tree-ssa-pre.c (valid_in_sets): Remove checking of trapping operations.
[gcc.git] / gcc / tree-ssa-pre.c
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
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 "tree.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.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 "alloc-pool.h"
40 #include "obstack.h"
41 #include "tree-pass.h"
42 #include "flags.h"
43 #include "bitmap.h"
44 #include "langhooks.h"
45 #include "cfgloop.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
48 #include "params.h"
49 #include "dbgcnt.h"
50
51 /* TODO:
52
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
60 */
61
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
66
67 /* Basic algorithm
68
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
79
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
93
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
96
97 An expression is partially redundant (excluding partial
98 anticipation) if:
99
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
103
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
106
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
110
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
113
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
117
118 /* Representations of value numbers:
119
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
128
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
134
135 /* Representation of expressions on value numbers:
136
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
141
142 /* Representation of sets:
143
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
148
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
152
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
155
156 enum pre_expr_kind
157 {
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
162 };
163
164 typedef union pre_expr_union_d
165 {
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
171
172 typedef struct pre_expr_d
173 {
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
178
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
183
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
186 {
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
189
190 if (e1->kind != e2->kind)
191 return false;
192
193 switch (e1->kind)
194 {
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 gcc_unreachable ();
207 }
208 }
209
210 static hashval_t
211 pre_expr_hash (const void *p1)
212 {
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
215 {
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 gcc_unreachable ();
226 }
227 }
228
229
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
232
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_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 if (expr->kind == NAME)
251 {
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
260 }
261 else
262 {
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
264 gcc_assert (!*slot);
265 *slot = expr;
266 }
267 return next_expression_id - 1;
268 }
269
270 /* Return the expression id for tree EXPR. */
271
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
274 {
275 return expr->id;
276 }
277
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
280 {
281 void **slot;
282
283 if (expr->kind == NAME)
284 {
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
287 return 0;
288 return VEC_index (unsigned, name_to_id, version);
289 }
290 else
291 {
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
293 if (!slot)
294 return 0;
295 return ((pre_expr)*slot)->id;
296 }
297 }
298
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
301
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
304 {
305 unsigned int id = lookup_expression_id (expr);
306 if (id == 0)
307 return alloc_expression_id (expr);
308 return expr->id = id;
309 }
310
311 /* Return the expression that has expression id ID */
312
313 static inline pre_expr
314 expression_for_id (unsigned int id)
315 {
316 return VEC_index (pre_expr, expressions, id);
317 }
318
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
321
322 static void
323 clear_expression_ids (void)
324 {
325 VEC_free (pre_expr, heap, expressions);
326 }
327
328 static alloc_pool pre_expr_pool;
329
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
331
332 static pre_expr
333 get_or_alloc_expr_for_name (tree name)
334 {
335 struct pre_expr_d expr;
336 pre_expr result;
337 unsigned int result_id;
338
339 expr.kind = NAME;
340 expr.id = 0;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
343 if (result_id != 0)
344 return expression_for_id (result_id);
345
346 result = (pre_expr) pool_alloc (pre_expr_pool);
347 result->kind = NAME;
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
350 return result;
351 }
352
353 static bool in_fre = false;
354
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
358 {
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
362
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
365
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
368
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
373
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
376 {
377 /* The EXP_GEN set, which represents expressions/values generated in
378 a basic block. */
379 bitmap_set_t exp_gen;
380
381 /* The PHI_GEN set, which represents PHI results generated in a
382 basic block. */
383 bitmap_set_t phi_gen;
384
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
388
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
392
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
396
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
399 bitmap_set_t pa_in;
400
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
405
406 /* A cache for value_dies_in_block_x. */
407 bitmap expr_dies;
408
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
411
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
415
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
418 } *bb_value_sets_t;
419
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
431
432
433 /* Basic block list in postorder. */
434 static int *postorder;
435
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
438 static struct
439 {
440 /* The number of RHS computations eliminated by PRE. */
441 int eliminations;
442
443 /* The number of new expressions/temporaries generated by PRE. */
444 int insertions;
445
446 /* The number of inserts found due to partial anticipation */
447 int pa_insert;
448
449 /* The number of new PHI nodes added by PRE. */
450 int phis;
451
452 /* The number of values found constant. */
453 int constified;
454
455 } pre_stats;
456
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
465 unsigned int, bool);
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
468 gimple, tree);
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
470 gimple);
471 static unsigned int get_expr_value_id (pre_expr);
472
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
475
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
478
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
483 static tree pretemp;
484 static tree storetemp;
485 static tree prephitemp;
486
487 /* Set of blocks with statements that have had their EH properties changed. */
488 static bitmap need_eh_cleanup;
489
490 /* Set of blocks with statements that have had their AB properties changed. */
491 static bitmap need_ab_cleanup;
492
493 /* The phi_translate_table caches phi translations for a given
494 expression and predecessor. */
495
496 static htab_t phi_translate_table;
497
498 /* A three tuple {e, pred, v} used to cache phi translations in the
499 phi_translate_table. */
500
501 typedef struct expr_pred_trans_d
502 {
503 /* The expression. */
504 pre_expr e;
505
506 /* The predecessor block along which we translated the expression. */
507 basic_block pred;
508
509 /* The value that resulted from the translation. */
510 pre_expr v;
511
512 /* The hashcode for the expression, pred pair. This is cached for
513 speed reasons. */
514 hashval_t hashcode;
515 } *expr_pred_trans_t;
516 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
517
518 /* Return the hash value for a phi translation table entry. */
519
520 static hashval_t
521 expr_pred_trans_hash (const void *p)
522 {
523 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
524 return ve->hashcode;
525 }
526
527 /* Return true if two phi translation table entries are the same.
528 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
529
530 static int
531 expr_pred_trans_eq (const void *p1, const void *p2)
532 {
533 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
534 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
537
538 /* If they are not translations for the same basic block, they can't
539 be equal. */
540 if (b1 != b2)
541 return false;
542 return pre_expr_eq (ve1->e, ve2->e);
543 }
544
545 /* Search in the phi translation table for the translation of
546 expression E in basic block PRED.
547 Return the translated value, if found, NULL otherwise. */
548
549 static inline pre_expr
550 phi_trans_lookup (pre_expr e, basic_block pred)
551 {
552 void **slot;
553 struct expr_pred_trans_d ept;
554
555 ept.e = e;
556 ept.pred = pred;
557 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
558 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
559 NO_INSERT);
560 if (!slot)
561 return NULL;
562 else
563 return ((expr_pred_trans_t) *slot)->v;
564 }
565
566
567 /* Add the tuple mapping from {expression E, basic block PRED} to
568 value V, to the phi translation table. */
569
570 static inline void
571 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
572 {
573 void **slot;
574 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
575 new_pair->e = e;
576 new_pair->pred = pred;
577 new_pair->v = v;
578 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 pred->index);
580
581 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
582 new_pair->hashcode, INSERT);
583 free (*slot);
584 *slot = (void *) new_pair;
585 }
586
587
588 /* Add expression E to the expression set of value id V. */
589
590 static void
591 add_to_value (unsigned int v, pre_expr e)
592 {
593 bitmap_set_t set;
594
595 gcc_assert (get_expr_value_id (e) == v);
596
597 if (v >= VEC_length (bitmap_set_t, value_expressions))
598 {
599 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
600 v + 1);
601 }
602
603 set = VEC_index (bitmap_set_t, value_expressions, v);
604 if (!set)
605 {
606 set = bitmap_set_new ();
607 VEC_replace (bitmap_set_t, value_expressions, v, set);
608 }
609
610 bitmap_insert_into_set_1 (set, e, v, true);
611 }
612
613 /* Create a new bitmap set and return it. */
614
615 static bitmap_set_t
616 bitmap_set_new (void)
617 {
618 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
619 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
620 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
621 return ret;
622 }
623
624 /* Return the value id for a PRE expression EXPR. */
625
626 static unsigned int
627 get_expr_value_id (pre_expr expr)
628 {
629 switch (expr->kind)
630 {
631 case CONSTANT:
632 {
633 unsigned int id;
634 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
635 if (id == 0)
636 {
637 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
638 add_to_value (id, expr);
639 }
640 return id;
641 }
642 case NAME:
643 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 case NARY:
645 return PRE_EXPR_NARY (expr)->value_id;
646 case REFERENCE:
647 return PRE_EXPR_REFERENCE (expr)->value_id;
648 default:
649 gcc_unreachable ();
650 }
651 }
652
653 /* Remove an expression EXPR from a bitmapped set. */
654
655 static void
656 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
657 {
658 unsigned int val = get_expr_value_id (expr);
659 if (!value_id_constant_p (val))
660 {
661 bitmap_clear_bit (&set->values, val);
662 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
663 }
664 }
665
666 static void
667 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
668 unsigned int val, bool allow_constants)
669 {
670 if (allow_constants || !value_id_constant_p (val))
671 {
672 /* We specifically expect this and only this function to be able to
673 insert constants into a set. */
674 bitmap_set_bit (&set->values, val);
675 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
676 }
677 }
678
679 /* Insert an expression EXPR into a bitmapped set. */
680
681 static void
682 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
683 {
684 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
685 }
686
687 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
688
689 static void
690 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
691 {
692 bitmap_copy (&dest->expressions, &orig->expressions);
693 bitmap_copy (&dest->values, &orig->values);
694 }
695
696
697 /* Free memory used up by SET. */
698 static void
699 bitmap_set_free (bitmap_set_t set)
700 {
701 bitmap_clear (&set->expressions);
702 bitmap_clear (&set->values);
703 }
704
705
706 /* Generate an topological-ordered array of bitmap set SET. */
707
708 static VEC(pre_expr, heap) *
709 sorted_array_from_bitmap_set (bitmap_set_t set)
710 {
711 unsigned int i, j;
712 bitmap_iterator bi, bj;
713 VEC(pre_expr, heap) *result;
714
715 /* Pre-allocate roughly enough space for the array. */
716 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
717
718 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
719 {
720 /* The number of expressions having a given value is usually
721 relatively small. Thus, rather than making a vector of all
722 the expressions and sorting it by value-id, we walk the values
723 and check in the reverse mapping that tells us what expressions
724 have a given value, to filter those in our set. As a result,
725 the expressions are inserted in value-id order, which means
726 topological order.
727
728 If this is somehow a significant lose for some cases, we can
729 choose which set to walk based on the set size. */
730 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
731 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
732 {
733 if (bitmap_bit_p (&set->expressions, j))
734 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
735 }
736 }
737
738 return result;
739 }
740
741 /* Perform bitmapped set operation DEST &= ORIG. */
742
743 static void
744 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
745 {
746 bitmap_iterator bi;
747 unsigned int i;
748
749 if (dest != orig)
750 {
751 bitmap_head temp;
752 bitmap_initialize (&temp, &grand_bitmap_obstack);
753
754 bitmap_and_into (&dest->values, &orig->values);
755 bitmap_copy (&temp, &dest->expressions);
756 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
757 {
758 pre_expr expr = expression_for_id (i);
759 unsigned int value_id = get_expr_value_id (expr);
760 if (!bitmap_bit_p (&dest->values, value_id))
761 bitmap_clear_bit (&dest->expressions, i);
762 }
763 bitmap_clear (&temp);
764 }
765 }
766
767 /* Subtract all values and expressions contained in ORIG from DEST. */
768
769 static bitmap_set_t
770 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
771 {
772 bitmap_set_t result = bitmap_set_new ();
773 bitmap_iterator bi;
774 unsigned int i;
775
776 bitmap_and_compl (&result->expressions, &dest->expressions,
777 &orig->expressions);
778
779 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
780 {
781 pre_expr expr = expression_for_id (i);
782 unsigned int value_id = get_expr_value_id (expr);
783 bitmap_set_bit (&result->values, value_id);
784 }
785
786 return result;
787 }
788
789 /* Subtract all the values in bitmap set B from bitmap set A. */
790
791 static void
792 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
793 {
794 unsigned int i;
795 bitmap_iterator bi;
796 bitmap_head temp;
797
798 bitmap_initialize (&temp, &grand_bitmap_obstack);
799
800 bitmap_copy (&temp, &a->expressions);
801 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
802 {
803 pre_expr expr = expression_for_id (i);
804 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
805 bitmap_remove_from_set (a, expr);
806 }
807 bitmap_clear (&temp);
808 }
809
810
811 /* Return true if bitmapped set SET contains the value VALUE_ID. */
812
813 static bool
814 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
815 {
816 if (value_id_constant_p (value_id))
817 return true;
818
819 if (!set || bitmap_empty_p (&set->expressions))
820 return false;
821
822 return bitmap_bit_p (&set->values, value_id);
823 }
824
825 static inline bool
826 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
827 {
828 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
829 }
830
831 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
832
833 static void
834 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
835 const pre_expr expr)
836 {
837 bitmap_set_t exprset;
838 unsigned int i;
839 bitmap_iterator bi;
840
841 if (value_id_constant_p (lookfor))
842 return;
843
844 if (!bitmap_set_contains_value (set, lookfor))
845 return;
846
847 /* The number of expressions having a given value is usually
848 significantly less than the total number of expressions in SET.
849 Thus, rather than check, for each expression in SET, whether it
850 has the value LOOKFOR, we walk the reverse mapping that tells us
851 what expressions have a given value, and see if any of those
852 expressions are in our set. For large testcases, this is about
853 5-10x faster than walking the bitmap. If this is somehow a
854 significant lose for some cases, we can choose which set to walk
855 based on the set size. */
856 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
857 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
858 {
859 if (bitmap_clear_bit (&set->expressions, i))
860 {
861 bitmap_set_bit (&set->expressions, get_expression_id (expr));
862 return;
863 }
864 }
865 }
866
867 /* Return true if two bitmap sets are equal. */
868
869 static bool
870 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
871 {
872 return bitmap_equal_p (&a->values, &b->values);
873 }
874
875 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
876 and add it otherwise. */
877
878 static void
879 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
880 {
881 unsigned int val = get_expr_value_id (expr);
882
883 if (bitmap_set_contains_value (set, val))
884 bitmap_set_replace_value (set, val, expr);
885 else
886 bitmap_insert_into_set (set, expr);
887 }
888
889 /* Insert EXPR into SET if EXPR's value is not already present in
890 SET. */
891
892 static void
893 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
894 {
895 unsigned int val = get_expr_value_id (expr);
896
897 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
898
899 /* Constant values are always considered to be part of the set. */
900 if (value_id_constant_p (val))
901 return;
902
903 /* If the value membership changed, add the expression. */
904 if (bitmap_set_bit (&set->values, val))
905 bitmap_set_bit (&set->expressions, expr->id);
906 }
907
908 /* Print out EXPR to outfile. */
909
910 static void
911 print_pre_expr (FILE *outfile, const pre_expr expr)
912 {
913 switch (expr->kind)
914 {
915 case CONSTANT:
916 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
917 break;
918 case NAME:
919 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
920 break;
921 case NARY:
922 {
923 unsigned int i;
924 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
925 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
926 for (i = 0; i < nary->length; i++)
927 {
928 print_generic_expr (outfile, nary->op[i], 0);
929 if (i != (unsigned) nary->length - 1)
930 fprintf (outfile, ",");
931 }
932 fprintf (outfile, "}");
933 }
934 break;
935
936 case REFERENCE:
937 {
938 vn_reference_op_t vro;
939 unsigned int i;
940 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
941 fprintf (outfile, "{");
942 for (i = 0;
943 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
944 i++)
945 {
946 bool closebrace = false;
947 if (vro->opcode != SSA_NAME
948 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
949 {
950 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
951 if (vro->op0)
952 {
953 fprintf (outfile, "<");
954 closebrace = true;
955 }
956 }
957 if (vro->op0)
958 {
959 print_generic_expr (outfile, vro->op0, 0);
960 if (vro->op1)
961 {
962 fprintf (outfile, ",");
963 print_generic_expr (outfile, vro->op1, 0);
964 }
965 if (vro->op2)
966 {
967 fprintf (outfile, ",");
968 print_generic_expr (outfile, vro->op2, 0);
969 }
970 }
971 if (closebrace)
972 fprintf (outfile, ">");
973 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
974 fprintf (outfile, ",");
975 }
976 fprintf (outfile, "}");
977 if (ref->vuse)
978 {
979 fprintf (outfile, "@");
980 print_generic_expr (outfile, ref->vuse, 0);
981 }
982 }
983 break;
984 }
985 }
986 void debug_pre_expr (pre_expr);
987
988 /* Like print_pre_expr but always prints to stderr. */
989 DEBUG_FUNCTION void
990 debug_pre_expr (pre_expr e)
991 {
992 print_pre_expr (stderr, e);
993 fprintf (stderr, "\n");
994 }
995
996 /* Print out SET to OUTFILE. */
997
998 static void
999 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1000 const char *setname, int blockindex)
1001 {
1002 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1003 if (set)
1004 {
1005 bool first = true;
1006 unsigned i;
1007 bitmap_iterator bi;
1008
1009 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1010 {
1011 const pre_expr expr = expression_for_id (i);
1012
1013 if (!first)
1014 fprintf (outfile, ", ");
1015 first = false;
1016 print_pre_expr (outfile, expr);
1017
1018 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1019 }
1020 }
1021 fprintf (outfile, " }\n");
1022 }
1023
1024 void debug_bitmap_set (bitmap_set_t);
1025
1026 DEBUG_FUNCTION void
1027 debug_bitmap_set (bitmap_set_t set)
1028 {
1029 print_bitmap_set (stderr, set, "debug", 0);
1030 }
1031
1032 void debug_bitmap_sets_for (basic_block);
1033
1034 DEBUG_FUNCTION void
1035 debug_bitmap_sets_for (basic_block bb)
1036 {
1037 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1038 if (!in_fre)
1039 {
1040 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1041 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1042 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1043 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1044 if (do_partial_partial)
1045 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1046 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1047 }
1048 }
1049
1050 /* Print out the expressions that have VAL to OUTFILE. */
1051
1052 static void
1053 print_value_expressions (FILE *outfile, unsigned int val)
1054 {
1055 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1056 if (set)
1057 {
1058 char s[10];
1059 sprintf (s, "%04d", val);
1060 print_bitmap_set (outfile, set, s, 0);
1061 }
1062 }
1063
1064
1065 DEBUG_FUNCTION void
1066 debug_value_expressions (unsigned int val)
1067 {
1068 print_value_expressions (stderr, val);
1069 }
1070
1071 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1072 represent it. */
1073
1074 static pre_expr
1075 get_or_alloc_expr_for_constant (tree constant)
1076 {
1077 unsigned int result_id;
1078 unsigned int value_id;
1079 struct pre_expr_d expr;
1080 pre_expr newexpr;
1081
1082 expr.kind = CONSTANT;
1083 PRE_EXPR_CONSTANT (&expr) = constant;
1084 result_id = lookup_expression_id (&expr);
1085 if (result_id != 0)
1086 return expression_for_id (result_id);
1087
1088 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1089 newexpr->kind = CONSTANT;
1090 PRE_EXPR_CONSTANT (newexpr) = constant;
1091 alloc_expression_id (newexpr);
1092 value_id = get_or_alloc_constant_value_id (constant);
1093 add_to_value (value_id, newexpr);
1094 return newexpr;
1095 }
1096
1097 /* Given a value id V, find the actual tree representing the constant
1098 value if there is one, and return it. Return NULL if we can't find
1099 a constant. */
1100
1101 static tree
1102 get_constant_for_value_id (unsigned int v)
1103 {
1104 if (value_id_constant_p (v))
1105 {
1106 unsigned int i;
1107 bitmap_iterator bi;
1108 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1109
1110 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1111 {
1112 pre_expr expr = expression_for_id (i);
1113 if (expr->kind == CONSTANT)
1114 return PRE_EXPR_CONSTANT (expr);
1115 }
1116 }
1117 return NULL;
1118 }
1119
1120 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1121 Currently only supports constants and SSA_NAMES. */
1122 static pre_expr
1123 get_or_alloc_expr_for (tree t)
1124 {
1125 if (TREE_CODE (t) == SSA_NAME)
1126 return get_or_alloc_expr_for_name (t);
1127 else if (is_gimple_min_invariant (t))
1128 return get_or_alloc_expr_for_constant (t);
1129 else
1130 {
1131 /* More complex expressions can result from SCCVN expression
1132 simplification that inserts values for them. As they all
1133 do not have VOPs the get handled by the nary ops struct. */
1134 vn_nary_op_t result;
1135 unsigned int result_id;
1136 vn_nary_op_lookup (t, &result);
1137 if (result != NULL)
1138 {
1139 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1140 e->kind = NARY;
1141 PRE_EXPR_NARY (e) = result;
1142 result_id = lookup_expression_id (e);
1143 if (result_id != 0)
1144 {
1145 pool_free (pre_expr_pool, e);
1146 e = expression_for_id (result_id);
1147 return e;
1148 }
1149 alloc_expression_id (e);
1150 return e;
1151 }
1152 }
1153 return NULL;
1154 }
1155
1156 /* Return the folded version of T if T, when folded, is a gimple
1157 min_invariant. Otherwise, return T. */
1158
1159 static pre_expr
1160 fully_constant_expression (pre_expr e)
1161 {
1162 switch (e->kind)
1163 {
1164 case CONSTANT:
1165 return e;
1166 case NARY:
1167 {
1168 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1169 switch (TREE_CODE_CLASS (nary->opcode))
1170 {
1171 case tcc_binary:
1172 case tcc_comparison:
1173 {
1174 /* We have to go from trees to pre exprs to value ids to
1175 constants. */
1176 tree naryop0 = nary->op[0];
1177 tree naryop1 = nary->op[1];
1178 tree result;
1179 if (!is_gimple_min_invariant (naryop0))
1180 {
1181 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1182 unsigned int vrep0 = get_expr_value_id (rep0);
1183 tree const0 = get_constant_for_value_id (vrep0);
1184 if (const0)
1185 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1186 }
1187 if (!is_gimple_min_invariant (naryop1))
1188 {
1189 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1190 unsigned int vrep1 = get_expr_value_id (rep1);
1191 tree const1 = get_constant_for_value_id (vrep1);
1192 if (const1)
1193 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1194 }
1195 result = fold_binary (nary->opcode, nary->type,
1196 naryop0, naryop1);
1197 if (result && is_gimple_min_invariant (result))
1198 return get_or_alloc_expr_for_constant (result);
1199 /* We might have simplified the expression to a
1200 SSA_NAME for example from x_1 * 1. But we cannot
1201 insert a PHI for x_1 unconditionally as x_1 might
1202 not be available readily. */
1203 return e;
1204 }
1205 case tcc_reference:
1206 if (nary->opcode != REALPART_EXPR
1207 && nary->opcode != IMAGPART_EXPR
1208 && nary->opcode != VIEW_CONVERT_EXPR)
1209 return e;
1210 /* Fallthrough. */
1211 case tcc_unary:
1212 {
1213 /* We have to go from trees to pre exprs to value ids to
1214 constants. */
1215 tree naryop0 = nary->op[0];
1216 tree const0, result;
1217 if (is_gimple_min_invariant (naryop0))
1218 const0 = naryop0;
1219 else
1220 {
1221 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1222 unsigned int vrep0 = get_expr_value_id (rep0);
1223 const0 = get_constant_for_value_id (vrep0);
1224 }
1225 result = NULL;
1226 if (const0)
1227 {
1228 tree type1 = TREE_TYPE (nary->op[0]);
1229 const0 = fold_convert (type1, const0);
1230 result = fold_unary (nary->opcode, nary->type, const0);
1231 }
1232 if (result && is_gimple_min_invariant (result))
1233 return get_or_alloc_expr_for_constant (result);
1234 return e;
1235 }
1236 default:
1237 return e;
1238 }
1239 }
1240 case REFERENCE:
1241 {
1242 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1243 tree folded;
1244 if ((folded = fully_constant_vn_reference_p (ref)))
1245 return get_or_alloc_expr_for_constant (folded);
1246 return e;
1247 }
1248 default:
1249 return e;
1250 }
1251 return e;
1252 }
1253
1254 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1255 it has the value it would have in BLOCK. Set *SAME_VALID to true
1256 in case the new vuse doesn't change the value id of the OPERANDS. */
1257
1258 static tree
1259 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1260 alias_set_type set, tree type, tree vuse,
1261 basic_block phiblock,
1262 basic_block block, bool *same_valid)
1263 {
1264 gimple phi = SSA_NAME_DEF_STMT (vuse);
1265 ao_ref ref;
1266 edge e = NULL;
1267 bool use_oracle;
1268
1269 *same_valid = true;
1270
1271 if (gimple_bb (phi) != phiblock)
1272 return vuse;
1273
1274 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1275
1276 /* Use the alias-oracle to find either the PHI node in this block,
1277 the first VUSE used in this block that is equivalent to vuse or
1278 the first VUSE which definition in this block kills the value. */
1279 if (gimple_code (phi) == GIMPLE_PHI)
1280 e = find_edge (block, phiblock);
1281 else if (use_oracle)
1282 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1283 {
1284 vuse = gimple_vuse (phi);
1285 phi = SSA_NAME_DEF_STMT (vuse);
1286 if (gimple_bb (phi) != phiblock)
1287 return vuse;
1288 if (gimple_code (phi) == GIMPLE_PHI)
1289 {
1290 e = find_edge (block, phiblock);
1291 break;
1292 }
1293 }
1294 else
1295 return NULL_TREE;
1296
1297 if (e)
1298 {
1299 if (use_oracle)
1300 {
1301 bitmap visited = NULL;
1302 /* Try to find a vuse that dominates this phi node by skipping
1303 non-clobbering statements. */
1304 vuse = get_continuation_for_phi (phi, &ref, &visited);
1305 if (visited)
1306 BITMAP_FREE (visited);
1307 }
1308 else
1309 vuse = NULL_TREE;
1310 if (!vuse)
1311 {
1312 /* If we didn't find any, the value ID can't stay the same,
1313 but return the translated vuse. */
1314 *same_valid = false;
1315 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1316 }
1317 /* ??? We would like to return vuse here as this is the canonical
1318 upmost vdef that this reference is associated with. But during
1319 insertion of the references into the hash tables we only ever
1320 directly insert with their direct gimple_vuse, hence returning
1321 something else would make us not find the other expression. */
1322 return PHI_ARG_DEF (phi, e->dest_idx);
1323 }
1324
1325 return NULL_TREE;
1326 }
1327
1328 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1329 SET2. This is used to avoid making a set consisting of the union
1330 of PA_IN and ANTIC_IN during insert. */
1331
1332 static inline pre_expr
1333 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1334 {
1335 pre_expr result;
1336
1337 result = bitmap_find_leader (set1, val, NULL);
1338 if (!result && set2)
1339 result = bitmap_find_leader (set2, val, NULL);
1340 return result;
1341 }
1342
1343 /* Get the tree type for our PRE expression e. */
1344
1345 static tree
1346 get_expr_type (const pre_expr e)
1347 {
1348 switch (e->kind)
1349 {
1350 case NAME:
1351 return TREE_TYPE (PRE_EXPR_NAME (e));
1352 case CONSTANT:
1353 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1354 case REFERENCE:
1355 return PRE_EXPR_REFERENCE (e)->type;
1356 case NARY:
1357 return PRE_EXPR_NARY (e)->type;
1358 }
1359 gcc_unreachable();
1360 }
1361
1362 /* Get a representative SSA_NAME for a given expression.
1363 Since all of our sub-expressions are treated as values, we require
1364 them to be SSA_NAME's for simplicity.
1365 Prior versions of GVNPRE used to use "value handles" here, so that
1366 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1367 either case, the operands are really values (IE we do not expect
1368 them to be usable without finding leaders). */
1369
1370 static tree
1371 get_representative_for (const pre_expr e)
1372 {
1373 tree exprtype;
1374 tree name;
1375 unsigned int value_id = get_expr_value_id (e);
1376
1377 switch (e->kind)
1378 {
1379 case NAME:
1380 return PRE_EXPR_NAME (e);
1381 case CONSTANT:
1382 return PRE_EXPR_CONSTANT (e);
1383 case NARY:
1384 case REFERENCE:
1385 {
1386 /* Go through all of the expressions representing this value
1387 and pick out an SSA_NAME. */
1388 unsigned int i;
1389 bitmap_iterator bi;
1390 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1391 value_id);
1392 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1393 {
1394 pre_expr rep = expression_for_id (i);
1395 if (rep->kind == NAME)
1396 return PRE_EXPR_NAME (rep);
1397 }
1398 }
1399 break;
1400 }
1401 /* If we reached here we couldn't find an SSA_NAME. This can
1402 happen when we've discovered a value that has never appeared in
1403 the program as set to an SSA_NAME, most likely as the result of
1404 phi translation. */
1405 if (dump_file)
1406 {
1407 fprintf (dump_file,
1408 "Could not find SSA_NAME representative for expression:");
1409 print_pre_expr (dump_file, e);
1410 fprintf (dump_file, "\n");
1411 }
1412
1413 exprtype = get_expr_type (e);
1414
1415 /* Build and insert the assignment of the end result to the temporary
1416 that we will return. */
1417 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1418 {
1419 pretemp = create_tmp_reg (exprtype, "pretmp");
1420 add_referenced_var (pretemp);
1421 }
1422
1423 name = make_ssa_name (pretemp, gimple_build_nop ());
1424 VN_INFO_GET (name)->value_id = value_id;
1425 if (e->kind == CONSTANT)
1426 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1427 else
1428 VN_INFO (name)->valnum = name;
1429
1430 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1431 if (dump_file)
1432 {
1433 fprintf (dump_file, "Created SSA_NAME representative ");
1434 print_generic_expr (dump_file, name, 0);
1435 fprintf (dump_file, " for expression:");
1436 print_pre_expr (dump_file, e);
1437 fprintf (dump_file, "\n");
1438 }
1439
1440 return name;
1441 }
1442
1443
1444
1445 static pre_expr
1446 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1447 basic_block pred, basic_block phiblock);
1448
1449 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1450 the phis in PRED. Return NULL if we can't find a leader for each part
1451 of the translated expression. */
1452
1453 static pre_expr
1454 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1455 basic_block pred, basic_block phiblock)
1456 {
1457 switch (expr->kind)
1458 {
1459 case NARY:
1460 {
1461 unsigned int i;
1462 bool changed = false;
1463 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1464 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1465 sizeof_vn_nary_op (nary->length));
1466 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1467
1468 for (i = 0; i < newnary->length; i++)
1469 {
1470 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1471 continue;
1472 else
1473 {
1474 pre_expr leader, result;
1475 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1476 leader = find_leader_in_sets (op_val_id, set1, set2);
1477 result = phi_translate (leader, set1, set2, pred, phiblock);
1478 if (result && result != leader)
1479 {
1480 tree name = get_representative_for (result);
1481 if (!name)
1482 return NULL;
1483 newnary->op[i] = name;
1484 }
1485 else if (!result)
1486 return NULL;
1487
1488 changed |= newnary->op[i] != nary->op[i];
1489 }
1490 }
1491 if (changed)
1492 {
1493 pre_expr constant;
1494 unsigned int new_val_id;
1495
1496 tree result = vn_nary_op_lookup_pieces (newnary->length,
1497 newnary->opcode,
1498 newnary->type,
1499 &newnary->op[0],
1500 &nary);
1501 if (result && is_gimple_min_invariant (result))
1502 return get_or_alloc_expr_for_constant (result);
1503
1504 expr = (pre_expr) pool_alloc (pre_expr_pool);
1505 expr->kind = NARY;
1506 expr->id = 0;
1507 if (nary)
1508 {
1509 PRE_EXPR_NARY (expr) = nary;
1510 constant = fully_constant_expression (expr);
1511 if (constant != expr)
1512 return constant;
1513
1514 new_val_id = nary->value_id;
1515 get_or_alloc_expression_id (expr);
1516 }
1517 else
1518 {
1519 new_val_id = get_next_value_id ();
1520 VEC_safe_grow_cleared (bitmap_set_t, heap,
1521 value_expressions,
1522 get_max_value_id() + 1);
1523 nary = vn_nary_op_insert_pieces (newnary->length,
1524 newnary->opcode,
1525 newnary->type,
1526 &newnary->op[0],
1527 result, new_val_id);
1528 PRE_EXPR_NARY (expr) = nary;
1529 constant = fully_constant_expression (expr);
1530 if (constant != expr)
1531 return constant;
1532 get_or_alloc_expression_id (expr);
1533 }
1534 add_to_value (new_val_id, expr);
1535 }
1536 return expr;
1537 }
1538 break;
1539
1540 case REFERENCE:
1541 {
1542 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1543 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1544 tree vuse = ref->vuse;
1545 tree newvuse = vuse;
1546 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1547 bool changed = false, same_valid = true;
1548 unsigned int i, j, n;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1551
1552 for (i = 0, j = 0;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1554 {
1555 pre_expr opresult;
1556 pre_expr leader;
1557 tree op[3];
1558 tree type = operand->type;
1559 vn_reference_op_s newop = *operand;
1560 op[0] = operand->op0;
1561 op[1] = operand->op1;
1562 op[2] = operand->op2;
1563 for (n = 0; n < 3; ++n)
1564 {
1565 unsigned int op_val_id;
1566 if (!op[n])
1567 continue;
1568 if (TREE_CODE (op[n]) != SSA_NAME)
1569 {
1570 /* We can't possibly insert these. */
1571 if (n != 0
1572 && !is_gimple_min_invariant (op[n]))
1573 break;
1574 continue;
1575 }
1576 op_val_id = VN_INFO (op[n])->value_id;
1577 leader = find_leader_in_sets (op_val_id, set1, set2);
1578 if (!leader)
1579 break;
1580 /* Make sure we do not recursively translate ourselves
1581 like for translating a[n_1] with the leader for
1582 n_1 being a[n_1]. */
1583 if (get_expression_id (leader) != get_expression_id (expr))
1584 {
1585 opresult = phi_translate (leader, set1, set2,
1586 pred, phiblock);
1587 if (!opresult)
1588 break;
1589 if (opresult != leader)
1590 {
1591 tree name = get_representative_for (opresult);
1592 if (!name)
1593 break;
1594 changed |= name != op[n];
1595 op[n] = name;
1596 }
1597 }
1598 }
1599 if (n != 3)
1600 {
1601 if (newoperands)
1602 VEC_free (vn_reference_op_s, heap, newoperands);
1603 return NULL;
1604 }
1605 if (!newoperands)
1606 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1607 /* We may have changed from an SSA_NAME to a constant */
1608 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1609 newop.opcode = TREE_CODE (op[0]);
1610 newop.type = type;
1611 newop.op0 = op[0];
1612 newop.op1 = op[1];
1613 newop.op2 = op[2];
1614 /* If it transforms a non-constant ARRAY_REF into a constant
1615 one, adjust the constant offset. */
1616 if (newop.opcode == ARRAY_REF
1617 && newop.off == -1
1618 && TREE_CODE (op[0]) == INTEGER_CST
1619 && TREE_CODE (op[1]) == INTEGER_CST
1620 && TREE_CODE (op[2]) == INTEGER_CST)
1621 {
1622 double_int off = tree_to_double_int (op[0]);
1623 off = double_int_add (off,
1624 double_int_neg
1625 (tree_to_double_int (op[1])));
1626 off = double_int_mul (off, tree_to_double_int (op[2]));
1627 if (double_int_fits_in_shwi_p (off))
1628 newop.off = off.low;
1629 }
1630 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1631 /* If it transforms from an SSA_NAME to an address, fold with
1632 a preceding indirect reference. */
1633 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1634 && VEC_index (vn_reference_op_s,
1635 newoperands, j - 1)->opcode == MEM_REF)
1636 vn_reference_fold_indirect (&newoperands, &j);
1637 }
1638 if (i != VEC_length (vn_reference_op_s, operands))
1639 {
1640 if (newoperands)
1641 VEC_free (vn_reference_op_s, heap, newoperands);
1642 return NULL;
1643 }
1644
1645 if (vuse)
1646 {
1647 newvuse = translate_vuse_through_block (newoperands,
1648 ref->set, ref->type,
1649 vuse, phiblock, pred,
1650 &same_valid);
1651 if (newvuse == NULL_TREE)
1652 {
1653 VEC_free (vn_reference_op_s, heap, newoperands);
1654 return NULL;
1655 }
1656 }
1657
1658 if (changed || newvuse != vuse)
1659 {
1660 unsigned int new_val_id;
1661 pre_expr constant;
1662 bool converted = false;
1663
1664 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1665 ref->type,
1666 newoperands,
1667 &newref, VN_WALK);
1668 if (result)
1669 VEC_free (vn_reference_op_s, heap, newoperands);
1670
1671 if (result
1672 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1673 {
1674 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1675 converted = true;
1676 }
1677 else if (!result && newref
1678 && !useless_type_conversion_p (ref->type, newref->type))
1679 {
1680 VEC_free (vn_reference_op_s, heap, newoperands);
1681 return NULL;
1682 }
1683
1684 if (result && is_gimple_min_invariant (result))
1685 {
1686 gcc_assert (!newoperands);
1687 return get_or_alloc_expr_for_constant (result);
1688 }
1689
1690 expr = (pre_expr) pool_alloc (pre_expr_pool);
1691 expr->kind = REFERENCE;
1692 expr->id = 0;
1693
1694 if (converted)
1695 {
1696 vn_nary_op_t nary;
1697 tree nresult;
1698
1699 gcc_assert (CONVERT_EXPR_P (result)
1700 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1701
1702 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1703 TREE_TYPE (result),
1704 &TREE_OPERAND (result, 0),
1705 &nary);
1706 if (nresult && is_gimple_min_invariant (nresult))
1707 return get_or_alloc_expr_for_constant (nresult);
1708
1709 expr->kind = NARY;
1710 if (nary)
1711 {
1712 PRE_EXPR_NARY (expr) = nary;
1713 constant = fully_constant_expression (expr);
1714 if (constant != expr)
1715 return constant;
1716
1717 new_val_id = nary->value_id;
1718 get_or_alloc_expression_id (expr);
1719 }
1720 else
1721 {
1722 new_val_id = get_next_value_id ();
1723 VEC_safe_grow_cleared (bitmap_set_t, heap,
1724 value_expressions,
1725 get_max_value_id() + 1);
1726 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1727 TREE_TYPE (result),
1728 &TREE_OPERAND (result, 0),
1729 NULL_TREE,
1730 new_val_id);
1731 PRE_EXPR_NARY (expr) = nary;
1732 constant = fully_constant_expression (expr);
1733 if (constant != expr)
1734 return constant;
1735 get_or_alloc_expression_id (expr);
1736 }
1737 }
1738 else if (newref)
1739 {
1740 PRE_EXPR_REFERENCE (expr) = newref;
1741 constant = fully_constant_expression (expr);
1742 if (constant != expr)
1743 return constant;
1744
1745 new_val_id = newref->value_id;
1746 get_or_alloc_expression_id (expr);
1747 }
1748 else
1749 {
1750 if (changed || !same_valid)
1751 {
1752 new_val_id = get_next_value_id ();
1753 VEC_safe_grow_cleared (bitmap_set_t, heap,
1754 value_expressions,
1755 get_max_value_id() + 1);
1756 }
1757 else
1758 new_val_id = ref->value_id;
1759 newref = vn_reference_insert_pieces (newvuse, ref->set,
1760 ref->type,
1761 newoperands,
1762 result, new_val_id);
1763 newoperands = NULL;
1764 PRE_EXPR_REFERENCE (expr) = newref;
1765 constant = fully_constant_expression (expr);
1766 if (constant != expr)
1767 return constant;
1768 get_or_alloc_expression_id (expr);
1769 }
1770 add_to_value (new_val_id, expr);
1771 }
1772 VEC_free (vn_reference_op_s, heap, newoperands);
1773 return expr;
1774 }
1775 break;
1776
1777 case NAME:
1778 {
1779 gimple phi = NULL;
1780 edge e;
1781 gimple def_stmt;
1782 tree name = PRE_EXPR_NAME (expr);
1783
1784 def_stmt = SSA_NAME_DEF_STMT (name);
1785 if (gimple_code (def_stmt) == GIMPLE_PHI
1786 && gimple_bb (def_stmt) == phiblock)
1787 phi = def_stmt;
1788 else
1789 return expr;
1790
1791 e = find_edge (pred, gimple_bb (phi));
1792 if (e)
1793 {
1794 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1795 pre_expr newexpr;
1796
1797 if (TREE_CODE (def) == SSA_NAME)
1798 def = VN_INFO (def)->valnum;
1799
1800 /* Handle constant. */
1801 if (is_gimple_min_invariant (def))
1802 return get_or_alloc_expr_for_constant (def);
1803
1804 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1805 return NULL;
1806
1807 newexpr = get_or_alloc_expr_for_name (def);
1808 return newexpr;
1809 }
1810 }
1811 return expr;
1812
1813 default:
1814 gcc_unreachable ();
1815 }
1816 }
1817
1818 /* Wrapper around phi_translate_1 providing caching functionality. */
1819
1820 static pre_expr
1821 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1822 basic_block pred, basic_block phiblock)
1823 {
1824 pre_expr phitrans;
1825
1826 if (!expr)
1827 return NULL;
1828
1829 /* Constants contain no values that need translation. */
1830 if (expr->kind == CONSTANT)
1831 return expr;
1832
1833 if (value_id_constant_p (get_expr_value_id (expr)))
1834 return expr;
1835
1836 if (expr->kind != NAME)
1837 {
1838 phitrans = phi_trans_lookup (expr, pred);
1839 if (phitrans)
1840 return phitrans;
1841 }
1842
1843 /* Translate. */
1844 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1845
1846 /* Don't add empty translations to the cache. Neither add
1847 translations of NAMEs as those are cheap to translate. */
1848 if (phitrans
1849 && expr->kind != NAME)
1850 phi_trans_add (expr, phitrans, pred);
1851
1852 return phitrans;
1853 }
1854
1855
1856 /* For each expression in SET, translate the values through phi nodes
1857 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1858 expressions in DEST. */
1859
1860 static void
1861 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1862 basic_block phiblock)
1863 {
1864 VEC (pre_expr, heap) *exprs;
1865 pre_expr expr;
1866 int i;
1867
1868 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1869 {
1870 bitmap_set_copy (dest, set);
1871 return;
1872 }
1873
1874 exprs = sorted_array_from_bitmap_set (set);
1875 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1876 {
1877 pre_expr translated;
1878 translated = phi_translate (expr, set, NULL, pred, phiblock);
1879 if (!translated)
1880 continue;
1881
1882 /* We might end up with multiple expressions from SET being
1883 translated to the same value. In this case we do not want
1884 to retain the NARY or REFERENCE expression but prefer a NAME
1885 which would be the leader. */
1886 if (translated->kind == NAME)
1887 bitmap_value_replace_in_set (dest, translated);
1888 else
1889 bitmap_value_insert_into_set (dest, translated);
1890 }
1891 VEC_free (pre_expr, heap, exprs);
1892 }
1893
1894 /* Find the leader for a value (i.e., the name representing that
1895 value) in a given set, and return it. If STMT is non-NULL it
1896 makes sure the defining statement for the leader dominates it.
1897 Return NULL if no leader is found. */
1898
1899 static pre_expr
1900 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1901 {
1902 if (value_id_constant_p (val))
1903 {
1904 unsigned int i;
1905 bitmap_iterator bi;
1906 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1907
1908 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1909 {
1910 pre_expr expr = expression_for_id (i);
1911 if (expr->kind == CONSTANT)
1912 return expr;
1913 }
1914 }
1915 if (bitmap_set_contains_value (set, val))
1916 {
1917 /* Rather than walk the entire bitmap of expressions, and see
1918 whether any of them has the value we are looking for, we look
1919 at the reverse mapping, which tells us the set of expressions
1920 that have a given value (IE value->expressions with that
1921 value) and see if any of those expressions are in our set.
1922 The number of expressions per value is usually significantly
1923 less than the number of expressions in the set. In fact, for
1924 large testcases, doing it this way is roughly 5-10x faster
1925 than walking the bitmap.
1926 If this is somehow a significant lose for some cases, we can
1927 choose which set to walk based on which set is smaller. */
1928 unsigned int i;
1929 bitmap_iterator bi;
1930 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1931
1932 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1933 &set->expressions, 0, i, bi)
1934 {
1935 pre_expr val = expression_for_id (i);
1936 /* At the point where stmt is not null, there should always
1937 be an SSA_NAME first in the list of expressions. */
1938 if (stmt)
1939 {
1940 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1941 if (gimple_code (def_stmt) != GIMPLE_PHI
1942 && gimple_bb (def_stmt) == gimple_bb (stmt)
1943 /* PRE insertions are at the end of the basic-block
1944 and have UID 0. */
1945 && (gimple_uid (def_stmt) == 0
1946 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1947 continue;
1948 }
1949 return val;
1950 }
1951 }
1952 return NULL;
1953 }
1954
1955 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1956 BLOCK by seeing if it is not killed in the block. Note that we are
1957 only determining whether there is a store that kills it. Because
1958 of the order in which clean iterates over values, we are guaranteed
1959 that altered operands will have caused us to be eliminated from the
1960 ANTIC_IN set already. */
1961
1962 static bool
1963 value_dies_in_block_x (pre_expr expr, basic_block block)
1964 {
1965 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1966 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1967 gimple def;
1968 gimple_stmt_iterator gsi;
1969 unsigned id = get_expression_id (expr);
1970 bool res = false;
1971 ao_ref ref;
1972
1973 if (!vuse)
1974 return false;
1975
1976 /* Lookup a previously calculated result. */
1977 if (EXPR_DIES (block)
1978 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1979 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1980
1981 /* A memory expression {e, VUSE} dies in the block if there is a
1982 statement that may clobber e. If, starting statement walk from the
1983 top of the basic block, a statement uses VUSE there can be no kill
1984 inbetween that use and the original statement that loaded {e, VUSE},
1985 so we can stop walking. */
1986 ref.base = NULL_TREE;
1987 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1988 {
1989 tree def_vuse, def_vdef;
1990 def = gsi_stmt (gsi);
1991 def_vuse = gimple_vuse (def);
1992 def_vdef = gimple_vdef (def);
1993
1994 /* Not a memory statement. */
1995 if (!def_vuse)
1996 continue;
1997
1998 /* Not a may-def. */
1999 if (!def_vdef)
2000 {
2001 /* A load with the same VUSE, we're done. */
2002 if (def_vuse == vuse)
2003 break;
2004
2005 continue;
2006 }
2007
2008 /* Init ref only if we really need it. */
2009 if (ref.base == NULL_TREE
2010 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2011 refx->operands))
2012 {
2013 res = true;
2014 break;
2015 }
2016 /* If the statement may clobber expr, it dies. */
2017 if (stmt_may_clobber_ref_p_1 (def, &ref))
2018 {
2019 res = true;
2020 break;
2021 }
2022 }
2023
2024 /* Remember the result. */
2025 if (!EXPR_DIES (block))
2026 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2027 bitmap_set_bit (EXPR_DIES (block), id * 2);
2028 if (res)
2029 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2030
2031 return res;
2032 }
2033
2034
2035 /* Determine if OP is valid in SET1 U SET2, which it is when the union
2036 contains its value-id. */
2037
2038 static bool
2039 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
2040 {
2041 if (op && TREE_CODE (op) == SSA_NAME)
2042 {
2043 unsigned int value_id = VN_INFO (op)->value_id;
2044 if (!bitmap_set_contains_value (set1, value_id)
2045 || (set2 && !bitmap_set_contains_value (set2, value_id)))
2046 return false;
2047 }
2048 return true;
2049 }
2050
2051 /* Determine if the expression EXPR is valid in SET1 U SET2.
2052 ONLY SET2 CAN BE NULL.
2053 This means that we have a leader for each part of the expression
2054 (if it consists of values), or the expression is an SSA_NAME.
2055 For loads/calls, we also see if the vuse is killed in this block. */
2056
2057 static bool
2058 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2059 basic_block block)
2060 {
2061 switch (expr->kind)
2062 {
2063 case NAME:
2064 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2065 case NARY:
2066 {
2067 unsigned int i;
2068 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2069 for (i = 0; i < nary->length; i++)
2070 if (!op_valid_in_sets (set1, set2, nary->op[i]))
2071 return false;
2072 return true;
2073 }
2074 break;
2075 case REFERENCE:
2076 {
2077 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2078 vn_reference_op_t vro;
2079 unsigned int i;
2080
2081 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2082 {
2083 if (!op_valid_in_sets (set1, set2, vro->op0)
2084 || !op_valid_in_sets (set1, set2, vro->op1)
2085 || !op_valid_in_sets (set1, set2, vro->op2))
2086 return false;
2087 }
2088 return true;
2089 }
2090 default:
2091 gcc_unreachable ();
2092 }
2093 }
2094
2095 /* Clean the set of expressions that are no longer valid in SET1 or
2096 SET2. This means expressions that are made up of values we have no
2097 leaders for in SET1 or SET2. This version is used for partial
2098 anticipation, which means it is not valid in either ANTIC_IN or
2099 PA_IN. */
2100
2101 static void
2102 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2103 {
2104 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2105 pre_expr expr;
2106 int i;
2107
2108 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2109 {
2110 if (!valid_in_sets (set1, set2, expr, block))
2111 bitmap_remove_from_set (set1, expr);
2112 }
2113 VEC_free (pre_expr, heap, exprs);
2114 }
2115
2116 /* Clean the set of expressions that are no longer valid in SET. This
2117 means expressions that are made up of values we have no leaders for
2118 in SET. */
2119
2120 static void
2121 clean (bitmap_set_t set, basic_block block)
2122 {
2123 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2124 pre_expr expr;
2125 int i;
2126
2127 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2128 {
2129 if (!valid_in_sets (set, NULL, expr, block))
2130 bitmap_remove_from_set (set, expr);
2131 }
2132 VEC_free (pre_expr, heap, exprs);
2133 }
2134
2135 /* Clean the set of expressions that are no longer valid in SET because
2136 they are clobbered in BLOCK or because they trap and may not be executed. */
2137
2138 static void
2139 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2140 {
2141 bitmap_iterator bi;
2142 unsigned i;
2143
2144 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2145 {
2146 pre_expr expr = expression_for_id (i);
2147 if (expr->kind == REFERENCE)
2148 {
2149 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2150 if (ref->vuse)
2151 {
2152 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2153 if (!gimple_nop_p (def_stmt)
2154 && ((gimple_bb (def_stmt) != block
2155 && !dominated_by_p (CDI_DOMINATORS,
2156 block, gimple_bb (def_stmt)))
2157 || (gimple_bb (def_stmt) == block
2158 && value_dies_in_block_x (expr, block))))
2159 bitmap_remove_from_set (set, expr);
2160 }
2161 }
2162 else if (expr->kind == NARY)
2163 {
2164 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2165 /* If the NARY may trap make sure the block does not contain
2166 a possible exit point.
2167 ??? This is overly conservative if we translate AVAIL_OUT
2168 as the available expression might be after the exit point. */
2169 if (BB_MAY_NOTRETURN (block)
2170 && vn_nary_may_trap (nary))
2171 bitmap_remove_from_set (set, expr);
2172 }
2173 }
2174 }
2175
2176 static sbitmap has_abnormal_preds;
2177
2178 /* List of blocks that may have changed during ANTIC computation and
2179 thus need to be iterated over. */
2180
2181 static sbitmap changed_blocks;
2182
2183 /* Decide whether to defer a block for a later iteration, or PHI
2184 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2185 should defer the block, and true if we processed it. */
2186
2187 static bool
2188 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2189 basic_block block, basic_block phiblock)
2190 {
2191 if (!BB_VISITED (phiblock))
2192 {
2193 SET_BIT (changed_blocks, block->index);
2194 BB_VISITED (block) = 0;
2195 BB_DEFERRED (block) = 1;
2196 return false;
2197 }
2198 else
2199 phi_translate_set (dest, source, block, phiblock);
2200 return true;
2201 }
2202
2203 /* Compute the ANTIC set for BLOCK.
2204
2205 If succs(BLOCK) > 1 then
2206 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2207 else if succs(BLOCK) == 1 then
2208 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2209
2210 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2211 */
2212
2213 static bool
2214 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2215 {
2216 bool changed = false;
2217 bitmap_set_t S, old, ANTIC_OUT;
2218 bitmap_iterator bi;
2219 unsigned int bii;
2220 edge e;
2221 edge_iterator ei;
2222
2223 old = ANTIC_OUT = S = NULL;
2224 BB_VISITED (block) = 1;
2225
2226 /* If any edges from predecessors are abnormal, antic_in is empty,
2227 so do nothing. */
2228 if (block_has_abnormal_pred_edge)
2229 goto maybe_dump_sets;
2230
2231 old = ANTIC_IN (block);
2232 ANTIC_OUT = bitmap_set_new ();
2233
2234 /* If the block has no successors, ANTIC_OUT is empty. */
2235 if (EDGE_COUNT (block->succs) == 0)
2236 ;
2237 /* If we have one successor, we could have some phi nodes to
2238 translate through. */
2239 else if (single_succ_p (block))
2240 {
2241 basic_block succ_bb = single_succ (block);
2242
2243 /* We trade iterations of the dataflow equations for having to
2244 phi translate the maximal set, which is incredibly slow
2245 (since the maximal set often has 300+ members, even when you
2246 have a small number of blocks).
2247 Basically, we defer the computation of ANTIC for this block
2248 until we have processed it's successor, which will inevitably
2249 have a *much* smaller set of values to phi translate once
2250 clean has been run on it.
2251 The cost of doing this is that we technically perform more
2252 iterations, however, they are lower cost iterations.
2253
2254 Timings for PRE on tramp3d-v4:
2255 without maximal set fix: 11 seconds
2256 with maximal set fix/without deferring: 26 seconds
2257 with maximal set fix/with deferring: 11 seconds
2258 */
2259
2260 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2261 block, succ_bb))
2262 {
2263 changed = true;
2264 goto maybe_dump_sets;
2265 }
2266 }
2267 /* If we have multiple successors, we take the intersection of all of
2268 them. Note that in the case of loop exit phi nodes, we may have
2269 phis to translate through. */
2270 else
2271 {
2272 VEC(basic_block, heap) * worklist;
2273 size_t i;
2274 basic_block bprime, first = NULL;
2275
2276 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2277 FOR_EACH_EDGE (e, ei, block->succs)
2278 {
2279 if (!first
2280 && BB_VISITED (e->dest))
2281 first = e->dest;
2282 else if (BB_VISITED (e->dest))
2283 VEC_quick_push (basic_block, worklist, e->dest);
2284 }
2285
2286 /* Of multiple successors we have to have visited one already. */
2287 if (!first)
2288 {
2289 SET_BIT (changed_blocks, block->index);
2290 BB_VISITED (block) = 0;
2291 BB_DEFERRED (block) = 1;
2292 changed = true;
2293 VEC_free (basic_block, heap, worklist);
2294 goto maybe_dump_sets;
2295 }
2296
2297 if (!gimple_seq_empty_p (phi_nodes (first)))
2298 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2299 else
2300 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2301
2302 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2303 {
2304 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2305 {
2306 bitmap_set_t tmp = bitmap_set_new ();
2307 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2308 bitmap_set_and (ANTIC_OUT, tmp);
2309 bitmap_set_free (tmp);
2310 }
2311 else
2312 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2313 }
2314 VEC_free (basic_block, heap, worklist);
2315 }
2316
2317 /* Prune expressions that are clobbered in block and thus become
2318 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2319 prune_clobbered_mems (ANTIC_OUT, block);
2320
2321 /* Generate ANTIC_OUT - TMP_GEN. */
2322 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2323
2324 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2325 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2326 TMP_GEN (block));
2327
2328 /* Then union in the ANTIC_OUT - TMP_GEN values,
2329 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2330 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2331 bitmap_value_insert_into_set (ANTIC_IN (block),
2332 expression_for_id (bii));
2333
2334 clean (ANTIC_IN (block), block);
2335
2336 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2337 {
2338 changed = true;
2339 SET_BIT (changed_blocks, block->index);
2340 FOR_EACH_EDGE (e, ei, block->preds)
2341 SET_BIT (changed_blocks, e->src->index);
2342 }
2343 else
2344 RESET_BIT (changed_blocks, block->index);
2345
2346 maybe_dump_sets:
2347 if (dump_file && (dump_flags & TDF_DETAILS))
2348 {
2349 if (!BB_DEFERRED (block) || BB_VISITED (block))
2350 {
2351 if (ANTIC_OUT)
2352 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2353
2354 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2355 block->index);
2356
2357 if (S)
2358 print_bitmap_set (dump_file, S, "S", block->index);
2359 }
2360 else
2361 {
2362 fprintf (dump_file,
2363 "Block %d was deferred for a future iteration.\n",
2364 block->index);
2365 }
2366 }
2367 if (old)
2368 bitmap_set_free (old);
2369 if (S)
2370 bitmap_set_free (S);
2371 if (ANTIC_OUT)
2372 bitmap_set_free (ANTIC_OUT);
2373 return changed;
2374 }
2375
2376 /* Compute PARTIAL_ANTIC for BLOCK.
2377
2378 If succs(BLOCK) > 1 then
2379 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2380 in ANTIC_OUT for all succ(BLOCK)
2381 else if succs(BLOCK) == 1 then
2382 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2383
2384 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2385 - ANTIC_IN[BLOCK])
2386
2387 */
2388 static bool
2389 compute_partial_antic_aux (basic_block block,
2390 bool block_has_abnormal_pred_edge)
2391 {
2392 bool changed = false;
2393 bitmap_set_t old_PA_IN;
2394 bitmap_set_t PA_OUT;
2395 edge e;
2396 edge_iterator ei;
2397 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2398
2399 old_PA_IN = PA_OUT = NULL;
2400
2401 /* If any edges from predecessors are abnormal, antic_in is empty,
2402 so do nothing. */
2403 if (block_has_abnormal_pred_edge)
2404 goto maybe_dump_sets;
2405
2406 /* If there are too many partially anticipatable values in the
2407 block, phi_translate_set can take an exponential time: stop
2408 before the translation starts. */
2409 if (max_pa
2410 && single_succ_p (block)
2411 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2412 goto maybe_dump_sets;
2413
2414 old_PA_IN = PA_IN (block);
2415 PA_OUT = bitmap_set_new ();
2416
2417 /* If the block has no successors, ANTIC_OUT is empty. */
2418 if (EDGE_COUNT (block->succs) == 0)
2419 ;
2420 /* If we have one successor, we could have some phi nodes to
2421 translate through. Note that we can't phi translate across DFS
2422 back edges in partial antic, because it uses a union operation on
2423 the successors. For recurrences like IV's, we will end up
2424 generating a new value in the set on each go around (i + 3 (VH.1)
2425 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2426 else if (single_succ_p (block))
2427 {
2428 basic_block succ = single_succ (block);
2429 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2430 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2431 }
2432 /* If we have multiple successors, we take the union of all of
2433 them. */
2434 else
2435 {
2436 VEC(basic_block, heap) * worklist;
2437 size_t i;
2438 basic_block bprime;
2439
2440 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2441 FOR_EACH_EDGE (e, ei, block->succs)
2442 {
2443 if (e->flags & EDGE_DFS_BACK)
2444 continue;
2445 VEC_quick_push (basic_block, worklist, e->dest);
2446 }
2447 if (VEC_length (basic_block, worklist) > 0)
2448 {
2449 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2450 {
2451 unsigned int i;
2452 bitmap_iterator bi;
2453
2454 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2455 bitmap_value_insert_into_set (PA_OUT,
2456 expression_for_id (i));
2457 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2458 {
2459 bitmap_set_t pa_in = bitmap_set_new ();
2460 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2461 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2462 bitmap_value_insert_into_set (PA_OUT,
2463 expression_for_id (i));
2464 bitmap_set_free (pa_in);
2465 }
2466 else
2467 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2468 bitmap_value_insert_into_set (PA_OUT,
2469 expression_for_id (i));
2470 }
2471 }
2472 VEC_free (basic_block, heap, worklist);
2473 }
2474
2475 /* Prune expressions that are clobbered in block and thus become
2476 invalid if translated from PA_OUT to PA_IN. */
2477 prune_clobbered_mems (PA_OUT, block);
2478
2479 /* PA_IN starts with PA_OUT - TMP_GEN.
2480 Then we subtract things from ANTIC_IN. */
2481 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2482
2483 /* For partial antic, we want to put back in the phi results, since
2484 we will properly avoid making them partially antic over backedges. */
2485 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2486 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2487
2488 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2489 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2490
2491 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2492
2493 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2494 {
2495 changed = true;
2496 SET_BIT (changed_blocks, block->index);
2497 FOR_EACH_EDGE (e, ei, block->preds)
2498 SET_BIT (changed_blocks, e->src->index);
2499 }
2500 else
2501 RESET_BIT (changed_blocks, block->index);
2502
2503 maybe_dump_sets:
2504 if (dump_file && (dump_flags & TDF_DETAILS))
2505 {
2506 if (PA_OUT)
2507 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2508
2509 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2510 }
2511 if (old_PA_IN)
2512 bitmap_set_free (old_PA_IN);
2513 if (PA_OUT)
2514 bitmap_set_free (PA_OUT);
2515 return changed;
2516 }
2517
2518 /* Compute ANTIC and partial ANTIC sets. */
2519
2520 static void
2521 compute_antic (void)
2522 {
2523 bool changed = true;
2524 int num_iterations = 0;
2525 basic_block block;
2526 int i;
2527
2528 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2529 We pre-build the map of blocks with incoming abnormal edges here. */
2530 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2531 sbitmap_zero (has_abnormal_preds);
2532
2533 FOR_EACH_BB (block)
2534 {
2535 edge_iterator ei;
2536 edge e;
2537
2538 FOR_EACH_EDGE (e, ei, block->preds)
2539 {
2540 e->flags &= ~EDGE_DFS_BACK;
2541 if (e->flags & EDGE_ABNORMAL)
2542 {
2543 SET_BIT (has_abnormal_preds, block->index);
2544 break;
2545 }
2546 }
2547
2548 BB_VISITED (block) = 0;
2549 BB_DEFERRED (block) = 0;
2550
2551 /* While we are here, give empty ANTIC_IN sets to each block. */
2552 ANTIC_IN (block) = bitmap_set_new ();
2553 PA_IN (block) = bitmap_set_new ();
2554 }
2555
2556 /* At the exit block we anticipate nothing. */
2557 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2558 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2559 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2560
2561 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2562 sbitmap_ones (changed_blocks);
2563 while (changed)
2564 {
2565 if (dump_file && (dump_flags & TDF_DETAILS))
2566 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2567 /* ??? We need to clear our PHI translation cache here as the
2568 ANTIC sets shrink and we restrict valid translations to
2569 those having operands with leaders in ANTIC. Same below
2570 for PA ANTIC computation. */
2571 num_iterations++;
2572 changed = false;
2573 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2574 {
2575 if (TEST_BIT (changed_blocks, postorder[i]))
2576 {
2577 basic_block block = BASIC_BLOCK (postorder[i]);
2578 changed |= compute_antic_aux (block,
2579 TEST_BIT (has_abnormal_preds,
2580 block->index));
2581 }
2582 }
2583 /* Theoretically possible, but *highly* unlikely. */
2584 gcc_checking_assert (num_iterations < 500);
2585 }
2586
2587 statistics_histogram_event (cfun, "compute_antic iterations",
2588 num_iterations);
2589
2590 if (do_partial_partial)
2591 {
2592 sbitmap_ones (changed_blocks);
2593 mark_dfs_back_edges ();
2594 num_iterations = 0;
2595 changed = true;
2596 while (changed)
2597 {
2598 if (dump_file && (dump_flags & TDF_DETAILS))
2599 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2600 num_iterations++;
2601 changed = false;
2602 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2603 {
2604 if (TEST_BIT (changed_blocks, postorder[i]))
2605 {
2606 basic_block block = BASIC_BLOCK (postorder[i]);
2607 changed
2608 |= compute_partial_antic_aux (block,
2609 TEST_BIT (has_abnormal_preds,
2610 block->index));
2611 }
2612 }
2613 /* Theoretically possible, but *highly* unlikely. */
2614 gcc_checking_assert (num_iterations < 500);
2615 }
2616 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2617 num_iterations);
2618 }
2619 sbitmap_free (has_abnormal_preds);
2620 sbitmap_free (changed_blocks);
2621 }
2622
2623 /* Return true if we can value number the call in STMT. This is true
2624 if we have a pure or constant call to a real function. */
2625
2626 static bool
2627 can_value_number_call (gimple stmt)
2628 {
2629 if (gimple_call_internal_p (stmt))
2630 return false;
2631 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2632 return true;
2633 return false;
2634 }
2635
2636 /* Return true if OP is a tree which we can perform PRE on.
2637 This may not match the operations we can value number, but in
2638 a perfect world would. */
2639
2640 static bool
2641 can_PRE_operation (tree op)
2642 {
2643 return UNARY_CLASS_P (op)
2644 || BINARY_CLASS_P (op)
2645 || COMPARISON_CLASS_P (op)
2646 || TREE_CODE (op) == MEM_REF
2647 || TREE_CODE (op) == COMPONENT_REF
2648 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2649 || TREE_CODE (op) == CALL_EXPR
2650 || TREE_CODE (op) == ARRAY_REF;
2651 }
2652
2653
2654 /* Inserted expressions are placed onto this worklist, which is used
2655 for performing quick dead code elimination of insertions we made
2656 that didn't turn out to be necessary. */
2657 static bitmap inserted_exprs;
2658
2659 /* Pool allocated fake store expressions are placed onto this
2660 worklist, which, after performing dead code elimination, is walked
2661 to see which expressions need to be put into GC'able memory */
2662 static VEC(gimple, heap) *need_creation;
2663
2664 /* The actual worker for create_component_ref_by_pieces. */
2665
2666 static tree
2667 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2668 unsigned int *operand, gimple_seq *stmts,
2669 gimple domstmt)
2670 {
2671 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2672 *operand);
2673 tree genop;
2674 ++*operand;
2675 switch (currop->opcode)
2676 {
2677 case CALL_EXPR:
2678 {
2679 tree folded, sc = NULL_TREE;
2680 unsigned int nargs = 0;
2681 tree fn, *args;
2682 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2683 fn = currop->op0;
2684 else
2685 {
2686 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2687 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2688 if (!fn)
2689 return NULL_TREE;
2690 }
2691 if (currop->op1)
2692 {
2693 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2694 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2695 if (!sc)
2696 return NULL_TREE;
2697 }
2698 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2699 ref->operands) - 1);
2700 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2701 {
2702 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2703 operand, stmts,
2704 domstmt);
2705 if (!args[nargs])
2706 {
2707 free (args);
2708 return NULL_TREE;
2709 }
2710 nargs++;
2711 }
2712 folded = build_call_array (currop->type,
2713 (TREE_CODE (fn) == FUNCTION_DECL
2714 ? build_fold_addr_expr (fn) : fn),
2715 nargs, args);
2716 free (args);
2717 if (sc)
2718 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2719 return folded;
2720 }
2721 break;
2722 case MEM_REF:
2723 {
2724 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2725 stmts, domstmt);
2726 tree offset = currop->op0;
2727 if (!baseop)
2728 return NULL_TREE;
2729 if (TREE_CODE (baseop) == ADDR_EXPR
2730 && handled_component_p (TREE_OPERAND (baseop, 0)))
2731 {
2732 HOST_WIDE_INT off;
2733 tree base;
2734 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2735 &off);
2736 gcc_assert (base);
2737 offset = int_const_binop (PLUS_EXPR, offset,
2738 build_int_cst (TREE_TYPE (offset),
2739 off));
2740 baseop = build_fold_addr_expr (base);
2741 }
2742 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2743 }
2744 break;
2745 case TARGET_MEM_REF:
2746 {
2747 pre_expr op0expr, op1expr;
2748 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2749 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2750 ++*operand);
2751 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2752 stmts, domstmt);
2753 if (!baseop)
2754 return NULL_TREE;
2755 if (currop->op0)
2756 {
2757 op0expr = get_or_alloc_expr_for (currop->op0);
2758 genop0 = find_or_generate_expression (block, op0expr,
2759 stmts, domstmt);
2760 if (!genop0)
2761 return NULL_TREE;
2762 }
2763 if (nextop->op0)
2764 {
2765 op1expr = get_or_alloc_expr_for (nextop->op0);
2766 genop1 = find_or_generate_expression (block, op1expr,
2767 stmts, domstmt);
2768 if (!genop1)
2769 return NULL_TREE;
2770 }
2771 return build5 (TARGET_MEM_REF, currop->type,
2772 baseop, currop->op2, genop0, currop->op1, genop1);
2773 }
2774 break;
2775 case ADDR_EXPR:
2776 if (currop->op0)
2777 {
2778 gcc_assert (is_gimple_min_invariant (currop->op0));
2779 return currop->op0;
2780 }
2781 /* Fallthrough. */
2782 case REALPART_EXPR:
2783 case IMAGPART_EXPR:
2784 case VIEW_CONVERT_EXPR:
2785 {
2786 tree folded;
2787 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2788 operand,
2789 stmts, domstmt);
2790 if (!genop0)
2791 return NULL_TREE;
2792 folded = fold_build1 (currop->opcode, currop->type,
2793 genop0);
2794 return folded;
2795 }
2796 break;
2797 case WITH_SIZE_EXPR:
2798 {
2799 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2800 stmts, domstmt);
2801 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2802 tree genop1;
2803
2804 if (!genop0)
2805 return NULL_TREE;
2806
2807 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2808 if (!genop1)
2809 return NULL_TREE;
2810
2811 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2812 }
2813 break;
2814 case BIT_FIELD_REF:
2815 {
2816 tree folded;
2817 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2818 stmts, domstmt);
2819 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2820 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2821 tree genop1;
2822 tree genop2;
2823
2824 if (!genop0)
2825 return NULL_TREE;
2826 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2827 if (!genop1)
2828 return NULL_TREE;
2829 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2830 if (!genop2)
2831 return NULL_TREE;
2832 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2833 genop2);
2834 return folded;
2835 }
2836
2837 /* For array ref vn_reference_op's, operand 1 of the array ref
2838 is op0 of the reference op and operand 3 of the array ref is
2839 op1. */
2840 case ARRAY_RANGE_REF:
2841 case ARRAY_REF:
2842 {
2843 tree genop0;
2844 tree genop1 = currop->op0;
2845 pre_expr op1expr;
2846 tree genop2 = currop->op1;
2847 pre_expr op2expr;
2848 tree genop3 = currop->op2;
2849 pre_expr op3expr;
2850 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2851 stmts, domstmt);
2852 if (!genop0)
2853 return NULL_TREE;
2854 op1expr = get_or_alloc_expr_for (genop1);
2855 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2856 if (!genop1)
2857 return NULL_TREE;
2858 if (genop2)
2859 {
2860 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2861 /* Drop zero minimum index if redundant. */
2862 if (integer_zerop (genop2)
2863 && (!domain_type
2864 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2865 genop2 = NULL_TREE;
2866 else
2867 {
2868 op2expr = get_or_alloc_expr_for (genop2);
2869 genop2 = find_or_generate_expression (block, op2expr, stmts,
2870 domstmt);
2871 if (!genop2)
2872 return NULL_TREE;
2873 }
2874 }
2875 if (genop3)
2876 {
2877 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2878 /* We can't always put a size in units of the element alignment
2879 here as the element alignment may be not visible. See
2880 PR43783. Simply drop the element size for constant
2881 sizes. */
2882 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2883 genop3 = NULL_TREE;
2884 else
2885 {
2886 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2887 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2888 op3expr = get_or_alloc_expr_for (genop3);
2889 genop3 = find_or_generate_expression (block, op3expr, stmts,
2890 domstmt);
2891 if (!genop3)
2892 return NULL_TREE;
2893 }
2894 }
2895 return build4 (currop->opcode, currop->type, genop0, genop1,
2896 genop2, genop3);
2897 }
2898 case COMPONENT_REF:
2899 {
2900 tree op0;
2901 tree op1;
2902 tree genop2 = currop->op1;
2903 pre_expr op2expr;
2904 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2905 stmts, domstmt);
2906 if (!op0)
2907 return NULL_TREE;
2908 /* op1 should be a FIELD_DECL, which are represented by
2909 themselves. */
2910 op1 = currop->op0;
2911 if (genop2)
2912 {
2913 op2expr = get_or_alloc_expr_for (genop2);
2914 genop2 = find_or_generate_expression (block, op2expr, stmts,
2915 domstmt);
2916 if (!genop2)
2917 return NULL_TREE;
2918 }
2919
2920 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2921 genop2);
2922 }
2923 break;
2924 case SSA_NAME:
2925 {
2926 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2927 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2928 return genop;
2929 }
2930 case STRING_CST:
2931 case INTEGER_CST:
2932 case COMPLEX_CST:
2933 case VECTOR_CST:
2934 case REAL_CST:
2935 case CONSTRUCTOR:
2936 case VAR_DECL:
2937 case PARM_DECL:
2938 case CONST_DECL:
2939 case RESULT_DECL:
2940 case FUNCTION_DECL:
2941 return currop->op0;
2942
2943 default:
2944 gcc_unreachable ();
2945 }
2946 }
2947
2948 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2949 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2950 trying to rename aggregates into ssa form directly, which is a no no.
2951
2952 Thus, this routine doesn't create temporaries, it just builds a
2953 single access expression for the array, calling
2954 find_or_generate_expression to build the innermost pieces.
2955
2956 This function is a subroutine of create_expression_by_pieces, and
2957 should not be called on it's own unless you really know what you
2958 are doing. */
2959
2960 static tree
2961 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2962 gimple_seq *stmts, gimple domstmt)
2963 {
2964 unsigned int op = 0;
2965 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2966 }
2967
2968 /* Find a leader for an expression, or generate one using
2969 create_expression_by_pieces if it's ANTIC but
2970 complex.
2971 BLOCK is the basic_block we are looking for leaders in.
2972 EXPR is the expression to find a leader or generate for.
2973 STMTS is the statement list to put the inserted expressions on.
2974 Returns the SSA_NAME of the LHS of the generated expression or the
2975 leader.
2976 DOMSTMT if non-NULL is a statement that should be dominated by
2977 all uses in the generated expression. If DOMSTMT is non-NULL this
2978 routine can fail and return NULL_TREE. Otherwise it will assert
2979 on failure. */
2980
2981 static tree
2982 find_or_generate_expression (basic_block block, pre_expr expr,
2983 gimple_seq *stmts, gimple domstmt)
2984 {
2985 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2986 get_expr_value_id (expr), domstmt);
2987 tree genop = NULL;
2988 if (leader)
2989 {
2990 if (leader->kind == NAME)
2991 genop = PRE_EXPR_NAME (leader);
2992 else if (leader->kind == CONSTANT)
2993 genop = PRE_EXPR_CONSTANT (leader);
2994 }
2995
2996 /* If it's still NULL, it must be a complex expression, so generate
2997 it recursively. Not so if inserting expressions for values generated
2998 by SCCVN. */
2999 if (genop == NULL
3000 && !domstmt)
3001 {
3002 bitmap_set_t exprset;
3003 unsigned int lookfor = get_expr_value_id (expr);
3004 bool handled = false;
3005 bitmap_iterator bi;
3006 unsigned int i;
3007
3008 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3009 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3010 {
3011 pre_expr temp = expression_for_id (i);
3012 if (temp->kind != NAME)
3013 {
3014 handled = true;
3015 genop = create_expression_by_pieces (block, temp, stmts,
3016 domstmt,
3017 get_expr_type (expr));
3018 break;
3019 }
3020 }
3021 if (!handled && domstmt)
3022 return NULL_TREE;
3023
3024 gcc_assert (handled);
3025 }
3026 return genop;
3027 }
3028
3029 #define NECESSARY GF_PLF_1
3030
3031 /* Create an expression in pieces, so that we can handle very complex
3032 expressions that may be ANTIC, but not necessary GIMPLE.
3033 BLOCK is the basic block the expression will be inserted into,
3034 EXPR is the expression to insert (in value form)
3035 STMTS is a statement list to append the necessary insertions into.
3036
3037 This function will die if we hit some value that shouldn't be
3038 ANTIC but is (IE there is no leader for it, or its components).
3039 This function may also generate expressions that are themselves
3040 partially or fully redundant. Those that are will be either made
3041 fully redundant during the next iteration of insert (for partially
3042 redundant ones), or eliminated by eliminate (for fully redundant
3043 ones).
3044
3045 If DOMSTMT is non-NULL then we make sure that all uses in the
3046 expressions dominate that statement. In this case the function
3047 can return NULL_TREE to signal failure. */
3048
3049 static tree
3050 create_expression_by_pieces (basic_block block, pre_expr expr,
3051 gimple_seq *stmts, gimple domstmt, tree type)
3052 {
3053 tree temp, name;
3054 tree folded;
3055 gimple_seq forced_stmts = NULL;
3056 unsigned int value_id;
3057 gimple_stmt_iterator gsi;
3058 tree exprtype = type ? type : get_expr_type (expr);
3059 pre_expr nameexpr;
3060 gimple newstmt;
3061
3062 switch (expr->kind)
3063 {
3064 /* We may hit the NAME/CONSTANT case if we have to convert types
3065 that value numbering saw through. */
3066 case NAME:
3067 folded = PRE_EXPR_NAME (expr);
3068 break;
3069 case CONSTANT:
3070 folded = PRE_EXPR_CONSTANT (expr);
3071 break;
3072 case REFERENCE:
3073 {
3074 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3075 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3076 }
3077 break;
3078 case NARY:
3079 {
3080 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3081 tree genop[4];
3082 unsigned i;
3083 for (i = 0; i < nary->length; ++i)
3084 {
3085 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3086 genop[i] = find_or_generate_expression (block, op,
3087 stmts, domstmt);
3088 if (!genop[i])
3089 return NULL_TREE;
3090 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3091 may have conversions stripped. */
3092 if (nary->opcode == POINTER_PLUS_EXPR)
3093 {
3094 if (i == 0)
3095 genop[i] = fold_convert (nary->type, genop[i]);
3096 else if (i == 1)
3097 genop[i] = convert_to_ptrofftype (genop[i]);
3098 }
3099 else
3100 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3101 }
3102 if (nary->opcode == CONSTRUCTOR)
3103 {
3104 VEC(constructor_elt,gc) *elts = NULL;
3105 for (i = 0; i < nary->length; ++i)
3106 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3107 folded = build_constructor (nary->type, elts);
3108 }
3109 else
3110 {
3111 switch (nary->length)
3112 {
3113 case 1:
3114 folded = fold_build1 (nary->opcode, nary->type,
3115 genop[0]);
3116 break;
3117 case 2:
3118 folded = fold_build2 (nary->opcode, nary->type,
3119 genop[0], genop[1]);
3120 break;
3121 case 3:
3122 folded = fold_build3 (nary->opcode, nary->type,
3123 genop[0], genop[1], genop[3]);
3124 break;
3125 default:
3126 gcc_unreachable ();
3127 }
3128 }
3129 }
3130 break;
3131 default:
3132 return NULL_TREE;
3133 }
3134
3135 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3136 folded = fold_convert (exprtype, folded);
3137
3138 /* Force the generated expression to be a sequence of GIMPLE
3139 statements.
3140 We have to call unshare_expr because force_gimple_operand may
3141 modify the tree we pass to it. */
3142 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3143 false, NULL);
3144
3145 /* If we have any intermediate expressions to the value sets, add them
3146 to the value sets and chain them in the instruction stream. */
3147 if (forced_stmts)
3148 {
3149 gsi = gsi_start (forced_stmts);
3150 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3151 {
3152 gimple stmt = gsi_stmt (gsi);
3153 tree forcedname = gimple_get_lhs (stmt);
3154 pre_expr nameexpr;
3155
3156 if (TREE_CODE (forcedname) == SSA_NAME)
3157 {
3158 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3159 VN_INFO_GET (forcedname)->valnum = forcedname;
3160 VN_INFO (forcedname)->value_id = get_next_value_id ();
3161 nameexpr = get_or_alloc_expr_for_name (forcedname);
3162 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3163 if (!in_fre)
3164 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3165 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3166 }
3167 mark_symbols_for_renaming (stmt);
3168 }
3169 gimple_seq_add_seq (stmts, forced_stmts);
3170 }
3171
3172 /* Build and insert the assignment of the end result to the temporary
3173 that we will return. */
3174 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3175 pretemp = create_tmp_reg (exprtype, "pretmp");
3176
3177 temp = pretemp;
3178 add_referenced_var (temp);
3179
3180 newstmt = gimple_build_assign (temp, folded);
3181 name = make_ssa_name (temp, newstmt);
3182 gimple_assign_set_lhs (newstmt, name);
3183 gimple_set_plf (newstmt, NECESSARY, false);
3184
3185 gimple_seq_add_stmt (stmts, newstmt);
3186 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3187
3188 /* All the symbols in NEWEXPR should be put into SSA form. */
3189 mark_symbols_for_renaming (newstmt);
3190
3191 /* Fold the last statement. */
3192 gsi = gsi_last (*stmts);
3193 if (fold_stmt_inplace (&gsi))
3194 update_stmt (gsi_stmt (gsi));
3195
3196 /* Add a value number to the temporary.
3197 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3198 we are creating the expression by pieces, and this particular piece of
3199 the expression may have been represented. There is no harm in replacing
3200 here. */
3201 VN_INFO_GET (name)->valnum = name;
3202 value_id = get_expr_value_id (expr);
3203 VN_INFO (name)->value_id = value_id;
3204 nameexpr = get_or_alloc_expr_for_name (name);
3205 add_to_value (value_id, nameexpr);
3206 if (NEW_SETS (block))
3207 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3208 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3209
3210 pre_stats.insertions++;
3211 if (dump_file && (dump_flags & TDF_DETAILS))
3212 {
3213 fprintf (dump_file, "Inserted ");
3214 print_gimple_stmt (dump_file, newstmt, 0, 0);
3215 fprintf (dump_file, " in predecessor %d\n", block->index);
3216 }
3217
3218 return name;
3219 }
3220
3221
3222 /* Returns true if we want to inhibit the insertions of PHI nodes
3223 for the given EXPR for basic block BB (a member of a loop).
3224 We want to do this, when we fear that the induction variable we
3225 create might inhibit vectorization. */
3226
3227 static bool
3228 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3229 {
3230 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3231 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3232 vn_reference_op_t op;
3233 unsigned i;
3234
3235 /* If we aren't going to vectorize we don't inhibit anything. */
3236 if (!flag_tree_vectorize)
3237 return false;
3238
3239 /* Otherwise we inhibit the insertion when the address of the
3240 memory reference is a simple induction variable. In other
3241 cases the vectorizer won't do anything anyway (either it's
3242 loop invariant or a complicated expression). */
3243 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3244 {
3245 switch (op->opcode)
3246 {
3247 case CALL_EXPR:
3248 /* Calls are not a problem. */
3249 return false;
3250
3251 case ARRAY_REF:
3252 case ARRAY_RANGE_REF:
3253 if (TREE_CODE (op->op0) != SSA_NAME)
3254 break;
3255 /* Fallthru. */
3256 case SSA_NAME:
3257 {
3258 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3259 affine_iv iv;
3260 /* Default defs are loop invariant. */
3261 if (!defbb)
3262 break;
3263 /* Defined outside this loop, also loop invariant. */
3264 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3265 break;
3266 /* If it's a simple induction variable inhibit insertion,
3267 the vectorizer might be interested in this one. */
3268 if (simple_iv (bb->loop_father, bb->loop_father,
3269 op->op0, &iv, true))
3270 return true;
3271 /* No simple IV, vectorizer can't do anything, hence no
3272 reason to inhibit the transformation for this operand. */
3273 break;
3274 }
3275 default:
3276 break;
3277 }
3278 }
3279 return false;
3280 }
3281
3282 /* Insert the to-be-made-available values of expression EXPRNUM for each
3283 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3284 merge the result with a phi node, given the same value number as
3285 NODE. Return true if we have inserted new stuff. */
3286
3287 static bool
3288 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3289 pre_expr *avail)
3290 {
3291 pre_expr expr = expression_for_id (exprnum);
3292 pre_expr newphi;
3293 unsigned int val = get_expr_value_id (expr);
3294 edge pred;
3295 bool insertions = false;
3296 bool nophi = false;
3297 basic_block bprime;
3298 pre_expr eprime;
3299 edge_iterator ei;
3300 tree type = get_expr_type (expr);
3301 tree temp;
3302 gimple phi;
3303
3304 /* Make sure we aren't creating an induction variable. */
3305 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3306 {
3307 bool firstinsideloop = false;
3308 bool secondinsideloop = false;
3309 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3310 EDGE_PRED (block, 0)->src);
3311 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3312 EDGE_PRED (block, 1)->src);
3313 /* Induction variables only have one edge inside the loop. */
3314 if ((firstinsideloop ^ secondinsideloop)
3315 && (expr->kind != REFERENCE
3316 || inhibit_phi_insertion (block, expr)))
3317 {
3318 if (dump_file && (dump_flags & TDF_DETAILS))
3319 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3320 nophi = true;
3321 }
3322 }
3323
3324 /* Make the necessary insertions. */
3325 FOR_EACH_EDGE (pred, ei, block->preds)
3326 {
3327 gimple_seq stmts = NULL;
3328 tree builtexpr;
3329 bprime = pred->src;
3330 eprime = avail[bprime->index];
3331
3332 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3333 {
3334 builtexpr = create_expression_by_pieces (bprime,
3335 eprime,
3336 &stmts, NULL,
3337 type);
3338 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3339 gsi_insert_seq_on_edge (pred, stmts);
3340 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3341 insertions = true;
3342 }
3343 else if (eprime->kind == CONSTANT)
3344 {
3345 /* Constants may not have the right type, fold_convert
3346 should give us back a constant with the right type.
3347 */
3348 tree constant = PRE_EXPR_CONSTANT (eprime);
3349 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3350 {
3351 tree builtexpr = fold_convert (type, constant);
3352 if (!is_gimple_min_invariant (builtexpr))
3353 {
3354 tree forcedexpr = force_gimple_operand (builtexpr,
3355 &stmts, true,
3356 NULL);
3357 if (!is_gimple_min_invariant (forcedexpr))
3358 {
3359 if (forcedexpr != builtexpr)
3360 {
3361 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3362 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3363 }
3364 if (stmts)
3365 {
3366 gimple_stmt_iterator gsi;
3367 gsi = gsi_start (stmts);
3368 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3369 {
3370 gimple stmt = gsi_stmt (gsi);
3371 tree lhs = gimple_get_lhs (stmt);
3372 if (TREE_CODE (lhs) == SSA_NAME)
3373 bitmap_set_bit (inserted_exprs,
3374 SSA_NAME_VERSION (lhs));
3375 gimple_set_plf (stmt, NECESSARY, false);
3376 }
3377 gsi_insert_seq_on_edge (pred, stmts);
3378 }
3379 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3380 }
3381 }
3382 else
3383 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3384 }
3385 }
3386 else if (eprime->kind == NAME)
3387 {
3388 /* We may have to do a conversion because our value
3389 numbering can look through types in certain cases, but
3390 our IL requires all operands of a phi node have the same
3391 type. */
3392 tree name = PRE_EXPR_NAME (eprime);
3393 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3394 {
3395 tree builtexpr;
3396 tree forcedexpr;
3397 builtexpr = fold_convert (type, name);
3398 forcedexpr = force_gimple_operand (builtexpr,
3399 &stmts, true,
3400 NULL);
3401
3402 if (forcedexpr != name)
3403 {
3404 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3405 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3406 }
3407
3408 if (stmts)
3409 {
3410 gimple_stmt_iterator gsi;
3411 gsi = gsi_start (stmts);
3412 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3413 {
3414 gimple stmt = gsi_stmt (gsi);
3415 tree lhs = gimple_get_lhs (stmt);
3416 if (TREE_CODE (lhs) == SSA_NAME)
3417 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3418 gimple_set_plf (stmt, NECESSARY, false);
3419 }
3420 gsi_insert_seq_on_edge (pred, stmts);
3421 }
3422 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3423 }
3424 }
3425 }
3426 /* If we didn't want a phi node, and we made insertions, we still have
3427 inserted new stuff, and thus return true. If we didn't want a phi node,
3428 and didn't make insertions, we haven't added anything new, so return
3429 false. */
3430 if (nophi && insertions)
3431 return true;
3432 else if (nophi && !insertions)
3433 return false;
3434
3435 /* Now build a phi for the new variable. */
3436 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3437 prephitemp = create_tmp_var (type, "prephitmp");
3438
3439 temp = prephitemp;
3440 add_referenced_var (temp);
3441
3442 if (TREE_CODE (type) == COMPLEX_TYPE
3443 || TREE_CODE (type) == VECTOR_TYPE)
3444 DECL_GIMPLE_REG_P (temp) = 1;
3445 phi = create_phi_node (temp, block);
3446
3447 gimple_set_plf (phi, NECESSARY, false);
3448 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3449 VN_INFO (gimple_phi_result (phi))->value_id = val;
3450 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3451 FOR_EACH_EDGE (pred, ei, block->preds)
3452 {
3453 pre_expr ae = avail[pred->src->index];
3454 gcc_assert (get_expr_type (ae) == type
3455 || useless_type_conversion_p (type, get_expr_type (ae)));
3456 if (ae->kind == CONSTANT)
3457 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3458 else
3459 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3460 UNKNOWN_LOCATION);
3461 }
3462
3463 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3464 add_to_value (val, newphi);
3465
3466 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3467 this insertion, since we test for the existence of this value in PHI_GEN
3468 before proceeding with the partial redundancy checks in insert_aux.
3469
3470 The value may exist in AVAIL_OUT, in particular, it could be represented
3471 by the expression we are trying to eliminate, in which case we want the
3472 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3473 inserted there.
3474
3475 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3476 this block, because if it did, it would have existed in our dominator's
3477 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3478 */
3479
3480 bitmap_insert_into_set (PHI_GEN (block), newphi);
3481 bitmap_value_replace_in_set (AVAIL_OUT (block),
3482 newphi);
3483 bitmap_insert_into_set (NEW_SETS (block),
3484 newphi);
3485
3486 if (dump_file && (dump_flags & TDF_DETAILS))
3487 {
3488 fprintf (dump_file, "Created phi ");
3489 print_gimple_stmt (dump_file, phi, 0, 0);
3490 fprintf (dump_file, " in block %d\n", block->index);
3491 }
3492 pre_stats.phis++;
3493 return true;
3494 }
3495
3496
3497
3498 /* Perform insertion of partially redundant values.
3499 For BLOCK, do the following:
3500 1. Propagate the NEW_SETS of the dominator into the current block.
3501 If the block has multiple predecessors,
3502 2a. Iterate over the ANTIC expressions for the block to see if
3503 any of them are partially redundant.
3504 2b. If so, insert them into the necessary predecessors to make
3505 the expression fully redundant.
3506 2c. Insert a new PHI merging the values of the predecessors.
3507 2d. Insert the new PHI, and the new expressions, into the
3508 NEW_SETS set.
3509 3. Recursively call ourselves on the dominator children of BLOCK.
3510
3511 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3512 do_regular_insertion and do_partial_insertion.
3513
3514 */
3515
3516 static bool
3517 do_regular_insertion (basic_block block, basic_block dom)
3518 {
3519 bool new_stuff = false;
3520 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3521 pre_expr expr;
3522 int i;
3523
3524 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3525 {
3526 if (expr->kind != NAME)
3527 {
3528 pre_expr *avail;
3529 unsigned int val;
3530 bool by_some = false;
3531 bool cant_insert = false;
3532 bool all_same = true;
3533 pre_expr first_s = NULL;
3534 edge pred;
3535 basic_block bprime;
3536 pre_expr eprime = NULL;
3537 edge_iterator ei;
3538 pre_expr edoubleprime = NULL;
3539 bool do_insertion = false;
3540
3541 val = get_expr_value_id (expr);
3542 if (bitmap_set_contains_value (PHI_GEN (block), val))
3543 continue;
3544 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3545 {
3546 if (dump_file && (dump_flags & TDF_DETAILS))
3547 fprintf (dump_file, "Found fully redundant value\n");
3548 continue;
3549 }
3550
3551 avail = XCNEWVEC (pre_expr, last_basic_block);
3552 FOR_EACH_EDGE (pred, ei, block->preds)
3553 {
3554 unsigned int vprime;
3555
3556 /* We should never run insertion for the exit block
3557 and so not come across fake pred edges. */
3558 gcc_assert (!(pred->flags & EDGE_FAKE));
3559 bprime = pred->src;
3560 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3561 bprime, block);
3562
3563 /* eprime will generally only be NULL if the
3564 value of the expression, translated
3565 through the PHI for this predecessor, is
3566 undefined. If that is the case, we can't
3567 make the expression fully redundant,
3568 because its value is undefined along a
3569 predecessor path. We can thus break out
3570 early because it doesn't matter what the
3571 rest of the results are. */
3572 if (eprime == NULL)
3573 {
3574 cant_insert = true;
3575 break;
3576 }
3577
3578 eprime = fully_constant_expression (eprime);
3579 vprime = get_expr_value_id (eprime);
3580 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3581 vprime, NULL);
3582 if (edoubleprime == NULL)
3583 {
3584 avail[bprime->index] = eprime;
3585 all_same = false;
3586 }
3587 else
3588 {
3589 avail[bprime->index] = edoubleprime;
3590 by_some = true;
3591 /* We want to perform insertions to remove a redundancy on
3592 a path in the CFG we want to optimize for speed. */
3593 if (optimize_edge_for_speed_p (pred))
3594 do_insertion = true;
3595 if (first_s == NULL)
3596 first_s = edoubleprime;
3597 else if (!pre_expr_eq (first_s, edoubleprime))
3598 all_same = false;
3599 }
3600 }
3601 /* If we can insert it, it's not the same value
3602 already existing along every predecessor, and
3603 it's defined by some predecessor, it is
3604 partially redundant. */
3605 if (!cant_insert && !all_same && by_some)
3606 {
3607 if (!do_insertion)
3608 {
3609 if (dump_file && (dump_flags & TDF_DETAILS))
3610 {
3611 fprintf (dump_file, "Skipping partial redundancy for "
3612 "expression ");
3613 print_pre_expr (dump_file, expr);
3614 fprintf (dump_file, " (%04d), no redundancy on to be "
3615 "optimized for speed edge\n", val);
3616 }
3617 }
3618 else if (dbg_cnt (treepre_insert))
3619 {
3620 if (dump_file && (dump_flags & TDF_DETAILS))
3621 {
3622 fprintf (dump_file, "Found partial redundancy for "
3623 "expression ");
3624 print_pre_expr (dump_file, expr);
3625 fprintf (dump_file, " (%04d)\n",
3626 get_expr_value_id (expr));
3627 }
3628 if (insert_into_preds_of_block (block,
3629 get_expression_id (expr),
3630 avail))
3631 new_stuff = true;
3632 }
3633 }
3634 /* If all edges produce the same value and that value is
3635 an invariant, then the PHI has the same value on all
3636 edges. Note this. */
3637 else if (!cant_insert && all_same && eprime
3638 && (edoubleprime->kind == CONSTANT
3639 || edoubleprime->kind == NAME)
3640 && !value_id_constant_p (val))
3641 {
3642 unsigned int j;
3643 bitmap_iterator bi;
3644 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3645 value_expressions, val);
3646
3647 unsigned int new_val = get_expr_value_id (edoubleprime);
3648 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3649 {
3650 pre_expr expr = expression_for_id (j);
3651
3652 if (expr->kind == NAME)
3653 {
3654 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3655 /* Just reset the value id and valnum so it is
3656 the same as the constant we have discovered. */
3657 if (edoubleprime->kind == CONSTANT)
3658 {
3659 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3660 pre_stats.constified++;
3661 }
3662 else
3663 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3664 info->value_id = new_val;
3665 }
3666 }
3667 }
3668 free (avail);
3669 }
3670 }
3671
3672 VEC_free (pre_expr, heap, exprs);
3673 return new_stuff;
3674 }
3675
3676
3677 /* Perform insertion for partially anticipatable expressions. There
3678 is only one case we will perform insertion for these. This case is
3679 if the expression is partially anticipatable, and fully available.
3680 In this case, we know that putting it earlier will enable us to
3681 remove the later computation. */
3682
3683
3684 static bool
3685 do_partial_partial_insertion (basic_block block, basic_block dom)
3686 {
3687 bool new_stuff = false;
3688 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3689 pre_expr expr;
3690 int i;
3691
3692 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3693 {
3694 if (expr->kind != NAME)
3695 {
3696 pre_expr *avail;
3697 unsigned int val;
3698 bool by_all = true;
3699 bool cant_insert = false;
3700 edge pred;
3701 basic_block bprime;
3702 pre_expr eprime = NULL;
3703 edge_iterator ei;
3704
3705 val = get_expr_value_id (expr);
3706 if (bitmap_set_contains_value (PHI_GEN (block), val))
3707 continue;
3708 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3709 continue;
3710
3711 avail = XCNEWVEC (pre_expr, last_basic_block);
3712 FOR_EACH_EDGE (pred, ei, block->preds)
3713 {
3714 unsigned int vprime;
3715 pre_expr edoubleprime;
3716
3717 /* We should never run insertion for the exit block
3718 and so not come across fake pred edges. */
3719 gcc_assert (!(pred->flags & EDGE_FAKE));
3720 bprime = pred->src;
3721 eprime = phi_translate (expr, ANTIC_IN (block),
3722 PA_IN (block),
3723 bprime, block);
3724
3725 /* eprime will generally only be NULL if the
3726 value of the expression, translated
3727 through the PHI for this predecessor, is
3728 undefined. If that is the case, we can't
3729 make the expression fully redundant,
3730 because its value is undefined along a
3731 predecessor path. We can thus break out
3732 early because it doesn't matter what the
3733 rest of the results are. */
3734 if (eprime == NULL)
3735 {
3736 cant_insert = true;
3737 break;
3738 }
3739
3740 eprime = fully_constant_expression (eprime);
3741 vprime = get_expr_value_id (eprime);
3742 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3743 vprime, NULL);
3744 if (edoubleprime == NULL)
3745 {
3746 by_all = false;
3747 break;
3748 }
3749 else
3750 avail[bprime->index] = edoubleprime;
3751 }
3752
3753 /* If we can insert it, it's not the same value
3754 already existing along every predecessor, and
3755 it's defined by some predecessor, it is
3756 partially redundant. */
3757 if (!cant_insert && by_all)
3758 {
3759 edge succ;
3760 bool do_insertion = false;
3761
3762 /* Insert only if we can remove a later expression on a path
3763 that we want to optimize for speed.
3764 The phi node that we will be inserting in BLOCK is not free,
3765 and inserting it for the sake of !optimize_for_speed successor
3766 may cause regressions on the speed path. */
3767 FOR_EACH_EDGE (succ, ei, block->succs)
3768 {
3769 if (bitmap_set_contains_value (PA_IN (succ->dest), val))
3770 {
3771 if (optimize_edge_for_speed_p (succ))
3772 do_insertion = true;
3773 }
3774 }
3775
3776 if (!do_insertion)
3777 {
3778 if (dump_file && (dump_flags & TDF_DETAILS))
3779 {
3780 fprintf (dump_file, "Skipping partial partial redundancy "
3781 "for expression ");
3782 print_pre_expr (dump_file, expr);
3783 fprintf (dump_file, " (%04d), not partially anticipated "
3784 "on any to be optimized for speed edges\n", val);
3785 }
3786 }
3787 else if (dbg_cnt (treepre_insert))
3788 {
3789 pre_stats.pa_insert++;
3790 if (dump_file && (dump_flags & TDF_DETAILS))
3791 {
3792 fprintf (dump_file, "Found partial partial redundancy "
3793 "for expression ");
3794 print_pre_expr (dump_file, expr);
3795 fprintf (dump_file, " (%04d)\n",
3796 get_expr_value_id (expr));
3797 }
3798 if (insert_into_preds_of_block (block,
3799 get_expression_id (expr),
3800 avail))
3801 new_stuff = true;
3802 }
3803 }
3804 free (avail);
3805 }
3806 }
3807
3808 VEC_free (pre_expr, heap, exprs);
3809 return new_stuff;
3810 }
3811
3812 static bool
3813 insert_aux (basic_block block)
3814 {
3815 basic_block son;
3816 bool new_stuff = false;
3817
3818 if (block)
3819 {
3820 basic_block dom;
3821 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3822 if (dom)
3823 {
3824 unsigned i;
3825 bitmap_iterator bi;
3826 bitmap_set_t newset = NEW_SETS (dom);
3827 if (newset)
3828 {
3829 /* Note that we need to value_replace both NEW_SETS, and
3830 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3831 represented by some non-simple expression here that we want
3832 to replace it with. */
3833 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3834 {
3835 pre_expr expr = expression_for_id (i);
3836 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3837 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3838 }
3839 }
3840 if (!single_pred_p (block))
3841 {
3842 new_stuff |= do_regular_insertion (block, dom);
3843 if (do_partial_partial)
3844 new_stuff |= do_partial_partial_insertion (block, dom);
3845 }
3846 }
3847 }
3848 for (son = first_dom_son (CDI_DOMINATORS, block);
3849 son;
3850 son = next_dom_son (CDI_DOMINATORS, son))
3851 {
3852 new_stuff |= insert_aux (son);
3853 }
3854
3855 return new_stuff;
3856 }
3857
3858 /* Perform insertion of partially redundant values. */
3859
3860 static void
3861 insert (void)
3862 {
3863 bool new_stuff = true;
3864 basic_block bb;
3865 int num_iterations = 0;
3866
3867 FOR_ALL_BB (bb)
3868 NEW_SETS (bb) = bitmap_set_new ();
3869
3870 while (new_stuff)
3871 {
3872 num_iterations++;
3873 if (dump_file && dump_flags & TDF_DETAILS)
3874 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3875 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3876 }
3877 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3878 }
3879
3880
3881 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3882
3883 static void
3884 add_to_exp_gen (basic_block block, tree op)
3885 {
3886 if (!in_fre)
3887 {
3888 pre_expr result;
3889 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3890 return;
3891 result = get_or_alloc_expr_for_name (op);
3892 bitmap_value_insert_into_set (EXP_GEN (block), result);
3893 }
3894 }
3895
3896 /* Create value ids for PHI in BLOCK. */
3897
3898 static void
3899 make_values_for_phi (gimple phi, basic_block block)
3900 {
3901 tree result = gimple_phi_result (phi);
3902
3903 /* We have no need for virtual phis, as they don't represent
3904 actual computations. */
3905 if (is_gimple_reg (result))
3906 {
3907 pre_expr e = get_or_alloc_expr_for_name (result);
3908 add_to_value (get_expr_value_id (e), e);
3909 bitmap_insert_into_set (PHI_GEN (block), e);
3910 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3911 if (!in_fre)
3912 {
3913 unsigned i;
3914 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3915 {
3916 tree arg = gimple_phi_arg_def (phi, i);
3917 if (TREE_CODE (arg) == SSA_NAME)
3918 {
3919 e = get_or_alloc_expr_for_name (arg);
3920 add_to_value (get_expr_value_id (e), e);
3921 }
3922 }
3923 }
3924 }
3925 }
3926
3927 /* Compute the AVAIL set for all basic blocks.
3928
3929 This function performs value numbering of the statements in each basic
3930 block. The AVAIL sets are built from information we glean while doing
3931 this value numbering, since the AVAIL sets contain only one entry per
3932 value.
3933
3934 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3935 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3936
3937 static void
3938 compute_avail (void)
3939 {
3940
3941 basic_block block, son;
3942 basic_block *worklist;
3943 size_t sp = 0;
3944 unsigned i;
3945
3946 /* We pretend that default definitions are defined in the entry block.
3947 This includes function arguments and the static chain decl. */
3948 for (i = 1; i < num_ssa_names; ++i)
3949 {
3950 tree name = ssa_name (i);
3951 pre_expr e;
3952 if (!name
3953 || !SSA_NAME_IS_DEFAULT_DEF (name)
3954 || has_zero_uses (name)
3955 || !is_gimple_reg (name))
3956 continue;
3957
3958 e = get_or_alloc_expr_for_name (name);
3959 add_to_value (get_expr_value_id (e), e);
3960 if (!in_fre)
3961 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3962 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3963 }
3964
3965 /* Allocate the worklist. */
3966 worklist = XNEWVEC (basic_block, n_basic_blocks);
3967
3968 /* Seed the algorithm by putting the dominator children of the entry
3969 block on the worklist. */
3970 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3971 son;
3972 son = next_dom_son (CDI_DOMINATORS, son))
3973 worklist[sp++] = son;
3974
3975 /* Loop until the worklist is empty. */
3976 while (sp)
3977 {
3978 gimple_stmt_iterator gsi;
3979 gimple stmt;
3980 basic_block dom;
3981 unsigned int stmt_uid = 1;
3982
3983 /* Pick a block from the worklist. */
3984 block = worklist[--sp];
3985
3986 /* Initially, the set of available values in BLOCK is that of
3987 its immediate dominator. */
3988 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3989 if (dom)
3990 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3991
3992 /* Generate values for PHI nodes. */
3993 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3994 make_values_for_phi (gsi_stmt (gsi), block);
3995
3996 BB_MAY_NOTRETURN (block) = 0;
3997
3998 /* Now compute value numbers and populate value sets with all
3999 the expressions computed in BLOCK. */
4000 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
4001 {
4002 ssa_op_iter iter;
4003 tree op;
4004
4005 stmt = gsi_stmt (gsi);
4006 gimple_set_uid (stmt, stmt_uid++);
4007
4008 /* Cache whether the basic-block has any non-visible side-effect
4009 or control flow.
4010 If this isn't a call or it is the last stmt in the
4011 basic-block then the CFG represents things correctly. */
4012 if (is_gimple_call (stmt)
4013 && !stmt_ends_bb_p (stmt))
4014 {
4015 /* Non-looping const functions always return normally.
4016 Otherwise the call might not return or have side-effects
4017 that forbids hoisting possibly trapping expressions
4018 before it. */
4019 int flags = gimple_call_flags (stmt);
4020 if (!(flags & ECF_CONST)
4021 || (flags & ECF_LOOPING_CONST_OR_PURE))
4022 BB_MAY_NOTRETURN (block) = 1;
4023 }
4024
4025 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
4026 {
4027 pre_expr e = get_or_alloc_expr_for_name (op);
4028
4029 add_to_value (get_expr_value_id (e), e);
4030 if (!in_fre)
4031 bitmap_insert_into_set (TMP_GEN (block), e);
4032 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4033 }
4034
4035 if (gimple_has_volatile_ops (stmt)
4036 || stmt_could_throw_p (stmt))
4037 continue;
4038
4039 switch (gimple_code (stmt))
4040 {
4041 case GIMPLE_RETURN:
4042 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4043 add_to_exp_gen (block, op);
4044 continue;
4045
4046 case GIMPLE_CALL:
4047 {
4048 vn_reference_t ref;
4049 unsigned int i;
4050 vn_reference_op_t vro;
4051 pre_expr result = NULL;
4052 VEC(vn_reference_op_s, heap) *ops = NULL;
4053
4054 if (!can_value_number_call (stmt))
4055 continue;
4056
4057 copy_reference_ops_from_call (stmt, &ops);
4058 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4059 gimple_expr_type (stmt),
4060 ops, &ref, VN_NOWALK);
4061 VEC_free (vn_reference_op_s, heap, ops);
4062 if (!ref)
4063 continue;
4064
4065 for (i = 0; VEC_iterate (vn_reference_op_s,
4066 ref->operands, i,
4067 vro); i++)
4068 {
4069 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4070 add_to_exp_gen (block, vro->op0);
4071 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4072 add_to_exp_gen (block, vro->op1);
4073 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4074 add_to_exp_gen (block, vro->op2);
4075 }
4076
4077 /* If the value of the call is not invalidated in
4078 this block until it is computed, add the expression
4079 to EXP_GEN. */
4080 if (!gimple_vuse (stmt)
4081 || gimple_code
4082 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
4083 || gimple_bb (SSA_NAME_DEF_STMT
4084 (gimple_vuse (stmt))) != block)
4085 {
4086 result = (pre_expr) pool_alloc (pre_expr_pool);
4087 result->kind = REFERENCE;
4088 result->id = 0;
4089 PRE_EXPR_REFERENCE (result) = ref;
4090
4091 get_or_alloc_expression_id (result);
4092 add_to_value (get_expr_value_id (result), result);
4093 if (!in_fre)
4094 bitmap_value_insert_into_set (EXP_GEN (block), result);
4095 }
4096 continue;
4097 }
4098
4099 case GIMPLE_ASSIGN:
4100 {
4101 pre_expr result = NULL;
4102 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4103 {
4104 case tcc_unary:
4105 case tcc_binary:
4106 case tcc_comparison:
4107 {
4108 vn_nary_op_t nary;
4109 unsigned int i;
4110
4111 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4112 gimple_assign_rhs_code (stmt),
4113 gimple_expr_type (stmt),
4114 gimple_assign_rhs1_ptr (stmt),
4115 &nary);
4116
4117 if (!nary)
4118 continue;
4119
4120 for (i = 0; i < nary->length; i++)
4121 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4122 add_to_exp_gen (block, nary->op[i]);
4123
4124 /* If the NARY traps and there was a preceeding
4125 point in the block that might not return avoid
4126 adding the nary to EXP_GEN. */
4127 if (BB_MAY_NOTRETURN (block)
4128 && vn_nary_may_trap (nary))
4129 continue;
4130
4131 result = (pre_expr) pool_alloc (pre_expr_pool);
4132 result->kind = NARY;
4133 result->id = 0;
4134 PRE_EXPR_NARY (result) = nary;
4135 break;
4136 }
4137
4138 case tcc_declaration:
4139 case tcc_reference:
4140 {
4141 vn_reference_t ref;
4142 unsigned int i;
4143 vn_reference_op_t vro;
4144
4145 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4146 gimple_vuse (stmt),
4147 VN_WALK, &ref);
4148 if (!ref)
4149 continue;
4150
4151 for (i = 0; VEC_iterate (vn_reference_op_s,
4152 ref->operands, i,
4153 vro); i++)
4154 {
4155 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4156 add_to_exp_gen (block, vro->op0);
4157 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4158 add_to_exp_gen (block, vro->op1);
4159 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4160 add_to_exp_gen (block, vro->op2);
4161 }
4162
4163 /* If the value of the reference is not invalidated in
4164 this block until it is computed, add the expression
4165 to EXP_GEN. */
4166 if (gimple_vuse (stmt))
4167 {
4168 gimple def_stmt;
4169 bool ok = true;
4170 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
4171 while (!gimple_nop_p (def_stmt)
4172 && gimple_code (def_stmt) != GIMPLE_PHI
4173 && gimple_bb (def_stmt) == block)
4174 {
4175 if (stmt_may_clobber_ref_p
4176 (def_stmt, gimple_assign_rhs1 (stmt)))
4177 {
4178 ok = false;
4179 break;
4180 }
4181 def_stmt
4182 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
4183 }
4184 if (!ok)
4185 continue;
4186 }
4187
4188 result = (pre_expr) pool_alloc (pre_expr_pool);
4189 result->kind = REFERENCE;
4190 result->id = 0;
4191 PRE_EXPR_REFERENCE (result) = ref;
4192 break;
4193 }
4194
4195 default:
4196 /* For any other statement that we don't
4197 recognize, simply add all referenced
4198 SSA_NAMEs to EXP_GEN. */
4199 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4200 add_to_exp_gen (block, op);
4201 continue;
4202 }
4203
4204 get_or_alloc_expression_id (result);
4205 add_to_value (get_expr_value_id (result), result);
4206 if (!in_fre)
4207 bitmap_value_insert_into_set (EXP_GEN (block), result);
4208
4209 continue;
4210 }
4211 default:
4212 break;
4213 }
4214 }
4215
4216 /* Put the dominator children of BLOCK on the worklist of blocks
4217 to compute available sets for. */
4218 for (son = first_dom_son (CDI_DOMINATORS, block);
4219 son;
4220 son = next_dom_son (CDI_DOMINATORS, son))
4221 worklist[sp++] = son;
4222 }
4223
4224 free (worklist);
4225 }
4226
4227 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4228 than the available expressions for it. The insertion point is
4229 right before the first use in STMT. Returns the SSA_NAME that should
4230 be used for replacement. */
4231
4232 static tree
4233 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4234 {
4235 basic_block bb = gimple_bb (stmt);
4236 gimple_stmt_iterator gsi;
4237 gimple_seq stmts = NULL;
4238 tree expr;
4239 pre_expr e;
4240
4241 /* First create a value expression from the expression we want
4242 to insert and associate it with the value handle for SSA_VN. */
4243 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4244 if (e == NULL)
4245 return NULL_TREE;
4246
4247 /* Then use create_expression_by_pieces to generate a valid
4248 expression to insert at this point of the IL stream. */
4249 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4250 if (expr == NULL_TREE)
4251 return NULL_TREE;
4252 gsi = gsi_for_stmt (stmt);
4253 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4254
4255 return expr;
4256 }
4257
4258 /* Eliminate fully redundant computations. */
4259
4260 static unsigned int
4261 eliminate (void)
4262 {
4263 VEC (gimple, heap) *to_remove = NULL;
4264 VEC (gimple, heap) *to_update = NULL;
4265 basic_block b;
4266 unsigned int todo = 0;
4267 gimple_stmt_iterator gsi;
4268 gimple stmt;
4269 unsigned i;
4270
4271 FOR_EACH_BB (b)
4272 {
4273 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4274 {
4275 tree lhs = NULL_TREE;
4276 tree rhs = NULL_TREE;
4277
4278 stmt = gsi_stmt (gsi);
4279
4280 if (gimple_has_lhs (stmt))
4281 lhs = gimple_get_lhs (stmt);
4282
4283 if (gimple_assign_single_p (stmt))
4284 rhs = gimple_assign_rhs1 (stmt);
4285
4286 /* Lookup the RHS of the expression, see if we have an
4287 available computation for it. If so, replace the RHS with
4288 the available computation.
4289
4290 See PR43491.
4291 We don't replace global register variable when it is a the RHS of
4292 a single assign. We do replace local register variable since gcc
4293 does not guarantee local variable will be allocated in register. */
4294 if (gimple_has_lhs (stmt)
4295 && TREE_CODE (lhs) == SSA_NAME
4296 && !gimple_assign_ssa_name_copy_p (stmt)
4297 && (!gimple_assign_single_p (stmt)
4298 || (!is_gimple_min_invariant (rhs)
4299 && (gimple_assign_rhs_code (stmt) != VAR_DECL
4300 || !is_global_var (rhs)
4301 || !DECL_HARD_REGISTER (rhs))))
4302 && !gimple_has_volatile_ops (stmt)
4303 && !has_zero_uses (lhs))
4304 {
4305 tree sprime = NULL;
4306 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4307 pre_expr sprimeexpr;
4308 gimple orig_stmt = stmt;
4309
4310 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4311 get_expr_value_id (lhsexpr),
4312 NULL);
4313
4314 if (sprimeexpr)
4315 {
4316 if (sprimeexpr->kind == CONSTANT)
4317 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4318 else if (sprimeexpr->kind == NAME)
4319 sprime = PRE_EXPR_NAME (sprimeexpr);
4320 else
4321 gcc_unreachable ();
4322 }
4323
4324 /* If there is no existing leader but SCCVN knows this
4325 value is constant, use that constant. */
4326 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4327 {
4328 sprime = VN_INFO (lhs)->valnum;
4329 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4330 TREE_TYPE (sprime)))
4331 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4332
4333 if (dump_file && (dump_flags & TDF_DETAILS))
4334 {
4335 fprintf (dump_file, "Replaced ");
4336 print_gimple_expr (dump_file, stmt, 0, 0);
4337 fprintf (dump_file, " with ");
4338 print_generic_expr (dump_file, sprime, 0);
4339 fprintf (dump_file, " in ");
4340 print_gimple_stmt (dump_file, stmt, 0, 0);
4341 }
4342 pre_stats.eliminations++;
4343 propagate_tree_value_into_stmt (&gsi, sprime);
4344 stmt = gsi_stmt (gsi);
4345 update_stmt (stmt);
4346
4347 /* If we removed EH side-effects from the statement, clean
4348 its EH information. */
4349 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4350 {
4351 bitmap_set_bit (need_eh_cleanup,
4352 gimple_bb (stmt)->index);
4353 if (dump_file && (dump_flags & TDF_DETAILS))
4354 fprintf (dump_file, " Removed EH side-effects.\n");
4355 }
4356 continue;
4357 }
4358
4359 /* If there is no existing usable leader but SCCVN thinks
4360 it has an expression it wants to use as replacement,
4361 insert that. */
4362 if (!sprime || sprime == lhs)
4363 {
4364 tree val = VN_INFO (lhs)->valnum;
4365 if (val != VN_TOP
4366 && TREE_CODE (val) == SSA_NAME
4367 && VN_INFO (val)->needs_insertion
4368 && can_PRE_operation (vn_get_expr_for (val)))
4369 sprime = do_SCCVN_insertion (stmt, val);
4370 }
4371 if (sprime
4372 && sprime != lhs
4373 && (rhs == NULL_TREE
4374 || TREE_CODE (rhs) != SSA_NAME
4375 || may_propagate_copy (rhs, sprime)))
4376 {
4377 bool can_make_abnormal_goto
4378 = is_gimple_call (stmt)
4379 && stmt_can_make_abnormal_goto (stmt);
4380
4381 gcc_assert (sprime != rhs);
4382
4383 if (dump_file && (dump_flags & TDF_DETAILS))
4384 {
4385 fprintf (dump_file, "Replaced ");
4386 print_gimple_expr (dump_file, stmt, 0, 0);
4387 fprintf (dump_file, " with ");
4388 print_generic_expr (dump_file, sprime, 0);
4389 fprintf (dump_file, " in ");
4390 print_gimple_stmt (dump_file, stmt, 0, 0);
4391 }
4392
4393 if (TREE_CODE (sprime) == SSA_NAME)
4394 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4395 NECESSARY, true);
4396 /* We need to make sure the new and old types actually match,
4397 which may require adding a simple cast, which fold_convert
4398 will do for us. */
4399 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4400 && !useless_type_conversion_p (gimple_expr_type (stmt),
4401 TREE_TYPE (sprime)))
4402 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4403
4404 pre_stats.eliminations++;
4405 propagate_tree_value_into_stmt (&gsi, sprime);
4406 stmt = gsi_stmt (gsi);
4407 update_stmt (stmt);
4408
4409 /* If we removed EH side-effects from the statement, clean
4410 its EH information. */
4411 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4412 {
4413 bitmap_set_bit (need_eh_cleanup,
4414 gimple_bb (stmt)->index);
4415 if (dump_file && (dump_flags & TDF_DETAILS))
4416 fprintf (dump_file, " Removed EH side-effects.\n");
4417 }
4418
4419 /* Likewise for AB side-effects. */
4420 if (can_make_abnormal_goto
4421 && !stmt_can_make_abnormal_goto (stmt))
4422 {
4423 bitmap_set_bit (need_ab_cleanup,
4424 gimple_bb (stmt)->index);
4425 if (dump_file && (dump_flags & TDF_DETAILS))
4426 fprintf (dump_file, " Removed AB side-effects.\n");
4427 }
4428 }
4429 }
4430 /* If the statement is a scalar store, see if the expression
4431 has the same value number as its rhs. If so, the store is
4432 dead. */
4433 else if (gimple_assign_single_p (stmt)
4434 && !gimple_has_volatile_ops (stmt)
4435 && !is_gimple_reg (gimple_assign_lhs (stmt))
4436 && (TREE_CODE (rhs) == SSA_NAME
4437 || is_gimple_min_invariant (rhs)))
4438 {
4439 tree val;
4440 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4441 gimple_vuse (stmt), VN_WALK, NULL);
4442 if (TREE_CODE (rhs) == SSA_NAME)
4443 rhs = VN_INFO (rhs)->valnum;
4444 if (val
4445 && operand_equal_p (val, rhs, 0))
4446 {
4447 if (dump_file && (dump_flags & TDF_DETAILS))
4448 {
4449 fprintf (dump_file, "Deleted redundant store ");
4450 print_gimple_stmt (dump_file, stmt, 0, 0);
4451 }
4452
4453 /* Queue stmt for removal. */
4454 VEC_safe_push (gimple, heap, to_remove, stmt);
4455 }
4456 }
4457 /* Visit COND_EXPRs and fold the comparison with the
4458 available value-numbers. */
4459 else if (gimple_code (stmt) == GIMPLE_COND)
4460 {
4461 tree op0 = gimple_cond_lhs (stmt);
4462 tree op1 = gimple_cond_rhs (stmt);
4463 tree result;
4464
4465 if (TREE_CODE (op0) == SSA_NAME)
4466 op0 = VN_INFO (op0)->valnum;
4467 if (TREE_CODE (op1) == SSA_NAME)
4468 op1 = VN_INFO (op1)->valnum;
4469 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4470 op0, op1);
4471 if (result && TREE_CODE (result) == INTEGER_CST)
4472 {
4473 if (integer_zerop (result))
4474 gimple_cond_make_false (stmt);
4475 else
4476 gimple_cond_make_true (stmt);
4477 update_stmt (stmt);
4478 todo = TODO_cleanup_cfg;
4479 }
4480 }
4481 /* Visit indirect calls and turn them into direct calls if
4482 possible. */
4483 if (is_gimple_call (stmt))
4484 {
4485 tree orig_fn = gimple_call_fn (stmt);
4486 tree fn;
4487 if (!orig_fn)
4488 continue;
4489 if (TREE_CODE (orig_fn) == SSA_NAME)
4490 fn = VN_INFO (orig_fn)->valnum;
4491 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4492 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4493 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4494 else
4495 continue;
4496 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4497 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4498 TREE_TYPE (fn)))
4499 {
4500 bool can_make_abnormal_goto
4501 = stmt_can_make_abnormal_goto (stmt);
4502 bool was_noreturn = gimple_call_noreturn_p (stmt);
4503
4504 if (dump_file && (dump_flags & TDF_DETAILS))
4505 {
4506 fprintf (dump_file, "Replacing call target with ");
4507 print_generic_expr (dump_file, fn, 0);
4508 fprintf (dump_file, " in ");
4509 print_gimple_stmt (dump_file, stmt, 0, 0);
4510 }
4511
4512 gimple_call_set_fn (stmt, fn);
4513 VEC_safe_push (gimple, heap, to_update, stmt);
4514
4515 /* When changing a call into a noreturn call, cfg cleanup
4516 is needed to fix up the noreturn call. */
4517 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4518 todo |= TODO_cleanup_cfg;
4519
4520 /* If we removed EH side-effects from the statement, clean
4521 its EH information. */
4522 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4523 {
4524 bitmap_set_bit (need_eh_cleanup,
4525 gimple_bb (stmt)->index);
4526 if (dump_file && (dump_flags & TDF_DETAILS))
4527 fprintf (dump_file, " Removed EH side-effects.\n");
4528 }
4529
4530 /* Likewise for AB side-effects. */
4531 if (can_make_abnormal_goto
4532 && !stmt_can_make_abnormal_goto (stmt))
4533 {
4534 bitmap_set_bit (need_ab_cleanup,
4535 gimple_bb (stmt)->index);
4536 if (dump_file && (dump_flags & TDF_DETAILS))
4537 fprintf (dump_file, " Removed AB side-effects.\n");
4538 }
4539
4540 /* Changing an indirect call to a direct call may
4541 have exposed different semantics. This may
4542 require an SSA update. */
4543 todo |= TODO_update_ssa_only_virtuals;
4544 }
4545 }
4546 }
4547
4548 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4549 {
4550 gimple stmt, phi = gsi_stmt (gsi);
4551 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4552 pre_expr sprimeexpr, resexpr;
4553 gimple_stmt_iterator gsi2;
4554
4555 /* We want to perform redundant PHI elimination. Do so by
4556 replacing the PHI with a single copy if possible.
4557 Do not touch inserted, single-argument or virtual PHIs. */
4558 if (gimple_phi_num_args (phi) == 1
4559 || !is_gimple_reg (res))
4560 {
4561 gsi_next (&gsi);
4562 continue;
4563 }
4564
4565 resexpr = get_or_alloc_expr_for_name (res);
4566 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4567 get_expr_value_id (resexpr), NULL);
4568 if (sprimeexpr)
4569 {
4570 if (sprimeexpr->kind == CONSTANT)
4571 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4572 else if (sprimeexpr->kind == NAME)
4573 sprime = PRE_EXPR_NAME (sprimeexpr);
4574 else
4575 gcc_unreachable ();
4576 }
4577 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4578 {
4579 sprime = VN_INFO (res)->valnum;
4580 if (!useless_type_conversion_p (TREE_TYPE (res),
4581 TREE_TYPE (sprime)))
4582 sprime = fold_convert (TREE_TYPE (res), sprime);
4583 }
4584 if (!sprime
4585 || sprime == res)
4586 {
4587 gsi_next (&gsi);
4588 continue;
4589 }
4590
4591 if (dump_file && (dump_flags & TDF_DETAILS))
4592 {
4593 fprintf (dump_file, "Replaced redundant PHI node defining ");
4594 print_generic_expr (dump_file, res, 0);
4595 fprintf (dump_file, " with ");
4596 print_generic_expr (dump_file, sprime, 0);
4597 fprintf (dump_file, "\n");
4598 }
4599
4600 remove_phi_node (&gsi, false);
4601
4602 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4603 && TREE_CODE (sprime) == SSA_NAME)
4604 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4605
4606 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4607 sprime = fold_convert (TREE_TYPE (res), sprime);
4608 stmt = gimple_build_assign (res, sprime);
4609 SSA_NAME_DEF_STMT (res) = stmt;
4610 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4611
4612 gsi2 = gsi_after_labels (b);
4613 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4614 /* Queue the copy for eventual removal. */
4615 VEC_safe_push (gimple, heap, to_remove, stmt);
4616 /* If we inserted this PHI node ourself, it's not an elimination. */
4617 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4618 pre_stats.phis--;
4619 else
4620 pre_stats.eliminations++;
4621 }
4622 }
4623
4624 /* We cannot remove stmts during BB walk, especially not release SSA
4625 names there as this confuses the VN machinery. The stmts ending
4626 up in to_remove are either stores or simple copies. */
4627 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4628 {
4629 tree lhs = gimple_assign_lhs (stmt);
4630 tree rhs = gimple_assign_rhs1 (stmt);
4631 use_operand_p use_p;
4632 gimple use_stmt;
4633
4634 /* If there is a single use only, propagate the equivalency
4635 instead of keeping the copy. */
4636 if (TREE_CODE (lhs) == SSA_NAME
4637 && TREE_CODE (rhs) == SSA_NAME
4638 && single_imm_use (lhs, &use_p, &use_stmt)
4639 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4640 {
4641 SET_USE (use_p, rhs);
4642 update_stmt (use_stmt);
4643 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4644 && TREE_CODE (rhs) == SSA_NAME)
4645 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4646 }
4647
4648 /* If this is a store or a now unused copy, remove it. */
4649 if (TREE_CODE (lhs) != SSA_NAME
4650 || has_zero_uses (lhs))
4651 {
4652 basic_block bb = gimple_bb (stmt);
4653 gsi = gsi_for_stmt (stmt);
4654 unlink_stmt_vdef (stmt);
4655 if (gsi_remove (&gsi, true))
4656 bitmap_set_bit (need_eh_cleanup, bb->index);
4657 if (TREE_CODE (lhs) == SSA_NAME)
4658 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4659 release_defs (stmt);
4660 }
4661 }
4662 VEC_free (gimple, heap, to_remove);
4663
4664 /* We cannot update call statements with virtual operands during
4665 SSA walk. This might remove them which in turn makes our
4666 VN lattice invalid. */
4667 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4668 update_stmt (stmt);
4669 VEC_free (gimple, heap, to_update);
4670
4671 return todo;
4672 }
4673
4674 /* Borrow a bit of tree-ssa-dce.c for the moment.
4675 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4676 this may be a bit faster, and we may want critical edges kept split. */
4677
4678 /* If OP's defining statement has not already been determined to be necessary,
4679 mark that statement necessary. Return the stmt, if it is newly
4680 necessary. */
4681
4682 static inline gimple
4683 mark_operand_necessary (tree op)
4684 {
4685 gimple stmt;
4686
4687 gcc_assert (op);
4688
4689 if (TREE_CODE (op) != SSA_NAME)
4690 return NULL;
4691
4692 stmt = SSA_NAME_DEF_STMT (op);
4693 gcc_assert (stmt);
4694
4695 if (gimple_plf (stmt, NECESSARY)
4696 || gimple_nop_p (stmt))
4697 return NULL;
4698
4699 gimple_set_plf (stmt, NECESSARY, true);
4700 return stmt;
4701 }
4702
4703 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4704 to insert PHI nodes sometimes, and because value numbering of casts isn't
4705 perfect, we sometimes end up inserting dead code. This simple DCE-like
4706 pass removes any insertions we made that weren't actually used. */
4707
4708 static void
4709 remove_dead_inserted_code (void)
4710 {
4711 bitmap worklist;
4712 unsigned i;
4713 bitmap_iterator bi;
4714 gimple t;
4715
4716 worklist = BITMAP_ALLOC (NULL);
4717 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4718 {
4719 t = SSA_NAME_DEF_STMT (ssa_name (i));
4720 if (gimple_plf (t, NECESSARY))
4721 bitmap_set_bit (worklist, i);
4722 }
4723 while (!bitmap_empty_p (worklist))
4724 {
4725 i = bitmap_first_set_bit (worklist);
4726 bitmap_clear_bit (worklist, i);
4727 t = SSA_NAME_DEF_STMT (ssa_name (i));
4728
4729 /* PHI nodes are somewhat special in that each PHI alternative has
4730 data and control dependencies. All the statements feeding the
4731 PHI node's arguments are always necessary. */
4732 if (gimple_code (t) == GIMPLE_PHI)
4733 {
4734 unsigned k;
4735
4736 for (k = 0; k < gimple_phi_num_args (t); k++)
4737 {
4738 tree arg = PHI_ARG_DEF (t, k);
4739 if (TREE_CODE (arg) == SSA_NAME)
4740 {
4741 gimple n = mark_operand_necessary (arg);
4742 if (n)
4743 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4744 }
4745 }
4746 }
4747 else
4748 {
4749 /* Propagate through the operands. Examine all the USE, VUSE and
4750 VDEF operands in this statement. Mark all the statements
4751 which feed this statement's uses as necessary. */
4752 ssa_op_iter iter;
4753 tree use;
4754
4755 /* The operands of VDEF expressions are also needed as they
4756 represent potential definitions that may reach this
4757 statement (VDEF operands allow us to follow def-def
4758 links). */
4759
4760 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4761 {
4762 gimple n = mark_operand_necessary (use);
4763 if (n)
4764 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4765 }
4766 }
4767 }
4768
4769 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4770 {
4771 t = SSA_NAME_DEF_STMT (ssa_name (i));
4772 if (!gimple_plf (t, NECESSARY))
4773 {
4774 gimple_stmt_iterator gsi;
4775
4776 if (dump_file && (dump_flags & TDF_DETAILS))
4777 {
4778 fprintf (dump_file, "Removing unnecessary insertion:");
4779 print_gimple_stmt (dump_file, t, 0, 0);
4780 }
4781
4782 gsi = gsi_for_stmt (t);
4783 if (gimple_code (t) == GIMPLE_PHI)
4784 remove_phi_node (&gsi, true);
4785 else
4786 {
4787 gsi_remove (&gsi, true);
4788 release_defs (t);
4789 }
4790 }
4791 }
4792 BITMAP_FREE (worklist);
4793 }
4794
4795 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4796 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4797 the number of visited blocks. */
4798
4799 static int
4800 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4801 {
4802 edge_iterator *stack;
4803 int sp;
4804 int post_order_num = 0;
4805 sbitmap visited;
4806
4807 if (include_entry_exit)
4808 post_order[post_order_num++] = EXIT_BLOCK;
4809
4810 /* Allocate stack for back-tracking up CFG. */
4811 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4812 sp = 0;
4813
4814 /* Allocate bitmap to track nodes that have been visited. */
4815 visited = sbitmap_alloc (last_basic_block);
4816
4817 /* None of the nodes in the CFG have been visited yet. */
4818 sbitmap_zero (visited);
4819
4820 /* Push the last edge on to the stack. */
4821 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4822
4823 while (sp)
4824 {
4825 edge_iterator ei;
4826 basic_block src;
4827 basic_block dest;
4828
4829 /* Look at the edge on the top of the stack. */
4830 ei = stack[sp - 1];
4831 src = ei_edge (ei)->src;
4832 dest = ei_edge (ei)->dest;
4833
4834 /* Check if the edge destination has been visited yet. */
4835 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4836 {
4837 /* Mark that we have visited the destination. */
4838 SET_BIT (visited, src->index);
4839
4840 if (EDGE_COUNT (src->preds) > 0)
4841 /* Since the DEST node has been visited for the first
4842 time, check its successors. */
4843 stack[sp++] = ei_start (src->preds);
4844 else
4845 post_order[post_order_num++] = src->index;
4846 }
4847 else
4848 {
4849 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4850 post_order[post_order_num++] = dest->index;
4851
4852 if (!ei_one_before_end_p (ei))
4853 ei_next (&stack[sp - 1]);
4854 else
4855 sp--;
4856 }
4857 }
4858
4859 if (include_entry_exit)
4860 post_order[post_order_num++] = ENTRY_BLOCK;
4861
4862 free (stack);
4863 sbitmap_free (visited);
4864 return post_order_num;
4865 }
4866
4867
4868 /* Initialize data structures used by PRE. */
4869
4870 static void
4871 init_pre (bool do_fre)
4872 {
4873 basic_block bb;
4874
4875 next_expression_id = 1;
4876 expressions = NULL;
4877 VEC_safe_push (pre_expr, heap, expressions, NULL);
4878 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4879 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4880 get_max_value_id() + 1);
4881 name_to_id = NULL;
4882
4883 in_fre = do_fre;
4884
4885 inserted_exprs = BITMAP_ALLOC (NULL);
4886 need_creation = NULL;
4887 pretemp = NULL_TREE;
4888 storetemp = NULL_TREE;
4889 prephitemp = NULL_TREE;
4890
4891 connect_infinite_loops_to_exit ();
4892 memset (&pre_stats, 0, sizeof (pre_stats));
4893
4894
4895 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4896 my_rev_post_order_compute (postorder, false);
4897
4898 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4899
4900 calculate_dominance_info (CDI_POST_DOMINATORS);
4901 calculate_dominance_info (CDI_DOMINATORS);
4902
4903 bitmap_obstack_initialize (&grand_bitmap_obstack);
4904 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4905 expr_pred_trans_eq, free);
4906 expression_to_id = htab_create (num_ssa_names * 3,
4907 pre_expr_hash,
4908 pre_expr_eq, NULL);
4909 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4910 sizeof (struct bitmap_set), 30);
4911 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4912 sizeof (struct pre_expr_d), 30);
4913 FOR_ALL_BB (bb)
4914 {
4915 EXP_GEN (bb) = bitmap_set_new ();
4916 PHI_GEN (bb) = bitmap_set_new ();
4917 TMP_GEN (bb) = bitmap_set_new ();
4918 AVAIL_OUT (bb) = bitmap_set_new ();
4919 }
4920
4921 need_eh_cleanup = BITMAP_ALLOC (NULL);
4922 need_ab_cleanup = BITMAP_ALLOC (NULL);
4923 }
4924
4925
4926 /* Deallocate data structures used by PRE. */
4927
4928 static void
4929 fini_pre (bool do_fre)
4930 {
4931 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4932 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4933
4934 free (postorder);
4935 VEC_free (bitmap_set_t, heap, value_expressions);
4936 BITMAP_FREE (inserted_exprs);
4937 VEC_free (gimple, heap, need_creation);
4938 bitmap_obstack_release (&grand_bitmap_obstack);
4939 free_alloc_pool (bitmap_set_pool);
4940 free_alloc_pool (pre_expr_pool);
4941 htab_delete (phi_translate_table);
4942 htab_delete (expression_to_id);
4943 VEC_free (unsigned, heap, name_to_id);
4944
4945 free_aux_for_blocks ();
4946
4947 free_dominance_info (CDI_POST_DOMINATORS);
4948
4949 if (do_eh_cleanup)
4950 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4951
4952 if (do_ab_cleanup)
4953 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4954
4955 BITMAP_FREE (need_eh_cleanup);
4956 BITMAP_FREE (need_ab_cleanup);
4957
4958 if (do_eh_cleanup || do_ab_cleanup)
4959 cleanup_tree_cfg ();
4960
4961 if (!do_fre)
4962 loop_optimizer_finalize ();
4963 }
4964
4965 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4966 only wants to do full redundancy elimination. */
4967
4968 static unsigned int
4969 execute_pre (bool do_fre)
4970 {
4971 unsigned int todo = 0;
4972
4973 do_partial_partial =
4974 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4975
4976 /* This has to happen before SCCVN runs because
4977 loop_optimizer_init may create new phis, etc. */
4978 if (!do_fre)
4979 loop_optimizer_init (LOOPS_NORMAL);
4980
4981 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4982 {
4983 if (!do_fre)
4984 loop_optimizer_finalize ();
4985
4986 return 0;
4987 }
4988
4989 init_pre (do_fre);
4990 scev_initialize ();
4991
4992 /* Collect and value number expressions computed in each basic block. */
4993 compute_avail ();
4994
4995 if (dump_file && (dump_flags & TDF_DETAILS))
4996 {
4997 basic_block bb;
4998
4999 FOR_ALL_BB (bb)
5000 {
5001 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
5002 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
5003 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
5004 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
5005 }
5006 }
5007
5008 /* Insert can get quite slow on an incredibly large number of basic
5009 blocks due to some quadratic behavior. Until this behavior is
5010 fixed, don't run it when he have an incredibly large number of
5011 bb's. If we aren't going to run insert, there is no point in
5012 computing ANTIC, either, even though it's plenty fast. */
5013 if (!do_fre && n_basic_blocks < 4000)
5014 {
5015 compute_antic ();
5016 insert ();
5017 }
5018
5019 /* Make sure to remove fake edges before committing our inserts.
5020 This makes sure we don't end up with extra critical edges that
5021 we would need to split. */
5022 remove_fake_exit_edges ();
5023 gsi_commit_edge_inserts ();
5024
5025 /* Remove all the redundant expressions. */
5026 todo |= eliminate ();
5027
5028 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
5029 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
5030 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
5031 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
5032 statistics_counter_event (cfun, "Constified", pre_stats.constified);
5033
5034 clear_expression_ids ();
5035 if (!do_fre)
5036 {
5037 remove_dead_inserted_code ();
5038 todo |= TODO_verify_flow;
5039 }
5040
5041 scev_finalize ();
5042 fini_pre (do_fre);
5043
5044 if (!do_fre)
5045 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5046 case we can merge the block with the remaining predecessor of the block.
5047 It should either:
5048 - call merge_blocks after each tail merge iteration
5049 - call merge_blocks after all tail merge iterations
5050 - mark TODO_cleanup_cfg when necessary
5051 - share the cfg cleanup with fini_pre. */
5052 todo |= tail_merge_optimize (todo);
5053 free_scc_vn ();
5054
5055 return todo;
5056 }
5057
5058 /* Gate and execute functions for PRE. */
5059
5060 static unsigned int
5061 do_pre (void)
5062 {
5063 return execute_pre (false);
5064 }
5065
5066 static bool
5067 gate_pre (void)
5068 {
5069 return flag_tree_pre != 0;
5070 }
5071
5072 struct gimple_opt_pass pass_pre =
5073 {
5074 {
5075 GIMPLE_PASS,
5076 "pre", /* name */
5077 gate_pre, /* gate */
5078 do_pre, /* execute */
5079 NULL, /* sub */
5080 NULL, /* next */
5081 0, /* static_pass_number */
5082 TV_TREE_PRE, /* tv_id */
5083 PROP_no_crit_edges | PROP_cfg
5084 | PROP_ssa, /* properties_required */
5085 0, /* properties_provided */
5086 0, /* properties_destroyed */
5087 TODO_rebuild_alias, /* todo_flags_start */
5088 TODO_update_ssa_only_virtuals | TODO_ggc_collect
5089 | TODO_verify_ssa /* todo_flags_finish */
5090 }
5091 };
5092
5093
5094 /* Gate and execute functions for FRE. */
5095
5096 static unsigned int
5097 execute_fre (void)
5098 {
5099 return execute_pre (true);
5100 }
5101
5102 static bool
5103 gate_fre (void)
5104 {
5105 return flag_tree_fre != 0;
5106 }
5107
5108 struct gimple_opt_pass pass_fre =
5109 {
5110 {
5111 GIMPLE_PASS,
5112 "fre", /* name */
5113 gate_fre, /* gate */
5114 execute_fre, /* execute */
5115 NULL, /* sub */
5116 NULL, /* next */
5117 0, /* static_pass_number */
5118 TV_TREE_FRE, /* tv_id */
5119 PROP_cfg | PROP_ssa, /* properties_required */
5120 0, /* properties_provided */
5121 0, /* properties_destroyed */
5122 0, /* todo_flags_start */
5123 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
5124 }
5125 };