stor-layout.c (bit_from_pos): Distribute conversion to bitsizetype into a PLUS_EXPR...
[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
1663 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1664 ref->type,
1665 newoperands,
1666 &newref, VN_WALK);
1667 if (result)
1668 VEC_free (vn_reference_op_s, heap, newoperands);
1669
1670 /* We can always insert constants, so if we have a partial
1671 redundant constant load of another type try to translate it
1672 to a constant of appropriate type. */
1673 if (result && is_gimple_min_invariant (result))
1674 {
1675 tree tem = result;
1676 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1677 {
1678 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1679 if (tem && !is_gimple_min_invariant (tem))
1680 tem = NULL_TREE;
1681 }
1682 if (tem)
1683 return get_or_alloc_expr_for_constant (tem);
1684 }
1685
1686 /* If we'd have to convert things we would need to validate
1687 if we can insert the translated expression. So fail
1688 here for now - we cannot insert an alias with a different
1689 type in the VN tables either, as that would assert. */
1690 if (result
1691 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1692 return NULL;
1693 else if (!result && newref
1694 && !useless_type_conversion_p (ref->type, newref->type))
1695 {
1696 VEC_free (vn_reference_op_s, heap, newoperands);
1697 return NULL;
1698 }
1699
1700 expr = (pre_expr) pool_alloc (pre_expr_pool);
1701 expr->kind = REFERENCE;
1702 expr->id = 0;
1703
1704 if (newref)
1705 {
1706 PRE_EXPR_REFERENCE (expr) = newref;
1707 constant = fully_constant_expression (expr);
1708 if (constant != expr)
1709 return constant;
1710
1711 new_val_id = newref->value_id;
1712 get_or_alloc_expression_id (expr);
1713 }
1714 else
1715 {
1716 if (changed || !same_valid)
1717 {
1718 new_val_id = get_next_value_id ();
1719 VEC_safe_grow_cleared (bitmap_set_t, heap,
1720 value_expressions,
1721 get_max_value_id() + 1);
1722 }
1723 else
1724 new_val_id = ref->value_id;
1725 newref = vn_reference_insert_pieces (newvuse, ref->set,
1726 ref->type,
1727 newoperands,
1728 result, new_val_id);
1729 newoperands = NULL;
1730 PRE_EXPR_REFERENCE (expr) = newref;
1731 constant = fully_constant_expression (expr);
1732 if (constant != expr)
1733 return constant;
1734 get_or_alloc_expression_id (expr);
1735 }
1736 add_to_value (new_val_id, expr);
1737 }
1738 VEC_free (vn_reference_op_s, heap, newoperands);
1739 return expr;
1740 }
1741 break;
1742
1743 case NAME:
1744 {
1745 gimple phi = NULL;
1746 edge e;
1747 gimple def_stmt;
1748 tree name = PRE_EXPR_NAME (expr);
1749
1750 def_stmt = SSA_NAME_DEF_STMT (name);
1751 if (gimple_code (def_stmt) == GIMPLE_PHI
1752 && gimple_bb (def_stmt) == phiblock)
1753 phi = def_stmt;
1754 else
1755 return expr;
1756
1757 e = find_edge (pred, gimple_bb (phi));
1758 if (e)
1759 {
1760 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1761 pre_expr newexpr;
1762
1763 if (TREE_CODE (def) == SSA_NAME)
1764 def = VN_INFO (def)->valnum;
1765
1766 /* Handle constant. */
1767 if (is_gimple_min_invariant (def))
1768 return get_or_alloc_expr_for_constant (def);
1769
1770 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1771 return NULL;
1772
1773 newexpr = get_or_alloc_expr_for_name (def);
1774 return newexpr;
1775 }
1776 }
1777 return expr;
1778
1779 default:
1780 gcc_unreachable ();
1781 }
1782 }
1783
1784 /* Wrapper around phi_translate_1 providing caching functionality. */
1785
1786 static pre_expr
1787 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1788 basic_block pred, basic_block phiblock)
1789 {
1790 pre_expr phitrans;
1791
1792 if (!expr)
1793 return NULL;
1794
1795 /* Constants contain no values that need translation. */
1796 if (expr->kind == CONSTANT)
1797 return expr;
1798
1799 if (value_id_constant_p (get_expr_value_id (expr)))
1800 return expr;
1801
1802 if (expr->kind != NAME)
1803 {
1804 phitrans = phi_trans_lookup (expr, pred);
1805 if (phitrans)
1806 return phitrans;
1807 }
1808
1809 /* Translate. */
1810 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1811
1812 /* Don't add empty translations to the cache. Neither add
1813 translations of NAMEs as those are cheap to translate. */
1814 if (phitrans
1815 && expr->kind != NAME)
1816 phi_trans_add (expr, phitrans, pred);
1817
1818 return phitrans;
1819 }
1820
1821
1822 /* For each expression in SET, translate the values through phi nodes
1823 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1824 expressions in DEST. */
1825
1826 static void
1827 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1828 basic_block phiblock)
1829 {
1830 VEC (pre_expr, heap) *exprs;
1831 pre_expr expr;
1832 int i;
1833
1834 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1835 {
1836 bitmap_set_copy (dest, set);
1837 return;
1838 }
1839
1840 exprs = sorted_array_from_bitmap_set (set);
1841 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1842 {
1843 pre_expr translated;
1844 translated = phi_translate (expr, set, NULL, pred, phiblock);
1845 if (!translated)
1846 continue;
1847
1848 /* We might end up with multiple expressions from SET being
1849 translated to the same value. In this case we do not want
1850 to retain the NARY or REFERENCE expression but prefer a NAME
1851 which would be the leader. */
1852 if (translated->kind == NAME)
1853 bitmap_value_replace_in_set (dest, translated);
1854 else
1855 bitmap_value_insert_into_set (dest, translated);
1856 }
1857 VEC_free (pre_expr, heap, exprs);
1858 }
1859
1860 /* Find the leader for a value (i.e., the name representing that
1861 value) in a given set, and return it. If STMT is non-NULL it
1862 makes sure the defining statement for the leader dominates it.
1863 Return NULL if no leader is found. */
1864
1865 static pre_expr
1866 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1867 {
1868 if (value_id_constant_p (val))
1869 {
1870 unsigned int i;
1871 bitmap_iterator bi;
1872 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1873
1874 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1875 {
1876 pre_expr expr = expression_for_id (i);
1877 if (expr->kind == CONSTANT)
1878 return expr;
1879 }
1880 }
1881 if (bitmap_set_contains_value (set, val))
1882 {
1883 /* Rather than walk the entire bitmap of expressions, and see
1884 whether any of them has the value we are looking for, we look
1885 at the reverse mapping, which tells us the set of expressions
1886 that have a given value (IE value->expressions with that
1887 value) and see if any of those expressions are in our set.
1888 The number of expressions per value is usually significantly
1889 less than the number of expressions in the set. In fact, for
1890 large testcases, doing it this way is roughly 5-10x faster
1891 than walking the bitmap.
1892 If this is somehow a significant lose for some cases, we can
1893 choose which set to walk based on which set is smaller. */
1894 unsigned int i;
1895 bitmap_iterator bi;
1896 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1897
1898 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1899 &set->expressions, 0, i, bi)
1900 {
1901 pre_expr val = expression_for_id (i);
1902 /* At the point where stmt is not null, there should always
1903 be an SSA_NAME first in the list of expressions. */
1904 if (stmt)
1905 {
1906 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1907 if (gimple_code (def_stmt) != GIMPLE_PHI
1908 && gimple_bb (def_stmt) == gimple_bb (stmt)
1909 /* PRE insertions are at the end of the basic-block
1910 and have UID 0. */
1911 && (gimple_uid (def_stmt) == 0
1912 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1913 continue;
1914 }
1915 return val;
1916 }
1917 }
1918 return NULL;
1919 }
1920
1921 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1922 BLOCK by seeing if it is not killed in the block. Note that we are
1923 only determining whether there is a store that kills it. Because
1924 of the order in which clean iterates over values, we are guaranteed
1925 that altered operands will have caused us to be eliminated from the
1926 ANTIC_IN set already. */
1927
1928 static bool
1929 value_dies_in_block_x (pre_expr expr, basic_block block)
1930 {
1931 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1932 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1933 gimple def;
1934 gimple_stmt_iterator gsi;
1935 unsigned id = get_expression_id (expr);
1936 bool res = false;
1937 ao_ref ref;
1938
1939 if (!vuse)
1940 return false;
1941
1942 /* Lookup a previously calculated result. */
1943 if (EXPR_DIES (block)
1944 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1945 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1946
1947 /* A memory expression {e, VUSE} dies in the block if there is a
1948 statement that may clobber e. If, starting statement walk from the
1949 top of the basic block, a statement uses VUSE there can be no kill
1950 inbetween that use and the original statement that loaded {e, VUSE},
1951 so we can stop walking. */
1952 ref.base = NULL_TREE;
1953 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1954 {
1955 tree def_vuse, def_vdef;
1956 def = gsi_stmt (gsi);
1957 def_vuse = gimple_vuse (def);
1958 def_vdef = gimple_vdef (def);
1959
1960 /* Not a memory statement. */
1961 if (!def_vuse)
1962 continue;
1963
1964 /* Not a may-def. */
1965 if (!def_vdef)
1966 {
1967 /* A load with the same VUSE, we're done. */
1968 if (def_vuse == vuse)
1969 break;
1970
1971 continue;
1972 }
1973
1974 /* Init ref only if we really need it. */
1975 if (ref.base == NULL_TREE
1976 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1977 refx->operands))
1978 {
1979 res = true;
1980 break;
1981 }
1982 /* If the statement may clobber expr, it dies. */
1983 if (stmt_may_clobber_ref_p_1 (def, &ref))
1984 {
1985 res = true;
1986 break;
1987 }
1988 }
1989
1990 /* Remember the result. */
1991 if (!EXPR_DIES (block))
1992 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1993 bitmap_set_bit (EXPR_DIES (block), id * 2);
1994 if (res)
1995 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1996
1997 return res;
1998 }
1999
2000
2001 /* Determine if OP is valid in SET1 U SET2, which it is when the union
2002 contains its value-id. */
2003
2004 static bool
2005 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
2006 {
2007 if (op && TREE_CODE (op) == SSA_NAME)
2008 {
2009 unsigned int value_id = VN_INFO (op)->value_id;
2010 if (!bitmap_set_contains_value (set1, value_id)
2011 || (set2 && !bitmap_set_contains_value (set2, value_id)))
2012 return false;
2013 }
2014 return true;
2015 }
2016
2017 /* Determine if the expression EXPR is valid in SET1 U SET2.
2018 ONLY SET2 CAN BE NULL.
2019 This means that we have a leader for each part of the expression
2020 (if it consists of values), or the expression is an SSA_NAME.
2021 For loads/calls, we also see if the vuse is killed in this block. */
2022
2023 static bool
2024 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2025 basic_block block)
2026 {
2027 switch (expr->kind)
2028 {
2029 case NAME:
2030 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2031 case NARY:
2032 {
2033 unsigned int i;
2034 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2035 for (i = 0; i < nary->length; i++)
2036 if (!op_valid_in_sets (set1, set2, nary->op[i]))
2037 return false;
2038 return true;
2039 }
2040 break;
2041 case REFERENCE:
2042 {
2043 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2044 vn_reference_op_t vro;
2045 unsigned int i;
2046
2047 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2048 {
2049 if (!op_valid_in_sets (set1, set2, vro->op0)
2050 || !op_valid_in_sets (set1, set2, vro->op1)
2051 || !op_valid_in_sets (set1, set2, vro->op2))
2052 return false;
2053 }
2054 return true;
2055 }
2056 default:
2057 gcc_unreachable ();
2058 }
2059 }
2060
2061 /* Clean the set of expressions that are no longer valid in SET1 or
2062 SET2. This means expressions that are made up of values we have no
2063 leaders for in SET1 or SET2. This version is used for partial
2064 anticipation, which means it is not valid in either ANTIC_IN or
2065 PA_IN. */
2066
2067 static void
2068 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2069 {
2070 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2071 pre_expr expr;
2072 int i;
2073
2074 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2075 {
2076 if (!valid_in_sets (set1, set2, expr, block))
2077 bitmap_remove_from_set (set1, expr);
2078 }
2079 VEC_free (pre_expr, heap, exprs);
2080 }
2081
2082 /* Clean the set of expressions that are no longer valid in SET. This
2083 means expressions that are made up of values we have no leaders for
2084 in SET. */
2085
2086 static void
2087 clean (bitmap_set_t set, basic_block block)
2088 {
2089 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2090 pre_expr expr;
2091 int i;
2092
2093 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2094 {
2095 if (!valid_in_sets (set, NULL, expr, block))
2096 bitmap_remove_from_set (set, expr);
2097 }
2098 VEC_free (pre_expr, heap, exprs);
2099 }
2100
2101 /* Clean the set of expressions that are no longer valid in SET because
2102 they are clobbered in BLOCK or because they trap and may not be executed. */
2103
2104 static void
2105 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2106 {
2107 bitmap_iterator bi;
2108 unsigned i;
2109
2110 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2111 {
2112 pre_expr expr = expression_for_id (i);
2113 if (expr->kind == REFERENCE)
2114 {
2115 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2116 if (ref->vuse)
2117 {
2118 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2119 if (!gimple_nop_p (def_stmt)
2120 && ((gimple_bb (def_stmt) != block
2121 && !dominated_by_p (CDI_DOMINATORS,
2122 block, gimple_bb (def_stmt)))
2123 || (gimple_bb (def_stmt) == block
2124 && value_dies_in_block_x (expr, block))))
2125 bitmap_remove_from_set (set, expr);
2126 }
2127 }
2128 else if (expr->kind == NARY)
2129 {
2130 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2131 /* If the NARY may trap make sure the block does not contain
2132 a possible exit point.
2133 ??? This is overly conservative if we translate AVAIL_OUT
2134 as the available expression might be after the exit point. */
2135 if (BB_MAY_NOTRETURN (block)
2136 && vn_nary_may_trap (nary))
2137 bitmap_remove_from_set (set, expr);
2138 }
2139 }
2140 }
2141
2142 static sbitmap has_abnormal_preds;
2143
2144 /* List of blocks that may have changed during ANTIC computation and
2145 thus need to be iterated over. */
2146
2147 static sbitmap changed_blocks;
2148
2149 /* Decide whether to defer a block for a later iteration, or PHI
2150 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2151 should defer the block, and true if we processed it. */
2152
2153 static bool
2154 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2155 basic_block block, basic_block phiblock)
2156 {
2157 if (!BB_VISITED (phiblock))
2158 {
2159 SET_BIT (changed_blocks, block->index);
2160 BB_VISITED (block) = 0;
2161 BB_DEFERRED (block) = 1;
2162 return false;
2163 }
2164 else
2165 phi_translate_set (dest, source, block, phiblock);
2166 return true;
2167 }
2168
2169 /* Compute the ANTIC set for BLOCK.
2170
2171 If succs(BLOCK) > 1 then
2172 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2173 else if succs(BLOCK) == 1 then
2174 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2175
2176 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2177 */
2178
2179 static bool
2180 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2181 {
2182 bool changed = false;
2183 bitmap_set_t S, old, ANTIC_OUT;
2184 bitmap_iterator bi;
2185 unsigned int bii;
2186 edge e;
2187 edge_iterator ei;
2188
2189 old = ANTIC_OUT = S = NULL;
2190 BB_VISITED (block) = 1;
2191
2192 /* If any edges from predecessors are abnormal, antic_in is empty,
2193 so do nothing. */
2194 if (block_has_abnormal_pred_edge)
2195 goto maybe_dump_sets;
2196
2197 old = ANTIC_IN (block);
2198 ANTIC_OUT = bitmap_set_new ();
2199
2200 /* If the block has no successors, ANTIC_OUT is empty. */
2201 if (EDGE_COUNT (block->succs) == 0)
2202 ;
2203 /* If we have one successor, we could have some phi nodes to
2204 translate through. */
2205 else if (single_succ_p (block))
2206 {
2207 basic_block succ_bb = single_succ (block);
2208
2209 /* We trade iterations of the dataflow equations for having to
2210 phi translate the maximal set, which is incredibly slow
2211 (since the maximal set often has 300+ members, even when you
2212 have a small number of blocks).
2213 Basically, we defer the computation of ANTIC for this block
2214 until we have processed it's successor, which will inevitably
2215 have a *much* smaller set of values to phi translate once
2216 clean has been run on it.
2217 The cost of doing this is that we technically perform more
2218 iterations, however, they are lower cost iterations.
2219
2220 Timings for PRE on tramp3d-v4:
2221 without maximal set fix: 11 seconds
2222 with maximal set fix/without deferring: 26 seconds
2223 with maximal set fix/with deferring: 11 seconds
2224 */
2225
2226 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2227 block, succ_bb))
2228 {
2229 changed = true;
2230 goto maybe_dump_sets;
2231 }
2232 }
2233 /* If we have multiple successors, we take the intersection of all of
2234 them. Note that in the case of loop exit phi nodes, we may have
2235 phis to translate through. */
2236 else
2237 {
2238 VEC(basic_block, heap) * worklist;
2239 size_t i;
2240 basic_block bprime, first = NULL;
2241
2242 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2243 FOR_EACH_EDGE (e, ei, block->succs)
2244 {
2245 if (!first
2246 && BB_VISITED (e->dest))
2247 first = e->dest;
2248 else if (BB_VISITED (e->dest))
2249 VEC_quick_push (basic_block, worklist, e->dest);
2250 }
2251
2252 /* Of multiple successors we have to have visited one already. */
2253 if (!first)
2254 {
2255 SET_BIT (changed_blocks, block->index);
2256 BB_VISITED (block) = 0;
2257 BB_DEFERRED (block) = 1;
2258 changed = true;
2259 VEC_free (basic_block, heap, worklist);
2260 goto maybe_dump_sets;
2261 }
2262
2263 if (!gimple_seq_empty_p (phi_nodes (first)))
2264 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2265 else
2266 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2267
2268 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2269 {
2270 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2271 {
2272 bitmap_set_t tmp = bitmap_set_new ();
2273 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2274 bitmap_set_and (ANTIC_OUT, tmp);
2275 bitmap_set_free (tmp);
2276 }
2277 else
2278 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2279 }
2280 VEC_free (basic_block, heap, worklist);
2281 }
2282
2283 /* Prune expressions that are clobbered in block and thus become
2284 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2285 prune_clobbered_mems (ANTIC_OUT, block);
2286
2287 /* Generate ANTIC_OUT - TMP_GEN. */
2288 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2289
2290 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2291 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2292 TMP_GEN (block));
2293
2294 /* Then union in the ANTIC_OUT - TMP_GEN values,
2295 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2296 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2297 bitmap_value_insert_into_set (ANTIC_IN (block),
2298 expression_for_id (bii));
2299
2300 clean (ANTIC_IN (block), block);
2301
2302 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2303 {
2304 changed = true;
2305 SET_BIT (changed_blocks, block->index);
2306 FOR_EACH_EDGE (e, ei, block->preds)
2307 SET_BIT (changed_blocks, e->src->index);
2308 }
2309 else
2310 RESET_BIT (changed_blocks, block->index);
2311
2312 maybe_dump_sets:
2313 if (dump_file && (dump_flags & TDF_DETAILS))
2314 {
2315 if (!BB_DEFERRED (block) || BB_VISITED (block))
2316 {
2317 if (ANTIC_OUT)
2318 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2319
2320 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2321 block->index);
2322
2323 if (S)
2324 print_bitmap_set (dump_file, S, "S", block->index);
2325 }
2326 else
2327 {
2328 fprintf (dump_file,
2329 "Block %d was deferred for a future iteration.\n",
2330 block->index);
2331 }
2332 }
2333 if (old)
2334 bitmap_set_free (old);
2335 if (S)
2336 bitmap_set_free (S);
2337 if (ANTIC_OUT)
2338 bitmap_set_free (ANTIC_OUT);
2339 return changed;
2340 }
2341
2342 /* Compute PARTIAL_ANTIC for BLOCK.
2343
2344 If succs(BLOCK) > 1 then
2345 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2346 in ANTIC_OUT for all succ(BLOCK)
2347 else if succs(BLOCK) == 1 then
2348 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2349
2350 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2351 - ANTIC_IN[BLOCK])
2352
2353 */
2354 static bool
2355 compute_partial_antic_aux (basic_block block,
2356 bool block_has_abnormal_pred_edge)
2357 {
2358 bool changed = false;
2359 bitmap_set_t old_PA_IN;
2360 bitmap_set_t PA_OUT;
2361 edge e;
2362 edge_iterator ei;
2363 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2364
2365 old_PA_IN = PA_OUT = NULL;
2366
2367 /* If any edges from predecessors are abnormal, antic_in is empty,
2368 so do nothing. */
2369 if (block_has_abnormal_pred_edge)
2370 goto maybe_dump_sets;
2371
2372 /* If there are too many partially anticipatable values in the
2373 block, phi_translate_set can take an exponential time: stop
2374 before the translation starts. */
2375 if (max_pa
2376 && single_succ_p (block)
2377 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2378 goto maybe_dump_sets;
2379
2380 old_PA_IN = PA_IN (block);
2381 PA_OUT = bitmap_set_new ();
2382
2383 /* If the block has no successors, ANTIC_OUT is empty. */
2384 if (EDGE_COUNT (block->succs) == 0)
2385 ;
2386 /* If we have one successor, we could have some phi nodes to
2387 translate through. Note that we can't phi translate across DFS
2388 back edges in partial antic, because it uses a union operation on
2389 the successors. For recurrences like IV's, we will end up
2390 generating a new value in the set on each go around (i + 3 (VH.1)
2391 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2392 else if (single_succ_p (block))
2393 {
2394 basic_block succ = single_succ (block);
2395 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2396 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2397 }
2398 /* If we have multiple successors, we take the union of all of
2399 them. */
2400 else
2401 {
2402 VEC(basic_block, heap) * worklist;
2403 size_t i;
2404 basic_block bprime;
2405
2406 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2407 FOR_EACH_EDGE (e, ei, block->succs)
2408 {
2409 if (e->flags & EDGE_DFS_BACK)
2410 continue;
2411 VEC_quick_push (basic_block, worklist, e->dest);
2412 }
2413 if (VEC_length (basic_block, worklist) > 0)
2414 {
2415 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2416 {
2417 unsigned int i;
2418 bitmap_iterator bi;
2419
2420 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2421 bitmap_value_insert_into_set (PA_OUT,
2422 expression_for_id (i));
2423 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2424 {
2425 bitmap_set_t pa_in = bitmap_set_new ();
2426 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2427 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2428 bitmap_value_insert_into_set (PA_OUT,
2429 expression_for_id (i));
2430 bitmap_set_free (pa_in);
2431 }
2432 else
2433 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2434 bitmap_value_insert_into_set (PA_OUT,
2435 expression_for_id (i));
2436 }
2437 }
2438 VEC_free (basic_block, heap, worklist);
2439 }
2440
2441 /* Prune expressions that are clobbered in block and thus become
2442 invalid if translated from PA_OUT to PA_IN. */
2443 prune_clobbered_mems (PA_OUT, block);
2444
2445 /* PA_IN starts with PA_OUT - TMP_GEN.
2446 Then we subtract things from ANTIC_IN. */
2447 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2448
2449 /* For partial antic, we want to put back in the phi results, since
2450 we will properly avoid making them partially antic over backedges. */
2451 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2452 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2453
2454 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2455 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2456
2457 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2458
2459 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2460 {
2461 changed = true;
2462 SET_BIT (changed_blocks, block->index);
2463 FOR_EACH_EDGE (e, ei, block->preds)
2464 SET_BIT (changed_blocks, e->src->index);
2465 }
2466 else
2467 RESET_BIT (changed_blocks, block->index);
2468
2469 maybe_dump_sets:
2470 if (dump_file && (dump_flags & TDF_DETAILS))
2471 {
2472 if (PA_OUT)
2473 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2474
2475 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2476 }
2477 if (old_PA_IN)
2478 bitmap_set_free (old_PA_IN);
2479 if (PA_OUT)
2480 bitmap_set_free (PA_OUT);
2481 return changed;
2482 }
2483
2484 /* Compute ANTIC and partial ANTIC sets. */
2485
2486 static void
2487 compute_antic (void)
2488 {
2489 bool changed = true;
2490 int num_iterations = 0;
2491 basic_block block;
2492 int i;
2493
2494 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2495 We pre-build the map of blocks with incoming abnormal edges here. */
2496 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2497 sbitmap_zero (has_abnormal_preds);
2498
2499 FOR_EACH_BB (block)
2500 {
2501 edge_iterator ei;
2502 edge e;
2503
2504 FOR_EACH_EDGE (e, ei, block->preds)
2505 {
2506 e->flags &= ~EDGE_DFS_BACK;
2507 if (e->flags & EDGE_ABNORMAL)
2508 {
2509 SET_BIT (has_abnormal_preds, block->index);
2510 break;
2511 }
2512 }
2513
2514 BB_VISITED (block) = 0;
2515 BB_DEFERRED (block) = 0;
2516
2517 /* While we are here, give empty ANTIC_IN sets to each block. */
2518 ANTIC_IN (block) = bitmap_set_new ();
2519 PA_IN (block) = bitmap_set_new ();
2520 }
2521
2522 /* At the exit block we anticipate nothing. */
2523 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2524 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2525 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2526
2527 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2528 sbitmap_ones (changed_blocks);
2529 while (changed)
2530 {
2531 if (dump_file && (dump_flags & TDF_DETAILS))
2532 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2533 /* ??? We need to clear our PHI translation cache here as the
2534 ANTIC sets shrink and we restrict valid translations to
2535 those having operands with leaders in ANTIC. Same below
2536 for PA ANTIC computation. */
2537 num_iterations++;
2538 changed = false;
2539 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2540 {
2541 if (TEST_BIT (changed_blocks, postorder[i]))
2542 {
2543 basic_block block = BASIC_BLOCK (postorder[i]);
2544 changed |= compute_antic_aux (block,
2545 TEST_BIT (has_abnormal_preds,
2546 block->index));
2547 }
2548 }
2549 /* Theoretically possible, but *highly* unlikely. */
2550 gcc_checking_assert (num_iterations < 500);
2551 }
2552
2553 statistics_histogram_event (cfun, "compute_antic iterations",
2554 num_iterations);
2555
2556 if (do_partial_partial)
2557 {
2558 sbitmap_ones (changed_blocks);
2559 mark_dfs_back_edges ();
2560 num_iterations = 0;
2561 changed = true;
2562 while (changed)
2563 {
2564 if (dump_file && (dump_flags & TDF_DETAILS))
2565 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2566 num_iterations++;
2567 changed = false;
2568 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2569 {
2570 if (TEST_BIT (changed_blocks, postorder[i]))
2571 {
2572 basic_block block = BASIC_BLOCK (postorder[i]);
2573 changed
2574 |= compute_partial_antic_aux (block,
2575 TEST_BIT (has_abnormal_preds,
2576 block->index));
2577 }
2578 }
2579 /* Theoretically possible, but *highly* unlikely. */
2580 gcc_checking_assert (num_iterations < 500);
2581 }
2582 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2583 num_iterations);
2584 }
2585 sbitmap_free (has_abnormal_preds);
2586 sbitmap_free (changed_blocks);
2587 }
2588
2589 /* Return true if OP is a tree which we can perform PRE on.
2590 This may not match the operations we can value number, but in
2591 a perfect world would. */
2592
2593 static bool
2594 can_PRE_operation (tree op)
2595 {
2596 return UNARY_CLASS_P (op)
2597 || BINARY_CLASS_P (op)
2598 || COMPARISON_CLASS_P (op)
2599 || TREE_CODE (op) == MEM_REF
2600 || TREE_CODE (op) == COMPONENT_REF
2601 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2602 || TREE_CODE (op) == CALL_EXPR
2603 || TREE_CODE (op) == ARRAY_REF;
2604 }
2605
2606
2607 /* Inserted expressions are placed onto this worklist, which is used
2608 for performing quick dead code elimination of insertions we made
2609 that didn't turn out to be necessary. */
2610 static bitmap inserted_exprs;
2611
2612 /* Pool allocated fake store expressions are placed onto this
2613 worklist, which, after performing dead code elimination, is walked
2614 to see which expressions need to be put into GC'able memory */
2615 static VEC(gimple, heap) *need_creation;
2616
2617 /* The actual worker for create_component_ref_by_pieces. */
2618
2619 static tree
2620 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2621 unsigned int *operand, gimple_seq *stmts,
2622 gimple domstmt)
2623 {
2624 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2625 *operand);
2626 tree genop;
2627 ++*operand;
2628 switch (currop->opcode)
2629 {
2630 case CALL_EXPR:
2631 {
2632 tree folded, sc = NULL_TREE;
2633 unsigned int nargs = 0;
2634 tree fn, *args;
2635 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2636 fn = currop->op0;
2637 else
2638 {
2639 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2640 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2641 if (!fn)
2642 return NULL_TREE;
2643 }
2644 if (currop->op1)
2645 {
2646 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2647 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2648 if (!sc)
2649 return NULL_TREE;
2650 }
2651 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2652 ref->operands) - 1);
2653 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2654 {
2655 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2656 operand, stmts,
2657 domstmt);
2658 if (!args[nargs])
2659 {
2660 free (args);
2661 return NULL_TREE;
2662 }
2663 nargs++;
2664 }
2665 folded = build_call_array (currop->type,
2666 (TREE_CODE (fn) == FUNCTION_DECL
2667 ? build_fold_addr_expr (fn) : fn),
2668 nargs, args);
2669 free (args);
2670 if (sc)
2671 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2672 return folded;
2673 }
2674 break;
2675 case MEM_REF:
2676 {
2677 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2678 stmts, domstmt);
2679 tree offset = currop->op0;
2680 if (!baseop)
2681 return NULL_TREE;
2682 if (TREE_CODE (baseop) == ADDR_EXPR
2683 && handled_component_p (TREE_OPERAND (baseop, 0)))
2684 {
2685 HOST_WIDE_INT off;
2686 tree base;
2687 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2688 &off);
2689 gcc_assert (base);
2690 offset = int_const_binop (PLUS_EXPR, offset,
2691 build_int_cst (TREE_TYPE (offset),
2692 off));
2693 baseop = build_fold_addr_expr (base);
2694 }
2695 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2696 }
2697 break;
2698 case TARGET_MEM_REF:
2699 {
2700 pre_expr op0expr, op1expr;
2701 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2702 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2703 ++*operand);
2704 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2705 stmts, domstmt);
2706 if (!baseop)
2707 return NULL_TREE;
2708 if (currop->op0)
2709 {
2710 op0expr = get_or_alloc_expr_for (currop->op0);
2711 genop0 = find_or_generate_expression (block, op0expr,
2712 stmts, domstmt);
2713 if (!genop0)
2714 return NULL_TREE;
2715 }
2716 if (nextop->op0)
2717 {
2718 op1expr = get_or_alloc_expr_for (nextop->op0);
2719 genop1 = find_or_generate_expression (block, op1expr,
2720 stmts, domstmt);
2721 if (!genop1)
2722 return NULL_TREE;
2723 }
2724 return build5 (TARGET_MEM_REF, currop->type,
2725 baseop, currop->op2, genop0, currop->op1, genop1);
2726 }
2727 break;
2728 case ADDR_EXPR:
2729 if (currop->op0)
2730 {
2731 gcc_assert (is_gimple_min_invariant (currop->op0));
2732 return currop->op0;
2733 }
2734 /* Fallthrough. */
2735 case REALPART_EXPR:
2736 case IMAGPART_EXPR:
2737 case VIEW_CONVERT_EXPR:
2738 {
2739 tree folded;
2740 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2741 operand,
2742 stmts, domstmt);
2743 if (!genop0)
2744 return NULL_TREE;
2745 folded = fold_build1 (currop->opcode, currop->type,
2746 genop0);
2747 return folded;
2748 }
2749 break;
2750 case WITH_SIZE_EXPR:
2751 {
2752 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2753 stmts, domstmt);
2754 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2755 tree genop1;
2756
2757 if (!genop0)
2758 return NULL_TREE;
2759
2760 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2761 if (!genop1)
2762 return NULL_TREE;
2763
2764 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2765 }
2766 break;
2767 case BIT_FIELD_REF:
2768 {
2769 tree folded;
2770 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2771 stmts, domstmt);
2772 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2773 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2774 tree genop1;
2775 tree genop2;
2776
2777 if (!genop0)
2778 return NULL_TREE;
2779 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2780 if (!genop1)
2781 return NULL_TREE;
2782 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2783 if (!genop2)
2784 return NULL_TREE;
2785 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2786 genop2);
2787 return folded;
2788 }
2789
2790 /* For array ref vn_reference_op's, operand 1 of the array ref
2791 is op0 of the reference op and operand 3 of the array ref is
2792 op1. */
2793 case ARRAY_RANGE_REF:
2794 case ARRAY_REF:
2795 {
2796 tree genop0;
2797 tree genop1 = currop->op0;
2798 pre_expr op1expr;
2799 tree genop2 = currop->op1;
2800 pre_expr op2expr;
2801 tree genop3 = currop->op2;
2802 pre_expr op3expr;
2803 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2804 stmts, domstmt);
2805 if (!genop0)
2806 return NULL_TREE;
2807 op1expr = get_or_alloc_expr_for (genop1);
2808 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2809 if (!genop1)
2810 return NULL_TREE;
2811 if (genop2)
2812 {
2813 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2814 /* Drop zero minimum index if redundant. */
2815 if (integer_zerop (genop2)
2816 && (!domain_type
2817 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2818 genop2 = NULL_TREE;
2819 else
2820 {
2821 op2expr = get_or_alloc_expr_for (genop2);
2822 genop2 = find_or_generate_expression (block, op2expr, stmts,
2823 domstmt);
2824 if (!genop2)
2825 return NULL_TREE;
2826 }
2827 }
2828 if (genop3)
2829 {
2830 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2831 /* We can't always put a size in units of the element alignment
2832 here as the element alignment may be not visible. See
2833 PR43783. Simply drop the element size for constant
2834 sizes. */
2835 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2836 genop3 = NULL_TREE;
2837 else
2838 {
2839 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2840 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2841 op3expr = get_or_alloc_expr_for (genop3);
2842 genop3 = find_or_generate_expression (block, op3expr, stmts,
2843 domstmt);
2844 if (!genop3)
2845 return NULL_TREE;
2846 }
2847 }
2848 return build4 (currop->opcode, currop->type, genop0, genop1,
2849 genop2, genop3);
2850 }
2851 case COMPONENT_REF:
2852 {
2853 tree op0;
2854 tree op1;
2855 tree genop2 = currop->op1;
2856 pre_expr op2expr;
2857 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2858 stmts, domstmt);
2859 if (!op0)
2860 return NULL_TREE;
2861 /* op1 should be a FIELD_DECL, which are represented by
2862 themselves. */
2863 op1 = currop->op0;
2864 if (genop2)
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 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2874 genop2);
2875 }
2876 break;
2877 case SSA_NAME:
2878 {
2879 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2880 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2881 return genop;
2882 }
2883 case STRING_CST:
2884 case INTEGER_CST:
2885 case COMPLEX_CST:
2886 case VECTOR_CST:
2887 case REAL_CST:
2888 case CONSTRUCTOR:
2889 case VAR_DECL:
2890 case PARM_DECL:
2891 case CONST_DECL:
2892 case RESULT_DECL:
2893 case FUNCTION_DECL:
2894 return currop->op0;
2895
2896 default:
2897 gcc_unreachable ();
2898 }
2899 }
2900
2901 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2902 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2903 trying to rename aggregates into ssa form directly, which is a no no.
2904
2905 Thus, this routine doesn't create temporaries, it just builds a
2906 single access expression for the array, calling
2907 find_or_generate_expression to build the innermost pieces.
2908
2909 This function is a subroutine of create_expression_by_pieces, and
2910 should not be called on it's own unless you really know what you
2911 are doing. */
2912
2913 static tree
2914 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2915 gimple_seq *stmts, gimple domstmt)
2916 {
2917 unsigned int op = 0;
2918 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2919 }
2920
2921 /* Find a leader for an expression, or generate one using
2922 create_expression_by_pieces if it's ANTIC but
2923 complex.
2924 BLOCK is the basic_block we are looking for leaders in.
2925 EXPR is the expression to find a leader or generate for.
2926 STMTS is the statement list to put the inserted expressions on.
2927 Returns the SSA_NAME of the LHS of the generated expression or the
2928 leader.
2929 DOMSTMT if non-NULL is a statement that should be dominated by
2930 all uses in the generated expression. If DOMSTMT is non-NULL this
2931 routine can fail and return NULL_TREE. Otherwise it will assert
2932 on failure. */
2933
2934 static tree
2935 find_or_generate_expression (basic_block block, pre_expr expr,
2936 gimple_seq *stmts, gimple domstmt)
2937 {
2938 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2939 get_expr_value_id (expr), domstmt);
2940 tree genop = NULL;
2941 if (leader)
2942 {
2943 if (leader->kind == NAME)
2944 genop = PRE_EXPR_NAME (leader);
2945 else if (leader->kind == CONSTANT)
2946 genop = PRE_EXPR_CONSTANT (leader);
2947 }
2948
2949 /* If it's still NULL, it must be a complex expression, so generate
2950 it recursively. Not so if inserting expressions for values generated
2951 by SCCVN. */
2952 if (genop == NULL
2953 && !domstmt)
2954 {
2955 bitmap_set_t exprset;
2956 unsigned int lookfor = get_expr_value_id (expr);
2957 bool handled = false;
2958 bitmap_iterator bi;
2959 unsigned int i;
2960
2961 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2962 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2963 {
2964 pre_expr temp = expression_for_id (i);
2965 if (temp->kind != NAME)
2966 {
2967 handled = true;
2968 genop = create_expression_by_pieces (block, temp, stmts,
2969 domstmt,
2970 get_expr_type (expr));
2971 break;
2972 }
2973 }
2974 if (!handled && domstmt)
2975 return NULL_TREE;
2976
2977 gcc_assert (handled);
2978 }
2979 return genop;
2980 }
2981
2982 #define NECESSARY GF_PLF_1
2983
2984 /* Create an expression in pieces, so that we can handle very complex
2985 expressions that may be ANTIC, but not necessary GIMPLE.
2986 BLOCK is the basic block the expression will be inserted into,
2987 EXPR is the expression to insert (in value form)
2988 STMTS is a statement list to append the necessary insertions into.
2989
2990 This function will die if we hit some value that shouldn't be
2991 ANTIC but is (IE there is no leader for it, or its components).
2992 This function may also generate expressions that are themselves
2993 partially or fully redundant. Those that are will be either made
2994 fully redundant during the next iteration of insert (for partially
2995 redundant ones), or eliminated by eliminate (for fully redundant
2996 ones).
2997
2998 If DOMSTMT is non-NULL then we make sure that all uses in the
2999 expressions dominate that statement. In this case the function
3000 can return NULL_TREE to signal failure. */
3001
3002 static tree
3003 create_expression_by_pieces (basic_block block, pre_expr expr,
3004 gimple_seq *stmts, gimple domstmt, tree type)
3005 {
3006 tree temp, name;
3007 tree folded;
3008 gimple_seq forced_stmts = NULL;
3009 unsigned int value_id;
3010 gimple_stmt_iterator gsi;
3011 tree exprtype = type ? type : get_expr_type (expr);
3012 pre_expr nameexpr;
3013 gimple newstmt;
3014
3015 switch (expr->kind)
3016 {
3017 /* We may hit the NAME/CONSTANT case if we have to convert types
3018 that value numbering saw through. */
3019 case NAME:
3020 folded = PRE_EXPR_NAME (expr);
3021 break;
3022 case CONSTANT:
3023 folded = PRE_EXPR_CONSTANT (expr);
3024 break;
3025 case REFERENCE:
3026 {
3027 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3028 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3029 }
3030 break;
3031 case NARY:
3032 {
3033 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3034 tree genop[4];
3035 unsigned i;
3036 for (i = 0; i < nary->length; ++i)
3037 {
3038 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3039 genop[i] = find_or_generate_expression (block, op,
3040 stmts, domstmt);
3041 if (!genop[i])
3042 return NULL_TREE;
3043 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3044 may have conversions stripped. */
3045 if (nary->opcode == POINTER_PLUS_EXPR)
3046 {
3047 if (i == 0)
3048 genop[i] = fold_convert (nary->type, genop[i]);
3049 else if (i == 1)
3050 genop[i] = convert_to_ptrofftype (genop[i]);
3051 }
3052 else
3053 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3054 }
3055 if (nary->opcode == CONSTRUCTOR)
3056 {
3057 VEC(constructor_elt,gc) *elts = NULL;
3058 for (i = 0; i < nary->length; ++i)
3059 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3060 folded = build_constructor (nary->type, elts);
3061 }
3062 else
3063 {
3064 switch (nary->length)
3065 {
3066 case 1:
3067 folded = fold_build1 (nary->opcode, nary->type,
3068 genop[0]);
3069 break;
3070 case 2:
3071 folded = fold_build2 (nary->opcode, nary->type,
3072 genop[0], genop[1]);
3073 break;
3074 case 3:
3075 folded = fold_build3 (nary->opcode, nary->type,
3076 genop[0], genop[1], genop[3]);
3077 break;
3078 default:
3079 gcc_unreachable ();
3080 }
3081 }
3082 }
3083 break;
3084 default:
3085 return NULL_TREE;
3086 }
3087
3088 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3089 folded = fold_convert (exprtype, folded);
3090
3091 /* Force the generated expression to be a sequence of GIMPLE
3092 statements.
3093 We have to call unshare_expr because force_gimple_operand may
3094 modify the tree we pass to it. */
3095 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3096 false, NULL);
3097
3098 /* If we have any intermediate expressions to the value sets, add them
3099 to the value sets and chain them in the instruction stream. */
3100 if (forced_stmts)
3101 {
3102 gsi = gsi_start (forced_stmts);
3103 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3104 {
3105 gimple stmt = gsi_stmt (gsi);
3106 tree forcedname = gimple_get_lhs (stmt);
3107 pre_expr nameexpr;
3108
3109 if (TREE_CODE (forcedname) == SSA_NAME)
3110 {
3111 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3112 VN_INFO_GET (forcedname)->valnum = forcedname;
3113 VN_INFO (forcedname)->value_id = get_next_value_id ();
3114 nameexpr = get_or_alloc_expr_for_name (forcedname);
3115 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3116 if (!in_fre)
3117 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3118 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3119 }
3120 mark_symbols_for_renaming (stmt);
3121 }
3122 gimple_seq_add_seq (stmts, forced_stmts);
3123 }
3124
3125 /* Build and insert the assignment of the end result to the temporary
3126 that we will return. */
3127 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3128 pretemp = create_tmp_reg (exprtype, "pretmp");
3129
3130 temp = pretemp;
3131 add_referenced_var (temp);
3132
3133 newstmt = gimple_build_assign (temp, folded);
3134 name = make_ssa_name (temp, newstmt);
3135 gimple_assign_set_lhs (newstmt, name);
3136 gimple_set_plf (newstmt, NECESSARY, false);
3137
3138 gimple_seq_add_stmt (stmts, newstmt);
3139 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3140
3141 /* All the symbols in NEWEXPR should be put into SSA form. */
3142 mark_symbols_for_renaming (newstmt);
3143
3144 /* Fold the last statement. */
3145 gsi = gsi_last (*stmts);
3146 if (fold_stmt_inplace (&gsi))
3147 update_stmt (gsi_stmt (gsi));
3148
3149 /* Add a value number to the temporary.
3150 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3151 we are creating the expression by pieces, and this particular piece of
3152 the expression may have been represented. There is no harm in replacing
3153 here. */
3154 VN_INFO_GET (name)->valnum = name;
3155 value_id = get_expr_value_id (expr);
3156 VN_INFO (name)->value_id = value_id;
3157 nameexpr = get_or_alloc_expr_for_name (name);
3158 add_to_value (value_id, nameexpr);
3159 if (NEW_SETS (block))
3160 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3161 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3162
3163 pre_stats.insertions++;
3164 if (dump_file && (dump_flags & TDF_DETAILS))
3165 {
3166 fprintf (dump_file, "Inserted ");
3167 print_gimple_stmt (dump_file, newstmt, 0, 0);
3168 fprintf (dump_file, " in predecessor %d\n", block->index);
3169 }
3170
3171 return name;
3172 }
3173
3174
3175 /* Returns true if we want to inhibit the insertions of PHI nodes
3176 for the given EXPR for basic block BB (a member of a loop).
3177 We want to do this, when we fear that the induction variable we
3178 create might inhibit vectorization. */
3179
3180 static bool
3181 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3182 {
3183 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3184 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3185 vn_reference_op_t op;
3186 unsigned i;
3187
3188 /* If we aren't going to vectorize we don't inhibit anything. */
3189 if (!flag_tree_vectorize)
3190 return false;
3191
3192 /* Otherwise we inhibit the insertion when the address of the
3193 memory reference is a simple induction variable. In other
3194 cases the vectorizer won't do anything anyway (either it's
3195 loop invariant or a complicated expression). */
3196 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3197 {
3198 switch (op->opcode)
3199 {
3200 case CALL_EXPR:
3201 /* Calls are not a problem. */
3202 return false;
3203
3204 case ARRAY_REF:
3205 case ARRAY_RANGE_REF:
3206 if (TREE_CODE (op->op0) != SSA_NAME)
3207 break;
3208 /* Fallthru. */
3209 case SSA_NAME:
3210 {
3211 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3212 affine_iv iv;
3213 /* Default defs are loop invariant. */
3214 if (!defbb)
3215 break;
3216 /* Defined outside this loop, also loop invariant. */
3217 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3218 break;
3219 /* If it's a simple induction variable inhibit insertion,
3220 the vectorizer might be interested in this one. */
3221 if (simple_iv (bb->loop_father, bb->loop_father,
3222 op->op0, &iv, true))
3223 return true;
3224 /* No simple IV, vectorizer can't do anything, hence no
3225 reason to inhibit the transformation for this operand. */
3226 break;
3227 }
3228 default:
3229 break;
3230 }
3231 }
3232 return false;
3233 }
3234
3235 /* Insert the to-be-made-available values of expression EXPRNUM for each
3236 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3237 merge the result with a phi node, given the same value number as
3238 NODE. Return true if we have inserted new stuff. */
3239
3240 static bool
3241 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3242 pre_expr *avail)
3243 {
3244 pre_expr expr = expression_for_id (exprnum);
3245 pre_expr newphi;
3246 unsigned int val = get_expr_value_id (expr);
3247 edge pred;
3248 bool insertions = false;
3249 bool nophi = false;
3250 basic_block bprime;
3251 pre_expr eprime;
3252 edge_iterator ei;
3253 tree type = get_expr_type (expr);
3254 tree temp;
3255 gimple phi;
3256
3257 /* Make sure we aren't creating an induction variable. */
3258 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3259 {
3260 bool firstinsideloop = false;
3261 bool secondinsideloop = false;
3262 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3263 EDGE_PRED (block, 0)->src);
3264 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3265 EDGE_PRED (block, 1)->src);
3266 /* Induction variables only have one edge inside the loop. */
3267 if ((firstinsideloop ^ secondinsideloop)
3268 && (expr->kind != REFERENCE
3269 || inhibit_phi_insertion (block, expr)))
3270 {
3271 if (dump_file && (dump_flags & TDF_DETAILS))
3272 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3273 nophi = true;
3274 }
3275 }
3276
3277 /* Make the necessary insertions. */
3278 FOR_EACH_EDGE (pred, ei, block->preds)
3279 {
3280 gimple_seq stmts = NULL;
3281 tree builtexpr;
3282 bprime = pred->src;
3283 eprime = avail[bprime->index];
3284
3285 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3286 {
3287 builtexpr = create_expression_by_pieces (bprime,
3288 eprime,
3289 &stmts, NULL,
3290 type);
3291 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3292 gsi_insert_seq_on_edge (pred, stmts);
3293 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3294 insertions = true;
3295 }
3296 else if (eprime->kind == CONSTANT)
3297 {
3298 /* Constants may not have the right type, fold_convert
3299 should give us back a constant with the right type.
3300 */
3301 tree constant = PRE_EXPR_CONSTANT (eprime);
3302 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3303 {
3304 tree builtexpr = fold_convert (type, constant);
3305 if (!is_gimple_min_invariant (builtexpr))
3306 {
3307 tree forcedexpr = force_gimple_operand (builtexpr,
3308 &stmts, true,
3309 NULL);
3310 if (!is_gimple_min_invariant (forcedexpr))
3311 {
3312 if (forcedexpr != builtexpr)
3313 {
3314 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3315 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3316 }
3317 if (stmts)
3318 {
3319 gimple_stmt_iterator gsi;
3320 gsi = gsi_start (stmts);
3321 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3322 {
3323 gimple stmt = gsi_stmt (gsi);
3324 tree lhs = gimple_get_lhs (stmt);
3325 if (TREE_CODE (lhs) == SSA_NAME)
3326 bitmap_set_bit (inserted_exprs,
3327 SSA_NAME_VERSION (lhs));
3328 gimple_set_plf (stmt, NECESSARY, false);
3329 }
3330 gsi_insert_seq_on_edge (pred, stmts);
3331 }
3332 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3333 }
3334 }
3335 else
3336 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3337 }
3338 }
3339 else if (eprime->kind == NAME)
3340 {
3341 /* We may have to do a conversion because our value
3342 numbering can look through types in certain cases, but
3343 our IL requires all operands of a phi node have the same
3344 type. */
3345 tree name = PRE_EXPR_NAME (eprime);
3346 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3347 {
3348 tree builtexpr;
3349 tree forcedexpr;
3350 builtexpr = fold_convert (type, name);
3351 forcedexpr = force_gimple_operand (builtexpr,
3352 &stmts, true,
3353 NULL);
3354
3355 if (forcedexpr != name)
3356 {
3357 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3358 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3359 }
3360
3361 if (stmts)
3362 {
3363 gimple_stmt_iterator gsi;
3364 gsi = gsi_start (stmts);
3365 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3366 {
3367 gimple stmt = gsi_stmt (gsi);
3368 tree lhs = gimple_get_lhs (stmt);
3369 if (TREE_CODE (lhs) == SSA_NAME)
3370 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3371 gimple_set_plf (stmt, NECESSARY, false);
3372 }
3373 gsi_insert_seq_on_edge (pred, stmts);
3374 }
3375 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3376 }
3377 }
3378 }
3379 /* If we didn't want a phi node, and we made insertions, we still have
3380 inserted new stuff, and thus return true. If we didn't want a phi node,
3381 and didn't make insertions, we haven't added anything new, so return
3382 false. */
3383 if (nophi && insertions)
3384 return true;
3385 else if (nophi && !insertions)
3386 return false;
3387
3388 /* Now build a phi for the new variable. */
3389 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3390 prephitemp = create_tmp_var (type, "prephitmp");
3391
3392 temp = prephitemp;
3393 add_referenced_var (temp);
3394
3395 if (TREE_CODE (type) == COMPLEX_TYPE
3396 || TREE_CODE (type) == VECTOR_TYPE)
3397 DECL_GIMPLE_REG_P (temp) = 1;
3398 phi = create_phi_node (temp, block);
3399
3400 gimple_set_plf (phi, NECESSARY, false);
3401 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3402 VN_INFO (gimple_phi_result (phi))->value_id = val;
3403 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3404 FOR_EACH_EDGE (pred, ei, block->preds)
3405 {
3406 pre_expr ae = avail[pred->src->index];
3407 gcc_assert (get_expr_type (ae) == type
3408 || useless_type_conversion_p (type, get_expr_type (ae)));
3409 if (ae->kind == CONSTANT)
3410 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3411 else
3412 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3413 UNKNOWN_LOCATION);
3414 }
3415
3416 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3417 add_to_value (val, newphi);
3418
3419 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3420 this insertion, since we test for the existence of this value in PHI_GEN
3421 before proceeding with the partial redundancy checks in insert_aux.
3422
3423 The value may exist in AVAIL_OUT, in particular, it could be represented
3424 by the expression we are trying to eliminate, in which case we want the
3425 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3426 inserted there.
3427
3428 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3429 this block, because if it did, it would have existed in our dominator's
3430 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3431 */
3432
3433 bitmap_insert_into_set (PHI_GEN (block), newphi);
3434 bitmap_value_replace_in_set (AVAIL_OUT (block),
3435 newphi);
3436 bitmap_insert_into_set (NEW_SETS (block),
3437 newphi);
3438
3439 if (dump_file && (dump_flags & TDF_DETAILS))
3440 {
3441 fprintf (dump_file, "Created phi ");
3442 print_gimple_stmt (dump_file, phi, 0, 0);
3443 fprintf (dump_file, " in block %d\n", block->index);
3444 }
3445 pre_stats.phis++;
3446 return true;
3447 }
3448
3449
3450
3451 /* Perform insertion of partially redundant values.
3452 For BLOCK, do the following:
3453 1. Propagate the NEW_SETS of the dominator into the current block.
3454 If the block has multiple predecessors,
3455 2a. Iterate over the ANTIC expressions for the block to see if
3456 any of them are partially redundant.
3457 2b. If so, insert them into the necessary predecessors to make
3458 the expression fully redundant.
3459 2c. Insert a new PHI merging the values of the predecessors.
3460 2d. Insert the new PHI, and the new expressions, into the
3461 NEW_SETS set.
3462 3. Recursively call ourselves on the dominator children of BLOCK.
3463
3464 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3465 do_regular_insertion and do_partial_insertion.
3466
3467 */
3468
3469 static bool
3470 do_regular_insertion (basic_block block, basic_block dom)
3471 {
3472 bool new_stuff = false;
3473 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3474 pre_expr expr;
3475 int i;
3476
3477 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3478 {
3479 if (expr->kind != NAME)
3480 {
3481 pre_expr *avail;
3482 unsigned int val;
3483 bool by_some = false;
3484 bool cant_insert = false;
3485 bool all_same = true;
3486 pre_expr first_s = NULL;
3487 edge pred;
3488 basic_block bprime;
3489 pre_expr eprime = NULL;
3490 edge_iterator ei;
3491 pre_expr edoubleprime = NULL;
3492 bool do_insertion = false;
3493
3494 val = get_expr_value_id (expr);
3495 if (bitmap_set_contains_value (PHI_GEN (block), val))
3496 continue;
3497 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3498 {
3499 if (dump_file && (dump_flags & TDF_DETAILS))
3500 fprintf (dump_file, "Found fully redundant value\n");
3501 continue;
3502 }
3503
3504 avail = XCNEWVEC (pre_expr, last_basic_block);
3505 FOR_EACH_EDGE (pred, ei, block->preds)
3506 {
3507 unsigned int vprime;
3508
3509 /* We should never run insertion for the exit block
3510 and so not come across fake pred edges. */
3511 gcc_assert (!(pred->flags & EDGE_FAKE));
3512 bprime = pred->src;
3513 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3514 bprime, block);
3515
3516 /* eprime will generally only be NULL if the
3517 value of the expression, translated
3518 through the PHI for this predecessor, is
3519 undefined. If that is the case, we can't
3520 make the expression fully redundant,
3521 because its value is undefined along a
3522 predecessor path. We can thus break out
3523 early because it doesn't matter what the
3524 rest of the results are. */
3525 if (eprime == NULL)
3526 {
3527 cant_insert = true;
3528 break;
3529 }
3530
3531 eprime = fully_constant_expression (eprime);
3532 vprime = get_expr_value_id (eprime);
3533 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3534 vprime, NULL);
3535 if (edoubleprime == NULL)
3536 {
3537 avail[bprime->index] = eprime;
3538 all_same = false;
3539 }
3540 else
3541 {
3542 avail[bprime->index] = edoubleprime;
3543 by_some = true;
3544 /* We want to perform insertions to remove a redundancy on
3545 a path in the CFG we want to optimize for speed. */
3546 if (optimize_edge_for_speed_p (pred))
3547 do_insertion = true;
3548 if (first_s == NULL)
3549 first_s = edoubleprime;
3550 else if (!pre_expr_eq (first_s, edoubleprime))
3551 all_same = false;
3552 }
3553 }
3554 /* If we can insert it, it's not the same value
3555 already existing along every predecessor, and
3556 it's defined by some predecessor, it is
3557 partially redundant. */
3558 if (!cant_insert && !all_same && by_some)
3559 {
3560 if (!do_insertion)
3561 {
3562 if (dump_file && (dump_flags & TDF_DETAILS))
3563 {
3564 fprintf (dump_file, "Skipping partial redundancy for "
3565 "expression ");
3566 print_pre_expr (dump_file, expr);
3567 fprintf (dump_file, " (%04d), no redundancy on to be "
3568 "optimized for speed edge\n", val);
3569 }
3570 }
3571 else if (dbg_cnt (treepre_insert))
3572 {
3573 if (dump_file && (dump_flags & TDF_DETAILS))
3574 {
3575 fprintf (dump_file, "Found partial redundancy for "
3576 "expression ");
3577 print_pre_expr (dump_file, expr);
3578 fprintf (dump_file, " (%04d)\n",
3579 get_expr_value_id (expr));
3580 }
3581 if (insert_into_preds_of_block (block,
3582 get_expression_id (expr),
3583 avail))
3584 new_stuff = true;
3585 }
3586 }
3587 /* If all edges produce the same value and that value is
3588 an invariant, then the PHI has the same value on all
3589 edges. Note this. */
3590 else if (!cant_insert && all_same && eprime
3591 && (edoubleprime->kind == CONSTANT
3592 || edoubleprime->kind == NAME)
3593 && !value_id_constant_p (val))
3594 {
3595 unsigned int j;
3596 bitmap_iterator bi;
3597 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3598 value_expressions, val);
3599
3600 unsigned int new_val = get_expr_value_id (edoubleprime);
3601 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3602 {
3603 pre_expr expr = expression_for_id (j);
3604
3605 if (expr->kind == NAME)
3606 {
3607 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3608 /* Just reset the value id and valnum so it is
3609 the same as the constant we have discovered. */
3610 if (edoubleprime->kind == CONSTANT)
3611 {
3612 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3613 pre_stats.constified++;
3614 }
3615 else
3616 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3617 info->value_id = new_val;
3618 }
3619 }
3620 }
3621 free (avail);
3622 }
3623 }
3624
3625 VEC_free (pre_expr, heap, exprs);
3626 return new_stuff;
3627 }
3628
3629
3630 /* Perform insertion for partially anticipatable expressions. There
3631 is only one case we will perform insertion for these. This case is
3632 if the expression is partially anticipatable, and fully available.
3633 In this case, we know that putting it earlier will enable us to
3634 remove the later computation. */
3635
3636
3637 static bool
3638 do_partial_partial_insertion (basic_block block, basic_block dom)
3639 {
3640 bool new_stuff = false;
3641 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3642 pre_expr expr;
3643 int i;
3644
3645 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3646 {
3647 if (expr->kind != NAME)
3648 {
3649 pre_expr *avail;
3650 unsigned int val;
3651 bool by_all = true;
3652 bool cant_insert = false;
3653 edge pred;
3654 basic_block bprime;
3655 pre_expr eprime = NULL;
3656 edge_iterator ei;
3657
3658 val = get_expr_value_id (expr);
3659 if (bitmap_set_contains_value (PHI_GEN (block), val))
3660 continue;
3661 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3662 continue;
3663
3664 avail = XCNEWVEC (pre_expr, last_basic_block);
3665 FOR_EACH_EDGE (pred, ei, block->preds)
3666 {
3667 unsigned int vprime;
3668 pre_expr edoubleprime;
3669
3670 /* We should never run insertion for the exit block
3671 and so not come across fake pred edges. */
3672 gcc_assert (!(pred->flags & EDGE_FAKE));
3673 bprime = pred->src;
3674 eprime = phi_translate (expr, ANTIC_IN (block),
3675 PA_IN (block),
3676 bprime, block);
3677
3678 /* eprime will generally only be NULL if the
3679 value of the expression, translated
3680 through the PHI for this predecessor, is
3681 undefined. If that is the case, we can't
3682 make the expression fully redundant,
3683 because its value is undefined along a
3684 predecessor path. We can thus break out
3685 early because it doesn't matter what the
3686 rest of the results are. */
3687 if (eprime == NULL)
3688 {
3689 cant_insert = true;
3690 break;
3691 }
3692
3693 eprime = fully_constant_expression (eprime);
3694 vprime = get_expr_value_id (eprime);
3695 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3696 vprime, NULL);
3697 if (edoubleprime == NULL)
3698 {
3699 by_all = false;
3700 break;
3701 }
3702 else
3703 avail[bprime->index] = edoubleprime;
3704 }
3705
3706 /* If we can insert it, it's not the same value
3707 already existing along every predecessor, and
3708 it's defined by some predecessor, it is
3709 partially redundant. */
3710 if (!cant_insert && by_all)
3711 {
3712 edge succ;
3713 bool do_insertion = false;
3714
3715 /* Insert only if we can remove a later expression on a path
3716 that we want to optimize for speed.
3717 The phi node that we will be inserting in BLOCK is not free,
3718 and inserting it for the sake of !optimize_for_speed successor
3719 may cause regressions on the speed path. */
3720 FOR_EACH_EDGE (succ, ei, block->succs)
3721 {
3722 if (bitmap_set_contains_value (PA_IN (succ->dest), val))
3723 {
3724 if (optimize_edge_for_speed_p (succ))
3725 do_insertion = true;
3726 }
3727 }
3728
3729 if (!do_insertion)
3730 {
3731 if (dump_file && (dump_flags & TDF_DETAILS))
3732 {
3733 fprintf (dump_file, "Skipping partial partial redundancy "
3734 "for expression ");
3735 print_pre_expr (dump_file, expr);
3736 fprintf (dump_file, " (%04d), not partially anticipated "
3737 "on any to be optimized for speed edges\n", val);
3738 }
3739 }
3740 else if (dbg_cnt (treepre_insert))
3741 {
3742 pre_stats.pa_insert++;
3743 if (dump_file && (dump_flags & TDF_DETAILS))
3744 {
3745 fprintf (dump_file, "Found partial partial redundancy "
3746 "for expression ");
3747 print_pre_expr (dump_file, expr);
3748 fprintf (dump_file, " (%04d)\n",
3749 get_expr_value_id (expr));
3750 }
3751 if (insert_into_preds_of_block (block,
3752 get_expression_id (expr),
3753 avail))
3754 new_stuff = true;
3755 }
3756 }
3757 free (avail);
3758 }
3759 }
3760
3761 VEC_free (pre_expr, heap, exprs);
3762 return new_stuff;
3763 }
3764
3765 static bool
3766 insert_aux (basic_block block)
3767 {
3768 basic_block son;
3769 bool new_stuff = false;
3770
3771 if (block)
3772 {
3773 basic_block dom;
3774 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3775 if (dom)
3776 {
3777 unsigned i;
3778 bitmap_iterator bi;
3779 bitmap_set_t newset = NEW_SETS (dom);
3780 if (newset)
3781 {
3782 /* Note that we need to value_replace both NEW_SETS, and
3783 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3784 represented by some non-simple expression here that we want
3785 to replace it with. */
3786 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3787 {
3788 pre_expr expr = expression_for_id (i);
3789 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3790 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3791 }
3792 }
3793 if (!single_pred_p (block))
3794 {
3795 new_stuff |= do_regular_insertion (block, dom);
3796 if (do_partial_partial)
3797 new_stuff |= do_partial_partial_insertion (block, dom);
3798 }
3799 }
3800 }
3801 for (son = first_dom_son (CDI_DOMINATORS, block);
3802 son;
3803 son = next_dom_son (CDI_DOMINATORS, son))
3804 {
3805 new_stuff |= insert_aux (son);
3806 }
3807
3808 return new_stuff;
3809 }
3810
3811 /* Perform insertion of partially redundant values. */
3812
3813 static void
3814 insert (void)
3815 {
3816 bool new_stuff = true;
3817 basic_block bb;
3818 int num_iterations = 0;
3819
3820 FOR_ALL_BB (bb)
3821 NEW_SETS (bb) = bitmap_set_new ();
3822
3823 while (new_stuff)
3824 {
3825 num_iterations++;
3826 if (dump_file && dump_flags & TDF_DETAILS)
3827 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3828 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3829 }
3830 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3831 }
3832
3833
3834 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3835
3836 static void
3837 add_to_exp_gen (basic_block block, tree op)
3838 {
3839 if (!in_fre)
3840 {
3841 pre_expr result;
3842 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3843 return;
3844 result = get_or_alloc_expr_for_name (op);
3845 bitmap_value_insert_into_set (EXP_GEN (block), result);
3846 }
3847 }
3848
3849 /* Create value ids for PHI in BLOCK. */
3850
3851 static void
3852 make_values_for_phi (gimple phi, basic_block block)
3853 {
3854 tree result = gimple_phi_result (phi);
3855
3856 /* We have no need for virtual phis, as they don't represent
3857 actual computations. */
3858 if (is_gimple_reg (result))
3859 {
3860 pre_expr e = get_or_alloc_expr_for_name (result);
3861 add_to_value (get_expr_value_id (e), e);
3862 bitmap_insert_into_set (PHI_GEN (block), e);
3863 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3864 if (!in_fre)
3865 {
3866 unsigned i;
3867 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3868 {
3869 tree arg = gimple_phi_arg_def (phi, i);
3870 if (TREE_CODE (arg) == SSA_NAME)
3871 {
3872 e = get_or_alloc_expr_for_name (arg);
3873 add_to_value (get_expr_value_id (e), e);
3874 }
3875 }
3876 }
3877 }
3878 }
3879
3880 /* Compute the AVAIL set for all basic blocks.
3881
3882 This function performs value numbering of the statements in each basic
3883 block. The AVAIL sets are built from information we glean while doing
3884 this value numbering, since the AVAIL sets contain only one entry per
3885 value.
3886
3887 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3888 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3889
3890 static void
3891 compute_avail (void)
3892 {
3893
3894 basic_block block, son;
3895 basic_block *worklist;
3896 size_t sp = 0;
3897 unsigned i;
3898
3899 /* We pretend that default definitions are defined in the entry block.
3900 This includes function arguments and the static chain decl. */
3901 for (i = 1; i < num_ssa_names; ++i)
3902 {
3903 tree name = ssa_name (i);
3904 pre_expr e;
3905 if (!name
3906 || !SSA_NAME_IS_DEFAULT_DEF (name)
3907 || has_zero_uses (name)
3908 || !is_gimple_reg (name))
3909 continue;
3910
3911 e = get_or_alloc_expr_for_name (name);
3912 add_to_value (get_expr_value_id (e), e);
3913 if (!in_fre)
3914 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3915 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3916 }
3917
3918 /* Allocate the worklist. */
3919 worklist = XNEWVEC (basic_block, n_basic_blocks);
3920
3921 /* Seed the algorithm by putting the dominator children of the entry
3922 block on the worklist. */
3923 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3924 son;
3925 son = next_dom_son (CDI_DOMINATORS, son))
3926 worklist[sp++] = son;
3927
3928 /* Loop until the worklist is empty. */
3929 while (sp)
3930 {
3931 gimple_stmt_iterator gsi;
3932 gimple stmt;
3933 basic_block dom;
3934 unsigned int stmt_uid = 1;
3935
3936 /* Pick a block from the worklist. */
3937 block = worklist[--sp];
3938
3939 /* Initially, the set of available values in BLOCK is that of
3940 its immediate dominator. */
3941 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3942 if (dom)
3943 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3944
3945 /* Generate values for PHI nodes. */
3946 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3947 make_values_for_phi (gsi_stmt (gsi), block);
3948
3949 BB_MAY_NOTRETURN (block) = 0;
3950
3951 /* Now compute value numbers and populate value sets with all
3952 the expressions computed in BLOCK. */
3953 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3954 {
3955 ssa_op_iter iter;
3956 tree op;
3957
3958 stmt = gsi_stmt (gsi);
3959 gimple_set_uid (stmt, stmt_uid++);
3960
3961 /* Cache whether the basic-block has any non-visible side-effect
3962 or control flow.
3963 If this isn't a call or it is the last stmt in the
3964 basic-block then the CFG represents things correctly. */
3965 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3966 {
3967 /* Non-looping const functions always return normally.
3968 Otherwise the call might not return or have side-effects
3969 that forbids hoisting possibly trapping expressions
3970 before it. */
3971 int flags = gimple_call_flags (stmt);
3972 if (!(flags & ECF_CONST)
3973 || (flags & ECF_LOOPING_CONST_OR_PURE))
3974 BB_MAY_NOTRETURN (block) = 1;
3975 }
3976
3977 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3978 {
3979 pre_expr e = get_or_alloc_expr_for_name (op);
3980
3981 add_to_value (get_expr_value_id (e), e);
3982 if (!in_fre)
3983 bitmap_insert_into_set (TMP_GEN (block), e);
3984 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3985 }
3986
3987 if (gimple_has_side_effects (stmt) || stmt_could_throw_p (stmt))
3988 continue;
3989
3990 switch (gimple_code (stmt))
3991 {
3992 case GIMPLE_RETURN:
3993 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3994 add_to_exp_gen (block, op);
3995 continue;
3996
3997 case GIMPLE_CALL:
3998 {
3999 vn_reference_t ref;
4000 unsigned int i;
4001 vn_reference_op_t vro;
4002 pre_expr result = NULL;
4003 VEC(vn_reference_op_s, heap) *ops = NULL;
4004
4005 /* We can value number only calls to real functions. */
4006 if (gimple_call_internal_p (stmt))
4007 continue;
4008
4009 copy_reference_ops_from_call (stmt, &ops);
4010 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4011 gimple_expr_type (stmt),
4012 ops, &ref, VN_NOWALK);
4013 VEC_free (vn_reference_op_s, heap, ops);
4014 if (!ref)
4015 continue;
4016
4017 for (i = 0; VEC_iterate (vn_reference_op_s,
4018 ref->operands, i,
4019 vro); i++)
4020 {
4021 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4022 add_to_exp_gen (block, vro->op0);
4023 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4024 add_to_exp_gen (block, vro->op1);
4025 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4026 add_to_exp_gen (block, vro->op2);
4027 }
4028
4029 /* If the value of the call is not invalidated in
4030 this block until it is computed, add the expression
4031 to EXP_GEN. */
4032 if (!gimple_vuse (stmt)
4033 || gimple_code
4034 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
4035 || gimple_bb (SSA_NAME_DEF_STMT
4036 (gimple_vuse (stmt))) != block)
4037 {
4038 result = (pre_expr) pool_alloc (pre_expr_pool);
4039 result->kind = REFERENCE;
4040 result->id = 0;
4041 PRE_EXPR_REFERENCE (result) = ref;
4042
4043 get_or_alloc_expression_id (result);
4044 add_to_value (get_expr_value_id (result), result);
4045 if (!in_fre)
4046 bitmap_value_insert_into_set (EXP_GEN (block), result);
4047 }
4048 continue;
4049 }
4050
4051 case GIMPLE_ASSIGN:
4052 {
4053 pre_expr result = NULL;
4054 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4055 {
4056 case tcc_unary:
4057 case tcc_binary:
4058 case tcc_comparison:
4059 {
4060 vn_nary_op_t nary;
4061 unsigned int i;
4062
4063 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4064 gimple_assign_rhs_code (stmt),
4065 gimple_expr_type (stmt),
4066 gimple_assign_rhs1_ptr (stmt),
4067 &nary);
4068
4069 if (!nary)
4070 continue;
4071
4072 for (i = 0; i < nary->length; i++)
4073 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4074 add_to_exp_gen (block, nary->op[i]);
4075
4076 /* If the NARY traps and there was a preceeding
4077 point in the block that might not return avoid
4078 adding the nary to EXP_GEN. */
4079 if (BB_MAY_NOTRETURN (block)
4080 && vn_nary_may_trap (nary))
4081 continue;
4082
4083 result = (pre_expr) pool_alloc (pre_expr_pool);
4084 result->kind = NARY;
4085 result->id = 0;
4086 PRE_EXPR_NARY (result) = nary;
4087 break;
4088 }
4089
4090 case tcc_declaration:
4091 case tcc_reference:
4092 {
4093 vn_reference_t ref;
4094 unsigned int i;
4095 vn_reference_op_t vro;
4096
4097 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4098 gimple_vuse (stmt),
4099 VN_WALK, &ref);
4100 if (!ref)
4101 continue;
4102
4103 for (i = 0; VEC_iterate (vn_reference_op_s,
4104 ref->operands, i,
4105 vro); i++)
4106 {
4107 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4108 add_to_exp_gen (block, vro->op0);
4109 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4110 add_to_exp_gen (block, vro->op1);
4111 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4112 add_to_exp_gen (block, vro->op2);
4113 }
4114
4115 /* If the value of the reference is not invalidated in
4116 this block until it is computed, add the expression
4117 to EXP_GEN. */
4118 if (gimple_vuse (stmt))
4119 {
4120 gimple def_stmt;
4121 bool ok = true;
4122 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
4123 while (!gimple_nop_p (def_stmt)
4124 && gimple_code (def_stmt) != GIMPLE_PHI
4125 && gimple_bb (def_stmt) == block)
4126 {
4127 if (stmt_may_clobber_ref_p
4128 (def_stmt, gimple_assign_rhs1 (stmt)))
4129 {
4130 ok = false;
4131 break;
4132 }
4133 def_stmt
4134 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
4135 }
4136 if (!ok)
4137 continue;
4138 }
4139
4140 result = (pre_expr) pool_alloc (pre_expr_pool);
4141 result->kind = REFERENCE;
4142 result->id = 0;
4143 PRE_EXPR_REFERENCE (result) = ref;
4144 break;
4145 }
4146
4147 default:
4148 /* For any other statement that we don't
4149 recognize, simply add all referenced
4150 SSA_NAMEs to EXP_GEN. */
4151 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4152 add_to_exp_gen (block, op);
4153 continue;
4154 }
4155
4156 get_or_alloc_expression_id (result);
4157 add_to_value (get_expr_value_id (result), result);
4158 if (!in_fre)
4159 bitmap_value_insert_into_set (EXP_GEN (block), result);
4160
4161 continue;
4162 }
4163 default:
4164 break;
4165 }
4166 }
4167
4168 /* Put the dominator children of BLOCK on the worklist of blocks
4169 to compute available sets for. */
4170 for (son = first_dom_son (CDI_DOMINATORS, block);
4171 son;
4172 son = next_dom_son (CDI_DOMINATORS, son))
4173 worklist[sp++] = son;
4174 }
4175
4176 free (worklist);
4177 }
4178
4179 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4180 than the available expressions for it. The insertion point is
4181 right before the first use in STMT. Returns the SSA_NAME that should
4182 be used for replacement. */
4183
4184 static tree
4185 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4186 {
4187 basic_block bb = gimple_bb (stmt);
4188 gimple_stmt_iterator gsi;
4189 gimple_seq stmts = NULL;
4190 tree expr;
4191 pre_expr e;
4192
4193 /* First create a value expression from the expression we want
4194 to insert and associate it with the value handle for SSA_VN. */
4195 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4196 if (e == NULL)
4197 return NULL_TREE;
4198
4199 /* Then use create_expression_by_pieces to generate a valid
4200 expression to insert at this point of the IL stream. */
4201 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4202 if (expr == NULL_TREE)
4203 return NULL_TREE;
4204 gsi = gsi_for_stmt (stmt);
4205 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4206
4207 return expr;
4208 }
4209
4210 /* Eliminate fully redundant computations. */
4211
4212 static unsigned int
4213 eliminate (void)
4214 {
4215 VEC (gimple, heap) *to_remove = NULL;
4216 VEC (gimple, heap) *to_update = NULL;
4217 basic_block b;
4218 unsigned int todo = 0;
4219 gimple_stmt_iterator gsi;
4220 gimple stmt;
4221 unsigned i;
4222
4223 FOR_EACH_BB (b)
4224 {
4225 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4226 {
4227 tree lhs = NULL_TREE;
4228 tree rhs = NULL_TREE;
4229
4230 stmt = gsi_stmt (gsi);
4231
4232 if (gimple_has_lhs (stmt))
4233 lhs = gimple_get_lhs (stmt);
4234
4235 if (gimple_assign_single_p (stmt))
4236 rhs = gimple_assign_rhs1 (stmt);
4237
4238 /* Lookup the RHS of the expression, see if we have an
4239 available computation for it. If so, replace the RHS with
4240 the available computation.
4241
4242 See PR43491.
4243 We don't replace global register variable when it is a the RHS of
4244 a single assign. We do replace local register variable since gcc
4245 does not guarantee local variable will be allocated in register. */
4246 if (gimple_has_lhs (stmt)
4247 && TREE_CODE (lhs) == SSA_NAME
4248 && !gimple_assign_ssa_name_copy_p (stmt)
4249 && (!gimple_assign_single_p (stmt)
4250 || (!is_gimple_min_invariant (rhs)
4251 && (gimple_assign_rhs_code (stmt) != VAR_DECL
4252 || !is_global_var (rhs)
4253 || !DECL_HARD_REGISTER (rhs))))
4254 && !gimple_has_volatile_ops (stmt)
4255 && !has_zero_uses (lhs))
4256 {
4257 tree sprime = NULL;
4258 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4259 pre_expr sprimeexpr;
4260 gimple orig_stmt = stmt;
4261
4262 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4263 get_expr_value_id (lhsexpr),
4264 NULL);
4265
4266 if (sprimeexpr)
4267 {
4268 if (sprimeexpr->kind == CONSTANT)
4269 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4270 else if (sprimeexpr->kind == NAME)
4271 sprime = PRE_EXPR_NAME (sprimeexpr);
4272 else
4273 gcc_unreachable ();
4274 }
4275
4276 /* If there is no existing leader but SCCVN knows this
4277 value is constant, use that constant. */
4278 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4279 {
4280 sprime = VN_INFO (lhs)->valnum;
4281 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4282 TREE_TYPE (sprime)))
4283 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4284
4285 if (dump_file && (dump_flags & TDF_DETAILS))
4286 {
4287 fprintf (dump_file, "Replaced ");
4288 print_gimple_expr (dump_file, stmt, 0, 0);
4289 fprintf (dump_file, " with ");
4290 print_generic_expr (dump_file, sprime, 0);
4291 fprintf (dump_file, " in ");
4292 print_gimple_stmt (dump_file, stmt, 0, 0);
4293 }
4294 pre_stats.eliminations++;
4295 propagate_tree_value_into_stmt (&gsi, sprime);
4296 stmt = gsi_stmt (gsi);
4297 update_stmt (stmt);
4298
4299 /* If we removed EH side-effects from the statement, clean
4300 its EH information. */
4301 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4302 {
4303 bitmap_set_bit (need_eh_cleanup,
4304 gimple_bb (stmt)->index);
4305 if (dump_file && (dump_flags & TDF_DETAILS))
4306 fprintf (dump_file, " Removed EH side-effects.\n");
4307 }
4308 continue;
4309 }
4310
4311 /* If there is no existing usable leader but SCCVN thinks
4312 it has an expression it wants to use as replacement,
4313 insert that. */
4314 if (!sprime || sprime == lhs)
4315 {
4316 tree val = VN_INFO (lhs)->valnum;
4317 if (val != VN_TOP
4318 && TREE_CODE (val) == SSA_NAME
4319 && VN_INFO (val)->needs_insertion
4320 && can_PRE_operation (vn_get_expr_for (val)))
4321 sprime = do_SCCVN_insertion (stmt, val);
4322 }
4323 if (sprime
4324 && sprime != lhs
4325 && (rhs == NULL_TREE
4326 || TREE_CODE (rhs) != SSA_NAME
4327 || may_propagate_copy (rhs, sprime)))
4328 {
4329 bool can_make_abnormal_goto
4330 = is_gimple_call (stmt)
4331 && stmt_can_make_abnormal_goto (stmt);
4332
4333 gcc_assert (sprime != rhs);
4334
4335 if (dump_file && (dump_flags & TDF_DETAILS))
4336 {
4337 fprintf (dump_file, "Replaced ");
4338 print_gimple_expr (dump_file, stmt, 0, 0);
4339 fprintf (dump_file, " with ");
4340 print_generic_expr (dump_file, sprime, 0);
4341 fprintf (dump_file, " in ");
4342 print_gimple_stmt (dump_file, stmt, 0, 0);
4343 }
4344
4345 if (TREE_CODE (sprime) == SSA_NAME)
4346 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4347 NECESSARY, true);
4348 /* We need to make sure the new and old types actually match,
4349 which may require adding a simple cast, which fold_convert
4350 will do for us. */
4351 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4352 && !useless_type_conversion_p (gimple_expr_type (stmt),
4353 TREE_TYPE (sprime)))
4354 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4355
4356 pre_stats.eliminations++;
4357 propagate_tree_value_into_stmt (&gsi, sprime);
4358 stmt = gsi_stmt (gsi);
4359 update_stmt (stmt);
4360
4361 /* If we removed EH side-effects from the statement, clean
4362 its EH information. */
4363 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4364 {
4365 bitmap_set_bit (need_eh_cleanup,
4366 gimple_bb (stmt)->index);
4367 if (dump_file && (dump_flags & TDF_DETAILS))
4368 fprintf (dump_file, " Removed EH side-effects.\n");
4369 }
4370
4371 /* Likewise for AB side-effects. */
4372 if (can_make_abnormal_goto
4373 && !stmt_can_make_abnormal_goto (stmt))
4374 {
4375 bitmap_set_bit (need_ab_cleanup,
4376 gimple_bb (stmt)->index);
4377 if (dump_file && (dump_flags & TDF_DETAILS))
4378 fprintf (dump_file, " Removed AB side-effects.\n");
4379 }
4380 }
4381 }
4382 /* If the statement is a scalar store, see if the expression
4383 has the same value number as its rhs. If so, the store is
4384 dead. */
4385 else if (gimple_assign_single_p (stmt)
4386 && !gimple_has_volatile_ops (stmt)
4387 && !is_gimple_reg (gimple_assign_lhs (stmt))
4388 && (TREE_CODE (rhs) == SSA_NAME
4389 || is_gimple_min_invariant (rhs)))
4390 {
4391 tree val;
4392 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4393 gimple_vuse (stmt), VN_WALK, NULL);
4394 if (TREE_CODE (rhs) == SSA_NAME)
4395 rhs = VN_INFO (rhs)->valnum;
4396 if (val
4397 && operand_equal_p (val, rhs, 0))
4398 {
4399 if (dump_file && (dump_flags & TDF_DETAILS))
4400 {
4401 fprintf (dump_file, "Deleted redundant store ");
4402 print_gimple_stmt (dump_file, stmt, 0, 0);
4403 }
4404
4405 /* Queue stmt for removal. */
4406 VEC_safe_push (gimple, heap, to_remove, stmt);
4407 }
4408 }
4409 /* Visit COND_EXPRs and fold the comparison with the
4410 available value-numbers. */
4411 else if (gimple_code (stmt) == GIMPLE_COND)
4412 {
4413 tree op0 = gimple_cond_lhs (stmt);
4414 tree op1 = gimple_cond_rhs (stmt);
4415 tree result;
4416
4417 if (TREE_CODE (op0) == SSA_NAME)
4418 op0 = VN_INFO (op0)->valnum;
4419 if (TREE_CODE (op1) == SSA_NAME)
4420 op1 = VN_INFO (op1)->valnum;
4421 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4422 op0, op1);
4423 if (result && TREE_CODE (result) == INTEGER_CST)
4424 {
4425 if (integer_zerop (result))
4426 gimple_cond_make_false (stmt);
4427 else
4428 gimple_cond_make_true (stmt);
4429 update_stmt (stmt);
4430 todo = TODO_cleanup_cfg;
4431 }
4432 }
4433 /* Visit indirect calls and turn them into direct calls if
4434 possible. */
4435 if (is_gimple_call (stmt))
4436 {
4437 tree orig_fn = gimple_call_fn (stmt);
4438 tree fn;
4439 if (!orig_fn)
4440 continue;
4441 if (TREE_CODE (orig_fn) == SSA_NAME)
4442 fn = VN_INFO (orig_fn)->valnum;
4443 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4444 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4445 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4446 else
4447 continue;
4448 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4449 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4450 TREE_TYPE (fn)))
4451 {
4452 bool can_make_abnormal_goto
4453 = stmt_can_make_abnormal_goto (stmt);
4454 bool was_noreturn = gimple_call_noreturn_p (stmt);
4455
4456 if (dump_file && (dump_flags & TDF_DETAILS))
4457 {
4458 fprintf (dump_file, "Replacing call target with ");
4459 print_generic_expr (dump_file, fn, 0);
4460 fprintf (dump_file, " in ");
4461 print_gimple_stmt (dump_file, stmt, 0, 0);
4462 }
4463
4464 gimple_call_set_fn (stmt, fn);
4465 VEC_safe_push (gimple, heap, to_update, stmt);
4466
4467 /* When changing a call into a noreturn call, cfg cleanup
4468 is needed to fix up the noreturn call. */
4469 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4470 todo |= TODO_cleanup_cfg;
4471
4472 /* If we removed EH side-effects from the statement, clean
4473 its EH information. */
4474 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4475 {
4476 bitmap_set_bit (need_eh_cleanup,
4477 gimple_bb (stmt)->index);
4478 if (dump_file && (dump_flags & TDF_DETAILS))
4479 fprintf (dump_file, " Removed EH side-effects.\n");
4480 }
4481
4482 /* Likewise for AB side-effects. */
4483 if (can_make_abnormal_goto
4484 && !stmt_can_make_abnormal_goto (stmt))
4485 {
4486 bitmap_set_bit (need_ab_cleanup,
4487 gimple_bb (stmt)->index);
4488 if (dump_file && (dump_flags & TDF_DETAILS))
4489 fprintf (dump_file, " Removed AB side-effects.\n");
4490 }
4491
4492 /* Changing an indirect call to a direct call may
4493 have exposed different semantics. This may
4494 require an SSA update. */
4495 todo |= TODO_update_ssa_only_virtuals;
4496 }
4497 }
4498 }
4499
4500 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4501 {
4502 gimple stmt, phi = gsi_stmt (gsi);
4503 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4504 pre_expr sprimeexpr, resexpr;
4505 gimple_stmt_iterator gsi2;
4506
4507 /* We want to perform redundant PHI elimination. Do so by
4508 replacing the PHI with a single copy if possible.
4509 Do not touch inserted, single-argument or virtual PHIs. */
4510 if (gimple_phi_num_args (phi) == 1
4511 || !is_gimple_reg (res))
4512 {
4513 gsi_next (&gsi);
4514 continue;
4515 }
4516
4517 resexpr = get_or_alloc_expr_for_name (res);
4518 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4519 get_expr_value_id (resexpr), NULL);
4520 if (sprimeexpr)
4521 {
4522 if (sprimeexpr->kind == CONSTANT)
4523 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4524 else if (sprimeexpr->kind == NAME)
4525 sprime = PRE_EXPR_NAME (sprimeexpr);
4526 else
4527 gcc_unreachable ();
4528 }
4529 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4530 {
4531 sprime = VN_INFO (res)->valnum;
4532 if (!useless_type_conversion_p (TREE_TYPE (res),
4533 TREE_TYPE (sprime)))
4534 sprime = fold_convert (TREE_TYPE (res), sprime);
4535 }
4536 if (!sprime
4537 || sprime == res)
4538 {
4539 gsi_next (&gsi);
4540 continue;
4541 }
4542
4543 if (dump_file && (dump_flags & TDF_DETAILS))
4544 {
4545 fprintf (dump_file, "Replaced redundant PHI node defining ");
4546 print_generic_expr (dump_file, res, 0);
4547 fprintf (dump_file, " with ");
4548 print_generic_expr (dump_file, sprime, 0);
4549 fprintf (dump_file, "\n");
4550 }
4551
4552 remove_phi_node (&gsi, false);
4553
4554 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4555 && TREE_CODE (sprime) == SSA_NAME)
4556 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4557
4558 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4559 sprime = fold_convert (TREE_TYPE (res), sprime);
4560 stmt = gimple_build_assign (res, sprime);
4561 SSA_NAME_DEF_STMT (res) = stmt;
4562 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4563
4564 gsi2 = gsi_after_labels (b);
4565 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4566 /* Queue the copy for eventual removal. */
4567 VEC_safe_push (gimple, heap, to_remove, stmt);
4568 /* If we inserted this PHI node ourself, it's not an elimination. */
4569 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4570 pre_stats.phis--;
4571 else
4572 pre_stats.eliminations++;
4573 }
4574 }
4575
4576 /* We cannot remove stmts during BB walk, especially not release SSA
4577 names there as this confuses the VN machinery. The stmts ending
4578 up in to_remove are either stores or simple copies. */
4579 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4580 {
4581 tree lhs = gimple_assign_lhs (stmt);
4582 tree rhs = gimple_assign_rhs1 (stmt);
4583 use_operand_p use_p;
4584 gimple use_stmt;
4585
4586 /* If there is a single use only, propagate the equivalency
4587 instead of keeping the copy. */
4588 if (TREE_CODE (lhs) == SSA_NAME
4589 && TREE_CODE (rhs) == SSA_NAME
4590 && single_imm_use (lhs, &use_p, &use_stmt)
4591 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4592 {
4593 SET_USE (use_p, rhs);
4594 update_stmt (use_stmt);
4595 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4596 && TREE_CODE (rhs) == SSA_NAME)
4597 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4598 }
4599
4600 /* If this is a store or a now unused copy, remove it. */
4601 if (TREE_CODE (lhs) != SSA_NAME
4602 || has_zero_uses (lhs))
4603 {
4604 basic_block bb = gimple_bb (stmt);
4605 gsi = gsi_for_stmt (stmt);
4606 unlink_stmt_vdef (stmt);
4607 if (gsi_remove (&gsi, true))
4608 bitmap_set_bit (need_eh_cleanup, bb->index);
4609 if (TREE_CODE (lhs) == SSA_NAME)
4610 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4611 release_defs (stmt);
4612 }
4613 }
4614 VEC_free (gimple, heap, to_remove);
4615
4616 /* We cannot update call statements with virtual operands during
4617 SSA walk. This might remove them which in turn makes our
4618 VN lattice invalid. */
4619 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4620 update_stmt (stmt);
4621 VEC_free (gimple, heap, to_update);
4622
4623 return todo;
4624 }
4625
4626 /* Borrow a bit of tree-ssa-dce.c for the moment.
4627 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4628 this may be a bit faster, and we may want critical edges kept split. */
4629
4630 /* If OP's defining statement has not already been determined to be necessary,
4631 mark that statement necessary. Return the stmt, if it is newly
4632 necessary. */
4633
4634 static inline gimple
4635 mark_operand_necessary (tree op)
4636 {
4637 gimple stmt;
4638
4639 gcc_assert (op);
4640
4641 if (TREE_CODE (op) != SSA_NAME)
4642 return NULL;
4643
4644 stmt = SSA_NAME_DEF_STMT (op);
4645 gcc_assert (stmt);
4646
4647 if (gimple_plf (stmt, NECESSARY)
4648 || gimple_nop_p (stmt))
4649 return NULL;
4650
4651 gimple_set_plf (stmt, NECESSARY, true);
4652 return stmt;
4653 }
4654
4655 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4656 to insert PHI nodes sometimes, and because value numbering of casts isn't
4657 perfect, we sometimes end up inserting dead code. This simple DCE-like
4658 pass removes any insertions we made that weren't actually used. */
4659
4660 static void
4661 remove_dead_inserted_code (void)
4662 {
4663 bitmap worklist;
4664 unsigned i;
4665 bitmap_iterator bi;
4666 gimple t;
4667
4668 worklist = BITMAP_ALLOC (NULL);
4669 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4670 {
4671 t = SSA_NAME_DEF_STMT (ssa_name (i));
4672 if (gimple_plf (t, NECESSARY))
4673 bitmap_set_bit (worklist, i);
4674 }
4675 while (!bitmap_empty_p (worklist))
4676 {
4677 i = bitmap_first_set_bit (worklist);
4678 bitmap_clear_bit (worklist, i);
4679 t = SSA_NAME_DEF_STMT (ssa_name (i));
4680
4681 /* PHI nodes are somewhat special in that each PHI alternative has
4682 data and control dependencies. All the statements feeding the
4683 PHI node's arguments are always necessary. */
4684 if (gimple_code (t) == GIMPLE_PHI)
4685 {
4686 unsigned k;
4687
4688 for (k = 0; k < gimple_phi_num_args (t); k++)
4689 {
4690 tree arg = PHI_ARG_DEF (t, k);
4691 if (TREE_CODE (arg) == SSA_NAME)
4692 {
4693 gimple n = mark_operand_necessary (arg);
4694 if (n)
4695 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4696 }
4697 }
4698 }
4699 else
4700 {
4701 /* Propagate through the operands. Examine all the USE, VUSE and
4702 VDEF operands in this statement. Mark all the statements
4703 which feed this statement's uses as necessary. */
4704 ssa_op_iter iter;
4705 tree use;
4706
4707 /* The operands of VDEF expressions are also needed as they
4708 represent potential definitions that may reach this
4709 statement (VDEF operands allow us to follow def-def
4710 links). */
4711
4712 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4713 {
4714 gimple n = mark_operand_necessary (use);
4715 if (n)
4716 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4717 }
4718 }
4719 }
4720
4721 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4722 {
4723 t = SSA_NAME_DEF_STMT (ssa_name (i));
4724 if (!gimple_plf (t, NECESSARY))
4725 {
4726 gimple_stmt_iterator gsi;
4727
4728 if (dump_file && (dump_flags & TDF_DETAILS))
4729 {
4730 fprintf (dump_file, "Removing unnecessary insertion:");
4731 print_gimple_stmt (dump_file, t, 0, 0);
4732 }
4733
4734 gsi = gsi_for_stmt (t);
4735 if (gimple_code (t) == GIMPLE_PHI)
4736 remove_phi_node (&gsi, true);
4737 else
4738 {
4739 gsi_remove (&gsi, true);
4740 release_defs (t);
4741 }
4742 }
4743 }
4744 BITMAP_FREE (worklist);
4745 }
4746
4747 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4748 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4749 the number of visited blocks. */
4750
4751 static int
4752 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4753 {
4754 edge_iterator *stack;
4755 int sp;
4756 int post_order_num = 0;
4757 sbitmap visited;
4758
4759 if (include_entry_exit)
4760 post_order[post_order_num++] = EXIT_BLOCK;
4761
4762 /* Allocate stack for back-tracking up CFG. */
4763 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4764 sp = 0;
4765
4766 /* Allocate bitmap to track nodes that have been visited. */
4767 visited = sbitmap_alloc (last_basic_block);
4768
4769 /* None of the nodes in the CFG have been visited yet. */
4770 sbitmap_zero (visited);
4771
4772 /* Push the last edge on to the stack. */
4773 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4774
4775 while (sp)
4776 {
4777 edge_iterator ei;
4778 basic_block src;
4779 basic_block dest;
4780
4781 /* Look at the edge on the top of the stack. */
4782 ei = stack[sp - 1];
4783 src = ei_edge (ei)->src;
4784 dest = ei_edge (ei)->dest;
4785
4786 /* Check if the edge destination has been visited yet. */
4787 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4788 {
4789 /* Mark that we have visited the destination. */
4790 SET_BIT (visited, src->index);
4791
4792 if (EDGE_COUNT (src->preds) > 0)
4793 /* Since the DEST node has been visited for the first
4794 time, check its successors. */
4795 stack[sp++] = ei_start (src->preds);
4796 else
4797 post_order[post_order_num++] = src->index;
4798 }
4799 else
4800 {
4801 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4802 post_order[post_order_num++] = dest->index;
4803
4804 if (!ei_one_before_end_p (ei))
4805 ei_next (&stack[sp - 1]);
4806 else
4807 sp--;
4808 }
4809 }
4810
4811 if (include_entry_exit)
4812 post_order[post_order_num++] = ENTRY_BLOCK;
4813
4814 free (stack);
4815 sbitmap_free (visited);
4816 return post_order_num;
4817 }
4818
4819
4820 /* Initialize data structures used by PRE. */
4821
4822 static void
4823 init_pre (bool do_fre)
4824 {
4825 basic_block bb;
4826
4827 next_expression_id = 1;
4828 expressions = NULL;
4829 VEC_safe_push (pre_expr, heap, expressions, NULL);
4830 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4831 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4832 get_max_value_id() + 1);
4833 name_to_id = NULL;
4834
4835 in_fre = do_fre;
4836
4837 inserted_exprs = BITMAP_ALLOC (NULL);
4838 need_creation = NULL;
4839 pretemp = NULL_TREE;
4840 storetemp = NULL_TREE;
4841 prephitemp = NULL_TREE;
4842
4843 connect_infinite_loops_to_exit ();
4844 memset (&pre_stats, 0, sizeof (pre_stats));
4845
4846
4847 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4848 my_rev_post_order_compute (postorder, false);
4849
4850 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4851
4852 calculate_dominance_info (CDI_POST_DOMINATORS);
4853 calculate_dominance_info (CDI_DOMINATORS);
4854
4855 bitmap_obstack_initialize (&grand_bitmap_obstack);
4856 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4857 expr_pred_trans_eq, free);
4858 expression_to_id = htab_create (num_ssa_names * 3,
4859 pre_expr_hash,
4860 pre_expr_eq, NULL);
4861 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4862 sizeof (struct bitmap_set), 30);
4863 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4864 sizeof (struct pre_expr_d), 30);
4865 FOR_ALL_BB (bb)
4866 {
4867 EXP_GEN (bb) = bitmap_set_new ();
4868 PHI_GEN (bb) = bitmap_set_new ();
4869 TMP_GEN (bb) = bitmap_set_new ();
4870 AVAIL_OUT (bb) = bitmap_set_new ();
4871 }
4872
4873 need_eh_cleanup = BITMAP_ALLOC (NULL);
4874 need_ab_cleanup = BITMAP_ALLOC (NULL);
4875 }
4876
4877
4878 /* Deallocate data structures used by PRE. */
4879
4880 static void
4881 fini_pre (bool do_fre)
4882 {
4883 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4884 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4885
4886 free (postorder);
4887 VEC_free (bitmap_set_t, heap, value_expressions);
4888 BITMAP_FREE (inserted_exprs);
4889 VEC_free (gimple, heap, need_creation);
4890 bitmap_obstack_release (&grand_bitmap_obstack);
4891 free_alloc_pool (bitmap_set_pool);
4892 free_alloc_pool (pre_expr_pool);
4893 htab_delete (phi_translate_table);
4894 htab_delete (expression_to_id);
4895 VEC_free (unsigned, heap, name_to_id);
4896
4897 free_aux_for_blocks ();
4898
4899 free_dominance_info (CDI_POST_DOMINATORS);
4900
4901 if (do_eh_cleanup)
4902 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4903
4904 if (do_ab_cleanup)
4905 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4906
4907 BITMAP_FREE (need_eh_cleanup);
4908 BITMAP_FREE (need_ab_cleanup);
4909
4910 if (do_eh_cleanup || do_ab_cleanup)
4911 cleanup_tree_cfg ();
4912
4913 if (!do_fre)
4914 loop_optimizer_finalize ();
4915 }
4916
4917 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4918 only wants to do full redundancy elimination. */
4919
4920 static unsigned int
4921 execute_pre (bool do_fre)
4922 {
4923 unsigned int todo = 0;
4924
4925 do_partial_partial =
4926 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4927
4928 /* This has to happen before SCCVN runs because
4929 loop_optimizer_init may create new phis, etc. */
4930 if (!do_fre)
4931 loop_optimizer_init (LOOPS_NORMAL);
4932
4933 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4934 {
4935 if (!do_fre)
4936 loop_optimizer_finalize ();
4937
4938 return 0;
4939 }
4940
4941 init_pre (do_fre);
4942 scev_initialize ();
4943
4944 /* Collect and value number expressions computed in each basic block. */
4945 compute_avail ();
4946
4947 if (dump_file && (dump_flags & TDF_DETAILS))
4948 {
4949 basic_block bb;
4950
4951 FOR_ALL_BB (bb)
4952 {
4953 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4954 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4955 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4956 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4957 }
4958 }
4959
4960 /* Insert can get quite slow on an incredibly large number of basic
4961 blocks due to some quadratic behavior. Until this behavior is
4962 fixed, don't run it when he have an incredibly large number of
4963 bb's. If we aren't going to run insert, there is no point in
4964 computing ANTIC, either, even though it's plenty fast. */
4965 if (!do_fre && n_basic_blocks < 4000)
4966 {
4967 compute_antic ();
4968 insert ();
4969 }
4970
4971 /* Make sure to remove fake edges before committing our inserts.
4972 This makes sure we don't end up with extra critical edges that
4973 we would need to split. */
4974 remove_fake_exit_edges ();
4975 gsi_commit_edge_inserts ();
4976
4977 /* Remove all the redundant expressions. */
4978 todo |= eliminate ();
4979
4980 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4981 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4982 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4983 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4984 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4985
4986 clear_expression_ids ();
4987 if (!do_fre)
4988 {
4989 remove_dead_inserted_code ();
4990 todo |= TODO_verify_flow;
4991 }
4992
4993 scev_finalize ();
4994 fini_pre (do_fre);
4995
4996 if (!do_fre)
4997 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4998 case we can merge the block with the remaining predecessor of the block.
4999 It should either:
5000 - call merge_blocks after each tail merge iteration
5001 - call merge_blocks after all tail merge iterations
5002 - mark TODO_cleanup_cfg when necessary
5003 - share the cfg cleanup with fini_pre. */
5004 todo |= tail_merge_optimize (todo);
5005 free_scc_vn ();
5006
5007 return todo;
5008 }
5009
5010 /* Gate and execute functions for PRE. */
5011
5012 static unsigned int
5013 do_pre (void)
5014 {
5015 return execute_pre (false);
5016 }
5017
5018 static bool
5019 gate_pre (void)
5020 {
5021 return flag_tree_pre != 0;
5022 }
5023
5024 struct gimple_opt_pass pass_pre =
5025 {
5026 {
5027 GIMPLE_PASS,
5028 "pre", /* name */
5029 gate_pre, /* gate */
5030 do_pre, /* execute */
5031 NULL, /* sub */
5032 NULL, /* next */
5033 0, /* static_pass_number */
5034 TV_TREE_PRE, /* tv_id */
5035 PROP_no_crit_edges | PROP_cfg
5036 | PROP_ssa, /* properties_required */
5037 0, /* properties_provided */
5038 0, /* properties_destroyed */
5039 TODO_rebuild_alias, /* todo_flags_start */
5040 TODO_update_ssa_only_virtuals | TODO_ggc_collect
5041 | TODO_verify_ssa /* todo_flags_finish */
5042 }
5043 };
5044
5045
5046 /* Gate and execute functions for FRE. */
5047
5048 static unsigned int
5049 execute_fre (void)
5050 {
5051 return execute_pre (true);
5052 }
5053
5054 static bool
5055 gate_fre (void)
5056 {
5057 return flag_tree_fre != 0;
5058 }
5059
5060 struct gimple_opt_pass pass_fre =
5061 {
5062 {
5063 GIMPLE_PASS,
5064 "fre", /* name */
5065 gate_fre, /* gate */
5066 execute_fre, /* execute */
5067 NULL, /* sub */
5068 NULL, /* next */
5069 0, /* static_pass_number */
5070 TV_TREE_FRE, /* tv_id */
5071 PROP_cfg | PROP_ssa, /* properties_required */
5072 0, /* properties_provided */
5073 0, /* properties_destroyed */
5074 0, /* todo_flags_start */
5075 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
5076 }
5077 };