tree-ssa-pre.c (debug_bitmap_sets_for): New function.
[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 /* If the NARY may trap make sure the block does not contain
2073 a possible exit point.
2074 ??? This is overly conservative if we translate AVAIL_OUT
2075 as the available expression might be after the exit point. */
2076 if (BB_MAY_NOTRETURN (block)
2077 && vn_nary_may_trap (nary))
2078 return false;
2079 return true;
2080 }
2081 break;
2082 case REFERENCE:
2083 {
2084 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2085 vn_reference_op_t vro;
2086 unsigned int i;
2087
2088 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2089 {
2090 if (!op_valid_in_sets (set1, set2, vro->op0)
2091 || !op_valid_in_sets (set1, set2, vro->op1)
2092 || !op_valid_in_sets (set1, set2, vro->op2))
2093 return false;
2094 }
2095 return true;
2096 }
2097 default:
2098 gcc_unreachable ();
2099 }
2100 }
2101
2102 /* Clean the set of expressions that are no longer valid in SET1 or
2103 SET2. This means expressions that are made up of values we have no
2104 leaders for in SET1 or SET2. This version is used for partial
2105 anticipation, which means it is not valid in either ANTIC_IN or
2106 PA_IN. */
2107
2108 static void
2109 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2110 {
2111 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2112 pre_expr expr;
2113 int i;
2114
2115 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2116 {
2117 if (!valid_in_sets (set1, set2, expr, block))
2118 bitmap_remove_from_set (set1, expr);
2119 }
2120 VEC_free (pre_expr, heap, exprs);
2121 }
2122
2123 /* Clean the set of expressions that are no longer valid in SET. This
2124 means expressions that are made up of values we have no leaders for
2125 in SET. */
2126
2127 static void
2128 clean (bitmap_set_t set, basic_block block)
2129 {
2130 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2131 pre_expr expr;
2132 int i;
2133
2134 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2135 {
2136 if (!valid_in_sets (set, NULL, expr, block))
2137 bitmap_remove_from_set (set, expr);
2138 }
2139 VEC_free (pre_expr, heap, exprs);
2140 }
2141
2142 /* Clean the set of expressions that are no longer valid in SET because
2143 they are clobbered in BLOCK. */
2144
2145 static void
2146 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2147 {
2148 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2149 pre_expr expr;
2150 int i;
2151
2152 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2153 {
2154 vn_reference_t ref;
2155 if (expr->kind != REFERENCE)
2156 continue;
2157
2158 ref = PRE_EXPR_REFERENCE (expr);
2159 if (ref->vuse)
2160 {
2161 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2162 if (!gimple_nop_p (def_stmt)
2163 && ((gimple_bb (def_stmt) != block
2164 && !dominated_by_p (CDI_DOMINATORS,
2165 block, gimple_bb (def_stmt)))
2166 || (gimple_bb (def_stmt) == block
2167 && value_dies_in_block_x (expr, block))))
2168 bitmap_remove_from_set (set, expr);
2169 }
2170 }
2171 VEC_free (pre_expr, heap, exprs);
2172 }
2173
2174 static sbitmap has_abnormal_preds;
2175
2176 /* List of blocks that may have changed during ANTIC computation and
2177 thus need to be iterated over. */
2178
2179 static sbitmap changed_blocks;
2180
2181 /* Decide whether to defer a block for a later iteration, or PHI
2182 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2183 should defer the block, and true if we processed it. */
2184
2185 static bool
2186 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2187 basic_block block, basic_block phiblock)
2188 {
2189 if (!BB_VISITED (phiblock))
2190 {
2191 SET_BIT (changed_blocks, block->index);
2192 BB_VISITED (block) = 0;
2193 BB_DEFERRED (block) = 1;
2194 return false;
2195 }
2196 else
2197 phi_translate_set (dest, source, block, phiblock);
2198 return true;
2199 }
2200
2201 /* Compute the ANTIC set for BLOCK.
2202
2203 If succs(BLOCK) > 1 then
2204 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2205 else if succs(BLOCK) == 1 then
2206 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2207
2208 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2209 */
2210
2211 static bool
2212 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2213 {
2214 bool changed = false;
2215 bitmap_set_t S, old, ANTIC_OUT;
2216 bitmap_iterator bi;
2217 unsigned int bii;
2218 edge e;
2219 edge_iterator ei;
2220
2221 old = ANTIC_OUT = S = NULL;
2222 BB_VISITED (block) = 1;
2223
2224 /* If any edges from predecessors are abnormal, antic_in is empty,
2225 so do nothing. */
2226 if (block_has_abnormal_pred_edge)
2227 goto maybe_dump_sets;
2228
2229 old = ANTIC_IN (block);
2230 ANTIC_OUT = bitmap_set_new ();
2231
2232 /* If the block has no successors, ANTIC_OUT is empty. */
2233 if (EDGE_COUNT (block->succs) == 0)
2234 ;
2235 /* If we have one successor, we could have some phi nodes to
2236 translate through. */
2237 else if (single_succ_p (block))
2238 {
2239 basic_block succ_bb = single_succ (block);
2240
2241 /* We trade iterations of the dataflow equations for having to
2242 phi translate the maximal set, which is incredibly slow
2243 (since the maximal set often has 300+ members, even when you
2244 have a small number of blocks).
2245 Basically, we defer the computation of ANTIC for this block
2246 until we have processed it's successor, which will inevitably
2247 have a *much* smaller set of values to phi translate once
2248 clean has been run on it.
2249 The cost of doing this is that we technically perform more
2250 iterations, however, they are lower cost iterations.
2251
2252 Timings for PRE on tramp3d-v4:
2253 without maximal set fix: 11 seconds
2254 with maximal set fix/without deferring: 26 seconds
2255 with maximal set fix/with deferring: 11 seconds
2256 */
2257
2258 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2259 block, succ_bb))
2260 {
2261 changed = true;
2262 goto maybe_dump_sets;
2263 }
2264 }
2265 /* If we have multiple successors, we take the intersection of all of
2266 them. Note that in the case of loop exit phi nodes, we may have
2267 phis to translate through. */
2268 else
2269 {
2270 VEC(basic_block, heap) * worklist;
2271 size_t i;
2272 basic_block bprime, first = NULL;
2273
2274 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2275 FOR_EACH_EDGE (e, ei, block->succs)
2276 {
2277 if (!first
2278 && BB_VISITED (e->dest))
2279 first = e->dest;
2280 else if (BB_VISITED (e->dest))
2281 VEC_quick_push (basic_block, worklist, e->dest);
2282 }
2283
2284 /* Of multiple successors we have to have visited one already. */
2285 if (!first)
2286 {
2287 SET_BIT (changed_blocks, block->index);
2288 BB_VISITED (block) = 0;
2289 BB_DEFERRED (block) = 1;
2290 changed = true;
2291 VEC_free (basic_block, heap, worklist);
2292 goto maybe_dump_sets;
2293 }
2294
2295 if (!gimple_seq_empty_p (phi_nodes (first)))
2296 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2297 else
2298 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2299
2300 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2301 {
2302 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2303 {
2304 bitmap_set_t tmp = bitmap_set_new ();
2305 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2306 bitmap_set_and (ANTIC_OUT, tmp);
2307 bitmap_set_free (tmp);
2308 }
2309 else
2310 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2311 }
2312 VEC_free (basic_block, heap, worklist);
2313 }
2314
2315 /* Prune expressions that are clobbered in block and thus become
2316 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2317 prune_clobbered_mems (ANTIC_OUT, block);
2318
2319 /* Generate ANTIC_OUT - TMP_GEN. */
2320 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2321
2322 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2323 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2324 TMP_GEN (block));
2325
2326 /* Then union in the ANTIC_OUT - TMP_GEN values,
2327 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2328 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2329 bitmap_value_insert_into_set (ANTIC_IN (block),
2330 expression_for_id (bii));
2331
2332 clean (ANTIC_IN (block), block);
2333
2334 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2335 {
2336 changed = true;
2337 SET_BIT (changed_blocks, block->index);
2338 FOR_EACH_EDGE (e, ei, block->preds)
2339 SET_BIT (changed_blocks, e->src->index);
2340 }
2341 else
2342 RESET_BIT (changed_blocks, block->index);
2343
2344 maybe_dump_sets:
2345 if (dump_file && (dump_flags & TDF_DETAILS))
2346 {
2347 if (!BB_DEFERRED (block) || BB_VISITED (block))
2348 {
2349 if (ANTIC_OUT)
2350 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2351
2352 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2353 block->index);
2354
2355 if (S)
2356 print_bitmap_set (dump_file, S, "S", block->index);
2357 }
2358 else
2359 {
2360 fprintf (dump_file,
2361 "Block %d was deferred for a future iteration.\n",
2362 block->index);
2363 }
2364 }
2365 if (old)
2366 bitmap_set_free (old);
2367 if (S)
2368 bitmap_set_free (S);
2369 if (ANTIC_OUT)
2370 bitmap_set_free (ANTIC_OUT);
2371 return changed;
2372 }
2373
2374 /* Compute PARTIAL_ANTIC for BLOCK.
2375
2376 If succs(BLOCK) > 1 then
2377 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2378 in ANTIC_OUT for all succ(BLOCK)
2379 else if succs(BLOCK) == 1 then
2380 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2381
2382 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2383 - ANTIC_IN[BLOCK])
2384
2385 */
2386 static bool
2387 compute_partial_antic_aux (basic_block block,
2388 bool block_has_abnormal_pred_edge)
2389 {
2390 bool changed = false;
2391 bitmap_set_t old_PA_IN;
2392 bitmap_set_t PA_OUT;
2393 edge e;
2394 edge_iterator ei;
2395 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2396
2397 old_PA_IN = PA_OUT = NULL;
2398
2399 /* If any edges from predecessors are abnormal, antic_in is empty,
2400 so do nothing. */
2401 if (block_has_abnormal_pred_edge)
2402 goto maybe_dump_sets;
2403
2404 /* If there are too many partially anticipatable values in the
2405 block, phi_translate_set can take an exponential time: stop
2406 before the translation starts. */
2407 if (max_pa
2408 && single_succ_p (block)
2409 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2410 goto maybe_dump_sets;
2411
2412 old_PA_IN = PA_IN (block);
2413 PA_OUT = bitmap_set_new ();
2414
2415 /* If the block has no successors, ANTIC_OUT is empty. */
2416 if (EDGE_COUNT (block->succs) == 0)
2417 ;
2418 /* If we have one successor, we could have some phi nodes to
2419 translate through. Note that we can't phi translate across DFS
2420 back edges in partial antic, because it uses a union operation on
2421 the successors. For recurrences like IV's, we will end up
2422 generating a new value in the set on each go around (i + 3 (VH.1)
2423 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2424 else if (single_succ_p (block))
2425 {
2426 basic_block succ = single_succ (block);
2427 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2428 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2429 }
2430 /* If we have multiple successors, we take the union of all of
2431 them. */
2432 else
2433 {
2434 VEC(basic_block, heap) * worklist;
2435 size_t i;
2436 basic_block bprime;
2437
2438 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2439 FOR_EACH_EDGE (e, ei, block->succs)
2440 {
2441 if (e->flags & EDGE_DFS_BACK)
2442 continue;
2443 VEC_quick_push (basic_block, worklist, e->dest);
2444 }
2445 if (VEC_length (basic_block, worklist) > 0)
2446 {
2447 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2448 {
2449 unsigned int i;
2450 bitmap_iterator bi;
2451
2452 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2453 bitmap_value_insert_into_set (PA_OUT,
2454 expression_for_id (i));
2455 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2456 {
2457 bitmap_set_t pa_in = bitmap_set_new ();
2458 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2459 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2460 bitmap_value_insert_into_set (PA_OUT,
2461 expression_for_id (i));
2462 bitmap_set_free (pa_in);
2463 }
2464 else
2465 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2466 bitmap_value_insert_into_set (PA_OUT,
2467 expression_for_id (i));
2468 }
2469 }
2470 VEC_free (basic_block, heap, worklist);
2471 }
2472
2473 /* Prune expressions that are clobbered in block and thus become
2474 invalid if translated from PA_OUT to PA_IN. */
2475 prune_clobbered_mems (PA_OUT, block);
2476
2477 /* PA_IN starts with PA_OUT - TMP_GEN.
2478 Then we subtract things from ANTIC_IN. */
2479 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2480
2481 /* For partial antic, we want to put back in the phi results, since
2482 we will properly avoid making them partially antic over backedges. */
2483 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2484 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2485
2486 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2487 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2488
2489 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2490
2491 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2492 {
2493 changed = true;
2494 SET_BIT (changed_blocks, block->index);
2495 FOR_EACH_EDGE (e, ei, block->preds)
2496 SET_BIT (changed_blocks, e->src->index);
2497 }
2498 else
2499 RESET_BIT (changed_blocks, block->index);
2500
2501 maybe_dump_sets:
2502 if (dump_file && (dump_flags & TDF_DETAILS))
2503 {
2504 if (PA_OUT)
2505 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2506
2507 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2508 }
2509 if (old_PA_IN)
2510 bitmap_set_free (old_PA_IN);
2511 if (PA_OUT)
2512 bitmap_set_free (PA_OUT);
2513 return changed;
2514 }
2515
2516 /* Compute ANTIC and partial ANTIC sets. */
2517
2518 static void
2519 compute_antic (void)
2520 {
2521 bool changed = true;
2522 int num_iterations = 0;
2523 basic_block block;
2524 int i;
2525
2526 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2527 We pre-build the map of blocks with incoming abnormal edges here. */
2528 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2529 sbitmap_zero (has_abnormal_preds);
2530
2531 FOR_EACH_BB (block)
2532 {
2533 edge_iterator ei;
2534 edge e;
2535
2536 FOR_EACH_EDGE (e, ei, block->preds)
2537 {
2538 e->flags &= ~EDGE_DFS_BACK;
2539 if (e->flags & EDGE_ABNORMAL)
2540 {
2541 SET_BIT (has_abnormal_preds, block->index);
2542 break;
2543 }
2544 }
2545
2546 BB_VISITED (block) = 0;
2547 BB_DEFERRED (block) = 0;
2548
2549 /* While we are here, give empty ANTIC_IN sets to each block. */
2550 ANTIC_IN (block) = bitmap_set_new ();
2551 PA_IN (block) = bitmap_set_new ();
2552 }
2553
2554 /* At the exit block we anticipate nothing. */
2555 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2556 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2557 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2558
2559 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2560 sbitmap_ones (changed_blocks);
2561 while (changed)
2562 {
2563 if (dump_file && (dump_flags & TDF_DETAILS))
2564 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2565 /* ??? We need to clear our PHI translation cache here as the
2566 ANTIC sets shrink and we restrict valid translations to
2567 those having operands with leaders in ANTIC. Same below
2568 for PA ANTIC computation. */
2569 num_iterations++;
2570 changed = false;
2571 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2572 {
2573 if (TEST_BIT (changed_blocks, postorder[i]))
2574 {
2575 basic_block block = BASIC_BLOCK (postorder[i]);
2576 changed |= compute_antic_aux (block,
2577 TEST_BIT (has_abnormal_preds,
2578 block->index));
2579 }
2580 }
2581 /* Theoretically possible, but *highly* unlikely. */
2582 gcc_checking_assert (num_iterations < 500);
2583 }
2584
2585 statistics_histogram_event (cfun, "compute_antic iterations",
2586 num_iterations);
2587
2588 if (do_partial_partial)
2589 {
2590 sbitmap_ones (changed_blocks);
2591 mark_dfs_back_edges ();
2592 num_iterations = 0;
2593 changed = true;
2594 while (changed)
2595 {
2596 if (dump_file && (dump_flags & TDF_DETAILS))
2597 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2598 num_iterations++;
2599 changed = false;
2600 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2601 {
2602 if (TEST_BIT (changed_blocks, postorder[i]))
2603 {
2604 basic_block block = BASIC_BLOCK (postorder[i]);
2605 changed
2606 |= compute_partial_antic_aux (block,
2607 TEST_BIT (has_abnormal_preds,
2608 block->index));
2609 }
2610 }
2611 /* Theoretically possible, but *highly* unlikely. */
2612 gcc_checking_assert (num_iterations < 500);
2613 }
2614 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2615 num_iterations);
2616 }
2617 sbitmap_free (has_abnormal_preds);
2618 sbitmap_free (changed_blocks);
2619 }
2620
2621 /* Return true if we can value number the call in STMT. This is true
2622 if we have a pure or constant call to a real function. */
2623
2624 static bool
2625 can_value_number_call (gimple stmt)
2626 {
2627 if (gimple_call_internal_p (stmt))
2628 return false;
2629 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2630 return true;
2631 return false;
2632 }
2633
2634 /* Return true if OP is a tree which we can perform PRE on.
2635 This may not match the operations we can value number, but in
2636 a perfect world would. */
2637
2638 static bool
2639 can_PRE_operation (tree op)
2640 {
2641 return UNARY_CLASS_P (op)
2642 || BINARY_CLASS_P (op)
2643 || COMPARISON_CLASS_P (op)
2644 || TREE_CODE (op) == MEM_REF
2645 || TREE_CODE (op) == COMPONENT_REF
2646 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2647 || TREE_CODE (op) == CALL_EXPR
2648 || TREE_CODE (op) == ARRAY_REF;
2649 }
2650
2651
2652 /* Inserted expressions are placed onto this worklist, which is used
2653 for performing quick dead code elimination of insertions we made
2654 that didn't turn out to be necessary. */
2655 static bitmap inserted_exprs;
2656
2657 /* Pool allocated fake store expressions are placed onto this
2658 worklist, which, after performing dead code elimination, is walked
2659 to see which expressions need to be put into GC'able memory */
2660 static VEC(gimple, heap) *need_creation;
2661
2662 /* The actual worker for create_component_ref_by_pieces. */
2663
2664 static tree
2665 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2666 unsigned int *operand, gimple_seq *stmts,
2667 gimple domstmt)
2668 {
2669 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2670 *operand);
2671 tree genop;
2672 ++*operand;
2673 switch (currop->opcode)
2674 {
2675 case CALL_EXPR:
2676 {
2677 tree folded, sc = NULL_TREE;
2678 unsigned int nargs = 0;
2679 tree fn, *args;
2680 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2681 fn = currop->op0;
2682 else
2683 {
2684 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2685 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2686 if (!fn)
2687 return NULL_TREE;
2688 }
2689 if (currop->op1)
2690 {
2691 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2692 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2693 if (!sc)
2694 return NULL_TREE;
2695 }
2696 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2697 ref->operands) - 1);
2698 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2699 {
2700 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2701 operand, stmts,
2702 domstmt);
2703 if (!args[nargs])
2704 {
2705 free (args);
2706 return NULL_TREE;
2707 }
2708 nargs++;
2709 }
2710 folded = build_call_array (currop->type,
2711 (TREE_CODE (fn) == FUNCTION_DECL
2712 ? build_fold_addr_expr (fn) : fn),
2713 nargs, args);
2714 free (args);
2715 if (sc)
2716 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2717 return folded;
2718 }
2719 break;
2720 case MEM_REF:
2721 {
2722 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2723 stmts, domstmt);
2724 tree offset = currop->op0;
2725 if (!baseop)
2726 return NULL_TREE;
2727 if (TREE_CODE (baseop) == ADDR_EXPR
2728 && handled_component_p (TREE_OPERAND (baseop, 0)))
2729 {
2730 HOST_WIDE_INT off;
2731 tree base;
2732 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2733 &off);
2734 gcc_assert (base);
2735 offset = int_const_binop (PLUS_EXPR, offset,
2736 build_int_cst (TREE_TYPE (offset),
2737 off));
2738 baseop = build_fold_addr_expr (base);
2739 }
2740 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2741 }
2742 break;
2743 case TARGET_MEM_REF:
2744 {
2745 pre_expr op0expr, op1expr;
2746 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2747 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2748 ++*operand);
2749 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2750 stmts, domstmt);
2751 if (!baseop)
2752 return NULL_TREE;
2753 if (currop->op0)
2754 {
2755 op0expr = get_or_alloc_expr_for (currop->op0);
2756 genop0 = find_or_generate_expression (block, op0expr,
2757 stmts, domstmt);
2758 if (!genop0)
2759 return NULL_TREE;
2760 }
2761 if (nextop->op0)
2762 {
2763 op1expr = get_or_alloc_expr_for (nextop->op0);
2764 genop1 = find_or_generate_expression (block, op1expr,
2765 stmts, domstmt);
2766 if (!genop1)
2767 return NULL_TREE;
2768 }
2769 return build5 (TARGET_MEM_REF, currop->type,
2770 baseop, currop->op2, genop0, currop->op1, genop1);
2771 }
2772 break;
2773 case ADDR_EXPR:
2774 if (currop->op0)
2775 {
2776 gcc_assert (is_gimple_min_invariant (currop->op0));
2777 return currop->op0;
2778 }
2779 /* Fallthrough. */
2780 case REALPART_EXPR:
2781 case IMAGPART_EXPR:
2782 case VIEW_CONVERT_EXPR:
2783 {
2784 tree folded;
2785 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2786 operand,
2787 stmts, domstmt);
2788 if (!genop0)
2789 return NULL_TREE;
2790 folded = fold_build1 (currop->opcode, currop->type,
2791 genop0);
2792 return folded;
2793 }
2794 break;
2795 case WITH_SIZE_EXPR:
2796 {
2797 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2798 stmts, domstmt);
2799 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2800 tree genop1;
2801
2802 if (!genop0)
2803 return NULL_TREE;
2804
2805 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2806 if (!genop1)
2807 return NULL_TREE;
2808
2809 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2810 }
2811 break;
2812 case BIT_FIELD_REF:
2813 {
2814 tree folded;
2815 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2816 stmts, domstmt);
2817 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2818 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2819 tree genop1;
2820 tree genop2;
2821
2822 if (!genop0)
2823 return NULL_TREE;
2824 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2825 if (!genop1)
2826 return NULL_TREE;
2827 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2828 if (!genop2)
2829 return NULL_TREE;
2830 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2831 genop2);
2832 return folded;
2833 }
2834
2835 /* For array ref vn_reference_op's, operand 1 of the array ref
2836 is op0 of the reference op and operand 3 of the array ref is
2837 op1. */
2838 case ARRAY_RANGE_REF:
2839 case ARRAY_REF:
2840 {
2841 tree genop0;
2842 tree genop1 = currop->op0;
2843 pre_expr op1expr;
2844 tree genop2 = currop->op1;
2845 pre_expr op2expr;
2846 tree genop3 = currop->op2;
2847 pre_expr op3expr;
2848 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2849 stmts, domstmt);
2850 if (!genop0)
2851 return NULL_TREE;
2852 op1expr = get_or_alloc_expr_for (genop1);
2853 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2854 if (!genop1)
2855 return NULL_TREE;
2856 if (genop2)
2857 {
2858 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2859 /* Drop zero minimum index if redundant. */
2860 if (integer_zerop (genop2)
2861 && (!domain_type
2862 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2863 genop2 = NULL_TREE;
2864 else
2865 {
2866 op2expr = get_or_alloc_expr_for (genop2);
2867 genop2 = find_or_generate_expression (block, op2expr, stmts,
2868 domstmt);
2869 if (!genop2)
2870 return NULL_TREE;
2871 }
2872 }
2873 if (genop3)
2874 {
2875 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2876 /* We can't always put a size in units of the element alignment
2877 here as the element alignment may be not visible. See
2878 PR43783. Simply drop the element size for constant
2879 sizes. */
2880 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2881 genop3 = NULL_TREE;
2882 else
2883 {
2884 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2885 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2886 op3expr = get_or_alloc_expr_for (genop3);
2887 genop3 = find_or_generate_expression (block, op3expr, stmts,
2888 domstmt);
2889 if (!genop3)
2890 return NULL_TREE;
2891 }
2892 }
2893 return build4 (currop->opcode, currop->type, genop0, genop1,
2894 genop2, genop3);
2895 }
2896 case COMPONENT_REF:
2897 {
2898 tree op0;
2899 tree op1;
2900 tree genop2 = currop->op1;
2901 pre_expr op2expr;
2902 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2903 stmts, domstmt);
2904 if (!op0)
2905 return NULL_TREE;
2906 /* op1 should be a FIELD_DECL, which are represented by
2907 themselves. */
2908 op1 = currop->op0;
2909 if (genop2)
2910 {
2911 op2expr = get_or_alloc_expr_for (genop2);
2912 genop2 = find_or_generate_expression (block, op2expr, stmts,
2913 domstmt);
2914 if (!genop2)
2915 return NULL_TREE;
2916 }
2917
2918 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2919 genop2);
2920 }
2921 break;
2922 case SSA_NAME:
2923 {
2924 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2925 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2926 return genop;
2927 }
2928 case STRING_CST:
2929 case INTEGER_CST:
2930 case COMPLEX_CST:
2931 case VECTOR_CST:
2932 case REAL_CST:
2933 case CONSTRUCTOR:
2934 case VAR_DECL:
2935 case PARM_DECL:
2936 case CONST_DECL:
2937 case RESULT_DECL:
2938 case FUNCTION_DECL:
2939 return currop->op0;
2940
2941 default:
2942 gcc_unreachable ();
2943 }
2944 }
2945
2946 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2947 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2948 trying to rename aggregates into ssa form directly, which is a no no.
2949
2950 Thus, this routine doesn't create temporaries, it just builds a
2951 single access expression for the array, calling
2952 find_or_generate_expression to build the innermost pieces.
2953
2954 This function is a subroutine of create_expression_by_pieces, and
2955 should not be called on it's own unless you really know what you
2956 are doing. */
2957
2958 static tree
2959 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2960 gimple_seq *stmts, gimple domstmt)
2961 {
2962 unsigned int op = 0;
2963 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2964 }
2965
2966 /* Find a leader for an expression, or generate one using
2967 create_expression_by_pieces if it's ANTIC but
2968 complex.
2969 BLOCK is the basic_block we are looking for leaders in.
2970 EXPR is the expression to find a leader or generate for.
2971 STMTS is the statement list to put the inserted expressions on.
2972 Returns the SSA_NAME of the LHS of the generated expression or the
2973 leader.
2974 DOMSTMT if non-NULL is a statement that should be dominated by
2975 all uses in the generated expression. If DOMSTMT is non-NULL this
2976 routine can fail and return NULL_TREE. Otherwise it will assert
2977 on failure. */
2978
2979 static tree
2980 find_or_generate_expression (basic_block block, pre_expr expr,
2981 gimple_seq *stmts, gimple domstmt)
2982 {
2983 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2984 get_expr_value_id (expr), domstmt);
2985 tree genop = NULL;
2986 if (leader)
2987 {
2988 if (leader->kind == NAME)
2989 genop = PRE_EXPR_NAME (leader);
2990 else if (leader->kind == CONSTANT)
2991 genop = PRE_EXPR_CONSTANT (leader);
2992 }
2993
2994 /* If it's still NULL, it must be a complex expression, so generate
2995 it recursively. Not so if inserting expressions for values generated
2996 by SCCVN. */
2997 if (genop == NULL
2998 && !domstmt)
2999 {
3000 bitmap_set_t exprset;
3001 unsigned int lookfor = get_expr_value_id (expr);
3002 bool handled = false;
3003 bitmap_iterator bi;
3004 unsigned int i;
3005
3006 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3007 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3008 {
3009 pre_expr temp = expression_for_id (i);
3010 if (temp->kind != NAME)
3011 {
3012 handled = true;
3013 genop = create_expression_by_pieces (block, temp, stmts,
3014 domstmt,
3015 get_expr_type (expr));
3016 break;
3017 }
3018 }
3019 if (!handled && domstmt)
3020 return NULL_TREE;
3021
3022 gcc_assert (handled);
3023 }
3024 return genop;
3025 }
3026
3027 #define NECESSARY GF_PLF_1
3028
3029 /* Create an expression in pieces, so that we can handle very complex
3030 expressions that may be ANTIC, but not necessary GIMPLE.
3031 BLOCK is the basic block the expression will be inserted into,
3032 EXPR is the expression to insert (in value form)
3033 STMTS is a statement list to append the necessary insertions into.
3034
3035 This function will die if we hit some value that shouldn't be
3036 ANTIC but is (IE there is no leader for it, or its components).
3037 This function may also generate expressions that are themselves
3038 partially or fully redundant. Those that are will be either made
3039 fully redundant during the next iteration of insert (for partially
3040 redundant ones), or eliminated by eliminate (for fully redundant
3041 ones).
3042
3043 If DOMSTMT is non-NULL then we make sure that all uses in the
3044 expressions dominate that statement. In this case the function
3045 can return NULL_TREE to signal failure. */
3046
3047 static tree
3048 create_expression_by_pieces (basic_block block, pre_expr expr,
3049 gimple_seq *stmts, gimple domstmt, tree type)
3050 {
3051 tree temp, name;
3052 tree folded;
3053 gimple_seq forced_stmts = NULL;
3054 unsigned int value_id;
3055 gimple_stmt_iterator gsi;
3056 tree exprtype = type ? type : get_expr_type (expr);
3057 pre_expr nameexpr;
3058 gimple newstmt;
3059
3060 switch (expr->kind)
3061 {
3062 /* We may hit the NAME/CONSTANT case if we have to convert types
3063 that value numbering saw through. */
3064 case NAME:
3065 folded = PRE_EXPR_NAME (expr);
3066 break;
3067 case CONSTANT:
3068 folded = PRE_EXPR_CONSTANT (expr);
3069 break;
3070 case REFERENCE:
3071 {
3072 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3073 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3074 }
3075 break;
3076 case NARY:
3077 {
3078 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3079 tree genop[4];
3080 unsigned i;
3081 for (i = 0; i < nary->length; ++i)
3082 {
3083 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3084 genop[i] = find_or_generate_expression (block, op,
3085 stmts, domstmt);
3086 if (!genop[i])
3087 return NULL_TREE;
3088 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3089 may have conversions stripped. */
3090 if (nary->opcode == POINTER_PLUS_EXPR)
3091 {
3092 if (i == 0)
3093 genop[i] = fold_convert (nary->type, genop[i]);
3094 else if (i == 1)
3095 genop[i] = convert_to_ptrofftype (genop[i]);
3096 }
3097 else
3098 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3099 }
3100 if (nary->opcode == CONSTRUCTOR)
3101 {
3102 VEC(constructor_elt,gc) *elts = NULL;
3103 for (i = 0; i < nary->length; ++i)
3104 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3105 folded = build_constructor (nary->type, elts);
3106 }
3107 else
3108 {
3109 switch (nary->length)
3110 {
3111 case 1:
3112 folded = fold_build1 (nary->opcode, nary->type,
3113 genop[0]);
3114 break;
3115 case 2:
3116 folded = fold_build2 (nary->opcode, nary->type,
3117 genop[0], genop[1]);
3118 break;
3119 case 3:
3120 folded = fold_build3 (nary->opcode, nary->type,
3121 genop[0], genop[1], genop[3]);
3122 break;
3123 default:
3124 gcc_unreachable ();
3125 }
3126 }
3127 }
3128 break;
3129 default:
3130 return NULL_TREE;
3131 }
3132
3133 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3134 folded = fold_convert (exprtype, folded);
3135
3136 /* Force the generated expression to be a sequence of GIMPLE
3137 statements.
3138 We have to call unshare_expr because force_gimple_operand may
3139 modify the tree we pass to it. */
3140 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3141 false, NULL);
3142
3143 /* If we have any intermediate expressions to the value sets, add them
3144 to the value sets and chain them in the instruction stream. */
3145 if (forced_stmts)
3146 {
3147 gsi = gsi_start (forced_stmts);
3148 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3149 {
3150 gimple stmt = gsi_stmt (gsi);
3151 tree forcedname = gimple_get_lhs (stmt);
3152 pre_expr nameexpr;
3153
3154 if (TREE_CODE (forcedname) == SSA_NAME)
3155 {
3156 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3157 VN_INFO_GET (forcedname)->valnum = forcedname;
3158 VN_INFO (forcedname)->value_id = get_next_value_id ();
3159 nameexpr = get_or_alloc_expr_for_name (forcedname);
3160 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3161 if (!in_fre)
3162 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3163 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3164 }
3165 mark_symbols_for_renaming (stmt);
3166 }
3167 gimple_seq_add_seq (stmts, forced_stmts);
3168 }
3169
3170 /* Build and insert the assignment of the end result to the temporary
3171 that we will return. */
3172 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3173 pretemp = create_tmp_reg (exprtype, "pretmp");
3174
3175 temp = pretemp;
3176 add_referenced_var (temp);
3177
3178 newstmt = gimple_build_assign (temp, folded);
3179 name = make_ssa_name (temp, newstmt);
3180 gimple_assign_set_lhs (newstmt, name);
3181 gimple_set_plf (newstmt, NECESSARY, false);
3182
3183 gimple_seq_add_stmt (stmts, newstmt);
3184 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3185
3186 /* All the symbols in NEWEXPR should be put into SSA form. */
3187 mark_symbols_for_renaming (newstmt);
3188
3189 /* Fold the last statement. */
3190 gsi = gsi_last (*stmts);
3191 if (fold_stmt_inplace (&gsi))
3192 update_stmt (gsi_stmt (gsi));
3193
3194 /* Add a value number to the temporary.
3195 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3196 we are creating the expression by pieces, and this particular piece of
3197 the expression may have been represented. There is no harm in replacing
3198 here. */
3199 VN_INFO_GET (name)->valnum = name;
3200 value_id = get_expr_value_id (expr);
3201 VN_INFO (name)->value_id = value_id;
3202 nameexpr = get_or_alloc_expr_for_name (name);
3203 add_to_value (value_id, nameexpr);
3204 if (NEW_SETS (block))
3205 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3206 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3207
3208 pre_stats.insertions++;
3209 if (dump_file && (dump_flags & TDF_DETAILS))
3210 {
3211 fprintf (dump_file, "Inserted ");
3212 print_gimple_stmt (dump_file, newstmt, 0, 0);
3213 fprintf (dump_file, " in predecessor %d\n", block->index);
3214 }
3215
3216 return name;
3217 }
3218
3219
3220 /* Returns true if we want to inhibit the insertions of PHI nodes
3221 for the given EXPR for basic block BB (a member of a loop).
3222 We want to do this, when we fear that the induction variable we
3223 create might inhibit vectorization. */
3224
3225 static bool
3226 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3227 {
3228 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3229 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3230 vn_reference_op_t op;
3231 unsigned i;
3232
3233 /* If we aren't going to vectorize we don't inhibit anything. */
3234 if (!flag_tree_vectorize)
3235 return false;
3236
3237 /* Otherwise we inhibit the insertion when the address of the
3238 memory reference is a simple induction variable. In other
3239 cases the vectorizer won't do anything anyway (either it's
3240 loop invariant or a complicated expression). */
3241 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3242 {
3243 switch (op->opcode)
3244 {
3245 case CALL_EXPR:
3246 /* Calls are not a problem. */
3247 return false;
3248
3249 case ARRAY_REF:
3250 case ARRAY_RANGE_REF:
3251 if (TREE_CODE (op->op0) != SSA_NAME)
3252 break;
3253 /* Fallthru. */
3254 case SSA_NAME:
3255 {
3256 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3257 affine_iv iv;
3258 /* Default defs are loop invariant. */
3259 if (!defbb)
3260 break;
3261 /* Defined outside this loop, also loop invariant. */
3262 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3263 break;
3264 /* If it's a simple induction variable inhibit insertion,
3265 the vectorizer might be interested in this one. */
3266 if (simple_iv (bb->loop_father, bb->loop_father,
3267 op->op0, &iv, true))
3268 return true;
3269 /* No simple IV, vectorizer can't do anything, hence no
3270 reason to inhibit the transformation for this operand. */
3271 break;
3272 }
3273 default:
3274 break;
3275 }
3276 }
3277 return false;
3278 }
3279
3280 /* Insert the to-be-made-available values of expression EXPRNUM for each
3281 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3282 merge the result with a phi node, given the same value number as
3283 NODE. Return true if we have inserted new stuff. */
3284
3285 static bool
3286 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3287 pre_expr *avail)
3288 {
3289 pre_expr expr = expression_for_id (exprnum);
3290 pre_expr newphi;
3291 unsigned int val = get_expr_value_id (expr);
3292 edge pred;
3293 bool insertions = false;
3294 bool nophi = false;
3295 basic_block bprime;
3296 pre_expr eprime;
3297 edge_iterator ei;
3298 tree type = get_expr_type (expr);
3299 tree temp;
3300 gimple phi;
3301
3302 /* Make sure we aren't creating an induction variable. */
3303 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3304 {
3305 bool firstinsideloop = false;
3306 bool secondinsideloop = false;
3307 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3308 EDGE_PRED (block, 0)->src);
3309 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3310 EDGE_PRED (block, 1)->src);
3311 /* Induction variables only have one edge inside the loop. */
3312 if ((firstinsideloop ^ secondinsideloop)
3313 && (expr->kind != REFERENCE
3314 || inhibit_phi_insertion (block, expr)))
3315 {
3316 if (dump_file && (dump_flags & TDF_DETAILS))
3317 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3318 nophi = true;
3319 }
3320 }
3321
3322 /* Make the necessary insertions. */
3323 FOR_EACH_EDGE (pred, ei, block->preds)
3324 {
3325 gimple_seq stmts = NULL;
3326 tree builtexpr;
3327 bprime = pred->src;
3328 eprime = avail[bprime->index];
3329
3330 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3331 {
3332 builtexpr = create_expression_by_pieces (bprime,
3333 eprime,
3334 &stmts, NULL,
3335 type);
3336 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3337 gsi_insert_seq_on_edge (pred, stmts);
3338 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3339 insertions = true;
3340 }
3341 else if (eprime->kind == CONSTANT)
3342 {
3343 /* Constants may not have the right type, fold_convert
3344 should give us back a constant with the right type.
3345 */
3346 tree constant = PRE_EXPR_CONSTANT (eprime);
3347 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3348 {
3349 tree builtexpr = fold_convert (type, constant);
3350 if (!is_gimple_min_invariant (builtexpr))
3351 {
3352 tree forcedexpr = force_gimple_operand (builtexpr,
3353 &stmts, true,
3354 NULL);
3355 if (!is_gimple_min_invariant (forcedexpr))
3356 {
3357 if (forcedexpr != builtexpr)
3358 {
3359 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3360 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3361 }
3362 if (stmts)
3363 {
3364 gimple_stmt_iterator gsi;
3365 gsi = gsi_start (stmts);
3366 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3367 {
3368 gimple stmt = gsi_stmt (gsi);
3369 tree lhs = gimple_get_lhs (stmt);
3370 if (TREE_CODE (lhs) == SSA_NAME)
3371 bitmap_set_bit (inserted_exprs,
3372 SSA_NAME_VERSION (lhs));
3373 gimple_set_plf (stmt, NECESSARY, false);
3374 }
3375 gsi_insert_seq_on_edge (pred, stmts);
3376 }
3377 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3378 }
3379 }
3380 else
3381 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3382 }
3383 }
3384 else if (eprime->kind == NAME)
3385 {
3386 /* We may have to do a conversion because our value
3387 numbering can look through types in certain cases, but
3388 our IL requires all operands of a phi node have the same
3389 type. */
3390 tree name = PRE_EXPR_NAME (eprime);
3391 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3392 {
3393 tree builtexpr;
3394 tree forcedexpr;
3395 builtexpr = fold_convert (type, name);
3396 forcedexpr = force_gimple_operand (builtexpr,
3397 &stmts, true,
3398 NULL);
3399
3400 if (forcedexpr != name)
3401 {
3402 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3403 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3404 }
3405
3406 if (stmts)
3407 {
3408 gimple_stmt_iterator gsi;
3409 gsi = gsi_start (stmts);
3410 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3411 {
3412 gimple stmt = gsi_stmt (gsi);
3413 tree lhs = gimple_get_lhs (stmt);
3414 if (TREE_CODE (lhs) == SSA_NAME)
3415 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3416 gimple_set_plf (stmt, NECESSARY, false);
3417 }
3418 gsi_insert_seq_on_edge (pred, stmts);
3419 }
3420 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3421 }
3422 }
3423 }
3424 /* If we didn't want a phi node, and we made insertions, we still have
3425 inserted new stuff, and thus return true. If we didn't want a phi node,
3426 and didn't make insertions, we haven't added anything new, so return
3427 false. */
3428 if (nophi && insertions)
3429 return true;
3430 else if (nophi && !insertions)
3431 return false;
3432
3433 /* Now build a phi for the new variable. */
3434 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3435 prephitemp = create_tmp_var (type, "prephitmp");
3436
3437 temp = prephitemp;
3438 add_referenced_var (temp);
3439
3440 if (TREE_CODE (type) == COMPLEX_TYPE
3441 || TREE_CODE (type) == VECTOR_TYPE)
3442 DECL_GIMPLE_REG_P (temp) = 1;
3443 phi = create_phi_node (temp, block);
3444
3445 gimple_set_plf (phi, NECESSARY, false);
3446 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3447 VN_INFO (gimple_phi_result (phi))->value_id = val;
3448 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3449 FOR_EACH_EDGE (pred, ei, block->preds)
3450 {
3451 pre_expr ae = avail[pred->src->index];
3452 gcc_assert (get_expr_type (ae) == type
3453 || useless_type_conversion_p (type, get_expr_type (ae)));
3454 if (ae->kind == CONSTANT)
3455 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3456 else
3457 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3458 UNKNOWN_LOCATION);
3459 }
3460
3461 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3462 add_to_value (val, newphi);
3463
3464 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3465 this insertion, since we test for the existence of this value in PHI_GEN
3466 before proceeding with the partial redundancy checks in insert_aux.
3467
3468 The value may exist in AVAIL_OUT, in particular, it could be represented
3469 by the expression we are trying to eliminate, in which case we want the
3470 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3471 inserted there.
3472
3473 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3474 this block, because if it did, it would have existed in our dominator's
3475 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3476 */
3477
3478 bitmap_insert_into_set (PHI_GEN (block), newphi);
3479 bitmap_value_replace_in_set (AVAIL_OUT (block),
3480 newphi);
3481 bitmap_insert_into_set (NEW_SETS (block),
3482 newphi);
3483
3484 if (dump_file && (dump_flags & TDF_DETAILS))
3485 {
3486 fprintf (dump_file, "Created phi ");
3487 print_gimple_stmt (dump_file, phi, 0, 0);
3488 fprintf (dump_file, " in block %d\n", block->index);
3489 }
3490 pre_stats.phis++;
3491 return true;
3492 }
3493
3494
3495
3496 /* Perform insertion of partially redundant values.
3497 For BLOCK, do the following:
3498 1. Propagate the NEW_SETS of the dominator into the current block.
3499 If the block has multiple predecessors,
3500 2a. Iterate over the ANTIC expressions for the block to see if
3501 any of them are partially redundant.
3502 2b. If so, insert them into the necessary predecessors to make
3503 the expression fully redundant.
3504 2c. Insert a new PHI merging the values of the predecessors.
3505 2d. Insert the new PHI, and the new expressions, into the
3506 NEW_SETS set.
3507 3. Recursively call ourselves on the dominator children of BLOCK.
3508
3509 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3510 do_regular_insertion and do_partial_insertion.
3511
3512 */
3513
3514 static bool
3515 do_regular_insertion (basic_block block, basic_block dom)
3516 {
3517 bool new_stuff = false;
3518 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3519 pre_expr expr;
3520 int i;
3521
3522 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3523 {
3524 if (expr->kind != NAME)
3525 {
3526 pre_expr *avail;
3527 unsigned int val;
3528 bool by_some = false;
3529 bool cant_insert = false;
3530 bool all_same = true;
3531 pre_expr first_s = NULL;
3532 edge pred;
3533 basic_block bprime;
3534 pre_expr eprime = NULL;
3535 edge_iterator ei;
3536 pre_expr edoubleprime = NULL;
3537 bool do_insertion = false;
3538
3539 val = get_expr_value_id (expr);
3540 if (bitmap_set_contains_value (PHI_GEN (block), val))
3541 continue;
3542 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3543 {
3544 if (dump_file && (dump_flags & TDF_DETAILS))
3545 fprintf (dump_file, "Found fully redundant value\n");
3546 continue;
3547 }
3548
3549 avail = XCNEWVEC (pre_expr, last_basic_block);
3550 FOR_EACH_EDGE (pred, ei, block->preds)
3551 {
3552 unsigned int vprime;
3553
3554 /* We should never run insertion for the exit block
3555 and so not come across fake pred edges. */
3556 gcc_assert (!(pred->flags & EDGE_FAKE));
3557 bprime = pred->src;
3558 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3559 bprime, block);
3560
3561 /* eprime will generally only be NULL if the
3562 value of the expression, translated
3563 through the PHI for this predecessor, is
3564 undefined. If that is the case, we can't
3565 make the expression fully redundant,
3566 because its value is undefined along a
3567 predecessor path. We can thus break out
3568 early because it doesn't matter what the
3569 rest of the results are. */
3570 if (eprime == NULL)
3571 {
3572 cant_insert = true;
3573 break;
3574 }
3575
3576 eprime = fully_constant_expression (eprime);
3577 vprime = get_expr_value_id (eprime);
3578 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3579 vprime, NULL);
3580 if (edoubleprime == NULL)
3581 {
3582 avail[bprime->index] = eprime;
3583 all_same = false;
3584 }
3585 else
3586 {
3587 avail[bprime->index] = edoubleprime;
3588 by_some = true;
3589 /* We want to perform insertions to remove a redundancy on
3590 a path in the CFG we want to optimize for speed. */
3591 if (optimize_edge_for_speed_p (pred))
3592 do_insertion = true;
3593 if (first_s == NULL)
3594 first_s = edoubleprime;
3595 else if (!pre_expr_eq (first_s, edoubleprime))
3596 all_same = false;
3597 }
3598 }
3599 /* If we can insert it, it's not the same value
3600 already existing along every predecessor, and
3601 it's defined by some predecessor, it is
3602 partially redundant. */
3603 if (!cant_insert && !all_same && by_some)
3604 {
3605 if (!do_insertion)
3606 {
3607 if (dump_file && (dump_flags & TDF_DETAILS))
3608 {
3609 fprintf (dump_file, "Skipping partial redundancy for "
3610 "expression ");
3611 print_pre_expr (dump_file, expr);
3612 fprintf (dump_file, " (%04d), no redundancy on to be "
3613 "optimized for speed edge\n", val);
3614 }
3615 }
3616 else if (dbg_cnt (treepre_insert))
3617 {
3618 if (dump_file && (dump_flags & TDF_DETAILS))
3619 {
3620 fprintf (dump_file, "Found partial redundancy for "
3621 "expression ");
3622 print_pre_expr (dump_file, expr);
3623 fprintf (dump_file, " (%04d)\n",
3624 get_expr_value_id (expr));
3625 }
3626 if (insert_into_preds_of_block (block,
3627 get_expression_id (expr),
3628 avail))
3629 new_stuff = true;
3630 }
3631 }
3632 /* If all edges produce the same value and that value is
3633 an invariant, then the PHI has the same value on all
3634 edges. Note this. */
3635 else if (!cant_insert && all_same && eprime
3636 && (edoubleprime->kind == CONSTANT
3637 || edoubleprime->kind == NAME)
3638 && !value_id_constant_p (val))
3639 {
3640 unsigned int j;
3641 bitmap_iterator bi;
3642 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3643 value_expressions, val);
3644
3645 unsigned int new_val = get_expr_value_id (edoubleprime);
3646 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3647 {
3648 pre_expr expr = expression_for_id (j);
3649
3650 if (expr->kind == NAME)
3651 {
3652 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3653 /* Just reset the value id and valnum so it is
3654 the same as the constant we have discovered. */
3655 if (edoubleprime->kind == CONSTANT)
3656 {
3657 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3658 pre_stats.constified++;
3659 }
3660 else
3661 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3662 info->value_id = new_val;
3663 }
3664 }
3665 }
3666 free (avail);
3667 }
3668 }
3669
3670 VEC_free (pre_expr, heap, exprs);
3671 return new_stuff;
3672 }
3673
3674
3675 /* Perform insertion for partially anticipatable expressions. There
3676 is only one case we will perform insertion for these. This case is
3677 if the expression is partially anticipatable, and fully available.
3678 In this case, we know that putting it earlier will enable us to
3679 remove the later computation. */
3680
3681
3682 static bool
3683 do_partial_partial_insertion (basic_block block, basic_block dom)
3684 {
3685 bool new_stuff = false;
3686 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3687 pre_expr expr;
3688 int i;
3689
3690 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3691 {
3692 if (expr->kind != NAME)
3693 {
3694 pre_expr *avail;
3695 unsigned int val;
3696 bool by_all = true;
3697 bool cant_insert = false;
3698 edge pred;
3699 basic_block bprime;
3700 pre_expr eprime = NULL;
3701 edge_iterator ei;
3702
3703 val = get_expr_value_id (expr);
3704 if (bitmap_set_contains_value (PHI_GEN (block), val))
3705 continue;
3706 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3707 continue;
3708
3709 avail = XCNEWVEC (pre_expr, last_basic_block);
3710 FOR_EACH_EDGE (pred, ei, block->preds)
3711 {
3712 unsigned int vprime;
3713 pre_expr edoubleprime;
3714
3715 /* We should never run insertion for the exit block
3716 and so not come across fake pred edges. */
3717 gcc_assert (!(pred->flags & EDGE_FAKE));
3718 bprime = pred->src;
3719 eprime = phi_translate (expr, ANTIC_IN (block),
3720 PA_IN (block),
3721 bprime, block);
3722
3723 /* eprime will generally only be NULL if the
3724 value of the expression, translated
3725 through the PHI for this predecessor, is
3726 undefined. If that is the case, we can't
3727 make the expression fully redundant,
3728 because its value is undefined along a
3729 predecessor path. We can thus break out
3730 early because it doesn't matter what the
3731 rest of the results are. */
3732 if (eprime == NULL)
3733 {
3734 cant_insert = true;
3735 break;
3736 }
3737
3738 eprime = fully_constant_expression (eprime);
3739 vprime = get_expr_value_id (eprime);
3740 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3741 vprime, NULL);
3742 if (edoubleprime == NULL)
3743 {
3744 by_all = false;
3745 break;
3746 }
3747 else
3748 avail[bprime->index] = edoubleprime;
3749 }
3750
3751 /* If we can insert it, it's not the same value
3752 already existing along every predecessor, and
3753 it's defined by some predecessor, it is
3754 partially redundant. */
3755 if (!cant_insert && by_all)
3756 {
3757 edge succ;
3758 bool do_insertion = false;
3759
3760 /* Insert only if we can remove a later expression on a path
3761 that we want to optimize for speed.
3762 The phi node that we will be inserting in BLOCK is not free,
3763 and inserting it for the sake of !optimize_for_speed successor
3764 may cause regressions on the speed path. */
3765 FOR_EACH_EDGE (succ, ei, block->succs)
3766 {
3767 if (bitmap_set_contains_value (PA_IN (succ->dest), val))
3768 {
3769 if (optimize_edge_for_speed_p (succ))
3770 do_insertion = true;
3771 }
3772 }
3773
3774 if (!do_insertion)
3775 {
3776 if (dump_file && (dump_flags & TDF_DETAILS))
3777 {
3778 fprintf (dump_file, "Skipping partial partial redundancy "
3779 "for expression ");
3780 print_pre_expr (dump_file, expr);
3781 fprintf (dump_file, " (%04d), not partially anticipated "
3782 "on any to be optimized for speed edges\n", val);
3783 }
3784 }
3785 else if (dbg_cnt (treepre_insert))
3786 {
3787 pre_stats.pa_insert++;
3788 if (dump_file && (dump_flags & TDF_DETAILS))
3789 {
3790 fprintf (dump_file, "Found partial partial redundancy "
3791 "for expression ");
3792 print_pre_expr (dump_file, expr);
3793 fprintf (dump_file, " (%04d)\n",
3794 get_expr_value_id (expr));
3795 }
3796 if (insert_into_preds_of_block (block,
3797 get_expression_id (expr),
3798 avail))
3799 new_stuff = true;
3800 }
3801 }
3802 free (avail);
3803 }
3804 }
3805
3806 VEC_free (pre_expr, heap, exprs);
3807 return new_stuff;
3808 }
3809
3810 static bool
3811 insert_aux (basic_block block)
3812 {
3813 basic_block son;
3814 bool new_stuff = false;
3815
3816 if (block)
3817 {
3818 basic_block dom;
3819 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3820 if (dom)
3821 {
3822 unsigned i;
3823 bitmap_iterator bi;
3824 bitmap_set_t newset = NEW_SETS (dom);
3825 if (newset)
3826 {
3827 /* Note that we need to value_replace both NEW_SETS, and
3828 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3829 represented by some non-simple expression here that we want
3830 to replace it with. */
3831 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3832 {
3833 pre_expr expr = expression_for_id (i);
3834 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3835 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3836 }
3837 }
3838 if (!single_pred_p (block))
3839 {
3840 new_stuff |= do_regular_insertion (block, dom);
3841 if (do_partial_partial)
3842 new_stuff |= do_partial_partial_insertion (block, dom);
3843 }
3844 }
3845 }
3846 for (son = first_dom_son (CDI_DOMINATORS, block);
3847 son;
3848 son = next_dom_son (CDI_DOMINATORS, son))
3849 {
3850 new_stuff |= insert_aux (son);
3851 }
3852
3853 return new_stuff;
3854 }
3855
3856 /* Perform insertion of partially redundant values. */
3857
3858 static void
3859 insert (void)
3860 {
3861 bool new_stuff = true;
3862 basic_block bb;
3863 int num_iterations = 0;
3864
3865 FOR_ALL_BB (bb)
3866 NEW_SETS (bb) = bitmap_set_new ();
3867
3868 while (new_stuff)
3869 {
3870 num_iterations++;
3871 if (dump_file && dump_flags & TDF_DETAILS)
3872 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3873 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3874 }
3875 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3876 }
3877
3878
3879 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3880
3881 static void
3882 add_to_exp_gen (basic_block block, tree op)
3883 {
3884 if (!in_fre)
3885 {
3886 pre_expr result;
3887 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3888 return;
3889 result = get_or_alloc_expr_for_name (op);
3890 bitmap_value_insert_into_set (EXP_GEN (block), result);
3891 }
3892 }
3893
3894 /* Create value ids for PHI in BLOCK. */
3895
3896 static void
3897 make_values_for_phi (gimple phi, basic_block block)
3898 {
3899 tree result = gimple_phi_result (phi);
3900
3901 /* We have no need for virtual phis, as they don't represent
3902 actual computations. */
3903 if (is_gimple_reg (result))
3904 {
3905 pre_expr e = get_or_alloc_expr_for_name (result);
3906 add_to_value (get_expr_value_id (e), e);
3907 bitmap_insert_into_set (PHI_GEN (block), e);
3908 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3909 if (!in_fre)
3910 {
3911 unsigned i;
3912 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3913 {
3914 tree arg = gimple_phi_arg_def (phi, i);
3915 if (TREE_CODE (arg) == SSA_NAME)
3916 {
3917 e = get_or_alloc_expr_for_name (arg);
3918 add_to_value (get_expr_value_id (e), e);
3919 }
3920 }
3921 }
3922 }
3923 }
3924
3925 /* Compute the AVAIL set for all basic blocks.
3926
3927 This function performs value numbering of the statements in each basic
3928 block. The AVAIL sets are built from information we glean while doing
3929 this value numbering, since the AVAIL sets contain only one entry per
3930 value.
3931
3932 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3933 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3934
3935 static void
3936 compute_avail (void)
3937 {
3938
3939 basic_block block, son;
3940 basic_block *worklist;
3941 size_t sp = 0;
3942 unsigned i;
3943
3944 /* We pretend that default definitions are defined in the entry block.
3945 This includes function arguments and the static chain decl. */
3946 for (i = 1; i < num_ssa_names; ++i)
3947 {
3948 tree name = ssa_name (i);
3949 pre_expr e;
3950 if (!name
3951 || !SSA_NAME_IS_DEFAULT_DEF (name)
3952 || has_zero_uses (name)
3953 || !is_gimple_reg (name))
3954 continue;
3955
3956 e = get_or_alloc_expr_for_name (name);
3957 add_to_value (get_expr_value_id (e), e);
3958 if (!in_fre)
3959 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3960 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3961 }
3962
3963 /* Allocate the worklist. */
3964 worklist = XNEWVEC (basic_block, n_basic_blocks);
3965
3966 /* Seed the algorithm by putting the dominator children of the entry
3967 block on the worklist. */
3968 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3969 son;
3970 son = next_dom_son (CDI_DOMINATORS, son))
3971 worklist[sp++] = son;
3972
3973 /* Loop until the worklist is empty. */
3974 while (sp)
3975 {
3976 gimple_stmt_iterator gsi;
3977 gimple stmt;
3978 basic_block dom;
3979 unsigned int stmt_uid = 1;
3980
3981 /* Pick a block from the worklist. */
3982 block = worklist[--sp];
3983
3984 /* Initially, the set of available values in BLOCK is that of
3985 its immediate dominator. */
3986 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3987 if (dom)
3988 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3989
3990 /* Generate values for PHI nodes. */
3991 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3992 make_values_for_phi (gsi_stmt (gsi), block);
3993
3994 BB_MAY_NOTRETURN (block) = 0;
3995
3996 /* Now compute value numbers and populate value sets with all
3997 the expressions computed in BLOCK. */
3998 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3999 {
4000 ssa_op_iter iter;
4001 tree op;
4002
4003 stmt = gsi_stmt (gsi);
4004 gimple_set_uid (stmt, stmt_uid++);
4005
4006 /* Cache whether the basic-block has any non-visible side-effect
4007 or control flow.
4008 If this isn't a call or it is the last stmt in the
4009 basic-block then the CFG represents things correctly. */
4010 if (is_gimple_call (stmt)
4011 && !stmt_ends_bb_p (stmt))
4012 {
4013 /* Non-looping const functions always return normally.
4014 Otherwise the call might not return or have side-effects
4015 that forbids hoisting possibly trapping expressions
4016 before it. */
4017 int flags = gimple_call_flags (stmt);
4018 if (!(flags & ECF_CONST)
4019 || (flags & ECF_LOOPING_CONST_OR_PURE))
4020 BB_MAY_NOTRETURN (block) = 1;
4021 }
4022
4023 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
4024 {
4025 pre_expr e = get_or_alloc_expr_for_name (op);
4026
4027 add_to_value (get_expr_value_id (e), e);
4028 if (!in_fre)
4029 bitmap_insert_into_set (TMP_GEN (block), e);
4030 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4031 }
4032
4033 if (gimple_has_volatile_ops (stmt)
4034 || stmt_could_throw_p (stmt))
4035 continue;
4036
4037 switch (gimple_code (stmt))
4038 {
4039 case GIMPLE_RETURN:
4040 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4041 add_to_exp_gen (block, op);
4042 continue;
4043
4044 case GIMPLE_CALL:
4045 {
4046 vn_reference_t ref;
4047 unsigned int i;
4048 vn_reference_op_t vro;
4049 pre_expr result = NULL;
4050 VEC(vn_reference_op_s, heap) *ops = NULL;
4051
4052 if (!can_value_number_call (stmt))
4053 continue;
4054
4055 copy_reference_ops_from_call (stmt, &ops);
4056 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4057 gimple_expr_type (stmt),
4058 ops, &ref, VN_NOWALK);
4059 VEC_free (vn_reference_op_s, heap, ops);
4060 if (!ref)
4061 continue;
4062
4063 for (i = 0; VEC_iterate (vn_reference_op_s,
4064 ref->operands, i,
4065 vro); i++)
4066 {
4067 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4068 add_to_exp_gen (block, vro->op0);
4069 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4070 add_to_exp_gen (block, vro->op1);
4071 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4072 add_to_exp_gen (block, vro->op2);
4073 }
4074
4075 /* If the value of the call is not invalidated in
4076 this block until it is computed, add the expression
4077 to EXP_GEN. */
4078 if (!gimple_vuse (stmt)
4079 || gimple_code
4080 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
4081 || gimple_bb (SSA_NAME_DEF_STMT
4082 (gimple_vuse (stmt))) != block)
4083 {
4084 result = (pre_expr) pool_alloc (pre_expr_pool);
4085 result->kind = REFERENCE;
4086 result->id = 0;
4087 PRE_EXPR_REFERENCE (result) = ref;
4088
4089 get_or_alloc_expression_id (result);
4090 add_to_value (get_expr_value_id (result), result);
4091 if (!in_fre)
4092 bitmap_value_insert_into_set (EXP_GEN (block), result);
4093 }
4094 continue;
4095 }
4096
4097 case GIMPLE_ASSIGN:
4098 {
4099 pre_expr result = NULL;
4100 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4101 {
4102 case tcc_unary:
4103 case tcc_binary:
4104 case tcc_comparison:
4105 {
4106 vn_nary_op_t nary;
4107 unsigned int i;
4108
4109 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4110 gimple_assign_rhs_code (stmt),
4111 gimple_expr_type (stmt),
4112 gimple_assign_rhs1_ptr (stmt),
4113 &nary);
4114
4115 if (!nary)
4116 continue;
4117
4118 for (i = 0; i < nary->length; i++)
4119 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4120 add_to_exp_gen (block, nary->op[i]);
4121
4122 result = (pre_expr) pool_alloc (pre_expr_pool);
4123 result->kind = NARY;
4124 result->id = 0;
4125 PRE_EXPR_NARY (result) = nary;
4126 break;
4127 }
4128
4129 case tcc_declaration:
4130 case tcc_reference:
4131 {
4132 vn_reference_t ref;
4133 unsigned int i;
4134 vn_reference_op_t vro;
4135
4136 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4137 gimple_vuse (stmt),
4138 VN_WALK, &ref);
4139 if (!ref)
4140 continue;
4141
4142 for (i = 0; VEC_iterate (vn_reference_op_s,
4143 ref->operands, i,
4144 vro); i++)
4145 {
4146 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4147 add_to_exp_gen (block, vro->op0);
4148 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4149 add_to_exp_gen (block, vro->op1);
4150 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4151 add_to_exp_gen (block, vro->op2);
4152 }
4153
4154 /* If the value of the reference is not invalidated in
4155 this block until it is computed, add the expression
4156 to EXP_GEN. */
4157 if (gimple_vuse (stmt))
4158 {
4159 gimple def_stmt;
4160 bool ok = true;
4161 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
4162 while (!gimple_nop_p (def_stmt)
4163 && gimple_code (def_stmt) != GIMPLE_PHI
4164 && gimple_bb (def_stmt) == block)
4165 {
4166 if (stmt_may_clobber_ref_p
4167 (def_stmt, gimple_assign_rhs1 (stmt)))
4168 {
4169 ok = false;
4170 break;
4171 }
4172 def_stmt
4173 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
4174 }
4175 if (!ok)
4176 continue;
4177 }
4178
4179 result = (pre_expr) pool_alloc (pre_expr_pool);
4180 result->kind = REFERENCE;
4181 result->id = 0;
4182 PRE_EXPR_REFERENCE (result) = ref;
4183 break;
4184 }
4185
4186 default:
4187 /* For any other statement that we don't
4188 recognize, simply add all referenced
4189 SSA_NAMEs to EXP_GEN. */
4190 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4191 add_to_exp_gen (block, op);
4192 continue;
4193 }
4194
4195 get_or_alloc_expression_id (result);
4196 add_to_value (get_expr_value_id (result), result);
4197 if (!in_fre)
4198 bitmap_value_insert_into_set (EXP_GEN (block), result);
4199
4200 continue;
4201 }
4202 default:
4203 break;
4204 }
4205 }
4206
4207 /* Put the dominator children of BLOCK on the worklist of blocks
4208 to compute available sets for. */
4209 for (son = first_dom_son (CDI_DOMINATORS, block);
4210 son;
4211 son = next_dom_son (CDI_DOMINATORS, son))
4212 worklist[sp++] = son;
4213 }
4214
4215 free (worklist);
4216 }
4217
4218 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4219 than the available expressions for it. The insertion point is
4220 right before the first use in STMT. Returns the SSA_NAME that should
4221 be used for replacement. */
4222
4223 static tree
4224 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4225 {
4226 basic_block bb = gimple_bb (stmt);
4227 gimple_stmt_iterator gsi;
4228 gimple_seq stmts = NULL;
4229 tree expr;
4230 pre_expr e;
4231
4232 /* First create a value expression from the expression we want
4233 to insert and associate it with the value handle for SSA_VN. */
4234 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4235 if (e == NULL)
4236 return NULL_TREE;
4237
4238 /* Then use create_expression_by_pieces to generate a valid
4239 expression to insert at this point of the IL stream. */
4240 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4241 if (expr == NULL_TREE)
4242 return NULL_TREE;
4243 gsi = gsi_for_stmt (stmt);
4244 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4245
4246 return expr;
4247 }
4248
4249 /* Eliminate fully redundant computations. */
4250
4251 static unsigned int
4252 eliminate (void)
4253 {
4254 VEC (gimple, heap) *to_remove = NULL;
4255 VEC (gimple, heap) *to_update = NULL;
4256 basic_block b;
4257 unsigned int todo = 0;
4258 gimple_stmt_iterator gsi;
4259 gimple stmt;
4260 unsigned i;
4261
4262 FOR_EACH_BB (b)
4263 {
4264 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4265 {
4266 tree lhs = NULL_TREE;
4267 tree rhs = NULL_TREE;
4268
4269 stmt = gsi_stmt (gsi);
4270
4271 if (gimple_has_lhs (stmt))
4272 lhs = gimple_get_lhs (stmt);
4273
4274 if (gimple_assign_single_p (stmt))
4275 rhs = gimple_assign_rhs1 (stmt);
4276
4277 /* Lookup the RHS of the expression, see if we have an
4278 available computation for it. If so, replace the RHS with
4279 the available computation.
4280
4281 See PR43491.
4282 We don't replace global register variable when it is a the RHS of
4283 a single assign. We do replace local register variable since gcc
4284 does not guarantee local variable will be allocated in register. */
4285 if (gimple_has_lhs (stmt)
4286 && TREE_CODE (lhs) == SSA_NAME
4287 && !gimple_assign_ssa_name_copy_p (stmt)
4288 && (!gimple_assign_single_p (stmt)
4289 || (!is_gimple_min_invariant (rhs)
4290 && (gimple_assign_rhs_code (stmt) != VAR_DECL
4291 || !is_global_var (rhs)
4292 || !DECL_HARD_REGISTER (rhs))))
4293 && !gimple_has_volatile_ops (stmt)
4294 && !has_zero_uses (lhs))
4295 {
4296 tree sprime = NULL;
4297 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4298 pre_expr sprimeexpr;
4299 gimple orig_stmt = stmt;
4300
4301 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4302 get_expr_value_id (lhsexpr),
4303 NULL);
4304
4305 if (sprimeexpr)
4306 {
4307 if (sprimeexpr->kind == CONSTANT)
4308 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4309 else if (sprimeexpr->kind == NAME)
4310 sprime = PRE_EXPR_NAME (sprimeexpr);
4311 else
4312 gcc_unreachable ();
4313 }
4314
4315 /* If there is no existing leader but SCCVN knows this
4316 value is constant, use that constant. */
4317 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4318 {
4319 sprime = VN_INFO (lhs)->valnum;
4320 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4321 TREE_TYPE (sprime)))
4322 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4323
4324 if (dump_file && (dump_flags & TDF_DETAILS))
4325 {
4326 fprintf (dump_file, "Replaced ");
4327 print_gimple_expr (dump_file, stmt, 0, 0);
4328 fprintf (dump_file, " with ");
4329 print_generic_expr (dump_file, sprime, 0);
4330 fprintf (dump_file, " in ");
4331 print_gimple_stmt (dump_file, stmt, 0, 0);
4332 }
4333 pre_stats.eliminations++;
4334 propagate_tree_value_into_stmt (&gsi, sprime);
4335 stmt = gsi_stmt (gsi);
4336 update_stmt (stmt);
4337
4338 /* If we removed EH side-effects from the statement, clean
4339 its EH information. */
4340 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4341 {
4342 bitmap_set_bit (need_eh_cleanup,
4343 gimple_bb (stmt)->index);
4344 if (dump_file && (dump_flags & TDF_DETAILS))
4345 fprintf (dump_file, " Removed EH side-effects.\n");
4346 }
4347 continue;
4348 }
4349
4350 /* If there is no existing usable leader but SCCVN thinks
4351 it has an expression it wants to use as replacement,
4352 insert that. */
4353 if (!sprime || sprime == lhs)
4354 {
4355 tree val = VN_INFO (lhs)->valnum;
4356 if (val != VN_TOP
4357 && TREE_CODE (val) == SSA_NAME
4358 && VN_INFO (val)->needs_insertion
4359 && can_PRE_operation (vn_get_expr_for (val)))
4360 sprime = do_SCCVN_insertion (stmt, val);
4361 }
4362 if (sprime
4363 && sprime != lhs
4364 && (rhs == NULL_TREE
4365 || TREE_CODE (rhs) != SSA_NAME
4366 || may_propagate_copy (rhs, sprime)))
4367 {
4368 bool can_make_abnormal_goto
4369 = is_gimple_call (stmt)
4370 && stmt_can_make_abnormal_goto (stmt);
4371
4372 gcc_assert (sprime != rhs);
4373
4374 if (dump_file && (dump_flags & TDF_DETAILS))
4375 {
4376 fprintf (dump_file, "Replaced ");
4377 print_gimple_expr (dump_file, stmt, 0, 0);
4378 fprintf (dump_file, " with ");
4379 print_generic_expr (dump_file, sprime, 0);
4380 fprintf (dump_file, " in ");
4381 print_gimple_stmt (dump_file, stmt, 0, 0);
4382 }
4383
4384 if (TREE_CODE (sprime) == SSA_NAME)
4385 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4386 NECESSARY, true);
4387 /* We need to make sure the new and old types actually match,
4388 which may require adding a simple cast, which fold_convert
4389 will do for us. */
4390 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4391 && !useless_type_conversion_p (gimple_expr_type (stmt),
4392 TREE_TYPE (sprime)))
4393 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4394
4395 pre_stats.eliminations++;
4396 propagate_tree_value_into_stmt (&gsi, sprime);
4397 stmt = gsi_stmt (gsi);
4398 update_stmt (stmt);
4399
4400 /* If we removed EH side-effects from the statement, clean
4401 its EH information. */
4402 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4403 {
4404 bitmap_set_bit (need_eh_cleanup,
4405 gimple_bb (stmt)->index);
4406 if (dump_file && (dump_flags & TDF_DETAILS))
4407 fprintf (dump_file, " Removed EH side-effects.\n");
4408 }
4409
4410 /* Likewise for AB side-effects. */
4411 if (can_make_abnormal_goto
4412 && !stmt_can_make_abnormal_goto (stmt))
4413 {
4414 bitmap_set_bit (need_ab_cleanup,
4415 gimple_bb (stmt)->index);
4416 if (dump_file && (dump_flags & TDF_DETAILS))
4417 fprintf (dump_file, " Removed AB side-effects.\n");
4418 }
4419 }
4420 }
4421 /* If the statement is a scalar store, see if the expression
4422 has the same value number as its rhs. If so, the store is
4423 dead. */
4424 else if (gimple_assign_single_p (stmt)
4425 && !gimple_has_volatile_ops (stmt)
4426 && !is_gimple_reg (gimple_assign_lhs (stmt))
4427 && (TREE_CODE (rhs) == SSA_NAME
4428 || is_gimple_min_invariant (rhs)))
4429 {
4430 tree val;
4431 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4432 gimple_vuse (stmt), VN_WALK, NULL);
4433 if (TREE_CODE (rhs) == SSA_NAME)
4434 rhs = VN_INFO (rhs)->valnum;
4435 if (val
4436 && operand_equal_p (val, rhs, 0))
4437 {
4438 if (dump_file && (dump_flags & TDF_DETAILS))
4439 {
4440 fprintf (dump_file, "Deleted redundant store ");
4441 print_gimple_stmt (dump_file, stmt, 0, 0);
4442 }
4443
4444 /* Queue stmt for removal. */
4445 VEC_safe_push (gimple, heap, to_remove, stmt);
4446 }
4447 }
4448 /* Visit COND_EXPRs and fold the comparison with the
4449 available value-numbers. */
4450 else if (gimple_code (stmt) == GIMPLE_COND)
4451 {
4452 tree op0 = gimple_cond_lhs (stmt);
4453 tree op1 = gimple_cond_rhs (stmt);
4454 tree result;
4455
4456 if (TREE_CODE (op0) == SSA_NAME)
4457 op0 = VN_INFO (op0)->valnum;
4458 if (TREE_CODE (op1) == SSA_NAME)
4459 op1 = VN_INFO (op1)->valnum;
4460 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4461 op0, op1);
4462 if (result && TREE_CODE (result) == INTEGER_CST)
4463 {
4464 if (integer_zerop (result))
4465 gimple_cond_make_false (stmt);
4466 else
4467 gimple_cond_make_true (stmt);
4468 update_stmt (stmt);
4469 todo = TODO_cleanup_cfg;
4470 }
4471 }
4472 /* Visit indirect calls and turn them into direct calls if
4473 possible. */
4474 if (is_gimple_call (stmt))
4475 {
4476 tree orig_fn = gimple_call_fn (stmt);
4477 tree fn;
4478 if (!orig_fn)
4479 continue;
4480 if (TREE_CODE (orig_fn) == SSA_NAME)
4481 fn = VN_INFO (orig_fn)->valnum;
4482 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4483 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4484 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4485 else
4486 continue;
4487 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4488 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4489 TREE_TYPE (fn)))
4490 {
4491 bool can_make_abnormal_goto
4492 = stmt_can_make_abnormal_goto (stmt);
4493 bool was_noreturn = gimple_call_noreturn_p (stmt);
4494
4495 if (dump_file && (dump_flags & TDF_DETAILS))
4496 {
4497 fprintf (dump_file, "Replacing call target with ");
4498 print_generic_expr (dump_file, fn, 0);
4499 fprintf (dump_file, " in ");
4500 print_gimple_stmt (dump_file, stmt, 0, 0);
4501 }
4502
4503 gimple_call_set_fn (stmt, fn);
4504 VEC_safe_push (gimple, heap, to_update, stmt);
4505
4506 /* When changing a call into a noreturn call, cfg cleanup
4507 is needed to fix up the noreturn call. */
4508 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4509 todo |= TODO_cleanup_cfg;
4510
4511 /* If we removed EH side-effects from the statement, clean
4512 its EH information. */
4513 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4514 {
4515 bitmap_set_bit (need_eh_cleanup,
4516 gimple_bb (stmt)->index);
4517 if (dump_file && (dump_flags & TDF_DETAILS))
4518 fprintf (dump_file, " Removed EH side-effects.\n");
4519 }
4520
4521 /* Likewise for AB side-effects. */
4522 if (can_make_abnormal_goto
4523 && !stmt_can_make_abnormal_goto (stmt))
4524 {
4525 bitmap_set_bit (need_ab_cleanup,
4526 gimple_bb (stmt)->index);
4527 if (dump_file && (dump_flags & TDF_DETAILS))
4528 fprintf (dump_file, " Removed AB side-effects.\n");
4529 }
4530
4531 /* Changing an indirect call to a direct call may
4532 have exposed different semantics. This may
4533 require an SSA update. */
4534 todo |= TODO_update_ssa_only_virtuals;
4535 }
4536 }
4537 }
4538
4539 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4540 {
4541 gimple stmt, phi = gsi_stmt (gsi);
4542 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4543 pre_expr sprimeexpr, resexpr;
4544 gimple_stmt_iterator gsi2;
4545
4546 /* We want to perform redundant PHI elimination. Do so by
4547 replacing the PHI with a single copy if possible.
4548 Do not touch inserted, single-argument or virtual PHIs. */
4549 if (gimple_phi_num_args (phi) == 1
4550 || !is_gimple_reg (res))
4551 {
4552 gsi_next (&gsi);
4553 continue;
4554 }
4555
4556 resexpr = get_or_alloc_expr_for_name (res);
4557 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4558 get_expr_value_id (resexpr), NULL);
4559 if (sprimeexpr)
4560 {
4561 if (sprimeexpr->kind == CONSTANT)
4562 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4563 else if (sprimeexpr->kind == NAME)
4564 sprime = PRE_EXPR_NAME (sprimeexpr);
4565 else
4566 gcc_unreachable ();
4567 }
4568 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4569 {
4570 sprime = VN_INFO (res)->valnum;
4571 if (!useless_type_conversion_p (TREE_TYPE (res),
4572 TREE_TYPE (sprime)))
4573 sprime = fold_convert (TREE_TYPE (res), sprime);
4574 }
4575 if (!sprime
4576 || sprime == res)
4577 {
4578 gsi_next (&gsi);
4579 continue;
4580 }
4581
4582 if (dump_file && (dump_flags & TDF_DETAILS))
4583 {
4584 fprintf (dump_file, "Replaced redundant PHI node defining ");
4585 print_generic_expr (dump_file, res, 0);
4586 fprintf (dump_file, " with ");
4587 print_generic_expr (dump_file, sprime, 0);
4588 fprintf (dump_file, "\n");
4589 }
4590
4591 remove_phi_node (&gsi, false);
4592
4593 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4594 && TREE_CODE (sprime) == SSA_NAME)
4595 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4596
4597 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4598 sprime = fold_convert (TREE_TYPE (res), sprime);
4599 stmt = gimple_build_assign (res, sprime);
4600 SSA_NAME_DEF_STMT (res) = stmt;
4601 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4602
4603 gsi2 = gsi_after_labels (b);
4604 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4605 /* Queue the copy for eventual removal. */
4606 VEC_safe_push (gimple, heap, to_remove, stmt);
4607 /* If we inserted this PHI node ourself, it's not an elimination. */
4608 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4609 pre_stats.phis--;
4610 else
4611 pre_stats.eliminations++;
4612 }
4613 }
4614
4615 /* We cannot remove stmts during BB walk, especially not release SSA
4616 names there as this confuses the VN machinery. The stmts ending
4617 up in to_remove are either stores or simple copies. */
4618 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4619 {
4620 tree lhs = gimple_assign_lhs (stmt);
4621 tree rhs = gimple_assign_rhs1 (stmt);
4622 use_operand_p use_p;
4623 gimple use_stmt;
4624
4625 /* If there is a single use only, propagate the equivalency
4626 instead of keeping the copy. */
4627 if (TREE_CODE (lhs) == SSA_NAME
4628 && TREE_CODE (rhs) == SSA_NAME
4629 && single_imm_use (lhs, &use_p, &use_stmt)
4630 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4631 {
4632 SET_USE (use_p, rhs);
4633 update_stmt (use_stmt);
4634 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4635 && TREE_CODE (rhs) == SSA_NAME)
4636 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4637 }
4638
4639 /* If this is a store or a now unused copy, remove it. */
4640 if (TREE_CODE (lhs) != SSA_NAME
4641 || has_zero_uses (lhs))
4642 {
4643 basic_block bb = gimple_bb (stmt);
4644 gsi = gsi_for_stmt (stmt);
4645 unlink_stmt_vdef (stmt);
4646 if (gsi_remove (&gsi, true))
4647 bitmap_set_bit (need_eh_cleanup, bb->index);
4648 if (TREE_CODE (lhs) == SSA_NAME)
4649 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4650 release_defs (stmt);
4651 }
4652 }
4653 VEC_free (gimple, heap, to_remove);
4654
4655 /* We cannot update call statements with virtual operands during
4656 SSA walk. This might remove them which in turn makes our
4657 VN lattice invalid. */
4658 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4659 update_stmt (stmt);
4660 VEC_free (gimple, heap, to_update);
4661
4662 return todo;
4663 }
4664
4665 /* Borrow a bit of tree-ssa-dce.c for the moment.
4666 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4667 this may be a bit faster, and we may want critical edges kept split. */
4668
4669 /* If OP's defining statement has not already been determined to be necessary,
4670 mark that statement necessary. Return the stmt, if it is newly
4671 necessary. */
4672
4673 static inline gimple
4674 mark_operand_necessary (tree op)
4675 {
4676 gimple stmt;
4677
4678 gcc_assert (op);
4679
4680 if (TREE_CODE (op) != SSA_NAME)
4681 return NULL;
4682
4683 stmt = SSA_NAME_DEF_STMT (op);
4684 gcc_assert (stmt);
4685
4686 if (gimple_plf (stmt, NECESSARY)
4687 || gimple_nop_p (stmt))
4688 return NULL;
4689
4690 gimple_set_plf (stmt, NECESSARY, true);
4691 return stmt;
4692 }
4693
4694 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4695 to insert PHI nodes sometimes, and because value numbering of casts isn't
4696 perfect, we sometimes end up inserting dead code. This simple DCE-like
4697 pass removes any insertions we made that weren't actually used. */
4698
4699 static void
4700 remove_dead_inserted_code (void)
4701 {
4702 bitmap worklist;
4703 unsigned i;
4704 bitmap_iterator bi;
4705 gimple t;
4706
4707 worklist = BITMAP_ALLOC (NULL);
4708 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4709 {
4710 t = SSA_NAME_DEF_STMT (ssa_name (i));
4711 if (gimple_plf (t, NECESSARY))
4712 bitmap_set_bit (worklist, i);
4713 }
4714 while (!bitmap_empty_p (worklist))
4715 {
4716 i = bitmap_first_set_bit (worklist);
4717 bitmap_clear_bit (worklist, i);
4718 t = SSA_NAME_DEF_STMT (ssa_name (i));
4719
4720 /* PHI nodes are somewhat special in that each PHI alternative has
4721 data and control dependencies. All the statements feeding the
4722 PHI node's arguments are always necessary. */
4723 if (gimple_code (t) == GIMPLE_PHI)
4724 {
4725 unsigned k;
4726
4727 for (k = 0; k < gimple_phi_num_args (t); k++)
4728 {
4729 tree arg = PHI_ARG_DEF (t, k);
4730 if (TREE_CODE (arg) == SSA_NAME)
4731 {
4732 gimple n = mark_operand_necessary (arg);
4733 if (n)
4734 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4735 }
4736 }
4737 }
4738 else
4739 {
4740 /* Propagate through the operands. Examine all the USE, VUSE and
4741 VDEF operands in this statement. Mark all the statements
4742 which feed this statement's uses as necessary. */
4743 ssa_op_iter iter;
4744 tree use;
4745
4746 /* The operands of VDEF expressions are also needed as they
4747 represent potential definitions that may reach this
4748 statement (VDEF operands allow us to follow def-def
4749 links). */
4750
4751 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4752 {
4753 gimple n = mark_operand_necessary (use);
4754 if (n)
4755 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4756 }
4757 }
4758 }
4759
4760 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4761 {
4762 t = SSA_NAME_DEF_STMT (ssa_name (i));
4763 if (!gimple_plf (t, NECESSARY))
4764 {
4765 gimple_stmt_iterator gsi;
4766
4767 if (dump_file && (dump_flags & TDF_DETAILS))
4768 {
4769 fprintf (dump_file, "Removing unnecessary insertion:");
4770 print_gimple_stmt (dump_file, t, 0, 0);
4771 }
4772
4773 gsi = gsi_for_stmt (t);
4774 if (gimple_code (t) == GIMPLE_PHI)
4775 remove_phi_node (&gsi, true);
4776 else
4777 {
4778 gsi_remove (&gsi, true);
4779 release_defs (t);
4780 }
4781 }
4782 }
4783 BITMAP_FREE (worklist);
4784 }
4785
4786 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4787 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4788 the number of visited blocks. */
4789
4790 static int
4791 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4792 {
4793 edge_iterator *stack;
4794 int sp;
4795 int post_order_num = 0;
4796 sbitmap visited;
4797
4798 if (include_entry_exit)
4799 post_order[post_order_num++] = EXIT_BLOCK;
4800
4801 /* Allocate stack for back-tracking up CFG. */
4802 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4803 sp = 0;
4804
4805 /* Allocate bitmap to track nodes that have been visited. */
4806 visited = sbitmap_alloc (last_basic_block);
4807
4808 /* None of the nodes in the CFG have been visited yet. */
4809 sbitmap_zero (visited);
4810
4811 /* Push the last edge on to the stack. */
4812 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4813
4814 while (sp)
4815 {
4816 edge_iterator ei;
4817 basic_block src;
4818 basic_block dest;
4819
4820 /* Look at the edge on the top of the stack. */
4821 ei = stack[sp - 1];
4822 src = ei_edge (ei)->src;
4823 dest = ei_edge (ei)->dest;
4824
4825 /* Check if the edge destination has been visited yet. */
4826 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4827 {
4828 /* Mark that we have visited the destination. */
4829 SET_BIT (visited, src->index);
4830
4831 if (EDGE_COUNT (src->preds) > 0)
4832 /* Since the DEST node has been visited for the first
4833 time, check its successors. */
4834 stack[sp++] = ei_start (src->preds);
4835 else
4836 post_order[post_order_num++] = src->index;
4837 }
4838 else
4839 {
4840 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4841 post_order[post_order_num++] = dest->index;
4842
4843 if (!ei_one_before_end_p (ei))
4844 ei_next (&stack[sp - 1]);
4845 else
4846 sp--;
4847 }
4848 }
4849
4850 if (include_entry_exit)
4851 post_order[post_order_num++] = ENTRY_BLOCK;
4852
4853 free (stack);
4854 sbitmap_free (visited);
4855 return post_order_num;
4856 }
4857
4858
4859 /* Initialize data structures used by PRE. */
4860
4861 static void
4862 init_pre (bool do_fre)
4863 {
4864 basic_block bb;
4865
4866 next_expression_id = 1;
4867 expressions = NULL;
4868 VEC_safe_push (pre_expr, heap, expressions, NULL);
4869 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4870 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4871 get_max_value_id() + 1);
4872 name_to_id = NULL;
4873
4874 in_fre = do_fre;
4875
4876 inserted_exprs = BITMAP_ALLOC (NULL);
4877 need_creation = NULL;
4878 pretemp = NULL_TREE;
4879 storetemp = NULL_TREE;
4880 prephitemp = NULL_TREE;
4881
4882 connect_infinite_loops_to_exit ();
4883 memset (&pre_stats, 0, sizeof (pre_stats));
4884
4885
4886 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4887 my_rev_post_order_compute (postorder, false);
4888
4889 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4890
4891 calculate_dominance_info (CDI_POST_DOMINATORS);
4892 calculate_dominance_info (CDI_DOMINATORS);
4893
4894 bitmap_obstack_initialize (&grand_bitmap_obstack);
4895 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4896 expr_pred_trans_eq, free);
4897 expression_to_id = htab_create (num_ssa_names * 3,
4898 pre_expr_hash,
4899 pre_expr_eq, NULL);
4900 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4901 sizeof (struct bitmap_set), 30);
4902 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4903 sizeof (struct pre_expr_d), 30);
4904 FOR_ALL_BB (bb)
4905 {
4906 EXP_GEN (bb) = bitmap_set_new ();
4907 PHI_GEN (bb) = bitmap_set_new ();
4908 TMP_GEN (bb) = bitmap_set_new ();
4909 AVAIL_OUT (bb) = bitmap_set_new ();
4910 }
4911
4912 need_eh_cleanup = BITMAP_ALLOC (NULL);
4913 need_ab_cleanup = BITMAP_ALLOC (NULL);
4914 }
4915
4916
4917 /* Deallocate data structures used by PRE. */
4918
4919 static void
4920 fini_pre (bool do_fre)
4921 {
4922 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4923 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4924
4925 free (postorder);
4926 VEC_free (bitmap_set_t, heap, value_expressions);
4927 BITMAP_FREE (inserted_exprs);
4928 VEC_free (gimple, heap, need_creation);
4929 bitmap_obstack_release (&grand_bitmap_obstack);
4930 free_alloc_pool (bitmap_set_pool);
4931 free_alloc_pool (pre_expr_pool);
4932 htab_delete (phi_translate_table);
4933 htab_delete (expression_to_id);
4934 VEC_free (unsigned, heap, name_to_id);
4935
4936 free_aux_for_blocks ();
4937
4938 free_dominance_info (CDI_POST_DOMINATORS);
4939
4940 if (do_eh_cleanup)
4941 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4942
4943 if (do_ab_cleanup)
4944 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4945
4946 BITMAP_FREE (need_eh_cleanup);
4947 BITMAP_FREE (need_ab_cleanup);
4948
4949 if (do_eh_cleanup || do_ab_cleanup)
4950 cleanup_tree_cfg ();
4951
4952 if (!do_fre)
4953 loop_optimizer_finalize ();
4954 }
4955
4956 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4957 only wants to do full redundancy elimination. */
4958
4959 static unsigned int
4960 execute_pre (bool do_fre)
4961 {
4962 unsigned int todo = 0;
4963
4964 do_partial_partial =
4965 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4966
4967 /* This has to happen before SCCVN runs because
4968 loop_optimizer_init may create new phis, etc. */
4969 if (!do_fre)
4970 loop_optimizer_init (LOOPS_NORMAL);
4971
4972 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4973 {
4974 if (!do_fre)
4975 loop_optimizer_finalize ();
4976
4977 return 0;
4978 }
4979
4980 init_pre (do_fre);
4981 scev_initialize ();
4982
4983 /* Collect and value number expressions computed in each basic block. */
4984 compute_avail ();
4985
4986 if (dump_file && (dump_flags & TDF_DETAILS))
4987 {
4988 basic_block bb;
4989
4990 FOR_ALL_BB (bb)
4991 {
4992 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4993 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4994 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4995 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4996 }
4997 }
4998
4999 /* Insert can get quite slow on an incredibly large number of basic
5000 blocks due to some quadratic behavior. Until this behavior is
5001 fixed, don't run it when he have an incredibly large number of
5002 bb's. If we aren't going to run insert, there is no point in
5003 computing ANTIC, either, even though it's plenty fast. */
5004 if (!do_fre && n_basic_blocks < 4000)
5005 {
5006 compute_antic ();
5007 insert ();
5008 }
5009
5010 /* Make sure to remove fake edges before committing our inserts.
5011 This makes sure we don't end up with extra critical edges that
5012 we would need to split. */
5013 remove_fake_exit_edges ();
5014 gsi_commit_edge_inserts ();
5015
5016 /* Remove all the redundant expressions. */
5017 todo |= eliminate ();
5018
5019 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
5020 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
5021 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
5022 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
5023 statistics_counter_event (cfun, "Constified", pre_stats.constified);
5024
5025 clear_expression_ids ();
5026 if (!do_fre)
5027 {
5028 remove_dead_inserted_code ();
5029 todo |= TODO_verify_flow;
5030 }
5031
5032 scev_finalize ();
5033 fini_pre (do_fre);
5034
5035 if (!do_fre)
5036 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5037 case we can merge the block with the remaining predecessor of the block.
5038 It should either:
5039 - call merge_blocks after each tail merge iteration
5040 - call merge_blocks after all tail merge iterations
5041 - mark TODO_cleanup_cfg when necessary
5042 - share the cfg cleanup with fini_pre. */
5043 todo |= tail_merge_optimize (todo);
5044 free_scc_vn ();
5045
5046 return todo;
5047 }
5048
5049 /* Gate and execute functions for PRE. */
5050
5051 static unsigned int
5052 do_pre (void)
5053 {
5054 return execute_pre (false);
5055 }
5056
5057 static bool
5058 gate_pre (void)
5059 {
5060 return flag_tree_pre != 0;
5061 }
5062
5063 struct gimple_opt_pass pass_pre =
5064 {
5065 {
5066 GIMPLE_PASS,
5067 "pre", /* name */
5068 gate_pre, /* gate */
5069 do_pre, /* execute */
5070 NULL, /* sub */
5071 NULL, /* next */
5072 0, /* static_pass_number */
5073 TV_TREE_PRE, /* tv_id */
5074 PROP_no_crit_edges | PROP_cfg
5075 | PROP_ssa, /* properties_required */
5076 0, /* properties_provided */
5077 0, /* properties_destroyed */
5078 TODO_rebuild_alias, /* todo_flags_start */
5079 TODO_update_ssa_only_virtuals | TODO_ggc_collect
5080 | TODO_verify_ssa /* todo_flags_finish */
5081 }
5082 };
5083
5084
5085 /* Gate and execute functions for FRE. */
5086
5087 static unsigned int
5088 execute_fre (void)
5089 {
5090 return execute_pre (true);
5091 }
5092
5093 static bool
5094 gate_fre (void)
5095 {
5096 return flag_tree_fre != 0;
5097 }
5098
5099 struct gimple_opt_pass pass_fre =
5100 {
5101 {
5102 GIMPLE_PASS,
5103 "fre", /* name */
5104 gate_fre, /* gate */
5105 execute_fre, /* execute */
5106 NULL, /* sub */
5107 NULL, /* next */
5108 0, /* static_pass_number */
5109 TV_TREE_FRE, /* tv_id */
5110 PROP_cfg | PROP_ssa, /* properties_required */
5111 0, /* properties_provided */
5112 0, /* properties_destroyed */
5113 0, /* todo_flags_start */
5114 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
5115 }
5116 };