re PR tree-optimization/38785 (huge performance regression on EEMBC bitmnp01)
[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 /* Print out the expressions that have VAL to OUTFILE. */
1033
1034 static void
1035 print_value_expressions (FILE *outfile, unsigned int val)
1036 {
1037 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1038 if (set)
1039 {
1040 char s[10];
1041 sprintf (s, "%04d", val);
1042 print_bitmap_set (outfile, set, s, 0);
1043 }
1044 }
1045
1046
1047 DEBUG_FUNCTION void
1048 debug_value_expressions (unsigned int val)
1049 {
1050 print_value_expressions (stderr, val);
1051 }
1052
1053 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1054 represent it. */
1055
1056 static pre_expr
1057 get_or_alloc_expr_for_constant (tree constant)
1058 {
1059 unsigned int result_id;
1060 unsigned int value_id;
1061 struct pre_expr_d expr;
1062 pre_expr newexpr;
1063
1064 expr.kind = CONSTANT;
1065 PRE_EXPR_CONSTANT (&expr) = constant;
1066 result_id = lookup_expression_id (&expr);
1067 if (result_id != 0)
1068 return expression_for_id (result_id);
1069
1070 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1071 newexpr->kind = CONSTANT;
1072 PRE_EXPR_CONSTANT (newexpr) = constant;
1073 alloc_expression_id (newexpr);
1074 value_id = get_or_alloc_constant_value_id (constant);
1075 add_to_value (value_id, newexpr);
1076 return newexpr;
1077 }
1078
1079 /* Given a value id V, find the actual tree representing the constant
1080 value if there is one, and return it. Return NULL if we can't find
1081 a constant. */
1082
1083 static tree
1084 get_constant_for_value_id (unsigned int v)
1085 {
1086 if (value_id_constant_p (v))
1087 {
1088 unsigned int i;
1089 bitmap_iterator bi;
1090 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1091
1092 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1093 {
1094 pre_expr expr = expression_for_id (i);
1095 if (expr->kind == CONSTANT)
1096 return PRE_EXPR_CONSTANT (expr);
1097 }
1098 }
1099 return NULL;
1100 }
1101
1102 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1103 Currently only supports constants and SSA_NAMES. */
1104 static pre_expr
1105 get_or_alloc_expr_for (tree t)
1106 {
1107 if (TREE_CODE (t) == SSA_NAME)
1108 return get_or_alloc_expr_for_name (t);
1109 else if (is_gimple_min_invariant (t))
1110 return get_or_alloc_expr_for_constant (t);
1111 else
1112 {
1113 /* More complex expressions can result from SCCVN expression
1114 simplification that inserts values for them. As they all
1115 do not have VOPs the get handled by the nary ops struct. */
1116 vn_nary_op_t result;
1117 unsigned int result_id;
1118 vn_nary_op_lookup (t, &result);
1119 if (result != NULL)
1120 {
1121 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1122 e->kind = NARY;
1123 PRE_EXPR_NARY (e) = result;
1124 result_id = lookup_expression_id (e);
1125 if (result_id != 0)
1126 {
1127 pool_free (pre_expr_pool, e);
1128 e = expression_for_id (result_id);
1129 return e;
1130 }
1131 alloc_expression_id (e);
1132 return e;
1133 }
1134 }
1135 return NULL;
1136 }
1137
1138 /* Return the folded version of T if T, when folded, is a gimple
1139 min_invariant. Otherwise, return T. */
1140
1141 static pre_expr
1142 fully_constant_expression (pre_expr e)
1143 {
1144 switch (e->kind)
1145 {
1146 case CONSTANT:
1147 return e;
1148 case NARY:
1149 {
1150 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1151 switch (TREE_CODE_CLASS (nary->opcode))
1152 {
1153 case tcc_binary:
1154 case tcc_comparison:
1155 {
1156 /* We have to go from trees to pre exprs to value ids to
1157 constants. */
1158 tree naryop0 = nary->op[0];
1159 tree naryop1 = nary->op[1];
1160 tree result;
1161 if (!is_gimple_min_invariant (naryop0))
1162 {
1163 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1164 unsigned int vrep0 = get_expr_value_id (rep0);
1165 tree const0 = get_constant_for_value_id (vrep0);
1166 if (const0)
1167 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1168 }
1169 if (!is_gimple_min_invariant (naryop1))
1170 {
1171 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1172 unsigned int vrep1 = get_expr_value_id (rep1);
1173 tree const1 = get_constant_for_value_id (vrep1);
1174 if (const1)
1175 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1176 }
1177 result = fold_binary (nary->opcode, nary->type,
1178 naryop0, naryop1);
1179 if (result && is_gimple_min_invariant (result))
1180 return get_or_alloc_expr_for_constant (result);
1181 /* We might have simplified the expression to a
1182 SSA_NAME for example from x_1 * 1. But we cannot
1183 insert a PHI for x_1 unconditionally as x_1 might
1184 not be available readily. */
1185 return e;
1186 }
1187 case tcc_reference:
1188 if (nary->opcode != REALPART_EXPR
1189 && nary->opcode != IMAGPART_EXPR
1190 && nary->opcode != VIEW_CONVERT_EXPR)
1191 return e;
1192 /* Fallthrough. */
1193 case tcc_unary:
1194 {
1195 /* We have to go from trees to pre exprs to value ids to
1196 constants. */
1197 tree naryop0 = nary->op[0];
1198 tree const0, result;
1199 if (is_gimple_min_invariant (naryop0))
1200 const0 = naryop0;
1201 else
1202 {
1203 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1204 unsigned int vrep0 = get_expr_value_id (rep0);
1205 const0 = get_constant_for_value_id (vrep0);
1206 }
1207 result = NULL;
1208 if (const0)
1209 {
1210 tree type1 = TREE_TYPE (nary->op[0]);
1211 const0 = fold_convert (type1, const0);
1212 result = fold_unary (nary->opcode, nary->type, const0);
1213 }
1214 if (result && is_gimple_min_invariant (result))
1215 return get_or_alloc_expr_for_constant (result);
1216 return e;
1217 }
1218 default:
1219 return e;
1220 }
1221 }
1222 case REFERENCE:
1223 {
1224 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1225 tree folded;
1226 if ((folded = fully_constant_vn_reference_p (ref)))
1227 return get_or_alloc_expr_for_constant (folded);
1228 return e;
1229 }
1230 default:
1231 return e;
1232 }
1233 return e;
1234 }
1235
1236 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1237 it has the value it would have in BLOCK. Set *SAME_VALID to true
1238 in case the new vuse doesn't change the value id of the OPERANDS. */
1239
1240 static tree
1241 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1242 alias_set_type set, tree type, tree vuse,
1243 basic_block phiblock,
1244 basic_block block, bool *same_valid)
1245 {
1246 gimple phi = SSA_NAME_DEF_STMT (vuse);
1247 ao_ref ref;
1248 edge e = NULL;
1249 bool use_oracle;
1250
1251 *same_valid = true;
1252
1253 if (gimple_bb (phi) != phiblock)
1254 return vuse;
1255
1256 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1257
1258 /* Use the alias-oracle to find either the PHI node in this block,
1259 the first VUSE used in this block that is equivalent to vuse or
1260 the first VUSE which definition in this block kills the value. */
1261 if (gimple_code (phi) == GIMPLE_PHI)
1262 e = find_edge (block, phiblock);
1263 else if (use_oracle)
1264 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1265 {
1266 vuse = gimple_vuse (phi);
1267 phi = SSA_NAME_DEF_STMT (vuse);
1268 if (gimple_bb (phi) != phiblock)
1269 return vuse;
1270 if (gimple_code (phi) == GIMPLE_PHI)
1271 {
1272 e = find_edge (block, phiblock);
1273 break;
1274 }
1275 }
1276 else
1277 return NULL_TREE;
1278
1279 if (e)
1280 {
1281 if (use_oracle)
1282 {
1283 bitmap visited = NULL;
1284 /* Try to find a vuse that dominates this phi node by skipping
1285 non-clobbering statements. */
1286 vuse = get_continuation_for_phi (phi, &ref, &visited);
1287 if (visited)
1288 BITMAP_FREE (visited);
1289 }
1290 else
1291 vuse = NULL_TREE;
1292 if (!vuse)
1293 {
1294 /* If we didn't find any, the value ID can't stay the same,
1295 but return the translated vuse. */
1296 *same_valid = false;
1297 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1298 }
1299 /* ??? We would like to return vuse here as this is the canonical
1300 upmost vdef that this reference is associated with. But during
1301 insertion of the references into the hash tables we only ever
1302 directly insert with their direct gimple_vuse, hence returning
1303 something else would make us not find the other expression. */
1304 return PHI_ARG_DEF (phi, e->dest_idx);
1305 }
1306
1307 return NULL_TREE;
1308 }
1309
1310 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1311 SET2. This is used to avoid making a set consisting of the union
1312 of PA_IN and ANTIC_IN during insert. */
1313
1314 static inline pre_expr
1315 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1316 {
1317 pre_expr result;
1318
1319 result = bitmap_find_leader (set1, val, NULL);
1320 if (!result && set2)
1321 result = bitmap_find_leader (set2, val, NULL);
1322 return result;
1323 }
1324
1325 /* Get the tree type for our PRE expression e. */
1326
1327 static tree
1328 get_expr_type (const pre_expr e)
1329 {
1330 switch (e->kind)
1331 {
1332 case NAME:
1333 return TREE_TYPE (PRE_EXPR_NAME (e));
1334 case CONSTANT:
1335 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1336 case REFERENCE:
1337 return PRE_EXPR_REFERENCE (e)->type;
1338 case NARY:
1339 return PRE_EXPR_NARY (e)->type;
1340 }
1341 gcc_unreachable();
1342 }
1343
1344 /* Get a representative SSA_NAME for a given expression.
1345 Since all of our sub-expressions are treated as values, we require
1346 them to be SSA_NAME's for simplicity.
1347 Prior versions of GVNPRE used to use "value handles" here, so that
1348 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1349 either case, the operands are really values (IE we do not expect
1350 them to be usable without finding leaders). */
1351
1352 static tree
1353 get_representative_for (const pre_expr e)
1354 {
1355 tree exprtype;
1356 tree name;
1357 unsigned int value_id = get_expr_value_id (e);
1358
1359 switch (e->kind)
1360 {
1361 case NAME:
1362 return PRE_EXPR_NAME (e);
1363 case CONSTANT:
1364 return PRE_EXPR_CONSTANT (e);
1365 case NARY:
1366 case REFERENCE:
1367 {
1368 /* Go through all of the expressions representing this value
1369 and pick out an SSA_NAME. */
1370 unsigned int i;
1371 bitmap_iterator bi;
1372 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1373 value_id);
1374 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1375 {
1376 pre_expr rep = expression_for_id (i);
1377 if (rep->kind == NAME)
1378 return PRE_EXPR_NAME (rep);
1379 }
1380 }
1381 break;
1382 }
1383 /* If we reached here we couldn't find an SSA_NAME. This can
1384 happen when we've discovered a value that has never appeared in
1385 the program as set to an SSA_NAME, most likely as the result of
1386 phi translation. */
1387 if (dump_file)
1388 {
1389 fprintf (dump_file,
1390 "Could not find SSA_NAME representative for expression:");
1391 print_pre_expr (dump_file, e);
1392 fprintf (dump_file, "\n");
1393 }
1394
1395 exprtype = get_expr_type (e);
1396
1397 /* Build and insert the assignment of the end result to the temporary
1398 that we will return. */
1399 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1400 {
1401 pretemp = create_tmp_reg (exprtype, "pretmp");
1402 add_referenced_var (pretemp);
1403 }
1404
1405 name = make_ssa_name (pretemp, gimple_build_nop ());
1406 VN_INFO_GET (name)->value_id = value_id;
1407 if (e->kind == CONSTANT)
1408 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1409 else
1410 VN_INFO (name)->valnum = name;
1411
1412 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1413 if (dump_file)
1414 {
1415 fprintf (dump_file, "Created SSA_NAME representative ");
1416 print_generic_expr (dump_file, name, 0);
1417 fprintf (dump_file, " for expression:");
1418 print_pre_expr (dump_file, e);
1419 fprintf (dump_file, "\n");
1420 }
1421
1422 return name;
1423 }
1424
1425
1426
1427 static pre_expr
1428 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1429 basic_block pred, basic_block phiblock);
1430
1431 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1432 the phis in PRED. Return NULL if we can't find a leader for each part
1433 of the translated expression. */
1434
1435 static pre_expr
1436 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1437 basic_block pred, basic_block phiblock)
1438 {
1439 switch (expr->kind)
1440 {
1441 case NARY:
1442 {
1443 unsigned int i;
1444 bool changed = false;
1445 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1446 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1447 sizeof_vn_nary_op (nary->length));
1448 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1449
1450 for (i = 0; i < newnary->length; i++)
1451 {
1452 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1453 continue;
1454 else
1455 {
1456 pre_expr leader, result;
1457 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1458 leader = find_leader_in_sets (op_val_id, set1, set2);
1459 result = phi_translate (leader, set1, set2, pred, phiblock);
1460 if (result && result != leader)
1461 {
1462 tree name = get_representative_for (result);
1463 if (!name)
1464 return NULL;
1465 newnary->op[i] = name;
1466 }
1467 else if (!result)
1468 return NULL;
1469
1470 changed |= newnary->op[i] != nary->op[i];
1471 }
1472 }
1473 if (changed)
1474 {
1475 pre_expr constant;
1476 unsigned int new_val_id;
1477
1478 tree result = vn_nary_op_lookup_pieces (newnary->length,
1479 newnary->opcode,
1480 newnary->type,
1481 &newnary->op[0],
1482 &nary);
1483 if (result && is_gimple_min_invariant (result))
1484 return get_or_alloc_expr_for_constant (result);
1485
1486 expr = (pre_expr) pool_alloc (pre_expr_pool);
1487 expr->kind = NARY;
1488 expr->id = 0;
1489 if (nary)
1490 {
1491 PRE_EXPR_NARY (expr) = nary;
1492 constant = fully_constant_expression (expr);
1493 if (constant != expr)
1494 return constant;
1495
1496 new_val_id = nary->value_id;
1497 get_or_alloc_expression_id (expr);
1498 }
1499 else
1500 {
1501 new_val_id = get_next_value_id ();
1502 VEC_safe_grow_cleared (bitmap_set_t, heap,
1503 value_expressions,
1504 get_max_value_id() + 1);
1505 nary = vn_nary_op_insert_pieces (newnary->length,
1506 newnary->opcode,
1507 newnary->type,
1508 &newnary->op[0],
1509 result, new_val_id);
1510 PRE_EXPR_NARY (expr) = nary;
1511 constant = fully_constant_expression (expr);
1512 if (constant != expr)
1513 return constant;
1514 get_or_alloc_expression_id (expr);
1515 }
1516 add_to_value (new_val_id, expr);
1517 }
1518 return expr;
1519 }
1520 break;
1521
1522 case REFERENCE:
1523 {
1524 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1525 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1526 tree vuse = ref->vuse;
1527 tree newvuse = vuse;
1528 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1529 bool changed = false, same_valid = true;
1530 unsigned int i, j, n;
1531 vn_reference_op_t operand;
1532 vn_reference_t newref;
1533
1534 for (i = 0, j = 0;
1535 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1536 {
1537 pre_expr opresult;
1538 pre_expr leader;
1539 tree op[3];
1540 tree type = operand->type;
1541 vn_reference_op_s newop = *operand;
1542 op[0] = operand->op0;
1543 op[1] = operand->op1;
1544 op[2] = operand->op2;
1545 for (n = 0; n < 3; ++n)
1546 {
1547 unsigned int op_val_id;
1548 if (!op[n])
1549 continue;
1550 if (TREE_CODE (op[n]) != SSA_NAME)
1551 {
1552 /* We can't possibly insert these. */
1553 if (n != 0
1554 && !is_gimple_min_invariant (op[n]))
1555 break;
1556 continue;
1557 }
1558 op_val_id = VN_INFO (op[n])->value_id;
1559 leader = find_leader_in_sets (op_val_id, set1, set2);
1560 if (!leader)
1561 break;
1562 /* Make sure we do not recursively translate ourselves
1563 like for translating a[n_1] with the leader for
1564 n_1 being a[n_1]. */
1565 if (get_expression_id (leader) != get_expression_id (expr))
1566 {
1567 opresult = phi_translate (leader, set1, set2,
1568 pred, phiblock);
1569 if (!opresult)
1570 break;
1571 if (opresult != leader)
1572 {
1573 tree name = get_representative_for (opresult);
1574 if (!name)
1575 break;
1576 changed |= name != op[n];
1577 op[n] = name;
1578 }
1579 }
1580 }
1581 if (n != 3)
1582 {
1583 if (newoperands)
1584 VEC_free (vn_reference_op_s, heap, newoperands);
1585 return NULL;
1586 }
1587 if (!newoperands)
1588 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1589 /* We may have changed from an SSA_NAME to a constant */
1590 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1591 newop.opcode = TREE_CODE (op[0]);
1592 newop.type = type;
1593 newop.op0 = op[0];
1594 newop.op1 = op[1];
1595 newop.op2 = op[2];
1596 /* If it transforms a non-constant ARRAY_REF into a constant
1597 one, adjust the constant offset. */
1598 if (newop.opcode == ARRAY_REF
1599 && newop.off == -1
1600 && TREE_CODE (op[0]) == INTEGER_CST
1601 && TREE_CODE (op[1]) == INTEGER_CST
1602 && TREE_CODE (op[2]) == INTEGER_CST)
1603 {
1604 double_int off = tree_to_double_int (op[0]);
1605 off = double_int_add (off,
1606 double_int_neg
1607 (tree_to_double_int (op[1])));
1608 off = double_int_mul (off, tree_to_double_int (op[2]));
1609 if (double_int_fits_in_shwi_p (off))
1610 newop.off = off.low;
1611 }
1612 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1613 /* If it transforms from an SSA_NAME to an address, fold with
1614 a preceding indirect reference. */
1615 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1616 && VEC_index (vn_reference_op_s,
1617 newoperands, j - 1)->opcode == MEM_REF)
1618 vn_reference_fold_indirect (&newoperands, &j);
1619 }
1620 if (i != VEC_length (vn_reference_op_s, operands))
1621 {
1622 if (newoperands)
1623 VEC_free (vn_reference_op_s, heap, newoperands);
1624 return NULL;
1625 }
1626
1627 if (vuse)
1628 {
1629 newvuse = translate_vuse_through_block (newoperands,
1630 ref->set, ref->type,
1631 vuse, phiblock, pred,
1632 &same_valid);
1633 if (newvuse == NULL_TREE)
1634 {
1635 VEC_free (vn_reference_op_s, heap, newoperands);
1636 return NULL;
1637 }
1638 }
1639
1640 if (changed || newvuse != vuse)
1641 {
1642 unsigned int new_val_id;
1643 pre_expr constant;
1644 bool converted = false;
1645
1646 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1647 ref->type,
1648 newoperands,
1649 &newref, VN_WALK);
1650 if (result)
1651 VEC_free (vn_reference_op_s, heap, newoperands);
1652
1653 if (result
1654 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1655 {
1656 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1657 converted = true;
1658 }
1659 else if (!result && newref
1660 && !useless_type_conversion_p (ref->type, newref->type))
1661 {
1662 VEC_free (vn_reference_op_s, heap, newoperands);
1663 return NULL;
1664 }
1665
1666 if (result && is_gimple_min_invariant (result))
1667 {
1668 gcc_assert (!newoperands);
1669 return get_or_alloc_expr_for_constant (result);
1670 }
1671
1672 expr = (pre_expr) pool_alloc (pre_expr_pool);
1673 expr->kind = REFERENCE;
1674 expr->id = 0;
1675
1676 if (converted)
1677 {
1678 vn_nary_op_t nary;
1679 tree nresult;
1680
1681 gcc_assert (CONVERT_EXPR_P (result)
1682 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1683
1684 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1685 TREE_TYPE (result),
1686 &TREE_OPERAND (result, 0),
1687 &nary);
1688 if (nresult && is_gimple_min_invariant (nresult))
1689 return get_or_alloc_expr_for_constant (nresult);
1690
1691 expr->kind = NARY;
1692 if (nary)
1693 {
1694 PRE_EXPR_NARY (expr) = nary;
1695 constant = fully_constant_expression (expr);
1696 if (constant != expr)
1697 return constant;
1698
1699 new_val_id = nary->value_id;
1700 get_or_alloc_expression_id (expr);
1701 }
1702 else
1703 {
1704 new_val_id = get_next_value_id ();
1705 VEC_safe_grow_cleared (bitmap_set_t, heap,
1706 value_expressions,
1707 get_max_value_id() + 1);
1708 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1709 TREE_TYPE (result),
1710 &TREE_OPERAND (result, 0),
1711 NULL_TREE,
1712 new_val_id);
1713 PRE_EXPR_NARY (expr) = nary;
1714 constant = fully_constant_expression (expr);
1715 if (constant != expr)
1716 return constant;
1717 get_or_alloc_expression_id (expr);
1718 }
1719 }
1720 else if (newref)
1721 {
1722 PRE_EXPR_REFERENCE (expr) = newref;
1723 constant = fully_constant_expression (expr);
1724 if (constant != expr)
1725 return constant;
1726
1727 new_val_id = newref->value_id;
1728 get_or_alloc_expression_id (expr);
1729 }
1730 else
1731 {
1732 if (changed || !same_valid)
1733 {
1734 new_val_id = get_next_value_id ();
1735 VEC_safe_grow_cleared (bitmap_set_t, heap,
1736 value_expressions,
1737 get_max_value_id() + 1);
1738 }
1739 else
1740 new_val_id = ref->value_id;
1741 newref = vn_reference_insert_pieces (newvuse, ref->set,
1742 ref->type,
1743 newoperands,
1744 result, new_val_id);
1745 newoperands = NULL;
1746 PRE_EXPR_REFERENCE (expr) = newref;
1747 constant = fully_constant_expression (expr);
1748 if (constant != expr)
1749 return constant;
1750 get_or_alloc_expression_id (expr);
1751 }
1752 add_to_value (new_val_id, expr);
1753 }
1754 VEC_free (vn_reference_op_s, heap, newoperands);
1755 return expr;
1756 }
1757 break;
1758
1759 case NAME:
1760 {
1761 gimple phi = NULL;
1762 edge e;
1763 gimple def_stmt;
1764 tree name = PRE_EXPR_NAME (expr);
1765
1766 def_stmt = SSA_NAME_DEF_STMT (name);
1767 if (gimple_code (def_stmt) == GIMPLE_PHI
1768 && gimple_bb (def_stmt) == phiblock)
1769 phi = def_stmt;
1770 else
1771 return expr;
1772
1773 e = find_edge (pred, gimple_bb (phi));
1774 if (e)
1775 {
1776 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1777 pre_expr newexpr;
1778
1779 if (TREE_CODE (def) == SSA_NAME)
1780 def = VN_INFO (def)->valnum;
1781
1782 /* Handle constant. */
1783 if (is_gimple_min_invariant (def))
1784 return get_or_alloc_expr_for_constant (def);
1785
1786 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1787 return NULL;
1788
1789 newexpr = get_or_alloc_expr_for_name (def);
1790 return newexpr;
1791 }
1792 }
1793 return expr;
1794
1795 default:
1796 gcc_unreachable ();
1797 }
1798 }
1799
1800 /* Wrapper around phi_translate_1 providing caching functionality. */
1801
1802 static pre_expr
1803 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1804 basic_block pred, basic_block phiblock)
1805 {
1806 pre_expr phitrans;
1807
1808 if (!expr)
1809 return NULL;
1810
1811 /* Constants contain no values that need translation. */
1812 if (expr->kind == CONSTANT)
1813 return expr;
1814
1815 if (value_id_constant_p (get_expr_value_id (expr)))
1816 return expr;
1817
1818 if (expr->kind != NAME)
1819 {
1820 phitrans = phi_trans_lookup (expr, pred);
1821 if (phitrans)
1822 return phitrans;
1823 }
1824
1825 /* Translate. */
1826 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1827
1828 /* Don't add empty translations to the cache. Neither add
1829 translations of NAMEs as those are cheap to translate. */
1830 if (phitrans
1831 && expr->kind != NAME)
1832 phi_trans_add (expr, phitrans, pred);
1833
1834 return phitrans;
1835 }
1836
1837
1838 /* For each expression in SET, translate the values through phi nodes
1839 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1840 expressions in DEST. */
1841
1842 static void
1843 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1844 basic_block phiblock)
1845 {
1846 VEC (pre_expr, heap) *exprs;
1847 pre_expr expr;
1848 int i;
1849
1850 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1851 {
1852 bitmap_set_copy (dest, set);
1853 return;
1854 }
1855
1856 exprs = sorted_array_from_bitmap_set (set);
1857 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1858 {
1859 pre_expr translated;
1860 translated = phi_translate (expr, set, NULL, pred, phiblock);
1861 if (!translated)
1862 continue;
1863
1864 /* We might end up with multiple expressions from SET being
1865 translated to the same value. In this case we do not want
1866 to retain the NARY or REFERENCE expression but prefer a NAME
1867 which would be the leader. */
1868 if (translated->kind == NAME)
1869 bitmap_value_replace_in_set (dest, translated);
1870 else
1871 bitmap_value_insert_into_set (dest, translated);
1872 }
1873 VEC_free (pre_expr, heap, exprs);
1874 }
1875
1876 /* Find the leader for a value (i.e., the name representing that
1877 value) in a given set, and return it. If STMT is non-NULL it
1878 makes sure the defining statement for the leader dominates it.
1879 Return NULL if no leader is found. */
1880
1881 static pre_expr
1882 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1883 {
1884 if (value_id_constant_p (val))
1885 {
1886 unsigned int i;
1887 bitmap_iterator bi;
1888 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1889
1890 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1891 {
1892 pre_expr expr = expression_for_id (i);
1893 if (expr->kind == CONSTANT)
1894 return expr;
1895 }
1896 }
1897 if (bitmap_set_contains_value (set, val))
1898 {
1899 /* Rather than walk the entire bitmap of expressions, and see
1900 whether any of them has the value we are looking for, we look
1901 at the reverse mapping, which tells us the set of expressions
1902 that have a given value (IE value->expressions with that
1903 value) and see if any of those expressions are in our set.
1904 The number of expressions per value is usually significantly
1905 less than the number of expressions in the set. In fact, for
1906 large testcases, doing it this way is roughly 5-10x faster
1907 than walking the bitmap.
1908 If this is somehow a significant lose for some cases, we can
1909 choose which set to walk based on which set is smaller. */
1910 unsigned int i;
1911 bitmap_iterator bi;
1912 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1913
1914 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1915 &set->expressions, 0, i, bi)
1916 {
1917 pre_expr val = expression_for_id (i);
1918 /* At the point where stmt is not null, there should always
1919 be an SSA_NAME first in the list of expressions. */
1920 if (stmt)
1921 {
1922 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1923 if (gimple_code (def_stmt) != GIMPLE_PHI
1924 && gimple_bb (def_stmt) == gimple_bb (stmt)
1925 /* PRE insertions are at the end of the basic-block
1926 and have UID 0. */
1927 && (gimple_uid (def_stmt) == 0
1928 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1929 continue;
1930 }
1931 return val;
1932 }
1933 }
1934 return NULL;
1935 }
1936
1937 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1938 BLOCK by seeing if it is not killed in the block. Note that we are
1939 only determining whether there is a store that kills it. Because
1940 of the order in which clean iterates over values, we are guaranteed
1941 that altered operands will have caused us to be eliminated from the
1942 ANTIC_IN set already. */
1943
1944 static bool
1945 value_dies_in_block_x (pre_expr expr, basic_block block)
1946 {
1947 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1948 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1949 gimple def;
1950 gimple_stmt_iterator gsi;
1951 unsigned id = get_expression_id (expr);
1952 bool res = false;
1953 ao_ref ref;
1954
1955 if (!vuse)
1956 return false;
1957
1958 /* Lookup a previously calculated result. */
1959 if (EXPR_DIES (block)
1960 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1961 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1962
1963 /* A memory expression {e, VUSE} dies in the block if there is a
1964 statement that may clobber e. If, starting statement walk from the
1965 top of the basic block, a statement uses VUSE there can be no kill
1966 inbetween that use and the original statement that loaded {e, VUSE},
1967 so we can stop walking. */
1968 ref.base = NULL_TREE;
1969 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1970 {
1971 tree def_vuse, def_vdef;
1972 def = gsi_stmt (gsi);
1973 def_vuse = gimple_vuse (def);
1974 def_vdef = gimple_vdef (def);
1975
1976 /* Not a memory statement. */
1977 if (!def_vuse)
1978 continue;
1979
1980 /* Not a may-def. */
1981 if (!def_vdef)
1982 {
1983 /* A load with the same VUSE, we're done. */
1984 if (def_vuse == vuse)
1985 break;
1986
1987 continue;
1988 }
1989
1990 /* Init ref only if we really need it. */
1991 if (ref.base == NULL_TREE
1992 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1993 refx->operands))
1994 {
1995 res = true;
1996 break;
1997 }
1998 /* If the statement may clobber expr, it dies. */
1999 if (stmt_may_clobber_ref_p_1 (def, &ref))
2000 {
2001 res = true;
2002 break;
2003 }
2004 }
2005
2006 /* Remember the result. */
2007 if (!EXPR_DIES (block))
2008 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2009 bitmap_set_bit (EXPR_DIES (block), id * 2);
2010 if (res)
2011 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2012
2013 return res;
2014 }
2015
2016
2017 #define union_contains_value(SET1, SET2, VAL) \
2018 (bitmap_set_contains_value ((SET1), (VAL)) \
2019 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2020
2021 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2022 */
2023 static bool
2024 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2025 vn_reference_op_t vro)
2026 {
2027 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2028 {
2029 struct pre_expr_d temp;
2030 temp.kind = NAME;
2031 temp.id = 0;
2032 PRE_EXPR_NAME (&temp) = vro->op0;
2033 temp.id = lookup_expression_id (&temp);
2034 if (temp.id == 0)
2035 return false;
2036 if (!union_contains_value (set1, set2,
2037 get_expr_value_id (&temp)))
2038 return false;
2039 }
2040 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2041 {
2042 struct pre_expr_d temp;
2043 temp.kind = NAME;
2044 temp.id = 0;
2045 PRE_EXPR_NAME (&temp) = vro->op1;
2046 temp.id = lookup_expression_id (&temp);
2047 if (temp.id == 0)
2048 return false;
2049 if (!union_contains_value (set1, set2,
2050 get_expr_value_id (&temp)))
2051 return false;
2052 }
2053
2054 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2055 {
2056 struct pre_expr_d temp;
2057 temp.kind = NAME;
2058 temp.id = 0;
2059 PRE_EXPR_NAME (&temp) = vro->op2;
2060 temp.id = lookup_expression_id (&temp);
2061 if (temp.id == 0)
2062 return false;
2063 if (!union_contains_value (set1, set2,
2064 get_expr_value_id (&temp)))
2065 return false;
2066 }
2067
2068 return true;
2069 }
2070
2071 /* Determine if the expression EXPR is valid in SET1 U SET2.
2072 ONLY SET2 CAN BE NULL.
2073 This means that we have a leader for each part of the expression
2074 (if it consists of values), or the expression is an SSA_NAME.
2075 For loads/calls, we also see if the vuse is killed in this block. */
2076
2077 static bool
2078 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2079 basic_block block)
2080 {
2081 switch (expr->kind)
2082 {
2083 case NAME:
2084 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2085 case NARY:
2086 {
2087 unsigned int i;
2088 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2089 for (i = 0; i < nary->length; i++)
2090 {
2091 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2092 {
2093 struct pre_expr_d temp;
2094 temp.kind = NAME;
2095 temp.id = 0;
2096 PRE_EXPR_NAME (&temp) = nary->op[i];
2097 temp.id = lookup_expression_id (&temp);
2098 if (temp.id == 0)
2099 return false;
2100 if (!union_contains_value (set1, set2,
2101 get_expr_value_id (&temp)))
2102 return false;
2103 }
2104 }
2105 /* If the NARY may trap make sure the block does not contain
2106 a possible exit point.
2107 ??? This is overly conservative if we translate AVAIL_OUT
2108 as the available expression might be after the exit point. */
2109 if (BB_MAY_NOTRETURN (block)
2110 && vn_nary_may_trap (nary))
2111 return false;
2112 return true;
2113 }
2114 break;
2115 case REFERENCE:
2116 {
2117 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2118 vn_reference_op_t vro;
2119 unsigned int i;
2120
2121 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2122 {
2123 if (!vro_valid_in_sets (set1, set2, vro))
2124 return false;
2125 }
2126 return true;
2127 }
2128 default:
2129 gcc_unreachable ();
2130 }
2131 }
2132
2133 /* Clean the set of expressions that are no longer valid in SET1 or
2134 SET2. This means expressions that are made up of values we have no
2135 leaders for in SET1 or SET2. This version is used for partial
2136 anticipation, which means it is not valid in either ANTIC_IN or
2137 PA_IN. */
2138
2139 static void
2140 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2141 {
2142 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2143 pre_expr expr;
2144 int i;
2145
2146 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2147 {
2148 if (!valid_in_sets (set1, set2, expr, block))
2149 bitmap_remove_from_set (set1, expr);
2150 }
2151 VEC_free (pre_expr, heap, exprs);
2152 }
2153
2154 /* Clean the set of expressions that are no longer valid in SET. This
2155 means expressions that are made up of values we have no leaders for
2156 in SET. */
2157
2158 static void
2159 clean (bitmap_set_t set, basic_block block)
2160 {
2161 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2162 pre_expr expr;
2163 int i;
2164
2165 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2166 {
2167 if (!valid_in_sets (set, NULL, expr, block))
2168 bitmap_remove_from_set (set, expr);
2169 }
2170 VEC_free (pre_expr, heap, exprs);
2171 }
2172
2173 /* Clean the set of expressions that are no longer valid in SET because
2174 they are clobbered in BLOCK. */
2175
2176 static void
2177 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2178 {
2179 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2180 pre_expr expr;
2181 int i;
2182
2183 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2184 {
2185 vn_reference_t ref;
2186 if (expr->kind != REFERENCE)
2187 continue;
2188
2189 ref = PRE_EXPR_REFERENCE (expr);
2190 if (ref->vuse)
2191 {
2192 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2193 if (!gimple_nop_p (def_stmt)
2194 && ((gimple_bb (def_stmt) != block
2195 && !dominated_by_p (CDI_DOMINATORS,
2196 block, gimple_bb (def_stmt)))
2197 || (gimple_bb (def_stmt) == block
2198 && value_dies_in_block_x (expr, block))))
2199 bitmap_remove_from_set (set, expr);
2200 }
2201 }
2202 VEC_free (pre_expr, heap, exprs);
2203 }
2204
2205 static sbitmap has_abnormal_preds;
2206
2207 /* List of blocks that may have changed during ANTIC computation and
2208 thus need to be iterated over. */
2209
2210 static sbitmap changed_blocks;
2211
2212 /* Decide whether to defer a block for a later iteration, or PHI
2213 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2214 should defer the block, and true if we processed it. */
2215
2216 static bool
2217 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2218 basic_block block, basic_block phiblock)
2219 {
2220 if (!BB_VISITED (phiblock))
2221 {
2222 SET_BIT (changed_blocks, block->index);
2223 BB_VISITED (block) = 0;
2224 BB_DEFERRED (block) = 1;
2225 return false;
2226 }
2227 else
2228 phi_translate_set (dest, source, block, phiblock);
2229 return true;
2230 }
2231
2232 /* Compute the ANTIC set for BLOCK.
2233
2234 If succs(BLOCK) > 1 then
2235 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2236 else if succs(BLOCK) == 1 then
2237 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2238
2239 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2240 */
2241
2242 static bool
2243 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2244 {
2245 bool changed = false;
2246 bitmap_set_t S, old, ANTIC_OUT;
2247 bitmap_iterator bi;
2248 unsigned int bii;
2249 edge e;
2250 edge_iterator ei;
2251
2252 old = ANTIC_OUT = S = NULL;
2253 BB_VISITED (block) = 1;
2254
2255 /* If any edges from predecessors are abnormal, antic_in is empty,
2256 so do nothing. */
2257 if (block_has_abnormal_pred_edge)
2258 goto maybe_dump_sets;
2259
2260 old = ANTIC_IN (block);
2261 ANTIC_OUT = bitmap_set_new ();
2262
2263 /* If the block has no successors, ANTIC_OUT is empty. */
2264 if (EDGE_COUNT (block->succs) == 0)
2265 ;
2266 /* If we have one successor, we could have some phi nodes to
2267 translate through. */
2268 else if (single_succ_p (block))
2269 {
2270 basic_block succ_bb = single_succ (block);
2271
2272 /* We trade iterations of the dataflow equations for having to
2273 phi translate the maximal set, which is incredibly slow
2274 (since the maximal set often has 300+ members, even when you
2275 have a small number of blocks).
2276 Basically, we defer the computation of ANTIC for this block
2277 until we have processed it's successor, which will inevitably
2278 have a *much* smaller set of values to phi translate once
2279 clean has been run on it.
2280 The cost of doing this is that we technically perform more
2281 iterations, however, they are lower cost iterations.
2282
2283 Timings for PRE on tramp3d-v4:
2284 without maximal set fix: 11 seconds
2285 with maximal set fix/without deferring: 26 seconds
2286 with maximal set fix/with deferring: 11 seconds
2287 */
2288
2289 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2290 block, succ_bb))
2291 {
2292 changed = true;
2293 goto maybe_dump_sets;
2294 }
2295 }
2296 /* If we have multiple successors, we take the intersection of all of
2297 them. Note that in the case of loop exit phi nodes, we may have
2298 phis to translate through. */
2299 else
2300 {
2301 VEC(basic_block, heap) * worklist;
2302 size_t i;
2303 basic_block bprime, first = NULL;
2304
2305 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2306 FOR_EACH_EDGE (e, ei, block->succs)
2307 {
2308 if (!first
2309 && BB_VISITED (e->dest))
2310 first = e->dest;
2311 else if (BB_VISITED (e->dest))
2312 VEC_quick_push (basic_block, worklist, e->dest);
2313 }
2314
2315 /* Of multiple successors we have to have visited one already. */
2316 if (!first)
2317 {
2318 SET_BIT (changed_blocks, block->index);
2319 BB_VISITED (block) = 0;
2320 BB_DEFERRED (block) = 1;
2321 changed = true;
2322 VEC_free (basic_block, heap, worklist);
2323 goto maybe_dump_sets;
2324 }
2325
2326 if (!gimple_seq_empty_p (phi_nodes (first)))
2327 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2328 else
2329 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2330
2331 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2332 {
2333 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2334 {
2335 bitmap_set_t tmp = bitmap_set_new ();
2336 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2337 bitmap_set_and (ANTIC_OUT, tmp);
2338 bitmap_set_free (tmp);
2339 }
2340 else
2341 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2342 }
2343 VEC_free (basic_block, heap, worklist);
2344 }
2345
2346 /* Prune expressions that are clobbered in block and thus become
2347 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2348 prune_clobbered_mems (ANTIC_OUT, block);
2349
2350 /* Generate ANTIC_OUT - TMP_GEN. */
2351 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2352
2353 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2354 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2355 TMP_GEN (block));
2356
2357 /* Then union in the ANTIC_OUT - TMP_GEN values,
2358 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2359 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2360 bitmap_value_insert_into_set (ANTIC_IN (block),
2361 expression_for_id (bii));
2362
2363 clean (ANTIC_IN (block), block);
2364
2365 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2366 {
2367 changed = true;
2368 SET_BIT (changed_blocks, block->index);
2369 FOR_EACH_EDGE (e, ei, block->preds)
2370 SET_BIT (changed_blocks, e->src->index);
2371 }
2372 else
2373 RESET_BIT (changed_blocks, block->index);
2374
2375 maybe_dump_sets:
2376 if (dump_file && (dump_flags & TDF_DETAILS))
2377 {
2378 if (!BB_DEFERRED (block) || BB_VISITED (block))
2379 {
2380 if (ANTIC_OUT)
2381 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2382
2383 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2384 block->index);
2385
2386 if (S)
2387 print_bitmap_set (dump_file, S, "S", block->index);
2388 }
2389 else
2390 {
2391 fprintf (dump_file,
2392 "Block %d was deferred for a future iteration.\n",
2393 block->index);
2394 }
2395 }
2396 if (old)
2397 bitmap_set_free (old);
2398 if (S)
2399 bitmap_set_free (S);
2400 if (ANTIC_OUT)
2401 bitmap_set_free (ANTIC_OUT);
2402 return changed;
2403 }
2404
2405 /* Compute PARTIAL_ANTIC for BLOCK.
2406
2407 If succs(BLOCK) > 1 then
2408 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2409 in ANTIC_OUT for all succ(BLOCK)
2410 else if succs(BLOCK) == 1 then
2411 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2412
2413 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2414 - ANTIC_IN[BLOCK])
2415
2416 */
2417 static bool
2418 compute_partial_antic_aux (basic_block block,
2419 bool block_has_abnormal_pred_edge)
2420 {
2421 bool changed = false;
2422 bitmap_set_t old_PA_IN;
2423 bitmap_set_t PA_OUT;
2424 edge e;
2425 edge_iterator ei;
2426 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2427
2428 old_PA_IN = PA_OUT = NULL;
2429
2430 /* If any edges from predecessors are abnormal, antic_in is empty,
2431 so do nothing. */
2432 if (block_has_abnormal_pred_edge)
2433 goto maybe_dump_sets;
2434
2435 /* If there are too many partially anticipatable values in the
2436 block, phi_translate_set can take an exponential time: stop
2437 before the translation starts. */
2438 if (max_pa
2439 && single_succ_p (block)
2440 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2441 goto maybe_dump_sets;
2442
2443 old_PA_IN = PA_IN (block);
2444 PA_OUT = bitmap_set_new ();
2445
2446 /* If the block has no successors, ANTIC_OUT is empty. */
2447 if (EDGE_COUNT (block->succs) == 0)
2448 ;
2449 /* If we have one successor, we could have some phi nodes to
2450 translate through. Note that we can't phi translate across DFS
2451 back edges in partial antic, because it uses a union operation on
2452 the successors. For recurrences like IV's, we will end up
2453 generating a new value in the set on each go around (i + 3 (VH.1)
2454 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2455 else if (single_succ_p (block))
2456 {
2457 basic_block succ = single_succ (block);
2458 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2459 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2460 }
2461 /* If we have multiple successors, we take the union of all of
2462 them. */
2463 else
2464 {
2465 VEC(basic_block, heap) * worklist;
2466 size_t i;
2467 basic_block bprime;
2468
2469 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2470 FOR_EACH_EDGE (e, ei, block->succs)
2471 {
2472 if (e->flags & EDGE_DFS_BACK)
2473 continue;
2474 VEC_quick_push (basic_block, worklist, e->dest);
2475 }
2476 if (VEC_length (basic_block, worklist) > 0)
2477 {
2478 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2479 {
2480 unsigned int i;
2481 bitmap_iterator bi;
2482
2483 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2484 bitmap_value_insert_into_set (PA_OUT,
2485 expression_for_id (i));
2486 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2487 {
2488 bitmap_set_t pa_in = bitmap_set_new ();
2489 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2490 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2491 bitmap_value_insert_into_set (PA_OUT,
2492 expression_for_id (i));
2493 bitmap_set_free (pa_in);
2494 }
2495 else
2496 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2497 bitmap_value_insert_into_set (PA_OUT,
2498 expression_for_id (i));
2499 }
2500 }
2501 VEC_free (basic_block, heap, worklist);
2502 }
2503
2504 /* Prune expressions that are clobbered in block and thus become
2505 invalid if translated from PA_OUT to PA_IN. */
2506 prune_clobbered_mems (PA_OUT, block);
2507
2508 /* PA_IN starts with PA_OUT - TMP_GEN.
2509 Then we subtract things from ANTIC_IN. */
2510 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2511
2512 /* For partial antic, we want to put back in the phi results, since
2513 we will properly avoid making them partially antic over backedges. */
2514 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2515 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2516
2517 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2518 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2519
2520 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2521
2522 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2523 {
2524 changed = true;
2525 SET_BIT (changed_blocks, block->index);
2526 FOR_EACH_EDGE (e, ei, block->preds)
2527 SET_BIT (changed_blocks, e->src->index);
2528 }
2529 else
2530 RESET_BIT (changed_blocks, block->index);
2531
2532 maybe_dump_sets:
2533 if (dump_file && (dump_flags & TDF_DETAILS))
2534 {
2535 if (PA_OUT)
2536 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2537
2538 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2539 }
2540 if (old_PA_IN)
2541 bitmap_set_free (old_PA_IN);
2542 if (PA_OUT)
2543 bitmap_set_free (PA_OUT);
2544 return changed;
2545 }
2546
2547 /* Compute ANTIC and partial ANTIC sets. */
2548
2549 static void
2550 compute_antic (void)
2551 {
2552 bool changed = true;
2553 int num_iterations = 0;
2554 basic_block block;
2555 int i;
2556
2557 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2558 We pre-build the map of blocks with incoming abnormal edges here. */
2559 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2560 sbitmap_zero (has_abnormal_preds);
2561
2562 FOR_EACH_BB (block)
2563 {
2564 edge_iterator ei;
2565 edge e;
2566
2567 FOR_EACH_EDGE (e, ei, block->preds)
2568 {
2569 e->flags &= ~EDGE_DFS_BACK;
2570 if (e->flags & EDGE_ABNORMAL)
2571 {
2572 SET_BIT (has_abnormal_preds, block->index);
2573 break;
2574 }
2575 }
2576
2577 BB_VISITED (block) = 0;
2578 BB_DEFERRED (block) = 0;
2579
2580 /* While we are here, give empty ANTIC_IN sets to each block. */
2581 ANTIC_IN (block) = bitmap_set_new ();
2582 PA_IN (block) = bitmap_set_new ();
2583 }
2584
2585 /* At the exit block we anticipate nothing. */
2586 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2587 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2588 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2589
2590 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2591 sbitmap_ones (changed_blocks);
2592 while (changed)
2593 {
2594 if (dump_file && (dump_flags & TDF_DETAILS))
2595 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2596 /* ??? We need to clear our PHI translation cache here as the
2597 ANTIC sets shrink and we restrict valid translations to
2598 those having operands with leaders in ANTIC. Same below
2599 for PA ANTIC computation. */
2600 num_iterations++;
2601 changed = false;
2602 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2603 {
2604 if (TEST_BIT (changed_blocks, postorder[i]))
2605 {
2606 basic_block block = BASIC_BLOCK (postorder[i]);
2607 changed |= compute_antic_aux (block,
2608 TEST_BIT (has_abnormal_preds,
2609 block->index));
2610 }
2611 }
2612 /* Theoretically possible, but *highly* unlikely. */
2613 gcc_checking_assert (num_iterations < 500);
2614 }
2615
2616 statistics_histogram_event (cfun, "compute_antic iterations",
2617 num_iterations);
2618
2619 if (do_partial_partial)
2620 {
2621 sbitmap_ones (changed_blocks);
2622 mark_dfs_back_edges ();
2623 num_iterations = 0;
2624 changed = true;
2625 while (changed)
2626 {
2627 if (dump_file && (dump_flags & TDF_DETAILS))
2628 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2629 num_iterations++;
2630 changed = false;
2631 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2632 {
2633 if (TEST_BIT (changed_blocks, postorder[i]))
2634 {
2635 basic_block block = BASIC_BLOCK (postorder[i]);
2636 changed
2637 |= compute_partial_antic_aux (block,
2638 TEST_BIT (has_abnormal_preds,
2639 block->index));
2640 }
2641 }
2642 /* Theoretically possible, but *highly* unlikely. */
2643 gcc_checking_assert (num_iterations < 500);
2644 }
2645 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2646 num_iterations);
2647 }
2648 sbitmap_free (has_abnormal_preds);
2649 sbitmap_free (changed_blocks);
2650 }
2651
2652 /* Return true if we can value number the call in STMT. This is true
2653 if we have a pure or constant call to a real function. */
2654
2655 static bool
2656 can_value_number_call (gimple stmt)
2657 {
2658 if (gimple_call_internal_p (stmt))
2659 return false;
2660 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2661 return true;
2662 return false;
2663 }
2664
2665 /* Return true if OP is a tree which we can perform PRE on.
2666 This may not match the operations we can value number, but in
2667 a perfect world would. */
2668
2669 static bool
2670 can_PRE_operation (tree op)
2671 {
2672 return UNARY_CLASS_P (op)
2673 || BINARY_CLASS_P (op)
2674 || COMPARISON_CLASS_P (op)
2675 || TREE_CODE (op) == MEM_REF
2676 || TREE_CODE (op) == COMPONENT_REF
2677 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2678 || TREE_CODE (op) == CALL_EXPR
2679 || TREE_CODE (op) == ARRAY_REF;
2680 }
2681
2682
2683 /* Inserted expressions are placed onto this worklist, which is used
2684 for performing quick dead code elimination of insertions we made
2685 that didn't turn out to be necessary. */
2686 static bitmap inserted_exprs;
2687
2688 /* Pool allocated fake store expressions are placed onto this
2689 worklist, which, after performing dead code elimination, is walked
2690 to see which expressions need to be put into GC'able memory */
2691 static VEC(gimple, heap) *need_creation;
2692
2693 /* The actual worker for create_component_ref_by_pieces. */
2694
2695 static tree
2696 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2697 unsigned int *operand, gimple_seq *stmts,
2698 gimple domstmt)
2699 {
2700 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2701 *operand);
2702 tree genop;
2703 ++*operand;
2704 switch (currop->opcode)
2705 {
2706 case CALL_EXPR:
2707 {
2708 tree folded, sc = NULL_TREE;
2709 unsigned int nargs = 0;
2710 tree fn, *args;
2711 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2712 fn = currop->op0;
2713 else
2714 {
2715 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2716 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2717 if (!fn)
2718 return NULL_TREE;
2719 }
2720 if (currop->op1)
2721 {
2722 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2723 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2724 if (!sc)
2725 return NULL_TREE;
2726 }
2727 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2728 ref->operands) - 1);
2729 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2730 {
2731 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2732 operand, stmts,
2733 domstmt);
2734 if (!args[nargs])
2735 {
2736 free (args);
2737 return NULL_TREE;
2738 }
2739 nargs++;
2740 }
2741 folded = build_call_array (currop->type,
2742 (TREE_CODE (fn) == FUNCTION_DECL
2743 ? build_fold_addr_expr (fn) : fn),
2744 nargs, args);
2745 free (args);
2746 if (sc)
2747 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2748 return folded;
2749 }
2750 break;
2751 case MEM_REF:
2752 {
2753 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2754 stmts, domstmt);
2755 tree offset = currop->op0;
2756 if (!baseop)
2757 return NULL_TREE;
2758 if (TREE_CODE (baseop) == ADDR_EXPR
2759 && handled_component_p (TREE_OPERAND (baseop, 0)))
2760 {
2761 HOST_WIDE_INT off;
2762 tree base;
2763 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2764 &off);
2765 gcc_assert (base);
2766 offset = int_const_binop (PLUS_EXPR, offset,
2767 build_int_cst (TREE_TYPE (offset),
2768 off));
2769 baseop = build_fold_addr_expr (base);
2770 }
2771 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2772 }
2773 break;
2774 case TARGET_MEM_REF:
2775 {
2776 pre_expr op0expr, op1expr;
2777 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2778 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2779 ++*operand);
2780 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2781 stmts, domstmt);
2782 if (!baseop)
2783 return NULL_TREE;
2784 if (currop->op0)
2785 {
2786 op0expr = get_or_alloc_expr_for (currop->op0);
2787 genop0 = find_or_generate_expression (block, op0expr,
2788 stmts, domstmt);
2789 if (!genop0)
2790 return NULL_TREE;
2791 }
2792 if (nextop->op0)
2793 {
2794 op1expr = get_or_alloc_expr_for (nextop->op0);
2795 genop1 = find_or_generate_expression (block, op1expr,
2796 stmts, domstmt);
2797 if (!genop1)
2798 return NULL_TREE;
2799 }
2800 return build5 (TARGET_MEM_REF, currop->type,
2801 baseop, currop->op2, genop0, currop->op1, genop1);
2802 }
2803 break;
2804 case ADDR_EXPR:
2805 if (currop->op0)
2806 {
2807 gcc_assert (is_gimple_min_invariant (currop->op0));
2808 return currop->op0;
2809 }
2810 /* Fallthrough. */
2811 case REALPART_EXPR:
2812 case IMAGPART_EXPR:
2813 case VIEW_CONVERT_EXPR:
2814 {
2815 tree folded;
2816 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2817 operand,
2818 stmts, domstmt);
2819 if (!genop0)
2820 return NULL_TREE;
2821 folded = fold_build1 (currop->opcode, currop->type,
2822 genop0);
2823 return folded;
2824 }
2825 break;
2826 case WITH_SIZE_EXPR:
2827 {
2828 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2829 stmts, domstmt);
2830 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2831 tree genop1;
2832
2833 if (!genop0)
2834 return NULL_TREE;
2835
2836 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2837 if (!genop1)
2838 return NULL_TREE;
2839
2840 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2841 }
2842 break;
2843 case BIT_FIELD_REF:
2844 {
2845 tree folded;
2846 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2847 stmts, domstmt);
2848 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2849 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2850 tree genop1;
2851 tree genop2;
2852
2853 if (!genop0)
2854 return NULL_TREE;
2855 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2856 if (!genop1)
2857 return NULL_TREE;
2858 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2859 if (!genop2)
2860 return NULL_TREE;
2861 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2862 genop2);
2863 return folded;
2864 }
2865
2866 /* For array ref vn_reference_op's, operand 1 of the array ref
2867 is op0 of the reference op and operand 3 of the array ref is
2868 op1. */
2869 case ARRAY_RANGE_REF:
2870 case ARRAY_REF:
2871 {
2872 tree genop0;
2873 tree genop1 = currop->op0;
2874 pre_expr op1expr;
2875 tree genop2 = currop->op1;
2876 pre_expr op2expr;
2877 tree genop3 = currop->op2;
2878 pre_expr op3expr;
2879 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2880 stmts, domstmt);
2881 if (!genop0)
2882 return NULL_TREE;
2883 op1expr = get_or_alloc_expr_for (genop1);
2884 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2885 if (!genop1)
2886 return NULL_TREE;
2887 if (genop2)
2888 {
2889 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2890 /* Drop zero minimum index if redundant. */
2891 if (integer_zerop (genop2)
2892 && (!domain_type
2893 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2894 genop2 = NULL_TREE;
2895 else
2896 {
2897 op2expr = get_or_alloc_expr_for (genop2);
2898 genop2 = find_or_generate_expression (block, op2expr, stmts,
2899 domstmt);
2900 if (!genop2)
2901 return NULL_TREE;
2902 }
2903 }
2904 if (genop3)
2905 {
2906 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2907 /* We can't always put a size in units of the element alignment
2908 here as the element alignment may be not visible. See
2909 PR43783. Simply drop the element size for constant
2910 sizes. */
2911 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2912 genop3 = NULL_TREE;
2913 else
2914 {
2915 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2916 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2917 op3expr = get_or_alloc_expr_for (genop3);
2918 genop3 = find_or_generate_expression (block, op3expr, stmts,
2919 domstmt);
2920 if (!genop3)
2921 return NULL_TREE;
2922 }
2923 }
2924 return build4 (currop->opcode, currop->type, genop0, genop1,
2925 genop2, genop3);
2926 }
2927 case COMPONENT_REF:
2928 {
2929 tree op0;
2930 tree op1;
2931 tree genop2 = currop->op1;
2932 pre_expr op2expr;
2933 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2934 stmts, domstmt);
2935 if (!op0)
2936 return NULL_TREE;
2937 /* op1 should be a FIELD_DECL, which are represented by
2938 themselves. */
2939 op1 = currop->op0;
2940 if (genop2)
2941 {
2942 op2expr = get_or_alloc_expr_for (genop2);
2943 genop2 = find_or_generate_expression (block, op2expr, stmts,
2944 domstmt);
2945 if (!genop2)
2946 return NULL_TREE;
2947 }
2948
2949 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2950 genop2);
2951 }
2952 break;
2953 case SSA_NAME:
2954 {
2955 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2956 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2957 return genop;
2958 }
2959 case STRING_CST:
2960 case INTEGER_CST:
2961 case COMPLEX_CST:
2962 case VECTOR_CST:
2963 case REAL_CST:
2964 case CONSTRUCTOR:
2965 case VAR_DECL:
2966 case PARM_DECL:
2967 case CONST_DECL:
2968 case RESULT_DECL:
2969 case FUNCTION_DECL:
2970 return currop->op0;
2971
2972 default:
2973 gcc_unreachable ();
2974 }
2975 }
2976
2977 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2978 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2979 trying to rename aggregates into ssa form directly, which is a no no.
2980
2981 Thus, this routine doesn't create temporaries, it just builds a
2982 single access expression for the array, calling
2983 find_or_generate_expression to build the innermost pieces.
2984
2985 This function is a subroutine of create_expression_by_pieces, and
2986 should not be called on it's own unless you really know what you
2987 are doing. */
2988
2989 static tree
2990 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2991 gimple_seq *stmts, gimple domstmt)
2992 {
2993 unsigned int op = 0;
2994 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2995 }
2996
2997 /* Find a leader for an expression, or generate one using
2998 create_expression_by_pieces if it's ANTIC but
2999 complex.
3000 BLOCK is the basic_block we are looking for leaders in.
3001 EXPR is the expression to find a leader or generate for.
3002 STMTS is the statement list to put the inserted expressions on.
3003 Returns the SSA_NAME of the LHS of the generated expression or the
3004 leader.
3005 DOMSTMT if non-NULL is a statement that should be dominated by
3006 all uses in the generated expression. If DOMSTMT is non-NULL this
3007 routine can fail and return NULL_TREE. Otherwise it will assert
3008 on failure. */
3009
3010 static tree
3011 find_or_generate_expression (basic_block block, pre_expr expr,
3012 gimple_seq *stmts, gimple domstmt)
3013 {
3014 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
3015 get_expr_value_id (expr), domstmt);
3016 tree genop = NULL;
3017 if (leader)
3018 {
3019 if (leader->kind == NAME)
3020 genop = PRE_EXPR_NAME (leader);
3021 else if (leader->kind == CONSTANT)
3022 genop = PRE_EXPR_CONSTANT (leader);
3023 }
3024
3025 /* If it's still NULL, it must be a complex expression, so generate
3026 it recursively. Not so if inserting expressions for values generated
3027 by SCCVN. */
3028 if (genop == NULL
3029 && !domstmt)
3030 {
3031 bitmap_set_t exprset;
3032 unsigned int lookfor = get_expr_value_id (expr);
3033 bool handled = false;
3034 bitmap_iterator bi;
3035 unsigned int i;
3036
3037 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3038 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3039 {
3040 pre_expr temp = expression_for_id (i);
3041 if (temp->kind != NAME)
3042 {
3043 handled = true;
3044 genop = create_expression_by_pieces (block, temp, stmts,
3045 domstmt,
3046 get_expr_type (expr));
3047 break;
3048 }
3049 }
3050 if (!handled && domstmt)
3051 return NULL_TREE;
3052
3053 gcc_assert (handled);
3054 }
3055 return genop;
3056 }
3057
3058 #define NECESSARY GF_PLF_1
3059
3060 /* Create an expression in pieces, so that we can handle very complex
3061 expressions that may be ANTIC, but not necessary GIMPLE.
3062 BLOCK is the basic block the expression will be inserted into,
3063 EXPR is the expression to insert (in value form)
3064 STMTS is a statement list to append the necessary insertions into.
3065
3066 This function will die if we hit some value that shouldn't be
3067 ANTIC but is (IE there is no leader for it, or its components).
3068 This function may also generate expressions that are themselves
3069 partially or fully redundant. Those that are will be either made
3070 fully redundant during the next iteration of insert (for partially
3071 redundant ones), or eliminated by eliminate (for fully redundant
3072 ones).
3073
3074 If DOMSTMT is non-NULL then we make sure that all uses in the
3075 expressions dominate that statement. In this case the function
3076 can return NULL_TREE to signal failure. */
3077
3078 static tree
3079 create_expression_by_pieces (basic_block block, pre_expr expr,
3080 gimple_seq *stmts, gimple domstmt, tree type)
3081 {
3082 tree temp, name;
3083 tree folded;
3084 gimple_seq forced_stmts = NULL;
3085 unsigned int value_id;
3086 gimple_stmt_iterator gsi;
3087 tree exprtype = type ? type : get_expr_type (expr);
3088 pre_expr nameexpr;
3089 gimple newstmt;
3090
3091 switch (expr->kind)
3092 {
3093 /* We may hit the NAME/CONSTANT case if we have to convert types
3094 that value numbering saw through. */
3095 case NAME:
3096 folded = PRE_EXPR_NAME (expr);
3097 break;
3098 case CONSTANT:
3099 folded = PRE_EXPR_CONSTANT (expr);
3100 break;
3101 case REFERENCE:
3102 {
3103 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3104 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3105 }
3106 break;
3107 case NARY:
3108 {
3109 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3110 tree genop[4];
3111 unsigned i;
3112 for (i = 0; i < nary->length; ++i)
3113 {
3114 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3115 genop[i] = find_or_generate_expression (block, op,
3116 stmts, domstmt);
3117 if (!genop[i])
3118 return NULL_TREE;
3119 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3120 may have conversions stripped. */
3121 if (nary->opcode == POINTER_PLUS_EXPR)
3122 {
3123 if (i == 0)
3124 genop[i] = fold_convert (nary->type, genop[i]);
3125 else if (i == 1)
3126 genop[i] = convert_to_ptrofftype (genop[i]);
3127 }
3128 else
3129 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3130 }
3131 if (nary->opcode == CONSTRUCTOR)
3132 {
3133 VEC(constructor_elt,gc) *elts = NULL;
3134 for (i = 0; i < nary->length; ++i)
3135 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3136 folded = build_constructor (nary->type, elts);
3137 }
3138 else
3139 {
3140 switch (nary->length)
3141 {
3142 case 1:
3143 folded = fold_build1 (nary->opcode, nary->type,
3144 genop[0]);
3145 break;
3146 case 2:
3147 folded = fold_build2 (nary->opcode, nary->type,
3148 genop[0], genop[1]);
3149 break;
3150 case 3:
3151 folded = fold_build3 (nary->opcode, nary->type,
3152 genop[0], genop[1], genop[3]);
3153 break;
3154 default:
3155 gcc_unreachable ();
3156 }
3157 }
3158 }
3159 break;
3160 default:
3161 return NULL_TREE;
3162 }
3163
3164 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3165 folded = fold_convert (exprtype, folded);
3166
3167 /* Force the generated expression to be a sequence of GIMPLE
3168 statements.
3169 We have to call unshare_expr because force_gimple_operand may
3170 modify the tree we pass to it. */
3171 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3172 false, NULL);
3173
3174 /* If we have any intermediate expressions to the value sets, add them
3175 to the value sets and chain them in the instruction stream. */
3176 if (forced_stmts)
3177 {
3178 gsi = gsi_start (forced_stmts);
3179 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3180 {
3181 gimple stmt = gsi_stmt (gsi);
3182 tree forcedname = gimple_get_lhs (stmt);
3183 pre_expr nameexpr;
3184
3185 if (TREE_CODE (forcedname) == SSA_NAME)
3186 {
3187 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3188 VN_INFO_GET (forcedname)->valnum = forcedname;
3189 VN_INFO (forcedname)->value_id = get_next_value_id ();
3190 nameexpr = get_or_alloc_expr_for_name (forcedname);
3191 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3192 if (!in_fre)
3193 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3194 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3195 }
3196 mark_symbols_for_renaming (stmt);
3197 }
3198 gimple_seq_add_seq (stmts, forced_stmts);
3199 }
3200
3201 /* Build and insert the assignment of the end result to the temporary
3202 that we will return. */
3203 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3204 pretemp = create_tmp_reg (exprtype, "pretmp");
3205
3206 temp = pretemp;
3207 add_referenced_var (temp);
3208
3209 newstmt = gimple_build_assign (temp, folded);
3210 name = make_ssa_name (temp, newstmt);
3211 gimple_assign_set_lhs (newstmt, name);
3212 gimple_set_plf (newstmt, NECESSARY, false);
3213
3214 gimple_seq_add_stmt (stmts, newstmt);
3215 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3216
3217 /* All the symbols in NEWEXPR should be put into SSA form. */
3218 mark_symbols_for_renaming (newstmt);
3219
3220 /* Fold the last statement. */
3221 gsi = gsi_last (*stmts);
3222 if (fold_stmt_inplace (&gsi))
3223 update_stmt (gsi_stmt (gsi));
3224
3225 /* Add a value number to the temporary.
3226 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3227 we are creating the expression by pieces, and this particular piece of
3228 the expression may have been represented. There is no harm in replacing
3229 here. */
3230 VN_INFO_GET (name)->valnum = name;
3231 value_id = get_expr_value_id (expr);
3232 VN_INFO (name)->value_id = value_id;
3233 nameexpr = get_or_alloc_expr_for_name (name);
3234 add_to_value (value_id, nameexpr);
3235 if (NEW_SETS (block))
3236 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3237 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3238
3239 pre_stats.insertions++;
3240 if (dump_file && (dump_flags & TDF_DETAILS))
3241 {
3242 fprintf (dump_file, "Inserted ");
3243 print_gimple_stmt (dump_file, newstmt, 0, 0);
3244 fprintf (dump_file, " in predecessor %d\n", block->index);
3245 }
3246
3247 return name;
3248 }
3249
3250
3251 /* Returns true if we want to inhibit the insertions of PHI nodes
3252 for the given EXPR for basic block BB (a member of a loop).
3253 We want to do this, when we fear that the induction variable we
3254 create might inhibit vectorization. */
3255
3256 static bool
3257 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3258 {
3259 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3260 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3261 vn_reference_op_t op;
3262 unsigned i;
3263
3264 /* If we aren't going to vectorize we don't inhibit anything. */
3265 if (!flag_tree_vectorize)
3266 return false;
3267
3268 /* Otherwise we inhibit the insertion when the address of the
3269 memory reference is a simple induction variable. In other
3270 cases the vectorizer won't do anything anyway (either it's
3271 loop invariant or a complicated expression). */
3272 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3273 {
3274 switch (op->opcode)
3275 {
3276 case CALL_EXPR:
3277 /* Calls are not a problem. */
3278 return false;
3279
3280 case ARRAY_REF:
3281 case ARRAY_RANGE_REF:
3282 if (TREE_CODE (op->op0) != SSA_NAME)
3283 break;
3284 /* Fallthru. */
3285 case SSA_NAME:
3286 {
3287 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3288 affine_iv iv;
3289 /* Default defs are loop invariant. */
3290 if (!defbb)
3291 break;
3292 /* Defined outside this loop, also loop invariant. */
3293 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3294 break;
3295 /* If it's a simple induction variable inhibit insertion,
3296 the vectorizer might be interested in this one. */
3297 if (simple_iv (bb->loop_father, bb->loop_father,
3298 op->op0, &iv, true))
3299 return true;
3300 /* No simple IV, vectorizer can't do anything, hence no
3301 reason to inhibit the transformation for this operand. */
3302 break;
3303 }
3304 default:
3305 break;
3306 }
3307 }
3308 return false;
3309 }
3310
3311 /* Insert the to-be-made-available values of expression EXPRNUM for each
3312 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3313 merge the result with a phi node, given the same value number as
3314 NODE. Return true if we have inserted new stuff. */
3315
3316 static bool
3317 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3318 pre_expr *avail)
3319 {
3320 pre_expr expr = expression_for_id (exprnum);
3321 pre_expr newphi;
3322 unsigned int val = get_expr_value_id (expr);
3323 edge pred;
3324 bool insertions = false;
3325 bool nophi = false;
3326 basic_block bprime;
3327 pre_expr eprime;
3328 edge_iterator ei;
3329 tree type = get_expr_type (expr);
3330 tree temp;
3331 gimple phi;
3332
3333 if (dump_file && (dump_flags & TDF_DETAILS))
3334 {
3335 fprintf (dump_file, "Found partial redundancy for expression ");
3336 print_pre_expr (dump_file, expr);
3337 fprintf (dump_file, " (%04d)\n", val);
3338 }
3339
3340 /* Make sure we aren't creating an induction variable. */
3341 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3342 {
3343 bool firstinsideloop = false;
3344 bool secondinsideloop = false;
3345 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3346 EDGE_PRED (block, 0)->src);
3347 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3348 EDGE_PRED (block, 1)->src);
3349 /* Induction variables only have one edge inside the loop. */
3350 if ((firstinsideloop ^ secondinsideloop)
3351 && (expr->kind != REFERENCE
3352 || inhibit_phi_insertion (block, expr)))
3353 {
3354 if (dump_file && (dump_flags & TDF_DETAILS))
3355 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3356 nophi = true;
3357 }
3358 }
3359
3360 /* Make the necessary insertions. */
3361 FOR_EACH_EDGE (pred, ei, block->preds)
3362 {
3363 gimple_seq stmts = NULL;
3364 tree builtexpr;
3365 bprime = pred->src;
3366 eprime = avail[bprime->index];
3367
3368 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3369 {
3370 builtexpr = create_expression_by_pieces (bprime,
3371 eprime,
3372 &stmts, NULL,
3373 type);
3374 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3375 gsi_insert_seq_on_edge (pred, stmts);
3376 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3377 insertions = true;
3378 }
3379 else if (eprime->kind == CONSTANT)
3380 {
3381 /* Constants may not have the right type, fold_convert
3382 should give us back a constant with the right type.
3383 */
3384 tree constant = PRE_EXPR_CONSTANT (eprime);
3385 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3386 {
3387 tree builtexpr = fold_convert (type, constant);
3388 if (!is_gimple_min_invariant (builtexpr))
3389 {
3390 tree forcedexpr = force_gimple_operand (builtexpr,
3391 &stmts, true,
3392 NULL);
3393 if (!is_gimple_min_invariant (forcedexpr))
3394 {
3395 if (forcedexpr != builtexpr)
3396 {
3397 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3398 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3399 }
3400 if (stmts)
3401 {
3402 gimple_stmt_iterator gsi;
3403 gsi = gsi_start (stmts);
3404 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3405 {
3406 gimple stmt = gsi_stmt (gsi);
3407 tree lhs = gimple_get_lhs (stmt);
3408 if (TREE_CODE (lhs) == SSA_NAME)
3409 bitmap_set_bit (inserted_exprs,
3410 SSA_NAME_VERSION (lhs));
3411 gimple_set_plf (stmt, NECESSARY, false);
3412 }
3413 gsi_insert_seq_on_edge (pred, stmts);
3414 }
3415 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3416 }
3417 }
3418 else
3419 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3420 }
3421 }
3422 else if (eprime->kind == NAME)
3423 {
3424 /* We may have to do a conversion because our value
3425 numbering can look through types in certain cases, but
3426 our IL requires all operands of a phi node have the same
3427 type. */
3428 tree name = PRE_EXPR_NAME (eprime);
3429 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3430 {
3431 tree builtexpr;
3432 tree forcedexpr;
3433 builtexpr = fold_convert (type, name);
3434 forcedexpr = force_gimple_operand (builtexpr,
3435 &stmts, true,
3436 NULL);
3437
3438 if (forcedexpr != name)
3439 {
3440 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3441 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3442 }
3443
3444 if (stmts)
3445 {
3446 gimple_stmt_iterator gsi;
3447 gsi = gsi_start (stmts);
3448 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3449 {
3450 gimple stmt = gsi_stmt (gsi);
3451 tree lhs = gimple_get_lhs (stmt);
3452 if (TREE_CODE (lhs) == SSA_NAME)
3453 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3454 gimple_set_plf (stmt, NECESSARY, false);
3455 }
3456 gsi_insert_seq_on_edge (pred, stmts);
3457 }
3458 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3459 }
3460 }
3461 }
3462 /* If we didn't want a phi node, and we made insertions, we still have
3463 inserted new stuff, and thus return true. If we didn't want a phi node,
3464 and didn't make insertions, we haven't added anything new, so return
3465 false. */
3466 if (nophi && insertions)
3467 return true;
3468 else if (nophi && !insertions)
3469 return false;
3470
3471 /* Now build a phi for the new variable. */
3472 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3473 prephitemp = create_tmp_var (type, "prephitmp");
3474
3475 temp = prephitemp;
3476 add_referenced_var (temp);
3477
3478 if (TREE_CODE (type) == COMPLEX_TYPE
3479 || TREE_CODE (type) == VECTOR_TYPE)
3480 DECL_GIMPLE_REG_P (temp) = 1;
3481 phi = create_phi_node (temp, block);
3482
3483 gimple_set_plf (phi, NECESSARY, false);
3484 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3485 VN_INFO (gimple_phi_result (phi))->value_id = val;
3486 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3487 FOR_EACH_EDGE (pred, ei, block->preds)
3488 {
3489 pre_expr ae = avail[pred->src->index];
3490 gcc_assert (get_expr_type (ae) == type
3491 || useless_type_conversion_p (type, get_expr_type (ae)));
3492 if (ae->kind == CONSTANT)
3493 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3494 else
3495 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3496 UNKNOWN_LOCATION);
3497 }
3498
3499 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3500 add_to_value (val, newphi);
3501
3502 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3503 this insertion, since we test for the existence of this value in PHI_GEN
3504 before proceeding with the partial redundancy checks in insert_aux.
3505
3506 The value may exist in AVAIL_OUT, in particular, it could be represented
3507 by the expression we are trying to eliminate, in which case we want the
3508 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3509 inserted there.
3510
3511 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3512 this block, because if it did, it would have existed in our dominator's
3513 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3514 */
3515
3516 bitmap_insert_into_set (PHI_GEN (block), newphi);
3517 bitmap_value_replace_in_set (AVAIL_OUT (block),
3518 newphi);
3519 bitmap_insert_into_set (NEW_SETS (block),
3520 newphi);
3521
3522 if (dump_file && (dump_flags & TDF_DETAILS))
3523 {
3524 fprintf (dump_file, "Created phi ");
3525 print_gimple_stmt (dump_file, phi, 0, 0);
3526 fprintf (dump_file, " in block %d\n", block->index);
3527 }
3528 pre_stats.phis++;
3529 return true;
3530 }
3531
3532
3533
3534 /* Perform insertion of partially redundant values.
3535 For BLOCK, do the following:
3536 1. Propagate the NEW_SETS of the dominator into the current block.
3537 If the block has multiple predecessors,
3538 2a. Iterate over the ANTIC expressions for the block to see if
3539 any of them are partially redundant.
3540 2b. If so, insert them into the necessary predecessors to make
3541 the expression fully redundant.
3542 2c. Insert a new PHI merging the values of the predecessors.
3543 2d. Insert the new PHI, and the new expressions, into the
3544 NEW_SETS set.
3545 3. Recursively call ourselves on the dominator children of BLOCK.
3546
3547 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3548 do_regular_insertion and do_partial_insertion.
3549
3550 */
3551
3552 static bool
3553 do_regular_insertion (basic_block block, basic_block dom)
3554 {
3555 bool new_stuff = false;
3556 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3557 pre_expr expr;
3558 int i;
3559
3560 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3561 {
3562 if (expr->kind != NAME)
3563 {
3564 pre_expr *avail;
3565 unsigned int val;
3566 bool by_some = false;
3567 bool cant_insert = false;
3568 bool all_same = true;
3569 pre_expr first_s = NULL;
3570 edge pred;
3571 basic_block bprime;
3572 pre_expr eprime = NULL;
3573 edge_iterator ei;
3574 pre_expr edoubleprime = NULL;
3575 bool do_insertion = false;
3576
3577 val = get_expr_value_id (expr);
3578 if (bitmap_set_contains_value (PHI_GEN (block), val))
3579 continue;
3580 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3581 {
3582 if (dump_file && (dump_flags & TDF_DETAILS))
3583 fprintf (dump_file, "Found fully redundant value\n");
3584 continue;
3585 }
3586
3587 avail = XCNEWVEC (pre_expr, last_basic_block);
3588 FOR_EACH_EDGE (pred, ei, block->preds)
3589 {
3590 unsigned int vprime;
3591
3592 /* We should never run insertion for the exit block
3593 and so not come across fake pred edges. */
3594 gcc_assert (!(pred->flags & EDGE_FAKE));
3595 bprime = pred->src;
3596 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3597 bprime, block);
3598
3599 /* eprime will generally only be NULL if the
3600 value of the expression, translated
3601 through the PHI for this predecessor, is
3602 undefined. If that is the case, we can't
3603 make the expression fully redundant,
3604 because its value is undefined along a
3605 predecessor path. We can thus break out
3606 early because it doesn't matter what the
3607 rest of the results are. */
3608 if (eprime == NULL)
3609 {
3610 cant_insert = true;
3611 break;
3612 }
3613
3614 eprime = fully_constant_expression (eprime);
3615 vprime = get_expr_value_id (eprime);
3616 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3617 vprime, NULL);
3618 if (edoubleprime == NULL)
3619 {
3620 avail[bprime->index] = eprime;
3621 all_same = false;
3622 }
3623 else
3624 {
3625 avail[bprime->index] = edoubleprime;
3626 by_some = true;
3627 /* We want to perform insertions to remove a redundancy on
3628 a path in the CFG we want to optimize for speed. */
3629 if (optimize_edge_for_speed_p (pred))
3630 do_insertion = true;
3631 if (first_s == NULL)
3632 first_s = edoubleprime;
3633 else if (!pre_expr_eq (first_s, edoubleprime))
3634 all_same = false;
3635 }
3636 }
3637 /* If we can insert it, it's not the same value
3638 already existing along every predecessor, and
3639 it's defined by some predecessor, it is
3640 partially redundant. */
3641 if (!cant_insert && !all_same && by_some)
3642 {
3643 if (!do_insertion)
3644 {
3645 if (dump_file && (dump_flags & TDF_DETAILS))
3646 {
3647 fprintf (dump_file, "Skipping partial redundancy for "
3648 "expression ");
3649 print_pre_expr (dump_file, expr);
3650 fprintf (dump_file, " (%04d), no redundancy on to be "
3651 "optimized for speed edge\n", val);
3652 }
3653 }
3654 else if (dbg_cnt (treepre_insert)
3655 && insert_into_preds_of_block (block,
3656 get_expression_id (expr),
3657 avail))
3658 new_stuff = true;
3659 }
3660 /* If all edges produce the same value and that value is
3661 an invariant, then the PHI has the same value on all
3662 edges. Note this. */
3663 else if (!cant_insert && all_same && eprime
3664 && (edoubleprime->kind == CONSTANT
3665 || edoubleprime->kind == NAME)
3666 && !value_id_constant_p (val))
3667 {
3668 unsigned int j;
3669 bitmap_iterator bi;
3670 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3671 value_expressions, val);
3672
3673 unsigned int new_val = get_expr_value_id (edoubleprime);
3674 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3675 {
3676 pre_expr expr = expression_for_id (j);
3677
3678 if (expr->kind == NAME)
3679 {
3680 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3681 /* Just reset the value id and valnum so it is
3682 the same as the constant we have discovered. */
3683 if (edoubleprime->kind == CONSTANT)
3684 {
3685 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3686 pre_stats.constified++;
3687 }
3688 else
3689 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3690 info->value_id = new_val;
3691 }
3692 }
3693 }
3694 free (avail);
3695 }
3696 }
3697
3698 VEC_free (pre_expr, heap, exprs);
3699 return new_stuff;
3700 }
3701
3702
3703 /* Perform insertion for partially anticipatable expressions. There
3704 is only one case we will perform insertion for these. This case is
3705 if the expression is partially anticipatable, and fully available.
3706 In this case, we know that putting it earlier will enable us to
3707 remove the later computation. */
3708
3709
3710 static bool
3711 do_partial_partial_insertion (basic_block block, basic_block dom)
3712 {
3713 bool new_stuff = false;
3714 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3715 pre_expr expr;
3716 int i;
3717
3718 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3719 {
3720 if (expr->kind != NAME)
3721 {
3722 pre_expr *avail;
3723 unsigned int val;
3724 bool by_all = true;
3725 bool cant_insert = false;
3726 edge pred;
3727 basic_block bprime;
3728 pre_expr eprime = NULL;
3729 edge_iterator ei;
3730
3731 val = get_expr_value_id (expr);
3732 if (bitmap_set_contains_value (PHI_GEN (block), val))
3733 continue;
3734 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3735 continue;
3736
3737 avail = XCNEWVEC (pre_expr, last_basic_block);
3738 FOR_EACH_EDGE (pred, ei, block->preds)
3739 {
3740 unsigned int vprime;
3741 pre_expr edoubleprime;
3742
3743 /* We should never run insertion for the exit block
3744 and so not come across fake pred edges. */
3745 gcc_assert (!(pred->flags & EDGE_FAKE));
3746 bprime = pred->src;
3747 eprime = phi_translate (expr, ANTIC_IN (block),
3748 PA_IN (block),
3749 bprime, block);
3750
3751 /* eprime will generally only be NULL if the
3752 value of the expression, translated
3753 through the PHI for this predecessor, is
3754 undefined. If that is the case, we can't
3755 make the expression fully redundant,
3756 because its value is undefined along a
3757 predecessor path. We can thus break out
3758 early because it doesn't matter what the
3759 rest of the results are. */
3760 if (eprime == NULL)
3761 {
3762 cant_insert = true;
3763 break;
3764 }
3765
3766 eprime = fully_constant_expression (eprime);
3767 vprime = get_expr_value_id (eprime);
3768 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3769 vprime, NULL);
3770 if (edoubleprime == NULL)
3771 {
3772 by_all = false;
3773 break;
3774 }
3775 else
3776 avail[bprime->index] = edoubleprime;
3777 }
3778
3779 /* If we can insert it, it's not the same value
3780 already existing along every predecessor, and
3781 it's defined by some predecessor, it is
3782 partially redundant. */
3783 if (!cant_insert && by_all)
3784 {
3785 edge succ;
3786 bool do_insertion = false;
3787
3788 /* Insert only if we can remove a later expression on a path
3789 that we want to optimize for speed.
3790 The phi node that we will be inserting in BLOCK is not free,
3791 and inserting it for the sake of !optimize_for_speed successor
3792 may cause regressions on the speed path. */
3793 FOR_EACH_EDGE (succ, ei, block->succs)
3794 {
3795 if (bitmap_set_contains_value (PA_IN (succ->dest), val))
3796 {
3797 if (optimize_edge_for_speed_p (succ))
3798 do_insertion = true;
3799 }
3800 }
3801
3802 if (!do_insertion)
3803 {
3804 if (dump_file && (dump_flags & TDF_DETAILS))
3805 {
3806 fprintf (dump_file, "Skipping partial partial redundancy "
3807 "for expression ");
3808 print_pre_expr (dump_file, expr);
3809 fprintf (dump_file, " (%04d), not partially anticipated "
3810 "on any to be optimized for speed edges\n", val);
3811 }
3812 }
3813 else if (dbg_cnt (treepre_insert))
3814 {
3815 pre_stats.pa_insert++;
3816 if (insert_into_preds_of_block (block,
3817 get_expression_id (expr),
3818 avail))
3819 new_stuff = true;
3820 }
3821 }
3822 free (avail);
3823 }
3824 }
3825
3826 VEC_free (pre_expr, heap, exprs);
3827 return new_stuff;
3828 }
3829
3830 static bool
3831 insert_aux (basic_block block)
3832 {
3833 basic_block son;
3834 bool new_stuff = false;
3835
3836 if (block)
3837 {
3838 basic_block dom;
3839 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3840 if (dom)
3841 {
3842 unsigned i;
3843 bitmap_iterator bi;
3844 bitmap_set_t newset = NEW_SETS (dom);
3845 if (newset)
3846 {
3847 /* Note that we need to value_replace both NEW_SETS, and
3848 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3849 represented by some non-simple expression here that we want
3850 to replace it with. */
3851 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3852 {
3853 pre_expr expr = expression_for_id (i);
3854 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3855 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3856 }
3857 }
3858 if (!single_pred_p (block))
3859 {
3860 new_stuff |= do_regular_insertion (block, dom);
3861 if (do_partial_partial)
3862 new_stuff |= do_partial_partial_insertion (block, dom);
3863 }
3864 }
3865 }
3866 for (son = first_dom_son (CDI_DOMINATORS, block);
3867 son;
3868 son = next_dom_son (CDI_DOMINATORS, son))
3869 {
3870 new_stuff |= insert_aux (son);
3871 }
3872
3873 return new_stuff;
3874 }
3875
3876 /* Perform insertion of partially redundant values. */
3877
3878 static void
3879 insert (void)
3880 {
3881 bool new_stuff = true;
3882 basic_block bb;
3883 int num_iterations = 0;
3884
3885 FOR_ALL_BB (bb)
3886 NEW_SETS (bb) = bitmap_set_new ();
3887
3888 while (new_stuff)
3889 {
3890 num_iterations++;
3891 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3892 }
3893 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3894 }
3895
3896
3897 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3898
3899 static void
3900 add_to_exp_gen (basic_block block, tree op)
3901 {
3902 if (!in_fre)
3903 {
3904 pre_expr result;
3905 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3906 return;
3907 result = get_or_alloc_expr_for_name (op);
3908 bitmap_value_insert_into_set (EXP_GEN (block), result);
3909 }
3910 }
3911
3912 /* Create value ids for PHI in BLOCK. */
3913
3914 static void
3915 make_values_for_phi (gimple phi, basic_block block)
3916 {
3917 tree result = gimple_phi_result (phi);
3918
3919 /* We have no need for virtual phis, as they don't represent
3920 actual computations. */
3921 if (is_gimple_reg (result))
3922 {
3923 pre_expr e = get_or_alloc_expr_for_name (result);
3924 add_to_value (get_expr_value_id (e), e);
3925 bitmap_insert_into_set (PHI_GEN (block), e);
3926 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3927 if (!in_fre)
3928 {
3929 unsigned i;
3930 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3931 {
3932 tree arg = gimple_phi_arg_def (phi, i);
3933 if (TREE_CODE (arg) == SSA_NAME)
3934 {
3935 e = get_or_alloc_expr_for_name (arg);
3936 add_to_value (get_expr_value_id (e), e);
3937 }
3938 }
3939 }
3940 }
3941 }
3942
3943 /* Compute the AVAIL set for all basic blocks.
3944
3945 This function performs value numbering of the statements in each basic
3946 block. The AVAIL sets are built from information we glean while doing
3947 this value numbering, since the AVAIL sets contain only one entry per
3948 value.
3949
3950 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3951 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3952
3953 static void
3954 compute_avail (void)
3955 {
3956
3957 basic_block block, son;
3958 basic_block *worklist;
3959 size_t sp = 0;
3960 unsigned i;
3961
3962 /* We pretend that default definitions are defined in the entry block.
3963 This includes function arguments and the static chain decl. */
3964 for (i = 1; i < num_ssa_names; ++i)
3965 {
3966 tree name = ssa_name (i);
3967 pre_expr e;
3968 if (!name
3969 || !SSA_NAME_IS_DEFAULT_DEF (name)
3970 || has_zero_uses (name)
3971 || !is_gimple_reg (name))
3972 continue;
3973
3974 e = get_or_alloc_expr_for_name (name);
3975 add_to_value (get_expr_value_id (e), e);
3976 if (!in_fre)
3977 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3978 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3979 }
3980
3981 /* Allocate the worklist. */
3982 worklist = XNEWVEC (basic_block, n_basic_blocks);
3983
3984 /* Seed the algorithm by putting the dominator children of the entry
3985 block on the worklist. */
3986 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3987 son;
3988 son = next_dom_son (CDI_DOMINATORS, son))
3989 worklist[sp++] = son;
3990
3991 /* Loop until the worklist is empty. */
3992 while (sp)
3993 {
3994 gimple_stmt_iterator gsi;
3995 gimple stmt;
3996 basic_block dom;
3997 unsigned int stmt_uid = 1;
3998
3999 /* Pick a block from the worklist. */
4000 block = worklist[--sp];
4001
4002 /* Initially, the set of available values in BLOCK is that of
4003 its immediate dominator. */
4004 dom = get_immediate_dominator (CDI_DOMINATORS, block);
4005 if (dom)
4006 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
4007
4008 /* Generate values for PHI nodes. */
4009 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
4010 make_values_for_phi (gsi_stmt (gsi), block);
4011
4012 BB_MAY_NOTRETURN (block) = 0;
4013
4014 /* Now compute value numbers and populate value sets with all
4015 the expressions computed in BLOCK. */
4016 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
4017 {
4018 ssa_op_iter iter;
4019 tree op;
4020
4021 stmt = gsi_stmt (gsi);
4022 gimple_set_uid (stmt, stmt_uid++);
4023
4024 /* Cache whether the basic-block has any non-visible side-effect
4025 or control flow.
4026 If this isn't a call or it is the last stmt in the
4027 basic-block then the CFG represents things correctly. */
4028 if (is_gimple_call (stmt)
4029 && !stmt_ends_bb_p (stmt))
4030 {
4031 /* Non-looping const functions always return normally.
4032 Otherwise the call might not return or have side-effects
4033 that forbids hoisting possibly trapping expressions
4034 before it. */
4035 int flags = gimple_call_flags (stmt);
4036 if (!(flags & ECF_CONST)
4037 || (flags & ECF_LOOPING_CONST_OR_PURE))
4038 BB_MAY_NOTRETURN (block) = 1;
4039 }
4040
4041 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
4042 {
4043 pre_expr e = get_or_alloc_expr_for_name (op);
4044
4045 add_to_value (get_expr_value_id (e), e);
4046 if (!in_fre)
4047 bitmap_insert_into_set (TMP_GEN (block), e);
4048 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4049 }
4050
4051 if (gimple_has_volatile_ops (stmt)
4052 || stmt_could_throw_p (stmt))
4053 continue;
4054
4055 switch (gimple_code (stmt))
4056 {
4057 case GIMPLE_RETURN:
4058 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4059 add_to_exp_gen (block, op);
4060 continue;
4061
4062 case GIMPLE_CALL:
4063 {
4064 vn_reference_t ref;
4065 unsigned int i;
4066 vn_reference_op_t vro;
4067 pre_expr result = NULL;
4068 VEC(vn_reference_op_s, heap) *ops = NULL;
4069
4070 if (!can_value_number_call (stmt))
4071 continue;
4072
4073 copy_reference_ops_from_call (stmt, &ops);
4074 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4075 gimple_expr_type (stmt),
4076 ops, &ref, VN_NOWALK);
4077 VEC_free (vn_reference_op_s, heap, ops);
4078 if (!ref)
4079 continue;
4080
4081 for (i = 0; VEC_iterate (vn_reference_op_s,
4082 ref->operands, i,
4083 vro); i++)
4084 {
4085 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4086 add_to_exp_gen (block, vro->op0);
4087 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4088 add_to_exp_gen (block, vro->op1);
4089 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4090 add_to_exp_gen (block, vro->op2);
4091 }
4092
4093 /* If the value of the call is not invalidated in
4094 this block until it is computed, add the expression
4095 to EXP_GEN. */
4096 if (!gimple_vuse (stmt)
4097 || gimple_code
4098 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
4099 || gimple_bb (SSA_NAME_DEF_STMT
4100 (gimple_vuse (stmt))) != block)
4101 {
4102 result = (pre_expr) pool_alloc (pre_expr_pool);
4103 result->kind = REFERENCE;
4104 result->id = 0;
4105 PRE_EXPR_REFERENCE (result) = ref;
4106
4107 get_or_alloc_expression_id (result);
4108 add_to_value (get_expr_value_id (result), result);
4109 if (!in_fre)
4110 bitmap_value_insert_into_set (EXP_GEN (block), result);
4111 }
4112 continue;
4113 }
4114
4115 case GIMPLE_ASSIGN:
4116 {
4117 pre_expr result = NULL;
4118 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4119 {
4120 case tcc_unary:
4121 case tcc_binary:
4122 case tcc_comparison:
4123 {
4124 vn_nary_op_t nary;
4125 unsigned int i;
4126
4127 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4128 gimple_assign_rhs_code (stmt),
4129 gimple_expr_type (stmt),
4130 gimple_assign_rhs1_ptr (stmt),
4131 &nary);
4132
4133 if (!nary)
4134 continue;
4135
4136 for (i = 0; i < nary->length; i++)
4137 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4138 add_to_exp_gen (block, nary->op[i]);
4139
4140 result = (pre_expr) pool_alloc (pre_expr_pool);
4141 result->kind = NARY;
4142 result->id = 0;
4143 PRE_EXPR_NARY (result) = nary;
4144 break;
4145 }
4146
4147 case tcc_declaration:
4148 case tcc_reference:
4149 {
4150 vn_reference_t ref;
4151 unsigned int i;
4152 vn_reference_op_t vro;
4153
4154 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4155 gimple_vuse (stmt),
4156 VN_WALK, &ref);
4157 if (!ref)
4158 continue;
4159
4160 for (i = 0; VEC_iterate (vn_reference_op_s,
4161 ref->operands, i,
4162 vro); i++)
4163 {
4164 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4165 add_to_exp_gen (block, vro->op0);
4166 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4167 add_to_exp_gen (block, vro->op1);
4168 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4169 add_to_exp_gen (block, vro->op2);
4170 }
4171
4172 /* If the value of the reference is not invalidated in
4173 this block until it is computed, add the expression
4174 to EXP_GEN. */
4175 if (gimple_vuse (stmt))
4176 {
4177 gimple def_stmt;
4178 bool ok = true;
4179 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
4180 while (!gimple_nop_p (def_stmt)
4181 && gimple_code (def_stmt) != GIMPLE_PHI
4182 && gimple_bb (def_stmt) == block)
4183 {
4184 if (stmt_may_clobber_ref_p
4185 (def_stmt, gimple_assign_rhs1 (stmt)))
4186 {
4187 ok = false;
4188 break;
4189 }
4190 def_stmt
4191 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
4192 }
4193 if (!ok)
4194 continue;
4195 }
4196
4197 result = (pre_expr) pool_alloc (pre_expr_pool);
4198 result->kind = REFERENCE;
4199 result->id = 0;
4200 PRE_EXPR_REFERENCE (result) = ref;
4201 break;
4202 }
4203
4204 default:
4205 /* For any other statement that we don't
4206 recognize, simply add all referenced
4207 SSA_NAMEs to EXP_GEN. */
4208 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4209 add_to_exp_gen (block, op);
4210 continue;
4211 }
4212
4213 get_or_alloc_expression_id (result);
4214 add_to_value (get_expr_value_id (result), result);
4215 if (!in_fre)
4216 bitmap_value_insert_into_set (EXP_GEN (block), result);
4217
4218 continue;
4219 }
4220 default:
4221 break;
4222 }
4223 }
4224
4225 /* Put the dominator children of BLOCK on the worklist of blocks
4226 to compute available sets for. */
4227 for (son = first_dom_son (CDI_DOMINATORS, block);
4228 son;
4229 son = next_dom_son (CDI_DOMINATORS, son))
4230 worklist[sp++] = son;
4231 }
4232
4233 free (worklist);
4234 }
4235
4236 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4237 than the available expressions for it. The insertion point is
4238 right before the first use in STMT. Returns the SSA_NAME that should
4239 be used for replacement. */
4240
4241 static tree
4242 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4243 {
4244 basic_block bb = gimple_bb (stmt);
4245 gimple_stmt_iterator gsi;
4246 gimple_seq stmts = NULL;
4247 tree expr;
4248 pre_expr e;
4249
4250 /* First create a value expression from the expression we want
4251 to insert and associate it with the value handle for SSA_VN. */
4252 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4253 if (e == NULL)
4254 return NULL_TREE;
4255
4256 /* Then use create_expression_by_pieces to generate a valid
4257 expression to insert at this point of the IL stream. */
4258 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4259 if (expr == NULL_TREE)
4260 return NULL_TREE;
4261 gsi = gsi_for_stmt (stmt);
4262 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4263
4264 return expr;
4265 }
4266
4267 /* Eliminate fully redundant computations. */
4268
4269 static unsigned int
4270 eliminate (void)
4271 {
4272 VEC (gimple, heap) *to_remove = NULL;
4273 VEC (gimple, heap) *to_update = NULL;
4274 basic_block b;
4275 unsigned int todo = 0;
4276 gimple_stmt_iterator gsi;
4277 gimple stmt;
4278 unsigned i;
4279
4280 FOR_EACH_BB (b)
4281 {
4282 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4283 {
4284 tree lhs = NULL_TREE;
4285 tree rhs = NULL_TREE;
4286
4287 stmt = gsi_stmt (gsi);
4288
4289 if (gimple_has_lhs (stmt))
4290 lhs = gimple_get_lhs (stmt);
4291
4292 if (gimple_assign_single_p (stmt))
4293 rhs = gimple_assign_rhs1 (stmt);
4294
4295 /* Lookup the RHS of the expression, see if we have an
4296 available computation for it. If so, replace the RHS with
4297 the available computation.
4298
4299 See PR43491.
4300 We don't replace global register variable when it is a the RHS of
4301 a single assign. We do replace local register variable since gcc
4302 does not guarantee local variable will be allocated in register. */
4303 if (gimple_has_lhs (stmt)
4304 && TREE_CODE (lhs) == SSA_NAME
4305 && !gimple_assign_ssa_name_copy_p (stmt)
4306 && (!gimple_assign_single_p (stmt)
4307 || (!is_gimple_min_invariant (rhs)
4308 && (gimple_assign_rhs_code (stmt) != VAR_DECL
4309 || !is_global_var (rhs)
4310 || !DECL_HARD_REGISTER (rhs))))
4311 && !gimple_has_volatile_ops (stmt)
4312 && !has_zero_uses (lhs))
4313 {
4314 tree sprime = NULL;
4315 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4316 pre_expr sprimeexpr;
4317 gimple orig_stmt = stmt;
4318
4319 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4320 get_expr_value_id (lhsexpr),
4321 NULL);
4322
4323 if (sprimeexpr)
4324 {
4325 if (sprimeexpr->kind == CONSTANT)
4326 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4327 else if (sprimeexpr->kind == NAME)
4328 sprime = PRE_EXPR_NAME (sprimeexpr);
4329 else
4330 gcc_unreachable ();
4331 }
4332
4333 /* If there is no existing leader but SCCVN knows this
4334 value is constant, use that constant. */
4335 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4336 {
4337 sprime = VN_INFO (lhs)->valnum;
4338 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4339 TREE_TYPE (sprime)))
4340 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4341
4342 if (dump_file && (dump_flags & TDF_DETAILS))
4343 {
4344 fprintf (dump_file, "Replaced ");
4345 print_gimple_expr (dump_file, stmt, 0, 0);
4346 fprintf (dump_file, " with ");
4347 print_generic_expr (dump_file, sprime, 0);
4348 fprintf (dump_file, " in ");
4349 print_gimple_stmt (dump_file, stmt, 0, 0);
4350 }
4351 pre_stats.eliminations++;
4352 propagate_tree_value_into_stmt (&gsi, sprime);
4353 stmt = gsi_stmt (gsi);
4354 update_stmt (stmt);
4355
4356 /* If we removed EH side-effects from the statement, clean
4357 its EH information. */
4358 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4359 {
4360 bitmap_set_bit (need_eh_cleanup,
4361 gimple_bb (stmt)->index);
4362 if (dump_file && (dump_flags & TDF_DETAILS))
4363 fprintf (dump_file, " Removed EH side-effects.\n");
4364 }
4365 continue;
4366 }
4367
4368 /* If there is no existing usable leader but SCCVN thinks
4369 it has an expression it wants to use as replacement,
4370 insert that. */
4371 if (!sprime || sprime == lhs)
4372 {
4373 tree val = VN_INFO (lhs)->valnum;
4374 if (val != VN_TOP
4375 && TREE_CODE (val) == SSA_NAME
4376 && VN_INFO (val)->needs_insertion
4377 && can_PRE_operation (vn_get_expr_for (val)))
4378 sprime = do_SCCVN_insertion (stmt, val);
4379 }
4380 if (sprime
4381 && sprime != lhs
4382 && (rhs == NULL_TREE
4383 || TREE_CODE (rhs) != SSA_NAME
4384 || may_propagate_copy (rhs, sprime)))
4385 {
4386 bool can_make_abnormal_goto
4387 = is_gimple_call (stmt)
4388 && stmt_can_make_abnormal_goto (stmt);
4389
4390 gcc_assert (sprime != rhs);
4391
4392 if (dump_file && (dump_flags & TDF_DETAILS))
4393 {
4394 fprintf (dump_file, "Replaced ");
4395 print_gimple_expr (dump_file, stmt, 0, 0);
4396 fprintf (dump_file, " with ");
4397 print_generic_expr (dump_file, sprime, 0);
4398 fprintf (dump_file, " in ");
4399 print_gimple_stmt (dump_file, stmt, 0, 0);
4400 }
4401
4402 if (TREE_CODE (sprime) == SSA_NAME)
4403 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4404 NECESSARY, true);
4405 /* We need to make sure the new and old types actually match,
4406 which may require adding a simple cast, which fold_convert
4407 will do for us. */
4408 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4409 && !useless_type_conversion_p (gimple_expr_type (stmt),
4410 TREE_TYPE (sprime)))
4411 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4412
4413 pre_stats.eliminations++;
4414 propagate_tree_value_into_stmt (&gsi, sprime);
4415 stmt = gsi_stmt (gsi);
4416 update_stmt (stmt);
4417
4418 /* If we removed EH side-effects from the statement, clean
4419 its EH information. */
4420 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4421 {
4422 bitmap_set_bit (need_eh_cleanup,
4423 gimple_bb (stmt)->index);
4424 if (dump_file && (dump_flags & TDF_DETAILS))
4425 fprintf (dump_file, " Removed EH side-effects.\n");
4426 }
4427
4428 /* Likewise for AB side-effects. */
4429 if (can_make_abnormal_goto
4430 && !stmt_can_make_abnormal_goto (stmt))
4431 {
4432 bitmap_set_bit (need_ab_cleanup,
4433 gimple_bb (stmt)->index);
4434 if (dump_file && (dump_flags & TDF_DETAILS))
4435 fprintf (dump_file, " Removed AB side-effects.\n");
4436 }
4437 }
4438 }
4439 /* If the statement is a scalar store, see if the expression
4440 has the same value number as its rhs. If so, the store is
4441 dead. */
4442 else if (gimple_assign_single_p (stmt)
4443 && !gimple_has_volatile_ops (stmt)
4444 && !is_gimple_reg (gimple_assign_lhs (stmt))
4445 && (TREE_CODE (rhs) == SSA_NAME
4446 || is_gimple_min_invariant (rhs)))
4447 {
4448 tree val;
4449 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4450 gimple_vuse (stmt), VN_WALK, NULL);
4451 if (TREE_CODE (rhs) == SSA_NAME)
4452 rhs = VN_INFO (rhs)->valnum;
4453 if (val
4454 && operand_equal_p (val, rhs, 0))
4455 {
4456 if (dump_file && (dump_flags & TDF_DETAILS))
4457 {
4458 fprintf (dump_file, "Deleted redundant store ");
4459 print_gimple_stmt (dump_file, stmt, 0, 0);
4460 }
4461
4462 /* Queue stmt for removal. */
4463 VEC_safe_push (gimple, heap, to_remove, stmt);
4464 }
4465 }
4466 /* Visit COND_EXPRs and fold the comparison with the
4467 available value-numbers. */
4468 else if (gimple_code (stmt) == GIMPLE_COND)
4469 {
4470 tree op0 = gimple_cond_lhs (stmt);
4471 tree op1 = gimple_cond_rhs (stmt);
4472 tree result;
4473
4474 if (TREE_CODE (op0) == SSA_NAME)
4475 op0 = VN_INFO (op0)->valnum;
4476 if (TREE_CODE (op1) == SSA_NAME)
4477 op1 = VN_INFO (op1)->valnum;
4478 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4479 op0, op1);
4480 if (result && TREE_CODE (result) == INTEGER_CST)
4481 {
4482 if (integer_zerop (result))
4483 gimple_cond_make_false (stmt);
4484 else
4485 gimple_cond_make_true (stmt);
4486 update_stmt (stmt);
4487 todo = TODO_cleanup_cfg;
4488 }
4489 }
4490 /* Visit indirect calls and turn them into direct calls if
4491 possible. */
4492 if (is_gimple_call (stmt))
4493 {
4494 tree orig_fn = gimple_call_fn (stmt);
4495 tree fn;
4496 if (!orig_fn)
4497 continue;
4498 if (TREE_CODE (orig_fn) == SSA_NAME)
4499 fn = VN_INFO (orig_fn)->valnum;
4500 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4501 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4502 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4503 else
4504 continue;
4505 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4506 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4507 TREE_TYPE (fn)))
4508 {
4509 bool can_make_abnormal_goto
4510 = stmt_can_make_abnormal_goto (stmt);
4511 bool was_noreturn = gimple_call_noreturn_p (stmt);
4512
4513 if (dump_file && (dump_flags & TDF_DETAILS))
4514 {
4515 fprintf (dump_file, "Replacing call target with ");
4516 print_generic_expr (dump_file, fn, 0);
4517 fprintf (dump_file, " in ");
4518 print_gimple_stmt (dump_file, stmt, 0, 0);
4519 }
4520
4521 gimple_call_set_fn (stmt, fn);
4522 VEC_safe_push (gimple, heap, to_update, stmt);
4523
4524 /* When changing a call into a noreturn call, cfg cleanup
4525 is needed to fix up the noreturn call. */
4526 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4527 todo |= TODO_cleanup_cfg;
4528
4529 /* If we removed EH side-effects from the statement, clean
4530 its EH information. */
4531 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4532 {
4533 bitmap_set_bit (need_eh_cleanup,
4534 gimple_bb (stmt)->index);
4535 if (dump_file && (dump_flags & TDF_DETAILS))
4536 fprintf (dump_file, " Removed EH side-effects.\n");
4537 }
4538
4539 /* Likewise for AB side-effects. */
4540 if (can_make_abnormal_goto
4541 && !stmt_can_make_abnormal_goto (stmt))
4542 {
4543 bitmap_set_bit (need_ab_cleanup,
4544 gimple_bb (stmt)->index);
4545 if (dump_file && (dump_flags & TDF_DETAILS))
4546 fprintf (dump_file, " Removed AB side-effects.\n");
4547 }
4548
4549 /* Changing an indirect call to a direct call may
4550 have exposed different semantics. This may
4551 require an SSA update. */
4552 todo |= TODO_update_ssa_only_virtuals;
4553 }
4554 }
4555 }
4556
4557 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4558 {
4559 gimple stmt, phi = gsi_stmt (gsi);
4560 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4561 pre_expr sprimeexpr, resexpr;
4562 gimple_stmt_iterator gsi2;
4563
4564 /* We want to perform redundant PHI elimination. Do so by
4565 replacing the PHI with a single copy if possible.
4566 Do not touch inserted, single-argument or virtual PHIs. */
4567 if (gimple_phi_num_args (phi) == 1
4568 || !is_gimple_reg (res))
4569 {
4570 gsi_next (&gsi);
4571 continue;
4572 }
4573
4574 resexpr = get_or_alloc_expr_for_name (res);
4575 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4576 get_expr_value_id (resexpr), NULL);
4577 if (sprimeexpr)
4578 {
4579 if (sprimeexpr->kind == CONSTANT)
4580 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4581 else if (sprimeexpr->kind == NAME)
4582 sprime = PRE_EXPR_NAME (sprimeexpr);
4583 else
4584 gcc_unreachable ();
4585 }
4586 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4587 {
4588 sprime = VN_INFO (res)->valnum;
4589 if (!useless_type_conversion_p (TREE_TYPE (res),
4590 TREE_TYPE (sprime)))
4591 sprime = fold_convert (TREE_TYPE (res), sprime);
4592 }
4593 if (!sprime
4594 || sprime == res)
4595 {
4596 gsi_next (&gsi);
4597 continue;
4598 }
4599
4600 if (dump_file && (dump_flags & TDF_DETAILS))
4601 {
4602 fprintf (dump_file, "Replaced redundant PHI node defining ");
4603 print_generic_expr (dump_file, res, 0);
4604 fprintf (dump_file, " with ");
4605 print_generic_expr (dump_file, sprime, 0);
4606 fprintf (dump_file, "\n");
4607 }
4608
4609 remove_phi_node (&gsi, false);
4610
4611 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4612 && TREE_CODE (sprime) == SSA_NAME)
4613 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4614
4615 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4616 sprime = fold_convert (TREE_TYPE (res), sprime);
4617 stmt = gimple_build_assign (res, sprime);
4618 SSA_NAME_DEF_STMT (res) = stmt;
4619 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4620
4621 gsi2 = gsi_after_labels (b);
4622 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4623 /* Queue the copy for eventual removal. */
4624 VEC_safe_push (gimple, heap, to_remove, stmt);
4625 /* If we inserted this PHI node ourself, it's not an elimination. */
4626 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4627 pre_stats.phis--;
4628 else
4629 pre_stats.eliminations++;
4630 }
4631 }
4632
4633 /* We cannot remove stmts during BB walk, especially not release SSA
4634 names there as this confuses the VN machinery. The stmts ending
4635 up in to_remove are either stores or simple copies. */
4636 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4637 {
4638 tree lhs = gimple_assign_lhs (stmt);
4639 tree rhs = gimple_assign_rhs1 (stmt);
4640 use_operand_p use_p;
4641 gimple use_stmt;
4642
4643 /* If there is a single use only, propagate the equivalency
4644 instead of keeping the copy. */
4645 if (TREE_CODE (lhs) == SSA_NAME
4646 && TREE_CODE (rhs) == SSA_NAME
4647 && single_imm_use (lhs, &use_p, &use_stmt)
4648 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4649 {
4650 SET_USE (use_p, rhs);
4651 update_stmt (use_stmt);
4652 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4653 && TREE_CODE (rhs) == SSA_NAME)
4654 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4655 }
4656
4657 /* If this is a store or a now unused copy, remove it. */
4658 if (TREE_CODE (lhs) != SSA_NAME
4659 || has_zero_uses (lhs))
4660 {
4661 basic_block bb = gimple_bb (stmt);
4662 gsi = gsi_for_stmt (stmt);
4663 unlink_stmt_vdef (stmt);
4664 if (gsi_remove (&gsi, true))
4665 bitmap_set_bit (need_eh_cleanup, bb->index);
4666 if (TREE_CODE (lhs) == SSA_NAME)
4667 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4668 release_defs (stmt);
4669 }
4670 }
4671 VEC_free (gimple, heap, to_remove);
4672
4673 /* We cannot update call statements with virtual operands during
4674 SSA walk. This might remove them which in turn makes our
4675 VN lattice invalid. */
4676 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4677 update_stmt (stmt);
4678 VEC_free (gimple, heap, to_update);
4679
4680 return todo;
4681 }
4682
4683 /* Borrow a bit of tree-ssa-dce.c for the moment.
4684 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4685 this may be a bit faster, and we may want critical edges kept split. */
4686
4687 /* If OP's defining statement has not already been determined to be necessary,
4688 mark that statement necessary. Return the stmt, if it is newly
4689 necessary. */
4690
4691 static inline gimple
4692 mark_operand_necessary (tree op)
4693 {
4694 gimple stmt;
4695
4696 gcc_assert (op);
4697
4698 if (TREE_CODE (op) != SSA_NAME)
4699 return NULL;
4700
4701 stmt = SSA_NAME_DEF_STMT (op);
4702 gcc_assert (stmt);
4703
4704 if (gimple_plf (stmt, NECESSARY)
4705 || gimple_nop_p (stmt))
4706 return NULL;
4707
4708 gimple_set_plf (stmt, NECESSARY, true);
4709 return stmt;
4710 }
4711
4712 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4713 to insert PHI nodes sometimes, and because value numbering of casts isn't
4714 perfect, we sometimes end up inserting dead code. This simple DCE-like
4715 pass removes any insertions we made that weren't actually used. */
4716
4717 static void
4718 remove_dead_inserted_code (void)
4719 {
4720 bitmap worklist;
4721 unsigned i;
4722 bitmap_iterator bi;
4723 gimple t;
4724
4725 worklist = BITMAP_ALLOC (NULL);
4726 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4727 {
4728 t = SSA_NAME_DEF_STMT (ssa_name (i));
4729 if (gimple_plf (t, NECESSARY))
4730 bitmap_set_bit (worklist, i);
4731 }
4732 while (!bitmap_empty_p (worklist))
4733 {
4734 i = bitmap_first_set_bit (worklist);
4735 bitmap_clear_bit (worklist, i);
4736 t = SSA_NAME_DEF_STMT (ssa_name (i));
4737
4738 /* PHI nodes are somewhat special in that each PHI alternative has
4739 data and control dependencies. All the statements feeding the
4740 PHI node's arguments are always necessary. */
4741 if (gimple_code (t) == GIMPLE_PHI)
4742 {
4743 unsigned k;
4744
4745 for (k = 0; k < gimple_phi_num_args (t); k++)
4746 {
4747 tree arg = PHI_ARG_DEF (t, k);
4748 if (TREE_CODE (arg) == SSA_NAME)
4749 {
4750 gimple n = mark_operand_necessary (arg);
4751 if (n)
4752 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4753 }
4754 }
4755 }
4756 else
4757 {
4758 /* Propagate through the operands. Examine all the USE, VUSE and
4759 VDEF operands in this statement. Mark all the statements
4760 which feed this statement's uses as necessary. */
4761 ssa_op_iter iter;
4762 tree use;
4763
4764 /* The operands of VDEF expressions are also needed as they
4765 represent potential definitions that may reach this
4766 statement (VDEF operands allow us to follow def-def
4767 links). */
4768
4769 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4770 {
4771 gimple n = mark_operand_necessary (use);
4772 if (n)
4773 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4774 }
4775 }
4776 }
4777
4778 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4779 {
4780 t = SSA_NAME_DEF_STMT (ssa_name (i));
4781 if (!gimple_plf (t, NECESSARY))
4782 {
4783 gimple_stmt_iterator gsi;
4784
4785 if (dump_file && (dump_flags & TDF_DETAILS))
4786 {
4787 fprintf (dump_file, "Removing unnecessary insertion:");
4788 print_gimple_stmt (dump_file, t, 0, 0);
4789 }
4790
4791 gsi = gsi_for_stmt (t);
4792 if (gimple_code (t) == GIMPLE_PHI)
4793 remove_phi_node (&gsi, true);
4794 else
4795 {
4796 gsi_remove (&gsi, true);
4797 release_defs (t);
4798 }
4799 }
4800 }
4801 BITMAP_FREE (worklist);
4802 }
4803
4804 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4805 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4806 the number of visited blocks. */
4807
4808 static int
4809 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4810 {
4811 edge_iterator *stack;
4812 int sp;
4813 int post_order_num = 0;
4814 sbitmap visited;
4815
4816 if (include_entry_exit)
4817 post_order[post_order_num++] = EXIT_BLOCK;
4818
4819 /* Allocate stack for back-tracking up CFG. */
4820 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4821 sp = 0;
4822
4823 /* Allocate bitmap to track nodes that have been visited. */
4824 visited = sbitmap_alloc (last_basic_block);
4825
4826 /* None of the nodes in the CFG have been visited yet. */
4827 sbitmap_zero (visited);
4828
4829 /* Push the last edge on to the stack. */
4830 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4831
4832 while (sp)
4833 {
4834 edge_iterator ei;
4835 basic_block src;
4836 basic_block dest;
4837
4838 /* Look at the edge on the top of the stack. */
4839 ei = stack[sp - 1];
4840 src = ei_edge (ei)->src;
4841 dest = ei_edge (ei)->dest;
4842
4843 /* Check if the edge destination has been visited yet. */
4844 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4845 {
4846 /* Mark that we have visited the destination. */
4847 SET_BIT (visited, src->index);
4848
4849 if (EDGE_COUNT (src->preds) > 0)
4850 /* Since the DEST node has been visited for the first
4851 time, check its successors. */
4852 stack[sp++] = ei_start (src->preds);
4853 else
4854 post_order[post_order_num++] = src->index;
4855 }
4856 else
4857 {
4858 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4859 post_order[post_order_num++] = dest->index;
4860
4861 if (!ei_one_before_end_p (ei))
4862 ei_next (&stack[sp - 1]);
4863 else
4864 sp--;
4865 }
4866 }
4867
4868 if (include_entry_exit)
4869 post_order[post_order_num++] = ENTRY_BLOCK;
4870
4871 free (stack);
4872 sbitmap_free (visited);
4873 return post_order_num;
4874 }
4875
4876
4877 /* Initialize data structures used by PRE. */
4878
4879 static void
4880 init_pre (bool do_fre)
4881 {
4882 basic_block bb;
4883
4884 next_expression_id = 1;
4885 expressions = NULL;
4886 VEC_safe_push (pre_expr, heap, expressions, NULL);
4887 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4888 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4889 get_max_value_id() + 1);
4890 name_to_id = NULL;
4891
4892 in_fre = do_fre;
4893
4894 inserted_exprs = BITMAP_ALLOC (NULL);
4895 need_creation = NULL;
4896 pretemp = NULL_TREE;
4897 storetemp = NULL_TREE;
4898 prephitemp = NULL_TREE;
4899
4900 connect_infinite_loops_to_exit ();
4901 memset (&pre_stats, 0, sizeof (pre_stats));
4902
4903
4904 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4905 my_rev_post_order_compute (postorder, false);
4906
4907 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4908
4909 calculate_dominance_info (CDI_POST_DOMINATORS);
4910 calculate_dominance_info (CDI_DOMINATORS);
4911
4912 bitmap_obstack_initialize (&grand_bitmap_obstack);
4913 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4914 expr_pred_trans_eq, free);
4915 expression_to_id = htab_create (num_ssa_names * 3,
4916 pre_expr_hash,
4917 pre_expr_eq, NULL);
4918 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4919 sizeof (struct bitmap_set), 30);
4920 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4921 sizeof (struct pre_expr_d), 30);
4922 FOR_ALL_BB (bb)
4923 {
4924 EXP_GEN (bb) = bitmap_set_new ();
4925 PHI_GEN (bb) = bitmap_set_new ();
4926 TMP_GEN (bb) = bitmap_set_new ();
4927 AVAIL_OUT (bb) = bitmap_set_new ();
4928 }
4929
4930 need_eh_cleanup = BITMAP_ALLOC (NULL);
4931 need_ab_cleanup = BITMAP_ALLOC (NULL);
4932 }
4933
4934
4935 /* Deallocate data structures used by PRE. */
4936
4937 static void
4938 fini_pre (bool do_fre)
4939 {
4940 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4941 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4942
4943 free (postorder);
4944 VEC_free (bitmap_set_t, heap, value_expressions);
4945 BITMAP_FREE (inserted_exprs);
4946 VEC_free (gimple, heap, need_creation);
4947 bitmap_obstack_release (&grand_bitmap_obstack);
4948 free_alloc_pool (bitmap_set_pool);
4949 free_alloc_pool (pre_expr_pool);
4950 htab_delete (phi_translate_table);
4951 htab_delete (expression_to_id);
4952 VEC_free (unsigned, heap, name_to_id);
4953
4954 free_aux_for_blocks ();
4955
4956 free_dominance_info (CDI_POST_DOMINATORS);
4957
4958 if (do_eh_cleanup)
4959 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4960
4961 if (do_ab_cleanup)
4962 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4963
4964 BITMAP_FREE (need_eh_cleanup);
4965 BITMAP_FREE (need_ab_cleanup);
4966
4967 if (do_eh_cleanup || do_ab_cleanup)
4968 cleanup_tree_cfg ();
4969
4970 if (!do_fre)
4971 loop_optimizer_finalize ();
4972 }
4973
4974 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4975 only wants to do full redundancy elimination. */
4976
4977 static unsigned int
4978 execute_pre (bool do_fre)
4979 {
4980 unsigned int todo = 0;
4981
4982 do_partial_partial =
4983 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4984
4985 /* This has to happen before SCCVN runs because
4986 loop_optimizer_init may create new phis, etc. */
4987 if (!do_fre)
4988 loop_optimizer_init (LOOPS_NORMAL);
4989
4990 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4991 {
4992 if (!do_fre)
4993 loop_optimizer_finalize ();
4994
4995 return 0;
4996 }
4997
4998 init_pre (do_fre);
4999 scev_initialize ();
5000
5001 /* Collect and value number expressions computed in each basic block. */
5002 compute_avail ();
5003
5004 if (dump_file && (dump_flags & TDF_DETAILS))
5005 {
5006 basic_block bb;
5007
5008 FOR_ALL_BB (bb)
5009 {
5010 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
5011 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
5012 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
5013 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
5014 }
5015 }
5016
5017 /* Insert can get quite slow on an incredibly large number of basic
5018 blocks due to some quadratic behavior. Until this behavior is
5019 fixed, don't run it when he have an incredibly large number of
5020 bb's. If we aren't going to run insert, there is no point in
5021 computing ANTIC, either, even though it's plenty fast. */
5022 if (!do_fre && n_basic_blocks < 4000)
5023 {
5024 compute_antic ();
5025 insert ();
5026 }
5027
5028 /* Make sure to remove fake edges before committing our inserts.
5029 This makes sure we don't end up with extra critical edges that
5030 we would need to split. */
5031 remove_fake_exit_edges ();
5032 gsi_commit_edge_inserts ();
5033
5034 /* Remove all the redundant expressions. */
5035 todo |= eliminate ();
5036
5037 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
5038 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
5039 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
5040 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
5041 statistics_counter_event (cfun, "Constified", pre_stats.constified);
5042
5043 clear_expression_ids ();
5044 if (!do_fre)
5045 {
5046 remove_dead_inserted_code ();
5047 todo |= TODO_verify_flow;
5048 }
5049
5050 scev_finalize ();
5051 fini_pre (do_fre);
5052
5053 if (!do_fre)
5054 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5055 case we can merge the block with the remaining predecessor of the block.
5056 It should either:
5057 - call merge_blocks after each tail merge iteration
5058 - call merge_blocks after all tail merge iterations
5059 - mark TODO_cleanup_cfg when necessary
5060 - share the cfg cleanup with fini_pre. */
5061 todo |= tail_merge_optimize (todo);
5062 free_scc_vn ();
5063
5064 return todo;
5065 }
5066
5067 /* Gate and execute functions for PRE. */
5068
5069 static unsigned int
5070 do_pre (void)
5071 {
5072 return execute_pre (false);
5073 }
5074
5075 static bool
5076 gate_pre (void)
5077 {
5078 return flag_tree_pre != 0;
5079 }
5080
5081 struct gimple_opt_pass pass_pre =
5082 {
5083 {
5084 GIMPLE_PASS,
5085 "pre", /* name */
5086 gate_pre, /* gate */
5087 do_pre, /* execute */
5088 NULL, /* sub */
5089 NULL, /* next */
5090 0, /* static_pass_number */
5091 TV_TREE_PRE, /* tv_id */
5092 PROP_no_crit_edges | PROP_cfg
5093 | PROP_ssa, /* properties_required */
5094 0, /* properties_provided */
5095 0, /* properties_destroyed */
5096 TODO_rebuild_alias, /* todo_flags_start */
5097 TODO_update_ssa_only_virtuals | TODO_ggc_collect
5098 | TODO_verify_ssa /* todo_flags_finish */
5099 }
5100 };
5101
5102
5103 /* Gate and execute functions for FRE. */
5104
5105 static unsigned int
5106 execute_fre (void)
5107 {
5108 return execute_pre (true);
5109 }
5110
5111 static bool
5112 gate_fre (void)
5113 {
5114 return flag_tree_fre != 0;
5115 }
5116
5117 struct gimple_opt_pass pass_fre =
5118 {
5119 {
5120 GIMPLE_PASS,
5121 "fre", /* name */
5122 gate_fre, /* gate */
5123 execute_fre, /* execute */
5124 NULL, /* sub */
5125 NULL, /* next */
5126 0, /* static_pass_number */
5127 TV_TREE_FRE, /* tv_id */
5128 PROP_cfg | PROP_ssa, /* properties_required */
5129 0, /* properties_provided */
5130 0, /* properties_destroyed */
5131 0, /* todo_flags_start */
5132 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */
5133 }
5134 };