Eliminate ENTRY_BLOCK_PTR and EXIT_BLOCK_PTR macros
[gcc.git] / gcc / tree-ssa-pre.c
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2013 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 "tm.h"
26 #include "tree.h"
27 #include "basic-block.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "gimple.h"
31 #include "gimplify.h"
32 #include "gimple-iterator.h"
33 #include "gimplify-me.h"
34 #include "gimple-ssa.h"
35 #include "tree-cfg.h"
36 #include "tree-phinodes.h"
37 #include "ssa-iterators.h"
38 #include "stringpool.h"
39 #include "tree-ssanames.h"
40 #include "tree-ssa-loop.h"
41 #include "tree-into-ssa.h"
42 #include "expr.h"
43 #include "tree-dfa.h"
44 #include "tree-ssa.h"
45 #include "hash-table.h"
46 #include "tree-iterator.h"
47 #include "alloc-pool.h"
48 #include "obstack.h"
49 #include "tree-pass.h"
50 #include "flags.h"
51 #include "langhooks.h"
52 #include "cfgloop.h"
53 #include "tree-ssa-sccvn.h"
54 #include "tree-scalar-evolution.h"
55 #include "params.h"
56 #include "dbgcnt.h"
57 #include "domwalk.h"
58 #include "ipa-prop.h"
59 #include "tree-ssa-propagate.h"
60
61 /* TODO:
62
63 1. Avail sets can be shared by making an avail_find_leader that
64 walks up the dominator tree and looks in those avail sets.
65 This might affect code optimality, it's unclear right now.
66 2. Strength reduction can be performed by anticipating expressions
67 we can repair later on.
68 3. We can do back-substitution or smarter value numbering to catch
69 commutative expressions split up over multiple statements.
70 */
71
72 /* For ease of terminology, "expression node" in the below refers to
73 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
74 represent the actual statement containing the expressions we care about,
75 and we cache the value number by putting it in the expression. */
76
77 /* Basic algorithm
78
79 First we walk the statements to generate the AVAIL sets, the
80 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
81 generation of values/expressions by a given block. We use them
82 when computing the ANTIC sets. The AVAIL sets consist of
83 SSA_NAME's that represent values, so we know what values are
84 available in what blocks. AVAIL is a forward dataflow problem. In
85 SSA, values are never killed, so we don't need a kill set, or a
86 fixpoint iteration, in order to calculate the AVAIL sets. In
87 traditional parlance, AVAIL sets tell us the downsafety of the
88 expressions/values.
89
90 Next, we generate the ANTIC sets. These sets represent the
91 anticipatable expressions. ANTIC is a backwards dataflow
92 problem. An expression is anticipatable in a given block if it could
93 be generated in that block. This means that if we had to perform
94 an insertion in that block, of the value of that expression, we
95 could. Calculating the ANTIC sets requires phi translation of
96 expressions, because the flow goes backwards through phis. We must
97 iterate to a fixpoint of the ANTIC sets, because we have a kill
98 set. Even in SSA form, values are not live over the entire
99 function, only from their definition point onwards. So we have to
100 remove values from the ANTIC set once we go past the definition
101 point of the leaders that make them up.
102 compute_antic/compute_antic_aux performs this computation.
103
104 Third, we perform insertions to make partially redundant
105 expressions fully redundant.
106
107 An expression is partially redundant (excluding partial
108 anticipation) if:
109
110 1. It is AVAIL in some, but not all, of the predecessors of a
111 given block.
112 2. It is ANTIC in all the predecessors.
113
114 In order to make it fully redundant, we insert the expression into
115 the predecessors where it is not available, but is ANTIC.
116
117 For the partial anticipation case, we only perform insertion if it
118 is partially anticipated in some block, and fully available in all
119 of the predecessors.
120
121 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
122 performs these steps.
123
124 Fourth, we eliminate fully redundant expressions.
125 This is a simple statement walk that replaces redundant
126 calculations with the now available values. */
127
128 /* Representations of value numbers:
129
130 Value numbers are represented by a representative SSA_NAME. We
131 will create fake SSA_NAME's in situations where we need a
132 representative but do not have one (because it is a complex
133 expression). In order to facilitate storing the value numbers in
134 bitmaps, and keep the number of wasted SSA_NAME's down, we also
135 associate a value_id with each value number, and create full blown
136 ssa_name's only where we actually need them (IE in operands of
137 existing expressions).
138
139 Theoretically you could replace all the value_id's with
140 SSA_NAME_VERSION, but this would allocate a large number of
141 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
142 It would also require an additional indirection at each point we
143 use the value id. */
144
145 /* Representation of expressions on value numbers:
146
147 Expressions consisting of value numbers are represented the same
148 way as our VN internally represents them, with an additional
149 "pre_expr" wrapping around them in order to facilitate storing all
150 of the expressions in the same sets. */
151
152 /* Representation of sets:
153
154 The dataflow sets do not need to be sorted in any particular order
155 for the majority of their lifetime, are simply represented as two
156 bitmaps, one that keeps track of values present in the set, and one
157 that keeps track of expressions present in the set.
158
159 When we need them in topological order, we produce it on demand by
160 transforming the bitmap into an array and sorting it into topo
161 order. */
162
163 /* Type of expression, used to know which member of the PRE_EXPR union
164 is valid. */
165
166 enum pre_expr_kind
167 {
168 NAME,
169 NARY,
170 REFERENCE,
171 CONSTANT
172 };
173
174 typedef union pre_expr_union_d
175 {
176 tree name;
177 tree constant;
178 vn_nary_op_t nary;
179 vn_reference_t reference;
180 } pre_expr_union;
181
182 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
183 {
184 enum pre_expr_kind kind;
185 unsigned int id;
186 pre_expr_union u;
187
188 /* hash_table support. */
189 typedef pre_expr_d value_type;
190 typedef pre_expr_d compare_type;
191 static inline hashval_t hash (const pre_expr_d *);
192 static inline int equal (const pre_expr_d *, const pre_expr_d *);
193 } *pre_expr;
194
195 #define PRE_EXPR_NAME(e) (e)->u.name
196 #define PRE_EXPR_NARY(e) (e)->u.nary
197 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
198 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
199
200 /* Compare E1 and E1 for equality. */
201
202 inline int
203 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
204 {
205 if (e1->kind != e2->kind)
206 return false;
207
208 switch (e1->kind)
209 {
210 case CONSTANT:
211 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
212 PRE_EXPR_CONSTANT (e2));
213 case NAME:
214 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
215 case NARY:
216 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
217 case REFERENCE:
218 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
219 PRE_EXPR_REFERENCE (e2));
220 default:
221 gcc_unreachable ();
222 }
223 }
224
225 /* Hash E. */
226
227 inline hashval_t
228 pre_expr_d::hash (const value_type *e)
229 {
230 switch (e->kind)
231 {
232 case CONSTANT:
233 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
234 case NAME:
235 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
236 case NARY:
237 return PRE_EXPR_NARY (e)->hashcode;
238 case REFERENCE:
239 return PRE_EXPR_REFERENCE (e)->hashcode;
240 default:
241 gcc_unreachable ();
242 }
243 }
244
245 /* Next global expression id number. */
246 static unsigned int next_expression_id;
247
248 /* Mapping from expression to id number we can use in bitmap sets. */
249 static vec<pre_expr> expressions;
250 static hash_table <pre_expr_d> expression_to_id;
251 static vec<unsigned> name_to_id;
252
253 /* Allocate an expression id for EXPR. */
254
255 static inline unsigned int
256 alloc_expression_id (pre_expr expr)
257 {
258 struct pre_expr_d **slot;
259 /* Make sure we won't overflow. */
260 gcc_assert (next_expression_id + 1 > next_expression_id);
261 expr->id = next_expression_id++;
262 expressions.safe_push (expr);
263 if (expr->kind == NAME)
264 {
265 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
266 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
267 re-allocations by using vec::reserve upfront. There is no
268 vec::quick_grow_cleared unfortunately. */
269 unsigned old_len = name_to_id.length ();
270 name_to_id.reserve (num_ssa_names - old_len);
271 name_to_id.safe_grow_cleared (num_ssa_names);
272 gcc_assert (name_to_id[version] == 0);
273 name_to_id[version] = expr->id;
274 }
275 else
276 {
277 slot = expression_to_id.find_slot (expr, INSERT);
278 gcc_assert (!*slot);
279 *slot = expr;
280 }
281 return next_expression_id - 1;
282 }
283
284 /* Return the expression id for tree EXPR. */
285
286 static inline unsigned int
287 get_expression_id (const pre_expr expr)
288 {
289 return expr->id;
290 }
291
292 static inline unsigned int
293 lookup_expression_id (const pre_expr expr)
294 {
295 struct pre_expr_d **slot;
296
297 if (expr->kind == NAME)
298 {
299 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
300 if (name_to_id.length () <= version)
301 return 0;
302 return name_to_id[version];
303 }
304 else
305 {
306 slot = expression_to_id.find_slot (expr, NO_INSERT);
307 if (!slot)
308 return 0;
309 return ((pre_expr)*slot)->id;
310 }
311 }
312
313 /* Return the existing expression id for EXPR, or create one if one
314 does not exist yet. */
315
316 static inline unsigned int
317 get_or_alloc_expression_id (pre_expr expr)
318 {
319 unsigned int id = lookup_expression_id (expr);
320 if (id == 0)
321 return alloc_expression_id (expr);
322 return expr->id = id;
323 }
324
325 /* Return the expression that has expression id ID */
326
327 static inline pre_expr
328 expression_for_id (unsigned int id)
329 {
330 return expressions[id];
331 }
332
333 /* Free the expression id field in all of our expressions,
334 and then destroy the expressions array. */
335
336 static void
337 clear_expression_ids (void)
338 {
339 expressions.release ();
340 }
341
342 static alloc_pool pre_expr_pool;
343
344 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
345
346 static pre_expr
347 get_or_alloc_expr_for_name (tree name)
348 {
349 struct pre_expr_d expr;
350 pre_expr result;
351 unsigned int result_id;
352
353 expr.kind = NAME;
354 expr.id = 0;
355 PRE_EXPR_NAME (&expr) = name;
356 result_id = lookup_expression_id (&expr);
357 if (result_id != 0)
358 return expression_for_id (result_id);
359
360 result = (pre_expr) pool_alloc (pre_expr_pool);
361 result->kind = NAME;
362 PRE_EXPR_NAME (result) = name;
363 alloc_expression_id (result);
364 return result;
365 }
366
367 /* An unordered bitmap set. One bitmap tracks values, the other,
368 expressions. */
369 typedef struct bitmap_set
370 {
371 bitmap_head expressions;
372 bitmap_head values;
373 } *bitmap_set_t;
374
375 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
376 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
377
378 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
379 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
380
381 /* Mapping from value id to expressions with that value_id. */
382 static vec<bitmap> value_expressions;
383
384 /* Sets that we need to keep track of. */
385 typedef struct bb_bitmap_sets
386 {
387 /* The EXP_GEN set, which represents expressions/values generated in
388 a basic block. */
389 bitmap_set_t exp_gen;
390
391 /* The PHI_GEN set, which represents PHI results generated in a
392 basic block. */
393 bitmap_set_t phi_gen;
394
395 /* The TMP_GEN set, which represents results/temporaries generated
396 in a basic block. IE the LHS of an expression. */
397 bitmap_set_t tmp_gen;
398
399 /* The AVAIL_OUT set, which represents which values are available in
400 a given basic block. */
401 bitmap_set_t avail_out;
402
403 /* The ANTIC_IN set, which represents which values are anticipatable
404 in a given basic block. */
405 bitmap_set_t antic_in;
406
407 /* The PA_IN set, which represents which values are
408 partially anticipatable in a given basic block. */
409 bitmap_set_t pa_in;
410
411 /* The NEW_SETS set, which is used during insertion to augment the
412 AVAIL_OUT set of blocks with the new insertions performed during
413 the current iteration. */
414 bitmap_set_t new_sets;
415
416 /* A cache for value_dies_in_block_x. */
417 bitmap expr_dies;
418
419 /* True if we have visited this block during ANTIC calculation. */
420 unsigned int visited : 1;
421
422 /* True we have deferred processing this block during ANTIC
423 calculation until its successor is processed. */
424 unsigned int deferred : 1;
425
426 /* True when the block contains a call that might not return. */
427 unsigned int contains_may_not_return_call : 1;
428 } *bb_value_sets_t;
429
430 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
431 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
432 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
433 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
434 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
435 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
436 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
437 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
438 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
439 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
440 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
441
442
443 /* Basic block list in postorder. */
444 static int *postorder;
445 static int postorder_num;
446
447 /* This structure is used to keep track of statistics on what
448 optimization PRE was able to perform. */
449 static struct
450 {
451 /* The number of RHS computations eliminated by PRE. */
452 int eliminations;
453
454 /* The number of new expressions/temporaries generated by PRE. */
455 int insertions;
456
457 /* The number of inserts found due to partial anticipation */
458 int pa_insert;
459
460 /* The number of new PHI nodes added by PRE. */
461 int phis;
462 } pre_stats;
463
464 static bool do_partial_partial;
465 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
466 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
467 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
468 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
469 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
470 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
471 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
472 unsigned int, bool);
473 static bitmap_set_t bitmap_set_new (void);
474 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
475 tree);
476 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
477 static unsigned int get_expr_value_id (pre_expr);
478
479 /* We can add and remove elements and entries to and from sets
480 and hash tables, so we use alloc pools for them. */
481
482 static alloc_pool bitmap_set_pool;
483 static bitmap_obstack grand_bitmap_obstack;
484
485 /* Set of blocks with statements that have had their EH properties changed. */
486 static bitmap need_eh_cleanup;
487
488 /* Set of blocks with statements that have had their AB properties changed. */
489 static bitmap need_ab_cleanup;
490
491 /* A three tuple {e, pred, v} used to cache phi translations in the
492 phi_translate_table. */
493
494 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
495 {
496 /* The expression. */
497 pre_expr e;
498
499 /* The predecessor block along which we translated the expression. */
500 basic_block pred;
501
502 /* The value that resulted from the translation. */
503 pre_expr v;
504
505 /* The hashcode for the expression, pred pair. This is cached for
506 speed reasons. */
507 hashval_t hashcode;
508
509 /* hash_table support. */
510 typedef expr_pred_trans_d value_type;
511 typedef expr_pred_trans_d compare_type;
512 static inline hashval_t hash (const value_type *);
513 static inline int equal (const value_type *, const compare_type *);
514 } *expr_pred_trans_t;
515 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
516
517 inline hashval_t
518 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
519 {
520 return e->hashcode;
521 }
522
523 inline int
524 expr_pred_trans_d::equal (const value_type *ve1,
525 const compare_type *ve2)
526 {
527 basic_block b1 = ve1->pred;
528 basic_block b2 = ve2->pred;
529
530 /* If they are not translations for the same basic block, they can't
531 be equal. */
532 if (b1 != b2)
533 return false;
534 return pre_expr_d::equal (ve1->e, ve2->e);
535 }
536
537 /* The phi_translate_table caches phi translations for a given
538 expression and predecessor. */
539 static hash_table <expr_pred_trans_d> phi_translate_table;
540
541 /* Add the tuple mapping from {expression E, basic block PRED} to
542 the phi translation table and return whether it pre-existed. */
543
544 static inline bool
545 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
546 {
547 expr_pred_trans_t *slot;
548 expr_pred_trans_d tem;
549 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
550 pred->index);
551 tem.e = e;
552 tem.pred = pred;
553 tem.hashcode = hash;
554 slot = phi_translate_table.find_slot_with_hash (&tem, hash, INSERT);
555 if (*slot)
556 {
557 *entry = *slot;
558 return true;
559 }
560
561 *entry = *slot = XNEW (struct expr_pred_trans_d);
562 (*entry)->e = e;
563 (*entry)->pred = pred;
564 (*entry)->hashcode = hash;
565 return false;
566 }
567
568
569 /* Add expression E to the expression set of value id V. */
570
571 static void
572 add_to_value (unsigned int v, pre_expr e)
573 {
574 bitmap set;
575
576 gcc_checking_assert (get_expr_value_id (e) == v);
577
578 if (v >= value_expressions.length ())
579 {
580 value_expressions.safe_grow_cleared (v + 1);
581 }
582
583 set = value_expressions[v];
584 if (!set)
585 {
586 set = BITMAP_ALLOC (&grand_bitmap_obstack);
587 value_expressions[v] = set;
588 }
589
590 bitmap_set_bit (set, get_or_alloc_expression_id (e));
591 }
592
593 /* Create a new bitmap set and return it. */
594
595 static bitmap_set_t
596 bitmap_set_new (void)
597 {
598 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
599 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
600 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
601 return ret;
602 }
603
604 /* Return the value id for a PRE expression EXPR. */
605
606 static unsigned int
607 get_expr_value_id (pre_expr expr)
608 {
609 unsigned int id;
610 switch (expr->kind)
611 {
612 case CONSTANT:
613 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
614 break;
615 case NAME:
616 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
617 break;
618 case NARY:
619 id = PRE_EXPR_NARY (expr)->value_id;
620 break;
621 case REFERENCE:
622 id = PRE_EXPR_REFERENCE (expr)->value_id;
623 break;
624 default:
625 gcc_unreachable ();
626 }
627 /* ??? We cannot assert that expr has a value-id (it can be 0), because
628 we assign value-ids only to expressions that have a result
629 in set_hashtable_value_ids. */
630 return id;
631 }
632
633 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
634
635 static tree
636 sccvn_valnum_from_value_id (unsigned int val)
637 {
638 bitmap_iterator bi;
639 unsigned int i;
640 bitmap exprset = value_expressions[val];
641 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
642 {
643 pre_expr vexpr = expression_for_id (i);
644 if (vexpr->kind == NAME)
645 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
646 else if (vexpr->kind == CONSTANT)
647 return PRE_EXPR_CONSTANT (vexpr);
648 }
649 return NULL_TREE;
650 }
651
652 /* Remove an expression EXPR from a bitmapped set. */
653
654 static void
655 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
656 {
657 unsigned int val = get_expr_value_id (expr);
658 if (!value_id_constant_p (val))
659 {
660 bitmap_clear_bit (&set->values, val);
661 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
662 }
663 }
664
665 static void
666 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
667 unsigned int val, bool allow_constants)
668 {
669 if (allow_constants || !value_id_constant_p (val))
670 {
671 /* We specifically expect this and only this function to be able to
672 insert constants into a set. */
673 bitmap_set_bit (&set->values, val);
674 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
675 }
676 }
677
678 /* Insert an expression EXPR into a bitmapped set. */
679
680 static void
681 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
682 {
683 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
684 }
685
686 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
687
688 static void
689 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
690 {
691 bitmap_copy (&dest->expressions, &orig->expressions);
692 bitmap_copy (&dest->values, &orig->values);
693 }
694
695
696 /* Free memory used up by SET. */
697 static void
698 bitmap_set_free (bitmap_set_t set)
699 {
700 bitmap_clear (&set->expressions);
701 bitmap_clear (&set->values);
702 }
703
704
705 /* Generate an topological-ordered array of bitmap set SET. */
706
707 static vec<pre_expr>
708 sorted_array_from_bitmap_set (bitmap_set_t set)
709 {
710 unsigned int i, j;
711 bitmap_iterator bi, bj;
712 vec<pre_expr> result;
713
714 /* Pre-allocate roughly enough space for the array. */
715 result.create (bitmap_count_bits (&set->values));
716
717 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
718 {
719 /* The number of expressions having a given value is usually
720 relatively small. Thus, rather than making a vector of all
721 the expressions and sorting it by value-id, we walk the values
722 and check in the reverse mapping that tells us what expressions
723 have a given value, to filter those in our set. As a result,
724 the expressions are inserted in value-id order, which means
725 topological order.
726
727 If this is somehow a significant lose for some cases, we can
728 choose which set to walk based on the set size. */
729 bitmap exprset = value_expressions[i];
730 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
731 {
732 if (bitmap_bit_p (&set->expressions, j))
733 result.safe_push (expression_for_id (j));
734 }
735 }
736
737 return result;
738 }
739
740 /* Perform bitmapped set operation DEST &= ORIG. */
741
742 static void
743 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
744 {
745 bitmap_iterator bi;
746 unsigned int i;
747
748 if (dest != orig)
749 {
750 bitmap_head temp;
751 bitmap_initialize (&temp, &grand_bitmap_obstack);
752
753 bitmap_and_into (&dest->values, &orig->values);
754 bitmap_copy (&temp, &dest->expressions);
755 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
756 {
757 pre_expr expr = expression_for_id (i);
758 unsigned int value_id = get_expr_value_id (expr);
759 if (!bitmap_bit_p (&dest->values, value_id))
760 bitmap_clear_bit (&dest->expressions, i);
761 }
762 bitmap_clear (&temp);
763 }
764 }
765
766 /* Subtract all values and expressions contained in ORIG from DEST. */
767
768 static bitmap_set_t
769 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
770 {
771 bitmap_set_t result = bitmap_set_new ();
772 bitmap_iterator bi;
773 unsigned int i;
774
775 bitmap_and_compl (&result->expressions, &dest->expressions,
776 &orig->expressions);
777
778 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
779 {
780 pre_expr expr = expression_for_id (i);
781 unsigned int value_id = get_expr_value_id (expr);
782 bitmap_set_bit (&result->values, value_id);
783 }
784
785 return result;
786 }
787
788 /* Subtract all the values in bitmap set B from bitmap set A. */
789
790 static void
791 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
792 {
793 unsigned int i;
794 bitmap_iterator bi;
795 bitmap_head temp;
796
797 bitmap_initialize (&temp, &grand_bitmap_obstack);
798
799 bitmap_copy (&temp, &a->expressions);
800 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
801 {
802 pre_expr expr = expression_for_id (i);
803 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
804 bitmap_remove_from_set (a, expr);
805 }
806 bitmap_clear (&temp);
807 }
808
809
810 /* Return true if bitmapped set SET contains the value VALUE_ID. */
811
812 static bool
813 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
814 {
815 if (value_id_constant_p (value_id))
816 return true;
817
818 if (!set || bitmap_empty_p (&set->expressions))
819 return false;
820
821 return bitmap_bit_p (&set->values, value_id);
822 }
823
824 static inline bool
825 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
826 {
827 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
828 }
829
830 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
831
832 static void
833 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
834 const pre_expr expr)
835 {
836 bitmap exprset;
837 unsigned int i;
838 bitmap_iterator bi;
839
840 if (value_id_constant_p (lookfor))
841 return;
842
843 if (!bitmap_set_contains_value (set, lookfor))
844 return;
845
846 /* The number of expressions having a given value is usually
847 significantly less than the total number of expressions in SET.
848 Thus, rather than check, for each expression in SET, whether it
849 has the value LOOKFOR, we walk the reverse mapping that tells us
850 what expressions have a given value, and see if any of those
851 expressions are in our set. For large testcases, this is about
852 5-10x faster than walking the bitmap. If this is somehow a
853 significant lose for some cases, we can choose which set to walk
854 based on the set size. */
855 exprset = value_expressions[lookfor];
856 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
857 {
858 if (bitmap_clear_bit (&set->expressions, i))
859 {
860 bitmap_set_bit (&set->expressions, get_expression_id (expr));
861 return;
862 }
863 }
864
865 gcc_unreachable ();
866 }
867
868 /* Return true if two bitmap sets are equal. */
869
870 static bool
871 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
872 {
873 return bitmap_equal_p (&a->values, &b->values);
874 }
875
876 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
877 and add it otherwise. */
878
879 static void
880 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
881 {
882 unsigned int val = get_expr_value_id (expr);
883
884 if (bitmap_set_contains_value (set, val))
885 bitmap_set_replace_value (set, val, expr);
886 else
887 bitmap_insert_into_set (set, expr);
888 }
889
890 /* Insert EXPR into SET if EXPR's value is not already present in
891 SET. */
892
893 static void
894 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
895 {
896 unsigned int val = get_expr_value_id (expr);
897
898 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
899
900 /* Constant values are always considered to be part of the set. */
901 if (value_id_constant_p (val))
902 return;
903
904 /* If the value membership changed, add the expression. */
905 if (bitmap_set_bit (&set->values, val))
906 bitmap_set_bit (&set->expressions, expr->id);
907 }
908
909 /* Print out EXPR to outfile. */
910
911 static void
912 print_pre_expr (FILE *outfile, const pre_expr expr)
913 {
914 switch (expr->kind)
915 {
916 case CONSTANT:
917 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
918 break;
919 case NAME:
920 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
921 break;
922 case NARY:
923 {
924 unsigned int i;
925 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
926 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
927 for (i = 0; i < nary->length; i++)
928 {
929 print_generic_expr (outfile, nary->op[i], 0);
930 if (i != (unsigned) nary->length - 1)
931 fprintf (outfile, ",");
932 }
933 fprintf (outfile, "}");
934 }
935 break;
936
937 case REFERENCE:
938 {
939 vn_reference_op_t vro;
940 unsigned int i;
941 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
942 fprintf (outfile, "{");
943 for (i = 0;
944 ref->operands.iterate (i, &vro);
945 i++)
946 {
947 bool closebrace = false;
948 if (vro->opcode != SSA_NAME
949 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
950 {
951 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
952 if (vro->op0)
953 {
954 fprintf (outfile, "<");
955 closebrace = true;
956 }
957 }
958 if (vro->op0)
959 {
960 print_generic_expr (outfile, vro->op0, 0);
961 if (vro->op1)
962 {
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op1, 0);
965 }
966 if (vro->op2)
967 {
968 fprintf (outfile, ",");
969 print_generic_expr (outfile, vro->op2, 0);
970 }
971 }
972 if (closebrace)
973 fprintf (outfile, ">");
974 if (i != ref->operands.length () - 1)
975 fprintf (outfile, ",");
976 }
977 fprintf (outfile, "}");
978 if (ref->vuse)
979 {
980 fprintf (outfile, "@");
981 print_generic_expr (outfile, ref->vuse, 0);
982 }
983 }
984 break;
985 }
986 }
987 void debug_pre_expr (pre_expr);
988
989 /* Like print_pre_expr but always prints to stderr. */
990 DEBUG_FUNCTION void
991 debug_pre_expr (pre_expr e)
992 {
993 print_pre_expr (stderr, e);
994 fprintf (stderr, "\n");
995 }
996
997 /* Print out SET to OUTFILE. */
998
999 static void
1000 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1001 const char *setname, int blockindex)
1002 {
1003 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1004 if (set)
1005 {
1006 bool first = true;
1007 unsigned i;
1008 bitmap_iterator bi;
1009
1010 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1011 {
1012 const pre_expr expr = expression_for_id (i);
1013
1014 if (!first)
1015 fprintf (outfile, ", ");
1016 first = false;
1017 print_pre_expr (outfile, expr);
1018
1019 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1020 }
1021 }
1022 fprintf (outfile, " }\n");
1023 }
1024
1025 void debug_bitmap_set (bitmap_set_t);
1026
1027 DEBUG_FUNCTION void
1028 debug_bitmap_set (bitmap_set_t set)
1029 {
1030 print_bitmap_set (stderr, set, "debug", 0);
1031 }
1032
1033 void debug_bitmap_sets_for (basic_block);
1034
1035 DEBUG_FUNCTION void
1036 debug_bitmap_sets_for (basic_block bb)
1037 {
1038 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1039 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1040 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1041 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1042 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1043 if (do_partial_partial)
1044 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1045 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1046 }
1047
1048 /* Print out the expressions that have VAL to OUTFILE. */
1049
1050 static void
1051 print_value_expressions (FILE *outfile, unsigned int val)
1052 {
1053 bitmap set = value_expressions[val];
1054 if (set)
1055 {
1056 bitmap_set x;
1057 char s[10];
1058 sprintf (s, "%04d", val);
1059 x.expressions = *set;
1060 print_bitmap_set (outfile, &x, s, 0);
1061 }
1062 }
1063
1064
1065 DEBUG_FUNCTION void
1066 debug_value_expressions (unsigned int val)
1067 {
1068 print_value_expressions (stderr, val);
1069 }
1070
1071 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1072 represent it. */
1073
1074 static pre_expr
1075 get_or_alloc_expr_for_constant (tree constant)
1076 {
1077 unsigned int result_id;
1078 unsigned int value_id;
1079 struct pre_expr_d expr;
1080 pre_expr newexpr;
1081
1082 expr.kind = CONSTANT;
1083 PRE_EXPR_CONSTANT (&expr) = constant;
1084 result_id = lookup_expression_id (&expr);
1085 if (result_id != 0)
1086 return expression_for_id (result_id);
1087
1088 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1089 newexpr->kind = CONSTANT;
1090 PRE_EXPR_CONSTANT (newexpr) = constant;
1091 alloc_expression_id (newexpr);
1092 value_id = get_or_alloc_constant_value_id (constant);
1093 add_to_value (value_id, newexpr);
1094 return newexpr;
1095 }
1096
1097 /* Given a value id V, find the actual tree representing the constant
1098 value if there is one, and return it. Return NULL if we can't find
1099 a constant. */
1100
1101 static tree
1102 get_constant_for_value_id (unsigned int v)
1103 {
1104 if (value_id_constant_p (v))
1105 {
1106 unsigned int i;
1107 bitmap_iterator bi;
1108 bitmap exprset = value_expressions[v];
1109
1110 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1111 {
1112 pre_expr expr = expression_for_id (i);
1113 if (expr->kind == CONSTANT)
1114 return PRE_EXPR_CONSTANT (expr);
1115 }
1116 }
1117 return NULL;
1118 }
1119
1120 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1121 Currently only supports constants and SSA_NAMES. */
1122 static pre_expr
1123 get_or_alloc_expr_for (tree t)
1124 {
1125 if (TREE_CODE (t) == SSA_NAME)
1126 return get_or_alloc_expr_for_name (t);
1127 else if (is_gimple_min_invariant (t))
1128 return get_or_alloc_expr_for_constant (t);
1129 else
1130 {
1131 /* More complex expressions can result from SCCVN expression
1132 simplification that inserts values for them. As they all
1133 do not have VOPs the get handled by the nary ops struct. */
1134 vn_nary_op_t result;
1135 unsigned int result_id;
1136 vn_nary_op_lookup (t, &result);
1137 if (result != NULL)
1138 {
1139 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1140 e->kind = NARY;
1141 PRE_EXPR_NARY (e) = result;
1142 result_id = lookup_expression_id (e);
1143 if (result_id != 0)
1144 {
1145 pool_free (pre_expr_pool, e);
1146 e = expression_for_id (result_id);
1147 return e;
1148 }
1149 alloc_expression_id (e);
1150 return e;
1151 }
1152 }
1153 return NULL;
1154 }
1155
1156 /* Return the folded version of T if T, when folded, is a gimple
1157 min_invariant. Otherwise, return T. */
1158
1159 static pre_expr
1160 fully_constant_expression (pre_expr e)
1161 {
1162 switch (e->kind)
1163 {
1164 case CONSTANT:
1165 return e;
1166 case NARY:
1167 {
1168 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1169 switch (TREE_CODE_CLASS (nary->opcode))
1170 {
1171 case tcc_binary:
1172 case tcc_comparison:
1173 {
1174 /* We have to go from trees to pre exprs to value ids to
1175 constants. */
1176 tree naryop0 = nary->op[0];
1177 tree naryop1 = nary->op[1];
1178 tree result;
1179 if (!is_gimple_min_invariant (naryop0))
1180 {
1181 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1182 unsigned int vrep0 = get_expr_value_id (rep0);
1183 tree const0 = get_constant_for_value_id (vrep0);
1184 if (const0)
1185 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1186 }
1187 if (!is_gimple_min_invariant (naryop1))
1188 {
1189 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1190 unsigned int vrep1 = get_expr_value_id (rep1);
1191 tree const1 = get_constant_for_value_id (vrep1);
1192 if (const1)
1193 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1194 }
1195 result = fold_binary (nary->opcode, nary->type,
1196 naryop0, naryop1);
1197 if (result && is_gimple_min_invariant (result))
1198 return get_or_alloc_expr_for_constant (result);
1199 /* We might have simplified the expression to a
1200 SSA_NAME for example from x_1 * 1. But we cannot
1201 insert a PHI for x_1 unconditionally as x_1 might
1202 not be available readily. */
1203 return e;
1204 }
1205 case tcc_reference:
1206 if (nary->opcode != REALPART_EXPR
1207 && nary->opcode != IMAGPART_EXPR
1208 && nary->opcode != VIEW_CONVERT_EXPR)
1209 return e;
1210 /* Fallthrough. */
1211 case tcc_unary:
1212 {
1213 /* We have to go from trees to pre exprs to value ids to
1214 constants. */
1215 tree naryop0 = nary->op[0];
1216 tree const0, result;
1217 if (is_gimple_min_invariant (naryop0))
1218 const0 = naryop0;
1219 else
1220 {
1221 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1222 unsigned int vrep0 = get_expr_value_id (rep0);
1223 const0 = get_constant_for_value_id (vrep0);
1224 }
1225 result = NULL;
1226 if (const0)
1227 {
1228 tree type1 = TREE_TYPE (nary->op[0]);
1229 const0 = fold_convert (type1, const0);
1230 result = fold_unary (nary->opcode, nary->type, const0);
1231 }
1232 if (result && is_gimple_min_invariant (result))
1233 return get_or_alloc_expr_for_constant (result);
1234 return e;
1235 }
1236 default:
1237 return e;
1238 }
1239 }
1240 case REFERENCE:
1241 {
1242 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1243 tree folded;
1244 if ((folded = fully_constant_vn_reference_p (ref)))
1245 return get_or_alloc_expr_for_constant (folded);
1246 return e;
1247 }
1248 default:
1249 return e;
1250 }
1251 return e;
1252 }
1253
1254 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1255 it has the value it would have in BLOCK. Set *SAME_VALID to true
1256 in case the new vuse doesn't change the value id of the OPERANDS. */
1257
1258 static tree
1259 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1260 alias_set_type set, tree type, tree vuse,
1261 basic_block phiblock,
1262 basic_block block, bool *same_valid)
1263 {
1264 gimple phi = SSA_NAME_DEF_STMT (vuse);
1265 ao_ref ref;
1266 edge e = NULL;
1267 bool use_oracle;
1268
1269 *same_valid = true;
1270
1271 if (gimple_bb (phi) != phiblock)
1272 return vuse;
1273
1274 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1275
1276 /* Use the alias-oracle to find either the PHI node in this block,
1277 the first VUSE used in this block that is equivalent to vuse or
1278 the first VUSE which definition in this block kills the value. */
1279 if (gimple_code (phi) == GIMPLE_PHI)
1280 e = find_edge (block, phiblock);
1281 else if (use_oracle)
1282 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1283 {
1284 vuse = gimple_vuse (phi);
1285 phi = SSA_NAME_DEF_STMT (vuse);
1286 if (gimple_bb (phi) != phiblock)
1287 return vuse;
1288 if (gimple_code (phi) == GIMPLE_PHI)
1289 {
1290 e = find_edge (block, phiblock);
1291 break;
1292 }
1293 }
1294 else
1295 return NULL_TREE;
1296
1297 if (e)
1298 {
1299 if (use_oracle)
1300 {
1301 bitmap visited = NULL;
1302 unsigned int cnt;
1303 /* Try to find a vuse that dominates this phi node by skipping
1304 non-clobbering statements. */
1305 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false);
1306 if (visited)
1307 BITMAP_FREE (visited);
1308 }
1309 else
1310 vuse = NULL_TREE;
1311 if (!vuse)
1312 {
1313 /* If we didn't find any, the value ID can't stay the same,
1314 but return the translated vuse. */
1315 *same_valid = false;
1316 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1317 }
1318 /* ??? We would like to return vuse here as this is the canonical
1319 upmost vdef that this reference is associated with. But during
1320 insertion of the references into the hash tables we only ever
1321 directly insert with their direct gimple_vuse, hence returning
1322 something else would make us not find the other expression. */
1323 return PHI_ARG_DEF (phi, e->dest_idx);
1324 }
1325
1326 return NULL_TREE;
1327 }
1328
1329 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1330 SET2. This is used to avoid making a set consisting of the union
1331 of PA_IN and ANTIC_IN during insert. */
1332
1333 static inline pre_expr
1334 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1335 {
1336 pre_expr result;
1337
1338 result = bitmap_find_leader (set1, val);
1339 if (!result && set2)
1340 result = bitmap_find_leader (set2, val);
1341 return result;
1342 }
1343
1344 /* Get the tree type for our PRE expression e. */
1345
1346 static tree
1347 get_expr_type (const pre_expr e)
1348 {
1349 switch (e->kind)
1350 {
1351 case NAME:
1352 return TREE_TYPE (PRE_EXPR_NAME (e));
1353 case CONSTANT:
1354 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1355 case REFERENCE:
1356 return PRE_EXPR_REFERENCE (e)->type;
1357 case NARY:
1358 return PRE_EXPR_NARY (e)->type;
1359 }
1360 gcc_unreachable ();
1361 }
1362
1363 /* Get a representative SSA_NAME for a given expression.
1364 Since all of our sub-expressions are treated as values, we require
1365 them to be SSA_NAME's for simplicity.
1366 Prior versions of GVNPRE used to use "value handles" here, so that
1367 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1368 either case, the operands are really values (IE we do not expect
1369 them to be usable without finding leaders). */
1370
1371 static tree
1372 get_representative_for (const pre_expr e)
1373 {
1374 tree name;
1375 unsigned int value_id = get_expr_value_id (e);
1376
1377 switch (e->kind)
1378 {
1379 case NAME:
1380 return PRE_EXPR_NAME (e);
1381 case CONSTANT:
1382 return PRE_EXPR_CONSTANT (e);
1383 case NARY:
1384 case REFERENCE:
1385 {
1386 /* Go through all of the expressions representing this value
1387 and pick out an SSA_NAME. */
1388 unsigned int i;
1389 bitmap_iterator bi;
1390 bitmap exprs = value_expressions[value_id];
1391 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1392 {
1393 pre_expr rep = expression_for_id (i);
1394 if (rep->kind == NAME)
1395 return PRE_EXPR_NAME (rep);
1396 else if (rep->kind == CONSTANT)
1397 return PRE_EXPR_CONSTANT (rep);
1398 }
1399 }
1400 break;
1401 }
1402
1403 /* If we reached here we couldn't find an SSA_NAME. This can
1404 happen when we've discovered a value that has never appeared in
1405 the program as set to an SSA_NAME, as the result of phi translation.
1406 Create one here.
1407 ??? We should be able to re-use this when we insert the statement
1408 to compute it. */
1409 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1410 VN_INFO_GET (name)->value_id = value_id;
1411 VN_INFO (name)->valnum = name;
1412 /* ??? For now mark this SSA name for release by SCCVN. */
1413 VN_INFO (name)->needs_insertion = true;
1414 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1415 if (dump_file && (dump_flags & TDF_DETAILS))
1416 {
1417 fprintf (dump_file, "Created SSA_NAME representative ");
1418 print_generic_expr (dump_file, name, 0);
1419 fprintf (dump_file, " for expression:");
1420 print_pre_expr (dump_file, e);
1421 fprintf (dump_file, " (%04d)\n", value_id);
1422 }
1423
1424 return name;
1425 }
1426
1427
1428
1429 static pre_expr
1430 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1431 basic_block pred, basic_block phiblock);
1432
1433 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1434 the phis in PRED. Return NULL if we can't find a leader for each part
1435 of the translated expression. */
1436
1437 static pre_expr
1438 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1439 basic_block pred, basic_block phiblock)
1440 {
1441 switch (expr->kind)
1442 {
1443 case NARY:
1444 {
1445 unsigned int i;
1446 bool changed = false;
1447 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1448 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1449 sizeof_vn_nary_op (nary->length));
1450 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1451
1452 for (i = 0; i < newnary->length; i++)
1453 {
1454 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1455 continue;
1456 else
1457 {
1458 pre_expr leader, result;
1459 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1460 leader = find_leader_in_sets (op_val_id, set1, set2);
1461 result = phi_translate (leader, set1, set2, pred, phiblock);
1462 if (result && result != leader)
1463 {
1464 tree name = get_representative_for (result);
1465 if (!name)
1466 return NULL;
1467 newnary->op[i] = name;
1468 }
1469 else if (!result)
1470 return NULL;
1471
1472 changed |= newnary->op[i] != nary->op[i];
1473 }
1474 }
1475 if (changed)
1476 {
1477 pre_expr constant;
1478 unsigned int new_val_id;
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_alloc (pre_expr_pool);
1489 expr->kind = NARY;
1490 expr->id = 0;
1491 if (nary)
1492 {
1493 PRE_EXPR_NARY (expr) = nary;
1494 constant = fully_constant_expression (expr);
1495 if (constant != expr)
1496 return constant;
1497
1498 new_val_id = nary->value_id;
1499 get_or_alloc_expression_id (expr);
1500 }
1501 else
1502 {
1503 new_val_id = get_next_value_id ();
1504 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1505 nary = vn_nary_op_insert_pieces (newnary->length,
1506 newnary->opcode,
1507 newnary->type,
1508 &newnary->op[0],
1509 result, new_val_id);
1510 PRE_EXPR_NARY (expr) = nary;
1511 constant = fully_constant_expression (expr);
1512 if (constant != expr)
1513 return constant;
1514 get_or_alloc_expression_id (expr);
1515 }
1516 add_to_value (new_val_id, expr);
1517 }
1518 return expr;
1519 }
1520 break;
1521
1522 case REFERENCE:
1523 {
1524 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1525 vec<vn_reference_op_s> operands = ref->operands;
1526 tree vuse = ref->vuse;
1527 tree newvuse = vuse;
1528 vec<vn_reference_op_s> newoperands = vNULL;
1529 bool changed = false, same_valid = true;
1530 unsigned int i, j, n;
1531 vn_reference_op_t operand;
1532 vn_reference_t newref;
1533
1534 for (i = 0, j = 0;
1535 operands.iterate (i, &operand); i++, j++)
1536 {
1537 pre_expr opresult;
1538 pre_expr leader;
1539 tree op[3];
1540 tree type = operand->type;
1541 vn_reference_op_s newop = *operand;
1542 op[0] = operand->op0;
1543 op[1] = operand->op1;
1544 op[2] = operand->op2;
1545 for (n = 0; n < 3; ++n)
1546 {
1547 unsigned int op_val_id;
1548 if (!op[n])
1549 continue;
1550 if (TREE_CODE (op[n]) != SSA_NAME)
1551 {
1552 /* We can't possibly insert these. */
1553 if (n != 0
1554 && !is_gimple_min_invariant (op[n]))
1555 break;
1556 continue;
1557 }
1558 op_val_id = VN_INFO (op[n])->value_id;
1559 leader = find_leader_in_sets (op_val_id, set1, set2);
1560 if (!leader)
1561 break;
1562 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1563 if (!opresult)
1564 break;
1565 if (opresult != leader)
1566 {
1567 tree name = get_representative_for (opresult);
1568 if (!name)
1569 break;
1570 changed |= name != op[n];
1571 op[n] = name;
1572 }
1573 }
1574 if (n != 3)
1575 {
1576 newoperands.release ();
1577 return NULL;
1578 }
1579 if (!newoperands.exists ())
1580 newoperands = operands.copy ();
1581 /* We may have changed from an SSA_NAME to a constant */
1582 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1583 newop.opcode = TREE_CODE (op[0]);
1584 newop.type = type;
1585 newop.op0 = op[0];
1586 newop.op1 = op[1];
1587 newop.op2 = op[2];
1588 /* If it transforms a non-constant ARRAY_REF into a constant
1589 one, adjust the constant offset. */
1590 if (newop.opcode == ARRAY_REF
1591 && newop.off == -1
1592 && TREE_CODE (op[0]) == INTEGER_CST
1593 && TREE_CODE (op[1]) == INTEGER_CST
1594 && TREE_CODE (op[2]) == INTEGER_CST)
1595 {
1596 double_int off = tree_to_double_int (op[0]);
1597 off += -tree_to_double_int (op[1]);
1598 off *= tree_to_double_int (op[2]);
1599 if (off.fits_shwi ())
1600 newop.off = off.low;
1601 }
1602 newoperands[j] = newop;
1603 /* If it transforms from an SSA_NAME to an address, fold with
1604 a preceding indirect reference. */
1605 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1606 && newoperands[j - 1].opcode == MEM_REF)
1607 vn_reference_fold_indirect (&newoperands, &j);
1608 }
1609 if (i != operands.length ())
1610 {
1611 newoperands.release ();
1612 return NULL;
1613 }
1614
1615 if (vuse)
1616 {
1617 newvuse = translate_vuse_through_block (newoperands,
1618 ref->set, ref->type,
1619 vuse, phiblock, pred,
1620 &same_valid);
1621 if (newvuse == NULL_TREE)
1622 {
1623 newoperands.release ();
1624 return NULL;
1625 }
1626 }
1627
1628 if (changed || newvuse != vuse)
1629 {
1630 unsigned int new_val_id;
1631 pre_expr constant;
1632
1633 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1634 ref->type,
1635 newoperands,
1636 &newref, VN_WALK);
1637 if (result)
1638 newoperands.release ();
1639
1640 /* We can always insert constants, so if we have a partial
1641 redundant constant load of another type try to translate it
1642 to a constant of appropriate type. */
1643 if (result && is_gimple_min_invariant (result))
1644 {
1645 tree tem = result;
1646 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1647 {
1648 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1649 if (tem && !is_gimple_min_invariant (tem))
1650 tem = NULL_TREE;
1651 }
1652 if (tem)
1653 return get_or_alloc_expr_for_constant (tem);
1654 }
1655
1656 /* If we'd have to convert things we would need to validate
1657 if we can insert the translated expression. So fail
1658 here for now - we cannot insert an alias with a different
1659 type in the VN tables either, as that would assert. */
1660 if (result
1661 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1662 return NULL;
1663 else if (!result && newref
1664 && !useless_type_conversion_p (ref->type, newref->type))
1665 {
1666 newoperands.release ();
1667 return NULL;
1668 }
1669
1670 expr = (pre_expr) pool_alloc (pre_expr_pool);
1671 expr->kind = REFERENCE;
1672 expr->id = 0;
1673
1674 if (newref)
1675 {
1676 PRE_EXPR_REFERENCE (expr) = newref;
1677 constant = fully_constant_expression (expr);
1678 if (constant != expr)
1679 return constant;
1680
1681 new_val_id = newref->value_id;
1682 get_or_alloc_expression_id (expr);
1683 }
1684 else
1685 {
1686 if (changed || !same_valid)
1687 {
1688 new_val_id = get_next_value_id ();
1689 value_expressions.safe_grow_cleared
1690 (get_max_value_id () + 1);
1691 }
1692 else
1693 new_val_id = ref->value_id;
1694 newref = vn_reference_insert_pieces (newvuse, ref->set,
1695 ref->type,
1696 newoperands,
1697 result, new_val_id);
1698 newoperands.create (0);
1699 PRE_EXPR_REFERENCE (expr) = newref;
1700 constant = fully_constant_expression (expr);
1701 if (constant != expr)
1702 return constant;
1703 get_or_alloc_expression_id (expr);
1704 }
1705 add_to_value (new_val_id, expr);
1706 }
1707 newoperands.release ();
1708 return expr;
1709 }
1710 break;
1711
1712 case NAME:
1713 {
1714 tree name = PRE_EXPR_NAME (expr);
1715 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1716 /* If the SSA name is defined by a PHI node in this block,
1717 translate it. */
1718 if (gimple_code (def_stmt) == GIMPLE_PHI
1719 && gimple_bb (def_stmt) == phiblock)
1720 {
1721 edge e = find_edge (pred, gimple_bb (def_stmt));
1722 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1723
1724 /* Handle constant. */
1725 if (is_gimple_min_invariant (def))
1726 return get_or_alloc_expr_for_constant (def);
1727
1728 return get_or_alloc_expr_for_name (def);
1729 }
1730 /* Otherwise return it unchanged - it will get cleaned if its
1731 value is not available in PREDs AVAIL_OUT set of expressions. */
1732 return expr;
1733 }
1734
1735 default:
1736 gcc_unreachable ();
1737 }
1738 }
1739
1740 /* Wrapper around phi_translate_1 providing caching functionality. */
1741
1742 static pre_expr
1743 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1744 basic_block pred, basic_block phiblock)
1745 {
1746 expr_pred_trans_t slot = NULL;
1747 pre_expr phitrans;
1748
1749 if (!expr)
1750 return NULL;
1751
1752 /* Constants contain no values that need translation. */
1753 if (expr->kind == CONSTANT)
1754 return expr;
1755
1756 if (value_id_constant_p (get_expr_value_id (expr)))
1757 return expr;
1758
1759 /* Don't add translations of NAMEs as those are cheap to translate. */
1760 if (expr->kind != NAME)
1761 {
1762 if (phi_trans_add (&slot, expr, pred))
1763 return slot->v;
1764 /* Store NULL for the value we want to return in the case of
1765 recursing. */
1766 slot->v = NULL;
1767 }
1768
1769 /* Translate. */
1770 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1771
1772 if (slot)
1773 {
1774 if (phitrans)
1775 slot->v = phitrans;
1776 else
1777 /* Remove failed translations again, they cause insert
1778 iteration to not pick up new opportunities reliably. */
1779 phi_translate_table.remove_elt_with_hash (slot, slot->hashcode);
1780 }
1781
1782 return phitrans;
1783 }
1784
1785
1786 /* For each expression in SET, translate the values through phi nodes
1787 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1788 expressions in DEST. */
1789
1790 static void
1791 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1792 basic_block phiblock)
1793 {
1794 vec<pre_expr> exprs;
1795 pre_expr expr;
1796 int i;
1797
1798 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1799 {
1800 bitmap_set_copy (dest, set);
1801 return;
1802 }
1803
1804 exprs = sorted_array_from_bitmap_set (set);
1805 FOR_EACH_VEC_ELT (exprs, i, expr)
1806 {
1807 pre_expr translated;
1808 translated = phi_translate (expr, set, NULL, pred, phiblock);
1809 if (!translated)
1810 continue;
1811
1812 /* We might end up with multiple expressions from SET being
1813 translated to the same value. In this case we do not want
1814 to retain the NARY or REFERENCE expression but prefer a NAME
1815 which would be the leader. */
1816 if (translated->kind == NAME)
1817 bitmap_value_replace_in_set (dest, translated);
1818 else
1819 bitmap_value_insert_into_set (dest, translated);
1820 }
1821 exprs.release ();
1822 }
1823
1824 /* Find the leader for a value (i.e., the name representing that
1825 value) in a given set, and return it. Return NULL if no leader
1826 is found. */
1827
1828 static pre_expr
1829 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1830 {
1831 if (value_id_constant_p (val))
1832 {
1833 unsigned int i;
1834 bitmap_iterator bi;
1835 bitmap exprset = value_expressions[val];
1836
1837 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1838 {
1839 pre_expr expr = expression_for_id (i);
1840 if (expr->kind == CONSTANT)
1841 return expr;
1842 }
1843 }
1844 if (bitmap_set_contains_value (set, val))
1845 {
1846 /* Rather than walk the entire bitmap of expressions, and see
1847 whether any of them has the value we are looking for, we look
1848 at the reverse mapping, which tells us the set of expressions
1849 that have a given value (IE value->expressions with that
1850 value) and see if any of those expressions are in our set.
1851 The number of expressions per value is usually significantly
1852 less than the number of expressions in the set. In fact, for
1853 large testcases, doing it this way is roughly 5-10x faster
1854 than walking the bitmap.
1855 If this is somehow a significant lose for some cases, we can
1856 choose which set to walk based on which set is smaller. */
1857 unsigned int i;
1858 bitmap_iterator bi;
1859 bitmap exprset = value_expressions[val];
1860
1861 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1862 return expression_for_id (i);
1863 }
1864 return NULL;
1865 }
1866
1867 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1868 BLOCK by seeing if it is not killed in the block. Note that we are
1869 only determining whether there is a store that kills it. Because
1870 of the order in which clean iterates over values, we are guaranteed
1871 that altered operands will have caused us to be eliminated from the
1872 ANTIC_IN set already. */
1873
1874 static bool
1875 value_dies_in_block_x (pre_expr expr, basic_block block)
1876 {
1877 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1878 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1879 gimple def;
1880 gimple_stmt_iterator gsi;
1881 unsigned id = get_expression_id (expr);
1882 bool res = false;
1883 ao_ref ref;
1884
1885 if (!vuse)
1886 return false;
1887
1888 /* Lookup a previously calculated result. */
1889 if (EXPR_DIES (block)
1890 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1891 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1892
1893 /* A memory expression {e, VUSE} dies in the block if there is a
1894 statement that may clobber e. If, starting statement walk from the
1895 top of the basic block, a statement uses VUSE there can be no kill
1896 inbetween that use and the original statement that loaded {e, VUSE},
1897 so we can stop walking. */
1898 ref.base = NULL_TREE;
1899 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1900 {
1901 tree def_vuse, def_vdef;
1902 def = gsi_stmt (gsi);
1903 def_vuse = gimple_vuse (def);
1904 def_vdef = gimple_vdef (def);
1905
1906 /* Not a memory statement. */
1907 if (!def_vuse)
1908 continue;
1909
1910 /* Not a may-def. */
1911 if (!def_vdef)
1912 {
1913 /* A load with the same VUSE, we're done. */
1914 if (def_vuse == vuse)
1915 break;
1916
1917 continue;
1918 }
1919
1920 /* Init ref only if we really need it. */
1921 if (ref.base == NULL_TREE
1922 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1923 refx->operands))
1924 {
1925 res = true;
1926 break;
1927 }
1928 /* If the statement may clobber expr, it dies. */
1929 if (stmt_may_clobber_ref_p_1 (def, &ref))
1930 {
1931 res = true;
1932 break;
1933 }
1934 }
1935
1936 /* Remember the result. */
1937 if (!EXPR_DIES (block))
1938 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1939 bitmap_set_bit (EXPR_DIES (block), id * 2);
1940 if (res)
1941 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1942
1943 return res;
1944 }
1945
1946
1947 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1948 contains its value-id. */
1949
1950 static bool
1951 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1952 {
1953 if (op && TREE_CODE (op) == SSA_NAME)
1954 {
1955 unsigned int value_id = VN_INFO (op)->value_id;
1956 if (!(bitmap_set_contains_value (set1, value_id)
1957 || (set2 && bitmap_set_contains_value (set2, value_id))))
1958 return false;
1959 }
1960 return true;
1961 }
1962
1963 /* Determine if the expression EXPR is valid in SET1 U SET2.
1964 ONLY SET2 CAN BE NULL.
1965 This means that we have a leader for each part of the expression
1966 (if it consists of values), or the expression is an SSA_NAME.
1967 For loads/calls, we also see if the vuse is killed in this block. */
1968
1969 static bool
1970 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1971 basic_block block)
1972 {
1973 switch (expr->kind)
1974 {
1975 case NAME:
1976 return bitmap_find_leader (AVAIL_OUT (block),
1977 get_expr_value_id (expr)) != NULL;
1978 case NARY:
1979 {
1980 unsigned int i;
1981 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1982 for (i = 0; i < nary->length; i++)
1983 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1984 return false;
1985 return true;
1986 }
1987 break;
1988 case REFERENCE:
1989 {
1990 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1991 vn_reference_op_t vro;
1992 unsigned int i;
1993
1994 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1995 {
1996 if (!op_valid_in_sets (set1, set2, vro->op0)
1997 || !op_valid_in_sets (set1, set2, vro->op1)
1998 || !op_valid_in_sets (set1, set2, vro->op2))
1999 return false;
2000 }
2001 return true;
2002 }
2003 default:
2004 gcc_unreachable ();
2005 }
2006 }
2007
2008 /* Clean the set of expressions that are no longer valid in SET1 or
2009 SET2. This means expressions that are made up of values we have no
2010 leaders for in SET1 or SET2. This version is used for partial
2011 anticipation, which means it is not valid in either ANTIC_IN or
2012 PA_IN. */
2013
2014 static void
2015 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2016 {
2017 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2018 pre_expr expr;
2019 int i;
2020
2021 FOR_EACH_VEC_ELT (exprs, i, expr)
2022 {
2023 if (!valid_in_sets (set1, set2, expr, block))
2024 bitmap_remove_from_set (set1, expr);
2025 }
2026 exprs.release ();
2027 }
2028
2029 /* Clean the set of expressions that are no longer valid in SET. This
2030 means expressions that are made up of values we have no leaders for
2031 in SET. */
2032
2033 static void
2034 clean (bitmap_set_t set, basic_block block)
2035 {
2036 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2037 pre_expr expr;
2038 int i;
2039
2040 FOR_EACH_VEC_ELT (exprs, i, expr)
2041 {
2042 if (!valid_in_sets (set, NULL, expr, block))
2043 bitmap_remove_from_set (set, expr);
2044 }
2045 exprs.release ();
2046 }
2047
2048 /* Clean the set of expressions that are no longer valid in SET because
2049 they are clobbered in BLOCK or because they trap and may not be executed. */
2050
2051 static void
2052 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2053 {
2054 bitmap_iterator bi;
2055 unsigned i;
2056
2057 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2058 {
2059 pre_expr expr = expression_for_id (i);
2060 if (expr->kind == REFERENCE)
2061 {
2062 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2063 if (ref->vuse)
2064 {
2065 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2066 if (!gimple_nop_p (def_stmt)
2067 && ((gimple_bb (def_stmt) != block
2068 && !dominated_by_p (CDI_DOMINATORS,
2069 block, gimple_bb (def_stmt)))
2070 || (gimple_bb (def_stmt) == block
2071 && value_dies_in_block_x (expr, block))))
2072 bitmap_remove_from_set (set, expr);
2073 }
2074 }
2075 else if (expr->kind == NARY)
2076 {
2077 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2078 /* If the NARY may trap make sure the block does not contain
2079 a possible exit point.
2080 ??? This is overly conservative if we translate AVAIL_OUT
2081 as the available expression might be after the exit point. */
2082 if (BB_MAY_NOTRETURN (block)
2083 && vn_nary_may_trap (nary))
2084 bitmap_remove_from_set (set, expr);
2085 }
2086 }
2087 }
2088
2089 static sbitmap has_abnormal_preds;
2090
2091 /* List of blocks that may have changed during ANTIC computation and
2092 thus need to be iterated over. */
2093
2094 static sbitmap changed_blocks;
2095
2096 /* Decide whether to defer a block for a later iteration, or PHI
2097 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2098 should defer the block, and true if we processed it. */
2099
2100 static bool
2101 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2102 basic_block block, basic_block phiblock)
2103 {
2104 if (!BB_VISITED (phiblock))
2105 {
2106 bitmap_set_bit (changed_blocks, block->index);
2107 BB_VISITED (block) = 0;
2108 BB_DEFERRED (block) = 1;
2109 return false;
2110 }
2111 else
2112 phi_translate_set (dest, source, block, phiblock);
2113 return true;
2114 }
2115
2116 /* Compute the ANTIC set for BLOCK.
2117
2118 If succs(BLOCK) > 1 then
2119 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2120 else if succs(BLOCK) == 1 then
2121 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2122
2123 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2124 */
2125
2126 static bool
2127 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2128 {
2129 bool changed = false;
2130 bitmap_set_t S, old, ANTIC_OUT;
2131 bitmap_iterator bi;
2132 unsigned int bii;
2133 edge e;
2134 edge_iterator ei;
2135
2136 old = ANTIC_OUT = S = NULL;
2137 BB_VISITED (block) = 1;
2138
2139 /* If any edges from predecessors are abnormal, antic_in is empty,
2140 so do nothing. */
2141 if (block_has_abnormal_pred_edge)
2142 goto maybe_dump_sets;
2143
2144 old = ANTIC_IN (block);
2145 ANTIC_OUT = bitmap_set_new ();
2146
2147 /* If the block has no successors, ANTIC_OUT is empty. */
2148 if (EDGE_COUNT (block->succs) == 0)
2149 ;
2150 /* If we have one successor, we could have some phi nodes to
2151 translate through. */
2152 else if (single_succ_p (block))
2153 {
2154 basic_block succ_bb = single_succ (block);
2155
2156 /* We trade iterations of the dataflow equations for having to
2157 phi translate the maximal set, which is incredibly slow
2158 (since the maximal set often has 300+ members, even when you
2159 have a small number of blocks).
2160 Basically, we defer the computation of ANTIC for this block
2161 until we have processed it's successor, which will inevitably
2162 have a *much* smaller set of values to phi translate once
2163 clean has been run on it.
2164 The cost of doing this is that we technically perform more
2165 iterations, however, they are lower cost iterations.
2166
2167 Timings for PRE on tramp3d-v4:
2168 without maximal set fix: 11 seconds
2169 with maximal set fix/without deferring: 26 seconds
2170 with maximal set fix/with deferring: 11 seconds
2171 */
2172
2173 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2174 block, succ_bb))
2175 {
2176 changed = true;
2177 goto maybe_dump_sets;
2178 }
2179 }
2180 /* If we have multiple successors, we take the intersection of all of
2181 them. Note that in the case of loop exit phi nodes, we may have
2182 phis to translate through. */
2183 else
2184 {
2185 vec<basic_block> worklist;
2186 size_t i;
2187 basic_block bprime, first = NULL;
2188
2189 worklist.create (EDGE_COUNT (block->succs));
2190 FOR_EACH_EDGE (e, ei, block->succs)
2191 {
2192 if (!first
2193 && BB_VISITED (e->dest))
2194 first = e->dest;
2195 else if (BB_VISITED (e->dest))
2196 worklist.quick_push (e->dest);
2197 }
2198
2199 /* Of multiple successors we have to have visited one already. */
2200 if (!first)
2201 {
2202 bitmap_set_bit (changed_blocks, block->index);
2203 BB_VISITED (block) = 0;
2204 BB_DEFERRED (block) = 1;
2205 changed = true;
2206 worklist.release ();
2207 goto maybe_dump_sets;
2208 }
2209
2210 if (!gimple_seq_empty_p (phi_nodes (first)))
2211 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2212 else
2213 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2214
2215 FOR_EACH_VEC_ELT (worklist, i, bprime)
2216 {
2217 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2218 {
2219 bitmap_set_t tmp = bitmap_set_new ();
2220 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2221 bitmap_set_and (ANTIC_OUT, tmp);
2222 bitmap_set_free (tmp);
2223 }
2224 else
2225 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2226 }
2227 worklist.release ();
2228 }
2229
2230 /* Prune expressions that are clobbered in block and thus become
2231 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2232 prune_clobbered_mems (ANTIC_OUT, block);
2233
2234 /* Generate ANTIC_OUT - TMP_GEN. */
2235 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2236
2237 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2238 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2239 TMP_GEN (block));
2240
2241 /* Then union in the ANTIC_OUT - TMP_GEN values,
2242 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2243 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2244 bitmap_value_insert_into_set (ANTIC_IN (block),
2245 expression_for_id (bii));
2246
2247 clean (ANTIC_IN (block), block);
2248
2249 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2250 {
2251 changed = true;
2252 bitmap_set_bit (changed_blocks, block->index);
2253 FOR_EACH_EDGE (e, ei, block->preds)
2254 bitmap_set_bit (changed_blocks, e->src->index);
2255 }
2256 else
2257 bitmap_clear_bit (changed_blocks, block->index);
2258
2259 maybe_dump_sets:
2260 if (dump_file && (dump_flags & TDF_DETAILS))
2261 {
2262 if (!BB_DEFERRED (block) || BB_VISITED (block))
2263 {
2264 if (ANTIC_OUT)
2265 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2266
2267 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2268 block->index);
2269
2270 if (S)
2271 print_bitmap_set (dump_file, S, "S", block->index);
2272 }
2273 else
2274 {
2275 fprintf (dump_file,
2276 "Block %d was deferred for a future iteration.\n",
2277 block->index);
2278 }
2279 }
2280 if (old)
2281 bitmap_set_free (old);
2282 if (S)
2283 bitmap_set_free (S);
2284 if (ANTIC_OUT)
2285 bitmap_set_free (ANTIC_OUT);
2286 return changed;
2287 }
2288
2289 /* Compute PARTIAL_ANTIC for BLOCK.
2290
2291 If succs(BLOCK) > 1 then
2292 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2293 in ANTIC_OUT for all succ(BLOCK)
2294 else if succs(BLOCK) == 1 then
2295 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2296
2297 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2298 - ANTIC_IN[BLOCK])
2299
2300 */
2301 static bool
2302 compute_partial_antic_aux (basic_block block,
2303 bool block_has_abnormal_pred_edge)
2304 {
2305 bool changed = false;
2306 bitmap_set_t old_PA_IN;
2307 bitmap_set_t PA_OUT;
2308 edge e;
2309 edge_iterator ei;
2310 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2311
2312 old_PA_IN = PA_OUT = NULL;
2313
2314 /* If any edges from predecessors are abnormal, antic_in is empty,
2315 so do nothing. */
2316 if (block_has_abnormal_pred_edge)
2317 goto maybe_dump_sets;
2318
2319 /* If there are too many partially anticipatable values in the
2320 block, phi_translate_set can take an exponential time: stop
2321 before the translation starts. */
2322 if (max_pa
2323 && single_succ_p (block)
2324 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2325 goto maybe_dump_sets;
2326
2327 old_PA_IN = PA_IN (block);
2328 PA_OUT = bitmap_set_new ();
2329
2330 /* If the block has no successors, ANTIC_OUT is empty. */
2331 if (EDGE_COUNT (block->succs) == 0)
2332 ;
2333 /* If we have one successor, we could have some phi nodes to
2334 translate through. Note that we can't phi translate across DFS
2335 back edges in partial antic, because it uses a union operation on
2336 the successors. For recurrences like IV's, we will end up
2337 generating a new value in the set on each go around (i + 3 (VH.1)
2338 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2339 else if (single_succ_p (block))
2340 {
2341 basic_block succ = single_succ (block);
2342 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2343 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2344 }
2345 /* If we have multiple successors, we take the union of all of
2346 them. */
2347 else
2348 {
2349 vec<basic_block> worklist;
2350 size_t i;
2351 basic_block bprime;
2352
2353 worklist.create (EDGE_COUNT (block->succs));
2354 FOR_EACH_EDGE (e, ei, block->succs)
2355 {
2356 if (e->flags & EDGE_DFS_BACK)
2357 continue;
2358 worklist.quick_push (e->dest);
2359 }
2360 if (worklist.length () > 0)
2361 {
2362 FOR_EACH_VEC_ELT (worklist, i, bprime)
2363 {
2364 unsigned int i;
2365 bitmap_iterator bi;
2366
2367 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2368 bitmap_value_insert_into_set (PA_OUT,
2369 expression_for_id (i));
2370 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2371 {
2372 bitmap_set_t pa_in = bitmap_set_new ();
2373 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2374 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2375 bitmap_value_insert_into_set (PA_OUT,
2376 expression_for_id (i));
2377 bitmap_set_free (pa_in);
2378 }
2379 else
2380 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2381 bitmap_value_insert_into_set (PA_OUT,
2382 expression_for_id (i));
2383 }
2384 }
2385 worklist.release ();
2386 }
2387
2388 /* Prune expressions that are clobbered in block and thus become
2389 invalid if translated from PA_OUT to PA_IN. */
2390 prune_clobbered_mems (PA_OUT, block);
2391
2392 /* PA_IN starts with PA_OUT - TMP_GEN.
2393 Then we subtract things from ANTIC_IN. */
2394 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2395
2396 /* For partial antic, we want to put back in the phi results, since
2397 we will properly avoid making them partially antic over backedges. */
2398 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2399 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2400
2401 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2402 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2403
2404 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2405
2406 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2407 {
2408 changed = true;
2409 bitmap_set_bit (changed_blocks, block->index);
2410 FOR_EACH_EDGE (e, ei, block->preds)
2411 bitmap_set_bit (changed_blocks, e->src->index);
2412 }
2413 else
2414 bitmap_clear_bit (changed_blocks, block->index);
2415
2416 maybe_dump_sets:
2417 if (dump_file && (dump_flags & TDF_DETAILS))
2418 {
2419 if (PA_OUT)
2420 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2421
2422 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2423 }
2424 if (old_PA_IN)
2425 bitmap_set_free (old_PA_IN);
2426 if (PA_OUT)
2427 bitmap_set_free (PA_OUT);
2428 return changed;
2429 }
2430
2431 /* Compute ANTIC and partial ANTIC sets. */
2432
2433 static void
2434 compute_antic (void)
2435 {
2436 bool changed = true;
2437 int num_iterations = 0;
2438 basic_block block;
2439 int i;
2440
2441 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2442 We pre-build the map of blocks with incoming abnormal edges here. */
2443 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2444 bitmap_clear (has_abnormal_preds);
2445
2446 FOR_ALL_BB (block)
2447 {
2448 edge_iterator ei;
2449 edge e;
2450
2451 FOR_EACH_EDGE (e, ei, block->preds)
2452 {
2453 e->flags &= ~EDGE_DFS_BACK;
2454 if (e->flags & EDGE_ABNORMAL)
2455 {
2456 bitmap_set_bit (has_abnormal_preds, block->index);
2457 break;
2458 }
2459 }
2460
2461 BB_VISITED (block) = 0;
2462 BB_DEFERRED (block) = 0;
2463
2464 /* While we are here, give empty ANTIC_IN sets to each block. */
2465 ANTIC_IN (block) = bitmap_set_new ();
2466 PA_IN (block) = bitmap_set_new ();
2467 }
2468
2469 /* At the exit block we anticipate nothing. */
2470 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2471
2472 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2473 bitmap_ones (changed_blocks);
2474 while (changed)
2475 {
2476 if (dump_file && (dump_flags & TDF_DETAILS))
2477 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2478 /* ??? We need to clear our PHI translation cache here as the
2479 ANTIC sets shrink and we restrict valid translations to
2480 those having operands with leaders in ANTIC. Same below
2481 for PA ANTIC computation. */
2482 num_iterations++;
2483 changed = false;
2484 for (i = postorder_num - 1; i >= 0; i--)
2485 {
2486 if (bitmap_bit_p (changed_blocks, postorder[i]))
2487 {
2488 basic_block block = BASIC_BLOCK (postorder[i]);
2489 changed |= compute_antic_aux (block,
2490 bitmap_bit_p (has_abnormal_preds,
2491 block->index));
2492 }
2493 }
2494 /* Theoretically possible, but *highly* unlikely. */
2495 gcc_checking_assert (num_iterations < 500);
2496 }
2497
2498 statistics_histogram_event (cfun, "compute_antic iterations",
2499 num_iterations);
2500
2501 if (do_partial_partial)
2502 {
2503 bitmap_ones (changed_blocks);
2504 mark_dfs_back_edges ();
2505 num_iterations = 0;
2506 changed = true;
2507 while (changed)
2508 {
2509 if (dump_file && (dump_flags & TDF_DETAILS))
2510 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2511 num_iterations++;
2512 changed = false;
2513 for (i = postorder_num - 1 ; i >= 0; i--)
2514 {
2515 if (bitmap_bit_p (changed_blocks, postorder[i]))
2516 {
2517 basic_block block = BASIC_BLOCK (postorder[i]);
2518 changed
2519 |= compute_partial_antic_aux (block,
2520 bitmap_bit_p (has_abnormal_preds,
2521 block->index));
2522 }
2523 }
2524 /* Theoretically possible, but *highly* unlikely. */
2525 gcc_checking_assert (num_iterations < 500);
2526 }
2527 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2528 num_iterations);
2529 }
2530 sbitmap_free (has_abnormal_preds);
2531 sbitmap_free (changed_blocks);
2532 }
2533
2534
2535 /* Inserted expressions are placed onto this worklist, which is used
2536 for performing quick dead code elimination of insertions we made
2537 that didn't turn out to be necessary. */
2538 static bitmap inserted_exprs;
2539
2540 /* The actual worker for create_component_ref_by_pieces. */
2541
2542 static tree
2543 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2544 unsigned int *operand, gimple_seq *stmts)
2545 {
2546 vn_reference_op_t currop = &ref->operands[*operand];
2547 tree genop;
2548 ++*operand;
2549 switch (currop->opcode)
2550 {
2551 case CALL_EXPR:
2552 {
2553 tree folded, sc = NULL_TREE;
2554 unsigned int nargs = 0;
2555 tree fn, *args;
2556 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2557 fn = currop->op0;
2558 else
2559 fn = find_or_generate_expression (block, currop->op0, stmts);
2560 if (!fn)
2561 return NULL_TREE;
2562 if (currop->op1)
2563 {
2564 sc = find_or_generate_expression (block, currop->op1, stmts);
2565 if (!sc)
2566 return NULL_TREE;
2567 }
2568 args = XNEWVEC (tree, ref->operands.length () - 1);
2569 while (*operand < ref->operands.length ())
2570 {
2571 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2572 operand, stmts);
2573 if (!args[nargs])
2574 return NULL_TREE;
2575 nargs++;
2576 }
2577 folded = build_call_array (currop->type,
2578 (TREE_CODE (fn) == FUNCTION_DECL
2579 ? build_fold_addr_expr (fn) : fn),
2580 nargs, args);
2581 free (args);
2582 if (sc)
2583 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2584 return folded;
2585 }
2586
2587 case MEM_REF:
2588 {
2589 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2590 stmts);
2591 if (!baseop)
2592 return NULL_TREE;
2593 tree offset = currop->op0;
2594 if (TREE_CODE (baseop) == ADDR_EXPR
2595 && handled_component_p (TREE_OPERAND (baseop, 0)))
2596 {
2597 HOST_WIDE_INT off;
2598 tree base;
2599 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2600 &off);
2601 gcc_assert (base);
2602 offset = int_const_binop (PLUS_EXPR, offset,
2603 build_int_cst (TREE_TYPE (offset),
2604 off));
2605 baseop = build_fold_addr_expr (base);
2606 }
2607 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2608 }
2609
2610 case TARGET_MEM_REF:
2611 {
2612 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2613 vn_reference_op_t nextop = &ref->operands[++*operand];
2614 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2615 stmts);
2616 if (!baseop)
2617 return NULL_TREE;
2618 if (currop->op0)
2619 {
2620 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2621 if (!genop0)
2622 return NULL_TREE;
2623 }
2624 if (nextop->op0)
2625 {
2626 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2627 if (!genop1)
2628 return NULL_TREE;
2629 }
2630 return build5 (TARGET_MEM_REF, currop->type,
2631 baseop, currop->op2, genop0, currop->op1, genop1);
2632 }
2633
2634 case ADDR_EXPR:
2635 if (currop->op0)
2636 {
2637 gcc_assert (is_gimple_min_invariant (currop->op0));
2638 return currop->op0;
2639 }
2640 /* Fallthrough. */
2641 case REALPART_EXPR:
2642 case IMAGPART_EXPR:
2643 case VIEW_CONVERT_EXPR:
2644 {
2645 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2646 stmts);
2647 if (!genop0)
2648 return NULL_TREE;
2649 return fold_build1 (currop->opcode, currop->type, genop0);
2650 }
2651
2652 case WITH_SIZE_EXPR:
2653 {
2654 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2655 stmts);
2656 if (!genop0)
2657 return NULL_TREE;
2658 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2659 if (!genop1)
2660 return NULL_TREE;
2661 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2662 }
2663
2664 case BIT_FIELD_REF:
2665 {
2666 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2667 stmts);
2668 if (!genop0)
2669 return NULL_TREE;
2670 tree op1 = currop->op0;
2671 tree op2 = currop->op1;
2672 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2673 }
2674
2675 /* For array ref vn_reference_op's, operand 1 of the array ref
2676 is op0 of the reference op and operand 3 of the array ref is
2677 op1. */
2678 case ARRAY_RANGE_REF:
2679 case ARRAY_REF:
2680 {
2681 tree genop0;
2682 tree genop1 = currop->op0;
2683 tree genop2 = currop->op1;
2684 tree genop3 = currop->op2;
2685 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2686 stmts);
2687 if (!genop0)
2688 return NULL_TREE;
2689 genop1 = find_or_generate_expression (block, genop1, stmts);
2690 if (!genop1)
2691 return NULL_TREE;
2692 if (genop2)
2693 {
2694 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2695 /* Drop zero minimum index if redundant. */
2696 if (integer_zerop (genop2)
2697 && (!domain_type
2698 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2699 genop2 = NULL_TREE;
2700 else
2701 {
2702 genop2 = find_or_generate_expression (block, genop2, stmts);
2703 if (!genop2)
2704 return NULL_TREE;
2705 }
2706 }
2707 if (genop3)
2708 {
2709 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2710 /* We can't always put a size in units of the element alignment
2711 here as the element alignment may be not visible. See
2712 PR43783. Simply drop the element size for constant
2713 sizes. */
2714 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2715 genop3 = NULL_TREE;
2716 else
2717 {
2718 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2719 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2720 genop3 = find_or_generate_expression (block, genop3, stmts);
2721 if (!genop3)
2722 return NULL_TREE;
2723 }
2724 }
2725 return build4 (currop->opcode, currop->type, genop0, genop1,
2726 genop2, genop3);
2727 }
2728 case COMPONENT_REF:
2729 {
2730 tree op0;
2731 tree op1;
2732 tree genop2 = currop->op1;
2733 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2734 if (!op0)
2735 return NULL_TREE;
2736 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2737 op1 = currop->op0;
2738 if (genop2)
2739 {
2740 genop2 = find_or_generate_expression (block, genop2, stmts);
2741 if (!genop2)
2742 return NULL_TREE;
2743 }
2744 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2745 }
2746
2747 case SSA_NAME:
2748 {
2749 genop = find_or_generate_expression (block, currop->op0, stmts);
2750 return genop;
2751 }
2752 case STRING_CST:
2753 case INTEGER_CST:
2754 case COMPLEX_CST:
2755 case VECTOR_CST:
2756 case REAL_CST:
2757 case CONSTRUCTOR:
2758 case VAR_DECL:
2759 case PARM_DECL:
2760 case CONST_DECL:
2761 case RESULT_DECL:
2762 case FUNCTION_DECL:
2763 return currop->op0;
2764
2765 default:
2766 gcc_unreachable ();
2767 }
2768 }
2769
2770 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2771 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2772 trying to rename aggregates into ssa form directly, which is a no no.
2773
2774 Thus, this routine doesn't create temporaries, it just builds a
2775 single access expression for the array, calling
2776 find_or_generate_expression to build the innermost pieces.
2777
2778 This function is a subroutine of create_expression_by_pieces, and
2779 should not be called on it's own unless you really know what you
2780 are doing. */
2781
2782 static tree
2783 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2784 gimple_seq *stmts)
2785 {
2786 unsigned int op = 0;
2787 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2788 }
2789
2790 /* Find a simple leader for an expression, or generate one using
2791 create_expression_by_pieces from a NARY expression for the value.
2792 BLOCK is the basic_block we are looking for leaders in.
2793 OP is the tree expression to find a leader for or generate.
2794 Returns the leader or NULL_TREE on failure. */
2795
2796 static tree
2797 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2798 {
2799 pre_expr expr = get_or_alloc_expr_for (op);
2800 unsigned int lookfor = get_expr_value_id (expr);
2801 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2802 if (leader)
2803 {
2804 if (leader->kind == NAME)
2805 return PRE_EXPR_NAME (leader);
2806 else if (leader->kind == CONSTANT)
2807 return PRE_EXPR_CONSTANT (leader);
2808
2809 /* Defer. */
2810 return NULL_TREE;
2811 }
2812
2813 /* It must be a complex expression, so generate it recursively. Note
2814 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2815 where the insert algorithm fails to insert a required expression. */
2816 bitmap exprset = value_expressions[lookfor];
2817 bitmap_iterator bi;
2818 unsigned int i;
2819 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2820 {
2821 pre_expr temp = expression_for_id (i);
2822 /* We cannot insert random REFERENCE expressions at arbitrary
2823 places. We can insert NARYs which eventually re-materializes
2824 its operand values. */
2825 if (temp->kind == NARY)
2826 return create_expression_by_pieces (block, temp, stmts,
2827 get_expr_type (expr));
2828 }
2829
2830 /* Defer. */
2831 return NULL_TREE;
2832 }
2833
2834 #define NECESSARY GF_PLF_1
2835
2836 /* Create an expression in pieces, so that we can handle very complex
2837 expressions that may be ANTIC, but not necessary GIMPLE.
2838 BLOCK is the basic block the expression will be inserted into,
2839 EXPR is the expression to insert (in value form)
2840 STMTS is a statement list to append the necessary insertions into.
2841
2842 This function will die if we hit some value that shouldn't be
2843 ANTIC but is (IE there is no leader for it, or its components).
2844 The function returns NULL_TREE in case a different antic expression
2845 has to be inserted first.
2846 This function may also generate expressions that are themselves
2847 partially or fully redundant. Those that are will be either made
2848 fully redundant during the next iteration of insert (for partially
2849 redundant ones), or eliminated by eliminate (for fully redundant
2850 ones). */
2851
2852 static tree
2853 create_expression_by_pieces (basic_block block, pre_expr expr,
2854 gimple_seq *stmts, tree type)
2855 {
2856 tree name;
2857 tree folded;
2858 gimple_seq forced_stmts = NULL;
2859 unsigned int value_id;
2860 gimple_stmt_iterator gsi;
2861 tree exprtype = type ? type : get_expr_type (expr);
2862 pre_expr nameexpr;
2863 gimple newstmt;
2864
2865 switch (expr->kind)
2866 {
2867 /* We may hit the NAME/CONSTANT case if we have to convert types
2868 that value numbering saw through. */
2869 case NAME:
2870 folded = PRE_EXPR_NAME (expr);
2871 break;
2872 case CONSTANT:
2873 folded = PRE_EXPR_CONSTANT (expr);
2874 break;
2875 case REFERENCE:
2876 {
2877 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2878 folded = create_component_ref_by_pieces (block, ref, stmts);
2879 if (!folded)
2880 return NULL_TREE;
2881 }
2882 break;
2883 case NARY:
2884 {
2885 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2886 tree *genop = XALLOCAVEC (tree, nary->length);
2887 unsigned i;
2888 for (i = 0; i < nary->length; ++i)
2889 {
2890 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2891 if (!genop[i])
2892 return NULL_TREE;
2893 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2894 may have conversions stripped. */
2895 if (nary->opcode == POINTER_PLUS_EXPR)
2896 {
2897 if (i == 0)
2898 genop[i] = fold_convert (nary->type, genop[i]);
2899 else if (i == 1)
2900 genop[i] = convert_to_ptrofftype (genop[i]);
2901 }
2902 else
2903 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2904 }
2905 if (nary->opcode == CONSTRUCTOR)
2906 {
2907 vec<constructor_elt, va_gc> *elts = NULL;
2908 for (i = 0; i < nary->length; ++i)
2909 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2910 folded = build_constructor (nary->type, elts);
2911 }
2912 else
2913 {
2914 switch (nary->length)
2915 {
2916 case 1:
2917 folded = fold_build1 (nary->opcode, nary->type,
2918 genop[0]);
2919 break;
2920 case 2:
2921 folded = fold_build2 (nary->opcode, nary->type,
2922 genop[0], genop[1]);
2923 break;
2924 case 3:
2925 folded = fold_build3 (nary->opcode, nary->type,
2926 genop[0], genop[1], genop[2]);
2927 break;
2928 default:
2929 gcc_unreachable ();
2930 }
2931 }
2932 }
2933 break;
2934 default:
2935 gcc_unreachable ();
2936 }
2937
2938 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2939 folded = fold_convert (exprtype, folded);
2940
2941 /* Force the generated expression to be a sequence of GIMPLE
2942 statements.
2943 We have to call unshare_expr because force_gimple_operand may
2944 modify the tree we pass to it. */
2945 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2946 false, NULL);
2947
2948 /* If we have any intermediate expressions to the value sets, add them
2949 to the value sets and chain them in the instruction stream. */
2950 if (forced_stmts)
2951 {
2952 gsi = gsi_start (forced_stmts);
2953 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2954 {
2955 gimple stmt = gsi_stmt (gsi);
2956 tree forcedname = gimple_get_lhs (stmt);
2957 pre_expr nameexpr;
2958
2959 if (TREE_CODE (forcedname) == SSA_NAME)
2960 {
2961 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2962 VN_INFO_GET (forcedname)->valnum = forcedname;
2963 VN_INFO (forcedname)->value_id = get_next_value_id ();
2964 nameexpr = get_or_alloc_expr_for_name (forcedname);
2965 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2966 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2967 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2968 }
2969 }
2970 gimple_seq_add_seq (stmts, forced_stmts);
2971 }
2972
2973 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2974 newstmt = gimple_build_assign (name, folded);
2975 gimple_set_plf (newstmt, NECESSARY, false);
2976
2977 gimple_seq_add_stmt (stmts, newstmt);
2978 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2979
2980 /* Fold the last statement. */
2981 gsi = gsi_last (*stmts);
2982 if (fold_stmt_inplace (&gsi))
2983 update_stmt (gsi_stmt (gsi));
2984
2985 /* Add a value number to the temporary.
2986 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2987 we are creating the expression by pieces, and this particular piece of
2988 the expression may have been represented. There is no harm in replacing
2989 here. */
2990 value_id = get_expr_value_id (expr);
2991 VN_INFO_GET (name)->value_id = value_id;
2992 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2993 if (VN_INFO (name)->valnum == NULL_TREE)
2994 VN_INFO (name)->valnum = name;
2995 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2996 nameexpr = get_or_alloc_expr_for_name (name);
2997 add_to_value (value_id, nameexpr);
2998 if (NEW_SETS (block))
2999 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3000 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3001
3002 pre_stats.insertions++;
3003 if (dump_file && (dump_flags & TDF_DETAILS))
3004 {
3005 fprintf (dump_file, "Inserted ");
3006 print_gimple_stmt (dump_file, newstmt, 0, 0);
3007 fprintf (dump_file, " in predecessor %d (%04d)\n",
3008 block->index, value_id);
3009 }
3010
3011 return name;
3012 }
3013
3014
3015 /* Returns true if we want to inhibit the insertions of PHI nodes
3016 for the given EXPR for basic block BB (a member of a loop).
3017 We want to do this, when we fear that the induction variable we
3018 create might inhibit vectorization. */
3019
3020 static bool
3021 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3022 {
3023 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3024 vec<vn_reference_op_s> ops = vr->operands;
3025 vn_reference_op_t op;
3026 unsigned i;
3027
3028 /* If we aren't going to vectorize we don't inhibit anything. */
3029 if (!flag_tree_loop_vectorize)
3030 return false;
3031
3032 /* Otherwise we inhibit the insertion when the address of the
3033 memory reference is a simple induction variable. In other
3034 cases the vectorizer won't do anything anyway (either it's
3035 loop invariant or a complicated expression). */
3036 FOR_EACH_VEC_ELT (ops, i, op)
3037 {
3038 switch (op->opcode)
3039 {
3040 case CALL_EXPR:
3041 /* Calls are not a problem. */
3042 return false;
3043
3044 case ARRAY_REF:
3045 case ARRAY_RANGE_REF:
3046 if (TREE_CODE (op->op0) != SSA_NAME)
3047 break;
3048 /* Fallthru. */
3049 case SSA_NAME:
3050 {
3051 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3052 affine_iv iv;
3053 /* Default defs are loop invariant. */
3054 if (!defbb)
3055 break;
3056 /* Defined outside this loop, also loop invariant. */
3057 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3058 break;
3059 /* If it's a simple induction variable inhibit insertion,
3060 the vectorizer might be interested in this one. */
3061 if (simple_iv (bb->loop_father, bb->loop_father,
3062 op->op0, &iv, true))
3063 return true;
3064 /* No simple IV, vectorizer can't do anything, hence no
3065 reason to inhibit the transformation for this operand. */
3066 break;
3067 }
3068 default:
3069 break;
3070 }
3071 }
3072 return false;
3073 }
3074
3075 /* Insert the to-be-made-available values of expression EXPRNUM for each
3076 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3077 merge the result with a phi node, given the same value number as
3078 NODE. Return true if we have inserted new stuff. */
3079
3080 static bool
3081 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3082 vec<pre_expr> avail)
3083 {
3084 pre_expr expr = expression_for_id (exprnum);
3085 pre_expr newphi;
3086 unsigned int val = get_expr_value_id (expr);
3087 edge pred;
3088 bool insertions = false;
3089 bool nophi = false;
3090 basic_block bprime;
3091 pre_expr eprime;
3092 edge_iterator ei;
3093 tree type = get_expr_type (expr);
3094 tree temp;
3095 gimple phi;
3096
3097 /* Make sure we aren't creating an induction variable. */
3098 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3099 {
3100 bool firstinsideloop = false;
3101 bool secondinsideloop = false;
3102 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3103 EDGE_PRED (block, 0)->src);
3104 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3105 EDGE_PRED (block, 1)->src);
3106 /* Induction variables only have one edge inside the loop. */
3107 if ((firstinsideloop ^ secondinsideloop)
3108 && (expr->kind != REFERENCE
3109 || inhibit_phi_insertion (block, expr)))
3110 {
3111 if (dump_file && (dump_flags & TDF_DETAILS))
3112 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3113 nophi = true;
3114 }
3115 }
3116
3117 /* Make the necessary insertions. */
3118 FOR_EACH_EDGE (pred, ei, block->preds)
3119 {
3120 gimple_seq stmts = NULL;
3121 tree builtexpr;
3122 bprime = pred->src;
3123 eprime = avail[pred->dest_idx];
3124
3125 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3126 {
3127 builtexpr = create_expression_by_pieces (bprime, eprime,
3128 &stmts, type);
3129 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3130 gsi_insert_seq_on_edge (pred, stmts);
3131 if (!builtexpr)
3132 {
3133 /* We cannot insert a PHI node if we failed to insert
3134 on one edge. */
3135 nophi = true;
3136 continue;
3137 }
3138 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3139 insertions = true;
3140 }
3141 else if (eprime->kind == CONSTANT)
3142 {
3143 /* Constants may not have the right type, fold_convert
3144 should give us back a constant with the right type. */
3145 tree constant = PRE_EXPR_CONSTANT (eprime);
3146 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3147 {
3148 tree builtexpr = fold_convert (type, constant);
3149 if (!is_gimple_min_invariant (builtexpr))
3150 {
3151 tree forcedexpr = force_gimple_operand (builtexpr,
3152 &stmts, true,
3153 NULL);
3154 if (!is_gimple_min_invariant (forcedexpr))
3155 {
3156 if (forcedexpr != builtexpr)
3157 {
3158 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3159 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3160 }
3161 if (stmts)
3162 {
3163 gimple_stmt_iterator gsi;
3164 gsi = gsi_start (stmts);
3165 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3166 {
3167 gimple stmt = gsi_stmt (gsi);
3168 tree lhs = gimple_get_lhs (stmt);
3169 if (TREE_CODE (lhs) == SSA_NAME)
3170 bitmap_set_bit (inserted_exprs,
3171 SSA_NAME_VERSION (lhs));
3172 gimple_set_plf (stmt, NECESSARY, false);
3173 }
3174 gsi_insert_seq_on_edge (pred, stmts);
3175 }
3176 avail[pred->dest_idx]
3177 = get_or_alloc_expr_for_name (forcedexpr);
3178 }
3179 }
3180 else
3181 avail[pred->dest_idx]
3182 = get_or_alloc_expr_for_constant (builtexpr);
3183 }
3184 }
3185 else if (eprime->kind == NAME)
3186 {
3187 /* We may have to do a conversion because our value
3188 numbering can look through types in certain cases, but
3189 our IL requires all operands of a phi node have the same
3190 type. */
3191 tree name = PRE_EXPR_NAME (eprime);
3192 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3193 {
3194 tree builtexpr;
3195 tree forcedexpr;
3196 builtexpr = fold_convert (type, name);
3197 forcedexpr = force_gimple_operand (builtexpr,
3198 &stmts, true,
3199 NULL);
3200
3201 if (forcedexpr != name)
3202 {
3203 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3204 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3205 }
3206
3207 if (stmts)
3208 {
3209 gimple_stmt_iterator gsi;
3210 gsi = gsi_start (stmts);
3211 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3212 {
3213 gimple stmt = gsi_stmt (gsi);
3214 tree lhs = gimple_get_lhs (stmt);
3215 if (TREE_CODE (lhs) == SSA_NAME)
3216 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3217 gimple_set_plf (stmt, NECESSARY, false);
3218 }
3219 gsi_insert_seq_on_edge (pred, stmts);
3220 }
3221 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3222 }
3223 }
3224 }
3225 /* If we didn't want a phi node, and we made insertions, we still have
3226 inserted new stuff, and thus return true. If we didn't want a phi node,
3227 and didn't make insertions, we haven't added anything new, so return
3228 false. */
3229 if (nophi && insertions)
3230 return true;
3231 else if (nophi && !insertions)
3232 return false;
3233
3234 /* Now build a phi for the new variable. */
3235 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3236 phi = create_phi_node (temp, block);
3237
3238 gimple_set_plf (phi, NECESSARY, false);
3239 VN_INFO_GET (temp)->value_id = val;
3240 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3241 if (VN_INFO (temp)->valnum == NULL_TREE)
3242 VN_INFO (temp)->valnum = temp;
3243 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3244 FOR_EACH_EDGE (pred, ei, block->preds)
3245 {
3246 pre_expr ae = avail[pred->dest_idx];
3247 gcc_assert (get_expr_type (ae) == type
3248 || useless_type_conversion_p (type, get_expr_type (ae)));
3249 if (ae->kind == CONSTANT)
3250 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3251 pred, UNKNOWN_LOCATION);
3252 else
3253 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3254 }
3255
3256 newphi = get_or_alloc_expr_for_name (temp);
3257 add_to_value (val, newphi);
3258
3259 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3260 this insertion, since we test for the existence of this value in PHI_GEN
3261 before proceeding with the partial redundancy checks in insert_aux.
3262
3263 The value may exist in AVAIL_OUT, in particular, it could be represented
3264 by the expression we are trying to eliminate, in which case we want the
3265 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3266 inserted there.
3267
3268 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3269 this block, because if it did, it would have existed in our dominator's
3270 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3271 */
3272
3273 bitmap_insert_into_set (PHI_GEN (block), newphi);
3274 bitmap_value_replace_in_set (AVAIL_OUT (block),
3275 newphi);
3276 bitmap_insert_into_set (NEW_SETS (block),
3277 newphi);
3278
3279 if (dump_file && (dump_flags & TDF_DETAILS))
3280 {
3281 fprintf (dump_file, "Created phi ");
3282 print_gimple_stmt (dump_file, phi, 0, 0);
3283 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3284 }
3285 pre_stats.phis++;
3286 return true;
3287 }
3288
3289
3290
3291 /* Perform insertion of partially redundant values.
3292 For BLOCK, do the following:
3293 1. Propagate the NEW_SETS of the dominator into the current block.
3294 If the block has multiple predecessors,
3295 2a. Iterate over the ANTIC expressions for the block to see if
3296 any of them are partially redundant.
3297 2b. If so, insert them into the necessary predecessors to make
3298 the expression fully redundant.
3299 2c. Insert a new PHI merging the values of the predecessors.
3300 2d. Insert the new PHI, and the new expressions, into the
3301 NEW_SETS set.
3302 3. Recursively call ourselves on the dominator children of BLOCK.
3303
3304 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3305 do_regular_insertion and do_partial_insertion.
3306
3307 */
3308
3309 static bool
3310 do_regular_insertion (basic_block block, basic_block dom)
3311 {
3312 bool new_stuff = false;
3313 vec<pre_expr> exprs;
3314 pre_expr expr;
3315 vec<pre_expr> avail = vNULL;
3316 int i;
3317
3318 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3319 avail.safe_grow (EDGE_COUNT (block->preds));
3320
3321 FOR_EACH_VEC_ELT (exprs, i, expr)
3322 {
3323 if (expr->kind == NARY
3324 || expr->kind == REFERENCE)
3325 {
3326 unsigned int val;
3327 bool by_some = false;
3328 bool cant_insert = false;
3329 bool all_same = true;
3330 pre_expr first_s = NULL;
3331 edge pred;
3332 basic_block bprime;
3333 pre_expr eprime = NULL;
3334 edge_iterator ei;
3335 pre_expr edoubleprime = NULL;
3336 bool do_insertion = false;
3337
3338 val = get_expr_value_id (expr);
3339 if (bitmap_set_contains_value (PHI_GEN (block), val))
3340 continue;
3341 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3342 {
3343 if (dump_file && (dump_flags & TDF_DETAILS))
3344 {
3345 fprintf (dump_file, "Found fully redundant value: ");
3346 print_pre_expr (dump_file, expr);
3347 fprintf (dump_file, "\n");
3348 }
3349 continue;
3350 }
3351
3352 FOR_EACH_EDGE (pred, ei, block->preds)
3353 {
3354 unsigned int vprime;
3355
3356 /* We should never run insertion for the exit block
3357 and so not come across fake pred edges. */
3358 gcc_assert (!(pred->flags & EDGE_FAKE));
3359 bprime = pred->src;
3360 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3361 bprime, block);
3362
3363 /* eprime will generally only be NULL if the
3364 value of the expression, translated
3365 through the PHI for this predecessor, is
3366 undefined. If that is the case, we can't
3367 make the expression fully redundant,
3368 because its value is undefined along a
3369 predecessor path. We can thus break out
3370 early because it doesn't matter what the
3371 rest of the results are. */
3372 if (eprime == NULL)
3373 {
3374 avail[pred->dest_idx] = NULL;
3375 cant_insert = true;
3376 break;
3377 }
3378
3379 eprime = fully_constant_expression (eprime);
3380 vprime = get_expr_value_id (eprime);
3381 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3382 vprime);
3383 if (edoubleprime == NULL)
3384 {
3385 avail[pred->dest_idx] = eprime;
3386 all_same = false;
3387 }
3388 else
3389 {
3390 avail[pred->dest_idx] = edoubleprime;
3391 by_some = true;
3392 /* We want to perform insertions to remove a redundancy on
3393 a path in the CFG we want to optimize for speed. */
3394 if (optimize_edge_for_speed_p (pred))
3395 do_insertion = true;
3396 if (first_s == NULL)
3397 first_s = edoubleprime;
3398 else if (!pre_expr_d::equal (first_s, edoubleprime))
3399 all_same = false;
3400 }
3401 }
3402 /* If we can insert it, it's not the same value
3403 already existing along every predecessor, and
3404 it's defined by some predecessor, it is
3405 partially redundant. */
3406 if (!cant_insert && !all_same && by_some)
3407 {
3408 if (!do_insertion)
3409 {
3410 if (dump_file && (dump_flags & TDF_DETAILS))
3411 {
3412 fprintf (dump_file, "Skipping partial redundancy for "
3413 "expression ");
3414 print_pre_expr (dump_file, expr);
3415 fprintf (dump_file, " (%04d), no redundancy on to be "
3416 "optimized for speed edge\n", val);
3417 }
3418 }
3419 else if (dbg_cnt (treepre_insert))
3420 {
3421 if (dump_file && (dump_flags & TDF_DETAILS))
3422 {
3423 fprintf (dump_file, "Found partial redundancy for "
3424 "expression ");
3425 print_pre_expr (dump_file, expr);
3426 fprintf (dump_file, " (%04d)\n",
3427 get_expr_value_id (expr));
3428 }
3429 if (insert_into_preds_of_block (block,
3430 get_expression_id (expr),
3431 avail))
3432 new_stuff = true;
3433 }
3434 }
3435 /* If all edges produce the same value and that value is
3436 an invariant, then the PHI has the same value on all
3437 edges. Note this. */
3438 else if (!cant_insert && all_same)
3439 {
3440 gcc_assert (edoubleprime->kind == CONSTANT
3441 || edoubleprime->kind == NAME);
3442
3443 tree temp = make_temp_ssa_name (get_expr_type (expr),
3444 NULL, "pretmp");
3445 gimple assign = gimple_build_assign (temp,
3446 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3447 gimple_stmt_iterator gsi = gsi_after_labels (block);
3448 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3449
3450 gimple_set_plf (assign, NECESSARY, false);
3451 VN_INFO_GET (temp)->value_id = val;
3452 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3453 if (VN_INFO (temp)->valnum == NULL_TREE)
3454 VN_INFO (temp)->valnum = temp;
3455 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3456 pre_expr newe = get_or_alloc_expr_for_name (temp);
3457 add_to_value (val, newe);
3458 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3459 bitmap_insert_into_set (NEW_SETS (block), newe);
3460 }
3461 }
3462 }
3463
3464 exprs.release ();
3465 avail.release ();
3466 return new_stuff;
3467 }
3468
3469
3470 /* Perform insertion for partially anticipatable expressions. There
3471 is only one case we will perform insertion for these. This case is
3472 if the expression is partially anticipatable, and fully available.
3473 In this case, we know that putting it earlier will enable us to
3474 remove the later computation. */
3475
3476
3477 static bool
3478 do_partial_partial_insertion (basic_block block, basic_block dom)
3479 {
3480 bool new_stuff = false;
3481 vec<pre_expr> exprs;
3482 pre_expr expr;
3483 vec<pre_expr> avail = vNULL;
3484 int i;
3485
3486 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3487 avail.safe_grow (EDGE_COUNT (block->preds));
3488
3489 FOR_EACH_VEC_ELT (exprs, i, expr)
3490 {
3491 if (expr->kind == NARY
3492 || expr->kind == REFERENCE)
3493 {
3494 unsigned int val;
3495 bool by_all = true;
3496 bool cant_insert = false;
3497 edge pred;
3498 basic_block bprime;
3499 pre_expr eprime = NULL;
3500 edge_iterator ei;
3501
3502 val = get_expr_value_id (expr);
3503 if (bitmap_set_contains_value (PHI_GEN (block), val))
3504 continue;
3505 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3506 continue;
3507
3508 FOR_EACH_EDGE (pred, ei, block->preds)
3509 {
3510 unsigned int vprime;
3511 pre_expr edoubleprime;
3512
3513 /* We should never run insertion for the exit block
3514 and so not come across fake pred edges. */
3515 gcc_assert (!(pred->flags & EDGE_FAKE));
3516 bprime = pred->src;
3517 eprime = phi_translate (expr, ANTIC_IN (block),
3518 PA_IN (block),
3519 bprime, block);
3520
3521 /* eprime will generally only be NULL if the
3522 value of the expression, translated
3523 through the PHI for this predecessor, is
3524 undefined. If that is the case, we can't
3525 make the expression fully redundant,
3526 because its value is undefined along a
3527 predecessor path. We can thus break out
3528 early because it doesn't matter what the
3529 rest of the results are. */
3530 if (eprime == NULL)
3531 {
3532 avail[pred->dest_idx] = NULL;
3533 cant_insert = true;
3534 break;
3535 }
3536
3537 eprime = fully_constant_expression (eprime);
3538 vprime = get_expr_value_id (eprime);
3539 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3540 avail[pred->dest_idx] = edoubleprime;
3541 if (edoubleprime == NULL)
3542 {
3543 by_all = false;
3544 break;
3545 }
3546 }
3547
3548 /* If we can insert it, it's not the same value
3549 already existing along every predecessor, and
3550 it's defined by some predecessor, it is
3551 partially redundant. */
3552 if (!cant_insert && by_all)
3553 {
3554 edge succ;
3555 bool do_insertion = false;
3556
3557 /* Insert only if we can remove a later expression on a path
3558 that we want to optimize for speed.
3559 The phi node that we will be inserting in BLOCK is not free,
3560 and inserting it for the sake of !optimize_for_speed successor
3561 may cause regressions on the speed path. */
3562 FOR_EACH_EDGE (succ, ei, block->succs)
3563 {
3564 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3565 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3566 {
3567 if (optimize_edge_for_speed_p (succ))
3568 do_insertion = true;
3569 }
3570 }
3571
3572 if (!do_insertion)
3573 {
3574 if (dump_file && (dump_flags & TDF_DETAILS))
3575 {
3576 fprintf (dump_file, "Skipping partial partial redundancy "
3577 "for expression ");
3578 print_pre_expr (dump_file, expr);
3579 fprintf (dump_file, " (%04d), not (partially) anticipated "
3580 "on any to be optimized for speed edges\n", val);
3581 }
3582 }
3583 else if (dbg_cnt (treepre_insert))
3584 {
3585 pre_stats.pa_insert++;
3586 if (dump_file && (dump_flags & TDF_DETAILS))
3587 {
3588 fprintf (dump_file, "Found partial partial redundancy "
3589 "for expression ");
3590 print_pre_expr (dump_file, expr);
3591 fprintf (dump_file, " (%04d)\n",
3592 get_expr_value_id (expr));
3593 }
3594 if (insert_into_preds_of_block (block,
3595 get_expression_id (expr),
3596 avail))
3597 new_stuff = true;
3598 }
3599 }
3600 }
3601 }
3602
3603 exprs.release ();
3604 avail.release ();
3605 return new_stuff;
3606 }
3607
3608 static bool
3609 insert_aux (basic_block block)
3610 {
3611 basic_block son;
3612 bool new_stuff = false;
3613
3614 if (block)
3615 {
3616 basic_block dom;
3617 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3618 if (dom)
3619 {
3620 unsigned i;
3621 bitmap_iterator bi;
3622 bitmap_set_t newset = NEW_SETS (dom);
3623 if (newset)
3624 {
3625 /* Note that we need to value_replace both NEW_SETS, and
3626 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3627 represented by some non-simple expression here that we want
3628 to replace it with. */
3629 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3630 {
3631 pre_expr expr = expression_for_id (i);
3632 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3633 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3634 }
3635 }
3636 if (!single_pred_p (block))
3637 {
3638 new_stuff |= do_regular_insertion (block, dom);
3639 if (do_partial_partial)
3640 new_stuff |= do_partial_partial_insertion (block, dom);
3641 }
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);
3649 }
3650
3651 return new_stuff;
3652 }
3653
3654 /* Perform insertion of partially redundant 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 (bb)
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));
3672
3673 /* Clear the NEW sets before the next iteration. We have already
3674 fully propagated its contents. */
3675 if (new_stuff)
3676 FOR_ALL_BB (bb)
3677 bitmap_set_free (NEW_SETS (bb));
3678 }
3679 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3680 }
3681
3682
3683 /* Compute the AVAIL set for all basic blocks.
3684
3685 This function performs value numbering of the statements in each basic
3686 block. The AVAIL sets are built from information we glean while doing
3687 this value numbering, since the AVAIL sets contain only one entry per
3688 value.
3689
3690 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3691 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3692
3693 static void
3694 compute_avail (void)
3695 {
3696
3697 basic_block block, son;
3698 basic_block *worklist;
3699 size_t sp = 0;
3700 unsigned i;
3701
3702 /* We pretend that default definitions are defined in the entry block.
3703 This includes function arguments and the static chain decl. */
3704 for (i = 1; i < num_ssa_names; ++i)
3705 {
3706 tree name = ssa_name (i);
3707 pre_expr e;
3708 if (!name
3709 || !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 /* Loop until the worklist is empty. */
3740 while (sp)
3741 {
3742 gimple_stmt_iterator gsi;
3743 gimple stmt;
3744 basic_block dom;
3745
3746 /* Pick a block from the worklist. */
3747 block = worklist[--sp];
3748
3749 /* Initially, the set of available values in BLOCK is that of
3750 its immediate dominator. */
3751 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3752 if (dom)
3753 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3754
3755 /* Generate values for PHI nodes. */
3756 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3757 {
3758 tree result = gimple_phi_result (gsi_stmt (gsi));
3759
3760 /* We have no need for virtual phis, as they don't represent
3761 actual computations. */
3762 if (virtual_operand_p (result))
3763 continue;
3764
3765 pre_expr e = get_or_alloc_expr_for_name (result);
3766 add_to_value (get_expr_value_id (e), e);
3767 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3768 bitmap_insert_into_set (PHI_GEN (block), e);
3769 }
3770
3771 BB_MAY_NOTRETURN (block) = 0;
3772
3773 /* Now compute value numbers and populate value sets with all
3774 the expressions computed in BLOCK. */
3775 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3776 {
3777 ssa_op_iter iter;
3778 tree op;
3779
3780 stmt = gsi_stmt (gsi);
3781
3782 /* Cache whether the basic-block has any non-visible side-effect
3783 or control flow.
3784 If this isn't a call or it is the last stmt in the
3785 basic-block then the CFG represents things correctly. */
3786 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3787 {
3788 /* Non-looping const functions always return normally.
3789 Otherwise the call might not return or have side-effects
3790 that forbids hoisting possibly trapping expressions
3791 before it. */
3792 int flags = gimple_call_flags (stmt);
3793 if (!(flags & ECF_CONST)
3794 || (flags & ECF_LOOPING_CONST_OR_PURE))
3795 BB_MAY_NOTRETURN (block) = 1;
3796 }
3797
3798 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3799 {
3800 pre_expr e = get_or_alloc_expr_for_name (op);
3801
3802 add_to_value (get_expr_value_id (e), e);
3803 bitmap_insert_into_set (TMP_GEN (block), e);
3804 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3805 }
3806
3807 if (gimple_has_side_effects (stmt)
3808 || stmt_could_throw_p (stmt)
3809 || is_gimple_debug (stmt))
3810 continue;
3811
3812 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3813 {
3814 if (ssa_undefined_value_p (op))
3815 continue;
3816 pre_expr e = get_or_alloc_expr_for_name (op);
3817 bitmap_value_insert_into_set (EXP_GEN (block), e);
3818 }
3819
3820 switch (gimple_code (stmt))
3821 {
3822 case GIMPLE_RETURN:
3823 continue;
3824
3825 case GIMPLE_CALL:
3826 {
3827 vn_reference_t ref;
3828 pre_expr result = NULL;
3829 vec<vn_reference_op_s> ops = vNULL;
3830
3831 /* We can value number only calls to real functions. */
3832 if (gimple_call_internal_p (stmt))
3833 continue;
3834
3835 copy_reference_ops_from_call (stmt, &ops);
3836 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3837 gimple_expr_type (stmt),
3838 ops, &ref, VN_NOWALK);
3839 ops.release ();
3840 if (!ref)
3841 continue;
3842
3843 /* If the value of the call is not invalidated in
3844 this block until it is computed, add the expression
3845 to EXP_GEN. */
3846 if (!gimple_vuse (stmt)
3847 || gimple_code
3848 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3849 || gimple_bb (SSA_NAME_DEF_STMT
3850 (gimple_vuse (stmt))) != block)
3851 {
3852 result = (pre_expr) pool_alloc (pre_expr_pool);
3853 result->kind = REFERENCE;
3854 result->id = 0;
3855 PRE_EXPR_REFERENCE (result) = ref;
3856
3857 get_or_alloc_expression_id (result);
3858 add_to_value (get_expr_value_id (result), result);
3859 bitmap_value_insert_into_set (EXP_GEN (block), result);
3860 }
3861 continue;
3862 }
3863
3864 case GIMPLE_ASSIGN:
3865 {
3866 pre_expr result = NULL;
3867 switch (vn_get_stmt_kind (stmt))
3868 {
3869 case VN_NARY:
3870 {
3871 enum tree_code code = gimple_assign_rhs_code (stmt);
3872 vn_nary_op_t nary;
3873
3874 /* COND_EXPR and VEC_COND_EXPR are awkward in
3875 that they contain an embedded complex expression.
3876 Don't even try to shove those through PRE. */
3877 if (code == COND_EXPR
3878 || code == VEC_COND_EXPR)
3879 continue;
3880
3881 vn_nary_op_lookup_stmt (stmt, &nary);
3882 if (!nary)
3883 continue;
3884
3885 /* If the NARY traps and there was a preceding
3886 point in the block that might not return avoid
3887 adding the nary to EXP_GEN. */
3888 if (BB_MAY_NOTRETURN (block)
3889 && vn_nary_may_trap (nary))
3890 continue;
3891
3892 result = (pre_expr) pool_alloc (pre_expr_pool);
3893 result->kind = NARY;
3894 result->id = 0;
3895 PRE_EXPR_NARY (result) = nary;
3896 break;
3897 }
3898
3899 case VN_REFERENCE:
3900 {
3901 vn_reference_t ref;
3902 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3903 gimple_vuse (stmt),
3904 VN_WALK, &ref);
3905 if (!ref)
3906 continue;
3907
3908 /* If the value of the reference is not invalidated in
3909 this block until it is computed, add the expression
3910 to EXP_GEN. */
3911 if (gimple_vuse (stmt))
3912 {
3913 gimple def_stmt;
3914 bool ok = true;
3915 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3916 while (!gimple_nop_p (def_stmt)
3917 && gimple_code (def_stmt) != GIMPLE_PHI
3918 && gimple_bb (def_stmt) == block)
3919 {
3920 if (stmt_may_clobber_ref_p
3921 (def_stmt, gimple_assign_rhs1 (stmt)))
3922 {
3923 ok = false;
3924 break;
3925 }
3926 def_stmt
3927 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3928 }
3929 if (!ok)
3930 continue;
3931 }
3932
3933 result = (pre_expr) pool_alloc (pre_expr_pool);
3934 result->kind = REFERENCE;
3935 result->id = 0;
3936 PRE_EXPR_REFERENCE (result) = ref;
3937 break;
3938 }
3939
3940 default:
3941 continue;
3942 }
3943
3944 get_or_alloc_expression_id (result);
3945 add_to_value (get_expr_value_id (result), result);
3946 bitmap_value_insert_into_set (EXP_GEN (block), result);
3947 continue;
3948 }
3949 default:
3950 break;
3951 }
3952 }
3953
3954 if (dump_file && (dump_flags & TDF_DETAILS))
3955 {
3956 print_bitmap_set (dump_file, EXP_GEN (block),
3957 "exp_gen", block->index);
3958 print_bitmap_set (dump_file, PHI_GEN (block),
3959 "phi_gen", block->index);
3960 print_bitmap_set (dump_file, TMP_GEN (block),
3961 "tmp_gen", block->index);
3962 print_bitmap_set (dump_file, AVAIL_OUT (block),
3963 "avail_out", block->index);
3964 }
3965
3966 /* Put the dominator children of BLOCK on the worklist of blocks
3967 to compute available sets for. */
3968 for (son = first_dom_son (CDI_DOMINATORS, block);
3969 son;
3970 son = next_dom_son (CDI_DOMINATORS, son))
3971 worklist[sp++] = son;
3972 }
3973
3974 free (worklist);
3975 }
3976
3977
3978 /* Local state for the eliminate domwalk. */
3979 static vec<gimple> el_to_remove;
3980 static vec<gimple> el_to_update;
3981 static unsigned int el_todo;
3982 static vec<tree> el_avail;
3983 static vec<tree> el_avail_stack;
3984
3985 /* Return a leader for OP that is available at the current point of the
3986 eliminate domwalk. */
3987
3988 static tree
3989 eliminate_avail (tree op)
3990 {
3991 tree valnum = VN_INFO (op)->valnum;
3992 if (TREE_CODE (valnum) == SSA_NAME)
3993 {
3994 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3995 return valnum;
3996 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3997 return el_avail[SSA_NAME_VERSION (valnum)];
3998 }
3999 else if (is_gimple_min_invariant (valnum))
4000 return valnum;
4001 return NULL_TREE;
4002 }
4003
4004 /* At the current point of the eliminate domwalk make OP available. */
4005
4006 static void
4007 eliminate_push_avail (tree op)
4008 {
4009 tree valnum = VN_INFO (op)->valnum;
4010 if (TREE_CODE (valnum) == SSA_NAME)
4011 {
4012 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4013 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4014 el_avail[SSA_NAME_VERSION (valnum)] = op;
4015 el_avail_stack.safe_push (op);
4016 }
4017 }
4018
4019 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4020 the leader for the expression if insertion was successful. */
4021
4022 static tree
4023 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4024 {
4025 tree expr = vn_get_expr_for (val);
4026 if (!CONVERT_EXPR_P (expr)
4027 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
4028 return NULL_TREE;
4029
4030 tree op = TREE_OPERAND (expr, 0);
4031 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4032 if (!leader)
4033 return NULL_TREE;
4034
4035 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
4036 gimple tem = gimple_build_assign (res,
4037 fold_build1 (TREE_CODE (expr),
4038 TREE_TYPE (expr), leader));
4039 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
4040 VN_INFO_GET (res)->valnum = val;
4041
4042 if (TREE_CODE (leader) == SSA_NAME)
4043 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4044
4045 pre_stats.insertions++;
4046 if (dump_file && (dump_flags & TDF_DETAILS))
4047 {
4048 fprintf (dump_file, "Inserted ");
4049 print_gimple_stmt (dump_file, tem, 0, 0);
4050 }
4051
4052 return res;
4053 }
4054
4055 class eliminate_dom_walker : public dom_walker
4056 {
4057 public:
4058 eliminate_dom_walker (cdi_direction direction) : dom_walker (direction) {}
4059
4060 virtual void before_dom_children (basic_block);
4061 virtual void after_dom_children (basic_block);
4062 };
4063
4064 /* Perform elimination for the basic-block B during the domwalk. */
4065
4066 void
4067 eliminate_dom_walker::before_dom_children (basic_block b)
4068 {
4069 gimple_stmt_iterator gsi;
4070 gimple stmt;
4071
4072 /* Mark new bb. */
4073 el_avail_stack.safe_push (NULL_TREE);
4074
4075 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4076 {
4077 gimple stmt, phi = gsi_stmt (gsi);
4078 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4079 gimple_stmt_iterator gsi2;
4080
4081 /* We want to perform redundant PHI elimination. Do so by
4082 replacing the PHI with a single copy if possible.
4083 Do not touch inserted, single-argument or virtual PHIs. */
4084 if (gimple_phi_num_args (phi) == 1
4085 || virtual_operand_p (res))
4086 {
4087 gsi_next (&gsi);
4088 continue;
4089 }
4090
4091 sprime = eliminate_avail (res);
4092 if (!sprime
4093 || sprime == res)
4094 {
4095 eliminate_push_avail (res);
4096 gsi_next (&gsi);
4097 continue;
4098 }
4099 else if (is_gimple_min_invariant (sprime))
4100 {
4101 if (!useless_type_conversion_p (TREE_TYPE (res),
4102 TREE_TYPE (sprime)))
4103 sprime = fold_convert (TREE_TYPE (res), sprime);
4104 }
4105
4106 if (dump_file && (dump_flags & TDF_DETAILS))
4107 {
4108 fprintf (dump_file, "Replaced redundant PHI node defining ");
4109 print_generic_expr (dump_file, res, 0);
4110 fprintf (dump_file, " with ");
4111 print_generic_expr (dump_file, sprime, 0);
4112 fprintf (dump_file, "\n");
4113 }
4114
4115 remove_phi_node (&gsi, false);
4116
4117 if (inserted_exprs
4118 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4119 && TREE_CODE (sprime) == SSA_NAME)
4120 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4121
4122 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4123 sprime = fold_convert (TREE_TYPE (res), sprime);
4124 stmt = gimple_build_assign (res, sprime);
4125 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4126
4127 gsi2 = gsi_after_labels (b);
4128 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4129 /* Queue the copy for eventual removal. */
4130 el_to_remove.safe_push (stmt);
4131 /* If we inserted this PHI node ourself, it's not an elimination. */
4132 if (inserted_exprs
4133 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4134 pre_stats.phis--;
4135 else
4136 pre_stats.eliminations++;
4137 }
4138
4139 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4140 {
4141 tree lhs = NULL_TREE;
4142 tree rhs = NULL_TREE;
4143
4144 stmt = gsi_stmt (gsi);
4145
4146 if (gimple_has_lhs (stmt))
4147 lhs = gimple_get_lhs (stmt);
4148
4149 if (gimple_assign_single_p (stmt))
4150 rhs = gimple_assign_rhs1 (stmt);
4151
4152 /* Lookup the RHS of the expression, see if we have an
4153 available computation for it. If so, replace the RHS with
4154 the available computation. */
4155 if (gimple_has_lhs (stmt)
4156 && TREE_CODE (lhs) == SSA_NAME
4157 && !gimple_has_volatile_ops (stmt))
4158 {
4159 tree sprime;
4160 gimple orig_stmt = stmt;
4161
4162 sprime = eliminate_avail (lhs);
4163 /* If there is no usable leader mark lhs as leader for its value. */
4164 if (!sprime)
4165 eliminate_push_avail (lhs);
4166
4167 /* See PR43491. Do not replace a global register variable when
4168 it is a the RHS of an assignment. Do replace local register
4169 variables since gcc does not guarantee a local variable will
4170 be allocated in register.
4171 Do not perform copy propagation or undo constant propagation. */
4172 if (gimple_assign_single_p (stmt)
4173 && (TREE_CODE (rhs) == SSA_NAME
4174 || is_gimple_min_invariant (rhs)
4175 || (TREE_CODE (rhs) == VAR_DECL
4176 && is_global_var (rhs)
4177 && DECL_HARD_REGISTER (rhs))))
4178 continue;
4179
4180 if (!sprime)
4181 {
4182 /* If there is no existing usable leader but SCCVN thinks
4183 it has an expression it wants to use as replacement,
4184 insert that. */
4185 tree val = VN_INFO (lhs)->valnum;
4186 if (val != VN_TOP
4187 && TREE_CODE (val) == SSA_NAME
4188 && VN_INFO (val)->needs_insertion
4189 && VN_INFO (val)->expr != NULL_TREE
4190 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4191 eliminate_push_avail (sprime);
4192 }
4193 else if (is_gimple_min_invariant (sprime))
4194 {
4195 /* If there is no existing leader but SCCVN knows this
4196 value is constant, use that constant. */
4197 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4198 TREE_TYPE (sprime)))
4199 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4200
4201 if (dump_file && (dump_flags & TDF_DETAILS))
4202 {
4203 fprintf (dump_file, "Replaced ");
4204 print_gimple_expr (dump_file, stmt, 0, 0);
4205 fprintf (dump_file, " with ");
4206 print_generic_expr (dump_file, sprime, 0);
4207 fprintf (dump_file, " in ");
4208 print_gimple_stmt (dump_file, stmt, 0, 0);
4209 }
4210 pre_stats.eliminations++;
4211 propagate_tree_value_into_stmt (&gsi, sprime);
4212 stmt = gsi_stmt (gsi);
4213 update_stmt (stmt);
4214
4215 /* If we removed EH side-effects from the statement, clean
4216 its EH information. */
4217 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4218 {
4219 bitmap_set_bit (need_eh_cleanup,
4220 gimple_bb (stmt)->index);
4221 if (dump_file && (dump_flags & TDF_DETAILS))
4222 fprintf (dump_file, " Removed EH side-effects.\n");
4223 }
4224 continue;
4225 }
4226
4227 if (sprime
4228 && sprime != lhs
4229 && (rhs == NULL_TREE
4230 || TREE_CODE (rhs) != SSA_NAME
4231 || may_propagate_copy (rhs, sprime)))
4232 {
4233 bool can_make_abnormal_goto
4234 = is_gimple_call (stmt)
4235 && stmt_can_make_abnormal_goto (stmt);
4236
4237 gcc_assert (sprime != rhs);
4238
4239 if (dump_file && (dump_flags & TDF_DETAILS))
4240 {
4241 fprintf (dump_file, "Replaced ");
4242 print_gimple_expr (dump_file, stmt, 0, 0);
4243 fprintf (dump_file, " with ");
4244 print_generic_expr (dump_file, sprime, 0);
4245 fprintf (dump_file, " in ");
4246 print_gimple_stmt (dump_file, stmt, 0, 0);
4247 }
4248
4249 if (TREE_CODE (sprime) == SSA_NAME)
4250 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4251 NECESSARY, true);
4252 /* We need to make sure the new and old types actually match,
4253 which may require adding a simple cast, which fold_convert
4254 will do for us. */
4255 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4256 && !useless_type_conversion_p (gimple_expr_type (stmt),
4257 TREE_TYPE (sprime)))
4258 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4259
4260 pre_stats.eliminations++;
4261 propagate_tree_value_into_stmt (&gsi, sprime);
4262 stmt = gsi_stmt (gsi);
4263 update_stmt (stmt);
4264
4265 /* If we removed EH side-effects from the statement, clean
4266 its EH information. */
4267 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4268 {
4269 bitmap_set_bit (need_eh_cleanup,
4270 gimple_bb (stmt)->index);
4271 if (dump_file && (dump_flags & TDF_DETAILS))
4272 fprintf (dump_file, " Removed EH side-effects.\n");
4273 }
4274
4275 /* Likewise for AB side-effects. */
4276 if (can_make_abnormal_goto
4277 && !stmt_can_make_abnormal_goto (stmt))
4278 {
4279 bitmap_set_bit (need_ab_cleanup,
4280 gimple_bb (stmt)->index);
4281 if (dump_file && (dump_flags & TDF_DETAILS))
4282 fprintf (dump_file, " Removed AB side-effects.\n");
4283 }
4284 }
4285 }
4286 /* If the statement is a scalar store, see if the expression
4287 has the same value number as its rhs. If so, the store is
4288 dead. */
4289 else if (gimple_assign_single_p (stmt)
4290 && !gimple_has_volatile_ops (stmt)
4291 && !is_gimple_reg (gimple_assign_lhs (stmt))
4292 && (TREE_CODE (rhs) == SSA_NAME
4293 || is_gimple_min_invariant (rhs)))
4294 {
4295 tree val;
4296 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4297 gimple_vuse (stmt), VN_WALK, NULL);
4298 if (TREE_CODE (rhs) == SSA_NAME)
4299 rhs = VN_INFO (rhs)->valnum;
4300 if (val
4301 && operand_equal_p (val, rhs, 0))
4302 {
4303 if (dump_file && (dump_flags & TDF_DETAILS))
4304 {
4305 fprintf (dump_file, "Deleted redundant store ");
4306 print_gimple_stmt (dump_file, stmt, 0, 0);
4307 }
4308
4309 /* Queue stmt for removal. */
4310 el_to_remove.safe_push (stmt);
4311 }
4312 }
4313 /* Visit COND_EXPRs and fold the comparison with the
4314 available value-numbers. */
4315 else if (gimple_code (stmt) == GIMPLE_COND)
4316 {
4317 tree op0 = gimple_cond_lhs (stmt);
4318 tree op1 = gimple_cond_rhs (stmt);
4319 tree result;
4320
4321 if (TREE_CODE (op0) == SSA_NAME)
4322 op0 = VN_INFO (op0)->valnum;
4323 if (TREE_CODE (op1) == SSA_NAME)
4324 op1 = VN_INFO (op1)->valnum;
4325 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4326 op0, op1);
4327 if (result && TREE_CODE (result) == INTEGER_CST)
4328 {
4329 if (integer_zerop (result))
4330 gimple_cond_make_false (stmt);
4331 else
4332 gimple_cond_make_true (stmt);
4333 update_stmt (stmt);
4334 el_todo = TODO_cleanup_cfg;
4335 }
4336 }
4337 /* Visit indirect calls and turn them into direct calls if
4338 possible. */
4339 if (is_gimple_call (stmt))
4340 {
4341 tree orig_fn = gimple_call_fn (stmt);
4342 tree fn;
4343 if (!orig_fn)
4344 continue;
4345 if (TREE_CODE (orig_fn) == SSA_NAME)
4346 fn = VN_INFO (orig_fn)->valnum;
4347 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4348 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4349 {
4350 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4351 if (!gimple_call_addr_fndecl (fn))
4352 {
4353 fn = ipa_intraprocedural_devirtualization (stmt);
4354 if (fn)
4355 fn = build_fold_addr_expr (fn);
4356 }
4357 }
4358 else
4359 continue;
4360 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4361 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4362 TREE_TYPE (fn)))
4363 {
4364 bool can_make_abnormal_goto
4365 = stmt_can_make_abnormal_goto (stmt);
4366 bool was_noreturn = gimple_call_noreturn_p (stmt);
4367
4368 if (dump_file && (dump_flags & TDF_DETAILS))
4369 {
4370 fprintf (dump_file, "Replacing call target with ");
4371 print_generic_expr (dump_file, fn, 0);
4372 fprintf (dump_file, " in ");
4373 print_gimple_stmt (dump_file, stmt, 0, 0);
4374 }
4375
4376 gimple_call_set_fn (stmt, fn);
4377 el_to_update.safe_push (stmt);
4378
4379 /* When changing a call into a noreturn call, cfg cleanup
4380 is needed to fix up the noreturn call. */
4381 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4382 el_todo |= TODO_cleanup_cfg;
4383
4384 /* If we removed EH side-effects from the statement, clean
4385 its EH information. */
4386 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4387 {
4388 bitmap_set_bit (need_eh_cleanup,
4389 gimple_bb (stmt)->index);
4390 if (dump_file && (dump_flags & TDF_DETAILS))
4391 fprintf (dump_file, " Removed EH side-effects.\n");
4392 }
4393
4394 /* Likewise for AB side-effects. */
4395 if (can_make_abnormal_goto
4396 && !stmt_can_make_abnormal_goto (stmt))
4397 {
4398 bitmap_set_bit (need_ab_cleanup,
4399 gimple_bb (stmt)->index);
4400 if (dump_file && (dump_flags & TDF_DETAILS))
4401 fprintf (dump_file, " Removed AB side-effects.\n");
4402 }
4403
4404 /* Changing an indirect call to a direct call may
4405 have exposed different semantics. This may
4406 require an SSA update. */
4407 el_todo |= TODO_update_ssa_only_virtuals;
4408 }
4409 }
4410 }
4411 }
4412
4413 /* Make no longer available leaders no longer available. */
4414
4415 void
4416 eliminate_dom_walker::after_dom_children (basic_block)
4417 {
4418 tree entry;
4419 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4420 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4421 }
4422
4423 /* Eliminate fully redundant computations. */
4424
4425 static unsigned int
4426 eliminate (void)
4427 {
4428 gimple_stmt_iterator gsi;
4429 gimple stmt;
4430 unsigned i;
4431
4432 need_eh_cleanup = BITMAP_ALLOC (NULL);
4433 need_ab_cleanup = BITMAP_ALLOC (NULL);
4434
4435 el_to_remove.create (0);
4436 el_to_update.create (0);
4437 el_todo = 0;
4438 el_avail.create (0);
4439 el_avail_stack.create (0);
4440
4441 eliminate_dom_walker (CDI_DOMINATORS).walk (cfun->cfg->x_entry_block_ptr);
4442
4443 el_avail.release ();
4444 el_avail_stack.release ();
4445
4446 /* We cannot remove stmts during BB walk, especially not release SSA
4447 names there as this confuses the VN machinery. The stmts ending
4448 up in el_to_remove are either stores or simple copies. */
4449 FOR_EACH_VEC_ELT (el_to_remove, i, stmt)
4450 {
4451 tree lhs = gimple_assign_lhs (stmt);
4452 tree rhs = gimple_assign_rhs1 (stmt);
4453 use_operand_p use_p;
4454 gimple use_stmt;
4455
4456 /* If there is a single use only, propagate the equivalency
4457 instead of keeping the copy. */
4458 if (TREE_CODE (lhs) == SSA_NAME
4459 && TREE_CODE (rhs) == SSA_NAME
4460 && single_imm_use (lhs, &use_p, &use_stmt)
4461 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4462 {
4463 SET_USE (use_p, rhs);
4464 update_stmt (use_stmt);
4465 if (inserted_exprs
4466 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4467 && TREE_CODE (rhs) == SSA_NAME)
4468 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4469 }
4470
4471 /* If this is a store or a now unused copy, remove it. */
4472 if (TREE_CODE (lhs) != SSA_NAME
4473 || has_zero_uses (lhs))
4474 {
4475 basic_block bb = gimple_bb (stmt);
4476 gsi = gsi_for_stmt (stmt);
4477 unlink_stmt_vdef (stmt);
4478 if (gsi_remove (&gsi, true))
4479 bitmap_set_bit (need_eh_cleanup, bb->index);
4480 if (inserted_exprs
4481 && TREE_CODE (lhs) == SSA_NAME)
4482 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4483 release_defs (stmt);
4484 }
4485 }
4486 el_to_remove.release ();
4487
4488 /* We cannot update call statements with virtual operands during
4489 SSA walk. This might remove them which in turn makes our
4490 VN lattice invalid. */
4491 FOR_EACH_VEC_ELT (el_to_update, i, stmt)
4492 update_stmt (stmt);
4493 el_to_update.release ();
4494
4495 return el_todo;
4496 }
4497
4498 /* Perform CFG cleanups made necessary by elimination. */
4499
4500 static unsigned
4501 fini_eliminate (void)
4502 {
4503 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4504 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4505
4506 if (do_eh_cleanup)
4507 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4508
4509 if (do_ab_cleanup)
4510 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4511
4512 BITMAP_FREE (need_eh_cleanup);
4513 BITMAP_FREE (need_ab_cleanup);
4514
4515 if (do_eh_cleanup || do_ab_cleanup)
4516 return TODO_cleanup_cfg;
4517 return 0;
4518 }
4519
4520 /* Borrow a bit of tree-ssa-dce.c for the moment.
4521 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4522 this may be a bit faster, and we may want critical edges kept split. */
4523
4524 /* If OP's defining statement has not already been determined to be necessary,
4525 mark that statement necessary. Return the stmt, if it is newly
4526 necessary. */
4527
4528 static inline gimple
4529 mark_operand_necessary (tree op)
4530 {
4531 gimple stmt;
4532
4533 gcc_assert (op);
4534
4535 if (TREE_CODE (op) != SSA_NAME)
4536 return NULL;
4537
4538 stmt = SSA_NAME_DEF_STMT (op);
4539 gcc_assert (stmt);
4540
4541 if (gimple_plf (stmt, NECESSARY)
4542 || gimple_nop_p (stmt))
4543 return NULL;
4544
4545 gimple_set_plf (stmt, NECESSARY, true);
4546 return stmt;
4547 }
4548
4549 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4550 to insert PHI nodes sometimes, and because value numbering of casts isn't
4551 perfect, we sometimes end up inserting dead code. This simple DCE-like
4552 pass removes any insertions we made that weren't actually used. */
4553
4554 static void
4555 remove_dead_inserted_code (void)
4556 {
4557 bitmap worklist;
4558 unsigned i;
4559 bitmap_iterator bi;
4560 gimple t;
4561
4562 worklist = BITMAP_ALLOC (NULL);
4563 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4564 {
4565 t = SSA_NAME_DEF_STMT (ssa_name (i));
4566 if (gimple_plf (t, NECESSARY))
4567 bitmap_set_bit (worklist, i);
4568 }
4569 while (!bitmap_empty_p (worklist))
4570 {
4571 i = bitmap_first_set_bit (worklist);
4572 bitmap_clear_bit (worklist, i);
4573 t = SSA_NAME_DEF_STMT (ssa_name (i));
4574
4575 /* PHI nodes are somewhat special in that each PHI alternative has
4576 data and control dependencies. All the statements feeding the
4577 PHI node's arguments are always necessary. */
4578 if (gimple_code (t) == GIMPLE_PHI)
4579 {
4580 unsigned k;
4581
4582 for (k = 0; k < gimple_phi_num_args (t); k++)
4583 {
4584 tree arg = PHI_ARG_DEF (t, k);
4585 if (TREE_CODE (arg) == SSA_NAME)
4586 {
4587 gimple n = mark_operand_necessary (arg);
4588 if (n)
4589 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4590 }
4591 }
4592 }
4593 else
4594 {
4595 /* Propagate through the operands. Examine all the USE, VUSE and
4596 VDEF operands in this statement. Mark all the statements
4597 which feed this statement's uses as necessary. */
4598 ssa_op_iter iter;
4599 tree use;
4600
4601 /* The operands of VDEF expressions are also needed as they
4602 represent potential definitions that may reach this
4603 statement (VDEF operands allow us to follow def-def
4604 links). */
4605
4606 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4607 {
4608 gimple n = mark_operand_necessary (use);
4609 if (n)
4610 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4611 }
4612 }
4613 }
4614
4615 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4616 {
4617 t = SSA_NAME_DEF_STMT (ssa_name (i));
4618 if (!gimple_plf (t, NECESSARY))
4619 {
4620 gimple_stmt_iterator gsi;
4621
4622 if (dump_file && (dump_flags & TDF_DETAILS))
4623 {
4624 fprintf (dump_file, "Removing unnecessary insertion:");
4625 print_gimple_stmt (dump_file, t, 0, 0);
4626 }
4627
4628 gsi = gsi_for_stmt (t);
4629 if (gimple_code (t) == GIMPLE_PHI)
4630 remove_phi_node (&gsi, true);
4631 else
4632 {
4633 gsi_remove (&gsi, true);
4634 release_defs (t);
4635 }
4636 }
4637 }
4638 BITMAP_FREE (worklist);
4639 }
4640
4641
4642 /* Initialize data structures used by PRE. */
4643
4644 static void
4645 init_pre (void)
4646 {
4647 basic_block bb;
4648
4649 next_expression_id = 1;
4650 expressions.create (0);
4651 expressions.safe_push (NULL);
4652 value_expressions.create (get_max_value_id () + 1);
4653 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4654 name_to_id.create (0);
4655
4656 inserted_exprs = BITMAP_ALLOC (NULL);
4657
4658 connect_infinite_loops_to_exit ();
4659 memset (&pre_stats, 0, sizeof (pre_stats));
4660
4661 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4662 postorder_num = inverted_post_order_compute (postorder);
4663
4664 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4665
4666 calculate_dominance_info (CDI_POST_DOMINATORS);
4667 calculate_dominance_info (CDI_DOMINATORS);
4668
4669 bitmap_obstack_initialize (&grand_bitmap_obstack);
4670 phi_translate_table.create (5110);
4671 expression_to_id.create (num_ssa_names * 3);
4672 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4673 sizeof (struct bitmap_set), 30);
4674 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4675 sizeof (struct pre_expr_d), 30);
4676 FOR_ALL_BB (bb)
4677 {
4678 EXP_GEN (bb) = bitmap_set_new ();
4679 PHI_GEN (bb) = bitmap_set_new ();
4680 TMP_GEN (bb) = bitmap_set_new ();
4681 AVAIL_OUT (bb) = bitmap_set_new ();
4682 }
4683 }
4684
4685
4686 /* Deallocate data structures used by PRE. */
4687
4688 static void
4689 fini_pre ()
4690 {
4691 free (postorder);
4692 value_expressions.release ();
4693 BITMAP_FREE (inserted_exprs);
4694 bitmap_obstack_release (&grand_bitmap_obstack);
4695 free_alloc_pool (bitmap_set_pool);
4696 free_alloc_pool (pre_expr_pool);
4697 phi_translate_table.dispose ();
4698 expression_to_id.dispose ();
4699 name_to_id.release ();
4700
4701 free_aux_for_blocks ();
4702
4703 free_dominance_info (CDI_POST_DOMINATORS);
4704 }
4705
4706 /* Gate and execute functions for PRE. */
4707
4708 static unsigned int
4709 do_pre (void)
4710 {
4711 unsigned int todo = 0;
4712
4713 do_partial_partial =
4714 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4715
4716 /* This has to happen before SCCVN runs because
4717 loop_optimizer_init may create new phis, etc. */
4718 loop_optimizer_init (LOOPS_NORMAL);
4719
4720 if (!run_scc_vn (VN_WALK))
4721 {
4722 loop_optimizer_finalize ();
4723 return 0;
4724 }
4725
4726 init_pre ();
4727 scev_initialize ();
4728
4729 /* Collect and value number expressions computed in each basic block. */
4730 compute_avail ();
4731
4732 /* Insert can get quite slow on an incredibly large number of basic
4733 blocks due to some quadratic behavior. Until this behavior is
4734 fixed, don't run it when he have an incredibly large number of
4735 bb's. If we aren't going to run insert, there is no point in
4736 computing ANTIC, either, even though it's plenty fast. */
4737 if (n_basic_blocks_for_fn (cfun) < 4000)
4738 {
4739 compute_antic ();
4740 insert ();
4741 }
4742
4743 /* Make sure to remove fake edges before committing our inserts.
4744 This makes sure we don't end up with extra critical edges that
4745 we would need to split. */
4746 remove_fake_exit_edges ();
4747 gsi_commit_edge_inserts ();
4748
4749 /* Remove all the redundant expressions. */
4750 todo |= eliminate ();
4751
4752 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4753 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4754 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4755 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4756
4757 clear_expression_ids ();
4758 remove_dead_inserted_code ();
4759 todo |= TODO_verify_flow;
4760
4761 scev_finalize ();
4762 fini_pre ();
4763 todo |= fini_eliminate ();
4764 loop_optimizer_finalize ();
4765
4766 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4767 case we can merge the block with the remaining predecessor of the block.
4768 It should either:
4769 - call merge_blocks after each tail merge iteration
4770 - call merge_blocks after all tail merge iterations
4771 - mark TODO_cleanup_cfg when necessary
4772 - share the cfg cleanup with fini_pre. */
4773 todo |= tail_merge_optimize (todo);
4774
4775 free_scc_vn ();
4776
4777 /* Tail merging invalidates the virtual SSA web, together with
4778 cfg-cleanup opportunities exposed by PRE this will wreck the
4779 SSA updating machinery. So make sure to run update-ssa
4780 manually, before eventually scheduling cfg-cleanup as part of
4781 the todo. */
4782 update_ssa (TODO_update_ssa_only_virtuals);
4783
4784 return todo;
4785 }
4786
4787 static bool
4788 gate_pre (void)
4789 {
4790 return flag_tree_pre != 0;
4791 }
4792
4793 namespace {
4794
4795 const pass_data pass_data_pre =
4796 {
4797 GIMPLE_PASS, /* type */
4798 "pre", /* name */
4799 OPTGROUP_NONE, /* optinfo_flags */
4800 true, /* has_gate */
4801 true, /* has_execute */
4802 TV_TREE_PRE, /* tv_id */
4803 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4804 0, /* properties_provided */
4805 0, /* properties_destroyed */
4806 TODO_rebuild_alias, /* todo_flags_start */
4807 TODO_verify_ssa, /* todo_flags_finish */
4808 };
4809
4810 class pass_pre : public gimple_opt_pass
4811 {
4812 public:
4813 pass_pre (gcc::context *ctxt)
4814 : gimple_opt_pass (pass_data_pre, ctxt)
4815 {}
4816
4817 /* opt_pass methods: */
4818 bool gate () { return gate_pre (); }
4819 unsigned int execute () { return do_pre (); }
4820
4821 }; // class pass_pre
4822
4823 } // anon namespace
4824
4825 gimple_opt_pass *
4826 make_pass_pre (gcc::context *ctxt)
4827 {
4828 return new pass_pre (ctxt);
4829 }
4830
4831
4832 /* Gate and execute functions for FRE. */
4833
4834 static unsigned int
4835 execute_fre (void)
4836 {
4837 unsigned int todo = 0;
4838
4839 if (!run_scc_vn (VN_WALKREWRITE))
4840 return 0;
4841
4842 memset (&pre_stats, 0, sizeof (pre_stats));
4843
4844 /* Remove all the redundant expressions. */
4845 todo |= eliminate ();
4846
4847 todo |= fini_eliminate ();
4848
4849 free_scc_vn ();
4850
4851 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4852 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4853
4854 return todo;
4855 }
4856
4857 static bool
4858 gate_fre (void)
4859 {
4860 return flag_tree_fre != 0;
4861 }
4862
4863 namespace {
4864
4865 const pass_data pass_data_fre =
4866 {
4867 GIMPLE_PASS, /* type */
4868 "fre", /* name */
4869 OPTGROUP_NONE, /* optinfo_flags */
4870 true, /* has_gate */
4871 true, /* has_execute */
4872 TV_TREE_FRE, /* tv_id */
4873 ( PROP_cfg | PROP_ssa ), /* properties_required */
4874 0, /* properties_provided */
4875 0, /* properties_destroyed */
4876 0, /* todo_flags_start */
4877 TODO_verify_ssa, /* todo_flags_finish */
4878 };
4879
4880 class pass_fre : public gimple_opt_pass
4881 {
4882 public:
4883 pass_fre (gcc::context *ctxt)
4884 : gimple_opt_pass (pass_data_fre, ctxt)
4885 {}
4886
4887 /* opt_pass methods: */
4888 opt_pass * clone () { return new pass_fre (m_ctxt); }
4889 bool gate () { return gate_fre (); }
4890 unsigned int execute () { return execute_fre (); }
4891
4892 }; // class pass_fre
4893
4894 } // anon namespace
4895
4896 gimple_opt_pass *
4897 make_pass_fre (gcc::context *ctxt)
4898 {
4899 return new pass_fre (ctxt);
4900 }