Squash commit of EH in gimple
[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 /* Basic block list in postorder. */
405 static int *postorder;
406
407 /* This structure is used to keep track of statistics on what
408 optimization PRE was able to perform. */
409 static struct
410 {
411 /* The number of RHS computations eliminated by PRE. */
412 int eliminations;
413
414 /* The number of new expressions/temporaries generated by PRE. */
415 int insertions;
416
417 /* The number of inserts found due to partial anticipation */
418 int pa_insert;
419
420 /* The number of new PHI nodes added by PRE. */
421 int phis;
422
423 /* The number of values found constant. */
424 int constified;
425
426 } pre_stats;
427
428 static bool do_partial_partial;
429 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
430 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
431 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
432 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
433 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
434 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
435 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
436 static bitmap_set_t bitmap_set_new (void);
437 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
438 gimple, tree);
439 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
440 gimple);
441 static unsigned int get_expr_value_id (pre_expr);
442
443 /* We can add and remove elements and entries to and from sets
444 and hash tables, so we use alloc pools for them. */
445
446 static alloc_pool bitmap_set_pool;
447 static bitmap_obstack grand_bitmap_obstack;
448
449 /* To avoid adding 300 temporary variables when we only need one, we
450 only create one temporary variable, on demand, and build ssa names
451 off that. We do have to change the variable if the types don't
452 match the current variable's type. */
453 static tree pretemp;
454 static tree storetemp;
455 static tree prephitemp;
456
457 /* Set of blocks with statements that have had its EH information
458 cleaned up. */
459 static bitmap need_eh_cleanup;
460
461 /* Which expressions have been seen during a given phi translation. */
462 static bitmap seen_during_translate;
463
464 /* The phi_translate_table caches phi translations for a given
465 expression and predecessor. */
466
467 static htab_t phi_translate_table;
468
469 /* A three tuple {e, pred, v} used to cache phi translations in the
470 phi_translate_table. */
471
472 typedef struct expr_pred_trans_d
473 {
474 /* The expression. */
475 pre_expr e;
476
477 /* The predecessor block along which we translated the expression. */
478 basic_block pred;
479
480 /* The value that resulted from the translation. */
481 pre_expr v;
482
483 /* The hashcode for the expression, pred pair. This is cached for
484 speed reasons. */
485 hashval_t hashcode;
486 } *expr_pred_trans_t;
487 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
488
489 /* Return the hash value for a phi translation table entry. */
490
491 static hashval_t
492 expr_pred_trans_hash (const void *p)
493 {
494 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
495 return ve->hashcode;
496 }
497
498 /* Return true if two phi translation table entries are the same.
499 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
500
501 static int
502 expr_pred_trans_eq (const void *p1, const void *p2)
503 {
504 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
505 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
506 basic_block b1 = ve1->pred;
507 basic_block b2 = ve2->pred;
508
509 /* If they are not translations for the same basic block, they can't
510 be equal. */
511 if (b1 != b2)
512 return false;
513 return pre_expr_eq (ve1->e, ve2->e);
514 }
515
516 /* Search in the phi translation table for the translation of
517 expression E in basic block PRED.
518 Return the translated value, if found, NULL otherwise. */
519
520 static inline pre_expr
521 phi_trans_lookup (pre_expr e, basic_block pred)
522 {
523 void **slot;
524 struct expr_pred_trans_d ept;
525
526 ept.e = e;
527 ept.pred = pred;
528 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
529 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
530 NO_INSERT);
531 if (!slot)
532 return NULL;
533 else
534 return ((expr_pred_trans_t) *slot)->v;
535 }
536
537
538 /* Add the tuple mapping from {expression E, basic block PRED} to
539 value V, to the phi translation table. */
540
541 static inline void
542 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
543 {
544 void **slot;
545 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
546 new_pair->e = e;
547 new_pair->pred = pred;
548 new_pair->v = v;
549 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
550 pred->index);
551
552 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
553 new_pair->hashcode, INSERT);
554 if (*slot)
555 free (*slot);
556 *slot = (void *) new_pair;
557 }
558
559
560 /* Add expression E to the expression set of value id V. */
561
562 void
563 add_to_value (unsigned int v, pre_expr e)
564 {
565 bitmap_set_t set;
566
567 gcc_assert (get_expr_value_id (e) == v);
568
569 if (v >= VEC_length (bitmap_set_t, value_expressions))
570 {
571 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
572 v + 1);
573 }
574
575 set = VEC_index (bitmap_set_t, value_expressions, v);
576 if (!set)
577 {
578 set = bitmap_set_new ();
579 VEC_replace (bitmap_set_t, value_expressions, v, set);
580 }
581
582 bitmap_insert_into_set_1 (set, e, true);
583 }
584
585 /* Create a new bitmap set and return it. */
586
587 static bitmap_set_t
588 bitmap_set_new (void)
589 {
590 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
591 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
592 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
593 return ret;
594 }
595
596 /* Return the value id for a PRE expression EXPR. */
597
598 static unsigned int
599 get_expr_value_id (pre_expr expr)
600 {
601 switch (expr->kind)
602 {
603 case CONSTANT:
604 {
605 unsigned int id;
606 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
607 if (id == 0)
608 {
609 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
610 add_to_value (id, expr);
611 }
612 return id;
613 }
614 case NAME:
615 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
616 case NARY:
617 return PRE_EXPR_NARY (expr)->value_id;
618 case REFERENCE:
619 return PRE_EXPR_REFERENCE (expr)->value_id;
620 default:
621 gcc_unreachable ();
622 }
623 }
624
625 /* Remove an expression EXPR from a bitmapped set. */
626
627 static void
628 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
629 {
630 unsigned int val = get_expr_value_id (expr);
631 if (!value_id_constant_p (val))
632 {
633 bitmap_clear_bit (set->values, val);
634 bitmap_clear_bit (set->expressions, get_expression_id (expr));
635 }
636 }
637
638 static void
639 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
640 bool allow_constants)
641 {
642 unsigned int val = get_expr_value_id (expr);
643 if (allow_constants || !value_id_constant_p (val))
644 {
645 /* We specifically expect this and only this function to be able to
646 insert constants into a set. */
647 bitmap_set_bit (set->values, val);
648 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
649 }
650 }
651
652 /* Insert an expression EXPR into a bitmapped set. */
653
654 static void
655 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
656 {
657 bitmap_insert_into_set_1 (set, expr, false);
658 }
659
660 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
661
662 static void
663 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
664 {
665 bitmap_copy (dest->expressions, orig->expressions);
666 bitmap_copy (dest->values, orig->values);
667 }
668
669
670 /* Free memory used up by SET. */
671 static void
672 bitmap_set_free (bitmap_set_t set)
673 {
674 BITMAP_FREE (set->expressions);
675 BITMAP_FREE (set->values);
676 }
677
678
679 /* Generate an topological-ordered array of bitmap set SET. */
680
681 static VEC(pre_expr, heap) *
682 sorted_array_from_bitmap_set (bitmap_set_t set)
683 {
684 unsigned int i, j;
685 bitmap_iterator bi, bj;
686 VEC(pre_expr, heap) *result = NULL;
687
688 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
689 {
690 /* The number of expressions having a given value is usually
691 relatively small. Thus, rather than making a vector of all
692 the expressions and sorting it by value-id, we walk the values
693 and check in the reverse mapping that tells us what expressions
694 have a given value, to filter those in our set. As a result,
695 the expressions are inserted in value-id order, which means
696 topological order.
697
698 If this is somehow a significant lose for some cases, we can
699 choose which set to walk based on the set size. */
700 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
701 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
702 {
703 if (bitmap_bit_p (set->expressions, j))
704 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
705 }
706 }
707
708 return result;
709 }
710
711 /* Perform bitmapped set operation DEST &= ORIG. */
712
713 static void
714 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
715 {
716 bitmap_iterator bi;
717 unsigned int i;
718
719 if (dest != orig)
720 {
721 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
722
723 bitmap_and_into (dest->values, orig->values);
724 bitmap_copy (temp, dest->expressions);
725 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
726 {
727 pre_expr expr = expression_for_id (i);
728 unsigned int value_id = get_expr_value_id (expr);
729 if (!bitmap_bit_p (dest->values, value_id))
730 bitmap_clear_bit (dest->expressions, i);
731 }
732 BITMAP_FREE (temp);
733 }
734 }
735
736 /* Subtract all values and expressions contained in ORIG from DEST. */
737
738 static bitmap_set_t
739 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
740 {
741 bitmap_set_t result = bitmap_set_new ();
742 bitmap_iterator bi;
743 unsigned int i;
744
745 bitmap_and_compl (result->expressions, dest->expressions,
746 orig->expressions);
747
748 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
749 {
750 pre_expr expr = expression_for_id (i);
751 unsigned int value_id = get_expr_value_id (expr);
752 bitmap_set_bit (result->values, value_id);
753 }
754
755 return result;
756 }
757
758 /* Subtract all the values in bitmap set B from bitmap set A. */
759
760 static void
761 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
762 {
763 unsigned int i;
764 bitmap_iterator bi;
765 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
766
767 bitmap_copy (temp, a->expressions);
768 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
769 {
770 pre_expr expr = expression_for_id (i);
771 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
772 bitmap_remove_from_set (a, expr);
773 }
774 BITMAP_FREE (temp);
775 }
776
777
778 /* Return true if bitmapped set SET contains the value VALUE_ID. */
779
780 static bool
781 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
782 {
783 if (value_id_constant_p (value_id))
784 return true;
785
786 if (!set || bitmap_empty_p (set->expressions))
787 return false;
788
789 return bitmap_bit_p (set->values, value_id);
790 }
791
792 static inline bool
793 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
794 {
795 return bitmap_bit_p (set->expressions, get_expression_id (expr));
796 }
797
798 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
799
800 static void
801 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
802 const pre_expr expr)
803 {
804 bitmap_set_t exprset;
805 unsigned int i;
806 bitmap_iterator bi;
807
808 if (value_id_constant_p (lookfor))
809 return;
810
811 if (!bitmap_set_contains_value (set, lookfor))
812 return;
813
814 /* The number of expressions having a given value is usually
815 significantly less than the total number of expressions in SET.
816 Thus, rather than check, for each expression in SET, whether it
817 has the value LOOKFOR, we walk the reverse mapping that tells us
818 what expressions have a given value, and see if any of those
819 expressions are in our set. For large testcases, this is about
820 5-10x faster than walking the bitmap. If this is somehow a
821 significant lose for some cases, we can choose which set to walk
822 based on the set size. */
823 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
824 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
825 {
826 if (bitmap_bit_p (set->expressions, i))
827 {
828 bitmap_clear_bit (set->expressions, i);
829 bitmap_set_bit (set->expressions, get_expression_id (expr));
830 return;
831 }
832 }
833 }
834
835 /* Return true if two bitmap sets are equal. */
836
837 static bool
838 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
839 {
840 return bitmap_equal_p (a->values, b->values);
841 }
842
843 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
844 and add it otherwise. */
845
846 static void
847 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
848 {
849 unsigned int val = get_expr_value_id (expr);
850
851 if (bitmap_set_contains_value (set, val))
852 bitmap_set_replace_value (set, val, expr);
853 else
854 bitmap_insert_into_set (set, expr);
855 }
856
857 /* Insert EXPR into SET if EXPR's value is not already present in
858 SET. */
859
860 static void
861 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
862 {
863 unsigned int val = get_expr_value_id (expr);
864
865 if (value_id_constant_p (val))
866 return;
867
868 if (!bitmap_set_contains_value (set, val))
869 bitmap_insert_into_set (set, expr);
870 }
871
872 /* Print out EXPR to outfile. */
873
874 static void
875 print_pre_expr (FILE *outfile, const pre_expr expr)
876 {
877 switch (expr->kind)
878 {
879 case CONSTANT:
880 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
881 break;
882 case NAME:
883 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
884 break;
885 case NARY:
886 {
887 unsigned int i;
888 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
889 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
890 for (i = 0; i < nary->length; i++)
891 {
892 print_generic_expr (outfile, nary->op[i], 0);
893 if (i != (unsigned) nary->length - 1)
894 fprintf (outfile, ",");
895 }
896 fprintf (outfile, "}");
897 }
898 break;
899
900 case REFERENCE:
901 {
902 vn_reference_op_t vro;
903 unsigned int i;
904 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
905 fprintf (outfile, "{");
906 for (i = 0;
907 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
908 i++)
909 {
910 bool closebrace = false;
911 if (vro->opcode != SSA_NAME
912 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
913 {
914 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
915 if (vro->op0)
916 {
917 fprintf (outfile, "<");
918 closebrace = true;
919 }
920 }
921 if (vro->op0)
922 {
923 print_generic_expr (outfile, vro->op0, 0);
924 if (vro->op1)
925 {
926 fprintf (outfile, ",");
927 print_generic_expr (outfile, vro->op1, 0);
928 }
929 if (vro->op2)
930 {
931 fprintf (outfile, ",");
932 print_generic_expr (outfile, vro->op2, 0);
933 }
934 }
935 if (closebrace)
936 fprintf (outfile, ">");
937 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
938 fprintf (outfile, ",");
939 }
940 fprintf (outfile, "}");
941 if (ref->vuse)
942 {
943 fprintf (outfile, "@");
944 print_generic_expr (outfile, ref->vuse, 0);
945 }
946 }
947 break;
948 }
949 }
950 void debug_pre_expr (pre_expr);
951
952 /* Like print_pre_expr but always prints to stderr. */
953 void
954 debug_pre_expr (pre_expr e)
955 {
956 print_pre_expr (stderr, e);
957 fprintf (stderr, "\n");
958 }
959
960 /* Print out SET to OUTFILE. */
961
962 static void
963 print_bitmap_set (FILE *outfile, bitmap_set_t set,
964 const char *setname, int blockindex)
965 {
966 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
967 if (set)
968 {
969 bool first = true;
970 unsigned i;
971 bitmap_iterator bi;
972
973 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
974 {
975 const pre_expr expr = expression_for_id (i);
976
977 if (!first)
978 fprintf (outfile, ", ");
979 first = false;
980 print_pre_expr (outfile, expr);
981
982 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
983 }
984 }
985 fprintf (outfile, " }\n");
986 }
987
988 void debug_bitmap_set (bitmap_set_t);
989
990 void
991 debug_bitmap_set (bitmap_set_t set)
992 {
993 print_bitmap_set (stderr, set, "debug", 0);
994 }
995
996 /* Print out the expressions that have VAL to OUTFILE. */
997
998 void
999 print_value_expressions (FILE *outfile, unsigned int val)
1000 {
1001 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1002 if (set)
1003 {
1004 char s[10];
1005 sprintf (s, "%04d", val);
1006 print_bitmap_set (outfile, set, s, 0);
1007 }
1008 }
1009
1010
1011 void
1012 debug_value_expressions (unsigned int val)
1013 {
1014 print_value_expressions (stderr, val);
1015 }
1016
1017 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1018 represent it. */
1019
1020 static pre_expr
1021 get_or_alloc_expr_for_constant (tree constant)
1022 {
1023 unsigned int result_id;
1024 unsigned int value_id;
1025 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1026 newexpr->kind = CONSTANT;
1027 PRE_EXPR_CONSTANT (newexpr) = constant;
1028 result_id = lookup_expression_id (newexpr);
1029 if (result_id != 0)
1030 {
1031 pool_free (pre_expr_pool, newexpr);
1032 newexpr = expression_for_id (result_id);
1033 return newexpr;
1034 }
1035 value_id = get_or_alloc_constant_value_id (constant);
1036 get_or_alloc_expression_id (newexpr);
1037 add_to_value (value_id, newexpr);
1038 return newexpr;
1039 }
1040
1041 /* Given a value id V, find the actual tree representing the constant
1042 value if there is one, and return it. Return NULL if we can't find
1043 a constant. */
1044
1045 static tree
1046 get_constant_for_value_id (unsigned int v)
1047 {
1048 if (value_id_constant_p (v))
1049 {
1050 unsigned int i;
1051 bitmap_iterator bi;
1052 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1053
1054 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1055 {
1056 pre_expr expr = expression_for_id (i);
1057 if (expr->kind == CONSTANT)
1058 return PRE_EXPR_CONSTANT (expr);
1059 }
1060 }
1061 return NULL;
1062 }
1063
1064 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1065 Currently only supports constants and SSA_NAMES. */
1066 static pre_expr
1067 get_or_alloc_expr_for (tree t)
1068 {
1069 if (TREE_CODE (t) == SSA_NAME)
1070 return get_or_alloc_expr_for_name (t);
1071 else if (is_gimple_min_invariant (t))
1072 return get_or_alloc_expr_for_constant (t);
1073 else
1074 {
1075 /* More complex expressions can result from SCCVN expression
1076 simplification that inserts values for them. As they all
1077 do not have VOPs the get handled by the nary ops struct. */
1078 vn_nary_op_t result;
1079 unsigned int result_id;
1080 vn_nary_op_lookup (t, &result);
1081 if (result != NULL)
1082 {
1083 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1084 e->kind = NARY;
1085 PRE_EXPR_NARY (e) = result;
1086 result_id = lookup_expression_id (e);
1087 if (result_id != 0)
1088 {
1089 pool_free (pre_expr_pool, e);
1090 e = expression_for_id (result_id);
1091 return e;
1092 }
1093 alloc_expression_id (e);
1094 return e;
1095 }
1096 }
1097 return NULL;
1098 }
1099
1100 /* Return the folded version of T if T, when folded, is a gimple
1101 min_invariant. Otherwise, return T. */
1102
1103 static pre_expr
1104 fully_constant_expression (pre_expr e)
1105 {
1106 switch (e->kind)
1107 {
1108 case CONSTANT:
1109 return e;
1110 case NARY:
1111 {
1112 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1113 switch (TREE_CODE_CLASS (nary->opcode))
1114 {
1115 case tcc_expression:
1116 if (nary->opcode == TRUTH_NOT_EXPR)
1117 goto do_unary;
1118 if (nary->opcode != TRUTH_AND_EXPR
1119 && nary->opcode != TRUTH_OR_EXPR
1120 && nary->opcode != TRUTH_XOR_EXPR)
1121 return e;
1122 /* Fallthrough. */
1123 case tcc_binary:
1124 case tcc_comparison:
1125 {
1126 /* We have to go from trees to pre exprs to value ids to
1127 constants. */
1128 tree naryop0 = nary->op[0];
1129 tree naryop1 = nary->op[1];
1130 tree result;
1131 if (!is_gimple_min_invariant (naryop0))
1132 {
1133 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1134 unsigned int vrep0 = get_expr_value_id (rep0);
1135 tree const0 = get_constant_for_value_id (vrep0);
1136 if (const0)
1137 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1138 }
1139 if (!is_gimple_min_invariant (naryop1))
1140 {
1141 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1142 unsigned int vrep1 = get_expr_value_id (rep1);
1143 tree const1 = get_constant_for_value_id (vrep1);
1144 if (const1)
1145 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1146 }
1147 result = fold_binary (nary->opcode, nary->type,
1148 naryop0, naryop1);
1149 if (result && is_gimple_min_invariant (result))
1150 return get_or_alloc_expr_for_constant (result);
1151 /* We might have simplified the expression to a
1152 SSA_NAME for example from x_1 * 1. But we cannot
1153 insert a PHI for x_1 unconditionally as x_1 might
1154 not be available readily. */
1155 return e;
1156 }
1157 case tcc_reference:
1158 if (nary->opcode != REALPART_EXPR
1159 && nary->opcode != IMAGPART_EXPR
1160 && nary->opcode != VIEW_CONVERT_EXPR)
1161 return e;
1162 /* Fallthrough. */
1163 case tcc_unary:
1164 do_unary:
1165 {
1166 /* We have to go from trees to pre exprs to value ids to
1167 constants. */
1168 tree naryop0 = nary->op[0];
1169 tree const0, result;
1170 if (is_gimple_min_invariant (naryop0))
1171 const0 = naryop0;
1172 else
1173 {
1174 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1175 unsigned int vrep0 = get_expr_value_id (rep0);
1176 const0 = get_constant_for_value_id (vrep0);
1177 }
1178 result = NULL;
1179 if (const0)
1180 {
1181 tree type1 = TREE_TYPE (nary->op[0]);
1182 const0 = fold_convert (type1, const0);
1183 result = fold_unary (nary->opcode, nary->type, const0);
1184 }
1185 if (result && is_gimple_min_invariant (result))
1186 return get_or_alloc_expr_for_constant (result);
1187 return e;
1188 }
1189 default:
1190 return e;
1191 }
1192 }
1193 case REFERENCE:
1194 {
1195 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1196 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1197 vn_reference_op_t op;
1198
1199 /* Try to simplify the translated expression if it is
1200 a call to a builtin function with at most two arguments. */
1201 op = VEC_index (vn_reference_op_s, operands, 0);
1202 if (op->opcode == CALL_EXPR
1203 && TREE_CODE (op->op0) == ADDR_EXPR
1204 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1205 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1206 && VEC_length (vn_reference_op_s, operands) >= 2
1207 && VEC_length (vn_reference_op_s, operands) <= 3)
1208 {
1209 vn_reference_op_t arg0, arg1 = NULL;
1210 bool anyconst = false;
1211 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1212 if (VEC_length (vn_reference_op_s, operands) > 2)
1213 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1214 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1215 || (arg0->opcode == ADDR_EXPR
1216 && is_gimple_min_invariant (arg0->op0)))
1217 anyconst = true;
1218 if (arg1
1219 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1220 || (arg1->opcode == ADDR_EXPR
1221 && is_gimple_min_invariant (arg1->op0))))
1222 anyconst = true;
1223 if (anyconst)
1224 {
1225 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1226 arg1 ? 2 : 1,
1227 arg0->op0,
1228 arg1 ? arg1->op0 : NULL);
1229 if (folded
1230 && TREE_CODE (folded) == NOP_EXPR)
1231 folded = TREE_OPERAND (folded, 0);
1232 if (folded
1233 && is_gimple_min_invariant (folded))
1234 return get_or_alloc_expr_for_constant (folded);
1235 }
1236 }
1237 return e;
1238 }
1239 default:
1240 return e;
1241 }
1242 return e;
1243 }
1244
1245 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1246 it has the value it would have in BLOCK. */
1247
1248 static tree
1249 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1250 alias_set_type set, tree type, tree vuse,
1251 basic_block phiblock,
1252 basic_block block)
1253 {
1254 gimple phi = SSA_NAME_DEF_STMT (vuse);
1255 ao_ref ref;
1256
1257 if (gimple_bb (phi) != phiblock)
1258 return vuse;
1259
1260 if (gimple_code (phi) == GIMPLE_PHI)
1261 {
1262 edge e = find_edge (block, phiblock);
1263 return PHI_ARG_DEF (phi, e->dest_idx);
1264 }
1265
1266 if (!ao_ref_init_from_vn_reference (&ref, set, type, operands))
1267 return NULL_TREE;
1268
1269 /* Use the alias-oracle to find either the PHI node in this block,
1270 the first VUSE used in this block that is equivalent to vuse or
1271 the first VUSE which definition in this block kills the value. */
1272 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1273 {
1274 vuse = gimple_vuse (phi);
1275 phi = SSA_NAME_DEF_STMT (vuse);
1276 if (gimple_bb (phi) != phiblock)
1277 return vuse;
1278 if (gimple_code (phi) == GIMPLE_PHI)
1279 {
1280 edge e = find_edge (block, phiblock);
1281 return PHI_ARG_DEF (phi, e->dest_idx);
1282 }
1283 }
1284
1285 return NULL_TREE;
1286 }
1287
1288 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1289 SET2. This is used to avoid making a set consisting of the union
1290 of PA_IN and ANTIC_IN during insert. */
1291
1292 static inline pre_expr
1293 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1294 {
1295 pre_expr result;
1296
1297 result = bitmap_find_leader (set1, val, NULL);
1298 if (!result && set2)
1299 result = bitmap_find_leader (set2, val, NULL);
1300 return result;
1301 }
1302
1303 /* Get the tree type for our PRE expression e. */
1304
1305 static tree
1306 get_expr_type (const pre_expr e)
1307 {
1308 switch (e->kind)
1309 {
1310 case NAME:
1311 return TREE_TYPE (PRE_EXPR_NAME (e));
1312 case CONSTANT:
1313 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1314 case REFERENCE:
1315 return PRE_EXPR_REFERENCE (e)->type;
1316 case NARY:
1317 return PRE_EXPR_NARY (e)->type;
1318 }
1319 gcc_unreachable();
1320 }
1321
1322 /* Get a representative SSA_NAME for a given expression.
1323 Since all of our sub-expressions are treated as values, we require
1324 them to be SSA_NAME's for simplicity.
1325 Prior versions of GVNPRE used to use "value handles" here, so that
1326 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1327 either case, the operands are really values (IE we do not expect
1328 them to be usable without finding leaders). */
1329
1330 static tree
1331 get_representative_for (const pre_expr e)
1332 {
1333 tree exprtype;
1334 tree name;
1335 unsigned int value_id = get_expr_value_id (e);
1336
1337 switch (e->kind)
1338 {
1339 case NAME:
1340 return PRE_EXPR_NAME (e);
1341 case CONSTANT:
1342 return PRE_EXPR_CONSTANT (e);
1343 case NARY:
1344 case REFERENCE:
1345 {
1346 /* Go through all of the expressions representing this value
1347 and pick out an SSA_NAME. */
1348 unsigned int i;
1349 bitmap_iterator bi;
1350 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1351 value_id);
1352 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1353 {
1354 pre_expr rep = expression_for_id (i);
1355 if (rep->kind == NAME)
1356 return PRE_EXPR_NAME (rep);
1357 }
1358 }
1359 break;
1360 }
1361 /* If we reached here we couldn't find an SSA_NAME. This can
1362 happen when we've discovered a value that has never appeared in
1363 the program as set to an SSA_NAME, most likely as the result of
1364 phi translation. */
1365 if (dump_file)
1366 {
1367 fprintf (dump_file,
1368 "Could not find SSA_NAME representative for expression:");
1369 print_pre_expr (dump_file, e);
1370 fprintf (dump_file, "\n");
1371 }
1372
1373 exprtype = get_expr_type (e);
1374
1375 /* Build and insert the assignment of the end result to the temporary
1376 that we will return. */
1377 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1378 {
1379 pretemp = create_tmp_var (exprtype, "pretmp");
1380 get_var_ann (pretemp);
1381 }
1382
1383 name = make_ssa_name (pretemp, gimple_build_nop ());
1384 VN_INFO_GET (name)->value_id = value_id;
1385 if (e->kind == CONSTANT)
1386 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1387 else
1388 VN_INFO (name)->valnum = name;
1389
1390 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1391 if (dump_file)
1392 {
1393 fprintf (dump_file, "Created SSA_NAME representative ");
1394 print_generic_expr (dump_file, name, 0);
1395 fprintf (dump_file, " for expression:");
1396 print_pre_expr (dump_file, e);
1397 fprintf (dump_file, "\n");
1398 }
1399
1400 return name;
1401 }
1402
1403
1404
1405
1406 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1407 the phis in PRED. SEEN is a bitmap saying which expression we have
1408 translated since we started translation of the toplevel expression.
1409 Return NULL if we can't find a leader for each part of the
1410 translated expression. */
1411
1412 static pre_expr
1413 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1414 basic_block pred, basic_block phiblock, bitmap seen)
1415 {
1416 pre_expr oldexpr = expr;
1417 pre_expr phitrans;
1418
1419 if (!expr)
1420 return NULL;
1421
1422 if (value_id_constant_p (get_expr_value_id (expr)))
1423 return expr;
1424
1425 phitrans = phi_trans_lookup (expr, pred);
1426 if (phitrans)
1427 return phitrans;
1428
1429 /* Prevent cycles when we have recursively dependent leaders. This
1430 can only happen when phi translating the maximal set. */
1431 if (seen)
1432 {
1433 unsigned int expr_id = get_expression_id (expr);
1434 if (bitmap_bit_p (seen, expr_id))
1435 return NULL;
1436 bitmap_set_bit (seen, expr_id);
1437 }
1438
1439 switch (expr->kind)
1440 {
1441 /* Constants contain no values that need translation. */
1442 case CONSTANT:
1443 return expr;
1444
1445 case NARY:
1446 {
1447 unsigned int i;
1448 bool changed = false;
1449 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1450 struct vn_nary_op_s newnary;
1451 /* The NARY structure is only guaranteed to have been
1452 allocated to the nary->length operands. */
1453 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1454 - sizeof (tree) * (4 - nary->length)));
1455
1456 for (i = 0; i < newnary.length; i++)
1457 {
1458 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1459 continue;
1460 else
1461 {
1462 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1463 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1464 pre_expr result = phi_translate_1 (leader, set1, set2,
1465 pred, phiblock, seen);
1466 if (result && result != leader)
1467 {
1468 tree name = get_representative_for (result);
1469 if (!name)
1470 return NULL;
1471 newnary.op[i] = name;
1472 }
1473 else if (!result)
1474 return NULL;
1475
1476 changed |= newnary.op[i] != nary->op[i];
1477 }
1478 }
1479 if (changed)
1480 {
1481 pre_expr constant;
1482
1483 tree result = vn_nary_op_lookup_pieces (newnary.length,
1484 newnary.opcode,
1485 newnary.type,
1486 newnary.op[0],
1487 newnary.op[1],
1488 newnary.op[2],
1489 newnary.op[3],
1490 &nary);
1491 unsigned int new_val_id;
1492
1493 expr = (pre_expr) pool_alloc (pre_expr_pool);
1494 expr->kind = NARY;
1495 expr->id = 0;
1496 if (result && is_gimple_min_invariant (result))
1497 return get_or_alloc_expr_for_constant (result);
1498
1499
1500 if (nary)
1501 {
1502 PRE_EXPR_NARY (expr) = nary;
1503 constant = fully_constant_expression (expr);
1504 if (constant != expr)
1505 return constant;
1506
1507 new_val_id = nary->value_id;
1508 get_or_alloc_expression_id (expr);
1509 }
1510 else
1511 {
1512 new_val_id = get_next_value_id ();
1513 VEC_safe_grow_cleared (bitmap_set_t, heap,
1514 value_expressions,
1515 get_max_value_id() + 1);
1516 nary = vn_nary_op_insert_pieces (newnary.length,
1517 newnary.opcode,
1518 newnary.type,
1519 newnary.op[0],
1520 newnary.op[1],
1521 newnary.op[2],
1522 newnary.op[3],
1523 result, new_val_id);
1524 PRE_EXPR_NARY (expr) = nary;
1525 constant = fully_constant_expression (expr);
1526 if (constant != expr)
1527 return constant;
1528 get_or_alloc_expression_id (expr);
1529 }
1530 add_to_value (new_val_id, expr);
1531 }
1532 phi_trans_add (oldexpr, expr, pred);
1533 return expr;
1534 }
1535 break;
1536
1537 case REFERENCE:
1538 {
1539 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1540 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1541 tree vuse = ref->vuse;
1542 tree newvuse = vuse;
1543 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1544 bool changed = false;
1545 unsigned int i, j;
1546 vn_reference_op_t operand;
1547 vn_reference_t newref;
1548
1549 for (i = 0, j = 0;
1550 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1551 {
1552 pre_expr opresult;
1553 pre_expr leader;
1554 tree oldop0 = operand->op0;
1555 tree oldop1 = operand->op1;
1556 tree oldop2 = operand->op2;
1557 tree op0 = oldop0;
1558 tree op1 = oldop1;
1559 tree op2 = oldop2;
1560 tree type = operand->type;
1561 vn_reference_op_s newop = *operand;
1562
1563 if (op0 && TREE_CODE (op0) == SSA_NAME)
1564 {
1565 unsigned int op_val_id = VN_INFO (op0)->value_id;
1566 leader = find_leader_in_sets (op_val_id, set1, set2);
1567 opresult = phi_translate_1 (leader, set1, set2,
1568 pred, phiblock, seen);
1569 if (opresult && opresult != leader)
1570 {
1571 tree name = get_representative_for (opresult);
1572 if (!name)
1573 break;
1574 op0 = name;
1575 }
1576 else if (!opresult)
1577 break;
1578 }
1579 changed |= op0 != oldop0;
1580
1581 if (op1 && TREE_CODE (op1) == SSA_NAME)
1582 {
1583 unsigned int op_val_id = VN_INFO (op1)->value_id;
1584 leader = find_leader_in_sets (op_val_id, set1, set2);
1585 opresult = phi_translate_1 (leader, set1, set2,
1586 pred, phiblock, seen);
1587 if (opresult && opresult != leader)
1588 {
1589 tree name = get_representative_for (opresult);
1590 if (!name)
1591 break;
1592 op1 = name;
1593 }
1594 else if (!opresult)
1595 break;
1596 }
1597 /* We can't possibly insert these. */
1598 else if (op1 && !is_gimple_min_invariant (op1))
1599 break;
1600 changed |= op1 != oldop1;
1601 if (op2 && TREE_CODE (op2) == SSA_NAME)
1602 {
1603 unsigned int op_val_id = VN_INFO (op2)->value_id;
1604 leader = find_leader_in_sets (op_val_id, set1, set2);
1605 opresult = phi_translate_1 (leader, set1, set2,
1606 pred, phiblock, seen);
1607 if (opresult && opresult != leader)
1608 {
1609 tree name = get_representative_for (opresult);
1610 if (!name)
1611 break;
1612 op2 = name;
1613 }
1614 else if (!opresult)
1615 break;
1616 }
1617 /* We can't possibly insert these. */
1618 else if (op2 && !is_gimple_min_invariant (op2))
1619 break;
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 = NULL;
2199
2200 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2201 FOR_EACH_EDGE (e, ei, block->succs)
2202 {
2203 if (!first
2204 && BB_VISITED (e->dest))
2205 first = e->dest;
2206 else if (BB_VISITED (e->dest))
2207 VEC_quick_push (basic_block, worklist, e->dest);
2208 }
2209
2210 /* Of multiple successors we have to have visited one already. */
2211 if (!first)
2212 {
2213 SET_BIT (changed_blocks, block->index);
2214 BB_VISITED (block) = 0;
2215 BB_DEFERRED (block) = 1;
2216 changed = true;
2217 VEC_free (basic_block, heap, worklist);
2218 goto maybe_dump_sets;
2219 }
2220
2221 if (phi_nodes (first))
2222 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2223 else
2224 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2225
2226 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2227 {
2228 if (phi_nodes (bprime))
2229 {
2230 bitmap_set_t tmp = bitmap_set_new ();
2231 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2232 bitmap_set_and (ANTIC_OUT, tmp);
2233 bitmap_set_free (tmp);
2234 }
2235 else
2236 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2237 }
2238 VEC_free (basic_block, heap, worklist);
2239 }
2240
2241 /* Generate ANTIC_OUT - TMP_GEN. */
2242 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2243
2244 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2245 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2246 TMP_GEN (block));
2247
2248 /* Then union in the ANTIC_OUT - TMP_GEN values,
2249 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2250 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2251 bitmap_value_insert_into_set (ANTIC_IN (block),
2252 expression_for_id (bii));
2253
2254 clean (ANTIC_IN (block), block);
2255
2256 /* !old->expressions can happen when we deferred a block. */
2257 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2258 {
2259 changed = true;
2260 SET_BIT (changed_blocks, block->index);
2261 FOR_EACH_EDGE (e, ei, block->preds)
2262 SET_BIT (changed_blocks, e->src->index);
2263 }
2264 else
2265 RESET_BIT (changed_blocks, block->index);
2266
2267 maybe_dump_sets:
2268 if (dump_file && (dump_flags & TDF_DETAILS))
2269 {
2270 if (!BB_DEFERRED (block) || BB_VISITED (block))
2271 {
2272 if (ANTIC_OUT)
2273 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2274
2275 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2276 block->index);
2277
2278 if (S)
2279 print_bitmap_set (dump_file, S, "S", block->index);
2280 }
2281 else
2282 {
2283 fprintf (dump_file,
2284 "Block %d was deferred for a future iteration.\n",
2285 block->index);
2286 }
2287 }
2288 if (old)
2289 bitmap_set_free (old);
2290 if (S)
2291 bitmap_set_free (S);
2292 if (ANTIC_OUT)
2293 bitmap_set_free (ANTIC_OUT);
2294 return changed;
2295 }
2296
2297 /* Compute PARTIAL_ANTIC for BLOCK.
2298
2299 If succs(BLOCK) > 1 then
2300 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2301 in ANTIC_OUT for all succ(BLOCK)
2302 else if succs(BLOCK) == 1 then
2303 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2304
2305 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2306 - ANTIC_IN[BLOCK])
2307
2308 */
2309 static bool
2310 compute_partial_antic_aux (basic_block block,
2311 bool block_has_abnormal_pred_edge)
2312 {
2313 bool changed = false;
2314 bitmap_set_t old_PA_IN;
2315 bitmap_set_t PA_OUT;
2316 edge e;
2317 edge_iterator ei;
2318 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2319
2320 old_PA_IN = PA_OUT = NULL;
2321
2322 /* If any edges from predecessors are abnormal, antic_in is empty,
2323 so do nothing. */
2324 if (block_has_abnormal_pred_edge)
2325 goto maybe_dump_sets;
2326
2327 /* If there are too many partially anticipatable values in the
2328 block, phi_translate_set can take an exponential time: stop
2329 before the translation starts. */
2330 if (max_pa
2331 && single_succ_p (block)
2332 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2333 goto maybe_dump_sets;
2334
2335 old_PA_IN = PA_IN (block);
2336 PA_OUT = bitmap_set_new ();
2337
2338 /* If the block has no successors, ANTIC_OUT is empty. */
2339 if (EDGE_COUNT (block->succs) == 0)
2340 ;
2341 /* If we have one successor, we could have some phi nodes to
2342 translate through. Note that we can't phi translate across DFS
2343 back edges in partial antic, because it uses a union operation on
2344 the successors. For recurrences like IV's, we will end up
2345 generating a new value in the set on each go around (i + 3 (VH.1)
2346 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2347 else if (single_succ_p (block))
2348 {
2349 basic_block succ = single_succ (block);
2350 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2351 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2352 }
2353 /* If we have multiple successors, we take the union of all of
2354 them. */
2355 else
2356 {
2357 VEC(basic_block, heap) * worklist;
2358 size_t i;
2359 basic_block bprime;
2360
2361 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2362 FOR_EACH_EDGE (e, ei, block->succs)
2363 {
2364 if (e->flags & EDGE_DFS_BACK)
2365 continue;
2366 VEC_quick_push (basic_block, worklist, e->dest);
2367 }
2368 if (VEC_length (basic_block, worklist) > 0)
2369 {
2370 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2371 {
2372 unsigned int i;
2373 bitmap_iterator bi;
2374
2375 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2376 bitmap_value_insert_into_set (PA_OUT,
2377 expression_for_id (i));
2378 if (phi_nodes (bprime))
2379 {
2380 bitmap_set_t pa_in = bitmap_set_new ();
2381 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2382 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2383 bitmap_value_insert_into_set (PA_OUT,
2384 expression_for_id (i));
2385 bitmap_set_free (pa_in);
2386 }
2387 else
2388 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2389 bitmap_value_insert_into_set (PA_OUT,
2390 expression_for_id (i));
2391 }
2392 }
2393 VEC_free (basic_block, heap, worklist);
2394 }
2395
2396 /* PA_IN starts with PA_OUT - TMP_GEN.
2397 Then we subtract things from ANTIC_IN. */
2398 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2399
2400 /* For partial antic, we want to put back in the phi results, since
2401 we will properly avoid making them partially antic over backedges. */
2402 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2403 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2404
2405 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2406 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2407
2408 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2409
2410 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2411 {
2412 changed = true;
2413 SET_BIT (changed_blocks, block->index);
2414 FOR_EACH_EDGE (e, ei, block->preds)
2415 SET_BIT (changed_blocks, e->src->index);
2416 }
2417 else
2418 RESET_BIT (changed_blocks, block->index);
2419
2420 maybe_dump_sets:
2421 if (dump_file && (dump_flags & TDF_DETAILS))
2422 {
2423 if (PA_OUT)
2424 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2425
2426 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2427 }
2428 if (old_PA_IN)
2429 bitmap_set_free (old_PA_IN);
2430 if (PA_OUT)
2431 bitmap_set_free (PA_OUT);
2432 return changed;
2433 }
2434
2435 /* Compute ANTIC and partial ANTIC sets. */
2436
2437 static void
2438 compute_antic (void)
2439 {
2440 bool changed = true;
2441 int num_iterations = 0;
2442 basic_block block;
2443 int i;
2444
2445 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2446 We pre-build the map of blocks with incoming abnormal edges here. */
2447 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2448 sbitmap_zero (has_abnormal_preds);
2449
2450 FOR_EACH_BB (block)
2451 {
2452 edge_iterator ei;
2453 edge e;
2454
2455 FOR_EACH_EDGE (e, ei, block->preds)
2456 {
2457 e->flags &= ~EDGE_DFS_BACK;
2458 if (e->flags & EDGE_ABNORMAL)
2459 {
2460 SET_BIT (has_abnormal_preds, block->index);
2461 break;
2462 }
2463 }
2464
2465 BB_VISITED (block) = 0;
2466 BB_DEFERRED (block) = 0;
2467 /* While we are here, give empty ANTIC_IN sets to each block. */
2468 ANTIC_IN (block) = bitmap_set_new ();
2469 PA_IN (block) = bitmap_set_new ();
2470 }
2471
2472 /* At the exit block we anticipate nothing. */
2473 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2474 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2475 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2476
2477 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2478 sbitmap_ones (changed_blocks);
2479 while (changed)
2480 {
2481 if (dump_file && (dump_flags & TDF_DETAILS))
2482 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2483 num_iterations++;
2484 changed = false;
2485 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2486 {
2487 if (TEST_BIT (changed_blocks, postorder[i]))
2488 {
2489 basic_block block = BASIC_BLOCK (postorder[i]);
2490 changed |= compute_antic_aux (block,
2491 TEST_BIT (has_abnormal_preds,
2492 block->index));
2493 }
2494 }
2495 #ifdef ENABLE_CHECKING
2496 /* Theoretically possible, but *highly* unlikely. */
2497 gcc_assert (num_iterations < 500);
2498 #endif
2499 }
2500
2501 statistics_histogram_event (cfun, "compute_antic iterations",
2502 num_iterations);
2503
2504 if (do_partial_partial)
2505 {
2506 sbitmap_ones (changed_blocks);
2507 mark_dfs_back_edges ();
2508 num_iterations = 0;
2509 changed = true;
2510 while (changed)
2511 {
2512 if (dump_file && (dump_flags & TDF_DETAILS))
2513 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2514 num_iterations++;
2515 changed = false;
2516 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2517 {
2518 if (TEST_BIT (changed_blocks, postorder[i]))
2519 {
2520 basic_block block = BASIC_BLOCK (postorder[i]);
2521 changed
2522 |= compute_partial_antic_aux (block,
2523 TEST_BIT (has_abnormal_preds,
2524 block->index));
2525 }
2526 }
2527 #ifdef ENABLE_CHECKING
2528 /* Theoretically possible, but *highly* unlikely. */
2529 gcc_assert (num_iterations < 500);
2530 #endif
2531 }
2532 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2533 num_iterations);
2534 }
2535 sbitmap_free (has_abnormal_preds);
2536 sbitmap_free (changed_blocks);
2537 }
2538
2539 /* Return true if we can value number the call in STMT. This is true
2540 if we have a pure or constant call. */
2541
2542 static bool
2543 can_value_number_call (gimple stmt)
2544 {
2545 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2546 return true;
2547 return false;
2548 }
2549
2550 /* Return true if OP is a tree which we can perform PRE on.
2551 This may not match the operations we can value number, but in
2552 a perfect world would. */
2553
2554 static bool
2555 can_PRE_operation (tree op)
2556 {
2557 return UNARY_CLASS_P (op)
2558 || BINARY_CLASS_P (op)
2559 || COMPARISON_CLASS_P (op)
2560 || TREE_CODE (op) == INDIRECT_REF
2561 || TREE_CODE (op) == COMPONENT_REF
2562 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2563 || TREE_CODE (op) == CALL_EXPR
2564 || TREE_CODE (op) == ARRAY_REF;
2565 }
2566
2567
2568 /* Inserted expressions are placed onto this worklist, which is used
2569 for performing quick dead code elimination of insertions we made
2570 that didn't turn out to be necessary. */
2571 static VEC(gimple,heap) *inserted_exprs;
2572 static bitmap inserted_phi_names;
2573
2574 /* Pool allocated fake store expressions are placed onto this
2575 worklist, which, after performing dead code elimination, is walked
2576 to see which expressions need to be put into GC'able memory */
2577 static VEC(gimple, heap) *need_creation;
2578
2579 /* The actual worker for create_component_ref_by_pieces. */
2580
2581 static tree
2582 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2583 unsigned int *operand, gimple_seq *stmts,
2584 gimple domstmt)
2585 {
2586 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2587 *operand);
2588 tree genop;
2589 ++*operand;
2590 switch (currop->opcode)
2591 {
2592 case CALL_EXPR:
2593 {
2594 tree folded, sc = currop->op1;
2595 unsigned int nargs = 0;
2596 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2597 ref->operands) - 1);
2598 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2599 {
2600 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2601 operand, stmts,
2602 domstmt);
2603 nargs++;
2604 }
2605 folded = build_call_array (currop->type,
2606 TREE_CODE (currop->op0) == FUNCTION_DECL
2607 ? build_fold_addr_expr (currop->op0)
2608 : currop->op0,
2609 nargs, args);
2610 free (args);
2611 if (sc)
2612 {
2613 pre_expr scexpr = get_or_alloc_expr_for (sc);
2614 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2615 if (!sc)
2616 return NULL_TREE;
2617 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2618 }
2619 return folded;
2620 }
2621 break;
2622 case TARGET_MEM_REF:
2623 {
2624 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2625 *operand);
2626 pre_expr op0expr;
2627 tree genop0 = NULL_TREE;
2628 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2629 stmts, domstmt);
2630 if (!baseop)
2631 return NULL_TREE;
2632 if (currop->op0)
2633 {
2634 op0expr = get_or_alloc_expr_for (currop->op0);
2635 genop0 = find_or_generate_expression (block, op0expr,
2636 stmts, domstmt);
2637 if (!genop0)
2638 return NULL_TREE;
2639 }
2640 if (DECL_P (baseop))
2641 return build6 (TARGET_MEM_REF, currop->type,
2642 baseop, NULL_TREE,
2643 genop0, currop->op1, currop->op2,
2644 unshare_expr (nextop->op1));
2645 else
2646 return build6 (TARGET_MEM_REF, currop->type,
2647 NULL_TREE, baseop,
2648 genop0, currop->op1, currop->op2,
2649 unshare_expr (nextop->op1));
2650 }
2651 break;
2652 case ADDR_EXPR:
2653 if (currop->op0)
2654 {
2655 gcc_assert (is_gimple_min_invariant (currop->op0));
2656 return currop->op0;
2657 }
2658 /* Fallthrough. */
2659 case REALPART_EXPR:
2660 case IMAGPART_EXPR:
2661 case VIEW_CONVERT_EXPR:
2662 {
2663 tree folded;
2664 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2665 operand,
2666 stmts, domstmt);
2667 if (!genop0)
2668 return NULL_TREE;
2669 folded = fold_build1 (currop->opcode, currop->type,
2670 genop0);
2671 return folded;
2672 }
2673 break;
2674 case ALIGN_INDIRECT_REF:
2675 case MISALIGNED_INDIRECT_REF:
2676 case INDIRECT_REF:
2677 {
2678 tree folded;
2679 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2680 operand,
2681 stmts, domstmt);
2682 if (!genop1)
2683 return NULL_TREE;
2684 genop1 = fold_convert (build_pointer_type (currop->type),
2685 genop1);
2686
2687 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2688 folded = fold_build2 (currop->opcode, currop->type,
2689 genop1, currop->op1);
2690 else
2691 folded = fold_build1 (currop->opcode, currop->type,
2692 genop1);
2693 return folded;
2694 }
2695 break;
2696 case BIT_FIELD_REF:
2697 {
2698 tree folded;
2699 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2700 stmts, domstmt);
2701 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2702 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2703 tree genop1;
2704 tree genop2;
2705
2706 if (!genop0)
2707 return NULL_TREE;
2708 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2709 if (!genop1)
2710 return NULL_TREE;
2711 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2712 if (!genop2)
2713 return NULL_TREE;
2714 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2715 genop2);
2716 return folded;
2717 }
2718
2719 /* For array ref vn_reference_op's, operand 1 of the array ref
2720 is op0 of the reference op and operand 3 of the array ref is
2721 op1. */
2722 case ARRAY_RANGE_REF:
2723 case ARRAY_REF:
2724 {
2725 tree genop0;
2726 tree genop1 = currop->op0;
2727 pre_expr op1expr;
2728 tree genop2 = currop->op1;
2729 pre_expr op2expr;
2730 tree genop3 = currop->op2;
2731 pre_expr op3expr;
2732 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2733 stmts, domstmt);
2734 if (!genop0)
2735 return NULL_TREE;
2736 op1expr = get_or_alloc_expr_for (genop1);
2737 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2738 if (!genop1)
2739 return NULL_TREE;
2740 if (genop2)
2741 {
2742 op2expr = get_or_alloc_expr_for (genop2);
2743 genop2 = find_or_generate_expression (block, op2expr, stmts,
2744 domstmt);
2745 if (!genop2)
2746 return NULL_TREE;
2747 }
2748 if (genop3)
2749 {
2750 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2751 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2752 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2753 op3expr = get_or_alloc_expr_for (genop3);
2754 genop3 = find_or_generate_expression (block, op3expr, stmts,
2755 domstmt);
2756 if (!genop3)
2757 return NULL_TREE;
2758 }
2759 return build4 (currop->opcode, currop->type, genop0, genop1,
2760 genop2, genop3);
2761 }
2762 case COMPONENT_REF:
2763 {
2764 tree op0;
2765 tree op1;
2766 tree genop2 = currop->op1;
2767 pre_expr op2expr;
2768 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2769 stmts, domstmt);
2770 if (!op0)
2771 return NULL_TREE;
2772 /* op1 should be a FIELD_DECL, which are represented by
2773 themselves. */
2774 op1 = currop->op0;
2775 if (genop2)
2776 {
2777 op2expr = get_or_alloc_expr_for (genop2);
2778 genop2 = find_or_generate_expression (block, op2expr, stmts,
2779 domstmt);
2780 if (!genop2)
2781 return NULL_TREE;
2782 }
2783
2784 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2785 genop2);
2786 }
2787 break;
2788 case SSA_NAME:
2789 {
2790 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2791 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2792 return genop;
2793 }
2794 case STRING_CST:
2795 case INTEGER_CST:
2796 case COMPLEX_CST:
2797 case VECTOR_CST:
2798 case REAL_CST:
2799 case CONSTRUCTOR:
2800 case VAR_DECL:
2801 case PARM_DECL:
2802 case CONST_DECL:
2803 case RESULT_DECL:
2804 case FUNCTION_DECL:
2805 return currop->op0;
2806
2807 default:
2808 gcc_unreachable ();
2809 }
2810 }
2811
2812 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2813 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2814 trying to rename aggregates into ssa form directly, which is a no no.
2815
2816 Thus, this routine doesn't create temporaries, it just builds a
2817 single access expression for the array, calling
2818 find_or_generate_expression to build the innermost pieces.
2819
2820 This function is a subroutine of create_expression_by_pieces, and
2821 should not be called on it's own unless you really know what you
2822 are doing. */
2823
2824 static tree
2825 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2826 gimple_seq *stmts, gimple domstmt)
2827 {
2828 unsigned int op = 0;
2829 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2830 }
2831
2832 /* Find a leader for an expression, or generate one using
2833 create_expression_by_pieces if it's ANTIC but
2834 complex.
2835 BLOCK is the basic_block we are looking for leaders in.
2836 EXPR is the expression to find a leader or generate for.
2837 STMTS is the statement list to put the inserted expressions on.
2838 Returns the SSA_NAME of the LHS of the generated expression or the
2839 leader.
2840 DOMSTMT if non-NULL is a statement that should be dominated by
2841 all uses in the generated expression. If DOMSTMT is non-NULL this
2842 routine can fail and return NULL_TREE. Otherwise it will assert
2843 on failure. */
2844
2845 static tree
2846 find_or_generate_expression (basic_block block, pre_expr expr,
2847 gimple_seq *stmts, gimple domstmt)
2848 {
2849 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2850 get_expr_value_id (expr), domstmt);
2851 tree genop = NULL;
2852 if (leader)
2853 {
2854 if (leader->kind == NAME)
2855 genop = PRE_EXPR_NAME (leader);
2856 else if (leader->kind == CONSTANT)
2857 genop = PRE_EXPR_CONSTANT (leader);
2858 }
2859
2860 /* If it's still NULL, it must be a complex expression, so generate
2861 it recursively. Not so for FRE though. */
2862 if (genop == NULL
2863 && !in_fre)
2864 {
2865 bitmap_set_t exprset;
2866 unsigned int lookfor = get_expr_value_id (expr);
2867 bool handled = false;
2868 bitmap_iterator bi;
2869 unsigned int i;
2870
2871 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2872 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2873 {
2874 pre_expr temp = expression_for_id (i);
2875 if (temp->kind != NAME)
2876 {
2877 handled = true;
2878 genop = create_expression_by_pieces (block, temp, stmts,
2879 domstmt,
2880 get_expr_type (expr));
2881 break;
2882 }
2883 }
2884 if (!handled && domstmt)
2885 return NULL_TREE;
2886
2887 gcc_assert (handled);
2888 }
2889 return genop;
2890 }
2891
2892 #define NECESSARY GF_PLF_1
2893
2894 /* Create an expression in pieces, so that we can handle very complex
2895 expressions that may be ANTIC, but not necessary GIMPLE.
2896 BLOCK is the basic block the expression will be inserted into,
2897 EXPR is the expression to insert (in value form)
2898 STMTS is a statement list to append the necessary insertions into.
2899
2900 This function will die if we hit some value that shouldn't be
2901 ANTIC but is (IE there is no leader for it, or its components).
2902 This function may also generate expressions that are themselves
2903 partially or fully redundant. Those that are will be either made
2904 fully redundant during the next iteration of insert (for partially
2905 redundant ones), or eliminated by eliminate (for fully redundant
2906 ones).
2907
2908 If DOMSTMT is non-NULL then we make sure that all uses in the
2909 expressions dominate that statement. In this case the function
2910 can return NULL_TREE to signal failure. */
2911
2912 static tree
2913 create_expression_by_pieces (basic_block block, pre_expr expr,
2914 gimple_seq *stmts, gimple domstmt, tree type)
2915 {
2916 tree temp, name;
2917 tree folded;
2918 gimple_seq forced_stmts = NULL;
2919 unsigned int value_id;
2920 gimple_stmt_iterator gsi;
2921 tree exprtype = type ? type : get_expr_type (expr);
2922 pre_expr nameexpr;
2923 gimple newstmt;
2924
2925 switch (expr->kind)
2926 {
2927 /* We may hit the NAME/CONSTANT case if we have to convert types
2928 that value numbering saw through. */
2929 case NAME:
2930 folded = PRE_EXPR_NAME (expr);
2931 break;
2932 case CONSTANT:
2933 folded = PRE_EXPR_CONSTANT (expr);
2934 break;
2935 case REFERENCE:
2936 {
2937 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2938 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2939 }
2940 break;
2941 case NARY:
2942 {
2943 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2944 switch (nary->length)
2945 {
2946 case 2:
2947 {
2948 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2949 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2950 tree genop1 = find_or_generate_expression (block, op1,
2951 stmts, domstmt);
2952 tree genop2 = find_or_generate_expression (block, op2,
2953 stmts, domstmt);
2954 if (!genop1 || !genop2)
2955 return NULL_TREE;
2956 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2957 genop1);
2958 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2959 may be a constant with the wrong type. */
2960 if (nary->opcode == POINTER_PLUS_EXPR)
2961 genop2 = fold_convert (sizetype, genop2);
2962 else
2963 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2964
2965 folded = fold_build2 (nary->opcode, nary->type,
2966 genop1, genop2);
2967 }
2968 break;
2969 case 1:
2970 {
2971 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2972 tree genop1 = find_or_generate_expression (block, op1,
2973 stmts, domstmt);
2974 if (!genop1)
2975 return NULL_TREE;
2976 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2977
2978 folded = fold_build1 (nary->opcode, nary->type,
2979 genop1);
2980 }
2981 break;
2982 default:
2983 return NULL_TREE;
2984 }
2985 }
2986 break;
2987 default:
2988 return NULL_TREE;
2989 }
2990
2991 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2992 folded = fold_convert (exprtype, folded);
2993
2994 /* Force the generated expression to be a sequence of GIMPLE
2995 statements.
2996 We have to call unshare_expr because force_gimple_operand may
2997 modify the tree we pass to it. */
2998 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2999 false, NULL);
3000
3001 /* If we have any intermediate expressions to the value sets, add them
3002 to the value sets and chain them in the instruction stream. */
3003 if (forced_stmts)
3004 {
3005 gsi = gsi_start (forced_stmts);
3006 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3007 {
3008 gimple stmt = gsi_stmt (gsi);
3009 tree forcedname = gimple_get_lhs (stmt);
3010 pre_expr nameexpr;
3011
3012 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3013 if (TREE_CODE (forcedname) == SSA_NAME)
3014 {
3015 VN_INFO_GET (forcedname)->valnum = forcedname;
3016 VN_INFO (forcedname)->value_id = get_next_value_id ();
3017 nameexpr = get_or_alloc_expr_for_name (forcedname);
3018 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3019 if (!in_fre)
3020 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3021 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3022 }
3023 mark_symbols_for_renaming (stmt);
3024 }
3025 gimple_seq_add_seq (stmts, forced_stmts);
3026 }
3027
3028 /* Build and insert the assignment of the end result to the temporary
3029 that we will return. */
3030 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3031 {
3032 pretemp = create_tmp_var (exprtype, "pretmp");
3033 get_var_ann (pretemp);
3034 }
3035
3036 temp = pretemp;
3037 add_referenced_var (temp);
3038
3039 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3040 || TREE_CODE (exprtype) == VECTOR_TYPE)
3041 DECL_GIMPLE_REG_P (temp) = 1;
3042
3043 newstmt = gimple_build_assign (temp, folded);
3044 name = make_ssa_name (temp, newstmt);
3045 gimple_assign_set_lhs (newstmt, name);
3046 gimple_set_plf (newstmt, NECESSARY, false);
3047
3048 gimple_seq_add_stmt (stmts, newstmt);
3049 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3050
3051 /* All the symbols in NEWEXPR should be put into SSA form. */
3052 mark_symbols_for_renaming (newstmt);
3053
3054 /* Add a value number to the temporary.
3055 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3056 we are creating the expression by pieces, and this particular piece of
3057 the expression may have been represented. There is no harm in replacing
3058 here. */
3059 VN_INFO_GET (name)->valnum = name;
3060 value_id = get_expr_value_id (expr);
3061 VN_INFO (name)->value_id = value_id;
3062 nameexpr = get_or_alloc_expr_for_name (name);
3063 add_to_value (value_id, nameexpr);
3064 if (!in_fre)
3065 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3066 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3067
3068 pre_stats.insertions++;
3069 if (dump_file && (dump_flags & TDF_DETAILS))
3070 {
3071 fprintf (dump_file, "Inserted ");
3072 print_gimple_stmt (dump_file, newstmt, 0, 0);
3073 fprintf (dump_file, " in predecessor %d\n", block->index);
3074 }
3075
3076 return name;
3077 }
3078
3079
3080 /* Returns true if we want to inhibit the insertions of PHI nodes
3081 for the given EXPR for basic block BB (a member of a loop).
3082 We want to do this, when we fear that the induction variable we
3083 create might inhibit vectorization. */
3084
3085 static bool
3086 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3087 {
3088 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3089 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3090 vn_reference_op_t op;
3091 unsigned i;
3092
3093 /* If we aren't going to vectorize we don't inhibit anything. */
3094 if (!flag_tree_vectorize)
3095 return false;
3096
3097 /* Otherwise we inhibit the insertion when the address of the
3098 memory reference is a simple induction variable. In other
3099 cases the vectorizer won't do anything anyway (either it's
3100 loop invariant or a complicated expression). */
3101 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3102 {
3103 switch (op->opcode)
3104 {
3105 case ARRAY_REF:
3106 case ARRAY_RANGE_REF:
3107 if (TREE_CODE (op->op0) != SSA_NAME)
3108 break;
3109 /* Fallthru. */
3110 case SSA_NAME:
3111 {
3112 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3113 affine_iv iv;
3114 /* Default defs are loop invariant. */
3115 if (!defbb)
3116 break;
3117 /* Defined outside this loop, also loop invariant. */
3118 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3119 break;
3120 /* If it's a simple induction variable inhibit insertion,
3121 the vectorizer might be interested in this one. */
3122 if (simple_iv (bb->loop_father, bb->loop_father,
3123 op->op0, &iv, true))
3124 return true;
3125 /* No simple IV, vectorizer can't do anything, hence no
3126 reason to inhibit the transformation for this operand. */
3127 break;
3128 }
3129 default:
3130 break;
3131 }
3132 }
3133 return false;
3134 }
3135
3136 /* Insert the to-be-made-available values of expression EXPRNUM for each
3137 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3138 merge the result with a phi node, given the same value number as
3139 NODE. Return true if we have inserted new stuff. */
3140
3141 static bool
3142 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3143 pre_expr *avail)
3144 {
3145 pre_expr expr = expression_for_id (exprnum);
3146 pre_expr newphi;
3147 unsigned int val = get_expr_value_id (expr);
3148 edge pred;
3149 bool insertions = false;
3150 bool nophi = false;
3151 basic_block bprime;
3152 pre_expr eprime;
3153 edge_iterator ei;
3154 tree type = get_expr_type (expr);
3155 tree temp;
3156 gimple phi;
3157
3158 if (dump_file && (dump_flags & TDF_DETAILS))
3159 {
3160 fprintf (dump_file, "Found partial redundancy for expression ");
3161 print_pre_expr (dump_file, expr);
3162 fprintf (dump_file, " (%04d)\n", val);
3163 }
3164
3165 /* Make sure we aren't creating an induction variable. */
3166 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3167 {
3168 bool firstinsideloop = false;
3169 bool secondinsideloop = false;
3170 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3171 EDGE_PRED (block, 0)->src);
3172 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3173 EDGE_PRED (block, 1)->src);
3174 /* Induction variables only have one edge inside the loop. */
3175 if ((firstinsideloop ^ secondinsideloop)
3176 && (expr->kind != REFERENCE
3177 || inhibit_phi_insertion (block, expr)))
3178 {
3179 if (dump_file && (dump_flags & TDF_DETAILS))
3180 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3181 nophi = true;
3182 }
3183 }
3184
3185 /* Make sure we are not inserting trapping expressions. */
3186 FOR_EACH_EDGE (pred, ei, block->preds)
3187 {
3188 bprime = pred->src;
3189 eprime = avail[bprime->index];
3190 if (eprime->kind == NARY
3191 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3192 return false;
3193 }
3194
3195 /* Make the necessary insertions. */
3196 FOR_EACH_EDGE (pred, ei, block->preds)
3197 {
3198 gimple_seq stmts = NULL;
3199 tree builtexpr;
3200 bprime = pred->src;
3201 eprime = avail[bprime->index];
3202
3203 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3204 {
3205 builtexpr = create_expression_by_pieces (bprime,
3206 eprime,
3207 &stmts, NULL,
3208 type);
3209 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3210 gsi_insert_seq_on_edge (pred, stmts);
3211 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3212 insertions = true;
3213 }
3214 else if (eprime->kind == CONSTANT)
3215 {
3216 /* Constants may not have the right type, fold_convert
3217 should give us back a constant with the right type.
3218 */
3219 tree constant = PRE_EXPR_CONSTANT (eprime);
3220 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3221 {
3222 tree builtexpr = fold_convert (type, constant);
3223 if (!is_gimple_min_invariant (builtexpr))
3224 {
3225 tree forcedexpr = force_gimple_operand (builtexpr,
3226 &stmts, true,
3227 NULL);
3228 if (!is_gimple_min_invariant (forcedexpr))
3229 {
3230 if (forcedexpr != builtexpr)
3231 {
3232 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3233 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3234 }
3235 if (stmts)
3236 {
3237 gimple_stmt_iterator gsi;
3238 gsi = gsi_start (stmts);
3239 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3240 {
3241 gimple stmt = gsi_stmt (gsi);
3242 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3243 gimple_set_plf (stmt, NECESSARY, false);
3244 }
3245 gsi_insert_seq_on_edge (pred, stmts);
3246 }
3247 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3248 }
3249 }
3250 }
3251 }
3252 else if (eprime->kind == NAME)
3253 {
3254 /* We may have to do a conversion because our value
3255 numbering can look through types in certain cases, but
3256 our IL requires all operands of a phi node have the same
3257 type. */
3258 tree name = PRE_EXPR_NAME (eprime);
3259 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3260 {
3261 tree builtexpr;
3262 tree forcedexpr;
3263 builtexpr = fold_convert (type, name);
3264 forcedexpr = force_gimple_operand (builtexpr,
3265 &stmts, true,
3266 NULL);
3267
3268 if (forcedexpr != name)
3269 {
3270 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3271 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3272 }
3273
3274 if (stmts)
3275 {
3276 gimple_stmt_iterator gsi;
3277 gsi = gsi_start (stmts);
3278 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3279 {
3280 gimple stmt = gsi_stmt (gsi);
3281 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3282 gimple_set_plf (stmt, NECESSARY, false);
3283 }
3284 gsi_insert_seq_on_edge (pred, stmts);
3285 }
3286 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3287 }
3288 }
3289 }
3290 /* If we didn't want a phi node, and we made insertions, we still have
3291 inserted new stuff, and thus return true. If we didn't want a phi node,
3292 and didn't make insertions, we haven't added anything new, so return
3293 false. */
3294 if (nophi && insertions)
3295 return true;
3296 else if (nophi && !insertions)
3297 return false;
3298
3299 /* Now build a phi for the new variable. */
3300 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3301 {
3302 prephitemp = create_tmp_var (type, "prephitmp");
3303 get_var_ann (prephitemp);
3304 }
3305
3306 temp = prephitemp;
3307 add_referenced_var (temp);
3308
3309 if (TREE_CODE (type) == COMPLEX_TYPE
3310 || TREE_CODE (type) == VECTOR_TYPE)
3311 DECL_GIMPLE_REG_P (temp) = 1;
3312 phi = create_phi_node (temp, block);
3313
3314 gimple_set_plf (phi, NECESSARY, false);
3315 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3316 VN_INFO (gimple_phi_result (phi))->value_id = val;
3317 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3318 bitmap_set_bit (inserted_phi_names,
3319 SSA_NAME_VERSION (gimple_phi_result (phi)));
3320 FOR_EACH_EDGE (pred, ei, block->preds)
3321 {
3322 pre_expr ae = avail[pred->src->index];
3323 gcc_assert (get_expr_type (ae) == type
3324 || useless_type_conversion_p (type, get_expr_type (ae)));
3325 if (ae->kind == CONSTANT)
3326 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3327 else
3328 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3329 UNKNOWN_LOCATION);
3330 }
3331
3332 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3333 add_to_value (val, newphi);
3334
3335 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3336 this insertion, since we test for the existence of this value in PHI_GEN
3337 before proceeding with the partial redundancy checks in insert_aux.
3338
3339 The value may exist in AVAIL_OUT, in particular, it could be represented
3340 by the expression we are trying to eliminate, in which case we want the
3341 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3342 inserted there.
3343
3344 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3345 this block, because if it did, it would have existed in our dominator's
3346 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3347 */
3348
3349 bitmap_insert_into_set (PHI_GEN (block), newphi);
3350 bitmap_value_replace_in_set (AVAIL_OUT (block),
3351 newphi);
3352 bitmap_insert_into_set (NEW_SETS (block),
3353 newphi);
3354
3355 if (dump_file && (dump_flags & TDF_DETAILS))
3356 {
3357 fprintf (dump_file, "Created phi ");
3358 print_gimple_stmt (dump_file, phi, 0, 0);
3359 fprintf (dump_file, " in block %d\n", block->index);
3360 }
3361 pre_stats.phis++;
3362 return true;
3363 }
3364
3365
3366
3367 /* Perform insertion of partially redundant values.
3368 For BLOCK, do the following:
3369 1. Propagate the NEW_SETS of the dominator into the current block.
3370 If the block has multiple predecessors,
3371 2a. Iterate over the ANTIC expressions for the block to see if
3372 any of them are partially redundant.
3373 2b. If so, insert them into the necessary predecessors to make
3374 the expression fully redundant.
3375 2c. Insert a new PHI merging the values of the predecessors.
3376 2d. Insert the new PHI, and the new expressions, into the
3377 NEW_SETS set.
3378 3. Recursively call ourselves on the dominator children of BLOCK.
3379
3380 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3381 do_regular_insertion and do_partial_insertion.
3382
3383 */
3384
3385 static bool
3386 do_regular_insertion (basic_block block, basic_block dom)
3387 {
3388 bool new_stuff = false;
3389 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3390 pre_expr expr;
3391 int i;
3392
3393 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3394 {
3395 if (expr->kind != NAME)
3396 {
3397 pre_expr *avail;
3398 unsigned int val;
3399 bool by_some = false;
3400 bool cant_insert = false;
3401 bool all_same = true;
3402 pre_expr first_s = NULL;
3403 edge pred;
3404 basic_block bprime;
3405 pre_expr eprime = NULL;
3406 edge_iterator ei;
3407 pre_expr edoubleprime = NULL;
3408
3409 val = get_expr_value_id (expr);
3410 if (bitmap_set_contains_value (PHI_GEN (block), val))
3411 continue;
3412 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3413 {
3414 if (dump_file && (dump_flags & TDF_DETAILS))
3415 fprintf (dump_file, "Found fully redundant value\n");
3416 continue;
3417 }
3418
3419 avail = XCNEWVEC (pre_expr, last_basic_block);
3420 FOR_EACH_EDGE (pred, ei, block->preds)
3421 {
3422 unsigned int vprime;
3423
3424 /* We should never run insertion for the exit block
3425 and so not come across fake pred edges. */
3426 gcc_assert (!(pred->flags & EDGE_FAKE));
3427 bprime = pred->src;
3428 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3429 bprime, block);
3430
3431 /* eprime will generally only be NULL if the
3432 value of the expression, translated
3433 through the PHI for this predecessor, is
3434 undefined. If that is the case, we can't
3435 make the expression fully redundant,
3436 because its value is undefined along a
3437 predecessor path. We can thus break out
3438 early because it doesn't matter what the
3439 rest of the results are. */
3440 if (eprime == NULL)
3441 {
3442 cant_insert = true;
3443 break;
3444 }
3445
3446 eprime = fully_constant_expression (eprime);
3447 vprime = get_expr_value_id (eprime);
3448 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3449 vprime, NULL);
3450 if (edoubleprime == NULL)
3451 {
3452 avail[bprime->index] = eprime;
3453 all_same = false;
3454 }
3455 else
3456 {
3457 avail[bprime->index] = edoubleprime;
3458 by_some = true;
3459 if (first_s == NULL)
3460 first_s = edoubleprime;
3461 else if (!pre_expr_eq (first_s, edoubleprime))
3462 all_same = false;
3463 }
3464 }
3465 /* If we can insert it, it's not the same value
3466 already existing along every predecessor, and
3467 it's defined by some predecessor, it is
3468 partially redundant. */
3469 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3470 {
3471 if (insert_into_preds_of_block (block, get_expression_id (expr),
3472 avail))
3473 new_stuff = true;
3474 }
3475 /* If all edges produce the same value and that value is
3476 an invariant, then the PHI has the same value on all
3477 edges. Note this. */
3478 else if (!cant_insert && all_same && eprime
3479 && (edoubleprime->kind == CONSTANT
3480 || edoubleprime->kind == NAME)
3481 && !value_id_constant_p (val))
3482 {
3483 unsigned int j;
3484 bitmap_iterator bi;
3485 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3486 value_expressions, val);
3487
3488 unsigned int new_val = get_expr_value_id (edoubleprime);
3489 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3490 {
3491 pre_expr expr = expression_for_id (j);
3492
3493 if (expr->kind == NAME)
3494 {
3495 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3496 /* Just reset the value id and valnum so it is
3497 the same as the constant we have discovered. */
3498 if (edoubleprime->kind == CONSTANT)
3499 {
3500 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3501 pre_stats.constified++;
3502 }
3503 else
3504 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3505 info->value_id = new_val;
3506 }
3507 }
3508 }
3509 free (avail);
3510 }
3511 }
3512
3513 VEC_free (pre_expr, heap, exprs);
3514 return new_stuff;
3515 }
3516
3517
3518 /* Perform insertion for partially anticipatable expressions. There
3519 is only one case we will perform insertion for these. This case is
3520 if the expression is partially anticipatable, and fully available.
3521 In this case, we know that putting it earlier will enable us to
3522 remove the later computation. */
3523
3524
3525 static bool
3526 do_partial_partial_insertion (basic_block block, basic_block dom)
3527 {
3528 bool new_stuff = false;
3529 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3530 pre_expr expr;
3531 int i;
3532
3533 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3534 {
3535 if (expr->kind != NAME)
3536 {
3537 pre_expr *avail;
3538 unsigned int val;
3539 bool by_all = true;
3540 bool cant_insert = false;
3541 edge pred;
3542 basic_block bprime;
3543 pre_expr eprime = NULL;
3544 edge_iterator ei;
3545
3546 val = get_expr_value_id (expr);
3547 if (bitmap_set_contains_value (PHI_GEN (block), val))
3548 continue;
3549 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3550 continue;
3551
3552 avail = XCNEWVEC (pre_expr, last_basic_block);
3553 FOR_EACH_EDGE (pred, ei, block->preds)
3554 {
3555 unsigned int vprime;
3556 pre_expr edoubleprime;
3557
3558 /* We should never run insertion for the exit block
3559 and so not come across fake pred edges. */
3560 gcc_assert (!(pred->flags & EDGE_FAKE));
3561 bprime = pred->src;
3562 eprime = phi_translate (expr, ANTIC_IN (block),
3563 PA_IN (block),
3564 bprime, block);
3565
3566 /* eprime will generally only be NULL if the
3567 value of the expression, translated
3568 through the PHI for this predecessor, is
3569 undefined. If that is the case, we can't
3570 make the expression fully redundant,
3571 because its value is undefined along a
3572 predecessor path. We can thus break out
3573 early because it doesn't matter what the
3574 rest of the results are. */
3575 if (eprime == NULL)
3576 {
3577 cant_insert = true;
3578 break;
3579 }
3580
3581 eprime = fully_constant_expression (eprime);
3582 vprime = get_expr_value_id (eprime);
3583 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3584 vprime, NULL);
3585 if (edoubleprime == NULL)
3586 {
3587 by_all = false;
3588 break;
3589 }
3590 else
3591 avail[bprime->index] = edoubleprime;
3592
3593 }
3594
3595 /* If we can insert it, it's not the same value
3596 already existing along every predecessor, and
3597 it's defined by some predecessor, it is
3598 partially redundant. */
3599 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3600 {
3601 pre_stats.pa_insert++;
3602 if (insert_into_preds_of_block (block, get_expression_id (expr),
3603 avail))
3604 new_stuff = true;
3605 }
3606 free (avail);
3607 }
3608 }
3609
3610 VEC_free (pre_expr, heap, exprs);
3611 return new_stuff;
3612 }
3613
3614 static bool
3615 insert_aux (basic_block block)
3616 {
3617 basic_block son;
3618 bool new_stuff = false;
3619
3620 if (block)
3621 {
3622 basic_block dom;
3623 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3624 if (dom)
3625 {
3626 unsigned i;
3627 bitmap_iterator bi;
3628 bitmap_set_t newset = NEW_SETS (dom);
3629 if (newset)
3630 {
3631 /* Note that we need to value_replace both NEW_SETS, and
3632 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3633 represented by some non-simple expression here that we want
3634 to replace it with. */
3635 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3636 {
3637 pre_expr expr = expression_for_id (i);
3638 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3639 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3640 }
3641 }
3642 if (!single_pred_p (block))
3643 {
3644 new_stuff |= do_regular_insertion (block, dom);
3645 if (do_partial_partial)
3646 new_stuff |= do_partial_partial_insertion (block, dom);
3647 }
3648 }
3649 }
3650 for (son = first_dom_son (CDI_DOMINATORS, block);
3651 son;
3652 son = next_dom_son (CDI_DOMINATORS, son))
3653 {
3654 new_stuff |= insert_aux (son);
3655 }
3656
3657 return new_stuff;
3658 }
3659
3660 /* Perform insertion of partially redundant values. */
3661
3662 static void
3663 insert (void)
3664 {
3665 bool new_stuff = true;
3666 basic_block bb;
3667 int num_iterations = 0;
3668
3669 FOR_ALL_BB (bb)
3670 NEW_SETS (bb) = bitmap_set_new ();
3671
3672 while (new_stuff)
3673 {
3674 num_iterations++;
3675 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3676 }
3677 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3678 }
3679
3680
3681 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3682
3683 static void
3684 add_to_exp_gen (basic_block block, tree op)
3685 {
3686 if (!in_fre)
3687 {
3688 pre_expr result;
3689 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3690 return;
3691 result = get_or_alloc_expr_for_name (op);
3692 bitmap_value_insert_into_set (EXP_GEN (block), result);
3693 }
3694 }
3695
3696 /* Create value ids for PHI in BLOCK. */
3697
3698 static void
3699 make_values_for_phi (gimple phi, basic_block block)
3700 {
3701 tree result = gimple_phi_result (phi);
3702
3703 /* We have no need for virtual phis, as they don't represent
3704 actual computations. */
3705 if (is_gimple_reg (result))
3706 {
3707 pre_expr e = get_or_alloc_expr_for_name (result);
3708 add_to_value (get_expr_value_id (e), e);
3709 bitmap_insert_into_set (PHI_GEN (block), e);
3710 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3711 if (!in_fre)
3712 {
3713 unsigned i;
3714 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3715 {
3716 tree arg = gimple_phi_arg_def (phi, i);
3717 if (TREE_CODE (arg) == SSA_NAME)
3718 {
3719 e = get_or_alloc_expr_for_name (arg);
3720 add_to_value (get_expr_value_id (e), e);
3721 }
3722 }
3723 }
3724 }
3725 }
3726
3727 /* Compute the AVAIL set for all basic blocks.
3728
3729 This function performs value numbering of the statements in each basic
3730 block. The AVAIL sets are built from information we glean while doing
3731 this value numbering, since the AVAIL sets contain only one entry per
3732 value.
3733
3734 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3735 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3736
3737 static void
3738 compute_avail (void)
3739 {
3740
3741 basic_block block, son;
3742 basic_block *worklist;
3743 size_t sp = 0;
3744 unsigned i;
3745
3746 /* We pretend that default definitions are defined in the entry block.
3747 This includes function arguments and the static chain decl. */
3748 for (i = 1; i < num_ssa_names; ++i)
3749 {
3750 tree name = ssa_name (i);
3751 pre_expr e;
3752 if (!name
3753 || !SSA_NAME_IS_DEFAULT_DEF (name)
3754 || has_zero_uses (name)
3755 || !is_gimple_reg (name))
3756 continue;
3757
3758 e = get_or_alloc_expr_for_name (name);
3759 add_to_value (get_expr_value_id (e), e);
3760 if (!in_fre)
3761 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3762 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3763 }
3764
3765 /* Allocate the worklist. */
3766 worklist = XNEWVEC (basic_block, n_basic_blocks);
3767
3768 /* Seed the algorithm by putting the dominator children of the entry
3769 block on the worklist. */
3770 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3771 son;
3772 son = next_dom_son (CDI_DOMINATORS, son))
3773 worklist[sp++] = son;
3774
3775 /* Loop until the worklist is empty. */
3776 while (sp)
3777 {
3778 gimple_stmt_iterator gsi;
3779 gimple stmt;
3780 basic_block dom;
3781 unsigned int stmt_uid = 1;
3782
3783 /* Pick a block from the worklist. */
3784 block = worklist[--sp];
3785
3786 /* Initially, the set of available values in BLOCK is that of
3787 its immediate dominator. */
3788 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3789 if (dom)
3790 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3791
3792 /* Generate values for PHI nodes. */
3793 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3794 make_values_for_phi (gsi_stmt (gsi), block);
3795
3796 /* Now compute value numbers and populate value sets with all
3797 the expressions computed in BLOCK. */
3798 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3799 {
3800 ssa_op_iter iter;
3801 tree op;
3802
3803 stmt = gsi_stmt (gsi);
3804 gimple_set_uid (stmt, stmt_uid++);
3805
3806 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3807 {
3808 pre_expr e = get_or_alloc_expr_for_name (op);
3809
3810 add_to_value (get_expr_value_id (e), e);
3811 if (!in_fre)
3812 bitmap_insert_into_set (TMP_GEN (block), e);
3813 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3814 }
3815
3816 if (gimple_has_volatile_ops (stmt)
3817 || stmt_could_throw_p (stmt))
3818 continue;
3819
3820 switch (gimple_code (stmt))
3821 {
3822 case GIMPLE_RETURN:
3823 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3824 add_to_exp_gen (block, op);
3825 continue;
3826
3827 case GIMPLE_CALL:
3828 {
3829 vn_reference_t ref;
3830 unsigned int i;
3831 vn_reference_op_t vro;
3832 pre_expr result = NULL;
3833 VEC(vn_reference_op_s, heap) *ops = NULL;
3834
3835 if (!can_value_number_call (stmt))
3836 continue;
3837
3838 copy_reference_ops_from_call (stmt, &ops);
3839 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3840 gimple_expr_type (stmt),
3841 ops, &ref, false);
3842 VEC_free (vn_reference_op_s, heap, ops);
3843 if (!ref)
3844 continue;
3845
3846 for (i = 0; VEC_iterate (vn_reference_op_s,
3847 ref->operands, i,
3848 vro); i++)
3849 {
3850 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3851 add_to_exp_gen (block, vro->op0);
3852 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3853 add_to_exp_gen (block, vro->op1);
3854 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3855 add_to_exp_gen (block, vro->op2);
3856 }
3857 result = (pre_expr) pool_alloc (pre_expr_pool);
3858 result->kind = REFERENCE;
3859 result->id = 0;
3860 PRE_EXPR_REFERENCE (result) = ref;
3861
3862 get_or_alloc_expression_id (result);
3863 add_to_value (get_expr_value_id (result), result);
3864 if (!in_fre)
3865 bitmap_value_insert_into_set (EXP_GEN (block), result);
3866 continue;
3867 }
3868
3869 case GIMPLE_ASSIGN:
3870 {
3871 pre_expr result = NULL;
3872 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3873 {
3874 case tcc_unary:
3875 case tcc_binary:
3876 case tcc_comparison:
3877 {
3878 vn_nary_op_t nary;
3879 unsigned int i;
3880
3881 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3882 gimple_assign_rhs_code (stmt),
3883 gimple_expr_type (stmt),
3884 gimple_assign_rhs1 (stmt),
3885 gimple_assign_rhs2 (stmt),
3886 NULL_TREE, NULL_TREE, &nary);
3887
3888 if (!nary)
3889 continue;
3890
3891 for (i = 0; i < nary->length; i++)
3892 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3893 add_to_exp_gen (block, nary->op[i]);
3894
3895 result = (pre_expr) pool_alloc (pre_expr_pool);
3896 result->kind = NARY;
3897 result->id = 0;
3898 PRE_EXPR_NARY (result) = nary;
3899 break;
3900 }
3901
3902 case tcc_declaration:
3903 case tcc_reference:
3904 {
3905 vn_reference_t ref;
3906 unsigned int i;
3907 vn_reference_op_t vro;
3908
3909 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3910 gimple_vuse (stmt),
3911 false, &ref);
3912 if (!ref)
3913 continue;
3914
3915 for (i = 0; VEC_iterate (vn_reference_op_s,
3916 ref->operands, i,
3917 vro); i++)
3918 {
3919 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3920 add_to_exp_gen (block, vro->op0);
3921 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3922 add_to_exp_gen (block, vro->op1);
3923 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3924 add_to_exp_gen (block, vro->op2);
3925 }
3926 result = (pre_expr) pool_alloc (pre_expr_pool);
3927 result->kind = REFERENCE;
3928 result->id = 0;
3929 PRE_EXPR_REFERENCE (result) = ref;
3930 break;
3931 }
3932
3933 default:
3934 /* For any other statement that we don't
3935 recognize, simply add all referenced
3936 SSA_NAMEs to EXP_GEN. */
3937 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3938 add_to_exp_gen (block, op);
3939 continue;
3940 }
3941
3942 get_or_alloc_expression_id (result);
3943 add_to_value (get_expr_value_id (result), result);
3944 if (!in_fre)
3945 bitmap_value_insert_into_set (EXP_GEN (block), result);
3946
3947 continue;
3948 }
3949 default:
3950 break;
3951 }
3952 }
3953
3954 /* Put the dominator children of BLOCK on the worklist of blocks
3955 to compute available sets for. */
3956 for (son = first_dom_son (CDI_DOMINATORS, block);
3957 son;
3958 son = next_dom_son (CDI_DOMINATORS, son))
3959 worklist[sp++] = son;
3960 }
3961
3962 free (worklist);
3963 }
3964
3965 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3966 than the available expressions for it. The insertion point is
3967 right before the first use in STMT. Returns the SSA_NAME that should
3968 be used for replacement. */
3969
3970 static tree
3971 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3972 {
3973 basic_block bb = gimple_bb (stmt);
3974 gimple_stmt_iterator gsi;
3975 gimple_seq stmts = NULL;
3976 tree expr;
3977 pre_expr e;
3978
3979 /* First create a value expression from the expression we want
3980 to insert and associate it with the value handle for SSA_VN. */
3981 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3982 if (e == NULL)
3983 return NULL_TREE;
3984
3985 /* Then use create_expression_by_pieces to generate a valid
3986 expression to insert at this point of the IL stream. */
3987 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3988 if (expr == NULL_TREE)
3989 return NULL_TREE;
3990 gsi = gsi_for_stmt (stmt);
3991 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3992
3993 return expr;
3994 }
3995
3996 /* Eliminate fully redundant computations. */
3997
3998 static unsigned int
3999 eliminate (void)
4000 {
4001 VEC (gimple, heap) *to_remove = NULL;
4002 basic_block b;
4003 unsigned int todo = 0;
4004 gimple_stmt_iterator gsi;
4005 gimple stmt;
4006 unsigned i;
4007
4008 FOR_EACH_BB (b)
4009 {
4010 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4011 {
4012 stmt = gsi_stmt (gsi);
4013
4014 /* Lookup the RHS of the expression, see if we have an
4015 available computation for it. If so, replace the RHS with
4016 the available computation. */
4017 if (gimple_has_lhs (stmt)
4018 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4019 && !gimple_assign_ssa_name_copy_p (stmt)
4020 && (!gimple_assign_single_p (stmt)
4021 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4022 && !gimple_has_volatile_ops (stmt)
4023 && !has_zero_uses (gimple_get_lhs (stmt)))
4024 {
4025 tree lhs = gimple_get_lhs (stmt);
4026 tree rhs = NULL_TREE;
4027 tree sprime = NULL;
4028 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4029 pre_expr sprimeexpr;
4030
4031 if (gimple_assign_single_p (stmt))
4032 rhs = gimple_assign_rhs1 (stmt);
4033
4034 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4035 get_expr_value_id (lhsexpr),
4036 NULL);
4037
4038 if (sprimeexpr)
4039 {
4040 if (sprimeexpr->kind == CONSTANT)
4041 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4042 else if (sprimeexpr->kind == NAME)
4043 sprime = PRE_EXPR_NAME (sprimeexpr);
4044 else
4045 gcc_unreachable ();
4046 }
4047
4048 /* If there is no existing leader but SCCVN knows this
4049 value is constant, use that constant. */
4050 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4051 {
4052 sprime = VN_INFO (lhs)->valnum;
4053 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4054 TREE_TYPE (sprime)))
4055 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4056
4057 if (dump_file && (dump_flags & TDF_DETAILS))
4058 {
4059 fprintf (dump_file, "Replaced ");
4060 print_gimple_expr (dump_file, stmt, 0, 0);
4061 fprintf (dump_file, " with ");
4062 print_generic_expr (dump_file, sprime, 0);
4063 fprintf (dump_file, " in ");
4064 print_gimple_stmt (dump_file, stmt, 0, 0);
4065 }
4066 pre_stats.eliminations++;
4067 propagate_tree_value_into_stmt (&gsi, sprime);
4068 stmt = gsi_stmt (gsi);
4069 update_stmt (stmt);
4070 continue;
4071 }
4072
4073 /* If there is no existing usable leader but SCCVN thinks
4074 it has an expression it wants to use as replacement,
4075 insert that. */
4076 if (!sprime || sprime == lhs)
4077 {
4078 tree val = VN_INFO (lhs)->valnum;
4079 if (val != VN_TOP
4080 && TREE_CODE (val) == SSA_NAME
4081 && VN_INFO (val)->needs_insertion
4082 && can_PRE_operation (vn_get_expr_for (val)))
4083 sprime = do_SCCVN_insertion (stmt, val);
4084 }
4085 if (sprime
4086 && sprime != lhs
4087 && (rhs == NULL_TREE
4088 || TREE_CODE (rhs) != SSA_NAME
4089 || may_propagate_copy (rhs, sprime)))
4090 {
4091 gcc_assert (sprime != rhs);
4092
4093 if (dump_file && (dump_flags & TDF_DETAILS))
4094 {
4095 fprintf (dump_file, "Replaced ");
4096 print_gimple_expr (dump_file, stmt, 0, 0);
4097 fprintf (dump_file, " with ");
4098 print_generic_expr (dump_file, sprime, 0);
4099 fprintf (dump_file, " in ");
4100 print_gimple_stmt (dump_file, stmt, 0, 0);
4101 }
4102
4103 if (TREE_CODE (sprime) == SSA_NAME)
4104 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4105 NECESSARY, true);
4106 /* We need to make sure the new and old types actually match,
4107 which may require adding a simple cast, which fold_convert
4108 will do for us. */
4109 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4110 && !useless_type_conversion_p (gimple_expr_type (stmt),
4111 TREE_TYPE (sprime)))
4112 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4113
4114 pre_stats.eliminations++;
4115 propagate_tree_value_into_stmt (&gsi, sprime);
4116 stmt = gsi_stmt (gsi);
4117 update_stmt (stmt);
4118
4119 /* If we removed EH side effects from the statement, clean
4120 its EH information. */
4121 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4122 {
4123 bitmap_set_bit (need_eh_cleanup,
4124 gimple_bb (stmt)->index);
4125 if (dump_file && (dump_flags & TDF_DETAILS))
4126 fprintf (dump_file, " Removed EH side effects.\n");
4127 }
4128 }
4129 }
4130 /* If the statement is a scalar store, see if the expression
4131 has the same value number as its rhs. If so, the store is
4132 dead. */
4133 else if (gimple_assign_single_p (stmt)
4134 && !is_gimple_reg (gimple_assign_lhs (stmt))
4135 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4136 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4137 {
4138 tree rhs = gimple_assign_rhs1 (stmt);
4139 tree val;
4140 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4141 gimple_vuse (stmt), true, NULL);
4142 if (TREE_CODE (rhs) == SSA_NAME)
4143 rhs = VN_INFO (rhs)->valnum;
4144 if (val
4145 && operand_equal_p (val, rhs, 0))
4146 {
4147 if (dump_file && (dump_flags & TDF_DETAILS))
4148 {
4149 fprintf (dump_file, "Deleted redundant store ");
4150 print_gimple_stmt (dump_file, stmt, 0, 0);
4151 }
4152
4153 /* Queue stmt for removal. */
4154 VEC_safe_push (gimple, heap, to_remove, stmt);
4155 }
4156 }
4157 /* Visit COND_EXPRs and fold the comparison with the
4158 available value-numbers. */
4159 else if (gimple_code (stmt) == GIMPLE_COND)
4160 {
4161 tree op0 = gimple_cond_lhs (stmt);
4162 tree op1 = gimple_cond_rhs (stmt);
4163 tree result;
4164
4165 if (TREE_CODE (op0) == SSA_NAME)
4166 op0 = VN_INFO (op0)->valnum;
4167 if (TREE_CODE (op1) == SSA_NAME)
4168 op1 = VN_INFO (op1)->valnum;
4169 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4170 op0, op1);
4171 if (result && TREE_CODE (result) == INTEGER_CST)
4172 {
4173 if (integer_zerop (result))
4174 gimple_cond_make_false (stmt);
4175 else
4176 gimple_cond_make_true (stmt);
4177 update_stmt (stmt);
4178 todo = TODO_cleanup_cfg;
4179 }
4180 }
4181 /* Visit indirect calls and turn them into direct calls if
4182 possible. */
4183 if (gimple_code (stmt) == GIMPLE_CALL
4184 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4185 {
4186 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4187 if (TREE_CODE (fn) == ADDR_EXPR
4188 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4189 {
4190 if (dump_file && (dump_flags & TDF_DETAILS))
4191 {
4192 fprintf (dump_file, "Replacing call target with ");
4193 print_generic_expr (dump_file, fn, 0);
4194 fprintf (dump_file, " in ");
4195 print_gimple_stmt (dump_file, stmt, 0, 0);
4196 }
4197
4198 gimple_call_set_fn (stmt, fn);
4199 update_stmt (stmt);
4200 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4201 {
4202 bitmap_set_bit (need_eh_cleanup,
4203 gimple_bb (stmt)->index);
4204 if (dump_file && (dump_flags & TDF_DETAILS))
4205 fprintf (dump_file, " Removed EH side effects.\n");
4206 }
4207
4208 /* Changing an indirect call to a direct call may
4209 have exposed different semantics. This may
4210 require an SSA update. */
4211 todo |= TODO_update_ssa_only_virtuals;
4212 }
4213 }
4214 }
4215
4216 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4217 {
4218 gimple stmt, phi = gsi_stmt (gsi);
4219 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4220 pre_expr sprimeexpr, resexpr;
4221 gimple_stmt_iterator gsi2;
4222
4223 /* We want to perform redundant PHI elimination. Do so by
4224 replacing the PHI with a single copy if possible.
4225 Do not touch inserted, single-argument or virtual PHIs. */
4226 if (gimple_phi_num_args (phi) == 1
4227 || !is_gimple_reg (res)
4228 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4229 {
4230 gsi_next (&gsi);
4231 continue;
4232 }
4233
4234 resexpr = get_or_alloc_expr_for_name (res);
4235 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4236 get_expr_value_id (resexpr), NULL);
4237 if (sprimeexpr)
4238 {
4239 if (sprimeexpr->kind == CONSTANT)
4240 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4241 else if (sprimeexpr->kind == NAME)
4242 sprime = PRE_EXPR_NAME (sprimeexpr);
4243 else
4244 gcc_unreachable ();
4245 }
4246 if (!sprimeexpr
4247 || sprime == res)
4248 {
4249 gsi_next (&gsi);
4250 continue;
4251 }
4252
4253 if (dump_file && (dump_flags & TDF_DETAILS))
4254 {
4255 fprintf (dump_file, "Replaced redundant PHI node defining ");
4256 print_generic_expr (dump_file, res, 0);
4257 fprintf (dump_file, " with ");
4258 print_generic_expr (dump_file, sprime, 0);
4259 fprintf (dump_file, "\n");
4260 }
4261
4262 remove_phi_node (&gsi, false);
4263
4264 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4265 sprime = fold_convert (TREE_TYPE (res), sprime);
4266 stmt = gimple_build_assign (res, sprime);
4267 SSA_NAME_DEF_STMT (res) = stmt;
4268 if (TREE_CODE (sprime) == SSA_NAME)
4269 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4270 NECESSARY, true);
4271 gsi2 = gsi_after_labels (b);
4272 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4273 /* Queue the copy for eventual removal. */
4274 VEC_safe_push (gimple, heap, to_remove, stmt);
4275 pre_stats.eliminations++;
4276 }
4277 }
4278
4279 /* We cannot remove stmts during BB walk, especially not release SSA
4280 names there as this confuses the VN machinery. The stmts ending
4281 up in to_remove are either stores or simple copies. */
4282 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4283 {
4284 tree lhs = gimple_assign_lhs (stmt);
4285 use_operand_p use_p;
4286 gimple use_stmt;
4287
4288 /* If there is a single use only, propagate the equivalency
4289 instead of keeping the copy. */
4290 if (TREE_CODE (lhs) == SSA_NAME
4291 && single_imm_use (lhs, &use_p, &use_stmt)
4292 && may_propagate_copy (USE_FROM_PTR (use_p),
4293 gimple_assign_rhs1 (stmt)))
4294 {
4295 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4296 update_stmt (use_stmt);
4297 }
4298
4299 /* If this is a store or a now unused copy, remove it. */
4300 if (TREE_CODE (lhs) != SSA_NAME
4301 || has_zero_uses (lhs))
4302 {
4303 gsi = gsi_for_stmt (stmt);
4304 unlink_stmt_vdef (stmt);
4305 gsi_remove (&gsi, true);
4306 release_defs (stmt);
4307 }
4308 }
4309 VEC_free (gimple, heap, to_remove);
4310
4311 return todo;
4312 }
4313
4314 /* Borrow a bit of tree-ssa-dce.c for the moment.
4315 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4316 this may be a bit faster, and we may want critical edges kept split. */
4317
4318 /* If OP's defining statement has not already been determined to be necessary,
4319 mark that statement necessary. Return the stmt, if it is newly
4320 necessary. */
4321
4322 static inline gimple
4323 mark_operand_necessary (tree op)
4324 {
4325 gimple stmt;
4326
4327 gcc_assert (op);
4328
4329 if (TREE_CODE (op) != SSA_NAME)
4330 return NULL;
4331
4332 stmt = SSA_NAME_DEF_STMT (op);
4333 gcc_assert (stmt);
4334
4335 if (gimple_plf (stmt, NECESSARY)
4336 || gimple_nop_p (stmt))
4337 return NULL;
4338
4339 gimple_set_plf (stmt, NECESSARY, true);
4340 return stmt;
4341 }
4342
4343 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4344 to insert PHI nodes sometimes, and because value numbering of casts isn't
4345 perfect, we sometimes end up inserting dead code. This simple DCE-like
4346 pass removes any insertions we made that weren't actually used. */
4347
4348 static void
4349 remove_dead_inserted_code (void)
4350 {
4351 VEC(gimple,heap) *worklist = NULL;
4352 int i;
4353 gimple t;
4354
4355 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4356 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4357 {
4358 if (gimple_plf (t, NECESSARY))
4359 VEC_quick_push (gimple, worklist, t);
4360 }
4361 while (VEC_length (gimple, worklist) > 0)
4362 {
4363 t = VEC_pop (gimple, worklist);
4364
4365 /* PHI nodes are somewhat special in that each PHI alternative has
4366 data and control dependencies. All the statements feeding the
4367 PHI node's arguments are always necessary. */
4368 if (gimple_code (t) == GIMPLE_PHI)
4369 {
4370 unsigned k;
4371
4372 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4373 for (k = 0; k < gimple_phi_num_args (t); k++)
4374 {
4375 tree arg = PHI_ARG_DEF (t, k);
4376 if (TREE_CODE (arg) == SSA_NAME)
4377 {
4378 gimple n = mark_operand_necessary (arg);
4379 if (n)
4380 VEC_quick_push (gimple, worklist, n);
4381 }
4382 }
4383 }
4384 else
4385 {
4386 /* Propagate through the operands. Examine all the USE, VUSE and
4387 VDEF operands in this statement. Mark all the statements
4388 which feed this statement's uses as necessary. */
4389 ssa_op_iter iter;
4390 tree use;
4391
4392 /* The operands of VDEF expressions are also needed as they
4393 represent potential definitions that may reach this
4394 statement (VDEF operands allow us to follow def-def
4395 links). */
4396
4397 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4398 {
4399 gimple n = mark_operand_necessary (use);
4400 if (n)
4401 VEC_safe_push (gimple, heap, worklist, n);
4402 }
4403 }
4404 }
4405
4406 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4407 {
4408 if (!gimple_plf (t, NECESSARY))
4409 {
4410 gimple_stmt_iterator gsi;
4411
4412 if (dump_file && (dump_flags & TDF_DETAILS))
4413 {
4414 fprintf (dump_file, "Removing unnecessary insertion:");
4415 print_gimple_stmt (dump_file, t, 0, 0);
4416 }
4417
4418 gsi = gsi_for_stmt (t);
4419 if (gimple_code (t) == GIMPLE_PHI)
4420 remove_phi_node (&gsi, true);
4421 else
4422 gsi_remove (&gsi, true);
4423 release_defs (t);
4424 }
4425 }
4426 VEC_free (gimple, heap, worklist);
4427 }
4428
4429 /* Initialize data structures used by PRE. */
4430
4431 static void
4432 init_pre (bool do_fre)
4433 {
4434 basic_block bb;
4435
4436 next_expression_id = 1;
4437 expressions = NULL;
4438 VEC_safe_push (pre_expr, heap, expressions, NULL);
4439 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4440 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4441 get_max_value_id() + 1);
4442
4443 in_fre = do_fre;
4444
4445 inserted_exprs = NULL;
4446 need_creation = NULL;
4447 pretemp = NULL_TREE;
4448 storetemp = NULL_TREE;
4449 prephitemp = NULL_TREE;
4450
4451 connect_infinite_loops_to_exit ();
4452 memset (&pre_stats, 0, sizeof (pre_stats));
4453
4454
4455 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4456 post_order_compute (postorder, false, false);
4457
4458 FOR_ALL_BB (bb)
4459 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4460
4461 calculate_dominance_info (CDI_POST_DOMINATORS);
4462 calculate_dominance_info (CDI_DOMINATORS);
4463
4464 bitmap_obstack_initialize (&grand_bitmap_obstack);
4465 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4466 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4467 expr_pred_trans_eq, free);
4468 expression_to_id = htab_create (num_ssa_names * 3,
4469 pre_expr_hash,
4470 pre_expr_eq, NULL);
4471 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4472 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4473 sizeof (struct bitmap_set), 30);
4474 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4475 sizeof (struct pre_expr_d), 30);
4476 FOR_ALL_BB (bb)
4477 {
4478 EXP_GEN (bb) = bitmap_set_new ();
4479 PHI_GEN (bb) = bitmap_set_new ();
4480 TMP_GEN (bb) = bitmap_set_new ();
4481 AVAIL_OUT (bb) = bitmap_set_new ();
4482 }
4483
4484 need_eh_cleanup = BITMAP_ALLOC (NULL);
4485 }
4486
4487
4488 /* Deallocate data structures used by PRE. */
4489
4490 static void
4491 fini_pre (bool do_fre)
4492 {
4493 basic_block bb;
4494
4495 free (postorder);
4496 VEC_free (bitmap_set_t, heap, value_expressions);
4497 VEC_free (gimple, heap, inserted_exprs);
4498 VEC_free (gimple, heap, need_creation);
4499 bitmap_obstack_release (&grand_bitmap_obstack);
4500 free_alloc_pool (bitmap_set_pool);
4501 free_alloc_pool (pre_expr_pool);
4502 htab_delete (phi_translate_table);
4503 htab_delete (expression_to_id);
4504
4505 FOR_ALL_BB (bb)
4506 {
4507 free (bb->aux);
4508 bb->aux = NULL;
4509 }
4510
4511 free_dominance_info (CDI_POST_DOMINATORS);
4512
4513 if (!bitmap_empty_p (need_eh_cleanup))
4514 {
4515 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4516 cleanup_tree_cfg ();
4517 }
4518
4519 BITMAP_FREE (need_eh_cleanup);
4520
4521 if (!do_fre)
4522 loop_optimizer_finalize ();
4523 }
4524
4525 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4526 only wants to do full redundancy elimination. */
4527
4528 static unsigned int
4529 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4530 {
4531 unsigned int todo = 0;
4532
4533 do_partial_partial = optimize > 2;
4534
4535 /* This has to happen before SCCVN runs because
4536 loop_optimizer_init may create new phis, etc. */
4537 if (!do_fre)
4538 loop_optimizer_init (LOOPS_NORMAL);
4539
4540 if (!run_scc_vn (do_fre))
4541 {
4542 if (!do_fre)
4543 {
4544 remove_dead_inserted_code ();
4545 loop_optimizer_finalize ();
4546 }
4547
4548 return 0;
4549 }
4550 init_pre (do_fre);
4551 scev_initialize ();
4552
4553
4554 /* Collect and value number expressions computed in each basic block. */
4555 compute_avail ();
4556
4557 if (dump_file && (dump_flags & TDF_DETAILS))
4558 {
4559 basic_block bb;
4560
4561 FOR_ALL_BB (bb)
4562 {
4563 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4564 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4565 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4566 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4567 }
4568 }
4569
4570 /* Insert can get quite slow on an incredibly large number of basic
4571 blocks due to some quadratic behavior. Until this behavior is
4572 fixed, don't run it when he have an incredibly large number of
4573 bb's. If we aren't going to run insert, there is no point in
4574 computing ANTIC, either, even though it's plenty fast. */
4575 if (!do_fre && n_basic_blocks < 4000)
4576 {
4577 compute_antic ();
4578 insert ();
4579 }
4580
4581 /* Remove all the redundant expressions. */
4582 todo |= eliminate ();
4583
4584 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4585 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4586 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4587 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4588 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4589
4590 /* Make sure to remove fake edges before committing our inserts.
4591 This makes sure we don't end up with extra critical edges that
4592 we would need to split. */
4593 remove_fake_exit_edges ();
4594 gsi_commit_edge_inserts ();
4595
4596 clear_expression_ids ();
4597 free_scc_vn ();
4598 if (!do_fre)
4599 remove_dead_inserted_code ();
4600
4601 scev_finalize ();
4602 fini_pre (do_fre);
4603
4604 return todo;
4605 }
4606
4607 /* Gate and execute functions for PRE. */
4608
4609 static unsigned int
4610 do_pre (void)
4611 {
4612 return execute_pre (false);
4613 }
4614
4615 static bool
4616 gate_pre (void)
4617 {
4618 /* PRE tends to generate bigger code. */
4619 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4620 }
4621
4622 struct gimple_opt_pass pass_pre =
4623 {
4624 {
4625 GIMPLE_PASS,
4626 "pre", /* name */
4627 gate_pre, /* gate */
4628 do_pre, /* execute */
4629 NULL, /* sub */
4630 NULL, /* next */
4631 0, /* static_pass_number */
4632 TV_TREE_PRE, /* tv_id */
4633 PROP_no_crit_edges | PROP_cfg
4634 | PROP_ssa, /* properties_required */
4635 0, /* properties_provided */
4636 0, /* properties_destroyed */
4637 TODO_rebuild_alias, /* todo_flags_start */
4638 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4639 | TODO_verify_ssa /* todo_flags_finish */
4640 }
4641 };
4642
4643
4644 /* Gate and execute functions for FRE. */
4645
4646 static unsigned int
4647 execute_fre (void)
4648 {
4649 return execute_pre (true);
4650 }
4651
4652 static bool
4653 gate_fre (void)
4654 {
4655 return flag_tree_fre != 0;
4656 }
4657
4658 struct gimple_opt_pass pass_fre =
4659 {
4660 {
4661 GIMPLE_PASS,
4662 "fre", /* name */
4663 gate_fre, /* gate */
4664 execute_fre, /* execute */
4665 NULL, /* sub */
4666 NULL, /* next */
4667 0, /* static_pass_number */
4668 TV_TREE_FRE, /* tv_id */
4669 PROP_cfg | PROP_ssa, /* properties_required */
4670 0, /* properties_provided */
4671 0, /* properties_destroyed */
4672 0, /* todo_flags_start */
4673 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
4674 }
4675 };