* tree-ssa-pre.c: Fix a comment typo.
[gcc.git] / gcc / tree-ssa-pre.c
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 2, or (at your option)
11 any later version.
12
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to
20 the Free Software Foundation, 59 Temple Place - Suite 330,
21 Boston, MA 02111-1307, USA. */
22
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "errors.h"
28 #include "ggc.h"
29 #include "tree.h"
30 #include "basic-block.h"
31 #include "diagnostic.h"
32 #include "tree-inline.h"
33 #include "tree-flow.h"
34 #include "tree-gimple.h"
35 #include "tree-dump.h"
36 #include "timevar.h"
37 #include "fibheap.h"
38 #include "hashtab.h"
39 #include "tree-iterator.h"
40 #include "real.h"
41 #include "alloc-pool.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47
48 /* TODO:
49
50 1. Avail sets can be shared by making an avail_find_leader that
51 walks up the dominator tree and looks in those avail sets.
52 This might affect code optimality, it's unclear right now.
53 2. Load motion can be performed by value numbering the loads the
54 same as we do other expressions. This requires iterative
55 hashing the vuses into the values. Right now we simply assign
56 a new value every time we see a statement with a vuse.
57 3. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 4. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
61 */
62
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but MODIFY_EXPR, because MODIFY_EXPR's represent
65 the actual statement containing the expressions we care about, and
66 we cache the value number by putting it in the expression. */
67
68 /* Basic algorithm
69
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
79 expressions/values.
80
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem.An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
94
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
97
98 An expression is partially redundant (excluding partial
99 anticipation) if:
100
101 1. It is AVAIL in some, but not all, of the predecessors of a
102 given block.
103 2. It is ANTIC in all the predecessors.
104
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
107 insert/insert_aux performs this insertion.
108
109 Fourth, we eliminate fully redundant expressions.
110 This is a simple statement walk that replaces redundant
111 calculations with the now available values. */
112
113 /* Representations of value numbers:
114
115 Value numbers are represented using the "value handle" approach.
116 This means that each SSA_NAME (and for other reasons to be
117 disclosed in a moment, expression nodes) has a value handle that
118 can be retrieved through get_value_handle. This value handle, *is*
119 the value number of the SSA_NAME. You can pointer compare the
120 value handles for equivalence purposes.
121
122 For debugging reasons, the value handle is internally more than
123 just a number, it is a VAR_DECL named "value.x", where x is a
124 unique number for each value number in use. This allows
125 expressions with SSA_NAMES replaced by value handles to still be
126 pretty printed in a sane way. They simply print as "value.3 *
127 value.5", etc.
128
129 Expression nodes have value handles associated with them as a
130 cache. Otherwise, we'd have to look them up again in the hash
131 table This makes significant difference (factor of two or more) on
132 some test cases. They can be thrown away after the pass is
133 finished. */
134
135 /* Representation of expressions on value numbers:
136
137 In some portions of this code, you will notice we allocate "fake"
138 analogues to the expression we are value numbering, and replace the
139 operands with the values of the expression. Since we work on
140 values, and not just names, we canonicalize expressions to value
141 expressions for use in the ANTIC sets, the EXP_GEN set, etc.
142
143 This is theoretically unnecessary, it just saves a bunch of
144 repeated get_value_handle and find_leader calls in the remainder of
145 the code, trading off temporary memory usage for speed. The tree
146 nodes aren't actually creating more garbage, since they are
147 allocated in a special pools which are thrown away at the end of
148 this pass.
149
150 All of this also means that if you print the EXP_GEN or ANTIC sets,
151 you will see "value.5 + value.7" in the set, instead of "a_55 +
152 b_66" or something. The only thing that actually cares about
153 seeing the value leaders is phi translation, and it needs to be
154 able to find the leader for a value in an arbitrary block, so this
155 "value expression" form is perfect for it (otherwise you'd do
156 get_value_handle->find_leader->translate->get_value_handle->find_leader).*/
157
158
159 /* Representation of sets:
160
161 There are currently two types of sets used, hopefully to be unified soon.
162 The AVAIL sets do not need to be sorted in any particular order,
163 and thus, are simply represented as two bitmaps, one that keeps
164 track of values present in the set, and one that keeps track of
165 expressions present in the set.
166
167 The other sets are represented as doubly linked lists kept in topological
168 order, with an optional supporting bitmap of values present in the
169 set. The sets represent values, and the elements can be values or
170 expressions. The elements can appear in different sets, but each
171 element can only appear once in each set.
172
173 Since each node in the set represents a value, we also want to be
174 able to map expression, set pairs to something that tells us
175 whether the value is present is a set. We use a per-set bitmap for
176 that. The value handles also point to a linked list of the
177 expressions they represent via a tree annotation. This is mainly
178 useful only for debugging, since we don't do identity lookups. */
179
180
181 /* A value set element. Basically a single linked list of
182 expressions/values. */
183 typedef struct value_set_node
184 {
185 /* An expression. */
186 tree expr;
187
188 /* A pointer to the next element of the value set. */
189 struct value_set_node *next;
190 } *value_set_node_t;
191
192
193 /* A value set. This is a singly linked list of value_set_node
194 elements with a possible bitmap that tells us what values exist in
195 the set. This set must be kept in topologically sorted order. */
196 typedef struct value_set
197 {
198 /* The head of the list. Used for iterating over the list in
199 order. */
200 value_set_node_t head;
201
202 /* The tail of the list. Used for tail insertions, which are
203 necessary to keep the set in topologically sorted order because
204 of how the set is built. */
205 value_set_node_t tail;
206
207 /* The length of the list. */
208 size_t length;
209
210 /* True if the set is indexed, which means it contains a backing
211 bitmap for quick determination of whether certain values exist in the
212 set. */
213 bool indexed;
214
215 /* The bitmap of values that exist in the set. May be NULL in an
216 empty or non-indexed set. */
217 bitmap values;
218
219 } *value_set_t;
220
221
222 /* An unordered bitmap set. One bitmap tracks values, the other,
223 expressions. */
224 typedef struct bitmap_set
225 {
226 bitmap expressions;
227 bitmap values;
228 } *bitmap_set_t;
229
230 /* Sets that we need to keep track of. */
231 typedef struct bb_value_sets
232 {
233 /* The EXP_GEN set, which represents expressions/values generated in
234 a basic block. */
235 value_set_t exp_gen;
236
237 /* The PHI_GEN set, which represents PHI results generated in a
238 basic block. */
239 bitmap_set_t phi_gen;
240
241 /* The TMP_GEN set, which represents results/temporaries generated
242 in a basic block. IE the LHS of an expression. */
243 bitmap_set_t tmp_gen;
244
245 /* The AVAIL_OUT set, which represents which values are available in
246 a given basic block. */
247 bitmap_set_t avail_out;
248
249 /* The ANTIC_IN set, which represents which values are anticiptable
250 in a given basic block. */
251 value_set_t antic_in;
252
253 /* The NEW_SETS set, which is used during insertion to augment the
254 AVAIL_OUT set of blocks with the new insertions performed during
255 the current iteration. */
256 bitmap_set_t new_sets;
257 } *bb_value_sets_t;
258
259 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
260 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
261 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
262 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
263 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
264 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
265
266 /* This structure is used to keep track of statistics on what
267 optimization PRE was able to perform. */
268 static struct
269 {
270 /* The number of RHS computations eliminated by PRE. */
271 int eliminations;
272
273 /* The number of new expressions/temporaries generated by PRE. */
274 int insertions;
275
276 /* The number of new PHI nodes added by PRE. */
277 int phis;
278
279 /* The number of values found constant. */
280 int constified;
281
282 } pre_stats;
283
284
285 static tree bitmap_find_leader (bitmap_set_t, tree);
286 static tree find_leader (value_set_t, tree);
287 static void value_insert_into_set (value_set_t, tree);
288 static void bitmap_value_insert_into_set (bitmap_set_t, tree);
289 static void bitmap_value_replace_in_set (bitmap_set_t, tree);
290 static void insert_into_set (value_set_t, tree);
291 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
292 static bool bitmap_set_contains_value (bitmap_set_t, tree);
293 static bitmap_set_t bitmap_set_new (void);
294 static value_set_t set_new (bool);
295 static bool is_undefined_value (tree);
296 static tree create_expression_by_pieces (basic_block, tree, tree);
297
298
299 /* We can add and remove elements and entries to and from sets
300 and hash tables, so we use alloc pools for them. */
301
302 static alloc_pool value_set_pool;
303 static alloc_pool bitmap_set_pool;
304 static alloc_pool value_set_node_pool;
305 static alloc_pool binary_node_pool;
306 static alloc_pool unary_node_pool;
307 static alloc_pool reference_node_pool;
308 static alloc_pool comparison_node_pool;
309 static alloc_pool expression_node_pool;
310 static alloc_pool list_node_pool;
311 static bitmap_obstack grand_bitmap_obstack;
312
313 /* Set of blocks with statements that have had its EH information
314 cleaned up. */
315 static bitmap need_eh_cleanup;
316
317 /* The phi_translate_table caches phi translations for a given
318 expression and predecessor. */
319
320 static htab_t phi_translate_table;
321
322 /* A three tuple {e, pred, v} used to cache phi translations in the
323 phi_translate_table. */
324
325 typedef struct expr_pred_trans_d
326 {
327 /* The expression. */
328 tree e;
329
330 /* The predecessor block along which we translated the expression. */
331 basic_block pred;
332
333 /* The value that resulted from the translation. */
334 tree v;
335
336 /* The hashcode for the expression, pred pair. This is cached for
337 speed reasons. */
338 hashval_t hashcode;
339 } *expr_pred_trans_t;
340
341 /* Return the hash value for a phi translation table entry. */
342
343 static hashval_t
344 expr_pred_trans_hash (const void *p)
345 {
346 const expr_pred_trans_t ve = (expr_pred_trans_t) p;
347 return ve->hashcode;
348 }
349
350 /* Return true if two phi translation table entries are the same.
351 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
352
353 static int
354 expr_pred_trans_eq (const void *p1, const void *p2)
355 {
356 const expr_pred_trans_t ve1 = (expr_pred_trans_t) p1;
357 const expr_pred_trans_t ve2 = (expr_pred_trans_t) p2;
358 basic_block b1 = ve1->pred;
359 basic_block b2 = ve2->pred;
360
361
362 /* If they are not translations for the same basic block, they can't
363 be equal. */
364 if (b1 != b2)
365 return false;
366
367 /* If they are for the same basic block, determine if the
368 expressions are equal. */
369 if (expressions_equal_p (ve1->e, ve2->e))
370 return true;
371
372 return false;
373 }
374
375 /* Search in the phi translation table for the translation of
376 expression E in basic block PRED. Return the translated value, if
377 found, NULL otherwise. */
378
379 static inline tree
380 phi_trans_lookup (tree e, basic_block pred)
381 {
382 void **slot;
383 struct expr_pred_trans_d ept;
384 ept.e = e;
385 ept.pred = pred;
386 ept.hashcode = vn_compute (e, (unsigned long) pred, NULL);
387 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
388 NO_INSERT);
389 if (!slot)
390 return NULL;
391 else
392 return ((expr_pred_trans_t) *slot)->v;
393 }
394
395
396 /* Add the tuple mapping from {expression E, basic block PRED} to
397 value V, to the phi translation table. */
398
399 static inline void
400 phi_trans_add (tree e, tree v, basic_block pred)
401 {
402 void **slot;
403 expr_pred_trans_t new_pair = xmalloc (sizeof (*new_pair));
404 new_pair->e = e;
405 new_pair->pred = pred;
406 new_pair->v = v;
407 new_pair->hashcode = vn_compute (e, (unsigned long) pred, NULL);
408 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
409 new_pair->hashcode, INSERT);
410 if (*slot)
411 free (*slot);
412 *slot = (void *) new_pair;
413 }
414
415
416 /* Add expression E to the expression set of value V. */
417
418 void
419 add_to_value (tree v, tree e)
420 {
421 /* Constants have no expression sets. */
422 if (is_gimple_min_invariant (v))
423 return;
424
425 if (VALUE_HANDLE_EXPR_SET (v) == NULL)
426 VALUE_HANDLE_EXPR_SET (v) = set_new (false);
427
428 insert_into_set (VALUE_HANDLE_EXPR_SET (v), e);
429 }
430
431
432 /* Return true if value V exists in the bitmap for SET. */
433
434 static inline bool
435 value_exists_in_set_bitmap (value_set_t set, tree v)
436 {
437 if (!set->values)
438 return false;
439
440 return bitmap_bit_p (set->values, VALUE_HANDLE_ID (v));
441 }
442
443
444 /* Remove value V from the bitmap for SET. */
445
446 static void
447 value_remove_from_set_bitmap (value_set_t set, tree v)
448 {
449 gcc_assert (set->indexed);
450
451 if (!set->values)
452 return;
453
454 bitmap_clear_bit (set->values, VALUE_HANDLE_ID (v));
455 }
456
457
458 /* Insert the value number V into the bitmap of values existing in
459 SET. */
460
461 static inline void
462 value_insert_into_set_bitmap (value_set_t set, tree v)
463 {
464 gcc_assert (set->indexed);
465
466 if (set->values == NULL)
467 set->values = BITMAP_ALLOC (&grand_bitmap_obstack);
468
469 bitmap_set_bit (set->values, VALUE_HANDLE_ID (v));
470 }
471
472
473 /* Create a new bitmap set and return it. */
474
475 static bitmap_set_t
476 bitmap_set_new (void)
477 {
478 bitmap_set_t ret = pool_alloc (bitmap_set_pool);
479 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
480 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
481 return ret;
482 }
483
484 /* Create a new set. */
485
486 static value_set_t
487 set_new (bool indexed)
488 {
489 value_set_t ret;
490 ret = pool_alloc (value_set_pool);
491 ret->head = ret->tail = NULL;
492 ret->length = 0;
493 ret->indexed = indexed;
494 ret->values = NULL;
495 return ret;
496 }
497
498 /* Insert an expression EXPR into a bitmapped set. */
499
500 static void
501 bitmap_insert_into_set (bitmap_set_t set, tree expr)
502 {
503 tree val;
504 /* XXX: For now, we only let SSA_NAMES into the bitmap sets. */
505 gcc_assert (TREE_CODE (expr) == SSA_NAME);
506 val = get_value_handle (expr);
507
508 gcc_assert (val);
509 if (!is_gimple_min_invariant (val))
510 {
511 bitmap_set_bit (set->values, VALUE_HANDLE_ID (val));
512 bitmap_set_bit (set->expressions, SSA_NAME_VERSION (expr));
513 }
514 }
515
516 /* Insert EXPR into SET. */
517
518 static void
519 insert_into_set (value_set_t set, tree expr)
520 {
521 value_set_node_t newnode = pool_alloc (value_set_node_pool);
522 tree val = get_value_handle (expr);
523 gcc_assert (val);
524
525 if (is_gimple_min_invariant (val))
526 return;
527
528 /* For indexed sets, insert the value into the set value bitmap.
529 For all sets, add it to the linked list and increment the list
530 length. */
531 if (set->indexed)
532 value_insert_into_set_bitmap (set, val);
533
534 newnode->next = NULL;
535 newnode->expr = expr;
536 set->length ++;
537 if (set->head == NULL)
538 {
539 set->head = set->tail = newnode;
540 }
541 else
542 {
543 set->tail->next = newnode;
544 set->tail = newnode;
545 }
546 }
547
548 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
549
550 static void
551 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
552 {
553 bitmap_copy (dest->expressions, orig->expressions);
554 bitmap_copy (dest->values, orig->values);
555 }
556
557 /* Copy the set ORIG to the set DEST. */
558
559 static void
560 set_copy (value_set_t dest, value_set_t orig)
561 {
562 value_set_node_t node;
563
564 if (!orig || !orig->head)
565 return;
566
567 for (node = orig->head;
568 node;
569 node = node->next)
570 {
571 insert_into_set (dest, node->expr);
572 }
573 }
574
575 /* Remove EXPR from SET. */
576
577 static void
578 set_remove (value_set_t set, tree expr)
579 {
580 value_set_node_t node, prev;
581
582 /* Remove the value of EXPR from the bitmap, decrement the set
583 length, and remove it from the actual double linked list. */
584 value_remove_from_set_bitmap (set, get_value_handle (expr));
585 set->length--;
586 prev = NULL;
587 for (node = set->head;
588 node != NULL;
589 prev = node, node = node->next)
590 {
591 if (node->expr == expr)
592 {
593 if (prev == NULL)
594 set->head = node->next;
595 else
596 prev->next= node->next;
597
598 if (node == set->tail)
599 set->tail = prev;
600 pool_free (value_set_node_pool, node);
601 return;
602 }
603 }
604 }
605
606 /* Return true if SET contains the value VAL. */
607
608 static bool
609 set_contains_value (value_set_t set, tree val)
610 {
611 /* All constants are in every set. */
612 if (is_gimple_min_invariant (val))
613 return true;
614
615 if (set->length == 0)
616 return false;
617
618 return value_exists_in_set_bitmap (set, val);
619 }
620
621 /* Return true if bitmapped set SET contains the expression EXPR. */
622 static bool
623 bitmap_set_contains (bitmap_set_t set, tree expr)
624 {
625 /* All constants are in every set. */
626 if (is_gimple_min_invariant (get_value_handle (expr)))
627 return true;
628
629 /* XXX: Bitmapped sets only contain SSA_NAME's for now. */
630 if (TREE_CODE (expr) != SSA_NAME)
631 return false;
632 return bitmap_bit_p (set->expressions, SSA_NAME_VERSION (expr));
633 }
634
635
636 /* Return true if bitmapped set SET contains the value VAL. */
637
638 static bool
639 bitmap_set_contains_value (bitmap_set_t set, tree val)
640 {
641 if (is_gimple_min_invariant (val))
642 return true;
643 return bitmap_bit_p (set->values, VALUE_HANDLE_ID (val));
644 }
645
646 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
647
648 static void
649 bitmap_set_replace_value (bitmap_set_t set, tree lookfor, tree expr)
650 {
651 value_set_t exprset;
652 value_set_node_t node;
653 if (is_gimple_min_invariant (lookfor))
654 return;
655 if (!bitmap_set_contains_value (set, lookfor))
656 return;
657
658 /* The number of expressions having a given value is usually
659 significantly less than the total number of expressions in SET.
660 Thus, rather than check, for each expression in SET, whether it
661 has the value LOOKFOR, we walk the reverse mapping that tells us
662 what expressions have a given value, and see if any of those
663 expressions are in our set. For large testcases, this is about
664 5-10x faster than walking the bitmap. If this is somehow a
665 significant lose for some cases, we can choose which set to walk
666 based on the set size. */
667 exprset = VALUE_HANDLE_EXPR_SET (lookfor);
668 for (node = exprset->head; node; node = node->next)
669 {
670 if (TREE_CODE (node->expr) == SSA_NAME)
671 {
672 if (bitmap_bit_p (set->expressions, SSA_NAME_VERSION (node->expr)))
673 {
674 bitmap_clear_bit (set->expressions, SSA_NAME_VERSION (node->expr));
675 bitmap_set_bit (set->expressions, SSA_NAME_VERSION (expr));
676 return;
677 }
678 }
679 }
680 }
681
682 /* Subtract bitmapped set B from value set A, and return the new set. */
683
684 static value_set_t
685 bitmap_set_subtract_from_value_set (value_set_t a, bitmap_set_t b,
686 bool indexed)
687 {
688 value_set_t ret = set_new (indexed);
689 value_set_node_t node;
690 for (node = a->head;
691 node;
692 node = node->next)
693 {
694 if (!bitmap_set_contains (b, node->expr))
695 insert_into_set (ret, node->expr);
696 }
697 return ret;
698 }
699
700 /* Return true if two sets are equal. */
701
702 static bool
703 set_equal (value_set_t a, value_set_t b)
704 {
705 value_set_node_t node;
706
707 if (a->length != b->length)
708 return false;
709 for (node = a->head;
710 node;
711 node = node->next)
712 {
713 if (!set_contains_value (b, get_value_handle (node->expr)))
714 return false;
715 }
716 return true;
717 }
718
719 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
720 and add it otherwise. */
721
722 static void
723 bitmap_value_replace_in_set (bitmap_set_t set, tree expr)
724 {
725 tree val = get_value_handle (expr);
726 if (bitmap_set_contains_value (set, val))
727 bitmap_set_replace_value (set, val, expr);
728 else
729 bitmap_insert_into_set (set, expr);
730 }
731
732 /* Insert EXPR into SET if EXPR's value is not already present in
733 SET. */
734
735 static void
736 bitmap_value_insert_into_set (bitmap_set_t set, tree expr)
737 {
738 tree val = get_value_handle (expr);
739
740 if (is_gimple_min_invariant (val))
741 return;
742
743 if (!bitmap_set_contains_value (set, val))
744 bitmap_insert_into_set (set, expr);
745 }
746
747 /* Insert the value for EXPR into SET, if it doesn't exist already. */
748
749 static void
750 value_insert_into_set (value_set_t set, tree expr)
751 {
752 tree val = get_value_handle (expr);
753
754 /* Constant and invariant values exist everywhere, and thus,
755 actually keeping them in the sets is pointless. */
756 if (is_gimple_min_invariant (val))
757 return;
758
759 if (!set_contains_value (set, val))
760 insert_into_set (set, expr);
761 }
762
763
764 /* Print out SET to OUTFILE. */
765
766 static void
767 bitmap_print_value_set (FILE *outfile, bitmap_set_t set,
768 const char *setname, int blockindex)
769 {
770 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
771 if (set)
772 {
773 bool first = true;
774 unsigned i;
775 bitmap_iterator bi;
776
777 EXECUTE_IF_SET_IN_BITMAP (set->expressions, 0, i, bi)
778 {
779 if (!first)
780 fprintf (outfile, ", ");
781 first = false;
782 print_generic_expr (outfile, ssa_name (i), 0);
783
784 fprintf (outfile, " (");
785 print_generic_expr (outfile, get_value_handle (ssa_name (i)), 0);
786 fprintf (outfile, ") ");
787 }
788 }
789 fprintf (outfile, " }\n");
790 }
791 /* Print out the value_set SET to OUTFILE. */
792
793 static void
794 print_value_set (FILE *outfile, value_set_t set,
795 const char *setname, int blockindex)
796 {
797 value_set_node_t node;
798 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
799 if (set)
800 {
801 for (node = set->head;
802 node;
803 node = node->next)
804 {
805 print_generic_expr (outfile, node->expr, 0);
806
807 fprintf (outfile, " (");
808 print_generic_expr (outfile, get_value_handle (node->expr), 0);
809 fprintf (outfile, ") ");
810
811 if (node->next)
812 fprintf (outfile, ", ");
813 }
814 }
815
816 fprintf (outfile, " }\n");
817 }
818
819 /* Print out the expressions that have VAL to OUTFILE. */
820
821 void
822 print_value_expressions (FILE *outfile, tree val)
823 {
824 if (VALUE_HANDLE_EXPR_SET (val))
825 {
826 char s[10];
827 sprintf (s, "VH.%04d", VALUE_HANDLE_ID (val));
828 print_value_set (outfile, VALUE_HANDLE_EXPR_SET (val), s, 0);
829 }
830 }
831
832
833 void
834 debug_value_expressions (tree val)
835 {
836 print_value_expressions (stderr, val);
837 }
838
839
840 void debug_value_set (value_set_t, const char *, int);
841
842 void
843 debug_value_set (value_set_t set, const char *setname, int blockindex)
844 {
845 print_value_set (stderr, set, setname, blockindex);
846 }
847
848 /* Return the folded version of T if T, when folded, is a gimple
849 min_invariant. Otherwise, return T. */
850
851 static tree
852 fully_constant_expression (tree t)
853 {
854 tree folded;
855 folded = fold (t);
856 if (folded && is_gimple_min_invariant (folded))
857 return folded;
858 return t;
859 }
860
861 /* Return a copy of a chain of nodes, chained through the TREE_CHAIN field.
862 For example, this can copy a list made of TREE_LIST nodes.
863 Allocates the nodes in list_node_pool*/
864
865 static tree
866 pool_copy_list (tree list)
867 {
868 tree head;
869 tree prev, next;
870
871 if (list == 0)
872 return 0;
873 head = pool_alloc (list_node_pool);
874
875 memcpy (head, list, tree_size (list));
876 prev = head;
877
878 next = TREE_CHAIN (list);
879 while (next)
880 {
881 TREE_CHAIN (prev) = pool_alloc (list_node_pool);
882 memcpy (TREE_CHAIN (prev), next, tree_size (next));
883 prev = TREE_CHAIN (prev);
884 next = TREE_CHAIN (next);
885 }
886 return head;
887 }
888
889
890 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
891 the phis in PRED. Return NULL if we can't find a leader for each
892 part of the translated expression. */
893
894 static tree
895 phi_translate (tree expr, value_set_t set, basic_block pred,
896 basic_block phiblock)
897 {
898 tree phitrans = NULL;
899 tree oldexpr = expr;
900
901 if (expr == NULL)
902 return NULL;
903
904 if (is_gimple_min_invariant (expr))
905 return expr;
906
907 /* Phi translations of a given expression don't change. */
908 phitrans = phi_trans_lookup (expr, pred);
909 if (phitrans)
910 return phitrans;
911
912 switch (TREE_CODE_CLASS (TREE_CODE (expr)))
913 {
914 case tcc_expression:
915 {
916 if (TREE_CODE (expr) != CALL_EXPR)
917 return NULL;
918 else
919 {
920 tree oldop0 = TREE_OPERAND (expr, 0);
921 tree oldarglist = TREE_OPERAND (expr, 1);
922 tree oldop2 = TREE_OPERAND (expr, 2);
923 tree newop0;
924 tree newarglist;
925 tree newop2 = NULL;
926 tree oldwalker;
927 tree newwalker;
928 tree newexpr;
929 bool listchanged = false;
930
931 /* Call expressions are kind of weird because they have an
932 argument list. We don't want to value number the list
933 as one value number, because that doesn't make much
934 sense, and just breaks the support functions we call,
935 which expect TREE_OPERAND (call_expr, 2) to be a
936 TREE_LIST. */
937
938 newop0 = phi_translate (find_leader (set, oldop0),
939 set, pred, phiblock);
940 if (newop0 == NULL)
941 return NULL;
942 if (oldop2)
943 {
944 newop2 = phi_translate (find_leader (set, oldop2),
945 set, pred, phiblock);
946 if (newop2 == NULL)
947 return NULL;
948 }
949
950 /* phi translate the argument list piece by piece.
951
952 We could actually build the list piece by piece here,
953 but it's likely to not be worth the memory we will save,
954 unless you have millions of call arguments. */
955
956 newarglist = pool_copy_list (oldarglist);
957 for (oldwalker = oldarglist, newwalker = newarglist;
958 oldwalker && newwalker;
959 oldwalker = TREE_CHAIN (oldwalker),
960 newwalker = TREE_CHAIN (newwalker))
961 {
962
963 tree oldval = TREE_VALUE (oldwalker);
964 tree newval;
965 if (oldval)
966 {
967 newval = phi_translate (find_leader (set, oldval),
968 set, pred, phiblock);
969 if (newval == NULL)
970 return NULL;
971 if (newval != oldval)
972 {
973 listchanged = true;
974 TREE_VALUE (newwalker) = get_value_handle (newval);
975 }
976 }
977 }
978 if (listchanged)
979 vn_lookup_or_add (newarglist, NULL);
980
981 if (listchanged || (newop0 != oldop0) || (oldop2 != newop2))
982 {
983 newexpr = pool_alloc (expression_node_pool);
984 memcpy (newexpr, expr, tree_size (expr));
985 TREE_OPERAND (newexpr, 0) = newop0 == oldop0 ? oldop0 : get_value_handle (newop0);
986 TREE_OPERAND (newexpr, 1) = listchanged ? newarglist : oldarglist;
987 TREE_OPERAND (newexpr, 2) = newop2 == oldop2 ? oldop2 : get_value_handle (newop2);
988 create_tree_ann (newexpr);
989 vn_lookup_or_add (newexpr, NULL);
990 expr = newexpr;
991 phi_trans_add (oldexpr, newexpr, pred);
992 }
993 }
994 }
995 return expr;
996
997 case tcc_reference:
998 /* XXX: Until we have PRE of loads working, none will be ANTIC. */
999 return NULL;
1000
1001 case tcc_binary:
1002 case tcc_comparison:
1003 {
1004 tree oldop1 = TREE_OPERAND (expr, 0);
1005 tree oldop2 = TREE_OPERAND (expr, 1);
1006 tree newop1;
1007 tree newop2;
1008 tree newexpr;
1009
1010 newop1 = phi_translate (find_leader (set, oldop1),
1011 set, pred, phiblock);
1012 if (newop1 == NULL)
1013 return NULL;
1014 newop2 = phi_translate (find_leader (set, oldop2),
1015 set, pred, phiblock);
1016 if (newop2 == NULL)
1017 return NULL;
1018 if (newop1 != oldop1 || newop2 != oldop2)
1019 {
1020 tree t;
1021 newexpr = pool_alloc (binary_node_pool);
1022 memcpy (newexpr, expr, tree_size (expr));
1023 TREE_OPERAND (newexpr, 0) = newop1 == oldop1 ? oldop1 : get_value_handle (newop1);
1024 TREE_OPERAND (newexpr, 1) = newop2 == oldop2 ? oldop2 : get_value_handle (newop2);
1025 t = fully_constant_expression (newexpr);
1026 if (t != newexpr)
1027 {
1028 pool_free (binary_node_pool, newexpr);
1029 newexpr = t;
1030 }
1031 else
1032 {
1033 create_tree_ann (newexpr);
1034 vn_lookup_or_add (newexpr, NULL);
1035 }
1036 expr = newexpr;
1037 phi_trans_add (oldexpr, newexpr, pred);
1038 }
1039 }
1040 return expr;
1041
1042 case tcc_unary:
1043 {
1044 tree oldop1 = TREE_OPERAND (expr, 0);
1045 tree newop1;
1046 tree newexpr;
1047
1048 newop1 = phi_translate (find_leader (set, oldop1),
1049 set, pred, phiblock);
1050 if (newop1 == NULL)
1051 return NULL;
1052 if (newop1 != oldop1)
1053 {
1054 tree t;
1055 newexpr = pool_alloc (unary_node_pool);
1056 memcpy (newexpr, expr, tree_size (expr));
1057 TREE_OPERAND (newexpr, 0) = get_value_handle (newop1);
1058 t = fully_constant_expression (newexpr);
1059 if (t != newexpr)
1060 {
1061 pool_free (unary_node_pool, newexpr);
1062 newexpr = t;
1063 }
1064 else
1065 {
1066 create_tree_ann (newexpr);
1067 vn_lookup_or_add (newexpr, NULL);
1068 }
1069 expr = newexpr;
1070 phi_trans_add (oldexpr, newexpr, pred);
1071 }
1072 }
1073 return expr;
1074
1075 case tcc_exceptional:
1076 {
1077 tree phi = NULL;
1078 edge e;
1079 gcc_assert (TREE_CODE (expr) == SSA_NAME);
1080 if (TREE_CODE (SSA_NAME_DEF_STMT (expr)) == PHI_NODE)
1081 phi = SSA_NAME_DEF_STMT (expr);
1082 else
1083 return expr;
1084
1085 e = find_edge (pred, bb_for_stmt (phi));
1086 if (e)
1087 {
1088 if (is_undefined_value (PHI_ARG_DEF (phi, e->dest_idx)))
1089 return NULL;
1090 vn_lookup_or_add (PHI_ARG_DEF (phi, e->dest_idx), NULL);
1091 return PHI_ARG_DEF (phi, e->dest_idx);
1092 }
1093 }
1094 return expr;
1095
1096 default:
1097 gcc_unreachable ();
1098 }
1099 }
1100
1101 /* For each expression in SET, translate the value handles through phi nodes
1102 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1103 expressions in DEST. */
1104
1105 static void
1106 phi_translate_set (value_set_t dest, value_set_t set, basic_block pred,
1107 basic_block phiblock)
1108 {
1109 value_set_node_t node;
1110 for (node = set->head;
1111 node;
1112 node = node->next)
1113 {
1114 tree translated;
1115 translated = phi_translate (node->expr, set, pred, phiblock);
1116 phi_trans_add (node->expr, translated, pred);
1117
1118 if (translated != NULL)
1119 value_insert_into_set (dest, translated);
1120 }
1121 }
1122
1123 /* Find the leader for a value (i.e., the name representing that
1124 value) in a given set, and return it. Return NULL if no leader is
1125 found. */
1126
1127 static tree
1128 bitmap_find_leader (bitmap_set_t set, tree val)
1129 {
1130 if (val == NULL)
1131 return NULL;
1132
1133 if (is_gimple_min_invariant (val))
1134 return val;
1135 if (bitmap_set_contains_value (set, val))
1136 {
1137 /* Rather than walk the entire bitmap of expressions, and see
1138 whether any of them has the value we are looking for, we look
1139 at the reverse mapping, which tells us the set of expressions
1140 that have a given value (IE value->expressions with that
1141 value) and see if any of those expressions are in our set.
1142 The number of expressions per value is usually significantly
1143 less than the number of expressions in the set. In fact, for
1144 large testcases, doing it this way is roughly 5-10x faster
1145 than walking the bitmap.
1146 If this is somehow a significant lose for some cases, we can
1147 choose which set to walk based on which set is smaller. */
1148 value_set_t exprset;
1149 value_set_node_t node;
1150 exprset = VALUE_HANDLE_EXPR_SET (val);
1151 for (node = exprset->head; node; node = node->next)
1152 {
1153 if (TREE_CODE (node->expr) == SSA_NAME)
1154 {
1155 if (bitmap_bit_p (set->expressions,
1156 SSA_NAME_VERSION (node->expr)))
1157 return node->expr;
1158 }
1159 }
1160 }
1161 return NULL;
1162 }
1163
1164
1165 /* Find the leader for a value (i.e., the name representing that
1166 value) in a given set, and return it. Return NULL if no leader is
1167 found. */
1168
1169 static tree
1170 find_leader (value_set_t set, tree val)
1171 {
1172 value_set_node_t node;
1173
1174 if (val == NULL)
1175 return NULL;
1176
1177 /* Constants represent themselves. */
1178 if (is_gimple_min_invariant (val))
1179 return val;
1180
1181 if (set->length == 0)
1182 return NULL;
1183
1184 if (value_exists_in_set_bitmap (set, val))
1185 {
1186 for (node = set->head;
1187 node;
1188 node = node->next)
1189 {
1190 if (get_value_handle (node->expr) == val)
1191 return node->expr;
1192 }
1193 }
1194
1195 return NULL;
1196 }
1197
1198 /* Determine if the expression EXPR is valid in SET. This means that
1199 we have a leader for each part of the expression (if it consists of
1200 values), or the expression is an SSA_NAME.
1201
1202 NB: We never should run into a case where we have SSA_NAME +
1203 SSA_NAME or SSA_NAME + value. The sets valid_in_set is called on,
1204 the ANTIC sets, will only ever have SSA_NAME's or value expressions
1205 (IE VALUE1 + VALUE2, *VALUE1, VALUE1 < VALUE2) */
1206
1207 static bool
1208 valid_in_set (value_set_t set, tree expr)
1209 {
1210 switch (TREE_CODE_CLASS (TREE_CODE (expr)))
1211 {
1212 case tcc_binary:
1213 case tcc_comparison:
1214 {
1215 tree op1 = TREE_OPERAND (expr, 0);
1216 tree op2 = TREE_OPERAND (expr, 1);
1217 return set_contains_value (set, op1) && set_contains_value (set, op2);
1218 }
1219
1220 case tcc_unary:
1221 {
1222 tree op1 = TREE_OPERAND (expr, 0);
1223 return set_contains_value (set, op1);
1224 }
1225
1226 case tcc_expression:
1227 {
1228 if (TREE_CODE (expr) == CALL_EXPR)
1229 {
1230 tree op0 = TREE_OPERAND (expr, 0);
1231 tree arglist = TREE_OPERAND (expr, 1);
1232 tree op2 = TREE_OPERAND (expr, 2);
1233
1234 /* Check the non-list operands first. */
1235 if (!set_contains_value (set, op0)
1236 || (op2 && !set_contains_value (set, op2)))
1237 return false;
1238
1239 /* Now check the operands. */
1240 for (; arglist; arglist = TREE_CHAIN (arglist))
1241 {
1242 if (!set_contains_value (set, TREE_VALUE (arglist)))
1243 return false;
1244 }
1245 return true;
1246 }
1247 return false;
1248 }
1249
1250 case tcc_reference:
1251 /* XXX: Until PRE of loads works, no reference nodes are ANTIC. */
1252 return false;
1253
1254 case tcc_exceptional:
1255 gcc_assert (TREE_CODE (expr) == SSA_NAME);
1256 return true;
1257
1258 case tcc_declaration:
1259 /* VAR_DECL and PARM_DECL are never anticipatable. */
1260 return false;
1261
1262 default:
1263 /* No other cases should be encountered. */
1264 gcc_unreachable ();
1265 }
1266 }
1267
1268 /* Clean the set of expressions that are no longer valid in SET. This
1269 means expressions that are made up of values we have no leaders for
1270 in SET. */
1271
1272 static void
1273 clean (value_set_t set)
1274 {
1275 value_set_node_t node;
1276 value_set_node_t next;
1277 node = set->head;
1278 while (node)
1279 {
1280 next = node->next;
1281 if (!valid_in_set (set, node->expr))
1282 set_remove (set, node->expr);
1283 node = next;
1284 }
1285 }
1286
1287 DEF_VEC_P (basic_block);
1288 DEF_VEC_ALLOC_P (basic_block, heap);
1289 static sbitmap has_abnormal_preds;
1290
1291 /* Compute the ANTIC set for BLOCK.
1292
1293 If succs(BLOCK) > 1 then
1294 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
1295 else if succs(BLOCK) == 1 then
1296 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
1297
1298 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
1299
1300 XXX: It would be nice to either write a set_clear, and use it for
1301 ANTIC_OUT, or to mark the antic_out set as deleted at the end
1302 of this routine, so that the pool can hand the same memory back out
1303 again for the next ANTIC_OUT. */
1304
1305 static bool
1306 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
1307 {
1308 basic_block son;
1309 bool changed = false;
1310 value_set_t S, old, ANTIC_OUT;
1311 value_set_node_t node;
1312
1313 ANTIC_OUT = S = NULL;
1314
1315 /* If any edges from predecessors are abnormal, antic_in is empty,
1316 so do nothing. */
1317 if (block_has_abnormal_pred_edge)
1318 goto maybe_dump_sets;
1319
1320 old = set_new (false);
1321 set_copy (old, ANTIC_IN (block));
1322 ANTIC_OUT = set_new (true);
1323
1324 /* If the block has no successors, ANTIC_OUT is empty. */
1325 if (EDGE_COUNT (block->succs) == 0)
1326 ;
1327 /* If we have one successor, we could have some phi nodes to
1328 translate through. */
1329 else if (single_succ_p (block))
1330 {
1331 phi_translate_set (ANTIC_OUT, ANTIC_IN (single_succ (block)),
1332 block, single_succ (block));
1333 }
1334 /* If we have multiple successors, we take the intersection of all of
1335 them. */
1336 else
1337 {
1338 VEC(basic_block, heap) * worklist;
1339 edge e;
1340 size_t i;
1341 basic_block bprime, first;
1342 edge_iterator ei;
1343
1344 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
1345 FOR_EACH_EDGE (e, ei, block->succs)
1346 VEC_quick_push (basic_block, worklist, e->dest);
1347 first = VEC_index (basic_block, worklist, 0);
1348 set_copy (ANTIC_OUT, ANTIC_IN (first));
1349
1350 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
1351 {
1352 node = ANTIC_OUT->head;
1353 while (node)
1354 {
1355 tree val;
1356 value_set_node_t next = node->next;
1357 val = get_value_handle (node->expr);
1358 if (!set_contains_value (ANTIC_IN (bprime), val))
1359 set_remove (ANTIC_OUT, node->expr);
1360 node = next;
1361 }
1362 }
1363 VEC_free (basic_block, heap, worklist);
1364 }
1365
1366 /* Generate ANTIC_OUT - TMP_GEN. */
1367 S = bitmap_set_subtract_from_value_set (ANTIC_OUT, TMP_GEN (block), false);
1368
1369 /* Start ANTIC_IN with EXP_GEN - TMP_GEN */
1370 ANTIC_IN (block) = bitmap_set_subtract_from_value_set (EXP_GEN (block),
1371 TMP_GEN (block),
1372 true);
1373
1374 /* Then union in the ANTIC_OUT - TMP_GEN values,
1375 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
1376 for (node = S->head; node; node = node->next)
1377 value_insert_into_set (ANTIC_IN (block), node->expr);
1378
1379 clean (ANTIC_IN (block));
1380 if (!set_equal (old, ANTIC_IN (block)))
1381 changed = true;
1382
1383 maybe_dump_sets:
1384 if (dump_file && (dump_flags & TDF_DETAILS))
1385 {
1386 if (ANTIC_OUT)
1387 print_value_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
1388 print_value_set (dump_file, ANTIC_IN (block), "ANTIC_IN", block->index);
1389 if (S)
1390 print_value_set (dump_file, S, "S", block->index);
1391 }
1392
1393 for (son = first_dom_son (CDI_POST_DOMINATORS, block);
1394 son;
1395 son = next_dom_son (CDI_POST_DOMINATORS, son))
1396 {
1397 changed |= compute_antic_aux (son,
1398 TEST_BIT (has_abnormal_preds, son->index));
1399 }
1400 return changed;
1401 }
1402
1403 /* Compute ANTIC sets. */
1404
1405 static void
1406 compute_antic (void)
1407 {
1408 bool changed = true;
1409 int num_iterations = 0;
1410 basic_block block;
1411
1412 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
1413 We pre-build the map of blocks with incoming abnormal edges here. */
1414 has_abnormal_preds = sbitmap_alloc (last_basic_block);
1415 sbitmap_zero (has_abnormal_preds);
1416 FOR_EACH_BB (block)
1417 {
1418 edge_iterator ei;
1419 edge e;
1420
1421 FOR_EACH_EDGE (e, ei, block->preds)
1422 if (e->flags & EDGE_ABNORMAL)
1423 {
1424 SET_BIT (has_abnormal_preds, block->index);
1425 break;
1426 }
1427
1428 /* While we are here, give empty ANTIC_IN sets to each block. */
1429 ANTIC_IN (block) = set_new (true);
1430 }
1431 /* At the exit block we anticipate nothing. */
1432 ANTIC_IN (EXIT_BLOCK_PTR) = set_new (true);
1433
1434 while (changed)
1435 {
1436 num_iterations++;
1437 changed = false;
1438 changed = compute_antic_aux (EXIT_BLOCK_PTR, false);
1439 }
1440
1441 sbitmap_free (has_abnormal_preds);
1442
1443 if (dump_file && (dump_flags & TDF_STATS))
1444 fprintf (dump_file, "compute_antic required %d iterations\n", num_iterations);
1445 }
1446
1447 static VEC(tree,heap) *inserted_exprs;
1448 /* Find a leader for an expression, or generate one using
1449 create_expression_by_pieces if it's ANTIC but
1450 complex.
1451 BLOCK is the basic_block we are looking for leaders in.
1452 EXPR is the expression to find a leader or generate for.
1453 STMTS is the statement list to put the inserted expressions on.
1454 Returns the SSA_NAME of the LHS of the generated expression or the
1455 leader. */
1456
1457 static tree
1458 find_or_generate_expression (basic_block block, tree expr, tree stmts)
1459 {
1460 tree genop = bitmap_find_leader (AVAIL_OUT (block), expr);
1461
1462 /* If it's still NULL, it must be a complex expression, so generate
1463 it recursively. */
1464 if (genop == NULL)
1465 {
1466 genop = VALUE_HANDLE_EXPR_SET (expr)->head->expr;
1467 gcc_assert (UNARY_CLASS_P (genop)
1468 || BINARY_CLASS_P (genop)
1469 || COMPARISON_CLASS_P (genop)
1470 || REFERENCE_CLASS_P (genop));
1471 genop = create_expression_by_pieces (block, genop, stmts);
1472 }
1473 return genop;
1474 }
1475
1476 #define NECESSARY(stmt) stmt->common.asm_written_flag
1477 /* Create an expression in pieces, so that we can handle very complex
1478 expressions that may be ANTIC, but not necessary GIMPLE.
1479 BLOCK is the basic block the expression will be inserted into,
1480 EXPR is the expression to insert (in value form)
1481 STMTS is a statement list to append the necessary insertions into.
1482
1483 This function will die if we hit some value that shouldn't be
1484 ANTIC but is (IE there is no leader for it, or its components).
1485 This function may also generate expressions that are themselves
1486 partially or fully redundant. Those that are will be either made
1487 fully redundant during the next iteration of insert (for partially
1488 redundant ones), or eliminated by eliminate (for fully redundant
1489 ones). */
1490
1491 static tree
1492 create_expression_by_pieces (basic_block block, tree expr, tree stmts)
1493 {
1494 tree temp, name;
1495 tree folded, forced_stmts, newexpr;
1496 tree v;
1497 tree_stmt_iterator tsi;
1498
1499 switch (TREE_CODE_CLASS (TREE_CODE (expr)))
1500 {
1501 case tcc_expression:
1502 {
1503 tree op0, op2;
1504 tree arglist;
1505 tree genop0, genop2;
1506 tree genarglist;
1507 tree walker, genwalker;
1508
1509 gcc_assert (TREE_CODE (expr) == CALL_EXPR);
1510 genop2 = NULL;
1511
1512 op0 = TREE_OPERAND (expr, 0);
1513 arglist = TREE_OPERAND (expr, 1);
1514 op2 = TREE_OPERAND (expr, 2);
1515
1516 genop0 = find_or_generate_expression (block, op0, stmts);
1517 genarglist = copy_list (arglist);
1518 for (walker = arglist, genwalker = genarglist;
1519 genwalker && walker;
1520 genwalker = TREE_CHAIN (genwalker), walker = TREE_CHAIN (walker))
1521 {
1522 TREE_VALUE (genwalker) = find_or_generate_expression (block,
1523 TREE_VALUE (walker),
1524 stmts);
1525 }
1526
1527 if (op2)
1528 genop2 = find_or_generate_expression (block, op2, stmts);
1529 folded = fold (build (TREE_CODE (expr), TREE_TYPE (expr),
1530 genop0, genarglist, genop2));
1531 break;
1532
1533
1534 }
1535 break;
1536
1537 case tcc_binary:
1538 case tcc_comparison:
1539 {
1540 tree op1 = TREE_OPERAND (expr, 0);
1541 tree op2 = TREE_OPERAND (expr, 1);
1542 tree genop1 = find_or_generate_expression (block, op1, stmts);
1543 tree genop2 = find_or_generate_expression (block, op2, stmts);
1544 folded = fold (build (TREE_CODE (expr), TREE_TYPE (expr),
1545 genop1, genop2));
1546 break;
1547 }
1548
1549 case tcc_unary:
1550 {
1551 tree op1 = TREE_OPERAND (expr, 0);
1552 tree genop1 = find_or_generate_expression (block, op1, stmts);
1553 folded = fold (build (TREE_CODE (expr), TREE_TYPE (expr),
1554 genop1));
1555 break;
1556 }
1557
1558 default:
1559 gcc_unreachable ();
1560 }
1561
1562 /* Force the generated expression to be a sequence of GIMPLE
1563 statements.
1564 We have to call unshare_expr because force_gimple_operand may
1565 modify the tree we pass to it. */
1566 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
1567 false, NULL);
1568
1569 /* If we have any intermediate expressions to the value sets, add them
1570 to the value sets and chain them on in the instruction stream. */
1571 if (forced_stmts)
1572 {
1573 tsi = tsi_start (forced_stmts);
1574 for (; !tsi_end_p (tsi); tsi_next (&tsi))
1575 {
1576 tree stmt = tsi_stmt (tsi);
1577 tree forcedname = TREE_OPERAND (stmt, 0);
1578 tree forcedexpr = TREE_OPERAND (stmt, 1);
1579 tree val = vn_lookup_or_add (forcedexpr, NULL);
1580
1581 VEC_safe_push (tree, heap, inserted_exprs, stmt);
1582 vn_add (forcedname, val, NULL);
1583 bitmap_value_replace_in_set (NEW_SETS (block), forcedname);
1584 bitmap_value_replace_in_set (AVAIL_OUT (block), forcedname);
1585 }
1586 tsi = tsi_last (stmts);
1587 tsi_link_after (&tsi, forced_stmts, TSI_CONTINUE_LINKING);
1588 }
1589
1590 /* Build and insert the assignment of the end result to the temporary
1591 that we will return. */
1592 temp = create_tmp_var (TREE_TYPE (expr), "pretmp");
1593 add_referenced_tmp_var (temp);
1594 newexpr = build (MODIFY_EXPR, TREE_TYPE (expr), temp, newexpr);
1595 name = make_ssa_name (temp, newexpr);
1596 TREE_OPERAND (newexpr, 0) = name;
1597 NECESSARY (newexpr) = 0;
1598 tsi = tsi_last (stmts);
1599 tsi_link_after (&tsi, newexpr, TSI_CONTINUE_LINKING);
1600 VEC_safe_push (tree, heap, inserted_exprs, newexpr);
1601
1602 /* Add a value handle to the temporary.
1603 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
1604 we are creating the expression by pieces, and this particular piece of
1605 the expression may have been represented. There is no harm in replacing
1606 here. */
1607 v = get_value_handle (expr);
1608 vn_add (name, v, NULL);
1609 bitmap_value_replace_in_set (NEW_SETS (block), name);
1610 bitmap_value_replace_in_set (AVAIL_OUT (block), name);
1611
1612 pre_stats.insertions++;
1613 if (dump_file && (dump_flags & TDF_DETAILS))
1614 {
1615 fprintf (dump_file, "Inserted ");
1616 print_generic_expr (dump_file, newexpr, 0);
1617 fprintf (dump_file, " in predecessor %d\n", block->index);
1618 }
1619
1620 return name;
1621 }
1622
1623 /* Insert the to-be-made-available values of NODE for each predecessor, stored
1624 in AVAIL, into the predecessors of BLOCK, and merge the result with a phi
1625 node, given the same value handle as NODE. The prefix of the phi node is
1626 given with TMPNAME. Return true if we have inserted new stuff. */
1627
1628 static bool
1629 insert_into_preds_of_block (basic_block block, value_set_node_t node,
1630 tree *avail, const char *tmpname)
1631 {
1632 tree val = get_value_handle (node->expr);
1633 edge pred;
1634 bool insertions = false;
1635 bool nophi = false;
1636 basic_block bprime;
1637 tree eprime;
1638 edge_iterator ei;
1639 tree type = TREE_TYPE (avail[EDGE_PRED (block, 0)->src->index]);
1640 tree temp;
1641
1642 if (dump_file && (dump_flags & TDF_DETAILS))
1643 {
1644 fprintf (dump_file, "Found partial redundancy for expression ");
1645 print_generic_expr (dump_file, node->expr, 0);
1646 fprintf (dump_file, "\n");
1647 }
1648
1649 /* Make sure we aren't creating an induction variable. */
1650 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
1651 {
1652 bool firstinsideloop = false;
1653 bool secondinsideloop = false;
1654 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
1655 EDGE_PRED (block, 0)->src);
1656 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
1657 EDGE_PRED (block, 1)->src);
1658 /* Induction variables only have one edge inside the loop. */
1659 if (firstinsideloop ^ secondinsideloop)
1660 {
1661 if (dump_file && (dump_flags & TDF_DETAILS))
1662 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
1663 nophi = true;
1664 }
1665 }
1666
1667
1668 /* Make the necessary insertions. */
1669 FOR_EACH_EDGE (pred, ei, block->preds)
1670 {
1671 tree stmts = alloc_stmt_list ();
1672 tree builtexpr;
1673 bprime = pred->src;
1674 eprime = avail[bprime->index];
1675 if (BINARY_CLASS_P (eprime)
1676 || COMPARISON_CLASS_P (eprime)
1677 || UNARY_CLASS_P (eprime)
1678 || TREE_CODE (eprime) == CALL_EXPR)
1679 {
1680 builtexpr = create_expression_by_pieces (bprime,
1681 eprime,
1682 stmts);
1683 bsi_insert_on_edge (pred, stmts);
1684 avail[bprime->index] = builtexpr;
1685 insertions = true;
1686 }
1687 }
1688 /* If we didn't want a phi node, and we made insertions, we still have
1689 inserted new stuff, and thus return true. If we didn't want a phi node,
1690 and didn't make insertions, we haven't added anything new, so return
1691 false. */
1692 if (nophi && insertions)
1693 return true;
1694 else if (nophi && !insertions)
1695 return false;
1696
1697 /* Now build a phi for the new variable. */
1698 temp = create_tmp_var (type, tmpname);
1699 add_referenced_tmp_var (temp);
1700 temp = create_phi_node (temp, block);
1701 NECESSARY (temp) = 0;
1702 VEC_safe_push (tree, heap, inserted_exprs, temp);
1703 FOR_EACH_EDGE (pred, ei, block->preds)
1704 add_phi_arg (temp, avail[pred->src->index], pred);
1705
1706 vn_add (PHI_RESULT (temp), val, NULL);
1707
1708 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
1709 this insertion, since we test for the existence of this value in PHI_GEN
1710 before proceeding with the partial redundancy checks in insert_aux.
1711
1712 The value may exist in AVAIL_OUT, in particular, it could be represented
1713 by the expression we are trying to eliminate, in which case we want the
1714 replacement to occur. If it's not existing in AVAIL_OUT, we want it
1715 inserted there.
1716
1717 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
1718 this block, because if it did, it would have existed in our dominator's
1719 AVAIL_OUT, and would have been skipped due to the full redundancy check.
1720 */
1721
1722 bitmap_insert_into_set (PHI_GEN (block),
1723 PHI_RESULT (temp));
1724 bitmap_value_replace_in_set (AVAIL_OUT (block),
1725 PHI_RESULT (temp));
1726 bitmap_insert_into_set (NEW_SETS (block),
1727 PHI_RESULT (temp));
1728
1729 if (dump_file && (dump_flags & TDF_DETAILS))
1730 {
1731 fprintf (dump_file, "Created phi ");
1732 print_generic_expr (dump_file, temp, 0);
1733 fprintf (dump_file, " in block %d\n", block->index);
1734 }
1735 pre_stats.phis++;
1736 return true;
1737 }
1738
1739
1740
1741 /* Perform insertion of partially redundant values.
1742 For BLOCK, do the following:
1743 1. Propagate the NEW_SETS of the dominator into the current block.
1744 If the block has multiple predecessors,
1745 2a. Iterate over the ANTIC expressions for the block to see if
1746 any of them are partially redundant.
1747 2b. If so, insert them into the necessary predecessors to make
1748 the expression fully redundant.
1749 2c. Insert a new PHI merging the values of the predecessors.
1750 2d. Insert the new PHI, and the new expressions, into the
1751 NEW_SETS set.
1752 3. Recursively call ourselves on the dominator children of BLOCK.
1753
1754 */
1755
1756 static bool
1757 insert_aux (basic_block block)
1758 {
1759 basic_block son;
1760 bool new_stuff = false;
1761
1762 if (block)
1763 {
1764 basic_block dom;
1765 dom = get_immediate_dominator (CDI_DOMINATORS, block);
1766 if (dom)
1767 {
1768 unsigned i;
1769 bitmap_iterator bi;
1770 bitmap_set_t newset = NEW_SETS (dom);
1771 if (newset)
1772 {
1773 /* Note that we need to value_replace both NEW_SETS, and
1774 AVAIL_OUT. For both the case of NEW_SETS, the value may be
1775 represented by some non-simple expression here that we want
1776 to replace it with. */
1777 EXECUTE_IF_SET_IN_BITMAP (newset->expressions, 0, i, bi)
1778 {
1779 bitmap_value_replace_in_set (NEW_SETS (block), ssa_name (i));
1780 bitmap_value_replace_in_set (AVAIL_OUT (block), ssa_name (i));
1781 }
1782 }
1783 if (!single_pred_p (block))
1784 {
1785 value_set_node_t node;
1786 for (node = ANTIC_IN (block)->head;
1787 node;
1788 node = node->next)
1789 {
1790 if (BINARY_CLASS_P (node->expr)
1791 || COMPARISON_CLASS_P (node->expr)
1792 || UNARY_CLASS_P (node->expr)
1793 || TREE_CODE (node->expr) == CALL_EXPR)
1794 {
1795 tree *avail;
1796 tree val;
1797 bool by_some = false;
1798 bool cant_insert = false;
1799 bool all_same = true;
1800 tree first_s = NULL;
1801 edge pred;
1802 basic_block bprime;
1803 tree eprime = NULL_TREE;
1804 edge_iterator ei;
1805
1806 val = get_value_handle (node->expr);
1807 if (bitmap_set_contains_value (PHI_GEN (block), val))
1808 continue;
1809 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
1810 {
1811 if (dump_file && (dump_flags & TDF_DETAILS))
1812 fprintf (dump_file, "Found fully redundant value\n");
1813 continue;
1814 }
1815
1816 avail = xcalloc (last_basic_block, sizeof (tree));
1817 FOR_EACH_EDGE (pred, ei, block->preds)
1818 {
1819 tree vprime;
1820 tree edoubleprime;
1821
1822 /* This can happen in the very weird case
1823 that our fake infinite loop edges have caused a
1824 critical edge to appear. */
1825 if (EDGE_CRITICAL_P (pred))
1826 {
1827 cant_insert = true;
1828 break;
1829 }
1830 bprime = pred->src;
1831 eprime = phi_translate (node->expr,
1832 ANTIC_IN (block),
1833 bprime, block);
1834
1835 /* eprime will generally only be NULL if the
1836 value of the expression, translated
1837 through the PHI for this predecessor, is
1838 undefined. If that is the case, we can't
1839 make the expression fully redundant,
1840 because its value is undefined along a
1841 predecessor path. We can thus break out
1842 early because it doesn't matter what the
1843 rest of the results are. */
1844 if (eprime == NULL)
1845 {
1846 cant_insert = true;
1847 break;
1848 }
1849
1850 eprime = fully_constant_expression (eprime);
1851 vprime = get_value_handle (eprime);
1852 gcc_assert (vprime);
1853 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
1854 vprime);
1855 if (edoubleprime == NULL)
1856 {
1857 avail[bprime->index] = eprime;
1858 all_same = false;
1859 }
1860 else
1861 {
1862 avail[bprime->index] = edoubleprime;
1863 by_some = true;
1864 if (first_s == NULL)
1865 first_s = edoubleprime;
1866 else if (!operand_equal_p (first_s, edoubleprime,
1867 0))
1868 all_same = false;
1869 }
1870 }
1871 /* If we can insert it, it's not the same value
1872 already existing along every predecessor, and
1873 it's defined by some predecessor, it is
1874 partially redundant. */
1875 if (!cant_insert && !all_same && by_some)
1876 {
1877 if (insert_into_preds_of_block (block, node, avail,
1878 "prephitmp"))
1879 new_stuff = true;
1880 }
1881 /* If all edges produce the same value and that value is
1882 an invariant, then the PHI has the same value on all
1883 edges. Note this. */
1884 else if (!cant_insert && all_same && eprime
1885 && is_gimple_min_invariant (eprime)
1886 && !is_gimple_min_invariant (val))
1887 {
1888 value_set_t exprset = VALUE_HANDLE_EXPR_SET (val);
1889 value_set_node_t node;
1890 for (node = exprset->head; node; node = node->next)
1891 {
1892 if (TREE_CODE (node->expr) == SSA_NAME)
1893 {
1894 vn_add (node->expr, eprime, NULL);
1895 pre_stats.constified++;
1896 }
1897 }
1898 }
1899 free (avail);
1900 }
1901 }
1902 }
1903 }
1904 }
1905 for (son = first_dom_son (CDI_DOMINATORS, block);
1906 son;
1907 son = next_dom_son (CDI_DOMINATORS, son))
1908 {
1909 new_stuff |= insert_aux (son);
1910 }
1911
1912 return new_stuff;
1913 }
1914
1915 /* Perform insertion of partially redundant values. */
1916
1917 static void
1918 insert (void)
1919 {
1920 bool new_stuff = true;
1921 basic_block bb;
1922 int num_iterations = 0;
1923
1924 FOR_ALL_BB (bb)
1925 NEW_SETS (bb) = bitmap_set_new ();
1926
1927 while (new_stuff)
1928 {
1929 num_iterations++;
1930 new_stuff = false;
1931 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
1932 }
1933 if (num_iterations > 2 && dump_file && (dump_flags & TDF_STATS))
1934 fprintf (dump_file, "insert required %d iterations\n", num_iterations);
1935 }
1936
1937
1938 /* Return true if VAR is an SSA variable with no defining statement in
1939 this procedure, *AND* isn't a live-on-entry parameter. */
1940
1941 static bool
1942 is_undefined_value (tree expr)
1943 {
1944 return (TREE_CODE (expr) == SSA_NAME
1945 && IS_EMPTY_STMT (SSA_NAME_DEF_STMT (expr))
1946 /* PARM_DECLs and hard registers are always defined. */
1947 && TREE_CODE (SSA_NAME_VAR (expr)) != PARM_DECL);
1948 }
1949
1950
1951 /* Given an SSA variable VAR and an expression EXPR, compute the value
1952 number for EXPR and create a value handle (VAL) for it. If VAR and
1953 EXPR are not the same, associate VAL with VAR. Finally, add VAR to
1954 S1 and its value handle to S2.
1955
1956 VUSES represent the virtual use operands associated with EXPR (if
1957 any). They are used when computing the hash value for EXPR. */
1958
1959 static inline void
1960 add_to_sets (tree var, tree expr, tree stmt, bitmap_set_t s1,
1961 bitmap_set_t s2)
1962 {
1963 tree val = vn_lookup_or_add (expr, stmt);
1964
1965 /* VAR and EXPR may be the same when processing statements for which
1966 we are not computing value numbers (e.g., non-assignments, or
1967 statements that make aliased stores). In those cases, we are
1968 only interested in making VAR available as its own value. */
1969 if (var != expr)
1970 vn_add (var, val, NULL_TREE);
1971
1972 if (s1)
1973 bitmap_insert_into_set (s1, var);
1974 bitmap_value_insert_into_set (s2, var);
1975 }
1976
1977
1978 /* Given a unary or binary expression EXPR, create and return a new
1979 expression with the same structure as EXPR but with its operands
1980 replaced with the value handles of each of the operands of EXPR.
1981
1982 VUSES represent the virtual use operands associated with EXPR (if
1983 any). They are used when computing the hash value for EXPR.
1984 Insert EXPR's operands into the EXP_GEN set for BLOCK. */
1985
1986 static inline tree
1987 create_value_expr_from (tree expr, basic_block block, tree stmt)
1988 {
1989 int i;
1990 enum tree_code code = TREE_CODE (expr);
1991 tree vexpr;
1992 alloc_pool pool;
1993
1994 gcc_assert (TREE_CODE_CLASS (code) == tcc_unary
1995 || TREE_CODE_CLASS (code) == tcc_binary
1996 || TREE_CODE_CLASS (code) == tcc_comparison
1997 || TREE_CODE_CLASS (code) == tcc_reference
1998 || TREE_CODE_CLASS (code) == tcc_expression
1999 || TREE_CODE_CLASS (code) == tcc_exceptional);
2000
2001 if (TREE_CODE_CLASS (code) == tcc_unary)
2002 pool = unary_node_pool;
2003 else if (TREE_CODE_CLASS (code) == tcc_reference)
2004 pool = reference_node_pool;
2005 else if (TREE_CODE_CLASS (code) == tcc_binary)
2006 pool = binary_node_pool;
2007 else if (TREE_CODE_CLASS (code) == tcc_comparison)
2008 pool = comparison_node_pool;
2009 else if (TREE_CODE_CLASS (code) == tcc_exceptional)
2010 {
2011 gcc_assert (code == TREE_LIST);
2012 pool = list_node_pool;
2013 }
2014 else
2015 {
2016 gcc_assert (code == CALL_EXPR);
2017 pool = expression_node_pool;
2018 }
2019
2020 vexpr = pool_alloc (pool);
2021 memcpy (vexpr, expr, tree_size (expr));
2022
2023 /* This case is only for TREE_LIST's that appear as part of
2024 CALL_EXPR's. Anything else is a bug, but we can't easily verify
2025 this, hence this comment. TREE_LIST is not handled by the
2026 general case below is because they don't have a fixed length, or
2027 operands, so you can't access purpose/value/chain through
2028 TREE_OPERAND macros. */
2029
2030 if (code == TREE_LIST)
2031 {
2032 tree temp = NULL_TREE;
2033 if (TREE_CHAIN (vexpr))
2034 temp = create_value_expr_from (TREE_CHAIN (vexpr), block, stmt);
2035 TREE_CHAIN (vexpr) = temp ? temp : TREE_CHAIN (vexpr);
2036
2037 /* This is the equivalent of inserting op into EXP_GEN like we
2038 do below */
2039 if (!is_undefined_value (TREE_VALUE (vexpr)))
2040 value_insert_into_set (EXP_GEN (block), TREE_VALUE (vexpr));
2041
2042 TREE_VALUE (vexpr) = vn_lookup_or_add (TREE_VALUE (vexpr), NULL);
2043
2044 return vexpr;
2045 }
2046
2047 for (i = 0; i < TREE_CODE_LENGTH (code); i++)
2048 {
2049 tree val, op;
2050
2051 op = TREE_OPERAND (expr, i);
2052 if (op == NULL_TREE)
2053 continue;
2054
2055 /* If OP is a constant that has overflowed, do not value number
2056 this expression. */
2057 if (CONSTANT_CLASS_P (op)
2058 && TREE_OVERFLOW (op))
2059 {
2060 pool_free (pool, vexpr);
2061 return NULL;
2062 }
2063
2064 /* Recursively value-numberize reference ops and tree lists. */
2065 if (REFERENCE_CLASS_P (op))
2066 {
2067 tree tempop = create_value_expr_from (op, block, stmt);
2068 op = tempop ? tempop : op;
2069 val = vn_lookup_or_add (op, stmt);
2070 }
2071 else if (TREE_CODE (op) == TREE_LIST)
2072 {
2073 tree tempop;
2074
2075 gcc_assert (TREE_CODE (expr) == CALL_EXPR);
2076 tempop = create_value_expr_from (op, block, stmt);
2077
2078 op = tempop ? tempop : op;
2079 vn_lookup_or_add (op, NULL);
2080 /* Unlike everywhere else, we do *not* want to replace the
2081 TREE_LIST itself with a value number, because support
2082 functions we call will blow up. */
2083 val = op;
2084 }
2085 else
2086 /* Create a value handle for OP and add it to VEXPR. */
2087 val = vn_lookup_or_add (op, NULL);
2088
2089 if (!is_undefined_value (op) && TREE_CODE (op) != TREE_LIST)
2090 value_insert_into_set (EXP_GEN (block), op);
2091
2092 if (TREE_CODE (val) == VALUE_HANDLE)
2093 TREE_TYPE (val) = TREE_TYPE (TREE_OPERAND (vexpr, i));
2094
2095 TREE_OPERAND (vexpr, i) = val;
2096 }
2097
2098 return vexpr;
2099 }
2100
2101
2102 /* Return true if we can value number a call. This is true if we have
2103 a pure or constant call. */
2104 static bool
2105 can_value_number_call (tree stmt)
2106 {
2107 tree call = get_call_expr_in (stmt);
2108
2109 /* This is a temporary restriction until we translate vuses through
2110 phi nodes. */
2111 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
2112 return false;
2113 if (call_expr_flags (call) & (ECF_PURE | ECF_CONST))
2114 return true;
2115 return false;
2116 }
2117
2118 /* Compute the AVAIL set for all basic blocks.
2119
2120 This function performs value numbering of the statements in each basic
2121 block. The AVAIL sets are built from information we glean while doing
2122 this value numbering, since the AVAIL sets contain only one entry per
2123 value.
2124
2125 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
2126 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
2127
2128 static void
2129 compute_avail (void)
2130 {
2131 basic_block block, son;
2132 basic_block *worklist;
2133 size_t sp = 0;
2134 tree param;
2135
2136 /* For arguments with default definitions, we pretend they are
2137 defined in the entry block. */
2138 for (param = DECL_ARGUMENTS (current_function_decl);
2139 param;
2140 param = TREE_CHAIN (param))
2141 {
2142 if (default_def (param) != NULL)
2143 {
2144 tree def = default_def (param);
2145 vn_lookup_or_add (def, NULL);
2146 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), def);
2147 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), def);
2148 }
2149 }
2150
2151 /* Allocate the worklist. */
2152 worklist = xmalloc (sizeof (basic_block) * n_basic_blocks);
2153
2154 /* Seed the algorithm by putting the dominator children of the entry
2155 block on the worklist. */
2156 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
2157 son;
2158 son = next_dom_son (CDI_DOMINATORS, son))
2159 worklist[sp++] = son;
2160
2161 /* Loop until the worklist is empty. */
2162 while (sp)
2163 {
2164 block_stmt_iterator bsi;
2165 tree stmt, phi;
2166 basic_block dom;
2167
2168 /* Pick a block from the worklist. */
2169 block = worklist[--sp];
2170
2171 /* Initially, the set of available values in BLOCK is that of
2172 its immediate dominator. */
2173 dom = get_immediate_dominator (CDI_DOMINATORS, block);
2174 if (dom)
2175 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
2176
2177 /* Generate values for PHI nodes. */
2178 for (phi = phi_nodes (block); phi; phi = PHI_CHAIN (phi))
2179 /* We have no need for virtual phis, as they don't represent
2180 actual computations. */
2181 if (is_gimple_reg (PHI_RESULT (phi)))
2182 add_to_sets (PHI_RESULT (phi), PHI_RESULT (phi), NULL,
2183 PHI_GEN (block), AVAIL_OUT (block));
2184
2185 /* Now compute value numbers and populate value sets with all
2186 the expressions computed in BLOCK. */
2187 for (bsi = bsi_start (block); !bsi_end_p (bsi); bsi_next (&bsi))
2188 {
2189 stmt_ann_t ann;
2190 ssa_op_iter iter;
2191 tree op;
2192
2193 stmt = bsi_stmt (bsi);
2194 ann = stmt_ann (stmt);
2195
2196 /* We are only interested in assignments of the form
2197 X_i = EXPR, where EXPR represents an "interesting"
2198 computation, it has no volatile operands and X_i
2199 doesn't flow through an abnormal edge. */
2200 if (TREE_CODE (stmt) == MODIFY_EXPR
2201 && !ann->has_volatile_ops
2202 && TREE_CODE (TREE_OPERAND (stmt, 0)) == SSA_NAME
2203 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (stmt, 0)))
2204 {
2205 tree lhs = TREE_OPERAND (stmt, 0);
2206 tree rhs = TREE_OPERAND (stmt, 1);
2207
2208 STRIP_USELESS_TYPE_CONVERSION (rhs);
2209 if (UNARY_CLASS_P (rhs)
2210 || BINARY_CLASS_P (rhs)
2211 || COMPARISON_CLASS_P (rhs)
2212 || REFERENCE_CLASS_P (rhs)
2213 || (TREE_CODE (rhs) == CALL_EXPR
2214 && can_value_number_call (stmt)))
2215 {
2216 /* For binary, unary, and reference expressions,
2217 create a duplicate expression with the operands
2218 replaced with the value handles of the original
2219 RHS. */
2220 tree newt = create_value_expr_from (rhs, block, stmt);
2221 if (newt)
2222 {
2223 add_to_sets (lhs, newt, stmt, TMP_GEN (block),
2224 AVAIL_OUT (block));
2225 value_insert_into_set (EXP_GEN (block), newt);
2226 continue;
2227 }
2228 }
2229 else if (TREE_CODE (rhs) == SSA_NAME
2230 || is_gimple_min_invariant (rhs)
2231 || TREE_CODE (rhs) == ADDR_EXPR
2232 || TREE_INVARIANT (rhs)
2233 || DECL_P (rhs))
2234 {
2235 /* Compute a value number for the RHS of the statement
2236 and add its value to the AVAIL_OUT set for the block.
2237 Add the LHS to TMP_GEN. */
2238 add_to_sets (lhs, rhs, stmt, TMP_GEN (block),
2239 AVAIL_OUT (block));
2240
2241 if (TREE_CODE (rhs) == SSA_NAME
2242 && !is_undefined_value (rhs))
2243 value_insert_into_set (EXP_GEN (block), rhs);
2244 continue;
2245 }
2246 }
2247
2248 /* For any other statement that we don't recognize, simply
2249 make the names generated by the statement available in
2250 AVAIL_OUT and TMP_GEN. */
2251 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
2252 add_to_sets (op, op, NULL, TMP_GEN (block), AVAIL_OUT (block));
2253
2254 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
2255 add_to_sets (op, op, NULL, NULL , AVAIL_OUT (block));
2256 }
2257
2258 /* Put the dominator children of BLOCK on the worklist of blocks
2259 to compute available sets for. */
2260 for (son = first_dom_son (CDI_DOMINATORS, block);
2261 son;
2262 son = next_dom_son (CDI_DOMINATORS, son))
2263 worklist[sp++] = son;
2264 }
2265
2266 free (worklist);
2267 }
2268
2269
2270 /* Eliminate fully redundant computations. */
2271
2272 static void
2273 eliminate (void)
2274 {
2275 basic_block b;
2276
2277 FOR_EACH_BB (b)
2278 {
2279 block_stmt_iterator i;
2280
2281 for (i = bsi_start (b); !bsi_end_p (i); bsi_next (&i))
2282 {
2283 tree stmt = bsi_stmt (i);
2284
2285 /* Lookup the RHS of the expression, see if we have an
2286 available computation for it. If so, replace the RHS with
2287 the available computation. */
2288 if (TREE_CODE (stmt) == MODIFY_EXPR
2289 && TREE_CODE (TREE_OPERAND (stmt, 0)) == SSA_NAME
2290 && TREE_CODE (TREE_OPERAND (stmt ,1)) != SSA_NAME
2291 && !is_gimple_min_invariant (TREE_OPERAND (stmt, 1))
2292 && !stmt_ann (stmt)->has_volatile_ops)
2293 {
2294 tree lhs = TREE_OPERAND (stmt, 0);
2295 tree *rhs_p = &TREE_OPERAND (stmt, 1);
2296 tree sprime;
2297
2298 sprime = bitmap_find_leader (AVAIL_OUT (b),
2299 vn_lookup (lhs, NULL));
2300 if (sprime
2301 && sprime != lhs
2302 && (TREE_CODE (*rhs_p) != SSA_NAME
2303 || may_propagate_copy (*rhs_p, sprime)))
2304 {
2305 gcc_assert (sprime != *rhs_p);
2306
2307 if (dump_file && (dump_flags & TDF_DETAILS))
2308 {
2309 fprintf (dump_file, "Replaced ");
2310 print_generic_expr (dump_file, *rhs_p, 0);
2311 fprintf (dump_file, " with ");
2312 print_generic_expr (dump_file, sprime, 0);
2313 fprintf (dump_file, " in ");
2314 print_generic_stmt (dump_file, stmt, 0);
2315 }
2316 if (TREE_CODE (sprime) == SSA_NAME)
2317 NECESSARY (SSA_NAME_DEF_STMT (sprime)) = 1;
2318 pre_stats.eliminations++;
2319 propagate_tree_value (rhs_p, sprime);
2320 update_stmt (stmt);
2321
2322 /* If we removed EH side effects from the statement, clean
2323 its EH information. */
2324 if (maybe_clean_eh_stmt (stmt))
2325 {
2326 bitmap_set_bit (need_eh_cleanup,
2327 bb_for_stmt (stmt)->index);
2328 if (dump_file && (dump_flags & TDF_DETAILS))
2329 fprintf (dump_file, " Removed EH side effects.\n");
2330 }
2331 }
2332 }
2333 }
2334 }
2335 }
2336
2337 /* Borrow a bit of tree-ssa-dce.c for the moment.
2338 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
2339 this may be a bit faster, and we may want critical edges kept split. */
2340
2341 /* If OP's defining statement has not already been determined to be necessary,
2342 mark that statement necessary. Return the stmt, if it is newly
2343 necessary. */
2344
2345 static inline tree
2346 mark_operand_necessary (tree op)
2347 {
2348 tree stmt;
2349
2350 gcc_assert (op);
2351
2352 stmt = SSA_NAME_DEF_STMT (op);
2353 gcc_assert (stmt);
2354
2355 if (NECESSARY (stmt)
2356 || IS_EMPTY_STMT (stmt))
2357 return NULL;
2358
2359 NECESSARY (stmt) = 1;
2360 return stmt;
2361 }
2362
2363 /* Because we don't follow exactly the standard PRE algorithm, and decide not
2364 to insert PHI nodes sometimes, and because value numbering of casts isn't
2365 perfect, we sometimes end up inserting dead code. This simple DCE-like
2366 pass removes any insertions we made that weren't actually used. */
2367
2368 static void
2369 remove_dead_inserted_code (void)
2370 {
2371 VEC(tree,heap) *worklist = NULL;
2372 int i;
2373 tree t;
2374
2375 worklist = VEC_alloc (tree, heap, VEC_length (tree, inserted_exprs));
2376 for (i = 0; VEC_iterate (tree, inserted_exprs, i, t); i++)
2377 {
2378 if (NECESSARY (t))
2379 VEC_quick_push (tree, worklist, t);
2380 }
2381 while (VEC_length (tree, worklist) > 0)
2382 {
2383 t = VEC_pop (tree, worklist);
2384 if (TREE_CODE (t) == PHI_NODE)
2385 {
2386 /* PHI nodes are somewhat special in that each PHI alternative has
2387 data and control dependencies. All the statements feeding the
2388 PHI node's arguments are always necessary. In aggressive mode,
2389 we also consider the control dependent edges leading to the
2390 predecessor block associated with each PHI alternative as
2391 necessary. */
2392 int k;
2393
2394 VEC_reserve (tree, heap, worklist, PHI_NUM_ARGS (t));
2395 for (k = 0; k < PHI_NUM_ARGS (t); k++)
2396 {
2397 tree arg = PHI_ARG_DEF (t, k);
2398 if (TREE_CODE (arg) == SSA_NAME)
2399 {
2400 arg = mark_operand_necessary (arg);
2401 if (arg)
2402 VEC_quick_push (tree, worklist, arg);
2403 }
2404 }
2405 }
2406 else
2407 {
2408 /* Propagate through the operands. Examine all the USE, VUSE and
2409 V_MAY_DEF operands in this statement. Mark all the statements
2410 which feed this statement's uses as necessary. */
2411 ssa_op_iter iter;
2412 tree use;
2413
2414 /* The operands of V_MAY_DEF expressions are also needed as they
2415 represent potential definitions that may reach this
2416 statement (V_MAY_DEF operands allow us to follow def-def
2417 links). */
2418
2419 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
2420 {
2421 tree n = mark_operand_necessary (use);
2422 if (n)
2423 VEC_safe_push (tree, heap, worklist, n);
2424 }
2425 }
2426 }
2427 for (i = 0; VEC_iterate (tree, inserted_exprs, i, t); i++)
2428 {
2429 if (!NECESSARY (t))
2430 {
2431 block_stmt_iterator bsi;
2432 if (dump_file && (dump_flags & TDF_DETAILS))
2433 {
2434 fprintf (dump_file, "Removing unnecessary insertion:");
2435 print_generic_stmt (dump_file, t, 0);
2436 }
2437 if (TREE_CODE (t) == PHI_NODE)
2438 {
2439 remove_phi_node (t, NULL);
2440 }
2441 else
2442 {
2443 bsi = bsi_for_stmt (t);
2444 bsi_remove (&bsi);
2445 }
2446 }
2447 }
2448 VEC_free (tree, heap, worklist);
2449 }
2450 /* Initialize data structures used by PRE. */
2451
2452 static void
2453 init_pre (bool do_fre)
2454 {
2455 basic_block bb;
2456
2457 inserted_exprs = NULL;
2458 vn_init ();
2459 if (!do_fre)
2460 current_loops = loop_optimizer_init (dump_file);
2461 connect_infinite_loops_to_exit ();
2462 memset (&pre_stats, 0, sizeof (pre_stats));
2463
2464 /* If block 0 has more than one predecessor, it means that its PHI
2465 nodes will have arguments coming from block -1. This creates
2466 problems for several places in PRE that keep local arrays indexed
2467 by block number. To prevent this, we split the edge coming from
2468 ENTRY_BLOCK_PTR (FIXME, if ENTRY_BLOCK_PTR had an index number
2469 different than -1 we wouldn't have to hack this. tree-ssa-dce.c
2470 needs a similar change). */
2471 if (!single_pred_p (single_succ (ENTRY_BLOCK_PTR)))
2472 if (!(single_succ_edge (ENTRY_BLOCK_PTR)->flags & EDGE_ABNORMAL))
2473 split_edge (single_succ_edge (ENTRY_BLOCK_PTR));
2474
2475 FOR_ALL_BB (bb)
2476 bb->aux = xcalloc (1, sizeof (struct bb_value_sets));
2477
2478 bitmap_obstack_initialize (&grand_bitmap_obstack);
2479 phi_translate_table = htab_create (511, expr_pred_trans_hash,
2480 expr_pred_trans_eq, free);
2481 value_set_pool = create_alloc_pool ("Value sets",
2482 sizeof (struct value_set), 30);
2483 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
2484 sizeof (struct bitmap_set), 30);
2485 value_set_node_pool = create_alloc_pool ("Value set nodes",
2486 sizeof (struct value_set_node), 30);
2487 calculate_dominance_info (CDI_POST_DOMINATORS);
2488 calculate_dominance_info (CDI_DOMINATORS);
2489 binary_node_pool = create_alloc_pool ("Binary tree nodes",
2490 tree_code_size (PLUS_EXPR), 30);
2491 unary_node_pool = create_alloc_pool ("Unary tree nodes",
2492 tree_code_size (NEGATE_EXPR), 30);
2493 reference_node_pool = create_alloc_pool ("Reference tree nodes",
2494 tree_code_size (ARRAY_REF), 30);
2495 expression_node_pool = create_alloc_pool ("Expression tree nodes",
2496 tree_code_size (CALL_EXPR), 30);
2497 list_node_pool = create_alloc_pool ("List tree nodes",
2498 tree_code_size (TREE_LIST), 30);
2499 comparison_node_pool = create_alloc_pool ("Comparison tree nodes",
2500 tree_code_size (EQ_EXPR), 30);
2501 FOR_ALL_BB (bb)
2502 {
2503 EXP_GEN (bb) = set_new (true);
2504 PHI_GEN (bb) = bitmap_set_new ();
2505 TMP_GEN (bb) = bitmap_set_new ();
2506 AVAIL_OUT (bb) = bitmap_set_new ();
2507 }
2508
2509 need_eh_cleanup = BITMAP_ALLOC (NULL);
2510 }
2511
2512
2513 /* Deallocate data structures used by PRE. */
2514
2515 static void
2516 fini_pre (bool do_fre)
2517 {
2518 basic_block bb;
2519 unsigned int i;
2520
2521 VEC_free (tree, heap, inserted_exprs);
2522 bitmap_obstack_release (&grand_bitmap_obstack);
2523 free_alloc_pool (value_set_pool);
2524 free_alloc_pool (bitmap_set_pool);
2525 free_alloc_pool (value_set_node_pool);
2526 free_alloc_pool (binary_node_pool);
2527 free_alloc_pool (reference_node_pool);
2528 free_alloc_pool (unary_node_pool);
2529 free_alloc_pool (list_node_pool);
2530 free_alloc_pool (expression_node_pool);
2531 free_alloc_pool (comparison_node_pool);
2532 htab_delete (phi_translate_table);
2533 remove_fake_exit_edges ();
2534
2535 FOR_ALL_BB (bb)
2536 {
2537 free (bb->aux);
2538 bb->aux = NULL;
2539 }
2540
2541 free_dominance_info (CDI_POST_DOMINATORS);
2542 vn_delete ();
2543
2544 if (!bitmap_empty_p (need_eh_cleanup))
2545 {
2546 tree_purge_all_dead_eh_edges (need_eh_cleanup);
2547 cleanup_tree_cfg ();
2548 }
2549
2550 BITMAP_FREE (need_eh_cleanup);
2551
2552 /* Wipe out pointers to VALUE_HANDLEs. In the not terribly distant
2553 future we will want them to be persistent though. */
2554 for (i = 0; i < num_ssa_names; i++)
2555 {
2556 tree name = ssa_name (i);
2557
2558 if (!name)
2559 continue;
2560
2561 if (SSA_NAME_VALUE (name)
2562 && TREE_CODE (SSA_NAME_VALUE (name)) == VALUE_HANDLE)
2563 SSA_NAME_VALUE (name) = NULL;
2564 }
2565 if (!do_fre && current_loops)
2566 {
2567 loop_optimizer_finalize (current_loops, dump_file);
2568 current_loops = NULL;
2569 }
2570 }
2571
2572
2573 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
2574 only wants to do full redundancy elimination. */
2575
2576 static void
2577 execute_pre (bool do_fre)
2578 {
2579 init_pre (do_fre);
2580
2581 /* Collect and value number expressions computed in each basic block. */
2582 compute_avail ();
2583
2584 if (dump_file && (dump_flags & TDF_DETAILS))
2585 {
2586 basic_block bb;
2587
2588 FOR_ALL_BB (bb)
2589 {
2590 print_value_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
2591 bitmap_print_value_set (dump_file, TMP_GEN (bb), "tmp_gen",
2592 bb->index);
2593 bitmap_print_value_set (dump_file, AVAIL_OUT (bb), "avail_out",
2594 bb->index);
2595 }
2596 }
2597
2598 /* Insert can get quite slow on an incredibly large number of basic
2599 blocks due to some quadratic behavior. Until this behavior is
2600 fixed, don't run it when he have an incredibly large number of
2601 bb's. If we aren't going to run insert, there is no point in
2602 computing ANTIC, either, even though it's plenty fast. */
2603 if (!do_fre && n_basic_blocks < 4000)
2604 {
2605 compute_antic ();
2606 insert ();
2607 }
2608
2609 /* Remove all the redundant expressions. */
2610 eliminate ();
2611
2612
2613 if (dump_file && (dump_flags & TDF_STATS))
2614 {
2615 fprintf (dump_file, "Insertions: %d\n", pre_stats.insertions);
2616 fprintf (dump_file, "New PHIs: %d\n", pre_stats.phis);
2617 fprintf (dump_file, "Eliminated: %d\n", pre_stats.eliminations);
2618 fprintf (dump_file, "Constified: %d\n", pre_stats.constified);
2619 }
2620
2621 bsi_commit_edge_inserts ();
2622 if (!do_fre)
2623 remove_dead_inserted_code ();
2624 fini_pre (do_fre);
2625
2626 }
2627
2628
2629 /* Gate and execute functions for PRE. */
2630
2631 static void
2632 do_pre (void)
2633 {
2634 execute_pre (false);
2635 }
2636
2637 static bool
2638 gate_pre (void)
2639 {
2640 return flag_tree_pre != 0;
2641 }
2642
2643 struct tree_opt_pass pass_pre =
2644 {
2645 "pre", /* name */
2646 gate_pre, /* gate */
2647 do_pre, /* execute */
2648 NULL, /* sub */
2649 NULL, /* next */
2650 0, /* static_pass_number */
2651 TV_TREE_PRE, /* tv_id */
2652 PROP_no_crit_edges | PROP_cfg
2653 | PROP_ssa | PROP_alias, /* properties_required */
2654 0, /* properties_provided */
2655 0, /* properties_destroyed */
2656 0, /* todo_flags_start */
2657 TODO_update_ssa | TODO_dump_func | TODO_ggc_collect
2658 | TODO_verify_ssa, /* todo_flags_finish */
2659 0 /* letter */
2660 };
2661
2662
2663 /* Gate and execute functions for FRE. */
2664
2665 static void
2666 execute_fre (void)
2667 {
2668 execute_pre (true);
2669 }
2670
2671 static bool
2672 gate_fre (void)
2673 {
2674 return flag_tree_fre != 0;
2675 }
2676
2677 struct tree_opt_pass pass_fre =
2678 {
2679 "fre", /* name */
2680 gate_fre, /* gate */
2681 execute_fre, /* execute */
2682 NULL, /* sub */
2683 NULL, /* next */
2684 0, /* static_pass_number */
2685 TV_TREE_FRE, /* tv_id */
2686 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
2687 0, /* properties_provided */
2688 0, /* properties_destroyed */
2689 0, /* todo_flags_start */
2690 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa, /* todo_flags_finish */
2691 0 /* letter */
2692 };