645 son;
3646 son = next_dom_son (CDI_DOMINATORS, son))
3647 {
3648 new_stuff |= insert_aux (son, do_pre, do_hoist);
3649 }
3650
3651 return new_stuff;
3652 }
3653
3654 /* Perform insertion of partially redundant and hoistable values. */
3655
3656 static void
3657 insert (void)
3658 {
3659 bool new_stuff = true;
3660 basic_block bb;
3661 int num_iterations = 0;
3662
3663 FOR_ALL_BB_FN (bb, cfun)
3664 NEW_SETS (bb) = bitmap_set_new ();
3665
3666 while (new_stuff)
3667 {
3668 num_iterations++;
3669 if (dump_file && dump_flags & TDF_DETAILS)
3670 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3671 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun), flag_tree_pre,
3672 flag_code_hoisting);
3673
3674 /* Clear the NEW sets before the next iteration. We have already
3675 fully propagated its contents. */
3676 if (new_stuff)
3677 FOR_ALL_BB_FN (bb, cfun)
3678 bitmap_set_free (NEW_SETS (bb));
3679 }
3680 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3681 }
3682
3683
3684 /* Compute the AVAIL set for all basic blocks.
3685
3686 This function performs value numbering of the statements in each basic
3687 block. The AVAIL sets are built from information we glean while doing
3688 this value numbering, since the AVAIL sets contain only one entry per
3689 value.
3690
3691 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3692 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3693
3694 static void
3695 compute_avail (void)
3696 {
3697
3698 basic_block block, son;
3699 basic_block *worklist;
3700 size_t sp = 0;
3701 unsigned i;
3702 tree name;
3703
3704 /* We pretend that default definitions are defined in the entry block.
3705 This includes function arguments and the static chain decl. */
3706 FOR_EACH_SSA_NAME (i, name, cfun)
3707 {
3708 pre_expr e;
3709 if (!SSA_NAME_IS_DEFAULT_DEF (name)
3710 || has_zero_uses (name)
3711 || virtual_operand_p (name))
3712 continue;
3713
3714 e = get_or_alloc_expr_for_name (name);
3715 add_to_value (get_expr_value_id (e), e);
3716 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3717 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3718 e);
3719 }
3720
3721 if (dump_file && (dump_flags & TDF_DETAILS))
3722 {
3723 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3724 "tmp_gen", ENTRY_BLOCK);
3725 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3726 "avail_out", ENTRY_BLOCK);
3727 }
3728
3729 /* Allocate the worklist. */
3730 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3731
3732 /* Seed the algorithm by putting the dominator children of the entry
3733 block on the worklist. */
3734 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3735 son;
3736 son = next_dom_son (CDI_DOMINATORS, son))
3737 worklist[sp++] = son;
3738
3739 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3740 = ssa_default_def (cfun, gimple_vop (cfun));
3741
3742 /* Loop until the worklist is empty. */
3743 while (sp)
3744 {
3745 gimple *stmt;
3746 basic_block dom;
3747
3748 /* Pick a block from the worklist. */
3749 block = worklist[--sp];
3750
3751 /* Initially, the set of available values in BLOCK is that of
3752 its immediate dominator. */
3753 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3754 if (dom)
3755 {
3756 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3757 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3758 }
3759
3760 /* Generate values for PHI nodes. */
3761 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3762 gsi_next (&gsi))
3763 {
3764 tree result = gimple_phi_result (gsi.phi ());
3765
3766 /* We have no need for virtual phis, as they don't represent
3767 actual computations. */
3768 if (virtual_operand_p (result))
3769 {
3770 BB_LIVE_VOP_ON_EXIT (block) = result;
3771 continue;
3772 }
3773
3774 pre_expr e = get_or_alloc_expr_for_name (result);
3775 add_to_value (get_expr_value_id (e), e);
3776 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3777 bitmap_insert_into_set (PHI_GEN (block), e);
3778 }
3779
3780 BB_MAY_NOTRETURN (block) = 0;
3781
3782 /* Now compute value numbers and populate value sets with all
3783 the expressions computed in BLOCK. */
3784 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3785 gsi_next (&gsi))
3786 {
3787 ssa_op_iter iter;
3788 tree op;
3789
3790 stmt = gsi_stmt (gsi);
3791
3792 /* Cache whether the basic-block has any non-visible side-effect
3793 or control flow.
3794 If this isn't a call or it is the last stmt in the
3795 basic-block then the CFG represents things correctly. */
3796 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3797 {
3798 /* Non-looping const functions always return normally.
3799 Otherwise the call might not return or have side-effects
3800 that forbids hoisting possibly trapping expressions
3801 before it. */
3802 int flags = gimple_call_flags (stmt);
3803 if (!(flags & ECF_CONST)
3804 || (flags & ECF_LOOPING_CONST_OR_PURE))
3805 BB_MAY_NOTRETURN (block) = 1;
3806 }
3807
3808 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3809 {
3810 pre_expr e = get_or_alloc_expr_for_name (op);
3811
3812 add_to_value (get_expr_value_id (e), e);
3813 bitmap_insert_into_set (TMP_GEN (block), e);
3814 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3815 }
3816
3817 if (gimple_vdef (stmt))
3818 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3819
3820 if (gimple_has_side_effects (stmt)
3821 || stmt_could_throw_p (stmt)
3822 || is_gimple_debug (stmt))
3823 continue;
3824
3825 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3826 {
3827 if (ssa_undefined_value_p (op))
3828 continue;
3829 pre_expr e = get_or_alloc_expr_for_name (op);
3830 bitmap_value_insert_into_set (EXP_GEN (block), e);
3831 }
3832
3833 switch (gimple_code (stmt))
3834 {
3835 case GIMPLE_RETURN:
3836 continue;
3837
3838 case GIMPLE_CALL:
3839 {
3840 vn_reference_t ref;
3841 vn_reference_s ref1;
3842 pre_expr result = NULL;
3843
3844 /* We can value number only calls to real functions. */
3845 if (gimple_call_internal_p (stmt))
3846 continue;
3847
3848 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3849 if (!ref)
3850 continue;
3851
3852 /* If the value of the call is not invalidated in
3853 this block until it is computed, add the expression
3854 to EXP_GEN. */
3855 if (!gimple_vuse (stmt)
3856 || gimple_code
3857 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3858 || gimple_bb (SSA_NAME_DEF_STMT
3859 (gimple_vuse (stmt))) != block)
3860 {
3861 result = pre_expr_pool.allocate ();
3862 result->kind = REFERENCE;
3863 result->id = 0;
3864 PRE_EXPR_REFERENCE (result) = ref;
3865
3866 get_or_alloc_expression_id (result);
3867 add_to_value (get_expr_value_id (result), result);
3868 bitmap_value_insert_into_set (EXP_GEN (block), result);
3869 }
3870 continue;
3871 }
3872
3873 case GIMPLE_ASSIGN:
3874 {
3875 pre_expr result = NULL;
3876 switch (vn_get_stmt_kind (stmt))
3877 {
3878 case VN_NARY:
3879 {
3880 enum tree_code code = gimple_assign_rhs_code (stmt);
3881 vn_nary_op_t nary;
3882
3883 /* COND_EXPR and VEC_COND_EXPR are awkward in
3884 that they contain an embedded complex expression.
3885 Don't even try to shove those through PRE. */
3886 if (code == COND_EXPR
3887 || code == VEC_COND_EXPR)
3888 continue;
3889
3890 vn_nary_op_lookup_stmt (stmt, &nary);
3891 if (!nary)
3892 continue;
3893
3894 /* If the NARY traps and there was a preceding
3895 point in the block that might not return avoid
3896 adding the nary to EXP_GEN. */
3897 if (BB_MAY_NOTRETURN (block)
3898 && vn_nary_may_trap (nary))
3899 continue;
3900
3901 result = pre_expr_pool.allocate ();
3902 result->kind = NARY;
3903 result->id = 0;
3904 PRE_EXPR_NARY (result) = nary;
3905 break;
3906 }
3907
3908 case VN_REFERENCE:
3909 {
3910 tree rhs1 = gimple_assign_rhs1 (stmt);
3911 alias_set_type set = get_alias_set (rhs1);
3912 vec<vn_reference_op_s> operands
3913 = vn_reference_operands_for_lookup (rhs1);
3914 vn_reference_t ref;
3915 vn_reference_lookup_pieces (gimple_vuse (stmt), set,
3916 TREE_TYPE (rhs1),
3917 operands, &ref, VN_WALK);
3918 if (!ref)
3919 {
3920 operands.release ();
3921 continue;
3922 }
3923
3924 /* If the value of the reference is not invalidated in
3925 this block until it is computed, add the expression
3926 to EXP_GEN. */
3927 if (gimple_vuse (stmt))
3928 {
3929 gimple *def_stmt;
3930 bool ok = true;
3931 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3932 while (!gimple_nop_p (def_stmt)
3933 && gimple_code (def_stmt) != GIMPLE_PHI
3934 && gimple_bb (def_stmt) == block)
3935 {
3936 if (stmt_may_clobber_ref_p
3937 (def_stmt, gimple_assign_rhs1 (stmt)))
3938 {
3939 ok = false;
3940 break;
3941 }
3942 def_stmt
3943 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3944 }
3945 if (!ok)
3946 {
3947 operands.release ();
3948 continue;
3949 }
3950 }
3951
3952 /* If the load was value-numbered to another
3953 load make sure we do not use its expression
3954 for insertion if it wouldn't be a valid
3955 replacement. */
3956 /* At the momemt we have a testcase
3957 for hoist insertion of aligned vs. misaligned
3958 variants in gcc.dg/torture/pr65270-1.c thus
3959 with just alignment to be considered we can
3960 simply replace the expression in the hashtable
3961 with the most conservative one. */
3962 vn_reference_op_t ref1 = &ref->operands.last ();
3963 while (ref1->opcode != TARGET_MEM_REF
3964 && ref1->opcode != MEM_REF
3965 && ref1 != &ref->operands[0])
3966 --ref1;
3967 vn_reference_op_t ref2 = &operands.last ();
3968 while (ref2->opcode != TARGET_MEM_REF
3969 && ref2->opcode != MEM_REF
3970 && ref2 != &operands[0])
3971 --ref2;
3972 if ((ref1->opcode == TARGET_MEM_REF
3973 || ref1->opcode == MEM_REF)
3974 && (TYPE_ALIGN (ref1->type)
3975 > TYPE_ALIGN (ref2->type)))
3976 ref1->type
3977 = build_aligned_type (ref1->type,
3978 TYPE_ALIGN (ref2->type));
3979 /* TBAA behavior is an obvious part so make sure
3980 that the hashtable one covers this as well
3981 by adjusting the ref alias set and its base. */
3982 if (ref->set == set
3983 || alias_set_subset_of (set, ref->set))
3984 ;
3985 else if (alias_set_subset_of (ref->set, set))
3986 {
3987 ref->set = set;
3988 if (ref1->opcode == MEM_REF)
3989 ref1->op0 = wide_int_to_tree (TREE_TYPE (ref2->op0),
3990 ref1->op0);
3991 else
3992 ref1->op2 = wide_int_to_tree (TREE_TYPE (ref2->op2),
3993 ref1->op2);
3994 }
3995 else
3996 {
3997 ref->set = 0;
3998 if (ref1->opcode == MEM_REF)
3999 ref1->op0 = wide_int_to_tree (ptr_type_node,
4000 ref1->op0);
4001 else
4002 ref1->op2 = wide_int_to_tree (ptr_type_node,
4003 ref1->op2);
4004 }
4005 operands.release ();
4006
4007 result = pre_expr_pool.allocate ();
4008 result->kind = REFERENCE;
4009 result->id = 0;
4010 PRE_EXPR_REFERENCE (result) = ref;
4011 break;
4012 }
4013
4014 default:
4015 continue;
4016 }
4017
4018 get_or_alloc_expression_id (result);
4019 add_to_value (get_expr_value_id (result), result);
4020 bitmap_value_insert_into_set (EXP_GEN (block), result);
4021 continue;
4022 }
4023 default:
4024 break;
4025 }
4026 }
4027
4028 if (dump_file && (dump_flags & TDF_DETAILS))
4029 {
4030 print_bitmap_set (dump_file, EXP_GEN (block),
4031 "exp_gen", block->index);
4032 print_bitmap_set (dump_file, PHI_GEN (block),
4033 "phi_gen", block->index);
4034 print_bitmap_set (dump_file, TMP_GEN (block),
4035 "tmp_gen", block->index);
4036 print_bitmap_set (dump_file, AVAIL_OUT (block),
4037 "avail_out", block->index);
4038 }
4039
4040 /* Put the dominator children of BLOCK on the worklist of blocks
4041 to compute available sets for. */
4042 for (son = first_dom_son (CDI_DOMINATORS, block);
4043 son;
4044 son = next_dom_son (CDI_DOMINATORS, son))
4045 worklist[sp++] = son;
4046 }
4047
4048 free (worklist);
4049 }
4050
4051
4052 /* Local state for the eliminate domwalk. */
4053 static vec<gimple *> el_to_remove;
4054 static vec<gimple *> el_to_fixup;
4055 static unsigned int el_todo;
4056 static vec<tree> el_avail;
4057 static vec<tree> el_avail_stack;
4058
4059 /* Return a leader for OP that is available at the current point of the
4060 eliminate domwalk. */
4061
4062 static tree
4063 eliminate_avail (tree op)
4064 {
4065 tree valnum = VN_INFO (op)->valnum;
4066 if (TREE_CODE (valnum) == SSA_NAME)
4067 {
4068 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
4069 return valnum;
4070 if (el_avail.length () > SSA_NAME_VERSION (valnum))
4071 return el_avail[SSA_NAME_VERSION (valnum)];
4072 }
4073 else if (is_gimple_min_invariant (valnum))
4074 return valnum;
4075 return NULL_TREE;
4076 }
4077
4078 /* At the current point of the eliminate domwalk make OP available. */
4079
4080 static void
4081 eliminate_push_avail (tree op)
4082 {
4083 tree valnum = VN_INFO (op)->valnum;
4084 if (TREE_CODE (valnum) == SSA_NAME)
4085 {
4086 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4087 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4088 tree pushop = op;
4089 if (el_avail[SSA_NAME_VERSION (valnum)])
4090 pushop = el_avail[SSA_NAME_VERSION (valnum)];
4091 el_avail_stack.safe_push (pushop);
4092 el_avail[SSA_NAME_VERSION (valnum)] = op;
4093 }
4094 }
4095
4096 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4097 the leader for the expression if insertion was successful. */
4098
4099 static tree
4100 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4101 {
4102 gimple *stmt = gimple_seq_first_stmt (VN_INFO (val)->expr);
4103 if (!is_gimple_assign (stmt)
4104 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
4105 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
4106 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
4107 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
4108 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
4109 return NULL_TREE;
4110
4111 tree op = gimple_assign_rhs1 (stmt);
4112 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
4113 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4114 op = TREE_OPERAND (op, 0);
4115 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4116 if (!leader)
4117 return NULL_TREE;
4118
4119 gimple_seq stmts = NULL;
4120 tree res;
4121 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4122 res = gimple_build (&stmts, BIT_FIELD_REF,
4123 TREE_TYPE (val), leader,
4124 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
4125 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
4126 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
4127 res = gimple_build (&stmts, BIT_AND_EXPR,
4128 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
4129 else
4130 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
4131 TREE_TYPE (val), leader);
4132 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4133 VN_INFO_GET (res)->valnum = val;
4134
4135 if (TREE_CODE (leader) == SSA_NAME)
4136 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4137
4138 pre_stats.insertions++;
4139 if (dump_file && (dump_flags & TDF_DETAILS))
4140 {
4141 fprintf (dump_file, "Inserted ");
4142 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0, 0);
4143 }
4144
4145 return res;
4146 }
4147
4148 class eliminate_dom_walker : public dom_walker
4149 {
4150 public:
4151 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
4152 : dom_walker (direction), do_pre (do_pre_) {}
4153
4154 virtual edge before_dom_children (basic_block);
4155 virtual void after_dom_children (basic_block);
4156
4157 bool do_pre;
4158 };
4159
4160 /* Perform elimination for the basic-block B during the domwalk. */
4161
4162 edge
4163 eliminate_dom_walker::before_dom_children (basic_block b)
4164 {
4165 /* Mark new bb. */
4166 el_avail_stack.safe_push (NULL_TREE);
4167
4168 /* ??? If we do nothing for unreachable blocks then this will confuse
4169 tailmerging. Eventually we can reduce its reliance on SCCVN now
4170 that we fully copy/constant-propagate (most) things. */
4171
4172 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4173 {
4174 gphi *phi = gsi.phi ();
4175 tree res = PHI_RESULT (phi);
4176
4177 if (virtual_operand_p (res))
4178 {
4179 gsi_next (&gsi);
4180 continue;
4181 }
4182
4183 tree sprime = eliminate_avail (res);
4184 if (sprime
4185 && sprime != res)
4186 {
4187 if (dump_file && (dump_flags & TDF_DETAILS))
4188 {
4189 fprintf (dump_file, "Replaced redundant PHI node defining ");
4190 print_generic_expr (dump_file, res, 0);
4191 fprintf (dump_file, " with ");
4192 print_generic_expr (dump_file, sprime, 0);
4193 fprintf (dump_file, "\n");
4194 }
4195
4196 /* If we inserted this PHI node ourself, it's not an elimination. */
4197 if (inserted_exprs
4198 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4199 pre_stats.phis--;
4200 else
4201 pre_stats.eliminations++;
4202
4203 /* If we will propagate into all uses don't bother to do
4204 anything. */
4205 if (may_propagate_copy (res, sprime))
4206 {
4207 /* Mark the PHI for removal. */
4208 el_to_remove.safe_push (phi);
4209 gsi_next (&gsi);
4210 continue;
4211 }
4212
4213 remove_phi_node (&gsi, false);
4214
4215 if (inserted_exprs
4216 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4217 && TREE_CODE (sprime) == SSA_NAME)
4218 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4219
4220 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4221 sprime = fold_convert (TREE_TYPE (res), sprime);
4222 gimple *stmt = gimple_build_assign (res, sprime);
4223 /* ??? It cannot yet be necessary (DOM walk). */
4224 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4225
4226 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4227 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4228 continue;
4229 }
4230
4231 eliminate_push_avail (res);
4232 gsi_next (&gsi);
4233 }
4234
4235 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
4236 !gsi_end_p (gsi);
4237 gsi_next (&gsi))
4238 {
4239 tree sprime = NULL_TREE;
4240 gimple *stmt = gsi_stmt (gsi);
4241 tree lhs = gimple_get_lhs (stmt);
4242 if (lhs && TREE_CODE (lhs) == SSA_NAME
4243 && !gimple_has_volatile_ops (stmt)
4244 /* See PR43491. Do not replace a global register variable when
4245 it is a the RHS of an assignment. Do replace local register
4246 variables since gcc does not guarantee a local variable will
4247 be allocated in register.
4248 ??? The fix isn't effective here. This should instead
4249 be ensured by not value-numbering them the same but treating
4250 them like volatiles? */
4251 && !(gimple_assign_single_p (stmt)
4252 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4253 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4254 && is_global_var (gimple_assign_rhs1 (stmt)))))
4255 {
4256 sprime = eliminate_avail (lhs);
4257 if (!sprime)
4258 {
4259 /* If there is no existing usable leader but SCCVN thinks
4260 it has an expression it wants to use as replacement,
4261 insert that. */
4262 tree val = VN_INFO (lhs)->valnum;
4263 if (val != VN_TOP
4264 && TREE_CODE (val) == SSA_NAME
4265 && VN_INFO (val)->needs_insertion
4266 && VN_INFO (val)->expr != NULL
4267 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4268 eliminate_push_avail (sprime);
4269 }
4270
4271 /* If this now constitutes a copy duplicate points-to
4272 and range info appropriately. This is especially
4273 important for inserted code. See tree-ssa-copy.c
4274 for similar code. */
4275 if (sprime
4276 && TREE_CODE (sprime) == SSA_NAME)
4277 {
4278 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4279 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4280 && VN_INFO_PTR_INFO (lhs)
4281 && ! VN_INFO_PTR_INFO (sprime))
4282 {
4283 duplicate_ssa_name_ptr_info (sprime,
4284 VN_INFO_PTR_INFO (lhs));
4285 if (b != sprime_b)
4286 mark_ptr_info_alignment_unknown
4287 (SSA_NAME_PTR_INFO (sprime));
4288 }
4289 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
4290 && VN_INFO_RANGE_INFO (lhs)
4291 && ! VN_INFO_RANGE_INFO (sprime)
4292 && b == sprime_b)
4293 duplicate_ssa_name_range_info (sprime,
4294 VN_INFO_RANGE_TYPE (lhs),
4295 VN_INFO_RANGE_INFO (lhs));
4296 }
4297
4298 /* Inhibit the use of an inserted PHI on a loop header when
4299 the address of the memory reference is a simple induction
4300 variable. In other cases the vectorizer won't do anything
4301 anyway (either it's loop invariant or a complicated
4302 expression). */
4303 if (sprime
4304 && TREE_CODE (sprime) == SSA_NAME
4305 && do_pre
4306 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
4307 && loop_outer (b->loop_father)
4308 && has_zero_uses (sprime)
4309 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4310 && gimple_assign_load_p (stmt))
4311 {
4312 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4313 basic_block def_bb = gimple_bb (def_stmt);
4314 if (gimple_code (def_stmt) == GIMPLE_PHI
4315 && def_bb->loop_father->header == def_bb)
4316 {
4317 loop_p loop = def_bb->loop_father;
4318 ssa_op_iter iter;
4319 tree op;
4320 bool found = false;
4321 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4322 {
4323 affine_iv iv;
4324 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4325 if (def_bb
4326 && flow_bb_inside_loop_p (loop, def_bb)
4327 && simple_iv (loop, loop, op, &iv, true))
4328 {
4329 found = true;
4330 break;
4331 }
4332 }
4333 if (found)
4334 {
4335 if (dump_file && (dump_flags & TDF_DETAILS))
4336 {
4337 fprintf (dump_file, "Not replacing ");
4338 print_gimple_expr (dump_file, stmt, 0, 0);
4339 fprintf (dump_file, " with ");
4340 print_generic_expr (dump_file, sprime, 0);
4341 fprintf (dump_file, " which would add a loop"
4342 " carried dependence to loop %d\n",
4343 loop->num);
4344 }
4345 /* Don't keep sprime available. */
4346 sprime = NULL_TREE;
4347 }
4348 }
4349 }
4350
4351 if (sprime)
4352 {
4353 /* If we can propagate the value computed for LHS into
4354 all uses don't bother doing anything with this stmt. */
4355 if (may_propagate_copy (lhs, sprime))
4356 {
4357 /* Mark it for removal. */
4358 el_to_remove.safe_push (stmt);
4359
4360 /* ??? Don't count copy/constant propagations. */
4361 if (gimple_assign_single_p (stmt)
4362 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4363 || gimple_assign_rhs1 (stmt) == sprime))
4364 continue;
4365
4366 if (dump_file && (dump_flags & TDF_DETAILS))
4367 {
4368 fprintf (dump_file, "Replaced ");
4369 print_gimple_expr (dump_file, stmt, 0, 0);
4370 fprintf (dump_file, " with ");
4371 print_generic_expr (dump_file, sprime, 0);
4372 fprintf (dump_file, " in all uses of ");
4373 print_gimple_stmt (dump_file, stmt, 0, 0);
4374 }
4375
4376 pre_stats.eliminations++;
4377 continue;
4378 }
4379
4380 /* If this is an assignment from our leader (which
4381 happens in the case the value-number is a constant)
4382 then there is nothing to do. */
4383 if (gimple_assign_single_p (stmt)
4384 && sprime == gimple_assign_rhs1 (stmt))
4385 continue;
4386
4387 /* Else replace its RHS. */
4388 bool can_make_abnormal_goto
4389 = is_gimple_call (stmt)
4390 && stmt_can_make_abnormal_goto (stmt);
4391
4392 if (dump_file && (dump_flags & TDF_DETAILS))
4393 {
4394 fprintf (dump_file, "Replaced ");
4395 print_gimple_expr (dump_file, stmt, 0, 0);
4396 fprintf (dump_file, " with ");
4397 print_generic_expr (dump_file, sprime, 0);
4398 fprintf (dump_file, " in ");
4399 print_gimple_stmt (dump_file, stmt, 0, 0);
4400 }
4401
4402 if (TREE_CODE (sprime) == SSA_NAME)
4403 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4404 NECESSARY, true);
4405
4406 pre_stats.eliminations++;
4407 gimple *orig_stmt = stmt;
4408 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4409 TREE_TYPE (sprime)))
4410 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4411 tree vdef = gimple_vdef (stmt);
4412 tree vuse = gimple_vuse (stmt);
4413 propagate_tree_value_into_stmt (&gsi, sprime);
4414 stmt = gsi_stmt (gsi);
4415 update_stmt (stmt);
4416 if (vdef != gimple_vdef (stmt))
4417 VN_INFO (vdef)->valnum = vuse;
4418
4419 /* If we removed EH side-effects from the statement, clean
4420 its EH information. */
4421 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4422 {
4423 bitmap_set_bit (need_eh_cleanup,
4424 gimple_bb (stmt)->index);
4425 if (dump_file && (dump_flags & TDF_DETAILS))
4426 fprintf (dump_file, " Removed EH side-effects.\n");
4427 }
4428
4429 /* Likewise for AB side-effects. */
4430 if (can_make_abnormal_goto
4431 && !stmt_can_make_abnormal_goto (stmt))
4432 {
4433 bitmap_set_bit (need_ab_cleanup,
4434 gimple_bb (stmt)->index);
4435 if (dump_file && (dump_flags & TDF_DETAILS))
4436 fprintf (dump_file, " Removed AB side-effects.\n");
4437 }
4438
4439 continue;
4440 }
4441 }
4442
4443 /* If the statement is a scalar store, see if the expression
4444 has the same value number as its rhs. If so, the store is
4445 dead. */
4446 if (gimple_assign_single_p (stmt)
4447 && !gimple_has_volatile_ops (stmt)
4448 && !is_gimple_reg (gimple_assign_lhs (stmt))
4449 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4450 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4451 {
4452 tree val;
4453 tree rhs = gimple_assign_rhs1 (stmt);
4454 vn_reference_t vnresult;
4455 val = vn_reference_lookup (lhs, gimple_vuse (stmt), VN_WALKREWRITE,
4456 &vnresult, false);
4457 if (TREE_CODE (rhs) == SSA_NAME)
4458 rhs = VN_INFO (rhs)->valnum;
4459 if (val
4460 && operand_equal_p (val, rhs, 0))
4461 {
4462 /* We can only remove the later store if the former aliases
4463 at least all accesses the later one does or if the store
4464 was to readonly memory storing the same value. */
4465 alias_set_type set = get_alias_set (lhs);
4466 if (! vnresult
4467 || vnresult->set == set
4468 || alias_set_subset_of (set, vnresult->set))
4469 {
4470 if (dump_file && (dump_flags & TDF_DETAILS))
4471 {
4472 fprintf (dump_file, "Deleted redundant store ");
4473 print_gimple_stmt (dump_file, stmt, 0, 0);
4474 }
4475
4476 /* Queue stmt for removal. */
4477 el_to_remove.safe_push (stmt);
4478 continue;
4479 }
4480 }
4481 }
4482
4483 /* If this is a control statement value numbering left edges
4484 unexecuted on force the condition in a way consistent with
4485 that. */
4486 if (gcond *cond = dyn_cast <gcond *> (stmt))
4487 {
4488 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4489 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4490 {
4491 if (dump_file && (dump_flags & TDF_DETAILS))
4492 {
4493 fprintf (dump_file, "Removing unexecutable edge from ");
4494 print_gimple_stmt (dump_file, stmt, 0, 0);
4495 }
4496 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4497 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4498 gimple_cond_make_true (cond);
4499 else
4500 gimple_cond_make_false (cond);
4501 update_stmt (cond);
4502 el_todo |= TODO_cleanup_cfg;
4503 continue;
4504 }
4505 }
4506
4507 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4508 bool was_noreturn = (is_gimple_call (stmt)
4509 && gimple_call_noreturn_p (stmt));
4510 tree vdef = gimple_vdef (stmt);
4511 tree vuse = gimple_vuse (stmt);
4512
4513 /* If we didn't replace the whole stmt (or propagate the result
4514 into all uses), replace all uses on this stmt with their
4515 leaders. */
4516 use_operand_p use_p;
4517 ssa_op_iter iter;
4518 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4519 {
4520 tree use = USE_FROM_PTR (use_p);
4521 /* ??? The call code above leaves stmt operands un-updated. */
4522 if (TREE_CODE (use) != SSA_NAME)
4523 continue;
4524 tree sprime = eliminate_avail (use);
4525 if (sprime && sprime != use
4526 && may_propagate_copy (use, sprime)
4527 /* We substitute into debug stmts to avoid excessive
4528 debug temporaries created by removed stmts, but we need
4529 to avoid doing so for inserted sprimes as we never want
4530 to create debug temporaries for them. */
4531 && (!inserted_exprs
4532 || TREE_CODE (sprime) != SSA_NAME
4533 || !is_gimple_debug (stmt)
4534 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4535 {
4536 propagate_value (use_p, sprime);
4537 gimple_set_modified (stmt, true);
4538 if (TREE_CODE (sprime) == SSA_NAME
4539 && !is_gimple_debug (stmt))
4540 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4541 NECESSARY, true);
4542 }
4543 }
4544
4545 /* Visit indirect calls and turn them into direct calls if
4546 possible using the devirtualization machinery. */
4547 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4548 {
4549 tree fn = gimple_call_fn (call_stmt);
4550 if (fn
4551 && flag_devirtualize
4552 && virtual_method_call_p (fn))
4553 {
4554 tree otr_type = obj_type_ref_class (fn);
4555 tree instance;
4556 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4557 bool final;
4558
4559 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4560
4561 vec <cgraph_node *>targets
4562 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4563 tree_to_uhwi
4564 (OBJ_TYPE_REF_TOKEN (fn)),
4565 context,
4566 &final);
4567 if (dump_file)
4568 dump_possible_polymorphic_call_targets (dump_file,
4569 obj_type_ref_class (fn),
4570 tree_to_uhwi
4571 (OBJ_TYPE_REF_TOKEN (fn)),
4572 context);
4573 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4574 {
4575 tree fn;
4576 if (targets.length () == 1)
4577 fn = targets[0]->decl;
4578 else
4579 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4580 if (dump_enabled_p ())
4581 {
4582 location_t loc = gimple_location_safe (stmt);
4583 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4584 "converting indirect call to "
4585 "function %s\n",
4586 lang_hooks.decl_printable_name (fn, 2));
4587 }
4588 gimple_call_set_fndecl (call_stmt, fn);
4589 /* If changing the call to __builtin_unreachable
4590 or similar noreturn function, adjust gimple_call_fntype
4591 too. */
4592 if (gimple_call_noreturn_p (call_stmt)
4593 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
4594 && TYPE_ARG_TYPES (TREE_TYPE (fn))
4595 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
4596 == void_type_node))
4597 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
4598 maybe_remove_unused_call_args (cfun, call_stmt);
4599 gimple_set_modified (stmt, true);
4600 }
4601 }
4602 }
4603
4604 if (gimple_modified_p (stmt))
4605 {
4606 /* If a formerly non-invariant ADDR_EXPR is turned into an
4607 invariant one it was on a separate stmt. */
4608 if (gimple_assign_single_p (stmt)
4609 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4610 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4611 gimple *old_stmt = stmt;
4612 if (is_gimple_call (stmt))
4613 {
4614 /* ??? Only fold calls inplace for now, this may create new
4615 SSA names which in turn will confuse free_scc_vn SSA name
4616 release code. */
4617 fold_stmt_inplace (&gsi);
4618 /* When changing a call into a noreturn call, cfg cleanup
4619 is needed to fix up the noreturn call. */
4620 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4621 el_to_fixup.safe_push (stmt);
4622 }
4623 else
4624 {
4625 fold_stmt (&gsi);
4626 stmt = gsi_stmt (gsi);
4627 if ((gimple_code (stmt) == GIMPLE_COND
4628 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4629 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4630 || (gimple_code (stmt) == GIMPLE_SWITCH
4631 && TREE_CODE (gimple_switch_index (
4632 as_a <gswitch *> (stmt)))
4633 == INTEGER_CST))
4634 el_todo |= TODO_cleanup_cfg;
4635 }
4636 /* If we removed EH side-effects from the statement, clean
4637 its EH information. */
4638 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4639 {
4640 bitmap_set_bit (need_eh_cleanup,
4641 gimple_bb (stmt)->index);
4642 if (dump_file && (dump_flags & TDF_DETAILS))
4643 fprintf (dump_file, " Removed EH side-effects.\n");
4644 }
4645 /* Likewise for AB side-effects. */
4646 if (can_make_abnormal_goto
4647 && !stmt_can_make_abnormal_goto (stmt))
4648 {
4649 bitmap_set_bit (need_ab_cleanup,
4650 gimple_bb (stmt)->index);
4651 if (dump_file && (dump_flags & TDF_DETAILS))
4652 fprintf (dump_file, " Removed AB side-effects.\n");
4653 }
4654 update_stmt (stmt);
4655 if (vdef != gimple_vdef (stmt))
4656 VN_INFO (vdef)->valnum = vuse;
4657 }
4658
4659 /* Make new values available - for fully redundant LHS we
4660 continue with the next stmt above and skip this. */
4661 def_operand_p defp;
4662 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4663 eliminate_push_avail (DEF_FROM_PTR (defp));
4664 }
4665
4666 /* Replace destination PHI arguments. */
4667 edge_iterator ei;
4668 edge e;
4669 FOR_EACH_EDGE (e, ei, b->succs)
4670 {
4671 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4672 !gsi_end_p (gsi);
4673 gsi_next (&gsi))
4674 {
4675 gphi *phi = gsi.phi ();
4676 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4677 tree arg = USE_FROM_PTR (use_p);
4678 if (TREE_CODE (arg) != SSA_NAME
4679 || virtual_operand_p (arg))
4680 continue;
4681 tree sprime = eliminate_avail (arg);
4682 if (sprime && may_propagate_copy (arg, sprime))
4683 {
4684 propagate_value (use_p, sprime);
4685 if (TREE_CODE (sprime) == SSA_NAME)
4686 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4687 }
4688 }
4689 }
4690 return NULL;
4691 }
4692
4693 /* Make no longer available leaders no longer available. */
4694
4695 void
4696 eliminate_dom_walker::after_dom_children (basic_block)
4697 {
4698 tree entry;
4699 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4700 {
4701 tree valnum = VN_INFO (entry)->valnum;
4702 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4703 if (old == entry)
4704 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4705 else
4706 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4707 }
4708 }
4709
4710 /* Eliminate fully redundant computations. */
4711
4712 static unsigned int
4713 eliminate (bool do_pre)
4714 {
4715 gimple_stmt_iterator gsi;
4716 gimple *stmt;
4717
4718 need_eh_cleanup = BITMAP_ALLOC (NULL);
4719 need_ab_cleanup = BITMAP_ALLOC (NULL);
4720
4721 el_to_remove.create (0);
4722 el_to_fixup.create (0);
4723 el_todo = 0;
4724 el_avail.create (num_ssa_names);
4725 el_avail_stack.create (0);
4726
4727 eliminate_dom_walker (CDI_DOMINATORS,
4728 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4729
4730 el_avail.release ();
4731 el_avail_stack.release ();
4732
4733 /* We cannot remove stmts during BB walk, especially not release SSA
4734 names there as this confuses the VN machinery. The stmts ending
4735 up in el_to_remove are either stores or simple copies.
4736 Remove stmts in reverse order to make debug stmt creation possible. */
4737 while (!el_to_remove.is_empty ())
4738 {
4739 stmt = el_to_remove.pop ();
4740
4741 if (dump_file && (dump_flags & TDF_DETAILS))
4742 {
4743 fprintf (dump_file, "Removing dead stmt ");
4744 print_gimple_stmt (dump_file, stmt, 0, 0);
4745 }
4746
4747 tree lhs;
4748 if (gimple_code (stmt) == GIMPLE_PHI)
4749 lhs = gimple_phi_result (stmt);
4750 else
4751 lhs = gimple_get_lhs (stmt);
4752
4753 if (inserted_exprs
4754 && TREE_CODE (lhs) == SSA_NAME)
4755 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4756
4757 gsi = gsi_for_stmt (stmt);
4758 if (gimple_code (stmt) == GIMPLE_PHI)
4759 remove_phi_node (&gsi, true);
4760 else
4761 {
4762 basic_block bb = gimple_bb (stmt);
4763 unlink_stmt_vdef (stmt);
4764 if (gsi_remove (&gsi, true))
4765 bitmap_set_bit (need_eh_cleanup, bb->index);
4766 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4767 bitmap_set_bit (need_ab_cleanup, bb->index);
4768 release_defs (stmt);
4769 }
4770
4771 /* Removing a stmt may expose a forwarder block. */
4772 el_todo |= TODO_cleanup_cfg;
4773 }
4774 el_to_remove.release ();
4775
4776 /* Fixup stmts that became noreturn calls. This may require splitting
4777 blocks and thus isn't possible during the dominator walk. Do this
4778 in reverse order so we don't inadvertedly remove a stmt we want to
4779 fixup by visiting a dominating now noreturn call first. */
4780 while (!el_to_fixup.is_empty ())
4781 {
4782 stmt = el_to_fixup.pop ();
4783
4784 if (dump_file && (dump_flags & TDF_DETAILS))
4785 {
4786 fprintf (dump_file, "Fixing up noreturn call ");
4787 print_gimple_stmt (dump_file, stmt, 0, 0);
4788 }
4789
4790 if (fixup_noreturn_call (stmt))
4791 el_todo |= TODO_cleanup_cfg;
4792 }
4793 el_to_fixup.release ();
4794
4795 return el_todo;
4796 }
4797
4798 /* Perform CFG cleanups made necessary by elimination. */
4799
4800 static unsigned
4801 fini_eliminate (void)
4802 {
4803 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4804 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4805
4806 if (do_eh_cleanup)
4807 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4808
4809 if (do_ab_cleanup)
4810 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4811
4812 BITMAP_FREE (need_eh_cleanup);
4813 BITMAP_FREE (need_ab_cleanup);
4814
4815 if (do_eh_cleanup || do_ab_cleanup)
4816 return TODO_cleanup_cfg;
4817 return 0;
4818 }
4819
4820 /* Borrow a bit of tree-ssa-dce.c for the moment.
4821 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4822 this may be a bit faster, and we may want critical edges kept split. */
4823
4824 /* If OP's defining statement has not already been determined to be necessary,
4825 mark that statement necessary. Return the stmt, if it is newly
4826 necessary. */
4827
4828 static inline gimple *
4829 mark_operand_necessary (tree op)
4830 {
4831 gimple *stmt;
4832
4833 gcc_assert (op);
4834
4835 if (TREE_CODE (op) != SSA_NAME)
4836 return NULL;
4837
4838 stmt = SSA_NAME_DEF_STMT (op);
4839 gcc_assert (stmt);
4840
4841 if (gimple_plf (stmt, NECESSARY)
4842 || gimple_nop_p (stmt))
4843 return NULL;
4844
4845 gimple_set_plf (stmt, NECESSARY, true);
4846 return stmt;
4847 }
4848
4849 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4850 to insert PHI nodes sometimes, and because value numbering of casts isn't
4851 perfect, we sometimes end up inserting dead code. This simple DCE-like
4852 pass removes any insertions we made that weren't actually used. */
4853
4854 static void
4855 remove_dead_inserted_code (void)
4856 {
4857 bitmap worklist;
4858 unsigned i;
4859 bitmap_iterator bi;
4860 gimple *t;
4861
4862 worklist = BITMAP_ALLOC (NULL);
4863 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4864 {
4865 t = SSA_NAME_DEF_STMT (ssa_name (i));
4866 if (gimple_plf (t, NECESSARY))
4867 bitmap_set_bit (worklist, i);
4868 }
4869 while (!bitmap_empty_p (worklist))
4870 {
4871 i = bitmap_first_set_bit (worklist);
4872 bitmap_clear_bit (worklist, i);
4873 t = SSA_NAME_DEF_STMT (ssa_name (i));
4874
4875 /* PHI nodes are somewhat special in that each PHI alternative has
4876 data and control dependencies. All the statements feeding the
4877 PHI node's arguments are always necessary. */
4878 if (gimple_code (t) == GIMPLE_PHI)
4879 {
4880 unsigned k;
4881
4882 for (k = 0; k < gimple_phi_num_args (t); k++)
4883 {
4884 tree arg = PHI_ARG_DEF (t, k);
4885 if (TREE_CODE (arg) == SSA_NAME)
4886 {
4887 gimple *n = mark_operand_necessary (arg);
4888 if (n)
4889 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4890 }
4891 }
4892 }
4893 else
4894 {
4895 /* Propagate through the operands. Examine all the USE, VUSE and
4896 VDEF operands in this statement. Mark all the statements
4897 which feed this statement's uses as necessary. */
4898 ssa_op_iter iter;
4899 tree use;
4900
4901 /* The operands of VDEF expressions are also needed as they
4902 represent potential definitions that may reach this
4903 statement (VDEF operands allow us to follow def-def
4904 links). */
4905
4906 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4907 {
4908 gimple *n = mark_operand_necessary (use);
4909 if (n)
4910 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4911 }
4912 }
4913 }
4914
4915 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4916 {
4917 t = SSA_NAME_DEF_STMT (ssa_name (i));
4918 if (!gimple_plf (t, NECESSARY))
4919 {
4920 gimple_stmt_iterator gsi;
4921
4922 if (dump_file && (dump_flags & TDF_DETAILS))
4923 {
4924 fprintf (dump_file, "Removing unnecessary insertion:");
4925 print_gimple_stmt (dump_file, t, 0, 0);
4926 }
4927
4928 gsi = gsi_for_stmt (t);
4929 if (gimple_code (t) == GIMPLE_PHI)
4930 remove_phi_node (&gsi, true);
4931 else
4932 {
4933 gsi_remove (&gsi, true);
4934 release_defs (t);
4935 }
4936 }
4937 }
4938 BITMAP_FREE (worklist);
4939 }
4940
4941
4942 /* Initialize data structures used by PRE. */
4943
4944 static void
4945 init_pre (void)
4946 {
4947 basic_block bb;
4948
4949 next_expression_id = 1;
4950 expressions.create (0);
4951 expressions.safe_push (NULL);
4952 value_expressions.create (get_max_value_id () + 1);
4953 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4954 name_to_id.create (0);
4955
4956 inserted_exprs = BITMAP_ALLOC (NULL);
4957
4958 connect_infinite_loops_to_exit ();
4959 memset (&pre_stats, 0, sizeof (pre_stats));
4960
4961 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4962
4963 calculate_dominance_info (CDI_DOMINATORS);
4964
4965 bitmap_obstack_initialize (&grand_bitmap_obstack);
4966 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4967 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4968 FOR_ALL_BB_FN (bb, cfun)
4969 {
4970 EXP_GEN (bb) = bitmap_set_new ();
4971 PHI_GEN (bb) = bitmap_set_new ();
4972 TMP_GEN (bb) = bitmap_set_new ();
4973 AVAIL_OUT (bb) = bitmap_set_new ();
4974 }
4975 }
4976
4977
4978 /* Deallocate data structures used by PRE. */
4979
4980 static void
4981 fini_pre ()
4982 {
4983 value_expressions.release ();
4984 BITMAP_FREE (inserted_exprs);
4985 bitmap_obstack_release (&grand_bitmap_obstack);
4986 bitmap_set_pool.release ();
4987 pre_expr_pool.release ();
4988 delete phi_translate_table;
4989 phi_translate_table = NULL;
4990 delete expression_to_id;
4991 expression_to_id = NULL;
4992 name_to_id.release ();
4993
4994 free_aux_for_blocks ();
4995 }
4996
4997 namespace {
4998
4999 const pass_data pass_data_pre =
5000 {
5001 GIMPLE_PASS, /* type */
5002 "pre", /* name */
5003 OPTGROUP_NONE, /* optinfo_flags */
5004 TV_TREE_PRE, /* tv_id */
5005 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
5006 pass_pre. */
5007 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
5008 0, /* properties_provided */
5009 PROP_no_crit_edges, /* properties_destroyed */
5010 TODO_rebuild_alias, /* todo_flags_start */
5011 0, /* todo_flags_finish */
5012 };
5013
5014 class pass_pre : public gimple_opt_pass
5015 {
5016 public:
5017 pass_pre (gcc::context *ctxt)
5018 : gimple_opt_pass (pass_data_pre, ctxt)
5019 {}
5020
5021 /* opt_pass methods: */
5022 virtual bool gate (function *)
5023 { return flag_tree_pre != 0 || flag_code_hoisting != 0; }
5024 virtual unsigned int execute (function *);
5025
5026 }; // class pass_pre
5027
5028 unsigned int
5029 pass_pre::execute (function *fun)
5030 {
5031 unsigned int todo = 0;
5032
5033 do_partial_partial =
5034 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
5035
5036 /* This has to happen before SCCVN runs because
5037 loop_optimizer_init may create new phis, etc. */
5038 loop_optimizer_init (LOOPS_NORMAL);
5039
5040 if (!run_scc_vn (VN_WALK))
5041 {
5042 loop_optimizer_finalize ();
5043 return 0;
5044 }
5045
5046 init_pre ();
5047 scev_initialize ();
5048
5049 /* Collect and value number expressions computed in each basic block. */
5050 compute_avail ();
5051
5052 /* Insert can get quite slow on an incredibly large number of basic
5053 blocks due to some quadratic behavior. Until this behavior is
5054 fixed, don't run it when he have an incredibly large number of
5055 bb's. If we aren't going to run insert, there is no point in
5056 computing ANTIC, either, even though it's plenty fast. */
5057 if (n_basic_blocks_for_fn (fun) < 4000)
5058 {
5059 compute_antic ();
5060 insert ();
5061 }
5062
5063 /* Make sure to remove fake edges before committing our inserts.
5064 This makes sure we don't end up with extra critical edges that
5065 we would need to split. */
5066 remove_fake_exit_edges ();
5067 gsi_commit_edge_inserts ();
5068
5069 /* Eliminate folds statements which might (should not...) end up
5070 not keeping virtual operands up-to-date. */
5071 gcc_assert (!need_ssa_update_p (fun));
5072
5073 /* Remove all the redundant expressions. */
5074 todo |= eliminate (true);
5075
5076 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5077 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
5078 statistics_counter_event (fun, "HOIST inserted", pre_stats.hoist_insert);
5079 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
5080 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5081
5082 clear_expression_ids ();
5083 remove_dead_inserted_code ();
5084
5085 scev_finalize ();
5086 fini_pre ();
5087 todo |= fini_eliminate ();
5088 loop_optimizer_finalize ();
5089
5090 /* Restore SSA info before tail-merging as that resets it as well. */
5091 scc_vn_restore_ssa_info ();
5092
5093 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5094 case we can merge the block with the remaining predecessor of the block.
5095 It should either:
5096 - call merge_blocks after each tail merge iteration
5097 - call merge_blocks after all tail merge iterations
5098 - mark TODO_cleanup_cfg when necessary
5099 - share the cfg cleanup with fini_pre. */
5100 todo |= tail_merge_optimize (todo);
5101
5102 free_scc_vn ();
5103
5104 /* Tail merging invalidates the virtual SSA web, together with
5105 cfg-cleanup opportunities exposed by PRE this will wreck the
5106 SSA updating machinery. So make sure to run update-ssa
5107 manually, before eventually scheduling cfg-cleanup as part of
5108 the todo. */
5109 update_ssa (TODO_update_ssa_only_virtuals);
5110
5111 return todo;
5112 }
5113
5114 } // anon namespace
5115
5116 gimple_opt_pass *
5117 make_pass_pre (gcc::context *ctxt)
5118 {
5119 return new pass_pre (ctxt);
5120 }
5121
5122 namespace {
5123
5124 const pass_data pass_data_fre =
5125 {
5126 GIMPLE_PASS, /* type */
5127 "fre", /* name */
5128 OPTGROUP_NONE, /* optinfo_flags */
5129 TV_TREE_FRE, /* tv_id */
5130 ( PROP_cfg | PROP_ssa ), /* properties_required */
5131 0, /* properties_provided */
5132 0, /* properties_destroyed */
5133 0, /* todo_flags_start */
5134 0, /* todo_flags_finish */
5135 };
5136
5137 class pass_fre : public gimple_opt_pass
5138 {
5139 public:
5140 pass_fre (gcc::context *ctxt)
5141 : gimple_opt_pass (pass_data_fre, ctxt)
5142 {}
5143
5144 /* opt_pass methods: */
5145 opt_pass * clone () { return new pass_fre (m_ctxt); }
5146 virtual bool gate (function *) { return flag_tree_fre != 0; }
5147 virtual unsigned int execute (function *);
5148
5149 }; // class pass_fre
5150
5151 unsigned int
5152 pass_fre::execute (function *fun)
5153 {
5154 unsigned int todo = 0;
5155
5156 if (!run_scc_vn (VN_WALKREWRITE))
5157 return 0;
5158
5159 memset (&pre_stats, 0, sizeof (pre_stats));
5160
5161 /* Remove all the redundant expressions. */
5162 todo |= eliminate (false);
5163
5164 todo |= fini_eliminate ();
5165
5166 scc_vn_restore_ssa_info ();
5167 free_scc_vn ();
5168
5169 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5170 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5171
5172 return todo;
5173 }
5174
5175 } // anon namespace
5176
5177 gimple_opt_pass *
5178 make_pass_fre (gcc::context *ctxt)
5179 {
5180 return new pass_fre (ctxt);
5181 }