re PR tree-optimization/35229 (Vectorizer doesn't support dependence created by...
[gcc.git] / gcc / tree-ssa-pre.c
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
6
7 This file is part of GCC.
8
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
13
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
18
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
22
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "ggc.h"
28 #include "tree.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "real.h"
40 #include "alloc-pool.h"
41 #include "obstack.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
51
52 /* TODO:
53
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
61 */
62
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
67
68 /* Basic algorithm
69
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
79 expressions/values.
80
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
94
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
97
98 An expression is partially redundant (excluding partial
99 anticipation) if:
100
101 1. It is AVAIL in some, but not all, of the predecessors of a
102 given block.
103 2. It is ANTIC in all the predecessors.
104
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
107
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
110 of the predecessors.
111
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
114
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
118
119 /* Representations of value numbers:
120
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
129
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
134 use the value id. */
135
136 /* Representation of expressions on value numbers:
137
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
142
143 /* Representation of sets:
144
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
149
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
152 order. */
153
154 /* Type of expression, used to know which member of the PRE_EXPR union
155 is valid. */
156
157 enum pre_expr_kind
158 {
159 NAME,
160 NARY,
161 REFERENCE,
162 CONSTANT
163 };
164
165 typedef union pre_expr_union_d
166 {
167 tree name;
168 tree constant;
169 vn_nary_op_t nary;
170 vn_reference_t reference;
171 } pre_expr_union;
172
173 typedef struct pre_expr_d
174 {
175 enum pre_expr_kind kind;
176 unsigned int id;
177 pre_expr_union u;
178 } *pre_expr;
179
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184
185 static int
186 pre_expr_eq (const void *p1, const void *p2)
187 {
188 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
189 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190
191 if (e1->kind != e2->kind)
192 return false;
193
194 switch (e1->kind)
195 {
196 case CONSTANT:
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
198 PRE_EXPR_CONSTANT (e2));
199 case NAME:
200 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 case NARY:
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 case REFERENCE:
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
205 PRE_EXPR_REFERENCE (e2));
206 default:
207 abort();
208 }
209 }
210
211 static hashval_t
212 pre_expr_hash (const void *p1)
213 {
214 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
215 switch (e->kind)
216 {
217 case CONSTANT:
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 case NAME:
220 return iterative_hash_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e)), 0);
221 case NARY:
222 return PRE_EXPR_NARY (e)->hashcode;
223 case REFERENCE:
224 return PRE_EXPR_REFERENCE (e)->hashcode;
225 default:
226 abort ();
227 }
228 }
229
230
231 /* Next global expression id number. */
232 static unsigned int next_expression_id;
233
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr);
236 DEF_VEC_ALLOC_P (pre_expr, heap);
237 static VEC(pre_expr, heap) *expressions;
238 static htab_t expression_to_id;
239
240 /* Allocate an expression id for EXPR. */
241
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
244 {
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 slot = htab_find_slot (expression_to_id, expr, INSERT);
251 gcc_assert (!*slot);
252 *slot = expr;
253 return next_expression_id - 1;
254 }
255
256 /* Return the expression id for tree EXPR. */
257
258 static inline unsigned int
259 get_expression_id (const pre_expr expr)
260 {
261 return expr->id;
262 }
263
264 static inline unsigned int
265 lookup_expression_id (const pre_expr expr)
266 {
267 void **slot;
268
269 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
270 if (!slot)
271 return 0;
272 return ((pre_expr)*slot)->id;
273 }
274
275 /* Return the existing expression id for EXPR, or create one if one
276 does not exist yet. */
277
278 static inline unsigned int
279 get_or_alloc_expression_id (pre_expr expr)
280 {
281 unsigned int id = lookup_expression_id (expr);
282 if (id == 0)
283 return alloc_expression_id (expr);
284 return expr->id = id;
285 }
286
287 /* Return the expression that has expression id ID */
288
289 static inline pre_expr
290 expression_for_id (unsigned int id)
291 {
292 return VEC_index (pre_expr, expressions, id);
293 }
294
295 /* Free the expression id field in all of our expressions,
296 and then destroy the expressions array. */
297
298 static void
299 clear_expression_ids (void)
300 {
301 VEC_free (pre_expr, heap, expressions);
302 }
303
304 static alloc_pool pre_expr_pool;
305
306 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
307
308 static pre_expr
309 get_or_alloc_expr_for_name (tree name)
310 {
311 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
312 unsigned int result_id;
313
314 result->kind = NAME;
315 result->id = 0;
316 PRE_EXPR_NAME (result) = name;
317 result_id = lookup_expression_id (result);
318 if (result_id != 0)
319 {
320 pool_free (pre_expr_pool, result);
321 result = expression_for_id (result_id);
322 return result;
323 }
324 get_or_alloc_expression_id (result);
325 return result;
326 }
327
328 static bool in_fre = false;
329
330 /* An unordered bitmap set. One bitmap tracks values, the other,
331 expressions. */
332 typedef struct bitmap_set
333 {
334 bitmap expressions;
335 bitmap values;
336 } *bitmap_set_t;
337
338 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
339 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
340
341 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
342 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
343
344 /* Mapping from value id to expressions with that value_id. */
345 DEF_VEC_P (bitmap_set_t);
346 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
347 static VEC(bitmap_set_t, heap) *value_expressions;
348
349 /* Sets that we need to keep track of. */
350 typedef struct bb_bitmap_sets
351 {
352 /* The EXP_GEN set, which represents expressions/values generated in
353 a basic block. */
354 bitmap_set_t exp_gen;
355
356 /* The PHI_GEN set, which represents PHI results generated in a
357 basic block. */
358 bitmap_set_t phi_gen;
359
360 /* The TMP_GEN set, which represents results/temporaries generated
361 in a basic block. IE the LHS of an expression. */
362 bitmap_set_t tmp_gen;
363
364 /* The AVAIL_OUT set, which represents which values are available in
365 a given basic block. */
366 bitmap_set_t avail_out;
367
368 /* The ANTIC_IN set, which represents which values are anticipatable
369 in a given basic block. */
370 bitmap_set_t antic_in;
371
372 /* The PA_IN set, which represents which values are
373 partially anticipatable in a given basic block. */
374 bitmap_set_t pa_in;
375
376 /* The NEW_SETS set, which is used during insertion to augment the
377 AVAIL_OUT set of blocks with the new insertions performed during
378 the current iteration. */
379 bitmap_set_t new_sets;
380
381 /* A cache for value_dies_in_block_x. */
382 bitmap expr_dies;
383
384 /* True if we have visited this block during ANTIC calculation. */
385 unsigned int visited:1;
386
387 /* True we have deferred processing this block during ANTIC
388 calculation until its successor is processed. */
389 unsigned int deferred : 1;
390 } *bb_value_sets_t;
391
392 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
393 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
394 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
395 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
396 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
397 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
398 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
399 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
400 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
401 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
402
403
404 /* Maximal set of values, used to initialize the ANTIC problem, which
405 is an intersection problem. */
406 static bitmap_set_t maximal_set;
407
408 /* Basic block list in postorder. */
409 static int *postorder;
410
411 /* This structure is used to keep track of statistics on what
412 optimization PRE was able to perform. */
413 static struct
414 {
415 /* The number of RHS computations eliminated by PRE. */
416 int eliminations;
417
418 /* The number of new expressions/temporaries generated by PRE. */
419 int insertions;
420
421 /* The number of inserts found due to partial anticipation */
422 int pa_insert;
423
424 /* The number of new PHI nodes added by PRE. */
425 int phis;
426
427 /* The number of values found constant. */
428 int constified;
429
430 } pre_stats;
431
432 static bool do_partial_partial;
433 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
434 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
435 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
436 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
437 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
438 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
439 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
440 static bitmap_set_t bitmap_set_new (void);
441 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
442 gimple, tree);
443 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
444 gimple);
445 static unsigned int get_expr_value_id (pre_expr);
446
447 /* We can add and remove elements and entries to and from sets
448 and hash tables, so we use alloc pools for them. */
449
450 static alloc_pool bitmap_set_pool;
451 static bitmap_obstack grand_bitmap_obstack;
452
453 /* To avoid adding 300 temporary variables when we only need one, we
454 only create one temporary variable, on demand, and build ssa names
455 off that. We do have to change the variable if the types don't
456 match the current variable's type. */
457 static tree pretemp;
458 static tree storetemp;
459 static tree prephitemp;
460
461 /* Set of blocks with statements that have had its EH information
462 cleaned up. */
463 static bitmap need_eh_cleanup;
464
465 /* Which expressions have been seen during a given phi translation. */
466 static bitmap seen_during_translate;
467
468 /* The phi_translate_table caches phi translations for a given
469 expression and predecessor. */
470
471 static htab_t phi_translate_table;
472
473 /* A three tuple {e, pred, v} used to cache phi translations in the
474 phi_translate_table. */
475
476 typedef struct expr_pred_trans_d
477 {
478 /* The expression. */
479 pre_expr e;
480
481 /* The predecessor block along which we translated the expression. */
482 basic_block pred;
483
484 /* The value that resulted from the translation. */
485 pre_expr v;
486
487 /* The hashcode for the expression, pred pair. This is cached for
488 speed reasons. */
489 hashval_t hashcode;
490 } *expr_pred_trans_t;
491 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
492
493 /* Return the hash value for a phi translation table entry. */
494
495 static hashval_t
496 expr_pred_trans_hash (const void *p)
497 {
498 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
499 return ve->hashcode;
500 }
501
502 /* Return true if two phi translation table entries are the same.
503 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
504
505 static int
506 expr_pred_trans_eq (const void *p1, const void *p2)
507 {
508 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
509 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
510 basic_block b1 = ve1->pred;
511 basic_block b2 = ve2->pred;
512
513 /* If they are not translations for the same basic block, they can't
514 be equal. */
515 if (b1 != b2)
516 return false;
517 return pre_expr_eq (ve1->e, ve2->e);
518 }
519
520 /* Search in the phi translation table for the translation of
521 expression E in basic block PRED.
522 Return the translated value, if found, NULL otherwise. */
523
524 static inline pre_expr
525 phi_trans_lookup (pre_expr e, basic_block pred)
526 {
527 void **slot;
528 struct expr_pred_trans_d ept;
529
530 ept.e = e;
531 ept.pred = pred;
532 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
533 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
534 NO_INSERT);
535 if (!slot)
536 return NULL;
537 else
538 return ((expr_pred_trans_t) *slot)->v;
539 }
540
541
542 /* Add the tuple mapping from {expression E, basic block PRED} to
543 value V, to the phi translation table. */
544
545 static inline void
546 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
547 {
548 void **slot;
549 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
550 new_pair->e = e;
551 new_pair->pred = pred;
552 new_pair->v = v;
553 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
554 pred->index);
555
556 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
557 new_pair->hashcode, INSERT);
558 if (*slot)
559 free (*slot);
560 *slot = (void *) new_pair;
561 }
562
563
564 /* Add expression E to the expression set of value id V. */
565
566 void
567 add_to_value (unsigned int v, pre_expr e)
568 {
569 bitmap_set_t set;
570
571 gcc_assert (get_expr_value_id (e) == v);
572
573 if (v >= VEC_length (bitmap_set_t, value_expressions))
574 {
575 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
576 v + 1);
577 }
578
579 set = VEC_index (bitmap_set_t, value_expressions, v);
580 if (!set)
581 {
582 set = bitmap_set_new ();
583 VEC_replace (bitmap_set_t, value_expressions, v, set);
584 }
585
586 bitmap_insert_into_set_1 (set, e, true);
587 }
588
589 /* Create a new bitmap set and return it. */
590
591 static bitmap_set_t
592 bitmap_set_new (void)
593 {
594 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
595 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
596 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
597 return ret;
598 }
599
600 /* Return the value id for a PRE expression EXPR. */
601
602 static unsigned int
603 get_expr_value_id (pre_expr expr)
604 {
605 switch (expr->kind)
606 {
607 case CONSTANT:
608 {
609 unsigned int id;
610 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
611 if (id == 0)
612 {
613 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
614 add_to_value (id, expr);
615 }
616 return id;
617 }
618 case NAME:
619 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
620 case NARY:
621 return PRE_EXPR_NARY (expr)->value_id;
622 case REFERENCE:
623 return PRE_EXPR_REFERENCE (expr)->value_id;
624 default:
625 gcc_unreachable ();
626 }
627 }
628
629 /* Remove an expression EXPR from a bitmapped set. */
630
631 static void
632 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
633 {
634 unsigned int val = get_expr_value_id (expr);
635 if (!value_id_constant_p (val))
636 {
637 bitmap_clear_bit (set->values, val);
638 bitmap_clear_bit (set->expressions, get_expression_id (expr));
639 }
640 }
641
642 static void
643 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
644 bool allow_constants)
645 {
646 unsigned int val = get_expr_value_id (expr);
647 if (allow_constants || !value_id_constant_p (val))
648 {
649 /* We specifically expect this and only this function to be able to
650 insert constants into a set. */
651 bitmap_set_bit (set->values, val);
652 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
653 }
654 }
655
656 /* Insert an expression EXPR into a bitmapped set. */
657
658 static void
659 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
660 {
661 bitmap_insert_into_set_1 (set, expr, false);
662 }
663
664 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
665
666 static void
667 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
668 {
669 bitmap_copy (dest->expressions, orig->expressions);
670 bitmap_copy (dest->values, orig->values);
671 }
672
673
674 /* Free memory used up by SET. */
675 static void
676 bitmap_set_free (bitmap_set_t set)
677 {
678 BITMAP_FREE (set->expressions);
679 BITMAP_FREE (set->values);
680 }
681
682
683 /* Generate an topological-ordered array of bitmap set SET. */
684
685 static VEC(pre_expr, heap) *
686 sorted_array_from_bitmap_set (bitmap_set_t set)
687 {
688 unsigned int i, j;
689 bitmap_iterator bi, bj;
690 VEC(pre_expr, heap) *result = NULL;
691
692 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
693 {
694 /* The number of expressions having a given value is usually
695 relatively small. Thus, rather than making a vector of all
696 the expressions and sorting it by value-id, we walk the values
697 and check in the reverse mapping that tells us what expressions
698 have a given value, to filter those in our set. As a result,
699 the expressions are inserted in value-id order, which means
700 topological order.
701
702 If this is somehow a significant lose for some cases, we can
703 choose which set to walk based on the set size. */
704 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
705 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
706 {
707 if (bitmap_bit_p (set->expressions, j))
708 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
709 }
710 }
711
712 return result;
713 }
714
715 /* Perform bitmapped set operation DEST &= ORIG. */
716
717 static void
718 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
719 {
720 bitmap_iterator bi;
721 unsigned int i;
722
723 if (dest != orig)
724 {
725 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
726
727 bitmap_and_into (dest->values, orig->values);
728 bitmap_copy (temp, dest->expressions);
729 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
730 {
731 pre_expr expr = expression_for_id (i);
732 unsigned int value_id = get_expr_value_id (expr);
733 if (!bitmap_bit_p (dest->values, value_id))
734 bitmap_clear_bit (dest->expressions, i);
735 }
736 BITMAP_FREE (temp);
737 }
738 }
739
740 /* Subtract all values and expressions contained in ORIG from DEST. */
741
742 static bitmap_set_t
743 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
744 {
745 bitmap_set_t result = bitmap_set_new ();
746 bitmap_iterator bi;
747 unsigned int i;
748
749 bitmap_and_compl (result->expressions, dest->expressions,
750 orig->expressions);
751
752 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
753 {
754 pre_expr expr = expression_for_id (i);
755 unsigned int value_id = get_expr_value_id (expr);
756 bitmap_set_bit (result->values, value_id);
757 }
758
759 return result;
760 }
761
762 /* Subtract all the values in bitmap set B from bitmap set A. */
763
764 static void
765 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
766 {
767 unsigned int i;
768 bitmap_iterator bi;
769 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
770
771 bitmap_copy (temp, a->expressions);
772 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
773 {
774 pre_expr expr = expression_for_id (i);
775 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
776 bitmap_remove_from_set (a, expr);
777 }
778 BITMAP_FREE (temp);
779 }
780
781
782 /* Return true if bitmapped set SET contains the value VALUE_ID. */
783
784 static bool
785 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
786 {
787 if (value_id_constant_p (value_id))
788 return true;
789
790 if (!set || bitmap_empty_p (set->expressions))
791 return false;
792
793 return bitmap_bit_p (set->values, value_id);
794 }
795
796 static inline bool
797 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
798 {
799 return bitmap_bit_p (set->expressions, get_expression_id (expr));
800 }
801
802 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
803
804 static void
805 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
806 const pre_expr expr)
807 {
808 bitmap_set_t exprset;
809 unsigned int i;
810 bitmap_iterator bi;
811
812 if (value_id_constant_p (lookfor))
813 return;
814
815 if (!bitmap_set_contains_value (set, lookfor))
816 return;
817
818 /* The number of expressions having a given value is usually
819 significantly less than the total number of expressions in SET.
820 Thus, rather than check, for each expression in SET, whether it
821 has the value LOOKFOR, we walk the reverse mapping that tells us
822 what expressions have a given value, and see if any of those
823 expressions are in our set. For large testcases, this is about
824 5-10x faster than walking the bitmap. If this is somehow a
825 significant lose for some cases, we can choose which set to walk
826 based on the set size. */
827 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
828 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
829 {
830 if (bitmap_bit_p (set->expressions, i))
831 {
832 bitmap_clear_bit (set->expressions, i);
833 bitmap_set_bit (set->expressions, get_expression_id (expr));
834 return;
835 }
836 }
837 }
838
839 /* Return true if two bitmap sets are equal. */
840
841 static bool
842 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
843 {
844 return bitmap_equal_p (a->values, b->values);
845 }
846
847 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
848 and add it otherwise. */
849
850 static void
851 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
852 {
853 unsigned int val = get_expr_value_id (expr);
854
855 if (bitmap_set_contains_value (set, val))
856 bitmap_set_replace_value (set, val, expr);
857 else
858 bitmap_insert_into_set (set, expr);
859 }
860
861 /* Insert EXPR into SET if EXPR's value is not already present in
862 SET. */
863
864 static void
865 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
866 {
867 unsigned int val = get_expr_value_id (expr);
868
869 if (value_id_constant_p (val))
870 return;
871
872 if (!bitmap_set_contains_value (set, val))
873 bitmap_insert_into_set (set, expr);
874 }
875
876 /* Print out EXPR to outfile. */
877
878 static void
879 print_pre_expr (FILE *outfile, const pre_expr expr)
880 {
881 switch (expr->kind)
882 {
883 case CONSTANT:
884 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
885 break;
886 case NAME:
887 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
888 break;
889 case NARY:
890 {
891 unsigned int i;
892 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
893 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
894 for (i = 0; i < nary->length; i++)
895 {
896 print_generic_expr (outfile, nary->op[i], 0);
897 if (i != (unsigned) nary->length - 1)
898 fprintf (outfile, ",");
899 }
900 fprintf (outfile, "}");
901 }
902 break;
903
904 case REFERENCE:
905 {
906 vn_reference_op_t vro;
907 unsigned int i;
908 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
909 fprintf (outfile, "{");
910 for (i = 0;
911 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
912 i++)
913 {
914 bool closebrace = false;
915 if (vro->opcode != SSA_NAME
916 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
917 {
918 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
919 if (vro->op0)
920 {
921 fprintf (outfile, "<");
922 closebrace = true;
923 }
924 }
925 if (vro->op0)
926 {
927 print_generic_expr (outfile, vro->op0, 0);
928 if (vro->op1)
929 {
930 fprintf (outfile, ",");
931 print_generic_expr (outfile, vro->op1, 0);
932 }
933 if (vro->op2)
934 {
935 fprintf (outfile, ",");
936 print_generic_expr (outfile, vro->op2, 0);
937 }
938 }
939 if (closebrace)
940 fprintf (outfile, ">");
941 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
942 fprintf (outfile, ",");
943 }
944 fprintf (outfile, "}");
945 if (ref->vuse)
946 {
947 fprintf (outfile, "@");
948 print_generic_expr (outfile, ref->vuse, 0);
949 }
950 }
951 break;
952 }
953 }
954 void debug_pre_expr (pre_expr);
955
956 /* Like print_pre_expr but always prints to stderr. */
957 void
958 debug_pre_expr (pre_expr e)
959 {
960 print_pre_expr (stderr, e);
961 fprintf (stderr, "\n");
962 }
963
964 /* Print out SET to OUTFILE. */
965
966 static void
967 print_bitmap_set (FILE *outfile, bitmap_set_t set,
968 const char *setname, int blockindex)
969 {
970 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
971 if (set)
972 {
973 bool first = true;
974 unsigned i;
975 bitmap_iterator bi;
976
977 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
978 {
979 const pre_expr expr = expression_for_id (i);
980
981 if (!first)
982 fprintf (outfile, ", ");
983 first = false;
984 print_pre_expr (outfile, expr);
985
986 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
987 }
988 }
989 fprintf (outfile, " }\n");
990 }
991
992 void debug_bitmap_set (bitmap_set_t);
993
994 void
995 debug_bitmap_set (bitmap_set_t set)
996 {
997 print_bitmap_set (stderr, set, "debug", 0);
998 }
999
1000 /* Print out the expressions that have VAL to OUTFILE. */
1001
1002 void
1003 print_value_expressions (FILE *outfile, unsigned int val)
1004 {
1005 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1006 if (set)
1007 {
1008 char s[10];
1009 sprintf (s, "%04d", val);
1010 print_bitmap_set (outfile, set, s, 0);
1011 }
1012 }
1013
1014
1015 void
1016 debug_value_expressions (unsigned int val)
1017 {
1018 print_value_expressions (stderr, val);
1019 }
1020
1021 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1022 represent it. */
1023
1024 static pre_expr
1025 get_or_alloc_expr_for_constant (tree constant)
1026 {
1027 unsigned int result_id;
1028 unsigned int value_id;
1029 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1030 newexpr->kind = CONSTANT;
1031 PRE_EXPR_CONSTANT (newexpr) = constant;
1032 result_id = lookup_expression_id (newexpr);
1033 if (result_id != 0)
1034 {
1035 pool_free (pre_expr_pool, newexpr);
1036 newexpr = expression_for_id (result_id);
1037 return newexpr;
1038 }
1039 value_id = get_or_alloc_constant_value_id (constant);
1040 get_or_alloc_expression_id (newexpr);
1041 add_to_value (value_id, newexpr);
1042 return newexpr;
1043 }
1044
1045 /* Given a value id V, find the actual tree representing the constant
1046 value if there is one, and return it. Return NULL if we can't find
1047 a constant. */
1048
1049 static tree
1050 get_constant_for_value_id (unsigned int v)
1051 {
1052 if (value_id_constant_p (v))
1053 {
1054 unsigned int i;
1055 bitmap_iterator bi;
1056 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1057
1058 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1059 {
1060 pre_expr expr = expression_for_id (i);
1061 if (expr->kind == CONSTANT)
1062 return PRE_EXPR_CONSTANT (expr);
1063 }
1064 }
1065 return NULL;
1066 }
1067
1068 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1069 Currently only supports constants and SSA_NAMES. */
1070 static pre_expr
1071 get_or_alloc_expr_for (tree t)
1072 {
1073 if (TREE_CODE (t) == SSA_NAME)
1074 return get_or_alloc_expr_for_name (t);
1075 else if (is_gimple_min_invariant (t)
1076 || TREE_CODE (t) == EXC_PTR_EXPR
1077 || TREE_CODE (t) == FILTER_EXPR)
1078 return get_or_alloc_expr_for_constant (t);
1079 else
1080 {
1081 /* More complex expressions can result from SCCVN expression
1082 simplification that inserts values for them. As they all
1083 do not have VOPs the get handled by the nary ops struct. */
1084 vn_nary_op_t result;
1085 unsigned int result_id;
1086 vn_nary_op_lookup (t, &result);
1087 if (result != NULL)
1088 {
1089 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1090 e->kind = NARY;
1091 PRE_EXPR_NARY (e) = result;
1092 result_id = lookup_expression_id (e);
1093 if (result_id != 0)
1094 {
1095 pool_free (pre_expr_pool, e);
1096 e = expression_for_id (result_id);
1097 return e;
1098 }
1099 alloc_expression_id (e);
1100 return e;
1101 }
1102 }
1103 return NULL;
1104 }
1105
1106 /* Return the folded version of T if T, when folded, is a gimple
1107 min_invariant. Otherwise, return T. */
1108
1109 static pre_expr
1110 fully_constant_expression (pre_expr e)
1111 {
1112 switch (e->kind)
1113 {
1114 case CONSTANT:
1115 return e;
1116 case NARY:
1117 {
1118 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1119 switch (TREE_CODE_CLASS (nary->opcode))
1120 {
1121 case tcc_expression:
1122 if (nary->opcode == TRUTH_NOT_EXPR)
1123 goto do_unary;
1124 if (nary->opcode != TRUTH_AND_EXPR
1125 && nary->opcode != TRUTH_OR_EXPR
1126 && nary->opcode != TRUTH_XOR_EXPR)
1127 return e;
1128 /* Fallthrough. */
1129 case tcc_binary:
1130 case tcc_comparison:
1131 {
1132 /* We have to go from trees to pre exprs to value ids to
1133 constants. */
1134 tree naryop0 = nary->op[0];
1135 tree naryop1 = nary->op[1];
1136 tree result;
1137 if (!is_gimple_min_invariant (naryop0))
1138 {
1139 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1140 unsigned int vrep0 = get_expr_value_id (rep0);
1141 tree const0 = get_constant_for_value_id (vrep0);
1142 if (const0)
1143 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1144 }
1145 if (!is_gimple_min_invariant (naryop1))
1146 {
1147 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1148 unsigned int vrep1 = get_expr_value_id (rep1);
1149 tree const1 = get_constant_for_value_id (vrep1);
1150 if (const1)
1151 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1152 }
1153 result = fold_binary (nary->opcode, nary->type,
1154 naryop0, naryop1);
1155 if (result && is_gimple_min_invariant (result))
1156 return get_or_alloc_expr_for_constant (result);
1157 /* We might have simplified the expression to a
1158 SSA_NAME for example from x_1 * 1. But we cannot
1159 insert a PHI for x_1 unconditionally as x_1 might
1160 not be available readily. */
1161 return e;
1162 }
1163 case tcc_reference:
1164 if (nary->opcode != REALPART_EXPR
1165 && nary->opcode != IMAGPART_EXPR
1166 && nary->opcode != VIEW_CONVERT_EXPR)
1167 return e;
1168 /* Fallthrough. */
1169 case tcc_unary:
1170 do_unary:
1171 {
1172 /* We have to go from trees to pre exprs to value ids to
1173 constants. */
1174 tree naryop0 = nary->op[0];
1175 tree const0, result;
1176 if (is_gimple_min_invariant (naryop0))
1177 const0 = naryop0;
1178 else
1179 {
1180 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1181 unsigned int vrep0 = get_expr_value_id (rep0);
1182 const0 = get_constant_for_value_id (vrep0);
1183 }
1184 result = NULL;
1185 if (const0)
1186 {
1187 tree type1 = TREE_TYPE (nary->op[0]);
1188 const0 = fold_convert (type1, const0);
1189 result = fold_unary (nary->opcode, nary->type, const0);
1190 }
1191 if (result && is_gimple_min_invariant (result))
1192 return get_or_alloc_expr_for_constant (result);
1193 return e;
1194 }
1195 default:
1196 return e;
1197 }
1198 }
1199 case REFERENCE:
1200 {
1201 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1202 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1203 vn_reference_op_t op;
1204
1205 /* Try to simplify the translated expression if it is
1206 a call to a builtin function with at most two arguments. */
1207 op = VEC_index (vn_reference_op_s, operands, 0);
1208 if (op->opcode == CALL_EXPR
1209 && TREE_CODE (op->op0) == ADDR_EXPR
1210 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1211 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1212 && VEC_length (vn_reference_op_s, operands) >= 2
1213 && VEC_length (vn_reference_op_s, operands) <= 3)
1214 {
1215 vn_reference_op_t arg0, arg1 = NULL;
1216 bool anyconst = false;
1217 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1218 if (VEC_length (vn_reference_op_s, operands) > 2)
1219 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1220 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1221 || (arg0->opcode == ADDR_EXPR
1222 && is_gimple_min_invariant (arg0->op0)))
1223 anyconst = true;
1224 if (arg1
1225 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1226 || (arg1->opcode == ADDR_EXPR
1227 && is_gimple_min_invariant (arg1->op0))))
1228 anyconst = true;
1229 if (anyconst)
1230 {
1231 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1232 arg1 ? 2 : 1,
1233 arg0->op0,
1234 arg1 ? arg1->op0 : NULL);
1235 if (folded
1236 && TREE_CODE (folded) == NOP_EXPR)
1237 folded = TREE_OPERAND (folded, 0);
1238 if (folded
1239 && is_gimple_min_invariant (folded))
1240 return get_or_alloc_expr_for_constant (folded);
1241 }
1242 }
1243 return e;
1244 }
1245 default:
1246 return e;
1247 }
1248 return e;
1249 }
1250
1251 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1252 it has the value it would have in BLOCK. */
1253
1254 static tree
1255 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1256 alias_set_type set, tree type, tree vuse,
1257 basic_block phiblock,
1258 basic_block block)
1259 {
1260 gimple phi = SSA_NAME_DEF_STMT (vuse);
1261 ao_ref ref;
1262
1263 if (gimple_bb (phi) != phiblock)
1264 return vuse;
1265
1266 if (gimple_code (phi) == GIMPLE_PHI)
1267 {
1268 edge e = find_edge (block, phiblock);
1269 return PHI_ARG_DEF (phi, e->dest_idx);
1270 }
1271
1272 if (!ao_ref_init_from_vn_reference (&ref, set, type, operands))
1273 return NULL_TREE;
1274
1275 /* Use the alias-oracle to find either the PHI node in this block,
1276 the first VUSE used in this block that is equivalent to vuse or
1277 the first VUSE which definition in this block kills the value. */
1278 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1279 {
1280 vuse = gimple_vuse (phi);
1281 phi = SSA_NAME_DEF_STMT (vuse);
1282 if (gimple_bb (phi) != phiblock)
1283 return vuse;
1284 if (gimple_code (phi) == GIMPLE_PHI)
1285 {
1286 edge e = find_edge (block, phiblock);
1287 return PHI_ARG_DEF (phi, e->dest_idx);
1288 }
1289 }
1290
1291 return NULL_TREE;
1292 }
1293
1294 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1295 SET2. This is used to avoid making a set consisting of the union
1296 of PA_IN and ANTIC_IN during insert. */
1297
1298 static inline pre_expr
1299 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1300 {
1301 pre_expr result;
1302
1303 result = bitmap_find_leader (set1, val, NULL);
1304 if (!result && set2)
1305 result = bitmap_find_leader (set2, val, NULL);
1306 return result;
1307 }
1308
1309 /* Get the tree type for our PRE expression e. */
1310
1311 static tree
1312 get_expr_type (const pre_expr e)
1313 {
1314 switch (e->kind)
1315 {
1316 case NAME:
1317 return TREE_TYPE (PRE_EXPR_NAME (e));
1318 case CONSTANT:
1319 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1320 case REFERENCE:
1321 return PRE_EXPR_REFERENCE (e)->type;
1322 case NARY:
1323 return PRE_EXPR_NARY (e)->type;
1324 }
1325 gcc_unreachable();
1326 }
1327
1328 /* Get a representative SSA_NAME for a given expression.
1329 Since all of our sub-expressions are treated as values, we require
1330 them to be SSA_NAME's for simplicity.
1331 Prior versions of GVNPRE used to use "value handles" here, so that
1332 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1333 either case, the operands are really values (IE we do not expect
1334 them to be usable without finding leaders). */
1335
1336 static tree
1337 get_representative_for (const pre_expr e)
1338 {
1339 tree exprtype;
1340 tree name;
1341 unsigned int value_id = get_expr_value_id (e);
1342
1343 switch (e->kind)
1344 {
1345 case NAME:
1346 return PRE_EXPR_NAME (e);
1347 case CONSTANT:
1348 return PRE_EXPR_CONSTANT (e);
1349 case NARY:
1350 case REFERENCE:
1351 {
1352 /* Go through all of the expressions representing this value
1353 and pick out an SSA_NAME. */
1354 unsigned int i;
1355 bitmap_iterator bi;
1356 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1357 value_id);
1358 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1359 {
1360 pre_expr rep = expression_for_id (i);
1361 if (rep->kind == NAME)
1362 return PRE_EXPR_NAME (rep);
1363 }
1364 }
1365 break;
1366 }
1367 /* If we reached here we couldn't find an SSA_NAME. This can
1368 happen when we've discovered a value that has never appeared in
1369 the program as set to an SSA_NAME, most likely as the result of
1370 phi translation. */
1371 if (dump_file)
1372 {
1373 fprintf (dump_file,
1374 "Could not find SSA_NAME representative for expression:");
1375 print_pre_expr (dump_file, e);
1376 fprintf (dump_file, "\n");
1377 }
1378
1379 exprtype = get_expr_type (e);
1380
1381 /* Build and insert the assignment of the end result to the temporary
1382 that we will return. */
1383 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1384 {
1385 pretemp = create_tmp_var (exprtype, "pretmp");
1386 get_var_ann (pretemp);
1387 }
1388
1389 name = make_ssa_name (pretemp, gimple_build_nop ());
1390 VN_INFO_GET (name)->value_id = value_id;
1391 if (e->kind == CONSTANT)
1392 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1393 else
1394 VN_INFO (name)->valnum = name;
1395
1396 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1397 if (dump_file)
1398 {
1399 fprintf (dump_file, "Created SSA_NAME representative ");
1400 print_generic_expr (dump_file, name, 0);
1401 fprintf (dump_file, " for expression:");
1402 print_pre_expr (dump_file, e);
1403 fprintf (dump_file, "\n");
1404 }
1405
1406 return name;
1407 }
1408
1409
1410
1411
1412 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1413 the phis in PRED. SEEN is a bitmap saying which expression we have
1414 translated since we started translation of the toplevel expression.
1415 Return NULL if we can't find a leader for each part of the
1416 translated expression. */
1417
1418 static pre_expr
1419 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1420 basic_block pred, basic_block phiblock, bitmap seen)
1421 {
1422 pre_expr oldexpr = expr;
1423 pre_expr phitrans;
1424
1425 if (!expr)
1426 return NULL;
1427
1428 if (value_id_constant_p (get_expr_value_id (expr)))
1429 return expr;
1430
1431 phitrans = phi_trans_lookup (expr, pred);
1432 if (phitrans)
1433 return phitrans;
1434
1435 /* Prevent cycles when we have recursively dependent leaders. This
1436 can only happen when phi translating the maximal set. */
1437 if (seen)
1438 {
1439 unsigned int expr_id = get_expression_id (expr);
1440 if (bitmap_bit_p (seen, expr_id))
1441 return NULL;
1442 bitmap_set_bit (seen, expr_id);
1443 }
1444
1445 switch (expr->kind)
1446 {
1447 /* Constants contain no values that need translation. */
1448 case CONSTANT:
1449 return expr;
1450
1451 case NARY:
1452 {
1453 unsigned int i;
1454 bool changed = false;
1455 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1456 struct vn_nary_op_s newnary;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1460 - sizeof (tree) * (4 - nary->length)));
1461
1462 for (i = 0; i < newnary.length; i++)
1463 {
1464 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1465 continue;
1466 else
1467 {
1468 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1469 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1470 pre_expr result = phi_translate_1 (leader, set1, set2,
1471 pred, phiblock, seen);
1472 if (result && result != leader)
1473 {
1474 tree name = get_representative_for (result);
1475 if (!name)
1476 return NULL;
1477 newnary.op[i] = name;
1478 }
1479 else if (!result)
1480 return NULL;
1481
1482 changed |= newnary.op[i] != nary->op[i];
1483 }
1484 }
1485 if (changed)
1486 {
1487 pre_expr constant;
1488
1489 tree result = vn_nary_op_lookup_pieces (newnary.length,
1490 newnary.opcode,
1491 newnary.type,
1492 newnary.op[0],
1493 newnary.op[1],
1494 newnary.op[2],
1495 newnary.op[3],
1496 &nary);
1497 unsigned int new_val_id;
1498
1499 expr = (pre_expr) pool_alloc (pre_expr_pool);
1500 expr->kind = NARY;
1501 expr->id = 0;
1502 if (result && is_gimple_min_invariant (result))
1503 return get_or_alloc_expr_for_constant (result);
1504
1505
1506 if (nary)
1507 {
1508 PRE_EXPR_NARY (expr) = nary;
1509 constant = fully_constant_expression (expr);
1510 if (constant != expr)
1511 return constant;
1512
1513 new_val_id = nary->value_id;
1514 get_or_alloc_expression_id (expr);
1515 }
1516 else
1517 {
1518 new_val_id = get_next_value_id ();
1519 VEC_safe_grow_cleared (bitmap_set_t, heap,
1520 value_expressions,
1521 get_max_value_id() + 1);
1522 nary = vn_nary_op_insert_pieces (newnary.length,
1523 newnary.opcode,
1524 newnary.type,
1525 newnary.op[0],
1526 newnary.op[1],
1527 newnary.op[2],
1528 newnary.op[3],
1529 result, new_val_id);
1530 PRE_EXPR_NARY (expr) = nary;
1531 constant = fully_constant_expression (expr);
1532 if (constant != expr)
1533 return constant;
1534 get_or_alloc_expression_id (expr);
1535 }
1536 add_to_value (new_val_id, expr);
1537 }
1538 phi_trans_add (oldexpr, expr, pred);
1539 return expr;
1540 }
1541 break;
1542
1543 case REFERENCE:
1544 {
1545 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1546 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1547 tree vuse = ref->vuse;
1548 tree newvuse = vuse;
1549 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1550 bool changed = false;
1551 unsigned int i, j;
1552 vn_reference_op_t operand;
1553 vn_reference_t newref;
1554
1555 for (i = 0, j = 0;
1556 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1557 {
1558 pre_expr opresult;
1559 pre_expr leader;
1560 tree oldop0 = operand->op0;
1561 tree oldop1 = operand->op1;
1562 tree oldop2 = operand->op2;
1563 tree op0 = oldop0;
1564 tree op1 = oldop1;
1565 tree op2 = oldop2;
1566 tree type = operand->type;
1567 vn_reference_op_s newop = *operand;
1568
1569 if (op0 && TREE_CODE (op0) == SSA_NAME)
1570 {
1571 unsigned int op_val_id = VN_INFO (op0)->value_id;
1572 leader = find_leader_in_sets (op_val_id, set1, set2);
1573 opresult = phi_translate_1 (leader, set1, set2,
1574 pred, phiblock, seen);
1575 if (opresult && opresult != leader)
1576 {
1577 tree name = get_representative_for (opresult);
1578 if (!name)
1579 break;
1580 op0 = name;
1581 }
1582 else if (!opresult)
1583 break;
1584 }
1585 changed |= op0 != oldop0;
1586
1587 if (op1 && TREE_CODE (op1) == SSA_NAME)
1588 {
1589 unsigned int op_val_id = VN_INFO (op1)->value_id;
1590 leader = find_leader_in_sets (op_val_id, set1, set2);
1591 opresult = phi_translate_1 (leader, set1, set2,
1592 pred, phiblock, seen);
1593 if (opresult && opresult != leader)
1594 {
1595 tree name = get_representative_for (opresult);
1596 if (!name)
1597 break;
1598 op1 = name;
1599 }
1600 else if (!opresult)
1601 break;
1602 }
1603 changed |= op1 != oldop1;
1604 if (op2 && TREE_CODE (op2) == SSA_NAME)
1605 {
1606 unsigned int op_val_id = VN_INFO (op2)->value_id;
1607 leader = find_leader_in_sets (op_val_id, set1, set2);
1608 opresult = phi_translate_1 (leader, set1, set2,
1609 pred, phiblock, seen);
1610 if (opresult && opresult != leader)
1611 {
1612 tree name = get_representative_for (opresult);
1613 if (!name)
1614 break;
1615 op2 = name;
1616 }
1617 else if (!opresult)
1618 break;
1619 }
1620 changed |= op2 != oldop2;
1621
1622 if (!newoperands)
1623 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1626 newop.opcode = TREE_CODE (op0);
1627 newop.type = type;
1628 newop.op0 = op0;
1629 newop.op1 = op1;
1630 newop.op2 = op2;
1631 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1632 /* If it transforms from an SSA_NAME to an address, fold with
1633 a preceding indirect reference. */
1634 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1635 && VEC_index (vn_reference_op_s,
1636 newoperands, j - 1)->opcode == INDIRECT_REF)
1637 vn_reference_fold_indirect (&newoperands, &j);
1638 }
1639 if (i != VEC_length (vn_reference_op_s, operands))
1640 {
1641 if (newoperands)
1642 VEC_free (vn_reference_op_s, heap, newoperands);
1643 return NULL;
1644 }
1645
1646 if (vuse)
1647 {
1648 newvuse = translate_vuse_through_block (newoperands,
1649 ref->set, ref->type,
1650 vuse, phiblock, pred);
1651 if (newvuse == NULL_TREE)
1652 {
1653 VEC_free (vn_reference_op_s, heap, newoperands);
1654 return NULL;
1655 }
1656 }
1657 changed |= newvuse != vuse;
1658
1659 if (changed)
1660 {
1661 unsigned int new_val_id;
1662 pre_expr constant;
1663
1664 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1665 ref->type,
1666 newoperands,
1667 &newref, true);
1668 if (newref)
1669 VEC_free (vn_reference_op_s, heap, newoperands);
1670
1671 if (result && is_gimple_min_invariant (result))
1672 {
1673 gcc_assert (!newoperands);
1674 return get_or_alloc_expr_for_constant (result);
1675 }
1676
1677 expr = (pre_expr) pool_alloc (pre_expr_pool);
1678 expr->kind = REFERENCE;
1679 expr->id = 0;
1680
1681 if (newref)
1682 {
1683 PRE_EXPR_REFERENCE (expr) = newref;
1684 constant = fully_constant_expression (expr);
1685 if (constant != expr)
1686 return constant;
1687
1688 new_val_id = newref->value_id;
1689 get_or_alloc_expression_id (expr);
1690 }
1691 else
1692 {
1693 new_val_id = get_next_value_id ();
1694 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1695 get_max_value_id() + 1);
1696 newref = vn_reference_insert_pieces (newvuse, ref->set,
1697 ref->type,
1698 newoperands,
1699 result, new_val_id);
1700 newoperands = NULL;
1701 PRE_EXPR_REFERENCE (expr) = newref;
1702 constant = fully_constant_expression (expr);
1703 if (constant != expr)
1704 return constant;
1705 get_or_alloc_expression_id (expr);
1706 }
1707 add_to_value (new_val_id, expr);
1708 }
1709 VEC_free (vn_reference_op_s, heap, newoperands);
1710 phi_trans_add (oldexpr, expr, pred);
1711 return expr;
1712 }
1713 break;
1714
1715 case NAME:
1716 {
1717 gimple phi = NULL;
1718 edge e;
1719 gimple def_stmt;
1720 tree name = PRE_EXPR_NAME (expr);
1721
1722 def_stmt = SSA_NAME_DEF_STMT (name);
1723 if (gimple_code (def_stmt) == GIMPLE_PHI
1724 && gimple_bb (def_stmt) == phiblock)
1725 phi = def_stmt;
1726 else
1727 return expr;
1728
1729 e = find_edge (pred, gimple_bb (phi));
1730 if (e)
1731 {
1732 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1733 pre_expr newexpr;
1734
1735 if (TREE_CODE (def) == SSA_NAME)
1736 def = VN_INFO (def)->valnum;
1737
1738 /* Handle constant. */
1739 if (is_gimple_min_invariant (def))
1740 return get_or_alloc_expr_for_constant (def);
1741
1742 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1743 return NULL;
1744
1745 newexpr = get_or_alloc_expr_for_name (def);
1746 return newexpr;
1747 }
1748 }
1749 return expr;
1750
1751 default:
1752 gcc_unreachable ();
1753 }
1754 }
1755
1756 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1757 the phis in PRED.
1758 Return NULL if we can't find a leader for each part of the
1759 translated expression. */
1760
1761 static pre_expr
1762 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1763 basic_block pred, basic_block phiblock)
1764 {
1765 bitmap_clear (seen_during_translate);
1766 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1767 seen_during_translate);
1768 }
1769
1770 /* For each expression in SET, translate the values through phi nodes
1771 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1772 expressions in DEST. */
1773
1774 static void
1775 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1776 basic_block phiblock)
1777 {
1778 VEC (pre_expr, heap) *exprs;
1779 pre_expr expr;
1780 int i;
1781
1782 if (!phi_nodes (phiblock))
1783 {
1784 bitmap_set_copy (dest, set);
1785 return;
1786 }
1787
1788 exprs = sorted_array_from_bitmap_set (set);
1789 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1790 {
1791 pre_expr translated;
1792 translated = phi_translate (expr, set, NULL, pred, phiblock);
1793
1794 /* Don't add empty translations to the cache */
1795 if (translated)
1796 phi_trans_add (expr, translated, pred);
1797
1798 if (translated != NULL)
1799 bitmap_value_insert_into_set (dest, translated);
1800 }
1801 VEC_free (pre_expr, heap, exprs);
1802 }
1803
1804 /* Find the leader for a value (i.e., the name representing that
1805 value) in a given set, and return it. If STMT is non-NULL it
1806 makes sure the defining statement for the leader dominates it.
1807 Return NULL if no leader is found. */
1808
1809 static pre_expr
1810 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1811 {
1812 if (value_id_constant_p (val))
1813 {
1814 unsigned int i;
1815 bitmap_iterator bi;
1816 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1817
1818 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1819 {
1820 pre_expr expr = expression_for_id (i);
1821 if (expr->kind == CONSTANT)
1822 return expr;
1823 }
1824 }
1825 if (bitmap_set_contains_value (set, val))
1826 {
1827 /* Rather than walk the entire bitmap of expressions, and see
1828 whether any of them has the value we are looking for, we look
1829 at the reverse mapping, which tells us the set of expressions
1830 that have a given value (IE value->expressions with that
1831 value) and see if any of those expressions are in our set.
1832 The number of expressions per value is usually significantly
1833 less than the number of expressions in the set. In fact, for
1834 large testcases, doing it this way is roughly 5-10x faster
1835 than walking the bitmap.
1836 If this is somehow a significant lose for some cases, we can
1837 choose which set to walk based on which set is smaller. */
1838 unsigned int i;
1839 bitmap_iterator bi;
1840 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1841
1842 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1843 set->expressions, 0, i, bi)
1844 {
1845 pre_expr val = expression_for_id (i);
1846 /* At the point where stmt is not null, there should always
1847 be an SSA_NAME first in the list of expressions. */
1848 if (stmt)
1849 {
1850 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1851 if (gimple_code (def_stmt) != GIMPLE_PHI
1852 && gimple_bb (def_stmt) == gimple_bb (stmt)
1853 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1854 continue;
1855 }
1856 return val;
1857 }
1858 }
1859 return NULL;
1860 }
1861
1862 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1863 BLOCK by seeing if it is not killed in the block. Note that we are
1864 only determining whether there is a store that kills it. Because
1865 of the order in which clean iterates over values, we are guaranteed
1866 that altered operands will have caused us to be eliminated from the
1867 ANTIC_IN set already. */
1868
1869 static bool
1870 value_dies_in_block_x (pre_expr expr, basic_block block)
1871 {
1872 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1873 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1874 gimple def;
1875 gimple_stmt_iterator gsi;
1876 unsigned id = get_expression_id (expr);
1877 bool res = false;
1878 ao_ref ref;
1879
1880 if (!vuse)
1881 return false;
1882
1883 /* Lookup a previously calculated result. */
1884 if (EXPR_DIES (block)
1885 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1886 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1887
1888 /* A memory expression {e, VUSE} dies in the block if there is a
1889 statement that may clobber e. If, starting statement walk from the
1890 top of the basic block, a statement uses VUSE there can be no kill
1891 inbetween that use and the original statement that loaded {e, VUSE},
1892 so we can stop walking. */
1893 ref.base = NULL_TREE;
1894 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1895 {
1896 tree def_vuse, def_vdef;
1897 def = gsi_stmt (gsi);
1898 def_vuse = gimple_vuse (def);
1899 def_vdef = gimple_vdef (def);
1900
1901 /* Not a memory statement. */
1902 if (!def_vuse)
1903 continue;
1904
1905 /* Not a may-def. */
1906 if (!def_vdef)
1907 {
1908 /* A load with the same VUSE, we're done. */
1909 if (def_vuse == vuse)
1910 break;
1911
1912 continue;
1913 }
1914
1915 /* Init ref only if we really need it. */
1916 if (ref.base == NULL_TREE
1917 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1918 refx->operands))
1919 {
1920 res = true;
1921 break;
1922 }
1923 /* If the statement may clobber expr, it dies. */
1924 if (stmt_may_clobber_ref_p_1 (def, &ref))
1925 {
1926 res = true;
1927 break;
1928 }
1929 }
1930
1931 /* Remember the result. */
1932 if (!EXPR_DIES (block))
1933 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1934 bitmap_set_bit (EXPR_DIES (block), id * 2);
1935 if (res)
1936 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1937
1938 return res;
1939 }
1940
1941
1942 #define union_contains_value(SET1, SET2, VAL) \
1943 (bitmap_set_contains_value ((SET1), (VAL)) \
1944 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1945
1946 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1947 */
1948 static bool
1949 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1950 vn_reference_op_t vro)
1951 {
1952 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1953 {
1954 struct pre_expr_d temp;
1955 temp.kind = NAME;
1956 temp.id = 0;
1957 PRE_EXPR_NAME (&temp) = vro->op0;
1958 temp.id = lookup_expression_id (&temp);
1959 if (temp.id == 0)
1960 return false;
1961 if (!union_contains_value (set1, set2,
1962 get_expr_value_id (&temp)))
1963 return false;
1964 }
1965 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1966 {
1967 struct pre_expr_d temp;
1968 temp.kind = NAME;
1969 temp.id = 0;
1970 PRE_EXPR_NAME (&temp) = vro->op1;
1971 temp.id = lookup_expression_id (&temp);
1972 if (temp.id == 0)
1973 return false;
1974 if (!union_contains_value (set1, set2,
1975 get_expr_value_id (&temp)))
1976 return false;
1977 }
1978
1979 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1980 {
1981 struct pre_expr_d temp;
1982 temp.kind = NAME;
1983 temp.id = 0;
1984 PRE_EXPR_NAME (&temp) = vro->op2;
1985 temp.id = lookup_expression_id (&temp);
1986 if (temp.id == 0)
1987 return false;
1988 if (!union_contains_value (set1, set2,
1989 get_expr_value_id (&temp)))
1990 return false;
1991 }
1992
1993 return true;
1994 }
1995
1996 /* Determine if the expression EXPR is valid in SET1 U SET2.
1997 ONLY SET2 CAN BE NULL.
1998 This means that we have a leader for each part of the expression
1999 (if it consists of values), or the expression is an SSA_NAME.
2000 For loads/calls, we also see if the vuse is killed in this block. */
2001
2002 static bool
2003 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2004 basic_block block)
2005 {
2006 switch (expr->kind)
2007 {
2008 case NAME:
2009 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2010 case NARY:
2011 {
2012 unsigned int i;
2013 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2014 for (i = 0; i < nary->length; i++)
2015 {
2016 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2017 {
2018 struct pre_expr_d temp;
2019 temp.kind = NAME;
2020 temp.id = 0;
2021 PRE_EXPR_NAME (&temp) = nary->op[i];
2022 temp.id = lookup_expression_id (&temp);
2023 if (temp.id == 0)
2024 return false;
2025 if (!union_contains_value (set1, set2,
2026 get_expr_value_id (&temp)))
2027 return false;
2028 }
2029 }
2030 return true;
2031 }
2032 break;
2033 case REFERENCE:
2034 {
2035 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2036 vn_reference_op_t vro;
2037 unsigned int i;
2038
2039 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2040 {
2041 if (!vro_valid_in_sets (set1, set2, vro))
2042 return false;
2043 }
2044 if (ref->vuse)
2045 {
2046 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2047 if (!gimple_nop_p (def_stmt)
2048 && gimple_bb (def_stmt) != block
2049 && !dominated_by_p (CDI_DOMINATORS,
2050 block, gimple_bb (def_stmt)))
2051 return false;
2052 }
2053 return !value_dies_in_block_x (expr, block);
2054 }
2055 default:
2056 gcc_unreachable ();
2057 }
2058 }
2059
2060 /* Clean the set of expressions that are no longer valid in SET1 or
2061 SET2. This means expressions that are made up of values we have no
2062 leaders for in SET1 or SET2. This version is used for partial
2063 anticipation, which means it is not valid in either ANTIC_IN or
2064 PA_IN. */
2065
2066 static void
2067 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2068 {
2069 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2070 pre_expr expr;
2071 int i;
2072
2073 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2074 {
2075 if (!valid_in_sets (set1, set2, expr, block))
2076 bitmap_remove_from_set (set1, expr);
2077 }
2078 VEC_free (pre_expr, heap, exprs);
2079 }
2080
2081 /* Clean the set of expressions that are no longer valid in SET. This
2082 means expressions that are made up of values we have no leaders for
2083 in SET. */
2084
2085 static void
2086 clean (bitmap_set_t set, basic_block block)
2087 {
2088 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2089 pre_expr expr;
2090 int i;
2091
2092 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2093 {
2094 if (!valid_in_sets (set, NULL, expr, block))
2095 bitmap_remove_from_set (set, expr);
2096 }
2097 VEC_free (pre_expr, heap, exprs);
2098 }
2099
2100 static sbitmap has_abnormal_preds;
2101
2102 /* List of blocks that may have changed during ANTIC computation and
2103 thus need to be iterated over. */
2104
2105 static sbitmap changed_blocks;
2106
2107 /* Decide whether to defer a block for a later iteration, or PHI
2108 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2109 should defer the block, and true if we processed it. */
2110
2111 static bool
2112 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2113 basic_block block, basic_block phiblock)
2114 {
2115 if (!BB_VISITED (phiblock))
2116 {
2117 SET_BIT (changed_blocks, block->index);
2118 BB_VISITED (block) = 0;
2119 BB_DEFERRED (block) = 1;
2120 return false;
2121 }
2122 else
2123 phi_translate_set (dest, source, block, phiblock);
2124 return true;
2125 }
2126
2127 /* Compute the ANTIC set for BLOCK.
2128
2129 If succs(BLOCK) > 1 then
2130 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2131 else if succs(BLOCK) == 1 then
2132 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2133
2134 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2135 */
2136
2137 static bool
2138 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2139 {
2140 bool changed = false;
2141 bitmap_set_t S, old, ANTIC_OUT;
2142 bitmap_iterator bi;
2143 unsigned int bii;
2144 edge e;
2145 edge_iterator ei;
2146
2147 old = ANTIC_OUT = S = NULL;
2148 BB_VISITED (block) = 1;
2149
2150 /* If any edges from predecessors are abnormal, antic_in is empty,
2151 so do nothing. */
2152 if (block_has_abnormal_pred_edge)
2153 goto maybe_dump_sets;
2154
2155 old = ANTIC_IN (block);
2156 ANTIC_OUT = bitmap_set_new ();
2157
2158 /* If the block has no successors, ANTIC_OUT is empty. */
2159 if (EDGE_COUNT (block->succs) == 0)
2160 ;
2161 /* If we have one successor, we could have some phi nodes to
2162 translate through. */
2163 else if (single_succ_p (block))
2164 {
2165 basic_block succ_bb = single_succ (block);
2166
2167 /* We trade iterations of the dataflow equations for having to
2168 phi translate the maximal set, which is incredibly slow
2169 (since the maximal set often has 300+ members, even when you
2170 have a small number of blocks).
2171 Basically, we defer the computation of ANTIC for this block
2172 until we have processed it's successor, which will inevitably
2173 have a *much* smaller set of values to phi translate once
2174 clean has been run on it.
2175 The cost of doing this is that we technically perform more
2176 iterations, however, they are lower cost iterations.
2177
2178 Timings for PRE on tramp3d-v4:
2179 without maximal set fix: 11 seconds
2180 with maximal set fix/without deferring: 26 seconds
2181 with maximal set fix/with deferring: 11 seconds
2182 */
2183
2184 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2185 block, succ_bb))
2186 {
2187 changed = true;
2188 goto maybe_dump_sets;
2189 }
2190 }
2191 /* If we have multiple successors, we take the intersection of all of
2192 them. Note that in the case of loop exit phi nodes, we may have
2193 phis to translate through. */
2194 else
2195 {
2196 VEC(basic_block, heap) * worklist;
2197 size_t i;
2198 basic_block bprime, first;
2199
2200 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2201 FOR_EACH_EDGE (e, ei, block->succs)
2202 VEC_quick_push (basic_block, worklist, e->dest);
2203 first = VEC_index (basic_block, worklist, 0);
2204
2205 if (phi_nodes (first))
2206 {
2207 bitmap_set_t from = ANTIC_IN (first);
2208
2209 if (!BB_VISITED (first))
2210 from = maximal_set;
2211 phi_translate_set (ANTIC_OUT, from, block, first);
2212 }
2213 else
2214 {
2215 if (!BB_VISITED (first))
2216 bitmap_set_copy (ANTIC_OUT, maximal_set);
2217 else
2218 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2219 }
2220
2221 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
2222 {
2223 if (phi_nodes (bprime))
2224 {
2225 bitmap_set_t tmp = bitmap_set_new ();
2226 bitmap_set_t from = ANTIC_IN (bprime);
2227
2228 if (!BB_VISITED (bprime))
2229 from = maximal_set;
2230 phi_translate_set (tmp, from, block, bprime);
2231 bitmap_set_and (ANTIC_OUT, tmp);
2232 bitmap_set_free (tmp);
2233 }
2234 else
2235 {
2236 if (!BB_VISITED (bprime))
2237 bitmap_set_and (ANTIC_OUT, maximal_set);
2238 else
2239 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2240 }
2241 }
2242 VEC_free (basic_block, heap, worklist);
2243 }
2244
2245 /* Generate ANTIC_OUT - TMP_GEN. */
2246 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2247
2248 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2249 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2250 TMP_GEN (block));
2251
2252 /* Then union in the ANTIC_OUT - TMP_GEN values,
2253 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2254 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2255 bitmap_value_insert_into_set (ANTIC_IN (block),
2256 expression_for_id (bii));
2257
2258 clean (ANTIC_IN (block), block);
2259
2260 /* !old->expressions can happen when we deferred a block. */
2261 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2262 {
2263 changed = true;
2264 SET_BIT (changed_blocks, block->index);
2265 FOR_EACH_EDGE (e, ei, block->preds)
2266 SET_BIT (changed_blocks, e->src->index);
2267 }
2268 else
2269 RESET_BIT (changed_blocks, block->index);
2270
2271 maybe_dump_sets:
2272 if (dump_file && (dump_flags & TDF_DETAILS))
2273 {
2274 if (!BB_DEFERRED (block) || BB_VISITED (block))
2275 {
2276 if (ANTIC_OUT)
2277 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2278
2279 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2280 block->index);
2281
2282 if (S)
2283 print_bitmap_set (dump_file, S, "S", block->index);
2284 }
2285 else
2286 {
2287 fprintf (dump_file,
2288 "Block %d was deferred for a future iteration.\n",
2289 block->index);
2290 }
2291 }
2292 if (old)
2293 bitmap_set_free (old);
2294 if (S)
2295 bitmap_set_free (S);
2296 if (ANTIC_OUT)
2297 bitmap_set_free (ANTIC_OUT);
2298 return changed;
2299 }
2300
2301 /* Compute PARTIAL_ANTIC for BLOCK.
2302
2303 If succs(BLOCK) > 1 then
2304 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2305 in ANTIC_OUT for all succ(BLOCK)
2306 else if succs(BLOCK) == 1 then
2307 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2308
2309 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2310 - ANTIC_IN[BLOCK])
2311
2312 */
2313 static bool
2314 compute_partial_antic_aux (basic_block block,
2315 bool block_has_abnormal_pred_edge)
2316 {
2317 bool changed = false;
2318 bitmap_set_t old_PA_IN;
2319 bitmap_set_t PA_OUT;
2320 edge e;
2321 edge_iterator ei;
2322 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2323
2324 old_PA_IN = PA_OUT = NULL;
2325
2326 /* If any edges from predecessors are abnormal, antic_in is empty,
2327 so do nothing. */
2328 if (block_has_abnormal_pred_edge)
2329 goto maybe_dump_sets;
2330
2331 /* If there are too many partially anticipatable values in the
2332 block, phi_translate_set can take an exponential time: stop
2333 before the translation starts. */
2334 if (max_pa
2335 && single_succ_p (block)
2336 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2337 goto maybe_dump_sets;
2338
2339 old_PA_IN = PA_IN (block);
2340 PA_OUT = bitmap_set_new ();
2341
2342 /* If the block has no successors, ANTIC_OUT is empty. */
2343 if (EDGE_COUNT (block->succs) == 0)
2344 ;
2345 /* If we have one successor, we could have some phi nodes to
2346 translate through. Note that we can't phi translate across DFS
2347 back edges in partial antic, because it uses a union operation on
2348 the successors. For recurrences like IV's, we will end up
2349 generating a new value in the set on each go around (i + 3 (VH.1)
2350 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2351 else if (single_succ_p (block))
2352 {
2353 basic_block succ = single_succ (block);
2354 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2355 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2356 }
2357 /* If we have multiple successors, we take the union of all of
2358 them. */
2359 else
2360 {
2361 VEC(basic_block, heap) * worklist;
2362 size_t i;
2363 basic_block bprime;
2364
2365 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2366 FOR_EACH_EDGE (e, ei, block->succs)
2367 {
2368 if (e->flags & EDGE_DFS_BACK)
2369 continue;
2370 VEC_quick_push (basic_block, worklist, e->dest);
2371 }
2372 if (VEC_length (basic_block, worklist) > 0)
2373 {
2374 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2375 {
2376 unsigned int i;
2377 bitmap_iterator bi;
2378
2379 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2380 bitmap_value_insert_into_set (PA_OUT,
2381 expression_for_id (i));
2382 if (phi_nodes (bprime))
2383 {
2384 bitmap_set_t pa_in = bitmap_set_new ();
2385 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2386 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2387 bitmap_value_insert_into_set (PA_OUT,
2388 expression_for_id (i));
2389 bitmap_set_free (pa_in);
2390 }
2391 else
2392 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2393 bitmap_value_insert_into_set (PA_OUT,
2394 expression_for_id (i));
2395 }
2396 }
2397 VEC_free (basic_block, heap, worklist);
2398 }
2399
2400 /* PA_IN starts with PA_OUT - TMP_GEN.
2401 Then we subtract things from ANTIC_IN. */
2402 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2403
2404 /* For partial antic, we want to put back in the phi results, since
2405 we will properly avoid making them partially antic over backedges. */
2406 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2407 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2408
2409 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2410 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2411
2412 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2413
2414 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2415 {
2416 changed = true;
2417 SET_BIT (changed_blocks, block->index);
2418 FOR_EACH_EDGE (e, ei, block->preds)
2419 SET_BIT (changed_blocks, e->src->index);
2420 }
2421 else
2422 RESET_BIT (changed_blocks, block->index);
2423
2424 maybe_dump_sets:
2425 if (dump_file && (dump_flags & TDF_DETAILS))
2426 {
2427 if (PA_OUT)
2428 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2429
2430 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2431 }
2432 if (old_PA_IN)
2433 bitmap_set_free (old_PA_IN);
2434 if (PA_OUT)
2435 bitmap_set_free (PA_OUT);
2436 return changed;
2437 }
2438
2439 /* Compute ANTIC and partial ANTIC sets. */
2440
2441 static void
2442 compute_antic (void)
2443 {
2444 bool changed = true;
2445 int num_iterations = 0;
2446 basic_block block;
2447 int i;
2448
2449 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2450 We pre-build the map of blocks with incoming abnormal edges here. */
2451 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2452 sbitmap_zero (has_abnormal_preds);
2453
2454 FOR_EACH_BB (block)
2455 {
2456 edge_iterator ei;
2457 edge e;
2458
2459 FOR_EACH_EDGE (e, ei, block->preds)
2460 {
2461 e->flags &= ~EDGE_DFS_BACK;
2462 if (e->flags & EDGE_ABNORMAL)
2463 {
2464 SET_BIT (has_abnormal_preds, block->index);
2465 break;
2466 }
2467 }
2468
2469 BB_VISITED (block) = 0;
2470 BB_DEFERRED (block) = 0;
2471 /* While we are here, give empty ANTIC_IN sets to each block. */
2472 ANTIC_IN (block) = bitmap_set_new ();
2473 PA_IN (block) = bitmap_set_new ();
2474 }
2475
2476 /* At the exit block we anticipate nothing. */
2477 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2478 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2479 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2480
2481 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2482 sbitmap_ones (changed_blocks);
2483 while (changed)
2484 {
2485 if (dump_file && (dump_flags & TDF_DETAILS))
2486 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2487 num_iterations++;
2488 changed = false;
2489 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2490 {
2491 if (TEST_BIT (changed_blocks, postorder[i]))
2492 {
2493 basic_block block = BASIC_BLOCK (postorder[i]);
2494 changed |= compute_antic_aux (block,
2495 TEST_BIT (has_abnormal_preds,
2496 block->index));
2497 }
2498 }
2499 #ifdef ENABLE_CHECKING
2500 /* Theoretically possible, but *highly* unlikely. */
2501 gcc_assert (num_iterations < 500);
2502 #endif
2503 }
2504
2505 statistics_histogram_event (cfun, "compute_antic iterations",
2506 num_iterations);
2507
2508 if (do_partial_partial)
2509 {
2510 sbitmap_ones (changed_blocks);
2511 mark_dfs_back_edges ();
2512 num_iterations = 0;
2513 changed = true;
2514 while (changed)
2515 {
2516 if (dump_file && (dump_flags & TDF_DETAILS))
2517 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2518 num_iterations++;
2519 changed = false;
2520 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2521 {
2522 if (TEST_BIT (changed_blocks, postorder[i]))
2523 {
2524 basic_block block = BASIC_BLOCK (postorder[i]);
2525 changed
2526 |= compute_partial_antic_aux (block,
2527 TEST_BIT (has_abnormal_preds,
2528 block->index));
2529 }
2530 }
2531 #ifdef ENABLE_CHECKING
2532 /* Theoretically possible, but *highly* unlikely. */
2533 gcc_assert (num_iterations < 500);
2534 #endif
2535 }
2536 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2537 num_iterations);
2538 }
2539 sbitmap_free (has_abnormal_preds);
2540 sbitmap_free (changed_blocks);
2541 }
2542
2543 /* Return true if we can value number the call in STMT. This is true
2544 if we have a pure or constant call. */
2545
2546 static bool
2547 can_value_number_call (gimple stmt)
2548 {
2549 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2550 return true;
2551 return false;
2552 }
2553
2554 /* Return true if OP is an exception handler related operation, such as
2555 FILTER_EXPR or EXC_PTR_EXPR. */
2556
2557 static bool
2558 is_exception_related (gimple stmt)
2559 {
2560 return (is_gimple_assign (stmt)
2561 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2562 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2563 }
2564
2565 /* Return true if OP is a tree which we can perform PRE on.
2566 This may not match the operations we can value number, but in
2567 a perfect world would. */
2568
2569 static bool
2570 can_PRE_operation (tree op)
2571 {
2572 return UNARY_CLASS_P (op)
2573 || BINARY_CLASS_P (op)
2574 || COMPARISON_CLASS_P (op)
2575 || TREE_CODE (op) == INDIRECT_REF
2576 || TREE_CODE (op) == COMPONENT_REF
2577 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2578 || TREE_CODE (op) == CALL_EXPR
2579 || TREE_CODE (op) == ARRAY_REF;
2580 }
2581
2582
2583 /* Inserted expressions are placed onto this worklist, which is used
2584 for performing quick dead code elimination of insertions we made
2585 that didn't turn out to be necessary. */
2586 static VEC(gimple,heap) *inserted_exprs;
2587 static bitmap inserted_phi_names;
2588
2589 /* Pool allocated fake store expressions are placed onto this
2590 worklist, which, after performing dead code elimination, is walked
2591 to see which expressions need to be put into GC'able memory */
2592 static VEC(gimple, heap) *need_creation;
2593
2594 /* The actual worker for create_component_ref_by_pieces. */
2595
2596 static tree
2597 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2598 unsigned int *operand, gimple_seq *stmts,
2599 gimple domstmt)
2600 {
2601 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2602 *operand);
2603 tree genop;
2604 ++*operand;
2605 switch (currop->opcode)
2606 {
2607 case CALL_EXPR:
2608 {
2609 tree folded, sc = currop->op1;
2610 unsigned int nargs = 0;
2611 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2612 ref->operands) - 1);
2613 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2614 {
2615 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2616 operand, stmts,
2617 domstmt);
2618 nargs++;
2619 }
2620 folded = build_call_array (currop->type,
2621 TREE_CODE (currop->op0) == FUNCTION_DECL
2622 ? build_fold_addr_expr (currop->op0)
2623 : currop->op0,
2624 nargs, args);
2625 free (args);
2626 if (sc)
2627 {
2628 pre_expr scexpr = get_or_alloc_expr_for (sc);
2629 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2630 if (!sc)
2631 return NULL_TREE;
2632 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2633 }
2634 return folded;
2635 }
2636 break;
2637 case TARGET_MEM_REF:
2638 {
2639 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2640 *operand);
2641 pre_expr op0expr;
2642 tree genop0 = NULL_TREE;
2643 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2644 stmts, domstmt);
2645 if (!baseop)
2646 return NULL_TREE;
2647 if (currop->op0)
2648 {
2649 op0expr = get_or_alloc_expr_for (currop->op0);
2650 genop0 = find_or_generate_expression (block, op0expr,
2651 stmts, domstmt);
2652 if (!genop0)
2653 return NULL_TREE;
2654 }
2655 if (DECL_P (baseop))
2656 return build6 (TARGET_MEM_REF, currop->type,
2657 baseop, NULL_TREE,
2658 genop0, currop->op1, currop->op2,
2659 unshare_expr (nextop->op1));
2660 else
2661 return build6 (TARGET_MEM_REF, currop->type,
2662 NULL_TREE, baseop,
2663 genop0, currop->op1, currop->op2,
2664 unshare_expr (nextop->op1));
2665 }
2666 break;
2667 case ADDR_EXPR:
2668 if (currop->op0)
2669 {
2670 gcc_assert (is_gimple_min_invariant (currop->op0));
2671 return currop->op0;
2672 }
2673 /* Fallthrough. */
2674 case REALPART_EXPR:
2675 case IMAGPART_EXPR:
2676 case VIEW_CONVERT_EXPR:
2677 {
2678 tree folded;
2679 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2680 operand,
2681 stmts, domstmt);
2682 if (!genop0)
2683 return NULL_TREE;
2684 folded = fold_build1 (currop->opcode, currop->type,
2685 genop0);
2686 return folded;
2687 }
2688 break;
2689 case ALIGN_INDIRECT_REF:
2690 case MISALIGNED_INDIRECT_REF:
2691 case INDIRECT_REF:
2692 {
2693 tree folded;
2694 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2695 operand,
2696 stmts, domstmt);
2697 if (!genop1)
2698 return NULL_TREE;
2699 genop1 = fold_convert (build_pointer_type (currop->type),
2700 genop1);
2701
2702 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2703 folded = fold_build2 (currop->opcode, currop->type,
2704 genop1, currop->op1);
2705 else
2706 folded = fold_build1 (currop->opcode, currop->type,
2707 genop1);
2708 return folded;
2709 }
2710 break;
2711 case BIT_FIELD_REF:
2712 {
2713 tree folded;
2714 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2715 stmts, domstmt);
2716 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2717 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2718 tree genop1;
2719 tree genop2;
2720
2721 if (!genop0)
2722 return NULL_TREE;
2723 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2724 if (!genop1)
2725 return NULL_TREE;
2726 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2727 if (!genop2)
2728 return NULL_TREE;
2729 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2730 genop2);
2731 return folded;
2732 }
2733
2734 /* For array ref vn_reference_op's, operand 1 of the array ref
2735 is op0 of the reference op and operand 3 of the array ref is
2736 op1. */
2737 case ARRAY_RANGE_REF:
2738 case ARRAY_REF:
2739 {
2740 tree genop0;
2741 tree genop1 = currop->op0;
2742 pre_expr op1expr;
2743 tree genop2 = currop->op1;
2744 pre_expr op2expr;
2745 tree genop3;
2746 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2747 stmts, domstmt);
2748 if (!genop0)
2749 return NULL_TREE;
2750 op1expr = get_or_alloc_expr_for (genop1);
2751 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2752 if (!genop1)
2753 return NULL_TREE;
2754 if (genop2)
2755 {
2756 op2expr = get_or_alloc_expr_for (genop2);
2757 genop2 = find_or_generate_expression (block, op2expr, stmts,
2758 domstmt);
2759 if (!genop2)
2760 return NULL_TREE;
2761 }
2762
2763 genop3 = currop->op2;
2764 return build4 (currop->opcode, currop->type, genop0, genop1,
2765 genop2, genop3);
2766 }
2767 case COMPONENT_REF:
2768 {
2769 tree op0;
2770 tree op1;
2771 tree genop2 = currop->op1;
2772 pre_expr op2expr;
2773 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2774 stmts, domstmt);
2775 if (!op0)
2776 return NULL_TREE;
2777 /* op1 should be a FIELD_DECL, which are represented by
2778 themselves. */
2779 op1 = currop->op0;
2780 if (genop2)
2781 {
2782 op2expr = get_or_alloc_expr_for (genop2);
2783 genop2 = find_or_generate_expression (block, op2expr, stmts,
2784 domstmt);
2785 if (!genop2)
2786 return NULL_TREE;
2787 }
2788
2789 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2790 genop2);
2791 }
2792 break;
2793 case SSA_NAME:
2794 {
2795 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2796 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2797 return genop;
2798 }
2799 case STRING_CST:
2800 case INTEGER_CST:
2801 case COMPLEX_CST:
2802 case VECTOR_CST:
2803 case REAL_CST:
2804 case CONSTRUCTOR:
2805 case VAR_DECL:
2806 case PARM_DECL:
2807 case CONST_DECL:
2808 case RESULT_DECL:
2809 case FUNCTION_DECL:
2810 return currop->op0;
2811
2812 default:
2813 gcc_unreachable ();
2814 }
2815 }
2816
2817 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2818 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2819 trying to rename aggregates into ssa form directly, which is a no no.
2820
2821 Thus, this routine doesn't create temporaries, it just builds a
2822 single access expression for the array, calling
2823 find_or_generate_expression to build the innermost pieces.
2824
2825 This function is a subroutine of create_expression_by_pieces, and
2826 should not be called on it's own unless you really know what you
2827 are doing. */
2828
2829 static tree
2830 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2831 gimple_seq *stmts, gimple domstmt)
2832 {
2833 unsigned int op = 0;
2834 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2835 }
2836
2837 /* Find a leader for an expression, or generate one using
2838 create_expression_by_pieces if it's ANTIC but
2839 complex.
2840 BLOCK is the basic_block we are looking for leaders in.
2841 EXPR is the expression to find a leader or generate for.
2842 STMTS is the statement list to put the inserted expressions on.
2843 Returns the SSA_NAME of the LHS of the generated expression or the
2844 leader.
2845 DOMSTMT if non-NULL is a statement that should be dominated by
2846 all uses in the generated expression. If DOMSTMT is non-NULL this
2847 routine can fail and return NULL_TREE. Otherwise it will assert
2848 on failure. */
2849
2850 static tree
2851 find_or_generate_expression (basic_block block, pre_expr expr,
2852 gimple_seq *stmts, gimple domstmt)
2853 {
2854 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2855 get_expr_value_id (expr), domstmt);
2856 tree genop = NULL;
2857 if (leader)
2858 {
2859 if (leader->kind == NAME)
2860 genop = PRE_EXPR_NAME (leader);
2861 else if (leader->kind == CONSTANT)
2862 genop = PRE_EXPR_CONSTANT (leader);
2863 }
2864
2865 /* If it's still NULL, it must be a complex expression, so generate
2866 it recursively. Not so for FRE though. */
2867 if (genop == NULL
2868 && !in_fre)
2869 {
2870 bitmap_set_t exprset;
2871 unsigned int lookfor = get_expr_value_id (expr);
2872 bool handled = false;
2873 bitmap_iterator bi;
2874 unsigned int i;
2875
2876 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2877 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2878 {
2879 pre_expr temp = expression_for_id (i);
2880 if (temp->kind != NAME)
2881 {
2882 handled = true;
2883 genop = create_expression_by_pieces (block, temp, stmts,
2884 domstmt,
2885 get_expr_type (expr));
2886 break;
2887 }
2888 }
2889 if (!handled && domstmt)
2890 return NULL_TREE;
2891
2892 gcc_assert (handled);
2893 }
2894 return genop;
2895 }
2896
2897 #define NECESSARY GF_PLF_1
2898
2899 /* Create an expression in pieces, so that we can handle very complex
2900 expressions that may be ANTIC, but not necessary GIMPLE.
2901 BLOCK is the basic block the expression will be inserted into,
2902 EXPR is the expression to insert (in value form)
2903 STMTS is a statement list to append the necessary insertions into.
2904
2905 This function will die if we hit some value that shouldn't be
2906 ANTIC but is (IE there is no leader for it, or its components).
2907 This function may also generate expressions that are themselves
2908 partially or fully redundant. Those that are will be either made
2909 fully redundant during the next iteration of insert (for partially
2910 redundant ones), or eliminated by eliminate (for fully redundant
2911 ones).
2912
2913 If DOMSTMT is non-NULL then we make sure that all uses in the
2914 expressions dominate that statement. In this case the function
2915 can return NULL_TREE to signal failure. */
2916
2917 static tree
2918 create_expression_by_pieces (basic_block block, pre_expr expr,
2919 gimple_seq *stmts, gimple domstmt, tree type)
2920 {
2921 tree temp, name;
2922 tree folded;
2923 gimple_seq forced_stmts = NULL;
2924 unsigned int value_id;
2925 gimple_stmt_iterator gsi;
2926 tree exprtype = type ? type : get_expr_type (expr);
2927 pre_expr nameexpr;
2928 gimple newstmt;
2929
2930 switch (expr->kind)
2931 {
2932 /* We may hit the NAME/CONSTANT case if we have to convert types
2933 that value numbering saw through. */
2934 case NAME:
2935 folded = PRE_EXPR_NAME (expr);
2936 break;
2937 case CONSTANT:
2938 folded = PRE_EXPR_CONSTANT (expr);
2939 break;
2940 case REFERENCE:
2941 {
2942 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2943 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2944 }
2945 break;
2946 case NARY:
2947 {
2948 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2949 switch (nary->length)
2950 {
2951 case 2:
2952 {
2953 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2954 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2955 tree genop1 = find_or_generate_expression (block, op1,
2956 stmts, domstmt);
2957 tree genop2 = find_or_generate_expression (block, op2,
2958 stmts, domstmt);
2959 if (!genop1 || !genop2)
2960 return NULL_TREE;
2961 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2962 genop1);
2963 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2964 may be a constant with the wrong type. */
2965 if (nary->opcode == POINTER_PLUS_EXPR)
2966 genop2 = fold_convert (sizetype, genop2);
2967 else
2968 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2969
2970 folded = fold_build2 (nary->opcode, nary->type,
2971 genop1, genop2);
2972 }
2973 break;
2974 case 1:
2975 {
2976 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2977 tree genop1 = find_or_generate_expression (block, op1,
2978 stmts, domstmt);
2979 if (!genop1)
2980 return NULL_TREE;
2981 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2982
2983 folded = fold_build1 (nary->opcode, nary->type,
2984 genop1);
2985 }
2986 break;
2987 default:
2988 return NULL_TREE;
2989 }
2990 }
2991 break;
2992 default:
2993 return NULL_TREE;
2994 }
2995
2996 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2997 folded = fold_convert (exprtype, folded);
2998
2999 /* Force the generated expression to be a sequence of GIMPLE
3000 statements.
3001 We have to call unshare_expr because force_gimple_operand may
3002 modify the tree we pass to it. */
3003 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3004 false, NULL);
3005
3006 /* If we have any intermediate expressions to the value sets, add them
3007 to the value sets and chain them in the instruction stream. */
3008 if (forced_stmts)
3009 {
3010 gsi = gsi_start (forced_stmts);
3011 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3012 {
3013 gimple stmt = gsi_stmt (gsi);
3014 tree forcedname = gimple_get_lhs (stmt);
3015 pre_expr nameexpr;
3016
3017 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3018 if (TREE_CODE (forcedname) == SSA_NAME)
3019 {
3020 VN_INFO_GET (forcedname)->valnum = forcedname;
3021 VN_INFO (forcedname)->value_id = get_next_value_id ();
3022 nameexpr = get_or_alloc_expr_for_name (forcedname);
3023 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3024 if (!in_fre)
3025 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3026 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3027 }
3028 mark_symbols_for_renaming (stmt);
3029 }
3030 gimple_seq_add_seq (stmts, forced_stmts);
3031 }
3032
3033 /* Build and insert the assignment of the end result to the temporary
3034 that we will return. */
3035 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3036 {
3037 pretemp = create_tmp_var (exprtype, "pretmp");
3038 get_var_ann (pretemp);
3039 }
3040
3041 temp = pretemp;
3042 add_referenced_var (temp);
3043
3044 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3045 || TREE_CODE (exprtype) == VECTOR_TYPE)
3046 DECL_GIMPLE_REG_P (temp) = 1;
3047
3048 newstmt = gimple_build_assign (temp, folded);
3049 name = make_ssa_name (temp, newstmt);
3050 gimple_assign_set_lhs (newstmt, name);
3051 gimple_set_plf (newstmt, NECESSARY, false);
3052
3053 gimple_seq_add_stmt (stmts, newstmt);
3054 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3055
3056 /* All the symbols in NEWEXPR should be put into SSA form. */
3057 mark_symbols_for_renaming (newstmt);
3058
3059 /* Add a value number to the temporary.
3060 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3061 we are creating the expression by pieces, and this particular piece of
3062 the expression may have been represented. There is no harm in replacing
3063 here. */
3064 VN_INFO_GET (name)->valnum = name;
3065 value_id = get_expr_value_id (expr);
3066 VN_INFO (name)->value_id = value_id;
3067 nameexpr = get_or_alloc_expr_for_name (name);
3068 add_to_value (value_id, nameexpr);
3069 if (!in_fre)
3070 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3071 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3072
3073 pre_stats.insertions++;
3074 if (dump_file && (dump_flags & TDF_DETAILS))
3075 {
3076 fprintf (dump_file, "Inserted ");
3077 print_gimple_stmt (dump_file, newstmt, 0, 0);
3078 fprintf (dump_file, " in predecessor %d\n", block->index);
3079 }
3080
3081 return name;
3082 }
3083
3084
3085 /* Returns true if we want to inhibit the insertions of PHI nodes
3086 for the given EXPR for basic block BB (a member of a loop).
3087 We want to do this, when we fear that the induction variable we
3088 create might inhibit vectorization. */
3089
3090 static bool
3091 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3092 {
3093 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3094 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3095 vn_reference_op_t op;
3096 unsigned i;
3097
3098 /* If we aren't going to vectorize we don't inhibit anything. */
3099 if (!flag_tree_vectorize)
3100 return false;
3101
3102 /* Otherwise we inhibit the insertion when the address of the
3103 memory reference is a simple induction variable. In other
3104 cases the vectorizer won't do anything anyway (either it's
3105 loop invariant or a complicated expression). */
3106 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3107 {
3108 switch (op->opcode)
3109 {
3110 case ARRAY_REF:
3111 case ARRAY_RANGE_REF:
3112 if (TREE_CODE (op->op0) != SSA_NAME)
3113 break;
3114 /* Fallthru. */
3115 case SSA_NAME:
3116 {
3117 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3118 affine_iv iv;
3119 /* Default defs are loop invariant. */
3120 if (!defbb)
3121 break;
3122 /* Defined outside this loop, also loop invariant. */
3123 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3124 break;
3125 /* If it's a simple induction variable inhibit insertion,
3126 the vectorizer might be interested in this one. */
3127 if (simple_iv (bb->loop_father, bb->loop_father,
3128 op->op0, &iv, true))
3129 return true;
3130 /* No simple IV, vectorizer can't do anything, hence no
3131 reason to inhibit the transformation for this operand. */
3132 break;
3133 }
3134 default:
3135 break;
3136 }
3137 }
3138 return false;
3139 }
3140
3141 /* Insert the to-be-made-available values of expression EXPRNUM for each
3142 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3143 merge the result with a phi node, given the same value number as
3144 NODE. Return true if we have inserted new stuff. */
3145
3146 static bool
3147 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3148 pre_expr *avail)
3149 {
3150 pre_expr expr = expression_for_id (exprnum);
3151 pre_expr newphi;
3152 unsigned int val = get_expr_value_id (expr);
3153 edge pred;
3154 bool insertions = false;
3155 bool nophi = false;
3156 basic_block bprime;
3157 pre_expr eprime;
3158 edge_iterator ei;
3159 tree type = get_expr_type (expr);
3160 tree temp;
3161 gimple phi;
3162
3163 if (dump_file && (dump_flags & TDF_DETAILS))
3164 {
3165 fprintf (dump_file, "Found partial redundancy for expression ");
3166 print_pre_expr (dump_file, expr);
3167 fprintf (dump_file, " (%04d)\n", val);
3168 }
3169
3170 /* Make sure we aren't creating an induction variable. */
3171 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3172 {
3173 bool firstinsideloop = false;
3174 bool secondinsideloop = false;
3175 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3176 EDGE_PRED (block, 0)->src);
3177 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3178 EDGE_PRED (block, 1)->src);
3179 /* Induction variables only have one edge inside the loop. */
3180 if ((firstinsideloop ^ secondinsideloop)
3181 && (expr->kind != REFERENCE
3182 || inhibit_phi_insertion (block, expr)))
3183 {
3184 if (dump_file && (dump_flags & TDF_DETAILS))
3185 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3186 nophi = true;
3187 }
3188 }
3189
3190 /* Make sure we are not inserting trapping expressions. */
3191 FOR_EACH_EDGE (pred, ei, block->preds)
3192 {
3193 bprime = pred->src;
3194 eprime = avail[bprime->index];
3195 if (eprime->kind == NARY
3196 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3197 return false;
3198 }
3199
3200 /* Make the necessary insertions. */
3201 FOR_EACH_EDGE (pred, ei, block->preds)
3202 {
3203 gimple_seq stmts = NULL;
3204 tree builtexpr;
3205 bprime = pred->src;
3206 eprime = avail[bprime->index];
3207
3208 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3209 {
3210 builtexpr = create_expression_by_pieces (bprime,
3211 eprime,
3212 &stmts, NULL,
3213 type);
3214 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3215 gsi_insert_seq_on_edge (pred, stmts);
3216 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3217 insertions = true;
3218 }
3219 else if (eprime->kind == CONSTANT)
3220 {
3221 /* Constants may not have the right type, fold_convert
3222 should give us back a constant with the right type.
3223 */
3224 tree constant = PRE_EXPR_CONSTANT (eprime);
3225 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3226 {
3227 tree builtexpr = fold_convert (type, constant);
3228 if (!is_gimple_min_invariant (builtexpr))
3229 {
3230 tree forcedexpr = force_gimple_operand (builtexpr,
3231 &stmts, true,
3232 NULL);
3233 if (!is_gimple_min_invariant (forcedexpr))
3234 {
3235 if (forcedexpr != builtexpr)
3236 {
3237 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3238 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3239 }
3240 if (stmts)
3241 {
3242 gimple_stmt_iterator gsi;
3243 gsi = gsi_start (stmts);
3244 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3245 {
3246 gimple stmt = gsi_stmt (gsi);
3247 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3248 gimple_set_plf (stmt, NECESSARY, false);
3249 }
3250 gsi_insert_seq_on_edge (pred, stmts);
3251 }
3252 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3253 }
3254 }
3255 }
3256 }
3257 else if (eprime->kind == NAME)
3258 {
3259 /* We may have to do a conversion because our value
3260 numbering can look through types in certain cases, but
3261 our IL requires all operands of a phi node have the same
3262 type. */
3263 tree name = PRE_EXPR_NAME (eprime);
3264 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3265 {
3266 tree builtexpr;
3267 tree forcedexpr;
3268 builtexpr = fold_convert (type, name);
3269 forcedexpr = force_gimple_operand (builtexpr,
3270 &stmts, true,
3271 NULL);
3272
3273 if (forcedexpr != name)
3274 {
3275 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3276 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3277 }
3278
3279 if (stmts)
3280 {
3281 gimple_stmt_iterator gsi;
3282 gsi = gsi_start (stmts);
3283 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3284 {
3285 gimple stmt = gsi_stmt (gsi);
3286 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3287 gimple_set_plf (stmt, NECESSARY, false);
3288 }
3289 gsi_insert_seq_on_edge (pred, stmts);
3290 }
3291 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3292 }
3293 }
3294 }
3295 /* If we didn't want a phi node, and we made insertions, we still have
3296 inserted new stuff, and thus return true. If we didn't want a phi node,
3297 and didn't make insertions, we haven't added anything new, so return
3298 false. */
3299 if (nophi && insertions)
3300 return true;
3301 else if (nophi && !insertions)
3302 return false;
3303
3304 /* Now build a phi for the new variable. */
3305 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3306 {
3307 prephitemp = create_tmp_var (type, "prephitmp");
3308 get_var_ann (prephitemp);
3309 }
3310
3311 temp = prephitemp;
3312 add_referenced_var (temp);
3313
3314 if (TREE_CODE (type) == COMPLEX_TYPE
3315 || TREE_CODE (type) == VECTOR_TYPE)
3316 DECL_GIMPLE_REG_P (temp) = 1;
3317 phi = create_phi_node (temp, block);
3318
3319 gimple_set_plf (phi, NECESSARY, false);
3320 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3321 VN_INFO (gimple_phi_result (phi))->value_id = val;
3322 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3323 bitmap_set_bit (inserted_phi_names,
3324 SSA_NAME_VERSION (gimple_phi_result (phi)));
3325 FOR_EACH_EDGE (pred, ei, block->preds)
3326 {
3327 pre_expr ae = avail[pred->src->index];
3328 gcc_assert (get_expr_type (ae) == type
3329 || useless_type_conversion_p (type, get_expr_type (ae)));
3330 if (ae->kind == CONSTANT)
3331 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3332 else
3333 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3334 }
3335
3336 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3337 add_to_value (val, newphi);
3338
3339 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3340 this insertion, since we test for the existence of this value in PHI_GEN
3341 before proceeding with the partial redundancy checks in insert_aux.
3342
3343 The value may exist in AVAIL_OUT, in particular, it could be represented
3344 by the expression we are trying to eliminate, in which case we want the
3345 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3346 inserted there.
3347
3348 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3349 this block, because if it did, it would have existed in our dominator's
3350 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3351 */
3352
3353 bitmap_insert_into_set (PHI_GEN (block), newphi);
3354 bitmap_value_replace_in_set (AVAIL_OUT (block),
3355 newphi);
3356 bitmap_insert_into_set (NEW_SETS (block),
3357 newphi);
3358
3359 if (dump_file && (dump_flags & TDF_DETAILS))
3360 {
3361 fprintf (dump_file, "Created phi ");
3362 print_gimple_stmt (dump_file, phi, 0, 0);
3363 fprintf (dump_file, " in block %d\n", block->index);
3364 }
3365 pre_stats.phis++;
3366 return true;
3367 }
3368
3369
3370
3371 /* Perform insertion of partially redundant values.
3372 For BLOCK, do the following:
3373 1. Propagate the NEW_SETS of the dominator into the current block.
3374 If the block has multiple predecessors,
3375 2a. Iterate over the ANTIC expressions for the block to see if
3376 any of them are partially redundant.
3377 2b. If so, insert them into the necessary predecessors to make
3378 the expression fully redundant.
3379 2c. Insert a new PHI merging the values of the predecessors.
3380 2d. Insert the new PHI, and the new expressions, into the
3381 NEW_SETS set.
3382 3. Recursively call ourselves on the dominator children of BLOCK.
3383
3384 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3385 do_regular_insertion and do_partial_insertion.
3386
3387 */
3388
3389 static bool
3390 do_regular_insertion (basic_block block, basic_block dom)
3391 {
3392 bool new_stuff = false;
3393 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3394 pre_expr expr;
3395 int i;
3396
3397 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3398 {
3399 if (expr->kind != NAME)
3400 {
3401 pre_expr *avail;
3402 unsigned int val;
3403 bool by_some = false;
3404 bool cant_insert = false;
3405 bool all_same = true;
3406 pre_expr first_s = NULL;
3407 edge pred;
3408 basic_block bprime;
3409 pre_expr eprime = NULL;
3410 edge_iterator ei;
3411 pre_expr edoubleprime = NULL;
3412
3413 val = get_expr_value_id (expr);
3414 if (bitmap_set_contains_value (PHI_GEN (block), val))
3415 continue;
3416 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3417 {
3418 if (dump_file && (dump_flags & TDF_DETAILS))
3419 fprintf (dump_file, "Found fully redundant value\n");
3420 continue;
3421 }
3422
3423 avail = XCNEWVEC (pre_expr, last_basic_block);
3424 FOR_EACH_EDGE (pred, ei, block->preds)
3425 {
3426 unsigned int vprime;
3427
3428 /* We should never run insertion for the exit block
3429 and so not come across fake pred edges. */
3430 gcc_assert (!(pred->flags & EDGE_FAKE));
3431 bprime = pred->src;
3432 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3433 bprime, block);
3434
3435 /* eprime will generally only be NULL if the
3436 value of the expression, translated
3437 through the PHI for this predecessor, is
3438 undefined. If that is the case, we can't
3439 make the expression fully redundant,
3440 because its value is undefined along a
3441 predecessor path. We can thus break out
3442 early because it doesn't matter what the
3443 rest of the results are. */
3444 if (eprime == NULL)
3445 {
3446 cant_insert = true;
3447 break;
3448 }
3449
3450 eprime = fully_constant_expression (eprime);
3451 vprime = get_expr_value_id (eprime);
3452 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3453 vprime, NULL);
3454 if (edoubleprime == NULL)
3455 {
3456 avail[bprime->index] = eprime;
3457 all_same = false;
3458 }
3459 else
3460 {
3461 avail[bprime->index] = edoubleprime;
3462 by_some = true;
3463 if (first_s == NULL)
3464 first_s = edoubleprime;
3465 else if (!pre_expr_eq (first_s, edoubleprime))
3466 all_same = false;
3467 }
3468 }
3469 /* If we can insert it, it's not the same value
3470 already existing along every predecessor, and
3471 it's defined by some predecessor, it is
3472 partially redundant. */
3473 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3474 {
3475 if (insert_into_preds_of_block (block, get_expression_id (expr),
3476 avail))
3477 new_stuff = true;
3478 }
3479 /* If all edges produce the same value and that value is
3480 an invariant, then the PHI has the same value on all
3481 edges. Note this. */
3482 else if (!cant_insert && all_same && eprime
3483 && (edoubleprime->kind == CONSTANT
3484 || edoubleprime->kind == NAME)
3485 && !value_id_constant_p (val))
3486 {
3487 unsigned int j;
3488 bitmap_iterator bi;
3489 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3490 value_expressions, val);
3491
3492 unsigned int new_val = get_expr_value_id (edoubleprime);
3493 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3494 {
3495 pre_expr expr = expression_for_id (j);
3496
3497 if (expr->kind == NAME)
3498 {
3499 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3500 /* Just reset the value id and valnum so it is
3501 the same as the constant we have discovered. */
3502 if (edoubleprime->kind == CONSTANT)
3503 {
3504 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3505 pre_stats.constified++;
3506 }
3507 else
3508 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3509 info->value_id = new_val;
3510 }
3511 }
3512 }
3513 free (avail);
3514 }
3515 }
3516
3517 VEC_free (pre_expr, heap, exprs);
3518 return new_stuff;
3519 }
3520
3521
3522 /* Perform insertion for partially anticipatable expressions. There
3523 is only one case we will perform insertion for these. This case is
3524 if the expression is partially anticipatable, and fully available.
3525 In this case, we know that putting it earlier will enable us to
3526 remove the later computation. */
3527
3528
3529 static bool
3530 do_partial_partial_insertion (basic_block block, basic_block dom)
3531 {
3532 bool new_stuff = false;
3533 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3534 pre_expr expr;
3535 int i;
3536
3537 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3538 {
3539 if (expr->kind != NAME)
3540 {
3541 pre_expr *avail;
3542 unsigned int val;
3543 bool by_all = true;
3544 bool cant_insert = false;
3545 edge pred;
3546 basic_block bprime;
3547 pre_expr eprime = NULL;
3548 edge_iterator ei;
3549
3550 val = get_expr_value_id (expr);
3551 if (bitmap_set_contains_value (PHI_GEN (block), val))
3552 continue;
3553 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3554 continue;
3555
3556 avail = XCNEWVEC (pre_expr, last_basic_block);
3557 FOR_EACH_EDGE (pred, ei, block->preds)
3558 {
3559 unsigned int vprime;
3560 pre_expr edoubleprime;
3561
3562 /* We should never run insertion for the exit block
3563 and so not come across fake pred edges. */
3564 gcc_assert (!(pred->flags & EDGE_FAKE));
3565 bprime = pred->src;
3566 eprime = phi_translate (expr, ANTIC_IN (block),
3567 PA_IN (block),
3568 bprime, block);
3569
3570 /* eprime will generally only be NULL if the
3571 value of the expression, translated
3572 through the PHI for this predecessor, is
3573 undefined. If that is the case, we can't
3574 make the expression fully redundant,
3575 because its value is undefined along a
3576 predecessor path. We can thus break out
3577 early because it doesn't matter what the
3578 rest of the results are. */
3579 if (eprime == NULL)
3580 {
3581 cant_insert = true;
3582 break;
3583 }
3584
3585 eprime = fully_constant_expression (eprime);
3586 vprime = get_expr_value_id (eprime);
3587 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3588 vprime, NULL);
3589 if (edoubleprime == NULL)
3590 {
3591 by_all = false;
3592 break;
3593 }
3594 else
3595 avail[bprime->index] = edoubleprime;
3596
3597 }
3598
3599 /* If we can insert it, it's not the same value
3600 already existing along every predecessor, and
3601 it's defined by some predecessor, it is
3602 partially redundant. */
3603 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3604 {
3605 pre_stats.pa_insert++;
3606 if (insert_into_preds_of_block (block, get_expression_id (expr),
3607 avail))
3608 new_stuff = true;
3609 }
3610 free (avail);
3611 }
3612 }
3613
3614 VEC_free (pre_expr, heap, exprs);
3615 return new_stuff;
3616 }
3617
3618 static bool
3619 insert_aux (basic_block block)
3620 {
3621 basic_block son;
3622 bool new_stuff = false;
3623
3624 if (block)
3625 {
3626 basic_block dom;
3627 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3628 if (dom)
3629 {
3630 unsigned i;
3631 bitmap_iterator bi;
3632 bitmap_set_t newset = NEW_SETS (dom);
3633 if (newset)
3634 {
3635 /* Note that we need to value_replace both NEW_SETS, and
3636 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3637 represented by some non-simple expression here that we want
3638 to replace it with. */
3639 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3640 {
3641 pre_expr expr = expression_for_id (i);
3642 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3643 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3644 }
3645 }
3646 if (!single_pred_p (block))
3647 {
3648 new_stuff |= do_regular_insertion (block, dom);
3649 if (do_partial_partial)
3650 new_stuff |= do_partial_partial_insertion (block, dom);
3651 }
3652 }
3653 }
3654 for (son = first_dom_son (CDI_DOMINATORS, block);
3655 son;
3656 son = next_dom_son (CDI_DOMINATORS, son))
3657 {
3658 new_stuff |= insert_aux (son);
3659 }
3660
3661 return new_stuff;
3662 }
3663
3664 /* Perform insertion of partially redundant values. */
3665
3666 static void
3667 insert (void)
3668 {
3669 bool new_stuff = true;
3670 basic_block bb;
3671 int num_iterations = 0;
3672
3673 FOR_ALL_BB (bb)
3674 NEW_SETS (bb) = bitmap_set_new ();
3675
3676 while (new_stuff)
3677 {
3678 num_iterations++;
3679 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3680 }
3681 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3682 }
3683
3684
3685 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3686
3687 static void
3688 add_to_exp_gen (basic_block block, tree op)
3689 {
3690 if (!in_fre)
3691 {
3692 pre_expr result;
3693 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3694 return;
3695 result = get_or_alloc_expr_for_name (op);
3696 bitmap_value_insert_into_set (EXP_GEN (block), result);
3697 bitmap_value_insert_into_set (maximal_set, result);
3698 }
3699 }
3700
3701 /* Create value ids for PHI in BLOCK. */
3702
3703 static void
3704 make_values_for_phi (gimple phi, basic_block block)
3705 {
3706 tree result = gimple_phi_result (phi);
3707
3708 /* We have no need for virtual phis, as they don't represent
3709 actual computations. */
3710 if (is_gimple_reg (result))
3711 {
3712 pre_expr e = get_or_alloc_expr_for_name (result);
3713 add_to_value (get_expr_value_id (e), e);
3714 bitmap_insert_into_set (PHI_GEN (block), e);
3715 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3716 if (!in_fre)
3717 {
3718 unsigned i;
3719 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3720 {
3721 tree arg = gimple_phi_arg_def (phi, i);
3722 if (TREE_CODE (arg) == SSA_NAME)
3723 {
3724 e = get_or_alloc_expr_for_name (arg);
3725 add_to_value (get_expr_value_id (e), e);
3726 bitmap_value_insert_into_set (maximal_set, e);
3727 }
3728 }
3729 }
3730 }
3731 }
3732
3733 /* Compute the AVAIL set for all basic blocks.
3734
3735 This function performs value numbering of the statements in each basic
3736 block. The AVAIL sets are built from information we glean while doing
3737 this value numbering, since the AVAIL sets contain only one entry per
3738 value.
3739
3740 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3741 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3742
3743 static void
3744 compute_avail (void)
3745 {
3746
3747 basic_block block, son;
3748 basic_block *worklist;
3749 size_t sp = 0;
3750 unsigned i;
3751
3752 /* We pretend that default definitions are defined in the entry block.
3753 This includes function arguments and the static chain decl. */
3754 for (i = 1; i < num_ssa_names; ++i)
3755 {
3756 tree name = ssa_name (i);
3757 pre_expr e;
3758 if (!name
3759 || !SSA_NAME_IS_DEFAULT_DEF (name)
3760 || has_zero_uses (name)
3761 || !is_gimple_reg (name))
3762 continue;
3763
3764 e = get_or_alloc_expr_for_name (name);
3765 add_to_value (get_expr_value_id (e), e);
3766 if (!in_fre)
3767 {
3768 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3769 bitmap_value_insert_into_set (maximal_set, e);
3770 }
3771 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3772 }
3773
3774 /* Allocate the worklist. */
3775 worklist = XNEWVEC (basic_block, n_basic_blocks);
3776
3777 /* Seed the algorithm by putting the dominator children of the entry
3778 block on the worklist. */
3779 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3780 son;
3781 son = next_dom_son (CDI_DOMINATORS, son))
3782 worklist[sp++] = son;
3783
3784 /* Loop until the worklist is empty. */
3785 while (sp)
3786 {
3787 gimple_stmt_iterator gsi;
3788 gimple stmt;
3789 basic_block dom;
3790 unsigned int stmt_uid = 1;
3791
3792 /* Pick a block from the worklist. */
3793 block = worklist[--sp];
3794
3795 /* Initially, the set of available values in BLOCK is that of
3796 its immediate dominator. */
3797 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3798 if (dom)
3799 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3800
3801 /* Generate values for PHI nodes. */
3802 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3803 make_values_for_phi (gsi_stmt (gsi), block);
3804
3805 /* Now compute value numbers and populate value sets with all
3806 the expressions computed in BLOCK. */
3807 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3808 {
3809 ssa_op_iter iter;
3810 tree op;
3811
3812 stmt = gsi_stmt (gsi);
3813 gimple_set_uid (stmt, stmt_uid++);
3814
3815 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3816 {
3817 pre_expr e = get_or_alloc_expr_for_name (op);
3818
3819 add_to_value (get_expr_value_id (e), e);
3820 if (!in_fre)
3821 bitmap_insert_into_set (TMP_GEN (block), e);
3822 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3823 }
3824
3825 if (gimple_has_volatile_ops (stmt)
3826 || stmt_could_throw_p (stmt))
3827 continue;
3828
3829 switch (gimple_code (stmt))
3830 {
3831 case GIMPLE_RETURN:
3832 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3833 add_to_exp_gen (block, op);
3834 continue;
3835
3836 case GIMPLE_CALL:
3837 {
3838 vn_reference_t ref;
3839 unsigned int i;
3840 vn_reference_op_t vro;
3841 pre_expr result = NULL;
3842 VEC(vn_reference_op_s, heap) *ops = NULL;
3843
3844 if (!can_value_number_call (stmt))
3845 continue;
3846
3847 copy_reference_ops_from_call (stmt, &ops);
3848 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3849 gimple_expr_type (stmt),
3850 ops, &ref, false);
3851 VEC_free (vn_reference_op_s, heap, ops);
3852 if (!ref)
3853 continue;
3854
3855 for (i = 0; VEC_iterate (vn_reference_op_s,
3856 ref->operands, i,
3857 vro); i++)
3858 {
3859 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3860 add_to_exp_gen (block, vro->op0);
3861 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3862 add_to_exp_gen (block, vro->op1);
3863 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3864 add_to_exp_gen (block, vro->op2);
3865 }
3866 result = (pre_expr) pool_alloc (pre_expr_pool);
3867 result->kind = REFERENCE;
3868 result->id = 0;
3869 PRE_EXPR_REFERENCE (result) = ref;
3870
3871 get_or_alloc_expression_id (result);
3872 add_to_value (get_expr_value_id (result), result);
3873 if (!in_fre)
3874 {
3875 bitmap_value_insert_into_set (EXP_GEN (block),
3876 result);
3877 bitmap_value_insert_into_set (maximal_set, result);
3878 }
3879 continue;
3880 }
3881
3882 case GIMPLE_ASSIGN:
3883 {
3884 pre_expr result = NULL;
3885 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3886 {
3887 case tcc_unary:
3888 if (is_exception_related (stmt))
3889 continue;
3890 case tcc_binary:
3891 case tcc_comparison:
3892 {
3893 vn_nary_op_t nary;
3894 unsigned int i;
3895
3896 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3897 gimple_assign_rhs_code (stmt),
3898 gimple_expr_type (stmt),
3899 gimple_assign_rhs1 (stmt),
3900 gimple_assign_rhs2 (stmt),
3901 NULL_TREE, NULL_TREE, &nary);
3902
3903 if (!nary)
3904 continue;
3905
3906 for (i = 0; i < nary->length; i++)
3907 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3908 add_to_exp_gen (block, nary->op[i]);
3909
3910 result = (pre_expr) pool_alloc (pre_expr_pool);
3911 result->kind = NARY;
3912 result->id = 0;
3913 PRE_EXPR_NARY (result) = nary;
3914 break;
3915 }
3916
3917 case tcc_declaration:
3918 case tcc_reference:
3919 {
3920 vn_reference_t ref;
3921 unsigned int i;
3922 vn_reference_op_t vro;
3923
3924 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3925 gimple_vuse (stmt),
3926 false, &ref);
3927 if (!ref)
3928 continue;
3929
3930 for (i = 0; VEC_iterate (vn_reference_op_s,
3931 ref->operands, i,
3932 vro); i++)
3933 {
3934 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3935 add_to_exp_gen (block, vro->op0);
3936 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3937 add_to_exp_gen (block, vro->op1);
3938 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3939 add_to_exp_gen (block, vro->op2);
3940 }
3941 result = (pre_expr) pool_alloc (pre_expr_pool);
3942 result->kind = REFERENCE;
3943 result->id = 0;
3944 PRE_EXPR_REFERENCE (result) = ref;
3945 break;
3946 }
3947
3948 default:
3949 /* For any other statement that we don't
3950 recognize, simply add all referenced
3951 SSA_NAMEs to EXP_GEN. */
3952 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3953 add_to_exp_gen (block, op);
3954 continue;
3955 }
3956
3957 get_or_alloc_expression_id (result);
3958 add_to_value (get_expr_value_id (result), result);
3959 if (!in_fre)
3960 {
3961 bitmap_value_insert_into_set (EXP_GEN (block), result);
3962 bitmap_value_insert_into_set (maximal_set, result);
3963 }
3964
3965 continue;
3966 }
3967 default:
3968 break;
3969 }
3970 }
3971
3972 /* Put the dominator children of BLOCK on the worklist of blocks
3973 to compute available sets for. */
3974 for (son = first_dom_son (CDI_DOMINATORS, block);
3975 son;
3976 son = next_dom_son (CDI_DOMINATORS, son))
3977 worklist[sp++] = son;
3978 }
3979
3980 free (worklist);
3981 }
3982
3983 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3984 than the available expressions for it. The insertion point is
3985 right before the first use in STMT. Returns the SSA_NAME that should
3986 be used for replacement. */
3987
3988 static tree
3989 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3990 {
3991 basic_block bb = gimple_bb (stmt);
3992 gimple_stmt_iterator gsi;
3993 gimple_seq stmts = NULL;
3994 tree expr;
3995 pre_expr e;
3996
3997 /* First create a value expression from the expression we want
3998 to insert and associate it with the value handle for SSA_VN. */
3999 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4000 if (e == NULL)
4001 return NULL_TREE;
4002
4003 /* Then use create_expression_by_pieces to generate a valid
4004 expression to insert at this point of the IL stream. */
4005 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4006 if (expr == NULL_TREE)
4007 return NULL_TREE;
4008 gsi = gsi_for_stmt (stmt);
4009 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4010
4011 return expr;
4012 }
4013
4014 /* Eliminate fully redundant computations. */
4015
4016 static unsigned int
4017 eliminate (void)
4018 {
4019 VEC (gimple, heap) *to_remove = NULL;
4020 basic_block b;
4021 unsigned int todo = 0;
4022 gimple_stmt_iterator gsi;
4023 gimple stmt;
4024 unsigned i;
4025
4026 FOR_EACH_BB (b)
4027 {
4028 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4029 {
4030 stmt = gsi_stmt (gsi);
4031
4032 /* Lookup the RHS of the expression, see if we have an
4033 available computation for it. If so, replace the RHS with
4034 the available computation. */
4035 if (gimple_has_lhs (stmt)
4036 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4037 && !gimple_assign_ssa_name_copy_p (stmt)
4038 && (!gimple_assign_single_p (stmt)
4039 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4040 && !gimple_has_volatile_ops (stmt)
4041 && !has_zero_uses (gimple_get_lhs (stmt)))
4042 {
4043 tree lhs = gimple_get_lhs (stmt);
4044 tree rhs = NULL_TREE;
4045 tree sprime = NULL;
4046 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4047 pre_expr sprimeexpr;
4048
4049 if (gimple_assign_single_p (stmt))
4050 rhs = gimple_assign_rhs1 (stmt);
4051
4052 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4053 get_expr_value_id (lhsexpr),
4054 NULL);
4055
4056 if (sprimeexpr)
4057 {
4058 if (sprimeexpr->kind == CONSTANT)
4059 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4060 else if (sprimeexpr->kind == NAME)
4061 sprime = PRE_EXPR_NAME (sprimeexpr);
4062 else
4063 gcc_unreachable ();
4064 }
4065
4066 /* If there is no existing leader but SCCVN knows this
4067 value is constant, use that constant. */
4068 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4069 {
4070 sprime = VN_INFO (lhs)->valnum;
4071 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4072 TREE_TYPE (sprime)))
4073 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4074
4075 if (dump_file && (dump_flags & TDF_DETAILS))
4076 {
4077 fprintf (dump_file, "Replaced ");
4078 print_gimple_expr (dump_file, stmt, 0, 0);
4079 fprintf (dump_file, " with ");
4080 print_generic_expr (dump_file, sprime, 0);
4081 fprintf (dump_file, " in ");
4082 print_gimple_stmt (dump_file, stmt, 0, 0);
4083 }
4084 pre_stats.eliminations++;
4085 propagate_tree_value_into_stmt (&gsi, sprime);
4086 stmt = gsi_stmt (gsi);
4087 update_stmt (stmt);
4088 continue;
4089 }
4090
4091 /* If there is no existing usable leader but SCCVN thinks
4092 it has an expression it wants to use as replacement,
4093 insert that. */
4094 if (!sprime || sprime == lhs)
4095 {
4096 tree val = VN_INFO (lhs)->valnum;
4097 if (val != VN_TOP
4098 && TREE_CODE (val) == SSA_NAME
4099 && VN_INFO (val)->needs_insertion
4100 && can_PRE_operation (vn_get_expr_for (val)))
4101 sprime = do_SCCVN_insertion (stmt, val);
4102 }
4103 if (sprime
4104 && sprime != lhs
4105 && (rhs == NULL_TREE
4106 || TREE_CODE (rhs) != SSA_NAME
4107 || may_propagate_copy (rhs, sprime)))
4108 {
4109 gcc_assert (sprime != rhs);
4110
4111 if (dump_file && (dump_flags & TDF_DETAILS))
4112 {
4113 fprintf (dump_file, "Replaced ");
4114 print_gimple_expr (dump_file, stmt, 0, 0);
4115 fprintf (dump_file, " with ");
4116 print_generic_expr (dump_file, sprime, 0);
4117 fprintf (dump_file, " in ");
4118 print_gimple_stmt (dump_file, stmt, 0, 0);
4119 }
4120
4121 if (TREE_CODE (sprime) == SSA_NAME)
4122 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4123 NECESSARY, true);
4124 /* We need to make sure the new and old types actually match,
4125 which may require adding a simple cast, which fold_convert
4126 will do for us. */
4127 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4128 && !useless_type_conversion_p (gimple_expr_type (stmt),
4129 TREE_TYPE (sprime)))
4130 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4131
4132 pre_stats.eliminations++;
4133 propagate_tree_value_into_stmt (&gsi, sprime);
4134 stmt = gsi_stmt (gsi);
4135 update_stmt (stmt);
4136
4137 /* If we removed EH side effects from the statement, clean
4138 its EH information. */
4139 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4140 {
4141 bitmap_set_bit (need_eh_cleanup,
4142 gimple_bb (stmt)->index);
4143 if (dump_file && (dump_flags & TDF_DETAILS))
4144 fprintf (dump_file, " Removed EH side effects.\n");
4145 }
4146 }
4147 }
4148 /* If the statement is a scalar store, see if the expression
4149 has the same value number as its rhs. If so, the store is
4150 dead. */
4151 else if (gimple_assign_single_p (stmt)
4152 && !is_gimple_reg (gimple_assign_lhs (stmt))
4153 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4154 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4155 {
4156 tree rhs = gimple_assign_rhs1 (stmt);
4157 tree val;
4158 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4159 gimple_vuse (stmt), true, NULL);
4160 if (TREE_CODE (rhs) == SSA_NAME)
4161 rhs = VN_INFO (rhs)->valnum;
4162 if (val
4163 && operand_equal_p (val, rhs, 0))
4164 {
4165 if (dump_file && (dump_flags & TDF_DETAILS))
4166 {
4167 fprintf (dump_file, "Deleted redundant store ");
4168 print_gimple_stmt (dump_file, stmt, 0, 0);
4169 }
4170
4171 /* Queue stmt for removal. */
4172 VEC_safe_push (gimple, heap, to_remove, stmt);
4173 }
4174 }
4175 /* Visit COND_EXPRs and fold the comparison with the
4176 available value-numbers. */
4177 else if (gimple_code (stmt) == GIMPLE_COND)
4178 {
4179 tree op0 = gimple_cond_lhs (stmt);
4180 tree op1 = gimple_cond_rhs (stmt);
4181 tree result;
4182
4183 if (TREE_CODE (op0) == SSA_NAME)
4184 op0 = VN_INFO (op0)->valnum;
4185 if (TREE_CODE (op1) == SSA_NAME)
4186 op1 = VN_INFO (op1)->valnum;
4187 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4188 op0, op1);
4189 if (result && TREE_CODE (result) == INTEGER_CST)
4190 {
4191 if (integer_zerop (result))
4192 gimple_cond_make_false (stmt);
4193 else
4194 gimple_cond_make_true (stmt);
4195 update_stmt (stmt);
4196 todo = TODO_cleanup_cfg;
4197 }
4198 }
4199 /* Visit indirect calls and turn them into direct calls if
4200 possible. */
4201 if (gimple_code (stmt) == GIMPLE_CALL
4202 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4203 {
4204 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4205 if (TREE_CODE (fn) == ADDR_EXPR
4206 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4207 {
4208 if (dump_file && (dump_flags & TDF_DETAILS))
4209 {
4210 fprintf (dump_file, "Replacing call target with ");
4211 print_generic_expr (dump_file, fn, 0);
4212 fprintf (dump_file, " in ");
4213 print_gimple_stmt (dump_file, stmt, 0, 0);
4214 }
4215
4216 gimple_call_set_fn (stmt, fn);
4217 update_stmt (stmt);
4218 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4219 gimple_purge_dead_eh_edges (b);
4220
4221 /* Changing an indirect call to a direct call may
4222 have exposed different semantics. This may
4223 require an SSA update. */
4224 todo |= TODO_update_ssa_only_virtuals;
4225 }
4226 }
4227 }
4228
4229 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4230 {
4231 gimple stmt, phi = gsi_stmt (gsi);
4232 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4233 pre_expr sprimeexpr, resexpr;
4234 gimple_stmt_iterator gsi2;
4235
4236 /* We want to perform redundant PHI elimination. Do so by
4237 replacing the PHI with a single copy if possible.
4238 Do not touch inserted, single-argument or virtual PHIs. */
4239 if (gimple_phi_num_args (phi) == 1
4240 || !is_gimple_reg (res)
4241 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4242 {
4243 gsi_next (&gsi);
4244 continue;
4245 }
4246
4247 resexpr = get_or_alloc_expr_for_name (res);
4248 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4249 get_expr_value_id (resexpr), NULL);
4250 if (sprimeexpr)
4251 {
4252 if (sprimeexpr->kind == CONSTANT)
4253 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4254 else if (sprimeexpr->kind == NAME)
4255 sprime = PRE_EXPR_NAME (sprimeexpr);
4256 else
4257 gcc_unreachable ();
4258 }
4259 if (!sprimeexpr
4260 || sprime == res)
4261 {
4262 gsi_next (&gsi);
4263 continue;
4264 }
4265
4266 if (dump_file && (dump_flags & TDF_DETAILS))
4267 {
4268 fprintf (dump_file, "Replaced redundant PHI node defining ");
4269 print_generic_expr (dump_file, res, 0);
4270 fprintf (dump_file, " with ");
4271 print_generic_expr (dump_file, sprime, 0);
4272 fprintf (dump_file, "\n");
4273 }
4274
4275 remove_phi_node (&gsi, false);
4276
4277 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4278 sprime = fold_convert (TREE_TYPE (res), sprime);
4279 stmt = gimple_build_assign (res, sprime);
4280 SSA_NAME_DEF_STMT (res) = stmt;
4281 if (TREE_CODE (sprime) == SSA_NAME)
4282 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4283 NECESSARY, true);
4284 gsi2 = gsi_after_labels (b);
4285 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4286 /* Queue the copy for eventual removal. */
4287 VEC_safe_push (gimple, heap, to_remove, stmt);
4288 pre_stats.eliminations++;
4289 }
4290 }
4291
4292 /* We cannot remove stmts during BB walk, especially not release SSA
4293 names there as this confuses the VN machinery. The stmts ending
4294 up in to_remove are either stores or simple copies. */
4295 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4296 {
4297 tree lhs = gimple_assign_lhs (stmt);
4298 use_operand_p use_p;
4299 gimple use_stmt;
4300
4301 /* If there is a single use only, propagate the equivalency
4302 instead of keeping the copy. */
4303 if (TREE_CODE (lhs) == SSA_NAME
4304 && single_imm_use (lhs, &use_p, &use_stmt)
4305 && may_propagate_copy (USE_FROM_PTR (use_p),
4306 gimple_assign_rhs1 (stmt)))
4307 {
4308 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4309 update_stmt (use_stmt);
4310 }
4311
4312 /* If this is a store or a now unused copy, remove it. */
4313 if (TREE_CODE (lhs) != SSA_NAME
4314 || has_zero_uses (lhs))
4315 {
4316 gsi = gsi_for_stmt (stmt);
4317 unlink_stmt_vdef (stmt);
4318 gsi_remove (&gsi, true);
4319 release_defs (stmt);
4320 }
4321 }
4322 VEC_free (gimple, heap, to_remove);
4323
4324 return todo;
4325 }
4326
4327 /* Borrow a bit of tree-ssa-dce.c for the moment.
4328 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4329 this may be a bit faster, and we may want critical edges kept split. */
4330
4331 /* If OP's defining statement has not already been determined to be necessary,
4332 mark that statement necessary. Return the stmt, if it is newly
4333 necessary. */
4334
4335 static inline gimple
4336 mark_operand_necessary (tree op)
4337 {
4338 gimple stmt;
4339
4340 gcc_assert (op);
4341
4342 if (TREE_CODE (op) != SSA_NAME)
4343 return NULL;
4344
4345 stmt = SSA_NAME_DEF_STMT (op);
4346 gcc_assert (stmt);
4347
4348 if (gimple_plf (stmt, NECESSARY)
4349 || gimple_nop_p (stmt))
4350 return NULL;
4351
4352 gimple_set_plf (stmt, NECESSARY, true);
4353 return stmt;
4354 }
4355
4356 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4357 to insert PHI nodes sometimes, and because value numbering of casts isn't
4358 perfect, we sometimes end up inserting dead code. This simple DCE-like
4359 pass removes any insertions we made that weren't actually used. */
4360
4361 static void
4362 remove_dead_inserted_code (void)
4363 {
4364 VEC(gimple,heap) *worklist = NULL;
4365 int i;
4366 gimple t;
4367
4368 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4369 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4370 {
4371 if (gimple_plf (t, NECESSARY))
4372 VEC_quick_push (gimple, worklist, t);
4373 }
4374 while (VEC_length (gimple, worklist) > 0)
4375 {
4376 t = VEC_pop (gimple, worklist);
4377
4378 /* PHI nodes are somewhat special in that each PHI alternative has
4379 data and control dependencies. All the statements feeding the
4380 PHI node's arguments are always necessary. */
4381 if (gimple_code (t) == GIMPLE_PHI)
4382 {
4383 unsigned k;
4384
4385 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4386 for (k = 0; k < gimple_phi_num_args (t); k++)
4387 {
4388 tree arg = PHI_ARG_DEF (t, k);
4389 if (TREE_CODE (arg) == SSA_NAME)
4390 {
4391 gimple n = mark_operand_necessary (arg);
4392 if (n)
4393 VEC_quick_push (gimple, worklist, n);
4394 }
4395 }
4396 }
4397 else
4398 {
4399 /* Propagate through the operands. Examine all the USE, VUSE and
4400 VDEF operands in this statement. Mark all the statements
4401 which feed this statement's uses as necessary. */
4402 ssa_op_iter iter;
4403 tree use;
4404
4405 /* The operands of VDEF expressions are also needed as they
4406 represent potential definitions that may reach this
4407 statement (VDEF operands allow us to follow def-def
4408 links). */
4409
4410 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4411 {
4412 gimple n = mark_operand_necessary (use);
4413 if (n)
4414 VEC_safe_push (gimple, heap, worklist, n);
4415 }
4416 }
4417 }
4418
4419 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4420 {
4421 if (!gimple_plf (t, NECESSARY))
4422 {
4423 gimple_stmt_iterator gsi;
4424
4425 if (dump_file && (dump_flags & TDF_DETAILS))
4426 {
4427 fprintf (dump_file, "Removing unnecessary insertion:");
4428 print_gimple_stmt (dump_file, t, 0, 0);
4429 }
4430
4431 gsi = gsi_for_stmt (t);
4432 if (gimple_code (t) == GIMPLE_PHI)
4433 remove_phi_node (&gsi, true);
4434 else
4435 gsi_remove (&gsi, true);
4436 release_defs (t);
4437 }
4438 }
4439 VEC_free (gimple, heap, worklist);
4440 }
4441
4442 /* Initialize data structures used by PRE. */
4443
4444 static void
4445 init_pre (bool do_fre)
4446 {
4447 basic_block bb;
4448
4449 next_expression_id = 1;
4450 expressions = NULL;
4451 VEC_safe_push (pre_expr, heap, expressions, NULL);
4452 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4453 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4454 get_max_value_id() + 1);
4455
4456 in_fre = do_fre;
4457
4458 inserted_exprs = NULL;
4459 need_creation = NULL;
4460 pretemp = NULL_TREE;
4461 storetemp = NULL_TREE;
4462 prephitemp = NULL_TREE;
4463
4464 connect_infinite_loops_to_exit ();
4465 memset (&pre_stats, 0, sizeof (pre_stats));
4466
4467
4468 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4469 post_order_compute (postorder, false, false);
4470
4471 FOR_ALL_BB (bb)
4472 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4473
4474 calculate_dominance_info (CDI_POST_DOMINATORS);
4475 calculate_dominance_info (CDI_DOMINATORS);
4476
4477 bitmap_obstack_initialize (&grand_bitmap_obstack);
4478 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4479 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4480 expr_pred_trans_eq, free);
4481 expression_to_id = htab_create (num_ssa_names * 3,
4482 pre_expr_hash,
4483 pre_expr_eq, NULL);
4484 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4485 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4486 sizeof (struct bitmap_set), 30);
4487 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4488 sizeof (struct pre_expr_d), 30);
4489 FOR_ALL_BB (bb)
4490 {
4491 EXP_GEN (bb) = bitmap_set_new ();
4492 PHI_GEN (bb) = bitmap_set_new ();
4493 TMP_GEN (bb) = bitmap_set_new ();
4494 AVAIL_OUT (bb) = bitmap_set_new ();
4495 }
4496 maximal_set = in_fre ? NULL : bitmap_set_new ();
4497
4498 need_eh_cleanup = BITMAP_ALLOC (NULL);
4499 }
4500
4501
4502 /* Deallocate data structures used by PRE. */
4503
4504 static void
4505 fini_pre (bool do_fre)
4506 {
4507 basic_block bb;
4508
4509 free (postorder);
4510 VEC_free (bitmap_set_t, heap, value_expressions);
4511 VEC_free (gimple, heap, inserted_exprs);
4512 VEC_free (gimple, heap, need_creation);
4513 bitmap_obstack_release (&grand_bitmap_obstack);
4514 free_alloc_pool (bitmap_set_pool);
4515 free_alloc_pool (pre_expr_pool);
4516 htab_delete (phi_translate_table);
4517 htab_delete (expression_to_id);
4518
4519 FOR_ALL_BB (bb)
4520 {
4521 free (bb->aux);
4522 bb->aux = NULL;
4523 }
4524
4525 free_dominance_info (CDI_POST_DOMINATORS);
4526
4527 if (!bitmap_empty_p (need_eh_cleanup))
4528 {
4529 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4530 cleanup_tree_cfg ();
4531 }
4532
4533 BITMAP_FREE (need_eh_cleanup);
4534
4535 if (!do_fre)
4536 loop_optimizer_finalize ();
4537 }
4538
4539 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4540 only wants to do full redundancy elimination. */
4541
4542 static unsigned int
4543 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4544 {
4545 unsigned int todo = 0;
4546
4547 do_partial_partial = optimize > 2;
4548
4549 /* This has to happen before SCCVN runs because
4550 loop_optimizer_init may create new phis, etc. */
4551 if (!do_fre)
4552 loop_optimizer_init (LOOPS_NORMAL);
4553
4554 if (!run_scc_vn (do_fre))
4555 {
4556 if (!do_fre)
4557 {
4558 remove_dead_inserted_code ();
4559 loop_optimizer_finalize ();
4560 }
4561
4562 return 0;
4563 }
4564 init_pre (do_fre);
4565 scev_initialize ();
4566
4567
4568 /* Collect and value number expressions computed in each basic block. */
4569 compute_avail ();
4570
4571 if (dump_file && (dump_flags & TDF_DETAILS))
4572 {
4573 basic_block bb;
4574
4575 FOR_ALL_BB (bb)
4576 {
4577 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4578 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4579 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4580 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4581 }
4582
4583 print_bitmap_set (dump_file, maximal_set, "maximal", 0);
4584 }
4585
4586 /* Insert can get quite slow on an incredibly large number of basic
4587 blocks due to some quadratic behavior. Until this behavior is
4588 fixed, don't run it when he have an incredibly large number of
4589 bb's. If we aren't going to run insert, there is no point in
4590 computing ANTIC, either, even though it's plenty fast. */
4591 if (!do_fre && n_basic_blocks < 4000)
4592 {
4593 compute_antic ();
4594 insert ();
4595 }
4596
4597 /* Remove all the redundant expressions. */
4598 todo |= eliminate ();
4599
4600 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4601 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4602 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4603 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4604 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4605
4606 /* Make sure to remove fake edges before committing our inserts.
4607 This makes sure we don't end up with extra critical edges that
4608 we would need to split. */
4609 remove_fake_exit_edges ();
4610 gsi_commit_edge_inserts ();
4611
4612 clear_expression_ids ();
4613 free_scc_vn ();
4614 if (!do_fre)
4615 remove_dead_inserted_code ();
4616
4617 scev_finalize ();
4618 fini_pre (do_fre);
4619
4620 return todo;
4621 }
4622
4623 /* Gate and execute functions for PRE. */
4624
4625 static unsigned int
4626 do_pre (void)
4627 {
4628 return execute_pre (false);
4629 }
4630
4631 static bool
4632 gate_pre (void)
4633 {
4634 /* PRE tends to generate bigger code. */
4635 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4636 }
4637
4638 struct gimple_opt_pass pass_pre =
4639 {
4640 {
4641 GIMPLE_PASS,
4642 "pre", /* name */
4643 gate_pre, /* gate */
4644 do_pre, /* execute */
4645 NULL, /* sub */
4646 NULL, /* next */
4647 0, /* static_pass_number */
4648 TV_TREE_PRE, /* tv_id */
4649 PROP_no_crit_edges | PROP_cfg
4650 | PROP_ssa, /* properties_required */
4651 0, /* properties_provided */
4652 0, /* properties_destroyed */
4653 TODO_rebuild_alias, /* todo_flags_start */
4654 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4655 | TODO_verify_ssa /* todo_flags_finish */
4656 }
4657 };
4658
4659
4660 /* Gate and execute functions for FRE. */
4661
4662 static unsigned int
4663 execute_fre (void)
4664 {
4665 return execute_pre (true);
4666 }
4667
4668 static bool
4669 gate_fre (void)
4670 {
4671 return flag_tree_fre != 0;
4672 }
4673
4674 struct gimple_opt_pass pass_fre =
4675 {
4676 {
4677 GIMPLE_PASS,
4678 "fre", /* name */
4679 gate_fre, /* gate */
4680 execute_fre, /* execute */
4681 NULL, /* sub */
4682 NULL, /* next */
4683 0, /* static_pass_number */
4684 TV_TREE_FRE, /* tv_id */
4685 PROP_cfg | PROP_ssa, /* properties_required */
4686 0, /* properties_provided */
4687 0, /* properties_destroyed */
4688 0, /* todo_flags_start */
4689 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
4690 }
4691 };